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