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> | |
105 | #include <linux/rcupdate.h> | |
543537bd | 106 | #include <linux/string.h> |
138ae663 | 107 | #include <linux/uaccess.h> |
e498be7d | 108 | #include <linux/nodemask.h> |
dc85da15 | 109 | #include <linux/mempolicy.h> |
fc0abb14 | 110 | #include <linux/mutex.h> |
8a8b6502 | 111 | #include <linux/fault-inject.h> |
e7eebaf6 | 112 | #include <linux/rtmutex.h> |
6a2d7a95 | 113 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 114 | #include <linux/debugobjects.h> |
1da177e4 | 115 | |
1da177e4 LT |
116 | #include <asm/cacheflush.h> |
117 | #include <asm/tlbflush.h> | |
118 | #include <asm/page.h> | |
119 | ||
120 | /* | |
50953fe9 | 121 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
122 | * 0 for faster, smaller code (especially in the critical paths). |
123 | * | |
124 | * STATS - 1 to collect stats for /proc/slabinfo. | |
125 | * 0 for faster, smaller code (especially in the critical paths). | |
126 | * | |
127 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
128 | */ | |
129 | ||
130 | #ifdef CONFIG_DEBUG_SLAB | |
131 | #define DEBUG 1 | |
132 | #define STATS 1 | |
133 | #define FORCED_DEBUG 1 | |
134 | #else | |
135 | #define DEBUG 0 | |
136 | #define STATS 0 | |
137 | #define FORCED_DEBUG 0 | |
138 | #endif | |
139 | ||
1da177e4 LT |
140 | /* Shouldn't this be in a header file somewhere? */ |
141 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 142 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 143 | |
1da177e4 LT |
144 | #ifndef ARCH_KMALLOC_MINALIGN |
145 | /* | |
146 | * Enforce a minimum alignment for the kmalloc caches. | |
147 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
148 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
149 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
b46b8f19 DW |
150 | * alignment larger than the alignment of a 64-bit integer. |
151 | * ARCH_KMALLOC_MINALIGN allows that. | |
152 | * Note that increasing this value may disable some debug features. | |
1da177e4 | 153 | */ |
b46b8f19 | 154 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
1da177e4 LT |
155 | #endif |
156 | ||
157 | #ifndef ARCH_SLAB_MINALIGN | |
158 | /* | |
159 | * Enforce a minimum alignment for all caches. | |
160 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
161 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
162 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
163 | * some debug features. | |
164 | */ | |
165 | #define ARCH_SLAB_MINALIGN 0 | |
166 | #endif | |
167 | ||
168 | #ifndef ARCH_KMALLOC_FLAGS | |
169 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
170 | #endif | |
171 | ||
172 | /* Legal flag mask for kmem_cache_create(). */ | |
173 | #if DEBUG | |
50953fe9 | 174 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 175 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 176 | SLAB_CACHE_DMA | \ |
5af60839 | 177 | SLAB_STORE_USER | \ |
1da177e4 | 178 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a TG |
179 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
180 | SLAB_DEBUG_OBJECTS) | |
1da177e4 | 181 | #else |
ac2b898c | 182 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 183 | SLAB_CACHE_DMA | \ |
1da177e4 | 184 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a TG |
185 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
186 | SLAB_DEBUG_OBJECTS) | |
1da177e4 LT |
187 | #endif |
188 | ||
189 | /* | |
190 | * kmem_bufctl_t: | |
191 | * | |
192 | * Bufctl's are used for linking objs within a slab | |
193 | * linked offsets. | |
194 | * | |
195 | * This implementation relies on "struct page" for locating the cache & | |
196 | * slab an object belongs to. | |
197 | * This allows the bufctl structure to be small (one int), but limits | |
198 | * the number of objects a slab (not a cache) can contain when off-slab | |
199 | * bufctls are used. The limit is the size of the largest general cache | |
200 | * that does not use off-slab slabs. | |
201 | * For 32bit archs with 4 kB pages, is this 56. | |
202 | * This is not serious, as it is only for large objects, when it is unwise | |
203 | * to have too many per slab. | |
204 | * Note: This limit can be raised by introducing a general cache whose size | |
205 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
206 | */ | |
207 | ||
fa5b08d5 | 208 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
209 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
210 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
211 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
212 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 213 | |
1da177e4 LT |
214 | /* |
215 | * struct slab | |
216 | * | |
217 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
218 | * for a slab, or allocated from an general cache. | |
219 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
220 | */ | |
221 | struct slab { | |
b28a02de PE |
222 | struct list_head list; |
223 | unsigned long colouroff; | |
224 | void *s_mem; /* including colour offset */ | |
225 | unsigned int inuse; /* num of objs active in slab */ | |
226 | kmem_bufctl_t free; | |
227 | unsigned short nodeid; | |
1da177e4 LT |
228 | }; |
229 | ||
230 | /* | |
231 | * struct slab_rcu | |
232 | * | |
233 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
234 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
235 | * we need to approach a kernel structure obliquely, from its address | |
236 | * obtained without the usual locking. We can lock the structure to | |
237 | * stabilize it and check it's still at the given address, only if we | |
238 | * can be sure that the memory has not been meanwhile reused for some | |
239 | * other kind of object (which our subsystem's lock might corrupt). | |
240 | * | |
241 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
242 | * taking the spinlock within the structure expected at that address. | |
243 | * | |
244 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
245 | */ | |
246 | struct slab_rcu { | |
b28a02de | 247 | struct rcu_head head; |
343e0d7a | 248 | struct kmem_cache *cachep; |
b28a02de | 249 | void *addr; |
1da177e4 LT |
250 | }; |
251 | ||
252 | /* | |
253 | * struct array_cache | |
254 | * | |
1da177e4 LT |
255 | * Purpose: |
256 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
257 | * - reduce the number of linked list operations | |
258 | * - reduce spinlock operations | |
259 | * | |
260 | * The limit is stored in the per-cpu structure to reduce the data cache | |
261 | * footprint. | |
262 | * | |
263 | */ | |
264 | struct array_cache { | |
265 | unsigned int avail; | |
266 | unsigned int limit; | |
267 | unsigned int batchcount; | |
268 | unsigned int touched; | |
e498be7d | 269 | spinlock_t lock; |
bda5b655 | 270 | void *entry[]; /* |
a737b3e2 AM |
271 | * Must have this definition in here for the proper |
272 | * alignment of array_cache. Also simplifies accessing | |
273 | * the entries. | |
a737b3e2 | 274 | */ |
1da177e4 LT |
275 | }; |
276 | ||
a737b3e2 AM |
277 | /* |
278 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
279 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
280 | */ |
281 | #define BOOT_CPUCACHE_ENTRIES 1 | |
282 | struct arraycache_init { | |
283 | struct array_cache cache; | |
b28a02de | 284 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
285 | }; |
286 | ||
287 | /* | |
e498be7d | 288 | * The slab lists for all objects. |
1da177e4 LT |
289 | */ |
290 | struct kmem_list3 { | |
b28a02de PE |
291 | struct list_head slabs_partial; /* partial list first, better asm code */ |
292 | struct list_head slabs_full; | |
293 | struct list_head slabs_free; | |
294 | unsigned long free_objects; | |
b28a02de | 295 | unsigned int free_limit; |
2e1217cf | 296 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
297 | spinlock_t list_lock; |
298 | struct array_cache *shared; /* shared per node */ | |
299 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
300 | unsigned long next_reap; /* updated without locking */ |
301 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
302 | }; |
303 | ||
e498be7d CL |
304 | /* |
305 | * Need this for bootstrapping a per node allocator. | |
306 | */ | |
556a169d | 307 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
e498be7d CL |
308 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
309 | #define CACHE_CACHE 0 | |
556a169d PE |
310 | #define SIZE_AC MAX_NUMNODES |
311 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 312 | |
ed11d9eb CL |
313 | static int drain_freelist(struct kmem_cache *cache, |
314 | struct kmem_list3 *l3, int tofree); | |
315 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
316 | int node); | |
2ed3a4ef | 317 | static int enable_cpucache(struct kmem_cache *cachep); |
65f27f38 | 318 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 319 | |
e498be7d | 320 | /* |
a737b3e2 AM |
321 | * This function must be completely optimized away if a constant is passed to |
322 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 323 | */ |
7243cc05 | 324 | static __always_inline int index_of(const size_t size) |
e498be7d | 325 | { |
5ec8a847 SR |
326 | extern void __bad_size(void); |
327 | ||
e498be7d CL |
328 | if (__builtin_constant_p(size)) { |
329 | int i = 0; | |
330 | ||
331 | #define CACHE(x) \ | |
332 | if (size <=x) \ | |
333 | return i; \ | |
334 | else \ | |
335 | i++; | |
1c61fc40 | 336 | #include <linux/kmalloc_sizes.h> |
e498be7d | 337 | #undef CACHE |
5ec8a847 | 338 | __bad_size(); |
7243cc05 | 339 | } else |
5ec8a847 | 340 | __bad_size(); |
e498be7d CL |
341 | return 0; |
342 | } | |
343 | ||
e0a42726 IM |
344 | static int slab_early_init = 1; |
345 | ||
e498be7d CL |
346 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
347 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 348 | |
5295a74c | 349 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
350 | { |
351 | INIT_LIST_HEAD(&parent->slabs_full); | |
352 | INIT_LIST_HEAD(&parent->slabs_partial); | |
353 | INIT_LIST_HEAD(&parent->slabs_free); | |
354 | parent->shared = NULL; | |
355 | parent->alien = NULL; | |
2e1217cf | 356 | parent->colour_next = 0; |
e498be7d CL |
357 | spin_lock_init(&parent->list_lock); |
358 | parent->free_objects = 0; | |
359 | parent->free_touched = 0; | |
360 | } | |
361 | ||
a737b3e2 AM |
362 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
363 | do { \ | |
364 | INIT_LIST_HEAD(listp); \ | |
365 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
366 | } while (0) |
367 | ||
a737b3e2 AM |
368 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
369 | do { \ | |
e498be7d CL |
370 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
371 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
372 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
373 | } while (0) | |
1da177e4 LT |
374 | |
375 | /* | |
343e0d7a | 376 | * struct kmem_cache |
1da177e4 LT |
377 | * |
378 | * manages a cache. | |
379 | */ | |
b28a02de | 380 | |
2109a2d1 | 381 | struct kmem_cache { |
1da177e4 | 382 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 383 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 384 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
385 | unsigned int batchcount; |
386 | unsigned int limit; | |
387 | unsigned int shared; | |
b5d8ca7c | 388 | |
3dafccf2 | 389 | unsigned int buffer_size; |
6a2d7a95 | 390 | u32 reciprocal_buffer_size; |
b5d8ca7c | 391 | /* 3) touched by every alloc & free from the backend */ |
b5d8ca7c | 392 | |
a737b3e2 AM |
393 | unsigned int flags; /* constant flags */ |
394 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 395 | |
b5d8ca7c | 396 | /* 4) cache_grow/shrink */ |
1da177e4 | 397 | /* order of pgs per slab (2^n) */ |
b28a02de | 398 | unsigned int gfporder; |
1da177e4 LT |
399 | |
400 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 401 | gfp_t gfpflags; |
1da177e4 | 402 | |
a737b3e2 | 403 | size_t colour; /* cache colouring range */ |
b28a02de | 404 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 405 | struct kmem_cache *slabp_cache; |
b28a02de | 406 | unsigned int slab_size; |
a737b3e2 | 407 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
408 | |
409 | /* constructor func */ | |
51cc5068 | 410 | void (*ctor)(void *obj); |
1da177e4 | 411 | |
b5d8ca7c | 412 | /* 5) cache creation/removal */ |
b28a02de PE |
413 | const char *name; |
414 | struct list_head next; | |
1da177e4 | 415 | |
b5d8ca7c | 416 | /* 6) statistics */ |
1da177e4 | 417 | #if STATS |
b28a02de PE |
418 | unsigned long num_active; |
419 | unsigned long num_allocations; | |
420 | unsigned long high_mark; | |
421 | unsigned long grown; | |
422 | unsigned long reaped; | |
423 | unsigned long errors; | |
424 | unsigned long max_freeable; | |
425 | unsigned long node_allocs; | |
426 | unsigned long node_frees; | |
fb7faf33 | 427 | unsigned long node_overflow; |
b28a02de PE |
428 | atomic_t allochit; |
429 | atomic_t allocmiss; | |
430 | atomic_t freehit; | |
431 | atomic_t freemiss; | |
1da177e4 LT |
432 | #endif |
433 | #if DEBUG | |
3dafccf2 MS |
434 | /* |
435 | * If debugging is enabled, then the allocator can add additional | |
436 | * fields and/or padding to every object. buffer_size contains the total | |
437 | * object size including these internal fields, the following two | |
438 | * variables contain the offset to the user object and its size. | |
439 | */ | |
440 | int obj_offset; | |
441 | int obj_size; | |
1da177e4 | 442 | #endif |
8da3430d ED |
443 | /* |
444 | * We put nodelists[] at the end of kmem_cache, because we want to size | |
445 | * this array to nr_node_ids slots instead of MAX_NUMNODES | |
446 | * (see kmem_cache_init()) | |
447 | * We still use [MAX_NUMNODES] and not [1] or [0] because cache_cache | |
448 | * is statically defined, so we reserve the max number of nodes. | |
449 | */ | |
450 | struct kmem_list3 *nodelists[MAX_NUMNODES]; | |
451 | /* | |
452 | * Do not add fields after nodelists[] | |
453 | */ | |
1da177e4 LT |
454 | }; |
455 | ||
456 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
457 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
458 | ||
459 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
460 | /* |
461 | * Optimization question: fewer reaps means less probability for unnessary | |
462 | * cpucache drain/refill cycles. | |
1da177e4 | 463 | * |
dc6f3f27 | 464 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
465 | * which could lock up otherwise freeable slabs. |
466 | */ | |
467 | #define REAPTIMEOUT_CPUC (2*HZ) | |
468 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
469 | ||
470 | #if STATS | |
471 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
472 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
473 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
474 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 475 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
476 | #define STATS_SET_HIGH(x) \ |
477 | do { \ | |
478 | if ((x)->num_active > (x)->high_mark) \ | |
479 | (x)->high_mark = (x)->num_active; \ | |
480 | } while (0) | |
1da177e4 LT |
481 | #define STATS_INC_ERR(x) ((x)->errors++) |
482 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 483 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 484 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
485 | #define STATS_SET_FREEABLE(x, i) \ |
486 | do { \ | |
487 | if ((x)->max_freeable < i) \ | |
488 | (x)->max_freeable = i; \ | |
489 | } while (0) | |
1da177e4 LT |
490 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
491 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
492 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
493 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
494 | #else | |
495 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
496 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
497 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
498 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 499 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
500 | #define STATS_SET_HIGH(x) do { } while (0) |
501 | #define STATS_INC_ERR(x) do { } while (0) | |
502 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 503 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 504 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 505 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
506 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
507 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
508 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
509 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
510 | #endif | |
511 | ||
512 | #if DEBUG | |
1da177e4 | 513 | |
a737b3e2 AM |
514 | /* |
515 | * memory layout of objects: | |
1da177e4 | 516 | * 0 : objp |
3dafccf2 | 517 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
518 | * the end of an object is aligned with the end of the real |
519 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 520 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 521 | * redzone word. |
3dafccf2 MS |
522 | * cachep->obj_offset: The real object. |
523 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
524 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
525 | * [BYTES_PER_WORD long] | |
1da177e4 | 526 | */ |
343e0d7a | 527 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 528 | { |
3dafccf2 | 529 | return cachep->obj_offset; |
1da177e4 LT |
530 | } |
531 | ||
343e0d7a | 532 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 533 | { |
3dafccf2 | 534 | return cachep->obj_size; |
1da177e4 LT |
535 | } |
536 | ||
b46b8f19 | 537 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
538 | { |
539 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
540 | return (unsigned long long*) (objp + obj_offset(cachep) - |
541 | sizeof(unsigned long long)); | |
1da177e4 LT |
542 | } |
543 | ||
b46b8f19 | 544 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
545 | { |
546 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
547 | if (cachep->flags & SLAB_STORE_USER) | |
b46b8f19 DW |
548 | return (unsigned long long *)(objp + cachep->buffer_size - |
549 | sizeof(unsigned long long) - | |
87a927c7 | 550 | REDZONE_ALIGN); |
b46b8f19 DW |
551 | return (unsigned long long *) (objp + cachep->buffer_size - |
552 | sizeof(unsigned long long)); | |
1da177e4 LT |
553 | } |
554 | ||
343e0d7a | 555 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
556 | { |
557 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 558 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
559 | } |
560 | ||
561 | #else | |
562 | ||
3dafccf2 MS |
563 | #define obj_offset(x) 0 |
564 | #define obj_size(cachep) (cachep->buffer_size) | |
b46b8f19 DW |
565 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
566 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
567 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
568 | ||
569 | #endif | |
570 | ||
1da177e4 LT |
571 | /* |
572 | * Do not go above this order unless 0 objects fit into the slab. | |
573 | */ | |
574 | #define BREAK_GFP_ORDER_HI 1 | |
575 | #define BREAK_GFP_ORDER_LO 0 | |
576 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
577 | ||
a737b3e2 AM |
578 | /* |
579 | * Functions for storing/retrieving the cachep and or slab from the page | |
580 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
581 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 582 | */ |
065d41cb PE |
583 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
584 | { | |
585 | page->lru.next = (struct list_head *)cache; | |
586 | } | |
587 | ||
588 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
589 | { | |
d85f3385 | 590 | page = compound_head(page); |
ddc2e812 | 591 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
592 | return (struct kmem_cache *)page->lru.next; |
593 | } | |
594 | ||
595 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
596 | { | |
597 | page->lru.prev = (struct list_head *)slab; | |
598 | } | |
599 | ||
600 | static inline struct slab *page_get_slab(struct page *page) | |
601 | { | |
ddc2e812 | 602 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
603 | return (struct slab *)page->lru.prev; |
604 | } | |
1da177e4 | 605 | |
6ed5eb22 PE |
606 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
607 | { | |
b49af68f | 608 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
609 | return page_get_cache(page); |
610 | } | |
611 | ||
612 | static inline struct slab *virt_to_slab(const void *obj) | |
613 | { | |
b49af68f | 614 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
615 | return page_get_slab(page); |
616 | } | |
617 | ||
8fea4e96 PE |
618 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
619 | unsigned int idx) | |
620 | { | |
621 | return slab->s_mem + cache->buffer_size * idx; | |
622 | } | |
623 | ||
6a2d7a95 ED |
624 | /* |
625 | * We want to avoid an expensive divide : (offset / cache->buffer_size) | |
626 | * Using the fact that buffer_size is a constant for a particular cache, | |
627 | * we can replace (offset / cache->buffer_size) by | |
628 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
629 | */ | |
630 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
631 | const struct slab *slab, void *obj) | |
8fea4e96 | 632 | { |
6a2d7a95 ED |
633 | u32 offset = (obj - slab->s_mem); |
634 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
635 | } |
636 | ||
a737b3e2 AM |
637 | /* |
638 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
639 | */ | |
1da177e4 LT |
640 | struct cache_sizes malloc_sizes[] = { |
641 | #define CACHE(x) { .cs_size = (x) }, | |
642 | #include <linux/kmalloc_sizes.h> | |
643 | CACHE(ULONG_MAX) | |
644 | #undef CACHE | |
645 | }; | |
646 | EXPORT_SYMBOL(malloc_sizes); | |
647 | ||
648 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
649 | struct cache_names { | |
650 | char *name; | |
651 | char *name_dma; | |
652 | }; | |
653 | ||
654 | static struct cache_names __initdata cache_names[] = { | |
655 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
656 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 657 | {NULL,} |
1da177e4 LT |
658 | #undef CACHE |
659 | }; | |
660 | ||
661 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 662 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 663 | static struct arraycache_init initarray_generic = |
b28a02de | 664 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
665 | |
666 | /* internal cache of cache description objs */ | |
343e0d7a | 667 | static struct kmem_cache cache_cache = { |
b28a02de PE |
668 | .batchcount = 1, |
669 | .limit = BOOT_CPUCACHE_ENTRIES, | |
670 | .shared = 1, | |
343e0d7a | 671 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 672 | .name = "kmem_cache", |
1da177e4 LT |
673 | }; |
674 | ||
056c6241 RT |
675 | #define BAD_ALIEN_MAGIC 0x01020304ul |
676 | ||
f1aaee53 AV |
677 | #ifdef CONFIG_LOCKDEP |
678 | ||
679 | /* | |
680 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
681 | * for other slabs "off slab". | |
682 | * The locking for this is tricky in that it nests within the locks | |
683 | * of all other slabs in a few places; to deal with this special | |
684 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
685 | * |
686 | * We set lock class for alien array caches which are up during init. | |
687 | * The lock annotation will be lost if all cpus of a node goes down and | |
688 | * then comes back up during hotplug | |
f1aaee53 | 689 | */ |
056c6241 RT |
690 | static struct lock_class_key on_slab_l3_key; |
691 | static struct lock_class_key on_slab_alc_key; | |
692 | ||
693 | static inline void init_lock_keys(void) | |
f1aaee53 | 694 | |
f1aaee53 AV |
695 | { |
696 | int q; | |
056c6241 RT |
697 | struct cache_sizes *s = malloc_sizes; |
698 | ||
699 | while (s->cs_size != ULONG_MAX) { | |
700 | for_each_node(q) { | |
701 | struct array_cache **alc; | |
702 | int r; | |
703 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
704 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
705 | continue; | |
706 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
707 | alc = l3->alien; | |
708 | /* | |
709 | * FIXME: This check for BAD_ALIEN_MAGIC | |
710 | * should go away when common slab code is taught to | |
711 | * work even without alien caches. | |
712 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
713 | * for alloc_alien_cache, | |
714 | */ | |
715 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
716 | continue; | |
717 | for_each_node(r) { | |
718 | if (alc[r]) | |
719 | lockdep_set_class(&alc[r]->lock, | |
720 | &on_slab_alc_key); | |
721 | } | |
722 | } | |
723 | s++; | |
f1aaee53 AV |
724 | } |
725 | } | |
f1aaee53 | 726 | #else |
056c6241 | 727 | static inline void init_lock_keys(void) |
f1aaee53 AV |
728 | { |
729 | } | |
730 | #endif | |
731 | ||
8f5be20b | 732 | /* |
95402b38 | 733 | * Guard access to the cache-chain. |
8f5be20b | 734 | */ |
fc0abb14 | 735 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
736 | static struct list_head cache_chain; |
737 | ||
1da177e4 LT |
738 | /* |
739 | * chicken and egg problem: delay the per-cpu array allocation | |
740 | * until the general caches are up. | |
741 | */ | |
742 | static enum { | |
743 | NONE, | |
e498be7d CL |
744 | PARTIAL_AC, |
745 | PARTIAL_L3, | |
1da177e4 LT |
746 | FULL |
747 | } g_cpucache_up; | |
748 | ||
39d24e64 MK |
749 | /* |
750 | * used by boot code to determine if it can use slab based allocator | |
751 | */ | |
752 | int slab_is_available(void) | |
753 | { | |
754 | return g_cpucache_up == FULL; | |
755 | } | |
756 | ||
52bad64d | 757 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 758 | |
343e0d7a | 759 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
760 | { |
761 | return cachep->array[smp_processor_id()]; | |
762 | } | |
763 | ||
a737b3e2 AM |
764 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
765 | gfp_t gfpflags) | |
1da177e4 LT |
766 | { |
767 | struct cache_sizes *csizep = malloc_sizes; | |
768 | ||
769 | #if DEBUG | |
770 | /* This happens if someone tries to call | |
b28a02de PE |
771 | * kmem_cache_create(), or __kmalloc(), before |
772 | * the generic caches are initialized. | |
773 | */ | |
c7e43c78 | 774 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 775 | #endif |
6cb8f913 CL |
776 | if (!size) |
777 | return ZERO_SIZE_PTR; | |
778 | ||
1da177e4 LT |
779 | while (size > csizep->cs_size) |
780 | csizep++; | |
781 | ||
782 | /* | |
0abf40c1 | 783 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
784 | * has cs_{dma,}cachep==NULL. Thus no special case |
785 | * for large kmalloc calls required. | |
786 | */ | |
4b51d669 | 787 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
788 | if (unlikely(gfpflags & GFP_DMA)) |
789 | return csizep->cs_dmacachep; | |
4b51d669 | 790 | #endif |
1da177e4 LT |
791 | return csizep->cs_cachep; |
792 | } | |
793 | ||
b221385b | 794 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
795 | { |
796 | return __find_general_cachep(size, gfpflags); | |
797 | } | |
97e2bde4 | 798 | |
fbaccacf | 799 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 800 | { |
fbaccacf SR |
801 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
802 | } | |
1da177e4 | 803 | |
a737b3e2 AM |
804 | /* |
805 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
806 | */ | |
fbaccacf SR |
807 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
808 | size_t align, int flags, size_t *left_over, | |
809 | unsigned int *num) | |
810 | { | |
811 | int nr_objs; | |
812 | size_t mgmt_size; | |
813 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 814 | |
fbaccacf SR |
815 | /* |
816 | * The slab management structure can be either off the slab or | |
817 | * on it. For the latter case, the memory allocated for a | |
818 | * slab is used for: | |
819 | * | |
820 | * - The struct slab | |
821 | * - One kmem_bufctl_t for each object | |
822 | * - Padding to respect alignment of @align | |
823 | * - @buffer_size bytes for each object | |
824 | * | |
825 | * If the slab management structure is off the slab, then the | |
826 | * alignment will already be calculated into the size. Because | |
827 | * the slabs are all pages aligned, the objects will be at the | |
828 | * correct alignment when allocated. | |
829 | */ | |
830 | if (flags & CFLGS_OFF_SLAB) { | |
831 | mgmt_size = 0; | |
832 | nr_objs = slab_size / buffer_size; | |
833 | ||
834 | if (nr_objs > SLAB_LIMIT) | |
835 | nr_objs = SLAB_LIMIT; | |
836 | } else { | |
837 | /* | |
838 | * Ignore padding for the initial guess. The padding | |
839 | * is at most @align-1 bytes, and @buffer_size is at | |
840 | * least @align. In the worst case, this result will | |
841 | * be one greater than the number of objects that fit | |
842 | * into the memory allocation when taking the padding | |
843 | * into account. | |
844 | */ | |
845 | nr_objs = (slab_size - sizeof(struct slab)) / | |
846 | (buffer_size + sizeof(kmem_bufctl_t)); | |
847 | ||
848 | /* | |
849 | * This calculated number will be either the right | |
850 | * amount, or one greater than what we want. | |
851 | */ | |
852 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
853 | > slab_size) | |
854 | nr_objs--; | |
855 | ||
856 | if (nr_objs > SLAB_LIMIT) | |
857 | nr_objs = SLAB_LIMIT; | |
858 | ||
859 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
860 | } | |
861 | *num = nr_objs; | |
862 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
863 | } |
864 | ||
d40cee24 | 865 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 866 | |
a737b3e2 AM |
867 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
868 | char *msg) | |
1da177e4 LT |
869 | { |
870 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 871 | function, cachep->name, msg); |
1da177e4 LT |
872 | dump_stack(); |
873 | } | |
874 | ||
3395ee05 PM |
875 | /* |
876 | * By default on NUMA we use alien caches to stage the freeing of | |
877 | * objects allocated from other nodes. This causes massive memory | |
878 | * inefficiencies when using fake NUMA setup to split memory into a | |
879 | * large number of small nodes, so it can be disabled on the command | |
880 | * line | |
881 | */ | |
882 | ||
883 | static int use_alien_caches __read_mostly = 1; | |
1807a1aa | 884 | static int numa_platform __read_mostly = 1; |
3395ee05 PM |
885 | static int __init noaliencache_setup(char *s) |
886 | { | |
887 | use_alien_caches = 0; | |
888 | return 1; | |
889 | } | |
890 | __setup("noaliencache", noaliencache_setup); | |
891 | ||
8fce4d8e CL |
892 | #ifdef CONFIG_NUMA |
893 | /* | |
894 | * Special reaping functions for NUMA systems called from cache_reap(). | |
895 | * These take care of doing round robin flushing of alien caches (containing | |
896 | * objects freed on different nodes from which they were allocated) and the | |
897 | * flushing of remote pcps by calling drain_node_pages. | |
898 | */ | |
899 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
900 | ||
901 | static void init_reap_node(int cpu) | |
902 | { | |
903 | int node; | |
904 | ||
905 | node = next_node(cpu_to_node(cpu), node_online_map); | |
906 | if (node == MAX_NUMNODES) | |
442295c9 | 907 | node = first_node(node_online_map); |
8fce4d8e | 908 | |
7f6b8876 | 909 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
910 | } |
911 | ||
912 | static void next_reap_node(void) | |
913 | { | |
914 | int node = __get_cpu_var(reap_node); | |
915 | ||
8fce4d8e CL |
916 | node = next_node(node, node_online_map); |
917 | if (unlikely(node >= MAX_NUMNODES)) | |
918 | node = first_node(node_online_map); | |
919 | __get_cpu_var(reap_node) = node; | |
920 | } | |
921 | ||
922 | #else | |
923 | #define init_reap_node(cpu) do { } while (0) | |
924 | #define next_reap_node(void) do { } while (0) | |
925 | #endif | |
926 | ||
1da177e4 LT |
927 | /* |
928 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
929 | * via the workqueue/eventd. | |
930 | * Add the CPU number into the expiration time to minimize the possibility of | |
931 | * the CPUs getting into lockstep and contending for the global cache chain | |
932 | * lock. | |
933 | */ | |
897e679b | 934 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 935 | { |
52bad64d | 936 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
937 | |
938 | /* | |
939 | * When this gets called from do_initcalls via cpucache_init(), | |
940 | * init_workqueues() has already run, so keventd will be setup | |
941 | * at that time. | |
942 | */ | |
52bad64d | 943 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 944 | init_reap_node(cpu); |
65f27f38 | 945 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
946 | schedule_delayed_work_on(cpu, reap_work, |
947 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
948 | } |
949 | } | |
950 | ||
e498be7d | 951 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 952 | int batchcount) |
1da177e4 | 953 | { |
b28a02de | 954 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
955 | struct array_cache *nc = NULL; |
956 | ||
e498be7d | 957 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
958 | if (nc) { |
959 | nc->avail = 0; | |
960 | nc->limit = entries; | |
961 | nc->batchcount = batchcount; | |
962 | nc->touched = 0; | |
e498be7d | 963 | spin_lock_init(&nc->lock); |
1da177e4 LT |
964 | } |
965 | return nc; | |
966 | } | |
967 | ||
3ded175a CL |
968 | /* |
969 | * Transfer objects in one arraycache to another. | |
970 | * Locking must be handled by the caller. | |
971 | * | |
972 | * Return the number of entries transferred. | |
973 | */ | |
974 | static int transfer_objects(struct array_cache *to, | |
975 | struct array_cache *from, unsigned int max) | |
976 | { | |
977 | /* Figure out how many entries to transfer */ | |
978 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
979 | ||
980 | if (!nr) | |
981 | return 0; | |
982 | ||
983 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
984 | sizeof(void *) *nr); | |
985 | ||
986 | from->avail -= nr; | |
987 | to->avail += nr; | |
988 | to->touched = 1; | |
989 | return nr; | |
990 | } | |
991 | ||
765c4507 CL |
992 | #ifndef CONFIG_NUMA |
993 | ||
994 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
995 | #define reap_alien(cachep, l3) do { } while (0) | |
996 | ||
997 | static inline struct array_cache **alloc_alien_cache(int node, int limit) | |
998 | { | |
999 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1000 | } | |
1001 | ||
1002 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1003 | { | |
1004 | } | |
1005 | ||
1006 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1007 | { | |
1008 | return 0; | |
1009 | } | |
1010 | ||
1011 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1012 | gfp_t flags) | |
1013 | { | |
1014 | return NULL; | |
1015 | } | |
1016 | ||
8b98c169 | 1017 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1018 | gfp_t flags, int nodeid) |
1019 | { | |
1020 | return NULL; | |
1021 | } | |
1022 | ||
1023 | #else /* CONFIG_NUMA */ | |
1024 | ||
8b98c169 | 1025 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1026 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1027 | |
5295a74c | 1028 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
1029 | { |
1030 | struct array_cache **ac_ptr; | |
8ef82866 | 1031 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1032 | int i; |
1033 | ||
1034 | if (limit > 1) | |
1035 | limit = 12; | |
1036 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
1037 | if (ac_ptr) { | |
1038 | for_each_node(i) { | |
1039 | if (i == node || !node_online(i)) { | |
1040 | ac_ptr[i] = NULL; | |
1041 | continue; | |
1042 | } | |
1043 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
1044 | if (!ac_ptr[i]) { | |
cc550def | 1045 | for (i--; i >= 0; i--) |
e498be7d CL |
1046 | kfree(ac_ptr[i]); |
1047 | kfree(ac_ptr); | |
1048 | return NULL; | |
1049 | } | |
1050 | } | |
1051 | } | |
1052 | return ac_ptr; | |
1053 | } | |
1054 | ||
5295a74c | 1055 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1056 | { |
1057 | int i; | |
1058 | ||
1059 | if (!ac_ptr) | |
1060 | return; | |
e498be7d | 1061 | for_each_node(i) |
b28a02de | 1062 | kfree(ac_ptr[i]); |
e498be7d CL |
1063 | kfree(ac_ptr); |
1064 | } | |
1065 | ||
343e0d7a | 1066 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1067 | struct array_cache *ac, int node) |
e498be7d CL |
1068 | { |
1069 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1070 | ||
1071 | if (ac->avail) { | |
1072 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1073 | /* |
1074 | * Stuff objects into the remote nodes shared array first. | |
1075 | * That way we could avoid the overhead of putting the objects | |
1076 | * into the free lists and getting them back later. | |
1077 | */ | |
693f7d36 JS |
1078 | if (rl3->shared) |
1079 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1080 | |
ff69416e | 1081 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1082 | ac->avail = 0; |
1083 | spin_unlock(&rl3->list_lock); | |
1084 | } | |
1085 | } | |
1086 | ||
8fce4d8e CL |
1087 | /* |
1088 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1089 | */ | |
1090 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1091 | { | |
1092 | int node = __get_cpu_var(reap_node); | |
1093 | ||
1094 | if (l3->alien) { | |
1095 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1096 | |
1097 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1098 | __drain_alien_cache(cachep, ac, node); |
1099 | spin_unlock_irq(&ac->lock); | |
1100 | } | |
1101 | } | |
1102 | } | |
1103 | ||
a737b3e2 AM |
1104 | static void drain_alien_cache(struct kmem_cache *cachep, |
1105 | struct array_cache **alien) | |
e498be7d | 1106 | { |
b28a02de | 1107 | int i = 0; |
e498be7d CL |
1108 | struct array_cache *ac; |
1109 | unsigned long flags; | |
1110 | ||
1111 | for_each_online_node(i) { | |
4484ebf1 | 1112 | ac = alien[i]; |
e498be7d CL |
1113 | if (ac) { |
1114 | spin_lock_irqsave(&ac->lock, flags); | |
1115 | __drain_alien_cache(cachep, ac, i); | |
1116 | spin_unlock_irqrestore(&ac->lock, flags); | |
1117 | } | |
1118 | } | |
1119 | } | |
729bd0b7 | 1120 | |
873623df | 1121 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1122 | { |
1123 | struct slab *slabp = virt_to_slab(objp); | |
1124 | int nodeid = slabp->nodeid; | |
1125 | struct kmem_list3 *l3; | |
1126 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1127 | int node; |
1128 | ||
1129 | node = numa_node_id(); | |
729bd0b7 PE |
1130 | |
1131 | /* | |
1132 | * Make sure we are not freeing a object from another node to the array | |
1133 | * cache on this cpu. | |
1134 | */ | |
62918a03 | 1135 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1136 | return 0; |
1137 | ||
1ca4cb24 | 1138 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1139 | STATS_INC_NODEFREES(cachep); |
1140 | if (l3->alien && l3->alien[nodeid]) { | |
1141 | alien = l3->alien[nodeid]; | |
873623df | 1142 | spin_lock(&alien->lock); |
729bd0b7 PE |
1143 | if (unlikely(alien->avail == alien->limit)) { |
1144 | STATS_INC_ACOVERFLOW(cachep); | |
1145 | __drain_alien_cache(cachep, alien, nodeid); | |
1146 | } | |
1147 | alien->entry[alien->avail++] = objp; | |
1148 | spin_unlock(&alien->lock); | |
1149 | } else { | |
1150 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1151 | free_block(cachep, &objp, 1, nodeid); | |
1152 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1153 | } | |
1154 | return 1; | |
1155 | } | |
e498be7d CL |
1156 | #endif |
1157 | ||
fbf1e473 AM |
1158 | static void __cpuinit cpuup_canceled(long cpu) |
1159 | { | |
1160 | struct kmem_cache *cachep; | |
1161 | struct kmem_list3 *l3 = NULL; | |
1162 | int node = cpu_to_node(cpu); | |
c5f59f08 | 1163 | node_to_cpumask_ptr(mask, node); |
fbf1e473 AM |
1164 | |
1165 | list_for_each_entry(cachep, &cache_chain, next) { | |
1166 | struct array_cache *nc; | |
1167 | struct array_cache *shared; | |
1168 | struct array_cache **alien; | |
fbf1e473 | 1169 | |
fbf1e473 AM |
1170 | /* cpu is dead; no one can alloc from it. */ |
1171 | nc = cachep->array[cpu]; | |
1172 | cachep->array[cpu] = NULL; | |
1173 | l3 = cachep->nodelists[node]; | |
1174 | ||
1175 | if (!l3) | |
1176 | goto free_array_cache; | |
1177 | ||
1178 | spin_lock_irq(&l3->list_lock); | |
1179 | ||
1180 | /* Free limit for this kmem_list3 */ | |
1181 | l3->free_limit -= cachep->batchcount; | |
1182 | if (nc) | |
1183 | free_block(cachep, nc->entry, nc->avail, node); | |
1184 | ||
c5f59f08 | 1185 | if (!cpus_empty(*mask)) { |
fbf1e473 AM |
1186 | spin_unlock_irq(&l3->list_lock); |
1187 | goto free_array_cache; | |
1188 | } | |
1189 | ||
1190 | shared = l3->shared; | |
1191 | if (shared) { | |
1192 | free_block(cachep, shared->entry, | |
1193 | shared->avail, node); | |
1194 | l3->shared = NULL; | |
1195 | } | |
1196 | ||
1197 | alien = l3->alien; | |
1198 | l3->alien = NULL; | |
1199 | ||
1200 | spin_unlock_irq(&l3->list_lock); | |
1201 | ||
1202 | kfree(shared); | |
1203 | if (alien) { | |
1204 | drain_alien_cache(cachep, alien); | |
1205 | free_alien_cache(alien); | |
1206 | } | |
1207 | free_array_cache: | |
1208 | kfree(nc); | |
1209 | } | |
1210 | /* | |
1211 | * In the previous loop, all the objects were freed to | |
1212 | * the respective cache's slabs, now we can go ahead and | |
1213 | * shrink each nodelist to its limit. | |
1214 | */ | |
1215 | list_for_each_entry(cachep, &cache_chain, next) { | |
1216 | l3 = cachep->nodelists[node]; | |
1217 | if (!l3) | |
1218 | continue; | |
1219 | drain_freelist(cachep, l3, l3->free_objects); | |
1220 | } | |
1221 | } | |
1222 | ||
1223 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1224 | { |
343e0d7a | 1225 | struct kmem_cache *cachep; |
e498be7d CL |
1226 | struct kmem_list3 *l3 = NULL; |
1227 | int node = cpu_to_node(cpu); | |
ea02e3dd | 1228 | const int memsize = sizeof(struct kmem_list3); |
1da177e4 | 1229 | |
fbf1e473 AM |
1230 | /* |
1231 | * We need to do this right in the beginning since | |
1232 | * alloc_arraycache's are going to use this list. | |
1233 | * kmalloc_node allows us to add the slab to the right | |
1234 | * kmem_list3 and not this cpu's kmem_list3 | |
1235 | */ | |
1236 | ||
1237 | list_for_each_entry(cachep, &cache_chain, next) { | |
a737b3e2 | 1238 | /* |
fbf1e473 AM |
1239 | * Set up the size64 kmemlist for cpu before we can |
1240 | * begin anything. Make sure some other cpu on this | |
1241 | * node has not already allocated this | |
e498be7d | 1242 | */ |
fbf1e473 AM |
1243 | if (!cachep->nodelists[node]) { |
1244 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1245 | if (!l3) | |
1246 | goto bad; | |
1247 | kmem_list3_init(l3); | |
1248 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1249 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1250 | |
a737b3e2 | 1251 | /* |
fbf1e473 AM |
1252 | * The l3s don't come and go as CPUs come and |
1253 | * go. cache_chain_mutex is sufficient | |
1254 | * protection here. | |
e498be7d | 1255 | */ |
fbf1e473 | 1256 | cachep->nodelists[node] = l3; |
e498be7d CL |
1257 | } |
1258 | ||
fbf1e473 AM |
1259 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1260 | cachep->nodelists[node]->free_limit = | |
1261 | (1 + nr_cpus_node(node)) * | |
1262 | cachep->batchcount + cachep->num; | |
1263 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1264 | } | |
1265 | ||
1266 | /* | |
1267 | * Now we can go ahead with allocating the shared arrays and | |
1268 | * array caches | |
1269 | */ | |
1270 | list_for_each_entry(cachep, &cache_chain, next) { | |
1271 | struct array_cache *nc; | |
1272 | struct array_cache *shared = NULL; | |
1273 | struct array_cache **alien = NULL; | |
1274 | ||
1275 | nc = alloc_arraycache(node, cachep->limit, | |
1276 | cachep->batchcount); | |
1277 | if (!nc) | |
1278 | goto bad; | |
1279 | if (cachep->shared) { | |
1280 | shared = alloc_arraycache(node, | |
1281 | cachep->shared * cachep->batchcount, | |
1282 | 0xbaadf00d); | |
12d00f6a AM |
1283 | if (!shared) { |
1284 | kfree(nc); | |
1da177e4 | 1285 | goto bad; |
12d00f6a | 1286 | } |
fbf1e473 AM |
1287 | } |
1288 | if (use_alien_caches) { | |
1289 | alien = alloc_alien_cache(node, cachep->limit); | |
12d00f6a AM |
1290 | if (!alien) { |
1291 | kfree(shared); | |
1292 | kfree(nc); | |
fbf1e473 | 1293 | goto bad; |
12d00f6a | 1294 | } |
fbf1e473 AM |
1295 | } |
1296 | cachep->array[cpu] = nc; | |
1297 | l3 = cachep->nodelists[node]; | |
1298 | BUG_ON(!l3); | |
1299 | ||
1300 | spin_lock_irq(&l3->list_lock); | |
1301 | if (!l3->shared) { | |
1302 | /* | |
1303 | * We are serialised from CPU_DEAD or | |
1304 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1305 | */ | |
1306 | l3->shared = shared; | |
1307 | shared = NULL; | |
1308 | } | |
4484ebf1 | 1309 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1310 | if (!l3->alien) { |
1311 | l3->alien = alien; | |
1312 | alien = NULL; | |
1da177e4 | 1313 | } |
fbf1e473 AM |
1314 | #endif |
1315 | spin_unlock_irq(&l3->list_lock); | |
1316 | kfree(shared); | |
1317 | free_alien_cache(alien); | |
1318 | } | |
1319 | return 0; | |
1320 | bad: | |
12d00f6a | 1321 | cpuup_canceled(cpu); |
fbf1e473 AM |
1322 | return -ENOMEM; |
1323 | } | |
1324 | ||
1325 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1326 | unsigned long action, void *hcpu) | |
1327 | { | |
1328 | long cpu = (long)hcpu; | |
1329 | int err = 0; | |
1330 | ||
1331 | switch (action) { | |
fbf1e473 AM |
1332 | case CPU_UP_PREPARE: |
1333 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1334 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1335 | err = cpuup_prepare(cpu); |
95402b38 | 1336 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1337 | break; |
1338 | case CPU_ONLINE: | |
8bb78442 | 1339 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1340 | start_cpu_timer(cpu); |
1341 | break; | |
1342 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1343 | case CPU_DOWN_PREPARE: |
8bb78442 | 1344 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1345 | /* |
1346 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1347 | * held so that if cache_reap() is invoked it cannot do | |
1348 | * anything expensive but will only modify reap_work | |
1349 | * and reschedule the timer. | |
1350 | */ | |
1351 | cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); | |
1352 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1353 | per_cpu(reap_work, cpu).work.func = NULL; | |
1354 | break; | |
1355 | case CPU_DOWN_FAILED: | |
8bb78442 | 1356 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1357 | start_cpu_timer(cpu); |
1358 | break; | |
1da177e4 | 1359 | case CPU_DEAD: |
8bb78442 | 1360 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1361 | /* |
1362 | * Even if all the cpus of a node are down, we don't free the | |
1363 | * kmem_list3 of any cache. This to avoid a race between | |
1364 | * cpu_down, and a kmalloc allocation from another cpu for | |
1365 | * memory from the node of the cpu going down. The list3 | |
1366 | * structure is usually allocated from kmem_cache_create() and | |
1367 | * gets destroyed at kmem_cache_destroy(). | |
1368 | */ | |
183ff22b | 1369 | /* fall through */ |
8f5be20b | 1370 | #endif |
1da177e4 | 1371 | case CPU_UP_CANCELED: |
8bb78442 | 1372 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1373 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1374 | cpuup_canceled(cpu); |
fc0abb14 | 1375 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1376 | break; |
1da177e4 | 1377 | } |
fbf1e473 | 1378 | return err ? NOTIFY_BAD : NOTIFY_OK; |
1da177e4 LT |
1379 | } |
1380 | ||
74b85f37 CS |
1381 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1382 | &cpuup_callback, NULL, 0 | |
1383 | }; | |
1da177e4 | 1384 | |
e498be7d CL |
1385 | /* |
1386 | * swap the static kmem_list3 with kmalloced memory | |
1387 | */ | |
a737b3e2 AM |
1388 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1389 | int nodeid) | |
e498be7d CL |
1390 | { |
1391 | struct kmem_list3 *ptr; | |
1392 | ||
e498be7d CL |
1393 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); |
1394 | BUG_ON(!ptr); | |
1395 | ||
1396 | local_irq_disable(); | |
1397 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1398 | /* |
1399 | * Do not assume that spinlocks can be initialized via memcpy: | |
1400 | */ | |
1401 | spin_lock_init(&ptr->list_lock); | |
1402 | ||
e498be7d CL |
1403 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1404 | cachep->nodelists[nodeid] = ptr; | |
1405 | local_irq_enable(); | |
1406 | } | |
1407 | ||
556a169d PE |
1408 | /* |
1409 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1410 | * size of kmem_list3. | |
1411 | */ | |
1412 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1413 | { | |
1414 | int node; | |
1415 | ||
1416 | for_each_online_node(node) { | |
1417 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1418 | cachep->nodelists[node]->next_reap = jiffies + | |
1419 | REAPTIMEOUT_LIST3 + | |
1420 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1421 | } | |
1422 | } | |
1423 | ||
a737b3e2 AM |
1424 | /* |
1425 | * Initialisation. Called after the page allocator have been initialised and | |
1426 | * before smp_init(). | |
1da177e4 LT |
1427 | */ |
1428 | void __init kmem_cache_init(void) | |
1429 | { | |
1430 | size_t left_over; | |
1431 | struct cache_sizes *sizes; | |
1432 | struct cache_names *names; | |
e498be7d | 1433 | int i; |
07ed76b2 | 1434 | int order; |
1ca4cb24 | 1435 | int node; |
e498be7d | 1436 | |
1807a1aa | 1437 | if (num_possible_nodes() == 1) { |
62918a03 | 1438 | use_alien_caches = 0; |
1807a1aa SS |
1439 | numa_platform = 0; |
1440 | } | |
62918a03 | 1441 | |
e498be7d CL |
1442 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1443 | kmem_list3_init(&initkmem_list3[i]); | |
1444 | if (i < MAX_NUMNODES) | |
1445 | cache_cache.nodelists[i] = NULL; | |
1446 | } | |
556a169d | 1447 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1448 | |
1449 | /* | |
1450 | * Fragmentation resistance on low memory - only use bigger | |
1451 | * page orders on machines with more than 32MB of memory. | |
1452 | */ | |
1453 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1454 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1455 | ||
1da177e4 LT |
1456 | /* Bootstrap is tricky, because several objects are allocated |
1457 | * from caches that do not exist yet: | |
a737b3e2 AM |
1458 | * 1) initialize the cache_cache cache: it contains the struct |
1459 | * kmem_cache structures of all caches, except cache_cache itself: | |
1460 | * cache_cache is statically allocated. | |
e498be7d CL |
1461 | * Initially an __init data area is used for the head array and the |
1462 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1463 | * array at the end of the bootstrap. | |
1da177e4 | 1464 | * 2) Create the first kmalloc cache. |
343e0d7a | 1465 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1466 | * An __init data area is used for the head array. |
1467 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1468 | * head arrays. | |
1da177e4 LT |
1469 | * 4) Replace the __init data head arrays for cache_cache and the first |
1470 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1471 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1472 | * the other cache's with kmalloc allocated memory. | |
1473 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1474 | */ |
1475 | ||
1ca4cb24 PE |
1476 | node = numa_node_id(); |
1477 | ||
1da177e4 | 1478 | /* 1) create the cache_cache */ |
1da177e4 LT |
1479 | INIT_LIST_HEAD(&cache_chain); |
1480 | list_add(&cache_cache.next, &cache_chain); | |
1481 | cache_cache.colour_off = cache_line_size(); | |
1482 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1483 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1484 | |
8da3430d ED |
1485 | /* |
1486 | * struct kmem_cache size depends on nr_node_ids, which | |
1487 | * can be less than MAX_NUMNODES. | |
1488 | */ | |
1489 | cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + | |
1490 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1491 | #if DEBUG | |
1492 | cache_cache.obj_size = cache_cache.buffer_size; | |
1493 | #endif | |
a737b3e2 AM |
1494 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1495 | cache_line_size()); | |
6a2d7a95 ED |
1496 | cache_cache.reciprocal_buffer_size = |
1497 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1498 | |
07ed76b2 JS |
1499 | for (order = 0; order < MAX_ORDER; order++) { |
1500 | cache_estimate(order, cache_cache.buffer_size, | |
1501 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1502 | if (cache_cache.num) | |
1503 | break; | |
1504 | } | |
40094fa6 | 1505 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1506 | cache_cache.gfporder = order; |
b28a02de | 1507 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1508 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1509 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1510 | |
1511 | /* 2+3) create the kmalloc caches */ | |
1512 | sizes = malloc_sizes; | |
1513 | names = cache_names; | |
1514 | ||
a737b3e2 AM |
1515 | /* |
1516 | * Initialize the caches that provide memory for the array cache and the | |
1517 | * kmem_list3 structures first. Without this, further allocations will | |
1518 | * bug. | |
e498be7d CL |
1519 | */ |
1520 | ||
1521 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1522 | sizes[INDEX_AC].cs_size, |
1523 | ARCH_KMALLOC_MINALIGN, | |
1524 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1525 | NULL); |
e498be7d | 1526 | |
a737b3e2 | 1527 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1528 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1529 | kmem_cache_create(names[INDEX_L3].name, |
1530 | sizes[INDEX_L3].cs_size, | |
1531 | ARCH_KMALLOC_MINALIGN, | |
1532 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1533 | NULL); |
a737b3e2 | 1534 | } |
e498be7d | 1535 | |
e0a42726 IM |
1536 | slab_early_init = 0; |
1537 | ||
1da177e4 | 1538 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1539 | /* |
1540 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1541 | * This should be particularly beneficial on SMP boxes, as it |
1542 | * eliminates "false sharing". | |
1543 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1544 | * allow tighter packing of the smaller caches. |
1545 | */ | |
a737b3e2 | 1546 | if (!sizes->cs_cachep) { |
e498be7d | 1547 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1548 | sizes->cs_size, |
1549 | ARCH_KMALLOC_MINALIGN, | |
1550 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1551 | NULL); |
a737b3e2 | 1552 | } |
4b51d669 CL |
1553 | #ifdef CONFIG_ZONE_DMA |
1554 | sizes->cs_dmacachep = kmem_cache_create( | |
1555 | names->name_dma, | |
a737b3e2 AM |
1556 | sizes->cs_size, |
1557 | ARCH_KMALLOC_MINALIGN, | |
1558 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1559 | SLAB_PANIC, | |
20c2df83 | 1560 | NULL); |
4b51d669 | 1561 | #endif |
1da177e4 LT |
1562 | sizes++; |
1563 | names++; | |
1564 | } | |
1565 | /* 4) Replace the bootstrap head arrays */ | |
1566 | { | |
2b2d5493 | 1567 | struct array_cache *ptr; |
e498be7d | 1568 | |
1da177e4 | 1569 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1570 | |
1da177e4 | 1571 | local_irq_disable(); |
9a2dba4b PE |
1572 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1573 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1574 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1575 | /* |
1576 | * Do not assume that spinlocks can be initialized via memcpy: | |
1577 | */ | |
1578 | spin_lock_init(&ptr->lock); | |
1579 | ||
1da177e4 LT |
1580 | cache_cache.array[smp_processor_id()] = ptr; |
1581 | local_irq_enable(); | |
e498be7d | 1582 | |
1da177e4 | 1583 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1584 | |
1da177e4 | 1585 | local_irq_disable(); |
9a2dba4b | 1586 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1587 | != &initarray_generic.cache); |
9a2dba4b | 1588 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
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 | ||
e498be7d | 1595 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1596 | ptr; |
1da177e4 LT |
1597 | local_irq_enable(); |
1598 | } | |
e498be7d CL |
1599 | /* 5) Replace the bootstrap kmem_list3's */ |
1600 | { | |
1ca4cb24 PE |
1601 | int nid; |
1602 | ||
9c09a95c | 1603 | for_each_online_node(nid) { |
ec1f5eee | 1604 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1605 | |
e498be7d | 1606 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1607 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1608 | |
1609 | if (INDEX_AC != INDEX_L3) { | |
1610 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1611 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1612 | } |
1613 | } | |
1614 | } | |
1da177e4 | 1615 | |
e498be7d | 1616 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1617 | { |
343e0d7a | 1618 | struct kmem_cache *cachep; |
fc0abb14 | 1619 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1620 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1621 | if (enable_cpucache(cachep)) |
1622 | BUG(); | |
fc0abb14 | 1623 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1624 | } |
1625 | ||
056c6241 RT |
1626 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1627 | init_lock_keys(); | |
1628 | ||
1629 | ||
1da177e4 LT |
1630 | /* Done! */ |
1631 | g_cpucache_up = FULL; | |
1632 | ||
a737b3e2 AM |
1633 | /* |
1634 | * Register a cpu startup notifier callback that initializes | |
1635 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1636 | */ |
1637 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1638 | |
a737b3e2 AM |
1639 | /* |
1640 | * The reap timers are started later, with a module init call: That part | |
1641 | * of the kernel is not yet operational. | |
1da177e4 LT |
1642 | */ |
1643 | } | |
1644 | ||
1645 | static int __init cpucache_init(void) | |
1646 | { | |
1647 | int cpu; | |
1648 | ||
a737b3e2 AM |
1649 | /* |
1650 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1651 | */ |
e498be7d | 1652 | for_each_online_cpu(cpu) |
a737b3e2 | 1653 | start_cpu_timer(cpu); |
1da177e4 LT |
1654 | return 0; |
1655 | } | |
1da177e4 LT |
1656 | __initcall(cpucache_init); |
1657 | ||
1658 | /* | |
1659 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1660 | * | |
1661 | * If we requested dmaable memory, we will get it. Even if we | |
1662 | * did not request dmaable memory, we might get it, but that | |
1663 | * would be relatively rare and ignorable. | |
1664 | */ | |
343e0d7a | 1665 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1666 | { |
1667 | struct page *page; | |
e1b6aa6f | 1668 | int nr_pages; |
1da177e4 LT |
1669 | int i; |
1670 | ||
d6fef9da | 1671 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1672 | /* |
1673 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1674 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1675 | */ |
e1b6aa6f | 1676 | flags |= __GFP_COMP; |
d6fef9da | 1677 | #endif |
765c4507 | 1678 | |
3c517a61 | 1679 | flags |= cachep->gfpflags; |
e12ba74d MG |
1680 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1681 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f CH |
1682 | |
1683 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1684 | if (!page) |
1685 | return NULL; | |
1da177e4 | 1686 | |
e1b6aa6f | 1687 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1688 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1689 | add_zone_page_state(page_zone(page), |
1690 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1691 | else | |
1692 | add_zone_page_state(page_zone(page), | |
1693 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1694 | for (i = 0; i < nr_pages; i++) |
1695 | __SetPageSlab(page + i); | |
1696 | return page_address(page); | |
1da177e4 LT |
1697 | } |
1698 | ||
1699 | /* | |
1700 | * Interface to system's page release. | |
1701 | */ | |
343e0d7a | 1702 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1703 | { |
b28a02de | 1704 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1705 | struct page *page = virt_to_page(addr); |
1706 | const unsigned long nr_freed = i; | |
1707 | ||
972d1a7b CL |
1708 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1709 | sub_zone_page_state(page_zone(page), | |
1710 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1711 | else | |
1712 | sub_zone_page_state(page_zone(page), | |
1713 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1714 | while (i--) { |
f205b2fe NP |
1715 | BUG_ON(!PageSlab(page)); |
1716 | __ClearPageSlab(page); | |
1da177e4 LT |
1717 | page++; |
1718 | } | |
1da177e4 LT |
1719 | if (current->reclaim_state) |
1720 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1721 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1722 | } |
1723 | ||
1724 | static void kmem_rcu_free(struct rcu_head *head) | |
1725 | { | |
b28a02de | 1726 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1727 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1728 | |
1729 | kmem_freepages(cachep, slab_rcu->addr); | |
1730 | if (OFF_SLAB(cachep)) | |
1731 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1732 | } | |
1733 | ||
1734 | #if DEBUG | |
1735 | ||
1736 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1737 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1738 | unsigned long caller) |
1da177e4 | 1739 | { |
3dafccf2 | 1740 | int size = obj_size(cachep); |
1da177e4 | 1741 | |
3dafccf2 | 1742 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1743 | |
b28a02de | 1744 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1745 | return; |
1746 | ||
b28a02de PE |
1747 | *addr++ = 0x12345678; |
1748 | *addr++ = caller; | |
1749 | *addr++ = smp_processor_id(); | |
1750 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1751 | { |
1752 | unsigned long *sptr = &caller; | |
1753 | unsigned long svalue; | |
1754 | ||
1755 | while (!kstack_end(sptr)) { | |
1756 | svalue = *sptr++; | |
1757 | if (kernel_text_address(svalue)) { | |
b28a02de | 1758 | *addr++ = svalue; |
1da177e4 LT |
1759 | size -= sizeof(unsigned long); |
1760 | if (size <= sizeof(unsigned long)) | |
1761 | break; | |
1762 | } | |
1763 | } | |
1764 | ||
1765 | } | |
b28a02de | 1766 | *addr++ = 0x87654321; |
1da177e4 LT |
1767 | } |
1768 | #endif | |
1769 | ||
343e0d7a | 1770 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1771 | { |
3dafccf2 MS |
1772 | int size = obj_size(cachep); |
1773 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1774 | |
1775 | memset(addr, val, size); | |
b28a02de | 1776 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1777 | } |
1778 | ||
1779 | static void dump_line(char *data, int offset, int limit) | |
1780 | { | |
1781 | int i; | |
aa83aa40 DJ |
1782 | unsigned char error = 0; |
1783 | int bad_count = 0; | |
1784 | ||
1da177e4 | 1785 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1786 | for (i = 0; i < limit; i++) { |
1787 | if (data[offset + i] != POISON_FREE) { | |
1788 | error = data[offset + i]; | |
1789 | bad_count++; | |
1790 | } | |
b28a02de | 1791 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1792 | } |
1da177e4 | 1793 | printk("\n"); |
aa83aa40 DJ |
1794 | |
1795 | if (bad_count == 1) { | |
1796 | error ^= POISON_FREE; | |
1797 | if (!(error & (error - 1))) { | |
1798 | printk(KERN_ERR "Single bit error detected. Probably " | |
1799 | "bad RAM.\n"); | |
1800 | #ifdef CONFIG_X86 | |
1801 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1802 | "test tool.\n"); | |
1803 | #else | |
1804 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1805 | #endif | |
1806 | } | |
1807 | } | |
1da177e4 LT |
1808 | } |
1809 | #endif | |
1810 | ||
1811 | #if DEBUG | |
1812 | ||
343e0d7a | 1813 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1814 | { |
1815 | int i, size; | |
1816 | char *realobj; | |
1817 | ||
1818 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1819 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1820 | *dbg_redzone1(cachep, objp), |
1821 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1822 | } |
1823 | ||
1824 | if (cachep->flags & SLAB_STORE_USER) { | |
1825 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1826 | *dbg_userword(cachep, objp)); |
1da177e4 | 1827 | print_symbol("(%s)", |
a737b3e2 | 1828 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1829 | printk("\n"); |
1830 | } | |
3dafccf2 MS |
1831 | realobj = (char *)objp + obj_offset(cachep); |
1832 | size = obj_size(cachep); | |
b28a02de | 1833 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1834 | int limit; |
1835 | limit = 16; | |
b28a02de PE |
1836 | if (i + limit > size) |
1837 | limit = size - i; | |
1da177e4 LT |
1838 | dump_line(realobj, i, limit); |
1839 | } | |
1840 | } | |
1841 | ||
343e0d7a | 1842 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1843 | { |
1844 | char *realobj; | |
1845 | int size, i; | |
1846 | int lines = 0; | |
1847 | ||
3dafccf2 MS |
1848 | realobj = (char *)objp + obj_offset(cachep); |
1849 | size = obj_size(cachep); | |
1da177e4 | 1850 | |
b28a02de | 1851 | for (i = 0; i < size; i++) { |
1da177e4 | 1852 | char exp = POISON_FREE; |
b28a02de | 1853 | if (i == size - 1) |
1da177e4 LT |
1854 | exp = POISON_END; |
1855 | if (realobj[i] != exp) { | |
1856 | int limit; | |
1857 | /* Mismatch ! */ | |
1858 | /* Print header */ | |
1859 | if (lines == 0) { | |
b28a02de | 1860 | printk(KERN_ERR |
e94a40c5 DH |
1861 | "Slab corruption: %s start=%p, len=%d\n", |
1862 | cachep->name, realobj, size); | |
1da177e4 LT |
1863 | print_objinfo(cachep, objp, 0); |
1864 | } | |
1865 | /* Hexdump the affected line */ | |
b28a02de | 1866 | i = (i / 16) * 16; |
1da177e4 | 1867 | limit = 16; |
b28a02de PE |
1868 | if (i + limit > size) |
1869 | limit = size - i; | |
1da177e4 LT |
1870 | dump_line(realobj, i, limit); |
1871 | i += 16; | |
1872 | lines++; | |
1873 | /* Limit to 5 lines */ | |
1874 | if (lines > 5) | |
1875 | break; | |
1876 | } | |
1877 | } | |
1878 | if (lines != 0) { | |
1879 | /* Print some data about the neighboring objects, if they | |
1880 | * exist: | |
1881 | */ | |
6ed5eb22 | 1882 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1883 | unsigned int objnr; |
1da177e4 | 1884 | |
8fea4e96 | 1885 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1886 | if (objnr) { |
8fea4e96 | 1887 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1888 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1889 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1890 | realobj, size); |
1da177e4 LT |
1891 | print_objinfo(cachep, objp, 2); |
1892 | } | |
b28a02de | 1893 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1894 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1895 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1896 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1897 | realobj, size); |
1da177e4 LT |
1898 | print_objinfo(cachep, objp, 2); |
1899 | } | |
1900 | } | |
1901 | } | |
1902 | #endif | |
1903 | ||
12dd36fa | 1904 | #if DEBUG |
e79aec29 | 1905 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1906 | { |
1da177e4 LT |
1907 | int i; |
1908 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1909 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1910 | |
1911 | if (cachep->flags & SLAB_POISON) { | |
1912 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1913 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1914 | OFF_SLAB(cachep)) | |
b28a02de | 1915 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1916 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1917 | else |
1918 | check_poison_obj(cachep, objp); | |
1919 | #else | |
1920 | check_poison_obj(cachep, objp); | |
1921 | #endif | |
1922 | } | |
1923 | if (cachep->flags & SLAB_RED_ZONE) { | |
1924 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1925 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1926 | "was overwritten"); |
1da177e4 LT |
1927 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1928 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1929 | "was overwritten"); |
1da177e4 | 1930 | } |
1da177e4 | 1931 | } |
12dd36fa | 1932 | } |
1da177e4 | 1933 | #else |
e79aec29 | 1934 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1935 | { |
12dd36fa | 1936 | } |
1da177e4 LT |
1937 | #endif |
1938 | ||
911851e6 RD |
1939 | /** |
1940 | * slab_destroy - destroy and release all objects in a slab | |
1941 | * @cachep: cache pointer being destroyed | |
1942 | * @slabp: slab pointer being destroyed | |
1943 | * | |
12dd36fa | 1944 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1945 | * Before calling the slab must have been unlinked from the cache. The |
1946 | * cache-lock is not held/needed. | |
12dd36fa | 1947 | */ |
343e0d7a | 1948 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1949 | { |
1950 | void *addr = slabp->s_mem - slabp->colouroff; | |
1951 | ||
e79aec29 | 1952 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
1953 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1954 | struct slab_rcu *slab_rcu; | |
1955 | ||
b28a02de | 1956 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1957 | slab_rcu->cachep = cachep; |
1958 | slab_rcu->addr = addr; | |
1959 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1960 | } else { | |
1961 | kmem_freepages(cachep, addr); | |
873623df IM |
1962 | if (OFF_SLAB(cachep)) |
1963 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1964 | } |
1965 | } | |
1966 | ||
117f6eb1 CL |
1967 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1968 | { | |
1969 | int i; | |
1970 | struct kmem_list3 *l3; | |
1971 | ||
1972 | for_each_online_cpu(i) | |
1973 | kfree(cachep->array[i]); | |
1974 | ||
1975 | /* NUMA: free the list3 structures */ | |
1976 | for_each_online_node(i) { | |
1977 | l3 = cachep->nodelists[i]; | |
1978 | if (l3) { | |
1979 | kfree(l3->shared); | |
1980 | free_alien_cache(l3->alien); | |
1981 | kfree(l3); | |
1982 | } | |
1983 | } | |
1984 | kmem_cache_free(&cache_cache, cachep); | |
1985 | } | |
1986 | ||
1987 | ||
4d268eba | 1988 | /** |
a70773dd RD |
1989 | * calculate_slab_order - calculate size (page order) of slabs |
1990 | * @cachep: pointer to the cache that is being created | |
1991 | * @size: size of objects to be created in this cache. | |
1992 | * @align: required alignment for the objects. | |
1993 | * @flags: slab allocation flags | |
1994 | * | |
1995 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1996 | * |
1997 | * This could be made much more intelligent. For now, try to avoid using | |
1998 | * high order pages for slabs. When the gfp() functions are more friendly | |
1999 | * towards high-order requests, this should be changed. | |
2000 | */ | |
a737b3e2 | 2001 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2002 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2003 | { |
b1ab41c4 | 2004 | unsigned long offslab_limit; |
4d268eba | 2005 | size_t left_over = 0; |
9888e6fa | 2006 | int gfporder; |
4d268eba | 2007 | |
0aa817f0 | 2008 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2009 | unsigned int num; |
2010 | size_t remainder; | |
2011 | ||
9888e6fa | 2012 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2013 | if (!num) |
2014 | continue; | |
9888e6fa | 2015 | |
b1ab41c4 IM |
2016 | if (flags & CFLGS_OFF_SLAB) { |
2017 | /* | |
2018 | * Max number of objs-per-slab for caches which | |
2019 | * use off-slab slabs. Needed to avoid a possible | |
2020 | * looping condition in cache_grow(). | |
2021 | */ | |
2022 | offslab_limit = size - sizeof(struct slab); | |
2023 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2024 | ||
2025 | if (num > offslab_limit) | |
2026 | break; | |
2027 | } | |
4d268eba | 2028 | |
9888e6fa | 2029 | /* Found something acceptable - save it away */ |
4d268eba | 2030 | cachep->num = num; |
9888e6fa | 2031 | cachep->gfporder = gfporder; |
4d268eba PE |
2032 | left_over = remainder; |
2033 | ||
f78bb8ad LT |
2034 | /* |
2035 | * A VFS-reclaimable slab tends to have most allocations | |
2036 | * as GFP_NOFS and we really don't want to have to be allocating | |
2037 | * higher-order pages when we are unable to shrink dcache. | |
2038 | */ | |
2039 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2040 | break; | |
2041 | ||
4d268eba PE |
2042 | /* |
2043 | * Large number of objects is good, but very large slabs are | |
2044 | * currently bad for the gfp()s. | |
2045 | */ | |
9888e6fa | 2046 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2047 | break; |
2048 | ||
9888e6fa LT |
2049 | /* |
2050 | * Acceptable internal fragmentation? | |
2051 | */ | |
a737b3e2 | 2052 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2053 | break; |
2054 | } | |
2055 | return left_over; | |
2056 | } | |
2057 | ||
38bdc32a | 2058 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 2059 | { |
2ed3a4ef CL |
2060 | if (g_cpucache_up == FULL) |
2061 | return enable_cpucache(cachep); | |
2062 | ||
f30cf7d1 PE |
2063 | if (g_cpucache_up == NONE) { |
2064 | /* | |
2065 | * Note: the first kmem_cache_create must create the cache | |
2066 | * that's used by kmalloc(24), otherwise the creation of | |
2067 | * further caches will BUG(). | |
2068 | */ | |
2069 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2070 | ||
2071 | /* | |
2072 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2073 | * the first cache, then we need to set up all its list3s, | |
2074 | * otherwise the creation of further caches will BUG(). | |
2075 | */ | |
2076 | set_up_list3s(cachep, SIZE_AC); | |
2077 | if (INDEX_AC == INDEX_L3) | |
2078 | g_cpucache_up = PARTIAL_L3; | |
2079 | else | |
2080 | g_cpucache_up = PARTIAL_AC; | |
2081 | } else { | |
2082 | cachep->array[smp_processor_id()] = | |
2083 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
2084 | ||
2085 | if (g_cpucache_up == PARTIAL_AC) { | |
2086 | set_up_list3s(cachep, SIZE_L3); | |
2087 | g_cpucache_up = PARTIAL_L3; | |
2088 | } else { | |
2089 | int node; | |
556a169d | 2090 | for_each_online_node(node) { |
f30cf7d1 PE |
2091 | cachep->nodelists[node] = |
2092 | kmalloc_node(sizeof(struct kmem_list3), | |
2093 | GFP_KERNEL, node); | |
2094 | BUG_ON(!cachep->nodelists[node]); | |
2095 | kmem_list3_init(cachep->nodelists[node]); | |
2096 | } | |
2097 | } | |
2098 | } | |
2099 | cachep->nodelists[numa_node_id()]->next_reap = | |
2100 | jiffies + REAPTIMEOUT_LIST3 + | |
2101 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2102 | ||
2103 | cpu_cache_get(cachep)->avail = 0; | |
2104 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2105 | cpu_cache_get(cachep)->batchcount = 1; | |
2106 | cpu_cache_get(cachep)->touched = 0; | |
2107 | cachep->batchcount = 1; | |
2108 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2109 | return 0; |
f30cf7d1 PE |
2110 | } |
2111 | ||
1da177e4 LT |
2112 | /** |
2113 | * kmem_cache_create - Create a cache. | |
2114 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2115 | * @size: The size of objects to be created in this cache. | |
2116 | * @align: The required alignment for the objects. | |
2117 | * @flags: SLAB flags | |
2118 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2119 | * |
2120 | * Returns a ptr to the cache on success, NULL on failure. | |
2121 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2122 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2123 | * |
2124 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2125 | * the module calling this has to destroy the cache before getting unloaded. |
2126 | * | |
1da177e4 LT |
2127 | * The flags are |
2128 | * | |
2129 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2130 | * to catch references to uninitialised memory. | |
2131 | * | |
2132 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2133 | * for buffer overruns. | |
2134 | * | |
1da177e4 LT |
2135 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2136 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2137 | * as davem. | |
2138 | */ | |
343e0d7a | 2139 | struct kmem_cache * |
1da177e4 | 2140 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2141 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2142 | { |
2143 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2144 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2145 | |
2146 | /* | |
2147 | * Sanity checks... these are all serious usage bugs. | |
2148 | */ | |
a737b3e2 | 2149 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2150 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2151 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2152 | name); |
b28a02de PE |
2153 | BUG(); |
2154 | } | |
1da177e4 | 2155 | |
f0188f47 | 2156 | /* |
8f5be20b RT |
2157 | * We use cache_chain_mutex to ensure a consistent view of |
2158 | * cpu_online_map as well. Please see cpuup_callback | |
f0188f47 | 2159 | */ |
95402b38 | 2160 | get_online_cpus(); |
fc0abb14 | 2161 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2162 | |
7a7c381d | 2163 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2164 | char tmp; |
2165 | int res; | |
2166 | ||
2167 | /* | |
2168 | * This happens when the module gets unloaded and doesn't | |
2169 | * destroy its slab cache and no-one else reuses the vmalloc | |
2170 | * area of the module. Print a warning. | |
2171 | */ | |
138ae663 | 2172 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2173 | if (res) { |
b4169525 | 2174 | printk(KERN_ERR |
2175 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2176 | pc->buffer_size); |
4f12bb4f AM |
2177 | continue; |
2178 | } | |
2179 | ||
b28a02de | 2180 | if (!strcmp(pc->name, name)) { |
b4169525 | 2181 | printk(KERN_ERR |
2182 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2183 | dump_stack(); |
2184 | goto oops; | |
2185 | } | |
2186 | } | |
2187 | ||
1da177e4 LT |
2188 | #if DEBUG |
2189 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2190 | #if FORCED_DEBUG |
2191 | /* | |
2192 | * Enable redzoning and last user accounting, except for caches with | |
2193 | * large objects, if the increased size would increase the object size | |
2194 | * above the next power of two: caches with object sizes just above a | |
2195 | * power of two have a significant amount of internal fragmentation. | |
2196 | */ | |
87a927c7 DW |
2197 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2198 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2199 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2200 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2201 | flags |= SLAB_POISON; | |
2202 | #endif | |
2203 | if (flags & SLAB_DESTROY_BY_RCU) | |
2204 | BUG_ON(flags & SLAB_POISON); | |
2205 | #endif | |
1da177e4 | 2206 | /* |
a737b3e2 AM |
2207 | * Always checks flags, a caller might be expecting debug support which |
2208 | * isn't available. | |
1da177e4 | 2209 | */ |
40094fa6 | 2210 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2211 | |
a737b3e2 AM |
2212 | /* |
2213 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2214 | * unaligned accesses for some archs when redzoning is used, and makes |
2215 | * sure any on-slab bufctl's are also correctly aligned. | |
2216 | */ | |
b28a02de PE |
2217 | if (size & (BYTES_PER_WORD - 1)) { |
2218 | size += (BYTES_PER_WORD - 1); | |
2219 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2220 | } |
2221 | ||
a737b3e2 AM |
2222 | /* calculate the final buffer alignment: */ |
2223 | ||
1da177e4 LT |
2224 | /* 1) arch recommendation: can be overridden for debug */ |
2225 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2226 | /* |
2227 | * Default alignment: as specified by the arch code. Except if | |
2228 | * an object is really small, then squeeze multiple objects into | |
2229 | * one cacheline. | |
1da177e4 LT |
2230 | */ |
2231 | ralign = cache_line_size(); | |
b28a02de | 2232 | while (size <= ralign / 2) |
1da177e4 LT |
2233 | ralign /= 2; |
2234 | } else { | |
2235 | ralign = BYTES_PER_WORD; | |
2236 | } | |
ca5f9703 PE |
2237 | |
2238 | /* | |
87a927c7 DW |
2239 | * Redzoning and user store require word alignment or possibly larger. |
2240 | * Note this will be overridden by architecture or caller mandated | |
2241 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2242 | */ |
87a927c7 DW |
2243 | if (flags & SLAB_STORE_USER) |
2244 | ralign = BYTES_PER_WORD; | |
2245 | ||
2246 | if (flags & SLAB_RED_ZONE) { | |
2247 | ralign = REDZONE_ALIGN; | |
2248 | /* If redzoning, ensure that the second redzone is suitably | |
2249 | * aligned, by adjusting the object size accordingly. */ | |
2250 | size += REDZONE_ALIGN - 1; | |
2251 | size &= ~(REDZONE_ALIGN - 1); | |
2252 | } | |
ca5f9703 | 2253 | |
a44b56d3 | 2254 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2255 | if (ralign < ARCH_SLAB_MINALIGN) { |
2256 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2257 | } |
a44b56d3 | 2258 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2259 | if (ralign < align) { |
2260 | ralign = align; | |
1da177e4 | 2261 | } |
a44b56d3 | 2262 | /* disable debug if necessary */ |
b46b8f19 | 2263 | if (ralign > __alignof__(unsigned long long)) |
a44b56d3 | 2264 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2265 | /* |
ca5f9703 | 2266 | * 4) Store it. |
1da177e4 LT |
2267 | */ |
2268 | align = ralign; | |
2269 | ||
2270 | /* Get cache's description obj. */ | |
e94b1766 | 2271 | cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); |
1da177e4 | 2272 | if (!cachep) |
4f12bb4f | 2273 | goto oops; |
1da177e4 LT |
2274 | |
2275 | #if DEBUG | |
3dafccf2 | 2276 | cachep->obj_size = size; |
1da177e4 | 2277 | |
ca5f9703 PE |
2278 | /* |
2279 | * Both debugging options require word-alignment which is calculated | |
2280 | * into align above. | |
2281 | */ | |
1da177e4 | 2282 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2283 | /* add space for red zone words */ |
b46b8f19 DW |
2284 | cachep->obj_offset += sizeof(unsigned long long); |
2285 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2286 | } |
2287 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2288 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2289 | * the real object. But if the second red zone needs to be |
2290 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2291 | */ |
87a927c7 DW |
2292 | if (flags & SLAB_RED_ZONE) |
2293 | size += REDZONE_ALIGN; | |
2294 | else | |
2295 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2296 | } |
2297 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2298 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2299 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2300 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2301 | size = PAGE_SIZE; |
2302 | } | |
2303 | #endif | |
2304 | #endif | |
2305 | ||
e0a42726 IM |
2306 | /* |
2307 | * Determine if the slab management is 'on' or 'off' slab. | |
2308 | * (bootstrapping cannot cope with offslab caches so don't do | |
2309 | * it too early on.) | |
2310 | */ | |
2311 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2312 | /* |
2313 | * Size is large, assume best to place the slab management obj | |
2314 | * off-slab (should allow better packing of objs). | |
2315 | */ | |
2316 | flags |= CFLGS_OFF_SLAB; | |
2317 | ||
2318 | size = ALIGN(size, align); | |
2319 | ||
f78bb8ad | 2320 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2321 | |
2322 | if (!cachep->num) { | |
b4169525 | 2323 | printk(KERN_ERR |
2324 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2325 | kmem_cache_free(&cache_cache, cachep); |
2326 | cachep = NULL; | |
4f12bb4f | 2327 | goto oops; |
1da177e4 | 2328 | } |
b28a02de PE |
2329 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2330 | + sizeof(struct slab), align); | |
1da177e4 LT |
2331 | |
2332 | /* | |
2333 | * If the slab has been placed off-slab, and we have enough space then | |
2334 | * move it on-slab. This is at the expense of any extra colouring. | |
2335 | */ | |
2336 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2337 | flags &= ~CFLGS_OFF_SLAB; | |
2338 | left_over -= slab_size; | |
2339 | } | |
2340 | ||
2341 | if (flags & CFLGS_OFF_SLAB) { | |
2342 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2343 | slab_size = |
2344 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2345 | } |
2346 | ||
2347 | cachep->colour_off = cache_line_size(); | |
2348 | /* Offset must be a multiple of the alignment. */ | |
2349 | if (cachep->colour_off < align) | |
2350 | cachep->colour_off = align; | |
b28a02de | 2351 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2352 | cachep->slab_size = slab_size; |
2353 | cachep->flags = flags; | |
2354 | cachep->gfpflags = 0; | |
4b51d669 | 2355 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2356 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2357 | cachep->buffer_size = size; |
6a2d7a95 | 2358 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2359 | |
e5ac9c5a | 2360 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2361 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2362 | /* |
2363 | * This is a possibility for one of the malloc_sizes caches. | |
2364 | * But since we go off slab only for object size greater than | |
2365 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2366 | * this should not happen at all. | |
2367 | * But leave a BUG_ON for some lucky dude. | |
2368 | */ | |
6cb8f913 | 2369 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2370 | } |
1da177e4 | 2371 | cachep->ctor = ctor; |
1da177e4 LT |
2372 | cachep->name = name; |
2373 | ||
2ed3a4ef CL |
2374 | if (setup_cpu_cache(cachep)) { |
2375 | __kmem_cache_destroy(cachep); | |
2376 | cachep = NULL; | |
2377 | goto oops; | |
2378 | } | |
1da177e4 | 2379 | |
1da177e4 LT |
2380 | /* cache setup completed, link it into the list */ |
2381 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2382 | oops: |
1da177e4 LT |
2383 | if (!cachep && (flags & SLAB_PANIC)) |
2384 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2385 | name); |
fc0abb14 | 2386 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2387 | put_online_cpus(); |
1da177e4 LT |
2388 | return cachep; |
2389 | } | |
2390 | EXPORT_SYMBOL(kmem_cache_create); | |
2391 | ||
2392 | #if DEBUG | |
2393 | static void check_irq_off(void) | |
2394 | { | |
2395 | BUG_ON(!irqs_disabled()); | |
2396 | } | |
2397 | ||
2398 | static void check_irq_on(void) | |
2399 | { | |
2400 | BUG_ON(irqs_disabled()); | |
2401 | } | |
2402 | ||
343e0d7a | 2403 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2404 | { |
2405 | #ifdef CONFIG_SMP | |
2406 | check_irq_off(); | |
e498be7d | 2407 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2408 | #endif |
2409 | } | |
e498be7d | 2410 | |
343e0d7a | 2411 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2412 | { |
2413 | #ifdef CONFIG_SMP | |
2414 | check_irq_off(); | |
2415 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2416 | #endif | |
2417 | } | |
2418 | ||
1da177e4 LT |
2419 | #else |
2420 | #define check_irq_off() do { } while(0) | |
2421 | #define check_irq_on() do { } while(0) | |
2422 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2423 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2424 | #endif |
2425 | ||
aab2207c CL |
2426 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2427 | struct array_cache *ac, | |
2428 | int force, int node); | |
2429 | ||
1da177e4 LT |
2430 | static void do_drain(void *arg) |
2431 | { | |
a737b3e2 | 2432 | struct kmem_cache *cachep = arg; |
1da177e4 | 2433 | struct array_cache *ac; |
ff69416e | 2434 | int node = numa_node_id(); |
1da177e4 LT |
2435 | |
2436 | check_irq_off(); | |
9a2dba4b | 2437 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2438 | spin_lock(&cachep->nodelists[node]->list_lock); |
2439 | free_block(cachep, ac->entry, ac->avail, node); | |
2440 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2441 | ac->avail = 0; |
2442 | } | |
2443 | ||
343e0d7a | 2444 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2445 | { |
e498be7d CL |
2446 | struct kmem_list3 *l3; |
2447 | int node; | |
2448 | ||
15c8b6c1 | 2449 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2450 | check_irq_on(); |
b28a02de | 2451 | for_each_online_node(node) { |
e498be7d | 2452 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2453 | if (l3 && l3->alien) |
2454 | drain_alien_cache(cachep, l3->alien); | |
2455 | } | |
2456 | ||
2457 | for_each_online_node(node) { | |
2458 | l3 = cachep->nodelists[node]; | |
2459 | if (l3) | |
aab2207c | 2460 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2461 | } |
1da177e4 LT |
2462 | } |
2463 | ||
ed11d9eb CL |
2464 | /* |
2465 | * Remove slabs from the list of free slabs. | |
2466 | * Specify the number of slabs to drain in tofree. | |
2467 | * | |
2468 | * Returns the actual number of slabs released. | |
2469 | */ | |
2470 | static int drain_freelist(struct kmem_cache *cache, | |
2471 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2472 | { |
ed11d9eb CL |
2473 | struct list_head *p; |
2474 | int nr_freed; | |
1da177e4 | 2475 | struct slab *slabp; |
1da177e4 | 2476 | |
ed11d9eb CL |
2477 | nr_freed = 0; |
2478 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2479 | |
ed11d9eb | 2480 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2481 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2482 | if (p == &l3->slabs_free) { |
2483 | spin_unlock_irq(&l3->list_lock); | |
2484 | goto out; | |
2485 | } | |
1da177e4 | 2486 | |
ed11d9eb | 2487 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2488 | #if DEBUG |
40094fa6 | 2489 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2490 | #endif |
2491 | list_del(&slabp->list); | |
ed11d9eb CL |
2492 | /* |
2493 | * Safe to drop the lock. The slab is no longer linked | |
2494 | * to the cache. | |
2495 | */ | |
2496 | l3->free_objects -= cache->num; | |
e498be7d | 2497 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2498 | slab_destroy(cache, slabp); |
2499 | nr_freed++; | |
1da177e4 | 2500 | } |
ed11d9eb CL |
2501 | out: |
2502 | return nr_freed; | |
1da177e4 LT |
2503 | } |
2504 | ||
8f5be20b | 2505 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2506 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2507 | { |
2508 | int ret = 0, i = 0; | |
2509 | struct kmem_list3 *l3; | |
2510 | ||
2511 | drain_cpu_caches(cachep); | |
2512 | ||
2513 | check_irq_on(); | |
2514 | for_each_online_node(i) { | |
2515 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2516 | if (!l3) |
2517 | continue; | |
2518 | ||
2519 | drain_freelist(cachep, l3, l3->free_objects); | |
2520 | ||
2521 | ret += !list_empty(&l3->slabs_full) || | |
2522 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2523 | } |
2524 | return (ret ? 1 : 0); | |
2525 | } | |
2526 | ||
1da177e4 LT |
2527 | /** |
2528 | * kmem_cache_shrink - Shrink a cache. | |
2529 | * @cachep: The cache to shrink. | |
2530 | * | |
2531 | * Releases as many slabs as possible for a cache. | |
2532 | * To help debugging, a zero exit status indicates all slabs were released. | |
2533 | */ | |
343e0d7a | 2534 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2535 | { |
8f5be20b | 2536 | int ret; |
40094fa6 | 2537 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2538 | |
95402b38 | 2539 | get_online_cpus(); |
8f5be20b RT |
2540 | mutex_lock(&cache_chain_mutex); |
2541 | ret = __cache_shrink(cachep); | |
2542 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2543 | put_online_cpus(); |
8f5be20b | 2544 | return ret; |
1da177e4 LT |
2545 | } |
2546 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2547 | ||
2548 | /** | |
2549 | * kmem_cache_destroy - delete a cache | |
2550 | * @cachep: the cache to destroy | |
2551 | * | |
72fd4a35 | 2552 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2553 | * |
2554 | * It is expected this function will be called by a module when it is | |
2555 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2556 | * cache being allocated each time a module is loaded and unloaded, if the | |
2557 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2558 | * | |
2559 | * The cache must be empty before calling this function. | |
2560 | * | |
2561 | * The caller must guarantee that noone will allocate memory from the cache | |
2562 | * during the kmem_cache_destroy(). | |
2563 | */ | |
133d205a | 2564 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2565 | { |
40094fa6 | 2566 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2567 | |
1da177e4 | 2568 | /* Find the cache in the chain of caches. */ |
95402b38 | 2569 | get_online_cpus(); |
fc0abb14 | 2570 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2571 | /* |
2572 | * the chain is never empty, cache_cache is never destroyed | |
2573 | */ | |
2574 | list_del(&cachep->next); | |
1da177e4 LT |
2575 | if (__cache_shrink(cachep)) { |
2576 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2577 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2578 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2579 | put_online_cpus(); |
133d205a | 2580 | return; |
1da177e4 LT |
2581 | } |
2582 | ||
2583 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2584 | synchronize_rcu(); |
1da177e4 | 2585 | |
117f6eb1 | 2586 | __kmem_cache_destroy(cachep); |
8f5be20b | 2587 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2588 | put_online_cpus(); |
1da177e4 LT |
2589 | } |
2590 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2591 | ||
e5ac9c5a RT |
2592 | /* |
2593 | * Get the memory for a slab management obj. | |
2594 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2595 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2596 | * come from the same cache which is getting created because, | |
2597 | * when we are searching for an appropriate cache for these | |
2598 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2599 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2600 | * kmem_find_general_cachep till the initialization is complete. | |
2601 | * Hence we cannot have slabp_cache same as the original cache. | |
2602 | */ | |
343e0d7a | 2603 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2604 | int colour_off, gfp_t local_flags, |
2605 | int nodeid) | |
1da177e4 LT |
2606 | { |
2607 | struct slab *slabp; | |
b28a02de | 2608 | |
1da177e4 LT |
2609 | if (OFF_SLAB(cachep)) { |
2610 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2611 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
3c517a61 | 2612 | local_flags & ~GFP_THISNODE, nodeid); |
1da177e4 LT |
2613 | if (!slabp) |
2614 | return NULL; | |
2615 | } else { | |
b28a02de | 2616 | slabp = objp + colour_off; |
1da177e4 LT |
2617 | colour_off += cachep->slab_size; |
2618 | } | |
2619 | slabp->inuse = 0; | |
2620 | slabp->colouroff = colour_off; | |
b28a02de | 2621 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2622 | slabp->nodeid = nodeid; |
e51bfd0a | 2623 | slabp->free = 0; |
1da177e4 LT |
2624 | return slabp; |
2625 | } | |
2626 | ||
2627 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2628 | { | |
b28a02de | 2629 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2630 | } |
2631 | ||
343e0d7a | 2632 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2633 | struct slab *slabp) |
1da177e4 LT |
2634 | { |
2635 | int i; | |
2636 | ||
2637 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2638 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2639 | #if DEBUG |
2640 | /* need to poison the objs? */ | |
2641 | if (cachep->flags & SLAB_POISON) | |
2642 | poison_obj(cachep, objp, POISON_FREE); | |
2643 | if (cachep->flags & SLAB_STORE_USER) | |
2644 | *dbg_userword(cachep, objp) = NULL; | |
2645 | ||
2646 | if (cachep->flags & SLAB_RED_ZONE) { | |
2647 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2648 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2649 | } | |
2650 | /* | |
a737b3e2 AM |
2651 | * Constructors are not allowed to allocate memory from the same |
2652 | * cache which they are a constructor for. Otherwise, deadlock. | |
2653 | * They must also be threaded. | |
1da177e4 LT |
2654 | */ |
2655 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2656 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2657 | |
2658 | if (cachep->flags & SLAB_RED_ZONE) { | |
2659 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2660 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2661 | " end of an object"); |
1da177e4 LT |
2662 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2663 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2664 | " start of an object"); |
1da177e4 | 2665 | } |
a737b3e2 AM |
2666 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2667 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2668 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2669 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2670 | #else |
2671 | if (cachep->ctor) | |
51cc5068 | 2672 | cachep->ctor(objp); |
1da177e4 | 2673 | #endif |
b28a02de | 2674 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2675 | } |
b28a02de | 2676 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2677 | } |
2678 | ||
343e0d7a | 2679 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2680 | { |
4b51d669 CL |
2681 | if (CONFIG_ZONE_DMA_FLAG) { |
2682 | if (flags & GFP_DMA) | |
2683 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2684 | else | |
2685 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2686 | } | |
1da177e4 LT |
2687 | } |
2688 | ||
a737b3e2 AM |
2689 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2690 | int nodeid) | |
78d382d7 | 2691 | { |
8fea4e96 | 2692 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2693 | kmem_bufctl_t next; |
2694 | ||
2695 | slabp->inuse++; | |
2696 | next = slab_bufctl(slabp)[slabp->free]; | |
2697 | #if DEBUG | |
2698 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2699 | WARN_ON(slabp->nodeid != nodeid); | |
2700 | #endif | |
2701 | slabp->free = next; | |
2702 | ||
2703 | return objp; | |
2704 | } | |
2705 | ||
a737b3e2 AM |
2706 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2707 | void *objp, int nodeid) | |
78d382d7 | 2708 | { |
8fea4e96 | 2709 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2710 | |
2711 | #if DEBUG | |
2712 | /* Verify that the slab belongs to the intended node */ | |
2713 | WARN_ON(slabp->nodeid != nodeid); | |
2714 | ||
871751e2 | 2715 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2716 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2717 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2718 | BUG(); |
2719 | } | |
2720 | #endif | |
2721 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2722 | slabp->free = objnr; | |
2723 | slabp->inuse--; | |
2724 | } | |
2725 | ||
4776874f PE |
2726 | /* |
2727 | * Map pages beginning at addr to the given cache and slab. This is required | |
2728 | * for the slab allocator to be able to lookup the cache and slab of a | |
2729 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2730 | */ | |
2731 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2732 | void *addr) | |
1da177e4 | 2733 | { |
4776874f | 2734 | int nr_pages; |
1da177e4 LT |
2735 | struct page *page; |
2736 | ||
4776874f | 2737 | page = virt_to_page(addr); |
84097518 | 2738 | |
4776874f | 2739 | nr_pages = 1; |
84097518 | 2740 | if (likely(!PageCompound(page))) |
4776874f PE |
2741 | nr_pages <<= cache->gfporder; |
2742 | ||
1da177e4 | 2743 | do { |
4776874f PE |
2744 | page_set_cache(page, cache); |
2745 | page_set_slab(page, slab); | |
1da177e4 | 2746 | page++; |
4776874f | 2747 | } while (--nr_pages); |
1da177e4 LT |
2748 | } |
2749 | ||
2750 | /* | |
2751 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2752 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2753 | */ | |
3c517a61 CL |
2754 | static int cache_grow(struct kmem_cache *cachep, |
2755 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2756 | { |
b28a02de | 2757 | struct slab *slabp; |
b28a02de PE |
2758 | size_t offset; |
2759 | gfp_t local_flags; | |
e498be7d | 2760 | struct kmem_list3 *l3; |
1da177e4 | 2761 | |
a737b3e2 AM |
2762 | /* |
2763 | * Be lazy and only check for valid flags here, keeping it out of the | |
2764 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2765 | */ |
6cb06229 CL |
2766 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2767 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2768 | |
2e1217cf | 2769 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2770 | check_irq_off(); |
2e1217cf RT |
2771 | l3 = cachep->nodelists[nodeid]; |
2772 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2773 | |
2774 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2775 | offset = l3->colour_next; |
2776 | l3->colour_next++; | |
2777 | if (l3->colour_next >= cachep->colour) | |
2778 | l3->colour_next = 0; | |
2779 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2780 | |
2e1217cf | 2781 | offset *= cachep->colour_off; |
1da177e4 LT |
2782 | |
2783 | if (local_flags & __GFP_WAIT) | |
2784 | local_irq_enable(); | |
2785 | ||
2786 | /* | |
2787 | * The test for missing atomic flag is performed here, rather than | |
2788 | * the more obvious place, simply to reduce the critical path length | |
2789 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2790 | * will eventually be caught here (where it matters). | |
2791 | */ | |
2792 | kmem_flagcheck(cachep, flags); | |
2793 | ||
a737b3e2 AM |
2794 | /* |
2795 | * Get mem for the objs. Attempt to allocate a physical page from | |
2796 | * 'nodeid'. | |
e498be7d | 2797 | */ |
3c517a61 | 2798 | if (!objp) |
b8c1c5da | 2799 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2800 | if (!objp) |
1da177e4 LT |
2801 | goto failed; |
2802 | ||
2803 | /* Get slab management. */ | |
3c517a61 | 2804 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2805 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2806 | if (!slabp) |
1da177e4 LT |
2807 | goto opps1; |
2808 | ||
4776874f | 2809 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2810 | |
a35afb83 | 2811 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2812 | |
2813 | if (local_flags & __GFP_WAIT) | |
2814 | local_irq_disable(); | |
2815 | check_irq_off(); | |
e498be7d | 2816 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2817 | |
2818 | /* Make slab active. */ | |
e498be7d | 2819 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2820 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2821 | l3->free_objects += cachep->num; |
2822 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2823 | return 1; |
a737b3e2 | 2824 | opps1: |
1da177e4 | 2825 | kmem_freepages(cachep, objp); |
a737b3e2 | 2826 | failed: |
1da177e4 LT |
2827 | if (local_flags & __GFP_WAIT) |
2828 | local_irq_disable(); | |
2829 | return 0; | |
2830 | } | |
2831 | ||
2832 | #if DEBUG | |
2833 | ||
2834 | /* | |
2835 | * Perform extra freeing checks: | |
2836 | * - detect bad pointers. | |
2837 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2838 | */ |
2839 | static void kfree_debugcheck(const void *objp) | |
2840 | { | |
1da177e4 LT |
2841 | if (!virt_addr_valid(objp)) { |
2842 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2843 | (unsigned long)objp); |
2844 | BUG(); | |
1da177e4 | 2845 | } |
1da177e4 LT |
2846 | } |
2847 | ||
58ce1fd5 PE |
2848 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2849 | { | |
b46b8f19 | 2850 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2851 | |
2852 | redzone1 = *dbg_redzone1(cache, obj); | |
2853 | redzone2 = *dbg_redzone2(cache, obj); | |
2854 | ||
2855 | /* | |
2856 | * Redzone is ok. | |
2857 | */ | |
2858 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2859 | return; | |
2860 | ||
2861 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2862 | slab_error(cache, "double free detected"); | |
2863 | else | |
2864 | slab_error(cache, "memory outside object was overwritten"); | |
2865 | ||
b46b8f19 | 2866 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2867 | obj, redzone1, redzone2); |
2868 | } | |
2869 | ||
343e0d7a | 2870 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2871 | void *caller) |
1da177e4 LT |
2872 | { |
2873 | struct page *page; | |
2874 | unsigned int objnr; | |
2875 | struct slab *slabp; | |
2876 | ||
80cbd911 MW |
2877 | BUG_ON(virt_to_cache(objp) != cachep); |
2878 | ||
3dafccf2 | 2879 | objp -= obj_offset(cachep); |
1da177e4 | 2880 | kfree_debugcheck(objp); |
b49af68f | 2881 | page = virt_to_head_page(objp); |
1da177e4 | 2882 | |
065d41cb | 2883 | slabp = page_get_slab(page); |
1da177e4 LT |
2884 | |
2885 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2886 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2887 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2888 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2889 | } | |
2890 | if (cachep->flags & SLAB_STORE_USER) | |
2891 | *dbg_userword(cachep, objp) = caller; | |
2892 | ||
8fea4e96 | 2893 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2894 | |
2895 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2896 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2897 | |
871751e2 AV |
2898 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2899 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2900 | #endif | |
1da177e4 LT |
2901 | if (cachep->flags & SLAB_POISON) { |
2902 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2903 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2904 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2905 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2906 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2907 | } else { |
2908 | poison_obj(cachep, objp, POISON_FREE); | |
2909 | } | |
2910 | #else | |
2911 | poison_obj(cachep, objp, POISON_FREE); | |
2912 | #endif | |
2913 | } | |
2914 | return objp; | |
2915 | } | |
2916 | ||
343e0d7a | 2917 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2918 | { |
2919 | kmem_bufctl_t i; | |
2920 | int entries = 0; | |
b28a02de | 2921 | |
1da177e4 LT |
2922 | /* Check slab's freelist to see if this obj is there. */ |
2923 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2924 | entries++; | |
2925 | if (entries > cachep->num || i >= cachep->num) | |
2926 | goto bad; | |
2927 | } | |
2928 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2929 | bad: |
2930 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2931 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2932 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2933 | for (i = 0; |
264132bc | 2934 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2935 | i++) { |
a737b3e2 | 2936 | if (i % 16 == 0) |
1da177e4 | 2937 | printk("\n%03x:", i); |
b28a02de | 2938 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2939 | } |
2940 | printk("\n"); | |
2941 | BUG(); | |
2942 | } | |
2943 | } | |
2944 | #else | |
2945 | #define kfree_debugcheck(x) do { } while(0) | |
2946 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2947 | #define check_slabp(x,y) do { } while(0) | |
2948 | #endif | |
2949 | ||
343e0d7a | 2950 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2951 | { |
2952 | int batchcount; | |
2953 | struct kmem_list3 *l3; | |
2954 | struct array_cache *ac; | |
1ca4cb24 PE |
2955 | int node; |
2956 | ||
6d2144d3 | 2957 | retry: |
1da177e4 | 2958 | check_irq_off(); |
6d2144d3 | 2959 | node = numa_node_id(); |
9a2dba4b | 2960 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2961 | batchcount = ac->batchcount; |
2962 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2963 | /* |
2964 | * If there was little recent activity on this cache, then | |
2965 | * perform only a partial refill. Otherwise we could generate | |
2966 | * refill bouncing. | |
1da177e4 LT |
2967 | */ |
2968 | batchcount = BATCHREFILL_LIMIT; | |
2969 | } | |
1ca4cb24 | 2970 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2971 | |
2972 | BUG_ON(ac->avail > 0 || !l3); | |
2973 | spin_lock(&l3->list_lock); | |
1da177e4 | 2974 | |
3ded175a CL |
2975 | /* See if we can refill from the shared array */ |
2976 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2977 | goto alloc_done; | |
2978 | ||
1da177e4 LT |
2979 | while (batchcount > 0) { |
2980 | struct list_head *entry; | |
2981 | struct slab *slabp; | |
2982 | /* Get slab alloc is to come from. */ | |
2983 | entry = l3->slabs_partial.next; | |
2984 | if (entry == &l3->slabs_partial) { | |
2985 | l3->free_touched = 1; | |
2986 | entry = l3->slabs_free.next; | |
2987 | if (entry == &l3->slabs_free) | |
2988 | goto must_grow; | |
2989 | } | |
2990 | ||
2991 | slabp = list_entry(entry, struct slab, list); | |
2992 | check_slabp(cachep, slabp); | |
2993 | check_spinlock_acquired(cachep); | |
714b8171 PE |
2994 | |
2995 | /* | |
2996 | * The slab was either on partial or free list so | |
2997 | * there must be at least one object available for | |
2998 | * allocation. | |
2999 | */ | |
3000 | BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num); | |
3001 | ||
1da177e4 | 3002 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3003 | STATS_INC_ALLOCED(cachep); |
3004 | STATS_INC_ACTIVE(cachep); | |
3005 | STATS_SET_HIGH(cachep); | |
3006 | ||
78d382d7 | 3007 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3008 | node); |
1da177e4 LT |
3009 | } |
3010 | check_slabp(cachep, slabp); | |
3011 | ||
3012 | /* move slabp to correct slabp list: */ | |
3013 | list_del(&slabp->list); | |
3014 | if (slabp->free == BUFCTL_END) | |
3015 | list_add(&slabp->list, &l3->slabs_full); | |
3016 | else | |
3017 | list_add(&slabp->list, &l3->slabs_partial); | |
3018 | } | |
3019 | ||
a737b3e2 | 3020 | must_grow: |
1da177e4 | 3021 | l3->free_objects -= ac->avail; |
a737b3e2 | 3022 | alloc_done: |
e498be7d | 3023 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3024 | |
3025 | if (unlikely(!ac->avail)) { | |
3026 | int x; | |
3c517a61 | 3027 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3028 | |
a737b3e2 | 3029 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3030 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3031 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3032 | return NULL; |
3033 | ||
a737b3e2 | 3034 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3035 | goto retry; |
3036 | } | |
3037 | ac->touched = 1; | |
e498be7d | 3038 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3039 | } |
3040 | ||
a737b3e2 AM |
3041 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3042 | gfp_t flags) | |
1da177e4 LT |
3043 | { |
3044 | might_sleep_if(flags & __GFP_WAIT); | |
3045 | #if DEBUG | |
3046 | kmem_flagcheck(cachep, flags); | |
3047 | #endif | |
3048 | } | |
3049 | ||
3050 | #if DEBUG | |
a737b3e2 AM |
3051 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3052 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3053 | { |
b28a02de | 3054 | if (!objp) |
1da177e4 | 3055 | return objp; |
b28a02de | 3056 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3057 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3058 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3059 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3060 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3061 | else |
3062 | check_poison_obj(cachep, objp); | |
3063 | #else | |
3064 | check_poison_obj(cachep, objp); | |
3065 | #endif | |
3066 | poison_obj(cachep, objp, POISON_INUSE); | |
3067 | } | |
3068 | if (cachep->flags & SLAB_STORE_USER) | |
3069 | *dbg_userword(cachep, objp) = caller; | |
3070 | ||
3071 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3072 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3073 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3074 | slab_error(cachep, "double free, or memory outside" | |
3075 | " object was overwritten"); | |
b28a02de | 3076 | printk(KERN_ERR |
b46b8f19 | 3077 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3078 | objp, *dbg_redzone1(cachep, objp), |
3079 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3080 | } |
3081 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3082 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3083 | } | |
871751e2 AV |
3084 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3085 | { | |
3086 | struct slab *slabp; | |
3087 | unsigned objnr; | |
3088 | ||
b49af68f | 3089 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3090 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3091 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3092 | } | |
3093 | #endif | |
3dafccf2 | 3094 | objp += obj_offset(cachep); |
4f104934 | 3095 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3096 | cachep->ctor(objp); |
a44b56d3 KH |
3097 | #if ARCH_SLAB_MINALIGN |
3098 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3099 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3100 | objp, ARCH_SLAB_MINALIGN); | |
3101 | } | |
3102 | #endif | |
1da177e4 LT |
3103 | return objp; |
3104 | } | |
3105 | #else | |
3106 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3107 | #endif | |
3108 | ||
8a8b6502 AM |
3109 | #ifdef CONFIG_FAILSLAB |
3110 | ||
3111 | static struct failslab_attr { | |
3112 | ||
3113 | struct fault_attr attr; | |
3114 | ||
3115 | u32 ignore_gfp_wait; | |
3116 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3117 | struct dentry *ignore_gfp_wait_file; | |
3118 | #endif | |
3119 | ||
3120 | } failslab = { | |
3121 | .attr = FAULT_ATTR_INITIALIZER, | |
6b1b60f4 | 3122 | .ignore_gfp_wait = 1, |
8a8b6502 AM |
3123 | }; |
3124 | ||
3125 | static int __init setup_failslab(char *str) | |
3126 | { | |
3127 | return setup_fault_attr(&failslab.attr, str); | |
3128 | } | |
3129 | __setup("failslab=", setup_failslab); | |
3130 | ||
3131 | static int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3132 | { | |
3133 | if (cachep == &cache_cache) | |
3134 | return 0; | |
3135 | if (flags & __GFP_NOFAIL) | |
3136 | return 0; | |
3137 | if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) | |
3138 | return 0; | |
3139 | ||
3140 | return should_fail(&failslab.attr, obj_size(cachep)); | |
3141 | } | |
3142 | ||
3143 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3144 | ||
3145 | static int __init failslab_debugfs(void) | |
3146 | { | |
3147 | mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
3148 | struct dentry *dir; | |
3149 | int err; | |
3150 | ||
824ebef1 | 3151 | err = init_fault_attr_dentries(&failslab.attr, "failslab"); |
8a8b6502 AM |
3152 | if (err) |
3153 | return err; | |
3154 | dir = failslab.attr.dentries.dir; | |
3155 | ||
3156 | failslab.ignore_gfp_wait_file = | |
3157 | debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
3158 | &failslab.ignore_gfp_wait); | |
3159 | ||
3160 | if (!failslab.ignore_gfp_wait_file) { | |
3161 | err = -ENOMEM; | |
3162 | debugfs_remove(failslab.ignore_gfp_wait_file); | |
3163 | cleanup_fault_attr_dentries(&failslab.attr); | |
3164 | } | |
3165 | ||
3166 | return err; | |
3167 | } | |
3168 | ||
3169 | late_initcall(failslab_debugfs); | |
3170 | ||
3171 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
3172 | ||
3173 | #else /* CONFIG_FAILSLAB */ | |
3174 | ||
3175 | static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3176 | { | |
3177 | return 0; | |
3178 | } | |
3179 | ||
3180 | #endif /* CONFIG_FAILSLAB */ | |
3181 | ||
343e0d7a | 3182 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3183 | { |
b28a02de | 3184 | void *objp; |
1da177e4 LT |
3185 | struct array_cache *ac; |
3186 | ||
5c382300 | 3187 | check_irq_off(); |
8a8b6502 | 3188 | |
9a2dba4b | 3189 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3190 | if (likely(ac->avail)) { |
3191 | STATS_INC_ALLOCHIT(cachep); | |
3192 | ac->touched = 1; | |
e498be7d | 3193 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3194 | } else { |
3195 | STATS_INC_ALLOCMISS(cachep); | |
3196 | objp = cache_alloc_refill(cachep, flags); | |
3197 | } | |
5c382300 AK |
3198 | return objp; |
3199 | } | |
3200 | ||
e498be7d | 3201 | #ifdef CONFIG_NUMA |
c61afb18 | 3202 | /* |
b2455396 | 3203 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3204 | * |
3205 | * If we are in_interrupt, then process context, including cpusets and | |
3206 | * mempolicy, may not apply and should not be used for allocation policy. | |
3207 | */ | |
3208 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3209 | { | |
3210 | int nid_alloc, nid_here; | |
3211 | ||
765c4507 | 3212 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3213 | return NULL; |
3214 | nid_alloc = nid_here = numa_node_id(); | |
3215 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3216 | nid_alloc = cpuset_mem_spread_node(); | |
3217 | else if (current->mempolicy) | |
3218 | nid_alloc = slab_node(current->mempolicy); | |
3219 | if (nid_alloc != nid_here) | |
8b98c169 | 3220 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3221 | return NULL; |
3222 | } | |
3223 | ||
765c4507 CL |
3224 | /* |
3225 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3226 | * certain node and fall back is permitted. First we scan all the |
3227 | * available nodelists for available objects. If that fails then we | |
3228 | * perform an allocation without specifying a node. This allows the page | |
3229 | * allocator to do its reclaim / fallback magic. We then insert the | |
3230 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3231 | */ |
8c8cc2c1 | 3232 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3233 | { |
8c8cc2c1 PE |
3234 | struct zonelist *zonelist; |
3235 | gfp_t local_flags; | |
dd1a239f | 3236 | struct zoneref *z; |
54a6eb5c MG |
3237 | struct zone *zone; |
3238 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3239 | void *obj = NULL; |
3c517a61 | 3240 | int nid; |
8c8cc2c1 PE |
3241 | |
3242 | if (flags & __GFP_THISNODE) | |
3243 | return NULL; | |
3244 | ||
0e88460d | 3245 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
6cb06229 | 3246 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3247 | |
3c517a61 CL |
3248 | retry: |
3249 | /* | |
3250 | * Look through allowed nodes for objects available | |
3251 | * from existing per node queues. | |
3252 | */ | |
54a6eb5c MG |
3253 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3254 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3255 | |
54a6eb5c | 3256 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3257 | cache->nodelists[nid] && |
481c5346 | 3258 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3259 | obj = ____cache_alloc_node(cache, |
3260 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3261 | if (obj) |
3262 | break; | |
3263 | } | |
3c517a61 CL |
3264 | } |
3265 | ||
cfce6604 | 3266 | if (!obj) { |
3c517a61 CL |
3267 | /* |
3268 | * This allocation will be performed within the constraints | |
3269 | * of the current cpuset / memory policy requirements. | |
3270 | * We may trigger various forms of reclaim on the allowed | |
3271 | * set and go into memory reserves if necessary. | |
3272 | */ | |
dd47ea75 CL |
3273 | if (local_flags & __GFP_WAIT) |
3274 | local_irq_enable(); | |
3275 | kmem_flagcheck(cache, flags); | |
9ac33b2b | 3276 | obj = kmem_getpages(cache, local_flags, -1); |
dd47ea75 CL |
3277 | if (local_flags & __GFP_WAIT) |
3278 | local_irq_disable(); | |
3c517a61 CL |
3279 | if (obj) { |
3280 | /* | |
3281 | * Insert into the appropriate per node queues | |
3282 | */ | |
3283 | nid = page_to_nid(virt_to_page(obj)); | |
3284 | if (cache_grow(cache, flags, nid, obj)) { | |
3285 | obj = ____cache_alloc_node(cache, | |
3286 | flags | GFP_THISNODE, nid); | |
3287 | if (!obj) | |
3288 | /* | |
3289 | * Another processor may allocate the | |
3290 | * objects in the slab since we are | |
3291 | * not holding any locks. | |
3292 | */ | |
3293 | goto retry; | |
3294 | } else { | |
b6a60451 | 3295 | /* cache_grow already freed obj */ |
3c517a61 CL |
3296 | obj = NULL; |
3297 | } | |
3298 | } | |
aedb0eb1 | 3299 | } |
765c4507 CL |
3300 | return obj; |
3301 | } | |
3302 | ||
e498be7d CL |
3303 | /* |
3304 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3305 | */ |
8b98c169 | 3306 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3307 | int nodeid) |
e498be7d CL |
3308 | { |
3309 | struct list_head *entry; | |
b28a02de PE |
3310 | struct slab *slabp; |
3311 | struct kmem_list3 *l3; | |
3312 | void *obj; | |
b28a02de PE |
3313 | int x; |
3314 | ||
3315 | l3 = cachep->nodelists[nodeid]; | |
3316 | BUG_ON(!l3); | |
3317 | ||
a737b3e2 | 3318 | retry: |
ca3b9b91 | 3319 | check_irq_off(); |
b28a02de PE |
3320 | spin_lock(&l3->list_lock); |
3321 | entry = l3->slabs_partial.next; | |
3322 | if (entry == &l3->slabs_partial) { | |
3323 | l3->free_touched = 1; | |
3324 | entry = l3->slabs_free.next; | |
3325 | if (entry == &l3->slabs_free) | |
3326 | goto must_grow; | |
3327 | } | |
3328 | ||
3329 | slabp = list_entry(entry, struct slab, list); | |
3330 | check_spinlock_acquired_node(cachep, nodeid); | |
3331 | check_slabp(cachep, slabp); | |
3332 | ||
3333 | STATS_INC_NODEALLOCS(cachep); | |
3334 | STATS_INC_ACTIVE(cachep); | |
3335 | STATS_SET_HIGH(cachep); | |
3336 | ||
3337 | BUG_ON(slabp->inuse == cachep->num); | |
3338 | ||
78d382d7 | 3339 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3340 | check_slabp(cachep, slabp); |
3341 | l3->free_objects--; | |
3342 | /* move slabp to correct slabp list: */ | |
3343 | list_del(&slabp->list); | |
3344 | ||
a737b3e2 | 3345 | if (slabp->free == BUFCTL_END) |
b28a02de | 3346 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3347 | else |
b28a02de | 3348 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3349 | |
b28a02de PE |
3350 | spin_unlock(&l3->list_lock); |
3351 | goto done; | |
e498be7d | 3352 | |
a737b3e2 | 3353 | must_grow: |
b28a02de | 3354 | spin_unlock(&l3->list_lock); |
3c517a61 | 3355 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3356 | if (x) |
3357 | goto retry; | |
1da177e4 | 3358 | |
8c8cc2c1 | 3359 | return fallback_alloc(cachep, flags); |
e498be7d | 3360 | |
a737b3e2 | 3361 | done: |
b28a02de | 3362 | return obj; |
e498be7d | 3363 | } |
8c8cc2c1 PE |
3364 | |
3365 | /** | |
3366 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3367 | * @cachep: The cache to allocate from. | |
3368 | * @flags: See kmalloc(). | |
3369 | * @nodeid: node number of the target node. | |
3370 | * @caller: return address of caller, used for debug information | |
3371 | * | |
3372 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3373 | * node, which can improve the performance for cpu bound structures. | |
3374 | * | |
3375 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3376 | */ | |
3377 | static __always_inline void * | |
3378 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3379 | void *caller) | |
3380 | { | |
3381 | unsigned long save_flags; | |
3382 | void *ptr; | |
3383 | ||
824ebef1 AM |
3384 | if (should_failslab(cachep, flags)) |
3385 | return NULL; | |
3386 | ||
8c8cc2c1 PE |
3387 | cache_alloc_debugcheck_before(cachep, flags); |
3388 | local_irq_save(save_flags); | |
3389 | ||
3390 | if (unlikely(nodeid == -1)) | |
3391 | nodeid = numa_node_id(); | |
3392 | ||
3393 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3394 | /* Node not bootstrapped yet */ | |
3395 | ptr = fallback_alloc(cachep, flags); | |
3396 | goto out; | |
3397 | } | |
3398 | ||
3399 | if (nodeid == numa_node_id()) { | |
3400 | /* | |
3401 | * Use the locally cached objects if possible. | |
3402 | * However ____cache_alloc does not allow fallback | |
3403 | * to other nodes. It may fail while we still have | |
3404 | * objects on other nodes available. | |
3405 | */ | |
3406 | ptr = ____cache_alloc(cachep, flags); | |
3407 | if (ptr) | |
3408 | goto out; | |
3409 | } | |
3410 | /* ___cache_alloc_node can fall back to other nodes */ | |
3411 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3412 | out: | |
3413 | local_irq_restore(save_flags); | |
3414 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
3415 | ||
d07dbea4 CL |
3416 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3417 | memset(ptr, 0, obj_size(cachep)); | |
3418 | ||
8c8cc2c1 PE |
3419 | return ptr; |
3420 | } | |
3421 | ||
3422 | static __always_inline void * | |
3423 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3424 | { | |
3425 | void *objp; | |
3426 | ||
3427 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3428 | objp = alternate_node_alloc(cache, flags); | |
3429 | if (objp) | |
3430 | goto out; | |
3431 | } | |
3432 | objp = ____cache_alloc(cache, flags); | |
3433 | ||
3434 | /* | |
3435 | * We may just have run out of memory on the local node. | |
3436 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3437 | */ | |
3438 | if (!objp) | |
3439 | objp = ____cache_alloc_node(cache, flags, numa_node_id()); | |
3440 | ||
3441 | out: | |
3442 | return objp; | |
3443 | } | |
3444 | #else | |
3445 | ||
3446 | static __always_inline void * | |
3447 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3448 | { | |
3449 | return ____cache_alloc(cachep, flags); | |
3450 | } | |
3451 | ||
3452 | #endif /* CONFIG_NUMA */ | |
3453 | ||
3454 | static __always_inline void * | |
3455 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3456 | { | |
3457 | unsigned long save_flags; | |
3458 | void *objp; | |
3459 | ||
824ebef1 AM |
3460 | if (should_failslab(cachep, flags)) |
3461 | return NULL; | |
3462 | ||
8c8cc2c1 PE |
3463 | cache_alloc_debugcheck_before(cachep, flags); |
3464 | local_irq_save(save_flags); | |
3465 | objp = __do_cache_alloc(cachep, flags); | |
3466 | local_irq_restore(save_flags); | |
3467 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
3468 | prefetchw(objp); | |
3469 | ||
d07dbea4 CL |
3470 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3471 | memset(objp, 0, obj_size(cachep)); | |
3472 | ||
8c8cc2c1 PE |
3473 | return objp; |
3474 | } | |
e498be7d CL |
3475 | |
3476 | /* | |
3477 | * Caller needs to acquire correct kmem_list's list_lock | |
3478 | */ | |
343e0d7a | 3479 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3480 | int node) |
1da177e4 LT |
3481 | { |
3482 | int i; | |
e498be7d | 3483 | struct kmem_list3 *l3; |
1da177e4 LT |
3484 | |
3485 | for (i = 0; i < nr_objects; i++) { | |
3486 | void *objp = objpp[i]; | |
3487 | struct slab *slabp; | |
1da177e4 | 3488 | |
6ed5eb22 | 3489 | slabp = virt_to_slab(objp); |
ff69416e | 3490 | l3 = cachep->nodelists[node]; |
1da177e4 | 3491 | list_del(&slabp->list); |
ff69416e | 3492 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3493 | check_slabp(cachep, slabp); |
78d382d7 | 3494 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3495 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3496 | l3->free_objects++; |
1da177e4 LT |
3497 | check_slabp(cachep, slabp); |
3498 | ||
3499 | /* fixup slab chains */ | |
3500 | if (slabp->inuse == 0) { | |
e498be7d CL |
3501 | if (l3->free_objects > l3->free_limit) { |
3502 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3503 | /* No need to drop any previously held |
3504 | * lock here, even if we have a off-slab slab | |
3505 | * descriptor it is guaranteed to come from | |
3506 | * a different cache, refer to comments before | |
3507 | * alloc_slabmgmt. | |
3508 | */ | |
1da177e4 LT |
3509 | slab_destroy(cachep, slabp); |
3510 | } else { | |
e498be7d | 3511 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3512 | } |
3513 | } else { | |
3514 | /* Unconditionally move a slab to the end of the | |
3515 | * partial list on free - maximum time for the | |
3516 | * other objects to be freed, too. | |
3517 | */ | |
e498be7d | 3518 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3519 | } |
3520 | } | |
3521 | } | |
3522 | ||
343e0d7a | 3523 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3524 | { |
3525 | int batchcount; | |
e498be7d | 3526 | struct kmem_list3 *l3; |
ff69416e | 3527 | int node = numa_node_id(); |
1da177e4 LT |
3528 | |
3529 | batchcount = ac->batchcount; | |
3530 | #if DEBUG | |
3531 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3532 | #endif | |
3533 | check_irq_off(); | |
ff69416e | 3534 | l3 = cachep->nodelists[node]; |
873623df | 3535 | spin_lock(&l3->list_lock); |
e498be7d CL |
3536 | if (l3->shared) { |
3537 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3538 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3539 | if (max) { |
3540 | if (batchcount > max) | |
3541 | batchcount = max; | |
e498be7d | 3542 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3543 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3544 | shared_array->avail += batchcount; |
3545 | goto free_done; | |
3546 | } | |
3547 | } | |
3548 | ||
ff69416e | 3549 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3550 | free_done: |
1da177e4 LT |
3551 | #if STATS |
3552 | { | |
3553 | int i = 0; | |
3554 | struct list_head *p; | |
3555 | ||
e498be7d CL |
3556 | p = l3->slabs_free.next; |
3557 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3558 | struct slab *slabp; |
3559 | ||
3560 | slabp = list_entry(p, struct slab, list); | |
3561 | BUG_ON(slabp->inuse); | |
3562 | ||
3563 | i++; | |
3564 | p = p->next; | |
3565 | } | |
3566 | STATS_SET_FREEABLE(cachep, i); | |
3567 | } | |
3568 | #endif | |
e498be7d | 3569 | spin_unlock(&l3->list_lock); |
1da177e4 | 3570 | ac->avail -= batchcount; |
a737b3e2 | 3571 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3572 | } |
3573 | ||
3574 | /* | |
a737b3e2 AM |
3575 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3576 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3577 | */ |
873623df | 3578 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3579 | { |
9a2dba4b | 3580 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3581 | |
3582 | check_irq_off(); | |
3583 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3584 | ||
1807a1aa SS |
3585 | /* |
3586 | * Skip calling cache_free_alien() when the platform is not numa. | |
3587 | * This will avoid cache misses that happen while accessing slabp (which | |
3588 | * is per page memory reference) to get nodeid. Instead use a global | |
3589 | * variable to skip the call, which is mostly likely to be present in | |
3590 | * the cache. | |
3591 | */ | |
3592 | if (numa_platform && cache_free_alien(cachep, objp)) | |
729bd0b7 PE |
3593 | return; |
3594 | ||
1da177e4 LT |
3595 | if (likely(ac->avail < ac->limit)) { |
3596 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3597 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3598 | return; |
3599 | } else { | |
3600 | STATS_INC_FREEMISS(cachep); | |
3601 | cache_flusharray(cachep, ac); | |
e498be7d | 3602 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3603 | } |
3604 | } | |
3605 | ||
3606 | /** | |
3607 | * kmem_cache_alloc - Allocate an object | |
3608 | * @cachep: The cache to allocate from. | |
3609 | * @flags: See kmalloc(). | |
3610 | * | |
3611 | * Allocate an object from this cache. The flags are only relevant | |
3612 | * if the cache has no available objects. | |
3613 | */ | |
343e0d7a | 3614 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3615 | { |
7fd6b141 | 3616 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3617 | } |
3618 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3619 | ||
3620 | /** | |
7682486b | 3621 | * kmem_ptr_validate - check if an untrusted pointer might be a slab entry. |
1da177e4 LT |
3622 | * @cachep: the cache we're checking against |
3623 | * @ptr: pointer to validate | |
3624 | * | |
7682486b | 3625 | * This verifies that the untrusted pointer looks sane; |
1da177e4 LT |
3626 | * it is _not_ a guarantee that the pointer is actually |
3627 | * part of the slab cache in question, but it at least | |
3628 | * validates that the pointer can be dereferenced and | |
3629 | * looks half-way sane. | |
3630 | * | |
3631 | * Currently only used for dentry validation. | |
3632 | */ | |
b7f869a2 | 3633 | int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) |
1da177e4 | 3634 | { |
b28a02de | 3635 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3636 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3637 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3638 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3639 | struct page *page; |
3640 | ||
3641 | if (unlikely(addr < min_addr)) | |
3642 | goto out; | |
3643 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3644 | goto out; | |
3645 | if (unlikely(addr & align_mask)) | |
3646 | goto out; | |
3647 | if (unlikely(!kern_addr_valid(addr))) | |
3648 | goto out; | |
3649 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3650 | goto out; | |
3651 | page = virt_to_page(ptr); | |
3652 | if (unlikely(!PageSlab(page))) | |
3653 | goto out; | |
065d41cb | 3654 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3655 | goto out; |
3656 | return 1; | |
a737b3e2 | 3657 | out: |
1da177e4 LT |
3658 | return 0; |
3659 | } | |
3660 | ||
3661 | #ifdef CONFIG_NUMA | |
8b98c169 CH |
3662 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3663 | { | |
3664 | return __cache_alloc_node(cachep, flags, nodeid, | |
3665 | __builtin_return_address(0)); | |
3666 | } | |
1da177e4 LT |
3667 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3668 | ||
8b98c169 CH |
3669 | static __always_inline void * |
3670 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3671 | { |
343e0d7a | 3672 | struct kmem_cache *cachep; |
97e2bde4 MS |
3673 | |
3674 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3675 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3676 | return cachep; | |
97e2bde4 MS |
3677 | return kmem_cache_alloc_node(cachep, flags, node); |
3678 | } | |
8b98c169 CH |
3679 | |
3680 | #ifdef CONFIG_DEBUG_SLAB | |
3681 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3682 | { | |
3683 | return __do_kmalloc_node(size, flags, node, | |
3684 | __builtin_return_address(0)); | |
3685 | } | |
dbe5e69d | 3686 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3687 | |
3688 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
3689 | int node, void *caller) | |
3690 | { | |
3691 | return __do_kmalloc_node(size, flags, node, caller); | |
3692 | } | |
3693 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3694 | #else | |
3695 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3696 | { | |
3697 | return __do_kmalloc_node(size, flags, node, NULL); | |
3698 | } | |
3699 | EXPORT_SYMBOL(__kmalloc_node); | |
3700 | #endif /* CONFIG_DEBUG_SLAB */ | |
3701 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3702 | |
3703 | /** | |
800590f5 | 3704 | * __do_kmalloc - allocate memory |
1da177e4 | 3705 | * @size: how many bytes of memory are required. |
800590f5 | 3706 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3707 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3708 | */ |
7fd6b141 PE |
3709 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3710 | void *caller) | |
1da177e4 | 3711 | { |
343e0d7a | 3712 | struct kmem_cache *cachep; |
1da177e4 | 3713 | |
97e2bde4 MS |
3714 | /* If you want to save a few bytes .text space: replace |
3715 | * __ with kmem_. | |
3716 | * Then kmalloc uses the uninlined functions instead of the inline | |
3717 | * functions. | |
3718 | */ | |
3719 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3720 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3721 | return cachep; | |
7fd6b141 PE |
3722 | return __cache_alloc(cachep, flags, caller); |
3723 | } | |
3724 | ||
7fd6b141 | 3725 | |
1d2c8eea | 3726 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3727 | void *__kmalloc(size_t size, gfp_t flags) |
3728 | { | |
871751e2 | 3729 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3730 | } |
3731 | EXPORT_SYMBOL(__kmalloc); | |
3732 | ||
7fd6b141 PE |
3733 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3734 | { | |
3735 | return __do_kmalloc(size, flags, caller); | |
3736 | } | |
3737 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3738 | |
3739 | #else | |
3740 | void *__kmalloc(size_t size, gfp_t flags) | |
3741 | { | |
3742 | return __do_kmalloc(size, flags, NULL); | |
3743 | } | |
3744 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3745 | #endif |
3746 | ||
1da177e4 LT |
3747 | /** |
3748 | * kmem_cache_free - Deallocate an object | |
3749 | * @cachep: The cache the allocation was from. | |
3750 | * @objp: The previously allocated object. | |
3751 | * | |
3752 | * Free an object which was previously allocated from this | |
3753 | * cache. | |
3754 | */ | |
343e0d7a | 3755 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3756 | { |
3757 | unsigned long flags; | |
3758 | ||
3759 | local_irq_save(flags); | |
898552c9 | 3760 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
3ac7fe5a TG |
3761 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3762 | debug_check_no_obj_freed(objp, obj_size(cachep)); | |
873623df | 3763 | __cache_free(cachep, objp); |
1da177e4 LT |
3764 | local_irq_restore(flags); |
3765 | } | |
3766 | EXPORT_SYMBOL(kmem_cache_free); | |
3767 | ||
1da177e4 LT |
3768 | /** |
3769 | * kfree - free previously allocated memory | |
3770 | * @objp: pointer returned by kmalloc. | |
3771 | * | |
80e93eff PE |
3772 | * If @objp is NULL, no operation is performed. |
3773 | * | |
1da177e4 LT |
3774 | * Don't free memory not originally allocated by kmalloc() |
3775 | * or you will run into trouble. | |
3776 | */ | |
3777 | void kfree(const void *objp) | |
3778 | { | |
343e0d7a | 3779 | struct kmem_cache *c; |
1da177e4 LT |
3780 | unsigned long flags; |
3781 | ||
6cb8f913 | 3782 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3783 | return; |
3784 | local_irq_save(flags); | |
3785 | kfree_debugcheck(objp); | |
6ed5eb22 | 3786 | c = virt_to_cache(objp); |
f9b8404c | 3787 | debug_check_no_locks_freed(objp, obj_size(c)); |
3ac7fe5a | 3788 | debug_check_no_obj_freed(objp, obj_size(c)); |
873623df | 3789 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3790 | local_irq_restore(flags); |
3791 | } | |
3792 | EXPORT_SYMBOL(kfree); | |
3793 | ||
343e0d7a | 3794 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3795 | { |
3dafccf2 | 3796 | return obj_size(cachep); |
1da177e4 LT |
3797 | } |
3798 | EXPORT_SYMBOL(kmem_cache_size); | |
3799 | ||
343e0d7a | 3800 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3801 | { |
3802 | return cachep->name; | |
3803 | } | |
3804 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3805 | ||
e498be7d | 3806 | /* |
183ff22b | 3807 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3808 | */ |
343e0d7a | 3809 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3810 | { |
3811 | int node; | |
3812 | struct kmem_list3 *l3; | |
cafeb02e | 3813 | struct array_cache *new_shared; |
3395ee05 | 3814 | struct array_cache **new_alien = NULL; |
e498be7d | 3815 | |
9c09a95c | 3816 | for_each_online_node(node) { |
cafeb02e | 3817 | |
3395ee05 PM |
3818 | if (use_alien_caches) { |
3819 | new_alien = alloc_alien_cache(node, cachep->limit); | |
3820 | if (!new_alien) | |
3821 | goto fail; | |
3822 | } | |
cafeb02e | 3823 | |
63109846 ED |
3824 | new_shared = NULL; |
3825 | if (cachep->shared) { | |
3826 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3827 | cachep->shared*cachep->batchcount, |
a737b3e2 | 3828 | 0xbaadf00d); |
63109846 ED |
3829 | if (!new_shared) { |
3830 | free_alien_cache(new_alien); | |
3831 | goto fail; | |
3832 | } | |
0718dc2a | 3833 | } |
cafeb02e | 3834 | |
a737b3e2 AM |
3835 | l3 = cachep->nodelists[node]; |
3836 | if (l3) { | |
cafeb02e CL |
3837 | struct array_cache *shared = l3->shared; |
3838 | ||
e498be7d CL |
3839 | spin_lock_irq(&l3->list_lock); |
3840 | ||
cafeb02e | 3841 | if (shared) |
0718dc2a CL |
3842 | free_block(cachep, shared->entry, |
3843 | shared->avail, node); | |
e498be7d | 3844 | |
cafeb02e CL |
3845 | l3->shared = new_shared; |
3846 | if (!l3->alien) { | |
e498be7d CL |
3847 | l3->alien = new_alien; |
3848 | new_alien = NULL; | |
3849 | } | |
b28a02de | 3850 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3851 | cachep->batchcount + cachep->num; |
e498be7d | 3852 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3853 | kfree(shared); |
e498be7d CL |
3854 | free_alien_cache(new_alien); |
3855 | continue; | |
3856 | } | |
a737b3e2 | 3857 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3858 | if (!l3) { |
3859 | free_alien_cache(new_alien); | |
3860 | kfree(new_shared); | |
e498be7d | 3861 | goto fail; |
0718dc2a | 3862 | } |
e498be7d CL |
3863 | |
3864 | kmem_list3_init(l3); | |
3865 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3866 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3867 | l3->shared = new_shared; |
e498be7d | 3868 | l3->alien = new_alien; |
b28a02de | 3869 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3870 | cachep->batchcount + cachep->num; |
e498be7d CL |
3871 | cachep->nodelists[node] = l3; |
3872 | } | |
cafeb02e | 3873 | return 0; |
0718dc2a | 3874 | |
a737b3e2 | 3875 | fail: |
0718dc2a CL |
3876 | if (!cachep->next.next) { |
3877 | /* Cache is not active yet. Roll back what we did */ | |
3878 | node--; | |
3879 | while (node >= 0) { | |
3880 | if (cachep->nodelists[node]) { | |
3881 | l3 = cachep->nodelists[node]; | |
3882 | ||
3883 | kfree(l3->shared); | |
3884 | free_alien_cache(l3->alien); | |
3885 | kfree(l3); | |
3886 | cachep->nodelists[node] = NULL; | |
3887 | } | |
3888 | node--; | |
3889 | } | |
3890 | } | |
cafeb02e | 3891 | return -ENOMEM; |
e498be7d CL |
3892 | } |
3893 | ||
1da177e4 | 3894 | struct ccupdate_struct { |
343e0d7a | 3895 | struct kmem_cache *cachep; |
1da177e4 LT |
3896 | struct array_cache *new[NR_CPUS]; |
3897 | }; | |
3898 | ||
3899 | static void do_ccupdate_local(void *info) | |
3900 | { | |
a737b3e2 | 3901 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3902 | struct array_cache *old; |
3903 | ||
3904 | check_irq_off(); | |
9a2dba4b | 3905 | old = cpu_cache_get(new->cachep); |
e498be7d | 3906 | |
1da177e4 LT |
3907 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3908 | new->new[smp_processor_id()] = old; | |
3909 | } | |
3910 | ||
b5d8ca7c | 3911 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3912 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3913 | int batchcount, int shared) | |
1da177e4 | 3914 | { |
d2e7b7d0 | 3915 | struct ccupdate_struct *new; |
2ed3a4ef | 3916 | int i; |
1da177e4 | 3917 | |
d2e7b7d0 SS |
3918 | new = kzalloc(sizeof(*new), GFP_KERNEL); |
3919 | if (!new) | |
3920 | return -ENOMEM; | |
3921 | ||
e498be7d | 3922 | for_each_online_cpu(i) { |
d2e7b7d0 | 3923 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
a737b3e2 | 3924 | batchcount); |
d2e7b7d0 | 3925 | if (!new->new[i]) { |
b28a02de | 3926 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3927 | kfree(new->new[i]); |
3928 | kfree(new); | |
e498be7d | 3929 | return -ENOMEM; |
1da177e4 LT |
3930 | } |
3931 | } | |
d2e7b7d0 | 3932 | new->cachep = cachep; |
1da177e4 | 3933 | |
15c8b6c1 | 3934 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 3935 | |
1da177e4 | 3936 | check_irq_on(); |
1da177e4 LT |
3937 | cachep->batchcount = batchcount; |
3938 | cachep->limit = limit; | |
e498be7d | 3939 | cachep->shared = shared; |
1da177e4 | 3940 | |
e498be7d | 3941 | for_each_online_cpu(i) { |
d2e7b7d0 | 3942 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3943 | if (!ccold) |
3944 | continue; | |
e498be7d | 3945 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3946 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3947 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3948 | kfree(ccold); |
3949 | } | |
d2e7b7d0 | 3950 | kfree(new); |
2ed3a4ef | 3951 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3952 | } |
3953 | ||
b5d8ca7c | 3954 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3955 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3956 | { |
3957 | int err; | |
3958 | int limit, shared; | |
3959 | ||
a737b3e2 AM |
3960 | /* |
3961 | * The head array serves three purposes: | |
1da177e4 LT |
3962 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3963 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3964 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3965 | * bufctl chains: array operations are cheaper. |
3966 | * The numbers are guessed, we should auto-tune as described by | |
3967 | * Bonwick. | |
3968 | */ | |
3dafccf2 | 3969 | if (cachep->buffer_size > 131072) |
1da177e4 | 3970 | limit = 1; |
3dafccf2 | 3971 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3972 | limit = 8; |
3dafccf2 | 3973 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3974 | limit = 24; |
3dafccf2 | 3975 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3976 | limit = 54; |
3977 | else | |
3978 | limit = 120; | |
3979 | ||
a737b3e2 AM |
3980 | /* |
3981 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3982 | * allocation behaviour: Most allocs on one cpu, most free operations |
3983 | * on another cpu. For these cases, an efficient object passing between | |
3984 | * cpus is necessary. This is provided by a shared array. The array | |
3985 | * replaces Bonwick's magazine layer. | |
3986 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3987 | * to a larger limit. Thus disabled by default. | |
3988 | */ | |
3989 | shared = 0; | |
364fbb29 | 3990 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 3991 | shared = 8; |
1da177e4 LT |
3992 | |
3993 | #if DEBUG | |
a737b3e2 AM |
3994 | /* |
3995 | * With debugging enabled, large batchcount lead to excessively long | |
3996 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3997 | */ |
3998 | if (limit > 32) | |
3999 | limit = 32; | |
4000 | #endif | |
b28a02de | 4001 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
4002 | if (err) |
4003 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4004 | cachep->name, -err); |
2ed3a4ef | 4005 | return err; |
1da177e4 LT |
4006 | } |
4007 | ||
1b55253a CL |
4008 | /* |
4009 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4010 | * necessary. Note that the l3 listlock also protects the array_cache |
4011 | * if drain_array() is used on the shared array. | |
1b55253a CL |
4012 | */ |
4013 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4014 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
4015 | { |
4016 | int tofree; | |
4017 | ||
1b55253a CL |
4018 | if (!ac || !ac->avail) |
4019 | return; | |
1da177e4 LT |
4020 | if (ac->touched && !force) { |
4021 | ac->touched = 0; | |
b18e7e65 | 4022 | } else { |
1b55253a | 4023 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4024 | if (ac->avail) { |
4025 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4026 | if (tofree > ac->avail) | |
4027 | tofree = (ac->avail + 1) / 2; | |
4028 | free_block(cachep, ac->entry, tofree, node); | |
4029 | ac->avail -= tofree; | |
4030 | memmove(ac->entry, &(ac->entry[tofree]), | |
4031 | sizeof(void *) * ac->avail); | |
4032 | } | |
1b55253a | 4033 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4034 | } |
4035 | } | |
4036 | ||
4037 | /** | |
4038 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4039 | * @w: work descriptor |
1da177e4 LT |
4040 | * |
4041 | * Called from workqueue/eventd every few seconds. | |
4042 | * Purpose: | |
4043 | * - clear the per-cpu caches for this CPU. | |
4044 | * - return freeable pages to the main free memory pool. | |
4045 | * | |
a737b3e2 AM |
4046 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4047 | * again on the next iteration. | |
1da177e4 | 4048 | */ |
7c5cae36 | 4049 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4050 | { |
7a7c381d | 4051 | struct kmem_cache *searchp; |
e498be7d | 4052 | struct kmem_list3 *l3; |
aab2207c | 4053 | int node = numa_node_id(); |
7c5cae36 CL |
4054 | struct delayed_work *work = |
4055 | container_of(w, struct delayed_work, work); | |
1da177e4 | 4056 | |
7c5cae36 | 4057 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4058 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4059 | goto out; |
1da177e4 | 4060 | |
7a7c381d | 4061 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4062 | check_irq_on(); |
4063 | ||
35386e3b CL |
4064 | /* |
4065 | * We only take the l3 lock if absolutely necessary and we | |
4066 | * have established with reasonable certainty that | |
4067 | * we can do some work if the lock was obtained. | |
4068 | */ | |
aab2207c | 4069 | l3 = searchp->nodelists[node]; |
35386e3b | 4070 | |
8fce4d8e | 4071 | reap_alien(searchp, l3); |
1da177e4 | 4072 | |
aab2207c | 4073 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4074 | |
35386e3b CL |
4075 | /* |
4076 | * These are racy checks but it does not matter | |
4077 | * if we skip one check or scan twice. | |
4078 | */ | |
e498be7d | 4079 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4080 | goto next; |
1da177e4 | 4081 | |
e498be7d | 4082 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4083 | |
aab2207c | 4084 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4085 | |
ed11d9eb | 4086 | if (l3->free_touched) |
e498be7d | 4087 | l3->free_touched = 0; |
ed11d9eb CL |
4088 | else { |
4089 | int freed; | |
1da177e4 | 4090 | |
ed11d9eb CL |
4091 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4092 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4093 | STATS_ADD_REAPED(searchp, freed); | |
4094 | } | |
35386e3b | 4095 | next: |
1da177e4 LT |
4096 | cond_resched(); |
4097 | } | |
4098 | check_irq_on(); | |
fc0abb14 | 4099 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4100 | next_reap_node(); |
7c5cae36 | 4101 | out: |
a737b3e2 | 4102 | /* Set up the next iteration */ |
7c5cae36 | 4103 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4104 | } |
4105 | ||
158a9624 | 4106 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4107 | |
85289f98 | 4108 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4109 | { |
85289f98 PE |
4110 | /* |
4111 | * Output format version, so at least we can change it | |
4112 | * without _too_ many complaints. | |
4113 | */ | |
1da177e4 | 4114 | #if STATS |
85289f98 | 4115 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4116 | #else |
85289f98 | 4117 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4118 | #endif |
85289f98 PE |
4119 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4120 | "<objperslab> <pagesperslab>"); | |
4121 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4122 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4123 | #if STATS |
85289f98 | 4124 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4125 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4126 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4127 | #endif |
85289f98 PE |
4128 | seq_putc(m, '\n'); |
4129 | } | |
4130 | ||
4131 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4132 | { | |
4133 | loff_t n = *pos; | |
85289f98 | 4134 | |
fc0abb14 | 4135 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4136 | if (!n) |
4137 | print_slabinfo_header(m); | |
b92151ba PE |
4138 | |
4139 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4140 | } |
4141 | ||
4142 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4143 | { | |
b92151ba | 4144 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4145 | } |
4146 | ||
4147 | static void s_stop(struct seq_file *m, void *p) | |
4148 | { | |
fc0abb14 | 4149 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4150 | } |
4151 | ||
4152 | static int s_show(struct seq_file *m, void *p) | |
4153 | { | |
b92151ba | 4154 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
b28a02de PE |
4155 | struct slab *slabp; |
4156 | unsigned long active_objs; | |
4157 | unsigned long num_objs; | |
4158 | unsigned long active_slabs = 0; | |
4159 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4160 | const char *name; |
1da177e4 | 4161 | char *error = NULL; |
e498be7d CL |
4162 | int node; |
4163 | struct kmem_list3 *l3; | |
1da177e4 | 4164 | |
1da177e4 LT |
4165 | active_objs = 0; |
4166 | num_slabs = 0; | |
e498be7d CL |
4167 | for_each_online_node(node) { |
4168 | l3 = cachep->nodelists[node]; | |
4169 | if (!l3) | |
4170 | continue; | |
4171 | ||
ca3b9b91 RT |
4172 | check_irq_on(); |
4173 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4174 | |
7a7c381d | 4175 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4176 | if (slabp->inuse != cachep->num && !error) |
4177 | error = "slabs_full accounting error"; | |
4178 | active_objs += cachep->num; | |
4179 | active_slabs++; | |
4180 | } | |
7a7c381d | 4181 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4182 | if (slabp->inuse == cachep->num && !error) |
4183 | error = "slabs_partial inuse accounting error"; | |
4184 | if (!slabp->inuse && !error) | |
4185 | error = "slabs_partial/inuse accounting error"; | |
4186 | active_objs += slabp->inuse; | |
4187 | active_slabs++; | |
4188 | } | |
7a7c381d | 4189 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4190 | if (slabp->inuse && !error) |
4191 | error = "slabs_free/inuse accounting error"; | |
4192 | num_slabs++; | |
4193 | } | |
4194 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4195 | if (l3->shared) |
4196 | shared_avail += l3->shared->avail; | |
e498be7d | 4197 | |
ca3b9b91 | 4198 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4199 | } |
b28a02de PE |
4200 | num_slabs += active_slabs; |
4201 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4202 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4203 | error = "free_objects accounting error"; |
4204 | ||
b28a02de | 4205 | name = cachep->name; |
1da177e4 LT |
4206 | if (error) |
4207 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4208 | ||
4209 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4210 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4211 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4212 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4213 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4214 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4215 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4216 | #if STATS |
b28a02de | 4217 | { /* list3 stats */ |
1da177e4 LT |
4218 | unsigned long high = cachep->high_mark; |
4219 | unsigned long allocs = cachep->num_allocations; | |
4220 | unsigned long grown = cachep->grown; | |
4221 | unsigned long reaped = cachep->reaped; | |
4222 | unsigned long errors = cachep->errors; | |
4223 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4224 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4225 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4226 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4227 | |
e498be7d | 4228 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4229 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4230 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4231 | node_frees, overflows); |
1da177e4 LT |
4232 | } |
4233 | /* cpu stats */ | |
4234 | { | |
4235 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4236 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4237 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4238 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4239 | ||
4240 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4241 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4242 | } |
4243 | #endif | |
4244 | seq_putc(m, '\n'); | |
1da177e4 LT |
4245 | return 0; |
4246 | } | |
4247 | ||
4248 | /* | |
4249 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4250 | * | |
4251 | * Output layout: | |
4252 | * cache-name | |
4253 | * num-active-objs | |
4254 | * total-objs | |
4255 | * object size | |
4256 | * num-active-slabs | |
4257 | * total-slabs | |
4258 | * num-pages-per-slab | |
4259 | * + further values on SMP and with statistics enabled | |
4260 | */ | |
4261 | ||
7b3c3a50 | 4262 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4263 | .start = s_start, |
4264 | .next = s_next, | |
4265 | .stop = s_stop, | |
4266 | .show = s_show, | |
1da177e4 LT |
4267 | }; |
4268 | ||
4269 | #define MAX_SLABINFO_WRITE 128 | |
4270 | /** | |
4271 | * slabinfo_write - Tuning for the slab allocator | |
4272 | * @file: unused | |
4273 | * @buffer: user buffer | |
4274 | * @count: data length | |
4275 | * @ppos: unused | |
4276 | */ | |
b28a02de PE |
4277 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4278 | size_t count, loff_t *ppos) | |
1da177e4 | 4279 | { |
b28a02de | 4280 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4281 | int limit, batchcount, shared, res; |
7a7c381d | 4282 | struct kmem_cache *cachep; |
b28a02de | 4283 | |
1da177e4 LT |
4284 | if (count > MAX_SLABINFO_WRITE) |
4285 | return -EINVAL; | |
4286 | if (copy_from_user(&kbuf, buffer, count)) | |
4287 | return -EFAULT; | |
b28a02de | 4288 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4289 | |
4290 | tmp = strchr(kbuf, ' '); | |
4291 | if (!tmp) | |
4292 | return -EINVAL; | |
4293 | *tmp = '\0'; | |
4294 | tmp++; | |
4295 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4296 | return -EINVAL; | |
4297 | ||
4298 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4299 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4300 | res = -EINVAL; |
7a7c381d | 4301 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4302 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4303 | if (limit < 1 || batchcount < 1 || |
4304 | batchcount > limit || shared < 0) { | |
e498be7d | 4305 | res = 0; |
1da177e4 | 4306 | } else { |
e498be7d | 4307 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4308 | batchcount, shared); |
1da177e4 LT |
4309 | } |
4310 | break; | |
4311 | } | |
4312 | } | |
fc0abb14 | 4313 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4314 | if (res >= 0) |
4315 | res = count; | |
4316 | return res; | |
4317 | } | |
871751e2 | 4318 | |
7b3c3a50 AD |
4319 | static int slabinfo_open(struct inode *inode, struct file *file) |
4320 | { | |
4321 | return seq_open(file, &slabinfo_op); | |
4322 | } | |
4323 | ||
4324 | static const struct file_operations proc_slabinfo_operations = { | |
4325 | .open = slabinfo_open, | |
4326 | .read = seq_read, | |
4327 | .write = slabinfo_write, | |
4328 | .llseek = seq_lseek, | |
4329 | .release = seq_release, | |
4330 | }; | |
4331 | ||
871751e2 AV |
4332 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4333 | ||
4334 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4335 | { | |
871751e2 | 4336 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4337 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4338 | } |
4339 | ||
4340 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4341 | { | |
4342 | unsigned long *p; | |
4343 | int l; | |
4344 | if (!v) | |
4345 | return 1; | |
4346 | l = n[1]; | |
4347 | p = n + 2; | |
4348 | while (l) { | |
4349 | int i = l/2; | |
4350 | unsigned long *q = p + 2 * i; | |
4351 | if (*q == v) { | |
4352 | q[1]++; | |
4353 | return 1; | |
4354 | } | |
4355 | if (*q > v) { | |
4356 | l = i; | |
4357 | } else { | |
4358 | p = q + 2; | |
4359 | l -= i + 1; | |
4360 | } | |
4361 | } | |
4362 | if (++n[1] == n[0]) | |
4363 | return 0; | |
4364 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4365 | p[0] = v; | |
4366 | p[1] = 1; | |
4367 | return 1; | |
4368 | } | |
4369 | ||
4370 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4371 | { | |
4372 | void *p; | |
4373 | int i; | |
4374 | if (n[0] == n[1]) | |
4375 | return; | |
4376 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4377 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4378 | continue; | |
4379 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4380 | return; | |
4381 | } | |
4382 | } | |
4383 | ||
4384 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4385 | { | |
4386 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4387 | unsigned long offset, size; |
9281acea | 4388 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4389 | |
a5c43dae | 4390 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4391 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4392 | if (modname[0]) |
871751e2 AV |
4393 | seq_printf(m, " [%s]", modname); |
4394 | return; | |
4395 | } | |
4396 | #endif | |
4397 | seq_printf(m, "%p", (void *)address); | |
4398 | } | |
4399 | ||
4400 | static int leaks_show(struct seq_file *m, void *p) | |
4401 | { | |
b92151ba | 4402 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4403 | struct slab *slabp; |
4404 | struct kmem_list3 *l3; | |
4405 | const char *name; | |
4406 | unsigned long *n = m->private; | |
4407 | int node; | |
4408 | int i; | |
4409 | ||
4410 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4411 | return 0; | |
4412 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4413 | return 0; | |
4414 | ||
4415 | /* OK, we can do it */ | |
4416 | ||
4417 | n[1] = 0; | |
4418 | ||
4419 | for_each_online_node(node) { | |
4420 | l3 = cachep->nodelists[node]; | |
4421 | if (!l3) | |
4422 | continue; | |
4423 | ||
4424 | check_irq_on(); | |
4425 | spin_lock_irq(&l3->list_lock); | |
4426 | ||
7a7c381d | 4427 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4428 | handle_slab(n, cachep, slabp); |
7a7c381d | 4429 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4430 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4431 | spin_unlock_irq(&l3->list_lock); |
4432 | } | |
4433 | name = cachep->name; | |
4434 | if (n[0] == n[1]) { | |
4435 | /* Increase the buffer size */ | |
4436 | mutex_unlock(&cache_chain_mutex); | |
4437 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4438 | if (!m->private) { | |
4439 | /* Too bad, we are really out */ | |
4440 | m->private = n; | |
4441 | mutex_lock(&cache_chain_mutex); | |
4442 | return -ENOMEM; | |
4443 | } | |
4444 | *(unsigned long *)m->private = n[0] * 2; | |
4445 | kfree(n); | |
4446 | mutex_lock(&cache_chain_mutex); | |
4447 | /* Now make sure this entry will be retried */ | |
4448 | m->count = m->size; | |
4449 | return 0; | |
4450 | } | |
4451 | for (i = 0; i < n[1]; i++) { | |
4452 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4453 | show_symbol(m, n[2*i+2]); | |
4454 | seq_putc(m, '\n'); | |
4455 | } | |
d2e7b7d0 | 4456 | |
871751e2 AV |
4457 | return 0; |
4458 | } | |
4459 | ||
a0ec95a8 | 4460 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4461 | .start = leaks_start, |
4462 | .next = s_next, | |
4463 | .stop = s_stop, | |
4464 | .show = leaks_show, | |
4465 | }; | |
a0ec95a8 AD |
4466 | |
4467 | static int slabstats_open(struct inode *inode, struct file *file) | |
4468 | { | |
4469 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4470 | int ret = -ENOMEM; | |
4471 | if (n) { | |
4472 | ret = seq_open(file, &slabstats_op); | |
4473 | if (!ret) { | |
4474 | struct seq_file *m = file->private_data; | |
4475 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4476 | m->private = n; | |
4477 | n = NULL; | |
4478 | } | |
4479 | kfree(n); | |
4480 | } | |
4481 | return ret; | |
4482 | } | |
4483 | ||
4484 | static const struct file_operations proc_slabstats_operations = { | |
4485 | .open = slabstats_open, | |
4486 | .read = seq_read, | |
4487 | .llseek = seq_lseek, | |
4488 | .release = seq_release_private, | |
4489 | }; | |
4490 | #endif | |
4491 | ||
4492 | static int __init slab_proc_init(void) | |
4493 | { | |
7b3c3a50 | 4494 | proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4495 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4496 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4497 | #endif |
a0ec95a8 AD |
4498 | return 0; |
4499 | } | |
4500 | module_init(slab_proc_init); | |
1da177e4 LT |
4501 | #endif |
4502 | ||
00e145b6 MS |
4503 | /** |
4504 | * ksize - get the actual amount of memory allocated for a given object | |
4505 | * @objp: Pointer to the object | |
4506 | * | |
4507 | * kmalloc may internally round up allocations and return more memory | |
4508 | * than requested. ksize() can be used to determine the actual amount of | |
4509 | * memory allocated. The caller may use this additional memory, even though | |
4510 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4511 | * The caller must guarantee that objp points to a valid object previously | |
4512 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4513 | * must not be freed during the duration of the call. | |
4514 | */ | |
fd76bab2 | 4515 | size_t ksize(const void *objp) |
1da177e4 | 4516 | { |
ef8b4520 CL |
4517 | BUG_ON(!objp); |
4518 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4519 | return 0; |
1da177e4 | 4520 | |
6ed5eb22 | 4521 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4522 | } |