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 | 209 | |
1da177e4 LT |
210 | /* |
211 | * struct slab | |
212 | * | |
213 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
214 | * for a slab, or allocated from an general cache. | |
215 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
216 | */ | |
217 | struct slab { | |
b28a02de PE |
218 | struct list_head list; |
219 | unsigned long colouroff; | |
220 | void *s_mem; /* including colour offset */ | |
221 | unsigned int inuse; /* num of objs active in slab */ | |
222 | kmem_bufctl_t free; | |
223 | unsigned short nodeid; | |
1da177e4 LT |
224 | }; |
225 | ||
226 | /* | |
227 | * struct slab_rcu | |
228 | * | |
229 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
230 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
231 | * we need to approach a kernel structure obliquely, from its address | |
232 | * obtained without the usual locking. We can lock the structure to | |
233 | * stabilize it and check it's still at the given address, only if we | |
234 | * can be sure that the memory has not been meanwhile reused for some | |
235 | * other kind of object (which our subsystem's lock might corrupt). | |
236 | * | |
237 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
238 | * taking the spinlock within the structure expected at that address. | |
239 | * | |
240 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
241 | */ | |
242 | struct slab_rcu { | |
b28a02de | 243 | struct rcu_head head; |
343e0d7a | 244 | struct kmem_cache *cachep; |
b28a02de | 245 | void *addr; |
1da177e4 LT |
246 | }; |
247 | ||
248 | /* | |
249 | * struct array_cache | |
250 | * | |
1da177e4 LT |
251 | * Purpose: |
252 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
253 | * - reduce the number of linked list operations | |
254 | * - reduce spinlock operations | |
255 | * | |
256 | * The limit is stored in the per-cpu structure to reduce the data cache | |
257 | * footprint. | |
258 | * | |
259 | */ | |
260 | struct array_cache { | |
261 | unsigned int avail; | |
262 | unsigned int limit; | |
263 | unsigned int batchcount; | |
264 | unsigned int touched; | |
e498be7d | 265 | spinlock_t lock; |
a737b3e2 AM |
266 | void *entry[0]; /* |
267 | * Must have this definition in here for the proper | |
268 | * alignment of array_cache. Also simplifies accessing | |
269 | * the entries. | |
270 | * [0] is for gcc 2.95. It should really be []. | |
271 | */ | |
1da177e4 LT |
272 | }; |
273 | ||
a737b3e2 AM |
274 | /* |
275 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
276 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
277 | */ |
278 | #define BOOT_CPUCACHE_ENTRIES 1 | |
279 | struct arraycache_init { | |
280 | struct array_cache cache; | |
b28a02de | 281 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
282 | }; |
283 | ||
284 | /* | |
e498be7d | 285 | * The slab lists for all objects. |
1da177e4 LT |
286 | */ |
287 | struct kmem_list3 { | |
b28a02de PE |
288 | struct list_head slabs_partial; /* partial list first, better asm code */ |
289 | struct list_head slabs_full; | |
290 | struct list_head slabs_free; | |
291 | unsigned long free_objects; | |
b28a02de | 292 | unsigned int free_limit; |
2e1217cf | 293 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
294 | spinlock_t list_lock; |
295 | struct array_cache *shared; /* shared per node */ | |
296 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
297 | unsigned long next_reap; /* updated without locking */ |
298 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
299 | }; |
300 | ||
e498be7d CL |
301 | /* |
302 | * Need this for bootstrapping a per node allocator. | |
303 | */ | |
304 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
305 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
306 | #define CACHE_CACHE 0 | |
307 | #define SIZE_AC 1 | |
308 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
309 | ||
310 | /* | |
a737b3e2 AM |
311 | * This function must be completely optimized away if a constant is passed to |
312 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 313 | */ |
7243cc05 | 314 | static __always_inline int index_of(const size_t size) |
e498be7d | 315 | { |
5ec8a847 SR |
316 | extern void __bad_size(void); |
317 | ||
e498be7d CL |
318 | if (__builtin_constant_p(size)) { |
319 | int i = 0; | |
320 | ||
321 | #define CACHE(x) \ | |
322 | if (size <=x) \ | |
323 | return i; \ | |
324 | else \ | |
325 | i++; | |
326 | #include "linux/kmalloc_sizes.h" | |
327 | #undef CACHE | |
5ec8a847 | 328 | __bad_size(); |
7243cc05 | 329 | } else |
5ec8a847 | 330 | __bad_size(); |
e498be7d CL |
331 | return 0; |
332 | } | |
333 | ||
334 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) | |
335 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 336 | |
5295a74c | 337 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
338 | { |
339 | INIT_LIST_HEAD(&parent->slabs_full); | |
340 | INIT_LIST_HEAD(&parent->slabs_partial); | |
341 | INIT_LIST_HEAD(&parent->slabs_free); | |
342 | parent->shared = NULL; | |
343 | parent->alien = NULL; | |
2e1217cf | 344 | parent->colour_next = 0; |
e498be7d CL |
345 | spin_lock_init(&parent->list_lock); |
346 | parent->free_objects = 0; | |
347 | parent->free_touched = 0; | |
348 | } | |
349 | ||
a737b3e2 AM |
350 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
351 | do { \ | |
352 | INIT_LIST_HEAD(listp); \ | |
353 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
354 | } while (0) |
355 | ||
a737b3e2 AM |
356 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
357 | do { \ | |
e498be7d CL |
358 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
359 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
360 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
361 | } while (0) | |
1da177e4 LT |
362 | |
363 | /* | |
343e0d7a | 364 | * struct kmem_cache |
1da177e4 LT |
365 | * |
366 | * manages a cache. | |
367 | */ | |
b28a02de | 368 | |
2109a2d1 | 369 | struct kmem_cache { |
1da177e4 | 370 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 371 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 372 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
373 | unsigned int batchcount; |
374 | unsigned int limit; | |
375 | unsigned int shared; | |
b5d8ca7c | 376 | |
3dafccf2 | 377 | unsigned int buffer_size; |
b5d8ca7c | 378 | /* 3) touched by every alloc & free from the backend */ |
b28a02de | 379 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
b5d8ca7c | 380 | |
a737b3e2 AM |
381 | unsigned int flags; /* constant flags */ |
382 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 383 | |
b5d8ca7c | 384 | /* 4) cache_grow/shrink */ |
1da177e4 | 385 | /* order of pgs per slab (2^n) */ |
b28a02de | 386 | unsigned int gfporder; |
1da177e4 LT |
387 | |
388 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 389 | gfp_t gfpflags; |
1da177e4 | 390 | |
a737b3e2 | 391 | size_t colour; /* cache colouring range */ |
b28a02de | 392 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 393 | struct kmem_cache *slabp_cache; |
b28a02de | 394 | unsigned int slab_size; |
a737b3e2 | 395 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
396 | |
397 | /* constructor func */ | |
343e0d7a | 398 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
399 | |
400 | /* de-constructor func */ | |
343e0d7a | 401 | void (*dtor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 | 402 | |
b5d8ca7c | 403 | /* 5) cache creation/removal */ |
b28a02de PE |
404 | const char *name; |
405 | struct list_head next; | |
1da177e4 | 406 | |
b5d8ca7c | 407 | /* 6) statistics */ |
1da177e4 | 408 | #if STATS |
b28a02de PE |
409 | unsigned long num_active; |
410 | unsigned long num_allocations; | |
411 | unsigned long high_mark; | |
412 | unsigned long grown; | |
413 | unsigned long reaped; | |
414 | unsigned long errors; | |
415 | unsigned long max_freeable; | |
416 | unsigned long node_allocs; | |
417 | unsigned long node_frees; | |
fb7faf33 | 418 | unsigned long node_overflow; |
b28a02de PE |
419 | atomic_t allochit; |
420 | atomic_t allocmiss; | |
421 | atomic_t freehit; | |
422 | atomic_t freemiss; | |
1da177e4 LT |
423 | #endif |
424 | #if DEBUG | |
3dafccf2 MS |
425 | /* |
426 | * If debugging is enabled, then the allocator can add additional | |
427 | * fields and/or padding to every object. buffer_size contains the total | |
428 | * object size including these internal fields, the following two | |
429 | * variables contain the offset to the user object and its size. | |
430 | */ | |
431 | int obj_offset; | |
432 | int obj_size; | |
1da177e4 LT |
433 | #endif |
434 | }; | |
435 | ||
436 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
437 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
438 | ||
439 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
440 | /* |
441 | * Optimization question: fewer reaps means less probability for unnessary | |
442 | * cpucache drain/refill cycles. | |
1da177e4 | 443 | * |
dc6f3f27 | 444 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
445 | * which could lock up otherwise freeable slabs. |
446 | */ | |
447 | #define REAPTIMEOUT_CPUC (2*HZ) | |
448 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
449 | ||
450 | #if STATS | |
451 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
452 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
453 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
454 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
455 | #define STATS_INC_REAPED(x) ((x)->reaped++) | |
a737b3e2 AM |
456 | #define STATS_SET_HIGH(x) \ |
457 | do { \ | |
458 | if ((x)->num_active > (x)->high_mark) \ | |
459 | (x)->high_mark = (x)->num_active; \ | |
460 | } while (0) | |
1da177e4 LT |
461 | #define STATS_INC_ERR(x) ((x)->errors++) |
462 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 463 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 464 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
465 | #define STATS_SET_FREEABLE(x, i) \ |
466 | do { \ | |
467 | if ((x)->max_freeable < i) \ | |
468 | (x)->max_freeable = i; \ | |
469 | } while (0) | |
1da177e4 LT |
470 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
471 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
472 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
473 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
474 | #else | |
475 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
476 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
477 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
478 | #define STATS_INC_GROWN(x) do { } while (0) | |
479 | #define STATS_INC_REAPED(x) do { } while (0) | |
480 | #define STATS_SET_HIGH(x) do { } while (0) | |
481 | #define STATS_INC_ERR(x) do { } while (0) | |
482 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 483 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 484 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 485 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
486 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
487 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
488 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
489 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
490 | #endif | |
491 | ||
492 | #if DEBUG | |
a737b3e2 AM |
493 | /* |
494 | * Magic nums for obj red zoning. | |
1da177e4 LT |
495 | * Placed in the first word before and the first word after an obj. |
496 | */ | |
497 | #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ | |
498 | #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ | |
499 | ||
500 | /* ...and for poisoning */ | |
501 | #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ | |
502 | #define POISON_FREE 0x6b /* for use-after-free poisoning */ | |
503 | #define POISON_END 0xa5 /* end-byte of poisoning */ | |
504 | ||
a737b3e2 AM |
505 | /* |
506 | * memory layout of objects: | |
1da177e4 | 507 | * 0 : objp |
3dafccf2 | 508 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
509 | * the end of an object is aligned with the end of the real |
510 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 511 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 512 | * redzone word. |
3dafccf2 MS |
513 | * cachep->obj_offset: The real object. |
514 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
515 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
516 | * [BYTES_PER_WORD long] | |
1da177e4 | 517 | */ |
343e0d7a | 518 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 519 | { |
3dafccf2 | 520 | return cachep->obj_offset; |
1da177e4 LT |
521 | } |
522 | ||
343e0d7a | 523 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 524 | { |
3dafccf2 | 525 | return cachep->obj_size; |
1da177e4 LT |
526 | } |
527 | ||
343e0d7a | 528 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
529 | { |
530 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 531 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
532 | } |
533 | ||
343e0d7a | 534 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
535 | { |
536 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
537 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 538 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 539 | 2 * BYTES_PER_WORD); |
3dafccf2 | 540 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
541 | } |
542 | ||
343e0d7a | 543 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
544 | { |
545 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 546 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
547 | } |
548 | ||
549 | #else | |
550 | ||
3dafccf2 MS |
551 | #define obj_offset(x) 0 |
552 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
553 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
554 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
555 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
556 | ||
557 | #endif | |
558 | ||
559 | /* | |
a737b3e2 AM |
560 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
561 | * order. | |
1da177e4 LT |
562 | */ |
563 | #if defined(CONFIG_LARGE_ALLOCS) | |
564 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
565 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
566 | #elif defined(CONFIG_MMU) | |
567 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
568 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
569 | #else | |
570 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
571 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
572 | #endif | |
573 | ||
574 | /* | |
575 | * Do not go above this order unless 0 objects fit into the slab. | |
576 | */ | |
577 | #define BREAK_GFP_ORDER_HI 1 | |
578 | #define BREAK_GFP_ORDER_LO 0 | |
579 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
580 | ||
a737b3e2 AM |
581 | /* |
582 | * Functions for storing/retrieving the cachep and or slab from the page | |
583 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
584 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 585 | */ |
065d41cb PE |
586 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
587 | { | |
588 | page->lru.next = (struct list_head *)cache; | |
589 | } | |
590 | ||
591 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
592 | { | |
84097518 NP |
593 | if (unlikely(PageCompound(page))) |
594 | page = (struct page *)page_private(page); | |
065d41cb PE |
595 | return (struct kmem_cache *)page->lru.next; |
596 | } | |
597 | ||
598 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
599 | { | |
600 | page->lru.prev = (struct list_head *)slab; | |
601 | } | |
602 | ||
603 | static inline struct slab *page_get_slab(struct page *page) | |
604 | { | |
84097518 NP |
605 | if (unlikely(PageCompound(page))) |
606 | page = (struct page *)page_private(page); | |
065d41cb PE |
607 | return (struct slab *)page->lru.prev; |
608 | } | |
1da177e4 | 609 | |
6ed5eb22 PE |
610 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
611 | { | |
612 | struct page *page = virt_to_page(obj); | |
613 | return page_get_cache(page); | |
614 | } | |
615 | ||
616 | static inline struct slab *virt_to_slab(const void *obj) | |
617 | { | |
618 | struct page *page = virt_to_page(obj); | |
619 | return page_get_slab(page); | |
620 | } | |
621 | ||
8fea4e96 PE |
622 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
623 | unsigned int idx) | |
624 | { | |
625 | return slab->s_mem + cache->buffer_size * idx; | |
626 | } | |
627 | ||
628 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
629 | struct slab *slab, void *obj) | |
630 | { | |
631 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
632 | } | |
633 | ||
a737b3e2 AM |
634 | /* |
635 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
636 | */ | |
1da177e4 LT |
637 | struct cache_sizes malloc_sizes[] = { |
638 | #define CACHE(x) { .cs_size = (x) }, | |
639 | #include <linux/kmalloc_sizes.h> | |
640 | CACHE(ULONG_MAX) | |
641 | #undef CACHE | |
642 | }; | |
643 | EXPORT_SYMBOL(malloc_sizes); | |
644 | ||
645 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
646 | struct cache_names { | |
647 | char *name; | |
648 | char *name_dma; | |
649 | }; | |
650 | ||
651 | static struct cache_names __initdata cache_names[] = { | |
652 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
653 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 654 | {NULL,} |
1da177e4 LT |
655 | #undef CACHE |
656 | }; | |
657 | ||
658 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 659 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 660 | static struct arraycache_init initarray_generic = |
b28a02de | 661 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
662 | |
663 | /* internal cache of cache description objs */ | |
343e0d7a | 664 | static struct kmem_cache cache_cache = { |
b28a02de PE |
665 | .batchcount = 1, |
666 | .limit = BOOT_CPUCACHE_ENTRIES, | |
667 | .shared = 1, | |
343e0d7a | 668 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 669 | .name = "kmem_cache", |
1da177e4 | 670 | #if DEBUG |
343e0d7a | 671 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
672 | #endif |
673 | }; | |
674 | ||
675 | /* Guard access to the cache-chain. */ | |
fc0abb14 | 676 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
677 | static struct list_head cache_chain; |
678 | ||
679 | /* | |
a737b3e2 AM |
680 | * vm_enough_memory() looks at this to determine how many slab-allocated pages |
681 | * are possibly freeable under pressure | |
1da177e4 LT |
682 | * |
683 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
684 | */ | |
685 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
686 | |
687 | /* | |
688 | * chicken and egg problem: delay the per-cpu array allocation | |
689 | * until the general caches are up. | |
690 | */ | |
691 | static enum { | |
692 | NONE, | |
e498be7d CL |
693 | PARTIAL_AC, |
694 | PARTIAL_L3, | |
1da177e4 LT |
695 | FULL |
696 | } g_cpucache_up; | |
697 | ||
39d24e64 MK |
698 | /* |
699 | * used by boot code to determine if it can use slab based allocator | |
700 | */ | |
701 | int slab_is_available(void) | |
702 | { | |
703 | return g_cpucache_up == FULL; | |
704 | } | |
705 | ||
1da177e4 LT |
706 | static DEFINE_PER_CPU(struct work_struct, reap_work); |
707 | ||
a737b3e2 AM |
708 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, |
709 | int node); | |
343e0d7a | 710 | static void enable_cpucache(struct kmem_cache *cachep); |
b28a02de | 711 | static void cache_reap(void *unused); |
343e0d7a | 712 | static int __node_shrink(struct kmem_cache *cachep, int node); |
1da177e4 | 713 | |
343e0d7a | 714 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
715 | { |
716 | return cachep->array[smp_processor_id()]; | |
717 | } | |
718 | ||
a737b3e2 AM |
719 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
720 | gfp_t gfpflags) | |
1da177e4 LT |
721 | { |
722 | struct cache_sizes *csizep = malloc_sizes; | |
723 | ||
724 | #if DEBUG | |
725 | /* This happens if someone tries to call | |
b28a02de PE |
726 | * kmem_cache_create(), or __kmalloc(), before |
727 | * the generic caches are initialized. | |
728 | */ | |
c7e43c78 | 729 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
730 | #endif |
731 | while (size > csizep->cs_size) | |
732 | csizep++; | |
733 | ||
734 | /* | |
0abf40c1 | 735 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
736 | * has cs_{dma,}cachep==NULL. Thus no special case |
737 | * for large kmalloc calls required. | |
738 | */ | |
739 | if (unlikely(gfpflags & GFP_DMA)) | |
740 | return csizep->cs_dmacachep; | |
741 | return csizep->cs_cachep; | |
742 | } | |
743 | ||
343e0d7a | 744 | struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
745 | { |
746 | return __find_general_cachep(size, gfpflags); | |
747 | } | |
748 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
749 | ||
fbaccacf | 750 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 751 | { |
fbaccacf SR |
752 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
753 | } | |
1da177e4 | 754 | |
a737b3e2 AM |
755 | /* |
756 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
757 | */ | |
fbaccacf SR |
758 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
759 | size_t align, int flags, size_t *left_over, | |
760 | unsigned int *num) | |
761 | { | |
762 | int nr_objs; | |
763 | size_t mgmt_size; | |
764 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 765 | |
fbaccacf SR |
766 | /* |
767 | * The slab management structure can be either off the slab or | |
768 | * on it. For the latter case, the memory allocated for a | |
769 | * slab is used for: | |
770 | * | |
771 | * - The struct slab | |
772 | * - One kmem_bufctl_t for each object | |
773 | * - Padding to respect alignment of @align | |
774 | * - @buffer_size bytes for each object | |
775 | * | |
776 | * If the slab management structure is off the slab, then the | |
777 | * alignment will already be calculated into the size. Because | |
778 | * the slabs are all pages aligned, the objects will be at the | |
779 | * correct alignment when allocated. | |
780 | */ | |
781 | if (flags & CFLGS_OFF_SLAB) { | |
782 | mgmt_size = 0; | |
783 | nr_objs = slab_size / buffer_size; | |
784 | ||
785 | if (nr_objs > SLAB_LIMIT) | |
786 | nr_objs = SLAB_LIMIT; | |
787 | } else { | |
788 | /* | |
789 | * Ignore padding for the initial guess. The padding | |
790 | * is at most @align-1 bytes, and @buffer_size is at | |
791 | * least @align. In the worst case, this result will | |
792 | * be one greater than the number of objects that fit | |
793 | * into the memory allocation when taking the padding | |
794 | * into account. | |
795 | */ | |
796 | nr_objs = (slab_size - sizeof(struct slab)) / | |
797 | (buffer_size + sizeof(kmem_bufctl_t)); | |
798 | ||
799 | /* | |
800 | * This calculated number will be either the right | |
801 | * amount, or one greater than what we want. | |
802 | */ | |
803 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
804 | > slab_size) | |
805 | nr_objs--; | |
806 | ||
807 | if (nr_objs > SLAB_LIMIT) | |
808 | nr_objs = SLAB_LIMIT; | |
809 | ||
810 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
811 | } | |
812 | *num = nr_objs; | |
813 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
814 | } |
815 | ||
816 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
817 | ||
a737b3e2 AM |
818 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
819 | char *msg) | |
1da177e4 LT |
820 | { |
821 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 822 | function, cachep->name, msg); |
1da177e4 LT |
823 | dump_stack(); |
824 | } | |
825 | ||
8fce4d8e CL |
826 | #ifdef CONFIG_NUMA |
827 | /* | |
828 | * Special reaping functions for NUMA systems called from cache_reap(). | |
829 | * These take care of doing round robin flushing of alien caches (containing | |
830 | * objects freed on different nodes from which they were allocated) and the | |
831 | * flushing of remote pcps by calling drain_node_pages. | |
832 | */ | |
833 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
834 | ||
835 | static void init_reap_node(int cpu) | |
836 | { | |
837 | int node; | |
838 | ||
839 | node = next_node(cpu_to_node(cpu), node_online_map); | |
840 | if (node == MAX_NUMNODES) | |
442295c9 | 841 | node = first_node(node_online_map); |
8fce4d8e CL |
842 | |
843 | __get_cpu_var(reap_node) = node; | |
844 | } | |
845 | ||
846 | static void next_reap_node(void) | |
847 | { | |
848 | int node = __get_cpu_var(reap_node); | |
849 | ||
850 | /* | |
851 | * Also drain per cpu pages on remote zones | |
852 | */ | |
853 | if (node != numa_node_id()) | |
854 | drain_node_pages(node); | |
855 | ||
856 | node = next_node(node, node_online_map); | |
857 | if (unlikely(node >= MAX_NUMNODES)) | |
858 | node = first_node(node_online_map); | |
859 | __get_cpu_var(reap_node) = node; | |
860 | } | |
861 | ||
862 | #else | |
863 | #define init_reap_node(cpu) do { } while (0) | |
864 | #define next_reap_node(void) do { } while (0) | |
865 | #endif | |
866 | ||
1da177e4 LT |
867 | /* |
868 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
869 | * via the workqueue/eventd. | |
870 | * Add the CPU number into the expiration time to minimize the possibility of | |
871 | * the CPUs getting into lockstep and contending for the global cache chain | |
872 | * lock. | |
873 | */ | |
874 | static void __devinit start_cpu_timer(int cpu) | |
875 | { | |
876 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
877 | ||
878 | /* | |
879 | * When this gets called from do_initcalls via cpucache_init(), | |
880 | * init_workqueues() has already run, so keventd will be setup | |
881 | * at that time. | |
882 | */ | |
883 | if (keventd_up() && reap_work->func == NULL) { | |
8fce4d8e | 884 | init_reap_node(cpu); |
1da177e4 LT |
885 | INIT_WORK(reap_work, cache_reap, NULL); |
886 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
887 | } | |
888 | } | |
889 | ||
e498be7d | 890 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 891 | int batchcount) |
1da177e4 | 892 | { |
b28a02de | 893 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
894 | struct array_cache *nc = NULL; |
895 | ||
e498be7d | 896 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
897 | if (nc) { |
898 | nc->avail = 0; | |
899 | nc->limit = entries; | |
900 | nc->batchcount = batchcount; | |
901 | nc->touched = 0; | |
e498be7d | 902 | spin_lock_init(&nc->lock); |
1da177e4 LT |
903 | } |
904 | return nc; | |
905 | } | |
906 | ||
3ded175a CL |
907 | /* |
908 | * Transfer objects in one arraycache to another. | |
909 | * Locking must be handled by the caller. | |
910 | * | |
911 | * Return the number of entries transferred. | |
912 | */ | |
913 | static int transfer_objects(struct array_cache *to, | |
914 | struct array_cache *from, unsigned int max) | |
915 | { | |
916 | /* Figure out how many entries to transfer */ | |
917 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
918 | ||
919 | if (!nr) | |
920 | return 0; | |
921 | ||
922 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
923 | sizeof(void *) *nr); | |
924 | ||
925 | from->avail -= nr; | |
926 | to->avail += nr; | |
927 | to->touched = 1; | |
928 | return nr; | |
929 | } | |
930 | ||
e498be7d | 931 | #ifdef CONFIG_NUMA |
343e0d7a | 932 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 933 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 934 | |
5295a74c | 935 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
936 | { |
937 | struct array_cache **ac_ptr; | |
b28a02de | 938 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
939 | int i; |
940 | ||
941 | if (limit > 1) | |
942 | limit = 12; | |
943 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
944 | if (ac_ptr) { | |
945 | for_each_node(i) { | |
946 | if (i == node || !node_online(i)) { | |
947 | ac_ptr[i] = NULL; | |
948 | continue; | |
949 | } | |
950 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
951 | if (!ac_ptr[i]) { | |
b28a02de | 952 | for (i--; i <= 0; i--) |
e498be7d CL |
953 | kfree(ac_ptr[i]); |
954 | kfree(ac_ptr); | |
955 | return NULL; | |
956 | } | |
957 | } | |
958 | } | |
959 | return ac_ptr; | |
960 | } | |
961 | ||
5295a74c | 962 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
963 | { |
964 | int i; | |
965 | ||
966 | if (!ac_ptr) | |
967 | return; | |
e498be7d | 968 | for_each_node(i) |
b28a02de | 969 | kfree(ac_ptr[i]); |
e498be7d CL |
970 | kfree(ac_ptr); |
971 | } | |
972 | ||
343e0d7a | 973 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 974 | struct array_cache *ac, int node) |
e498be7d CL |
975 | { |
976 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
977 | ||
978 | if (ac->avail) { | |
979 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
980 | /* |
981 | * Stuff objects into the remote nodes shared array first. | |
982 | * That way we could avoid the overhead of putting the objects | |
983 | * into the free lists and getting them back later. | |
984 | */ | |
693f7d36 | 985 | if (rl3->shared) |
986 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 987 | |
ff69416e | 988 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
989 | ac->avail = 0; |
990 | spin_unlock(&rl3->list_lock); | |
991 | } | |
992 | } | |
993 | ||
8fce4d8e CL |
994 | /* |
995 | * Called from cache_reap() to regularly drain alien caches round robin. | |
996 | */ | |
997 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
998 | { | |
999 | int node = __get_cpu_var(reap_node); | |
1000 | ||
1001 | if (l3->alien) { | |
1002 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1003 | |
1004 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1005 | __drain_alien_cache(cachep, ac, node); |
1006 | spin_unlock_irq(&ac->lock); | |
1007 | } | |
1008 | } | |
1009 | } | |
1010 | ||
a737b3e2 AM |
1011 | static void drain_alien_cache(struct kmem_cache *cachep, |
1012 | struct array_cache **alien) | |
e498be7d | 1013 | { |
b28a02de | 1014 | int i = 0; |
e498be7d CL |
1015 | struct array_cache *ac; |
1016 | unsigned long flags; | |
1017 | ||
1018 | for_each_online_node(i) { | |
4484ebf1 | 1019 | ac = alien[i]; |
e498be7d CL |
1020 | if (ac) { |
1021 | spin_lock_irqsave(&ac->lock, flags); | |
1022 | __drain_alien_cache(cachep, ac, i); | |
1023 | spin_unlock_irqrestore(&ac->lock, flags); | |
1024 | } | |
1025 | } | |
1026 | } | |
729bd0b7 PE |
1027 | |
1028 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1029 | { | |
1030 | struct slab *slabp = virt_to_slab(objp); | |
1031 | int nodeid = slabp->nodeid; | |
1032 | struct kmem_list3 *l3; | |
1033 | struct array_cache *alien = NULL; | |
1034 | ||
1035 | /* | |
1036 | * Make sure we are not freeing a object from another node to the array | |
1037 | * cache on this cpu. | |
1038 | */ | |
1039 | if (likely(slabp->nodeid == numa_node_id())) | |
1040 | return 0; | |
1041 | ||
1042 | l3 = cachep->nodelists[numa_node_id()]; | |
1043 | STATS_INC_NODEFREES(cachep); | |
1044 | if (l3->alien && l3->alien[nodeid]) { | |
1045 | alien = l3->alien[nodeid]; | |
1046 | spin_lock(&alien->lock); | |
1047 | if (unlikely(alien->avail == alien->limit)) { | |
1048 | STATS_INC_ACOVERFLOW(cachep); | |
1049 | __drain_alien_cache(cachep, alien, nodeid); | |
1050 | } | |
1051 | alien->entry[alien->avail++] = objp; | |
1052 | spin_unlock(&alien->lock); | |
1053 | } else { | |
1054 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1055 | free_block(cachep, &objp, 1, nodeid); | |
1056 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1057 | } | |
1058 | return 1; | |
1059 | } | |
1060 | ||
e498be7d | 1061 | #else |
7a21ef6f | 1062 | |
4484ebf1 | 1063 | #define drain_alien_cache(cachep, alien) do { } while (0) |
8fce4d8e | 1064 | #define reap_alien(cachep, l3) do { } while (0) |
4484ebf1 | 1065 | |
7a21ef6f LT |
1066 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
1067 | { | |
1068 | return (struct array_cache **) 0x01020304ul; | |
1069 | } | |
1070 | ||
4484ebf1 RT |
1071 | static inline void free_alien_cache(struct array_cache **ac_ptr) |
1072 | { | |
1073 | } | |
7a21ef6f | 1074 | |
729bd0b7 PE |
1075 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
1076 | { | |
1077 | return 0; | |
1078 | } | |
1079 | ||
e498be7d CL |
1080 | #endif |
1081 | ||
83d722f7 | 1082 | static int cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1083 | unsigned long action, void *hcpu) |
1da177e4 LT |
1084 | { |
1085 | long cpu = (long)hcpu; | |
343e0d7a | 1086 | struct kmem_cache *cachep; |
e498be7d CL |
1087 | struct kmem_list3 *l3 = NULL; |
1088 | int node = cpu_to_node(cpu); | |
1089 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1090 | |
1091 | switch (action) { | |
1092 | case CPU_UP_PREPARE: | |
fc0abb14 | 1093 | mutex_lock(&cache_chain_mutex); |
a737b3e2 AM |
1094 | /* |
1095 | * We need to do this right in the beginning since | |
e498be7d CL |
1096 | * alloc_arraycache's are going to use this list. |
1097 | * kmalloc_node allows us to add the slab to the right | |
1098 | * kmem_list3 and not this cpu's kmem_list3 | |
1099 | */ | |
1100 | ||
1da177e4 | 1101 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1102 | /* |
1103 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1104 | * begin anything. Make sure some other cpu on this |
1105 | * node has not already allocated this | |
1106 | */ | |
1107 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1108 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1109 | if (!l3) | |
e498be7d CL |
1110 | goto bad; |
1111 | kmem_list3_init(l3); | |
1112 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1113 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1114 | |
4484ebf1 RT |
1115 | /* |
1116 | * The l3s don't come and go as CPUs come and | |
1117 | * go. cache_chain_mutex is sufficient | |
1118 | * protection here. | |
1119 | */ | |
e498be7d CL |
1120 | cachep->nodelists[node] = l3; |
1121 | } | |
1da177e4 | 1122 | |
e498be7d CL |
1123 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1124 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1125 | (1 + nr_cpus_node(node)) * |
1126 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1127 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1128 | } | |
1129 | ||
a737b3e2 AM |
1130 | /* |
1131 | * Now we can go ahead with allocating the shared arrays and | |
1132 | * array caches | |
1133 | */ | |
e498be7d | 1134 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1135 | struct array_cache *nc; |
4484ebf1 RT |
1136 | struct array_cache *shared; |
1137 | struct array_cache **alien; | |
cd105df4 | 1138 | |
e498be7d | 1139 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1140 | cachep->batchcount); |
1da177e4 LT |
1141 | if (!nc) |
1142 | goto bad; | |
4484ebf1 RT |
1143 | shared = alloc_arraycache(node, |
1144 | cachep->shared * cachep->batchcount, | |
1145 | 0xbaadf00d); | |
1146 | if (!shared) | |
1147 | goto bad; | |
7a21ef6f | 1148 | |
4484ebf1 RT |
1149 | alien = alloc_alien_cache(node, cachep->limit); |
1150 | if (!alien) | |
1151 | goto bad; | |
1da177e4 | 1152 | cachep->array[cpu] = nc; |
e498be7d CL |
1153 | l3 = cachep->nodelists[node]; |
1154 | BUG_ON(!l3); | |
e498be7d | 1155 | |
4484ebf1 RT |
1156 | spin_lock_irq(&l3->list_lock); |
1157 | if (!l3->shared) { | |
1158 | /* | |
1159 | * We are serialised from CPU_DEAD or | |
1160 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1161 | */ | |
1162 | l3->shared = shared; | |
1163 | shared = NULL; | |
e498be7d | 1164 | } |
4484ebf1 RT |
1165 | #ifdef CONFIG_NUMA |
1166 | if (!l3->alien) { | |
1167 | l3->alien = alien; | |
1168 | alien = NULL; | |
1169 | } | |
1170 | #endif | |
1171 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1172 | kfree(shared); |
1173 | free_alien_cache(alien); | |
1da177e4 | 1174 | } |
fc0abb14 | 1175 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1176 | break; |
1177 | case CPU_ONLINE: | |
1178 | start_cpu_timer(cpu); | |
1179 | break; | |
1180 | #ifdef CONFIG_HOTPLUG_CPU | |
1181 | case CPU_DEAD: | |
4484ebf1 RT |
1182 | /* |
1183 | * Even if all the cpus of a node are down, we don't free the | |
1184 | * kmem_list3 of any cache. This to avoid a race between | |
1185 | * cpu_down, and a kmalloc allocation from another cpu for | |
1186 | * memory from the node of the cpu going down. The list3 | |
1187 | * structure is usually allocated from kmem_cache_create() and | |
1188 | * gets destroyed at kmem_cache_destroy(). | |
1189 | */ | |
1da177e4 LT |
1190 | /* fall thru */ |
1191 | case CPU_UP_CANCELED: | |
fc0abb14 | 1192 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1193 | list_for_each_entry(cachep, &cache_chain, next) { |
1194 | struct array_cache *nc; | |
4484ebf1 RT |
1195 | struct array_cache *shared; |
1196 | struct array_cache **alien; | |
e498be7d | 1197 | cpumask_t mask; |
1da177e4 | 1198 | |
e498be7d | 1199 | mask = node_to_cpumask(node); |
1da177e4 LT |
1200 | /* cpu is dead; no one can alloc from it. */ |
1201 | nc = cachep->array[cpu]; | |
1202 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1203 | l3 = cachep->nodelists[node]; |
1204 | ||
1205 | if (!l3) | |
4484ebf1 | 1206 | goto free_array_cache; |
e498be7d | 1207 | |
ca3b9b91 | 1208 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1209 | |
1210 | /* Free limit for this kmem_list3 */ | |
1211 | l3->free_limit -= cachep->batchcount; | |
1212 | if (nc) | |
ff69416e | 1213 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1214 | |
1215 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1216 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1217 | goto free_array_cache; |
b28a02de | 1218 | } |
e498be7d | 1219 | |
4484ebf1 RT |
1220 | shared = l3->shared; |
1221 | if (shared) { | |
e498be7d | 1222 | free_block(cachep, l3->shared->entry, |
b28a02de | 1223 | l3->shared->avail, node); |
e498be7d CL |
1224 | l3->shared = NULL; |
1225 | } | |
e498be7d | 1226 | |
4484ebf1 RT |
1227 | alien = l3->alien; |
1228 | l3->alien = NULL; | |
1229 | ||
1230 | spin_unlock_irq(&l3->list_lock); | |
1231 | ||
1232 | kfree(shared); | |
1233 | if (alien) { | |
1234 | drain_alien_cache(cachep, alien); | |
1235 | free_alien_cache(alien); | |
e498be7d | 1236 | } |
4484ebf1 | 1237 | free_array_cache: |
1da177e4 LT |
1238 | kfree(nc); |
1239 | } | |
4484ebf1 RT |
1240 | /* |
1241 | * In the previous loop, all the objects were freed to | |
1242 | * the respective cache's slabs, now we can go ahead and | |
1243 | * shrink each nodelist to its limit. | |
1244 | */ | |
1245 | list_for_each_entry(cachep, &cache_chain, next) { | |
1246 | l3 = cachep->nodelists[node]; | |
1247 | if (!l3) | |
1248 | continue; | |
1249 | spin_lock_irq(&l3->list_lock); | |
1250 | /* free slabs belonging to this node */ | |
1251 | __node_shrink(cachep, node); | |
1252 | spin_unlock_irq(&l3->list_lock); | |
1253 | } | |
fc0abb14 | 1254 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1255 | break; |
1256 | #endif | |
1257 | } | |
1258 | return NOTIFY_OK; | |
a737b3e2 | 1259 | bad: |
fc0abb14 | 1260 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1261 | return NOTIFY_BAD; |
1262 | } | |
1263 | ||
1264 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
1265 | ||
e498be7d CL |
1266 | /* |
1267 | * swap the static kmem_list3 with kmalloced memory | |
1268 | */ | |
a737b3e2 AM |
1269 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1270 | int nodeid) | |
e498be7d CL |
1271 | { |
1272 | struct kmem_list3 *ptr; | |
1273 | ||
1274 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1275 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1276 | BUG_ON(!ptr); | |
1277 | ||
1278 | local_irq_disable(); | |
1279 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1280 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1281 | cachep->nodelists[nodeid] = ptr; | |
1282 | local_irq_enable(); | |
1283 | } | |
1284 | ||
a737b3e2 AM |
1285 | /* |
1286 | * Initialisation. Called after the page allocator have been initialised and | |
1287 | * before smp_init(). | |
1da177e4 LT |
1288 | */ |
1289 | void __init kmem_cache_init(void) | |
1290 | { | |
1291 | size_t left_over; | |
1292 | struct cache_sizes *sizes; | |
1293 | struct cache_names *names; | |
e498be7d | 1294 | int i; |
07ed76b2 | 1295 | int order; |
e498be7d CL |
1296 | |
1297 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1298 | kmem_list3_init(&initkmem_list3[i]); | |
1299 | if (i < MAX_NUMNODES) | |
1300 | cache_cache.nodelists[i] = NULL; | |
1301 | } | |
1da177e4 LT |
1302 | |
1303 | /* | |
1304 | * Fragmentation resistance on low memory - only use bigger | |
1305 | * page orders on machines with more than 32MB of memory. | |
1306 | */ | |
1307 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1308 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1309 | ||
1da177e4 LT |
1310 | /* Bootstrap is tricky, because several objects are allocated |
1311 | * from caches that do not exist yet: | |
a737b3e2 AM |
1312 | * 1) initialize the cache_cache cache: it contains the struct |
1313 | * kmem_cache structures of all caches, except cache_cache itself: | |
1314 | * cache_cache is statically allocated. | |
e498be7d CL |
1315 | * Initially an __init data area is used for the head array and the |
1316 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1317 | * array at the end of the bootstrap. | |
1da177e4 | 1318 | * 2) Create the first kmalloc cache. |
343e0d7a | 1319 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1320 | * An __init data area is used for the head array. |
1321 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1322 | * head arrays. | |
1da177e4 LT |
1323 | * 4) Replace the __init data head arrays for cache_cache and the first |
1324 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1325 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1326 | * the other cache's with kmalloc allocated memory. | |
1327 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1328 | */ |
1329 | ||
1330 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1331 | INIT_LIST_HEAD(&cache_chain); |
1332 | list_add(&cache_cache.next, &cache_chain); | |
1333 | cache_cache.colour_off = cache_line_size(); | |
1334 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1335 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1336 | |
a737b3e2 AM |
1337 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1338 | cache_line_size()); | |
1da177e4 | 1339 | |
07ed76b2 JS |
1340 | for (order = 0; order < MAX_ORDER; order++) { |
1341 | cache_estimate(order, cache_cache.buffer_size, | |
1342 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1343 | if (cache_cache.num) | |
1344 | break; | |
1345 | } | |
40094fa6 | 1346 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1347 | cache_cache.gfporder = order; |
b28a02de | 1348 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1349 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1350 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1351 | |
1352 | /* 2+3) create the kmalloc caches */ | |
1353 | sizes = malloc_sizes; | |
1354 | names = cache_names; | |
1355 | ||
a737b3e2 AM |
1356 | /* |
1357 | * Initialize the caches that provide memory for the array cache and the | |
1358 | * kmem_list3 structures first. Without this, further allocations will | |
1359 | * bug. | |
e498be7d CL |
1360 | */ |
1361 | ||
1362 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1363 | sizes[INDEX_AC].cs_size, |
1364 | ARCH_KMALLOC_MINALIGN, | |
1365 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1366 | NULL, NULL); | |
e498be7d | 1367 | |
a737b3e2 | 1368 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1369 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1370 | kmem_cache_create(names[INDEX_L3].name, |
1371 | sizes[INDEX_L3].cs_size, | |
1372 | ARCH_KMALLOC_MINALIGN, | |
1373 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1374 | NULL, NULL); | |
1375 | } | |
e498be7d | 1376 | |
1da177e4 | 1377 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1378 | /* |
1379 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1380 | * This should be particularly beneficial on SMP boxes, as it |
1381 | * eliminates "false sharing". | |
1382 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1383 | * allow tighter packing of the smaller caches. |
1384 | */ | |
a737b3e2 | 1385 | if (!sizes->cs_cachep) { |
e498be7d | 1386 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1387 | sizes->cs_size, |
1388 | ARCH_KMALLOC_MINALIGN, | |
1389 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1390 | NULL, NULL); | |
1391 | } | |
1da177e4 | 1392 | |
1da177e4 | 1393 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, |
a737b3e2 AM |
1394 | sizes->cs_size, |
1395 | ARCH_KMALLOC_MINALIGN, | |
1396 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1397 | SLAB_PANIC, | |
1398 | NULL, NULL); | |
1da177e4 LT |
1399 | sizes++; |
1400 | names++; | |
1401 | } | |
1402 | /* 4) Replace the bootstrap head arrays */ | |
1403 | { | |
b28a02de | 1404 | void *ptr; |
e498be7d | 1405 | |
1da177e4 | 1406 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1407 | |
1da177e4 | 1408 | local_irq_disable(); |
9a2dba4b PE |
1409 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1410 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1411 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1412 | cache_cache.array[smp_processor_id()] = ptr; |
1413 | local_irq_enable(); | |
e498be7d | 1414 | |
1da177e4 | 1415 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1416 | |
1da177e4 | 1417 | local_irq_disable(); |
9a2dba4b | 1418 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1419 | != &initarray_generic.cache); |
9a2dba4b | 1420 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1421 | sizeof(struct arraycache_init)); |
e498be7d | 1422 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1423 | ptr; |
1da177e4 LT |
1424 | local_irq_enable(); |
1425 | } | |
e498be7d CL |
1426 | /* 5) Replace the bootstrap kmem_list3's */ |
1427 | { | |
1428 | int node; | |
1429 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1430 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1431 | numa_node_id()); |
e498be7d CL |
1432 | |
1433 | for_each_online_node(node) { | |
1434 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1435 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1436 | |
1437 | if (INDEX_AC != INDEX_L3) { | |
1438 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1439 | &initkmem_list3[SIZE_L3 + node], |
1440 | node); | |
e498be7d CL |
1441 | } |
1442 | } | |
1443 | } | |
1da177e4 | 1444 | |
e498be7d | 1445 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1446 | { |
343e0d7a | 1447 | struct kmem_cache *cachep; |
fc0abb14 | 1448 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1449 | list_for_each_entry(cachep, &cache_chain, next) |
a737b3e2 | 1450 | enable_cpucache(cachep); |
fc0abb14 | 1451 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1452 | } |
1453 | ||
1454 | /* Done! */ | |
1455 | g_cpucache_up = FULL; | |
1456 | ||
a737b3e2 AM |
1457 | /* |
1458 | * Register a cpu startup notifier callback that initializes | |
1459 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1460 | */ |
1461 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1462 | |
a737b3e2 AM |
1463 | /* |
1464 | * The reap timers are started later, with a module init call: That part | |
1465 | * of the kernel is not yet operational. | |
1da177e4 LT |
1466 | */ |
1467 | } | |
1468 | ||
1469 | static int __init cpucache_init(void) | |
1470 | { | |
1471 | int cpu; | |
1472 | ||
a737b3e2 AM |
1473 | /* |
1474 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1475 | */ |
e498be7d | 1476 | for_each_online_cpu(cpu) |
a737b3e2 | 1477 | start_cpu_timer(cpu); |
1da177e4 LT |
1478 | return 0; |
1479 | } | |
1da177e4 LT |
1480 | __initcall(cpucache_init); |
1481 | ||
1482 | /* | |
1483 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1484 | * | |
1485 | * If we requested dmaable memory, we will get it. Even if we | |
1486 | * did not request dmaable memory, we might get it, but that | |
1487 | * would be relatively rare and ignorable. | |
1488 | */ | |
343e0d7a | 1489 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1490 | { |
1491 | struct page *page; | |
e1b6aa6f | 1492 | int nr_pages; |
1da177e4 LT |
1493 | int i; |
1494 | ||
d6fef9da | 1495 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1496 | /* |
1497 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1498 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1499 | */ |
e1b6aa6f | 1500 | flags |= __GFP_COMP; |
d6fef9da | 1501 | #endif |
e1b6aa6f CH |
1502 | flags |= cachep->gfpflags; |
1503 | ||
1504 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1505 | if (!page) |
1506 | return NULL; | |
1da177e4 | 1507 | |
e1b6aa6f | 1508 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1509 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
e1b6aa6f CH |
1510 | atomic_add(nr_pages, &slab_reclaim_pages); |
1511 | add_page_state(nr_slab, nr_pages); | |
1512 | for (i = 0; i < nr_pages; i++) | |
1513 | __SetPageSlab(page + i); | |
1514 | return page_address(page); | |
1da177e4 LT |
1515 | } |
1516 | ||
1517 | /* | |
1518 | * Interface to system's page release. | |
1519 | */ | |
343e0d7a | 1520 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1521 | { |
b28a02de | 1522 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1523 | struct page *page = virt_to_page(addr); |
1524 | const unsigned long nr_freed = i; | |
1525 | ||
1526 | while (i--) { | |
f205b2fe NP |
1527 | BUG_ON(!PageSlab(page)); |
1528 | __ClearPageSlab(page); | |
1da177e4 LT |
1529 | page++; |
1530 | } | |
1531 | sub_page_state(nr_slab, nr_freed); | |
1532 | if (current->reclaim_state) | |
1533 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1534 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1535 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1536 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1537 | } |
1538 | ||
1539 | static void kmem_rcu_free(struct rcu_head *head) | |
1540 | { | |
b28a02de | 1541 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1542 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1543 | |
1544 | kmem_freepages(cachep, slab_rcu->addr); | |
1545 | if (OFF_SLAB(cachep)) | |
1546 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1547 | } | |
1548 | ||
1549 | #if DEBUG | |
1550 | ||
1551 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1552 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1553 | unsigned long caller) |
1da177e4 | 1554 | { |
3dafccf2 | 1555 | int size = obj_size(cachep); |
1da177e4 | 1556 | |
3dafccf2 | 1557 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1558 | |
b28a02de | 1559 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1560 | return; |
1561 | ||
b28a02de PE |
1562 | *addr++ = 0x12345678; |
1563 | *addr++ = caller; | |
1564 | *addr++ = smp_processor_id(); | |
1565 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1566 | { |
1567 | unsigned long *sptr = &caller; | |
1568 | unsigned long svalue; | |
1569 | ||
1570 | while (!kstack_end(sptr)) { | |
1571 | svalue = *sptr++; | |
1572 | if (kernel_text_address(svalue)) { | |
b28a02de | 1573 | *addr++ = svalue; |
1da177e4 LT |
1574 | size -= sizeof(unsigned long); |
1575 | if (size <= sizeof(unsigned long)) | |
1576 | break; | |
1577 | } | |
1578 | } | |
1579 | ||
1580 | } | |
b28a02de | 1581 | *addr++ = 0x87654321; |
1da177e4 LT |
1582 | } |
1583 | #endif | |
1584 | ||
343e0d7a | 1585 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1586 | { |
3dafccf2 MS |
1587 | int size = obj_size(cachep); |
1588 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1589 | |
1590 | memset(addr, val, size); | |
b28a02de | 1591 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1592 | } |
1593 | ||
1594 | static void dump_line(char *data, int offset, int limit) | |
1595 | { | |
1596 | int i; | |
1597 | printk(KERN_ERR "%03x:", offset); | |
a737b3e2 | 1598 | for (i = 0; i < limit; i++) |
b28a02de | 1599 | printk(" %02x", (unsigned char)data[offset + i]); |
1da177e4 LT |
1600 | printk("\n"); |
1601 | } | |
1602 | #endif | |
1603 | ||
1604 | #if DEBUG | |
1605 | ||
343e0d7a | 1606 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1607 | { |
1608 | int i, size; | |
1609 | char *realobj; | |
1610 | ||
1611 | if (cachep->flags & SLAB_RED_ZONE) { | |
1612 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
a737b3e2 AM |
1613 | *dbg_redzone1(cachep, objp), |
1614 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1615 | } |
1616 | ||
1617 | if (cachep->flags & SLAB_STORE_USER) { | |
1618 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1619 | *dbg_userword(cachep, objp)); |
1da177e4 | 1620 | print_symbol("(%s)", |
a737b3e2 | 1621 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1622 | printk("\n"); |
1623 | } | |
3dafccf2 MS |
1624 | realobj = (char *)objp + obj_offset(cachep); |
1625 | size = obj_size(cachep); | |
b28a02de | 1626 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1627 | int limit; |
1628 | limit = 16; | |
b28a02de PE |
1629 | if (i + limit > size) |
1630 | limit = size - i; | |
1da177e4 LT |
1631 | dump_line(realobj, i, limit); |
1632 | } | |
1633 | } | |
1634 | ||
343e0d7a | 1635 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1636 | { |
1637 | char *realobj; | |
1638 | int size, i; | |
1639 | int lines = 0; | |
1640 | ||
3dafccf2 MS |
1641 | realobj = (char *)objp + obj_offset(cachep); |
1642 | size = obj_size(cachep); | |
1da177e4 | 1643 | |
b28a02de | 1644 | for (i = 0; i < size; i++) { |
1da177e4 | 1645 | char exp = POISON_FREE; |
b28a02de | 1646 | if (i == size - 1) |
1da177e4 LT |
1647 | exp = POISON_END; |
1648 | if (realobj[i] != exp) { | |
1649 | int limit; | |
1650 | /* Mismatch ! */ | |
1651 | /* Print header */ | |
1652 | if (lines == 0) { | |
b28a02de | 1653 | printk(KERN_ERR |
a737b3e2 AM |
1654 | "Slab corruption: start=%p, len=%d\n", |
1655 | realobj, size); | |
1da177e4 LT |
1656 | print_objinfo(cachep, objp, 0); |
1657 | } | |
1658 | /* Hexdump the affected line */ | |
b28a02de | 1659 | i = (i / 16) * 16; |
1da177e4 | 1660 | limit = 16; |
b28a02de PE |
1661 | if (i + limit > size) |
1662 | limit = size - i; | |
1da177e4 LT |
1663 | dump_line(realobj, i, limit); |
1664 | i += 16; | |
1665 | lines++; | |
1666 | /* Limit to 5 lines */ | |
1667 | if (lines > 5) | |
1668 | break; | |
1669 | } | |
1670 | } | |
1671 | if (lines != 0) { | |
1672 | /* Print some data about the neighboring objects, if they | |
1673 | * exist: | |
1674 | */ | |
6ed5eb22 | 1675 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1676 | unsigned int objnr; |
1da177e4 | 1677 | |
8fea4e96 | 1678 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1679 | if (objnr) { |
8fea4e96 | 1680 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1681 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1682 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1683 | realobj, size); |
1da177e4 LT |
1684 | print_objinfo(cachep, objp, 2); |
1685 | } | |
b28a02de | 1686 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1687 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1688 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1689 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1690 | realobj, size); |
1da177e4 LT |
1691 | print_objinfo(cachep, objp, 2); |
1692 | } | |
1693 | } | |
1694 | } | |
1695 | #endif | |
1696 | ||
12dd36fa MD |
1697 | #if DEBUG |
1698 | /** | |
911851e6 RD |
1699 | * slab_destroy_objs - destroy a slab and its objects |
1700 | * @cachep: cache pointer being destroyed | |
1701 | * @slabp: slab pointer being destroyed | |
1702 | * | |
1703 | * Call the registered destructor for each object in a slab that is being | |
1704 | * destroyed. | |
1da177e4 | 1705 | */ |
343e0d7a | 1706 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1707 | { |
1da177e4 LT |
1708 | int i; |
1709 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1710 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1711 | |
1712 | if (cachep->flags & SLAB_POISON) { | |
1713 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1714 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1715 | OFF_SLAB(cachep)) | |
b28a02de | 1716 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1717 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1718 | else |
1719 | check_poison_obj(cachep, objp); | |
1720 | #else | |
1721 | check_poison_obj(cachep, objp); | |
1722 | #endif | |
1723 | } | |
1724 | if (cachep->flags & SLAB_RED_ZONE) { | |
1725 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1726 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1727 | "was overwritten"); |
1da177e4 LT |
1728 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1729 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1730 | "was overwritten"); |
1da177e4 LT |
1731 | } |
1732 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1733 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1734 | } |
12dd36fa | 1735 | } |
1da177e4 | 1736 | #else |
343e0d7a | 1737 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1738 | { |
1da177e4 LT |
1739 | if (cachep->dtor) { |
1740 | int i; | |
1741 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1742 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1743 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1744 | } |
1745 | } | |
12dd36fa | 1746 | } |
1da177e4 LT |
1747 | #endif |
1748 | ||
911851e6 RD |
1749 | /** |
1750 | * slab_destroy - destroy and release all objects in a slab | |
1751 | * @cachep: cache pointer being destroyed | |
1752 | * @slabp: slab pointer being destroyed | |
1753 | * | |
12dd36fa | 1754 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1755 | * Before calling the slab must have been unlinked from the cache. The |
1756 | * cache-lock is not held/needed. | |
12dd36fa | 1757 | */ |
343e0d7a | 1758 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1759 | { |
1760 | void *addr = slabp->s_mem - slabp->colouroff; | |
1761 | ||
1762 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1763 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1764 | struct slab_rcu *slab_rcu; | |
1765 | ||
b28a02de | 1766 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1767 | slab_rcu->cachep = cachep; |
1768 | slab_rcu->addr = addr; | |
1769 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1770 | } else { | |
1771 | kmem_freepages(cachep, addr); | |
1772 | if (OFF_SLAB(cachep)) | |
1773 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1774 | } | |
1775 | } | |
1776 | ||
a737b3e2 AM |
1777 | /* |
1778 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1779 | * size of kmem_list3. | |
1780 | */ | |
343e0d7a | 1781 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1782 | { |
1783 | int node; | |
1784 | ||
1785 | for_each_online_node(node) { | |
b28a02de | 1786 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1787 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1788 | REAPTIMEOUT_LIST3 + |
1789 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1790 | } |
1791 | } | |
1792 | ||
4d268eba | 1793 | /** |
a70773dd RD |
1794 | * calculate_slab_order - calculate size (page order) of slabs |
1795 | * @cachep: pointer to the cache that is being created | |
1796 | * @size: size of objects to be created in this cache. | |
1797 | * @align: required alignment for the objects. | |
1798 | * @flags: slab allocation flags | |
1799 | * | |
1800 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1801 | * |
1802 | * This could be made much more intelligent. For now, try to avoid using | |
1803 | * high order pages for slabs. When the gfp() functions are more friendly | |
1804 | * towards high-order requests, this should be changed. | |
1805 | */ | |
a737b3e2 | 1806 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1807 | size_t size, size_t align, unsigned long flags) |
4d268eba | 1808 | { |
b1ab41c4 | 1809 | unsigned long offslab_limit; |
4d268eba | 1810 | size_t left_over = 0; |
9888e6fa | 1811 | int gfporder; |
4d268eba | 1812 | |
a737b3e2 | 1813 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1814 | unsigned int num; |
1815 | size_t remainder; | |
1816 | ||
9888e6fa | 1817 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1818 | if (!num) |
1819 | continue; | |
9888e6fa | 1820 | |
b1ab41c4 IM |
1821 | if (flags & CFLGS_OFF_SLAB) { |
1822 | /* | |
1823 | * Max number of objs-per-slab for caches which | |
1824 | * use off-slab slabs. Needed to avoid a possible | |
1825 | * looping condition in cache_grow(). | |
1826 | */ | |
1827 | offslab_limit = size - sizeof(struct slab); | |
1828 | offslab_limit /= sizeof(kmem_bufctl_t); | |
1829 | ||
1830 | if (num > offslab_limit) | |
1831 | break; | |
1832 | } | |
4d268eba | 1833 | |
9888e6fa | 1834 | /* Found something acceptable - save it away */ |
4d268eba | 1835 | cachep->num = num; |
9888e6fa | 1836 | cachep->gfporder = gfporder; |
4d268eba PE |
1837 | left_over = remainder; |
1838 | ||
f78bb8ad LT |
1839 | /* |
1840 | * A VFS-reclaimable slab tends to have most allocations | |
1841 | * as GFP_NOFS and we really don't want to have to be allocating | |
1842 | * higher-order pages when we are unable to shrink dcache. | |
1843 | */ | |
1844 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1845 | break; | |
1846 | ||
4d268eba PE |
1847 | /* |
1848 | * Large number of objects is good, but very large slabs are | |
1849 | * currently bad for the gfp()s. | |
1850 | */ | |
9888e6fa | 1851 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1852 | break; |
1853 | ||
9888e6fa LT |
1854 | /* |
1855 | * Acceptable internal fragmentation? | |
1856 | */ | |
a737b3e2 | 1857 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
1858 | break; |
1859 | } | |
1860 | return left_over; | |
1861 | } | |
1862 | ||
f30cf7d1 PE |
1863 | static void setup_cpu_cache(struct kmem_cache *cachep) |
1864 | { | |
1865 | if (g_cpucache_up == FULL) { | |
1866 | enable_cpucache(cachep); | |
1867 | return; | |
1868 | } | |
1869 | if (g_cpucache_up == NONE) { | |
1870 | /* | |
1871 | * Note: the first kmem_cache_create must create the cache | |
1872 | * that's used by kmalloc(24), otherwise the creation of | |
1873 | * further caches will BUG(). | |
1874 | */ | |
1875 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
1876 | ||
1877 | /* | |
1878 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
1879 | * the first cache, then we need to set up all its list3s, | |
1880 | * otherwise the creation of further caches will BUG(). | |
1881 | */ | |
1882 | set_up_list3s(cachep, SIZE_AC); | |
1883 | if (INDEX_AC == INDEX_L3) | |
1884 | g_cpucache_up = PARTIAL_L3; | |
1885 | else | |
1886 | g_cpucache_up = PARTIAL_AC; | |
1887 | } else { | |
1888 | cachep->array[smp_processor_id()] = | |
1889 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
1890 | ||
1891 | if (g_cpucache_up == PARTIAL_AC) { | |
1892 | set_up_list3s(cachep, SIZE_L3); | |
1893 | g_cpucache_up = PARTIAL_L3; | |
1894 | } else { | |
1895 | int node; | |
1896 | for_each_online_node(node) { | |
1897 | cachep->nodelists[node] = | |
1898 | kmalloc_node(sizeof(struct kmem_list3), | |
1899 | GFP_KERNEL, node); | |
1900 | BUG_ON(!cachep->nodelists[node]); | |
1901 | kmem_list3_init(cachep->nodelists[node]); | |
1902 | } | |
1903 | } | |
1904 | } | |
1905 | cachep->nodelists[numa_node_id()]->next_reap = | |
1906 | jiffies + REAPTIMEOUT_LIST3 + | |
1907 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1908 | ||
1909 | cpu_cache_get(cachep)->avail = 0; | |
1910 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1911 | cpu_cache_get(cachep)->batchcount = 1; | |
1912 | cpu_cache_get(cachep)->touched = 0; | |
1913 | cachep->batchcount = 1; | |
1914 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
1915 | } | |
1916 | ||
1da177e4 LT |
1917 | /** |
1918 | * kmem_cache_create - Create a cache. | |
1919 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1920 | * @size: The size of objects to be created in this cache. | |
1921 | * @align: The required alignment for the objects. | |
1922 | * @flags: SLAB flags | |
1923 | * @ctor: A constructor for the objects. | |
1924 | * @dtor: A destructor for the objects. | |
1925 | * | |
1926 | * Returns a ptr to the cache on success, NULL on failure. | |
1927 | * Cannot be called within a int, but can be interrupted. | |
1928 | * The @ctor is run when new pages are allocated by the cache | |
1929 | * and the @dtor is run before the pages are handed back. | |
1930 | * | |
1931 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
1932 | * the module calling this has to destroy the cache before getting unloaded. |
1933 | * | |
1da177e4 LT |
1934 | * The flags are |
1935 | * | |
1936 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1937 | * to catch references to uninitialised memory. | |
1938 | * | |
1939 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1940 | * for buffer overruns. | |
1941 | * | |
1da177e4 LT |
1942 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
1943 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1944 | * as davem. | |
1945 | */ | |
343e0d7a | 1946 | struct kmem_cache * |
1da177e4 | 1947 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
1948 | unsigned long flags, |
1949 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 1950 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
1951 | { |
1952 | size_t left_over, slab_size, ralign; | |
7a7c381d | 1953 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
1954 | |
1955 | /* | |
1956 | * Sanity checks... these are all serious usage bugs. | |
1957 | */ | |
a737b3e2 | 1958 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
b28a02de | 1959 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { |
a737b3e2 AM |
1960 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
1961 | name); | |
b28a02de PE |
1962 | BUG(); |
1963 | } | |
1da177e4 | 1964 | |
f0188f47 RT |
1965 | /* |
1966 | * Prevent CPUs from coming and going. | |
1967 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
1968 | */ | |
1969 | lock_cpu_hotplug(); | |
1970 | ||
fc0abb14 | 1971 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 1972 | |
7a7c381d | 1973 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
1974 | mm_segment_t old_fs = get_fs(); |
1975 | char tmp; | |
1976 | int res; | |
1977 | ||
1978 | /* | |
1979 | * This happens when the module gets unloaded and doesn't | |
1980 | * destroy its slab cache and no-one else reuses the vmalloc | |
1981 | * area of the module. Print a warning. | |
1982 | */ | |
1983 | set_fs(KERNEL_DS); | |
1984 | res = __get_user(tmp, pc->name); | |
1985 | set_fs(old_fs); | |
1986 | if (res) { | |
1987 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 1988 | pc->buffer_size); |
4f12bb4f AM |
1989 | continue; |
1990 | } | |
1991 | ||
b28a02de | 1992 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1993 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1994 | dump_stack(); | |
1995 | goto oops; | |
1996 | } | |
1997 | } | |
1998 | ||
1da177e4 LT |
1999 | #if DEBUG |
2000 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
2001 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
2002 | /* No constructor, but inital state check requested */ | |
2003 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 2004 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
2005 | flags &= ~SLAB_DEBUG_INITIAL; |
2006 | } | |
1da177e4 LT |
2007 | #if FORCED_DEBUG |
2008 | /* | |
2009 | * Enable redzoning and last user accounting, except for caches with | |
2010 | * large objects, if the increased size would increase the object size | |
2011 | * above the next power of two: caches with object sizes just above a | |
2012 | * power of two have a significant amount of internal fragmentation. | |
2013 | */ | |
a737b3e2 | 2014 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 2015 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2016 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2017 | flags |= SLAB_POISON; | |
2018 | #endif | |
2019 | if (flags & SLAB_DESTROY_BY_RCU) | |
2020 | BUG_ON(flags & SLAB_POISON); | |
2021 | #endif | |
2022 | if (flags & SLAB_DESTROY_BY_RCU) | |
2023 | BUG_ON(dtor); | |
2024 | ||
2025 | /* | |
a737b3e2 AM |
2026 | * Always checks flags, a caller might be expecting debug support which |
2027 | * isn't available. | |
1da177e4 | 2028 | */ |
40094fa6 | 2029 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2030 | |
a737b3e2 AM |
2031 | /* |
2032 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2033 | * unaligned accesses for some archs when redzoning is used, and makes |
2034 | * sure any on-slab bufctl's are also correctly aligned. | |
2035 | */ | |
b28a02de PE |
2036 | if (size & (BYTES_PER_WORD - 1)) { |
2037 | size += (BYTES_PER_WORD - 1); | |
2038 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2039 | } |
2040 | ||
a737b3e2 AM |
2041 | /* calculate the final buffer alignment: */ |
2042 | ||
1da177e4 LT |
2043 | /* 1) arch recommendation: can be overridden for debug */ |
2044 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2045 | /* |
2046 | * Default alignment: as specified by the arch code. Except if | |
2047 | * an object is really small, then squeeze multiple objects into | |
2048 | * one cacheline. | |
1da177e4 LT |
2049 | */ |
2050 | ralign = cache_line_size(); | |
b28a02de | 2051 | while (size <= ralign / 2) |
1da177e4 LT |
2052 | ralign /= 2; |
2053 | } else { | |
2054 | ralign = BYTES_PER_WORD; | |
2055 | } | |
2056 | /* 2) arch mandated alignment: disables debug if necessary */ | |
2057 | if (ralign < ARCH_SLAB_MINALIGN) { | |
2058 | ralign = ARCH_SLAB_MINALIGN; | |
2059 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2060 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
2061 | } |
2062 | /* 3) caller mandated alignment: disables debug if necessary */ | |
2063 | if (ralign < align) { | |
2064 | ralign = align; | |
2065 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2066 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 | 2067 | } |
a737b3e2 AM |
2068 | /* |
2069 | * 4) Store it. Note that the debug code below can reduce | |
1da177e4 LT |
2070 | * the alignment to BYTES_PER_WORD. |
2071 | */ | |
2072 | align = ralign; | |
2073 | ||
2074 | /* Get cache's description obj. */ | |
c5e3b83e | 2075 | cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 2076 | if (!cachep) |
4f12bb4f | 2077 | goto oops; |
1da177e4 LT |
2078 | |
2079 | #if DEBUG | |
3dafccf2 | 2080 | cachep->obj_size = size; |
1da177e4 LT |
2081 | |
2082 | if (flags & SLAB_RED_ZONE) { | |
2083 | /* redzoning only works with word aligned caches */ | |
2084 | align = BYTES_PER_WORD; | |
2085 | ||
2086 | /* add space for red zone words */ | |
3dafccf2 | 2087 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 2088 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
2089 | } |
2090 | if (flags & SLAB_STORE_USER) { | |
2091 | /* user store requires word alignment and | |
2092 | * one word storage behind the end of the real | |
2093 | * object. | |
2094 | */ | |
2095 | align = BYTES_PER_WORD; | |
2096 | size += BYTES_PER_WORD; | |
2097 | } | |
2098 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2099 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2100 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2101 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2102 | size = PAGE_SIZE; |
2103 | } | |
2104 | #endif | |
2105 | #endif | |
2106 | ||
2107 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 2108 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
2109 | /* |
2110 | * Size is large, assume best to place the slab management obj | |
2111 | * off-slab (should allow better packing of objs). | |
2112 | */ | |
2113 | flags |= CFLGS_OFF_SLAB; | |
2114 | ||
2115 | size = ALIGN(size, align); | |
2116 | ||
f78bb8ad | 2117 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2118 | |
2119 | if (!cachep->num) { | |
2120 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
2121 | kmem_cache_free(&cache_cache, cachep); | |
2122 | cachep = NULL; | |
4f12bb4f | 2123 | goto oops; |
1da177e4 | 2124 | } |
b28a02de PE |
2125 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2126 | + sizeof(struct slab), align); | |
1da177e4 LT |
2127 | |
2128 | /* | |
2129 | * If the slab has been placed off-slab, and we have enough space then | |
2130 | * move it on-slab. This is at the expense of any extra colouring. | |
2131 | */ | |
2132 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2133 | flags &= ~CFLGS_OFF_SLAB; | |
2134 | left_over -= slab_size; | |
2135 | } | |
2136 | ||
2137 | if (flags & CFLGS_OFF_SLAB) { | |
2138 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2139 | slab_size = |
2140 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2141 | } |
2142 | ||
2143 | cachep->colour_off = cache_line_size(); | |
2144 | /* Offset must be a multiple of the alignment. */ | |
2145 | if (cachep->colour_off < align) | |
2146 | cachep->colour_off = align; | |
b28a02de | 2147 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2148 | cachep->slab_size = slab_size; |
2149 | cachep->flags = flags; | |
2150 | cachep->gfpflags = 0; | |
2151 | if (flags & SLAB_CACHE_DMA) | |
2152 | cachep->gfpflags |= GFP_DMA; | |
3dafccf2 | 2153 | cachep->buffer_size = size; |
1da177e4 LT |
2154 | |
2155 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 2156 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
2157 | cachep->ctor = ctor; |
2158 | cachep->dtor = dtor; | |
2159 | cachep->name = name; | |
2160 | ||
1da177e4 | 2161 | |
f30cf7d1 | 2162 | setup_cpu_cache(cachep); |
1da177e4 | 2163 | |
1da177e4 LT |
2164 | /* cache setup completed, link it into the list */ |
2165 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2166 | oops: |
1da177e4 LT |
2167 | if (!cachep && (flags & SLAB_PANIC)) |
2168 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2169 | name); |
fc0abb14 | 2170 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 2171 | unlock_cpu_hotplug(); |
1da177e4 LT |
2172 | return cachep; |
2173 | } | |
2174 | EXPORT_SYMBOL(kmem_cache_create); | |
2175 | ||
2176 | #if DEBUG | |
2177 | static void check_irq_off(void) | |
2178 | { | |
2179 | BUG_ON(!irqs_disabled()); | |
2180 | } | |
2181 | ||
2182 | static void check_irq_on(void) | |
2183 | { | |
2184 | BUG_ON(irqs_disabled()); | |
2185 | } | |
2186 | ||
343e0d7a | 2187 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2188 | { |
2189 | #ifdef CONFIG_SMP | |
2190 | check_irq_off(); | |
e498be7d | 2191 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2192 | #endif |
2193 | } | |
e498be7d | 2194 | |
343e0d7a | 2195 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2196 | { |
2197 | #ifdef CONFIG_SMP | |
2198 | check_irq_off(); | |
2199 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2200 | #endif | |
2201 | } | |
2202 | ||
1da177e4 LT |
2203 | #else |
2204 | #define check_irq_off() do { } while(0) | |
2205 | #define check_irq_on() do { } while(0) | |
2206 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2207 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2208 | #endif |
2209 | ||
aab2207c CL |
2210 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2211 | struct array_cache *ac, | |
2212 | int force, int node); | |
2213 | ||
1da177e4 LT |
2214 | static void do_drain(void *arg) |
2215 | { | |
a737b3e2 | 2216 | struct kmem_cache *cachep = arg; |
1da177e4 | 2217 | struct array_cache *ac; |
ff69416e | 2218 | int node = numa_node_id(); |
1da177e4 LT |
2219 | |
2220 | check_irq_off(); | |
9a2dba4b | 2221 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2222 | spin_lock(&cachep->nodelists[node]->list_lock); |
2223 | free_block(cachep, ac->entry, ac->avail, node); | |
2224 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2225 | ac->avail = 0; |
2226 | } | |
2227 | ||
343e0d7a | 2228 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2229 | { |
e498be7d CL |
2230 | struct kmem_list3 *l3; |
2231 | int node; | |
2232 | ||
a07fa394 | 2233 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2234 | check_irq_on(); |
b28a02de | 2235 | for_each_online_node(node) { |
e498be7d | 2236 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2237 | if (l3 && l3->alien) |
2238 | drain_alien_cache(cachep, l3->alien); | |
2239 | } | |
2240 | ||
2241 | for_each_online_node(node) { | |
2242 | l3 = cachep->nodelists[node]; | |
2243 | if (l3) | |
aab2207c | 2244 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2245 | } |
1da177e4 LT |
2246 | } |
2247 | ||
343e0d7a | 2248 | static int __node_shrink(struct kmem_cache *cachep, int node) |
1da177e4 LT |
2249 | { |
2250 | struct slab *slabp; | |
e498be7d | 2251 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
2252 | int ret; |
2253 | ||
e498be7d | 2254 | for (;;) { |
1da177e4 LT |
2255 | struct list_head *p; |
2256 | ||
e498be7d CL |
2257 | p = l3->slabs_free.prev; |
2258 | if (p == &l3->slabs_free) | |
1da177e4 LT |
2259 | break; |
2260 | ||
e498be7d | 2261 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 | 2262 | #if DEBUG |
40094fa6 | 2263 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2264 | #endif |
2265 | list_del(&slabp->list); | |
2266 | ||
e498be7d CL |
2267 | l3->free_objects -= cachep->num; |
2268 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 2269 | slab_destroy(cachep, slabp); |
e498be7d | 2270 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 2271 | } |
b28a02de | 2272 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
2273 | return ret; |
2274 | } | |
2275 | ||
343e0d7a | 2276 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2277 | { |
2278 | int ret = 0, i = 0; | |
2279 | struct kmem_list3 *l3; | |
2280 | ||
2281 | drain_cpu_caches(cachep); | |
2282 | ||
2283 | check_irq_on(); | |
2284 | for_each_online_node(i) { | |
2285 | l3 = cachep->nodelists[i]; | |
2286 | if (l3) { | |
2287 | spin_lock_irq(&l3->list_lock); | |
2288 | ret += __node_shrink(cachep, i); | |
2289 | spin_unlock_irq(&l3->list_lock); | |
2290 | } | |
2291 | } | |
2292 | return (ret ? 1 : 0); | |
2293 | } | |
2294 | ||
1da177e4 LT |
2295 | /** |
2296 | * kmem_cache_shrink - Shrink a cache. | |
2297 | * @cachep: The cache to shrink. | |
2298 | * | |
2299 | * Releases as many slabs as possible for a cache. | |
2300 | * To help debugging, a zero exit status indicates all slabs were released. | |
2301 | */ | |
343e0d7a | 2302 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2303 | { |
40094fa6 | 2304 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2305 | |
2306 | return __cache_shrink(cachep); | |
2307 | } | |
2308 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2309 | ||
2310 | /** | |
2311 | * kmem_cache_destroy - delete a cache | |
2312 | * @cachep: the cache to destroy | |
2313 | * | |
343e0d7a | 2314 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2315 | * Returns 0 on success. |
2316 | * | |
2317 | * It is expected this function will be called by a module when it is | |
2318 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2319 | * cache being allocated each time a module is loaded and unloaded, if the | |
2320 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2321 | * | |
2322 | * The cache must be empty before calling this function. | |
2323 | * | |
2324 | * The caller must guarantee that noone will allocate memory from the cache | |
2325 | * during the kmem_cache_destroy(). | |
2326 | */ | |
343e0d7a | 2327 | int kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 LT |
2328 | { |
2329 | int i; | |
e498be7d | 2330 | struct kmem_list3 *l3; |
1da177e4 | 2331 | |
40094fa6 | 2332 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2333 | |
2334 | /* Don't let CPUs to come and go */ | |
2335 | lock_cpu_hotplug(); | |
2336 | ||
2337 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2338 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2339 | /* |
2340 | * the chain is never empty, cache_cache is never destroyed | |
2341 | */ | |
2342 | list_del(&cachep->next); | |
fc0abb14 | 2343 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2344 | |
2345 | if (__cache_shrink(cachep)) { | |
2346 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2347 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2348 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2349 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2350 | unlock_cpu_hotplug(); |
2351 | return 1; | |
2352 | } | |
2353 | ||
2354 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2355 | synchronize_rcu(); |
1da177e4 | 2356 | |
e498be7d | 2357 | for_each_online_cpu(i) |
b28a02de | 2358 | kfree(cachep->array[i]); |
1da177e4 LT |
2359 | |
2360 | /* NUMA: free the list3 structures */ | |
e498be7d | 2361 | for_each_online_node(i) { |
a737b3e2 AM |
2362 | l3 = cachep->nodelists[i]; |
2363 | if (l3) { | |
e498be7d CL |
2364 | kfree(l3->shared); |
2365 | free_alien_cache(l3->alien); | |
2366 | kfree(l3); | |
2367 | } | |
2368 | } | |
1da177e4 | 2369 | kmem_cache_free(&cache_cache, cachep); |
1da177e4 | 2370 | unlock_cpu_hotplug(); |
1da177e4 LT |
2371 | return 0; |
2372 | } | |
2373 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2374 | ||
2375 | /* Get the memory for a slab management obj. */ | |
343e0d7a | 2376 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2377 | int colour_off, gfp_t local_flags, |
2378 | int nodeid) | |
1da177e4 LT |
2379 | { |
2380 | struct slab *slabp; | |
b28a02de | 2381 | |
1da177e4 LT |
2382 | if (OFF_SLAB(cachep)) { |
2383 | /* Slab management obj is off-slab. */ | |
5b74ada7 RT |
2384 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
2385 | local_flags, nodeid); | |
1da177e4 LT |
2386 | if (!slabp) |
2387 | return NULL; | |
2388 | } else { | |
b28a02de | 2389 | slabp = objp + colour_off; |
1da177e4 LT |
2390 | colour_off += cachep->slab_size; |
2391 | } | |
2392 | slabp->inuse = 0; | |
2393 | slabp->colouroff = colour_off; | |
b28a02de | 2394 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2395 | slabp->nodeid = nodeid; |
1da177e4 LT |
2396 | return slabp; |
2397 | } | |
2398 | ||
2399 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2400 | { | |
b28a02de | 2401 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2402 | } |
2403 | ||
343e0d7a | 2404 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2405 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2406 | { |
2407 | int i; | |
2408 | ||
2409 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2410 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2411 | #if DEBUG |
2412 | /* need to poison the objs? */ | |
2413 | if (cachep->flags & SLAB_POISON) | |
2414 | poison_obj(cachep, objp, POISON_FREE); | |
2415 | if (cachep->flags & SLAB_STORE_USER) | |
2416 | *dbg_userword(cachep, objp) = NULL; | |
2417 | ||
2418 | if (cachep->flags & SLAB_RED_ZONE) { | |
2419 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2420 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2421 | } | |
2422 | /* | |
a737b3e2 AM |
2423 | * Constructors are not allowed to allocate memory from the same |
2424 | * cache which they are a constructor for. Otherwise, deadlock. | |
2425 | * They must also be threaded. | |
1da177e4 LT |
2426 | */ |
2427 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2428 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2429 | ctor_flags); |
1da177e4 LT |
2430 | |
2431 | if (cachep->flags & SLAB_RED_ZONE) { | |
2432 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2433 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2434 | " end of an object"); |
1da177e4 LT |
2435 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2436 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2437 | " start of an object"); |
1da177e4 | 2438 | } |
a737b3e2 AM |
2439 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2440 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2441 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2442 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2443 | #else |
2444 | if (cachep->ctor) | |
2445 | cachep->ctor(objp, cachep, ctor_flags); | |
2446 | #endif | |
b28a02de | 2447 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2448 | } |
b28a02de | 2449 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2450 | slabp->free = 0; |
2451 | } | |
2452 | ||
343e0d7a | 2453 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2454 | { |
a737b3e2 AM |
2455 | if (flags & SLAB_DMA) |
2456 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2457 | else | |
2458 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
1da177e4 LT |
2459 | } |
2460 | ||
a737b3e2 AM |
2461 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2462 | int nodeid) | |
78d382d7 | 2463 | { |
8fea4e96 | 2464 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2465 | kmem_bufctl_t next; |
2466 | ||
2467 | slabp->inuse++; | |
2468 | next = slab_bufctl(slabp)[slabp->free]; | |
2469 | #if DEBUG | |
2470 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2471 | WARN_ON(slabp->nodeid != nodeid); | |
2472 | #endif | |
2473 | slabp->free = next; | |
2474 | ||
2475 | return objp; | |
2476 | } | |
2477 | ||
a737b3e2 AM |
2478 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2479 | void *objp, int nodeid) | |
78d382d7 | 2480 | { |
8fea4e96 | 2481 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2482 | |
2483 | #if DEBUG | |
2484 | /* Verify that the slab belongs to the intended node */ | |
2485 | WARN_ON(slabp->nodeid != nodeid); | |
2486 | ||
871751e2 | 2487 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2488 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2489 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2490 | BUG(); |
2491 | } | |
2492 | #endif | |
2493 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2494 | slabp->free = objnr; | |
2495 | slabp->inuse--; | |
2496 | } | |
2497 | ||
4776874f PE |
2498 | /* |
2499 | * Map pages beginning at addr to the given cache and slab. This is required | |
2500 | * for the slab allocator to be able to lookup the cache and slab of a | |
2501 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2502 | */ | |
2503 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2504 | void *addr) | |
1da177e4 | 2505 | { |
4776874f | 2506 | int nr_pages; |
1da177e4 LT |
2507 | struct page *page; |
2508 | ||
4776874f | 2509 | page = virt_to_page(addr); |
84097518 | 2510 | |
4776874f | 2511 | nr_pages = 1; |
84097518 | 2512 | if (likely(!PageCompound(page))) |
4776874f PE |
2513 | nr_pages <<= cache->gfporder; |
2514 | ||
1da177e4 | 2515 | do { |
4776874f PE |
2516 | page_set_cache(page, cache); |
2517 | page_set_slab(page, slab); | |
1da177e4 | 2518 | page++; |
4776874f | 2519 | } while (--nr_pages); |
1da177e4 LT |
2520 | } |
2521 | ||
2522 | /* | |
2523 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2524 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2525 | */ | |
343e0d7a | 2526 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2527 | { |
b28a02de PE |
2528 | struct slab *slabp; |
2529 | void *objp; | |
2530 | size_t offset; | |
2531 | gfp_t local_flags; | |
2532 | unsigned long ctor_flags; | |
e498be7d | 2533 | struct kmem_list3 *l3; |
1da177e4 | 2534 | |
a737b3e2 AM |
2535 | /* |
2536 | * Be lazy and only check for valid flags here, keeping it out of the | |
2537 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2538 | */ |
40094fa6 | 2539 | BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); |
1da177e4 LT |
2540 | if (flags & SLAB_NO_GROW) |
2541 | return 0; | |
2542 | ||
2543 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2544 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2545 | if (!(local_flags & __GFP_WAIT)) | |
2546 | /* | |
2547 | * Not allowed to sleep. Need to tell a constructor about | |
2548 | * this - it might need to know... | |
2549 | */ | |
2550 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2551 | ||
2e1217cf | 2552 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2553 | check_irq_off(); |
2e1217cf RT |
2554 | l3 = cachep->nodelists[nodeid]; |
2555 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2556 | |
2557 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2558 | offset = l3->colour_next; |
2559 | l3->colour_next++; | |
2560 | if (l3->colour_next >= cachep->colour) | |
2561 | l3->colour_next = 0; | |
2562 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2563 | |
2e1217cf | 2564 | offset *= cachep->colour_off; |
1da177e4 LT |
2565 | |
2566 | if (local_flags & __GFP_WAIT) | |
2567 | local_irq_enable(); | |
2568 | ||
2569 | /* | |
2570 | * The test for missing atomic flag is performed here, rather than | |
2571 | * the more obvious place, simply to reduce the critical path length | |
2572 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2573 | * will eventually be caught here (where it matters). | |
2574 | */ | |
2575 | kmem_flagcheck(cachep, flags); | |
2576 | ||
a737b3e2 AM |
2577 | /* |
2578 | * Get mem for the objs. Attempt to allocate a physical page from | |
2579 | * 'nodeid'. | |
e498be7d | 2580 | */ |
a737b3e2 AM |
2581 | objp = kmem_getpages(cachep, flags, nodeid); |
2582 | if (!objp) | |
1da177e4 LT |
2583 | goto failed; |
2584 | ||
2585 | /* Get slab management. */ | |
5b74ada7 | 2586 | slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); |
a737b3e2 | 2587 | if (!slabp) |
1da177e4 LT |
2588 | goto opps1; |
2589 | ||
e498be7d | 2590 | slabp->nodeid = nodeid; |
4776874f | 2591 | slab_map_pages(cachep, slabp, objp); |
1da177e4 LT |
2592 | |
2593 | cache_init_objs(cachep, slabp, ctor_flags); | |
2594 | ||
2595 | if (local_flags & __GFP_WAIT) | |
2596 | local_irq_disable(); | |
2597 | check_irq_off(); | |
e498be7d | 2598 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2599 | |
2600 | /* Make slab active. */ | |
e498be7d | 2601 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2602 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2603 | l3->free_objects += cachep->num; |
2604 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2605 | return 1; |
a737b3e2 | 2606 | opps1: |
1da177e4 | 2607 | kmem_freepages(cachep, objp); |
a737b3e2 | 2608 | failed: |
1da177e4 LT |
2609 | if (local_flags & __GFP_WAIT) |
2610 | local_irq_disable(); | |
2611 | return 0; | |
2612 | } | |
2613 | ||
2614 | #if DEBUG | |
2615 | ||
2616 | /* | |
2617 | * Perform extra freeing checks: | |
2618 | * - detect bad pointers. | |
2619 | * - POISON/RED_ZONE checking | |
2620 | * - destructor calls, for caches with POISON+dtor | |
2621 | */ | |
2622 | static void kfree_debugcheck(const void *objp) | |
2623 | { | |
2624 | struct page *page; | |
2625 | ||
2626 | if (!virt_addr_valid(objp)) { | |
2627 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2628 | (unsigned long)objp); |
2629 | BUG(); | |
1da177e4 LT |
2630 | } |
2631 | page = virt_to_page(objp); | |
2632 | if (!PageSlab(page)) { | |
b28a02de PE |
2633 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2634 | (unsigned long)objp); | |
1da177e4 LT |
2635 | BUG(); |
2636 | } | |
2637 | } | |
2638 | ||
343e0d7a | 2639 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2640 | void *caller) |
1da177e4 LT |
2641 | { |
2642 | struct page *page; | |
2643 | unsigned int objnr; | |
2644 | struct slab *slabp; | |
2645 | ||
3dafccf2 | 2646 | objp -= obj_offset(cachep); |
1da177e4 LT |
2647 | kfree_debugcheck(objp); |
2648 | page = virt_to_page(objp); | |
2649 | ||
065d41cb | 2650 | if (page_get_cache(page) != cachep) { |
a737b3e2 AM |
2651 | printk(KERN_ERR "mismatch in kmem_cache_free: expected " |
2652 | "cache %p, got %p\n", | |
b28a02de | 2653 | page_get_cache(page), cachep); |
1da177e4 | 2654 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2655 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2656 | page_get_cache(page)->name); | |
1da177e4 LT |
2657 | WARN_ON(1); |
2658 | } | |
065d41cb | 2659 | slabp = page_get_slab(page); |
1da177e4 LT |
2660 | |
2661 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2662 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || |
2663 | *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2664 | slab_error(cachep, "double free, or memory outside" | |
2665 | " object was overwritten"); | |
2666 | printk(KERN_ERR "%p: redzone 1:0x%lx, " | |
2667 | "redzone 2:0x%lx.\n", | |
b28a02de PE |
2668 | objp, *dbg_redzone1(cachep, objp), |
2669 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2670 | } |
2671 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2672 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2673 | } | |
2674 | if (cachep->flags & SLAB_STORE_USER) | |
2675 | *dbg_userword(cachep, objp) = caller; | |
2676 | ||
8fea4e96 | 2677 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2678 | |
2679 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2680 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2681 | |
2682 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
a737b3e2 AM |
2683 | /* |
2684 | * Need to call the slab's constructor so the caller can | |
2685 | * perform a verify of its state (debugging). Called without | |
2686 | * the cache-lock held. | |
1da177e4 | 2687 | */ |
3dafccf2 | 2688 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2689 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2690 | } |
2691 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2692 | /* we want to cache poison the object, | |
2693 | * call the destruction callback | |
2694 | */ | |
3dafccf2 | 2695 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 2696 | } |
871751e2 AV |
2697 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2698 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2699 | #endif | |
1da177e4 LT |
2700 | if (cachep->flags & SLAB_POISON) { |
2701 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2702 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2703 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2704 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2705 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2706 | } else { |
2707 | poison_obj(cachep, objp, POISON_FREE); | |
2708 | } | |
2709 | #else | |
2710 | poison_obj(cachep, objp, POISON_FREE); | |
2711 | #endif | |
2712 | } | |
2713 | return objp; | |
2714 | } | |
2715 | ||
343e0d7a | 2716 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2717 | { |
2718 | kmem_bufctl_t i; | |
2719 | int entries = 0; | |
b28a02de | 2720 | |
1da177e4 LT |
2721 | /* Check slab's freelist to see if this obj is there. */ |
2722 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2723 | entries++; | |
2724 | if (entries > cachep->num || i >= cachep->num) | |
2725 | goto bad; | |
2726 | } | |
2727 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2728 | bad: |
2729 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2730 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2731 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2732 | for (i = 0; |
264132bc | 2733 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2734 | i++) { |
a737b3e2 | 2735 | if (i % 16 == 0) |
1da177e4 | 2736 | printk("\n%03x:", i); |
b28a02de | 2737 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2738 | } |
2739 | printk("\n"); | |
2740 | BUG(); | |
2741 | } | |
2742 | } | |
2743 | #else | |
2744 | #define kfree_debugcheck(x) do { } while(0) | |
2745 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2746 | #define check_slabp(x,y) do { } while(0) | |
2747 | #endif | |
2748 | ||
343e0d7a | 2749 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2750 | { |
2751 | int batchcount; | |
2752 | struct kmem_list3 *l3; | |
2753 | struct array_cache *ac; | |
2754 | ||
2755 | check_irq_off(); | |
9a2dba4b | 2756 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2757 | retry: |
1da177e4 LT |
2758 | batchcount = ac->batchcount; |
2759 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2760 | /* |
2761 | * If there was little recent activity on this cache, then | |
2762 | * perform only a partial refill. Otherwise we could generate | |
2763 | * refill bouncing. | |
1da177e4 LT |
2764 | */ |
2765 | batchcount = BATCHREFILL_LIMIT; | |
2766 | } | |
e498be7d CL |
2767 | l3 = cachep->nodelists[numa_node_id()]; |
2768 | ||
2769 | BUG_ON(ac->avail > 0 || !l3); | |
2770 | spin_lock(&l3->list_lock); | |
1da177e4 | 2771 | |
3ded175a CL |
2772 | /* See if we can refill from the shared array */ |
2773 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2774 | goto alloc_done; | |
2775 | ||
1da177e4 LT |
2776 | while (batchcount > 0) { |
2777 | struct list_head *entry; | |
2778 | struct slab *slabp; | |
2779 | /* Get slab alloc is to come from. */ | |
2780 | entry = l3->slabs_partial.next; | |
2781 | if (entry == &l3->slabs_partial) { | |
2782 | l3->free_touched = 1; | |
2783 | entry = l3->slabs_free.next; | |
2784 | if (entry == &l3->slabs_free) | |
2785 | goto must_grow; | |
2786 | } | |
2787 | ||
2788 | slabp = list_entry(entry, struct slab, list); | |
2789 | check_slabp(cachep, slabp); | |
2790 | check_spinlock_acquired(cachep); | |
2791 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2792 | STATS_INC_ALLOCED(cachep); |
2793 | STATS_INC_ACTIVE(cachep); | |
2794 | STATS_SET_HIGH(cachep); | |
2795 | ||
78d382d7 MD |
2796 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2797 | numa_node_id()); | |
1da177e4 LT |
2798 | } |
2799 | check_slabp(cachep, slabp); | |
2800 | ||
2801 | /* move slabp to correct slabp list: */ | |
2802 | list_del(&slabp->list); | |
2803 | if (slabp->free == BUFCTL_END) | |
2804 | list_add(&slabp->list, &l3->slabs_full); | |
2805 | else | |
2806 | list_add(&slabp->list, &l3->slabs_partial); | |
2807 | } | |
2808 | ||
a737b3e2 | 2809 | must_grow: |
1da177e4 | 2810 | l3->free_objects -= ac->avail; |
a737b3e2 | 2811 | alloc_done: |
e498be7d | 2812 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2813 | |
2814 | if (unlikely(!ac->avail)) { | |
2815 | int x; | |
e498be7d CL |
2816 | x = cache_grow(cachep, flags, numa_node_id()); |
2817 | ||
a737b3e2 | 2818 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 2819 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2820 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
2821 | return NULL; |
2822 | ||
a737b3e2 | 2823 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
2824 | goto retry; |
2825 | } | |
2826 | ac->touched = 1; | |
e498be7d | 2827 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2828 | } |
2829 | ||
a737b3e2 AM |
2830 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
2831 | gfp_t flags) | |
1da177e4 LT |
2832 | { |
2833 | might_sleep_if(flags & __GFP_WAIT); | |
2834 | #if DEBUG | |
2835 | kmem_flagcheck(cachep, flags); | |
2836 | #endif | |
2837 | } | |
2838 | ||
2839 | #if DEBUG | |
a737b3e2 AM |
2840 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
2841 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 2842 | { |
b28a02de | 2843 | if (!objp) |
1da177e4 | 2844 | return objp; |
b28a02de | 2845 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2846 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2847 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2848 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2849 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2850 | else |
2851 | check_poison_obj(cachep, objp); | |
2852 | #else | |
2853 | check_poison_obj(cachep, objp); | |
2854 | #endif | |
2855 | poison_obj(cachep, objp, POISON_INUSE); | |
2856 | } | |
2857 | if (cachep->flags & SLAB_STORE_USER) | |
2858 | *dbg_userword(cachep, objp) = caller; | |
2859 | ||
2860 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2861 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
2862 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2863 | slab_error(cachep, "double free, or memory outside" | |
2864 | " object was overwritten"); | |
b28a02de | 2865 | printk(KERN_ERR |
a737b3e2 AM |
2866 | "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", |
2867 | objp, *dbg_redzone1(cachep, objp), | |
2868 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2869 | } |
2870 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2871 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2872 | } | |
871751e2 AV |
2873 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2874 | { | |
2875 | struct slab *slabp; | |
2876 | unsigned objnr; | |
2877 | ||
2878 | slabp = page_get_slab(virt_to_page(objp)); | |
2879 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; | |
2880 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
2881 | } | |
2882 | #endif | |
3dafccf2 | 2883 | objp += obj_offset(cachep); |
1da177e4 | 2884 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2885 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2886 | |
2887 | if (!(flags & __GFP_WAIT)) | |
2888 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2889 | ||
2890 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2891 | } |
1da177e4 LT |
2892 | return objp; |
2893 | } | |
2894 | #else | |
2895 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2896 | #endif | |
2897 | ||
343e0d7a | 2898 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2899 | { |
b28a02de | 2900 | void *objp; |
1da177e4 LT |
2901 | struct array_cache *ac; |
2902 | ||
dc85da15 | 2903 | #ifdef CONFIG_NUMA |
b2455396 | 2904 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { |
c61afb18 PJ |
2905 | objp = alternate_node_alloc(cachep, flags); |
2906 | if (objp != NULL) | |
2907 | return objp; | |
dc85da15 CL |
2908 | } |
2909 | #endif | |
2910 | ||
5c382300 | 2911 | check_irq_off(); |
9a2dba4b | 2912 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2913 | if (likely(ac->avail)) { |
2914 | STATS_INC_ALLOCHIT(cachep); | |
2915 | ac->touched = 1; | |
e498be7d | 2916 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2917 | } else { |
2918 | STATS_INC_ALLOCMISS(cachep); | |
2919 | objp = cache_alloc_refill(cachep, flags); | |
2920 | } | |
5c382300 AK |
2921 | return objp; |
2922 | } | |
2923 | ||
a737b3e2 AM |
2924 | static __always_inline void *__cache_alloc(struct kmem_cache *cachep, |
2925 | gfp_t flags, void *caller) | |
5c382300 AK |
2926 | { |
2927 | unsigned long save_flags; | |
b28a02de | 2928 | void *objp; |
5c382300 AK |
2929 | |
2930 | cache_alloc_debugcheck_before(cachep, flags); | |
2931 | ||
2932 | local_irq_save(save_flags); | |
2933 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2934 | local_irq_restore(save_flags); |
34342e86 | 2935 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 2936 | caller); |
34342e86 | 2937 | prefetchw(objp); |
1da177e4 LT |
2938 | return objp; |
2939 | } | |
2940 | ||
e498be7d | 2941 | #ifdef CONFIG_NUMA |
c61afb18 | 2942 | /* |
b2455396 | 2943 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
2944 | * |
2945 | * If we are in_interrupt, then process context, including cpusets and | |
2946 | * mempolicy, may not apply and should not be used for allocation policy. | |
2947 | */ | |
2948 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
2949 | { | |
2950 | int nid_alloc, nid_here; | |
2951 | ||
2952 | if (in_interrupt()) | |
2953 | return NULL; | |
2954 | nid_alloc = nid_here = numa_node_id(); | |
2955 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
2956 | nid_alloc = cpuset_mem_spread_node(); | |
2957 | else if (current->mempolicy) | |
2958 | nid_alloc = slab_node(current->mempolicy); | |
2959 | if (nid_alloc != nid_here) | |
2960 | return __cache_alloc_node(cachep, flags, nid_alloc); | |
2961 | return NULL; | |
2962 | } | |
2963 | ||
e498be7d CL |
2964 | /* |
2965 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2966 | */ |
a737b3e2 AM |
2967 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
2968 | int nodeid) | |
e498be7d CL |
2969 | { |
2970 | struct list_head *entry; | |
b28a02de PE |
2971 | struct slab *slabp; |
2972 | struct kmem_list3 *l3; | |
2973 | void *obj; | |
b28a02de PE |
2974 | int x; |
2975 | ||
2976 | l3 = cachep->nodelists[nodeid]; | |
2977 | BUG_ON(!l3); | |
2978 | ||
a737b3e2 | 2979 | retry: |
ca3b9b91 | 2980 | check_irq_off(); |
b28a02de PE |
2981 | spin_lock(&l3->list_lock); |
2982 | entry = l3->slabs_partial.next; | |
2983 | if (entry == &l3->slabs_partial) { | |
2984 | l3->free_touched = 1; | |
2985 | entry = l3->slabs_free.next; | |
2986 | if (entry == &l3->slabs_free) | |
2987 | goto must_grow; | |
2988 | } | |
2989 | ||
2990 | slabp = list_entry(entry, struct slab, list); | |
2991 | check_spinlock_acquired_node(cachep, nodeid); | |
2992 | check_slabp(cachep, slabp); | |
2993 | ||
2994 | STATS_INC_NODEALLOCS(cachep); | |
2995 | STATS_INC_ACTIVE(cachep); | |
2996 | STATS_SET_HIGH(cachep); | |
2997 | ||
2998 | BUG_ON(slabp->inuse == cachep->num); | |
2999 | ||
78d382d7 | 3000 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3001 | check_slabp(cachep, slabp); |
3002 | l3->free_objects--; | |
3003 | /* move slabp to correct slabp list: */ | |
3004 | list_del(&slabp->list); | |
3005 | ||
a737b3e2 | 3006 | if (slabp->free == BUFCTL_END) |
b28a02de | 3007 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3008 | else |
b28a02de | 3009 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3010 | |
b28a02de PE |
3011 | spin_unlock(&l3->list_lock); |
3012 | goto done; | |
e498be7d | 3013 | |
a737b3e2 | 3014 | must_grow: |
b28a02de PE |
3015 | spin_unlock(&l3->list_lock); |
3016 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 3017 | |
b28a02de PE |
3018 | if (!x) |
3019 | return NULL; | |
e498be7d | 3020 | |
b28a02de | 3021 | goto retry; |
a737b3e2 | 3022 | done: |
b28a02de | 3023 | return obj; |
e498be7d CL |
3024 | } |
3025 | #endif | |
3026 | ||
3027 | /* | |
3028 | * Caller needs to acquire correct kmem_list's list_lock | |
3029 | */ | |
343e0d7a | 3030 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3031 | int node) |
1da177e4 LT |
3032 | { |
3033 | int i; | |
e498be7d | 3034 | struct kmem_list3 *l3; |
1da177e4 LT |
3035 | |
3036 | for (i = 0; i < nr_objects; i++) { | |
3037 | void *objp = objpp[i]; | |
3038 | struct slab *slabp; | |
1da177e4 | 3039 | |
6ed5eb22 | 3040 | slabp = virt_to_slab(objp); |
ff69416e | 3041 | l3 = cachep->nodelists[node]; |
1da177e4 | 3042 | list_del(&slabp->list); |
ff69416e | 3043 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3044 | check_slabp(cachep, slabp); |
78d382d7 | 3045 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3046 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3047 | l3->free_objects++; |
1da177e4 LT |
3048 | check_slabp(cachep, slabp); |
3049 | ||
3050 | /* fixup slab chains */ | |
3051 | if (slabp->inuse == 0) { | |
e498be7d CL |
3052 | if (l3->free_objects > l3->free_limit) { |
3053 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
3054 | slab_destroy(cachep, slabp); |
3055 | } else { | |
e498be7d | 3056 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3057 | } |
3058 | } else { | |
3059 | /* Unconditionally move a slab to the end of the | |
3060 | * partial list on free - maximum time for the | |
3061 | * other objects to be freed, too. | |
3062 | */ | |
e498be7d | 3063 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3064 | } |
3065 | } | |
3066 | } | |
3067 | ||
343e0d7a | 3068 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3069 | { |
3070 | int batchcount; | |
e498be7d | 3071 | struct kmem_list3 *l3; |
ff69416e | 3072 | int node = numa_node_id(); |
1da177e4 LT |
3073 | |
3074 | batchcount = ac->batchcount; | |
3075 | #if DEBUG | |
3076 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3077 | #endif | |
3078 | check_irq_off(); | |
ff69416e | 3079 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3080 | spin_lock(&l3->list_lock); |
3081 | if (l3->shared) { | |
3082 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3083 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3084 | if (max) { |
3085 | if (batchcount > max) | |
3086 | batchcount = max; | |
e498be7d | 3087 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3088 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3089 | shared_array->avail += batchcount; |
3090 | goto free_done; | |
3091 | } | |
3092 | } | |
3093 | ||
ff69416e | 3094 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3095 | free_done: |
1da177e4 LT |
3096 | #if STATS |
3097 | { | |
3098 | int i = 0; | |
3099 | struct list_head *p; | |
3100 | ||
e498be7d CL |
3101 | p = l3->slabs_free.next; |
3102 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3103 | struct slab *slabp; |
3104 | ||
3105 | slabp = list_entry(p, struct slab, list); | |
3106 | BUG_ON(slabp->inuse); | |
3107 | ||
3108 | i++; | |
3109 | p = p->next; | |
3110 | } | |
3111 | STATS_SET_FREEABLE(cachep, i); | |
3112 | } | |
3113 | #endif | |
e498be7d | 3114 | spin_unlock(&l3->list_lock); |
1da177e4 | 3115 | ac->avail -= batchcount; |
a737b3e2 | 3116 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3117 | } |
3118 | ||
3119 | /* | |
a737b3e2 AM |
3120 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3121 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3122 | */ |
343e0d7a | 3123 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3124 | { |
9a2dba4b | 3125 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3126 | |
3127 | check_irq_off(); | |
3128 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3129 | ||
729bd0b7 PE |
3130 | if (cache_free_alien(cachep, objp)) |
3131 | return; | |
3132 | ||
1da177e4 LT |
3133 | if (likely(ac->avail < ac->limit)) { |
3134 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3135 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3136 | return; |
3137 | } else { | |
3138 | STATS_INC_FREEMISS(cachep); | |
3139 | cache_flusharray(cachep, ac); | |
e498be7d | 3140 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3141 | } |
3142 | } | |
3143 | ||
3144 | /** | |
3145 | * kmem_cache_alloc - Allocate an object | |
3146 | * @cachep: The cache to allocate from. | |
3147 | * @flags: See kmalloc(). | |
3148 | * | |
3149 | * Allocate an object from this cache. The flags are only relevant | |
3150 | * if the cache has no available objects. | |
3151 | */ | |
343e0d7a | 3152 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3153 | { |
7fd6b141 | 3154 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3155 | } |
3156 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3157 | ||
a8c0f9a4 PE |
3158 | /** |
3159 | * kmem_cache_alloc - Allocate an object. The memory is set to zero. | |
3160 | * @cache: The cache to allocate from. | |
3161 | * @flags: See kmalloc(). | |
3162 | * | |
3163 | * Allocate an object from this cache and set the allocated memory to zero. | |
3164 | * The flags are only relevant if the cache has no available objects. | |
3165 | */ | |
3166 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3167 | { | |
3168 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3169 | if (ret) | |
3170 | memset(ret, 0, obj_size(cache)); | |
3171 | return ret; | |
3172 | } | |
3173 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3174 | ||
1da177e4 LT |
3175 | /** |
3176 | * kmem_ptr_validate - check if an untrusted pointer might | |
3177 | * be a slab entry. | |
3178 | * @cachep: the cache we're checking against | |
3179 | * @ptr: pointer to validate | |
3180 | * | |
3181 | * This verifies that the untrusted pointer looks sane: | |
3182 | * it is _not_ a guarantee that the pointer is actually | |
3183 | * part of the slab cache in question, but it at least | |
3184 | * validates that the pointer can be dereferenced and | |
3185 | * looks half-way sane. | |
3186 | * | |
3187 | * Currently only used for dentry validation. | |
3188 | */ | |
343e0d7a | 3189 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3190 | { |
b28a02de | 3191 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3192 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3193 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3194 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3195 | struct page *page; |
3196 | ||
3197 | if (unlikely(addr < min_addr)) | |
3198 | goto out; | |
3199 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3200 | goto out; | |
3201 | if (unlikely(addr & align_mask)) | |
3202 | goto out; | |
3203 | if (unlikely(!kern_addr_valid(addr))) | |
3204 | goto out; | |
3205 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3206 | goto out; | |
3207 | page = virt_to_page(ptr); | |
3208 | if (unlikely(!PageSlab(page))) | |
3209 | goto out; | |
065d41cb | 3210 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3211 | goto out; |
3212 | return 1; | |
a737b3e2 | 3213 | out: |
1da177e4 LT |
3214 | return 0; |
3215 | } | |
3216 | ||
3217 | #ifdef CONFIG_NUMA | |
3218 | /** | |
3219 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3220 | * @cachep: The cache to allocate from. | |
3221 | * @flags: See kmalloc(). | |
3222 | * @nodeid: node number of the target node. | |
3223 | * | |
3224 | * Identical to kmem_cache_alloc, except that this function is slow | |
3225 | * and can sleep. And it will allocate memory on the given node, which | |
3226 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3227 | * New and improved: it will now make sure that the object gets |
3228 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3229 | */ |
343e0d7a | 3230 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3231 | { |
e498be7d CL |
3232 | unsigned long save_flags; |
3233 | void *ptr; | |
1da177e4 | 3234 | |
e498be7d CL |
3235 | cache_alloc_debugcheck_before(cachep, flags); |
3236 | local_irq_save(save_flags); | |
18f820f6 CL |
3237 | |
3238 | if (nodeid == -1 || nodeid == numa_node_id() || | |
a737b3e2 | 3239 | !cachep->nodelists[nodeid]) |
5c382300 AK |
3240 | ptr = ____cache_alloc(cachep, flags); |
3241 | else | |
3242 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3243 | local_irq_restore(save_flags); |
18f820f6 CL |
3244 | |
3245 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3246 | __builtin_return_address(0)); | |
1da177e4 | 3247 | |
e498be7d | 3248 | return ptr; |
1da177e4 LT |
3249 | } |
3250 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3251 | ||
dd0fc66f | 3252 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3253 | { |
343e0d7a | 3254 | struct kmem_cache *cachep; |
97e2bde4 MS |
3255 | |
3256 | cachep = kmem_find_general_cachep(size, flags); | |
3257 | if (unlikely(cachep == NULL)) | |
3258 | return NULL; | |
3259 | return kmem_cache_alloc_node(cachep, flags, node); | |
3260 | } | |
3261 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
3262 | #endif |
3263 | ||
3264 | /** | |
3265 | * kmalloc - allocate memory | |
3266 | * @size: how many bytes of memory are required. | |
3267 | * @flags: the type of memory to allocate. | |
911851e6 | 3268 | * @caller: function caller for debug tracking of the caller |
1da177e4 LT |
3269 | * |
3270 | * kmalloc is the normal method of allocating memory | |
3271 | * in the kernel. | |
3272 | * | |
3273 | * The @flags argument may be one of: | |
3274 | * | |
3275 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
3276 | * | |
3277 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
3278 | * | |
3279 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
3280 | * | |
3281 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
3282 | * must be suitable for DMA. This can mean different things on different | |
3283 | * platforms. For example, on i386, it means that the memory must come | |
3284 | * from the first 16MB. | |
3285 | */ | |
7fd6b141 PE |
3286 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3287 | void *caller) | |
1da177e4 | 3288 | { |
343e0d7a | 3289 | struct kmem_cache *cachep; |
1da177e4 | 3290 | |
97e2bde4 MS |
3291 | /* If you want to save a few bytes .text space: replace |
3292 | * __ with kmem_. | |
3293 | * Then kmalloc uses the uninlined functions instead of the inline | |
3294 | * functions. | |
3295 | */ | |
3296 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3297 | if (unlikely(cachep == NULL)) |
3298 | return NULL; | |
7fd6b141 PE |
3299 | return __cache_alloc(cachep, flags, caller); |
3300 | } | |
3301 | ||
7fd6b141 PE |
3302 | |
3303 | void *__kmalloc(size_t size, gfp_t flags) | |
3304 | { | |
871751e2 | 3305 | #ifndef CONFIG_DEBUG_SLAB |
7fd6b141 | 3306 | return __do_kmalloc(size, flags, NULL); |
871751e2 AV |
3307 | #else |
3308 | return __do_kmalloc(size, flags, __builtin_return_address(0)); | |
3309 | #endif | |
1da177e4 LT |
3310 | } |
3311 | EXPORT_SYMBOL(__kmalloc); | |
3312 | ||
871751e2 | 3313 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3314 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3315 | { | |
3316 | return __do_kmalloc(size, flags, caller); | |
3317 | } | |
3318 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
7fd6b141 PE |
3319 | #endif |
3320 | ||
1da177e4 LT |
3321 | #ifdef CONFIG_SMP |
3322 | /** | |
3323 | * __alloc_percpu - allocate one copy of the object for every present | |
3324 | * cpu in the system, zeroing them. | |
3325 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
3326 | * | |
3327 | * @size: how many bytes of memory are required. | |
1da177e4 | 3328 | */ |
f9f75005 | 3329 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
3330 | { |
3331 | int i; | |
b28a02de | 3332 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3333 | |
3334 | if (!pdata) | |
3335 | return NULL; | |
3336 | ||
e498be7d CL |
3337 | /* |
3338 | * Cannot use for_each_online_cpu since a cpu may come online | |
3339 | * and we have no way of figuring out how to fix the array | |
3340 | * that we have allocated then.... | |
3341 | */ | |
0a945022 | 3342 | for_each_possible_cpu(i) { |
e498be7d CL |
3343 | int node = cpu_to_node(i); |
3344 | ||
3345 | if (node_online(node)) | |
3346 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3347 | else | |
3348 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3349 | |
3350 | if (!pdata->ptrs[i]) | |
3351 | goto unwind_oom; | |
3352 | memset(pdata->ptrs[i], 0, size); | |
3353 | } | |
3354 | ||
3355 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3356 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3357 | |
a737b3e2 | 3358 | unwind_oom: |
1da177e4 LT |
3359 | while (--i >= 0) { |
3360 | if (!cpu_possible(i)) | |
3361 | continue; | |
3362 | kfree(pdata->ptrs[i]); | |
3363 | } | |
3364 | kfree(pdata); | |
3365 | return NULL; | |
3366 | } | |
3367 | EXPORT_SYMBOL(__alloc_percpu); | |
3368 | #endif | |
3369 | ||
3370 | /** | |
3371 | * kmem_cache_free - Deallocate an object | |
3372 | * @cachep: The cache the allocation was from. | |
3373 | * @objp: The previously allocated object. | |
3374 | * | |
3375 | * Free an object which was previously allocated from this | |
3376 | * cache. | |
3377 | */ | |
343e0d7a | 3378 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3379 | { |
3380 | unsigned long flags; | |
3381 | ||
3382 | local_irq_save(flags); | |
3383 | __cache_free(cachep, objp); | |
3384 | local_irq_restore(flags); | |
3385 | } | |
3386 | EXPORT_SYMBOL(kmem_cache_free); | |
3387 | ||
1da177e4 LT |
3388 | /** |
3389 | * kfree - free previously allocated memory | |
3390 | * @objp: pointer returned by kmalloc. | |
3391 | * | |
80e93eff PE |
3392 | * If @objp is NULL, no operation is performed. |
3393 | * | |
1da177e4 LT |
3394 | * Don't free memory not originally allocated by kmalloc() |
3395 | * or you will run into trouble. | |
3396 | */ | |
3397 | void kfree(const void *objp) | |
3398 | { | |
343e0d7a | 3399 | struct kmem_cache *c; |
1da177e4 LT |
3400 | unsigned long flags; |
3401 | ||
3402 | if (unlikely(!objp)) | |
3403 | return; | |
3404 | local_irq_save(flags); | |
3405 | kfree_debugcheck(objp); | |
6ed5eb22 | 3406 | c = virt_to_cache(objp); |
3dafccf2 | 3407 | mutex_debug_check_no_locks_freed(objp, obj_size(c)); |
b28a02de | 3408 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3409 | local_irq_restore(flags); |
3410 | } | |
3411 | EXPORT_SYMBOL(kfree); | |
3412 | ||
3413 | #ifdef CONFIG_SMP | |
3414 | /** | |
3415 | * free_percpu - free previously allocated percpu memory | |
3416 | * @objp: pointer returned by alloc_percpu. | |
3417 | * | |
3418 | * Don't free memory not originally allocated by alloc_percpu() | |
3419 | * The complemented objp is to check for that. | |
3420 | */ | |
b28a02de | 3421 | void free_percpu(const void *objp) |
1da177e4 LT |
3422 | { |
3423 | int i; | |
b28a02de | 3424 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3425 | |
e498be7d CL |
3426 | /* |
3427 | * We allocate for all cpus so we cannot use for online cpu here. | |
3428 | */ | |
0a945022 | 3429 | for_each_possible_cpu(i) |
b28a02de | 3430 | kfree(p->ptrs[i]); |
1da177e4 LT |
3431 | kfree(p); |
3432 | } | |
3433 | EXPORT_SYMBOL(free_percpu); | |
3434 | #endif | |
3435 | ||
343e0d7a | 3436 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3437 | { |
3dafccf2 | 3438 | return obj_size(cachep); |
1da177e4 LT |
3439 | } |
3440 | EXPORT_SYMBOL(kmem_cache_size); | |
3441 | ||
343e0d7a | 3442 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3443 | { |
3444 | return cachep->name; | |
3445 | } | |
3446 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3447 | ||
e498be7d | 3448 | /* |
0718dc2a | 3449 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3450 | */ |
343e0d7a | 3451 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3452 | { |
3453 | int node; | |
3454 | struct kmem_list3 *l3; | |
cafeb02e CL |
3455 | struct array_cache *new_shared; |
3456 | struct array_cache **new_alien; | |
e498be7d CL |
3457 | |
3458 | for_each_online_node(node) { | |
cafeb02e | 3459 | |
a737b3e2 AM |
3460 | new_alien = alloc_alien_cache(node, cachep->limit); |
3461 | if (!new_alien) | |
e498be7d | 3462 | goto fail; |
cafeb02e | 3463 | |
0718dc2a CL |
3464 | new_shared = alloc_arraycache(node, |
3465 | cachep->shared*cachep->batchcount, | |
a737b3e2 | 3466 | 0xbaadf00d); |
0718dc2a CL |
3467 | if (!new_shared) { |
3468 | free_alien_cache(new_alien); | |
e498be7d | 3469 | goto fail; |
0718dc2a | 3470 | } |
cafeb02e | 3471 | |
a737b3e2 AM |
3472 | l3 = cachep->nodelists[node]; |
3473 | if (l3) { | |
cafeb02e CL |
3474 | struct array_cache *shared = l3->shared; |
3475 | ||
e498be7d CL |
3476 | spin_lock_irq(&l3->list_lock); |
3477 | ||
cafeb02e | 3478 | if (shared) |
0718dc2a CL |
3479 | free_block(cachep, shared->entry, |
3480 | shared->avail, node); | |
e498be7d | 3481 | |
cafeb02e CL |
3482 | l3->shared = new_shared; |
3483 | if (!l3->alien) { | |
e498be7d CL |
3484 | l3->alien = new_alien; |
3485 | new_alien = NULL; | |
3486 | } | |
b28a02de | 3487 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3488 | cachep->batchcount + cachep->num; |
e498be7d | 3489 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3490 | kfree(shared); |
e498be7d CL |
3491 | free_alien_cache(new_alien); |
3492 | continue; | |
3493 | } | |
a737b3e2 | 3494 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3495 | if (!l3) { |
3496 | free_alien_cache(new_alien); | |
3497 | kfree(new_shared); | |
e498be7d | 3498 | goto fail; |
0718dc2a | 3499 | } |
e498be7d CL |
3500 | |
3501 | kmem_list3_init(l3); | |
3502 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3503 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3504 | l3->shared = new_shared; |
e498be7d | 3505 | l3->alien = new_alien; |
b28a02de | 3506 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3507 | cachep->batchcount + cachep->num; |
e498be7d CL |
3508 | cachep->nodelists[node] = l3; |
3509 | } | |
cafeb02e | 3510 | return 0; |
0718dc2a | 3511 | |
a737b3e2 | 3512 | fail: |
0718dc2a CL |
3513 | if (!cachep->next.next) { |
3514 | /* Cache is not active yet. Roll back what we did */ | |
3515 | node--; | |
3516 | while (node >= 0) { | |
3517 | if (cachep->nodelists[node]) { | |
3518 | l3 = cachep->nodelists[node]; | |
3519 | ||
3520 | kfree(l3->shared); | |
3521 | free_alien_cache(l3->alien); | |
3522 | kfree(l3); | |
3523 | cachep->nodelists[node] = NULL; | |
3524 | } | |
3525 | node--; | |
3526 | } | |
3527 | } | |
cafeb02e | 3528 | return -ENOMEM; |
e498be7d CL |
3529 | } |
3530 | ||
1da177e4 | 3531 | struct ccupdate_struct { |
343e0d7a | 3532 | struct kmem_cache *cachep; |
1da177e4 LT |
3533 | struct array_cache *new[NR_CPUS]; |
3534 | }; | |
3535 | ||
3536 | static void do_ccupdate_local(void *info) | |
3537 | { | |
a737b3e2 | 3538 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3539 | struct array_cache *old; |
3540 | ||
3541 | check_irq_off(); | |
9a2dba4b | 3542 | old = cpu_cache_get(new->cachep); |
e498be7d | 3543 | |
1da177e4 LT |
3544 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3545 | new->new[smp_processor_id()] = old; | |
3546 | } | |
3547 | ||
b5d8ca7c | 3548 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3549 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3550 | int batchcount, int shared) | |
1da177e4 LT |
3551 | { |
3552 | struct ccupdate_struct new; | |
e498be7d | 3553 | int i, err; |
1da177e4 | 3554 | |
b28a02de | 3555 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3556 | for_each_online_cpu(i) { |
a737b3e2 AM |
3557 | new.new[i] = alloc_arraycache(cpu_to_node(i), limit, |
3558 | batchcount); | |
e498be7d | 3559 | if (!new.new[i]) { |
b28a02de PE |
3560 | for (i--; i >= 0; i--) |
3561 | kfree(new.new[i]); | |
e498be7d | 3562 | return -ENOMEM; |
1da177e4 LT |
3563 | } |
3564 | } | |
3565 | new.cachep = cachep; | |
3566 | ||
a07fa394 | 3567 | on_each_cpu(do_ccupdate_local, (void *)&new, 1, 1); |
e498be7d | 3568 | |
1da177e4 | 3569 | check_irq_on(); |
1da177e4 LT |
3570 | cachep->batchcount = batchcount; |
3571 | cachep->limit = limit; | |
e498be7d | 3572 | cachep->shared = shared; |
1da177e4 | 3573 | |
e498be7d | 3574 | for_each_online_cpu(i) { |
1da177e4 LT |
3575 | struct array_cache *ccold = new.new[i]; |
3576 | if (!ccold) | |
3577 | continue; | |
e498be7d | 3578 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3579 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3580 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3581 | kfree(ccold); |
3582 | } | |
1da177e4 | 3583 | |
e498be7d CL |
3584 | err = alloc_kmemlist(cachep); |
3585 | if (err) { | |
3586 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3587 | cachep->name, -err); |
e498be7d | 3588 | BUG(); |
1da177e4 | 3589 | } |
1da177e4 LT |
3590 | return 0; |
3591 | } | |
3592 | ||
b5d8ca7c | 3593 | /* Called with cache_chain_mutex held always */ |
343e0d7a | 3594 | static void enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3595 | { |
3596 | int err; | |
3597 | int limit, shared; | |
3598 | ||
a737b3e2 AM |
3599 | /* |
3600 | * The head array serves three purposes: | |
1da177e4 LT |
3601 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3602 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3603 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3604 | * bufctl chains: array operations are cheaper. |
3605 | * The numbers are guessed, we should auto-tune as described by | |
3606 | * Bonwick. | |
3607 | */ | |
3dafccf2 | 3608 | if (cachep->buffer_size > 131072) |
1da177e4 | 3609 | limit = 1; |
3dafccf2 | 3610 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3611 | limit = 8; |
3dafccf2 | 3612 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3613 | limit = 24; |
3dafccf2 | 3614 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3615 | limit = 54; |
3616 | else | |
3617 | limit = 120; | |
3618 | ||
a737b3e2 AM |
3619 | /* |
3620 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3621 | * allocation behaviour: Most allocs on one cpu, most free operations |
3622 | * on another cpu. For these cases, an efficient object passing between | |
3623 | * cpus is necessary. This is provided by a shared array. The array | |
3624 | * replaces Bonwick's magazine layer. | |
3625 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3626 | * to a larger limit. Thus disabled by default. | |
3627 | */ | |
3628 | shared = 0; | |
3629 | #ifdef CONFIG_SMP | |
3dafccf2 | 3630 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3631 | shared = 8; |
3632 | #endif | |
3633 | ||
3634 | #if DEBUG | |
a737b3e2 AM |
3635 | /* |
3636 | * With debugging enabled, large batchcount lead to excessively long | |
3637 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3638 | */ |
3639 | if (limit > 32) | |
3640 | limit = 32; | |
3641 | #endif | |
b28a02de | 3642 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3643 | if (err) |
3644 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3645 | cachep->name, -err); |
1da177e4 LT |
3646 | } |
3647 | ||
1b55253a CL |
3648 | /* |
3649 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3650 | * necessary. Note that the l3 listlock also protects the array_cache |
3651 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3652 | */ |
3653 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
3654 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
3655 | { |
3656 | int tofree; | |
3657 | ||
1b55253a CL |
3658 | if (!ac || !ac->avail) |
3659 | return; | |
1da177e4 LT |
3660 | if (ac->touched && !force) { |
3661 | ac->touched = 0; | |
b18e7e65 | 3662 | } else { |
1b55253a | 3663 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
3664 | if (ac->avail) { |
3665 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3666 | if (tofree > ac->avail) | |
3667 | tofree = (ac->avail + 1) / 2; | |
3668 | free_block(cachep, ac->entry, tofree, node); | |
3669 | ac->avail -= tofree; | |
3670 | memmove(ac->entry, &(ac->entry[tofree]), | |
3671 | sizeof(void *) * ac->avail); | |
3672 | } | |
1b55253a | 3673 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
3674 | } |
3675 | } | |
3676 | ||
3677 | /** | |
3678 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3679 | * @unused: unused parameter |
1da177e4 LT |
3680 | * |
3681 | * Called from workqueue/eventd every few seconds. | |
3682 | * Purpose: | |
3683 | * - clear the per-cpu caches for this CPU. | |
3684 | * - return freeable pages to the main free memory pool. | |
3685 | * | |
a737b3e2 AM |
3686 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3687 | * again on the next iteration. | |
1da177e4 LT |
3688 | */ |
3689 | static void cache_reap(void *unused) | |
3690 | { | |
7a7c381d | 3691 | struct kmem_cache *searchp; |
e498be7d | 3692 | struct kmem_list3 *l3; |
aab2207c | 3693 | int node = numa_node_id(); |
1da177e4 | 3694 | |
fc0abb14 | 3695 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3696 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3697 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3698 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3699 | return; |
3700 | } | |
3701 | ||
7a7c381d | 3702 | list_for_each_entry(searchp, &cache_chain, next) { |
b28a02de | 3703 | struct list_head *p; |
1da177e4 LT |
3704 | int tofree; |
3705 | struct slab *slabp; | |
3706 | ||
1da177e4 LT |
3707 | check_irq_on(); |
3708 | ||
35386e3b CL |
3709 | /* |
3710 | * We only take the l3 lock if absolutely necessary and we | |
3711 | * have established with reasonable certainty that | |
3712 | * we can do some work if the lock was obtained. | |
3713 | */ | |
aab2207c | 3714 | l3 = searchp->nodelists[node]; |
35386e3b | 3715 | |
8fce4d8e | 3716 | reap_alien(searchp, l3); |
1da177e4 | 3717 | |
aab2207c | 3718 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 3719 | |
35386e3b CL |
3720 | /* |
3721 | * These are racy checks but it does not matter | |
3722 | * if we skip one check or scan twice. | |
3723 | */ | |
e498be7d | 3724 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 3725 | goto next; |
1da177e4 | 3726 | |
e498be7d | 3727 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3728 | |
aab2207c | 3729 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 3730 | |
e498be7d CL |
3731 | if (l3->free_touched) { |
3732 | l3->free_touched = 0; | |
35386e3b | 3733 | goto next; |
1da177e4 LT |
3734 | } |
3735 | ||
a737b3e2 AM |
3736 | tofree = (l3->free_limit + 5 * searchp->num - 1) / |
3737 | (5 * searchp->num); | |
1da177e4 | 3738 | do { |
35386e3b CL |
3739 | /* |
3740 | * Do not lock if there are no free blocks. | |
3741 | */ | |
3742 | if (list_empty(&l3->slabs_free)) | |
3743 | break; | |
3744 | ||
3745 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3746 | p = l3->slabs_free.next; |
35386e3b CL |
3747 | if (p == &(l3->slabs_free)) { |
3748 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3749 | break; |
35386e3b | 3750 | } |
1da177e4 LT |
3751 | |
3752 | slabp = list_entry(p, struct slab, list); | |
3753 | BUG_ON(slabp->inuse); | |
3754 | list_del(&slabp->list); | |
3755 | STATS_INC_REAPED(searchp); | |
3756 | ||
a737b3e2 AM |
3757 | /* |
3758 | * Safe to drop the lock. The slab is no longer linked | |
3759 | * to the cache. searchp cannot disappear, we hold | |
1da177e4 LT |
3760 | * cache_chain_lock |
3761 | */ | |
e498be7d CL |
3762 | l3->free_objects -= searchp->num; |
3763 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3764 | slab_destroy(searchp, slabp); |
b28a02de | 3765 | } while (--tofree > 0); |
35386e3b | 3766 | next: |
1da177e4 LT |
3767 | cond_resched(); |
3768 | } | |
3769 | check_irq_on(); | |
fc0abb14 | 3770 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 3771 | next_reap_node(); |
a737b3e2 | 3772 | /* Set up the next iteration */ |
cd61ef62 | 3773 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3774 | } |
3775 | ||
3776 | #ifdef CONFIG_PROC_FS | |
3777 | ||
85289f98 | 3778 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3779 | { |
85289f98 PE |
3780 | /* |
3781 | * Output format version, so at least we can change it | |
3782 | * without _too_ many complaints. | |
3783 | */ | |
1da177e4 | 3784 | #if STATS |
85289f98 | 3785 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3786 | #else |
85289f98 | 3787 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3788 | #endif |
85289f98 PE |
3789 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3790 | "<objperslab> <pagesperslab>"); | |
3791 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3792 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3793 | #if STATS |
85289f98 | 3794 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 3795 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 3796 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 3797 | #endif |
85289f98 PE |
3798 | seq_putc(m, '\n'); |
3799 | } | |
3800 | ||
3801 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3802 | { | |
3803 | loff_t n = *pos; | |
3804 | struct list_head *p; | |
3805 | ||
fc0abb14 | 3806 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3807 | if (!n) |
3808 | print_slabinfo_header(m); | |
1da177e4 LT |
3809 | p = cache_chain.next; |
3810 | while (n--) { | |
3811 | p = p->next; | |
3812 | if (p == &cache_chain) | |
3813 | return NULL; | |
3814 | } | |
343e0d7a | 3815 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3816 | } |
3817 | ||
3818 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3819 | { | |
343e0d7a | 3820 | struct kmem_cache *cachep = p; |
1da177e4 | 3821 | ++*pos; |
a737b3e2 AM |
3822 | return cachep->next.next == &cache_chain ? |
3823 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
3824 | } |
3825 | ||
3826 | static void s_stop(struct seq_file *m, void *p) | |
3827 | { | |
fc0abb14 | 3828 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3829 | } |
3830 | ||
3831 | static int s_show(struct seq_file *m, void *p) | |
3832 | { | |
343e0d7a | 3833 | struct kmem_cache *cachep = p; |
b28a02de PE |
3834 | struct slab *slabp; |
3835 | unsigned long active_objs; | |
3836 | unsigned long num_objs; | |
3837 | unsigned long active_slabs = 0; | |
3838 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3839 | const char *name; |
1da177e4 | 3840 | char *error = NULL; |
e498be7d CL |
3841 | int node; |
3842 | struct kmem_list3 *l3; | |
1da177e4 | 3843 | |
1da177e4 LT |
3844 | active_objs = 0; |
3845 | num_slabs = 0; | |
e498be7d CL |
3846 | for_each_online_node(node) { |
3847 | l3 = cachep->nodelists[node]; | |
3848 | if (!l3) | |
3849 | continue; | |
3850 | ||
ca3b9b91 RT |
3851 | check_irq_on(); |
3852 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3853 | |
7a7c381d | 3854 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
3855 | if (slabp->inuse != cachep->num && !error) |
3856 | error = "slabs_full accounting error"; | |
3857 | active_objs += cachep->num; | |
3858 | active_slabs++; | |
3859 | } | |
7a7c381d | 3860 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
3861 | if (slabp->inuse == cachep->num && !error) |
3862 | error = "slabs_partial inuse accounting error"; | |
3863 | if (!slabp->inuse && !error) | |
3864 | error = "slabs_partial/inuse accounting error"; | |
3865 | active_objs += slabp->inuse; | |
3866 | active_slabs++; | |
3867 | } | |
7a7c381d | 3868 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
3869 | if (slabp->inuse && !error) |
3870 | error = "slabs_free/inuse accounting error"; | |
3871 | num_slabs++; | |
3872 | } | |
3873 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3874 | if (l3->shared) |
3875 | shared_avail += l3->shared->avail; | |
e498be7d | 3876 | |
ca3b9b91 | 3877 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3878 | } |
b28a02de PE |
3879 | num_slabs += active_slabs; |
3880 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3881 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3882 | error = "free_objects accounting error"; |
3883 | ||
b28a02de | 3884 | name = cachep->name; |
1da177e4 LT |
3885 | if (error) |
3886 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3887 | ||
3888 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3889 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3890 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3891 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3892 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3893 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3894 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3895 | #if STATS |
b28a02de | 3896 | { /* list3 stats */ |
1da177e4 LT |
3897 | unsigned long high = cachep->high_mark; |
3898 | unsigned long allocs = cachep->num_allocations; | |
3899 | unsigned long grown = cachep->grown; | |
3900 | unsigned long reaped = cachep->reaped; | |
3901 | unsigned long errors = cachep->errors; | |
3902 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3903 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3904 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 3905 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 3906 | |
e498be7d | 3907 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 3908 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 3909 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 3910 | node_frees, overflows); |
1da177e4 LT |
3911 | } |
3912 | /* cpu stats */ | |
3913 | { | |
3914 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3915 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3916 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3917 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3918 | ||
3919 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3920 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3921 | } |
3922 | #endif | |
3923 | seq_putc(m, '\n'); | |
1da177e4 LT |
3924 | return 0; |
3925 | } | |
3926 | ||
3927 | /* | |
3928 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3929 | * | |
3930 | * Output layout: | |
3931 | * cache-name | |
3932 | * num-active-objs | |
3933 | * total-objs | |
3934 | * object size | |
3935 | * num-active-slabs | |
3936 | * total-slabs | |
3937 | * num-pages-per-slab | |
3938 | * + further values on SMP and with statistics enabled | |
3939 | */ | |
3940 | ||
3941 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3942 | .start = s_start, |
3943 | .next = s_next, | |
3944 | .stop = s_stop, | |
3945 | .show = s_show, | |
1da177e4 LT |
3946 | }; |
3947 | ||
3948 | #define MAX_SLABINFO_WRITE 128 | |
3949 | /** | |
3950 | * slabinfo_write - Tuning for the slab allocator | |
3951 | * @file: unused | |
3952 | * @buffer: user buffer | |
3953 | * @count: data length | |
3954 | * @ppos: unused | |
3955 | */ | |
b28a02de PE |
3956 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3957 | size_t count, loff_t *ppos) | |
1da177e4 | 3958 | { |
b28a02de | 3959 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 3960 | int limit, batchcount, shared, res; |
7a7c381d | 3961 | struct kmem_cache *cachep; |
b28a02de | 3962 | |
1da177e4 LT |
3963 | if (count > MAX_SLABINFO_WRITE) |
3964 | return -EINVAL; | |
3965 | if (copy_from_user(&kbuf, buffer, count)) | |
3966 | return -EFAULT; | |
b28a02de | 3967 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3968 | |
3969 | tmp = strchr(kbuf, ' '); | |
3970 | if (!tmp) | |
3971 | return -EINVAL; | |
3972 | *tmp = '\0'; | |
3973 | tmp++; | |
3974 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3975 | return -EINVAL; | |
3976 | ||
3977 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3978 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3979 | res = -EINVAL; |
7a7c381d | 3980 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 3981 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
3982 | if (limit < 1 || batchcount < 1 || |
3983 | batchcount > limit || shared < 0) { | |
e498be7d | 3984 | res = 0; |
1da177e4 | 3985 | } else { |
e498be7d | 3986 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3987 | batchcount, shared); |
1da177e4 LT |
3988 | } |
3989 | break; | |
3990 | } | |
3991 | } | |
fc0abb14 | 3992 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3993 | if (res >= 0) |
3994 | res = count; | |
3995 | return res; | |
3996 | } | |
871751e2 AV |
3997 | |
3998 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
3999 | ||
4000 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4001 | { | |
4002 | loff_t n = *pos; | |
4003 | struct list_head *p; | |
4004 | ||
4005 | mutex_lock(&cache_chain_mutex); | |
4006 | p = cache_chain.next; | |
4007 | while (n--) { | |
4008 | p = p->next; | |
4009 | if (p == &cache_chain) | |
4010 | return NULL; | |
4011 | } | |
4012 | return list_entry(p, struct kmem_cache, next); | |
4013 | } | |
4014 | ||
4015 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4016 | { | |
4017 | unsigned long *p; | |
4018 | int l; | |
4019 | if (!v) | |
4020 | return 1; | |
4021 | l = n[1]; | |
4022 | p = n + 2; | |
4023 | while (l) { | |
4024 | int i = l/2; | |
4025 | unsigned long *q = p + 2 * i; | |
4026 | if (*q == v) { | |
4027 | q[1]++; | |
4028 | return 1; | |
4029 | } | |
4030 | if (*q > v) { | |
4031 | l = i; | |
4032 | } else { | |
4033 | p = q + 2; | |
4034 | l -= i + 1; | |
4035 | } | |
4036 | } | |
4037 | if (++n[1] == n[0]) | |
4038 | return 0; | |
4039 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4040 | p[0] = v; | |
4041 | p[1] = 1; | |
4042 | return 1; | |
4043 | } | |
4044 | ||
4045 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4046 | { | |
4047 | void *p; | |
4048 | int i; | |
4049 | if (n[0] == n[1]) | |
4050 | return; | |
4051 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4052 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4053 | continue; | |
4054 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4055 | return; | |
4056 | } | |
4057 | } | |
4058 | ||
4059 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4060 | { | |
4061 | #ifdef CONFIG_KALLSYMS | |
4062 | char *modname; | |
4063 | const char *name; | |
4064 | unsigned long offset, size; | |
4065 | char namebuf[KSYM_NAME_LEN+1]; | |
4066 | ||
4067 | name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); | |
4068 | ||
4069 | if (name) { | |
4070 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4071 | if (modname) | |
4072 | seq_printf(m, " [%s]", modname); | |
4073 | return; | |
4074 | } | |
4075 | #endif | |
4076 | seq_printf(m, "%p", (void *)address); | |
4077 | } | |
4078 | ||
4079 | static int leaks_show(struct seq_file *m, void *p) | |
4080 | { | |
4081 | struct kmem_cache *cachep = p; | |
871751e2 AV |
4082 | struct slab *slabp; |
4083 | struct kmem_list3 *l3; | |
4084 | const char *name; | |
4085 | unsigned long *n = m->private; | |
4086 | int node; | |
4087 | int i; | |
4088 | ||
4089 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4090 | return 0; | |
4091 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4092 | return 0; | |
4093 | ||
4094 | /* OK, we can do it */ | |
4095 | ||
4096 | n[1] = 0; | |
4097 | ||
4098 | for_each_online_node(node) { | |
4099 | l3 = cachep->nodelists[node]; | |
4100 | if (!l3) | |
4101 | continue; | |
4102 | ||
4103 | check_irq_on(); | |
4104 | spin_lock_irq(&l3->list_lock); | |
4105 | ||
7a7c381d | 4106 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4107 | handle_slab(n, cachep, slabp); |
7a7c381d | 4108 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4109 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4110 | spin_unlock_irq(&l3->list_lock); |
4111 | } | |
4112 | name = cachep->name; | |
4113 | if (n[0] == n[1]) { | |
4114 | /* Increase the buffer size */ | |
4115 | mutex_unlock(&cache_chain_mutex); | |
4116 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4117 | if (!m->private) { | |
4118 | /* Too bad, we are really out */ | |
4119 | m->private = n; | |
4120 | mutex_lock(&cache_chain_mutex); | |
4121 | return -ENOMEM; | |
4122 | } | |
4123 | *(unsigned long *)m->private = n[0] * 2; | |
4124 | kfree(n); | |
4125 | mutex_lock(&cache_chain_mutex); | |
4126 | /* Now make sure this entry will be retried */ | |
4127 | m->count = m->size; | |
4128 | return 0; | |
4129 | } | |
4130 | for (i = 0; i < n[1]; i++) { | |
4131 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4132 | show_symbol(m, n[2*i+2]); | |
4133 | seq_putc(m, '\n'); | |
4134 | } | |
4135 | return 0; | |
4136 | } | |
4137 | ||
4138 | struct seq_operations slabstats_op = { | |
4139 | .start = leaks_start, | |
4140 | .next = s_next, | |
4141 | .stop = s_stop, | |
4142 | .show = leaks_show, | |
4143 | }; | |
4144 | #endif | |
1da177e4 LT |
4145 | #endif |
4146 | ||
00e145b6 MS |
4147 | /** |
4148 | * ksize - get the actual amount of memory allocated for a given object | |
4149 | * @objp: Pointer to the object | |
4150 | * | |
4151 | * kmalloc may internally round up allocations and return more memory | |
4152 | * than requested. ksize() can be used to determine the actual amount of | |
4153 | * memory allocated. The caller may use this additional memory, even though | |
4154 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4155 | * The caller must guarantee that objp points to a valid object previously | |
4156 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4157 | * must not be freed during the duration of the call. | |
4158 | */ | |
1da177e4 LT |
4159 | unsigned int ksize(const void *objp) |
4160 | { | |
00e145b6 MS |
4161 | if (unlikely(objp == NULL)) |
4162 | return 0; | |
1da177e4 | 4163 | |
6ed5eb22 | 4164 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4165 | } |