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 | |
183ff22b | 29 | * slabs and you must pass objects with the same initializations to |
1da177e4 LT |
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 | ||
1da177e4 LT |
89 | #include <linux/slab.h> |
90 | #include <linux/mm.h> | |
c9cf5528 | 91 | #include <linux/poison.h> |
1da177e4 LT |
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> |
a0ec95a8 | 98 | #include <linux/proc_fs.h> |
1da177e4 LT |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> | |
101 | #include <linux/kallsyms.h> | |
102 | #include <linux/cpu.h> | |
103 | #include <linux/sysctl.h> | |
104 | #include <linux/module.h> | |
02af61bb | 105 | #include <linux/kmemtrace.h> |
1da177e4 | 106 | #include <linux/rcupdate.h> |
543537bd | 107 | #include <linux/string.h> |
138ae663 | 108 | #include <linux/uaccess.h> |
e498be7d | 109 | #include <linux/nodemask.h> |
d5cff635 | 110 | #include <linux/kmemleak.h> |
dc85da15 | 111 | #include <linux/mempolicy.h> |
fc0abb14 | 112 | #include <linux/mutex.h> |
8a8b6502 | 113 | #include <linux/fault-inject.h> |
e7eebaf6 | 114 | #include <linux/rtmutex.h> |
6a2d7a95 | 115 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 116 | #include <linux/debugobjects.h> |
1da177e4 | 117 | |
1da177e4 LT |
118 | #include <asm/cacheflush.h> |
119 | #include <asm/tlbflush.h> | |
120 | #include <asm/page.h> | |
121 | ||
122 | /* | |
50953fe9 | 123 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
124 | * 0 for faster, smaller code (especially in the critical paths). |
125 | * | |
126 | * STATS - 1 to collect stats for /proc/slabinfo. | |
127 | * 0 for faster, smaller code (especially in the critical paths). | |
128 | * | |
129 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
130 | */ | |
131 | ||
132 | #ifdef CONFIG_DEBUG_SLAB | |
133 | #define DEBUG 1 | |
134 | #define STATS 1 | |
135 | #define FORCED_DEBUG 1 | |
136 | #else | |
137 | #define DEBUG 0 | |
138 | #define STATS 0 | |
139 | #define FORCED_DEBUG 0 | |
140 | #endif | |
141 | ||
1da177e4 LT |
142 | /* Shouldn't this be in a header file somewhere? */ |
143 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 144 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 145 | |
1da177e4 LT |
146 | #ifndef ARCH_KMALLOC_MINALIGN |
147 | /* | |
148 | * Enforce a minimum alignment for the kmalloc caches. | |
149 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
150 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
151 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
b46b8f19 DW |
152 | * alignment larger than the alignment of a 64-bit integer. |
153 | * ARCH_KMALLOC_MINALIGN allows that. | |
154 | * Note that increasing this value may disable some debug features. | |
1da177e4 | 155 | */ |
b46b8f19 | 156 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
1da177e4 LT |
157 | #endif |
158 | ||
159 | #ifndef ARCH_SLAB_MINALIGN | |
160 | /* | |
161 | * Enforce a minimum alignment for all caches. | |
162 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
163 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
164 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
165 | * some debug features. | |
166 | */ | |
167 | #define ARCH_SLAB_MINALIGN 0 | |
168 | #endif | |
169 | ||
170 | #ifndef ARCH_KMALLOC_FLAGS | |
171 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
172 | #endif | |
173 | ||
174 | /* Legal flag mask for kmem_cache_create(). */ | |
175 | #if DEBUG | |
50953fe9 | 176 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 177 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 178 | SLAB_CACHE_DMA | \ |
5af60839 | 179 | SLAB_STORE_USER | \ |
1da177e4 | 180 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 181 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
d5cff635 | 182 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE) |
1da177e4 | 183 | #else |
ac2b898c | 184 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 185 | SLAB_CACHE_DMA | \ |
1da177e4 | 186 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 187 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
d5cff635 | 188 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE) |
1da177e4 LT |
189 | #endif |
190 | ||
191 | /* | |
192 | * kmem_bufctl_t: | |
193 | * | |
194 | * Bufctl's are used for linking objs within a slab | |
195 | * linked offsets. | |
196 | * | |
197 | * This implementation relies on "struct page" for locating the cache & | |
198 | * slab an object belongs to. | |
199 | * This allows the bufctl structure to be small (one int), but limits | |
200 | * the number of objects a slab (not a cache) can contain when off-slab | |
201 | * bufctls are used. The limit is the size of the largest general cache | |
202 | * that does not use off-slab slabs. | |
203 | * For 32bit archs with 4 kB pages, is this 56. | |
204 | * This is not serious, as it is only for large objects, when it is unwise | |
205 | * to have too many per slab. | |
206 | * Note: This limit can be raised by introducing a general cache whose size | |
207 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
208 | */ | |
209 | ||
fa5b08d5 | 210 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
211 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
212 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
213 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
214 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 215 | |
1da177e4 LT |
216 | /* |
217 | * struct slab | |
218 | * | |
219 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
220 | * for a slab, or allocated from an general cache. | |
221 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
222 | */ | |
223 | struct slab { | |
b28a02de PE |
224 | struct list_head list; |
225 | unsigned long colouroff; | |
226 | void *s_mem; /* including colour offset */ | |
227 | unsigned int inuse; /* num of objs active in slab */ | |
228 | kmem_bufctl_t free; | |
229 | unsigned short nodeid; | |
1da177e4 LT |
230 | }; |
231 | ||
232 | /* | |
233 | * struct slab_rcu | |
234 | * | |
235 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
236 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
237 | * we need to approach a kernel structure obliquely, from its address | |
238 | * obtained without the usual locking. We can lock the structure to | |
239 | * stabilize it and check it's still at the given address, only if we | |
240 | * can be sure that the memory has not been meanwhile reused for some | |
241 | * other kind of object (which our subsystem's lock might corrupt). | |
242 | * | |
243 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
244 | * taking the spinlock within the structure expected at that address. | |
245 | * | |
246 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
247 | */ | |
248 | struct slab_rcu { | |
b28a02de | 249 | struct rcu_head head; |
343e0d7a | 250 | struct kmem_cache *cachep; |
b28a02de | 251 | void *addr; |
1da177e4 LT |
252 | }; |
253 | ||
254 | /* | |
255 | * struct array_cache | |
256 | * | |
1da177e4 LT |
257 | * Purpose: |
258 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
259 | * - reduce the number of linked list operations | |
260 | * - reduce spinlock operations | |
261 | * | |
262 | * The limit is stored in the per-cpu structure to reduce the data cache | |
263 | * footprint. | |
264 | * | |
265 | */ | |
266 | struct array_cache { | |
267 | unsigned int avail; | |
268 | unsigned int limit; | |
269 | unsigned int batchcount; | |
270 | unsigned int touched; | |
e498be7d | 271 | spinlock_t lock; |
bda5b655 | 272 | void *entry[]; /* |
a737b3e2 AM |
273 | * Must have this definition in here for the proper |
274 | * alignment of array_cache. Also simplifies accessing | |
275 | * the entries. | |
a737b3e2 | 276 | */ |
1da177e4 LT |
277 | }; |
278 | ||
a737b3e2 AM |
279 | /* |
280 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
281 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
282 | */ |
283 | #define BOOT_CPUCACHE_ENTRIES 1 | |
284 | struct arraycache_init { | |
285 | struct array_cache cache; | |
b28a02de | 286 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
287 | }; |
288 | ||
289 | /* | |
e498be7d | 290 | * The slab lists for all objects. |
1da177e4 LT |
291 | */ |
292 | struct kmem_list3 { | |
b28a02de PE |
293 | struct list_head slabs_partial; /* partial list first, better asm code */ |
294 | struct list_head slabs_full; | |
295 | struct list_head slabs_free; | |
296 | unsigned long free_objects; | |
b28a02de | 297 | unsigned int free_limit; |
2e1217cf | 298 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
299 | spinlock_t list_lock; |
300 | struct array_cache *shared; /* shared per node */ | |
301 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
302 | unsigned long next_reap; /* updated without locking */ |
303 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
304 | }; |
305 | ||
e498be7d CL |
306 | /* |
307 | * Need this for bootstrapping a per node allocator. | |
308 | */ | |
556a169d | 309 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
e498be7d CL |
310 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
311 | #define CACHE_CACHE 0 | |
556a169d PE |
312 | #define SIZE_AC MAX_NUMNODES |
313 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 314 | |
ed11d9eb CL |
315 | static int drain_freelist(struct kmem_cache *cache, |
316 | struct kmem_list3 *l3, int tofree); | |
317 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
318 | int node); | |
83b519e8 | 319 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 320 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 321 | |
e498be7d | 322 | /* |
a737b3e2 AM |
323 | * This function must be completely optimized away if a constant is passed to |
324 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 325 | */ |
7243cc05 | 326 | static __always_inline int index_of(const size_t size) |
e498be7d | 327 | { |
5ec8a847 SR |
328 | extern void __bad_size(void); |
329 | ||
e498be7d CL |
330 | if (__builtin_constant_p(size)) { |
331 | int i = 0; | |
332 | ||
333 | #define CACHE(x) \ | |
334 | if (size <=x) \ | |
335 | return i; \ | |
336 | else \ | |
337 | i++; | |
1c61fc40 | 338 | #include <linux/kmalloc_sizes.h> |
e498be7d | 339 | #undef CACHE |
5ec8a847 | 340 | __bad_size(); |
7243cc05 | 341 | } else |
5ec8a847 | 342 | __bad_size(); |
e498be7d CL |
343 | return 0; |
344 | } | |
345 | ||
e0a42726 IM |
346 | static int slab_early_init = 1; |
347 | ||
e498be7d CL |
348 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
349 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 350 | |
5295a74c | 351 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
352 | { |
353 | INIT_LIST_HEAD(&parent->slabs_full); | |
354 | INIT_LIST_HEAD(&parent->slabs_partial); | |
355 | INIT_LIST_HEAD(&parent->slabs_free); | |
356 | parent->shared = NULL; | |
357 | parent->alien = NULL; | |
2e1217cf | 358 | parent->colour_next = 0; |
e498be7d CL |
359 | spin_lock_init(&parent->list_lock); |
360 | parent->free_objects = 0; | |
361 | parent->free_touched = 0; | |
362 | } | |
363 | ||
a737b3e2 AM |
364 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
365 | do { \ | |
366 | INIT_LIST_HEAD(listp); \ | |
367 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
368 | } while (0) |
369 | ||
a737b3e2 AM |
370 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
371 | do { \ | |
e498be7d CL |
372 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
373 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
374 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
375 | } while (0) | |
1da177e4 LT |
376 | |
377 | /* | |
343e0d7a | 378 | * struct kmem_cache |
1da177e4 LT |
379 | * |
380 | * manages a cache. | |
381 | */ | |
b28a02de | 382 | |
2109a2d1 | 383 | struct kmem_cache { |
1da177e4 | 384 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 385 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 386 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
387 | unsigned int batchcount; |
388 | unsigned int limit; | |
389 | unsigned int shared; | |
b5d8ca7c | 390 | |
3dafccf2 | 391 | unsigned int buffer_size; |
6a2d7a95 | 392 | u32 reciprocal_buffer_size; |
b5d8ca7c | 393 | /* 3) touched by every alloc & free from the backend */ |
b5d8ca7c | 394 | |
a737b3e2 AM |
395 | unsigned int flags; /* constant flags */ |
396 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 397 | |
b5d8ca7c | 398 | /* 4) cache_grow/shrink */ |
1da177e4 | 399 | /* order of pgs per slab (2^n) */ |
b28a02de | 400 | unsigned int gfporder; |
1da177e4 LT |
401 | |
402 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 403 | gfp_t gfpflags; |
1da177e4 | 404 | |
a737b3e2 | 405 | size_t colour; /* cache colouring range */ |
b28a02de | 406 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 407 | struct kmem_cache *slabp_cache; |
b28a02de | 408 | unsigned int slab_size; |
a737b3e2 | 409 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
410 | |
411 | /* constructor func */ | |
51cc5068 | 412 | void (*ctor)(void *obj); |
1da177e4 | 413 | |
b5d8ca7c | 414 | /* 5) cache creation/removal */ |
b28a02de PE |
415 | const char *name; |
416 | struct list_head next; | |
1da177e4 | 417 | |
b5d8ca7c | 418 | /* 6) statistics */ |
1da177e4 | 419 | #if STATS |
b28a02de PE |
420 | unsigned long num_active; |
421 | unsigned long num_allocations; | |
422 | unsigned long high_mark; | |
423 | unsigned long grown; | |
424 | unsigned long reaped; | |
425 | unsigned long errors; | |
426 | unsigned long max_freeable; | |
427 | unsigned long node_allocs; | |
428 | unsigned long node_frees; | |
fb7faf33 | 429 | unsigned long node_overflow; |
b28a02de PE |
430 | atomic_t allochit; |
431 | atomic_t allocmiss; | |
432 | atomic_t freehit; | |
433 | atomic_t freemiss; | |
1da177e4 LT |
434 | #endif |
435 | #if DEBUG | |
3dafccf2 MS |
436 | /* |
437 | * If debugging is enabled, then the allocator can add additional | |
438 | * fields and/or padding to every object. buffer_size contains the total | |
439 | * object size including these internal fields, the following two | |
440 | * variables contain the offset to the user object and its size. | |
441 | */ | |
442 | int obj_offset; | |
443 | int obj_size; | |
1da177e4 | 444 | #endif |
8da3430d ED |
445 | /* |
446 | * We put nodelists[] at the end of kmem_cache, because we want to size | |
447 | * this array to nr_node_ids slots instead of MAX_NUMNODES | |
448 | * (see kmem_cache_init()) | |
449 | * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache | |
450 | * is statically defined, so we reserve the max number of nodes. | |
451 | */ | |
452 | struct kmem_list3 *nodelists[MAX_NUMNODES]; | |
453 | /* | |
454 | * Do not add fields after nodelists[] | |
455 | */ | |
1da177e4 LT |
456 | }; |
457 | ||
458 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
459 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
460 | ||
461 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
462 | /* |
463 | * Optimization question: fewer reaps means less probability for unnessary | |
464 | * cpucache drain/refill cycles. | |
1da177e4 | 465 | * |
dc6f3f27 | 466 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
467 | * which could lock up otherwise freeable slabs. |
468 | */ | |
469 | #define REAPTIMEOUT_CPUC (2*HZ) | |
470 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
471 | ||
472 | #if STATS | |
473 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
474 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
475 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
476 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 477 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
478 | #define STATS_SET_HIGH(x) \ |
479 | do { \ | |
480 | if ((x)->num_active > (x)->high_mark) \ | |
481 | (x)->high_mark = (x)->num_active; \ | |
482 | } while (0) | |
1da177e4 LT |
483 | #define STATS_INC_ERR(x) ((x)->errors++) |
484 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 485 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 486 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
487 | #define STATS_SET_FREEABLE(x, i) \ |
488 | do { \ | |
489 | if ((x)->max_freeable < i) \ | |
490 | (x)->max_freeable = i; \ | |
491 | } while (0) | |
1da177e4 LT |
492 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
493 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
494 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
495 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
496 | #else | |
497 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
498 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
499 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
500 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 501 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
502 | #define STATS_SET_HIGH(x) do { } while (0) |
503 | #define STATS_INC_ERR(x) do { } while (0) | |
504 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 505 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 506 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 507 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
508 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
509 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
510 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
511 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
512 | #endif | |
513 | ||
514 | #if DEBUG | |
1da177e4 | 515 | |
a737b3e2 AM |
516 | /* |
517 | * memory layout of objects: | |
1da177e4 | 518 | * 0 : objp |
3dafccf2 | 519 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
520 | * the end of an object is aligned with the end of the real |
521 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 522 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 523 | * redzone word. |
3dafccf2 MS |
524 | * cachep->obj_offset: The real object. |
525 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
526 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
527 | * [BYTES_PER_WORD long] | |
1da177e4 | 528 | */ |
343e0d7a | 529 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 530 | { |
3dafccf2 | 531 | return cachep->obj_offset; |
1da177e4 LT |
532 | } |
533 | ||
343e0d7a | 534 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 535 | { |
3dafccf2 | 536 | return cachep->obj_size; |
1da177e4 LT |
537 | } |
538 | ||
b46b8f19 | 539 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
540 | { |
541 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
542 | return (unsigned long long*) (objp + obj_offset(cachep) - |
543 | sizeof(unsigned long long)); | |
1da177e4 LT |
544 | } |
545 | ||
b46b8f19 | 546 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
547 | { |
548 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
549 | if (cachep->flags & SLAB_STORE_USER) | |
b46b8f19 DW |
550 | return (unsigned long long *)(objp + cachep->buffer_size - |
551 | sizeof(unsigned long long) - | |
87a927c7 | 552 | REDZONE_ALIGN); |
b46b8f19 DW |
553 | return (unsigned long long *) (objp + cachep->buffer_size - |
554 | sizeof(unsigned long long)); | |
1da177e4 LT |
555 | } |
556 | ||
343e0d7a | 557 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
558 | { |
559 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 560 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
561 | } |
562 | ||
563 | #else | |
564 | ||
3dafccf2 MS |
565 | #define obj_offset(x) 0 |
566 | #define obj_size(cachep) (cachep->buffer_size) | |
b46b8f19 DW |
567 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
568 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
569 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
570 | ||
571 | #endif | |
572 | ||
36555751 EGM |
573 | #ifdef CONFIG_KMEMTRACE |
574 | size_t slab_buffer_size(struct kmem_cache *cachep) | |
575 | { | |
576 | return cachep->buffer_size; | |
577 | } | |
578 | EXPORT_SYMBOL(slab_buffer_size); | |
579 | #endif | |
580 | ||
1da177e4 LT |
581 | /* |
582 | * Do not go above this order unless 0 objects fit into the slab. | |
583 | */ | |
584 | #define BREAK_GFP_ORDER_HI 1 | |
585 | #define BREAK_GFP_ORDER_LO 0 | |
586 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
587 | ||
a737b3e2 AM |
588 | /* |
589 | * Functions for storing/retrieving the cachep and or slab from the page | |
590 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
591 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 592 | */ |
065d41cb PE |
593 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
594 | { | |
595 | page->lru.next = (struct list_head *)cache; | |
596 | } | |
597 | ||
598 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
599 | { | |
d85f3385 | 600 | page = compound_head(page); |
ddc2e812 | 601 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
602 | return (struct kmem_cache *)page->lru.next; |
603 | } | |
604 | ||
605 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
606 | { | |
607 | page->lru.prev = (struct list_head *)slab; | |
608 | } | |
609 | ||
610 | static inline struct slab *page_get_slab(struct page *page) | |
611 | { | |
ddc2e812 | 612 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
613 | return (struct slab *)page->lru.prev; |
614 | } | |
1da177e4 | 615 | |
6ed5eb22 PE |
616 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
617 | { | |
b49af68f | 618 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
619 | return page_get_cache(page); |
620 | } | |
621 | ||
622 | static inline struct slab *virt_to_slab(const void *obj) | |
623 | { | |
b49af68f | 624 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
625 | return page_get_slab(page); |
626 | } | |
627 | ||
8fea4e96 PE |
628 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
629 | unsigned int idx) | |
630 | { | |
631 | return slab->s_mem + cache->buffer_size * idx; | |
632 | } | |
633 | ||
6a2d7a95 ED |
634 | /* |
635 | * We want to avoid an expensive divide : (offset / cache->buffer_size) | |
636 | * Using the fact that buffer_size is a constant for a particular cache, | |
637 | * we can replace (offset / cache->buffer_size) by | |
638 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
639 | */ | |
640 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
641 | const struct slab *slab, void *obj) | |
8fea4e96 | 642 | { |
6a2d7a95 ED |
643 | u32 offset = (obj - slab->s_mem); |
644 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
645 | } |
646 | ||
a737b3e2 AM |
647 | /* |
648 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
649 | */ | |
1da177e4 LT |
650 | struct cache_sizes malloc_sizes[] = { |
651 | #define CACHE(x) { .cs_size = (x) }, | |
652 | #include <linux/kmalloc_sizes.h> | |
653 | CACHE(ULONG_MAX) | |
654 | #undef CACHE | |
655 | }; | |
656 | EXPORT_SYMBOL(malloc_sizes); | |
657 | ||
658 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
659 | struct cache_names { | |
660 | char *name; | |
661 | char *name_dma; | |
662 | }; | |
663 | ||
664 | static struct cache_names __initdata cache_names[] = { | |
665 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
666 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 667 | {NULL,} |
1da177e4 LT |
668 | #undef CACHE |
669 | }; | |
670 | ||
671 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 672 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 673 | static struct arraycache_init initarray_generic = |
b28a02de | 674 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
675 | |
676 | /* internal cache of cache description objs */ | |
343e0d7a | 677 | static struct kmem_cache cache_cache = { |
b28a02de PE |
678 | .batchcount = 1, |
679 | .limit = BOOT_CPUCACHE_ENTRIES, | |
680 | .shared = 1, | |
343e0d7a | 681 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 682 | .name = "kmem_cache", |
1da177e4 LT |
683 | }; |
684 | ||
056c6241 RT |
685 | #define BAD_ALIEN_MAGIC 0x01020304ul |
686 | ||
f1aaee53 AV |
687 | #ifdef CONFIG_LOCKDEP |
688 | ||
689 | /* | |
690 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
691 | * for other slabs "off slab". | |
692 | * The locking for this is tricky in that it nests within the locks | |
693 | * of all other slabs in a few places; to deal with this special | |
694 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
695 | * |
696 | * We set lock class for alien array caches which are up during init. | |
697 | * The lock annotation will be lost if all cpus of a node goes down and | |
698 | * then comes back up during hotplug | |
f1aaee53 | 699 | */ |
056c6241 RT |
700 | static struct lock_class_key on_slab_l3_key; |
701 | static struct lock_class_key on_slab_alc_key; | |
702 | ||
703 | static inline void init_lock_keys(void) | |
f1aaee53 | 704 | |
f1aaee53 AV |
705 | { |
706 | int q; | |
056c6241 RT |
707 | struct cache_sizes *s = malloc_sizes; |
708 | ||
709 | while (s->cs_size != ULONG_MAX) { | |
710 | for_each_node(q) { | |
711 | struct array_cache **alc; | |
712 | int r; | |
713 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
714 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
715 | continue; | |
716 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
717 | alc = l3->alien; | |
718 | /* | |
719 | * FIXME: This check for BAD_ALIEN_MAGIC | |
720 | * should go away when common slab code is taught to | |
721 | * work even without alien caches. | |
722 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
723 | * for alloc_alien_cache, | |
724 | */ | |
725 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
726 | continue; | |
727 | for_each_node(r) { | |
728 | if (alc[r]) | |
729 | lockdep_set_class(&alc[r]->lock, | |
730 | &on_slab_alc_key); | |
731 | } | |
732 | } | |
733 | s++; | |
f1aaee53 AV |
734 | } |
735 | } | |
f1aaee53 | 736 | #else |
056c6241 | 737 | static inline void init_lock_keys(void) |
f1aaee53 AV |
738 | { |
739 | } | |
740 | #endif | |
741 | ||
8f5be20b | 742 | /* |
95402b38 | 743 | * Guard access to the cache-chain. |
8f5be20b | 744 | */ |
fc0abb14 | 745 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
746 | static struct list_head cache_chain; |
747 | ||
1da177e4 LT |
748 | /* |
749 | * chicken and egg problem: delay the per-cpu array allocation | |
750 | * until the general caches are up. | |
751 | */ | |
752 | static enum { | |
753 | NONE, | |
e498be7d CL |
754 | PARTIAL_AC, |
755 | PARTIAL_L3, | |
1da177e4 LT |
756 | FULL |
757 | } g_cpucache_up; | |
758 | ||
39d24e64 MK |
759 | /* |
760 | * used by boot code to determine if it can use slab based allocator | |
761 | */ | |
762 | int slab_is_available(void) | |
763 | { | |
764 | return g_cpucache_up == FULL; | |
765 | } | |
766 | ||
52bad64d | 767 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 768 | |
343e0d7a | 769 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
770 | { |
771 | return cachep->array[smp_processor_id()]; | |
772 | } | |
773 | ||
a737b3e2 AM |
774 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
775 | gfp_t gfpflags) | |
1da177e4 LT |
776 | { |
777 | struct cache_sizes *csizep = malloc_sizes; | |
778 | ||
779 | #if DEBUG | |
780 | /* This happens if someone tries to call | |
b28a02de PE |
781 | * kmem_cache_create(), or __kmalloc(), before |
782 | * the generic caches are initialized. | |
783 | */ | |
c7e43c78 | 784 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 785 | #endif |
6cb8f913 CL |
786 | if (!size) |
787 | return ZERO_SIZE_PTR; | |
788 | ||
1da177e4 LT |
789 | while (size > csizep->cs_size) |
790 | csizep++; | |
791 | ||
792 | /* | |
0abf40c1 | 793 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
794 | * has cs_{dma,}cachep==NULL. Thus no special case |
795 | * for large kmalloc calls required. | |
796 | */ | |
4b51d669 | 797 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
798 | if (unlikely(gfpflags & GFP_DMA)) |
799 | return csizep->cs_dmacachep; | |
4b51d669 | 800 | #endif |
1da177e4 LT |
801 | return csizep->cs_cachep; |
802 | } | |
803 | ||
b221385b | 804 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
805 | { |
806 | return __find_general_cachep(size, gfpflags); | |
807 | } | |
97e2bde4 | 808 | |
fbaccacf | 809 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 810 | { |
fbaccacf SR |
811 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
812 | } | |
1da177e4 | 813 | |
a737b3e2 AM |
814 | /* |
815 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
816 | */ | |
fbaccacf SR |
817 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
818 | size_t align, int flags, size_t *left_over, | |
819 | unsigned int *num) | |
820 | { | |
821 | int nr_objs; | |
822 | size_t mgmt_size; | |
823 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 824 | |
fbaccacf SR |
825 | /* |
826 | * The slab management structure can be either off the slab or | |
827 | * on it. For the latter case, the memory allocated for a | |
828 | * slab is used for: | |
829 | * | |
830 | * - The struct slab | |
831 | * - One kmem_bufctl_t for each object | |
832 | * - Padding to respect alignment of @align | |
833 | * - @buffer_size bytes for each object | |
834 | * | |
835 | * If the slab management structure is off the slab, then the | |
836 | * alignment will already be calculated into the size. Because | |
837 | * the slabs are all pages aligned, the objects will be at the | |
838 | * correct alignment when allocated. | |
839 | */ | |
840 | if (flags & CFLGS_OFF_SLAB) { | |
841 | mgmt_size = 0; | |
842 | nr_objs = slab_size / buffer_size; | |
843 | ||
844 | if (nr_objs > SLAB_LIMIT) | |
845 | nr_objs = SLAB_LIMIT; | |
846 | } else { | |
847 | /* | |
848 | * Ignore padding for the initial guess. The padding | |
849 | * is at most @align-1 bytes, and @buffer_size is at | |
850 | * least @align. In the worst case, this result will | |
851 | * be one greater than the number of objects that fit | |
852 | * into the memory allocation when taking the padding | |
853 | * into account. | |
854 | */ | |
855 | nr_objs = (slab_size - sizeof(struct slab)) / | |
856 | (buffer_size + sizeof(kmem_bufctl_t)); | |
857 | ||
858 | /* | |
859 | * This calculated number will be either the right | |
860 | * amount, or one greater than what we want. | |
861 | */ | |
862 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
863 | > slab_size) | |
864 | nr_objs--; | |
865 | ||
866 | if (nr_objs > SLAB_LIMIT) | |
867 | nr_objs = SLAB_LIMIT; | |
868 | ||
869 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
870 | } | |
871 | *num = nr_objs; | |
872 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
873 | } |
874 | ||
d40cee24 | 875 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 876 | |
a737b3e2 AM |
877 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
878 | char *msg) | |
1da177e4 LT |
879 | { |
880 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 881 | function, cachep->name, msg); |
1da177e4 LT |
882 | dump_stack(); |
883 | } | |
884 | ||
3395ee05 PM |
885 | /* |
886 | * By default on NUMA we use alien caches to stage the freeing of | |
887 | * objects allocated from other nodes. This causes massive memory | |
888 | * inefficiencies when using fake NUMA setup to split memory into a | |
889 | * large number of small nodes, so it can be disabled on the command | |
890 | * line | |
891 | */ | |
892 | ||
893 | static int use_alien_caches __read_mostly = 1; | |
1807a1aa | 894 | static int numa_platform __read_mostly = 1; |
3395ee05 PM |
895 | static int __init noaliencache_setup(char *s) |
896 | { | |
897 | use_alien_caches = 0; | |
898 | return 1; | |
899 | } | |
900 | __setup("noaliencache", noaliencache_setup); | |
901 | ||
8fce4d8e CL |
902 | #ifdef CONFIG_NUMA |
903 | /* | |
904 | * Special reaping functions for NUMA systems called from cache_reap(). | |
905 | * These take care of doing round robin flushing of alien caches (containing | |
906 | * objects freed on different nodes from which they were allocated) and the | |
907 | * flushing of remote pcps by calling drain_node_pages. | |
908 | */ | |
909 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
910 | ||
911 | static void init_reap_node(int cpu) | |
912 | { | |
913 | int node; | |
914 | ||
915 | node = next_node(cpu_to_node(cpu), node_online_map); | |
916 | if (node == MAX_NUMNODES) | |
442295c9 | 917 | node = first_node(node_online_map); |
8fce4d8e | 918 | |
7f6b8876 | 919 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
920 | } |
921 | ||
922 | static void next_reap_node(void) | |
923 | { | |
924 | int node = __get_cpu_var(reap_node); | |
925 | ||
8fce4d8e CL |
926 | node = next_node(node, node_online_map); |
927 | if (unlikely(node >= MAX_NUMNODES)) | |
928 | node = first_node(node_online_map); | |
929 | __get_cpu_var(reap_node) = node; | |
930 | } | |
931 | ||
932 | #else | |
933 | #define init_reap_node(cpu) do { } while (0) | |
934 | #define next_reap_node(void) do { } while (0) | |
935 | #endif | |
936 | ||
1da177e4 LT |
937 | /* |
938 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
939 | * via the workqueue/eventd. | |
940 | * Add the CPU number into the expiration time to minimize the possibility of | |
941 | * the CPUs getting into lockstep and contending for the global cache chain | |
942 | * lock. | |
943 | */ | |
897e679b | 944 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 945 | { |
52bad64d | 946 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
947 | |
948 | /* | |
949 | * When this gets called from do_initcalls via cpucache_init(), | |
950 | * init_workqueues() has already run, so keventd will be setup | |
951 | * at that time. | |
952 | */ | |
52bad64d | 953 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 954 | init_reap_node(cpu); |
65f27f38 | 955 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
956 | schedule_delayed_work_on(cpu, reap_work, |
957 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
958 | } |
959 | } | |
960 | ||
e498be7d | 961 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 962 | int batchcount, gfp_t gfp) |
1da177e4 | 963 | { |
b28a02de | 964 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
965 | struct array_cache *nc = NULL; |
966 | ||
83b519e8 | 967 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
968 | /* |
969 | * The array_cache structures contain pointers to free object. | |
970 | * However, when such objects are allocated or transfered to another | |
971 | * cache the pointers are not cleared and they could be counted as | |
972 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
973 | * not scan such objects. | |
974 | */ | |
975 | kmemleak_no_scan(nc); | |
1da177e4 LT |
976 | if (nc) { |
977 | nc->avail = 0; | |
978 | nc->limit = entries; | |
979 | nc->batchcount = batchcount; | |
980 | nc->touched = 0; | |
e498be7d | 981 | spin_lock_init(&nc->lock); |
1da177e4 LT |
982 | } |
983 | return nc; | |
984 | } | |
985 | ||
3ded175a CL |
986 | /* |
987 | * Transfer objects in one arraycache to another. | |
988 | * Locking must be handled by the caller. | |
989 | * | |
990 | * Return the number of entries transferred. | |
991 | */ | |
992 | static int transfer_objects(struct array_cache *to, | |
993 | struct array_cache *from, unsigned int max) | |
994 | { | |
995 | /* Figure out how many entries to transfer */ | |
996 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
997 | ||
998 | if (!nr) | |
999 | return 0; | |
1000 | ||
1001 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1002 | sizeof(void *) *nr); | |
1003 | ||
1004 | from->avail -= nr; | |
1005 | to->avail += nr; | |
1006 | to->touched = 1; | |
1007 | return nr; | |
1008 | } | |
1009 | ||
765c4507 CL |
1010 | #ifndef CONFIG_NUMA |
1011 | ||
1012 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1013 | #define reap_alien(cachep, l3) do { } while (0) | |
1014 | ||
83b519e8 | 1015 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1016 | { |
1017 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1018 | } | |
1019 | ||
1020 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1021 | { | |
1022 | } | |
1023 | ||
1024 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1025 | { | |
1026 | return 0; | |
1027 | } | |
1028 | ||
1029 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1030 | gfp_t flags) | |
1031 | { | |
1032 | return NULL; | |
1033 | } | |
1034 | ||
8b98c169 | 1035 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1036 | gfp_t flags, int nodeid) |
1037 | { | |
1038 | return NULL; | |
1039 | } | |
1040 | ||
1041 | #else /* CONFIG_NUMA */ | |
1042 | ||
8b98c169 | 1043 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1044 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1045 | |
83b519e8 | 1046 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1047 | { |
1048 | struct array_cache **ac_ptr; | |
8ef82866 | 1049 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1050 | int i; |
1051 | ||
1052 | if (limit > 1) | |
1053 | limit = 12; | |
83b519e8 | 1054 | ac_ptr = kmalloc_node(memsize, gfp, node); |
e498be7d CL |
1055 | if (ac_ptr) { |
1056 | for_each_node(i) { | |
1057 | if (i == node || !node_online(i)) { | |
1058 | ac_ptr[i] = NULL; | |
1059 | continue; | |
1060 | } | |
83b519e8 | 1061 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1062 | if (!ac_ptr[i]) { |
cc550def | 1063 | for (i--; i >= 0; i--) |
e498be7d CL |
1064 | kfree(ac_ptr[i]); |
1065 | kfree(ac_ptr); | |
1066 | return NULL; | |
1067 | } | |
1068 | } | |
1069 | } | |
1070 | return ac_ptr; | |
1071 | } | |
1072 | ||
5295a74c | 1073 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1074 | { |
1075 | int i; | |
1076 | ||
1077 | if (!ac_ptr) | |
1078 | return; | |
e498be7d | 1079 | for_each_node(i) |
b28a02de | 1080 | kfree(ac_ptr[i]); |
e498be7d CL |
1081 | kfree(ac_ptr); |
1082 | } | |
1083 | ||
343e0d7a | 1084 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1085 | struct array_cache *ac, int node) |
e498be7d CL |
1086 | { |
1087 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1088 | ||
1089 | if (ac->avail) { | |
1090 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1091 | /* |
1092 | * Stuff objects into the remote nodes shared array first. | |
1093 | * That way we could avoid the overhead of putting the objects | |
1094 | * into the free lists and getting them back later. | |
1095 | */ | |
693f7d36 | 1096 | if (rl3->shared) |
1097 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1098 | |
ff69416e | 1099 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1100 | ac->avail = 0; |
1101 | spin_unlock(&rl3->list_lock); | |
1102 | } | |
1103 | } | |
1104 | ||
8fce4d8e CL |
1105 | /* |
1106 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1107 | */ | |
1108 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1109 | { | |
1110 | int node = __get_cpu_var(reap_node); | |
1111 | ||
1112 | if (l3->alien) { | |
1113 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1114 | |
1115 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1116 | __drain_alien_cache(cachep, ac, node); |
1117 | spin_unlock_irq(&ac->lock); | |
1118 | } | |
1119 | } | |
1120 | } | |
1121 | ||
a737b3e2 AM |
1122 | static void drain_alien_cache(struct kmem_cache *cachep, |
1123 | struct array_cache **alien) | |
e498be7d | 1124 | { |
b28a02de | 1125 | int i = 0; |
e498be7d CL |
1126 | struct array_cache *ac; |
1127 | unsigned long flags; | |
1128 | ||
1129 | for_each_online_node(i) { | |
4484ebf1 | 1130 | ac = alien[i]; |
e498be7d CL |
1131 | if (ac) { |
1132 | spin_lock_irqsave(&ac->lock, flags); | |
1133 | __drain_alien_cache(cachep, ac, i); | |
1134 | spin_unlock_irqrestore(&ac->lock, flags); | |
1135 | } | |
1136 | } | |
1137 | } | |
729bd0b7 | 1138 | |
873623df | 1139 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1140 | { |
1141 | struct slab *slabp = virt_to_slab(objp); | |
1142 | int nodeid = slabp->nodeid; | |
1143 | struct kmem_list3 *l3; | |
1144 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1145 | int node; |
1146 | ||
1147 | node = numa_node_id(); | |
729bd0b7 PE |
1148 | |
1149 | /* | |
1150 | * Make sure we are not freeing a object from another node to the array | |
1151 | * cache on this cpu. | |
1152 | */ | |
62918a03 | 1153 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1154 | return 0; |
1155 | ||
1ca4cb24 | 1156 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1157 | STATS_INC_NODEFREES(cachep); |
1158 | if (l3->alien && l3->alien[nodeid]) { | |
1159 | alien = l3->alien[nodeid]; | |
873623df | 1160 | spin_lock(&alien->lock); |
729bd0b7 PE |
1161 | if (unlikely(alien->avail == alien->limit)) { |
1162 | STATS_INC_ACOVERFLOW(cachep); | |
1163 | __drain_alien_cache(cachep, alien, nodeid); | |
1164 | } | |
1165 | alien->entry[alien->avail++] = objp; | |
1166 | spin_unlock(&alien->lock); | |
1167 | } else { | |
1168 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1169 | free_block(cachep, &objp, 1, nodeid); | |
1170 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1171 | } | |
1172 | return 1; | |
1173 | } | |
e498be7d CL |
1174 | #endif |
1175 | ||
fbf1e473 AM |
1176 | static void __cpuinit cpuup_canceled(long cpu) |
1177 | { | |
1178 | struct kmem_cache *cachep; | |
1179 | struct kmem_list3 *l3 = NULL; | |
1180 | int node = cpu_to_node(cpu); | |
a70f7302 | 1181 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 AM |
1182 | |
1183 | list_for_each_entry(cachep, &cache_chain, next) { | |
1184 | struct array_cache *nc; | |
1185 | struct array_cache *shared; | |
1186 | struct array_cache **alien; | |
fbf1e473 | 1187 | |
fbf1e473 AM |
1188 | /* cpu is dead; no one can alloc from it. */ |
1189 | nc = cachep->array[cpu]; | |
1190 | cachep->array[cpu] = NULL; | |
1191 | l3 = cachep->nodelists[node]; | |
1192 | ||
1193 | if (!l3) | |
1194 | goto free_array_cache; | |
1195 | ||
1196 | spin_lock_irq(&l3->list_lock); | |
1197 | ||
1198 | /* Free limit for this kmem_list3 */ | |
1199 | l3->free_limit -= cachep->batchcount; | |
1200 | if (nc) | |
1201 | free_block(cachep, nc->entry, nc->avail, node); | |
1202 | ||
c5f59f08 | 1203 | if (!cpus_empty(*mask)) { |
fbf1e473 AM |
1204 | spin_unlock_irq(&l3->list_lock); |
1205 | goto free_array_cache; | |
1206 | } | |
1207 | ||
1208 | shared = l3->shared; | |
1209 | if (shared) { | |
1210 | free_block(cachep, shared->entry, | |
1211 | shared->avail, node); | |
1212 | l3->shared = NULL; | |
1213 | } | |
1214 | ||
1215 | alien = l3->alien; | |
1216 | l3->alien = NULL; | |
1217 | ||
1218 | spin_unlock_irq(&l3->list_lock); | |
1219 | ||
1220 | kfree(shared); | |
1221 | if (alien) { | |
1222 | drain_alien_cache(cachep, alien); | |
1223 | free_alien_cache(alien); | |
1224 | } | |
1225 | free_array_cache: | |
1226 | kfree(nc); | |
1227 | } | |
1228 | /* | |
1229 | * In the previous loop, all the objects were freed to | |
1230 | * the respective cache's slabs, now we can go ahead and | |
1231 | * shrink each nodelist to its limit. | |
1232 | */ | |
1233 | list_for_each_entry(cachep, &cache_chain, next) { | |
1234 | l3 = cachep->nodelists[node]; | |
1235 | if (!l3) | |
1236 | continue; | |
1237 | drain_freelist(cachep, l3, l3->free_objects); | |
1238 | } | |
1239 | } | |
1240 | ||
1241 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1242 | { |
343e0d7a | 1243 | struct kmem_cache *cachep; |
e498be7d CL |
1244 | struct kmem_list3 *l3 = NULL; |
1245 | int node = cpu_to_node(cpu); | |
ea02e3dd | 1246 | const int memsize = sizeof(struct kmem_list3); |
1da177e4 | 1247 | |
fbf1e473 AM |
1248 | /* |
1249 | * We need to do this right in the beginning since | |
1250 | * alloc_arraycache's are going to use this list. | |
1251 | * kmalloc_node allows us to add the slab to the right | |
1252 | * kmem_list3 and not this cpu's kmem_list3 | |
1253 | */ | |
1254 | ||
1255 | list_for_each_entry(cachep, &cache_chain, next) { | |
a737b3e2 | 1256 | /* |
fbf1e473 AM |
1257 | * Set up the size64 kmemlist for cpu before we can |
1258 | * begin anything. Make sure some other cpu on this | |
1259 | * node has not already allocated this | |
e498be7d | 1260 | */ |
fbf1e473 AM |
1261 | if (!cachep->nodelists[node]) { |
1262 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1263 | if (!l3) | |
1264 | goto bad; | |
1265 | kmem_list3_init(l3); | |
1266 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1267 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1268 | |
a737b3e2 | 1269 | /* |
fbf1e473 AM |
1270 | * The l3s don't come and go as CPUs come and |
1271 | * go. cache_chain_mutex is sufficient | |
1272 | * protection here. | |
e498be7d | 1273 | */ |
fbf1e473 | 1274 | cachep->nodelists[node] = l3; |
e498be7d CL |
1275 | } |
1276 | ||
fbf1e473 AM |
1277 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1278 | cachep->nodelists[node]->free_limit = | |
1279 | (1 + nr_cpus_node(node)) * | |
1280 | cachep->batchcount + cachep->num; | |
1281 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1282 | } | |
1283 | ||
1284 | /* | |
1285 | * Now we can go ahead with allocating the shared arrays and | |
1286 | * array caches | |
1287 | */ | |
1288 | list_for_each_entry(cachep, &cache_chain, next) { | |
1289 | struct array_cache *nc; | |
1290 | struct array_cache *shared = NULL; | |
1291 | struct array_cache **alien = NULL; | |
1292 | ||
1293 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1294 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1295 | if (!nc) |
1296 | goto bad; | |
1297 | if (cachep->shared) { | |
1298 | shared = alloc_arraycache(node, | |
1299 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1300 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1301 | if (!shared) { |
1302 | kfree(nc); | |
1da177e4 | 1303 | goto bad; |
12d00f6a | 1304 | } |
fbf1e473 AM |
1305 | } |
1306 | if (use_alien_caches) { | |
83b519e8 | 1307 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1308 | if (!alien) { |
1309 | kfree(shared); | |
1310 | kfree(nc); | |
fbf1e473 | 1311 | goto bad; |
12d00f6a | 1312 | } |
fbf1e473 AM |
1313 | } |
1314 | cachep->array[cpu] = nc; | |
1315 | l3 = cachep->nodelists[node]; | |
1316 | BUG_ON(!l3); | |
1317 | ||
1318 | spin_lock_irq(&l3->list_lock); | |
1319 | if (!l3->shared) { | |
1320 | /* | |
1321 | * We are serialised from CPU_DEAD or | |
1322 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1323 | */ | |
1324 | l3->shared = shared; | |
1325 | shared = NULL; | |
1326 | } | |
4484ebf1 | 1327 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1328 | if (!l3->alien) { |
1329 | l3->alien = alien; | |
1330 | alien = NULL; | |
1da177e4 | 1331 | } |
fbf1e473 AM |
1332 | #endif |
1333 | spin_unlock_irq(&l3->list_lock); | |
1334 | kfree(shared); | |
1335 | free_alien_cache(alien); | |
1336 | } | |
1337 | return 0; | |
1338 | bad: | |
12d00f6a | 1339 | cpuup_canceled(cpu); |
fbf1e473 AM |
1340 | return -ENOMEM; |
1341 | } | |
1342 | ||
1343 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1344 | unsigned long action, void *hcpu) | |
1345 | { | |
1346 | long cpu = (long)hcpu; | |
1347 | int err = 0; | |
1348 | ||
1349 | switch (action) { | |
fbf1e473 AM |
1350 | case CPU_UP_PREPARE: |
1351 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1352 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1353 | err = cpuup_prepare(cpu); |
95402b38 | 1354 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1355 | break; |
1356 | case CPU_ONLINE: | |
8bb78442 | 1357 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1358 | start_cpu_timer(cpu); |
1359 | break; | |
1360 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1361 | case CPU_DOWN_PREPARE: |
8bb78442 | 1362 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1363 | /* |
1364 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1365 | * held so that if cache_reap() is invoked it cannot do | |
1366 | * anything expensive but will only modify reap_work | |
1367 | * and reschedule the timer. | |
1368 | */ | |
1369 | cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); | |
1370 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1371 | per_cpu(reap_work, cpu).work.func = NULL; | |
1372 | break; | |
1373 | case CPU_DOWN_FAILED: | |
8bb78442 | 1374 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1375 | start_cpu_timer(cpu); |
1376 | break; | |
1da177e4 | 1377 | case CPU_DEAD: |
8bb78442 | 1378 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1379 | /* |
1380 | * Even if all the cpus of a node are down, we don't free the | |
1381 | * kmem_list3 of any cache. This to avoid a race between | |
1382 | * cpu_down, and a kmalloc allocation from another cpu for | |
1383 | * memory from the node of the cpu going down. The list3 | |
1384 | * structure is usually allocated from kmem_cache_create() and | |
1385 | * gets destroyed at kmem_cache_destroy(). | |
1386 | */ | |
183ff22b | 1387 | /* fall through */ |
8f5be20b | 1388 | #endif |
1da177e4 | 1389 | case CPU_UP_CANCELED: |
8bb78442 | 1390 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1391 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1392 | cpuup_canceled(cpu); |
fc0abb14 | 1393 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1394 | break; |
1da177e4 | 1395 | } |
fbf1e473 | 1396 | return err ? NOTIFY_BAD : NOTIFY_OK; |
1da177e4 LT |
1397 | } |
1398 | ||
74b85f37 CS |
1399 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1400 | &cpuup_callback, NULL, 0 | |
1401 | }; | |
1da177e4 | 1402 | |
e498be7d CL |
1403 | /* |
1404 | * swap the static kmem_list3 with kmalloced memory | |
1405 | */ | |
a737b3e2 AM |
1406 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1407 | int nodeid) | |
e498be7d CL |
1408 | { |
1409 | struct kmem_list3 *ptr; | |
1410 | ||
83b519e8 | 1411 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1412 | BUG_ON(!ptr); |
1413 | ||
e498be7d | 1414 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1415 | /* |
1416 | * Do not assume that spinlocks can be initialized via memcpy: | |
1417 | */ | |
1418 | spin_lock_init(&ptr->list_lock); | |
1419 | ||
e498be7d CL |
1420 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1421 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1422 | } |
1423 | ||
556a169d PE |
1424 | /* |
1425 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1426 | * size of kmem_list3. | |
1427 | */ | |
1428 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1429 | { | |
1430 | int node; | |
1431 | ||
1432 | for_each_online_node(node) { | |
1433 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1434 | cachep->nodelists[node]->next_reap = jiffies + | |
1435 | REAPTIMEOUT_LIST3 + | |
1436 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1437 | } | |
1438 | } | |
1439 | ||
a737b3e2 AM |
1440 | /* |
1441 | * Initialisation. Called after the page allocator have been initialised and | |
1442 | * before smp_init(). | |
1da177e4 LT |
1443 | */ |
1444 | void __init kmem_cache_init(void) | |
1445 | { | |
1446 | size_t left_over; | |
1447 | struct cache_sizes *sizes; | |
1448 | struct cache_names *names; | |
e498be7d | 1449 | int i; |
07ed76b2 | 1450 | int order; |
1ca4cb24 | 1451 | int node; |
e498be7d | 1452 | |
1807a1aa | 1453 | if (num_possible_nodes() == 1) { |
62918a03 | 1454 | use_alien_caches = 0; |
1807a1aa SS |
1455 | numa_platform = 0; |
1456 | } | |
62918a03 | 1457 | |
e498be7d CL |
1458 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1459 | kmem_list3_init(&initkmem_list3[i]); | |
1460 | if (i < MAX_NUMNODES) | |
1461 | cache_cache.nodelists[i] = NULL; | |
1462 | } | |
556a169d | 1463 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1464 | |
1465 | /* | |
1466 | * Fragmentation resistance on low memory - only use bigger | |
1467 | * page orders on machines with more than 32MB of memory. | |
1468 | */ | |
1469 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1470 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1471 | ||
1da177e4 LT |
1472 | /* Bootstrap is tricky, because several objects are allocated |
1473 | * from caches that do not exist yet: | |
a737b3e2 AM |
1474 | * 1) initialize the cache_cache cache: it contains the struct |
1475 | * kmem_cache structures of all caches, except cache_cache itself: | |
1476 | * cache_cache is statically allocated. | |
e498be7d CL |
1477 | * Initially an __init data area is used for the head array and the |
1478 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1479 | * array at the end of the bootstrap. | |
1da177e4 | 1480 | * 2) Create the first kmalloc cache. |
343e0d7a | 1481 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1482 | * An __init data area is used for the head array. |
1483 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1484 | * head arrays. | |
1da177e4 LT |
1485 | * 4) Replace the __init data head arrays for cache_cache and the first |
1486 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1487 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1488 | * the other cache's with kmalloc allocated memory. | |
1489 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1490 | */ |
1491 | ||
1ca4cb24 PE |
1492 | node = numa_node_id(); |
1493 | ||
1da177e4 | 1494 | /* 1) create the cache_cache */ |
1da177e4 LT |
1495 | INIT_LIST_HEAD(&cache_chain); |
1496 | list_add(&cache_cache.next, &cache_chain); | |
1497 | cache_cache.colour_off = cache_line_size(); | |
1498 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1499 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1500 | |
8da3430d ED |
1501 | /* |
1502 | * struct kmem_cache size depends on nr_node_ids, which | |
1503 | * can be less than MAX_NUMNODES. | |
1504 | */ | |
1505 | cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + | |
1506 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1507 | #if DEBUG | |
1508 | cache_cache.obj_size = cache_cache.buffer_size; | |
1509 | #endif | |
a737b3e2 AM |
1510 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1511 | cache_line_size()); | |
6a2d7a95 ED |
1512 | cache_cache.reciprocal_buffer_size = |
1513 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1514 | |
07ed76b2 JS |
1515 | for (order = 0; order < MAX_ORDER; order++) { |
1516 | cache_estimate(order, cache_cache.buffer_size, | |
1517 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1518 | if (cache_cache.num) | |
1519 | break; | |
1520 | } | |
40094fa6 | 1521 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1522 | cache_cache.gfporder = order; |
b28a02de | 1523 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1524 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1525 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1526 | |
1527 | /* 2+3) create the kmalloc caches */ | |
1528 | sizes = malloc_sizes; | |
1529 | names = cache_names; | |
1530 | ||
a737b3e2 AM |
1531 | /* |
1532 | * Initialize the caches that provide memory for the array cache and the | |
1533 | * kmem_list3 structures first. Without this, further allocations will | |
1534 | * bug. | |
e498be7d CL |
1535 | */ |
1536 | ||
1537 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1538 | sizes[INDEX_AC].cs_size, |
1539 | ARCH_KMALLOC_MINALIGN, | |
1540 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1541 | NULL); |
e498be7d | 1542 | |
a737b3e2 | 1543 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1544 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1545 | kmem_cache_create(names[INDEX_L3].name, |
1546 | sizes[INDEX_L3].cs_size, | |
1547 | ARCH_KMALLOC_MINALIGN, | |
1548 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1549 | NULL); |
a737b3e2 | 1550 | } |
e498be7d | 1551 | |
e0a42726 IM |
1552 | slab_early_init = 0; |
1553 | ||
1da177e4 | 1554 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1555 | /* |
1556 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1557 | * This should be particularly beneficial on SMP boxes, as it |
1558 | * eliminates "false sharing". | |
1559 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1560 | * allow tighter packing of the smaller caches. |
1561 | */ | |
a737b3e2 | 1562 | if (!sizes->cs_cachep) { |
e498be7d | 1563 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1564 | sizes->cs_size, |
1565 | ARCH_KMALLOC_MINALIGN, | |
1566 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1567 | NULL); |
a737b3e2 | 1568 | } |
4b51d669 CL |
1569 | #ifdef CONFIG_ZONE_DMA |
1570 | sizes->cs_dmacachep = kmem_cache_create( | |
1571 | names->name_dma, | |
a737b3e2 AM |
1572 | sizes->cs_size, |
1573 | ARCH_KMALLOC_MINALIGN, | |
1574 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1575 | SLAB_PANIC, | |
20c2df83 | 1576 | NULL); |
4b51d669 | 1577 | #endif |
1da177e4 LT |
1578 | sizes++; |
1579 | names++; | |
1580 | } | |
1581 | /* 4) Replace the bootstrap head arrays */ | |
1582 | { | |
2b2d5493 | 1583 | struct array_cache *ptr; |
e498be7d | 1584 | |
83b519e8 | 1585 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1586 | |
9a2dba4b PE |
1587 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1588 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1589 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1590 | /* |
1591 | * Do not assume that spinlocks can be initialized via memcpy: | |
1592 | */ | |
1593 | spin_lock_init(&ptr->lock); | |
1594 | ||
1da177e4 | 1595 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1596 | |
83b519e8 | 1597 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1598 | |
9a2dba4b | 1599 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1600 | != &initarray_generic.cache); |
9a2dba4b | 1601 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1602 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1603 | /* |
1604 | * Do not assume that spinlocks can be initialized via memcpy: | |
1605 | */ | |
1606 | spin_lock_init(&ptr->lock); | |
1607 | ||
e498be7d | 1608 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1609 | ptr; |
1da177e4 | 1610 | } |
e498be7d CL |
1611 | /* 5) Replace the bootstrap kmem_list3's */ |
1612 | { | |
1ca4cb24 PE |
1613 | int nid; |
1614 | ||
9c09a95c | 1615 | for_each_online_node(nid) { |
ec1f5eee | 1616 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1617 | |
e498be7d | 1618 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1619 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1620 | |
1621 | if (INDEX_AC != INDEX_L3) { | |
1622 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1623 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1624 | } |
1625 | } | |
1626 | } | |
1da177e4 | 1627 | |
e498be7d | 1628 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1629 | { |
343e0d7a | 1630 | struct kmem_cache *cachep; |
fc0abb14 | 1631 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1632 | list_for_each_entry(cachep, &cache_chain, next) |
83b519e8 | 1633 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
2ed3a4ef | 1634 | BUG(); |
fc0abb14 | 1635 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1636 | } |
1637 | ||
056c6241 RT |
1638 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1639 | init_lock_keys(); | |
1640 | ||
1641 | ||
1da177e4 LT |
1642 | /* Done! */ |
1643 | g_cpucache_up = FULL; | |
1644 | ||
a737b3e2 AM |
1645 | /* |
1646 | * Register a cpu startup notifier callback that initializes | |
1647 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1648 | */ |
1649 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1650 | |
a737b3e2 AM |
1651 | /* |
1652 | * The reap timers are started later, with a module init call: That part | |
1653 | * of the kernel is not yet operational. | |
1da177e4 LT |
1654 | */ |
1655 | } | |
1656 | ||
1657 | static int __init cpucache_init(void) | |
1658 | { | |
1659 | int cpu; | |
1660 | ||
a737b3e2 AM |
1661 | /* |
1662 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1663 | */ |
e498be7d | 1664 | for_each_online_cpu(cpu) |
a737b3e2 | 1665 | start_cpu_timer(cpu); |
1da177e4 LT |
1666 | return 0; |
1667 | } | |
1da177e4 LT |
1668 | __initcall(cpucache_init); |
1669 | ||
1670 | /* | |
1671 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1672 | * | |
1673 | * If we requested dmaable memory, we will get it. Even if we | |
1674 | * did not request dmaable memory, we might get it, but that | |
1675 | * would be relatively rare and ignorable. | |
1676 | */ | |
343e0d7a | 1677 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1678 | { |
1679 | struct page *page; | |
e1b6aa6f | 1680 | int nr_pages; |
1da177e4 LT |
1681 | int i; |
1682 | ||
d6fef9da | 1683 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1684 | /* |
1685 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1686 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1687 | */ |
e1b6aa6f | 1688 | flags |= __GFP_COMP; |
d6fef9da | 1689 | #endif |
765c4507 | 1690 | |
3c517a61 | 1691 | flags |= cachep->gfpflags; |
e12ba74d MG |
1692 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1693 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f CH |
1694 | |
1695 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1696 | if (!page) |
1697 | return NULL; | |
1da177e4 | 1698 | |
e1b6aa6f | 1699 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1700 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1701 | add_zone_page_state(page_zone(page), |
1702 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1703 | else | |
1704 | add_zone_page_state(page_zone(page), | |
1705 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1706 | for (i = 0; i < nr_pages; i++) |
1707 | __SetPageSlab(page + i); | |
1708 | return page_address(page); | |
1da177e4 LT |
1709 | } |
1710 | ||
1711 | /* | |
1712 | * Interface to system's page release. | |
1713 | */ | |
343e0d7a | 1714 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1715 | { |
b28a02de | 1716 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1717 | struct page *page = virt_to_page(addr); |
1718 | const unsigned long nr_freed = i; | |
1719 | ||
972d1a7b CL |
1720 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1721 | sub_zone_page_state(page_zone(page), | |
1722 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1723 | else | |
1724 | sub_zone_page_state(page_zone(page), | |
1725 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1726 | while (i--) { |
f205b2fe NP |
1727 | BUG_ON(!PageSlab(page)); |
1728 | __ClearPageSlab(page); | |
1da177e4 LT |
1729 | page++; |
1730 | } | |
1da177e4 LT |
1731 | if (current->reclaim_state) |
1732 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1733 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1734 | } |
1735 | ||
1736 | static void kmem_rcu_free(struct rcu_head *head) | |
1737 | { | |
b28a02de | 1738 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1739 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1740 | |
1741 | kmem_freepages(cachep, slab_rcu->addr); | |
1742 | if (OFF_SLAB(cachep)) | |
1743 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1744 | } | |
1745 | ||
1746 | #if DEBUG | |
1747 | ||
1748 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1749 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1750 | unsigned long caller) |
1da177e4 | 1751 | { |
3dafccf2 | 1752 | int size = obj_size(cachep); |
1da177e4 | 1753 | |
3dafccf2 | 1754 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1755 | |
b28a02de | 1756 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1757 | return; |
1758 | ||
b28a02de PE |
1759 | *addr++ = 0x12345678; |
1760 | *addr++ = caller; | |
1761 | *addr++ = smp_processor_id(); | |
1762 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1763 | { |
1764 | unsigned long *sptr = &caller; | |
1765 | unsigned long svalue; | |
1766 | ||
1767 | while (!kstack_end(sptr)) { | |
1768 | svalue = *sptr++; | |
1769 | if (kernel_text_address(svalue)) { | |
b28a02de | 1770 | *addr++ = svalue; |
1da177e4 LT |
1771 | size -= sizeof(unsigned long); |
1772 | if (size <= sizeof(unsigned long)) | |
1773 | break; | |
1774 | } | |
1775 | } | |
1776 | ||
1777 | } | |
b28a02de | 1778 | *addr++ = 0x87654321; |
1da177e4 LT |
1779 | } |
1780 | #endif | |
1781 | ||
343e0d7a | 1782 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1783 | { |
3dafccf2 MS |
1784 | int size = obj_size(cachep); |
1785 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1786 | |
1787 | memset(addr, val, size); | |
b28a02de | 1788 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1789 | } |
1790 | ||
1791 | static void dump_line(char *data, int offset, int limit) | |
1792 | { | |
1793 | int i; | |
aa83aa40 DJ |
1794 | unsigned char error = 0; |
1795 | int bad_count = 0; | |
1796 | ||
1da177e4 | 1797 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1798 | for (i = 0; i < limit; i++) { |
1799 | if (data[offset + i] != POISON_FREE) { | |
1800 | error = data[offset + i]; | |
1801 | bad_count++; | |
1802 | } | |
b28a02de | 1803 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1804 | } |
1da177e4 | 1805 | printk("\n"); |
aa83aa40 DJ |
1806 | |
1807 | if (bad_count == 1) { | |
1808 | error ^= POISON_FREE; | |
1809 | if (!(error & (error - 1))) { | |
1810 | printk(KERN_ERR "Single bit error detected. Probably " | |
1811 | "bad RAM.\n"); | |
1812 | #ifdef CONFIG_X86 | |
1813 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1814 | "test tool.\n"); | |
1815 | #else | |
1816 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1817 | #endif | |
1818 | } | |
1819 | } | |
1da177e4 LT |
1820 | } |
1821 | #endif | |
1822 | ||
1823 | #if DEBUG | |
1824 | ||
343e0d7a | 1825 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1826 | { |
1827 | int i, size; | |
1828 | char *realobj; | |
1829 | ||
1830 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1831 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1832 | *dbg_redzone1(cachep, objp), |
1833 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1834 | } |
1835 | ||
1836 | if (cachep->flags & SLAB_STORE_USER) { | |
1837 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1838 | *dbg_userword(cachep, objp)); |
1da177e4 | 1839 | print_symbol("(%s)", |
a737b3e2 | 1840 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1841 | printk("\n"); |
1842 | } | |
3dafccf2 MS |
1843 | realobj = (char *)objp + obj_offset(cachep); |
1844 | size = obj_size(cachep); | |
b28a02de | 1845 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1846 | int limit; |
1847 | limit = 16; | |
b28a02de PE |
1848 | if (i + limit > size) |
1849 | limit = size - i; | |
1da177e4 LT |
1850 | dump_line(realobj, i, limit); |
1851 | } | |
1852 | } | |
1853 | ||
343e0d7a | 1854 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1855 | { |
1856 | char *realobj; | |
1857 | int size, i; | |
1858 | int lines = 0; | |
1859 | ||
3dafccf2 MS |
1860 | realobj = (char *)objp + obj_offset(cachep); |
1861 | size = obj_size(cachep); | |
1da177e4 | 1862 | |
b28a02de | 1863 | for (i = 0; i < size; i++) { |
1da177e4 | 1864 | char exp = POISON_FREE; |
b28a02de | 1865 | if (i == size - 1) |
1da177e4 LT |
1866 | exp = POISON_END; |
1867 | if (realobj[i] != exp) { | |
1868 | int limit; | |
1869 | /* Mismatch ! */ | |
1870 | /* Print header */ | |
1871 | if (lines == 0) { | |
b28a02de | 1872 | printk(KERN_ERR |
e94a40c5 DH |
1873 | "Slab corruption: %s start=%p, len=%d\n", |
1874 | cachep->name, realobj, size); | |
1da177e4 LT |
1875 | print_objinfo(cachep, objp, 0); |
1876 | } | |
1877 | /* Hexdump the affected line */ | |
b28a02de | 1878 | i = (i / 16) * 16; |
1da177e4 | 1879 | limit = 16; |
b28a02de PE |
1880 | if (i + limit > size) |
1881 | limit = size - i; | |
1da177e4 LT |
1882 | dump_line(realobj, i, limit); |
1883 | i += 16; | |
1884 | lines++; | |
1885 | /* Limit to 5 lines */ | |
1886 | if (lines > 5) | |
1887 | break; | |
1888 | } | |
1889 | } | |
1890 | if (lines != 0) { | |
1891 | /* Print some data about the neighboring objects, if they | |
1892 | * exist: | |
1893 | */ | |
6ed5eb22 | 1894 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1895 | unsigned int objnr; |
1da177e4 | 1896 | |
8fea4e96 | 1897 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1898 | if (objnr) { |
8fea4e96 | 1899 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1900 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1901 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1902 | realobj, size); |
1da177e4 LT |
1903 | print_objinfo(cachep, objp, 2); |
1904 | } | |
b28a02de | 1905 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1906 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1907 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1908 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1909 | realobj, size); |
1da177e4 LT |
1910 | print_objinfo(cachep, objp, 2); |
1911 | } | |
1912 | } | |
1913 | } | |
1914 | #endif | |
1915 | ||
12dd36fa | 1916 | #if DEBUG |
e79aec29 | 1917 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1918 | { |
1da177e4 LT |
1919 | int i; |
1920 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1921 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1922 | |
1923 | if (cachep->flags & SLAB_POISON) { | |
1924 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1925 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1926 | OFF_SLAB(cachep)) | |
b28a02de | 1927 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1928 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1929 | else |
1930 | check_poison_obj(cachep, objp); | |
1931 | #else | |
1932 | check_poison_obj(cachep, objp); | |
1933 | #endif | |
1934 | } | |
1935 | if (cachep->flags & SLAB_RED_ZONE) { | |
1936 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1937 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1938 | "was overwritten"); |
1da177e4 LT |
1939 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1940 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1941 | "was overwritten"); |
1da177e4 | 1942 | } |
1da177e4 | 1943 | } |
12dd36fa | 1944 | } |
1da177e4 | 1945 | #else |
e79aec29 | 1946 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1947 | { |
12dd36fa | 1948 | } |
1da177e4 LT |
1949 | #endif |
1950 | ||
911851e6 RD |
1951 | /** |
1952 | * slab_destroy - destroy and release all objects in a slab | |
1953 | * @cachep: cache pointer being destroyed | |
1954 | * @slabp: slab pointer being destroyed | |
1955 | * | |
12dd36fa | 1956 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1957 | * Before calling the slab must have been unlinked from the cache. The |
1958 | * cache-lock is not held/needed. | |
12dd36fa | 1959 | */ |
343e0d7a | 1960 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1961 | { |
1962 | void *addr = slabp->s_mem - slabp->colouroff; | |
1963 | ||
e79aec29 | 1964 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
1965 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1966 | struct slab_rcu *slab_rcu; | |
1967 | ||
b28a02de | 1968 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1969 | slab_rcu->cachep = cachep; |
1970 | slab_rcu->addr = addr; | |
1971 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1972 | } else { | |
1973 | kmem_freepages(cachep, addr); | |
873623df IM |
1974 | if (OFF_SLAB(cachep)) |
1975 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1976 | } |
1977 | } | |
1978 | ||
117f6eb1 CL |
1979 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1980 | { | |
1981 | int i; | |
1982 | struct kmem_list3 *l3; | |
1983 | ||
1984 | for_each_online_cpu(i) | |
1985 | kfree(cachep->array[i]); | |
1986 | ||
1987 | /* NUMA: free the list3 structures */ | |
1988 | for_each_online_node(i) { | |
1989 | l3 = cachep->nodelists[i]; | |
1990 | if (l3) { | |
1991 | kfree(l3->shared); | |
1992 | free_alien_cache(l3->alien); | |
1993 | kfree(l3); | |
1994 | } | |
1995 | } | |
1996 | kmem_cache_free(&cache_cache, cachep); | |
1997 | } | |
1998 | ||
1999 | ||
4d268eba | 2000 | /** |
a70773dd RD |
2001 | * calculate_slab_order - calculate size (page order) of slabs |
2002 | * @cachep: pointer to the cache that is being created | |
2003 | * @size: size of objects to be created in this cache. | |
2004 | * @align: required alignment for the objects. | |
2005 | * @flags: slab allocation flags | |
2006 | * | |
2007 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2008 | * |
2009 | * This could be made much more intelligent. For now, try to avoid using | |
2010 | * high order pages for slabs. When the gfp() functions are more friendly | |
2011 | * towards high-order requests, this should be changed. | |
2012 | */ | |
a737b3e2 | 2013 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2014 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2015 | { |
b1ab41c4 | 2016 | unsigned long offslab_limit; |
4d268eba | 2017 | size_t left_over = 0; |
9888e6fa | 2018 | int gfporder; |
4d268eba | 2019 | |
0aa817f0 | 2020 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2021 | unsigned int num; |
2022 | size_t remainder; | |
2023 | ||
9888e6fa | 2024 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2025 | if (!num) |
2026 | continue; | |
9888e6fa | 2027 | |
b1ab41c4 IM |
2028 | if (flags & CFLGS_OFF_SLAB) { |
2029 | /* | |
2030 | * Max number of objs-per-slab for caches which | |
2031 | * use off-slab slabs. Needed to avoid a possible | |
2032 | * looping condition in cache_grow(). | |
2033 | */ | |
2034 | offslab_limit = size - sizeof(struct slab); | |
2035 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2036 | ||
2037 | if (num > offslab_limit) | |
2038 | break; | |
2039 | } | |
4d268eba | 2040 | |
9888e6fa | 2041 | /* Found something acceptable - save it away */ |
4d268eba | 2042 | cachep->num = num; |
9888e6fa | 2043 | cachep->gfporder = gfporder; |
4d268eba PE |
2044 | left_over = remainder; |
2045 | ||
f78bb8ad LT |
2046 | /* |
2047 | * A VFS-reclaimable slab tends to have most allocations | |
2048 | * as GFP_NOFS and we really don't want to have to be allocating | |
2049 | * higher-order pages when we are unable to shrink dcache. | |
2050 | */ | |
2051 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2052 | break; | |
2053 | ||
4d268eba PE |
2054 | /* |
2055 | * Large number of objects is good, but very large slabs are | |
2056 | * currently bad for the gfp()s. | |
2057 | */ | |
9888e6fa | 2058 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2059 | break; |
2060 | ||
9888e6fa LT |
2061 | /* |
2062 | * Acceptable internal fragmentation? | |
2063 | */ | |
a737b3e2 | 2064 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2065 | break; |
2066 | } | |
2067 | return left_over; | |
2068 | } | |
2069 | ||
83b519e8 | 2070 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2071 | { |
2ed3a4ef | 2072 | if (g_cpucache_up == FULL) |
83b519e8 | 2073 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2074 | |
f30cf7d1 PE |
2075 | if (g_cpucache_up == NONE) { |
2076 | /* | |
2077 | * Note: the first kmem_cache_create must create the cache | |
2078 | * that's used by kmalloc(24), otherwise the creation of | |
2079 | * further caches will BUG(). | |
2080 | */ | |
2081 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2082 | ||
2083 | /* | |
2084 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2085 | * the first cache, then we need to set up all its list3s, | |
2086 | * otherwise the creation of further caches will BUG(). | |
2087 | */ | |
2088 | set_up_list3s(cachep, SIZE_AC); | |
2089 | if (INDEX_AC == INDEX_L3) | |
2090 | g_cpucache_up = PARTIAL_L3; | |
2091 | else | |
2092 | g_cpucache_up = PARTIAL_AC; | |
2093 | } else { | |
2094 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2095 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 PE |
2096 | |
2097 | if (g_cpucache_up == PARTIAL_AC) { | |
2098 | set_up_list3s(cachep, SIZE_L3); | |
2099 | g_cpucache_up = PARTIAL_L3; | |
2100 | } else { | |
2101 | int node; | |
556a169d | 2102 | for_each_online_node(node) { |
f30cf7d1 PE |
2103 | cachep->nodelists[node] = |
2104 | kmalloc_node(sizeof(struct kmem_list3), | |
2105 | GFP_KERNEL, node); | |
2106 | BUG_ON(!cachep->nodelists[node]); | |
2107 | kmem_list3_init(cachep->nodelists[node]); | |
2108 | } | |
2109 | } | |
2110 | } | |
2111 | cachep->nodelists[numa_node_id()]->next_reap = | |
2112 | jiffies + REAPTIMEOUT_LIST3 + | |
2113 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2114 | ||
2115 | cpu_cache_get(cachep)->avail = 0; | |
2116 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2117 | cpu_cache_get(cachep)->batchcount = 1; | |
2118 | cpu_cache_get(cachep)->touched = 0; | |
2119 | cachep->batchcount = 1; | |
2120 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2121 | return 0; |
f30cf7d1 PE |
2122 | } |
2123 | ||
1da177e4 LT |
2124 | /** |
2125 | * kmem_cache_create - Create a cache. | |
2126 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2127 | * @size: The size of objects to be created in this cache. | |
2128 | * @align: The required alignment for the objects. | |
2129 | * @flags: SLAB flags | |
2130 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2131 | * |
2132 | * Returns a ptr to the cache on success, NULL on failure. | |
2133 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2134 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2135 | * |
2136 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 | 2137 | * the module calling this has to destroy the cache before getting unloaded. |
249da166 CM |
2138 | * Note that kmem_cache_name() is not guaranteed to return the same pointer, |
2139 | * therefore applications must manage it themselves. | |
a737b3e2 | 2140 | * |
1da177e4 LT |
2141 | * The flags are |
2142 | * | |
2143 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2144 | * to catch references to uninitialised memory. | |
2145 | * | |
2146 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2147 | * for buffer overruns. | |
2148 | * | |
1da177e4 LT |
2149 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2150 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2151 | * as davem. | |
2152 | */ | |
343e0d7a | 2153 | struct kmem_cache * |
1da177e4 | 2154 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2155 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2156 | { |
2157 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2158 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2159 | gfp_t gfp; |
1da177e4 LT |
2160 | |
2161 | /* | |
2162 | * Sanity checks... these are all serious usage bugs. | |
2163 | */ | |
a737b3e2 | 2164 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2165 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2166 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2167 | name); |
b28a02de PE |
2168 | BUG(); |
2169 | } | |
1da177e4 | 2170 | |
f0188f47 | 2171 | /* |
8f5be20b | 2172 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2173 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2174 | */ |
83b519e8 PE |
2175 | if (slab_is_available()) { |
2176 | get_online_cpus(); | |
2177 | mutex_lock(&cache_chain_mutex); | |
2178 | } | |
4f12bb4f | 2179 | |
7a7c381d | 2180 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2181 | char tmp; |
2182 | int res; | |
2183 | ||
2184 | /* | |
2185 | * This happens when the module gets unloaded and doesn't | |
2186 | * destroy its slab cache and no-one else reuses the vmalloc | |
2187 | * area of the module. Print a warning. | |
2188 | */ | |
138ae663 | 2189 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2190 | if (res) { |
b4169525 | 2191 | printk(KERN_ERR |
2192 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2193 | pc->buffer_size); |
4f12bb4f AM |
2194 | continue; |
2195 | } | |
2196 | ||
b28a02de | 2197 | if (!strcmp(pc->name, name)) { |
b4169525 | 2198 | printk(KERN_ERR |
2199 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2200 | dump_stack(); |
2201 | goto oops; | |
2202 | } | |
2203 | } | |
2204 | ||
1da177e4 LT |
2205 | #if DEBUG |
2206 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2207 | #if FORCED_DEBUG |
2208 | /* | |
2209 | * Enable redzoning and last user accounting, except for caches with | |
2210 | * large objects, if the increased size would increase the object size | |
2211 | * above the next power of two: caches with object sizes just above a | |
2212 | * power of two have a significant amount of internal fragmentation. | |
2213 | */ | |
87a927c7 DW |
2214 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2215 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2216 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2217 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2218 | flags |= SLAB_POISON; | |
2219 | #endif | |
2220 | if (flags & SLAB_DESTROY_BY_RCU) | |
2221 | BUG_ON(flags & SLAB_POISON); | |
2222 | #endif | |
1da177e4 | 2223 | /* |
a737b3e2 AM |
2224 | * Always checks flags, a caller might be expecting debug support which |
2225 | * isn't available. | |
1da177e4 | 2226 | */ |
40094fa6 | 2227 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2228 | |
a737b3e2 AM |
2229 | /* |
2230 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2231 | * unaligned accesses for some archs when redzoning is used, and makes |
2232 | * sure any on-slab bufctl's are also correctly aligned. | |
2233 | */ | |
b28a02de PE |
2234 | if (size & (BYTES_PER_WORD - 1)) { |
2235 | size += (BYTES_PER_WORD - 1); | |
2236 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2237 | } |
2238 | ||
a737b3e2 AM |
2239 | /* calculate the final buffer alignment: */ |
2240 | ||
1da177e4 LT |
2241 | /* 1) arch recommendation: can be overridden for debug */ |
2242 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2243 | /* |
2244 | * Default alignment: as specified by the arch code. Except if | |
2245 | * an object is really small, then squeeze multiple objects into | |
2246 | * one cacheline. | |
1da177e4 LT |
2247 | */ |
2248 | ralign = cache_line_size(); | |
b28a02de | 2249 | while (size <= ralign / 2) |
1da177e4 LT |
2250 | ralign /= 2; |
2251 | } else { | |
2252 | ralign = BYTES_PER_WORD; | |
2253 | } | |
ca5f9703 PE |
2254 | |
2255 | /* | |
87a927c7 DW |
2256 | * Redzoning and user store require word alignment or possibly larger. |
2257 | * Note this will be overridden by architecture or caller mandated | |
2258 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2259 | */ |
87a927c7 DW |
2260 | if (flags & SLAB_STORE_USER) |
2261 | ralign = BYTES_PER_WORD; | |
2262 | ||
2263 | if (flags & SLAB_RED_ZONE) { | |
2264 | ralign = REDZONE_ALIGN; | |
2265 | /* If redzoning, ensure that the second redzone is suitably | |
2266 | * aligned, by adjusting the object size accordingly. */ | |
2267 | size += REDZONE_ALIGN - 1; | |
2268 | size &= ~(REDZONE_ALIGN - 1); | |
2269 | } | |
ca5f9703 | 2270 | |
a44b56d3 | 2271 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2272 | if (ralign < ARCH_SLAB_MINALIGN) { |
2273 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2274 | } |
a44b56d3 | 2275 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2276 | if (ralign < align) { |
2277 | ralign = align; | |
1da177e4 | 2278 | } |
a44b56d3 | 2279 | /* disable debug if necessary */ |
b46b8f19 | 2280 | if (ralign > __alignof__(unsigned long long)) |
a44b56d3 | 2281 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2282 | /* |
ca5f9703 | 2283 | * 4) Store it. |
1da177e4 LT |
2284 | */ |
2285 | align = ralign; | |
2286 | ||
83b519e8 PE |
2287 | if (slab_is_available()) |
2288 | gfp = GFP_KERNEL; | |
2289 | else | |
2290 | gfp = GFP_NOWAIT; | |
2291 | ||
1da177e4 | 2292 | /* Get cache's description obj. */ |
83b519e8 | 2293 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2294 | if (!cachep) |
4f12bb4f | 2295 | goto oops; |
1da177e4 LT |
2296 | |
2297 | #if DEBUG | |
3dafccf2 | 2298 | cachep->obj_size = size; |
1da177e4 | 2299 | |
ca5f9703 PE |
2300 | /* |
2301 | * Both debugging options require word-alignment which is calculated | |
2302 | * into align above. | |
2303 | */ | |
1da177e4 | 2304 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2305 | /* add space for red zone words */ |
b46b8f19 DW |
2306 | cachep->obj_offset += sizeof(unsigned long long); |
2307 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2308 | } |
2309 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2310 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2311 | * the real object. But if the second red zone needs to be |
2312 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2313 | */ |
87a927c7 DW |
2314 | if (flags & SLAB_RED_ZONE) |
2315 | size += REDZONE_ALIGN; | |
2316 | else | |
2317 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2318 | } |
2319 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2320 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2321 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2322 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2323 | size = PAGE_SIZE; |
2324 | } | |
2325 | #endif | |
2326 | #endif | |
2327 | ||
e0a42726 IM |
2328 | /* |
2329 | * Determine if the slab management is 'on' or 'off' slab. | |
2330 | * (bootstrapping cannot cope with offslab caches so don't do | |
2331 | * it too early on.) | |
2332 | */ | |
2333 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2334 | /* |
2335 | * Size is large, assume best to place the slab management obj | |
2336 | * off-slab (should allow better packing of objs). | |
2337 | */ | |
2338 | flags |= CFLGS_OFF_SLAB; | |
2339 | ||
2340 | size = ALIGN(size, align); | |
2341 | ||
f78bb8ad | 2342 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2343 | |
2344 | if (!cachep->num) { | |
b4169525 | 2345 | printk(KERN_ERR |
2346 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2347 | kmem_cache_free(&cache_cache, cachep); |
2348 | cachep = NULL; | |
4f12bb4f | 2349 | goto oops; |
1da177e4 | 2350 | } |
b28a02de PE |
2351 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2352 | + sizeof(struct slab), align); | |
1da177e4 LT |
2353 | |
2354 | /* | |
2355 | * If the slab has been placed off-slab, and we have enough space then | |
2356 | * move it on-slab. This is at the expense of any extra colouring. | |
2357 | */ | |
2358 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2359 | flags &= ~CFLGS_OFF_SLAB; | |
2360 | left_over -= slab_size; | |
2361 | } | |
2362 | ||
2363 | if (flags & CFLGS_OFF_SLAB) { | |
2364 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2365 | slab_size = |
2366 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2367 | } |
2368 | ||
2369 | cachep->colour_off = cache_line_size(); | |
2370 | /* Offset must be a multiple of the alignment. */ | |
2371 | if (cachep->colour_off < align) | |
2372 | cachep->colour_off = align; | |
b28a02de | 2373 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2374 | cachep->slab_size = slab_size; |
2375 | cachep->flags = flags; | |
2376 | cachep->gfpflags = 0; | |
4b51d669 | 2377 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2378 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2379 | cachep->buffer_size = size; |
6a2d7a95 | 2380 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2381 | |
e5ac9c5a | 2382 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2383 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2384 | /* |
2385 | * This is a possibility for one of the malloc_sizes caches. | |
2386 | * But since we go off slab only for object size greater than | |
2387 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2388 | * this should not happen at all. | |
2389 | * But leave a BUG_ON for some lucky dude. | |
2390 | */ | |
6cb8f913 | 2391 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2392 | } |
1da177e4 | 2393 | cachep->ctor = ctor; |
1da177e4 LT |
2394 | cachep->name = name; |
2395 | ||
83b519e8 | 2396 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef CL |
2397 | __kmem_cache_destroy(cachep); |
2398 | cachep = NULL; | |
2399 | goto oops; | |
2400 | } | |
1da177e4 | 2401 | |
1da177e4 LT |
2402 | /* cache setup completed, link it into the list */ |
2403 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2404 | oops: |
1da177e4 LT |
2405 | if (!cachep && (flags & SLAB_PANIC)) |
2406 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2407 | name); |
83b519e8 PE |
2408 | if (slab_is_available()) { |
2409 | mutex_unlock(&cache_chain_mutex); | |
2410 | put_online_cpus(); | |
2411 | } | |
1da177e4 LT |
2412 | return cachep; |
2413 | } | |
2414 | EXPORT_SYMBOL(kmem_cache_create); | |
2415 | ||
2416 | #if DEBUG | |
2417 | static void check_irq_off(void) | |
2418 | { | |
2419 | BUG_ON(!irqs_disabled()); | |
2420 | } | |
2421 | ||
2422 | static void check_irq_on(void) | |
2423 | { | |
2424 | BUG_ON(irqs_disabled()); | |
2425 | } | |
2426 | ||
343e0d7a | 2427 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2428 | { |
2429 | #ifdef CONFIG_SMP | |
2430 | check_irq_off(); | |
e498be7d | 2431 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2432 | #endif |
2433 | } | |
e498be7d | 2434 | |
343e0d7a | 2435 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2436 | { |
2437 | #ifdef CONFIG_SMP | |
2438 | check_irq_off(); | |
2439 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2440 | #endif | |
2441 | } | |
2442 | ||
1da177e4 LT |
2443 | #else |
2444 | #define check_irq_off() do { } while(0) | |
2445 | #define check_irq_on() do { } while(0) | |
2446 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2447 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2448 | #endif |
2449 | ||
aab2207c CL |
2450 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2451 | struct array_cache *ac, | |
2452 | int force, int node); | |
2453 | ||
1da177e4 LT |
2454 | static void do_drain(void *arg) |
2455 | { | |
a737b3e2 | 2456 | struct kmem_cache *cachep = arg; |
1da177e4 | 2457 | struct array_cache *ac; |
ff69416e | 2458 | int node = numa_node_id(); |
1da177e4 LT |
2459 | |
2460 | check_irq_off(); | |
9a2dba4b | 2461 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2462 | spin_lock(&cachep->nodelists[node]->list_lock); |
2463 | free_block(cachep, ac->entry, ac->avail, node); | |
2464 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2465 | ac->avail = 0; |
2466 | } | |
2467 | ||
343e0d7a | 2468 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2469 | { |
e498be7d CL |
2470 | struct kmem_list3 *l3; |
2471 | int node; | |
2472 | ||
15c8b6c1 | 2473 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2474 | check_irq_on(); |
b28a02de | 2475 | for_each_online_node(node) { |
e498be7d | 2476 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2477 | if (l3 && l3->alien) |
2478 | drain_alien_cache(cachep, l3->alien); | |
2479 | } | |
2480 | ||
2481 | for_each_online_node(node) { | |
2482 | l3 = cachep->nodelists[node]; | |
2483 | if (l3) | |
aab2207c | 2484 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2485 | } |
1da177e4 LT |
2486 | } |
2487 | ||
ed11d9eb CL |
2488 | /* |
2489 | * Remove slabs from the list of free slabs. | |
2490 | * Specify the number of slabs to drain in tofree. | |
2491 | * | |
2492 | * Returns the actual number of slabs released. | |
2493 | */ | |
2494 | static int drain_freelist(struct kmem_cache *cache, | |
2495 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2496 | { |
ed11d9eb CL |
2497 | struct list_head *p; |
2498 | int nr_freed; | |
1da177e4 | 2499 | struct slab *slabp; |
1da177e4 | 2500 | |
ed11d9eb CL |
2501 | nr_freed = 0; |
2502 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2503 | |
ed11d9eb | 2504 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2505 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2506 | if (p == &l3->slabs_free) { |
2507 | spin_unlock_irq(&l3->list_lock); | |
2508 | goto out; | |
2509 | } | |
1da177e4 | 2510 | |
ed11d9eb | 2511 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2512 | #if DEBUG |
40094fa6 | 2513 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2514 | #endif |
2515 | list_del(&slabp->list); | |
ed11d9eb CL |
2516 | /* |
2517 | * Safe to drop the lock. The slab is no longer linked | |
2518 | * to the cache. | |
2519 | */ | |
2520 | l3->free_objects -= cache->num; | |
e498be7d | 2521 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2522 | slab_destroy(cache, slabp); |
2523 | nr_freed++; | |
1da177e4 | 2524 | } |
ed11d9eb CL |
2525 | out: |
2526 | return nr_freed; | |
1da177e4 LT |
2527 | } |
2528 | ||
8f5be20b | 2529 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2530 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2531 | { |
2532 | int ret = 0, i = 0; | |
2533 | struct kmem_list3 *l3; | |
2534 | ||
2535 | drain_cpu_caches(cachep); | |
2536 | ||
2537 | check_irq_on(); | |
2538 | for_each_online_node(i) { | |
2539 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2540 | if (!l3) |
2541 | continue; | |
2542 | ||
2543 | drain_freelist(cachep, l3, l3->free_objects); | |
2544 | ||
2545 | ret += !list_empty(&l3->slabs_full) || | |
2546 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2547 | } |
2548 | return (ret ? 1 : 0); | |
2549 | } | |
2550 | ||
1da177e4 LT |
2551 | /** |
2552 | * kmem_cache_shrink - Shrink a cache. | |
2553 | * @cachep: The cache to shrink. | |
2554 | * | |
2555 | * Releases as many slabs as possible for a cache. | |
2556 | * To help debugging, a zero exit status indicates all slabs were released. | |
2557 | */ | |
343e0d7a | 2558 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2559 | { |
8f5be20b | 2560 | int ret; |
40094fa6 | 2561 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2562 | |
95402b38 | 2563 | get_online_cpus(); |
8f5be20b RT |
2564 | mutex_lock(&cache_chain_mutex); |
2565 | ret = __cache_shrink(cachep); | |
2566 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2567 | put_online_cpus(); |
8f5be20b | 2568 | return ret; |
1da177e4 LT |
2569 | } |
2570 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2571 | ||
2572 | /** | |
2573 | * kmem_cache_destroy - delete a cache | |
2574 | * @cachep: the cache to destroy | |
2575 | * | |
72fd4a35 | 2576 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2577 | * |
2578 | * It is expected this function will be called by a module when it is | |
2579 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2580 | * cache being allocated each time a module is loaded and unloaded, if the | |
2581 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2582 | * | |
2583 | * The cache must be empty before calling this function. | |
2584 | * | |
2585 | * The caller must guarantee that noone will allocate memory from the cache | |
2586 | * during the kmem_cache_destroy(). | |
2587 | */ | |
133d205a | 2588 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2589 | { |
40094fa6 | 2590 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2591 | |
1da177e4 | 2592 | /* Find the cache in the chain of caches. */ |
95402b38 | 2593 | get_online_cpus(); |
fc0abb14 | 2594 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2595 | /* |
2596 | * the chain is never empty, cache_cache is never destroyed | |
2597 | */ | |
2598 | list_del(&cachep->next); | |
1da177e4 LT |
2599 | if (__cache_shrink(cachep)) { |
2600 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2601 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2602 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2603 | put_online_cpus(); |
133d205a | 2604 | return; |
1da177e4 LT |
2605 | } |
2606 | ||
2607 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2608 | synchronize_rcu(); |
1da177e4 | 2609 | |
117f6eb1 | 2610 | __kmem_cache_destroy(cachep); |
8f5be20b | 2611 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2612 | put_online_cpus(); |
1da177e4 LT |
2613 | } |
2614 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2615 | ||
e5ac9c5a RT |
2616 | /* |
2617 | * Get the memory for a slab management obj. | |
2618 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2619 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2620 | * come from the same cache which is getting created because, | |
2621 | * when we are searching for an appropriate cache for these | |
2622 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2623 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2624 | * kmem_find_general_cachep till the initialization is complete. | |
2625 | * Hence we cannot have slabp_cache same as the original cache. | |
2626 | */ | |
343e0d7a | 2627 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2628 | int colour_off, gfp_t local_flags, |
2629 | int nodeid) | |
1da177e4 LT |
2630 | { |
2631 | struct slab *slabp; | |
b28a02de | 2632 | |
1da177e4 LT |
2633 | if (OFF_SLAB(cachep)) { |
2634 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2635 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2636 | local_flags, nodeid); |
d5cff635 CM |
2637 | /* |
2638 | * If the first object in the slab is leaked (it's allocated | |
2639 | * but no one has a reference to it), we want to make sure | |
2640 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2641 | * to the object. Otherwise we will not report the leak. | |
2642 | */ | |
2643 | kmemleak_scan_area(slabp, offsetof(struct slab, list), | |
2644 | sizeof(struct list_head), local_flags); | |
1da177e4 LT |
2645 | if (!slabp) |
2646 | return NULL; | |
2647 | } else { | |
b28a02de | 2648 | slabp = objp + colour_off; |
1da177e4 LT |
2649 | colour_off += cachep->slab_size; |
2650 | } | |
2651 | slabp->inuse = 0; | |
2652 | slabp->colouroff = colour_off; | |
b28a02de | 2653 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2654 | slabp->nodeid = nodeid; |
e51bfd0a | 2655 | slabp->free = 0; |
1da177e4 LT |
2656 | return slabp; |
2657 | } | |
2658 | ||
2659 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2660 | { | |
b28a02de | 2661 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2662 | } |
2663 | ||
343e0d7a | 2664 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2665 | struct slab *slabp) |
1da177e4 LT |
2666 | { |
2667 | int i; | |
2668 | ||
2669 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2670 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2671 | #if DEBUG |
2672 | /* need to poison the objs? */ | |
2673 | if (cachep->flags & SLAB_POISON) | |
2674 | poison_obj(cachep, objp, POISON_FREE); | |
2675 | if (cachep->flags & SLAB_STORE_USER) | |
2676 | *dbg_userword(cachep, objp) = NULL; | |
2677 | ||
2678 | if (cachep->flags & SLAB_RED_ZONE) { | |
2679 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2680 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2681 | } | |
2682 | /* | |
a737b3e2 AM |
2683 | * Constructors are not allowed to allocate memory from the same |
2684 | * cache which they are a constructor for. Otherwise, deadlock. | |
2685 | * They must also be threaded. | |
1da177e4 LT |
2686 | */ |
2687 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2688 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2689 | |
2690 | if (cachep->flags & SLAB_RED_ZONE) { | |
2691 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2692 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2693 | " end of an object"); |
1da177e4 LT |
2694 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2695 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2696 | " start of an object"); |
1da177e4 | 2697 | } |
a737b3e2 AM |
2698 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2699 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2700 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2701 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2702 | #else |
2703 | if (cachep->ctor) | |
51cc5068 | 2704 | cachep->ctor(objp); |
1da177e4 | 2705 | #endif |
b28a02de | 2706 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2707 | } |
b28a02de | 2708 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2709 | } |
2710 | ||
343e0d7a | 2711 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2712 | { |
4b51d669 CL |
2713 | if (CONFIG_ZONE_DMA_FLAG) { |
2714 | if (flags & GFP_DMA) | |
2715 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2716 | else | |
2717 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2718 | } | |
1da177e4 LT |
2719 | } |
2720 | ||
a737b3e2 AM |
2721 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2722 | int nodeid) | |
78d382d7 | 2723 | { |
8fea4e96 | 2724 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2725 | kmem_bufctl_t next; |
2726 | ||
2727 | slabp->inuse++; | |
2728 | next = slab_bufctl(slabp)[slabp->free]; | |
2729 | #if DEBUG | |
2730 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2731 | WARN_ON(slabp->nodeid != nodeid); | |
2732 | #endif | |
2733 | slabp->free = next; | |
2734 | ||
2735 | return objp; | |
2736 | } | |
2737 | ||
a737b3e2 AM |
2738 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2739 | void *objp, int nodeid) | |
78d382d7 | 2740 | { |
8fea4e96 | 2741 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2742 | |
2743 | #if DEBUG | |
2744 | /* Verify that the slab belongs to the intended node */ | |
2745 | WARN_ON(slabp->nodeid != nodeid); | |
2746 | ||
871751e2 | 2747 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2748 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2749 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2750 | BUG(); |
2751 | } | |
2752 | #endif | |
2753 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2754 | slabp->free = objnr; | |
2755 | slabp->inuse--; | |
2756 | } | |
2757 | ||
4776874f PE |
2758 | /* |
2759 | * Map pages beginning at addr to the given cache and slab. This is required | |
2760 | * for the slab allocator to be able to lookup the cache and slab of a | |
2761 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2762 | */ | |
2763 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2764 | void *addr) | |
1da177e4 | 2765 | { |
4776874f | 2766 | int nr_pages; |
1da177e4 LT |
2767 | struct page *page; |
2768 | ||
4776874f | 2769 | page = virt_to_page(addr); |
84097518 | 2770 | |
4776874f | 2771 | nr_pages = 1; |
84097518 | 2772 | if (likely(!PageCompound(page))) |
4776874f PE |
2773 | nr_pages <<= cache->gfporder; |
2774 | ||
1da177e4 | 2775 | do { |
4776874f PE |
2776 | page_set_cache(page, cache); |
2777 | page_set_slab(page, slab); | |
1da177e4 | 2778 | page++; |
4776874f | 2779 | } while (--nr_pages); |
1da177e4 LT |
2780 | } |
2781 | ||
2782 | /* | |
2783 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2784 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2785 | */ | |
3c517a61 CL |
2786 | static int cache_grow(struct kmem_cache *cachep, |
2787 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2788 | { |
b28a02de | 2789 | struct slab *slabp; |
b28a02de PE |
2790 | size_t offset; |
2791 | gfp_t local_flags; | |
e498be7d | 2792 | struct kmem_list3 *l3; |
1da177e4 | 2793 | |
a737b3e2 AM |
2794 | /* |
2795 | * Be lazy and only check for valid flags here, keeping it out of the | |
2796 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2797 | */ |
6cb06229 CL |
2798 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2799 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2800 | |
2e1217cf | 2801 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2802 | check_irq_off(); |
2e1217cf RT |
2803 | l3 = cachep->nodelists[nodeid]; |
2804 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2805 | |
2806 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2807 | offset = l3->colour_next; |
2808 | l3->colour_next++; | |
2809 | if (l3->colour_next >= cachep->colour) | |
2810 | l3->colour_next = 0; | |
2811 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2812 | |
2e1217cf | 2813 | offset *= cachep->colour_off; |
1da177e4 LT |
2814 | |
2815 | if (local_flags & __GFP_WAIT) | |
2816 | local_irq_enable(); | |
2817 | ||
2818 | /* | |
2819 | * The test for missing atomic flag is performed here, rather than | |
2820 | * the more obvious place, simply to reduce the critical path length | |
2821 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2822 | * will eventually be caught here (where it matters). | |
2823 | */ | |
2824 | kmem_flagcheck(cachep, flags); | |
2825 | ||
a737b3e2 AM |
2826 | /* |
2827 | * Get mem for the objs. Attempt to allocate a physical page from | |
2828 | * 'nodeid'. | |
e498be7d | 2829 | */ |
3c517a61 | 2830 | if (!objp) |
b8c1c5da | 2831 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2832 | if (!objp) |
1da177e4 LT |
2833 | goto failed; |
2834 | ||
2835 | /* Get slab management. */ | |
3c517a61 | 2836 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2837 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2838 | if (!slabp) |
1da177e4 LT |
2839 | goto opps1; |
2840 | ||
4776874f | 2841 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2842 | |
a35afb83 | 2843 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2844 | |
2845 | if (local_flags & __GFP_WAIT) | |
2846 | local_irq_disable(); | |
2847 | check_irq_off(); | |
e498be7d | 2848 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2849 | |
2850 | /* Make slab active. */ | |
e498be7d | 2851 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2852 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2853 | l3->free_objects += cachep->num; |
2854 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2855 | return 1; |
a737b3e2 | 2856 | opps1: |
1da177e4 | 2857 | kmem_freepages(cachep, objp); |
a737b3e2 | 2858 | failed: |
1da177e4 LT |
2859 | if (local_flags & __GFP_WAIT) |
2860 | local_irq_disable(); | |
2861 | return 0; | |
2862 | } | |
2863 | ||
2864 | #if DEBUG | |
2865 | ||
2866 | /* | |
2867 | * Perform extra freeing checks: | |
2868 | * - detect bad pointers. | |
2869 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2870 | */ |
2871 | static void kfree_debugcheck(const void *objp) | |
2872 | { | |
1da177e4 LT |
2873 | if (!virt_addr_valid(objp)) { |
2874 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2875 | (unsigned long)objp); |
2876 | BUG(); | |
1da177e4 | 2877 | } |
1da177e4 LT |
2878 | } |
2879 | ||
58ce1fd5 PE |
2880 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2881 | { | |
b46b8f19 | 2882 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2883 | |
2884 | redzone1 = *dbg_redzone1(cache, obj); | |
2885 | redzone2 = *dbg_redzone2(cache, obj); | |
2886 | ||
2887 | /* | |
2888 | * Redzone is ok. | |
2889 | */ | |
2890 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2891 | return; | |
2892 | ||
2893 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2894 | slab_error(cache, "double free detected"); | |
2895 | else | |
2896 | slab_error(cache, "memory outside object was overwritten"); | |
2897 | ||
b46b8f19 | 2898 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2899 | obj, redzone1, redzone2); |
2900 | } | |
2901 | ||
343e0d7a | 2902 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2903 | void *caller) |
1da177e4 LT |
2904 | { |
2905 | struct page *page; | |
2906 | unsigned int objnr; | |
2907 | struct slab *slabp; | |
2908 | ||
80cbd911 MW |
2909 | BUG_ON(virt_to_cache(objp) != cachep); |
2910 | ||
3dafccf2 | 2911 | objp -= obj_offset(cachep); |
1da177e4 | 2912 | kfree_debugcheck(objp); |
b49af68f | 2913 | page = virt_to_head_page(objp); |
1da177e4 | 2914 | |
065d41cb | 2915 | slabp = page_get_slab(page); |
1da177e4 LT |
2916 | |
2917 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2918 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2919 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2920 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2921 | } | |
2922 | if (cachep->flags & SLAB_STORE_USER) | |
2923 | *dbg_userword(cachep, objp) = caller; | |
2924 | ||
8fea4e96 | 2925 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2926 | |
2927 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2928 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2929 | |
871751e2 AV |
2930 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2931 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2932 | #endif | |
1da177e4 LT |
2933 | if (cachep->flags & SLAB_POISON) { |
2934 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2935 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2936 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2937 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2938 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2939 | } else { |
2940 | poison_obj(cachep, objp, POISON_FREE); | |
2941 | } | |
2942 | #else | |
2943 | poison_obj(cachep, objp, POISON_FREE); | |
2944 | #endif | |
2945 | } | |
2946 | return objp; | |
2947 | } | |
2948 | ||
343e0d7a | 2949 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2950 | { |
2951 | kmem_bufctl_t i; | |
2952 | int entries = 0; | |
b28a02de | 2953 | |
1da177e4 LT |
2954 | /* Check slab's freelist to see if this obj is there. */ |
2955 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2956 | entries++; | |
2957 | if (entries > cachep->num || i >= cachep->num) | |
2958 | goto bad; | |
2959 | } | |
2960 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2961 | bad: |
2962 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2963 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2964 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2965 | for (i = 0; |
264132bc | 2966 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2967 | i++) { |
a737b3e2 | 2968 | if (i % 16 == 0) |
1da177e4 | 2969 | printk("\n%03x:", i); |
b28a02de | 2970 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2971 | } |
2972 | printk("\n"); | |
2973 | BUG(); | |
2974 | } | |
2975 | } | |
2976 | #else | |
2977 | #define kfree_debugcheck(x) do { } while(0) | |
2978 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2979 | #define check_slabp(x,y) do { } while(0) | |
2980 | #endif | |
2981 | ||
343e0d7a | 2982 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2983 | { |
2984 | int batchcount; | |
2985 | struct kmem_list3 *l3; | |
2986 | struct array_cache *ac; | |
1ca4cb24 PE |
2987 | int node; |
2988 | ||
6d2144d3 | 2989 | retry: |
1da177e4 | 2990 | check_irq_off(); |
6d2144d3 | 2991 | node = numa_node_id(); |
9a2dba4b | 2992 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2993 | batchcount = ac->batchcount; |
2994 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2995 | /* |
2996 | * If there was little recent activity on this cache, then | |
2997 | * perform only a partial refill. Otherwise we could generate | |
2998 | * refill bouncing. | |
1da177e4 LT |
2999 | */ |
3000 | batchcount = BATCHREFILL_LIMIT; | |
3001 | } | |
1ca4cb24 | 3002 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3003 | |
3004 | BUG_ON(ac->avail > 0 || !l3); | |
3005 | spin_lock(&l3->list_lock); | |
1da177e4 | 3006 | |
3ded175a CL |
3007 | /* See if we can refill from the shared array */ |
3008 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
3009 | goto alloc_done; | |
3010 | ||
1da177e4 LT |
3011 | while (batchcount > 0) { |
3012 | struct list_head *entry; | |
3013 | struct slab *slabp; | |
3014 | /* Get slab alloc is to come from. */ | |
3015 | entry = l3->slabs_partial.next; | |
3016 | if (entry == &l3->slabs_partial) { | |
3017 | l3->free_touched = 1; | |
3018 | entry = l3->slabs_free.next; | |
3019 | if (entry == &l3->slabs_free) | |
3020 | goto must_grow; | |
3021 | } | |
3022 | ||
3023 | slabp = list_entry(entry, struct slab, list); | |
3024 | check_slabp(cachep, slabp); | |
3025 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3026 | |
3027 | /* | |
3028 | * The slab was either on partial or free list so | |
3029 | * there must be at least one object available for | |
3030 | * allocation. | |
3031 | */ | |
249b9f33 | 3032 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3033 | |
1da177e4 | 3034 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3035 | STATS_INC_ALLOCED(cachep); |
3036 | STATS_INC_ACTIVE(cachep); | |
3037 | STATS_SET_HIGH(cachep); | |
3038 | ||
78d382d7 | 3039 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3040 | node); |
1da177e4 LT |
3041 | } |
3042 | check_slabp(cachep, slabp); | |
3043 | ||
3044 | /* move slabp to correct slabp list: */ | |
3045 | list_del(&slabp->list); | |
3046 | if (slabp->free == BUFCTL_END) | |
3047 | list_add(&slabp->list, &l3->slabs_full); | |
3048 | else | |
3049 | list_add(&slabp->list, &l3->slabs_partial); | |
3050 | } | |
3051 | ||
a737b3e2 | 3052 | must_grow: |
1da177e4 | 3053 | l3->free_objects -= ac->avail; |
a737b3e2 | 3054 | alloc_done: |
e498be7d | 3055 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3056 | |
3057 | if (unlikely(!ac->avail)) { | |
3058 | int x; | |
3c517a61 | 3059 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3060 | |
a737b3e2 | 3061 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3062 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3063 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3064 | return NULL; |
3065 | ||
a737b3e2 | 3066 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3067 | goto retry; |
3068 | } | |
3069 | ac->touched = 1; | |
e498be7d | 3070 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3071 | } |
3072 | ||
a737b3e2 AM |
3073 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3074 | gfp_t flags) | |
1da177e4 LT |
3075 | { |
3076 | might_sleep_if(flags & __GFP_WAIT); | |
3077 | #if DEBUG | |
3078 | kmem_flagcheck(cachep, flags); | |
3079 | #endif | |
3080 | } | |
3081 | ||
3082 | #if DEBUG | |
a737b3e2 AM |
3083 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3084 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3085 | { |
b28a02de | 3086 | if (!objp) |
1da177e4 | 3087 | return objp; |
b28a02de | 3088 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3089 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3090 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3091 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3092 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3093 | else |
3094 | check_poison_obj(cachep, objp); | |
3095 | #else | |
3096 | check_poison_obj(cachep, objp); | |
3097 | #endif | |
3098 | poison_obj(cachep, objp, POISON_INUSE); | |
3099 | } | |
3100 | if (cachep->flags & SLAB_STORE_USER) | |
3101 | *dbg_userword(cachep, objp) = caller; | |
3102 | ||
3103 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3104 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3105 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3106 | slab_error(cachep, "double free, or memory outside" | |
3107 | " object was overwritten"); | |
b28a02de | 3108 | printk(KERN_ERR |
b46b8f19 | 3109 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3110 | objp, *dbg_redzone1(cachep, objp), |
3111 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3112 | } |
3113 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3114 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3115 | } | |
871751e2 AV |
3116 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3117 | { | |
3118 | struct slab *slabp; | |
3119 | unsigned objnr; | |
3120 | ||
b49af68f | 3121 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3122 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3123 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3124 | } | |
3125 | #endif | |
3dafccf2 | 3126 | objp += obj_offset(cachep); |
4f104934 | 3127 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3128 | cachep->ctor(objp); |
a44b56d3 KH |
3129 | #if ARCH_SLAB_MINALIGN |
3130 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3131 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3132 | objp, ARCH_SLAB_MINALIGN); | |
3133 | } | |
3134 | #endif | |
1da177e4 LT |
3135 | return objp; |
3136 | } | |
3137 | #else | |
3138 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3139 | #endif | |
3140 | ||
773ff60e | 3141 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3142 | { |
3143 | if (cachep == &cache_cache) | |
773ff60e | 3144 | return false; |
8a8b6502 | 3145 | |
773ff60e | 3146 | return should_failslab(obj_size(cachep), flags); |
8a8b6502 AM |
3147 | } |
3148 | ||
343e0d7a | 3149 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3150 | { |
b28a02de | 3151 | void *objp; |
1da177e4 LT |
3152 | struct array_cache *ac; |
3153 | ||
5c382300 | 3154 | check_irq_off(); |
8a8b6502 | 3155 | |
9a2dba4b | 3156 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3157 | if (likely(ac->avail)) { |
3158 | STATS_INC_ALLOCHIT(cachep); | |
3159 | ac->touched = 1; | |
e498be7d | 3160 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3161 | } else { |
3162 | STATS_INC_ALLOCMISS(cachep); | |
3163 | objp = cache_alloc_refill(cachep, flags); | |
3164 | } | |
d5cff635 CM |
3165 | /* |
3166 | * To avoid a false negative, if an object that is in one of the | |
3167 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3168 | * treat the array pointers as a reference to the object. | |
3169 | */ | |
3170 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3171 | return objp; |
3172 | } | |
3173 | ||
e498be7d | 3174 | #ifdef CONFIG_NUMA |
c61afb18 | 3175 | /* |
b2455396 | 3176 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3177 | * |
3178 | * If we are in_interrupt, then process context, including cpusets and | |
3179 | * mempolicy, may not apply and should not be used for allocation policy. | |
3180 | */ | |
3181 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3182 | { | |
3183 | int nid_alloc, nid_here; | |
3184 | ||
765c4507 | 3185 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3186 | return NULL; |
3187 | nid_alloc = nid_here = numa_node_id(); | |
3188 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3189 | nid_alloc = cpuset_mem_spread_node(); | |
3190 | else if (current->mempolicy) | |
3191 | nid_alloc = slab_node(current->mempolicy); | |
3192 | if (nid_alloc != nid_here) | |
8b98c169 | 3193 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3194 | return NULL; |
3195 | } | |
3196 | ||
765c4507 CL |
3197 | /* |
3198 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3199 | * certain node and fall back is permitted. First we scan all the |
3200 | * available nodelists for available objects. If that fails then we | |
3201 | * perform an allocation without specifying a node. This allows the page | |
3202 | * allocator to do its reclaim / fallback magic. We then insert the | |
3203 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3204 | */ |
8c8cc2c1 | 3205 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3206 | { |
8c8cc2c1 PE |
3207 | struct zonelist *zonelist; |
3208 | gfp_t local_flags; | |
dd1a239f | 3209 | struct zoneref *z; |
54a6eb5c MG |
3210 | struct zone *zone; |
3211 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3212 | void *obj = NULL; |
3c517a61 | 3213 | int nid; |
8c8cc2c1 PE |
3214 | |
3215 | if (flags & __GFP_THISNODE) | |
3216 | return NULL; | |
3217 | ||
0e88460d | 3218 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
6cb06229 | 3219 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3220 | |
3c517a61 CL |
3221 | retry: |
3222 | /* | |
3223 | * Look through allowed nodes for objects available | |
3224 | * from existing per node queues. | |
3225 | */ | |
54a6eb5c MG |
3226 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3227 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3228 | |
54a6eb5c | 3229 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3230 | cache->nodelists[nid] && |
481c5346 | 3231 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3232 | obj = ____cache_alloc_node(cache, |
3233 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3234 | if (obj) |
3235 | break; | |
3236 | } | |
3c517a61 CL |
3237 | } |
3238 | ||
cfce6604 | 3239 | if (!obj) { |
3c517a61 CL |
3240 | /* |
3241 | * This allocation will be performed within the constraints | |
3242 | * of the current cpuset / memory policy requirements. | |
3243 | * We may trigger various forms of reclaim on the allowed | |
3244 | * set and go into memory reserves if necessary. | |
3245 | */ | |
dd47ea75 CL |
3246 | if (local_flags & __GFP_WAIT) |
3247 | local_irq_enable(); | |
3248 | kmem_flagcheck(cache, flags); | |
9ac33b2b | 3249 | obj = kmem_getpages(cache, local_flags, -1); |
dd47ea75 CL |
3250 | if (local_flags & __GFP_WAIT) |
3251 | local_irq_disable(); | |
3c517a61 CL |
3252 | if (obj) { |
3253 | /* | |
3254 | * Insert into the appropriate per node queues | |
3255 | */ | |
3256 | nid = page_to_nid(virt_to_page(obj)); | |
3257 | if (cache_grow(cache, flags, nid, obj)) { | |
3258 | obj = ____cache_alloc_node(cache, | |
3259 | flags | GFP_THISNODE, nid); | |
3260 | if (!obj) | |
3261 | /* | |
3262 | * Another processor may allocate the | |
3263 | * objects in the slab since we are | |
3264 | * not holding any locks. | |
3265 | */ | |
3266 | goto retry; | |
3267 | } else { | |
b6a60451 | 3268 | /* cache_grow already freed obj */ |
3c517a61 CL |
3269 | obj = NULL; |
3270 | } | |
3271 | } | |
aedb0eb1 | 3272 | } |
765c4507 CL |
3273 | return obj; |
3274 | } | |
3275 | ||
e498be7d CL |
3276 | /* |
3277 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3278 | */ |
8b98c169 | 3279 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3280 | int nodeid) |
e498be7d CL |
3281 | { |
3282 | struct list_head *entry; | |
b28a02de PE |
3283 | struct slab *slabp; |
3284 | struct kmem_list3 *l3; | |
3285 | void *obj; | |
b28a02de PE |
3286 | int x; |
3287 | ||
3288 | l3 = cachep->nodelists[nodeid]; | |
3289 | BUG_ON(!l3); | |
3290 | ||
a737b3e2 | 3291 | retry: |
ca3b9b91 | 3292 | check_irq_off(); |
b28a02de PE |
3293 | spin_lock(&l3->list_lock); |
3294 | entry = l3->slabs_partial.next; | |
3295 | if (entry == &l3->slabs_partial) { | |
3296 | l3->free_touched = 1; | |
3297 | entry = l3->slabs_free.next; | |
3298 | if (entry == &l3->slabs_free) | |
3299 | goto must_grow; | |
3300 | } | |
3301 | ||
3302 | slabp = list_entry(entry, struct slab, list); | |
3303 | check_spinlock_acquired_node(cachep, nodeid); | |
3304 | check_slabp(cachep, slabp); | |
3305 | ||
3306 | STATS_INC_NODEALLOCS(cachep); | |
3307 | STATS_INC_ACTIVE(cachep); | |
3308 | STATS_SET_HIGH(cachep); | |
3309 | ||
3310 | BUG_ON(slabp->inuse == cachep->num); | |
3311 | ||
78d382d7 | 3312 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3313 | check_slabp(cachep, slabp); |
3314 | l3->free_objects--; | |
3315 | /* move slabp to correct slabp list: */ | |
3316 | list_del(&slabp->list); | |
3317 | ||
a737b3e2 | 3318 | if (slabp->free == BUFCTL_END) |
b28a02de | 3319 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3320 | else |
b28a02de | 3321 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3322 | |
b28a02de PE |
3323 | spin_unlock(&l3->list_lock); |
3324 | goto done; | |
e498be7d | 3325 | |
a737b3e2 | 3326 | must_grow: |
b28a02de | 3327 | spin_unlock(&l3->list_lock); |
3c517a61 | 3328 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3329 | if (x) |
3330 | goto retry; | |
1da177e4 | 3331 | |
8c8cc2c1 | 3332 | return fallback_alloc(cachep, flags); |
e498be7d | 3333 | |
a737b3e2 | 3334 | done: |
b28a02de | 3335 | return obj; |
e498be7d | 3336 | } |
8c8cc2c1 PE |
3337 | |
3338 | /** | |
3339 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3340 | * @cachep: The cache to allocate from. | |
3341 | * @flags: See kmalloc(). | |
3342 | * @nodeid: node number of the target node. | |
3343 | * @caller: return address of caller, used for debug information | |
3344 | * | |
3345 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3346 | * node, which can improve the performance for cpu bound structures. | |
3347 | * | |
3348 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3349 | */ | |
3350 | static __always_inline void * | |
3351 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3352 | void *caller) | |
3353 | { | |
3354 | unsigned long save_flags; | |
3355 | void *ptr; | |
3356 | ||
cf40bd16 NP |
3357 | lockdep_trace_alloc(flags); |
3358 | ||
773ff60e | 3359 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3360 | return NULL; |
3361 | ||
8c8cc2c1 PE |
3362 | cache_alloc_debugcheck_before(cachep, flags); |
3363 | local_irq_save(save_flags); | |
3364 | ||
3365 | if (unlikely(nodeid == -1)) | |
3366 | nodeid = numa_node_id(); | |
3367 | ||
3368 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3369 | /* Node not bootstrapped yet */ | |
3370 | ptr = fallback_alloc(cachep, flags); | |
3371 | goto out; | |
3372 | } | |
3373 | ||
3374 | if (nodeid == numa_node_id()) { | |
3375 | /* | |
3376 | * Use the locally cached objects if possible. | |
3377 | * However ____cache_alloc does not allow fallback | |
3378 | * to other nodes. It may fail while we still have | |
3379 | * objects on other nodes available. | |
3380 | */ | |
3381 | ptr = ____cache_alloc(cachep, flags); | |
3382 | if (ptr) | |
3383 | goto out; | |
3384 | } | |
3385 | /* ___cache_alloc_node can fall back to other nodes */ | |
3386 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3387 | out: | |
3388 | local_irq_restore(save_flags); | |
3389 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
d5cff635 CM |
3390 | kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags, |
3391 | flags); | |
8c8cc2c1 | 3392 | |
d07dbea4 CL |
3393 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3394 | memset(ptr, 0, obj_size(cachep)); | |
3395 | ||
8c8cc2c1 PE |
3396 | return ptr; |
3397 | } | |
3398 | ||
3399 | static __always_inline void * | |
3400 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3401 | { | |
3402 | void *objp; | |
3403 | ||
3404 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3405 | objp = alternate_node_alloc(cache, flags); | |
3406 | if (objp) | |
3407 | goto out; | |
3408 | } | |
3409 | objp = ____cache_alloc(cache, flags); | |
3410 | ||
3411 | /* | |
3412 | * We may just have run out of memory on the local node. | |
3413 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3414 | */ | |
3415 | if (!objp) | |
3416 | objp = ____cache_alloc_node(cache, flags, numa_node_id()); | |
3417 | ||
3418 | out: | |
3419 | return objp; | |
3420 | } | |
3421 | #else | |
3422 | ||
3423 | static __always_inline void * | |
3424 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3425 | { | |
3426 | return ____cache_alloc(cachep, flags); | |
3427 | } | |
3428 | ||
3429 | #endif /* CONFIG_NUMA */ | |
3430 | ||
3431 | static __always_inline void * | |
3432 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3433 | { | |
3434 | unsigned long save_flags; | |
3435 | void *objp; | |
3436 | ||
cf40bd16 NP |
3437 | lockdep_trace_alloc(flags); |
3438 | ||
773ff60e | 3439 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3440 | return NULL; |
3441 | ||
8c8cc2c1 PE |
3442 | cache_alloc_debugcheck_before(cachep, flags); |
3443 | local_irq_save(save_flags); | |
3444 | objp = __do_cache_alloc(cachep, flags); | |
3445 | local_irq_restore(save_flags); | |
3446 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
d5cff635 CM |
3447 | kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags, |
3448 | flags); | |
8c8cc2c1 PE |
3449 | prefetchw(objp); |
3450 | ||
d07dbea4 CL |
3451 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3452 | memset(objp, 0, obj_size(cachep)); | |
3453 | ||
8c8cc2c1 PE |
3454 | return objp; |
3455 | } | |
e498be7d CL |
3456 | |
3457 | /* | |
3458 | * Caller needs to acquire correct kmem_list's list_lock | |
3459 | */ | |
343e0d7a | 3460 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3461 | int node) |
1da177e4 LT |
3462 | { |
3463 | int i; | |
e498be7d | 3464 | struct kmem_list3 *l3; |
1da177e4 LT |
3465 | |
3466 | for (i = 0; i < nr_objects; i++) { | |
3467 | void *objp = objpp[i]; | |
3468 | struct slab *slabp; | |
1da177e4 | 3469 | |
6ed5eb22 | 3470 | slabp = virt_to_slab(objp); |
ff69416e | 3471 | l3 = cachep->nodelists[node]; |
1da177e4 | 3472 | list_del(&slabp->list); |
ff69416e | 3473 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3474 | check_slabp(cachep, slabp); |
78d382d7 | 3475 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3476 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3477 | l3->free_objects++; |
1da177e4 LT |
3478 | check_slabp(cachep, slabp); |
3479 | ||
3480 | /* fixup slab chains */ | |
3481 | if (slabp->inuse == 0) { | |
e498be7d CL |
3482 | if (l3->free_objects > l3->free_limit) { |
3483 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3484 | /* No need to drop any previously held |
3485 | * lock here, even if we have a off-slab slab | |
3486 | * descriptor it is guaranteed to come from | |
3487 | * a different cache, refer to comments before | |
3488 | * alloc_slabmgmt. | |
3489 | */ | |
1da177e4 LT |
3490 | slab_destroy(cachep, slabp); |
3491 | } else { | |
e498be7d | 3492 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3493 | } |
3494 | } else { | |
3495 | /* Unconditionally move a slab to the end of the | |
3496 | * partial list on free - maximum time for the | |
3497 | * other objects to be freed, too. | |
3498 | */ | |
e498be7d | 3499 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3500 | } |
3501 | } | |
3502 | } | |
3503 | ||
343e0d7a | 3504 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3505 | { |
3506 | int batchcount; | |
e498be7d | 3507 | struct kmem_list3 *l3; |
ff69416e | 3508 | int node = numa_node_id(); |
1da177e4 LT |
3509 | |
3510 | batchcount = ac->batchcount; | |
3511 | #if DEBUG | |
3512 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3513 | #endif | |
3514 | check_irq_off(); | |
ff69416e | 3515 | l3 = cachep->nodelists[node]; |
873623df | 3516 | spin_lock(&l3->list_lock); |
e498be7d CL |
3517 | if (l3->shared) { |
3518 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3519 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3520 | if (max) { |
3521 | if (batchcount > max) | |
3522 | batchcount = max; | |
e498be7d | 3523 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3524 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3525 | shared_array->avail += batchcount; |
3526 | goto free_done; | |
3527 | } | |
3528 | } | |
3529 | ||
ff69416e | 3530 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3531 | free_done: |
1da177e4 LT |
3532 | #if STATS |
3533 | { | |
3534 | int i = 0; | |
3535 | struct list_head *p; | |
3536 | ||
e498be7d CL |
3537 | p = l3->slabs_free.next; |
3538 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3539 | struct slab *slabp; |
3540 | ||
3541 | slabp = list_entry(p, struct slab, list); | |
3542 | BUG_ON(slabp->inuse); | |
3543 | ||
3544 | i++; | |
3545 | p = p->next; | |
3546 | } | |
3547 | STATS_SET_FREEABLE(cachep, i); | |
3548 | } | |
3549 | #endif | |
e498be7d | 3550 | spin_unlock(&l3->list_lock); |
1da177e4 | 3551 | ac->avail -= batchcount; |
a737b3e2 | 3552 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3553 | } |
3554 | ||
3555 | /* | |
a737b3e2 AM |
3556 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3557 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3558 | */ |
873623df | 3559 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3560 | { |
9a2dba4b | 3561 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3562 | |
3563 | check_irq_off(); | |
d5cff635 | 3564 | kmemleak_free_recursive(objp, cachep->flags); |
1da177e4 LT |
3565 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); |
3566 | ||
1807a1aa SS |
3567 | /* |
3568 | * Skip calling cache_free_alien() when the platform is not numa. | |
3569 | * This will avoid cache misses that happen while accessing slabp (which | |
3570 | * is per page memory reference) to get nodeid. Instead use a global | |
3571 | * variable to skip the call, which is mostly likely to be present in | |
3572 | * the cache. | |
3573 | */ | |
3574 | if (numa_platform && cache_free_alien(cachep, objp)) | |
729bd0b7 PE |
3575 | return; |
3576 | ||
1da177e4 LT |
3577 | if (likely(ac->avail < ac->limit)) { |
3578 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3579 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3580 | return; |
3581 | } else { | |
3582 | STATS_INC_FREEMISS(cachep); | |
3583 | cache_flusharray(cachep, ac); | |
e498be7d | 3584 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3585 | } |
3586 | } | |
3587 | ||
3588 | /** | |
3589 | * kmem_cache_alloc - Allocate an object | |
3590 | * @cachep: The cache to allocate from. | |
3591 | * @flags: See kmalloc(). | |
3592 | * | |
3593 | * Allocate an object from this cache. The flags are only relevant | |
3594 | * if the cache has no available objects. | |
3595 | */ | |
343e0d7a | 3596 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3597 | { |
36555751 EGM |
3598 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3599 | ||
ca2b84cb EGM |
3600 | trace_kmem_cache_alloc(_RET_IP_, ret, |
3601 | obj_size(cachep), cachep->buffer_size, flags); | |
36555751 EGM |
3602 | |
3603 | return ret; | |
1da177e4 LT |
3604 | } |
3605 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3606 | ||
36555751 EGM |
3607 | #ifdef CONFIG_KMEMTRACE |
3608 | void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags) | |
3609 | { | |
3610 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3611 | } | |
3612 | EXPORT_SYMBOL(kmem_cache_alloc_notrace); | |
3613 | #endif | |
3614 | ||
1da177e4 | 3615 | /** |
7682486b | 3616 | * kmem_ptr_validate - check if an untrusted pointer might be a slab entry. |
1da177e4 LT |
3617 | * @cachep: the cache we're checking against |
3618 | * @ptr: pointer to validate | |
3619 | * | |
7682486b | 3620 | * This verifies that the untrusted pointer looks sane; |
1da177e4 LT |
3621 | * it is _not_ a guarantee that the pointer is actually |
3622 | * part of the slab cache in question, but it at least | |
3623 | * validates that the pointer can be dereferenced and | |
3624 | * looks half-way sane. | |
3625 | * | |
3626 | * Currently only used for dentry validation. | |
3627 | */ | |
b7f869a2 | 3628 | int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) |
1da177e4 | 3629 | { |
b28a02de | 3630 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3631 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3632 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3633 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3634 | struct page *page; |
3635 | ||
3636 | if (unlikely(addr < min_addr)) | |
3637 | goto out; | |
3638 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3639 | goto out; | |
3640 | if (unlikely(addr & align_mask)) | |
3641 | goto out; | |
3642 | if (unlikely(!kern_addr_valid(addr))) | |
3643 | goto out; | |
3644 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3645 | goto out; | |
3646 | page = virt_to_page(ptr); | |
3647 | if (unlikely(!PageSlab(page))) | |
3648 | goto out; | |
065d41cb | 3649 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3650 | goto out; |
3651 | return 1; | |
a737b3e2 | 3652 | out: |
1da177e4 LT |
3653 | return 0; |
3654 | } | |
3655 | ||
3656 | #ifdef CONFIG_NUMA | |
8b98c169 CH |
3657 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3658 | { | |
36555751 EGM |
3659 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3660 | __builtin_return_address(0)); | |
3661 | ||
ca2b84cb EGM |
3662 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
3663 | obj_size(cachep), cachep->buffer_size, | |
3664 | flags, nodeid); | |
36555751 EGM |
3665 | |
3666 | return ret; | |
8b98c169 | 3667 | } |
1da177e4 LT |
3668 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3669 | ||
36555751 EGM |
3670 | #ifdef CONFIG_KMEMTRACE |
3671 | void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep, | |
3672 | gfp_t flags, | |
3673 | int nodeid) | |
3674 | { | |
3675 | return __cache_alloc_node(cachep, flags, nodeid, | |
3676 | __builtin_return_address(0)); | |
3677 | } | |
3678 | EXPORT_SYMBOL(kmem_cache_alloc_node_notrace); | |
3679 | #endif | |
3680 | ||
8b98c169 CH |
3681 | static __always_inline void * |
3682 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3683 | { |
343e0d7a | 3684 | struct kmem_cache *cachep; |
36555751 | 3685 | void *ret; |
97e2bde4 MS |
3686 | |
3687 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3688 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3689 | return cachep; | |
36555751 EGM |
3690 | ret = kmem_cache_alloc_node_notrace(cachep, flags, node); |
3691 | ||
ca2b84cb EGM |
3692 | trace_kmalloc_node((unsigned long) caller, ret, |
3693 | size, cachep->buffer_size, flags, node); | |
36555751 EGM |
3694 | |
3695 | return ret; | |
97e2bde4 | 3696 | } |
8b98c169 | 3697 | |
36555751 | 3698 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE) |
8b98c169 CH |
3699 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3700 | { | |
3701 | return __do_kmalloc_node(size, flags, node, | |
3702 | __builtin_return_address(0)); | |
3703 | } | |
dbe5e69d | 3704 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3705 | |
3706 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3707 | int node, unsigned long caller) |
8b98c169 | 3708 | { |
ce71e27c | 3709 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3710 | } |
3711 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3712 | #else | |
3713 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3714 | { | |
3715 | return __do_kmalloc_node(size, flags, node, NULL); | |
3716 | } | |
3717 | EXPORT_SYMBOL(__kmalloc_node); | |
3718 | #endif /* CONFIG_DEBUG_SLAB */ | |
3719 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3720 | |
3721 | /** | |
800590f5 | 3722 | * __do_kmalloc - allocate memory |
1da177e4 | 3723 | * @size: how many bytes of memory are required. |
800590f5 | 3724 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3725 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3726 | */ |
7fd6b141 PE |
3727 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3728 | void *caller) | |
1da177e4 | 3729 | { |
343e0d7a | 3730 | struct kmem_cache *cachep; |
36555751 | 3731 | void *ret; |
1da177e4 | 3732 | |
97e2bde4 MS |
3733 | /* If you want to save a few bytes .text space: replace |
3734 | * __ with kmem_. | |
3735 | * Then kmalloc uses the uninlined functions instead of the inline | |
3736 | * functions. | |
3737 | */ | |
3738 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3739 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3740 | return cachep; | |
36555751 EGM |
3741 | ret = __cache_alloc(cachep, flags, caller); |
3742 | ||
ca2b84cb EGM |
3743 | trace_kmalloc((unsigned long) caller, ret, |
3744 | size, cachep->buffer_size, flags); | |
36555751 EGM |
3745 | |
3746 | return ret; | |
7fd6b141 PE |
3747 | } |
3748 | ||
7fd6b141 | 3749 | |
36555751 | 3750 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE) |
7fd6b141 PE |
3751 | void *__kmalloc(size_t size, gfp_t flags) |
3752 | { | |
871751e2 | 3753 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3754 | } |
3755 | EXPORT_SYMBOL(__kmalloc); | |
3756 | ||
ce71e27c | 3757 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3758 | { |
ce71e27c | 3759 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3760 | } |
3761 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3762 | |
3763 | #else | |
3764 | void *__kmalloc(size_t size, gfp_t flags) | |
3765 | { | |
3766 | return __do_kmalloc(size, flags, NULL); | |
3767 | } | |
3768 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3769 | #endif |
3770 | ||
1da177e4 LT |
3771 | /** |
3772 | * kmem_cache_free - Deallocate an object | |
3773 | * @cachep: The cache the allocation was from. | |
3774 | * @objp: The previously allocated object. | |
3775 | * | |
3776 | * Free an object which was previously allocated from this | |
3777 | * cache. | |
3778 | */ | |
343e0d7a | 3779 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3780 | { |
3781 | unsigned long flags; | |
3782 | ||
3783 | local_irq_save(flags); | |
898552c9 | 3784 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
3ac7fe5a TG |
3785 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3786 | debug_check_no_obj_freed(objp, obj_size(cachep)); | |
873623df | 3787 | __cache_free(cachep, objp); |
1da177e4 | 3788 | local_irq_restore(flags); |
36555751 | 3789 | |
ca2b84cb | 3790 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3791 | } |
3792 | EXPORT_SYMBOL(kmem_cache_free); | |
3793 | ||
1da177e4 LT |
3794 | /** |
3795 | * kfree - free previously allocated memory | |
3796 | * @objp: pointer returned by kmalloc. | |
3797 | * | |
80e93eff PE |
3798 | * If @objp is NULL, no operation is performed. |
3799 | * | |
1da177e4 LT |
3800 | * Don't free memory not originally allocated by kmalloc() |
3801 | * or you will run into trouble. | |
3802 | */ | |
3803 | void kfree(const void *objp) | |
3804 | { | |
343e0d7a | 3805 | struct kmem_cache *c; |
1da177e4 LT |
3806 | unsigned long flags; |
3807 | ||
2121db74 PE |
3808 | trace_kfree(_RET_IP_, objp); |
3809 | ||
6cb8f913 | 3810 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3811 | return; |
3812 | local_irq_save(flags); | |
3813 | kfree_debugcheck(objp); | |
6ed5eb22 | 3814 | c = virt_to_cache(objp); |
f9b8404c | 3815 | debug_check_no_locks_freed(objp, obj_size(c)); |
3ac7fe5a | 3816 | debug_check_no_obj_freed(objp, obj_size(c)); |
873623df | 3817 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3818 | local_irq_restore(flags); |
3819 | } | |
3820 | EXPORT_SYMBOL(kfree); | |
3821 | ||
343e0d7a | 3822 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3823 | { |
3dafccf2 | 3824 | return obj_size(cachep); |
1da177e4 LT |
3825 | } |
3826 | EXPORT_SYMBOL(kmem_cache_size); | |
3827 | ||
343e0d7a | 3828 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3829 | { |
3830 | return cachep->name; | |
3831 | } | |
3832 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3833 | ||
e498be7d | 3834 | /* |
183ff22b | 3835 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3836 | */ |
83b519e8 | 3837 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3838 | { |
3839 | int node; | |
3840 | struct kmem_list3 *l3; | |
cafeb02e | 3841 | struct array_cache *new_shared; |
3395ee05 | 3842 | struct array_cache **new_alien = NULL; |
e498be7d | 3843 | |
9c09a95c | 3844 | for_each_online_node(node) { |
cafeb02e | 3845 | |
3395ee05 | 3846 | if (use_alien_caches) { |
83b519e8 | 3847 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3848 | if (!new_alien) |
3849 | goto fail; | |
3850 | } | |
cafeb02e | 3851 | |
63109846 ED |
3852 | new_shared = NULL; |
3853 | if (cachep->shared) { | |
3854 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3855 | cachep->shared*cachep->batchcount, |
83b519e8 | 3856 | 0xbaadf00d, gfp); |
63109846 ED |
3857 | if (!new_shared) { |
3858 | free_alien_cache(new_alien); | |
3859 | goto fail; | |
3860 | } | |
0718dc2a | 3861 | } |
cafeb02e | 3862 | |
a737b3e2 AM |
3863 | l3 = cachep->nodelists[node]; |
3864 | if (l3) { | |
cafeb02e CL |
3865 | struct array_cache *shared = l3->shared; |
3866 | ||
e498be7d CL |
3867 | spin_lock_irq(&l3->list_lock); |
3868 | ||
cafeb02e | 3869 | if (shared) |
0718dc2a CL |
3870 | free_block(cachep, shared->entry, |
3871 | shared->avail, node); | |
e498be7d | 3872 | |
cafeb02e CL |
3873 | l3->shared = new_shared; |
3874 | if (!l3->alien) { | |
e498be7d CL |
3875 | l3->alien = new_alien; |
3876 | new_alien = NULL; | |
3877 | } | |
b28a02de | 3878 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3879 | cachep->batchcount + cachep->num; |
e498be7d | 3880 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3881 | kfree(shared); |
e498be7d CL |
3882 | free_alien_cache(new_alien); |
3883 | continue; | |
3884 | } | |
83b519e8 | 3885 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3886 | if (!l3) { |
3887 | free_alien_cache(new_alien); | |
3888 | kfree(new_shared); | |
e498be7d | 3889 | goto fail; |
0718dc2a | 3890 | } |
e498be7d CL |
3891 | |
3892 | kmem_list3_init(l3); | |
3893 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3894 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3895 | l3->shared = new_shared; |
e498be7d | 3896 | l3->alien = new_alien; |
b28a02de | 3897 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3898 | cachep->batchcount + cachep->num; |
e498be7d CL |
3899 | cachep->nodelists[node] = l3; |
3900 | } | |
cafeb02e | 3901 | return 0; |
0718dc2a | 3902 | |
a737b3e2 | 3903 | fail: |
0718dc2a CL |
3904 | if (!cachep->next.next) { |
3905 | /* Cache is not active yet. Roll back what we did */ | |
3906 | node--; | |
3907 | while (node >= 0) { | |
3908 | if (cachep->nodelists[node]) { | |
3909 | l3 = cachep->nodelists[node]; | |
3910 | ||
3911 | kfree(l3->shared); | |
3912 | free_alien_cache(l3->alien); | |
3913 | kfree(l3); | |
3914 | cachep->nodelists[node] = NULL; | |
3915 | } | |
3916 | node--; | |
3917 | } | |
3918 | } | |
cafeb02e | 3919 | return -ENOMEM; |
e498be7d CL |
3920 | } |
3921 | ||
1da177e4 | 3922 | struct ccupdate_struct { |
343e0d7a | 3923 | struct kmem_cache *cachep; |
1da177e4 LT |
3924 | struct array_cache *new[NR_CPUS]; |
3925 | }; | |
3926 | ||
3927 | static void do_ccupdate_local(void *info) | |
3928 | { | |
a737b3e2 | 3929 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3930 | struct array_cache *old; |
3931 | ||
3932 | check_irq_off(); | |
9a2dba4b | 3933 | old = cpu_cache_get(new->cachep); |
e498be7d | 3934 | |
1da177e4 LT |
3935 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3936 | new->new[smp_processor_id()] = old; | |
3937 | } | |
3938 | ||
b5d8ca7c | 3939 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 3940 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 3941 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 3942 | { |
d2e7b7d0 | 3943 | struct ccupdate_struct *new; |
2ed3a4ef | 3944 | int i; |
1da177e4 | 3945 | |
83b519e8 | 3946 | new = kzalloc(sizeof(*new), gfp); |
d2e7b7d0 SS |
3947 | if (!new) |
3948 | return -ENOMEM; | |
3949 | ||
e498be7d | 3950 | for_each_online_cpu(i) { |
d2e7b7d0 | 3951 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
83b519e8 | 3952 | batchcount, gfp); |
d2e7b7d0 | 3953 | if (!new->new[i]) { |
b28a02de | 3954 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3955 | kfree(new->new[i]); |
3956 | kfree(new); | |
e498be7d | 3957 | return -ENOMEM; |
1da177e4 LT |
3958 | } |
3959 | } | |
d2e7b7d0 | 3960 | new->cachep = cachep; |
1da177e4 | 3961 | |
15c8b6c1 | 3962 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 3963 | |
1da177e4 | 3964 | check_irq_on(); |
1da177e4 LT |
3965 | cachep->batchcount = batchcount; |
3966 | cachep->limit = limit; | |
e498be7d | 3967 | cachep->shared = shared; |
1da177e4 | 3968 | |
e498be7d | 3969 | for_each_online_cpu(i) { |
d2e7b7d0 | 3970 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3971 | if (!ccold) |
3972 | continue; | |
e498be7d | 3973 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3974 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3975 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3976 | kfree(ccold); |
3977 | } | |
d2e7b7d0 | 3978 | kfree(new); |
83b519e8 | 3979 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
3980 | } |
3981 | ||
b5d8ca7c | 3982 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 3983 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
3984 | { |
3985 | int err; | |
3986 | int limit, shared; | |
3987 | ||
a737b3e2 AM |
3988 | /* |
3989 | * The head array serves three purposes: | |
1da177e4 LT |
3990 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3991 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3992 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3993 | * bufctl chains: array operations are cheaper. |
3994 | * The numbers are guessed, we should auto-tune as described by | |
3995 | * Bonwick. | |
3996 | */ | |
3dafccf2 | 3997 | if (cachep->buffer_size > 131072) |
1da177e4 | 3998 | limit = 1; |
3dafccf2 | 3999 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 4000 | limit = 8; |
3dafccf2 | 4001 | else if (cachep->buffer_size > 1024) |
1da177e4 | 4002 | limit = 24; |
3dafccf2 | 4003 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
4004 | limit = 54; |
4005 | else | |
4006 | limit = 120; | |
4007 | ||
a737b3e2 AM |
4008 | /* |
4009 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4010 | * allocation behaviour: Most allocs on one cpu, most free operations |
4011 | * on another cpu. For these cases, an efficient object passing between | |
4012 | * cpus is necessary. This is provided by a shared array. The array | |
4013 | * replaces Bonwick's magazine layer. | |
4014 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4015 | * to a larger limit. Thus disabled by default. | |
4016 | */ | |
4017 | shared = 0; | |
364fbb29 | 4018 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4019 | shared = 8; |
1da177e4 LT |
4020 | |
4021 | #if DEBUG | |
a737b3e2 AM |
4022 | /* |
4023 | * With debugging enabled, large batchcount lead to excessively long | |
4024 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4025 | */ |
4026 | if (limit > 32) | |
4027 | limit = 32; | |
4028 | #endif | |
83b519e8 | 4029 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4030 | if (err) |
4031 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4032 | cachep->name, -err); |
2ed3a4ef | 4033 | return err; |
1da177e4 LT |
4034 | } |
4035 | ||
1b55253a CL |
4036 | /* |
4037 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4038 | * necessary. Note that the l3 listlock also protects the array_cache |
4039 | * if drain_array() is used on the shared array. | |
1b55253a CL |
4040 | */ |
4041 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4042 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
4043 | { |
4044 | int tofree; | |
4045 | ||
1b55253a CL |
4046 | if (!ac || !ac->avail) |
4047 | return; | |
1da177e4 LT |
4048 | if (ac->touched && !force) { |
4049 | ac->touched = 0; | |
b18e7e65 | 4050 | } else { |
1b55253a | 4051 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4052 | if (ac->avail) { |
4053 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4054 | if (tofree > ac->avail) | |
4055 | tofree = (ac->avail + 1) / 2; | |
4056 | free_block(cachep, ac->entry, tofree, node); | |
4057 | ac->avail -= tofree; | |
4058 | memmove(ac->entry, &(ac->entry[tofree]), | |
4059 | sizeof(void *) * ac->avail); | |
4060 | } | |
1b55253a | 4061 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4062 | } |
4063 | } | |
4064 | ||
4065 | /** | |
4066 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4067 | * @w: work descriptor |
1da177e4 LT |
4068 | * |
4069 | * Called from workqueue/eventd every few seconds. | |
4070 | * Purpose: | |
4071 | * - clear the per-cpu caches for this CPU. | |
4072 | * - return freeable pages to the main free memory pool. | |
4073 | * | |
a737b3e2 AM |
4074 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4075 | * again on the next iteration. | |
1da177e4 | 4076 | */ |
7c5cae36 | 4077 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4078 | { |
7a7c381d | 4079 | struct kmem_cache *searchp; |
e498be7d | 4080 | struct kmem_list3 *l3; |
aab2207c | 4081 | int node = numa_node_id(); |
bf6aede7 | 4082 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4083 | |
7c5cae36 | 4084 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4085 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4086 | goto out; |
1da177e4 | 4087 | |
7a7c381d | 4088 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4089 | check_irq_on(); |
4090 | ||
35386e3b CL |
4091 | /* |
4092 | * We only take the l3 lock if absolutely necessary and we | |
4093 | * have established with reasonable certainty that | |
4094 | * we can do some work if the lock was obtained. | |
4095 | */ | |
aab2207c | 4096 | l3 = searchp->nodelists[node]; |
35386e3b | 4097 | |
8fce4d8e | 4098 | reap_alien(searchp, l3); |
1da177e4 | 4099 | |
aab2207c | 4100 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4101 | |
35386e3b CL |
4102 | /* |
4103 | * These are racy checks but it does not matter | |
4104 | * if we skip one check or scan twice. | |
4105 | */ | |
e498be7d | 4106 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4107 | goto next; |
1da177e4 | 4108 | |
e498be7d | 4109 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4110 | |
aab2207c | 4111 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4112 | |
ed11d9eb | 4113 | if (l3->free_touched) |
e498be7d | 4114 | l3->free_touched = 0; |
ed11d9eb CL |
4115 | else { |
4116 | int freed; | |
1da177e4 | 4117 | |
ed11d9eb CL |
4118 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4119 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4120 | STATS_ADD_REAPED(searchp, freed); | |
4121 | } | |
35386e3b | 4122 | next: |
1da177e4 LT |
4123 | cond_resched(); |
4124 | } | |
4125 | check_irq_on(); | |
fc0abb14 | 4126 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4127 | next_reap_node(); |
7c5cae36 | 4128 | out: |
a737b3e2 | 4129 | /* Set up the next iteration */ |
7c5cae36 | 4130 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4131 | } |
4132 | ||
158a9624 | 4133 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4134 | |
85289f98 | 4135 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4136 | { |
85289f98 PE |
4137 | /* |
4138 | * Output format version, so at least we can change it | |
4139 | * without _too_ many complaints. | |
4140 | */ | |
1da177e4 | 4141 | #if STATS |
85289f98 | 4142 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4143 | #else |
85289f98 | 4144 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4145 | #endif |
85289f98 PE |
4146 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4147 | "<objperslab> <pagesperslab>"); | |
4148 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4149 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4150 | #if STATS |
85289f98 | 4151 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4152 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4153 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4154 | #endif |
85289f98 PE |
4155 | seq_putc(m, '\n'); |
4156 | } | |
4157 | ||
4158 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4159 | { | |
4160 | loff_t n = *pos; | |
85289f98 | 4161 | |
fc0abb14 | 4162 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4163 | if (!n) |
4164 | print_slabinfo_header(m); | |
b92151ba PE |
4165 | |
4166 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4167 | } |
4168 | ||
4169 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4170 | { | |
b92151ba | 4171 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4172 | } |
4173 | ||
4174 | static void s_stop(struct seq_file *m, void *p) | |
4175 | { | |
fc0abb14 | 4176 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4177 | } |
4178 | ||
4179 | static int s_show(struct seq_file *m, void *p) | |
4180 | { | |
b92151ba | 4181 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
b28a02de PE |
4182 | struct slab *slabp; |
4183 | unsigned long active_objs; | |
4184 | unsigned long num_objs; | |
4185 | unsigned long active_slabs = 0; | |
4186 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4187 | const char *name; |
1da177e4 | 4188 | char *error = NULL; |
e498be7d CL |
4189 | int node; |
4190 | struct kmem_list3 *l3; | |
1da177e4 | 4191 | |
1da177e4 LT |
4192 | active_objs = 0; |
4193 | num_slabs = 0; | |
e498be7d CL |
4194 | for_each_online_node(node) { |
4195 | l3 = cachep->nodelists[node]; | |
4196 | if (!l3) | |
4197 | continue; | |
4198 | ||
ca3b9b91 RT |
4199 | check_irq_on(); |
4200 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4201 | |
7a7c381d | 4202 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4203 | if (slabp->inuse != cachep->num && !error) |
4204 | error = "slabs_full accounting error"; | |
4205 | active_objs += cachep->num; | |
4206 | active_slabs++; | |
4207 | } | |
7a7c381d | 4208 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4209 | if (slabp->inuse == cachep->num && !error) |
4210 | error = "slabs_partial inuse accounting error"; | |
4211 | if (!slabp->inuse && !error) | |
4212 | error = "slabs_partial/inuse accounting error"; | |
4213 | active_objs += slabp->inuse; | |
4214 | active_slabs++; | |
4215 | } | |
7a7c381d | 4216 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4217 | if (slabp->inuse && !error) |
4218 | error = "slabs_free/inuse accounting error"; | |
4219 | num_slabs++; | |
4220 | } | |
4221 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4222 | if (l3->shared) |
4223 | shared_avail += l3->shared->avail; | |
e498be7d | 4224 | |
ca3b9b91 | 4225 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4226 | } |
b28a02de PE |
4227 | num_slabs += active_slabs; |
4228 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4229 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4230 | error = "free_objects accounting error"; |
4231 | ||
b28a02de | 4232 | name = cachep->name; |
1da177e4 LT |
4233 | if (error) |
4234 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4235 | ||
4236 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4237 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4238 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4239 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4240 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4241 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4242 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4243 | #if STATS |
b28a02de | 4244 | { /* list3 stats */ |
1da177e4 LT |
4245 | unsigned long high = cachep->high_mark; |
4246 | unsigned long allocs = cachep->num_allocations; | |
4247 | unsigned long grown = cachep->grown; | |
4248 | unsigned long reaped = cachep->reaped; | |
4249 | unsigned long errors = cachep->errors; | |
4250 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4251 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4252 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4253 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4254 | |
e498be7d | 4255 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4256 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4257 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4258 | node_frees, overflows); |
1da177e4 LT |
4259 | } |
4260 | /* cpu stats */ | |
4261 | { | |
4262 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4263 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4264 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4265 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4266 | ||
4267 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4268 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4269 | } |
4270 | #endif | |
4271 | seq_putc(m, '\n'); | |
1da177e4 LT |
4272 | return 0; |
4273 | } | |
4274 | ||
4275 | /* | |
4276 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4277 | * | |
4278 | * Output layout: | |
4279 | * cache-name | |
4280 | * num-active-objs | |
4281 | * total-objs | |
4282 | * object size | |
4283 | * num-active-slabs | |
4284 | * total-slabs | |
4285 | * num-pages-per-slab | |
4286 | * + further values on SMP and with statistics enabled | |
4287 | */ | |
4288 | ||
7b3c3a50 | 4289 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4290 | .start = s_start, |
4291 | .next = s_next, | |
4292 | .stop = s_stop, | |
4293 | .show = s_show, | |
1da177e4 LT |
4294 | }; |
4295 | ||
4296 | #define MAX_SLABINFO_WRITE 128 | |
4297 | /** | |
4298 | * slabinfo_write - Tuning for the slab allocator | |
4299 | * @file: unused | |
4300 | * @buffer: user buffer | |
4301 | * @count: data length | |
4302 | * @ppos: unused | |
4303 | */ | |
b28a02de PE |
4304 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4305 | size_t count, loff_t *ppos) | |
1da177e4 | 4306 | { |
b28a02de | 4307 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4308 | int limit, batchcount, shared, res; |
7a7c381d | 4309 | struct kmem_cache *cachep; |
b28a02de | 4310 | |
1da177e4 LT |
4311 | if (count > MAX_SLABINFO_WRITE) |
4312 | return -EINVAL; | |
4313 | if (copy_from_user(&kbuf, buffer, count)) | |
4314 | return -EFAULT; | |
b28a02de | 4315 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4316 | |
4317 | tmp = strchr(kbuf, ' '); | |
4318 | if (!tmp) | |
4319 | return -EINVAL; | |
4320 | *tmp = '\0'; | |
4321 | tmp++; | |
4322 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4323 | return -EINVAL; | |
4324 | ||
4325 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4326 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4327 | res = -EINVAL; |
7a7c381d | 4328 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4329 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4330 | if (limit < 1 || batchcount < 1 || |
4331 | batchcount > limit || shared < 0) { | |
e498be7d | 4332 | res = 0; |
1da177e4 | 4333 | } else { |
e498be7d | 4334 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4335 | batchcount, shared, |
4336 | GFP_KERNEL); | |
1da177e4 LT |
4337 | } |
4338 | break; | |
4339 | } | |
4340 | } | |
fc0abb14 | 4341 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4342 | if (res >= 0) |
4343 | res = count; | |
4344 | return res; | |
4345 | } | |
871751e2 | 4346 | |
7b3c3a50 AD |
4347 | static int slabinfo_open(struct inode *inode, struct file *file) |
4348 | { | |
4349 | return seq_open(file, &slabinfo_op); | |
4350 | } | |
4351 | ||
4352 | static const struct file_operations proc_slabinfo_operations = { | |
4353 | .open = slabinfo_open, | |
4354 | .read = seq_read, | |
4355 | .write = slabinfo_write, | |
4356 | .llseek = seq_lseek, | |
4357 | .release = seq_release, | |
4358 | }; | |
4359 | ||
871751e2 AV |
4360 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4361 | ||
4362 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4363 | { | |
871751e2 | 4364 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4365 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4366 | } |
4367 | ||
4368 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4369 | { | |
4370 | unsigned long *p; | |
4371 | int l; | |
4372 | if (!v) | |
4373 | return 1; | |
4374 | l = n[1]; | |
4375 | p = n + 2; | |
4376 | while (l) { | |
4377 | int i = l/2; | |
4378 | unsigned long *q = p + 2 * i; | |
4379 | if (*q == v) { | |
4380 | q[1]++; | |
4381 | return 1; | |
4382 | } | |
4383 | if (*q > v) { | |
4384 | l = i; | |
4385 | } else { | |
4386 | p = q + 2; | |
4387 | l -= i + 1; | |
4388 | } | |
4389 | } | |
4390 | if (++n[1] == n[0]) | |
4391 | return 0; | |
4392 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4393 | p[0] = v; | |
4394 | p[1] = 1; | |
4395 | return 1; | |
4396 | } | |
4397 | ||
4398 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4399 | { | |
4400 | void *p; | |
4401 | int i; | |
4402 | if (n[0] == n[1]) | |
4403 | return; | |
4404 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4405 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4406 | continue; | |
4407 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4408 | return; | |
4409 | } | |
4410 | } | |
4411 | ||
4412 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4413 | { | |
4414 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4415 | unsigned long offset, size; |
9281acea | 4416 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4417 | |
a5c43dae | 4418 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4419 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4420 | if (modname[0]) |
871751e2 AV |
4421 | seq_printf(m, " [%s]", modname); |
4422 | return; | |
4423 | } | |
4424 | #endif | |
4425 | seq_printf(m, "%p", (void *)address); | |
4426 | } | |
4427 | ||
4428 | static int leaks_show(struct seq_file *m, void *p) | |
4429 | { | |
b92151ba | 4430 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4431 | struct slab *slabp; |
4432 | struct kmem_list3 *l3; | |
4433 | const char *name; | |
4434 | unsigned long *n = m->private; | |
4435 | int node; | |
4436 | int i; | |
4437 | ||
4438 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4439 | return 0; | |
4440 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4441 | return 0; | |
4442 | ||
4443 | /* OK, we can do it */ | |
4444 | ||
4445 | n[1] = 0; | |
4446 | ||
4447 | for_each_online_node(node) { | |
4448 | l3 = cachep->nodelists[node]; | |
4449 | if (!l3) | |
4450 | continue; | |
4451 | ||
4452 | check_irq_on(); | |
4453 | spin_lock_irq(&l3->list_lock); | |
4454 | ||
7a7c381d | 4455 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4456 | handle_slab(n, cachep, slabp); |
7a7c381d | 4457 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4458 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4459 | spin_unlock_irq(&l3->list_lock); |
4460 | } | |
4461 | name = cachep->name; | |
4462 | if (n[0] == n[1]) { | |
4463 | /* Increase the buffer size */ | |
4464 | mutex_unlock(&cache_chain_mutex); | |
4465 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4466 | if (!m->private) { | |
4467 | /* Too bad, we are really out */ | |
4468 | m->private = n; | |
4469 | mutex_lock(&cache_chain_mutex); | |
4470 | return -ENOMEM; | |
4471 | } | |
4472 | *(unsigned long *)m->private = n[0] * 2; | |
4473 | kfree(n); | |
4474 | mutex_lock(&cache_chain_mutex); | |
4475 | /* Now make sure this entry will be retried */ | |
4476 | m->count = m->size; | |
4477 | return 0; | |
4478 | } | |
4479 | for (i = 0; i < n[1]; i++) { | |
4480 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4481 | show_symbol(m, n[2*i+2]); | |
4482 | seq_putc(m, '\n'); | |
4483 | } | |
d2e7b7d0 | 4484 | |
871751e2 AV |
4485 | return 0; |
4486 | } | |
4487 | ||
a0ec95a8 | 4488 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4489 | .start = leaks_start, |
4490 | .next = s_next, | |
4491 | .stop = s_stop, | |
4492 | .show = leaks_show, | |
4493 | }; | |
a0ec95a8 AD |
4494 | |
4495 | static int slabstats_open(struct inode *inode, struct file *file) | |
4496 | { | |
4497 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4498 | int ret = -ENOMEM; | |
4499 | if (n) { | |
4500 | ret = seq_open(file, &slabstats_op); | |
4501 | if (!ret) { | |
4502 | struct seq_file *m = file->private_data; | |
4503 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4504 | m->private = n; | |
4505 | n = NULL; | |
4506 | } | |
4507 | kfree(n); | |
4508 | } | |
4509 | return ret; | |
4510 | } | |
4511 | ||
4512 | static const struct file_operations proc_slabstats_operations = { | |
4513 | .open = slabstats_open, | |
4514 | .read = seq_read, | |
4515 | .llseek = seq_lseek, | |
4516 | .release = seq_release_private, | |
4517 | }; | |
4518 | #endif | |
4519 | ||
4520 | static int __init slab_proc_init(void) | |
4521 | { | |
7b3c3a50 | 4522 | proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4523 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4524 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4525 | #endif |
a0ec95a8 AD |
4526 | return 0; |
4527 | } | |
4528 | module_init(slab_proc_init); | |
1da177e4 LT |
4529 | #endif |
4530 | ||
00e145b6 MS |
4531 | /** |
4532 | * ksize - get the actual amount of memory allocated for a given object | |
4533 | * @objp: Pointer to the object | |
4534 | * | |
4535 | * kmalloc may internally round up allocations and return more memory | |
4536 | * than requested. ksize() can be used to determine the actual amount of | |
4537 | * memory allocated. The caller may use this additional memory, even though | |
4538 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4539 | * The caller must guarantee that objp points to a valid object previously | |
4540 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4541 | * must not be freed during the duration of the call. | |
4542 | */ | |
fd76bab2 | 4543 | size_t ksize(const void *objp) |
1da177e4 | 4544 | { |
ef8b4520 CL |
4545 | BUG_ON(!objp); |
4546 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4547 | return 0; |
1da177e4 | 4548 | |
6ed5eb22 | 4549 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4550 | } |
b1aabecd | 4551 | EXPORT_SYMBOL(ksize); |