Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
29 | * slabs and you must pass objects with the same intializations to | |
30 | * kmem_cache_free. | |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
54 | * it's changed with a smp_call_function(). |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
fc0abb14 | 71 | * The global cache-chain is protected by the mutex 'cache_chain_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
89 | #include <linux/config.h> | |
90 | #include <linux/slab.h> | |
91 | #include <linux/mm.h> | |
92 | #include <linux/swap.h> | |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
101a5001 | 97 | #include <linux/cpuset.h> |
1da177e4 LT |
98 | #include <linux/seq_file.h> |
99 | #include <linux/notifier.h> | |
100 | #include <linux/kallsyms.h> | |
101 | #include <linux/cpu.h> | |
102 | #include <linux/sysctl.h> | |
103 | #include <linux/module.h> | |
104 | #include <linux/rcupdate.h> | |
543537bd | 105 | #include <linux/string.h> |
e498be7d | 106 | #include <linux/nodemask.h> |
dc85da15 | 107 | #include <linux/mempolicy.h> |
fc0abb14 | 108 | #include <linux/mutex.h> |
1da177e4 LT |
109 | |
110 | #include <asm/uaccess.h> | |
111 | #include <asm/cacheflush.h> | |
112 | #include <asm/tlbflush.h> | |
113 | #include <asm/page.h> | |
114 | ||
115 | /* | |
116 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
117 | * SLAB_RED_ZONE & SLAB_POISON. | |
118 | * 0 for faster, smaller code (especially in the critical paths). | |
119 | * | |
120 | * STATS - 1 to collect stats for /proc/slabinfo. | |
121 | * 0 for faster, smaller code (especially in the critical paths). | |
122 | * | |
123 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
124 | */ | |
125 | ||
126 | #ifdef CONFIG_DEBUG_SLAB | |
127 | #define DEBUG 1 | |
128 | #define STATS 1 | |
129 | #define FORCED_DEBUG 1 | |
130 | #else | |
131 | #define DEBUG 0 | |
132 | #define STATS 0 | |
133 | #define FORCED_DEBUG 0 | |
134 | #endif | |
135 | ||
1da177e4 LT |
136 | /* Shouldn't this be in a header file somewhere? */ |
137 | #define BYTES_PER_WORD sizeof(void *) | |
138 | ||
139 | #ifndef cache_line_size | |
140 | #define cache_line_size() L1_CACHE_BYTES | |
141 | #endif | |
142 | ||
143 | #ifndef ARCH_KMALLOC_MINALIGN | |
144 | /* | |
145 | * Enforce a minimum alignment for the kmalloc caches. | |
146 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
147 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
148 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
149 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
150 | * Note that this flag disables some debug features. | |
151 | */ | |
152 | #define ARCH_KMALLOC_MINALIGN 0 | |
153 | #endif | |
154 | ||
155 | #ifndef ARCH_SLAB_MINALIGN | |
156 | /* | |
157 | * Enforce a minimum alignment for all caches. | |
158 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
159 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
160 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
161 | * some debug features. | |
162 | */ | |
163 | #define ARCH_SLAB_MINALIGN 0 | |
164 | #endif | |
165 | ||
166 | #ifndef ARCH_KMALLOC_FLAGS | |
167 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
168 | #endif | |
169 | ||
170 | /* Legal flag mask for kmem_cache_create(). */ | |
171 | #if DEBUG | |
172 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
173 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
ac2b898c | 174 | SLAB_CACHE_DMA | \ |
1da177e4 LT |
175 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ |
176 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 177 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 178 | #else |
ac2b898c | 179 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
1da177e4 LT |
180 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ |
181 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 182 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 LT |
183 | #endif |
184 | ||
185 | /* | |
186 | * kmem_bufctl_t: | |
187 | * | |
188 | * Bufctl's are used for linking objs within a slab | |
189 | * linked offsets. | |
190 | * | |
191 | * This implementation relies on "struct page" for locating the cache & | |
192 | * slab an object belongs to. | |
193 | * This allows the bufctl structure to be small (one int), but limits | |
194 | * the number of objects a slab (not a cache) can contain when off-slab | |
195 | * bufctls are used. The limit is the size of the largest general cache | |
196 | * that does not use off-slab slabs. | |
197 | * For 32bit archs with 4 kB pages, is this 56. | |
198 | * This is not serious, as it is only for large objects, when it is unwise | |
199 | * to have too many per slab. | |
200 | * Note: This limit can be raised by introducing a general cache whose size | |
201 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
202 | */ | |
203 | ||
fa5b08d5 | 204 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
205 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
206 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
207 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
208 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 LT |
209 | |
210 | /* Max number of objs-per-slab for caches which use off-slab slabs. | |
211 | * Needed to avoid a possible looping condition in cache_grow(). | |
212 | */ | |
213 | static unsigned long offslab_limit; | |
214 | ||
215 | /* | |
216 | * struct slab | |
217 | * | |
218 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
219 | * for a slab, or allocated from an general cache. | |
220 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
221 | */ | |
222 | struct slab { | |
b28a02de PE |
223 | struct list_head list; |
224 | unsigned long colouroff; | |
225 | void *s_mem; /* including colour offset */ | |
226 | unsigned int inuse; /* num of objs active in slab */ | |
227 | kmem_bufctl_t free; | |
228 | unsigned short nodeid; | |
1da177e4 LT |
229 | }; |
230 | ||
231 | /* | |
232 | * struct slab_rcu | |
233 | * | |
234 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
235 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
236 | * we need to approach a kernel structure obliquely, from its address | |
237 | * obtained without the usual locking. We can lock the structure to | |
238 | * stabilize it and check it's still at the given address, only if we | |
239 | * can be sure that the memory has not been meanwhile reused for some | |
240 | * other kind of object (which our subsystem's lock might corrupt). | |
241 | * | |
242 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
243 | * taking the spinlock within the structure expected at that address. | |
244 | * | |
245 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
246 | */ | |
247 | struct slab_rcu { | |
b28a02de | 248 | struct rcu_head head; |
343e0d7a | 249 | struct kmem_cache *cachep; |
b28a02de | 250 | void *addr; |
1da177e4 LT |
251 | }; |
252 | ||
253 | /* | |
254 | * struct array_cache | |
255 | * | |
1da177e4 LT |
256 | * Purpose: |
257 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
258 | * - reduce the number of linked list operations | |
259 | * - reduce spinlock operations | |
260 | * | |
261 | * The limit is stored in the per-cpu structure to reduce the data cache | |
262 | * footprint. | |
263 | * | |
264 | */ | |
265 | struct array_cache { | |
266 | unsigned int avail; | |
267 | unsigned int limit; | |
268 | unsigned int batchcount; | |
269 | unsigned int touched; | |
e498be7d | 270 | spinlock_t lock; |
a737b3e2 AM |
271 | void *entry[0]; /* |
272 | * Must have this definition in here for the proper | |
273 | * alignment of array_cache. Also simplifies accessing | |
274 | * the entries. | |
275 | * [0] is for gcc 2.95. It should really be []. | |
276 | */ | |
1da177e4 LT |
277 | }; |
278 | ||
a737b3e2 AM |
279 | /* |
280 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
281 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
282 | */ |
283 | #define BOOT_CPUCACHE_ENTRIES 1 | |
284 | struct arraycache_init { | |
285 | struct array_cache cache; | |
b28a02de | 286 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
287 | }; |
288 | ||
289 | /* | |
e498be7d | 290 | * The slab lists for all objects. |
1da177e4 LT |
291 | */ |
292 | struct kmem_list3 { | |
b28a02de PE |
293 | struct list_head slabs_partial; /* partial list first, better asm code */ |
294 | struct list_head slabs_full; | |
295 | struct list_head slabs_free; | |
296 | unsigned long free_objects; | |
b28a02de | 297 | unsigned int free_limit; |
2e1217cf | 298 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
299 | spinlock_t list_lock; |
300 | struct array_cache *shared; /* shared per node */ | |
301 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
302 | unsigned long next_reap; /* updated without locking */ |
303 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
304 | }; |
305 | ||
e498be7d CL |
306 | /* |
307 | * Need this for bootstrapping a per node allocator. | |
308 | */ | |
309 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
310 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
311 | #define CACHE_CACHE 0 | |
312 | #define SIZE_AC 1 | |
313 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
314 | ||
315 | /* | |
a737b3e2 AM |
316 | * This function must be completely optimized away if a constant is passed to |
317 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 318 | */ |
7243cc05 | 319 | static __always_inline int index_of(const size_t size) |
e498be7d | 320 | { |
5ec8a847 SR |
321 | extern void __bad_size(void); |
322 | ||
e498be7d CL |
323 | if (__builtin_constant_p(size)) { |
324 | int i = 0; | |
325 | ||
326 | #define CACHE(x) \ | |
327 | if (size <=x) \ | |
328 | return i; \ | |
329 | else \ | |
330 | i++; | |
331 | #include "linux/kmalloc_sizes.h" | |
332 | #undef CACHE | |
5ec8a847 | 333 | __bad_size(); |
7243cc05 | 334 | } else |
5ec8a847 | 335 | __bad_size(); |
e498be7d CL |
336 | return 0; |
337 | } | |
338 | ||
339 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) | |
340 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 341 | |
5295a74c | 342 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
343 | { |
344 | INIT_LIST_HEAD(&parent->slabs_full); | |
345 | INIT_LIST_HEAD(&parent->slabs_partial); | |
346 | INIT_LIST_HEAD(&parent->slabs_free); | |
347 | parent->shared = NULL; | |
348 | parent->alien = NULL; | |
2e1217cf | 349 | parent->colour_next = 0; |
e498be7d CL |
350 | spin_lock_init(&parent->list_lock); |
351 | parent->free_objects = 0; | |
352 | parent->free_touched = 0; | |
353 | } | |
354 | ||
a737b3e2 AM |
355 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
356 | do { \ | |
357 | INIT_LIST_HEAD(listp); \ | |
358 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
359 | } while (0) |
360 | ||
a737b3e2 AM |
361 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
362 | do { \ | |
e498be7d CL |
363 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
364 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
365 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
366 | } while (0) | |
1da177e4 LT |
367 | |
368 | /* | |
343e0d7a | 369 | * struct kmem_cache |
1da177e4 LT |
370 | * |
371 | * manages a cache. | |
372 | */ | |
b28a02de | 373 | |
2109a2d1 | 374 | struct kmem_cache { |
1da177e4 | 375 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 376 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 377 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
378 | unsigned int batchcount; |
379 | unsigned int limit; | |
380 | unsigned int shared; | |
b5d8ca7c | 381 | |
3dafccf2 | 382 | unsigned int buffer_size; |
b5d8ca7c | 383 | /* 3) touched by every alloc & free from the backend */ |
b28a02de | 384 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
b5d8ca7c | 385 | |
a737b3e2 AM |
386 | unsigned int flags; /* constant flags */ |
387 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 388 | |
b5d8ca7c | 389 | /* 4) cache_grow/shrink */ |
1da177e4 | 390 | /* order of pgs per slab (2^n) */ |
b28a02de | 391 | unsigned int gfporder; |
1da177e4 LT |
392 | |
393 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 394 | gfp_t gfpflags; |
1da177e4 | 395 | |
a737b3e2 | 396 | size_t colour; /* cache colouring range */ |
b28a02de | 397 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 398 | struct kmem_cache *slabp_cache; |
b28a02de | 399 | unsigned int slab_size; |
a737b3e2 | 400 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
401 | |
402 | /* constructor func */ | |
343e0d7a | 403 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
404 | |
405 | /* de-constructor func */ | |
343e0d7a | 406 | void (*dtor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 | 407 | |
b5d8ca7c | 408 | /* 5) cache creation/removal */ |
b28a02de PE |
409 | const char *name; |
410 | struct list_head next; | |
1da177e4 | 411 | |
b5d8ca7c | 412 | /* 6) statistics */ |
1da177e4 | 413 | #if STATS |
b28a02de PE |
414 | unsigned long num_active; |
415 | unsigned long num_allocations; | |
416 | unsigned long high_mark; | |
417 | unsigned long grown; | |
418 | unsigned long reaped; | |
419 | unsigned long errors; | |
420 | unsigned long max_freeable; | |
421 | unsigned long node_allocs; | |
422 | unsigned long node_frees; | |
fb7faf33 | 423 | unsigned long node_overflow; |
b28a02de PE |
424 | atomic_t allochit; |
425 | atomic_t allocmiss; | |
426 | atomic_t freehit; | |
427 | atomic_t freemiss; | |
1da177e4 LT |
428 | #endif |
429 | #if DEBUG | |
3dafccf2 MS |
430 | /* |
431 | * If debugging is enabled, then the allocator can add additional | |
432 | * fields and/or padding to every object. buffer_size contains the total | |
433 | * object size including these internal fields, the following two | |
434 | * variables contain the offset to the user object and its size. | |
435 | */ | |
436 | int obj_offset; | |
437 | int obj_size; | |
1da177e4 LT |
438 | #endif |
439 | }; | |
440 | ||
441 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
442 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
443 | ||
444 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
445 | /* |
446 | * Optimization question: fewer reaps means less probability for unnessary | |
447 | * cpucache drain/refill cycles. | |
1da177e4 | 448 | * |
dc6f3f27 | 449 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
450 | * which could lock up otherwise freeable slabs. |
451 | */ | |
452 | #define REAPTIMEOUT_CPUC (2*HZ) | |
453 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
454 | ||
455 | #if STATS | |
456 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
457 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
458 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
459 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
460 | #define STATS_INC_REAPED(x) ((x)->reaped++) | |
a737b3e2 AM |
461 | #define STATS_SET_HIGH(x) \ |
462 | do { \ | |
463 | if ((x)->num_active > (x)->high_mark) \ | |
464 | (x)->high_mark = (x)->num_active; \ | |
465 | } while (0) | |
1da177e4 LT |
466 | #define STATS_INC_ERR(x) ((x)->errors++) |
467 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 468 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 469 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
470 | #define STATS_SET_FREEABLE(x, i) \ |
471 | do { \ | |
472 | if ((x)->max_freeable < i) \ | |
473 | (x)->max_freeable = i; \ | |
474 | } while (0) | |
1da177e4 LT |
475 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
476 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
477 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
478 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
479 | #else | |
480 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
481 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
482 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
483 | #define STATS_INC_GROWN(x) do { } while (0) | |
484 | #define STATS_INC_REAPED(x) do { } while (0) | |
485 | #define STATS_SET_HIGH(x) do { } while (0) | |
486 | #define STATS_INC_ERR(x) do { } while (0) | |
487 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 488 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 489 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 490 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
491 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
492 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
493 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
494 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
495 | #endif | |
496 | ||
497 | #if DEBUG | |
a737b3e2 AM |
498 | /* |
499 | * Magic nums for obj red zoning. | |
1da177e4 LT |
500 | * Placed in the first word before and the first word after an obj. |
501 | */ | |
502 | #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ | |
503 | #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ | |
504 | ||
505 | /* ...and for poisoning */ | |
506 | #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ | |
507 | #define POISON_FREE 0x6b /* for use-after-free poisoning */ | |
508 | #define POISON_END 0xa5 /* end-byte of poisoning */ | |
509 | ||
a737b3e2 AM |
510 | /* |
511 | * memory layout of objects: | |
1da177e4 | 512 | * 0 : objp |
3dafccf2 | 513 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
514 | * the end of an object is aligned with the end of the real |
515 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 516 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 517 | * redzone word. |
3dafccf2 MS |
518 | * cachep->obj_offset: The real object. |
519 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
520 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
521 | * [BYTES_PER_WORD long] | |
1da177e4 | 522 | */ |
343e0d7a | 523 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 524 | { |
3dafccf2 | 525 | return cachep->obj_offset; |
1da177e4 LT |
526 | } |
527 | ||
343e0d7a | 528 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 529 | { |
3dafccf2 | 530 | return cachep->obj_size; |
1da177e4 LT |
531 | } |
532 | ||
343e0d7a | 533 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
534 | { |
535 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 536 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
537 | } |
538 | ||
343e0d7a | 539 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
540 | { |
541 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
542 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 543 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 544 | 2 * BYTES_PER_WORD); |
3dafccf2 | 545 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
546 | } |
547 | ||
343e0d7a | 548 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
549 | { |
550 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 551 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
552 | } |
553 | ||
554 | #else | |
555 | ||
3dafccf2 MS |
556 | #define obj_offset(x) 0 |
557 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
558 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
559 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
560 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
561 | ||
562 | #endif | |
563 | ||
564 | /* | |
a737b3e2 AM |
565 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
566 | * order. | |
1da177e4 LT |
567 | */ |
568 | #if defined(CONFIG_LARGE_ALLOCS) | |
569 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
570 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
571 | #elif defined(CONFIG_MMU) | |
572 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
573 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
574 | #else | |
575 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
576 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
577 | #endif | |
578 | ||
579 | /* | |
580 | * Do not go above this order unless 0 objects fit into the slab. | |
581 | */ | |
582 | #define BREAK_GFP_ORDER_HI 1 | |
583 | #define BREAK_GFP_ORDER_LO 0 | |
584 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
585 | ||
a737b3e2 AM |
586 | /* |
587 | * Functions for storing/retrieving the cachep and or slab from the page | |
588 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
589 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 590 | */ |
065d41cb PE |
591 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
592 | { | |
593 | page->lru.next = (struct list_head *)cache; | |
594 | } | |
595 | ||
596 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
597 | { | |
84097518 NP |
598 | if (unlikely(PageCompound(page))) |
599 | page = (struct page *)page_private(page); | |
065d41cb PE |
600 | return (struct kmem_cache *)page->lru.next; |
601 | } | |
602 | ||
603 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
604 | { | |
605 | page->lru.prev = (struct list_head *)slab; | |
606 | } | |
607 | ||
608 | static inline struct slab *page_get_slab(struct page *page) | |
609 | { | |
84097518 NP |
610 | if (unlikely(PageCompound(page))) |
611 | page = (struct page *)page_private(page); | |
065d41cb PE |
612 | return (struct slab *)page->lru.prev; |
613 | } | |
1da177e4 | 614 | |
6ed5eb22 PE |
615 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
616 | { | |
617 | struct page *page = virt_to_page(obj); | |
618 | return page_get_cache(page); | |
619 | } | |
620 | ||
621 | static inline struct slab *virt_to_slab(const void *obj) | |
622 | { | |
623 | struct page *page = virt_to_page(obj); | |
624 | return page_get_slab(page); | |
625 | } | |
626 | ||
8fea4e96 PE |
627 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
628 | unsigned int idx) | |
629 | { | |
630 | return slab->s_mem + cache->buffer_size * idx; | |
631 | } | |
632 | ||
633 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
634 | struct slab *slab, void *obj) | |
635 | { | |
636 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
637 | } | |
638 | ||
a737b3e2 AM |
639 | /* |
640 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
641 | */ | |
1da177e4 LT |
642 | struct cache_sizes malloc_sizes[] = { |
643 | #define CACHE(x) { .cs_size = (x) }, | |
644 | #include <linux/kmalloc_sizes.h> | |
645 | CACHE(ULONG_MAX) | |
646 | #undef CACHE | |
647 | }; | |
648 | EXPORT_SYMBOL(malloc_sizes); | |
649 | ||
650 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
651 | struct cache_names { | |
652 | char *name; | |
653 | char *name_dma; | |
654 | }; | |
655 | ||
656 | static struct cache_names __initdata cache_names[] = { | |
657 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
658 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 659 | {NULL,} |
1da177e4 LT |
660 | #undef CACHE |
661 | }; | |
662 | ||
663 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 664 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 665 | static struct arraycache_init initarray_generic = |
b28a02de | 666 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
667 | |
668 | /* internal cache of cache description objs */ | |
343e0d7a | 669 | static struct kmem_cache cache_cache = { |
b28a02de PE |
670 | .batchcount = 1, |
671 | .limit = BOOT_CPUCACHE_ENTRIES, | |
672 | .shared = 1, | |
343e0d7a | 673 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 674 | .name = "kmem_cache", |
1da177e4 | 675 | #if DEBUG |
343e0d7a | 676 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
677 | #endif |
678 | }; | |
679 | ||
680 | /* Guard access to the cache-chain. */ | |
fc0abb14 | 681 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
682 | static struct list_head cache_chain; |
683 | ||
684 | /* | |
a737b3e2 AM |
685 | * vm_enough_memory() looks at this to determine how many slab-allocated pages |
686 | * are possibly freeable under pressure | |
1da177e4 LT |
687 | * |
688 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
689 | */ | |
690 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
691 | |
692 | /* | |
693 | * chicken and egg problem: delay the per-cpu array allocation | |
694 | * until the general caches are up. | |
695 | */ | |
696 | static enum { | |
697 | NONE, | |
e498be7d CL |
698 | PARTIAL_AC, |
699 | PARTIAL_L3, | |
1da177e4 LT |
700 | FULL |
701 | } g_cpucache_up; | |
702 | ||
39d24e64 MK |
703 | /* |
704 | * used by boot code to determine if it can use slab based allocator | |
705 | */ | |
706 | int slab_is_available(void) | |
707 | { | |
708 | return g_cpucache_up == FULL; | |
709 | } | |
710 | ||
1da177e4 LT |
711 | static DEFINE_PER_CPU(struct work_struct, reap_work); |
712 | ||
a737b3e2 AM |
713 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, |
714 | int node); | |
343e0d7a | 715 | static void enable_cpucache(struct kmem_cache *cachep); |
b28a02de | 716 | static void cache_reap(void *unused); |
343e0d7a | 717 | static int __node_shrink(struct kmem_cache *cachep, int node); |
1da177e4 | 718 | |
343e0d7a | 719 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
720 | { |
721 | return cachep->array[smp_processor_id()]; | |
722 | } | |
723 | ||
a737b3e2 AM |
724 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
725 | gfp_t gfpflags) | |
1da177e4 LT |
726 | { |
727 | struct cache_sizes *csizep = malloc_sizes; | |
728 | ||
729 | #if DEBUG | |
730 | /* This happens if someone tries to call | |
b28a02de PE |
731 | * kmem_cache_create(), or __kmalloc(), before |
732 | * the generic caches are initialized. | |
733 | */ | |
c7e43c78 | 734 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
735 | #endif |
736 | while (size > csizep->cs_size) | |
737 | csizep++; | |
738 | ||
739 | /* | |
0abf40c1 | 740 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
741 | * has cs_{dma,}cachep==NULL. Thus no special case |
742 | * for large kmalloc calls required. | |
743 | */ | |
744 | if (unlikely(gfpflags & GFP_DMA)) | |
745 | return csizep->cs_dmacachep; | |
746 | return csizep->cs_cachep; | |
747 | } | |
748 | ||
343e0d7a | 749 | struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
750 | { |
751 | return __find_general_cachep(size, gfpflags); | |
752 | } | |
753 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
754 | ||
fbaccacf | 755 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 756 | { |
fbaccacf SR |
757 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
758 | } | |
1da177e4 | 759 | |
a737b3e2 AM |
760 | /* |
761 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
762 | */ | |
fbaccacf SR |
763 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
764 | size_t align, int flags, size_t *left_over, | |
765 | unsigned int *num) | |
766 | { | |
767 | int nr_objs; | |
768 | size_t mgmt_size; | |
769 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 770 | |
fbaccacf SR |
771 | /* |
772 | * The slab management structure can be either off the slab or | |
773 | * on it. For the latter case, the memory allocated for a | |
774 | * slab is used for: | |
775 | * | |
776 | * - The struct slab | |
777 | * - One kmem_bufctl_t for each object | |
778 | * - Padding to respect alignment of @align | |
779 | * - @buffer_size bytes for each object | |
780 | * | |
781 | * If the slab management structure is off the slab, then the | |
782 | * alignment will already be calculated into the size. Because | |
783 | * the slabs are all pages aligned, the objects will be at the | |
784 | * correct alignment when allocated. | |
785 | */ | |
786 | if (flags & CFLGS_OFF_SLAB) { | |
787 | mgmt_size = 0; | |
788 | nr_objs = slab_size / buffer_size; | |
789 | ||
790 | if (nr_objs > SLAB_LIMIT) | |
791 | nr_objs = SLAB_LIMIT; | |
792 | } else { | |
793 | /* | |
794 | * Ignore padding for the initial guess. The padding | |
795 | * is at most @align-1 bytes, and @buffer_size is at | |
796 | * least @align. In the worst case, this result will | |
797 | * be one greater than the number of objects that fit | |
798 | * into the memory allocation when taking the padding | |
799 | * into account. | |
800 | */ | |
801 | nr_objs = (slab_size - sizeof(struct slab)) / | |
802 | (buffer_size + sizeof(kmem_bufctl_t)); | |
803 | ||
804 | /* | |
805 | * This calculated number will be either the right | |
806 | * amount, or one greater than what we want. | |
807 | */ | |
808 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
809 | > slab_size) | |
810 | nr_objs--; | |
811 | ||
812 | if (nr_objs > SLAB_LIMIT) | |
813 | nr_objs = SLAB_LIMIT; | |
814 | ||
815 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
816 | } | |
817 | *num = nr_objs; | |
818 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
819 | } |
820 | ||
821 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
822 | ||
a737b3e2 AM |
823 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
824 | char *msg) | |
1da177e4 LT |
825 | { |
826 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 827 | function, cachep->name, msg); |
1da177e4 LT |
828 | dump_stack(); |
829 | } | |
830 | ||
8fce4d8e CL |
831 | #ifdef CONFIG_NUMA |
832 | /* | |
833 | * Special reaping functions for NUMA systems called from cache_reap(). | |
834 | * These take care of doing round robin flushing of alien caches (containing | |
835 | * objects freed on different nodes from which they were allocated) and the | |
836 | * flushing of remote pcps by calling drain_node_pages. | |
837 | */ | |
838 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
839 | ||
840 | static void init_reap_node(int cpu) | |
841 | { | |
842 | int node; | |
843 | ||
844 | node = next_node(cpu_to_node(cpu), node_online_map); | |
845 | if (node == MAX_NUMNODES) | |
442295c9 | 846 | node = first_node(node_online_map); |
8fce4d8e CL |
847 | |
848 | __get_cpu_var(reap_node) = node; | |
849 | } | |
850 | ||
851 | static void next_reap_node(void) | |
852 | { | |
853 | int node = __get_cpu_var(reap_node); | |
854 | ||
855 | /* | |
856 | * Also drain per cpu pages on remote zones | |
857 | */ | |
858 | if (node != numa_node_id()) | |
859 | drain_node_pages(node); | |
860 | ||
861 | node = next_node(node, node_online_map); | |
862 | if (unlikely(node >= MAX_NUMNODES)) | |
863 | node = first_node(node_online_map); | |
864 | __get_cpu_var(reap_node) = node; | |
865 | } | |
866 | ||
867 | #else | |
868 | #define init_reap_node(cpu) do { } while (0) | |
869 | #define next_reap_node(void) do { } while (0) | |
870 | #endif | |
871 | ||
1da177e4 LT |
872 | /* |
873 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
874 | * via the workqueue/eventd. | |
875 | * Add the CPU number into the expiration time to minimize the possibility of | |
876 | * the CPUs getting into lockstep and contending for the global cache chain | |
877 | * lock. | |
878 | */ | |
879 | static void __devinit start_cpu_timer(int cpu) | |
880 | { | |
881 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
882 | ||
883 | /* | |
884 | * When this gets called from do_initcalls via cpucache_init(), | |
885 | * init_workqueues() has already run, so keventd will be setup | |
886 | * at that time. | |
887 | */ | |
888 | if (keventd_up() && reap_work->func == NULL) { | |
8fce4d8e | 889 | init_reap_node(cpu); |
1da177e4 LT |
890 | INIT_WORK(reap_work, cache_reap, NULL); |
891 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
892 | } | |
893 | } | |
894 | ||
e498be7d | 895 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 896 | int batchcount) |
1da177e4 | 897 | { |
b28a02de | 898 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
899 | struct array_cache *nc = NULL; |
900 | ||
e498be7d | 901 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
902 | if (nc) { |
903 | nc->avail = 0; | |
904 | nc->limit = entries; | |
905 | nc->batchcount = batchcount; | |
906 | nc->touched = 0; | |
e498be7d | 907 | spin_lock_init(&nc->lock); |
1da177e4 LT |
908 | } |
909 | return nc; | |
910 | } | |
911 | ||
3ded175a CL |
912 | /* |
913 | * Transfer objects in one arraycache to another. | |
914 | * Locking must be handled by the caller. | |
915 | * | |
916 | * Return the number of entries transferred. | |
917 | */ | |
918 | static int transfer_objects(struct array_cache *to, | |
919 | struct array_cache *from, unsigned int max) | |
920 | { | |
921 | /* Figure out how many entries to transfer */ | |
922 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
923 | ||
924 | if (!nr) | |
925 | return 0; | |
926 | ||
927 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
928 | sizeof(void *) *nr); | |
929 | ||
930 | from->avail -= nr; | |
931 | to->avail += nr; | |
932 | to->touched = 1; | |
933 | return nr; | |
934 | } | |
935 | ||
e498be7d | 936 | #ifdef CONFIG_NUMA |
343e0d7a | 937 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 938 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 939 | |
5295a74c | 940 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
941 | { |
942 | struct array_cache **ac_ptr; | |
b28a02de | 943 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
944 | int i; |
945 | ||
946 | if (limit > 1) | |
947 | limit = 12; | |
948 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
949 | if (ac_ptr) { | |
950 | for_each_node(i) { | |
951 | if (i == node || !node_online(i)) { | |
952 | ac_ptr[i] = NULL; | |
953 | continue; | |
954 | } | |
955 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
956 | if (!ac_ptr[i]) { | |
b28a02de | 957 | for (i--; i <= 0; i--) |
e498be7d CL |
958 | kfree(ac_ptr[i]); |
959 | kfree(ac_ptr); | |
960 | return NULL; | |
961 | } | |
962 | } | |
963 | } | |
964 | return ac_ptr; | |
965 | } | |
966 | ||
5295a74c | 967 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
968 | { |
969 | int i; | |
970 | ||
971 | if (!ac_ptr) | |
972 | return; | |
e498be7d | 973 | for_each_node(i) |
b28a02de | 974 | kfree(ac_ptr[i]); |
e498be7d CL |
975 | kfree(ac_ptr); |
976 | } | |
977 | ||
343e0d7a | 978 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 979 | struct array_cache *ac, int node) |
e498be7d CL |
980 | { |
981 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
982 | ||
983 | if (ac->avail) { | |
984 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
985 | /* |
986 | * Stuff objects into the remote nodes shared array first. | |
987 | * That way we could avoid the overhead of putting the objects | |
988 | * into the free lists and getting them back later. | |
989 | */ | |
693f7d36 | 990 | if (rl3->shared) |
991 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 992 | |
ff69416e | 993 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
994 | ac->avail = 0; |
995 | spin_unlock(&rl3->list_lock); | |
996 | } | |
997 | } | |
998 | ||
8fce4d8e CL |
999 | /* |
1000 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1001 | */ | |
1002 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1003 | { | |
1004 | int node = __get_cpu_var(reap_node); | |
1005 | ||
1006 | if (l3->alien) { | |
1007 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1008 | |
1009 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1010 | __drain_alien_cache(cachep, ac, node); |
1011 | spin_unlock_irq(&ac->lock); | |
1012 | } | |
1013 | } | |
1014 | } | |
1015 | ||
a737b3e2 AM |
1016 | static void drain_alien_cache(struct kmem_cache *cachep, |
1017 | struct array_cache **alien) | |
e498be7d | 1018 | { |
b28a02de | 1019 | int i = 0; |
e498be7d CL |
1020 | struct array_cache *ac; |
1021 | unsigned long flags; | |
1022 | ||
1023 | for_each_online_node(i) { | |
4484ebf1 | 1024 | ac = alien[i]; |
e498be7d CL |
1025 | if (ac) { |
1026 | spin_lock_irqsave(&ac->lock, flags); | |
1027 | __drain_alien_cache(cachep, ac, i); | |
1028 | spin_unlock_irqrestore(&ac->lock, flags); | |
1029 | } | |
1030 | } | |
1031 | } | |
1032 | #else | |
7a21ef6f | 1033 | |
4484ebf1 | 1034 | #define drain_alien_cache(cachep, alien) do { } while (0) |
8fce4d8e | 1035 | #define reap_alien(cachep, l3) do { } while (0) |
4484ebf1 | 1036 | |
7a21ef6f LT |
1037 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
1038 | { | |
1039 | return (struct array_cache **) 0x01020304ul; | |
1040 | } | |
1041 | ||
4484ebf1 RT |
1042 | static inline void free_alien_cache(struct array_cache **ac_ptr) |
1043 | { | |
1044 | } | |
7a21ef6f | 1045 | |
e498be7d CL |
1046 | #endif |
1047 | ||
83d722f7 | 1048 | static int cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1049 | unsigned long action, void *hcpu) |
1da177e4 LT |
1050 | { |
1051 | long cpu = (long)hcpu; | |
343e0d7a | 1052 | struct kmem_cache *cachep; |
e498be7d CL |
1053 | struct kmem_list3 *l3 = NULL; |
1054 | int node = cpu_to_node(cpu); | |
1055 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1056 | |
1057 | switch (action) { | |
1058 | case CPU_UP_PREPARE: | |
fc0abb14 | 1059 | mutex_lock(&cache_chain_mutex); |
a737b3e2 AM |
1060 | /* |
1061 | * We need to do this right in the beginning since | |
e498be7d CL |
1062 | * alloc_arraycache's are going to use this list. |
1063 | * kmalloc_node allows us to add the slab to the right | |
1064 | * kmem_list3 and not this cpu's kmem_list3 | |
1065 | */ | |
1066 | ||
1da177e4 | 1067 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1068 | /* |
1069 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1070 | * begin anything. Make sure some other cpu on this |
1071 | * node has not already allocated this | |
1072 | */ | |
1073 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1074 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1075 | if (!l3) | |
e498be7d CL |
1076 | goto bad; |
1077 | kmem_list3_init(l3); | |
1078 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1079 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1080 | |
4484ebf1 RT |
1081 | /* |
1082 | * The l3s don't come and go as CPUs come and | |
1083 | * go. cache_chain_mutex is sufficient | |
1084 | * protection here. | |
1085 | */ | |
e498be7d CL |
1086 | cachep->nodelists[node] = l3; |
1087 | } | |
1da177e4 | 1088 | |
e498be7d CL |
1089 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1090 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1091 | (1 + nr_cpus_node(node)) * |
1092 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1093 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1094 | } | |
1095 | ||
a737b3e2 AM |
1096 | /* |
1097 | * Now we can go ahead with allocating the shared arrays and | |
1098 | * array caches | |
1099 | */ | |
e498be7d | 1100 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1101 | struct array_cache *nc; |
4484ebf1 RT |
1102 | struct array_cache *shared; |
1103 | struct array_cache **alien; | |
cd105df4 | 1104 | |
e498be7d | 1105 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1106 | cachep->batchcount); |
1da177e4 LT |
1107 | if (!nc) |
1108 | goto bad; | |
4484ebf1 RT |
1109 | shared = alloc_arraycache(node, |
1110 | cachep->shared * cachep->batchcount, | |
1111 | 0xbaadf00d); | |
1112 | if (!shared) | |
1113 | goto bad; | |
7a21ef6f | 1114 | |
4484ebf1 RT |
1115 | alien = alloc_alien_cache(node, cachep->limit); |
1116 | if (!alien) | |
1117 | goto bad; | |
1da177e4 | 1118 | cachep->array[cpu] = nc; |
e498be7d CL |
1119 | l3 = cachep->nodelists[node]; |
1120 | BUG_ON(!l3); | |
e498be7d | 1121 | |
4484ebf1 RT |
1122 | spin_lock_irq(&l3->list_lock); |
1123 | if (!l3->shared) { | |
1124 | /* | |
1125 | * We are serialised from CPU_DEAD or | |
1126 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1127 | */ | |
1128 | l3->shared = shared; | |
1129 | shared = NULL; | |
e498be7d | 1130 | } |
4484ebf1 RT |
1131 | #ifdef CONFIG_NUMA |
1132 | if (!l3->alien) { | |
1133 | l3->alien = alien; | |
1134 | alien = NULL; | |
1135 | } | |
1136 | #endif | |
1137 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1138 | kfree(shared); |
1139 | free_alien_cache(alien); | |
1da177e4 | 1140 | } |
fc0abb14 | 1141 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1142 | break; |
1143 | case CPU_ONLINE: | |
1144 | start_cpu_timer(cpu); | |
1145 | break; | |
1146 | #ifdef CONFIG_HOTPLUG_CPU | |
1147 | case CPU_DEAD: | |
4484ebf1 RT |
1148 | /* |
1149 | * Even if all the cpus of a node are down, we don't free the | |
1150 | * kmem_list3 of any cache. This to avoid a race between | |
1151 | * cpu_down, and a kmalloc allocation from another cpu for | |
1152 | * memory from the node of the cpu going down. The list3 | |
1153 | * structure is usually allocated from kmem_cache_create() and | |
1154 | * gets destroyed at kmem_cache_destroy(). | |
1155 | */ | |
1da177e4 LT |
1156 | /* fall thru */ |
1157 | case CPU_UP_CANCELED: | |
fc0abb14 | 1158 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1159 | list_for_each_entry(cachep, &cache_chain, next) { |
1160 | struct array_cache *nc; | |
4484ebf1 RT |
1161 | struct array_cache *shared; |
1162 | struct array_cache **alien; | |
e498be7d | 1163 | cpumask_t mask; |
1da177e4 | 1164 | |
e498be7d | 1165 | mask = node_to_cpumask(node); |
1da177e4 LT |
1166 | /* cpu is dead; no one can alloc from it. */ |
1167 | nc = cachep->array[cpu]; | |
1168 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1169 | l3 = cachep->nodelists[node]; |
1170 | ||
1171 | if (!l3) | |
4484ebf1 | 1172 | goto free_array_cache; |
e498be7d | 1173 | |
ca3b9b91 | 1174 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1175 | |
1176 | /* Free limit for this kmem_list3 */ | |
1177 | l3->free_limit -= cachep->batchcount; | |
1178 | if (nc) | |
ff69416e | 1179 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1180 | |
1181 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1182 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1183 | goto free_array_cache; |
b28a02de | 1184 | } |
e498be7d | 1185 | |
4484ebf1 RT |
1186 | shared = l3->shared; |
1187 | if (shared) { | |
e498be7d | 1188 | free_block(cachep, l3->shared->entry, |
b28a02de | 1189 | l3->shared->avail, node); |
e498be7d CL |
1190 | l3->shared = NULL; |
1191 | } | |
e498be7d | 1192 | |
4484ebf1 RT |
1193 | alien = l3->alien; |
1194 | l3->alien = NULL; | |
1195 | ||
1196 | spin_unlock_irq(&l3->list_lock); | |
1197 | ||
1198 | kfree(shared); | |
1199 | if (alien) { | |
1200 | drain_alien_cache(cachep, alien); | |
1201 | free_alien_cache(alien); | |
e498be7d | 1202 | } |
4484ebf1 | 1203 | free_array_cache: |
1da177e4 LT |
1204 | kfree(nc); |
1205 | } | |
4484ebf1 RT |
1206 | /* |
1207 | * In the previous loop, all the objects were freed to | |
1208 | * the respective cache's slabs, now we can go ahead and | |
1209 | * shrink each nodelist to its limit. | |
1210 | */ | |
1211 | list_for_each_entry(cachep, &cache_chain, next) { | |
1212 | l3 = cachep->nodelists[node]; | |
1213 | if (!l3) | |
1214 | continue; | |
1215 | spin_lock_irq(&l3->list_lock); | |
1216 | /* free slabs belonging to this node */ | |
1217 | __node_shrink(cachep, node); | |
1218 | spin_unlock_irq(&l3->list_lock); | |
1219 | } | |
fc0abb14 | 1220 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1221 | break; |
1222 | #endif | |
1223 | } | |
1224 | return NOTIFY_OK; | |
a737b3e2 | 1225 | bad: |
fc0abb14 | 1226 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1227 | return NOTIFY_BAD; |
1228 | } | |
1229 | ||
1230 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
1231 | ||
e498be7d CL |
1232 | /* |
1233 | * swap the static kmem_list3 with kmalloced memory | |
1234 | */ | |
a737b3e2 AM |
1235 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1236 | int nodeid) | |
e498be7d CL |
1237 | { |
1238 | struct kmem_list3 *ptr; | |
1239 | ||
1240 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1241 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1242 | BUG_ON(!ptr); | |
1243 | ||
1244 | local_irq_disable(); | |
1245 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1246 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1247 | cachep->nodelists[nodeid] = ptr; | |
1248 | local_irq_enable(); | |
1249 | } | |
1250 | ||
a737b3e2 AM |
1251 | /* |
1252 | * Initialisation. Called after the page allocator have been initialised and | |
1253 | * before smp_init(). | |
1da177e4 LT |
1254 | */ |
1255 | void __init kmem_cache_init(void) | |
1256 | { | |
1257 | size_t left_over; | |
1258 | struct cache_sizes *sizes; | |
1259 | struct cache_names *names; | |
e498be7d | 1260 | int i; |
07ed76b2 | 1261 | int order; |
e498be7d CL |
1262 | |
1263 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1264 | kmem_list3_init(&initkmem_list3[i]); | |
1265 | if (i < MAX_NUMNODES) | |
1266 | cache_cache.nodelists[i] = NULL; | |
1267 | } | |
1da177e4 LT |
1268 | |
1269 | /* | |
1270 | * Fragmentation resistance on low memory - only use bigger | |
1271 | * page orders on machines with more than 32MB of memory. | |
1272 | */ | |
1273 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1274 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1275 | ||
1da177e4 LT |
1276 | /* Bootstrap is tricky, because several objects are allocated |
1277 | * from caches that do not exist yet: | |
a737b3e2 AM |
1278 | * 1) initialize the cache_cache cache: it contains the struct |
1279 | * kmem_cache structures of all caches, except cache_cache itself: | |
1280 | * cache_cache is statically allocated. | |
e498be7d CL |
1281 | * Initially an __init data area is used for the head array and the |
1282 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1283 | * array at the end of the bootstrap. | |
1da177e4 | 1284 | * 2) Create the first kmalloc cache. |
343e0d7a | 1285 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1286 | * An __init data area is used for the head array. |
1287 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1288 | * head arrays. | |
1da177e4 LT |
1289 | * 4) Replace the __init data head arrays for cache_cache and the first |
1290 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1291 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1292 | * the other cache's with kmalloc allocated memory. | |
1293 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1294 | */ |
1295 | ||
1296 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1297 | INIT_LIST_HEAD(&cache_chain); |
1298 | list_add(&cache_cache.next, &cache_chain); | |
1299 | cache_cache.colour_off = cache_line_size(); | |
1300 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1301 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1302 | |
a737b3e2 AM |
1303 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1304 | cache_line_size()); | |
1da177e4 | 1305 | |
07ed76b2 JS |
1306 | for (order = 0; order < MAX_ORDER; order++) { |
1307 | cache_estimate(order, cache_cache.buffer_size, | |
1308 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1309 | if (cache_cache.num) | |
1310 | break; | |
1311 | } | |
40094fa6 | 1312 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1313 | cache_cache.gfporder = order; |
b28a02de | 1314 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1315 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1316 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1317 | |
1318 | /* 2+3) create the kmalloc caches */ | |
1319 | sizes = malloc_sizes; | |
1320 | names = cache_names; | |
1321 | ||
a737b3e2 AM |
1322 | /* |
1323 | * Initialize the caches that provide memory for the array cache and the | |
1324 | * kmem_list3 structures first. Without this, further allocations will | |
1325 | * bug. | |
e498be7d CL |
1326 | */ |
1327 | ||
1328 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1329 | sizes[INDEX_AC].cs_size, |
1330 | ARCH_KMALLOC_MINALIGN, | |
1331 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1332 | NULL, NULL); | |
e498be7d | 1333 | |
a737b3e2 | 1334 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1335 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1336 | kmem_cache_create(names[INDEX_L3].name, |
1337 | sizes[INDEX_L3].cs_size, | |
1338 | ARCH_KMALLOC_MINALIGN, | |
1339 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1340 | NULL, NULL); | |
1341 | } | |
e498be7d | 1342 | |
1da177e4 | 1343 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1344 | /* |
1345 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1346 | * This should be particularly beneficial on SMP boxes, as it |
1347 | * eliminates "false sharing". | |
1348 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1349 | * allow tighter packing of the smaller caches. |
1350 | */ | |
a737b3e2 | 1351 | if (!sizes->cs_cachep) { |
e498be7d | 1352 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1353 | sizes->cs_size, |
1354 | ARCH_KMALLOC_MINALIGN, | |
1355 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1356 | NULL, NULL); | |
1357 | } | |
1da177e4 LT |
1358 | |
1359 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
1360 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
b28a02de | 1361 | offslab_limit = sizes->cs_size - sizeof(struct slab); |
1da177e4 LT |
1362 | offslab_limit /= sizeof(kmem_bufctl_t); |
1363 | } | |
1364 | ||
1365 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
a737b3e2 AM |
1366 | sizes->cs_size, |
1367 | ARCH_KMALLOC_MINALIGN, | |
1368 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1369 | SLAB_PANIC, | |
1370 | NULL, NULL); | |
1da177e4 LT |
1371 | sizes++; |
1372 | names++; | |
1373 | } | |
1374 | /* 4) Replace the bootstrap head arrays */ | |
1375 | { | |
b28a02de | 1376 | void *ptr; |
e498be7d | 1377 | |
1da177e4 | 1378 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1379 | |
1da177e4 | 1380 | local_irq_disable(); |
9a2dba4b PE |
1381 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1382 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1383 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1384 | cache_cache.array[smp_processor_id()] = ptr; |
1385 | local_irq_enable(); | |
e498be7d | 1386 | |
1da177e4 | 1387 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1388 | |
1da177e4 | 1389 | local_irq_disable(); |
9a2dba4b | 1390 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1391 | != &initarray_generic.cache); |
9a2dba4b | 1392 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1393 | sizeof(struct arraycache_init)); |
e498be7d | 1394 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1395 | ptr; |
1da177e4 LT |
1396 | local_irq_enable(); |
1397 | } | |
e498be7d CL |
1398 | /* 5) Replace the bootstrap kmem_list3's */ |
1399 | { | |
1400 | int node; | |
1401 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1402 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1403 | numa_node_id()); |
e498be7d CL |
1404 | |
1405 | for_each_online_node(node) { | |
1406 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1407 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1408 | |
1409 | if (INDEX_AC != INDEX_L3) { | |
1410 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1411 | &initkmem_list3[SIZE_L3 + node], |
1412 | node); | |
e498be7d CL |
1413 | } |
1414 | } | |
1415 | } | |
1da177e4 | 1416 | |
e498be7d | 1417 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1418 | { |
343e0d7a | 1419 | struct kmem_cache *cachep; |
fc0abb14 | 1420 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1421 | list_for_each_entry(cachep, &cache_chain, next) |
a737b3e2 | 1422 | enable_cpucache(cachep); |
fc0abb14 | 1423 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1424 | } |
1425 | ||
1426 | /* Done! */ | |
1427 | g_cpucache_up = FULL; | |
1428 | ||
a737b3e2 AM |
1429 | /* |
1430 | * Register a cpu startup notifier callback that initializes | |
1431 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1432 | */ |
1433 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1434 | |
a737b3e2 AM |
1435 | /* |
1436 | * The reap timers are started later, with a module init call: That part | |
1437 | * of the kernel is not yet operational. | |
1da177e4 LT |
1438 | */ |
1439 | } | |
1440 | ||
1441 | static int __init cpucache_init(void) | |
1442 | { | |
1443 | int cpu; | |
1444 | ||
a737b3e2 AM |
1445 | /* |
1446 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1447 | */ |
e498be7d | 1448 | for_each_online_cpu(cpu) |
a737b3e2 | 1449 | start_cpu_timer(cpu); |
1da177e4 LT |
1450 | return 0; |
1451 | } | |
1da177e4 LT |
1452 | __initcall(cpucache_init); |
1453 | ||
1454 | /* | |
1455 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1456 | * | |
1457 | * If we requested dmaable memory, we will get it. Even if we | |
1458 | * did not request dmaable memory, we might get it, but that | |
1459 | * would be relatively rare and ignorable. | |
1460 | */ | |
343e0d7a | 1461 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1462 | { |
1463 | struct page *page; | |
1464 | void *addr; | |
1465 | int i; | |
1466 | ||
1467 | flags |= cachep->gfpflags; | |
d6fef9da LY |
1468 | #ifndef CONFIG_MMU |
1469 | /* nommu uses slab's for process anonymous memory allocations, so | |
1470 | * requires __GFP_COMP to properly refcount higher order allocations" | |
1471 | */ | |
1472 | page = alloc_pages_node(nodeid, (flags | __GFP_COMP), cachep->gfporder); | |
1473 | #else | |
50c85a19 | 1474 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); |
d6fef9da | 1475 | #endif |
1da177e4 LT |
1476 | if (!page) |
1477 | return NULL; | |
1478 | addr = page_address(page); | |
1479 | ||
1480 | i = (1 << cachep->gfporder); | |
1481 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1482 | atomic_add(i, &slab_reclaim_pages); | |
1483 | add_page_state(nr_slab, i); | |
1484 | while (i--) { | |
f205b2fe | 1485 | __SetPageSlab(page); |
1da177e4 LT |
1486 | page++; |
1487 | } | |
1488 | return addr; | |
1489 | } | |
1490 | ||
1491 | /* | |
1492 | * Interface to system's page release. | |
1493 | */ | |
343e0d7a | 1494 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1495 | { |
b28a02de | 1496 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1497 | struct page *page = virt_to_page(addr); |
1498 | const unsigned long nr_freed = i; | |
1499 | ||
1500 | while (i--) { | |
f205b2fe NP |
1501 | BUG_ON(!PageSlab(page)); |
1502 | __ClearPageSlab(page); | |
1da177e4 LT |
1503 | page++; |
1504 | } | |
1505 | sub_page_state(nr_slab, nr_freed); | |
1506 | if (current->reclaim_state) | |
1507 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1508 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1509 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1510 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1511 | } |
1512 | ||
1513 | static void kmem_rcu_free(struct rcu_head *head) | |
1514 | { | |
b28a02de | 1515 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1516 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1517 | |
1518 | kmem_freepages(cachep, slab_rcu->addr); | |
1519 | if (OFF_SLAB(cachep)) | |
1520 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1521 | } | |
1522 | ||
1523 | #if DEBUG | |
1524 | ||
1525 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1526 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1527 | unsigned long caller) |
1da177e4 | 1528 | { |
3dafccf2 | 1529 | int size = obj_size(cachep); |
1da177e4 | 1530 | |
3dafccf2 | 1531 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1532 | |
b28a02de | 1533 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1534 | return; |
1535 | ||
b28a02de PE |
1536 | *addr++ = 0x12345678; |
1537 | *addr++ = caller; | |
1538 | *addr++ = smp_processor_id(); | |
1539 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1540 | { |
1541 | unsigned long *sptr = &caller; | |
1542 | unsigned long svalue; | |
1543 | ||
1544 | while (!kstack_end(sptr)) { | |
1545 | svalue = *sptr++; | |
1546 | if (kernel_text_address(svalue)) { | |
b28a02de | 1547 | *addr++ = svalue; |
1da177e4 LT |
1548 | size -= sizeof(unsigned long); |
1549 | if (size <= sizeof(unsigned long)) | |
1550 | break; | |
1551 | } | |
1552 | } | |
1553 | ||
1554 | } | |
b28a02de | 1555 | *addr++ = 0x87654321; |
1da177e4 LT |
1556 | } |
1557 | #endif | |
1558 | ||
343e0d7a | 1559 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1560 | { |
3dafccf2 MS |
1561 | int size = obj_size(cachep); |
1562 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1563 | |
1564 | memset(addr, val, size); | |
b28a02de | 1565 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1566 | } |
1567 | ||
1568 | static void dump_line(char *data, int offset, int limit) | |
1569 | { | |
1570 | int i; | |
1571 | printk(KERN_ERR "%03x:", offset); | |
a737b3e2 | 1572 | for (i = 0; i < limit; i++) |
b28a02de | 1573 | printk(" %02x", (unsigned char)data[offset + i]); |
1da177e4 LT |
1574 | printk("\n"); |
1575 | } | |
1576 | #endif | |
1577 | ||
1578 | #if DEBUG | |
1579 | ||
343e0d7a | 1580 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1581 | { |
1582 | int i, size; | |
1583 | char *realobj; | |
1584 | ||
1585 | if (cachep->flags & SLAB_RED_ZONE) { | |
1586 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
a737b3e2 AM |
1587 | *dbg_redzone1(cachep, objp), |
1588 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1589 | } |
1590 | ||
1591 | if (cachep->flags & SLAB_STORE_USER) { | |
1592 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1593 | *dbg_userword(cachep, objp)); |
1da177e4 | 1594 | print_symbol("(%s)", |
a737b3e2 | 1595 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1596 | printk("\n"); |
1597 | } | |
3dafccf2 MS |
1598 | realobj = (char *)objp + obj_offset(cachep); |
1599 | size = obj_size(cachep); | |
b28a02de | 1600 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1601 | int limit; |
1602 | limit = 16; | |
b28a02de PE |
1603 | if (i + limit > size) |
1604 | limit = size - i; | |
1da177e4 LT |
1605 | dump_line(realobj, i, limit); |
1606 | } | |
1607 | } | |
1608 | ||
343e0d7a | 1609 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1610 | { |
1611 | char *realobj; | |
1612 | int size, i; | |
1613 | int lines = 0; | |
1614 | ||
3dafccf2 MS |
1615 | realobj = (char *)objp + obj_offset(cachep); |
1616 | size = obj_size(cachep); | |
1da177e4 | 1617 | |
b28a02de | 1618 | for (i = 0; i < size; i++) { |
1da177e4 | 1619 | char exp = POISON_FREE; |
b28a02de | 1620 | if (i == size - 1) |
1da177e4 LT |
1621 | exp = POISON_END; |
1622 | if (realobj[i] != exp) { | |
1623 | int limit; | |
1624 | /* Mismatch ! */ | |
1625 | /* Print header */ | |
1626 | if (lines == 0) { | |
b28a02de | 1627 | printk(KERN_ERR |
a737b3e2 AM |
1628 | "Slab corruption: start=%p, len=%d\n", |
1629 | realobj, size); | |
1da177e4 LT |
1630 | print_objinfo(cachep, objp, 0); |
1631 | } | |
1632 | /* Hexdump the affected line */ | |
b28a02de | 1633 | i = (i / 16) * 16; |
1da177e4 | 1634 | limit = 16; |
b28a02de PE |
1635 | if (i + limit > size) |
1636 | limit = size - i; | |
1da177e4 LT |
1637 | dump_line(realobj, i, limit); |
1638 | i += 16; | |
1639 | lines++; | |
1640 | /* Limit to 5 lines */ | |
1641 | if (lines > 5) | |
1642 | break; | |
1643 | } | |
1644 | } | |
1645 | if (lines != 0) { | |
1646 | /* Print some data about the neighboring objects, if they | |
1647 | * exist: | |
1648 | */ | |
6ed5eb22 | 1649 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1650 | unsigned int objnr; |
1da177e4 | 1651 | |
8fea4e96 | 1652 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1653 | if (objnr) { |
8fea4e96 | 1654 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1655 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1656 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1657 | realobj, size); |
1da177e4 LT |
1658 | print_objinfo(cachep, objp, 2); |
1659 | } | |
b28a02de | 1660 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1661 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1662 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1663 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1664 | realobj, size); |
1da177e4 LT |
1665 | print_objinfo(cachep, objp, 2); |
1666 | } | |
1667 | } | |
1668 | } | |
1669 | #endif | |
1670 | ||
12dd36fa MD |
1671 | #if DEBUG |
1672 | /** | |
911851e6 RD |
1673 | * slab_destroy_objs - destroy a slab and its objects |
1674 | * @cachep: cache pointer being destroyed | |
1675 | * @slabp: slab pointer being destroyed | |
1676 | * | |
1677 | * Call the registered destructor for each object in a slab that is being | |
1678 | * destroyed. | |
1da177e4 | 1679 | */ |
343e0d7a | 1680 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1681 | { |
1da177e4 LT |
1682 | int i; |
1683 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1684 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1685 | |
1686 | if (cachep->flags & SLAB_POISON) { | |
1687 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1688 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1689 | OFF_SLAB(cachep)) | |
b28a02de | 1690 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1691 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1692 | else |
1693 | check_poison_obj(cachep, objp); | |
1694 | #else | |
1695 | check_poison_obj(cachep, objp); | |
1696 | #endif | |
1697 | } | |
1698 | if (cachep->flags & SLAB_RED_ZONE) { | |
1699 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1700 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1701 | "was overwritten"); |
1da177e4 LT |
1702 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1703 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1704 | "was overwritten"); |
1da177e4 LT |
1705 | } |
1706 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1707 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1708 | } |
12dd36fa | 1709 | } |
1da177e4 | 1710 | #else |
343e0d7a | 1711 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1712 | { |
1da177e4 LT |
1713 | if (cachep->dtor) { |
1714 | int i; | |
1715 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1716 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1717 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1718 | } |
1719 | } | |
12dd36fa | 1720 | } |
1da177e4 LT |
1721 | #endif |
1722 | ||
911851e6 RD |
1723 | /** |
1724 | * slab_destroy - destroy and release all objects in a slab | |
1725 | * @cachep: cache pointer being destroyed | |
1726 | * @slabp: slab pointer being destroyed | |
1727 | * | |
12dd36fa | 1728 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1729 | * Before calling the slab must have been unlinked from the cache. The |
1730 | * cache-lock is not held/needed. | |
12dd36fa | 1731 | */ |
343e0d7a | 1732 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1733 | { |
1734 | void *addr = slabp->s_mem - slabp->colouroff; | |
1735 | ||
1736 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1737 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1738 | struct slab_rcu *slab_rcu; | |
1739 | ||
b28a02de | 1740 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1741 | slab_rcu->cachep = cachep; |
1742 | slab_rcu->addr = addr; | |
1743 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1744 | } else { | |
1745 | kmem_freepages(cachep, addr); | |
1746 | if (OFF_SLAB(cachep)) | |
1747 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1748 | } | |
1749 | } | |
1750 | ||
a737b3e2 AM |
1751 | /* |
1752 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1753 | * size of kmem_list3. | |
1754 | */ | |
343e0d7a | 1755 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1756 | { |
1757 | int node; | |
1758 | ||
1759 | for_each_online_node(node) { | |
b28a02de | 1760 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1761 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1762 | REAPTIMEOUT_LIST3 + |
1763 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1764 | } |
1765 | } | |
1766 | ||
4d268eba | 1767 | /** |
a70773dd RD |
1768 | * calculate_slab_order - calculate size (page order) of slabs |
1769 | * @cachep: pointer to the cache that is being created | |
1770 | * @size: size of objects to be created in this cache. | |
1771 | * @align: required alignment for the objects. | |
1772 | * @flags: slab allocation flags | |
1773 | * | |
1774 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1775 | * |
1776 | * This could be made much more intelligent. For now, try to avoid using | |
1777 | * high order pages for slabs. When the gfp() functions are more friendly | |
1778 | * towards high-order requests, this should be changed. | |
1779 | */ | |
a737b3e2 | 1780 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1781 | size_t size, size_t align, unsigned long flags) |
4d268eba PE |
1782 | { |
1783 | size_t left_over = 0; | |
9888e6fa | 1784 | int gfporder; |
4d268eba | 1785 | |
a737b3e2 | 1786 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1787 | unsigned int num; |
1788 | size_t remainder; | |
1789 | ||
9888e6fa | 1790 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1791 | if (!num) |
1792 | continue; | |
9888e6fa | 1793 | |
4d268eba | 1794 | /* More than offslab_limit objects will cause problems */ |
9888e6fa | 1795 | if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit) |
4d268eba PE |
1796 | break; |
1797 | ||
9888e6fa | 1798 | /* Found something acceptable - save it away */ |
4d268eba | 1799 | cachep->num = num; |
9888e6fa | 1800 | cachep->gfporder = gfporder; |
4d268eba PE |
1801 | left_over = remainder; |
1802 | ||
f78bb8ad LT |
1803 | /* |
1804 | * A VFS-reclaimable slab tends to have most allocations | |
1805 | * as GFP_NOFS and we really don't want to have to be allocating | |
1806 | * higher-order pages when we are unable to shrink dcache. | |
1807 | */ | |
1808 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1809 | break; | |
1810 | ||
4d268eba PE |
1811 | /* |
1812 | * Large number of objects is good, but very large slabs are | |
1813 | * currently bad for the gfp()s. | |
1814 | */ | |
9888e6fa | 1815 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1816 | break; |
1817 | ||
9888e6fa LT |
1818 | /* |
1819 | * Acceptable internal fragmentation? | |
1820 | */ | |
a737b3e2 | 1821 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
1822 | break; |
1823 | } | |
1824 | return left_over; | |
1825 | } | |
1826 | ||
f30cf7d1 PE |
1827 | static void setup_cpu_cache(struct kmem_cache *cachep) |
1828 | { | |
1829 | if (g_cpucache_up == FULL) { | |
1830 | enable_cpucache(cachep); | |
1831 | return; | |
1832 | } | |
1833 | if (g_cpucache_up == NONE) { | |
1834 | /* | |
1835 | * Note: the first kmem_cache_create must create the cache | |
1836 | * that's used by kmalloc(24), otherwise the creation of | |
1837 | * further caches will BUG(). | |
1838 | */ | |
1839 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
1840 | ||
1841 | /* | |
1842 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
1843 | * the first cache, then we need to set up all its list3s, | |
1844 | * otherwise the creation of further caches will BUG(). | |
1845 | */ | |
1846 | set_up_list3s(cachep, SIZE_AC); | |
1847 | if (INDEX_AC == INDEX_L3) | |
1848 | g_cpucache_up = PARTIAL_L3; | |
1849 | else | |
1850 | g_cpucache_up = PARTIAL_AC; | |
1851 | } else { | |
1852 | cachep->array[smp_processor_id()] = | |
1853 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
1854 | ||
1855 | if (g_cpucache_up == PARTIAL_AC) { | |
1856 | set_up_list3s(cachep, SIZE_L3); | |
1857 | g_cpucache_up = PARTIAL_L3; | |
1858 | } else { | |
1859 | int node; | |
1860 | for_each_online_node(node) { | |
1861 | cachep->nodelists[node] = | |
1862 | kmalloc_node(sizeof(struct kmem_list3), | |
1863 | GFP_KERNEL, node); | |
1864 | BUG_ON(!cachep->nodelists[node]); | |
1865 | kmem_list3_init(cachep->nodelists[node]); | |
1866 | } | |
1867 | } | |
1868 | } | |
1869 | cachep->nodelists[numa_node_id()]->next_reap = | |
1870 | jiffies + REAPTIMEOUT_LIST3 + | |
1871 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1872 | ||
1873 | cpu_cache_get(cachep)->avail = 0; | |
1874 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1875 | cpu_cache_get(cachep)->batchcount = 1; | |
1876 | cpu_cache_get(cachep)->touched = 0; | |
1877 | cachep->batchcount = 1; | |
1878 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
1879 | } | |
1880 | ||
1da177e4 LT |
1881 | /** |
1882 | * kmem_cache_create - Create a cache. | |
1883 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1884 | * @size: The size of objects to be created in this cache. | |
1885 | * @align: The required alignment for the objects. | |
1886 | * @flags: SLAB flags | |
1887 | * @ctor: A constructor for the objects. | |
1888 | * @dtor: A destructor for the objects. | |
1889 | * | |
1890 | * Returns a ptr to the cache on success, NULL on failure. | |
1891 | * Cannot be called within a int, but can be interrupted. | |
1892 | * The @ctor is run when new pages are allocated by the cache | |
1893 | * and the @dtor is run before the pages are handed back. | |
1894 | * | |
1895 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
1896 | * the module calling this has to destroy the cache before getting unloaded. |
1897 | * | |
1da177e4 LT |
1898 | * The flags are |
1899 | * | |
1900 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1901 | * to catch references to uninitialised memory. | |
1902 | * | |
1903 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1904 | * for buffer overruns. | |
1905 | * | |
1da177e4 LT |
1906 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
1907 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1908 | * as davem. | |
1909 | */ | |
343e0d7a | 1910 | struct kmem_cache * |
1da177e4 | 1911 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
1912 | unsigned long flags, |
1913 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 1914 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
1915 | { |
1916 | size_t left_over, slab_size, ralign; | |
343e0d7a | 1917 | struct kmem_cache *cachep = NULL; |
4f12bb4f | 1918 | struct list_head *p; |
1da177e4 LT |
1919 | |
1920 | /* | |
1921 | * Sanity checks... these are all serious usage bugs. | |
1922 | */ | |
a737b3e2 | 1923 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
b28a02de | 1924 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { |
a737b3e2 AM |
1925 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
1926 | name); | |
b28a02de PE |
1927 | BUG(); |
1928 | } | |
1da177e4 | 1929 | |
f0188f47 RT |
1930 | /* |
1931 | * Prevent CPUs from coming and going. | |
1932 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
1933 | */ | |
1934 | lock_cpu_hotplug(); | |
1935 | ||
fc0abb14 | 1936 | mutex_lock(&cache_chain_mutex); |
4f12bb4f AM |
1937 | |
1938 | list_for_each(p, &cache_chain) { | |
343e0d7a | 1939 | struct kmem_cache *pc = list_entry(p, struct kmem_cache, next); |
4f12bb4f AM |
1940 | mm_segment_t old_fs = get_fs(); |
1941 | char tmp; | |
1942 | int res; | |
1943 | ||
1944 | /* | |
1945 | * This happens when the module gets unloaded and doesn't | |
1946 | * destroy its slab cache and no-one else reuses the vmalloc | |
1947 | * area of the module. Print a warning. | |
1948 | */ | |
1949 | set_fs(KERNEL_DS); | |
1950 | res = __get_user(tmp, pc->name); | |
1951 | set_fs(old_fs); | |
1952 | if (res) { | |
1953 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 1954 | pc->buffer_size); |
4f12bb4f AM |
1955 | continue; |
1956 | } | |
1957 | ||
b28a02de | 1958 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1959 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1960 | dump_stack(); | |
1961 | goto oops; | |
1962 | } | |
1963 | } | |
1964 | ||
1da177e4 LT |
1965 | #if DEBUG |
1966 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1967 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1968 | /* No constructor, but inital state check requested */ | |
1969 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 1970 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
1971 | flags &= ~SLAB_DEBUG_INITIAL; |
1972 | } | |
1da177e4 LT |
1973 | #if FORCED_DEBUG |
1974 | /* | |
1975 | * Enable redzoning and last user accounting, except for caches with | |
1976 | * large objects, if the increased size would increase the object size | |
1977 | * above the next power of two: caches with object sizes just above a | |
1978 | * power of two have a significant amount of internal fragmentation. | |
1979 | */ | |
a737b3e2 | 1980 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 1981 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
1982 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
1983 | flags |= SLAB_POISON; | |
1984 | #endif | |
1985 | if (flags & SLAB_DESTROY_BY_RCU) | |
1986 | BUG_ON(flags & SLAB_POISON); | |
1987 | #endif | |
1988 | if (flags & SLAB_DESTROY_BY_RCU) | |
1989 | BUG_ON(dtor); | |
1990 | ||
1991 | /* | |
a737b3e2 AM |
1992 | * Always checks flags, a caller might be expecting debug support which |
1993 | * isn't available. | |
1da177e4 | 1994 | */ |
40094fa6 | 1995 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 1996 | |
a737b3e2 AM |
1997 | /* |
1998 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
1999 | * unaligned accesses for some archs when redzoning is used, and makes |
2000 | * sure any on-slab bufctl's are also correctly aligned. | |
2001 | */ | |
b28a02de PE |
2002 | if (size & (BYTES_PER_WORD - 1)) { |
2003 | size += (BYTES_PER_WORD - 1); | |
2004 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2005 | } |
2006 | ||
a737b3e2 AM |
2007 | /* calculate the final buffer alignment: */ |
2008 | ||
1da177e4 LT |
2009 | /* 1) arch recommendation: can be overridden for debug */ |
2010 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2011 | /* |
2012 | * Default alignment: as specified by the arch code. Except if | |
2013 | * an object is really small, then squeeze multiple objects into | |
2014 | * one cacheline. | |
1da177e4 LT |
2015 | */ |
2016 | ralign = cache_line_size(); | |
b28a02de | 2017 | while (size <= ralign / 2) |
1da177e4 LT |
2018 | ralign /= 2; |
2019 | } else { | |
2020 | ralign = BYTES_PER_WORD; | |
2021 | } | |
2022 | /* 2) arch mandated alignment: disables debug if necessary */ | |
2023 | if (ralign < ARCH_SLAB_MINALIGN) { | |
2024 | ralign = ARCH_SLAB_MINALIGN; | |
2025 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2026 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
2027 | } |
2028 | /* 3) caller mandated alignment: disables debug if necessary */ | |
2029 | if (ralign < align) { | |
2030 | ralign = align; | |
2031 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2032 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 | 2033 | } |
a737b3e2 AM |
2034 | /* |
2035 | * 4) Store it. Note that the debug code below can reduce | |
1da177e4 LT |
2036 | * the alignment to BYTES_PER_WORD. |
2037 | */ | |
2038 | align = ralign; | |
2039 | ||
2040 | /* Get cache's description obj. */ | |
c5e3b83e | 2041 | cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 2042 | if (!cachep) |
4f12bb4f | 2043 | goto oops; |
1da177e4 LT |
2044 | |
2045 | #if DEBUG | |
3dafccf2 | 2046 | cachep->obj_size = size; |
1da177e4 LT |
2047 | |
2048 | if (flags & SLAB_RED_ZONE) { | |
2049 | /* redzoning only works with word aligned caches */ | |
2050 | align = BYTES_PER_WORD; | |
2051 | ||
2052 | /* add space for red zone words */ | |
3dafccf2 | 2053 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 2054 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
2055 | } |
2056 | if (flags & SLAB_STORE_USER) { | |
2057 | /* user store requires word alignment and | |
2058 | * one word storage behind the end of the real | |
2059 | * object. | |
2060 | */ | |
2061 | align = BYTES_PER_WORD; | |
2062 | size += BYTES_PER_WORD; | |
2063 | } | |
2064 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2065 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2066 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2067 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2068 | size = PAGE_SIZE; |
2069 | } | |
2070 | #endif | |
2071 | #endif | |
2072 | ||
2073 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 2074 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
2075 | /* |
2076 | * Size is large, assume best to place the slab management obj | |
2077 | * off-slab (should allow better packing of objs). | |
2078 | */ | |
2079 | flags |= CFLGS_OFF_SLAB; | |
2080 | ||
2081 | size = ALIGN(size, align); | |
2082 | ||
f78bb8ad | 2083 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2084 | |
2085 | if (!cachep->num) { | |
2086 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
2087 | kmem_cache_free(&cache_cache, cachep); | |
2088 | cachep = NULL; | |
4f12bb4f | 2089 | goto oops; |
1da177e4 | 2090 | } |
b28a02de PE |
2091 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2092 | + sizeof(struct slab), align); | |
1da177e4 LT |
2093 | |
2094 | /* | |
2095 | * If the slab has been placed off-slab, and we have enough space then | |
2096 | * move it on-slab. This is at the expense of any extra colouring. | |
2097 | */ | |
2098 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2099 | flags &= ~CFLGS_OFF_SLAB; | |
2100 | left_over -= slab_size; | |
2101 | } | |
2102 | ||
2103 | if (flags & CFLGS_OFF_SLAB) { | |
2104 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2105 | slab_size = |
2106 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2107 | } |
2108 | ||
2109 | cachep->colour_off = cache_line_size(); | |
2110 | /* Offset must be a multiple of the alignment. */ | |
2111 | if (cachep->colour_off < align) | |
2112 | cachep->colour_off = align; | |
b28a02de | 2113 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2114 | cachep->slab_size = slab_size; |
2115 | cachep->flags = flags; | |
2116 | cachep->gfpflags = 0; | |
2117 | if (flags & SLAB_CACHE_DMA) | |
2118 | cachep->gfpflags |= GFP_DMA; | |
3dafccf2 | 2119 | cachep->buffer_size = size; |
1da177e4 LT |
2120 | |
2121 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 2122 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
2123 | cachep->ctor = ctor; |
2124 | cachep->dtor = dtor; | |
2125 | cachep->name = name; | |
2126 | ||
1da177e4 | 2127 | |
f30cf7d1 | 2128 | setup_cpu_cache(cachep); |
1da177e4 | 2129 | |
1da177e4 LT |
2130 | /* cache setup completed, link it into the list */ |
2131 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2132 | oops: |
1da177e4 LT |
2133 | if (!cachep && (flags & SLAB_PANIC)) |
2134 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2135 | name); |
fc0abb14 | 2136 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 2137 | unlock_cpu_hotplug(); |
1da177e4 LT |
2138 | return cachep; |
2139 | } | |
2140 | EXPORT_SYMBOL(kmem_cache_create); | |
2141 | ||
2142 | #if DEBUG | |
2143 | static void check_irq_off(void) | |
2144 | { | |
2145 | BUG_ON(!irqs_disabled()); | |
2146 | } | |
2147 | ||
2148 | static void check_irq_on(void) | |
2149 | { | |
2150 | BUG_ON(irqs_disabled()); | |
2151 | } | |
2152 | ||
343e0d7a | 2153 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2154 | { |
2155 | #ifdef CONFIG_SMP | |
2156 | check_irq_off(); | |
e498be7d | 2157 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2158 | #endif |
2159 | } | |
e498be7d | 2160 | |
343e0d7a | 2161 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2162 | { |
2163 | #ifdef CONFIG_SMP | |
2164 | check_irq_off(); | |
2165 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2166 | #endif | |
2167 | } | |
2168 | ||
1da177e4 LT |
2169 | #else |
2170 | #define check_irq_off() do { } while(0) | |
2171 | #define check_irq_on() do { } while(0) | |
2172 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2173 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2174 | #endif |
2175 | ||
aab2207c CL |
2176 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2177 | struct array_cache *ac, | |
2178 | int force, int node); | |
2179 | ||
1da177e4 LT |
2180 | static void do_drain(void *arg) |
2181 | { | |
a737b3e2 | 2182 | struct kmem_cache *cachep = arg; |
1da177e4 | 2183 | struct array_cache *ac; |
ff69416e | 2184 | int node = numa_node_id(); |
1da177e4 LT |
2185 | |
2186 | check_irq_off(); | |
9a2dba4b | 2187 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2188 | spin_lock(&cachep->nodelists[node]->list_lock); |
2189 | free_block(cachep, ac->entry, ac->avail, node); | |
2190 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2191 | ac->avail = 0; |
2192 | } | |
2193 | ||
343e0d7a | 2194 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2195 | { |
e498be7d CL |
2196 | struct kmem_list3 *l3; |
2197 | int node; | |
2198 | ||
a07fa394 | 2199 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2200 | check_irq_on(); |
b28a02de | 2201 | for_each_online_node(node) { |
e498be7d | 2202 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2203 | if (l3 && l3->alien) |
2204 | drain_alien_cache(cachep, l3->alien); | |
2205 | } | |
2206 | ||
2207 | for_each_online_node(node) { | |
2208 | l3 = cachep->nodelists[node]; | |
2209 | if (l3) | |
aab2207c | 2210 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2211 | } |
1da177e4 LT |
2212 | } |
2213 | ||
343e0d7a | 2214 | static int __node_shrink(struct kmem_cache *cachep, int node) |
1da177e4 LT |
2215 | { |
2216 | struct slab *slabp; | |
e498be7d | 2217 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
2218 | int ret; |
2219 | ||
e498be7d | 2220 | for (;;) { |
1da177e4 LT |
2221 | struct list_head *p; |
2222 | ||
e498be7d CL |
2223 | p = l3->slabs_free.prev; |
2224 | if (p == &l3->slabs_free) | |
1da177e4 LT |
2225 | break; |
2226 | ||
e498be7d | 2227 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 | 2228 | #if DEBUG |
40094fa6 | 2229 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2230 | #endif |
2231 | list_del(&slabp->list); | |
2232 | ||
e498be7d CL |
2233 | l3->free_objects -= cachep->num; |
2234 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 2235 | slab_destroy(cachep, slabp); |
e498be7d | 2236 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 2237 | } |
b28a02de | 2238 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
2239 | return ret; |
2240 | } | |
2241 | ||
343e0d7a | 2242 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2243 | { |
2244 | int ret = 0, i = 0; | |
2245 | struct kmem_list3 *l3; | |
2246 | ||
2247 | drain_cpu_caches(cachep); | |
2248 | ||
2249 | check_irq_on(); | |
2250 | for_each_online_node(i) { | |
2251 | l3 = cachep->nodelists[i]; | |
2252 | if (l3) { | |
2253 | spin_lock_irq(&l3->list_lock); | |
2254 | ret += __node_shrink(cachep, i); | |
2255 | spin_unlock_irq(&l3->list_lock); | |
2256 | } | |
2257 | } | |
2258 | return (ret ? 1 : 0); | |
2259 | } | |
2260 | ||
1da177e4 LT |
2261 | /** |
2262 | * kmem_cache_shrink - Shrink a cache. | |
2263 | * @cachep: The cache to shrink. | |
2264 | * | |
2265 | * Releases as many slabs as possible for a cache. | |
2266 | * To help debugging, a zero exit status indicates all slabs were released. | |
2267 | */ | |
343e0d7a | 2268 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2269 | { |
40094fa6 | 2270 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2271 | |
2272 | return __cache_shrink(cachep); | |
2273 | } | |
2274 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2275 | ||
2276 | /** | |
2277 | * kmem_cache_destroy - delete a cache | |
2278 | * @cachep: the cache to destroy | |
2279 | * | |
343e0d7a | 2280 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2281 | * Returns 0 on success. |
2282 | * | |
2283 | * It is expected this function will be called by a module when it is | |
2284 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2285 | * cache being allocated each time a module is loaded and unloaded, if the | |
2286 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2287 | * | |
2288 | * The cache must be empty before calling this function. | |
2289 | * | |
2290 | * The caller must guarantee that noone will allocate memory from the cache | |
2291 | * during the kmem_cache_destroy(). | |
2292 | */ | |
343e0d7a | 2293 | int kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 LT |
2294 | { |
2295 | int i; | |
e498be7d | 2296 | struct kmem_list3 *l3; |
1da177e4 | 2297 | |
40094fa6 | 2298 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2299 | |
2300 | /* Don't let CPUs to come and go */ | |
2301 | lock_cpu_hotplug(); | |
2302 | ||
2303 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2304 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2305 | /* |
2306 | * the chain is never empty, cache_cache is never destroyed | |
2307 | */ | |
2308 | list_del(&cachep->next); | |
fc0abb14 | 2309 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2310 | |
2311 | if (__cache_shrink(cachep)) { | |
2312 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2313 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2314 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2315 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2316 | unlock_cpu_hotplug(); |
2317 | return 1; | |
2318 | } | |
2319 | ||
2320 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2321 | synchronize_rcu(); |
1da177e4 | 2322 | |
e498be7d | 2323 | for_each_online_cpu(i) |
b28a02de | 2324 | kfree(cachep->array[i]); |
1da177e4 LT |
2325 | |
2326 | /* NUMA: free the list3 structures */ | |
e498be7d | 2327 | for_each_online_node(i) { |
a737b3e2 AM |
2328 | l3 = cachep->nodelists[i]; |
2329 | if (l3) { | |
e498be7d CL |
2330 | kfree(l3->shared); |
2331 | free_alien_cache(l3->alien); | |
2332 | kfree(l3); | |
2333 | } | |
2334 | } | |
1da177e4 | 2335 | kmem_cache_free(&cache_cache, cachep); |
1da177e4 | 2336 | unlock_cpu_hotplug(); |
1da177e4 LT |
2337 | return 0; |
2338 | } | |
2339 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2340 | ||
2341 | /* Get the memory for a slab management obj. */ | |
343e0d7a | 2342 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2343 | int colour_off, gfp_t local_flags, |
2344 | int nodeid) | |
1da177e4 LT |
2345 | { |
2346 | struct slab *slabp; | |
b28a02de | 2347 | |
1da177e4 LT |
2348 | if (OFF_SLAB(cachep)) { |
2349 | /* Slab management obj is off-slab. */ | |
5b74ada7 RT |
2350 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
2351 | local_flags, nodeid); | |
1da177e4 LT |
2352 | if (!slabp) |
2353 | return NULL; | |
2354 | } else { | |
b28a02de | 2355 | slabp = objp + colour_off; |
1da177e4 LT |
2356 | colour_off += cachep->slab_size; |
2357 | } | |
2358 | slabp->inuse = 0; | |
2359 | slabp->colouroff = colour_off; | |
b28a02de | 2360 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2361 | slabp->nodeid = nodeid; |
1da177e4 LT |
2362 | return slabp; |
2363 | } | |
2364 | ||
2365 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2366 | { | |
b28a02de | 2367 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2368 | } |
2369 | ||
343e0d7a | 2370 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2371 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2372 | { |
2373 | int i; | |
2374 | ||
2375 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2376 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2377 | #if DEBUG |
2378 | /* need to poison the objs? */ | |
2379 | if (cachep->flags & SLAB_POISON) | |
2380 | poison_obj(cachep, objp, POISON_FREE); | |
2381 | if (cachep->flags & SLAB_STORE_USER) | |
2382 | *dbg_userword(cachep, objp) = NULL; | |
2383 | ||
2384 | if (cachep->flags & SLAB_RED_ZONE) { | |
2385 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2386 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2387 | } | |
2388 | /* | |
a737b3e2 AM |
2389 | * Constructors are not allowed to allocate memory from the same |
2390 | * cache which they are a constructor for. Otherwise, deadlock. | |
2391 | * They must also be threaded. | |
1da177e4 LT |
2392 | */ |
2393 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2394 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2395 | ctor_flags); |
1da177e4 LT |
2396 | |
2397 | if (cachep->flags & SLAB_RED_ZONE) { | |
2398 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2399 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2400 | " end of an object"); |
1da177e4 LT |
2401 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2402 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2403 | " start of an object"); |
1da177e4 | 2404 | } |
a737b3e2 AM |
2405 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2406 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2407 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2408 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2409 | #else |
2410 | if (cachep->ctor) | |
2411 | cachep->ctor(objp, cachep, ctor_flags); | |
2412 | #endif | |
b28a02de | 2413 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2414 | } |
b28a02de | 2415 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2416 | slabp->free = 0; |
2417 | } | |
2418 | ||
343e0d7a | 2419 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2420 | { |
a737b3e2 AM |
2421 | if (flags & SLAB_DMA) |
2422 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2423 | else | |
2424 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
1da177e4 LT |
2425 | } |
2426 | ||
a737b3e2 AM |
2427 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2428 | int nodeid) | |
78d382d7 | 2429 | { |
8fea4e96 | 2430 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2431 | kmem_bufctl_t next; |
2432 | ||
2433 | slabp->inuse++; | |
2434 | next = slab_bufctl(slabp)[slabp->free]; | |
2435 | #if DEBUG | |
2436 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2437 | WARN_ON(slabp->nodeid != nodeid); | |
2438 | #endif | |
2439 | slabp->free = next; | |
2440 | ||
2441 | return objp; | |
2442 | } | |
2443 | ||
a737b3e2 AM |
2444 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2445 | void *objp, int nodeid) | |
78d382d7 | 2446 | { |
8fea4e96 | 2447 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2448 | |
2449 | #if DEBUG | |
2450 | /* Verify that the slab belongs to the intended node */ | |
2451 | WARN_ON(slabp->nodeid != nodeid); | |
2452 | ||
871751e2 | 2453 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2454 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2455 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2456 | BUG(); |
2457 | } | |
2458 | #endif | |
2459 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2460 | slabp->free = objnr; | |
2461 | slabp->inuse--; | |
2462 | } | |
2463 | ||
a737b3e2 AM |
2464 | static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, |
2465 | void *objp) | |
1da177e4 LT |
2466 | { |
2467 | int i; | |
2468 | struct page *page; | |
2469 | ||
2470 | /* Nasty!!!!!! I hope this is OK. */ | |
1da177e4 | 2471 | page = virt_to_page(objp); |
84097518 NP |
2472 | |
2473 | i = 1; | |
2474 | if (likely(!PageCompound(page))) | |
2475 | i <<= cachep->gfporder; | |
1da177e4 | 2476 | do { |
065d41cb PE |
2477 | page_set_cache(page, cachep); |
2478 | page_set_slab(page, slabp); | |
1da177e4 LT |
2479 | page++; |
2480 | } while (--i); | |
2481 | } | |
2482 | ||
2483 | /* | |
2484 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2485 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2486 | */ | |
343e0d7a | 2487 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2488 | { |
b28a02de PE |
2489 | struct slab *slabp; |
2490 | void *objp; | |
2491 | size_t offset; | |
2492 | gfp_t local_flags; | |
2493 | unsigned long ctor_flags; | |
e498be7d | 2494 | struct kmem_list3 *l3; |
1da177e4 | 2495 | |
a737b3e2 AM |
2496 | /* |
2497 | * Be lazy and only check for valid flags here, keeping it out of the | |
2498 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2499 | */ |
40094fa6 | 2500 | BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); |
1da177e4 LT |
2501 | if (flags & SLAB_NO_GROW) |
2502 | return 0; | |
2503 | ||
2504 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2505 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2506 | if (!(local_flags & __GFP_WAIT)) | |
2507 | /* | |
2508 | * Not allowed to sleep. Need to tell a constructor about | |
2509 | * this - it might need to know... | |
2510 | */ | |
2511 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2512 | ||
2e1217cf | 2513 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2514 | check_irq_off(); |
2e1217cf RT |
2515 | l3 = cachep->nodelists[nodeid]; |
2516 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2517 | |
2518 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2519 | offset = l3->colour_next; |
2520 | l3->colour_next++; | |
2521 | if (l3->colour_next >= cachep->colour) | |
2522 | l3->colour_next = 0; | |
2523 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2524 | |
2e1217cf | 2525 | offset *= cachep->colour_off; |
1da177e4 LT |
2526 | |
2527 | if (local_flags & __GFP_WAIT) | |
2528 | local_irq_enable(); | |
2529 | ||
2530 | /* | |
2531 | * The test for missing atomic flag is performed here, rather than | |
2532 | * the more obvious place, simply to reduce the critical path length | |
2533 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2534 | * will eventually be caught here (where it matters). | |
2535 | */ | |
2536 | kmem_flagcheck(cachep, flags); | |
2537 | ||
a737b3e2 AM |
2538 | /* |
2539 | * Get mem for the objs. Attempt to allocate a physical page from | |
2540 | * 'nodeid'. | |
e498be7d | 2541 | */ |
a737b3e2 AM |
2542 | objp = kmem_getpages(cachep, flags, nodeid); |
2543 | if (!objp) | |
1da177e4 LT |
2544 | goto failed; |
2545 | ||
2546 | /* Get slab management. */ | |
5b74ada7 | 2547 | slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); |
a737b3e2 | 2548 | if (!slabp) |
1da177e4 LT |
2549 | goto opps1; |
2550 | ||
e498be7d | 2551 | slabp->nodeid = nodeid; |
1da177e4 LT |
2552 | set_slab_attr(cachep, slabp, objp); |
2553 | ||
2554 | cache_init_objs(cachep, slabp, ctor_flags); | |
2555 | ||
2556 | if (local_flags & __GFP_WAIT) | |
2557 | local_irq_disable(); | |
2558 | check_irq_off(); | |
e498be7d | 2559 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2560 | |
2561 | /* Make slab active. */ | |
e498be7d | 2562 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2563 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2564 | l3->free_objects += cachep->num; |
2565 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2566 | return 1; |
a737b3e2 | 2567 | opps1: |
1da177e4 | 2568 | kmem_freepages(cachep, objp); |
a737b3e2 | 2569 | failed: |
1da177e4 LT |
2570 | if (local_flags & __GFP_WAIT) |
2571 | local_irq_disable(); | |
2572 | return 0; | |
2573 | } | |
2574 | ||
2575 | #if DEBUG | |
2576 | ||
2577 | /* | |
2578 | * Perform extra freeing checks: | |
2579 | * - detect bad pointers. | |
2580 | * - POISON/RED_ZONE checking | |
2581 | * - destructor calls, for caches with POISON+dtor | |
2582 | */ | |
2583 | static void kfree_debugcheck(const void *objp) | |
2584 | { | |
2585 | struct page *page; | |
2586 | ||
2587 | if (!virt_addr_valid(objp)) { | |
2588 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2589 | (unsigned long)objp); |
2590 | BUG(); | |
1da177e4 LT |
2591 | } |
2592 | page = virt_to_page(objp); | |
2593 | if (!PageSlab(page)) { | |
b28a02de PE |
2594 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2595 | (unsigned long)objp); | |
1da177e4 LT |
2596 | BUG(); |
2597 | } | |
2598 | } | |
2599 | ||
343e0d7a | 2600 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2601 | void *caller) |
1da177e4 LT |
2602 | { |
2603 | struct page *page; | |
2604 | unsigned int objnr; | |
2605 | struct slab *slabp; | |
2606 | ||
3dafccf2 | 2607 | objp -= obj_offset(cachep); |
1da177e4 LT |
2608 | kfree_debugcheck(objp); |
2609 | page = virt_to_page(objp); | |
2610 | ||
065d41cb | 2611 | if (page_get_cache(page) != cachep) { |
a737b3e2 AM |
2612 | printk(KERN_ERR "mismatch in kmem_cache_free: expected " |
2613 | "cache %p, got %p\n", | |
b28a02de | 2614 | page_get_cache(page), cachep); |
1da177e4 | 2615 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2616 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2617 | page_get_cache(page)->name); | |
1da177e4 LT |
2618 | WARN_ON(1); |
2619 | } | |
065d41cb | 2620 | slabp = page_get_slab(page); |
1da177e4 LT |
2621 | |
2622 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2623 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || |
2624 | *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2625 | slab_error(cachep, "double free, or memory outside" | |
2626 | " object was overwritten"); | |
2627 | printk(KERN_ERR "%p: redzone 1:0x%lx, " | |
2628 | "redzone 2:0x%lx.\n", | |
b28a02de PE |
2629 | objp, *dbg_redzone1(cachep, objp), |
2630 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2631 | } |
2632 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2633 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2634 | } | |
2635 | if (cachep->flags & SLAB_STORE_USER) | |
2636 | *dbg_userword(cachep, objp) = caller; | |
2637 | ||
8fea4e96 | 2638 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2639 | |
2640 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2641 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2642 | |
2643 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
a737b3e2 AM |
2644 | /* |
2645 | * Need to call the slab's constructor so the caller can | |
2646 | * perform a verify of its state (debugging). Called without | |
2647 | * the cache-lock held. | |
1da177e4 | 2648 | */ |
3dafccf2 | 2649 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2650 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2651 | } |
2652 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2653 | /* we want to cache poison the object, | |
2654 | * call the destruction callback | |
2655 | */ | |
3dafccf2 | 2656 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 2657 | } |
871751e2 AV |
2658 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2659 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2660 | #endif | |
1da177e4 LT |
2661 | if (cachep->flags & SLAB_POISON) { |
2662 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2663 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2664 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2665 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2666 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2667 | } else { |
2668 | poison_obj(cachep, objp, POISON_FREE); | |
2669 | } | |
2670 | #else | |
2671 | poison_obj(cachep, objp, POISON_FREE); | |
2672 | #endif | |
2673 | } | |
2674 | return objp; | |
2675 | } | |
2676 | ||
343e0d7a | 2677 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2678 | { |
2679 | kmem_bufctl_t i; | |
2680 | int entries = 0; | |
b28a02de | 2681 | |
1da177e4 LT |
2682 | /* Check slab's freelist to see if this obj is there. */ |
2683 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2684 | entries++; | |
2685 | if (entries > cachep->num || i >= cachep->num) | |
2686 | goto bad; | |
2687 | } | |
2688 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2689 | bad: |
2690 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2691 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2692 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2693 | for (i = 0; |
264132bc | 2694 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2695 | i++) { |
a737b3e2 | 2696 | if (i % 16 == 0) |
1da177e4 | 2697 | printk("\n%03x:", i); |
b28a02de | 2698 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2699 | } |
2700 | printk("\n"); | |
2701 | BUG(); | |
2702 | } | |
2703 | } | |
2704 | #else | |
2705 | #define kfree_debugcheck(x) do { } while(0) | |
2706 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2707 | #define check_slabp(x,y) do { } while(0) | |
2708 | #endif | |
2709 | ||
343e0d7a | 2710 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2711 | { |
2712 | int batchcount; | |
2713 | struct kmem_list3 *l3; | |
2714 | struct array_cache *ac; | |
2715 | ||
2716 | check_irq_off(); | |
9a2dba4b | 2717 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2718 | retry: |
1da177e4 LT |
2719 | batchcount = ac->batchcount; |
2720 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2721 | /* |
2722 | * If there was little recent activity on this cache, then | |
2723 | * perform only a partial refill. Otherwise we could generate | |
2724 | * refill bouncing. | |
1da177e4 LT |
2725 | */ |
2726 | batchcount = BATCHREFILL_LIMIT; | |
2727 | } | |
e498be7d CL |
2728 | l3 = cachep->nodelists[numa_node_id()]; |
2729 | ||
2730 | BUG_ON(ac->avail > 0 || !l3); | |
2731 | spin_lock(&l3->list_lock); | |
1da177e4 | 2732 | |
3ded175a CL |
2733 | /* See if we can refill from the shared array */ |
2734 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2735 | goto alloc_done; | |
2736 | ||
1da177e4 LT |
2737 | while (batchcount > 0) { |
2738 | struct list_head *entry; | |
2739 | struct slab *slabp; | |
2740 | /* Get slab alloc is to come from. */ | |
2741 | entry = l3->slabs_partial.next; | |
2742 | if (entry == &l3->slabs_partial) { | |
2743 | l3->free_touched = 1; | |
2744 | entry = l3->slabs_free.next; | |
2745 | if (entry == &l3->slabs_free) | |
2746 | goto must_grow; | |
2747 | } | |
2748 | ||
2749 | slabp = list_entry(entry, struct slab, list); | |
2750 | check_slabp(cachep, slabp); | |
2751 | check_spinlock_acquired(cachep); | |
2752 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2753 | STATS_INC_ALLOCED(cachep); |
2754 | STATS_INC_ACTIVE(cachep); | |
2755 | STATS_SET_HIGH(cachep); | |
2756 | ||
78d382d7 MD |
2757 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2758 | numa_node_id()); | |
1da177e4 LT |
2759 | } |
2760 | check_slabp(cachep, slabp); | |
2761 | ||
2762 | /* move slabp to correct slabp list: */ | |
2763 | list_del(&slabp->list); | |
2764 | if (slabp->free == BUFCTL_END) | |
2765 | list_add(&slabp->list, &l3->slabs_full); | |
2766 | else | |
2767 | list_add(&slabp->list, &l3->slabs_partial); | |
2768 | } | |
2769 | ||
a737b3e2 | 2770 | must_grow: |
1da177e4 | 2771 | l3->free_objects -= ac->avail; |
a737b3e2 | 2772 | alloc_done: |
e498be7d | 2773 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2774 | |
2775 | if (unlikely(!ac->avail)) { | |
2776 | int x; | |
e498be7d CL |
2777 | x = cache_grow(cachep, flags, numa_node_id()); |
2778 | ||
a737b3e2 | 2779 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 2780 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2781 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
2782 | return NULL; |
2783 | ||
a737b3e2 | 2784 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
2785 | goto retry; |
2786 | } | |
2787 | ac->touched = 1; | |
e498be7d | 2788 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2789 | } |
2790 | ||
a737b3e2 AM |
2791 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
2792 | gfp_t flags) | |
1da177e4 LT |
2793 | { |
2794 | might_sleep_if(flags & __GFP_WAIT); | |
2795 | #if DEBUG | |
2796 | kmem_flagcheck(cachep, flags); | |
2797 | #endif | |
2798 | } | |
2799 | ||
2800 | #if DEBUG | |
a737b3e2 AM |
2801 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
2802 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 2803 | { |
b28a02de | 2804 | if (!objp) |
1da177e4 | 2805 | return objp; |
b28a02de | 2806 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2807 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2808 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2809 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2810 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2811 | else |
2812 | check_poison_obj(cachep, objp); | |
2813 | #else | |
2814 | check_poison_obj(cachep, objp); | |
2815 | #endif | |
2816 | poison_obj(cachep, objp, POISON_INUSE); | |
2817 | } | |
2818 | if (cachep->flags & SLAB_STORE_USER) | |
2819 | *dbg_userword(cachep, objp) = caller; | |
2820 | ||
2821 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2822 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
2823 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2824 | slab_error(cachep, "double free, or memory outside" | |
2825 | " object was overwritten"); | |
b28a02de | 2826 | printk(KERN_ERR |
a737b3e2 AM |
2827 | "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", |
2828 | objp, *dbg_redzone1(cachep, objp), | |
2829 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2830 | } |
2831 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2832 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2833 | } | |
871751e2 AV |
2834 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2835 | { | |
2836 | struct slab *slabp; | |
2837 | unsigned objnr; | |
2838 | ||
2839 | slabp = page_get_slab(virt_to_page(objp)); | |
2840 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; | |
2841 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
2842 | } | |
2843 | #endif | |
3dafccf2 | 2844 | objp += obj_offset(cachep); |
1da177e4 | 2845 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2846 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2847 | |
2848 | if (!(flags & __GFP_WAIT)) | |
2849 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2850 | ||
2851 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2852 | } |
1da177e4 LT |
2853 | return objp; |
2854 | } | |
2855 | #else | |
2856 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2857 | #endif | |
2858 | ||
343e0d7a | 2859 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2860 | { |
b28a02de | 2861 | void *objp; |
1da177e4 LT |
2862 | struct array_cache *ac; |
2863 | ||
dc85da15 | 2864 | #ifdef CONFIG_NUMA |
b2455396 | 2865 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { |
c61afb18 PJ |
2866 | objp = alternate_node_alloc(cachep, flags); |
2867 | if (objp != NULL) | |
2868 | return objp; | |
dc85da15 CL |
2869 | } |
2870 | #endif | |
2871 | ||
5c382300 | 2872 | check_irq_off(); |
9a2dba4b | 2873 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2874 | if (likely(ac->avail)) { |
2875 | STATS_INC_ALLOCHIT(cachep); | |
2876 | ac->touched = 1; | |
e498be7d | 2877 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2878 | } else { |
2879 | STATS_INC_ALLOCMISS(cachep); | |
2880 | objp = cache_alloc_refill(cachep, flags); | |
2881 | } | |
5c382300 AK |
2882 | return objp; |
2883 | } | |
2884 | ||
a737b3e2 AM |
2885 | static __always_inline void *__cache_alloc(struct kmem_cache *cachep, |
2886 | gfp_t flags, void *caller) | |
5c382300 AK |
2887 | { |
2888 | unsigned long save_flags; | |
b28a02de | 2889 | void *objp; |
5c382300 AK |
2890 | |
2891 | cache_alloc_debugcheck_before(cachep, flags); | |
2892 | ||
2893 | local_irq_save(save_flags); | |
2894 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2895 | local_irq_restore(save_flags); |
34342e86 | 2896 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 2897 | caller); |
34342e86 | 2898 | prefetchw(objp); |
1da177e4 LT |
2899 | return objp; |
2900 | } | |
2901 | ||
e498be7d | 2902 | #ifdef CONFIG_NUMA |
c61afb18 | 2903 | /* |
b2455396 | 2904 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
2905 | * |
2906 | * If we are in_interrupt, then process context, including cpusets and | |
2907 | * mempolicy, may not apply and should not be used for allocation policy. | |
2908 | */ | |
2909 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
2910 | { | |
2911 | int nid_alloc, nid_here; | |
2912 | ||
2913 | if (in_interrupt()) | |
2914 | return NULL; | |
2915 | nid_alloc = nid_here = numa_node_id(); | |
2916 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
2917 | nid_alloc = cpuset_mem_spread_node(); | |
2918 | else if (current->mempolicy) | |
2919 | nid_alloc = slab_node(current->mempolicy); | |
2920 | if (nid_alloc != nid_here) | |
2921 | return __cache_alloc_node(cachep, flags, nid_alloc); | |
2922 | return NULL; | |
2923 | } | |
2924 | ||
e498be7d CL |
2925 | /* |
2926 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2927 | */ |
a737b3e2 AM |
2928 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
2929 | int nodeid) | |
e498be7d CL |
2930 | { |
2931 | struct list_head *entry; | |
b28a02de PE |
2932 | struct slab *slabp; |
2933 | struct kmem_list3 *l3; | |
2934 | void *obj; | |
b28a02de PE |
2935 | int x; |
2936 | ||
2937 | l3 = cachep->nodelists[nodeid]; | |
2938 | BUG_ON(!l3); | |
2939 | ||
a737b3e2 | 2940 | retry: |
ca3b9b91 | 2941 | check_irq_off(); |
b28a02de PE |
2942 | spin_lock(&l3->list_lock); |
2943 | entry = l3->slabs_partial.next; | |
2944 | if (entry == &l3->slabs_partial) { | |
2945 | l3->free_touched = 1; | |
2946 | entry = l3->slabs_free.next; | |
2947 | if (entry == &l3->slabs_free) | |
2948 | goto must_grow; | |
2949 | } | |
2950 | ||
2951 | slabp = list_entry(entry, struct slab, list); | |
2952 | check_spinlock_acquired_node(cachep, nodeid); | |
2953 | check_slabp(cachep, slabp); | |
2954 | ||
2955 | STATS_INC_NODEALLOCS(cachep); | |
2956 | STATS_INC_ACTIVE(cachep); | |
2957 | STATS_SET_HIGH(cachep); | |
2958 | ||
2959 | BUG_ON(slabp->inuse == cachep->num); | |
2960 | ||
78d382d7 | 2961 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
2962 | check_slabp(cachep, slabp); |
2963 | l3->free_objects--; | |
2964 | /* move slabp to correct slabp list: */ | |
2965 | list_del(&slabp->list); | |
2966 | ||
a737b3e2 | 2967 | if (slabp->free == BUFCTL_END) |
b28a02de | 2968 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 2969 | else |
b28a02de | 2970 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 2971 | |
b28a02de PE |
2972 | spin_unlock(&l3->list_lock); |
2973 | goto done; | |
e498be7d | 2974 | |
a737b3e2 | 2975 | must_grow: |
b28a02de PE |
2976 | spin_unlock(&l3->list_lock); |
2977 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 2978 | |
b28a02de PE |
2979 | if (!x) |
2980 | return NULL; | |
e498be7d | 2981 | |
b28a02de | 2982 | goto retry; |
a737b3e2 | 2983 | done: |
b28a02de | 2984 | return obj; |
e498be7d CL |
2985 | } |
2986 | #endif | |
2987 | ||
2988 | /* | |
2989 | * Caller needs to acquire correct kmem_list's list_lock | |
2990 | */ | |
343e0d7a | 2991 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 2992 | int node) |
1da177e4 LT |
2993 | { |
2994 | int i; | |
e498be7d | 2995 | struct kmem_list3 *l3; |
1da177e4 LT |
2996 | |
2997 | for (i = 0; i < nr_objects; i++) { | |
2998 | void *objp = objpp[i]; | |
2999 | struct slab *slabp; | |
1da177e4 | 3000 | |
6ed5eb22 | 3001 | slabp = virt_to_slab(objp); |
ff69416e | 3002 | l3 = cachep->nodelists[node]; |
1da177e4 | 3003 | list_del(&slabp->list); |
ff69416e | 3004 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3005 | check_slabp(cachep, slabp); |
78d382d7 | 3006 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3007 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3008 | l3->free_objects++; |
1da177e4 LT |
3009 | check_slabp(cachep, slabp); |
3010 | ||
3011 | /* fixup slab chains */ | |
3012 | if (slabp->inuse == 0) { | |
e498be7d CL |
3013 | if (l3->free_objects > l3->free_limit) { |
3014 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
3015 | slab_destroy(cachep, slabp); |
3016 | } else { | |
e498be7d | 3017 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3018 | } |
3019 | } else { | |
3020 | /* Unconditionally move a slab to the end of the | |
3021 | * partial list on free - maximum time for the | |
3022 | * other objects to be freed, too. | |
3023 | */ | |
e498be7d | 3024 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3025 | } |
3026 | } | |
3027 | } | |
3028 | ||
343e0d7a | 3029 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3030 | { |
3031 | int batchcount; | |
e498be7d | 3032 | struct kmem_list3 *l3; |
ff69416e | 3033 | int node = numa_node_id(); |
1da177e4 LT |
3034 | |
3035 | batchcount = ac->batchcount; | |
3036 | #if DEBUG | |
3037 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3038 | #endif | |
3039 | check_irq_off(); | |
ff69416e | 3040 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3041 | spin_lock(&l3->list_lock); |
3042 | if (l3->shared) { | |
3043 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3044 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3045 | if (max) { |
3046 | if (batchcount > max) | |
3047 | batchcount = max; | |
e498be7d | 3048 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3049 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3050 | shared_array->avail += batchcount; |
3051 | goto free_done; | |
3052 | } | |
3053 | } | |
3054 | ||
ff69416e | 3055 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3056 | free_done: |
1da177e4 LT |
3057 | #if STATS |
3058 | { | |
3059 | int i = 0; | |
3060 | struct list_head *p; | |
3061 | ||
e498be7d CL |
3062 | p = l3->slabs_free.next; |
3063 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3064 | struct slab *slabp; |
3065 | ||
3066 | slabp = list_entry(p, struct slab, list); | |
3067 | BUG_ON(slabp->inuse); | |
3068 | ||
3069 | i++; | |
3070 | p = p->next; | |
3071 | } | |
3072 | STATS_SET_FREEABLE(cachep, i); | |
3073 | } | |
3074 | #endif | |
e498be7d | 3075 | spin_unlock(&l3->list_lock); |
1da177e4 | 3076 | ac->avail -= batchcount; |
a737b3e2 | 3077 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3078 | } |
3079 | ||
3080 | /* | |
a737b3e2 AM |
3081 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3082 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3083 | */ |
343e0d7a | 3084 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3085 | { |
9a2dba4b | 3086 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3087 | |
3088 | check_irq_off(); | |
3089 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3090 | ||
e498be7d CL |
3091 | /* Make sure we are not freeing a object from another |
3092 | * node to the array cache on this cpu. | |
3093 | */ | |
3094 | #ifdef CONFIG_NUMA | |
3095 | { | |
3096 | struct slab *slabp; | |
6ed5eb22 | 3097 | slabp = virt_to_slab(objp); |
e498be7d CL |
3098 | if (unlikely(slabp->nodeid != numa_node_id())) { |
3099 | struct array_cache *alien = NULL; | |
3100 | int nodeid = slabp->nodeid; | |
a737b3e2 | 3101 | struct kmem_list3 *l3; |
e498be7d | 3102 | |
a737b3e2 | 3103 | l3 = cachep->nodelists[numa_node_id()]; |
e498be7d CL |
3104 | STATS_INC_NODEFREES(cachep); |
3105 | if (l3->alien && l3->alien[nodeid]) { | |
3106 | alien = l3->alien[nodeid]; | |
3107 | spin_lock(&alien->lock); | |
fb7faf33 RT |
3108 | if (unlikely(alien->avail == alien->limit)) { |
3109 | STATS_INC_ACOVERFLOW(cachep); | |
e498be7d | 3110 | __drain_alien_cache(cachep, |
b28a02de | 3111 | alien, nodeid); |
fb7faf33 | 3112 | } |
e498be7d CL |
3113 | alien->entry[alien->avail++] = objp; |
3114 | spin_unlock(&alien->lock); | |
3115 | } else { | |
3116 | spin_lock(&(cachep->nodelists[nodeid])-> | |
b28a02de | 3117 | list_lock); |
ff69416e | 3118 | free_block(cachep, &objp, 1, nodeid); |
e498be7d | 3119 | spin_unlock(&(cachep->nodelists[nodeid])-> |
b28a02de | 3120 | list_lock); |
e498be7d CL |
3121 | } |
3122 | return; | |
3123 | } | |
3124 | } | |
3125 | #endif | |
1da177e4 LT |
3126 | if (likely(ac->avail < ac->limit)) { |
3127 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3128 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3129 | return; |
3130 | } else { | |
3131 | STATS_INC_FREEMISS(cachep); | |
3132 | cache_flusharray(cachep, ac); | |
e498be7d | 3133 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3134 | } |
3135 | } | |
3136 | ||
3137 | /** | |
3138 | * kmem_cache_alloc - Allocate an object | |
3139 | * @cachep: The cache to allocate from. | |
3140 | * @flags: See kmalloc(). | |
3141 | * | |
3142 | * Allocate an object from this cache. The flags are only relevant | |
3143 | * if the cache has no available objects. | |
3144 | */ | |
343e0d7a | 3145 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3146 | { |
7fd6b141 | 3147 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3148 | } |
3149 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3150 | ||
a8c0f9a4 PE |
3151 | /** |
3152 | * kmem_cache_alloc - Allocate an object. The memory is set to zero. | |
3153 | * @cache: The cache to allocate from. | |
3154 | * @flags: See kmalloc(). | |
3155 | * | |
3156 | * Allocate an object from this cache and set the allocated memory to zero. | |
3157 | * The flags are only relevant if the cache has no available objects. | |
3158 | */ | |
3159 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3160 | { | |
3161 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3162 | if (ret) | |
3163 | memset(ret, 0, obj_size(cache)); | |
3164 | return ret; | |
3165 | } | |
3166 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3167 | ||
1da177e4 LT |
3168 | /** |
3169 | * kmem_ptr_validate - check if an untrusted pointer might | |
3170 | * be a slab entry. | |
3171 | * @cachep: the cache we're checking against | |
3172 | * @ptr: pointer to validate | |
3173 | * | |
3174 | * This verifies that the untrusted pointer looks sane: | |
3175 | * it is _not_ a guarantee that the pointer is actually | |
3176 | * part of the slab cache in question, but it at least | |
3177 | * validates that the pointer can be dereferenced and | |
3178 | * looks half-way sane. | |
3179 | * | |
3180 | * Currently only used for dentry validation. | |
3181 | */ | |
343e0d7a | 3182 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3183 | { |
b28a02de | 3184 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3185 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3186 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3187 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3188 | struct page *page; |
3189 | ||
3190 | if (unlikely(addr < min_addr)) | |
3191 | goto out; | |
3192 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3193 | goto out; | |
3194 | if (unlikely(addr & align_mask)) | |
3195 | goto out; | |
3196 | if (unlikely(!kern_addr_valid(addr))) | |
3197 | goto out; | |
3198 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3199 | goto out; | |
3200 | page = virt_to_page(ptr); | |
3201 | if (unlikely(!PageSlab(page))) | |
3202 | goto out; | |
065d41cb | 3203 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3204 | goto out; |
3205 | return 1; | |
a737b3e2 | 3206 | out: |
1da177e4 LT |
3207 | return 0; |
3208 | } | |
3209 | ||
3210 | #ifdef CONFIG_NUMA | |
3211 | /** | |
3212 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3213 | * @cachep: The cache to allocate from. | |
3214 | * @flags: See kmalloc(). | |
3215 | * @nodeid: node number of the target node. | |
3216 | * | |
3217 | * Identical to kmem_cache_alloc, except that this function is slow | |
3218 | * and can sleep. And it will allocate memory on the given node, which | |
3219 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3220 | * New and improved: it will now make sure that the object gets |
3221 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3222 | */ |
343e0d7a | 3223 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3224 | { |
e498be7d CL |
3225 | unsigned long save_flags; |
3226 | void *ptr; | |
1da177e4 | 3227 | |
e498be7d CL |
3228 | cache_alloc_debugcheck_before(cachep, flags); |
3229 | local_irq_save(save_flags); | |
18f820f6 CL |
3230 | |
3231 | if (nodeid == -1 || nodeid == numa_node_id() || | |
a737b3e2 | 3232 | !cachep->nodelists[nodeid]) |
5c382300 AK |
3233 | ptr = ____cache_alloc(cachep, flags); |
3234 | else | |
3235 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3236 | local_irq_restore(save_flags); |
18f820f6 CL |
3237 | |
3238 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3239 | __builtin_return_address(0)); | |
1da177e4 | 3240 | |
e498be7d | 3241 | return ptr; |
1da177e4 LT |
3242 | } |
3243 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3244 | ||
dd0fc66f | 3245 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3246 | { |
343e0d7a | 3247 | struct kmem_cache *cachep; |
97e2bde4 MS |
3248 | |
3249 | cachep = kmem_find_general_cachep(size, flags); | |
3250 | if (unlikely(cachep == NULL)) | |
3251 | return NULL; | |
3252 | return kmem_cache_alloc_node(cachep, flags, node); | |
3253 | } | |
3254 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
3255 | #endif |
3256 | ||
3257 | /** | |
3258 | * kmalloc - allocate memory | |
3259 | * @size: how many bytes of memory are required. | |
3260 | * @flags: the type of memory to allocate. | |
911851e6 | 3261 | * @caller: function caller for debug tracking of the caller |
1da177e4 LT |
3262 | * |
3263 | * kmalloc is the normal method of allocating memory | |
3264 | * in the kernel. | |
3265 | * | |
3266 | * The @flags argument may be one of: | |
3267 | * | |
3268 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
3269 | * | |
3270 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
3271 | * | |
3272 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
3273 | * | |
3274 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
3275 | * must be suitable for DMA. This can mean different things on different | |
3276 | * platforms. For example, on i386, it means that the memory must come | |
3277 | * from the first 16MB. | |
3278 | */ | |
7fd6b141 PE |
3279 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3280 | void *caller) | |
1da177e4 | 3281 | { |
343e0d7a | 3282 | struct kmem_cache *cachep; |
1da177e4 | 3283 | |
97e2bde4 MS |
3284 | /* If you want to save a few bytes .text space: replace |
3285 | * __ with kmem_. | |
3286 | * Then kmalloc uses the uninlined functions instead of the inline | |
3287 | * functions. | |
3288 | */ | |
3289 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3290 | if (unlikely(cachep == NULL)) |
3291 | return NULL; | |
7fd6b141 PE |
3292 | return __cache_alloc(cachep, flags, caller); |
3293 | } | |
3294 | ||
7fd6b141 PE |
3295 | |
3296 | void *__kmalloc(size_t size, gfp_t flags) | |
3297 | { | |
871751e2 | 3298 | #ifndef CONFIG_DEBUG_SLAB |
7fd6b141 | 3299 | return __do_kmalloc(size, flags, NULL); |
871751e2 AV |
3300 | #else |
3301 | return __do_kmalloc(size, flags, __builtin_return_address(0)); | |
3302 | #endif | |
1da177e4 LT |
3303 | } |
3304 | EXPORT_SYMBOL(__kmalloc); | |
3305 | ||
871751e2 | 3306 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3307 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3308 | { | |
3309 | return __do_kmalloc(size, flags, caller); | |
3310 | } | |
3311 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
7fd6b141 PE |
3312 | #endif |
3313 | ||
1da177e4 LT |
3314 | #ifdef CONFIG_SMP |
3315 | /** | |
3316 | * __alloc_percpu - allocate one copy of the object for every present | |
3317 | * cpu in the system, zeroing them. | |
3318 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
3319 | * | |
3320 | * @size: how many bytes of memory are required. | |
1da177e4 | 3321 | */ |
f9f75005 | 3322 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
3323 | { |
3324 | int i; | |
b28a02de | 3325 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3326 | |
3327 | if (!pdata) | |
3328 | return NULL; | |
3329 | ||
e498be7d CL |
3330 | /* |
3331 | * Cannot use for_each_online_cpu since a cpu may come online | |
3332 | * and we have no way of figuring out how to fix the array | |
3333 | * that we have allocated then.... | |
3334 | */ | |
0a945022 | 3335 | for_each_possible_cpu(i) { |
e498be7d CL |
3336 | int node = cpu_to_node(i); |
3337 | ||
3338 | if (node_online(node)) | |
3339 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3340 | else | |
3341 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3342 | |
3343 | if (!pdata->ptrs[i]) | |
3344 | goto unwind_oom; | |
3345 | memset(pdata->ptrs[i], 0, size); | |
3346 | } | |
3347 | ||
3348 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3349 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3350 | |
a737b3e2 | 3351 | unwind_oom: |
1da177e4 LT |
3352 | while (--i >= 0) { |
3353 | if (!cpu_possible(i)) | |
3354 | continue; | |
3355 | kfree(pdata->ptrs[i]); | |
3356 | } | |
3357 | kfree(pdata); | |
3358 | return NULL; | |
3359 | } | |
3360 | EXPORT_SYMBOL(__alloc_percpu); | |
3361 | #endif | |
3362 | ||
3363 | /** | |
3364 | * kmem_cache_free - Deallocate an object | |
3365 | * @cachep: The cache the allocation was from. | |
3366 | * @objp: The previously allocated object. | |
3367 | * | |
3368 | * Free an object which was previously allocated from this | |
3369 | * cache. | |
3370 | */ | |
343e0d7a | 3371 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3372 | { |
3373 | unsigned long flags; | |
3374 | ||
3375 | local_irq_save(flags); | |
3376 | __cache_free(cachep, objp); | |
3377 | local_irq_restore(flags); | |
3378 | } | |
3379 | EXPORT_SYMBOL(kmem_cache_free); | |
3380 | ||
1da177e4 LT |
3381 | /** |
3382 | * kfree - free previously allocated memory | |
3383 | * @objp: pointer returned by kmalloc. | |
3384 | * | |
80e93eff PE |
3385 | * If @objp is NULL, no operation is performed. |
3386 | * | |
1da177e4 LT |
3387 | * Don't free memory not originally allocated by kmalloc() |
3388 | * or you will run into trouble. | |
3389 | */ | |
3390 | void kfree(const void *objp) | |
3391 | { | |
343e0d7a | 3392 | struct kmem_cache *c; |
1da177e4 LT |
3393 | unsigned long flags; |
3394 | ||
3395 | if (unlikely(!objp)) | |
3396 | return; | |
3397 | local_irq_save(flags); | |
3398 | kfree_debugcheck(objp); | |
6ed5eb22 | 3399 | c = virt_to_cache(objp); |
3dafccf2 | 3400 | mutex_debug_check_no_locks_freed(objp, obj_size(c)); |
b28a02de | 3401 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3402 | local_irq_restore(flags); |
3403 | } | |
3404 | EXPORT_SYMBOL(kfree); | |
3405 | ||
3406 | #ifdef CONFIG_SMP | |
3407 | /** | |
3408 | * free_percpu - free previously allocated percpu memory | |
3409 | * @objp: pointer returned by alloc_percpu. | |
3410 | * | |
3411 | * Don't free memory not originally allocated by alloc_percpu() | |
3412 | * The complemented objp is to check for that. | |
3413 | */ | |
b28a02de | 3414 | void free_percpu(const void *objp) |
1da177e4 LT |
3415 | { |
3416 | int i; | |
b28a02de | 3417 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3418 | |
e498be7d CL |
3419 | /* |
3420 | * We allocate for all cpus so we cannot use for online cpu here. | |
3421 | */ | |
0a945022 | 3422 | for_each_possible_cpu(i) |
b28a02de | 3423 | kfree(p->ptrs[i]); |
1da177e4 LT |
3424 | kfree(p); |
3425 | } | |
3426 | EXPORT_SYMBOL(free_percpu); | |
3427 | #endif | |
3428 | ||
343e0d7a | 3429 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3430 | { |
3dafccf2 | 3431 | return obj_size(cachep); |
1da177e4 LT |
3432 | } |
3433 | EXPORT_SYMBOL(kmem_cache_size); | |
3434 | ||
343e0d7a | 3435 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3436 | { |
3437 | return cachep->name; | |
3438 | } | |
3439 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3440 | ||
e498be7d | 3441 | /* |
0718dc2a | 3442 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3443 | */ |
343e0d7a | 3444 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3445 | { |
3446 | int node; | |
3447 | struct kmem_list3 *l3; | |
cafeb02e CL |
3448 | struct array_cache *new_shared; |
3449 | struct array_cache **new_alien; | |
e498be7d CL |
3450 | |
3451 | for_each_online_node(node) { | |
cafeb02e | 3452 | |
a737b3e2 AM |
3453 | new_alien = alloc_alien_cache(node, cachep->limit); |
3454 | if (!new_alien) | |
e498be7d | 3455 | goto fail; |
cafeb02e | 3456 | |
0718dc2a CL |
3457 | new_shared = alloc_arraycache(node, |
3458 | cachep->shared*cachep->batchcount, | |
a737b3e2 | 3459 | 0xbaadf00d); |
0718dc2a CL |
3460 | if (!new_shared) { |
3461 | free_alien_cache(new_alien); | |
e498be7d | 3462 | goto fail; |
0718dc2a | 3463 | } |
cafeb02e | 3464 | |
a737b3e2 AM |
3465 | l3 = cachep->nodelists[node]; |
3466 | if (l3) { | |
cafeb02e CL |
3467 | struct array_cache *shared = l3->shared; |
3468 | ||
e498be7d CL |
3469 | spin_lock_irq(&l3->list_lock); |
3470 | ||
cafeb02e | 3471 | if (shared) |
0718dc2a CL |
3472 | free_block(cachep, shared->entry, |
3473 | shared->avail, node); | |
e498be7d | 3474 | |
cafeb02e CL |
3475 | l3->shared = new_shared; |
3476 | if (!l3->alien) { | |
e498be7d CL |
3477 | l3->alien = new_alien; |
3478 | new_alien = NULL; | |
3479 | } | |
b28a02de | 3480 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3481 | cachep->batchcount + cachep->num; |
e498be7d | 3482 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3483 | kfree(shared); |
e498be7d CL |
3484 | free_alien_cache(new_alien); |
3485 | continue; | |
3486 | } | |
a737b3e2 | 3487 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3488 | if (!l3) { |
3489 | free_alien_cache(new_alien); | |
3490 | kfree(new_shared); | |
e498be7d | 3491 | goto fail; |
0718dc2a | 3492 | } |
e498be7d CL |
3493 | |
3494 | kmem_list3_init(l3); | |
3495 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3496 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3497 | l3->shared = new_shared; |
e498be7d | 3498 | l3->alien = new_alien; |
b28a02de | 3499 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3500 | cachep->batchcount + cachep->num; |
e498be7d CL |
3501 | cachep->nodelists[node] = l3; |
3502 | } | |
cafeb02e | 3503 | return 0; |
0718dc2a | 3504 | |
a737b3e2 | 3505 | fail: |
0718dc2a CL |
3506 | if (!cachep->next.next) { |
3507 | /* Cache is not active yet. Roll back what we did */ | |
3508 | node--; | |
3509 | while (node >= 0) { | |
3510 | if (cachep->nodelists[node]) { | |
3511 | l3 = cachep->nodelists[node]; | |
3512 | ||
3513 | kfree(l3->shared); | |
3514 | free_alien_cache(l3->alien); | |
3515 | kfree(l3); | |
3516 | cachep->nodelists[node] = NULL; | |
3517 | } | |
3518 | node--; | |
3519 | } | |
3520 | } | |
cafeb02e | 3521 | return -ENOMEM; |
e498be7d CL |
3522 | } |
3523 | ||
1da177e4 | 3524 | struct ccupdate_struct { |
343e0d7a | 3525 | struct kmem_cache *cachep; |
1da177e4 LT |
3526 | struct array_cache *new[NR_CPUS]; |
3527 | }; | |
3528 | ||
3529 | static void do_ccupdate_local(void *info) | |
3530 | { | |
a737b3e2 | 3531 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3532 | struct array_cache *old; |
3533 | ||
3534 | check_irq_off(); | |
9a2dba4b | 3535 | old = cpu_cache_get(new->cachep); |
e498be7d | 3536 | |
1da177e4 LT |
3537 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3538 | new->new[smp_processor_id()] = old; | |
3539 | } | |
3540 | ||
b5d8ca7c | 3541 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3542 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3543 | int batchcount, int shared) | |
1da177e4 LT |
3544 | { |
3545 | struct ccupdate_struct new; | |
e498be7d | 3546 | int i, err; |
1da177e4 | 3547 | |
b28a02de | 3548 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3549 | for_each_online_cpu(i) { |
a737b3e2 AM |
3550 | new.new[i] = alloc_arraycache(cpu_to_node(i), limit, |
3551 | batchcount); | |
e498be7d | 3552 | if (!new.new[i]) { |
b28a02de PE |
3553 | for (i--; i >= 0; i--) |
3554 | kfree(new.new[i]); | |
e498be7d | 3555 | return -ENOMEM; |
1da177e4 LT |
3556 | } |
3557 | } | |
3558 | new.cachep = cachep; | |
3559 | ||
a07fa394 | 3560 | on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1); |
e498be7d | 3561 | |
1da177e4 | 3562 | check_irq_on(); |
1da177e4 LT |
3563 | cachep->batchcount = batchcount; |
3564 | cachep->limit = limit; | |
e498be7d | 3565 | cachep->shared = shared; |
1da177e4 | 3566 | |
e498be7d | 3567 | for_each_online_cpu(i) { |
1da177e4 LT |
3568 | struct array_cache *ccold = new.new[i]; |
3569 | if (!ccold) | |
3570 | continue; | |
e498be7d | 3571 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3572 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3573 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3574 | kfree(ccold); |
3575 | } | |
1da177e4 | 3576 | |
e498be7d CL |
3577 | err = alloc_kmemlist(cachep); |
3578 | if (err) { | |
3579 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3580 | cachep->name, -err); |
e498be7d | 3581 | BUG(); |
1da177e4 | 3582 | } |
1da177e4 LT |
3583 | return 0; |
3584 | } | |
3585 | ||
b5d8ca7c | 3586 | /* Called with cache_chain_mutex held always */ |
343e0d7a | 3587 | static void enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3588 | { |
3589 | int err; | |
3590 | int limit, shared; | |
3591 | ||
a737b3e2 AM |
3592 | /* |
3593 | * The head array serves three purposes: | |
1da177e4 LT |
3594 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3595 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3596 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3597 | * bufctl chains: array operations are cheaper. |
3598 | * The numbers are guessed, we should auto-tune as described by | |
3599 | * Bonwick. | |
3600 | */ | |
3dafccf2 | 3601 | if (cachep->buffer_size > 131072) |
1da177e4 | 3602 | limit = 1; |
3dafccf2 | 3603 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3604 | limit = 8; |
3dafccf2 | 3605 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3606 | limit = 24; |
3dafccf2 | 3607 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3608 | limit = 54; |
3609 | else | |
3610 | limit = 120; | |
3611 | ||
a737b3e2 AM |
3612 | /* |
3613 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3614 | * allocation behaviour: Most allocs on one cpu, most free operations |
3615 | * on another cpu. For these cases, an efficient object passing between | |
3616 | * cpus is necessary. This is provided by a shared array. The array | |
3617 | * replaces Bonwick's magazine layer. | |
3618 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3619 | * to a larger limit. Thus disabled by default. | |
3620 | */ | |
3621 | shared = 0; | |
3622 | #ifdef CONFIG_SMP | |
3dafccf2 | 3623 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3624 | shared = 8; |
3625 | #endif | |
3626 | ||
3627 | #if DEBUG | |
a737b3e2 AM |
3628 | /* |
3629 | * With debugging enabled, large batchcount lead to excessively long | |
3630 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3631 | */ |
3632 | if (limit > 32) | |
3633 | limit = 32; | |
3634 | #endif | |
b28a02de | 3635 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3636 | if (err) |
3637 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3638 | cachep->name, -err); |
1da177e4 LT |
3639 | } |
3640 | ||
1b55253a CL |
3641 | /* |
3642 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3643 | * necessary. Note that the l3 listlock also protects the array_cache |
3644 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3645 | */ |
3646 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
3647 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
3648 | { |
3649 | int tofree; | |
3650 | ||
1b55253a CL |
3651 | if (!ac || !ac->avail) |
3652 | return; | |
1da177e4 LT |
3653 | if (ac->touched && !force) { |
3654 | ac->touched = 0; | |
b18e7e65 | 3655 | } else { |
1b55253a | 3656 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
3657 | if (ac->avail) { |
3658 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3659 | if (tofree > ac->avail) | |
3660 | tofree = (ac->avail + 1) / 2; | |
3661 | free_block(cachep, ac->entry, tofree, node); | |
3662 | ac->avail -= tofree; | |
3663 | memmove(ac->entry, &(ac->entry[tofree]), | |
3664 | sizeof(void *) * ac->avail); | |
3665 | } | |
1b55253a | 3666 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
3667 | } |
3668 | } | |
3669 | ||
3670 | /** | |
3671 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3672 | * @unused: unused parameter |
1da177e4 LT |
3673 | * |
3674 | * Called from workqueue/eventd every few seconds. | |
3675 | * Purpose: | |
3676 | * - clear the per-cpu caches for this CPU. | |
3677 | * - return freeable pages to the main free memory pool. | |
3678 | * | |
a737b3e2 AM |
3679 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3680 | * again on the next iteration. | |
1da177e4 LT |
3681 | */ |
3682 | static void cache_reap(void *unused) | |
3683 | { | |
3684 | struct list_head *walk; | |
e498be7d | 3685 | struct kmem_list3 *l3; |
aab2207c | 3686 | int node = numa_node_id(); |
1da177e4 | 3687 | |
fc0abb14 | 3688 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3689 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3690 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3691 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3692 | return; |
3693 | } | |
3694 | ||
3695 | list_for_each(walk, &cache_chain) { | |
343e0d7a | 3696 | struct kmem_cache *searchp; |
b28a02de | 3697 | struct list_head *p; |
1da177e4 LT |
3698 | int tofree; |
3699 | struct slab *slabp; | |
3700 | ||
343e0d7a | 3701 | searchp = list_entry(walk, struct kmem_cache, next); |
1da177e4 LT |
3702 | check_irq_on(); |
3703 | ||
35386e3b CL |
3704 | /* |
3705 | * We only take the l3 lock if absolutely necessary and we | |
3706 | * have established with reasonable certainty that | |
3707 | * we can do some work if the lock was obtained. | |
3708 | */ | |
aab2207c | 3709 | l3 = searchp->nodelists[node]; |
35386e3b | 3710 | |
8fce4d8e | 3711 | reap_alien(searchp, l3); |
1da177e4 | 3712 | |
aab2207c | 3713 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 3714 | |
35386e3b CL |
3715 | /* |
3716 | * These are racy checks but it does not matter | |
3717 | * if we skip one check or scan twice. | |
3718 | */ | |
e498be7d | 3719 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 3720 | goto next; |
1da177e4 | 3721 | |
e498be7d | 3722 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3723 | |
aab2207c | 3724 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 3725 | |
e498be7d CL |
3726 | if (l3->free_touched) { |
3727 | l3->free_touched = 0; | |
35386e3b | 3728 | goto next; |
1da177e4 LT |
3729 | } |
3730 | ||
a737b3e2 AM |
3731 | tofree = (l3->free_limit + 5 * searchp->num - 1) / |
3732 | (5 * searchp->num); | |
1da177e4 | 3733 | do { |
35386e3b CL |
3734 | /* |
3735 | * Do not lock if there are no free blocks. | |
3736 | */ | |
3737 | if (list_empty(&l3->slabs_free)) | |
3738 | break; | |
3739 | ||
3740 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3741 | p = l3->slabs_free.next; |
35386e3b CL |
3742 | if (p == &(l3->slabs_free)) { |
3743 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3744 | break; |
35386e3b | 3745 | } |
1da177e4 LT |
3746 | |
3747 | slabp = list_entry(p, struct slab, list); | |
3748 | BUG_ON(slabp->inuse); | |
3749 | list_del(&slabp->list); | |
3750 | STATS_INC_REAPED(searchp); | |
3751 | ||
a737b3e2 AM |
3752 | /* |
3753 | * Safe to drop the lock. The slab is no longer linked | |
3754 | * to the cache. searchp cannot disappear, we hold | |
1da177e4 LT |
3755 | * cache_chain_lock |
3756 | */ | |
e498be7d CL |
3757 | l3->free_objects -= searchp->num; |
3758 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3759 | slab_destroy(searchp, slabp); |
b28a02de | 3760 | } while (--tofree > 0); |
35386e3b | 3761 | next: |
1da177e4 LT |
3762 | cond_resched(); |
3763 | } | |
3764 | check_irq_on(); | |
fc0abb14 | 3765 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 3766 | next_reap_node(); |
a737b3e2 | 3767 | /* Set up the next iteration */ |
cd61ef62 | 3768 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3769 | } |
3770 | ||
3771 | #ifdef CONFIG_PROC_FS | |
3772 | ||
85289f98 | 3773 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3774 | { |
85289f98 PE |
3775 | /* |
3776 | * Output format version, so at least we can change it | |
3777 | * without _too_ many complaints. | |
3778 | */ | |
1da177e4 | 3779 | #if STATS |
85289f98 | 3780 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3781 | #else |
85289f98 | 3782 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3783 | #endif |
85289f98 PE |
3784 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3785 | "<objperslab> <pagesperslab>"); | |
3786 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3787 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3788 | #if STATS |
85289f98 | 3789 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 3790 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 3791 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 3792 | #endif |
85289f98 PE |
3793 | seq_putc(m, '\n'); |
3794 | } | |
3795 | ||
3796 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3797 | { | |
3798 | loff_t n = *pos; | |
3799 | struct list_head *p; | |
3800 | ||
fc0abb14 | 3801 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3802 | if (!n) |
3803 | print_slabinfo_header(m); | |
1da177e4 LT |
3804 | p = cache_chain.next; |
3805 | while (n--) { | |
3806 | p = p->next; | |
3807 | if (p == &cache_chain) | |
3808 | return NULL; | |
3809 | } | |
343e0d7a | 3810 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3811 | } |
3812 | ||
3813 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3814 | { | |
343e0d7a | 3815 | struct kmem_cache *cachep = p; |
1da177e4 | 3816 | ++*pos; |
a737b3e2 AM |
3817 | return cachep->next.next == &cache_chain ? |
3818 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
3819 | } |
3820 | ||
3821 | static void s_stop(struct seq_file *m, void *p) | |
3822 | { | |
fc0abb14 | 3823 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3824 | } |
3825 | ||
3826 | static int s_show(struct seq_file *m, void *p) | |
3827 | { | |
343e0d7a | 3828 | struct kmem_cache *cachep = p; |
1da177e4 | 3829 | struct list_head *q; |
b28a02de PE |
3830 | struct slab *slabp; |
3831 | unsigned long active_objs; | |
3832 | unsigned long num_objs; | |
3833 | unsigned long active_slabs = 0; | |
3834 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3835 | const char *name; |
1da177e4 | 3836 | char *error = NULL; |
e498be7d CL |
3837 | int node; |
3838 | struct kmem_list3 *l3; | |
1da177e4 | 3839 | |
1da177e4 LT |
3840 | active_objs = 0; |
3841 | num_slabs = 0; | |
e498be7d CL |
3842 | for_each_online_node(node) { |
3843 | l3 = cachep->nodelists[node]; | |
3844 | if (!l3) | |
3845 | continue; | |
3846 | ||
ca3b9b91 RT |
3847 | check_irq_on(); |
3848 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3849 | |
b28a02de | 3850 | list_for_each(q, &l3->slabs_full) { |
e498be7d CL |
3851 | slabp = list_entry(q, struct slab, list); |
3852 | if (slabp->inuse != cachep->num && !error) | |
3853 | error = "slabs_full accounting error"; | |
3854 | active_objs += cachep->num; | |
3855 | active_slabs++; | |
3856 | } | |
b28a02de | 3857 | list_for_each(q, &l3->slabs_partial) { |
e498be7d CL |
3858 | slabp = list_entry(q, struct slab, list); |
3859 | if (slabp->inuse == cachep->num && !error) | |
3860 | error = "slabs_partial inuse accounting error"; | |
3861 | if (!slabp->inuse && !error) | |
3862 | error = "slabs_partial/inuse accounting error"; | |
3863 | active_objs += slabp->inuse; | |
3864 | active_slabs++; | |
3865 | } | |
b28a02de | 3866 | list_for_each(q, &l3->slabs_free) { |
e498be7d CL |
3867 | slabp = list_entry(q, struct slab, list); |
3868 | if (slabp->inuse && !error) | |
3869 | error = "slabs_free/inuse accounting error"; | |
3870 | num_slabs++; | |
3871 | } | |
3872 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3873 | if (l3->shared) |
3874 | shared_avail += l3->shared->avail; | |
e498be7d | 3875 | |
ca3b9b91 | 3876 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3877 | } |
b28a02de PE |
3878 | num_slabs += active_slabs; |
3879 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3880 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3881 | error = "free_objects accounting error"; |
3882 | ||
b28a02de | 3883 | name = cachep->name; |
1da177e4 LT |
3884 | if (error) |
3885 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3886 | ||
3887 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3888 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3889 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3890 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3891 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3892 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3893 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3894 | #if STATS |
b28a02de | 3895 | { /* list3 stats */ |
1da177e4 LT |
3896 | unsigned long high = cachep->high_mark; |
3897 | unsigned long allocs = cachep->num_allocations; | |
3898 | unsigned long grown = cachep->grown; | |
3899 | unsigned long reaped = cachep->reaped; | |
3900 | unsigned long errors = cachep->errors; | |
3901 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3902 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3903 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 3904 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 3905 | |
e498be7d | 3906 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 3907 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 3908 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 3909 | node_frees, overflows); |
1da177e4 LT |
3910 | } |
3911 | /* cpu stats */ | |
3912 | { | |
3913 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3914 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3915 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3916 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3917 | ||
3918 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3919 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3920 | } |
3921 | #endif | |
3922 | seq_putc(m, '\n'); | |
1da177e4 LT |
3923 | return 0; |
3924 | } | |
3925 | ||
3926 | /* | |
3927 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3928 | * | |
3929 | * Output layout: | |
3930 | * cache-name | |
3931 | * num-active-objs | |
3932 | * total-objs | |
3933 | * object size | |
3934 | * num-active-slabs | |
3935 | * total-slabs | |
3936 | * num-pages-per-slab | |
3937 | * + further values on SMP and with statistics enabled | |
3938 | */ | |
3939 | ||
3940 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3941 | .start = s_start, |
3942 | .next = s_next, | |
3943 | .stop = s_stop, | |
3944 | .show = s_show, | |
1da177e4 LT |
3945 | }; |
3946 | ||
3947 | #define MAX_SLABINFO_WRITE 128 | |
3948 | /** | |
3949 | * slabinfo_write - Tuning for the slab allocator | |
3950 | * @file: unused | |
3951 | * @buffer: user buffer | |
3952 | * @count: data length | |
3953 | * @ppos: unused | |
3954 | */ | |
b28a02de PE |
3955 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3956 | size_t count, loff_t *ppos) | |
1da177e4 | 3957 | { |
b28a02de | 3958 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 LT |
3959 | int limit, batchcount, shared, res; |
3960 | struct list_head *p; | |
b28a02de | 3961 | |
1da177e4 LT |
3962 | if (count > MAX_SLABINFO_WRITE) |
3963 | return -EINVAL; | |
3964 | if (copy_from_user(&kbuf, buffer, count)) | |
3965 | return -EFAULT; | |
b28a02de | 3966 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3967 | |
3968 | tmp = strchr(kbuf, ' '); | |
3969 | if (!tmp) | |
3970 | return -EINVAL; | |
3971 | *tmp = '\0'; | |
3972 | tmp++; | |
3973 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3974 | return -EINVAL; | |
3975 | ||
3976 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3977 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3978 | res = -EINVAL; |
b28a02de | 3979 | list_for_each(p, &cache_chain) { |
a737b3e2 | 3980 | struct kmem_cache *cachep; |
1da177e4 | 3981 | |
a737b3e2 | 3982 | cachep = list_entry(p, struct kmem_cache, next); |
1da177e4 | 3983 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
3984 | if (limit < 1 || batchcount < 1 || |
3985 | batchcount > limit || shared < 0) { | |
e498be7d | 3986 | res = 0; |
1da177e4 | 3987 | } else { |
e498be7d | 3988 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3989 | batchcount, shared); |
1da177e4 LT |
3990 | } |
3991 | break; | |
3992 | } | |
3993 | } | |
fc0abb14 | 3994 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3995 | if (res >= 0) |
3996 | res = count; | |
3997 | return res; | |
3998 | } | |
871751e2 AV |
3999 | |
4000 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4001 | ||
4002 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4003 | { | |
4004 | loff_t n = *pos; | |
4005 | struct list_head *p; | |
4006 | ||
4007 | mutex_lock(&cache_chain_mutex); | |
4008 | p = cache_chain.next; | |
4009 | while (n--) { | |
4010 | p = p->next; | |
4011 | if (p == &cache_chain) | |
4012 | return NULL; | |
4013 | } | |
4014 | return list_entry(p, struct kmem_cache, next); | |
4015 | } | |
4016 | ||
4017 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4018 | { | |
4019 | unsigned long *p; | |
4020 | int l; | |
4021 | if (!v) | |
4022 | return 1; | |
4023 | l = n[1]; | |
4024 | p = n + 2; | |
4025 | while (l) { | |
4026 | int i = l/2; | |
4027 | unsigned long *q = p + 2 * i; | |
4028 | if (*q == v) { | |
4029 | q[1]++; | |
4030 | return 1; | |
4031 | } | |
4032 | if (*q > v) { | |
4033 | l = i; | |
4034 | } else { | |
4035 | p = q + 2; | |
4036 | l -= i + 1; | |
4037 | } | |
4038 | } | |
4039 | if (++n[1] == n[0]) | |
4040 | return 0; | |
4041 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4042 | p[0] = v; | |
4043 | p[1] = 1; | |
4044 | return 1; | |
4045 | } | |
4046 | ||
4047 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4048 | { | |
4049 | void *p; | |
4050 | int i; | |
4051 | if (n[0] == n[1]) | |
4052 | return; | |
4053 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4054 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4055 | continue; | |
4056 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4057 | return; | |
4058 | } | |
4059 | } | |
4060 | ||
4061 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4062 | { | |
4063 | #ifdef CONFIG_KALLSYMS | |
4064 | char *modname; | |
4065 | const char *name; | |
4066 | unsigned long offset, size; | |
4067 | char namebuf[KSYM_NAME_LEN+1]; | |
4068 | ||
4069 | name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); | |
4070 | ||
4071 | if (name) { | |
4072 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4073 | if (modname) | |
4074 | seq_printf(m, " [%s]", modname); | |
4075 | return; | |
4076 | } | |
4077 | #endif | |
4078 | seq_printf(m, "%p", (void *)address); | |
4079 | } | |
4080 | ||
4081 | static int leaks_show(struct seq_file *m, void *p) | |
4082 | { | |
4083 | struct kmem_cache *cachep = p; | |
4084 | struct list_head *q; | |
4085 | struct slab *slabp; | |
4086 | struct kmem_list3 *l3; | |
4087 | const char *name; | |
4088 | unsigned long *n = m->private; | |
4089 | int node; | |
4090 | int i; | |
4091 | ||
4092 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4093 | return 0; | |
4094 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4095 | return 0; | |
4096 | ||
4097 | /* OK, we can do it */ | |
4098 | ||
4099 | n[1] = 0; | |
4100 | ||
4101 | for_each_online_node(node) { | |
4102 | l3 = cachep->nodelists[node]; | |
4103 | if (!l3) | |
4104 | continue; | |
4105 | ||
4106 | check_irq_on(); | |
4107 | spin_lock_irq(&l3->list_lock); | |
4108 | ||
4109 | list_for_each(q, &l3->slabs_full) { | |
4110 | slabp = list_entry(q, struct slab, list); | |
4111 | handle_slab(n, cachep, slabp); | |
4112 | } | |
4113 | list_for_each(q, &l3->slabs_partial) { | |
4114 | slabp = list_entry(q, struct slab, list); | |
4115 | handle_slab(n, cachep, slabp); | |
4116 | } | |
4117 | spin_unlock_irq(&l3->list_lock); | |
4118 | } | |
4119 | name = cachep->name; | |
4120 | if (n[0] == n[1]) { | |
4121 | /* Increase the buffer size */ | |
4122 | mutex_unlock(&cache_chain_mutex); | |
4123 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4124 | if (!m->private) { | |
4125 | /* Too bad, we are really out */ | |
4126 | m->private = n; | |
4127 | mutex_lock(&cache_chain_mutex); | |
4128 | return -ENOMEM; | |
4129 | } | |
4130 | *(unsigned long *)m->private = n[0] * 2; | |
4131 | kfree(n); | |
4132 | mutex_lock(&cache_chain_mutex); | |
4133 | /* Now make sure this entry will be retried */ | |
4134 | m->count = m->size; | |
4135 | return 0; | |
4136 | } | |
4137 | for (i = 0; i < n[1]; i++) { | |
4138 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4139 | show_symbol(m, n[2*i+2]); | |
4140 | seq_putc(m, '\n'); | |
4141 | } | |
4142 | return 0; | |
4143 | } | |
4144 | ||
4145 | struct seq_operations slabstats_op = { | |
4146 | .start = leaks_start, | |
4147 | .next = s_next, | |
4148 | .stop = s_stop, | |
4149 | .show = leaks_show, | |
4150 | }; | |
4151 | #endif | |
1da177e4 LT |
4152 | #endif |
4153 | ||
00e145b6 MS |
4154 | /** |
4155 | * ksize - get the actual amount of memory allocated for a given object | |
4156 | * @objp: Pointer to the object | |
4157 | * | |
4158 | * kmalloc may internally round up allocations and return more memory | |
4159 | * than requested. ksize() can be used to determine the actual amount of | |
4160 | * memory allocated. The caller may use this additional memory, even though | |
4161 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4162 | * The caller must guarantee that objp points to a valid object previously | |
4163 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4164 | * must not be freed during the duration of the call. | |
4165 | */ | |
1da177e4 LT |
4166 | unsigned int ksize(const void *objp) |
4167 | { | |
00e145b6 MS |
4168 | if (unlikely(objp == NULL)) |
4169 | return 0; | |
1da177e4 | 4170 | |
6ed5eb22 | 4171 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4172 | } |