]>
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> |
138ae663 | 106 | #include <linux/uaccess.h> |
e498be7d | 107 | #include <linux/nodemask.h> |
dc85da15 | 108 | #include <linux/mempolicy.h> |
fc0abb14 | 109 | #include <linux/mutex.h> |
8a8b6502 | 110 | #include <linux/fault-inject.h> |
e7eebaf6 | 111 | #include <linux/rtmutex.h> |
1da177e4 | 112 | |
1da177e4 LT |
113 | #include <asm/cacheflush.h> |
114 | #include <asm/tlbflush.h> | |
115 | #include <asm/page.h> | |
116 | ||
117 | /* | |
118 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
119 | * SLAB_RED_ZONE & SLAB_POISON. | |
120 | * 0 for faster, smaller code (especially in the critical paths). | |
121 | * | |
122 | * STATS - 1 to collect stats for /proc/slabinfo. | |
123 | * 0 for faster, smaller code (especially in the critical paths). | |
124 | * | |
125 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
126 | */ | |
127 | ||
128 | #ifdef CONFIG_DEBUG_SLAB | |
129 | #define DEBUG 1 | |
130 | #define STATS 1 | |
131 | #define FORCED_DEBUG 1 | |
132 | #else | |
133 | #define DEBUG 0 | |
134 | #define STATS 0 | |
135 | #define FORCED_DEBUG 0 | |
136 | #endif | |
137 | ||
1da177e4 LT |
138 | /* Shouldn't this be in a header file somewhere? */ |
139 | #define BYTES_PER_WORD sizeof(void *) | |
140 | ||
141 | #ifndef cache_line_size | |
142 | #define cache_line_size() L1_CACHE_BYTES | |
143 | #endif | |
144 | ||
145 | #ifndef ARCH_KMALLOC_MINALIGN | |
146 | /* | |
147 | * Enforce a minimum alignment for the kmalloc caches. | |
148 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
149 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
150 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
151 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
152 | * Note that this flag disables some debug features. | |
153 | */ | |
154 | #define ARCH_KMALLOC_MINALIGN 0 | |
155 | #endif | |
156 | ||
157 | #ifndef ARCH_SLAB_MINALIGN | |
158 | /* | |
159 | * Enforce a minimum alignment for all caches. | |
160 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
161 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
162 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
163 | * some debug features. | |
164 | */ | |
165 | #define ARCH_SLAB_MINALIGN 0 | |
166 | #endif | |
167 | ||
168 | #ifndef ARCH_KMALLOC_FLAGS | |
169 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
170 | #endif | |
171 | ||
172 | /* Legal flag mask for kmem_cache_create(). */ | |
173 | #if DEBUG | |
174 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
175 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
ac2b898c | 176 | SLAB_CACHE_DMA | \ |
1da177e4 LT |
177 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ |
178 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 179 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 180 | #else |
ac2b898c | 181 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
1da177e4 LT |
182 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ |
183 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 184 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 LT |
185 | #endif |
186 | ||
187 | /* | |
188 | * kmem_bufctl_t: | |
189 | * | |
190 | * Bufctl's are used for linking objs within a slab | |
191 | * linked offsets. | |
192 | * | |
193 | * This implementation relies on "struct page" for locating the cache & | |
194 | * slab an object belongs to. | |
195 | * This allows the bufctl structure to be small (one int), but limits | |
196 | * the number of objects a slab (not a cache) can contain when off-slab | |
197 | * bufctls are used. The limit is the size of the largest general cache | |
198 | * that does not use off-slab slabs. | |
199 | * For 32bit archs with 4 kB pages, is this 56. | |
200 | * This is not serious, as it is only for large objects, when it is unwise | |
201 | * to have too many per slab. | |
202 | * Note: This limit can be raised by introducing a general cache whose size | |
203 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
204 | */ | |
205 | ||
fa5b08d5 | 206 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
207 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
208 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
209 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
210 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 211 | |
1da177e4 LT |
212 | /* |
213 | * struct slab | |
214 | * | |
215 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
216 | * for a slab, or allocated from an general cache. | |
217 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
218 | */ | |
219 | struct slab { | |
b28a02de PE |
220 | struct list_head list; |
221 | unsigned long colouroff; | |
222 | void *s_mem; /* including colour offset */ | |
223 | unsigned int inuse; /* num of objs active in slab */ | |
224 | kmem_bufctl_t free; | |
225 | unsigned short nodeid; | |
1da177e4 LT |
226 | }; |
227 | ||
228 | /* | |
229 | * struct slab_rcu | |
230 | * | |
231 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
232 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
233 | * we need to approach a kernel structure obliquely, from its address | |
234 | * obtained without the usual locking. We can lock the structure to | |
235 | * stabilize it and check it's still at the given address, only if we | |
236 | * can be sure that the memory has not been meanwhile reused for some | |
237 | * other kind of object (which our subsystem's lock might corrupt). | |
238 | * | |
239 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
240 | * taking the spinlock within the structure expected at that address. | |
241 | * | |
242 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
243 | */ | |
244 | struct slab_rcu { | |
b28a02de | 245 | struct rcu_head head; |
343e0d7a | 246 | struct kmem_cache *cachep; |
b28a02de | 247 | void *addr; |
1da177e4 LT |
248 | }; |
249 | ||
250 | /* | |
251 | * struct array_cache | |
252 | * | |
1da177e4 LT |
253 | * Purpose: |
254 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
255 | * - reduce the number of linked list operations | |
256 | * - reduce spinlock operations | |
257 | * | |
258 | * The limit is stored in the per-cpu structure to reduce the data cache | |
259 | * footprint. | |
260 | * | |
261 | */ | |
262 | struct array_cache { | |
263 | unsigned int avail; | |
264 | unsigned int limit; | |
265 | unsigned int batchcount; | |
266 | unsigned int touched; | |
e498be7d | 267 | spinlock_t lock; |
a737b3e2 AM |
268 | void *entry[0]; /* |
269 | * Must have this definition in here for the proper | |
270 | * alignment of array_cache. Also simplifies accessing | |
271 | * the entries. | |
272 | * [0] is for gcc 2.95. It should really be []. | |
273 | */ | |
1da177e4 LT |
274 | }; |
275 | ||
a737b3e2 AM |
276 | /* |
277 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
278 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
279 | */ |
280 | #define BOOT_CPUCACHE_ENTRIES 1 | |
281 | struct arraycache_init { | |
282 | struct array_cache cache; | |
b28a02de | 283 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
284 | }; |
285 | ||
286 | /* | |
e498be7d | 287 | * The slab lists for all objects. |
1da177e4 LT |
288 | */ |
289 | struct kmem_list3 { | |
b28a02de PE |
290 | struct list_head slabs_partial; /* partial list first, better asm code */ |
291 | struct list_head slabs_full; | |
292 | struct list_head slabs_free; | |
293 | unsigned long free_objects; | |
b28a02de | 294 | unsigned int free_limit; |
2e1217cf | 295 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
296 | spinlock_t list_lock; |
297 | struct array_cache *shared; /* shared per node */ | |
298 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
299 | unsigned long next_reap; /* updated without locking */ |
300 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
301 | }; |
302 | ||
e498be7d CL |
303 | /* |
304 | * Need this for bootstrapping a per node allocator. | |
305 | */ | |
306 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
307 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
308 | #define CACHE_CACHE 0 | |
309 | #define SIZE_AC 1 | |
310 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
311 | ||
ed11d9eb CL |
312 | static int drain_freelist(struct kmem_cache *cache, |
313 | struct kmem_list3 *l3, int tofree); | |
314 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
315 | int node); | |
2ed3a4ef | 316 | static int enable_cpucache(struct kmem_cache *cachep); |
65f27f38 | 317 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 318 | |
e498be7d | 319 | /* |
a737b3e2 AM |
320 | * This function must be completely optimized away if a constant is passed to |
321 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 322 | */ |
7243cc05 | 323 | static __always_inline int index_of(const size_t size) |
e498be7d | 324 | { |
5ec8a847 SR |
325 | extern void __bad_size(void); |
326 | ||
e498be7d CL |
327 | if (__builtin_constant_p(size)) { |
328 | int i = 0; | |
329 | ||
330 | #define CACHE(x) \ | |
331 | if (size <=x) \ | |
332 | return i; \ | |
333 | else \ | |
334 | i++; | |
335 | #include "linux/kmalloc_sizes.h" | |
336 | #undef CACHE | |
5ec8a847 | 337 | __bad_size(); |
7243cc05 | 338 | } else |
5ec8a847 | 339 | __bad_size(); |
e498be7d CL |
340 | return 0; |
341 | } | |
342 | ||
e0a42726 IM |
343 | static int slab_early_init = 1; |
344 | ||
e498be7d CL |
345 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
346 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 347 | |
5295a74c | 348 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
349 | { |
350 | INIT_LIST_HEAD(&parent->slabs_full); | |
351 | INIT_LIST_HEAD(&parent->slabs_partial); | |
352 | INIT_LIST_HEAD(&parent->slabs_free); | |
353 | parent->shared = NULL; | |
354 | parent->alien = NULL; | |
2e1217cf | 355 | parent->colour_next = 0; |
e498be7d CL |
356 | spin_lock_init(&parent->list_lock); |
357 | parent->free_objects = 0; | |
358 | parent->free_touched = 0; | |
359 | } | |
360 | ||
a737b3e2 AM |
361 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
362 | do { \ | |
363 | INIT_LIST_HEAD(listp); \ | |
364 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
365 | } while (0) |
366 | ||
a737b3e2 AM |
367 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
368 | do { \ | |
e498be7d CL |
369 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
370 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
371 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
372 | } while (0) | |
1da177e4 LT |
373 | |
374 | /* | |
343e0d7a | 375 | * struct kmem_cache |
1da177e4 LT |
376 | * |
377 | * manages a cache. | |
378 | */ | |
b28a02de | 379 | |
2109a2d1 | 380 | struct kmem_cache { |
1da177e4 | 381 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 382 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 383 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
384 | unsigned int batchcount; |
385 | unsigned int limit; | |
386 | unsigned int shared; | |
b5d8ca7c | 387 | |
3dafccf2 | 388 | unsigned int buffer_size; |
b5d8ca7c | 389 | /* 3) touched by every alloc & free from the backend */ |
b28a02de | 390 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
b5d8ca7c | 391 | |
a737b3e2 AM |
392 | unsigned int flags; /* constant flags */ |
393 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 394 | |
b5d8ca7c | 395 | /* 4) cache_grow/shrink */ |
1da177e4 | 396 | /* order of pgs per slab (2^n) */ |
b28a02de | 397 | unsigned int gfporder; |
1da177e4 LT |
398 | |
399 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 400 | gfp_t gfpflags; |
1da177e4 | 401 | |
a737b3e2 | 402 | size_t colour; /* cache colouring range */ |
b28a02de | 403 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 404 | struct kmem_cache *slabp_cache; |
b28a02de | 405 | unsigned int slab_size; |
a737b3e2 | 406 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
407 | |
408 | /* constructor func */ | |
343e0d7a | 409 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
410 | |
411 | /* de-constructor func */ | |
343e0d7a | 412 | void (*dtor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 | 413 | |
b5d8ca7c | 414 | /* 5) cache creation/removal */ |
b28a02de PE |
415 | const char *name; |
416 | struct list_head next; | |
1da177e4 | 417 | |
b5d8ca7c | 418 | /* 6) statistics */ |
1da177e4 | 419 | #if STATS |
b28a02de PE |
420 | unsigned long num_active; |
421 | unsigned long num_allocations; | |
422 | unsigned long high_mark; | |
423 | unsigned long grown; | |
424 | unsigned long reaped; | |
425 | unsigned long errors; | |
426 | unsigned long max_freeable; | |
427 | unsigned long node_allocs; | |
428 | unsigned long node_frees; | |
fb7faf33 | 429 | unsigned long node_overflow; |
b28a02de PE |
430 | atomic_t allochit; |
431 | atomic_t allocmiss; | |
432 | atomic_t freehit; | |
433 | atomic_t freemiss; | |
1da177e4 LT |
434 | #endif |
435 | #if DEBUG | |
3dafccf2 MS |
436 | /* |
437 | * If debugging is enabled, then the allocator can add additional | |
438 | * fields and/or padding to every object. buffer_size contains the total | |
439 | * object size including these internal fields, the following two | |
440 | * variables contain the offset to the user object and its size. | |
441 | */ | |
442 | int obj_offset; | |
443 | int obj_size; | |
1da177e4 LT |
444 | #endif |
445 | }; | |
446 | ||
447 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
448 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
449 | ||
450 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
451 | /* |
452 | * Optimization question: fewer reaps means less probability for unnessary | |
453 | * cpucache drain/refill cycles. | |
1da177e4 | 454 | * |
dc6f3f27 | 455 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
456 | * which could lock up otherwise freeable slabs. |
457 | */ | |
458 | #define REAPTIMEOUT_CPUC (2*HZ) | |
459 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
460 | ||
461 | #if STATS | |
462 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
463 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
464 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
465 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 466 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
467 | #define STATS_SET_HIGH(x) \ |
468 | do { \ | |
469 | if ((x)->num_active > (x)->high_mark) \ | |
470 | (x)->high_mark = (x)->num_active; \ | |
471 | } while (0) | |
1da177e4 LT |
472 | #define STATS_INC_ERR(x) ((x)->errors++) |
473 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 474 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 475 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
476 | #define STATS_SET_FREEABLE(x, i) \ |
477 | do { \ | |
478 | if ((x)->max_freeable < i) \ | |
479 | (x)->max_freeable = i; \ | |
480 | } while (0) | |
1da177e4 LT |
481 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
482 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
483 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
484 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
485 | #else | |
486 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
487 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
488 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
489 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 490 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
491 | #define STATS_SET_HIGH(x) do { } while (0) |
492 | #define STATS_INC_ERR(x) do { } while (0) | |
493 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 494 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 495 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 496 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
497 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
498 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
499 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
500 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
501 | #endif | |
502 | ||
503 | #if DEBUG | |
1da177e4 | 504 | |
a737b3e2 AM |
505 | /* |
506 | * memory layout of objects: | |
1da177e4 | 507 | * 0 : objp |
3dafccf2 | 508 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
509 | * the end of an object is aligned with the end of the real |
510 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 511 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 512 | * redzone word. |
3dafccf2 MS |
513 | * cachep->obj_offset: The real object. |
514 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
515 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
516 | * [BYTES_PER_WORD long] | |
1da177e4 | 517 | */ |
343e0d7a | 518 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 519 | { |
3dafccf2 | 520 | return cachep->obj_offset; |
1da177e4 LT |
521 | } |
522 | ||
343e0d7a | 523 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 524 | { |
3dafccf2 | 525 | return cachep->obj_size; |
1da177e4 LT |
526 | } |
527 | ||
343e0d7a | 528 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
529 | { |
530 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 531 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
532 | } |
533 | ||
343e0d7a | 534 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
535 | { |
536 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
537 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 538 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 539 | 2 * BYTES_PER_WORD); |
3dafccf2 | 540 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
541 | } |
542 | ||
343e0d7a | 543 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
544 | { |
545 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 546 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
547 | } |
548 | ||
549 | #else | |
550 | ||
3dafccf2 MS |
551 | #define obj_offset(x) 0 |
552 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
553 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
554 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
555 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
556 | ||
557 | #endif | |
558 | ||
559 | /* | |
a737b3e2 AM |
560 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
561 | * order. | |
1da177e4 LT |
562 | */ |
563 | #if defined(CONFIG_LARGE_ALLOCS) | |
564 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
565 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
566 | #elif defined(CONFIG_MMU) | |
567 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
568 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
569 | #else | |
570 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
571 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
572 | #endif | |
573 | ||
574 | /* | |
575 | * Do not go above this order unless 0 objects fit into the slab. | |
576 | */ | |
577 | #define BREAK_GFP_ORDER_HI 1 | |
578 | #define BREAK_GFP_ORDER_LO 0 | |
579 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
580 | ||
a737b3e2 AM |
581 | /* |
582 | * Functions for storing/retrieving the cachep and or slab from the page | |
583 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
584 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 585 | */ |
065d41cb PE |
586 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
587 | { | |
588 | page->lru.next = (struct list_head *)cache; | |
589 | } | |
590 | ||
591 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
592 | { | |
84097518 NP |
593 | if (unlikely(PageCompound(page))) |
594 | page = (struct page *)page_private(page); | |
ddc2e812 | 595 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
596 | return (struct kmem_cache *)page->lru.next; |
597 | } | |
598 | ||
599 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
600 | { | |
601 | page->lru.prev = (struct list_head *)slab; | |
602 | } | |
603 | ||
604 | static inline struct slab *page_get_slab(struct page *page) | |
605 | { | |
84097518 NP |
606 | if (unlikely(PageCompound(page))) |
607 | page = (struct page *)page_private(page); | |
ddc2e812 | 608 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
609 | return (struct slab *)page->lru.prev; |
610 | } | |
1da177e4 | 611 | |
6ed5eb22 PE |
612 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
613 | { | |
614 | struct page *page = virt_to_page(obj); | |
615 | return page_get_cache(page); | |
616 | } | |
617 | ||
618 | static inline struct slab *virt_to_slab(const void *obj) | |
619 | { | |
620 | struct page *page = virt_to_page(obj); | |
621 | return page_get_slab(page); | |
622 | } | |
623 | ||
8fea4e96 PE |
624 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
625 | unsigned int idx) | |
626 | { | |
627 | return slab->s_mem + cache->buffer_size * idx; | |
628 | } | |
629 | ||
630 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
631 | struct slab *slab, void *obj) | |
632 | { | |
633 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
634 | } | |
635 | ||
a737b3e2 AM |
636 | /* |
637 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
638 | */ | |
1da177e4 LT |
639 | struct cache_sizes malloc_sizes[] = { |
640 | #define CACHE(x) { .cs_size = (x) }, | |
641 | #include <linux/kmalloc_sizes.h> | |
642 | CACHE(ULONG_MAX) | |
643 | #undef CACHE | |
644 | }; | |
645 | EXPORT_SYMBOL(malloc_sizes); | |
646 | ||
647 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
648 | struct cache_names { | |
649 | char *name; | |
650 | char *name_dma; | |
651 | }; | |
652 | ||
653 | static struct cache_names __initdata cache_names[] = { | |
654 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
655 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 656 | {NULL,} |
1da177e4 LT |
657 | #undef CACHE |
658 | }; | |
659 | ||
660 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 661 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 662 | static struct arraycache_init initarray_generic = |
b28a02de | 663 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
664 | |
665 | /* internal cache of cache description objs */ | |
343e0d7a | 666 | static struct kmem_cache cache_cache = { |
b28a02de PE |
667 | .batchcount = 1, |
668 | .limit = BOOT_CPUCACHE_ENTRIES, | |
669 | .shared = 1, | |
343e0d7a | 670 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 671 | .name = "kmem_cache", |
1da177e4 | 672 | #if DEBUG |
343e0d7a | 673 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
674 | #endif |
675 | }; | |
676 | ||
056c6241 RT |
677 | #define BAD_ALIEN_MAGIC 0x01020304ul |
678 | ||
f1aaee53 AV |
679 | #ifdef CONFIG_LOCKDEP |
680 | ||
681 | /* | |
682 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
683 | * for other slabs "off slab". | |
684 | * The locking for this is tricky in that it nests within the locks | |
685 | * of all other slabs in a few places; to deal with this special | |
686 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
687 | * |
688 | * We set lock class for alien array caches which are up during init. | |
689 | * The lock annotation will be lost if all cpus of a node goes down and | |
690 | * then comes back up during hotplug | |
f1aaee53 | 691 | */ |
056c6241 RT |
692 | static struct lock_class_key on_slab_l3_key; |
693 | static struct lock_class_key on_slab_alc_key; | |
694 | ||
695 | static inline void init_lock_keys(void) | |
f1aaee53 | 696 | |
f1aaee53 AV |
697 | { |
698 | int q; | |
056c6241 RT |
699 | struct cache_sizes *s = malloc_sizes; |
700 | ||
701 | while (s->cs_size != ULONG_MAX) { | |
702 | for_each_node(q) { | |
703 | struct array_cache **alc; | |
704 | int r; | |
705 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
706 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
707 | continue; | |
708 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
709 | alc = l3->alien; | |
710 | /* | |
711 | * FIXME: This check for BAD_ALIEN_MAGIC | |
712 | * should go away when common slab code is taught to | |
713 | * work even without alien caches. | |
714 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
715 | * for alloc_alien_cache, | |
716 | */ | |
717 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
718 | continue; | |
719 | for_each_node(r) { | |
720 | if (alc[r]) | |
721 | lockdep_set_class(&alc[r]->lock, | |
722 | &on_slab_alc_key); | |
723 | } | |
724 | } | |
725 | s++; | |
f1aaee53 AV |
726 | } |
727 | } | |
f1aaee53 | 728 | #else |
056c6241 | 729 | static inline void init_lock_keys(void) |
f1aaee53 AV |
730 | { |
731 | } | |
732 | #endif | |
733 | ||
8f5be20b RT |
734 | /* |
735 | * 1. Guard access to the cache-chain. | |
736 | * 2. Protect sanity of cpu_online_map against cpu hotplug events | |
737 | */ | |
fc0abb14 | 738 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
739 | static struct list_head cache_chain; |
740 | ||
1da177e4 LT |
741 | /* |
742 | * chicken and egg problem: delay the per-cpu array allocation | |
743 | * until the general caches are up. | |
744 | */ | |
745 | static enum { | |
746 | NONE, | |
e498be7d CL |
747 | PARTIAL_AC, |
748 | PARTIAL_L3, | |
1da177e4 LT |
749 | FULL |
750 | } g_cpucache_up; | |
751 | ||
39d24e64 MK |
752 | /* |
753 | * used by boot code to determine if it can use slab based allocator | |
754 | */ | |
755 | int slab_is_available(void) | |
756 | { | |
757 | return g_cpucache_up == FULL; | |
758 | } | |
759 | ||
52bad64d | 760 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 761 | |
343e0d7a | 762 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
763 | { |
764 | return cachep->array[smp_processor_id()]; | |
765 | } | |
766 | ||
a737b3e2 AM |
767 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
768 | gfp_t gfpflags) | |
1da177e4 LT |
769 | { |
770 | struct cache_sizes *csizep = malloc_sizes; | |
771 | ||
772 | #if DEBUG | |
773 | /* This happens if someone tries to call | |
b28a02de PE |
774 | * kmem_cache_create(), or __kmalloc(), before |
775 | * the generic caches are initialized. | |
776 | */ | |
c7e43c78 | 777 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
778 | #endif |
779 | while (size > csizep->cs_size) | |
780 | csizep++; | |
781 | ||
782 | /* | |
0abf40c1 | 783 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
784 | * has cs_{dma,}cachep==NULL. Thus no special case |
785 | * for large kmalloc calls required. | |
786 | */ | |
787 | if (unlikely(gfpflags & GFP_DMA)) | |
788 | return csizep->cs_dmacachep; | |
789 | return csizep->cs_cachep; | |
790 | } | |
791 | ||
b221385b | 792 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
793 | { |
794 | return __find_general_cachep(size, gfpflags); | |
795 | } | |
97e2bde4 | 796 | |
fbaccacf | 797 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 798 | { |
fbaccacf SR |
799 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
800 | } | |
1da177e4 | 801 | |
a737b3e2 AM |
802 | /* |
803 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
804 | */ | |
fbaccacf SR |
805 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
806 | size_t align, int flags, size_t *left_over, | |
807 | unsigned int *num) | |
808 | { | |
809 | int nr_objs; | |
810 | size_t mgmt_size; | |
811 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 812 | |
fbaccacf SR |
813 | /* |
814 | * The slab management structure can be either off the slab or | |
815 | * on it. For the latter case, the memory allocated for a | |
816 | * slab is used for: | |
817 | * | |
818 | * - The struct slab | |
819 | * - One kmem_bufctl_t for each object | |
820 | * - Padding to respect alignment of @align | |
821 | * - @buffer_size bytes for each object | |
822 | * | |
823 | * If the slab management structure is off the slab, then the | |
824 | * alignment will already be calculated into the size. Because | |
825 | * the slabs are all pages aligned, the objects will be at the | |
826 | * correct alignment when allocated. | |
827 | */ | |
828 | if (flags & CFLGS_OFF_SLAB) { | |
829 | mgmt_size = 0; | |
830 | nr_objs = slab_size / buffer_size; | |
831 | ||
832 | if (nr_objs > SLAB_LIMIT) | |
833 | nr_objs = SLAB_LIMIT; | |
834 | } else { | |
835 | /* | |
836 | * Ignore padding for the initial guess. The padding | |
837 | * is at most @align-1 bytes, and @buffer_size is at | |
838 | * least @align. In the worst case, this result will | |
839 | * be one greater than the number of objects that fit | |
840 | * into the memory allocation when taking the padding | |
841 | * into account. | |
842 | */ | |
843 | nr_objs = (slab_size - sizeof(struct slab)) / | |
844 | (buffer_size + sizeof(kmem_bufctl_t)); | |
845 | ||
846 | /* | |
847 | * This calculated number will be either the right | |
848 | * amount, or one greater than what we want. | |
849 | */ | |
850 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
851 | > slab_size) | |
852 | nr_objs--; | |
853 | ||
854 | if (nr_objs > SLAB_LIMIT) | |
855 | nr_objs = SLAB_LIMIT; | |
856 | ||
857 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
858 | } | |
859 | *num = nr_objs; | |
860 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
861 | } |
862 | ||
863 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
864 | ||
a737b3e2 AM |
865 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
866 | char *msg) | |
1da177e4 LT |
867 | { |
868 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 869 | function, cachep->name, msg); |
1da177e4 LT |
870 | dump_stack(); |
871 | } | |
872 | ||
3395ee05 PM |
873 | /* |
874 | * By default on NUMA we use alien caches to stage the freeing of | |
875 | * objects allocated from other nodes. This causes massive memory | |
876 | * inefficiencies when using fake NUMA setup to split memory into a | |
877 | * large number of small nodes, so it can be disabled on the command | |
878 | * line | |
879 | */ | |
880 | ||
881 | static int use_alien_caches __read_mostly = 1; | |
882 | static int __init noaliencache_setup(char *s) | |
883 | { | |
884 | use_alien_caches = 0; | |
885 | return 1; | |
886 | } | |
887 | __setup("noaliencache", noaliencache_setup); | |
888 | ||
8fce4d8e CL |
889 | #ifdef CONFIG_NUMA |
890 | /* | |
891 | * Special reaping functions for NUMA systems called from cache_reap(). | |
892 | * These take care of doing round robin flushing of alien caches (containing | |
893 | * objects freed on different nodes from which they were allocated) and the | |
894 | * flushing of remote pcps by calling drain_node_pages. | |
895 | */ | |
896 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
897 | ||
898 | static void init_reap_node(int cpu) | |
899 | { | |
900 | int node; | |
901 | ||
902 | node = next_node(cpu_to_node(cpu), node_online_map); | |
903 | if (node == MAX_NUMNODES) | |
442295c9 | 904 | node = first_node(node_online_map); |
8fce4d8e | 905 | |
7f6b8876 | 906 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
907 | } |
908 | ||
909 | static void next_reap_node(void) | |
910 | { | |
911 | int node = __get_cpu_var(reap_node); | |
912 | ||
913 | /* | |
914 | * Also drain per cpu pages on remote zones | |
915 | */ | |
916 | if (node != numa_node_id()) | |
917 | drain_node_pages(node); | |
918 | ||
919 | node = next_node(node, node_online_map); | |
920 | if (unlikely(node >= MAX_NUMNODES)) | |
921 | node = first_node(node_online_map); | |
922 | __get_cpu_var(reap_node) = node; | |
923 | } | |
924 | ||
925 | #else | |
926 | #define init_reap_node(cpu) do { } while (0) | |
927 | #define next_reap_node(void) do { } while (0) | |
928 | #endif | |
929 | ||
1da177e4 LT |
930 | /* |
931 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
932 | * via the workqueue/eventd. | |
933 | * Add the CPU number into the expiration time to minimize the possibility of | |
934 | * the CPUs getting into lockstep and contending for the global cache chain | |
935 | * lock. | |
936 | */ | |
937 | static void __devinit start_cpu_timer(int cpu) | |
938 | { | |
52bad64d | 939 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
940 | |
941 | /* | |
942 | * When this gets called from do_initcalls via cpucache_init(), | |
943 | * init_workqueues() has already run, so keventd will be setup | |
944 | * at that time. | |
945 | */ | |
52bad64d | 946 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 947 | init_reap_node(cpu); |
65f27f38 | 948 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
949 | schedule_delayed_work_on(cpu, reap_work, |
950 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
951 | } |
952 | } | |
953 | ||
e498be7d | 954 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 955 | int batchcount) |
1da177e4 | 956 | { |
b28a02de | 957 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
958 | struct array_cache *nc = NULL; |
959 | ||
e498be7d | 960 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
961 | if (nc) { |
962 | nc->avail = 0; | |
963 | nc->limit = entries; | |
964 | nc->batchcount = batchcount; | |
965 | nc->touched = 0; | |
e498be7d | 966 | spin_lock_init(&nc->lock); |
1da177e4 LT |
967 | } |
968 | return nc; | |
969 | } | |
970 | ||
3ded175a CL |
971 | /* |
972 | * Transfer objects in one arraycache to another. | |
973 | * Locking must be handled by the caller. | |
974 | * | |
975 | * Return the number of entries transferred. | |
976 | */ | |
977 | static int transfer_objects(struct array_cache *to, | |
978 | struct array_cache *from, unsigned int max) | |
979 | { | |
980 | /* Figure out how many entries to transfer */ | |
981 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
982 | ||
983 | if (!nr) | |
984 | return 0; | |
985 | ||
986 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
987 | sizeof(void *) *nr); | |
988 | ||
989 | from->avail -= nr; | |
990 | to->avail += nr; | |
991 | to->touched = 1; | |
992 | return nr; | |
993 | } | |
994 | ||
765c4507 CL |
995 | #ifndef CONFIG_NUMA |
996 | ||
997 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
998 | #define reap_alien(cachep, l3) do { } while (0) | |
999 | ||
1000 | static inline struct array_cache **alloc_alien_cache(int node, int limit) | |
1001 | { | |
1002 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1003 | } | |
1004 | ||
1005 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1006 | { | |
1007 | } | |
1008 | ||
1009 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1010 | { | |
1011 | return 0; | |
1012 | } | |
1013 | ||
1014 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1015 | gfp_t flags) | |
1016 | { | |
1017 | return NULL; | |
1018 | } | |
1019 | ||
8b98c169 | 1020 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1021 | gfp_t flags, int nodeid) |
1022 | { | |
1023 | return NULL; | |
1024 | } | |
1025 | ||
1026 | #else /* CONFIG_NUMA */ | |
1027 | ||
8b98c169 | 1028 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1029 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1030 | |
5295a74c | 1031 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
1032 | { |
1033 | struct array_cache **ac_ptr; | |
b28a02de | 1034 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
1035 | int i; |
1036 | ||
1037 | if (limit > 1) | |
1038 | limit = 12; | |
1039 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
1040 | if (ac_ptr) { | |
1041 | for_each_node(i) { | |
1042 | if (i == node || !node_online(i)) { | |
1043 | ac_ptr[i] = NULL; | |
1044 | continue; | |
1045 | } | |
1046 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
1047 | if (!ac_ptr[i]) { | |
b28a02de | 1048 | for (i--; i <= 0; i--) |
e498be7d CL |
1049 | kfree(ac_ptr[i]); |
1050 | kfree(ac_ptr); | |
1051 | return NULL; | |
1052 | } | |
1053 | } | |
1054 | } | |
1055 | return ac_ptr; | |
1056 | } | |
1057 | ||
5295a74c | 1058 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1059 | { |
1060 | int i; | |
1061 | ||
1062 | if (!ac_ptr) | |
1063 | return; | |
e498be7d | 1064 | for_each_node(i) |
b28a02de | 1065 | kfree(ac_ptr[i]); |
e498be7d CL |
1066 | kfree(ac_ptr); |
1067 | } | |
1068 | ||
343e0d7a | 1069 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1070 | struct array_cache *ac, int node) |
e498be7d CL |
1071 | { |
1072 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1073 | ||
1074 | if (ac->avail) { | |
1075 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1076 | /* |
1077 | * Stuff objects into the remote nodes shared array first. | |
1078 | * That way we could avoid the overhead of putting the objects | |
1079 | * into the free lists and getting them back later. | |
1080 | */ | |
693f7d36 | 1081 | if (rl3->shared) |
1082 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1083 | |
ff69416e | 1084 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1085 | ac->avail = 0; |
1086 | spin_unlock(&rl3->list_lock); | |
1087 | } | |
1088 | } | |
1089 | ||
8fce4d8e CL |
1090 | /* |
1091 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1092 | */ | |
1093 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1094 | { | |
1095 | int node = __get_cpu_var(reap_node); | |
1096 | ||
1097 | if (l3->alien) { | |
1098 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1099 | |
1100 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1101 | __drain_alien_cache(cachep, ac, node); |
1102 | spin_unlock_irq(&ac->lock); | |
1103 | } | |
1104 | } | |
1105 | } | |
1106 | ||
a737b3e2 AM |
1107 | static void drain_alien_cache(struct kmem_cache *cachep, |
1108 | struct array_cache **alien) | |
e498be7d | 1109 | { |
b28a02de | 1110 | int i = 0; |
e498be7d CL |
1111 | struct array_cache *ac; |
1112 | unsigned long flags; | |
1113 | ||
1114 | for_each_online_node(i) { | |
4484ebf1 | 1115 | ac = alien[i]; |
e498be7d CL |
1116 | if (ac) { |
1117 | spin_lock_irqsave(&ac->lock, flags); | |
1118 | __drain_alien_cache(cachep, ac, i); | |
1119 | spin_unlock_irqrestore(&ac->lock, flags); | |
1120 | } | |
1121 | } | |
1122 | } | |
729bd0b7 | 1123 | |
873623df | 1124 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1125 | { |
1126 | struct slab *slabp = virt_to_slab(objp); | |
1127 | int nodeid = slabp->nodeid; | |
1128 | struct kmem_list3 *l3; | |
1129 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1130 | int node; |
1131 | ||
1132 | node = numa_node_id(); | |
729bd0b7 PE |
1133 | |
1134 | /* | |
1135 | * Make sure we are not freeing a object from another node to the array | |
1136 | * cache on this cpu. | |
1137 | */ | |
3395ee05 | 1138 | if (likely(slabp->nodeid == node) || unlikely(!use_alien_caches)) |
729bd0b7 PE |
1139 | return 0; |
1140 | ||
1ca4cb24 | 1141 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1142 | STATS_INC_NODEFREES(cachep); |
1143 | if (l3->alien && l3->alien[nodeid]) { | |
1144 | alien = l3->alien[nodeid]; | |
873623df | 1145 | spin_lock(&alien->lock); |
729bd0b7 PE |
1146 | if (unlikely(alien->avail == alien->limit)) { |
1147 | STATS_INC_ACOVERFLOW(cachep); | |
1148 | __drain_alien_cache(cachep, alien, nodeid); | |
1149 | } | |
1150 | alien->entry[alien->avail++] = objp; | |
1151 | spin_unlock(&alien->lock); | |
1152 | } else { | |
1153 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1154 | free_block(cachep, &objp, 1, nodeid); | |
1155 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1156 | } | |
1157 | return 1; | |
1158 | } | |
e498be7d CL |
1159 | #endif |
1160 | ||
8c78f307 | 1161 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1162 | unsigned long action, void *hcpu) |
1da177e4 LT |
1163 | { |
1164 | long cpu = (long)hcpu; | |
343e0d7a | 1165 | struct kmem_cache *cachep; |
e498be7d CL |
1166 | struct kmem_list3 *l3 = NULL; |
1167 | int node = cpu_to_node(cpu); | |
1168 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1169 | |
1170 | switch (action) { | |
1171 | case CPU_UP_PREPARE: | |
fc0abb14 | 1172 | mutex_lock(&cache_chain_mutex); |
a737b3e2 AM |
1173 | /* |
1174 | * We need to do this right in the beginning since | |
e498be7d CL |
1175 | * alloc_arraycache's are going to use this list. |
1176 | * kmalloc_node allows us to add the slab to the right | |
1177 | * kmem_list3 and not this cpu's kmem_list3 | |
1178 | */ | |
1179 | ||
1da177e4 | 1180 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1181 | /* |
1182 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1183 | * begin anything. Make sure some other cpu on this |
1184 | * node has not already allocated this | |
1185 | */ | |
1186 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1187 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1188 | if (!l3) | |
e498be7d CL |
1189 | goto bad; |
1190 | kmem_list3_init(l3); | |
1191 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1192 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1193 | |
4484ebf1 RT |
1194 | /* |
1195 | * The l3s don't come and go as CPUs come and | |
1196 | * go. cache_chain_mutex is sufficient | |
1197 | * protection here. | |
1198 | */ | |
e498be7d CL |
1199 | cachep->nodelists[node] = l3; |
1200 | } | |
1da177e4 | 1201 | |
e498be7d CL |
1202 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1203 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1204 | (1 + nr_cpus_node(node)) * |
1205 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1206 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1207 | } | |
1208 | ||
a737b3e2 AM |
1209 | /* |
1210 | * Now we can go ahead with allocating the shared arrays and | |
1211 | * array caches | |
1212 | */ | |
e498be7d | 1213 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1214 | struct array_cache *nc; |
4484ebf1 | 1215 | struct array_cache *shared; |
3395ee05 | 1216 | struct array_cache **alien = NULL; |
cd105df4 | 1217 | |
e498be7d | 1218 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1219 | cachep->batchcount); |
1da177e4 LT |
1220 | if (!nc) |
1221 | goto bad; | |
4484ebf1 RT |
1222 | shared = alloc_arraycache(node, |
1223 | cachep->shared * cachep->batchcount, | |
1224 | 0xbaadf00d); | |
1225 | if (!shared) | |
1226 | goto bad; | |
7a21ef6f | 1227 | |
3395ee05 PM |
1228 | if (use_alien_caches) { |
1229 | alien = alloc_alien_cache(node, cachep->limit); | |
1230 | if (!alien) | |
1231 | goto bad; | |
1232 | } | |
1da177e4 | 1233 | cachep->array[cpu] = nc; |
e498be7d CL |
1234 | l3 = cachep->nodelists[node]; |
1235 | BUG_ON(!l3); | |
e498be7d | 1236 | |
4484ebf1 RT |
1237 | spin_lock_irq(&l3->list_lock); |
1238 | if (!l3->shared) { | |
1239 | /* | |
1240 | * We are serialised from CPU_DEAD or | |
1241 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1242 | */ | |
1243 | l3->shared = shared; | |
1244 | shared = NULL; | |
e498be7d | 1245 | } |
4484ebf1 RT |
1246 | #ifdef CONFIG_NUMA |
1247 | if (!l3->alien) { | |
1248 | l3->alien = alien; | |
1249 | alien = NULL; | |
1250 | } | |
1251 | #endif | |
1252 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1253 | kfree(shared); |
1254 | free_alien_cache(alien); | |
1da177e4 | 1255 | } |
1da177e4 LT |
1256 | break; |
1257 | case CPU_ONLINE: | |
8f5be20b | 1258 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1259 | start_cpu_timer(cpu); |
1260 | break; | |
1261 | #ifdef CONFIG_HOTPLUG_CPU | |
8f5be20b RT |
1262 | case CPU_DOWN_PREPARE: |
1263 | mutex_lock(&cache_chain_mutex); | |
1264 | break; | |
1265 | case CPU_DOWN_FAILED: | |
1266 | mutex_unlock(&cache_chain_mutex); | |
1267 | break; | |
1da177e4 | 1268 | case CPU_DEAD: |
4484ebf1 RT |
1269 | /* |
1270 | * Even if all the cpus of a node are down, we don't free the | |
1271 | * kmem_list3 of any cache. This to avoid a race between | |
1272 | * cpu_down, and a kmalloc allocation from another cpu for | |
1273 | * memory from the node of the cpu going down. The list3 | |
1274 | * structure is usually allocated from kmem_cache_create() and | |
1275 | * gets destroyed at kmem_cache_destroy(). | |
1276 | */ | |
1da177e4 | 1277 | /* fall thru */ |
8f5be20b | 1278 | #endif |
1da177e4 | 1279 | case CPU_UP_CANCELED: |
1da177e4 LT |
1280 | list_for_each_entry(cachep, &cache_chain, next) { |
1281 | struct array_cache *nc; | |
4484ebf1 RT |
1282 | struct array_cache *shared; |
1283 | struct array_cache **alien; | |
e498be7d | 1284 | cpumask_t mask; |
1da177e4 | 1285 | |
e498be7d | 1286 | mask = node_to_cpumask(node); |
1da177e4 LT |
1287 | /* cpu is dead; no one can alloc from it. */ |
1288 | nc = cachep->array[cpu]; | |
1289 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1290 | l3 = cachep->nodelists[node]; |
1291 | ||
1292 | if (!l3) | |
4484ebf1 | 1293 | goto free_array_cache; |
e498be7d | 1294 | |
ca3b9b91 | 1295 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1296 | |
1297 | /* Free limit for this kmem_list3 */ | |
1298 | l3->free_limit -= cachep->batchcount; | |
1299 | if (nc) | |
ff69416e | 1300 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1301 | |
1302 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1303 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1304 | goto free_array_cache; |
b28a02de | 1305 | } |
e498be7d | 1306 | |
4484ebf1 RT |
1307 | shared = l3->shared; |
1308 | if (shared) { | |
e498be7d | 1309 | free_block(cachep, l3->shared->entry, |
b28a02de | 1310 | l3->shared->avail, node); |
e498be7d CL |
1311 | l3->shared = NULL; |
1312 | } | |
e498be7d | 1313 | |
4484ebf1 RT |
1314 | alien = l3->alien; |
1315 | l3->alien = NULL; | |
1316 | ||
1317 | spin_unlock_irq(&l3->list_lock); | |
1318 | ||
1319 | kfree(shared); | |
1320 | if (alien) { | |
1321 | drain_alien_cache(cachep, alien); | |
1322 | free_alien_cache(alien); | |
e498be7d | 1323 | } |
4484ebf1 | 1324 | free_array_cache: |
1da177e4 LT |
1325 | kfree(nc); |
1326 | } | |
4484ebf1 RT |
1327 | /* |
1328 | * In the previous loop, all the objects were freed to | |
1329 | * the respective cache's slabs, now we can go ahead and | |
1330 | * shrink each nodelist to its limit. | |
1331 | */ | |
1332 | list_for_each_entry(cachep, &cache_chain, next) { | |
1333 | l3 = cachep->nodelists[node]; | |
1334 | if (!l3) | |
1335 | continue; | |
ed11d9eb | 1336 | drain_freelist(cachep, l3, l3->free_objects); |
4484ebf1 | 1337 | } |
fc0abb14 | 1338 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1339 | break; |
1da177e4 LT |
1340 | } |
1341 | return NOTIFY_OK; | |
a737b3e2 | 1342 | bad: |
1da177e4 LT |
1343 | return NOTIFY_BAD; |
1344 | } | |
1345 | ||
74b85f37 CS |
1346 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1347 | &cpuup_callback, NULL, 0 | |
1348 | }; | |
1da177e4 | 1349 | |
e498be7d CL |
1350 | /* |
1351 | * swap the static kmem_list3 with kmalloced memory | |
1352 | */ | |
a737b3e2 AM |
1353 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1354 | int nodeid) | |
e498be7d CL |
1355 | { |
1356 | struct kmem_list3 *ptr; | |
1357 | ||
e498be7d CL |
1358 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); |
1359 | BUG_ON(!ptr); | |
1360 | ||
1361 | local_irq_disable(); | |
1362 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1363 | /* |
1364 | * Do not assume that spinlocks can be initialized via memcpy: | |
1365 | */ | |
1366 | spin_lock_init(&ptr->list_lock); | |
1367 | ||
e498be7d CL |
1368 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1369 | cachep->nodelists[nodeid] = ptr; | |
1370 | local_irq_enable(); | |
1371 | } | |
1372 | ||
a737b3e2 AM |
1373 | /* |
1374 | * Initialisation. Called after the page allocator have been initialised and | |
1375 | * before smp_init(). | |
1da177e4 LT |
1376 | */ |
1377 | void __init kmem_cache_init(void) | |
1378 | { | |
1379 | size_t left_over; | |
1380 | struct cache_sizes *sizes; | |
1381 | struct cache_names *names; | |
e498be7d | 1382 | int i; |
07ed76b2 | 1383 | int order; |
1ca4cb24 | 1384 | int node; |
e498be7d CL |
1385 | |
1386 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1387 | kmem_list3_init(&initkmem_list3[i]); | |
1388 | if (i < MAX_NUMNODES) | |
1389 | cache_cache.nodelists[i] = NULL; | |
1390 | } | |
1da177e4 LT |
1391 | |
1392 | /* | |
1393 | * Fragmentation resistance on low memory - only use bigger | |
1394 | * page orders on machines with more than 32MB of memory. | |
1395 | */ | |
1396 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1397 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1398 | ||
1da177e4 LT |
1399 | /* Bootstrap is tricky, because several objects are allocated |
1400 | * from caches that do not exist yet: | |
a737b3e2 AM |
1401 | * 1) initialize the cache_cache cache: it contains the struct |
1402 | * kmem_cache structures of all caches, except cache_cache itself: | |
1403 | * cache_cache is statically allocated. | |
e498be7d CL |
1404 | * Initially an __init data area is used for the head array and the |
1405 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1406 | * array at the end of the bootstrap. | |
1da177e4 | 1407 | * 2) Create the first kmalloc cache. |
343e0d7a | 1408 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1409 | * An __init data area is used for the head array. |
1410 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1411 | * head arrays. | |
1da177e4 LT |
1412 | * 4) Replace the __init data head arrays for cache_cache and the first |
1413 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1414 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1415 | * the other cache's with kmalloc allocated memory. | |
1416 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1417 | */ |
1418 | ||
1ca4cb24 PE |
1419 | node = numa_node_id(); |
1420 | ||
1da177e4 | 1421 | /* 1) create the cache_cache */ |
1da177e4 LT |
1422 | INIT_LIST_HEAD(&cache_chain); |
1423 | list_add(&cache_cache.next, &cache_chain); | |
1424 | cache_cache.colour_off = cache_line_size(); | |
1425 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
1ca4cb24 | 1426 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1427 | |
a737b3e2 AM |
1428 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1429 | cache_line_size()); | |
1da177e4 | 1430 | |
07ed76b2 JS |
1431 | for (order = 0; order < MAX_ORDER; order++) { |
1432 | cache_estimate(order, cache_cache.buffer_size, | |
1433 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1434 | if (cache_cache.num) | |
1435 | break; | |
1436 | } | |
40094fa6 | 1437 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1438 | cache_cache.gfporder = order; |
b28a02de | 1439 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1440 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1441 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1442 | |
1443 | /* 2+3) create the kmalloc caches */ | |
1444 | sizes = malloc_sizes; | |
1445 | names = cache_names; | |
1446 | ||
a737b3e2 AM |
1447 | /* |
1448 | * Initialize the caches that provide memory for the array cache and the | |
1449 | * kmem_list3 structures first. Without this, further allocations will | |
1450 | * bug. | |
e498be7d CL |
1451 | */ |
1452 | ||
1453 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1454 | sizes[INDEX_AC].cs_size, |
1455 | ARCH_KMALLOC_MINALIGN, | |
1456 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1457 | NULL, NULL); | |
e498be7d | 1458 | |
a737b3e2 | 1459 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1460 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1461 | kmem_cache_create(names[INDEX_L3].name, |
1462 | sizes[INDEX_L3].cs_size, | |
1463 | ARCH_KMALLOC_MINALIGN, | |
1464 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1465 | NULL, NULL); | |
1466 | } | |
e498be7d | 1467 | |
e0a42726 IM |
1468 | slab_early_init = 0; |
1469 | ||
1da177e4 | 1470 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1471 | /* |
1472 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1473 | * This should be particularly beneficial on SMP boxes, as it |
1474 | * eliminates "false sharing". | |
1475 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1476 | * allow tighter packing of the smaller caches. |
1477 | */ | |
a737b3e2 | 1478 | if (!sizes->cs_cachep) { |
e498be7d | 1479 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1480 | sizes->cs_size, |
1481 | ARCH_KMALLOC_MINALIGN, | |
1482 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1483 | NULL, NULL); | |
1484 | } | |
1da177e4 | 1485 | |
1da177e4 | 1486 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, |
a737b3e2 AM |
1487 | sizes->cs_size, |
1488 | ARCH_KMALLOC_MINALIGN, | |
1489 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1490 | SLAB_PANIC, | |
1491 | NULL, NULL); | |
1da177e4 LT |
1492 | sizes++; |
1493 | names++; | |
1494 | } | |
1495 | /* 4) Replace the bootstrap head arrays */ | |
1496 | { | |
2b2d5493 | 1497 | struct array_cache *ptr; |
e498be7d | 1498 | |
1da177e4 | 1499 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1500 | |
1da177e4 | 1501 | local_irq_disable(); |
9a2dba4b PE |
1502 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1503 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1504 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1505 | /* |
1506 | * Do not assume that spinlocks can be initialized via memcpy: | |
1507 | */ | |
1508 | spin_lock_init(&ptr->lock); | |
1509 | ||
1da177e4 LT |
1510 | cache_cache.array[smp_processor_id()] = ptr; |
1511 | local_irq_enable(); | |
e498be7d | 1512 | |
1da177e4 | 1513 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1514 | |
1da177e4 | 1515 | local_irq_disable(); |
9a2dba4b | 1516 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1517 | != &initarray_generic.cache); |
9a2dba4b | 1518 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1519 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1520 | /* |
1521 | * Do not assume that spinlocks can be initialized via memcpy: | |
1522 | */ | |
1523 | spin_lock_init(&ptr->lock); | |
1524 | ||
e498be7d | 1525 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1526 | ptr; |
1da177e4 LT |
1527 | local_irq_enable(); |
1528 | } | |
e498be7d CL |
1529 | /* 5) Replace the bootstrap kmem_list3's */ |
1530 | { | |
1ca4cb24 PE |
1531 | int nid; |
1532 | ||
e498be7d | 1533 | /* Replace the static kmem_list3 structures for the boot cpu */ |
1ca4cb24 | 1534 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); |
e498be7d | 1535 | |
1ca4cb24 | 1536 | for_each_online_node(nid) { |
e498be7d | 1537 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1538 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1539 | |
1540 | if (INDEX_AC != INDEX_L3) { | |
1541 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1542 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1543 | } |
1544 | } | |
1545 | } | |
1da177e4 | 1546 | |
e498be7d | 1547 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1548 | { |
343e0d7a | 1549 | struct kmem_cache *cachep; |
fc0abb14 | 1550 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1551 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1552 | if (enable_cpucache(cachep)) |
1553 | BUG(); | |
fc0abb14 | 1554 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1555 | } |
1556 | ||
056c6241 RT |
1557 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1558 | init_lock_keys(); | |
1559 | ||
1560 | ||
1da177e4 LT |
1561 | /* Done! */ |
1562 | g_cpucache_up = FULL; | |
1563 | ||
a737b3e2 AM |
1564 | /* |
1565 | * Register a cpu startup notifier callback that initializes | |
1566 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1567 | */ |
1568 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1569 | |
a737b3e2 AM |
1570 | /* |
1571 | * The reap timers are started later, with a module init call: That part | |
1572 | * of the kernel is not yet operational. | |
1da177e4 LT |
1573 | */ |
1574 | } | |
1575 | ||
1576 | static int __init cpucache_init(void) | |
1577 | { | |
1578 | int cpu; | |
1579 | ||
a737b3e2 AM |
1580 | /* |
1581 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1582 | */ |
e498be7d | 1583 | for_each_online_cpu(cpu) |
a737b3e2 | 1584 | start_cpu_timer(cpu); |
1da177e4 LT |
1585 | return 0; |
1586 | } | |
1da177e4 LT |
1587 | __initcall(cpucache_init); |
1588 | ||
1589 | /* | |
1590 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1591 | * | |
1592 | * If we requested dmaable memory, we will get it. Even if we | |
1593 | * did not request dmaable memory, we might get it, but that | |
1594 | * would be relatively rare and ignorable. | |
1595 | */ | |
343e0d7a | 1596 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1597 | { |
1598 | struct page *page; | |
e1b6aa6f | 1599 | int nr_pages; |
1da177e4 LT |
1600 | int i; |
1601 | ||
d6fef9da | 1602 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1603 | /* |
1604 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1605 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1606 | */ |
e1b6aa6f | 1607 | flags |= __GFP_COMP; |
d6fef9da | 1608 | #endif |
765c4507 | 1609 | |
3c517a61 | 1610 | flags |= cachep->gfpflags; |
e1b6aa6f CH |
1611 | |
1612 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1613 | if (!page) |
1614 | return NULL; | |
1da177e4 | 1615 | |
e1b6aa6f | 1616 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1617 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1618 | add_zone_page_state(page_zone(page), |
1619 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1620 | else | |
1621 | add_zone_page_state(page_zone(page), | |
1622 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1623 | for (i = 0; i < nr_pages; i++) |
1624 | __SetPageSlab(page + i); | |
1625 | return page_address(page); | |
1da177e4 LT |
1626 | } |
1627 | ||
1628 | /* | |
1629 | * Interface to system's page release. | |
1630 | */ | |
343e0d7a | 1631 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1632 | { |
b28a02de | 1633 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1634 | struct page *page = virt_to_page(addr); |
1635 | const unsigned long nr_freed = i; | |
1636 | ||
972d1a7b CL |
1637 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1638 | sub_zone_page_state(page_zone(page), | |
1639 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1640 | else | |
1641 | sub_zone_page_state(page_zone(page), | |
1642 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1643 | while (i--) { |
f205b2fe NP |
1644 | BUG_ON(!PageSlab(page)); |
1645 | __ClearPageSlab(page); | |
1da177e4 LT |
1646 | page++; |
1647 | } | |
1da177e4 LT |
1648 | if (current->reclaim_state) |
1649 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1650 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1651 | } |
1652 | ||
1653 | static void kmem_rcu_free(struct rcu_head *head) | |
1654 | { | |
b28a02de | 1655 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1656 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1657 | |
1658 | kmem_freepages(cachep, slab_rcu->addr); | |
1659 | if (OFF_SLAB(cachep)) | |
1660 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1661 | } | |
1662 | ||
1663 | #if DEBUG | |
1664 | ||
1665 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1666 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1667 | unsigned long caller) |
1da177e4 | 1668 | { |
3dafccf2 | 1669 | int size = obj_size(cachep); |
1da177e4 | 1670 | |
3dafccf2 | 1671 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1672 | |
b28a02de | 1673 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1674 | return; |
1675 | ||
b28a02de PE |
1676 | *addr++ = 0x12345678; |
1677 | *addr++ = caller; | |
1678 | *addr++ = smp_processor_id(); | |
1679 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1680 | { |
1681 | unsigned long *sptr = &caller; | |
1682 | unsigned long svalue; | |
1683 | ||
1684 | while (!kstack_end(sptr)) { | |
1685 | svalue = *sptr++; | |
1686 | if (kernel_text_address(svalue)) { | |
b28a02de | 1687 | *addr++ = svalue; |
1da177e4 LT |
1688 | size -= sizeof(unsigned long); |
1689 | if (size <= sizeof(unsigned long)) | |
1690 | break; | |
1691 | } | |
1692 | } | |
1693 | ||
1694 | } | |
b28a02de | 1695 | *addr++ = 0x87654321; |
1da177e4 LT |
1696 | } |
1697 | #endif | |
1698 | ||
343e0d7a | 1699 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1700 | { |
3dafccf2 MS |
1701 | int size = obj_size(cachep); |
1702 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1703 | |
1704 | memset(addr, val, size); | |
b28a02de | 1705 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1706 | } |
1707 | ||
1708 | static void dump_line(char *data, int offset, int limit) | |
1709 | { | |
1710 | int i; | |
aa83aa40 DJ |
1711 | unsigned char error = 0; |
1712 | int bad_count = 0; | |
1713 | ||
1da177e4 | 1714 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1715 | for (i = 0; i < limit; i++) { |
1716 | if (data[offset + i] != POISON_FREE) { | |
1717 | error = data[offset + i]; | |
1718 | bad_count++; | |
1719 | } | |
b28a02de | 1720 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1721 | } |
1da177e4 | 1722 | printk("\n"); |
aa83aa40 DJ |
1723 | |
1724 | if (bad_count == 1) { | |
1725 | error ^= POISON_FREE; | |
1726 | if (!(error & (error - 1))) { | |
1727 | printk(KERN_ERR "Single bit error detected. Probably " | |
1728 | "bad RAM.\n"); | |
1729 | #ifdef CONFIG_X86 | |
1730 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1731 | "test tool.\n"); | |
1732 | #else | |
1733 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1734 | #endif | |
1735 | } | |
1736 | } | |
1da177e4 LT |
1737 | } |
1738 | #endif | |
1739 | ||
1740 | #if DEBUG | |
1741 | ||
343e0d7a | 1742 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1743 | { |
1744 | int i, size; | |
1745 | char *realobj; | |
1746 | ||
1747 | if (cachep->flags & SLAB_RED_ZONE) { | |
1748 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
a737b3e2 AM |
1749 | *dbg_redzone1(cachep, objp), |
1750 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1751 | } |
1752 | ||
1753 | if (cachep->flags & SLAB_STORE_USER) { | |
1754 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1755 | *dbg_userword(cachep, objp)); |
1da177e4 | 1756 | print_symbol("(%s)", |
a737b3e2 | 1757 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1758 | printk("\n"); |
1759 | } | |
3dafccf2 MS |
1760 | realobj = (char *)objp + obj_offset(cachep); |
1761 | size = obj_size(cachep); | |
b28a02de | 1762 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1763 | int limit; |
1764 | limit = 16; | |
b28a02de PE |
1765 | if (i + limit > size) |
1766 | limit = size - i; | |
1da177e4 LT |
1767 | dump_line(realobj, i, limit); |
1768 | } | |
1769 | } | |
1770 | ||
343e0d7a | 1771 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1772 | { |
1773 | char *realobj; | |
1774 | int size, i; | |
1775 | int lines = 0; | |
1776 | ||
3dafccf2 MS |
1777 | realobj = (char *)objp + obj_offset(cachep); |
1778 | size = obj_size(cachep); | |
1da177e4 | 1779 | |
b28a02de | 1780 | for (i = 0; i < size; i++) { |
1da177e4 | 1781 | char exp = POISON_FREE; |
b28a02de | 1782 | if (i == size - 1) |
1da177e4 LT |
1783 | exp = POISON_END; |
1784 | if (realobj[i] != exp) { | |
1785 | int limit; | |
1786 | /* Mismatch ! */ | |
1787 | /* Print header */ | |
1788 | if (lines == 0) { | |
b28a02de | 1789 | printk(KERN_ERR |
a737b3e2 AM |
1790 | "Slab corruption: start=%p, len=%d\n", |
1791 | realobj, size); | |
1da177e4 LT |
1792 | print_objinfo(cachep, objp, 0); |
1793 | } | |
1794 | /* Hexdump the affected line */ | |
b28a02de | 1795 | i = (i / 16) * 16; |
1da177e4 | 1796 | limit = 16; |
b28a02de PE |
1797 | if (i + limit > size) |
1798 | limit = size - i; | |
1da177e4 LT |
1799 | dump_line(realobj, i, limit); |
1800 | i += 16; | |
1801 | lines++; | |
1802 | /* Limit to 5 lines */ | |
1803 | if (lines > 5) | |
1804 | break; | |
1805 | } | |
1806 | } | |
1807 | if (lines != 0) { | |
1808 | /* Print some data about the neighboring objects, if they | |
1809 | * exist: | |
1810 | */ | |
6ed5eb22 | 1811 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1812 | unsigned int objnr; |
1da177e4 | 1813 | |
8fea4e96 | 1814 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1815 | if (objnr) { |
8fea4e96 | 1816 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1817 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1818 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1819 | realobj, size); |
1da177e4 LT |
1820 | print_objinfo(cachep, objp, 2); |
1821 | } | |
b28a02de | 1822 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1823 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1824 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1825 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1826 | realobj, size); |
1da177e4 LT |
1827 | print_objinfo(cachep, objp, 2); |
1828 | } | |
1829 | } | |
1830 | } | |
1831 | #endif | |
1832 | ||
12dd36fa MD |
1833 | #if DEBUG |
1834 | /** | |
911851e6 RD |
1835 | * slab_destroy_objs - destroy a slab and its objects |
1836 | * @cachep: cache pointer being destroyed | |
1837 | * @slabp: slab pointer being destroyed | |
1838 | * | |
1839 | * Call the registered destructor for each object in a slab that is being | |
1840 | * destroyed. | |
1da177e4 | 1841 | */ |
343e0d7a | 1842 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1843 | { |
1da177e4 LT |
1844 | int i; |
1845 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1846 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1847 | |
1848 | if (cachep->flags & SLAB_POISON) { | |
1849 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1850 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1851 | OFF_SLAB(cachep)) | |
b28a02de | 1852 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1853 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1854 | else |
1855 | check_poison_obj(cachep, objp); | |
1856 | #else | |
1857 | check_poison_obj(cachep, objp); | |
1858 | #endif | |
1859 | } | |
1860 | if (cachep->flags & SLAB_RED_ZONE) { | |
1861 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1862 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1863 | "was overwritten"); |
1da177e4 LT |
1864 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1865 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1866 | "was overwritten"); |
1da177e4 LT |
1867 | } |
1868 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1869 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1870 | } |
12dd36fa | 1871 | } |
1da177e4 | 1872 | #else |
343e0d7a | 1873 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1874 | { |
1da177e4 LT |
1875 | if (cachep->dtor) { |
1876 | int i; | |
1877 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1878 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1879 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1880 | } |
1881 | } | |
12dd36fa | 1882 | } |
1da177e4 LT |
1883 | #endif |
1884 | ||
911851e6 RD |
1885 | /** |
1886 | * slab_destroy - destroy and release all objects in a slab | |
1887 | * @cachep: cache pointer being destroyed | |
1888 | * @slabp: slab pointer being destroyed | |
1889 | * | |
12dd36fa | 1890 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1891 | * Before calling the slab must have been unlinked from the cache. The |
1892 | * cache-lock is not held/needed. | |
12dd36fa | 1893 | */ |
343e0d7a | 1894 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1895 | { |
1896 | void *addr = slabp->s_mem - slabp->colouroff; | |
1897 | ||
1898 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1899 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1900 | struct slab_rcu *slab_rcu; | |
1901 | ||
b28a02de | 1902 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1903 | slab_rcu->cachep = cachep; |
1904 | slab_rcu->addr = addr; | |
1905 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1906 | } else { | |
1907 | kmem_freepages(cachep, addr); | |
873623df IM |
1908 | if (OFF_SLAB(cachep)) |
1909 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1910 | } |
1911 | } | |
1912 | ||
a737b3e2 AM |
1913 | /* |
1914 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1915 | * size of kmem_list3. | |
1916 | */ | |
343e0d7a | 1917 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1918 | { |
1919 | int node; | |
1920 | ||
1921 | for_each_online_node(node) { | |
b28a02de | 1922 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1923 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1924 | REAPTIMEOUT_LIST3 + |
1925 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1926 | } |
1927 | } | |
1928 | ||
117f6eb1 CL |
1929 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1930 | { | |
1931 | int i; | |
1932 | struct kmem_list3 *l3; | |
1933 | ||
1934 | for_each_online_cpu(i) | |
1935 | kfree(cachep->array[i]); | |
1936 | ||
1937 | /* NUMA: free the list3 structures */ | |
1938 | for_each_online_node(i) { | |
1939 | l3 = cachep->nodelists[i]; | |
1940 | if (l3) { | |
1941 | kfree(l3->shared); | |
1942 | free_alien_cache(l3->alien); | |
1943 | kfree(l3); | |
1944 | } | |
1945 | } | |
1946 | kmem_cache_free(&cache_cache, cachep); | |
1947 | } | |
1948 | ||
1949 | ||
4d268eba | 1950 | /** |
a70773dd RD |
1951 | * calculate_slab_order - calculate size (page order) of slabs |
1952 | * @cachep: pointer to the cache that is being created | |
1953 | * @size: size of objects to be created in this cache. | |
1954 | * @align: required alignment for the objects. | |
1955 | * @flags: slab allocation flags | |
1956 | * | |
1957 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1958 | * |
1959 | * This could be made much more intelligent. For now, try to avoid using | |
1960 | * high order pages for slabs. When the gfp() functions are more friendly | |
1961 | * towards high-order requests, this should be changed. | |
1962 | */ | |
a737b3e2 | 1963 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1964 | size_t size, size_t align, unsigned long flags) |
4d268eba | 1965 | { |
b1ab41c4 | 1966 | unsigned long offslab_limit; |
4d268eba | 1967 | size_t left_over = 0; |
9888e6fa | 1968 | int gfporder; |
4d268eba | 1969 | |
a737b3e2 | 1970 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1971 | unsigned int num; |
1972 | size_t remainder; | |
1973 | ||
9888e6fa | 1974 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1975 | if (!num) |
1976 | continue; | |
9888e6fa | 1977 | |
b1ab41c4 IM |
1978 | if (flags & CFLGS_OFF_SLAB) { |
1979 | /* | |
1980 | * Max number of objs-per-slab for caches which | |
1981 | * use off-slab slabs. Needed to avoid a possible | |
1982 | * looping condition in cache_grow(). | |
1983 | */ | |
1984 | offslab_limit = size - sizeof(struct slab); | |
1985 | offslab_limit /= sizeof(kmem_bufctl_t); | |
1986 | ||
1987 | if (num > offslab_limit) | |
1988 | break; | |
1989 | } | |
4d268eba | 1990 | |
9888e6fa | 1991 | /* Found something acceptable - save it away */ |
4d268eba | 1992 | cachep->num = num; |
9888e6fa | 1993 | cachep->gfporder = gfporder; |
4d268eba PE |
1994 | left_over = remainder; |
1995 | ||
f78bb8ad LT |
1996 | /* |
1997 | * A VFS-reclaimable slab tends to have most allocations | |
1998 | * as GFP_NOFS and we really don't want to have to be allocating | |
1999 | * higher-order pages when we are unable to shrink dcache. | |
2000 | */ | |
2001 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2002 | break; | |
2003 | ||
4d268eba PE |
2004 | /* |
2005 | * Large number of objects is good, but very large slabs are | |
2006 | * currently bad for the gfp()s. | |
2007 | */ | |
9888e6fa | 2008 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2009 | break; |
2010 | ||
9888e6fa LT |
2011 | /* |
2012 | * Acceptable internal fragmentation? | |
2013 | */ | |
a737b3e2 | 2014 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2015 | break; |
2016 | } | |
2017 | return left_over; | |
2018 | } | |
2019 | ||
2ed3a4ef | 2020 | static int setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 2021 | { |
2ed3a4ef CL |
2022 | if (g_cpucache_up == FULL) |
2023 | return enable_cpucache(cachep); | |
2024 | ||
f30cf7d1 PE |
2025 | if (g_cpucache_up == NONE) { |
2026 | /* | |
2027 | * Note: the first kmem_cache_create must create the cache | |
2028 | * that's used by kmalloc(24), otherwise the creation of | |
2029 | * further caches will BUG(). | |
2030 | */ | |
2031 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2032 | ||
2033 | /* | |
2034 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2035 | * the first cache, then we need to set up all its list3s, | |
2036 | * otherwise the creation of further caches will BUG(). | |
2037 | */ | |
2038 | set_up_list3s(cachep, SIZE_AC); | |
2039 | if (INDEX_AC == INDEX_L3) | |
2040 | g_cpucache_up = PARTIAL_L3; | |
2041 | else | |
2042 | g_cpucache_up = PARTIAL_AC; | |
2043 | } else { | |
2044 | cachep->array[smp_processor_id()] = | |
2045 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
2046 | ||
2047 | if (g_cpucache_up == PARTIAL_AC) { | |
2048 | set_up_list3s(cachep, SIZE_L3); | |
2049 | g_cpucache_up = PARTIAL_L3; | |
2050 | } else { | |
2051 | int node; | |
2052 | for_each_online_node(node) { | |
2053 | cachep->nodelists[node] = | |
2054 | kmalloc_node(sizeof(struct kmem_list3), | |
2055 | GFP_KERNEL, node); | |
2056 | BUG_ON(!cachep->nodelists[node]); | |
2057 | kmem_list3_init(cachep->nodelists[node]); | |
2058 | } | |
2059 | } | |
2060 | } | |
2061 | cachep->nodelists[numa_node_id()]->next_reap = | |
2062 | jiffies + REAPTIMEOUT_LIST3 + | |
2063 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2064 | ||
2065 | cpu_cache_get(cachep)->avail = 0; | |
2066 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2067 | cpu_cache_get(cachep)->batchcount = 1; | |
2068 | cpu_cache_get(cachep)->touched = 0; | |
2069 | cachep->batchcount = 1; | |
2070 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2071 | return 0; |
f30cf7d1 PE |
2072 | } |
2073 | ||
1da177e4 LT |
2074 | /** |
2075 | * kmem_cache_create - Create a cache. | |
2076 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2077 | * @size: The size of objects to be created in this cache. | |
2078 | * @align: The required alignment for the objects. | |
2079 | * @flags: SLAB flags | |
2080 | * @ctor: A constructor for the objects. | |
2081 | * @dtor: A destructor for the objects. | |
2082 | * | |
2083 | * Returns a ptr to the cache on success, NULL on failure. | |
2084 | * Cannot be called within a int, but can be interrupted. | |
2085 | * The @ctor is run when new pages are allocated by the cache | |
2086 | * and the @dtor is run before the pages are handed back. | |
2087 | * | |
2088 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2089 | * the module calling this has to destroy the cache before getting unloaded. |
2090 | * | |
1da177e4 LT |
2091 | * The flags are |
2092 | * | |
2093 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2094 | * to catch references to uninitialised memory. | |
2095 | * | |
2096 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2097 | * for buffer overruns. | |
2098 | * | |
1da177e4 LT |
2099 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2100 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2101 | * as davem. | |
2102 | */ | |
343e0d7a | 2103 | struct kmem_cache * |
1da177e4 | 2104 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
2105 | unsigned long flags, |
2106 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 2107 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
2108 | { |
2109 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2110 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2111 | |
2112 | /* | |
2113 | * Sanity checks... these are all serious usage bugs. | |
2114 | */ | |
a737b3e2 | 2115 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
b28a02de | 2116 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { |
a737b3e2 AM |
2117 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
2118 | name); | |
b28a02de PE |
2119 | BUG(); |
2120 | } | |
1da177e4 | 2121 | |
f0188f47 | 2122 | /* |
8f5be20b RT |
2123 | * We use cache_chain_mutex to ensure a consistent view of |
2124 | * cpu_online_map as well. Please see cpuup_callback | |
f0188f47 | 2125 | */ |
fc0abb14 | 2126 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2127 | |
7a7c381d | 2128 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2129 | char tmp; |
2130 | int res; | |
2131 | ||
2132 | /* | |
2133 | * This happens when the module gets unloaded and doesn't | |
2134 | * destroy its slab cache and no-one else reuses the vmalloc | |
2135 | * area of the module. Print a warning. | |
2136 | */ | |
138ae663 | 2137 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f AM |
2138 | if (res) { |
2139 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2140 | pc->buffer_size); |
4f12bb4f AM |
2141 | continue; |
2142 | } | |
2143 | ||
b28a02de | 2144 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
2145 | printk("kmem_cache_create: duplicate cache %s\n", name); |
2146 | dump_stack(); | |
2147 | goto oops; | |
2148 | } | |
2149 | } | |
2150 | ||
1da177e4 LT |
2151 | #if DEBUG |
2152 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
2153 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
2154 | /* No constructor, but inital state check requested */ | |
2155 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 2156 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
2157 | flags &= ~SLAB_DEBUG_INITIAL; |
2158 | } | |
1da177e4 LT |
2159 | #if FORCED_DEBUG |
2160 | /* | |
2161 | * Enable redzoning and last user accounting, except for caches with | |
2162 | * large objects, if the increased size would increase the object size | |
2163 | * above the next power of two: caches with object sizes just above a | |
2164 | * power of two have a significant amount of internal fragmentation. | |
2165 | */ | |
a737b3e2 | 2166 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 2167 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2168 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2169 | flags |= SLAB_POISON; | |
2170 | #endif | |
2171 | if (flags & SLAB_DESTROY_BY_RCU) | |
2172 | BUG_ON(flags & SLAB_POISON); | |
2173 | #endif | |
2174 | if (flags & SLAB_DESTROY_BY_RCU) | |
2175 | BUG_ON(dtor); | |
2176 | ||
2177 | /* | |
a737b3e2 AM |
2178 | * Always checks flags, a caller might be expecting debug support which |
2179 | * isn't available. | |
1da177e4 | 2180 | */ |
40094fa6 | 2181 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2182 | |
a737b3e2 AM |
2183 | /* |
2184 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2185 | * unaligned accesses for some archs when redzoning is used, and makes |
2186 | * sure any on-slab bufctl's are also correctly aligned. | |
2187 | */ | |
b28a02de PE |
2188 | if (size & (BYTES_PER_WORD - 1)) { |
2189 | size += (BYTES_PER_WORD - 1); | |
2190 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2191 | } |
2192 | ||
a737b3e2 AM |
2193 | /* calculate the final buffer alignment: */ |
2194 | ||
1da177e4 LT |
2195 | /* 1) arch recommendation: can be overridden for debug */ |
2196 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2197 | /* |
2198 | * Default alignment: as specified by the arch code. Except if | |
2199 | * an object is really small, then squeeze multiple objects into | |
2200 | * one cacheline. | |
1da177e4 LT |
2201 | */ |
2202 | ralign = cache_line_size(); | |
b28a02de | 2203 | while (size <= ralign / 2) |
1da177e4 LT |
2204 | ralign /= 2; |
2205 | } else { | |
2206 | ralign = BYTES_PER_WORD; | |
2207 | } | |
ca5f9703 PE |
2208 | |
2209 | /* | |
2210 | * Redzoning and user store require word alignment. Note this will be | |
2211 | * overridden by architecture or caller mandated alignment if either | |
2212 | * is greater than BYTES_PER_WORD. | |
2213 | */ | |
2214 | if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) | |
2215 | ralign = BYTES_PER_WORD; | |
2216 | ||
a44b56d3 | 2217 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2218 | if (ralign < ARCH_SLAB_MINALIGN) { |
2219 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2220 | } |
a44b56d3 | 2221 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2222 | if (ralign < align) { |
2223 | ralign = align; | |
1da177e4 | 2224 | } |
a44b56d3 KH |
2225 | /* disable debug if necessary */ |
2226 | if (ralign > BYTES_PER_WORD) | |
2227 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
a737b3e2 | 2228 | /* |
ca5f9703 | 2229 | * 4) Store it. |
1da177e4 LT |
2230 | */ |
2231 | align = ralign; | |
2232 | ||
2233 | /* Get cache's description obj. */ | |
e94b1766 | 2234 | cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); |
1da177e4 | 2235 | if (!cachep) |
4f12bb4f | 2236 | goto oops; |
1da177e4 LT |
2237 | |
2238 | #if DEBUG | |
3dafccf2 | 2239 | cachep->obj_size = size; |
1da177e4 | 2240 | |
ca5f9703 PE |
2241 | /* |
2242 | * Both debugging options require word-alignment which is calculated | |
2243 | * into align above. | |
2244 | */ | |
1da177e4 | 2245 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2246 | /* add space for red zone words */ |
3dafccf2 | 2247 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 2248 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
2249 | } |
2250 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 PE |
2251 | /* user store requires one word storage behind the end of |
2252 | * the real object. | |
1da177e4 | 2253 | */ |
1da177e4 LT |
2254 | size += BYTES_PER_WORD; |
2255 | } | |
2256 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2257 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2258 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2259 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2260 | size = PAGE_SIZE; |
2261 | } | |
2262 | #endif | |
2263 | #endif | |
2264 | ||
e0a42726 IM |
2265 | /* |
2266 | * Determine if the slab management is 'on' or 'off' slab. | |
2267 | * (bootstrapping cannot cope with offslab caches so don't do | |
2268 | * it too early on.) | |
2269 | */ | |
2270 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2271 | /* |
2272 | * Size is large, assume best to place the slab management obj | |
2273 | * off-slab (should allow better packing of objs). | |
2274 | */ | |
2275 | flags |= CFLGS_OFF_SLAB; | |
2276 | ||
2277 | size = ALIGN(size, align); | |
2278 | ||
f78bb8ad | 2279 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2280 | |
2281 | if (!cachep->num) { | |
2282 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
2283 | kmem_cache_free(&cache_cache, cachep); | |
2284 | cachep = NULL; | |
4f12bb4f | 2285 | goto oops; |
1da177e4 | 2286 | } |
b28a02de PE |
2287 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2288 | + sizeof(struct slab), align); | |
1da177e4 LT |
2289 | |
2290 | /* | |
2291 | * If the slab has been placed off-slab, and we have enough space then | |
2292 | * move it on-slab. This is at the expense of any extra colouring. | |
2293 | */ | |
2294 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2295 | flags &= ~CFLGS_OFF_SLAB; | |
2296 | left_over -= slab_size; | |
2297 | } | |
2298 | ||
2299 | if (flags & CFLGS_OFF_SLAB) { | |
2300 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2301 | slab_size = |
2302 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2303 | } |
2304 | ||
2305 | cachep->colour_off = cache_line_size(); | |
2306 | /* Offset must be a multiple of the alignment. */ | |
2307 | if (cachep->colour_off < align) | |
2308 | cachep->colour_off = align; | |
b28a02de | 2309 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2310 | cachep->slab_size = slab_size; |
2311 | cachep->flags = flags; | |
2312 | cachep->gfpflags = 0; | |
2313 | if (flags & SLAB_CACHE_DMA) | |
2314 | cachep->gfpflags |= GFP_DMA; | |
3dafccf2 | 2315 | cachep->buffer_size = size; |
1da177e4 | 2316 | |
e5ac9c5a | 2317 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2318 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2319 | /* |
2320 | * This is a possibility for one of the malloc_sizes caches. | |
2321 | * But since we go off slab only for object size greater than | |
2322 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2323 | * this should not happen at all. | |
2324 | * But leave a BUG_ON for some lucky dude. | |
2325 | */ | |
2326 | BUG_ON(!cachep->slabp_cache); | |
2327 | } | |
1da177e4 LT |
2328 | cachep->ctor = ctor; |
2329 | cachep->dtor = dtor; | |
2330 | cachep->name = name; | |
2331 | ||
2ed3a4ef CL |
2332 | if (setup_cpu_cache(cachep)) { |
2333 | __kmem_cache_destroy(cachep); | |
2334 | cachep = NULL; | |
2335 | goto oops; | |
2336 | } | |
1da177e4 | 2337 | |
1da177e4 LT |
2338 | /* cache setup completed, link it into the list */ |
2339 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2340 | oops: |
1da177e4 LT |
2341 | if (!cachep && (flags & SLAB_PANIC)) |
2342 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2343 | name); |
fc0abb14 | 2344 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2345 | return cachep; |
2346 | } | |
2347 | EXPORT_SYMBOL(kmem_cache_create); | |
2348 | ||
2349 | #if DEBUG | |
2350 | static void check_irq_off(void) | |
2351 | { | |
2352 | BUG_ON(!irqs_disabled()); | |
2353 | } | |
2354 | ||
2355 | static void check_irq_on(void) | |
2356 | { | |
2357 | BUG_ON(irqs_disabled()); | |
2358 | } | |
2359 | ||
343e0d7a | 2360 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2361 | { |
2362 | #ifdef CONFIG_SMP | |
2363 | check_irq_off(); | |
e498be7d | 2364 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2365 | #endif |
2366 | } | |
e498be7d | 2367 | |
343e0d7a | 2368 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2369 | { |
2370 | #ifdef CONFIG_SMP | |
2371 | check_irq_off(); | |
2372 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2373 | #endif | |
2374 | } | |
2375 | ||
1da177e4 LT |
2376 | #else |
2377 | #define check_irq_off() do { } while(0) | |
2378 | #define check_irq_on() do { } while(0) | |
2379 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2380 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2381 | #endif |
2382 | ||
aab2207c CL |
2383 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2384 | struct array_cache *ac, | |
2385 | int force, int node); | |
2386 | ||
1da177e4 LT |
2387 | static void do_drain(void *arg) |
2388 | { | |
a737b3e2 | 2389 | struct kmem_cache *cachep = arg; |
1da177e4 | 2390 | struct array_cache *ac; |
ff69416e | 2391 | int node = numa_node_id(); |
1da177e4 LT |
2392 | |
2393 | check_irq_off(); | |
9a2dba4b | 2394 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2395 | spin_lock(&cachep->nodelists[node]->list_lock); |
2396 | free_block(cachep, ac->entry, ac->avail, node); | |
2397 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2398 | ac->avail = 0; |
2399 | } | |
2400 | ||
343e0d7a | 2401 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2402 | { |
e498be7d CL |
2403 | struct kmem_list3 *l3; |
2404 | int node; | |
2405 | ||
a07fa394 | 2406 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2407 | check_irq_on(); |
b28a02de | 2408 | for_each_online_node(node) { |
e498be7d | 2409 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2410 | if (l3 && l3->alien) |
2411 | drain_alien_cache(cachep, l3->alien); | |
2412 | } | |
2413 | ||
2414 | for_each_online_node(node) { | |
2415 | l3 = cachep->nodelists[node]; | |
2416 | if (l3) | |
aab2207c | 2417 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2418 | } |
1da177e4 LT |
2419 | } |
2420 | ||
ed11d9eb CL |
2421 | /* |
2422 | * Remove slabs from the list of free slabs. | |
2423 | * Specify the number of slabs to drain in tofree. | |
2424 | * | |
2425 | * Returns the actual number of slabs released. | |
2426 | */ | |
2427 | static int drain_freelist(struct kmem_cache *cache, | |
2428 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2429 | { |
ed11d9eb CL |
2430 | struct list_head *p; |
2431 | int nr_freed; | |
1da177e4 | 2432 | struct slab *slabp; |
1da177e4 | 2433 | |
ed11d9eb CL |
2434 | nr_freed = 0; |
2435 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2436 | |
ed11d9eb | 2437 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2438 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2439 | if (p == &l3->slabs_free) { |
2440 | spin_unlock_irq(&l3->list_lock); | |
2441 | goto out; | |
2442 | } | |
1da177e4 | 2443 | |
ed11d9eb | 2444 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2445 | #if DEBUG |
40094fa6 | 2446 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2447 | #endif |
2448 | list_del(&slabp->list); | |
ed11d9eb CL |
2449 | /* |
2450 | * Safe to drop the lock. The slab is no longer linked | |
2451 | * to the cache. | |
2452 | */ | |
2453 | l3->free_objects -= cache->num; | |
e498be7d | 2454 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2455 | slab_destroy(cache, slabp); |
2456 | nr_freed++; | |
1da177e4 | 2457 | } |
ed11d9eb CL |
2458 | out: |
2459 | return nr_freed; | |
1da177e4 LT |
2460 | } |
2461 | ||
8f5be20b | 2462 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2463 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2464 | { |
2465 | int ret = 0, i = 0; | |
2466 | struct kmem_list3 *l3; | |
2467 | ||
2468 | drain_cpu_caches(cachep); | |
2469 | ||
2470 | check_irq_on(); | |
2471 | for_each_online_node(i) { | |
2472 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2473 | if (!l3) |
2474 | continue; | |
2475 | ||
2476 | drain_freelist(cachep, l3, l3->free_objects); | |
2477 | ||
2478 | ret += !list_empty(&l3->slabs_full) || | |
2479 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2480 | } |
2481 | return (ret ? 1 : 0); | |
2482 | } | |
2483 | ||
1da177e4 LT |
2484 | /** |
2485 | * kmem_cache_shrink - Shrink a cache. | |
2486 | * @cachep: The cache to shrink. | |
2487 | * | |
2488 | * Releases as many slabs as possible for a cache. | |
2489 | * To help debugging, a zero exit status indicates all slabs were released. | |
2490 | */ | |
343e0d7a | 2491 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2492 | { |
8f5be20b | 2493 | int ret; |
40094fa6 | 2494 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2495 | |
8f5be20b RT |
2496 | mutex_lock(&cache_chain_mutex); |
2497 | ret = __cache_shrink(cachep); | |
2498 | mutex_unlock(&cache_chain_mutex); | |
2499 | return ret; | |
1da177e4 LT |
2500 | } |
2501 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2502 | ||
2503 | /** | |
2504 | * kmem_cache_destroy - delete a cache | |
2505 | * @cachep: the cache to destroy | |
2506 | * | |
343e0d7a | 2507 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2508 | * |
2509 | * It is expected this function will be called by a module when it is | |
2510 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2511 | * cache being allocated each time a module is loaded and unloaded, if the | |
2512 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2513 | * | |
2514 | * The cache must be empty before calling this function. | |
2515 | * | |
2516 | * The caller must guarantee that noone will allocate memory from the cache | |
2517 | * during the kmem_cache_destroy(). | |
2518 | */ | |
133d205a | 2519 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2520 | { |
40094fa6 | 2521 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2522 | |
1da177e4 | 2523 | /* Find the cache in the chain of caches. */ |
fc0abb14 | 2524 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2525 | /* |
2526 | * the chain is never empty, cache_cache is never destroyed | |
2527 | */ | |
2528 | list_del(&cachep->next); | |
1da177e4 LT |
2529 | if (__cache_shrink(cachep)) { |
2530 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2531 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2532 | mutex_unlock(&cache_chain_mutex); |
133d205a | 2533 | return; |
1da177e4 LT |
2534 | } |
2535 | ||
2536 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2537 | synchronize_rcu(); |
1da177e4 | 2538 | |
117f6eb1 | 2539 | __kmem_cache_destroy(cachep); |
8f5be20b | 2540 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2541 | } |
2542 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2543 | ||
e5ac9c5a RT |
2544 | /* |
2545 | * Get the memory for a slab management obj. | |
2546 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2547 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2548 | * come from the same cache which is getting created because, | |
2549 | * when we are searching for an appropriate cache for these | |
2550 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2551 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2552 | * kmem_find_general_cachep till the initialization is complete. | |
2553 | * Hence we cannot have slabp_cache same as the original cache. | |
2554 | */ | |
343e0d7a | 2555 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2556 | int colour_off, gfp_t local_flags, |
2557 | int nodeid) | |
1da177e4 LT |
2558 | { |
2559 | struct slab *slabp; | |
b28a02de | 2560 | |
1da177e4 LT |
2561 | if (OFF_SLAB(cachep)) { |
2562 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2563 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
3c517a61 | 2564 | local_flags & ~GFP_THISNODE, nodeid); |
1da177e4 LT |
2565 | if (!slabp) |
2566 | return NULL; | |
2567 | } else { | |
b28a02de | 2568 | slabp = objp + colour_off; |
1da177e4 LT |
2569 | colour_off += cachep->slab_size; |
2570 | } | |
2571 | slabp->inuse = 0; | |
2572 | slabp->colouroff = colour_off; | |
b28a02de | 2573 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2574 | slabp->nodeid = nodeid; |
1da177e4 LT |
2575 | return slabp; |
2576 | } | |
2577 | ||
2578 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2579 | { | |
b28a02de | 2580 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2581 | } |
2582 | ||
343e0d7a | 2583 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2584 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2585 | { |
2586 | int i; | |
2587 | ||
2588 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2589 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2590 | #if DEBUG |
2591 | /* need to poison the objs? */ | |
2592 | if (cachep->flags & SLAB_POISON) | |
2593 | poison_obj(cachep, objp, POISON_FREE); | |
2594 | if (cachep->flags & SLAB_STORE_USER) | |
2595 | *dbg_userword(cachep, objp) = NULL; | |
2596 | ||
2597 | if (cachep->flags & SLAB_RED_ZONE) { | |
2598 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2599 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2600 | } | |
2601 | /* | |
a737b3e2 AM |
2602 | * Constructors are not allowed to allocate memory from the same |
2603 | * cache which they are a constructor for. Otherwise, deadlock. | |
2604 | * They must also be threaded. | |
1da177e4 LT |
2605 | */ |
2606 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2607 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2608 | ctor_flags); |
1da177e4 LT |
2609 | |
2610 | if (cachep->flags & SLAB_RED_ZONE) { | |
2611 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2612 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2613 | " end of an object"); |
1da177e4 LT |
2614 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2615 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2616 | " start of an object"); |
1da177e4 | 2617 | } |
a737b3e2 AM |
2618 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2619 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2620 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2621 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2622 | #else |
2623 | if (cachep->ctor) | |
2624 | cachep->ctor(objp, cachep, ctor_flags); | |
2625 | #endif | |
b28a02de | 2626 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2627 | } |
b28a02de | 2628 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2629 | slabp->free = 0; |
2630 | } | |
2631 | ||
343e0d7a | 2632 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2633 | { |
441e143e | 2634 | if (flags & GFP_DMA) |
a737b3e2 AM |
2635 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); |
2636 | else | |
2637 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
1da177e4 LT |
2638 | } |
2639 | ||
a737b3e2 AM |
2640 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2641 | int nodeid) | |
78d382d7 | 2642 | { |
8fea4e96 | 2643 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2644 | kmem_bufctl_t next; |
2645 | ||
2646 | slabp->inuse++; | |
2647 | next = slab_bufctl(slabp)[slabp->free]; | |
2648 | #if DEBUG | |
2649 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2650 | WARN_ON(slabp->nodeid != nodeid); | |
2651 | #endif | |
2652 | slabp->free = next; | |
2653 | ||
2654 | return objp; | |
2655 | } | |
2656 | ||
a737b3e2 AM |
2657 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2658 | void *objp, int nodeid) | |
78d382d7 | 2659 | { |
8fea4e96 | 2660 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2661 | |
2662 | #if DEBUG | |
2663 | /* Verify that the slab belongs to the intended node */ | |
2664 | WARN_ON(slabp->nodeid != nodeid); | |
2665 | ||
871751e2 | 2666 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2667 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2668 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2669 | BUG(); |
2670 | } | |
2671 | #endif | |
2672 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2673 | slabp->free = objnr; | |
2674 | slabp->inuse--; | |
2675 | } | |
2676 | ||
4776874f PE |
2677 | /* |
2678 | * Map pages beginning at addr to the given cache and slab. This is required | |
2679 | * for the slab allocator to be able to lookup the cache and slab of a | |
2680 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2681 | */ | |
2682 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2683 | void *addr) | |
1da177e4 | 2684 | { |
4776874f | 2685 | int nr_pages; |
1da177e4 LT |
2686 | struct page *page; |
2687 | ||
4776874f | 2688 | page = virt_to_page(addr); |
84097518 | 2689 | |
4776874f | 2690 | nr_pages = 1; |
84097518 | 2691 | if (likely(!PageCompound(page))) |
4776874f PE |
2692 | nr_pages <<= cache->gfporder; |
2693 | ||
1da177e4 | 2694 | do { |
4776874f PE |
2695 | page_set_cache(page, cache); |
2696 | page_set_slab(page, slab); | |
1da177e4 | 2697 | page++; |
4776874f | 2698 | } while (--nr_pages); |
1da177e4 LT |
2699 | } |
2700 | ||
2701 | /* | |
2702 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2703 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2704 | */ | |
3c517a61 CL |
2705 | static int cache_grow(struct kmem_cache *cachep, |
2706 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2707 | { |
b28a02de | 2708 | struct slab *slabp; |
b28a02de PE |
2709 | size_t offset; |
2710 | gfp_t local_flags; | |
2711 | unsigned long ctor_flags; | |
e498be7d | 2712 | struct kmem_list3 *l3; |
1da177e4 | 2713 | |
a737b3e2 AM |
2714 | /* |
2715 | * Be lazy and only check for valid flags here, keeping it out of the | |
2716 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2717 | */ |
441e143e | 2718 | BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK | __GFP_NO_GROW)); |
6e0eaa4b | 2719 | if (flags & __GFP_NO_GROW) |
1da177e4 LT |
2720 | return 0; |
2721 | ||
2722 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
a06d72c1 | 2723 | local_flags = (flags & GFP_LEVEL_MASK); |
1da177e4 LT |
2724 | if (!(local_flags & __GFP_WAIT)) |
2725 | /* | |
2726 | * Not allowed to sleep. Need to tell a constructor about | |
2727 | * this - it might need to know... | |
2728 | */ | |
2729 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2730 | ||
2e1217cf | 2731 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2732 | check_irq_off(); |
2e1217cf RT |
2733 | l3 = cachep->nodelists[nodeid]; |
2734 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2735 | |
2736 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2737 | offset = l3->colour_next; |
2738 | l3->colour_next++; | |
2739 | if (l3->colour_next >= cachep->colour) | |
2740 | l3->colour_next = 0; | |
2741 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2742 | |
2e1217cf | 2743 | offset *= cachep->colour_off; |
1da177e4 LT |
2744 | |
2745 | if (local_flags & __GFP_WAIT) | |
2746 | local_irq_enable(); | |
2747 | ||
2748 | /* | |
2749 | * The test for missing atomic flag is performed here, rather than | |
2750 | * the more obvious place, simply to reduce the critical path length | |
2751 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2752 | * will eventually be caught here (where it matters). | |
2753 | */ | |
2754 | kmem_flagcheck(cachep, flags); | |
2755 | ||
a737b3e2 AM |
2756 | /* |
2757 | * Get mem for the objs. Attempt to allocate a physical page from | |
2758 | * 'nodeid'. | |
e498be7d | 2759 | */ |
3c517a61 CL |
2760 | if (!objp) |
2761 | objp = kmem_getpages(cachep, flags, nodeid); | |
a737b3e2 | 2762 | if (!objp) |
1da177e4 LT |
2763 | goto failed; |
2764 | ||
2765 | /* Get slab management. */ | |
3c517a61 CL |
2766 | slabp = alloc_slabmgmt(cachep, objp, offset, |
2767 | local_flags & ~GFP_THISNODE, nodeid); | |
a737b3e2 | 2768 | if (!slabp) |
1da177e4 LT |
2769 | goto opps1; |
2770 | ||
e498be7d | 2771 | slabp->nodeid = nodeid; |
4776874f | 2772 | slab_map_pages(cachep, slabp, objp); |
1da177e4 LT |
2773 | |
2774 | cache_init_objs(cachep, slabp, ctor_flags); | |
2775 | ||
2776 | if (local_flags & __GFP_WAIT) | |
2777 | local_irq_disable(); | |
2778 | check_irq_off(); | |
e498be7d | 2779 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2780 | |
2781 | /* Make slab active. */ | |
e498be7d | 2782 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2783 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2784 | l3->free_objects += cachep->num; |
2785 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2786 | return 1; |
a737b3e2 | 2787 | opps1: |
1da177e4 | 2788 | kmem_freepages(cachep, objp); |
a737b3e2 | 2789 | failed: |
1da177e4 LT |
2790 | if (local_flags & __GFP_WAIT) |
2791 | local_irq_disable(); | |
2792 | return 0; | |
2793 | } | |
2794 | ||
2795 | #if DEBUG | |
2796 | ||
2797 | /* | |
2798 | * Perform extra freeing checks: | |
2799 | * - detect bad pointers. | |
2800 | * - POISON/RED_ZONE checking | |
2801 | * - destructor calls, for caches with POISON+dtor | |
2802 | */ | |
2803 | static void kfree_debugcheck(const void *objp) | |
2804 | { | |
2805 | struct page *page; | |
2806 | ||
2807 | if (!virt_addr_valid(objp)) { | |
2808 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2809 | (unsigned long)objp); |
2810 | BUG(); | |
1da177e4 LT |
2811 | } |
2812 | page = virt_to_page(objp); | |
2813 | if (!PageSlab(page)) { | |
b28a02de PE |
2814 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2815 | (unsigned long)objp); | |
1da177e4 LT |
2816 | BUG(); |
2817 | } | |
2818 | } | |
2819 | ||
58ce1fd5 PE |
2820 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2821 | { | |
2822 | unsigned long redzone1, redzone2; | |
2823 | ||
2824 | redzone1 = *dbg_redzone1(cache, obj); | |
2825 | redzone2 = *dbg_redzone2(cache, obj); | |
2826 | ||
2827 | /* | |
2828 | * Redzone is ok. | |
2829 | */ | |
2830 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2831 | return; | |
2832 | ||
2833 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2834 | slab_error(cache, "double free detected"); | |
2835 | else | |
2836 | slab_error(cache, "memory outside object was overwritten"); | |
2837 | ||
2838 | printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n", | |
2839 | obj, redzone1, redzone2); | |
2840 | } | |
2841 | ||
343e0d7a | 2842 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2843 | void *caller) |
1da177e4 LT |
2844 | { |
2845 | struct page *page; | |
2846 | unsigned int objnr; | |
2847 | struct slab *slabp; | |
2848 | ||
3dafccf2 | 2849 | objp -= obj_offset(cachep); |
1da177e4 LT |
2850 | kfree_debugcheck(objp); |
2851 | page = virt_to_page(objp); | |
2852 | ||
065d41cb | 2853 | slabp = page_get_slab(page); |
1da177e4 LT |
2854 | |
2855 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2856 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2857 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2858 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2859 | } | |
2860 | if (cachep->flags & SLAB_STORE_USER) | |
2861 | *dbg_userword(cachep, objp) = caller; | |
2862 | ||
8fea4e96 | 2863 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2864 | |
2865 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2866 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2867 | |
2868 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
a737b3e2 AM |
2869 | /* |
2870 | * Need to call the slab's constructor so the caller can | |
2871 | * perform a verify of its state (debugging). Called without | |
2872 | * the cache-lock held. | |
1da177e4 | 2873 | */ |
3dafccf2 | 2874 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2875 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2876 | } |
2877 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2878 | /* we want to cache poison the object, | |
2879 | * call the destruction callback | |
2880 | */ | |
3dafccf2 | 2881 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 2882 | } |
871751e2 AV |
2883 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2884 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2885 | #endif | |
1da177e4 LT |
2886 | if (cachep->flags & SLAB_POISON) { |
2887 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2888 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2889 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2890 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2891 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2892 | } else { |
2893 | poison_obj(cachep, objp, POISON_FREE); | |
2894 | } | |
2895 | #else | |
2896 | poison_obj(cachep, objp, POISON_FREE); | |
2897 | #endif | |
2898 | } | |
2899 | return objp; | |
2900 | } | |
2901 | ||
343e0d7a | 2902 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2903 | { |
2904 | kmem_bufctl_t i; | |
2905 | int entries = 0; | |
b28a02de | 2906 | |
1da177e4 LT |
2907 | /* Check slab's freelist to see if this obj is there. */ |
2908 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2909 | entries++; | |
2910 | if (entries > cachep->num || i >= cachep->num) | |
2911 | goto bad; | |
2912 | } | |
2913 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2914 | bad: |
2915 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2916 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2917 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2918 | for (i = 0; |
264132bc | 2919 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2920 | i++) { |
a737b3e2 | 2921 | if (i % 16 == 0) |
1da177e4 | 2922 | printk("\n%03x:", i); |
b28a02de | 2923 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2924 | } |
2925 | printk("\n"); | |
2926 | BUG(); | |
2927 | } | |
2928 | } | |
2929 | #else | |
2930 | #define kfree_debugcheck(x) do { } while(0) | |
2931 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2932 | #define check_slabp(x,y) do { } while(0) | |
2933 | #endif | |
2934 | ||
343e0d7a | 2935 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2936 | { |
2937 | int batchcount; | |
2938 | struct kmem_list3 *l3; | |
2939 | struct array_cache *ac; | |
1ca4cb24 PE |
2940 | int node; |
2941 | ||
2942 | node = numa_node_id(); | |
1da177e4 LT |
2943 | |
2944 | check_irq_off(); | |
9a2dba4b | 2945 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2946 | retry: |
1da177e4 LT |
2947 | batchcount = ac->batchcount; |
2948 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2949 | /* |
2950 | * If there was little recent activity on this cache, then | |
2951 | * perform only a partial refill. Otherwise we could generate | |
2952 | * refill bouncing. | |
1da177e4 LT |
2953 | */ |
2954 | batchcount = BATCHREFILL_LIMIT; | |
2955 | } | |
1ca4cb24 | 2956 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2957 | |
2958 | BUG_ON(ac->avail > 0 || !l3); | |
2959 | spin_lock(&l3->list_lock); | |
1da177e4 | 2960 | |
3ded175a CL |
2961 | /* See if we can refill from the shared array */ |
2962 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2963 | goto alloc_done; | |
2964 | ||
1da177e4 LT |
2965 | while (batchcount > 0) { |
2966 | struct list_head *entry; | |
2967 | struct slab *slabp; | |
2968 | /* Get slab alloc is to come from. */ | |
2969 | entry = l3->slabs_partial.next; | |
2970 | if (entry == &l3->slabs_partial) { | |
2971 | l3->free_touched = 1; | |
2972 | entry = l3->slabs_free.next; | |
2973 | if (entry == &l3->slabs_free) | |
2974 | goto must_grow; | |
2975 | } | |
2976 | ||
2977 | slabp = list_entry(entry, struct slab, list); | |
2978 | check_slabp(cachep, slabp); | |
2979 | check_spinlock_acquired(cachep); | |
2980 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2981 | STATS_INC_ALLOCED(cachep); |
2982 | STATS_INC_ACTIVE(cachep); | |
2983 | STATS_SET_HIGH(cachep); | |
2984 | ||
78d382d7 | 2985 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 2986 | node); |
1da177e4 LT |
2987 | } |
2988 | check_slabp(cachep, slabp); | |
2989 | ||
2990 | /* move slabp to correct slabp list: */ | |
2991 | list_del(&slabp->list); | |
2992 | if (slabp->free == BUFCTL_END) | |
2993 | list_add(&slabp->list, &l3->slabs_full); | |
2994 | else | |
2995 | list_add(&slabp->list, &l3->slabs_partial); | |
2996 | } | |
2997 | ||
a737b3e2 | 2998 | must_grow: |
1da177e4 | 2999 | l3->free_objects -= ac->avail; |
a737b3e2 | 3000 | alloc_done: |
e498be7d | 3001 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3002 | |
3003 | if (unlikely(!ac->avail)) { | |
3004 | int x; | |
3c517a61 | 3005 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3006 | |
a737b3e2 | 3007 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3008 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3009 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3010 | return NULL; |
3011 | ||
a737b3e2 | 3012 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3013 | goto retry; |
3014 | } | |
3015 | ac->touched = 1; | |
e498be7d | 3016 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3017 | } |
3018 | ||
a737b3e2 AM |
3019 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3020 | gfp_t flags) | |
1da177e4 LT |
3021 | { |
3022 | might_sleep_if(flags & __GFP_WAIT); | |
3023 | #if DEBUG | |
3024 | kmem_flagcheck(cachep, flags); | |
3025 | #endif | |
3026 | } | |
3027 | ||
3028 | #if DEBUG | |
a737b3e2 AM |
3029 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3030 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3031 | { |
b28a02de | 3032 | if (!objp) |
1da177e4 | 3033 | return objp; |
b28a02de | 3034 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3035 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3036 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3037 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3038 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3039 | else |
3040 | check_poison_obj(cachep, objp); | |
3041 | #else | |
3042 | check_poison_obj(cachep, objp); | |
3043 | #endif | |
3044 | poison_obj(cachep, objp, POISON_INUSE); | |
3045 | } | |
3046 | if (cachep->flags & SLAB_STORE_USER) | |
3047 | *dbg_userword(cachep, objp) = caller; | |
3048 | ||
3049 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3050 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3051 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3052 | slab_error(cachep, "double free, or memory outside" | |
3053 | " object was overwritten"); | |
b28a02de | 3054 | printk(KERN_ERR |
a737b3e2 AM |
3055 | "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", |
3056 | objp, *dbg_redzone1(cachep, objp), | |
3057 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3058 | } |
3059 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3060 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3061 | } | |
871751e2 AV |
3062 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3063 | { | |
3064 | struct slab *slabp; | |
3065 | unsigned objnr; | |
3066 | ||
3067 | slabp = page_get_slab(virt_to_page(objp)); | |
3068 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; | |
3069 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3070 | } | |
3071 | #endif | |
3dafccf2 | 3072 | objp += obj_offset(cachep); |
1da177e4 | 3073 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 3074 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
3075 | |
3076 | if (!(flags & __GFP_WAIT)) | |
3077 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
3078 | ||
3079 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 3080 | } |
a44b56d3 KH |
3081 | #if ARCH_SLAB_MINALIGN |
3082 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3083 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3084 | objp, ARCH_SLAB_MINALIGN); | |
3085 | } | |
3086 | #endif | |
1da177e4 LT |
3087 | return objp; |
3088 | } | |
3089 | #else | |
3090 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3091 | #endif | |
3092 | ||
8a8b6502 AM |
3093 | #ifdef CONFIG_FAILSLAB |
3094 | ||
3095 | static struct failslab_attr { | |
3096 | ||
3097 | struct fault_attr attr; | |
3098 | ||
3099 | u32 ignore_gfp_wait; | |
3100 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3101 | struct dentry *ignore_gfp_wait_file; | |
3102 | #endif | |
3103 | ||
3104 | } failslab = { | |
3105 | .attr = FAULT_ATTR_INITIALIZER, | |
6b1b60f4 | 3106 | .ignore_gfp_wait = 1, |
8a8b6502 AM |
3107 | }; |
3108 | ||
3109 | static int __init setup_failslab(char *str) | |
3110 | { | |
3111 | return setup_fault_attr(&failslab.attr, str); | |
3112 | } | |
3113 | __setup("failslab=", setup_failslab); | |
3114 | ||
3115 | static int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3116 | { | |
3117 | if (cachep == &cache_cache) | |
3118 | return 0; | |
3119 | if (flags & __GFP_NOFAIL) | |
3120 | return 0; | |
3121 | if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) | |
3122 | return 0; | |
3123 | ||
3124 | return should_fail(&failslab.attr, obj_size(cachep)); | |
3125 | } | |
3126 | ||
3127 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3128 | ||
3129 | static int __init failslab_debugfs(void) | |
3130 | { | |
3131 | mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
3132 | struct dentry *dir; | |
3133 | int err; | |
3134 | ||
3135 | err = init_fault_attr_dentries(&failslab.attr, "failslab"); | |
3136 | if (err) | |
3137 | return err; | |
3138 | dir = failslab.attr.dentries.dir; | |
3139 | ||
3140 | failslab.ignore_gfp_wait_file = | |
3141 | debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
3142 | &failslab.ignore_gfp_wait); | |
3143 | ||
3144 | if (!failslab.ignore_gfp_wait_file) { | |
3145 | err = -ENOMEM; | |
3146 | debugfs_remove(failslab.ignore_gfp_wait_file); | |
3147 | cleanup_fault_attr_dentries(&failslab.attr); | |
3148 | } | |
3149 | ||
3150 | return err; | |
3151 | } | |
3152 | ||
3153 | late_initcall(failslab_debugfs); | |
3154 | ||
3155 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
3156 | ||
3157 | #else /* CONFIG_FAILSLAB */ | |
3158 | ||
3159 | static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3160 | { | |
3161 | return 0; | |
3162 | } | |
3163 | ||
3164 | #endif /* CONFIG_FAILSLAB */ | |
3165 | ||
343e0d7a | 3166 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3167 | { |
b28a02de | 3168 | void *objp; |
1da177e4 LT |
3169 | struct array_cache *ac; |
3170 | ||
5c382300 | 3171 | check_irq_off(); |
8a8b6502 AM |
3172 | |
3173 | if (should_failslab(cachep, flags)) | |
3174 | return NULL; | |
3175 | ||
9a2dba4b | 3176 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3177 | if (likely(ac->avail)) { |
3178 | STATS_INC_ALLOCHIT(cachep); | |
3179 | ac->touched = 1; | |
e498be7d | 3180 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3181 | } else { |
3182 | STATS_INC_ALLOCMISS(cachep); | |
3183 | objp = cache_alloc_refill(cachep, flags); | |
3184 | } | |
5c382300 AK |
3185 | return objp; |
3186 | } | |
3187 | ||
a737b3e2 AM |
3188 | static __always_inline void *__cache_alloc(struct kmem_cache *cachep, |
3189 | gfp_t flags, void *caller) | |
5c382300 AK |
3190 | { |
3191 | unsigned long save_flags; | |
de3083ec | 3192 | void *objp = NULL; |
5c382300 AK |
3193 | |
3194 | cache_alloc_debugcheck_before(cachep, flags); | |
3195 | ||
3196 | local_irq_save(save_flags); | |
de3083ec | 3197 | |
765c4507 CL |
3198 | if (unlikely(NUMA_BUILD && |
3199 | current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) | |
de3083ec | 3200 | objp = alternate_node_alloc(cachep, flags); |
de3083ec CL |
3201 | |
3202 | if (!objp) | |
3203 | objp = ____cache_alloc(cachep, flags); | |
765c4507 CL |
3204 | /* |
3205 | * We may just have run out of memory on the local node. | |
8b98c169 | 3206 | * ____cache_alloc_node() knows how to locate memory on other nodes |
765c4507 CL |
3207 | */ |
3208 | if (NUMA_BUILD && !objp) | |
8b98c169 | 3209 | objp = ____cache_alloc_node(cachep, flags, numa_node_id()); |
1da177e4 | 3210 | local_irq_restore(save_flags); |
34342e86 | 3211 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 3212 | caller); |
34342e86 | 3213 | prefetchw(objp); |
1da177e4 LT |
3214 | return objp; |
3215 | } | |
3216 | ||
e498be7d | 3217 | #ifdef CONFIG_NUMA |
c61afb18 | 3218 | /* |
b2455396 | 3219 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3220 | * |
3221 | * If we are in_interrupt, then process context, including cpusets and | |
3222 | * mempolicy, may not apply and should not be used for allocation policy. | |
3223 | */ | |
3224 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3225 | { | |
3226 | int nid_alloc, nid_here; | |
3227 | ||
765c4507 | 3228 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3229 | return NULL; |
3230 | nid_alloc = nid_here = numa_node_id(); | |
3231 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3232 | nid_alloc = cpuset_mem_spread_node(); | |
3233 | else if (current->mempolicy) | |
3234 | nid_alloc = slab_node(current->mempolicy); | |
3235 | if (nid_alloc != nid_here) | |
8b98c169 | 3236 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3237 | return NULL; |
3238 | } | |
3239 | ||
765c4507 CL |
3240 | /* |
3241 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3242 | * certain node and fall back is permitted. First we scan all the |
3243 | * available nodelists for available objects. If that fails then we | |
3244 | * perform an allocation without specifying a node. This allows the page | |
3245 | * allocator to do its reclaim / fallback magic. We then insert the | |
3246 | * slab into the proper nodelist and then allocate from it. | |
765c4507 CL |
3247 | */ |
3248 | void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) | |
3249 | { | |
3250 | struct zonelist *zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
3251 | ->node_zonelists[gfp_zone(flags)]; | |
3252 | struct zone **z; | |
3253 | void *obj = NULL; | |
3c517a61 | 3254 | int nid; |
765c4507 | 3255 | |
3c517a61 CL |
3256 | retry: |
3257 | /* | |
3258 | * Look through allowed nodes for objects available | |
3259 | * from existing per node queues. | |
3260 | */ | |
aedb0eb1 | 3261 | for (z = zonelist->zones; *z && !obj; z++) { |
3c517a61 | 3262 | nid = zone_to_nid(*z); |
aedb0eb1 | 3263 | |
b8b50b65 | 3264 | if (cpuset_zone_allowed(*z, flags | __GFP_HARDWALL) && |
3c517a61 CL |
3265 | cache->nodelists[nid] && |
3266 | cache->nodelists[nid]->free_objects) | |
3267 | obj = ____cache_alloc_node(cache, | |
3268 | flags | GFP_THISNODE, nid); | |
3269 | } | |
3270 | ||
3271 | if (!obj) { | |
3272 | /* | |
3273 | * This allocation will be performed within the constraints | |
3274 | * of the current cpuset / memory policy requirements. | |
3275 | * We may trigger various forms of reclaim on the allowed | |
3276 | * set and go into memory reserves if necessary. | |
3277 | */ | |
3278 | obj = kmem_getpages(cache, flags, -1); | |
3279 | if (obj) { | |
3280 | /* | |
3281 | * Insert into the appropriate per node queues | |
3282 | */ | |
3283 | nid = page_to_nid(virt_to_page(obj)); | |
3284 | if (cache_grow(cache, flags, nid, obj)) { | |
3285 | obj = ____cache_alloc_node(cache, | |
3286 | flags | GFP_THISNODE, nid); | |
3287 | if (!obj) | |
3288 | /* | |
3289 | * Another processor may allocate the | |
3290 | * objects in the slab since we are | |
3291 | * not holding any locks. | |
3292 | */ | |
3293 | goto retry; | |
3294 | } else { | |
3295 | kmem_freepages(cache, obj); | |
3296 | obj = NULL; | |
3297 | } | |
3298 | } | |
aedb0eb1 | 3299 | } |
765c4507 CL |
3300 | return obj; |
3301 | } | |
3302 | ||
e498be7d CL |
3303 | /* |
3304 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3305 | */ |
8b98c169 | 3306 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3307 | int nodeid) |
e498be7d CL |
3308 | { |
3309 | struct list_head *entry; | |
b28a02de PE |
3310 | struct slab *slabp; |
3311 | struct kmem_list3 *l3; | |
3312 | void *obj; | |
b28a02de PE |
3313 | int x; |
3314 | ||
3315 | l3 = cachep->nodelists[nodeid]; | |
3316 | BUG_ON(!l3); | |
3317 | ||
a737b3e2 | 3318 | retry: |
ca3b9b91 | 3319 | check_irq_off(); |
b28a02de PE |
3320 | spin_lock(&l3->list_lock); |
3321 | entry = l3->slabs_partial.next; | |
3322 | if (entry == &l3->slabs_partial) { | |
3323 | l3->free_touched = 1; | |
3324 | entry = l3->slabs_free.next; | |
3325 | if (entry == &l3->slabs_free) | |
3326 | goto must_grow; | |
3327 | } | |
3328 | ||
3329 | slabp = list_entry(entry, struct slab, list); | |
3330 | check_spinlock_acquired_node(cachep, nodeid); | |
3331 | check_slabp(cachep, slabp); | |
3332 | ||
3333 | STATS_INC_NODEALLOCS(cachep); | |
3334 | STATS_INC_ACTIVE(cachep); | |
3335 | STATS_SET_HIGH(cachep); | |
3336 | ||
3337 | BUG_ON(slabp->inuse == cachep->num); | |
3338 | ||
78d382d7 | 3339 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3340 | check_slabp(cachep, slabp); |
3341 | l3->free_objects--; | |
3342 | /* move slabp to correct slabp list: */ | |
3343 | list_del(&slabp->list); | |
3344 | ||
a737b3e2 | 3345 | if (slabp->free == BUFCTL_END) |
b28a02de | 3346 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3347 | else |
b28a02de | 3348 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3349 | |
b28a02de PE |
3350 | spin_unlock(&l3->list_lock); |
3351 | goto done; | |
e498be7d | 3352 | |
a737b3e2 | 3353 | must_grow: |
b28a02de | 3354 | spin_unlock(&l3->list_lock); |
3c517a61 | 3355 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3356 | if (x) |
3357 | goto retry; | |
1da177e4 | 3358 | |
765c4507 CL |
3359 | if (!(flags & __GFP_THISNODE)) |
3360 | /* Unable to grow the cache. Fall back to other nodes. */ | |
3361 | return fallback_alloc(cachep, flags); | |
3362 | ||
3363 | return NULL; | |
e498be7d | 3364 | |
a737b3e2 | 3365 | done: |
b28a02de | 3366 | return obj; |
e498be7d CL |
3367 | } |
3368 | #endif | |
3369 | ||
3370 | /* | |
3371 | * Caller needs to acquire correct kmem_list's list_lock | |
3372 | */ | |
343e0d7a | 3373 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3374 | int node) |
1da177e4 LT |
3375 | { |
3376 | int i; | |
e498be7d | 3377 | struct kmem_list3 *l3; |
1da177e4 LT |
3378 | |
3379 | for (i = 0; i < nr_objects; i++) { | |
3380 | void *objp = objpp[i]; | |
3381 | struct slab *slabp; | |
1da177e4 | 3382 | |
6ed5eb22 | 3383 | slabp = virt_to_slab(objp); |
ff69416e | 3384 | l3 = cachep->nodelists[node]; |
1da177e4 | 3385 | list_del(&slabp->list); |
ff69416e | 3386 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3387 | check_slabp(cachep, slabp); |
78d382d7 | 3388 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3389 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3390 | l3->free_objects++; |
1da177e4 LT |
3391 | check_slabp(cachep, slabp); |
3392 | ||
3393 | /* fixup slab chains */ | |
3394 | if (slabp->inuse == 0) { | |
e498be7d CL |
3395 | if (l3->free_objects > l3->free_limit) { |
3396 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3397 | /* No need to drop any previously held |
3398 | * lock here, even if we have a off-slab slab | |
3399 | * descriptor it is guaranteed to come from | |
3400 | * a different cache, refer to comments before | |
3401 | * alloc_slabmgmt. | |
3402 | */ | |
1da177e4 LT |
3403 | slab_destroy(cachep, slabp); |
3404 | } else { | |
e498be7d | 3405 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3406 | } |
3407 | } else { | |
3408 | /* Unconditionally move a slab to the end of the | |
3409 | * partial list on free - maximum time for the | |
3410 | * other objects to be freed, too. | |
3411 | */ | |
e498be7d | 3412 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3413 | } |
3414 | } | |
3415 | } | |
3416 | ||
343e0d7a | 3417 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3418 | { |
3419 | int batchcount; | |
e498be7d | 3420 | struct kmem_list3 *l3; |
ff69416e | 3421 | int node = numa_node_id(); |
1da177e4 LT |
3422 | |
3423 | batchcount = ac->batchcount; | |
3424 | #if DEBUG | |
3425 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3426 | #endif | |
3427 | check_irq_off(); | |
ff69416e | 3428 | l3 = cachep->nodelists[node]; |
873623df | 3429 | spin_lock(&l3->list_lock); |
e498be7d CL |
3430 | if (l3->shared) { |
3431 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3432 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3433 | if (max) { |
3434 | if (batchcount > max) | |
3435 | batchcount = max; | |
e498be7d | 3436 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3437 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3438 | shared_array->avail += batchcount; |
3439 | goto free_done; | |
3440 | } | |
3441 | } | |
3442 | ||
ff69416e | 3443 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3444 | free_done: |
1da177e4 LT |
3445 | #if STATS |
3446 | { | |
3447 | int i = 0; | |
3448 | struct list_head *p; | |
3449 | ||
e498be7d CL |
3450 | p = l3->slabs_free.next; |
3451 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3452 | struct slab *slabp; |
3453 | ||
3454 | slabp = list_entry(p, struct slab, list); | |
3455 | BUG_ON(slabp->inuse); | |
3456 | ||
3457 | i++; | |
3458 | p = p->next; | |
3459 | } | |
3460 | STATS_SET_FREEABLE(cachep, i); | |
3461 | } | |
3462 | #endif | |
e498be7d | 3463 | spin_unlock(&l3->list_lock); |
1da177e4 | 3464 | ac->avail -= batchcount; |
a737b3e2 | 3465 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3466 | } |
3467 | ||
3468 | /* | |
a737b3e2 AM |
3469 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3470 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3471 | */ |
873623df | 3472 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3473 | { |
9a2dba4b | 3474 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3475 | |
3476 | check_irq_off(); | |
3477 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3478 | ||
873623df | 3479 | if (cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3480 | return; |
3481 | ||
1da177e4 LT |
3482 | if (likely(ac->avail < ac->limit)) { |
3483 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3484 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3485 | return; |
3486 | } else { | |
3487 | STATS_INC_FREEMISS(cachep); | |
3488 | cache_flusharray(cachep, ac); | |
e498be7d | 3489 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3490 | } |
3491 | } | |
3492 | ||
3493 | /** | |
3494 | * kmem_cache_alloc - Allocate an object | |
3495 | * @cachep: The cache to allocate from. | |
3496 | * @flags: See kmalloc(). | |
3497 | * | |
3498 | * Allocate an object from this cache. The flags are only relevant | |
3499 | * if the cache has no available objects. | |
3500 | */ | |
343e0d7a | 3501 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3502 | { |
7fd6b141 | 3503 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3504 | } |
3505 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3506 | ||
a8c0f9a4 | 3507 | /** |
b8008b2b | 3508 | * kmem_cache_zalloc - Allocate an object. The memory is set to zero. |
a8c0f9a4 PE |
3509 | * @cache: The cache to allocate from. |
3510 | * @flags: See kmalloc(). | |
3511 | * | |
3512 | * Allocate an object from this cache and set the allocated memory to zero. | |
3513 | * The flags are only relevant if the cache has no available objects. | |
3514 | */ | |
3515 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3516 | { | |
3517 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3518 | if (ret) | |
3519 | memset(ret, 0, obj_size(cache)); | |
3520 | return ret; | |
3521 | } | |
3522 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3523 | ||
1da177e4 LT |
3524 | /** |
3525 | * kmem_ptr_validate - check if an untrusted pointer might | |
3526 | * be a slab entry. | |
3527 | * @cachep: the cache we're checking against | |
3528 | * @ptr: pointer to validate | |
3529 | * | |
3530 | * This verifies that the untrusted pointer looks sane: | |
3531 | * it is _not_ a guarantee that the pointer is actually | |
3532 | * part of the slab cache in question, but it at least | |
3533 | * validates that the pointer can be dereferenced and | |
3534 | * looks half-way sane. | |
3535 | * | |
3536 | * Currently only used for dentry validation. | |
3537 | */ | |
343e0d7a | 3538 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3539 | { |
b28a02de | 3540 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3541 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3542 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3543 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3544 | struct page *page; |
3545 | ||
3546 | if (unlikely(addr < min_addr)) | |
3547 | goto out; | |
3548 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3549 | goto out; | |
3550 | if (unlikely(addr & align_mask)) | |
3551 | goto out; | |
3552 | if (unlikely(!kern_addr_valid(addr))) | |
3553 | goto out; | |
3554 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3555 | goto out; | |
3556 | page = virt_to_page(ptr); | |
3557 | if (unlikely(!PageSlab(page))) | |
3558 | goto out; | |
065d41cb | 3559 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3560 | goto out; |
3561 | return 1; | |
a737b3e2 | 3562 | out: |
1da177e4 LT |
3563 | return 0; |
3564 | } | |
3565 | ||
3566 | #ifdef CONFIG_NUMA | |
3567 | /** | |
3568 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3569 | * @cachep: The cache to allocate from. | |
3570 | * @flags: See kmalloc(). | |
3571 | * @nodeid: node number of the target node. | |
3572 | * | |
5bcd234d CL |
3573 | * Identical to kmem_cache_alloc but it will allocate memory on the given |
3574 | * node, which can improve the performance for cpu bound structures. | |
3575 | * | |
3576 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
1da177e4 | 3577 | */ |
8b98c169 CH |
3578 | static __always_inline void * |
3579 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, | |
3580 | int nodeid, void *caller) | |
1da177e4 | 3581 | { |
e498be7d | 3582 | unsigned long save_flags; |
5bcd234d | 3583 | void *ptr = NULL; |
1da177e4 | 3584 | |
e498be7d CL |
3585 | cache_alloc_debugcheck_before(cachep, flags); |
3586 | local_irq_save(save_flags); | |
18f820f6 | 3587 | |
5bcd234d CL |
3588 | if (unlikely(nodeid == -1)) |
3589 | nodeid = numa_node_id(); | |
18f820f6 | 3590 | |
5bcd234d CL |
3591 | if (likely(cachep->nodelists[nodeid])) { |
3592 | if (nodeid == numa_node_id()) { | |
3593 | /* | |
3594 | * Use the locally cached objects if possible. | |
3595 | * However ____cache_alloc does not allow fallback | |
3596 | * to other nodes. It may fail while we still have | |
3597 | * objects on other nodes available. | |
3598 | */ | |
3599 | ptr = ____cache_alloc(cachep, flags); | |
3600 | } | |
3601 | if (!ptr) { | |
3602 | /* ___cache_alloc_node can fall back to other nodes */ | |
3603 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3604 | } | |
3605 | } else { | |
3606 | /* Node not bootstrapped yet */ | |
3607 | if (!(flags & __GFP_THISNODE)) | |
3608 | ptr = fallback_alloc(cachep, flags); | |
3609 | } | |
3610 | ||
3611 | local_irq_restore(save_flags); | |
8b98c169 | 3612 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); |
1da177e4 | 3613 | |
e498be7d | 3614 | return ptr; |
1da177e4 | 3615 | } |
8b98c169 CH |
3616 | |
3617 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) | |
3618 | { | |
3619 | return __cache_alloc_node(cachep, flags, nodeid, | |
3620 | __builtin_return_address(0)); | |
3621 | } | |
1da177e4 LT |
3622 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3623 | ||
8b98c169 CH |
3624 | static __always_inline void * |
3625 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3626 | { |
343e0d7a | 3627 | struct kmem_cache *cachep; |
97e2bde4 MS |
3628 | |
3629 | cachep = kmem_find_general_cachep(size, flags); | |
3630 | if (unlikely(cachep == NULL)) | |
3631 | return NULL; | |
3632 | return kmem_cache_alloc_node(cachep, flags, node); | |
3633 | } | |
8b98c169 CH |
3634 | |
3635 | #ifdef CONFIG_DEBUG_SLAB | |
3636 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3637 | { | |
3638 | return __do_kmalloc_node(size, flags, node, | |
3639 | __builtin_return_address(0)); | |
3640 | } | |
dbe5e69d | 3641 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3642 | |
3643 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
3644 | int node, void *caller) | |
3645 | { | |
3646 | return __do_kmalloc_node(size, flags, node, caller); | |
3647 | } | |
3648 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3649 | #else | |
3650 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3651 | { | |
3652 | return __do_kmalloc_node(size, flags, node, NULL); | |
3653 | } | |
3654 | EXPORT_SYMBOL(__kmalloc_node); | |
3655 | #endif /* CONFIG_DEBUG_SLAB */ | |
3656 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3657 | |
3658 | /** | |
800590f5 | 3659 | * __do_kmalloc - allocate memory |
1da177e4 | 3660 | * @size: how many bytes of memory are required. |
800590f5 | 3661 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3662 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3663 | */ |
7fd6b141 PE |
3664 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3665 | void *caller) | |
1da177e4 | 3666 | { |
343e0d7a | 3667 | struct kmem_cache *cachep; |
1da177e4 | 3668 | |
97e2bde4 MS |
3669 | /* If you want to save a few bytes .text space: replace |
3670 | * __ with kmem_. | |
3671 | * Then kmalloc uses the uninlined functions instead of the inline | |
3672 | * functions. | |
3673 | */ | |
3674 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3675 | if (unlikely(cachep == NULL)) |
3676 | return NULL; | |
7fd6b141 PE |
3677 | return __cache_alloc(cachep, flags, caller); |
3678 | } | |
3679 | ||
7fd6b141 | 3680 | |
1d2c8eea | 3681 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3682 | void *__kmalloc(size_t size, gfp_t flags) |
3683 | { | |
871751e2 | 3684 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3685 | } |
3686 | EXPORT_SYMBOL(__kmalloc); | |
3687 | ||
7fd6b141 PE |
3688 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3689 | { | |
3690 | return __do_kmalloc(size, flags, caller); | |
3691 | } | |
3692 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3693 | |
3694 | #else | |
3695 | void *__kmalloc(size_t size, gfp_t flags) | |
3696 | { | |
3697 | return __do_kmalloc(size, flags, NULL); | |
3698 | } | |
3699 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3700 | #endif |
3701 | ||
1da177e4 LT |
3702 | /** |
3703 | * kmem_cache_free - Deallocate an object | |
3704 | * @cachep: The cache the allocation was from. | |
3705 | * @objp: The previously allocated object. | |
3706 | * | |
3707 | * Free an object which was previously allocated from this | |
3708 | * cache. | |
3709 | */ | |
343e0d7a | 3710 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3711 | { |
3712 | unsigned long flags; | |
3713 | ||
ddc2e812 PE |
3714 | BUG_ON(virt_to_cache(objp) != cachep); |
3715 | ||
1da177e4 | 3716 | local_irq_save(flags); |
873623df | 3717 | __cache_free(cachep, objp); |
1da177e4 LT |
3718 | local_irq_restore(flags); |
3719 | } | |
3720 | EXPORT_SYMBOL(kmem_cache_free); | |
3721 | ||
1da177e4 LT |
3722 | /** |
3723 | * kfree - free previously allocated memory | |
3724 | * @objp: pointer returned by kmalloc. | |
3725 | * | |
80e93eff PE |
3726 | * If @objp is NULL, no operation is performed. |
3727 | * | |
1da177e4 LT |
3728 | * Don't free memory not originally allocated by kmalloc() |
3729 | * or you will run into trouble. | |
3730 | */ | |
3731 | void kfree(const void *objp) | |
3732 | { | |
343e0d7a | 3733 | struct kmem_cache *c; |
1da177e4 LT |
3734 | unsigned long flags; |
3735 | ||
3736 | if (unlikely(!objp)) | |
3737 | return; | |
3738 | local_irq_save(flags); | |
3739 | kfree_debugcheck(objp); | |
6ed5eb22 | 3740 | c = virt_to_cache(objp); |
f9b8404c | 3741 | debug_check_no_locks_freed(objp, obj_size(c)); |
873623df | 3742 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3743 | local_irq_restore(flags); |
3744 | } | |
3745 | EXPORT_SYMBOL(kfree); | |
3746 | ||
343e0d7a | 3747 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3748 | { |
3dafccf2 | 3749 | return obj_size(cachep); |
1da177e4 LT |
3750 | } |
3751 | EXPORT_SYMBOL(kmem_cache_size); | |
3752 | ||
343e0d7a | 3753 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3754 | { |
3755 | return cachep->name; | |
3756 | } | |
3757 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3758 | ||
e498be7d | 3759 | /* |
0718dc2a | 3760 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3761 | */ |
343e0d7a | 3762 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3763 | { |
3764 | int node; | |
3765 | struct kmem_list3 *l3; | |
cafeb02e | 3766 | struct array_cache *new_shared; |
3395ee05 | 3767 | struct array_cache **new_alien = NULL; |
e498be7d CL |
3768 | |
3769 | for_each_online_node(node) { | |
cafeb02e | 3770 | |
3395ee05 PM |
3771 | if (use_alien_caches) { |
3772 | new_alien = alloc_alien_cache(node, cachep->limit); | |
3773 | if (!new_alien) | |
3774 | goto fail; | |
3775 | } | |
cafeb02e | 3776 | |
0718dc2a CL |
3777 | new_shared = alloc_arraycache(node, |
3778 | cachep->shared*cachep->batchcount, | |
a737b3e2 | 3779 | 0xbaadf00d); |
0718dc2a CL |
3780 | if (!new_shared) { |
3781 | free_alien_cache(new_alien); | |
e498be7d | 3782 | goto fail; |
0718dc2a | 3783 | } |
cafeb02e | 3784 | |
a737b3e2 AM |
3785 | l3 = cachep->nodelists[node]; |
3786 | if (l3) { | |
cafeb02e CL |
3787 | struct array_cache *shared = l3->shared; |
3788 | ||
e498be7d CL |
3789 | spin_lock_irq(&l3->list_lock); |
3790 | ||
cafeb02e | 3791 | if (shared) |
0718dc2a CL |
3792 | free_block(cachep, shared->entry, |
3793 | shared->avail, node); | |
e498be7d | 3794 | |
cafeb02e CL |
3795 | l3->shared = new_shared; |
3796 | if (!l3->alien) { | |
e498be7d CL |
3797 | l3->alien = new_alien; |
3798 | new_alien = NULL; | |
3799 | } | |
b28a02de | 3800 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3801 | cachep->batchcount + cachep->num; |
e498be7d | 3802 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3803 | kfree(shared); |
e498be7d CL |
3804 | free_alien_cache(new_alien); |
3805 | continue; | |
3806 | } | |
a737b3e2 | 3807 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3808 | if (!l3) { |
3809 | free_alien_cache(new_alien); | |
3810 | kfree(new_shared); | |
e498be7d | 3811 | goto fail; |
0718dc2a | 3812 | } |
e498be7d CL |
3813 | |
3814 | kmem_list3_init(l3); | |
3815 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3816 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3817 | l3->shared = new_shared; |
e498be7d | 3818 | l3->alien = new_alien; |
b28a02de | 3819 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3820 | cachep->batchcount + cachep->num; |
e498be7d CL |
3821 | cachep->nodelists[node] = l3; |
3822 | } | |
cafeb02e | 3823 | return 0; |
0718dc2a | 3824 | |
a737b3e2 | 3825 | fail: |
0718dc2a CL |
3826 | if (!cachep->next.next) { |
3827 | /* Cache is not active yet. Roll back what we did */ | |
3828 | node--; | |
3829 | while (node >= 0) { | |
3830 | if (cachep->nodelists[node]) { | |
3831 | l3 = cachep->nodelists[node]; | |
3832 | ||
3833 | kfree(l3->shared); | |
3834 | free_alien_cache(l3->alien); | |
3835 | kfree(l3); | |
3836 | cachep->nodelists[node] = NULL; | |
3837 | } | |
3838 | node--; | |
3839 | } | |
3840 | } | |
cafeb02e | 3841 | return -ENOMEM; |
e498be7d CL |
3842 | } |
3843 | ||
1da177e4 | 3844 | struct ccupdate_struct { |
343e0d7a | 3845 | struct kmem_cache *cachep; |
1da177e4 LT |
3846 | struct array_cache *new[NR_CPUS]; |
3847 | }; | |
3848 | ||
3849 | static void do_ccupdate_local(void *info) | |
3850 | { | |
a737b3e2 | 3851 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3852 | struct array_cache *old; |
3853 | ||
3854 | check_irq_off(); | |
9a2dba4b | 3855 | old = cpu_cache_get(new->cachep); |
e498be7d | 3856 | |
1da177e4 LT |
3857 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3858 | new->new[smp_processor_id()] = old; | |
3859 | } | |
3860 | ||
b5d8ca7c | 3861 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3862 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3863 | int batchcount, int shared) | |
1da177e4 | 3864 | { |
d2e7b7d0 | 3865 | struct ccupdate_struct *new; |
2ed3a4ef | 3866 | int i; |
1da177e4 | 3867 | |
d2e7b7d0 SS |
3868 | new = kzalloc(sizeof(*new), GFP_KERNEL); |
3869 | if (!new) | |
3870 | return -ENOMEM; | |
3871 | ||
e498be7d | 3872 | for_each_online_cpu(i) { |
d2e7b7d0 | 3873 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
a737b3e2 | 3874 | batchcount); |
d2e7b7d0 | 3875 | if (!new->new[i]) { |
b28a02de | 3876 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3877 | kfree(new->new[i]); |
3878 | kfree(new); | |
e498be7d | 3879 | return -ENOMEM; |
1da177e4 LT |
3880 | } |
3881 | } | |
d2e7b7d0 | 3882 | new->cachep = cachep; |
1da177e4 | 3883 | |
d2e7b7d0 | 3884 | on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); |
e498be7d | 3885 | |
1da177e4 | 3886 | check_irq_on(); |
1da177e4 LT |
3887 | cachep->batchcount = batchcount; |
3888 | cachep->limit = limit; | |
e498be7d | 3889 | cachep->shared = shared; |
1da177e4 | 3890 | |
e498be7d | 3891 | for_each_online_cpu(i) { |
d2e7b7d0 | 3892 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3893 | if (!ccold) |
3894 | continue; | |
e498be7d | 3895 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3896 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3897 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3898 | kfree(ccold); |
3899 | } | |
d2e7b7d0 | 3900 | kfree(new); |
2ed3a4ef | 3901 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3902 | } |
3903 | ||
b5d8ca7c | 3904 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3905 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3906 | { |
3907 | int err; | |
3908 | int limit, shared; | |
3909 | ||
a737b3e2 AM |
3910 | /* |
3911 | * The head array serves three purposes: | |
1da177e4 LT |
3912 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3913 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3914 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3915 | * bufctl chains: array operations are cheaper. |
3916 | * The numbers are guessed, we should auto-tune as described by | |
3917 | * Bonwick. | |
3918 | */ | |
3dafccf2 | 3919 | if (cachep->buffer_size > 131072) |
1da177e4 | 3920 | limit = 1; |
3dafccf2 | 3921 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3922 | limit = 8; |
3dafccf2 | 3923 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3924 | limit = 24; |
3dafccf2 | 3925 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3926 | limit = 54; |
3927 | else | |
3928 | limit = 120; | |
3929 | ||
a737b3e2 AM |
3930 | /* |
3931 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3932 | * allocation behaviour: Most allocs on one cpu, most free operations |
3933 | * on another cpu. For these cases, an efficient object passing between | |
3934 | * cpus is necessary. This is provided by a shared array. The array | |
3935 | * replaces Bonwick's magazine layer. | |
3936 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3937 | * to a larger limit. Thus disabled by default. | |
3938 | */ | |
3939 | shared = 0; | |
3940 | #ifdef CONFIG_SMP | |
3dafccf2 | 3941 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3942 | shared = 8; |
3943 | #endif | |
3944 | ||
3945 | #if DEBUG | |
a737b3e2 AM |
3946 | /* |
3947 | * With debugging enabled, large batchcount lead to excessively long | |
3948 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3949 | */ |
3950 | if (limit > 32) | |
3951 | limit = 32; | |
3952 | #endif | |
b28a02de | 3953 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3954 | if (err) |
3955 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3956 | cachep->name, -err); |
2ed3a4ef | 3957 | return err; |
1da177e4 LT |
3958 | } |
3959 | ||
1b55253a CL |
3960 | /* |
3961 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3962 | * necessary. Note that the l3 listlock also protects the array_cache |
3963 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3964 | */ |
3965 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
3966 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
3967 | { |
3968 | int tofree; | |
3969 | ||
1b55253a CL |
3970 | if (!ac || !ac->avail) |
3971 | return; | |
1da177e4 LT |
3972 | if (ac->touched && !force) { |
3973 | ac->touched = 0; | |
b18e7e65 | 3974 | } else { |
1b55253a | 3975 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
3976 | if (ac->avail) { |
3977 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3978 | if (tofree > ac->avail) | |
3979 | tofree = (ac->avail + 1) / 2; | |
3980 | free_block(cachep, ac->entry, tofree, node); | |
3981 | ac->avail -= tofree; | |
3982 | memmove(ac->entry, &(ac->entry[tofree]), | |
3983 | sizeof(void *) * ac->avail); | |
3984 | } | |
1b55253a | 3985 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
3986 | } |
3987 | } | |
3988 | ||
3989 | /** | |
3990 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3991 | * @unused: unused parameter |
1da177e4 LT |
3992 | * |
3993 | * Called from workqueue/eventd every few seconds. | |
3994 | * Purpose: | |
3995 | * - clear the per-cpu caches for this CPU. | |
3996 | * - return freeable pages to the main free memory pool. | |
3997 | * | |
a737b3e2 AM |
3998 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3999 | * again on the next iteration. | |
1da177e4 | 4000 | */ |
65f27f38 | 4001 | static void cache_reap(struct work_struct *unused) |
1da177e4 | 4002 | { |
7a7c381d | 4003 | struct kmem_cache *searchp; |
e498be7d | 4004 | struct kmem_list3 *l3; |
aab2207c | 4005 | int node = numa_node_id(); |
1da177e4 | 4006 | |
fc0abb14 | 4007 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 4008 | /* Give up. Setup the next iteration. */ |
b28a02de | 4009 | schedule_delayed_work(&__get_cpu_var(reap_work), |
2b284214 | 4010 | round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4011 | return; |
4012 | } | |
4013 | ||
7a7c381d | 4014 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4015 | check_irq_on(); |
4016 | ||
35386e3b CL |
4017 | /* |
4018 | * We only take the l3 lock if absolutely necessary and we | |
4019 | * have established with reasonable certainty that | |
4020 | * we can do some work if the lock was obtained. | |
4021 | */ | |
aab2207c | 4022 | l3 = searchp->nodelists[node]; |
35386e3b | 4023 | |
8fce4d8e | 4024 | reap_alien(searchp, l3); |
1da177e4 | 4025 | |
aab2207c | 4026 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4027 | |
35386e3b CL |
4028 | /* |
4029 | * These are racy checks but it does not matter | |
4030 | * if we skip one check or scan twice. | |
4031 | */ | |
e498be7d | 4032 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4033 | goto next; |
1da177e4 | 4034 | |
e498be7d | 4035 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4036 | |
aab2207c | 4037 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4038 | |
ed11d9eb | 4039 | if (l3->free_touched) |
e498be7d | 4040 | l3->free_touched = 0; |
ed11d9eb CL |
4041 | else { |
4042 | int freed; | |
1da177e4 | 4043 | |
ed11d9eb CL |
4044 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4045 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4046 | STATS_ADD_REAPED(searchp, freed); | |
4047 | } | |
35386e3b | 4048 | next: |
1da177e4 LT |
4049 | cond_resched(); |
4050 | } | |
4051 | check_irq_on(); | |
fc0abb14 | 4052 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4053 | next_reap_node(); |
2244b95a | 4054 | refresh_cpu_vm_stats(smp_processor_id()); |
a737b3e2 | 4055 | /* Set up the next iteration */ |
2b284214 AV |
4056 | schedule_delayed_work(&__get_cpu_var(reap_work), |
4057 | round_jiffies_relative(REAPTIMEOUT_CPUC)); | |
1da177e4 LT |
4058 | } |
4059 | ||
4060 | #ifdef CONFIG_PROC_FS | |
4061 | ||
85289f98 | 4062 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4063 | { |
85289f98 PE |
4064 | /* |
4065 | * Output format version, so at least we can change it | |
4066 | * without _too_ many complaints. | |
4067 | */ | |
1da177e4 | 4068 | #if STATS |
85289f98 | 4069 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4070 | #else |
85289f98 | 4071 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4072 | #endif |
85289f98 PE |
4073 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4074 | "<objperslab> <pagesperslab>"); | |
4075 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4076 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4077 | #if STATS |
85289f98 | 4078 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4079 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4080 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4081 | #endif |
85289f98 PE |
4082 | seq_putc(m, '\n'); |
4083 | } | |
4084 | ||
4085 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4086 | { | |
4087 | loff_t n = *pos; | |
4088 | struct list_head *p; | |
4089 | ||
fc0abb14 | 4090 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4091 | if (!n) |
4092 | print_slabinfo_header(m); | |
1da177e4 LT |
4093 | p = cache_chain.next; |
4094 | while (n--) { | |
4095 | p = p->next; | |
4096 | if (p == &cache_chain) | |
4097 | return NULL; | |
4098 | } | |
343e0d7a | 4099 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
4100 | } |
4101 | ||
4102 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4103 | { | |
343e0d7a | 4104 | struct kmem_cache *cachep = p; |
1da177e4 | 4105 | ++*pos; |
a737b3e2 AM |
4106 | return cachep->next.next == &cache_chain ? |
4107 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
4108 | } |
4109 | ||
4110 | static void s_stop(struct seq_file *m, void *p) | |
4111 | { | |
fc0abb14 | 4112 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4113 | } |
4114 | ||
4115 | static int s_show(struct seq_file *m, void *p) | |
4116 | { | |
343e0d7a | 4117 | struct kmem_cache *cachep = p; |
b28a02de PE |
4118 | struct slab *slabp; |
4119 | unsigned long active_objs; | |
4120 | unsigned long num_objs; | |
4121 | unsigned long active_slabs = 0; | |
4122 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4123 | const char *name; |
1da177e4 | 4124 | char *error = NULL; |
e498be7d CL |
4125 | int node; |
4126 | struct kmem_list3 *l3; | |
1da177e4 | 4127 | |
1da177e4 LT |
4128 | active_objs = 0; |
4129 | num_slabs = 0; | |
e498be7d CL |
4130 | for_each_online_node(node) { |
4131 | l3 = cachep->nodelists[node]; | |
4132 | if (!l3) | |
4133 | continue; | |
4134 | ||
ca3b9b91 RT |
4135 | check_irq_on(); |
4136 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4137 | |
7a7c381d | 4138 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4139 | if (slabp->inuse != cachep->num && !error) |
4140 | error = "slabs_full accounting error"; | |
4141 | active_objs += cachep->num; | |
4142 | active_slabs++; | |
4143 | } | |
7a7c381d | 4144 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4145 | if (slabp->inuse == cachep->num && !error) |
4146 | error = "slabs_partial inuse accounting error"; | |
4147 | if (!slabp->inuse && !error) | |
4148 | error = "slabs_partial/inuse accounting error"; | |
4149 | active_objs += slabp->inuse; | |
4150 | active_slabs++; | |
4151 | } | |
7a7c381d | 4152 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4153 | if (slabp->inuse && !error) |
4154 | error = "slabs_free/inuse accounting error"; | |
4155 | num_slabs++; | |
4156 | } | |
4157 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4158 | if (l3->shared) |
4159 | shared_avail += l3->shared->avail; | |
e498be7d | 4160 | |
ca3b9b91 | 4161 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4162 | } |
b28a02de PE |
4163 | num_slabs += active_slabs; |
4164 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4165 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4166 | error = "free_objects accounting error"; |
4167 | ||
b28a02de | 4168 | name = cachep->name; |
1da177e4 LT |
4169 | if (error) |
4170 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4171 | ||
4172 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4173 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4174 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4175 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4176 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4177 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4178 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4179 | #if STATS |
b28a02de | 4180 | { /* list3 stats */ |
1da177e4 LT |
4181 | unsigned long high = cachep->high_mark; |
4182 | unsigned long allocs = cachep->num_allocations; | |
4183 | unsigned long grown = cachep->grown; | |
4184 | unsigned long reaped = cachep->reaped; | |
4185 | unsigned long errors = cachep->errors; | |
4186 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4187 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4188 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4189 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4190 | |
e498be7d | 4191 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4192 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4193 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4194 | node_frees, overflows); |
1da177e4 LT |
4195 | } |
4196 | /* cpu stats */ | |
4197 | { | |
4198 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4199 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4200 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4201 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4202 | ||
4203 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4204 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4205 | } |
4206 | #endif | |
4207 | seq_putc(m, '\n'); | |
1da177e4 LT |
4208 | return 0; |
4209 | } | |
4210 | ||
4211 | /* | |
4212 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4213 | * | |
4214 | * Output layout: | |
4215 | * cache-name | |
4216 | * num-active-objs | |
4217 | * total-objs | |
4218 | * object size | |
4219 | * num-active-slabs | |
4220 | * total-slabs | |
4221 | * num-pages-per-slab | |
4222 | * + further values on SMP and with statistics enabled | |
4223 | */ | |
4224 | ||
15ad7cdc | 4225 | const struct seq_operations slabinfo_op = { |
b28a02de PE |
4226 | .start = s_start, |
4227 | .next = s_next, | |
4228 | .stop = s_stop, | |
4229 | .show = s_show, | |
1da177e4 LT |
4230 | }; |
4231 | ||
4232 | #define MAX_SLABINFO_WRITE 128 | |
4233 | /** | |
4234 | * slabinfo_write - Tuning for the slab allocator | |
4235 | * @file: unused | |
4236 | * @buffer: user buffer | |
4237 | * @count: data length | |
4238 | * @ppos: unused | |
4239 | */ | |
b28a02de PE |
4240 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4241 | size_t count, loff_t *ppos) | |
1da177e4 | 4242 | { |
b28a02de | 4243 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4244 | int limit, batchcount, shared, res; |
7a7c381d | 4245 | struct kmem_cache *cachep; |
b28a02de | 4246 | |
1da177e4 LT |
4247 | if (count > MAX_SLABINFO_WRITE) |
4248 | return -EINVAL; | |
4249 | if (copy_from_user(&kbuf, buffer, count)) | |
4250 | return -EFAULT; | |
b28a02de | 4251 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4252 | |
4253 | tmp = strchr(kbuf, ' '); | |
4254 | if (!tmp) | |
4255 | return -EINVAL; | |
4256 | *tmp = '\0'; | |
4257 | tmp++; | |
4258 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4259 | return -EINVAL; | |
4260 | ||
4261 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4262 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4263 | res = -EINVAL; |
7a7c381d | 4264 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4265 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4266 | if (limit < 1 || batchcount < 1 || |
4267 | batchcount > limit || shared < 0) { | |
e498be7d | 4268 | res = 0; |
1da177e4 | 4269 | } else { |
e498be7d | 4270 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4271 | batchcount, shared); |
1da177e4 LT |
4272 | } |
4273 | break; | |
4274 | } | |
4275 | } | |
fc0abb14 | 4276 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4277 | if (res >= 0) |
4278 | res = count; | |
4279 | return res; | |
4280 | } | |
871751e2 AV |
4281 | |
4282 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4283 | ||
4284 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4285 | { | |
4286 | loff_t n = *pos; | |
4287 | struct list_head *p; | |
4288 | ||
4289 | mutex_lock(&cache_chain_mutex); | |
4290 | p = cache_chain.next; | |
4291 | while (n--) { | |
4292 | p = p->next; | |
4293 | if (p == &cache_chain) | |
4294 | return NULL; | |
4295 | } | |
4296 | return list_entry(p, struct kmem_cache, next); | |
4297 | } | |
4298 | ||
4299 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4300 | { | |
4301 | unsigned long *p; | |
4302 | int l; | |
4303 | if (!v) | |
4304 | return 1; | |
4305 | l = n[1]; | |
4306 | p = n + 2; | |
4307 | while (l) { | |
4308 | int i = l/2; | |
4309 | unsigned long *q = p + 2 * i; | |
4310 | if (*q == v) { | |
4311 | q[1]++; | |
4312 | return 1; | |
4313 | } | |
4314 | if (*q > v) { | |
4315 | l = i; | |
4316 | } else { | |
4317 | p = q + 2; | |
4318 | l -= i + 1; | |
4319 | } | |
4320 | } | |
4321 | if (++n[1] == n[0]) | |
4322 | return 0; | |
4323 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4324 | p[0] = v; | |
4325 | p[1] = 1; | |
4326 | return 1; | |
4327 | } | |
4328 | ||
4329 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4330 | { | |
4331 | void *p; | |
4332 | int i; | |
4333 | if (n[0] == n[1]) | |
4334 | return; | |
4335 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4336 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4337 | continue; | |
4338 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4339 | return; | |
4340 | } | |
4341 | } | |
4342 | ||
4343 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4344 | { | |
4345 | #ifdef CONFIG_KALLSYMS | |
4346 | char *modname; | |
4347 | const char *name; | |
4348 | unsigned long offset, size; | |
4349 | char namebuf[KSYM_NAME_LEN+1]; | |
4350 | ||
4351 | name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); | |
4352 | ||
4353 | if (name) { | |
4354 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4355 | if (modname) | |
4356 | seq_printf(m, " [%s]", modname); | |
4357 | return; | |
4358 | } | |
4359 | #endif | |
4360 | seq_printf(m, "%p", (void *)address); | |
4361 | } | |
4362 | ||
4363 | static int leaks_show(struct seq_file *m, void *p) | |
4364 | { | |
4365 | struct kmem_cache *cachep = p; | |
871751e2 AV |
4366 | struct slab *slabp; |
4367 | struct kmem_list3 *l3; | |
4368 | const char *name; | |
4369 | unsigned long *n = m->private; | |
4370 | int node; | |
4371 | int i; | |
4372 | ||
4373 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4374 | return 0; | |
4375 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4376 | return 0; | |
4377 | ||
4378 | /* OK, we can do it */ | |
4379 | ||
4380 | n[1] = 0; | |
4381 | ||
4382 | for_each_online_node(node) { | |
4383 | l3 = cachep->nodelists[node]; | |
4384 | if (!l3) | |
4385 | continue; | |
4386 | ||
4387 | check_irq_on(); | |
4388 | spin_lock_irq(&l3->list_lock); | |
4389 | ||
7a7c381d | 4390 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4391 | handle_slab(n, cachep, slabp); |
7a7c381d | 4392 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4393 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4394 | spin_unlock_irq(&l3->list_lock); |
4395 | } | |
4396 | name = cachep->name; | |
4397 | if (n[0] == n[1]) { | |
4398 | /* Increase the buffer size */ | |
4399 | mutex_unlock(&cache_chain_mutex); | |
4400 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4401 | if (!m->private) { | |
4402 | /* Too bad, we are really out */ | |
4403 | m->private = n; | |
4404 | mutex_lock(&cache_chain_mutex); | |
4405 | return -ENOMEM; | |
4406 | } | |
4407 | *(unsigned long *)m->private = n[0] * 2; | |
4408 | kfree(n); | |
4409 | mutex_lock(&cache_chain_mutex); | |
4410 | /* Now make sure this entry will be retried */ | |
4411 | m->count = m->size; | |
4412 | return 0; | |
4413 | } | |
4414 | for (i = 0; i < n[1]; i++) { | |
4415 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4416 | show_symbol(m, n[2*i+2]); | |
4417 | seq_putc(m, '\n'); | |
4418 | } | |
d2e7b7d0 | 4419 | |
871751e2 AV |
4420 | return 0; |
4421 | } | |
4422 | ||
15ad7cdc | 4423 | const struct seq_operations slabstats_op = { |
871751e2 AV |
4424 | .start = leaks_start, |
4425 | .next = s_next, | |
4426 | .stop = s_stop, | |
4427 | .show = leaks_show, | |
4428 | }; | |
4429 | #endif | |
1da177e4 LT |
4430 | #endif |
4431 | ||
00e145b6 MS |
4432 | /** |
4433 | * ksize - get the actual amount of memory allocated for a given object | |
4434 | * @objp: Pointer to the object | |
4435 | * | |
4436 | * kmalloc may internally round up allocations and return more memory | |
4437 | * than requested. ksize() can be used to determine the actual amount of | |
4438 | * memory allocated. The caller may use this additional memory, even though | |
4439 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4440 | * The caller must guarantee that objp points to a valid object previously | |
4441 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4442 | * must not be freed during the duration of the call. | |
4443 | */ | |
1da177e4 LT |
4444 | unsigned int ksize(const void *objp) |
4445 | { | |
00e145b6 MS |
4446 | if (unlikely(objp == NULL)) |
4447 | return 0; | |
1da177e4 | 4448 | |
6ed5eb22 | 4449 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4450 | } |