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