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