]>
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> |
1da177e4 LT |
98 | #include <linux/seq_file.h> |
99 | #include <linux/notifier.h> | |
100 | #include <linux/kallsyms.h> | |
101 | #include <linux/cpu.h> | |
102 | #include <linux/sysctl.h> | |
103 | #include <linux/module.h> | |
104 | #include <linux/rcupdate.h> | |
543537bd | 105 | #include <linux/string.h> |
138ae663 | 106 | #include <linux/uaccess.h> |
e498be7d | 107 | #include <linux/nodemask.h> |
dc85da15 | 108 | #include <linux/mempolicy.h> |
fc0abb14 | 109 | #include <linux/mutex.h> |
8a8b6502 | 110 | #include <linux/fault-inject.h> |
e7eebaf6 | 111 | #include <linux/rtmutex.h> |
6a2d7a95 | 112 | #include <linux/reciprocal_div.h> |
1da177e4 | 113 | |
1da177e4 LT |
114 | #include <asm/cacheflush.h> |
115 | #include <asm/tlbflush.h> | |
116 | #include <asm/page.h> | |
117 | ||
118 | /* | |
50953fe9 | 119 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
120 | * 0 for faster, smaller code (especially in the critical paths). |
121 | * | |
122 | * STATS - 1 to collect stats for /proc/slabinfo. | |
123 | * 0 for faster, smaller code (especially in the critical paths). | |
124 | * | |
125 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
126 | */ | |
127 | ||
128 | #ifdef CONFIG_DEBUG_SLAB | |
129 | #define DEBUG 1 | |
130 | #define STATS 1 | |
131 | #define FORCED_DEBUG 1 | |
132 | #else | |
133 | #define DEBUG 0 | |
134 | #define STATS 0 | |
135 | #define FORCED_DEBUG 0 | |
136 | #endif | |
137 | ||
1da177e4 LT |
138 | /* Shouldn't this be in a header file somewhere? */ |
139 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 140 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 LT |
141 | |
142 | #ifndef cache_line_size | |
143 | #define cache_line_size() L1_CACHE_BYTES | |
144 | #endif | |
145 | ||
146 | #ifndef ARCH_KMALLOC_MINALIGN | |
147 | /* | |
148 | * Enforce a minimum alignment for the kmalloc caches. | |
149 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
150 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
151 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
b46b8f19 DW |
152 | * alignment larger than the alignment of a 64-bit integer. |
153 | * ARCH_KMALLOC_MINALIGN allows that. | |
154 | * Note that increasing this value may disable some debug features. | |
1da177e4 | 155 | */ |
b46b8f19 | 156 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
1da177e4 LT |
157 | #endif |
158 | ||
159 | #ifndef ARCH_SLAB_MINALIGN | |
160 | /* | |
161 | * Enforce a minimum alignment for all caches. | |
162 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
163 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
164 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
165 | * some debug features. | |
166 | */ | |
167 | #define ARCH_SLAB_MINALIGN 0 | |
168 | #endif | |
169 | ||
170 | #ifndef ARCH_KMALLOC_FLAGS | |
171 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
172 | #endif | |
173 | ||
174 | /* Legal flag mask for kmem_cache_create(). */ | |
175 | #if DEBUG | |
50953fe9 | 176 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 177 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 178 | SLAB_CACHE_DMA | \ |
5af60839 | 179 | SLAB_STORE_USER | \ |
1da177e4 | 180 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
101a5001 | 181 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 182 | #else |
ac2b898c | 183 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 184 | SLAB_CACHE_DMA | \ |
1da177e4 | 185 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
101a5001 | 186 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
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 | */ | |
307 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
308 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
309 | #define CACHE_CACHE 0 | |
310 | #define SIZE_AC 1 | |
311 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
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++; | |
336 | #include "linux/kmalloc_sizes.h" | |
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 */ | |
4ba9b9d0 | 410 | void (*ctor)(struct kmem_cache *, void *); |
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 RT |
732 | /* |
733 | * 1. Guard access to the cache-chain. | |
734 | * 2. Protect sanity of cpu_online_map against cpu hotplug events | |
735 | */ | |
fc0abb14 | 736 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
737 | static struct list_head cache_chain; |
738 | ||
1da177e4 LT |
739 | /* |
740 | * chicken and egg problem: delay the per-cpu array allocation | |
741 | * until the general caches are up. | |
742 | */ | |
743 | static enum { | |
744 | NONE, | |
e498be7d CL |
745 | PARTIAL_AC, |
746 | PARTIAL_L3, | |
1da177e4 LT |
747 | FULL |
748 | } g_cpucache_up; | |
749 | ||
39d24e64 MK |
750 | /* |
751 | * used by boot code to determine if it can use slab based allocator | |
752 | */ | |
753 | int slab_is_available(void) | |
754 | { | |
755 | return g_cpucache_up == FULL; | |
756 | } | |
757 | ||
52bad64d | 758 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 759 | |
343e0d7a | 760 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
761 | { |
762 | return cachep->array[smp_processor_id()]; | |
763 | } | |
764 | ||
a737b3e2 AM |
765 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
766 | gfp_t gfpflags) | |
1da177e4 LT |
767 | { |
768 | struct cache_sizes *csizep = malloc_sizes; | |
769 | ||
770 | #if DEBUG | |
771 | /* This happens if someone tries to call | |
b28a02de PE |
772 | * kmem_cache_create(), or __kmalloc(), before |
773 | * the generic caches are initialized. | |
774 | */ | |
c7e43c78 | 775 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 776 | #endif |
6cb8f913 CL |
777 | if (!size) |
778 | return ZERO_SIZE_PTR; | |
779 | ||
1da177e4 LT |
780 | while (size > csizep->cs_size) |
781 | csizep++; | |
782 | ||
783 | /* | |
0abf40c1 | 784 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
785 | * has cs_{dma,}cachep==NULL. Thus no special case |
786 | * for large kmalloc calls required. | |
787 | */ | |
4b51d669 | 788 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
789 | if (unlikely(gfpflags & GFP_DMA)) |
790 | return csizep->cs_dmacachep; | |
4b51d669 | 791 | #endif |
1da177e4 LT |
792 | return csizep->cs_cachep; |
793 | } | |
794 | ||
b221385b | 795 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
796 | { |
797 | return __find_general_cachep(size, gfpflags); | |
798 | } | |
97e2bde4 | 799 | |
fbaccacf | 800 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 801 | { |
fbaccacf SR |
802 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
803 | } | |
1da177e4 | 804 | |
a737b3e2 AM |
805 | /* |
806 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
807 | */ | |
fbaccacf SR |
808 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
809 | size_t align, int flags, size_t *left_over, | |
810 | unsigned int *num) | |
811 | { | |
812 | int nr_objs; | |
813 | size_t mgmt_size; | |
814 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 815 | |
fbaccacf SR |
816 | /* |
817 | * The slab management structure can be either off the slab or | |
818 | * on it. For the latter case, the memory allocated for a | |
819 | * slab is used for: | |
820 | * | |
821 | * - The struct slab | |
822 | * - One kmem_bufctl_t for each object | |
823 | * - Padding to respect alignment of @align | |
824 | * - @buffer_size bytes for each object | |
825 | * | |
826 | * If the slab management structure is off the slab, then the | |
827 | * alignment will already be calculated into the size. Because | |
828 | * the slabs are all pages aligned, the objects will be at the | |
829 | * correct alignment when allocated. | |
830 | */ | |
831 | if (flags & CFLGS_OFF_SLAB) { | |
832 | mgmt_size = 0; | |
833 | nr_objs = slab_size / buffer_size; | |
834 | ||
835 | if (nr_objs > SLAB_LIMIT) | |
836 | nr_objs = SLAB_LIMIT; | |
837 | } else { | |
838 | /* | |
839 | * Ignore padding for the initial guess. The padding | |
840 | * is at most @align-1 bytes, and @buffer_size is at | |
841 | * least @align. In the worst case, this result will | |
842 | * be one greater than the number of objects that fit | |
843 | * into the memory allocation when taking the padding | |
844 | * into account. | |
845 | */ | |
846 | nr_objs = (slab_size - sizeof(struct slab)) / | |
847 | (buffer_size + sizeof(kmem_bufctl_t)); | |
848 | ||
849 | /* | |
850 | * This calculated number will be either the right | |
851 | * amount, or one greater than what we want. | |
852 | */ | |
853 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
854 | > slab_size) | |
855 | nr_objs--; | |
856 | ||
857 | if (nr_objs > SLAB_LIMIT) | |
858 | nr_objs = SLAB_LIMIT; | |
859 | ||
860 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
861 | } | |
862 | *num = nr_objs; | |
863 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
864 | } |
865 | ||
866 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
867 | ||
a737b3e2 AM |
868 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
869 | char *msg) | |
1da177e4 LT |
870 | { |
871 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 872 | function, cachep->name, msg); |
1da177e4 LT |
873 | dump_stack(); |
874 | } | |
875 | ||
3395ee05 PM |
876 | /* |
877 | * By default on NUMA we use alien caches to stage the freeing of | |
878 | * objects allocated from other nodes. This causes massive memory | |
879 | * inefficiencies when using fake NUMA setup to split memory into a | |
880 | * large number of small nodes, so it can be disabled on the command | |
881 | * line | |
882 | */ | |
883 | ||
884 | static int use_alien_caches __read_mostly = 1; | |
1807a1aa | 885 | static int numa_platform __read_mostly = 1; |
3395ee05 PM |
886 | static int __init noaliencache_setup(char *s) |
887 | { | |
888 | use_alien_caches = 0; | |
889 | return 1; | |
890 | } | |
891 | __setup("noaliencache", noaliencache_setup); | |
892 | ||
8fce4d8e CL |
893 | #ifdef CONFIG_NUMA |
894 | /* | |
895 | * Special reaping functions for NUMA systems called from cache_reap(). | |
896 | * These take care of doing round robin flushing of alien caches (containing | |
897 | * objects freed on different nodes from which they were allocated) and the | |
898 | * flushing of remote pcps by calling drain_node_pages. | |
899 | */ | |
900 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
901 | ||
902 | static void init_reap_node(int cpu) | |
903 | { | |
904 | int node; | |
905 | ||
906 | node = next_node(cpu_to_node(cpu), node_online_map); | |
907 | if (node == MAX_NUMNODES) | |
442295c9 | 908 | node = first_node(node_online_map); |
8fce4d8e | 909 | |
7f6b8876 | 910 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
911 | } |
912 | ||
913 | static void next_reap_node(void) | |
914 | { | |
915 | int node = __get_cpu_var(reap_node); | |
916 | ||
8fce4d8e CL |
917 | node = next_node(node, node_online_map); |
918 | if (unlikely(node >= MAX_NUMNODES)) | |
919 | node = first_node(node_online_map); | |
920 | __get_cpu_var(reap_node) = node; | |
921 | } | |
922 | ||
923 | #else | |
924 | #define init_reap_node(cpu) do { } while (0) | |
925 | #define next_reap_node(void) do { } while (0) | |
926 | #endif | |
927 | ||
1da177e4 LT |
928 | /* |
929 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
930 | * via the workqueue/eventd. | |
931 | * Add the CPU number into the expiration time to minimize the possibility of | |
932 | * the CPUs getting into lockstep and contending for the global cache chain | |
933 | * lock. | |
934 | */ | |
897e679b | 935 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 936 | { |
52bad64d | 937 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
938 | |
939 | /* | |
940 | * When this gets called from do_initcalls via cpucache_init(), | |
941 | * init_workqueues() has already run, so keventd will be setup | |
942 | * at that time. | |
943 | */ | |
52bad64d | 944 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 945 | init_reap_node(cpu); |
65f27f38 | 946 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
947 | schedule_delayed_work_on(cpu, reap_work, |
948 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
949 | } |
950 | } | |
951 | ||
e498be7d | 952 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 953 | int batchcount) |
1da177e4 | 954 | { |
b28a02de | 955 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
956 | struct array_cache *nc = NULL; |
957 | ||
e498be7d | 958 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
959 | if (nc) { |
960 | nc->avail = 0; | |
961 | nc->limit = entries; | |
962 | nc->batchcount = batchcount; | |
963 | nc->touched = 0; | |
e498be7d | 964 | spin_lock_init(&nc->lock); |
1da177e4 LT |
965 | } |
966 | return nc; | |
967 | } | |
968 | ||
3ded175a CL |
969 | /* |
970 | * Transfer objects in one arraycache to another. | |
971 | * Locking must be handled by the caller. | |
972 | * | |
973 | * Return the number of entries transferred. | |
974 | */ | |
975 | static int transfer_objects(struct array_cache *to, | |
976 | struct array_cache *from, unsigned int max) | |
977 | { | |
978 | /* Figure out how many entries to transfer */ | |
979 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
980 | ||
981 | if (!nr) | |
982 | return 0; | |
983 | ||
984 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
985 | sizeof(void *) *nr); | |
986 | ||
987 | from->avail -= nr; | |
988 | to->avail += nr; | |
989 | to->touched = 1; | |
990 | return nr; | |
991 | } | |
992 | ||
765c4507 CL |
993 | #ifndef CONFIG_NUMA |
994 | ||
995 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
996 | #define reap_alien(cachep, l3) do { } while (0) | |
997 | ||
998 | static inline struct array_cache **alloc_alien_cache(int node, int limit) | |
999 | { | |
1000 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1001 | } | |
1002 | ||
1003 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1004 | { | |
1005 | } | |
1006 | ||
1007 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1008 | { | |
1009 | return 0; | |
1010 | } | |
1011 | ||
1012 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1013 | gfp_t flags) | |
1014 | { | |
1015 | return NULL; | |
1016 | } | |
1017 | ||
8b98c169 | 1018 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1019 | gfp_t flags, int nodeid) |
1020 | { | |
1021 | return NULL; | |
1022 | } | |
1023 | ||
1024 | #else /* CONFIG_NUMA */ | |
1025 | ||
8b98c169 | 1026 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1027 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1028 | |
5295a74c | 1029 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
1030 | { |
1031 | struct array_cache **ac_ptr; | |
8ef82866 | 1032 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1033 | int i; |
1034 | ||
1035 | if (limit > 1) | |
1036 | limit = 12; | |
1037 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
1038 | if (ac_ptr) { | |
1039 | for_each_node(i) { | |
1040 | if (i == node || !node_online(i)) { | |
1041 | ac_ptr[i] = NULL; | |
1042 | continue; | |
1043 | } | |
1044 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
1045 | if (!ac_ptr[i]) { | |
cc550def | 1046 | for (i--; i >= 0; i--) |
e498be7d CL |
1047 | kfree(ac_ptr[i]); |
1048 | kfree(ac_ptr); | |
1049 | return NULL; | |
1050 | } | |
1051 | } | |
1052 | } | |
1053 | return ac_ptr; | |
1054 | } | |
1055 | ||
5295a74c | 1056 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1057 | { |
1058 | int i; | |
1059 | ||
1060 | if (!ac_ptr) | |
1061 | return; | |
e498be7d | 1062 | for_each_node(i) |
b28a02de | 1063 | kfree(ac_ptr[i]); |
e498be7d CL |
1064 | kfree(ac_ptr); |
1065 | } | |
1066 | ||
343e0d7a | 1067 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1068 | struct array_cache *ac, int node) |
e498be7d CL |
1069 | { |
1070 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1071 | ||
1072 | if (ac->avail) { | |
1073 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1074 | /* |
1075 | * Stuff objects into the remote nodes shared array first. | |
1076 | * That way we could avoid the overhead of putting the objects | |
1077 | * into the free lists and getting them back later. | |
1078 | */ | |
693f7d36 JS |
1079 | if (rl3->shared) |
1080 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1081 | |
ff69416e | 1082 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1083 | ac->avail = 0; |
1084 | spin_unlock(&rl3->list_lock); | |
1085 | } | |
1086 | } | |
1087 | ||
8fce4d8e CL |
1088 | /* |
1089 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1090 | */ | |
1091 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1092 | { | |
1093 | int node = __get_cpu_var(reap_node); | |
1094 | ||
1095 | if (l3->alien) { | |
1096 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1097 | |
1098 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1099 | __drain_alien_cache(cachep, ac, node); |
1100 | spin_unlock_irq(&ac->lock); | |
1101 | } | |
1102 | } | |
1103 | } | |
1104 | ||
a737b3e2 AM |
1105 | static void drain_alien_cache(struct kmem_cache *cachep, |
1106 | struct array_cache **alien) | |
e498be7d | 1107 | { |
b28a02de | 1108 | int i = 0; |
e498be7d CL |
1109 | struct array_cache *ac; |
1110 | unsigned long flags; | |
1111 | ||
1112 | for_each_online_node(i) { | |
4484ebf1 | 1113 | ac = alien[i]; |
e498be7d CL |
1114 | if (ac) { |
1115 | spin_lock_irqsave(&ac->lock, flags); | |
1116 | __drain_alien_cache(cachep, ac, i); | |
1117 | spin_unlock_irqrestore(&ac->lock, flags); | |
1118 | } | |
1119 | } | |
1120 | } | |
729bd0b7 | 1121 | |
873623df | 1122 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1123 | { |
1124 | struct slab *slabp = virt_to_slab(objp); | |
1125 | int nodeid = slabp->nodeid; | |
1126 | struct kmem_list3 *l3; | |
1127 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1128 | int node; |
1129 | ||
1130 | node = numa_node_id(); | |
729bd0b7 PE |
1131 | |
1132 | /* | |
1133 | * Make sure we are not freeing a object from another node to the array | |
1134 | * cache on this cpu. | |
1135 | */ | |
62918a03 | 1136 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1137 | return 0; |
1138 | ||
1ca4cb24 | 1139 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1140 | STATS_INC_NODEFREES(cachep); |
1141 | if (l3->alien && l3->alien[nodeid]) { | |
1142 | alien = l3->alien[nodeid]; | |
873623df | 1143 | spin_lock(&alien->lock); |
729bd0b7 PE |
1144 | if (unlikely(alien->avail == alien->limit)) { |
1145 | STATS_INC_ACOVERFLOW(cachep); | |
1146 | __drain_alien_cache(cachep, alien, nodeid); | |
1147 | } | |
1148 | alien->entry[alien->avail++] = objp; | |
1149 | spin_unlock(&alien->lock); | |
1150 | } else { | |
1151 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1152 | free_block(cachep, &objp, 1, nodeid); | |
1153 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1154 | } | |
1155 | return 1; | |
1156 | } | |
e498be7d CL |
1157 | #endif |
1158 | ||
fbf1e473 AM |
1159 | static void __cpuinit cpuup_canceled(long cpu) |
1160 | { | |
1161 | struct kmem_cache *cachep; | |
1162 | struct kmem_list3 *l3 = NULL; | |
1163 | int node = cpu_to_node(cpu); | |
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; | |
1169 | cpumask_t mask; | |
1170 | ||
1171 | mask = node_to_cpumask(node); | |
1172 | /* cpu is dead; no one can alloc from it. */ | |
1173 | nc = cachep->array[cpu]; | |
1174 | cachep->array[cpu] = NULL; | |
1175 | l3 = cachep->nodelists[node]; | |
1176 | ||
1177 | if (!l3) | |
1178 | goto free_array_cache; | |
1179 | ||
1180 | spin_lock_irq(&l3->list_lock); | |
1181 | ||
1182 | /* Free limit for this kmem_list3 */ | |
1183 | l3->free_limit -= cachep->batchcount; | |
1184 | if (nc) | |
1185 | free_block(cachep, nc->entry, nc->avail, node); | |
1186 | ||
1187 | if (!cpus_empty(mask)) { | |
1188 | spin_unlock_irq(&l3->list_lock); | |
1189 | goto free_array_cache; | |
1190 | } | |
1191 | ||
1192 | shared = l3->shared; | |
1193 | if (shared) { | |
1194 | free_block(cachep, shared->entry, | |
1195 | shared->avail, node); | |
1196 | l3->shared = NULL; | |
1197 | } | |
1198 | ||
1199 | alien = l3->alien; | |
1200 | l3->alien = NULL; | |
1201 | ||
1202 | spin_unlock_irq(&l3->list_lock); | |
1203 | ||
1204 | kfree(shared); | |
1205 | if (alien) { | |
1206 | drain_alien_cache(cachep, alien); | |
1207 | free_alien_cache(alien); | |
1208 | } | |
1209 | free_array_cache: | |
1210 | kfree(nc); | |
1211 | } | |
1212 | /* | |
1213 | * In the previous loop, all the objects were freed to | |
1214 | * the respective cache's slabs, now we can go ahead and | |
1215 | * shrink each nodelist to its limit. | |
1216 | */ | |
1217 | list_for_each_entry(cachep, &cache_chain, next) { | |
1218 | l3 = cachep->nodelists[node]; | |
1219 | if (!l3) | |
1220 | continue; | |
1221 | drain_freelist(cachep, l3, l3->free_objects); | |
1222 | } | |
1223 | } | |
1224 | ||
1225 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1226 | { |
343e0d7a | 1227 | struct kmem_cache *cachep; |
e498be7d CL |
1228 | struct kmem_list3 *l3 = NULL; |
1229 | int node = cpu_to_node(cpu); | |
ea02e3dd | 1230 | const int memsize = sizeof(struct kmem_list3); |
1da177e4 | 1231 | |
fbf1e473 AM |
1232 | /* |
1233 | * We need to do this right in the beginning since | |
1234 | * alloc_arraycache's are going to use this list. | |
1235 | * kmalloc_node allows us to add the slab to the right | |
1236 | * kmem_list3 and not this cpu's kmem_list3 | |
1237 | */ | |
1238 | ||
1239 | list_for_each_entry(cachep, &cache_chain, next) { | |
a737b3e2 | 1240 | /* |
fbf1e473 AM |
1241 | * Set up the size64 kmemlist for cpu before we can |
1242 | * begin anything. Make sure some other cpu on this | |
1243 | * node has not already allocated this | |
e498be7d | 1244 | */ |
fbf1e473 AM |
1245 | if (!cachep->nodelists[node]) { |
1246 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1247 | if (!l3) | |
1248 | goto bad; | |
1249 | kmem_list3_init(l3); | |
1250 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1251 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1252 | |
a737b3e2 | 1253 | /* |
fbf1e473 AM |
1254 | * The l3s don't come and go as CPUs come and |
1255 | * go. cache_chain_mutex is sufficient | |
1256 | * protection here. | |
e498be7d | 1257 | */ |
fbf1e473 | 1258 | cachep->nodelists[node] = l3; |
e498be7d CL |
1259 | } |
1260 | ||
fbf1e473 AM |
1261 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1262 | cachep->nodelists[node]->free_limit = | |
1263 | (1 + nr_cpus_node(node)) * | |
1264 | cachep->batchcount + cachep->num; | |
1265 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1266 | } | |
1267 | ||
1268 | /* | |
1269 | * Now we can go ahead with allocating the shared arrays and | |
1270 | * array caches | |
1271 | */ | |
1272 | list_for_each_entry(cachep, &cache_chain, next) { | |
1273 | struct array_cache *nc; | |
1274 | struct array_cache *shared = NULL; | |
1275 | struct array_cache **alien = NULL; | |
1276 | ||
1277 | nc = alloc_arraycache(node, cachep->limit, | |
1278 | cachep->batchcount); | |
1279 | if (!nc) | |
1280 | goto bad; | |
1281 | if (cachep->shared) { | |
1282 | shared = alloc_arraycache(node, | |
1283 | cachep->shared * cachep->batchcount, | |
1284 | 0xbaadf00d); | |
12d00f6a AM |
1285 | if (!shared) { |
1286 | kfree(nc); | |
1da177e4 | 1287 | goto bad; |
12d00f6a | 1288 | } |
fbf1e473 AM |
1289 | } |
1290 | if (use_alien_caches) { | |
1291 | alien = alloc_alien_cache(node, cachep->limit); | |
12d00f6a AM |
1292 | if (!alien) { |
1293 | kfree(shared); | |
1294 | kfree(nc); | |
fbf1e473 | 1295 | goto bad; |
12d00f6a | 1296 | } |
fbf1e473 AM |
1297 | } |
1298 | cachep->array[cpu] = nc; | |
1299 | l3 = cachep->nodelists[node]; | |
1300 | BUG_ON(!l3); | |
1301 | ||
1302 | spin_lock_irq(&l3->list_lock); | |
1303 | if (!l3->shared) { | |
1304 | /* | |
1305 | * We are serialised from CPU_DEAD or | |
1306 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1307 | */ | |
1308 | l3->shared = shared; | |
1309 | shared = NULL; | |
1310 | } | |
4484ebf1 | 1311 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1312 | if (!l3->alien) { |
1313 | l3->alien = alien; | |
1314 | alien = NULL; | |
1da177e4 | 1315 | } |
fbf1e473 AM |
1316 | #endif |
1317 | spin_unlock_irq(&l3->list_lock); | |
1318 | kfree(shared); | |
1319 | free_alien_cache(alien); | |
1320 | } | |
1321 | return 0; | |
1322 | bad: | |
12d00f6a | 1323 | cpuup_canceled(cpu); |
fbf1e473 AM |
1324 | return -ENOMEM; |
1325 | } | |
1326 | ||
1327 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1328 | unsigned long action, void *hcpu) | |
1329 | { | |
1330 | long cpu = (long)hcpu; | |
1331 | int err = 0; | |
1332 | ||
1333 | switch (action) { | |
1334 | case CPU_LOCK_ACQUIRE: | |
1335 | mutex_lock(&cache_chain_mutex); | |
1336 | break; | |
1337 | case CPU_UP_PREPARE: | |
1338 | case CPU_UP_PREPARE_FROZEN: | |
1339 | err = cpuup_prepare(cpu); | |
1da177e4 LT |
1340 | break; |
1341 | case CPU_ONLINE: | |
8bb78442 | 1342 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1343 | start_cpu_timer(cpu); |
1344 | break; | |
1345 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1346 | case CPU_DOWN_PREPARE: |
8bb78442 | 1347 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1348 | /* |
1349 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1350 | * held so that if cache_reap() is invoked it cannot do | |
1351 | * anything expensive but will only modify reap_work | |
1352 | * and reschedule the timer. | |
1353 | */ | |
1354 | cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); | |
1355 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1356 | per_cpu(reap_work, cpu).work.func = NULL; | |
1357 | break; | |
1358 | case CPU_DOWN_FAILED: | |
8bb78442 | 1359 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1360 | start_cpu_timer(cpu); |
1361 | break; | |
1da177e4 | 1362 | case CPU_DEAD: |
8bb78442 | 1363 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1364 | /* |
1365 | * Even if all the cpus of a node are down, we don't free the | |
1366 | * kmem_list3 of any cache. This to avoid a race between | |
1367 | * cpu_down, and a kmalloc allocation from another cpu for | |
1368 | * memory from the node of the cpu going down. The list3 | |
1369 | * structure is usually allocated from kmem_cache_create() and | |
1370 | * gets destroyed at kmem_cache_destroy(). | |
1371 | */ | |
183ff22b | 1372 | /* fall through */ |
8f5be20b | 1373 | #endif |
1da177e4 | 1374 | case CPU_UP_CANCELED: |
8bb78442 | 1375 | case CPU_UP_CANCELED_FROZEN: |
fbf1e473 | 1376 | cpuup_canceled(cpu); |
38c3bd96 HC |
1377 | break; |
1378 | case CPU_LOCK_RELEASE: | |
fc0abb14 | 1379 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1380 | break; |
1da177e4 | 1381 | } |
fbf1e473 | 1382 | return err ? NOTIFY_BAD : NOTIFY_OK; |
1da177e4 LT |
1383 | } |
1384 | ||
74b85f37 CS |
1385 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1386 | &cpuup_callback, NULL, 0 | |
1387 | }; | |
1da177e4 | 1388 | |
e498be7d CL |
1389 | /* |
1390 | * swap the static kmem_list3 with kmalloced memory | |
1391 | */ | |
a737b3e2 AM |
1392 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1393 | int nodeid) | |
e498be7d CL |
1394 | { |
1395 | struct kmem_list3 *ptr; | |
1396 | ||
e498be7d CL |
1397 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); |
1398 | BUG_ON(!ptr); | |
1399 | ||
1400 | local_irq_disable(); | |
1401 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1402 | /* |
1403 | * Do not assume that spinlocks can be initialized via memcpy: | |
1404 | */ | |
1405 | spin_lock_init(&ptr->list_lock); | |
1406 | ||
e498be7d CL |
1407 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1408 | cachep->nodelists[nodeid] = ptr; | |
1409 | local_irq_enable(); | |
1410 | } | |
1411 | ||
a737b3e2 AM |
1412 | /* |
1413 | * Initialisation. Called after the page allocator have been initialised and | |
1414 | * before smp_init(). | |
1da177e4 LT |
1415 | */ |
1416 | void __init kmem_cache_init(void) | |
1417 | { | |
1418 | size_t left_over; | |
1419 | struct cache_sizes *sizes; | |
1420 | struct cache_names *names; | |
e498be7d | 1421 | int i; |
07ed76b2 | 1422 | int order; |
1ca4cb24 | 1423 | int node; |
e498be7d | 1424 | |
1807a1aa | 1425 | if (num_possible_nodes() == 1) { |
62918a03 | 1426 | use_alien_caches = 0; |
1807a1aa SS |
1427 | numa_platform = 0; |
1428 | } | |
62918a03 | 1429 | |
e498be7d CL |
1430 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1431 | kmem_list3_init(&initkmem_list3[i]); | |
1432 | if (i < MAX_NUMNODES) | |
1433 | cache_cache.nodelists[i] = NULL; | |
1434 | } | |
1da177e4 LT |
1435 | |
1436 | /* | |
1437 | * Fragmentation resistance on low memory - only use bigger | |
1438 | * page orders on machines with more than 32MB of memory. | |
1439 | */ | |
1440 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1441 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1442 | ||
1da177e4 LT |
1443 | /* Bootstrap is tricky, because several objects are allocated |
1444 | * from caches that do not exist yet: | |
a737b3e2 AM |
1445 | * 1) initialize the cache_cache cache: it contains the struct |
1446 | * kmem_cache structures of all caches, except cache_cache itself: | |
1447 | * cache_cache is statically allocated. | |
e498be7d CL |
1448 | * Initially an __init data area is used for the head array and the |
1449 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1450 | * array at the end of the bootstrap. | |
1da177e4 | 1451 | * 2) Create the first kmalloc cache. |
343e0d7a | 1452 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1453 | * An __init data area is used for the head array. |
1454 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1455 | * head arrays. | |
1da177e4 LT |
1456 | * 4) Replace the __init data head arrays for cache_cache and the first |
1457 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1458 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1459 | * the other cache's with kmalloc allocated memory. | |
1460 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1461 | */ |
1462 | ||
1ca4cb24 PE |
1463 | node = numa_node_id(); |
1464 | ||
1da177e4 | 1465 | /* 1) create the cache_cache */ |
1da177e4 LT |
1466 | INIT_LIST_HEAD(&cache_chain); |
1467 | list_add(&cache_cache.next, &cache_chain); | |
1468 | cache_cache.colour_off = cache_line_size(); | |
1469 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
1ca4cb24 | 1470 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1471 | |
8da3430d ED |
1472 | /* |
1473 | * struct kmem_cache size depends on nr_node_ids, which | |
1474 | * can be less than MAX_NUMNODES. | |
1475 | */ | |
1476 | cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + | |
1477 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1478 | #if DEBUG | |
1479 | cache_cache.obj_size = cache_cache.buffer_size; | |
1480 | #endif | |
a737b3e2 AM |
1481 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1482 | cache_line_size()); | |
6a2d7a95 ED |
1483 | cache_cache.reciprocal_buffer_size = |
1484 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1485 | |
07ed76b2 JS |
1486 | for (order = 0; order < MAX_ORDER; order++) { |
1487 | cache_estimate(order, cache_cache.buffer_size, | |
1488 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1489 | if (cache_cache.num) | |
1490 | break; | |
1491 | } | |
40094fa6 | 1492 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1493 | cache_cache.gfporder = order; |
b28a02de | 1494 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1495 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1496 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1497 | |
1498 | /* 2+3) create the kmalloc caches */ | |
1499 | sizes = malloc_sizes; | |
1500 | names = cache_names; | |
1501 | ||
a737b3e2 AM |
1502 | /* |
1503 | * Initialize the caches that provide memory for the array cache and the | |
1504 | * kmem_list3 structures first. Without this, further allocations will | |
1505 | * bug. | |
e498be7d CL |
1506 | */ |
1507 | ||
1508 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1509 | sizes[INDEX_AC].cs_size, |
1510 | ARCH_KMALLOC_MINALIGN, | |
1511 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1512 | NULL); |
e498be7d | 1513 | |
a737b3e2 | 1514 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1515 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1516 | kmem_cache_create(names[INDEX_L3].name, |
1517 | sizes[INDEX_L3].cs_size, | |
1518 | ARCH_KMALLOC_MINALIGN, | |
1519 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1520 | NULL); |
a737b3e2 | 1521 | } |
e498be7d | 1522 | |
e0a42726 IM |
1523 | slab_early_init = 0; |
1524 | ||
1da177e4 | 1525 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1526 | /* |
1527 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1528 | * This should be particularly beneficial on SMP boxes, as it |
1529 | * eliminates "false sharing". | |
1530 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1531 | * allow tighter packing of the smaller caches. |
1532 | */ | |
a737b3e2 | 1533 | if (!sizes->cs_cachep) { |
e498be7d | 1534 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1535 | sizes->cs_size, |
1536 | ARCH_KMALLOC_MINALIGN, | |
1537 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1538 | NULL); |
a737b3e2 | 1539 | } |
4b51d669 CL |
1540 | #ifdef CONFIG_ZONE_DMA |
1541 | sizes->cs_dmacachep = kmem_cache_create( | |
1542 | names->name_dma, | |
a737b3e2 AM |
1543 | sizes->cs_size, |
1544 | ARCH_KMALLOC_MINALIGN, | |
1545 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1546 | SLAB_PANIC, | |
20c2df83 | 1547 | NULL); |
4b51d669 | 1548 | #endif |
1da177e4 LT |
1549 | sizes++; |
1550 | names++; | |
1551 | } | |
1552 | /* 4) Replace the bootstrap head arrays */ | |
1553 | { | |
2b2d5493 | 1554 | struct array_cache *ptr; |
e498be7d | 1555 | |
1da177e4 | 1556 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1557 | |
1da177e4 | 1558 | local_irq_disable(); |
9a2dba4b PE |
1559 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1560 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1561 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1562 | /* |
1563 | * Do not assume that spinlocks can be initialized via memcpy: | |
1564 | */ | |
1565 | spin_lock_init(&ptr->lock); | |
1566 | ||
1da177e4 LT |
1567 | cache_cache.array[smp_processor_id()] = ptr; |
1568 | local_irq_enable(); | |
e498be7d | 1569 | |
1da177e4 | 1570 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1571 | |
1da177e4 | 1572 | local_irq_disable(); |
9a2dba4b | 1573 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1574 | != &initarray_generic.cache); |
9a2dba4b | 1575 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1576 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1577 | /* |
1578 | * Do not assume that spinlocks can be initialized via memcpy: | |
1579 | */ | |
1580 | spin_lock_init(&ptr->lock); | |
1581 | ||
e498be7d | 1582 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1583 | ptr; |
1da177e4 LT |
1584 | local_irq_enable(); |
1585 | } | |
e498be7d CL |
1586 | /* 5) Replace the bootstrap kmem_list3's */ |
1587 | { | |
1ca4cb24 PE |
1588 | int nid; |
1589 | ||
e498be7d | 1590 | /* Replace the static kmem_list3 structures for the boot cpu */ |
1ca4cb24 | 1591 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); |
e498be7d | 1592 | |
04231b30 | 1593 | for_each_node_state(nid, N_NORMAL_MEMORY) { |
e498be7d | 1594 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1595 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1596 | |
1597 | if (INDEX_AC != INDEX_L3) { | |
1598 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1599 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1600 | } |
1601 | } | |
1602 | } | |
1da177e4 | 1603 | |
e498be7d | 1604 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1605 | { |
343e0d7a | 1606 | struct kmem_cache *cachep; |
fc0abb14 | 1607 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1608 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1609 | if (enable_cpucache(cachep)) |
1610 | BUG(); | |
fc0abb14 | 1611 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1612 | } |
1613 | ||
056c6241 RT |
1614 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1615 | init_lock_keys(); | |
1616 | ||
1617 | ||
1da177e4 LT |
1618 | /* Done! */ |
1619 | g_cpucache_up = FULL; | |
1620 | ||
a737b3e2 AM |
1621 | /* |
1622 | * Register a cpu startup notifier callback that initializes | |
1623 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1624 | */ |
1625 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1626 | |
a737b3e2 AM |
1627 | /* |
1628 | * The reap timers are started later, with a module init call: That part | |
1629 | * of the kernel is not yet operational. | |
1da177e4 LT |
1630 | */ |
1631 | } | |
1632 | ||
1633 | static int __init cpucache_init(void) | |
1634 | { | |
1635 | int cpu; | |
1636 | ||
a737b3e2 AM |
1637 | /* |
1638 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1639 | */ |
e498be7d | 1640 | for_each_online_cpu(cpu) |
a737b3e2 | 1641 | start_cpu_timer(cpu); |
1da177e4 LT |
1642 | return 0; |
1643 | } | |
1da177e4 LT |
1644 | __initcall(cpucache_init); |
1645 | ||
1646 | /* | |
1647 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1648 | * | |
1649 | * If we requested dmaable memory, we will get it. Even if we | |
1650 | * did not request dmaable memory, we might get it, but that | |
1651 | * would be relatively rare and ignorable. | |
1652 | */ | |
343e0d7a | 1653 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1654 | { |
1655 | struct page *page; | |
e1b6aa6f | 1656 | int nr_pages; |
1da177e4 LT |
1657 | int i; |
1658 | ||
d6fef9da | 1659 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1660 | /* |
1661 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1662 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1663 | */ |
e1b6aa6f | 1664 | flags |= __GFP_COMP; |
d6fef9da | 1665 | #endif |
765c4507 | 1666 | |
3c517a61 | 1667 | flags |= cachep->gfpflags; |
e12ba74d MG |
1668 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1669 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f CH |
1670 | |
1671 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1672 | if (!page) |
1673 | return NULL; | |
1da177e4 | 1674 | |
e1b6aa6f | 1675 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1676 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1677 | add_zone_page_state(page_zone(page), |
1678 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1679 | else | |
1680 | add_zone_page_state(page_zone(page), | |
1681 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1682 | for (i = 0; i < nr_pages; i++) |
1683 | __SetPageSlab(page + i); | |
1684 | return page_address(page); | |
1da177e4 LT |
1685 | } |
1686 | ||
1687 | /* | |
1688 | * Interface to system's page release. | |
1689 | */ | |
343e0d7a | 1690 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1691 | { |
b28a02de | 1692 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1693 | struct page *page = virt_to_page(addr); |
1694 | const unsigned long nr_freed = i; | |
1695 | ||
972d1a7b CL |
1696 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1697 | sub_zone_page_state(page_zone(page), | |
1698 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1699 | else | |
1700 | sub_zone_page_state(page_zone(page), | |
1701 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1702 | while (i--) { |
f205b2fe NP |
1703 | BUG_ON(!PageSlab(page)); |
1704 | __ClearPageSlab(page); | |
1da177e4 LT |
1705 | page++; |
1706 | } | |
1da177e4 LT |
1707 | if (current->reclaim_state) |
1708 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1709 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1710 | } |
1711 | ||
1712 | static void kmem_rcu_free(struct rcu_head *head) | |
1713 | { | |
b28a02de | 1714 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1715 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1716 | |
1717 | kmem_freepages(cachep, slab_rcu->addr); | |
1718 | if (OFF_SLAB(cachep)) | |
1719 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1720 | } | |
1721 | ||
1722 | #if DEBUG | |
1723 | ||
1724 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1725 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1726 | unsigned long caller) |
1da177e4 | 1727 | { |
3dafccf2 | 1728 | int size = obj_size(cachep); |
1da177e4 | 1729 | |
3dafccf2 | 1730 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1731 | |
b28a02de | 1732 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1733 | return; |
1734 | ||
b28a02de PE |
1735 | *addr++ = 0x12345678; |
1736 | *addr++ = caller; | |
1737 | *addr++ = smp_processor_id(); | |
1738 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1739 | { |
1740 | unsigned long *sptr = &caller; | |
1741 | unsigned long svalue; | |
1742 | ||
1743 | while (!kstack_end(sptr)) { | |
1744 | svalue = *sptr++; | |
1745 | if (kernel_text_address(svalue)) { | |
b28a02de | 1746 | *addr++ = svalue; |
1da177e4 LT |
1747 | size -= sizeof(unsigned long); |
1748 | if (size <= sizeof(unsigned long)) | |
1749 | break; | |
1750 | } | |
1751 | } | |
1752 | ||
1753 | } | |
b28a02de | 1754 | *addr++ = 0x87654321; |
1da177e4 LT |
1755 | } |
1756 | #endif | |
1757 | ||
343e0d7a | 1758 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1759 | { |
3dafccf2 MS |
1760 | int size = obj_size(cachep); |
1761 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1762 | |
1763 | memset(addr, val, size); | |
b28a02de | 1764 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1765 | } |
1766 | ||
1767 | static void dump_line(char *data, int offset, int limit) | |
1768 | { | |
1769 | int i; | |
aa83aa40 DJ |
1770 | unsigned char error = 0; |
1771 | int bad_count = 0; | |
1772 | ||
1da177e4 | 1773 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1774 | for (i = 0; i < limit; i++) { |
1775 | if (data[offset + i] != POISON_FREE) { | |
1776 | error = data[offset + i]; | |
1777 | bad_count++; | |
1778 | } | |
b28a02de | 1779 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1780 | } |
1da177e4 | 1781 | printk("\n"); |
aa83aa40 DJ |
1782 | |
1783 | if (bad_count == 1) { | |
1784 | error ^= POISON_FREE; | |
1785 | if (!(error & (error - 1))) { | |
1786 | printk(KERN_ERR "Single bit error detected. Probably " | |
1787 | "bad RAM.\n"); | |
1788 | #ifdef CONFIG_X86 | |
1789 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1790 | "test tool.\n"); | |
1791 | #else | |
1792 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1793 | #endif | |
1794 | } | |
1795 | } | |
1da177e4 LT |
1796 | } |
1797 | #endif | |
1798 | ||
1799 | #if DEBUG | |
1800 | ||
343e0d7a | 1801 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1802 | { |
1803 | int i, size; | |
1804 | char *realobj; | |
1805 | ||
1806 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1807 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1808 | *dbg_redzone1(cachep, objp), |
1809 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1810 | } |
1811 | ||
1812 | if (cachep->flags & SLAB_STORE_USER) { | |
1813 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1814 | *dbg_userword(cachep, objp)); |
1da177e4 | 1815 | print_symbol("(%s)", |
a737b3e2 | 1816 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1817 | printk("\n"); |
1818 | } | |
3dafccf2 MS |
1819 | realobj = (char *)objp + obj_offset(cachep); |
1820 | size = obj_size(cachep); | |
b28a02de | 1821 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1822 | int limit; |
1823 | limit = 16; | |
b28a02de PE |
1824 | if (i + limit > size) |
1825 | limit = size - i; | |
1da177e4 LT |
1826 | dump_line(realobj, i, limit); |
1827 | } | |
1828 | } | |
1829 | ||
343e0d7a | 1830 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1831 | { |
1832 | char *realobj; | |
1833 | int size, i; | |
1834 | int lines = 0; | |
1835 | ||
3dafccf2 MS |
1836 | realobj = (char *)objp + obj_offset(cachep); |
1837 | size = obj_size(cachep); | |
1da177e4 | 1838 | |
b28a02de | 1839 | for (i = 0; i < size; i++) { |
1da177e4 | 1840 | char exp = POISON_FREE; |
b28a02de | 1841 | if (i == size - 1) |
1da177e4 LT |
1842 | exp = POISON_END; |
1843 | if (realobj[i] != exp) { | |
1844 | int limit; | |
1845 | /* Mismatch ! */ | |
1846 | /* Print header */ | |
1847 | if (lines == 0) { | |
b28a02de | 1848 | printk(KERN_ERR |
e94a40c5 DH |
1849 | "Slab corruption: %s start=%p, len=%d\n", |
1850 | cachep->name, realobj, size); | |
1da177e4 LT |
1851 | print_objinfo(cachep, objp, 0); |
1852 | } | |
1853 | /* Hexdump the affected line */ | |
b28a02de | 1854 | i = (i / 16) * 16; |
1da177e4 | 1855 | limit = 16; |
b28a02de PE |
1856 | if (i + limit > size) |
1857 | limit = size - i; | |
1da177e4 LT |
1858 | dump_line(realobj, i, limit); |
1859 | i += 16; | |
1860 | lines++; | |
1861 | /* Limit to 5 lines */ | |
1862 | if (lines > 5) | |
1863 | break; | |
1864 | } | |
1865 | } | |
1866 | if (lines != 0) { | |
1867 | /* Print some data about the neighboring objects, if they | |
1868 | * exist: | |
1869 | */ | |
6ed5eb22 | 1870 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1871 | unsigned int objnr; |
1da177e4 | 1872 | |
8fea4e96 | 1873 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1874 | if (objnr) { |
8fea4e96 | 1875 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1876 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1877 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1878 | realobj, size); |
1da177e4 LT |
1879 | print_objinfo(cachep, objp, 2); |
1880 | } | |
b28a02de | 1881 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1882 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1883 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1884 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1885 | realobj, size); |
1da177e4 LT |
1886 | print_objinfo(cachep, objp, 2); |
1887 | } | |
1888 | } | |
1889 | } | |
1890 | #endif | |
1891 | ||
12dd36fa MD |
1892 | #if DEBUG |
1893 | /** | |
911851e6 RD |
1894 | * slab_destroy_objs - destroy a slab and its objects |
1895 | * @cachep: cache pointer being destroyed | |
1896 | * @slabp: slab pointer being destroyed | |
1897 | * | |
1898 | * Call the registered destructor for each object in a slab that is being | |
1899 | * destroyed. | |
1da177e4 | 1900 | */ |
343e0d7a | 1901 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1902 | { |
1da177e4 LT |
1903 | int i; |
1904 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1905 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1906 | |
1907 | if (cachep->flags & SLAB_POISON) { | |
1908 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1909 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1910 | OFF_SLAB(cachep)) | |
b28a02de | 1911 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1912 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1913 | else |
1914 | check_poison_obj(cachep, objp); | |
1915 | #else | |
1916 | check_poison_obj(cachep, objp); | |
1917 | #endif | |
1918 | } | |
1919 | if (cachep->flags & SLAB_RED_ZONE) { | |
1920 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1921 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1922 | "was overwritten"); |
1da177e4 LT |
1923 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1924 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1925 | "was overwritten"); |
1da177e4 | 1926 | } |
1da177e4 | 1927 | } |
12dd36fa | 1928 | } |
1da177e4 | 1929 | #else |
343e0d7a | 1930 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1931 | { |
12dd36fa | 1932 | } |
1da177e4 LT |
1933 | #endif |
1934 | ||
911851e6 RD |
1935 | /** |
1936 | * slab_destroy - destroy and release all objects in a slab | |
1937 | * @cachep: cache pointer being destroyed | |
1938 | * @slabp: slab pointer being destroyed | |
1939 | * | |
12dd36fa | 1940 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1941 | * Before calling the slab must have been unlinked from the cache. The |
1942 | * cache-lock is not held/needed. | |
12dd36fa | 1943 | */ |
343e0d7a | 1944 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1945 | { |
1946 | void *addr = slabp->s_mem - slabp->colouroff; | |
1947 | ||
1948 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1949 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1950 | struct slab_rcu *slab_rcu; | |
1951 | ||
b28a02de | 1952 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1953 | slab_rcu->cachep = cachep; |
1954 | slab_rcu->addr = addr; | |
1955 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1956 | } else { | |
1957 | kmem_freepages(cachep, addr); | |
873623df IM |
1958 | if (OFF_SLAB(cachep)) |
1959 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1960 | } |
1961 | } | |
1962 | ||
a737b3e2 AM |
1963 | /* |
1964 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1965 | * size of kmem_list3. | |
1966 | */ | |
a3a02be7 | 1967 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1968 | { |
1969 | int node; | |
1970 | ||
04231b30 | 1971 | for_each_node_state(node, N_NORMAL_MEMORY) { |
b28a02de | 1972 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1973 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1974 | REAPTIMEOUT_LIST3 + |
1975 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1976 | } |
1977 | } | |
1978 | ||
117f6eb1 CL |
1979 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1980 | { | |
1981 | int i; | |
1982 | struct kmem_list3 *l3; | |
1983 | ||
1984 | for_each_online_cpu(i) | |
1985 | kfree(cachep->array[i]); | |
1986 | ||
1987 | /* NUMA: free the list3 structures */ | |
1988 | for_each_online_node(i) { | |
1989 | l3 = cachep->nodelists[i]; | |
1990 | if (l3) { | |
1991 | kfree(l3->shared); | |
1992 | free_alien_cache(l3->alien); | |
1993 | kfree(l3); | |
1994 | } | |
1995 | } | |
1996 | kmem_cache_free(&cache_cache, cachep); | |
1997 | } | |
1998 | ||
1999 | ||
4d268eba | 2000 | /** |
a70773dd RD |
2001 | * calculate_slab_order - calculate size (page order) of slabs |
2002 | * @cachep: pointer to the cache that is being created | |
2003 | * @size: size of objects to be created in this cache. | |
2004 | * @align: required alignment for the objects. | |
2005 | * @flags: slab allocation flags | |
2006 | * | |
2007 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2008 | * |
2009 | * This could be made much more intelligent. For now, try to avoid using | |
2010 | * high order pages for slabs. When the gfp() functions are more friendly | |
2011 | * towards high-order requests, this should be changed. | |
2012 | */ | |
a737b3e2 | 2013 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2014 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2015 | { |
b1ab41c4 | 2016 | unsigned long offslab_limit; |
4d268eba | 2017 | size_t left_over = 0; |
9888e6fa | 2018 | int gfporder; |
4d268eba | 2019 | |
0aa817f0 | 2020 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2021 | unsigned int num; |
2022 | size_t remainder; | |
2023 | ||
9888e6fa | 2024 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2025 | if (!num) |
2026 | continue; | |
9888e6fa | 2027 | |
b1ab41c4 IM |
2028 | if (flags & CFLGS_OFF_SLAB) { |
2029 | /* | |
2030 | * Max number of objs-per-slab for caches which | |
2031 | * use off-slab slabs. Needed to avoid a possible | |
2032 | * looping condition in cache_grow(). | |
2033 | */ | |
2034 | offslab_limit = size - sizeof(struct slab); | |
2035 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2036 | ||
2037 | if (num > offslab_limit) | |
2038 | break; | |
2039 | } | |
4d268eba | 2040 | |
9888e6fa | 2041 | /* Found something acceptable - save it away */ |
4d268eba | 2042 | cachep->num = num; |
9888e6fa | 2043 | cachep->gfporder = gfporder; |
4d268eba PE |
2044 | left_over = remainder; |
2045 | ||
f78bb8ad LT |
2046 | /* |
2047 | * A VFS-reclaimable slab tends to have most allocations | |
2048 | * as GFP_NOFS and we really don't want to have to be allocating | |
2049 | * higher-order pages when we are unable to shrink dcache. | |
2050 | */ | |
2051 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2052 | break; | |
2053 | ||
4d268eba PE |
2054 | /* |
2055 | * Large number of objects is good, but very large slabs are | |
2056 | * currently bad for the gfp()s. | |
2057 | */ | |
9888e6fa | 2058 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2059 | break; |
2060 | ||
9888e6fa LT |
2061 | /* |
2062 | * Acceptable internal fragmentation? | |
2063 | */ | |
a737b3e2 | 2064 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2065 | break; |
2066 | } | |
2067 | return left_over; | |
2068 | } | |
2069 | ||
38bdc32a | 2070 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 2071 | { |
2ed3a4ef CL |
2072 | if (g_cpucache_up == FULL) |
2073 | return enable_cpucache(cachep); | |
2074 | ||
f30cf7d1 PE |
2075 | if (g_cpucache_up == NONE) { |
2076 | /* | |
2077 | * Note: the first kmem_cache_create must create the cache | |
2078 | * that's used by kmalloc(24), otherwise the creation of | |
2079 | * further caches will BUG(). | |
2080 | */ | |
2081 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2082 | ||
2083 | /* | |
2084 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2085 | * the first cache, then we need to set up all its list3s, | |
2086 | * otherwise the creation of further caches will BUG(). | |
2087 | */ | |
2088 | set_up_list3s(cachep, SIZE_AC); | |
2089 | if (INDEX_AC == INDEX_L3) | |
2090 | g_cpucache_up = PARTIAL_L3; | |
2091 | else | |
2092 | g_cpucache_up = PARTIAL_AC; | |
2093 | } else { | |
2094 | cachep->array[smp_processor_id()] = | |
2095 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
2096 | ||
2097 | if (g_cpucache_up == PARTIAL_AC) { | |
2098 | set_up_list3s(cachep, SIZE_L3); | |
2099 | g_cpucache_up = PARTIAL_L3; | |
2100 | } else { | |
2101 | int node; | |
04231b30 | 2102 | for_each_node_state(node, N_NORMAL_MEMORY) { |
f30cf7d1 PE |
2103 | cachep->nodelists[node] = |
2104 | kmalloc_node(sizeof(struct kmem_list3), | |
2105 | GFP_KERNEL, node); | |
2106 | BUG_ON(!cachep->nodelists[node]); | |
2107 | kmem_list3_init(cachep->nodelists[node]); | |
2108 | } | |
2109 | } | |
2110 | } | |
2111 | cachep->nodelists[numa_node_id()]->next_reap = | |
2112 | jiffies + REAPTIMEOUT_LIST3 + | |
2113 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2114 | ||
2115 | cpu_cache_get(cachep)->avail = 0; | |
2116 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2117 | cpu_cache_get(cachep)->batchcount = 1; | |
2118 | cpu_cache_get(cachep)->touched = 0; | |
2119 | cachep->batchcount = 1; | |
2120 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2121 | return 0; |
f30cf7d1 PE |
2122 | } |
2123 | ||
1da177e4 LT |
2124 | /** |
2125 | * kmem_cache_create - Create a cache. | |
2126 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2127 | * @size: The size of objects to be created in this cache. | |
2128 | * @align: The required alignment for the objects. | |
2129 | * @flags: SLAB flags | |
2130 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2131 | * |
2132 | * Returns a ptr to the cache on success, NULL on failure. | |
2133 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2134 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2135 | * |
2136 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2137 | * the module calling this has to destroy the cache before getting unloaded. |
2138 | * | |
1da177e4 LT |
2139 | * The flags are |
2140 | * | |
2141 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2142 | * to catch references to uninitialised memory. | |
2143 | * | |
2144 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2145 | * for buffer overruns. | |
2146 | * | |
1da177e4 LT |
2147 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2148 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2149 | * as davem. | |
2150 | */ | |
343e0d7a | 2151 | struct kmem_cache * |
1da177e4 | 2152 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 | 2153 | unsigned long flags, |
4ba9b9d0 | 2154 | void (*ctor)(struct kmem_cache *, void *)) |
1da177e4 LT |
2155 | { |
2156 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2157 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2158 | |
2159 | /* | |
2160 | * Sanity checks... these are all serious usage bugs. | |
2161 | */ | |
a737b3e2 | 2162 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2163 | size > KMALLOC_MAX_SIZE) { |
a737b3e2 AM |
2164 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
2165 | name); | |
b28a02de PE |
2166 | BUG(); |
2167 | } | |
1da177e4 | 2168 | |
f0188f47 | 2169 | /* |
8f5be20b RT |
2170 | * We use cache_chain_mutex to ensure a consistent view of |
2171 | * cpu_online_map as well. Please see cpuup_callback | |
f0188f47 | 2172 | */ |
fc0abb14 | 2173 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2174 | |
7a7c381d | 2175 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2176 | char tmp; |
2177 | int res; | |
2178 | ||
2179 | /* | |
2180 | * This happens when the module gets unloaded and doesn't | |
2181 | * destroy its slab cache and no-one else reuses the vmalloc | |
2182 | * area of the module. Print a warning. | |
2183 | */ | |
138ae663 | 2184 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2185 | if (res) { |
b4169525 | 2186 | printk(KERN_ERR |
2187 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2188 | pc->buffer_size); |
4f12bb4f AM |
2189 | continue; |
2190 | } | |
2191 | ||
b28a02de | 2192 | if (!strcmp(pc->name, name)) { |
b4169525 | 2193 | printk(KERN_ERR |
2194 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2195 | dump_stack(); |
2196 | goto oops; | |
2197 | } | |
2198 | } | |
2199 | ||
1da177e4 LT |
2200 | #if DEBUG |
2201 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2202 | #if FORCED_DEBUG |
2203 | /* | |
2204 | * Enable redzoning and last user accounting, except for caches with | |
2205 | * large objects, if the increased size would increase the object size | |
2206 | * above the next power of two: caches with object sizes just above a | |
2207 | * power of two have a significant amount of internal fragmentation. | |
2208 | */ | |
87a927c7 DW |
2209 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2210 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2211 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2212 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2213 | flags |= SLAB_POISON; | |
2214 | #endif | |
2215 | if (flags & SLAB_DESTROY_BY_RCU) | |
2216 | BUG_ON(flags & SLAB_POISON); | |
2217 | #endif | |
1da177e4 | 2218 | /* |
a737b3e2 AM |
2219 | * Always checks flags, a caller might be expecting debug support which |
2220 | * isn't available. | |
1da177e4 | 2221 | */ |
40094fa6 | 2222 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2223 | |
a737b3e2 AM |
2224 | /* |
2225 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2226 | * unaligned accesses for some archs when redzoning is used, and makes |
2227 | * sure any on-slab bufctl's are also correctly aligned. | |
2228 | */ | |
b28a02de PE |
2229 | if (size & (BYTES_PER_WORD - 1)) { |
2230 | size += (BYTES_PER_WORD - 1); | |
2231 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2232 | } |
2233 | ||
a737b3e2 AM |
2234 | /* calculate the final buffer alignment: */ |
2235 | ||
1da177e4 LT |
2236 | /* 1) arch recommendation: can be overridden for debug */ |
2237 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2238 | /* |
2239 | * Default alignment: as specified by the arch code. Except if | |
2240 | * an object is really small, then squeeze multiple objects into | |
2241 | * one cacheline. | |
1da177e4 LT |
2242 | */ |
2243 | ralign = cache_line_size(); | |
b28a02de | 2244 | while (size <= ralign / 2) |
1da177e4 LT |
2245 | ralign /= 2; |
2246 | } else { | |
2247 | ralign = BYTES_PER_WORD; | |
2248 | } | |
ca5f9703 PE |
2249 | |
2250 | /* | |
87a927c7 DW |
2251 | * Redzoning and user store require word alignment or possibly larger. |
2252 | * Note this will be overridden by architecture or caller mandated | |
2253 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2254 | */ |
87a927c7 DW |
2255 | if (flags & SLAB_STORE_USER) |
2256 | ralign = BYTES_PER_WORD; | |
2257 | ||
2258 | if (flags & SLAB_RED_ZONE) { | |
2259 | ralign = REDZONE_ALIGN; | |
2260 | /* If redzoning, ensure that the second redzone is suitably | |
2261 | * aligned, by adjusting the object size accordingly. */ | |
2262 | size += REDZONE_ALIGN - 1; | |
2263 | size &= ~(REDZONE_ALIGN - 1); | |
2264 | } | |
ca5f9703 | 2265 | |
a44b56d3 | 2266 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2267 | if (ralign < ARCH_SLAB_MINALIGN) { |
2268 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2269 | } |
a44b56d3 | 2270 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2271 | if (ralign < align) { |
2272 | ralign = align; | |
1da177e4 | 2273 | } |
a44b56d3 | 2274 | /* disable debug if necessary */ |
b46b8f19 | 2275 | if (ralign > __alignof__(unsigned long long)) |
a44b56d3 | 2276 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2277 | /* |
ca5f9703 | 2278 | * 4) Store it. |
1da177e4 LT |
2279 | */ |
2280 | align = ralign; | |
2281 | ||
2282 | /* Get cache's description obj. */ | |
e94b1766 | 2283 | cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); |
1da177e4 | 2284 | if (!cachep) |
4f12bb4f | 2285 | goto oops; |
1da177e4 LT |
2286 | |
2287 | #if DEBUG | |
3dafccf2 | 2288 | cachep->obj_size = size; |
1da177e4 | 2289 | |
ca5f9703 PE |
2290 | /* |
2291 | * Both debugging options require word-alignment which is calculated | |
2292 | * into align above. | |
2293 | */ | |
1da177e4 | 2294 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2295 | /* add space for red zone words */ |
b46b8f19 DW |
2296 | cachep->obj_offset += sizeof(unsigned long long); |
2297 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2298 | } |
2299 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2300 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2301 | * the real object. But if the second red zone needs to be |
2302 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2303 | */ |
87a927c7 DW |
2304 | if (flags & SLAB_RED_ZONE) |
2305 | size += REDZONE_ALIGN; | |
2306 | else | |
2307 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2308 | } |
2309 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2310 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2311 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2312 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2313 | size = PAGE_SIZE; |
2314 | } | |
2315 | #endif | |
2316 | #endif | |
2317 | ||
e0a42726 IM |
2318 | /* |
2319 | * Determine if the slab management is 'on' or 'off' slab. | |
2320 | * (bootstrapping cannot cope with offslab caches so don't do | |
2321 | * it too early on.) | |
2322 | */ | |
2323 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2324 | /* |
2325 | * Size is large, assume best to place the slab management obj | |
2326 | * off-slab (should allow better packing of objs). | |
2327 | */ | |
2328 | flags |= CFLGS_OFF_SLAB; | |
2329 | ||
2330 | size = ALIGN(size, align); | |
2331 | ||
f78bb8ad | 2332 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2333 | |
2334 | if (!cachep->num) { | |
b4169525 | 2335 | printk(KERN_ERR |
2336 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2337 | kmem_cache_free(&cache_cache, cachep); |
2338 | cachep = NULL; | |
4f12bb4f | 2339 | goto oops; |
1da177e4 | 2340 | } |
b28a02de PE |
2341 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2342 | + sizeof(struct slab), align); | |
1da177e4 LT |
2343 | |
2344 | /* | |
2345 | * If the slab has been placed off-slab, and we have enough space then | |
2346 | * move it on-slab. This is at the expense of any extra colouring. | |
2347 | */ | |
2348 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2349 | flags &= ~CFLGS_OFF_SLAB; | |
2350 | left_over -= slab_size; | |
2351 | } | |
2352 | ||
2353 | if (flags & CFLGS_OFF_SLAB) { | |
2354 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2355 | slab_size = |
2356 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2357 | } |
2358 | ||
2359 | cachep->colour_off = cache_line_size(); | |
2360 | /* Offset must be a multiple of the alignment. */ | |
2361 | if (cachep->colour_off < align) | |
2362 | cachep->colour_off = align; | |
b28a02de | 2363 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2364 | cachep->slab_size = slab_size; |
2365 | cachep->flags = flags; | |
2366 | cachep->gfpflags = 0; | |
4b51d669 | 2367 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2368 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2369 | cachep->buffer_size = size; |
6a2d7a95 | 2370 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2371 | |
e5ac9c5a | 2372 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2373 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2374 | /* |
2375 | * This is a possibility for one of the malloc_sizes caches. | |
2376 | * But since we go off slab only for object size greater than | |
2377 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2378 | * this should not happen at all. | |
2379 | * But leave a BUG_ON for some lucky dude. | |
2380 | */ | |
6cb8f913 | 2381 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2382 | } |
1da177e4 | 2383 | cachep->ctor = ctor; |
1da177e4 LT |
2384 | cachep->name = name; |
2385 | ||
2ed3a4ef CL |
2386 | if (setup_cpu_cache(cachep)) { |
2387 | __kmem_cache_destroy(cachep); | |
2388 | cachep = NULL; | |
2389 | goto oops; | |
2390 | } | |
1da177e4 | 2391 | |
1da177e4 LT |
2392 | /* cache setup completed, link it into the list */ |
2393 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2394 | oops: |
1da177e4 LT |
2395 | if (!cachep && (flags & SLAB_PANIC)) |
2396 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2397 | name); |
fc0abb14 | 2398 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2399 | return cachep; |
2400 | } | |
2401 | EXPORT_SYMBOL(kmem_cache_create); | |
2402 | ||
2403 | #if DEBUG | |
2404 | static void check_irq_off(void) | |
2405 | { | |
2406 | BUG_ON(!irqs_disabled()); | |
2407 | } | |
2408 | ||
2409 | static void check_irq_on(void) | |
2410 | { | |
2411 | BUG_ON(irqs_disabled()); | |
2412 | } | |
2413 | ||
343e0d7a | 2414 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2415 | { |
2416 | #ifdef CONFIG_SMP | |
2417 | check_irq_off(); | |
e498be7d | 2418 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2419 | #endif |
2420 | } | |
e498be7d | 2421 | |
343e0d7a | 2422 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2423 | { |
2424 | #ifdef CONFIG_SMP | |
2425 | check_irq_off(); | |
2426 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2427 | #endif | |
2428 | } | |
2429 | ||
1da177e4 LT |
2430 | #else |
2431 | #define check_irq_off() do { } while(0) | |
2432 | #define check_irq_on() do { } while(0) | |
2433 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2434 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2435 | #endif |
2436 | ||
aab2207c CL |
2437 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2438 | struct array_cache *ac, | |
2439 | int force, int node); | |
2440 | ||
1da177e4 LT |
2441 | static void do_drain(void *arg) |
2442 | { | |
a737b3e2 | 2443 | struct kmem_cache *cachep = arg; |
1da177e4 | 2444 | struct array_cache *ac; |
ff69416e | 2445 | int node = numa_node_id(); |
1da177e4 LT |
2446 | |
2447 | check_irq_off(); | |
9a2dba4b | 2448 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2449 | spin_lock(&cachep->nodelists[node]->list_lock); |
2450 | free_block(cachep, ac->entry, ac->avail, node); | |
2451 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2452 | ac->avail = 0; |
2453 | } | |
2454 | ||
343e0d7a | 2455 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2456 | { |
e498be7d CL |
2457 | struct kmem_list3 *l3; |
2458 | int node; | |
2459 | ||
a07fa394 | 2460 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2461 | check_irq_on(); |
b28a02de | 2462 | for_each_online_node(node) { |
e498be7d | 2463 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2464 | if (l3 && l3->alien) |
2465 | drain_alien_cache(cachep, l3->alien); | |
2466 | } | |
2467 | ||
2468 | for_each_online_node(node) { | |
2469 | l3 = cachep->nodelists[node]; | |
2470 | if (l3) | |
aab2207c | 2471 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2472 | } |
1da177e4 LT |
2473 | } |
2474 | ||
ed11d9eb CL |
2475 | /* |
2476 | * Remove slabs from the list of free slabs. | |
2477 | * Specify the number of slabs to drain in tofree. | |
2478 | * | |
2479 | * Returns the actual number of slabs released. | |
2480 | */ | |
2481 | static int drain_freelist(struct kmem_cache *cache, | |
2482 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2483 | { |
ed11d9eb CL |
2484 | struct list_head *p; |
2485 | int nr_freed; | |
1da177e4 | 2486 | struct slab *slabp; |
1da177e4 | 2487 | |
ed11d9eb CL |
2488 | nr_freed = 0; |
2489 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2490 | |
ed11d9eb | 2491 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2492 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2493 | if (p == &l3->slabs_free) { |
2494 | spin_unlock_irq(&l3->list_lock); | |
2495 | goto out; | |
2496 | } | |
1da177e4 | 2497 | |
ed11d9eb | 2498 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2499 | #if DEBUG |
40094fa6 | 2500 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2501 | #endif |
2502 | list_del(&slabp->list); | |
ed11d9eb CL |
2503 | /* |
2504 | * Safe to drop the lock. The slab is no longer linked | |
2505 | * to the cache. | |
2506 | */ | |
2507 | l3->free_objects -= cache->num; | |
e498be7d | 2508 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2509 | slab_destroy(cache, slabp); |
2510 | nr_freed++; | |
1da177e4 | 2511 | } |
ed11d9eb CL |
2512 | out: |
2513 | return nr_freed; | |
1da177e4 LT |
2514 | } |
2515 | ||
8f5be20b | 2516 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2517 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2518 | { |
2519 | int ret = 0, i = 0; | |
2520 | struct kmem_list3 *l3; | |
2521 | ||
2522 | drain_cpu_caches(cachep); | |
2523 | ||
2524 | check_irq_on(); | |
2525 | for_each_online_node(i) { | |
2526 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2527 | if (!l3) |
2528 | continue; | |
2529 | ||
2530 | drain_freelist(cachep, l3, l3->free_objects); | |
2531 | ||
2532 | ret += !list_empty(&l3->slabs_full) || | |
2533 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2534 | } |
2535 | return (ret ? 1 : 0); | |
2536 | } | |
2537 | ||
1da177e4 LT |
2538 | /** |
2539 | * kmem_cache_shrink - Shrink a cache. | |
2540 | * @cachep: The cache to shrink. | |
2541 | * | |
2542 | * Releases as many slabs as possible for a cache. | |
2543 | * To help debugging, a zero exit status indicates all slabs were released. | |
2544 | */ | |
343e0d7a | 2545 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2546 | { |
8f5be20b | 2547 | int ret; |
40094fa6 | 2548 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2549 | |
8f5be20b RT |
2550 | mutex_lock(&cache_chain_mutex); |
2551 | ret = __cache_shrink(cachep); | |
2552 | mutex_unlock(&cache_chain_mutex); | |
2553 | return ret; | |
1da177e4 LT |
2554 | } |
2555 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2556 | ||
2557 | /** | |
2558 | * kmem_cache_destroy - delete a cache | |
2559 | * @cachep: the cache to destroy | |
2560 | * | |
72fd4a35 | 2561 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2562 | * |
2563 | * It is expected this function will be called by a module when it is | |
2564 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2565 | * cache being allocated each time a module is loaded and unloaded, if the | |
2566 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2567 | * | |
2568 | * The cache must be empty before calling this function. | |
2569 | * | |
2570 | * The caller must guarantee that noone will allocate memory from the cache | |
2571 | * during the kmem_cache_destroy(). | |
2572 | */ | |
133d205a | 2573 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2574 | { |
40094fa6 | 2575 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2576 | |
1da177e4 | 2577 | /* Find the cache in the chain of caches. */ |
fc0abb14 | 2578 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2579 | /* |
2580 | * the chain is never empty, cache_cache is never destroyed | |
2581 | */ | |
2582 | list_del(&cachep->next); | |
1da177e4 LT |
2583 | if (__cache_shrink(cachep)) { |
2584 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2585 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2586 | mutex_unlock(&cache_chain_mutex); |
133d205a | 2587 | return; |
1da177e4 LT |
2588 | } |
2589 | ||
2590 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2591 | synchronize_rcu(); |
1da177e4 | 2592 | |
117f6eb1 | 2593 | __kmem_cache_destroy(cachep); |
8f5be20b | 2594 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2595 | } |
2596 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2597 | ||
e5ac9c5a RT |
2598 | /* |
2599 | * Get the memory for a slab management obj. | |
2600 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2601 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2602 | * come from the same cache which is getting created because, | |
2603 | * when we are searching for an appropriate cache for these | |
2604 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2605 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2606 | * kmem_find_general_cachep till the initialization is complete. | |
2607 | * Hence we cannot have slabp_cache same as the original cache. | |
2608 | */ | |
343e0d7a | 2609 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2610 | int colour_off, gfp_t local_flags, |
2611 | int nodeid) | |
1da177e4 LT |
2612 | { |
2613 | struct slab *slabp; | |
b28a02de | 2614 | |
1da177e4 LT |
2615 | if (OFF_SLAB(cachep)) { |
2616 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2617 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
3c517a61 | 2618 | local_flags & ~GFP_THISNODE, nodeid); |
1da177e4 LT |
2619 | if (!slabp) |
2620 | return NULL; | |
2621 | } else { | |
b28a02de | 2622 | slabp = objp + colour_off; |
1da177e4 LT |
2623 | colour_off += cachep->slab_size; |
2624 | } | |
2625 | slabp->inuse = 0; | |
2626 | slabp->colouroff = colour_off; | |
b28a02de | 2627 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2628 | slabp->nodeid = nodeid; |
1da177e4 LT |
2629 | return slabp; |
2630 | } | |
2631 | ||
2632 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2633 | { | |
b28a02de | 2634 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2635 | } |
2636 | ||
343e0d7a | 2637 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2638 | struct slab *slabp) |
1da177e4 LT |
2639 | { |
2640 | int i; | |
2641 | ||
2642 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2643 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2644 | #if DEBUG |
2645 | /* need to poison the objs? */ | |
2646 | if (cachep->flags & SLAB_POISON) | |
2647 | poison_obj(cachep, objp, POISON_FREE); | |
2648 | if (cachep->flags & SLAB_STORE_USER) | |
2649 | *dbg_userword(cachep, objp) = NULL; | |
2650 | ||
2651 | if (cachep->flags & SLAB_RED_ZONE) { | |
2652 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2653 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2654 | } | |
2655 | /* | |
a737b3e2 AM |
2656 | * Constructors are not allowed to allocate memory from the same |
2657 | * cache which they are a constructor for. Otherwise, deadlock. | |
2658 | * They must also be threaded. | |
1da177e4 LT |
2659 | */ |
2660 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
4ba9b9d0 | 2661 | cachep->ctor(cachep, objp + obj_offset(cachep)); |
1da177e4 LT |
2662 | |
2663 | if (cachep->flags & SLAB_RED_ZONE) { | |
2664 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2665 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2666 | " end of an object"); |
1da177e4 LT |
2667 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2668 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2669 | " start of an object"); |
1da177e4 | 2670 | } |
a737b3e2 AM |
2671 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2672 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2673 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2674 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2675 | #else |
2676 | if (cachep->ctor) | |
4ba9b9d0 | 2677 | cachep->ctor(cachep, objp); |
1da177e4 | 2678 | #endif |
b28a02de | 2679 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2680 | } |
b28a02de | 2681 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2682 | slabp->free = 0; |
2683 | } | |
2684 | ||
343e0d7a | 2685 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2686 | { |
4b51d669 CL |
2687 | if (CONFIG_ZONE_DMA_FLAG) { |
2688 | if (flags & GFP_DMA) | |
2689 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2690 | else | |
2691 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2692 | } | |
1da177e4 LT |
2693 | } |
2694 | ||
a737b3e2 AM |
2695 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2696 | int nodeid) | |
78d382d7 | 2697 | { |
8fea4e96 | 2698 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2699 | kmem_bufctl_t next; |
2700 | ||
2701 | slabp->inuse++; | |
2702 | next = slab_bufctl(slabp)[slabp->free]; | |
2703 | #if DEBUG | |
2704 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2705 | WARN_ON(slabp->nodeid != nodeid); | |
2706 | #endif | |
2707 | slabp->free = next; | |
2708 | ||
2709 | return objp; | |
2710 | } | |
2711 | ||
a737b3e2 AM |
2712 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2713 | void *objp, int nodeid) | |
78d382d7 | 2714 | { |
8fea4e96 | 2715 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2716 | |
2717 | #if DEBUG | |
2718 | /* Verify that the slab belongs to the intended node */ | |
2719 | WARN_ON(slabp->nodeid != nodeid); | |
2720 | ||
871751e2 | 2721 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2722 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2723 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2724 | BUG(); |
2725 | } | |
2726 | #endif | |
2727 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2728 | slabp->free = objnr; | |
2729 | slabp->inuse--; | |
2730 | } | |
2731 | ||
4776874f PE |
2732 | /* |
2733 | * Map pages beginning at addr to the given cache and slab. This is required | |
2734 | * for the slab allocator to be able to lookup the cache and slab of a | |
2735 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2736 | */ | |
2737 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2738 | void *addr) | |
1da177e4 | 2739 | { |
4776874f | 2740 | int nr_pages; |
1da177e4 LT |
2741 | struct page *page; |
2742 | ||
4776874f | 2743 | page = virt_to_page(addr); |
84097518 | 2744 | |
4776874f | 2745 | nr_pages = 1; |
84097518 | 2746 | if (likely(!PageCompound(page))) |
4776874f PE |
2747 | nr_pages <<= cache->gfporder; |
2748 | ||
1da177e4 | 2749 | do { |
4776874f PE |
2750 | page_set_cache(page, cache); |
2751 | page_set_slab(page, slab); | |
1da177e4 | 2752 | page++; |
4776874f | 2753 | } while (--nr_pages); |
1da177e4 LT |
2754 | } |
2755 | ||
2756 | /* | |
2757 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2758 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2759 | */ | |
3c517a61 CL |
2760 | static int cache_grow(struct kmem_cache *cachep, |
2761 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2762 | { |
b28a02de | 2763 | struct slab *slabp; |
b28a02de PE |
2764 | size_t offset; |
2765 | gfp_t local_flags; | |
e498be7d | 2766 | struct kmem_list3 *l3; |
1da177e4 | 2767 | |
a737b3e2 AM |
2768 | /* |
2769 | * Be lazy and only check for valid flags here, keeping it out of the | |
2770 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2771 | */ |
6cb06229 CL |
2772 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2773 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2774 | |
2e1217cf | 2775 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2776 | check_irq_off(); |
2e1217cf RT |
2777 | l3 = cachep->nodelists[nodeid]; |
2778 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2779 | |
2780 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2781 | offset = l3->colour_next; |
2782 | l3->colour_next++; | |
2783 | if (l3->colour_next >= cachep->colour) | |
2784 | l3->colour_next = 0; | |
2785 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2786 | |
2e1217cf | 2787 | offset *= cachep->colour_off; |
1da177e4 LT |
2788 | |
2789 | if (local_flags & __GFP_WAIT) | |
2790 | local_irq_enable(); | |
2791 | ||
2792 | /* | |
2793 | * The test for missing atomic flag is performed here, rather than | |
2794 | * the more obvious place, simply to reduce the critical path length | |
2795 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2796 | * will eventually be caught here (where it matters). | |
2797 | */ | |
2798 | kmem_flagcheck(cachep, flags); | |
2799 | ||
a737b3e2 AM |
2800 | /* |
2801 | * Get mem for the objs. Attempt to allocate a physical page from | |
2802 | * 'nodeid'. | |
e498be7d | 2803 | */ |
3c517a61 | 2804 | if (!objp) |
b8c1c5da | 2805 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2806 | if (!objp) |
1da177e4 LT |
2807 | goto failed; |
2808 | ||
2809 | /* Get slab management. */ | |
3c517a61 | 2810 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2811 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2812 | if (!slabp) |
1da177e4 LT |
2813 | goto opps1; |
2814 | ||
e498be7d | 2815 | slabp->nodeid = nodeid; |
4776874f | 2816 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2817 | |
a35afb83 | 2818 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2819 | |
2820 | if (local_flags & __GFP_WAIT) | |
2821 | local_irq_disable(); | |
2822 | check_irq_off(); | |
e498be7d | 2823 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2824 | |
2825 | /* Make slab active. */ | |
e498be7d | 2826 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2827 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2828 | l3->free_objects += cachep->num; |
2829 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2830 | return 1; |
a737b3e2 | 2831 | opps1: |
1da177e4 | 2832 | kmem_freepages(cachep, objp); |
a737b3e2 | 2833 | failed: |
1da177e4 LT |
2834 | if (local_flags & __GFP_WAIT) |
2835 | local_irq_disable(); | |
2836 | return 0; | |
2837 | } | |
2838 | ||
2839 | #if DEBUG | |
2840 | ||
2841 | /* | |
2842 | * Perform extra freeing checks: | |
2843 | * - detect bad pointers. | |
2844 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2845 | */ |
2846 | static void kfree_debugcheck(const void *objp) | |
2847 | { | |
1da177e4 LT |
2848 | if (!virt_addr_valid(objp)) { |
2849 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2850 | (unsigned long)objp); |
2851 | BUG(); | |
1da177e4 | 2852 | } |
1da177e4 LT |
2853 | } |
2854 | ||
58ce1fd5 PE |
2855 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2856 | { | |
b46b8f19 | 2857 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2858 | |
2859 | redzone1 = *dbg_redzone1(cache, obj); | |
2860 | redzone2 = *dbg_redzone2(cache, obj); | |
2861 | ||
2862 | /* | |
2863 | * Redzone is ok. | |
2864 | */ | |
2865 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2866 | return; | |
2867 | ||
2868 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2869 | slab_error(cache, "double free detected"); | |
2870 | else | |
2871 | slab_error(cache, "memory outside object was overwritten"); | |
2872 | ||
b46b8f19 | 2873 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2874 | obj, redzone1, redzone2); |
2875 | } | |
2876 | ||
343e0d7a | 2877 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2878 | void *caller) |
1da177e4 LT |
2879 | { |
2880 | struct page *page; | |
2881 | unsigned int objnr; | |
2882 | struct slab *slabp; | |
2883 | ||
3dafccf2 | 2884 | objp -= obj_offset(cachep); |
1da177e4 | 2885 | kfree_debugcheck(objp); |
b49af68f | 2886 | page = virt_to_head_page(objp); |
1da177e4 | 2887 | |
065d41cb | 2888 | slabp = page_get_slab(page); |
1da177e4 LT |
2889 | |
2890 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2891 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2892 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2893 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2894 | } | |
2895 | if (cachep->flags & SLAB_STORE_USER) | |
2896 | *dbg_userword(cachep, objp) = caller; | |
2897 | ||
8fea4e96 | 2898 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2899 | |
2900 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2901 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2902 | |
871751e2 AV |
2903 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2904 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2905 | #endif | |
1da177e4 LT |
2906 | if (cachep->flags & SLAB_POISON) { |
2907 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2908 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2909 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2910 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2911 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2912 | } else { |
2913 | poison_obj(cachep, objp, POISON_FREE); | |
2914 | } | |
2915 | #else | |
2916 | poison_obj(cachep, objp, POISON_FREE); | |
2917 | #endif | |
2918 | } | |
2919 | return objp; | |
2920 | } | |
2921 | ||
343e0d7a | 2922 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2923 | { |
2924 | kmem_bufctl_t i; | |
2925 | int entries = 0; | |
b28a02de | 2926 | |
1da177e4 LT |
2927 | /* Check slab's freelist to see if this obj is there. */ |
2928 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2929 | entries++; | |
2930 | if (entries > cachep->num || i >= cachep->num) | |
2931 | goto bad; | |
2932 | } | |
2933 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2934 | bad: |
2935 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2936 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2937 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2938 | for (i = 0; |
264132bc | 2939 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2940 | i++) { |
a737b3e2 | 2941 | if (i % 16 == 0) |
1da177e4 | 2942 | printk("\n%03x:", i); |
b28a02de | 2943 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2944 | } |
2945 | printk("\n"); | |
2946 | BUG(); | |
2947 | } | |
2948 | } | |
2949 | #else | |
2950 | #define kfree_debugcheck(x) do { } while(0) | |
2951 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2952 | #define check_slabp(x,y) do { } while(0) | |
2953 | #endif | |
2954 | ||
343e0d7a | 2955 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2956 | { |
2957 | int batchcount; | |
2958 | struct kmem_list3 *l3; | |
2959 | struct array_cache *ac; | |
1ca4cb24 PE |
2960 | int node; |
2961 | ||
2962 | node = numa_node_id(); | |
1da177e4 LT |
2963 | |
2964 | check_irq_off(); | |
9a2dba4b | 2965 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2966 | retry: |
1da177e4 LT |
2967 | batchcount = ac->batchcount; |
2968 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2969 | /* |
2970 | * If there was little recent activity on this cache, then | |
2971 | * perform only a partial refill. Otherwise we could generate | |
2972 | * refill bouncing. | |
1da177e4 LT |
2973 | */ |
2974 | batchcount = BATCHREFILL_LIMIT; | |
2975 | } | |
1ca4cb24 | 2976 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2977 | |
2978 | BUG_ON(ac->avail > 0 || !l3); | |
2979 | spin_lock(&l3->list_lock); | |
1da177e4 | 2980 | |
3ded175a CL |
2981 | /* See if we can refill from the shared array */ |
2982 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2983 | goto alloc_done; | |
2984 | ||
1da177e4 LT |
2985 | while (batchcount > 0) { |
2986 | struct list_head *entry; | |
2987 | struct slab *slabp; | |
2988 | /* Get slab alloc is to come from. */ | |
2989 | entry = l3->slabs_partial.next; | |
2990 | if (entry == &l3->slabs_partial) { | |
2991 | l3->free_touched = 1; | |
2992 | entry = l3->slabs_free.next; | |
2993 | if (entry == &l3->slabs_free) | |
2994 | goto must_grow; | |
2995 | } | |
2996 | ||
2997 | slabp = list_entry(entry, struct slab, list); | |
2998 | check_slabp(cachep, slabp); | |
2999 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3000 | |
3001 | /* | |
3002 | * The slab was either on partial or free list so | |
3003 | * there must be at least one object available for | |
3004 | * allocation. | |
3005 | */ | |
3006 | BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num); | |
3007 | ||
1da177e4 | 3008 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3009 | STATS_INC_ALLOCED(cachep); |
3010 | STATS_INC_ACTIVE(cachep); | |
3011 | STATS_SET_HIGH(cachep); | |
3012 | ||
78d382d7 | 3013 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3014 | node); |
1da177e4 LT |
3015 | } |
3016 | check_slabp(cachep, slabp); | |
3017 | ||
3018 | /* move slabp to correct slabp list: */ | |
3019 | list_del(&slabp->list); | |
3020 | if (slabp->free == BUFCTL_END) | |
3021 | list_add(&slabp->list, &l3->slabs_full); | |
3022 | else | |
3023 | list_add(&slabp->list, &l3->slabs_partial); | |
3024 | } | |
3025 | ||
a737b3e2 | 3026 | must_grow: |
1da177e4 | 3027 | l3->free_objects -= ac->avail; |
a737b3e2 | 3028 | alloc_done: |
e498be7d | 3029 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3030 | |
3031 | if (unlikely(!ac->avail)) { | |
3032 | int x; | |
3c517a61 | 3033 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3034 | |
a737b3e2 | 3035 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3036 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3037 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3038 | return NULL; |
3039 | ||
a737b3e2 | 3040 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3041 | goto retry; |
3042 | } | |
3043 | ac->touched = 1; | |
e498be7d | 3044 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3045 | } |
3046 | ||
a737b3e2 AM |
3047 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3048 | gfp_t flags) | |
1da177e4 LT |
3049 | { |
3050 | might_sleep_if(flags & __GFP_WAIT); | |
3051 | #if DEBUG | |
3052 | kmem_flagcheck(cachep, flags); | |
3053 | #endif | |
3054 | } | |
3055 | ||
3056 | #if DEBUG | |
a737b3e2 AM |
3057 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3058 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3059 | { |
b28a02de | 3060 | if (!objp) |
1da177e4 | 3061 | return objp; |
b28a02de | 3062 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3063 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3064 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3065 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3066 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3067 | else |
3068 | check_poison_obj(cachep, objp); | |
3069 | #else | |
3070 | check_poison_obj(cachep, objp); | |
3071 | #endif | |
3072 | poison_obj(cachep, objp, POISON_INUSE); | |
3073 | } | |
3074 | if (cachep->flags & SLAB_STORE_USER) | |
3075 | *dbg_userword(cachep, objp) = caller; | |
3076 | ||
3077 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3078 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3079 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3080 | slab_error(cachep, "double free, or memory outside" | |
3081 | " object was overwritten"); | |
b28a02de | 3082 | printk(KERN_ERR |
b46b8f19 | 3083 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3084 | objp, *dbg_redzone1(cachep, objp), |
3085 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3086 | } |
3087 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3088 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3089 | } | |
871751e2 AV |
3090 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3091 | { | |
3092 | struct slab *slabp; | |
3093 | unsigned objnr; | |
3094 | ||
b49af68f | 3095 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3096 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3097 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3098 | } | |
3099 | #endif | |
3dafccf2 | 3100 | objp += obj_offset(cachep); |
4f104934 | 3101 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
4ba9b9d0 | 3102 | cachep->ctor(cachep, objp); |
a44b56d3 KH |
3103 | #if ARCH_SLAB_MINALIGN |
3104 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3105 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3106 | objp, ARCH_SLAB_MINALIGN); | |
3107 | } | |
3108 | #endif | |
1da177e4 LT |
3109 | return objp; |
3110 | } | |
3111 | #else | |
3112 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3113 | #endif | |
3114 | ||
8a8b6502 AM |
3115 | #ifdef CONFIG_FAILSLAB |
3116 | ||
3117 | static struct failslab_attr { | |
3118 | ||
3119 | struct fault_attr attr; | |
3120 | ||
3121 | u32 ignore_gfp_wait; | |
3122 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3123 | struct dentry *ignore_gfp_wait_file; | |
3124 | #endif | |
3125 | ||
3126 | } failslab = { | |
3127 | .attr = FAULT_ATTR_INITIALIZER, | |
6b1b60f4 | 3128 | .ignore_gfp_wait = 1, |
8a8b6502 AM |
3129 | }; |
3130 | ||
3131 | static int __init setup_failslab(char *str) | |
3132 | { | |
3133 | return setup_fault_attr(&failslab.attr, str); | |
3134 | } | |
3135 | __setup("failslab=", setup_failslab); | |
3136 | ||
3137 | static int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3138 | { | |
3139 | if (cachep == &cache_cache) | |
3140 | return 0; | |
3141 | if (flags & __GFP_NOFAIL) | |
3142 | return 0; | |
3143 | if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) | |
3144 | return 0; | |
3145 | ||
3146 | return should_fail(&failslab.attr, obj_size(cachep)); | |
3147 | } | |
3148 | ||
3149 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3150 | ||
3151 | static int __init failslab_debugfs(void) | |
3152 | { | |
3153 | mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
3154 | struct dentry *dir; | |
3155 | int err; | |
3156 | ||
824ebef1 | 3157 | err = init_fault_attr_dentries(&failslab.attr, "failslab"); |
8a8b6502 AM |
3158 | if (err) |
3159 | return err; | |
3160 | dir = failslab.attr.dentries.dir; | |
3161 | ||
3162 | failslab.ignore_gfp_wait_file = | |
3163 | debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
3164 | &failslab.ignore_gfp_wait); | |
3165 | ||
3166 | if (!failslab.ignore_gfp_wait_file) { | |
3167 | err = -ENOMEM; | |
3168 | debugfs_remove(failslab.ignore_gfp_wait_file); | |
3169 | cleanup_fault_attr_dentries(&failslab.attr); | |
3170 | } | |
3171 | ||
3172 | return err; | |
3173 | } | |
3174 | ||
3175 | late_initcall(failslab_debugfs); | |
3176 | ||
3177 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
3178 | ||
3179 | #else /* CONFIG_FAILSLAB */ | |
3180 | ||
3181 | static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3182 | { | |
3183 | return 0; | |
3184 | } | |
3185 | ||
3186 | #endif /* CONFIG_FAILSLAB */ | |
3187 | ||
343e0d7a | 3188 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3189 | { |
b28a02de | 3190 | void *objp; |
1da177e4 LT |
3191 | struct array_cache *ac; |
3192 | ||
5c382300 | 3193 | check_irq_off(); |
8a8b6502 | 3194 | |
9a2dba4b | 3195 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3196 | if (likely(ac->avail)) { |
3197 | STATS_INC_ALLOCHIT(cachep); | |
3198 | ac->touched = 1; | |
e498be7d | 3199 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3200 | } else { |
3201 | STATS_INC_ALLOCMISS(cachep); | |
3202 | objp = cache_alloc_refill(cachep, flags); | |
3203 | } | |
5c382300 AK |
3204 | return objp; |
3205 | } | |
3206 | ||
e498be7d | 3207 | #ifdef CONFIG_NUMA |
c61afb18 | 3208 | /* |
b2455396 | 3209 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3210 | * |
3211 | * If we are in_interrupt, then process context, including cpusets and | |
3212 | * mempolicy, may not apply and should not be used for allocation policy. | |
3213 | */ | |
3214 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3215 | { | |
3216 | int nid_alloc, nid_here; | |
3217 | ||
765c4507 | 3218 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3219 | return NULL; |
3220 | nid_alloc = nid_here = numa_node_id(); | |
3221 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3222 | nid_alloc = cpuset_mem_spread_node(); | |
3223 | else if (current->mempolicy) | |
3224 | nid_alloc = slab_node(current->mempolicy); | |
3225 | if (nid_alloc != nid_here) | |
8b98c169 | 3226 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3227 | return NULL; |
3228 | } | |
3229 | ||
765c4507 CL |
3230 | /* |
3231 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3232 | * certain node and fall back is permitted. First we scan all the |
3233 | * available nodelists for available objects. If that fails then we | |
3234 | * perform an allocation without specifying a node. This allows the page | |
3235 | * allocator to do its reclaim / fallback magic. We then insert the | |
3236 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3237 | */ |
8c8cc2c1 | 3238 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3239 | { |
8c8cc2c1 PE |
3240 | struct zonelist *zonelist; |
3241 | gfp_t local_flags; | |
765c4507 CL |
3242 | struct zone **z; |
3243 | void *obj = NULL; | |
3c517a61 | 3244 | int nid; |
8c8cc2c1 PE |
3245 | |
3246 | if (flags & __GFP_THISNODE) | |
3247 | return NULL; | |
3248 | ||
3249 | zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
3250 | ->node_zonelists[gfp_zone(flags)]; | |
6cb06229 | 3251 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3252 | |
3c517a61 CL |
3253 | retry: |
3254 | /* | |
3255 | * Look through allowed nodes for objects available | |
3256 | * from existing per node queues. | |
3257 | */ | |
aedb0eb1 | 3258 | for (z = zonelist->zones; *z && !obj; z++) { |
3c517a61 | 3259 | nid = zone_to_nid(*z); |
aedb0eb1 | 3260 | |
02a0e53d | 3261 | if (cpuset_zone_allowed_hardwall(*z, flags) && |
3c517a61 CL |
3262 | cache->nodelists[nid] && |
3263 | cache->nodelists[nid]->free_objects) | |
3264 | obj = ____cache_alloc_node(cache, | |
3265 | flags | GFP_THISNODE, nid); | |
3266 | } | |
3267 | ||
cfce6604 | 3268 | if (!obj) { |
3c517a61 CL |
3269 | /* |
3270 | * This allocation will be performed within the constraints | |
3271 | * of the current cpuset / memory policy requirements. | |
3272 | * We may trigger various forms of reclaim on the allowed | |
3273 | * set and go into memory reserves if necessary. | |
3274 | */ | |
dd47ea75 CL |
3275 | if (local_flags & __GFP_WAIT) |
3276 | local_irq_enable(); | |
3277 | kmem_flagcheck(cache, flags); | |
3c517a61 | 3278 | obj = kmem_getpages(cache, flags, -1); |
dd47ea75 CL |
3279 | if (local_flags & __GFP_WAIT) |
3280 | local_irq_disable(); | |
3c517a61 CL |
3281 | if (obj) { |
3282 | /* | |
3283 | * Insert into the appropriate per node queues | |
3284 | */ | |
3285 | nid = page_to_nid(virt_to_page(obj)); | |
3286 | if (cache_grow(cache, flags, nid, obj)) { | |
3287 | obj = ____cache_alloc_node(cache, | |
3288 | flags | GFP_THISNODE, nid); | |
3289 | if (!obj) | |
3290 | /* | |
3291 | * Another processor may allocate the | |
3292 | * objects in the slab since we are | |
3293 | * not holding any locks. | |
3294 | */ | |
3295 | goto retry; | |
3296 | } else { | |
b6a60451 | 3297 | /* cache_grow already freed obj */ |
3c517a61 CL |
3298 | obj = NULL; |
3299 | } | |
3300 | } | |
aedb0eb1 | 3301 | } |
765c4507 CL |
3302 | return obj; |
3303 | } | |
3304 | ||
e498be7d CL |
3305 | /* |
3306 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3307 | */ |
8b98c169 | 3308 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3309 | int nodeid) |
e498be7d CL |
3310 | { |
3311 | struct list_head *entry; | |
b28a02de PE |
3312 | struct slab *slabp; |
3313 | struct kmem_list3 *l3; | |
3314 | void *obj; | |
b28a02de PE |
3315 | int x; |
3316 | ||
3317 | l3 = cachep->nodelists[nodeid]; | |
3318 | BUG_ON(!l3); | |
3319 | ||
a737b3e2 | 3320 | retry: |
ca3b9b91 | 3321 | check_irq_off(); |
b28a02de PE |
3322 | spin_lock(&l3->list_lock); |
3323 | entry = l3->slabs_partial.next; | |
3324 | if (entry == &l3->slabs_partial) { | |
3325 | l3->free_touched = 1; | |
3326 | entry = l3->slabs_free.next; | |
3327 | if (entry == &l3->slabs_free) | |
3328 | goto must_grow; | |
3329 | } | |
3330 | ||
3331 | slabp = list_entry(entry, struct slab, list); | |
3332 | check_spinlock_acquired_node(cachep, nodeid); | |
3333 | check_slabp(cachep, slabp); | |
3334 | ||
3335 | STATS_INC_NODEALLOCS(cachep); | |
3336 | STATS_INC_ACTIVE(cachep); | |
3337 | STATS_SET_HIGH(cachep); | |
3338 | ||
3339 | BUG_ON(slabp->inuse == cachep->num); | |
3340 | ||
78d382d7 | 3341 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3342 | check_slabp(cachep, slabp); |
3343 | l3->free_objects--; | |
3344 | /* move slabp to correct slabp list: */ | |
3345 | list_del(&slabp->list); | |
3346 | ||
a737b3e2 | 3347 | if (slabp->free == BUFCTL_END) |
b28a02de | 3348 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3349 | else |
b28a02de | 3350 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3351 | |
b28a02de PE |
3352 | spin_unlock(&l3->list_lock); |
3353 | goto done; | |
e498be7d | 3354 | |
a737b3e2 | 3355 | must_grow: |
b28a02de | 3356 | spin_unlock(&l3->list_lock); |
3c517a61 | 3357 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3358 | if (x) |
3359 | goto retry; | |
1da177e4 | 3360 | |
8c8cc2c1 | 3361 | return fallback_alloc(cachep, flags); |
e498be7d | 3362 | |
a737b3e2 | 3363 | done: |
b28a02de | 3364 | return obj; |
e498be7d | 3365 | } |
8c8cc2c1 PE |
3366 | |
3367 | /** | |
3368 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3369 | * @cachep: The cache to allocate from. | |
3370 | * @flags: See kmalloc(). | |
3371 | * @nodeid: node number of the target node. | |
3372 | * @caller: return address of caller, used for debug information | |
3373 | * | |
3374 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3375 | * node, which can improve the performance for cpu bound structures. | |
3376 | * | |
3377 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3378 | */ | |
3379 | static __always_inline void * | |
3380 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3381 | void *caller) | |
3382 | { | |
3383 | unsigned long save_flags; | |
3384 | void *ptr; | |
3385 | ||
824ebef1 AM |
3386 | if (should_failslab(cachep, flags)) |
3387 | return NULL; | |
3388 | ||
8c8cc2c1 PE |
3389 | cache_alloc_debugcheck_before(cachep, flags); |
3390 | local_irq_save(save_flags); | |
3391 | ||
3392 | if (unlikely(nodeid == -1)) | |
3393 | nodeid = numa_node_id(); | |
3394 | ||
3395 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3396 | /* Node not bootstrapped yet */ | |
3397 | ptr = fallback_alloc(cachep, flags); | |
3398 | goto out; | |
3399 | } | |
3400 | ||
3401 | if (nodeid == numa_node_id()) { | |
3402 | /* | |
3403 | * Use the locally cached objects if possible. | |
3404 | * However ____cache_alloc does not allow fallback | |
3405 | * to other nodes. It may fail while we still have | |
3406 | * objects on other nodes available. | |
3407 | */ | |
3408 | ptr = ____cache_alloc(cachep, flags); | |
3409 | if (ptr) | |
3410 | goto out; | |
3411 | } | |
3412 | /* ___cache_alloc_node can fall back to other nodes */ | |
3413 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3414 | out: | |
3415 | local_irq_restore(save_flags); | |
3416 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
3417 | ||
d07dbea4 CL |
3418 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3419 | memset(ptr, 0, obj_size(cachep)); | |
3420 | ||
8c8cc2c1 PE |
3421 | return ptr; |
3422 | } | |
3423 | ||
3424 | static __always_inline void * | |
3425 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3426 | { | |
3427 | void *objp; | |
3428 | ||
3429 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3430 | objp = alternate_node_alloc(cache, flags); | |
3431 | if (objp) | |
3432 | goto out; | |
3433 | } | |
3434 | objp = ____cache_alloc(cache, flags); | |
3435 | ||
3436 | /* | |
3437 | * We may just have run out of memory on the local node. | |
3438 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3439 | */ | |
3440 | if (!objp) | |
3441 | objp = ____cache_alloc_node(cache, flags, numa_node_id()); | |
3442 | ||
3443 | out: | |
3444 | return objp; | |
3445 | } | |
3446 | #else | |
3447 | ||
3448 | static __always_inline void * | |
3449 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3450 | { | |
3451 | return ____cache_alloc(cachep, flags); | |
3452 | } | |
3453 | ||
3454 | #endif /* CONFIG_NUMA */ | |
3455 | ||
3456 | static __always_inline void * | |
3457 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3458 | { | |
3459 | unsigned long save_flags; | |
3460 | void *objp; | |
3461 | ||
824ebef1 AM |
3462 | if (should_failslab(cachep, flags)) |
3463 | return NULL; | |
3464 | ||
8c8cc2c1 PE |
3465 | cache_alloc_debugcheck_before(cachep, flags); |
3466 | local_irq_save(save_flags); | |
3467 | objp = __do_cache_alloc(cachep, flags); | |
3468 | local_irq_restore(save_flags); | |
3469 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
3470 | prefetchw(objp); | |
3471 | ||
d07dbea4 CL |
3472 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3473 | memset(objp, 0, obj_size(cachep)); | |
3474 | ||
8c8cc2c1 PE |
3475 | return objp; |
3476 | } | |
e498be7d CL |
3477 | |
3478 | /* | |
3479 | * Caller needs to acquire correct kmem_list's list_lock | |
3480 | */ | |
343e0d7a | 3481 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3482 | int node) |
1da177e4 LT |
3483 | { |
3484 | int i; | |
e498be7d | 3485 | struct kmem_list3 *l3; |
1da177e4 LT |
3486 | |
3487 | for (i = 0; i < nr_objects; i++) { | |
3488 | void *objp = objpp[i]; | |
3489 | struct slab *slabp; | |
1da177e4 | 3490 | |
6ed5eb22 | 3491 | slabp = virt_to_slab(objp); |
ff69416e | 3492 | l3 = cachep->nodelists[node]; |
1da177e4 | 3493 | list_del(&slabp->list); |
ff69416e | 3494 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3495 | check_slabp(cachep, slabp); |
78d382d7 | 3496 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3497 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3498 | l3->free_objects++; |
1da177e4 LT |
3499 | check_slabp(cachep, slabp); |
3500 | ||
3501 | /* fixup slab chains */ | |
3502 | if (slabp->inuse == 0) { | |
e498be7d CL |
3503 | if (l3->free_objects > l3->free_limit) { |
3504 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3505 | /* No need to drop any previously held |
3506 | * lock here, even if we have a off-slab slab | |
3507 | * descriptor it is guaranteed to come from | |
3508 | * a different cache, refer to comments before | |
3509 | * alloc_slabmgmt. | |
3510 | */ | |
1da177e4 LT |
3511 | slab_destroy(cachep, slabp); |
3512 | } else { | |
e498be7d | 3513 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3514 | } |
3515 | } else { | |
3516 | /* Unconditionally move a slab to the end of the | |
3517 | * partial list on free - maximum time for the | |
3518 | * other objects to be freed, too. | |
3519 | */ | |
e498be7d | 3520 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3521 | } |
3522 | } | |
3523 | } | |
3524 | ||
343e0d7a | 3525 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3526 | { |
3527 | int batchcount; | |
e498be7d | 3528 | struct kmem_list3 *l3; |
ff69416e | 3529 | int node = numa_node_id(); |
1da177e4 LT |
3530 | |
3531 | batchcount = ac->batchcount; | |
3532 | #if DEBUG | |
3533 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3534 | #endif | |
3535 | check_irq_off(); | |
ff69416e | 3536 | l3 = cachep->nodelists[node]; |
873623df | 3537 | spin_lock(&l3->list_lock); |
e498be7d CL |
3538 | if (l3->shared) { |
3539 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3540 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3541 | if (max) { |
3542 | if (batchcount > max) | |
3543 | batchcount = max; | |
e498be7d | 3544 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3545 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3546 | shared_array->avail += batchcount; |
3547 | goto free_done; | |
3548 | } | |
3549 | } | |
3550 | ||
ff69416e | 3551 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3552 | free_done: |
1da177e4 LT |
3553 | #if STATS |
3554 | { | |
3555 | int i = 0; | |
3556 | struct list_head *p; | |
3557 | ||
e498be7d CL |
3558 | p = l3->slabs_free.next; |
3559 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3560 | struct slab *slabp; |
3561 | ||
3562 | slabp = list_entry(p, struct slab, list); | |
3563 | BUG_ON(slabp->inuse); | |
3564 | ||
3565 | i++; | |
3566 | p = p->next; | |
3567 | } | |
3568 | STATS_SET_FREEABLE(cachep, i); | |
3569 | } | |
3570 | #endif | |
e498be7d | 3571 | spin_unlock(&l3->list_lock); |
1da177e4 | 3572 | ac->avail -= batchcount; |
a737b3e2 | 3573 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3574 | } |
3575 | ||
3576 | /* | |
a737b3e2 AM |
3577 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3578 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3579 | */ |
873623df | 3580 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3581 | { |
9a2dba4b | 3582 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3583 | |
3584 | check_irq_off(); | |
3585 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3586 | ||
1807a1aa SS |
3587 | /* |
3588 | * Skip calling cache_free_alien() when the platform is not numa. | |
3589 | * This will avoid cache misses that happen while accessing slabp (which | |
3590 | * is per page memory reference) to get nodeid. Instead use a global | |
3591 | * variable to skip the call, which is mostly likely to be present in | |
3592 | * the cache. | |
3593 | */ | |
3594 | if (numa_platform && cache_free_alien(cachep, objp)) | |
729bd0b7 PE |
3595 | return; |
3596 | ||
1da177e4 LT |
3597 | if (likely(ac->avail < ac->limit)) { |
3598 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3599 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3600 | return; |
3601 | } else { | |
3602 | STATS_INC_FREEMISS(cachep); | |
3603 | cache_flusharray(cachep, ac); | |
e498be7d | 3604 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3605 | } |
3606 | } | |
3607 | ||
3608 | /** | |
3609 | * kmem_cache_alloc - Allocate an object | |
3610 | * @cachep: The cache to allocate from. | |
3611 | * @flags: See kmalloc(). | |
3612 | * | |
3613 | * Allocate an object from this cache. The flags are only relevant | |
3614 | * if the cache has no available objects. | |
3615 | */ | |
343e0d7a | 3616 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3617 | { |
7fd6b141 | 3618 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3619 | } |
3620 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3621 | ||
3622 | /** | |
3623 | * kmem_ptr_validate - check if an untrusted pointer might | |
3624 | * be a slab entry. | |
3625 | * @cachep: the cache we're checking against | |
3626 | * @ptr: pointer to validate | |
3627 | * | |
3628 | * This verifies that the untrusted pointer looks sane: | |
3629 | * it is _not_ a guarantee that the pointer is actually | |
3630 | * part of the slab cache in question, but it at least | |
3631 | * validates that the pointer can be dereferenced and | |
3632 | * looks half-way sane. | |
3633 | * | |
3634 | * Currently only used for dentry validation. | |
3635 | */ | |
b7f869a2 | 3636 | int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) |
1da177e4 | 3637 | { |
b28a02de | 3638 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3639 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3640 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3641 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3642 | struct page *page; |
3643 | ||
3644 | if (unlikely(addr < min_addr)) | |
3645 | goto out; | |
3646 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3647 | goto out; | |
3648 | if (unlikely(addr & align_mask)) | |
3649 | goto out; | |
3650 | if (unlikely(!kern_addr_valid(addr))) | |
3651 | goto out; | |
3652 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3653 | goto out; | |
3654 | page = virt_to_page(ptr); | |
3655 | if (unlikely(!PageSlab(page))) | |
3656 | goto out; | |
065d41cb | 3657 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3658 | goto out; |
3659 | return 1; | |
a737b3e2 | 3660 | out: |
1da177e4 LT |
3661 | return 0; |
3662 | } | |
3663 | ||
3664 | #ifdef CONFIG_NUMA | |
8b98c169 CH |
3665 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3666 | { | |
3667 | return __cache_alloc_node(cachep, flags, nodeid, | |
3668 | __builtin_return_address(0)); | |
3669 | } | |
1da177e4 LT |
3670 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3671 | ||
8b98c169 CH |
3672 | static __always_inline void * |
3673 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3674 | { |
343e0d7a | 3675 | struct kmem_cache *cachep; |
97e2bde4 MS |
3676 | |
3677 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3678 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3679 | return cachep; | |
97e2bde4 MS |
3680 | return kmem_cache_alloc_node(cachep, flags, node); |
3681 | } | |
8b98c169 CH |
3682 | |
3683 | #ifdef CONFIG_DEBUG_SLAB | |
3684 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3685 | { | |
3686 | return __do_kmalloc_node(size, flags, node, | |
3687 | __builtin_return_address(0)); | |
3688 | } | |
dbe5e69d | 3689 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3690 | |
3691 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
3692 | int node, void *caller) | |
3693 | { | |
3694 | return __do_kmalloc_node(size, flags, node, caller); | |
3695 | } | |
3696 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3697 | #else | |
3698 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3699 | { | |
3700 | return __do_kmalloc_node(size, flags, node, NULL); | |
3701 | } | |
3702 | EXPORT_SYMBOL(__kmalloc_node); | |
3703 | #endif /* CONFIG_DEBUG_SLAB */ | |
3704 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3705 | |
3706 | /** | |
800590f5 | 3707 | * __do_kmalloc - allocate memory |
1da177e4 | 3708 | * @size: how many bytes of memory are required. |
800590f5 | 3709 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3710 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3711 | */ |
7fd6b141 PE |
3712 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3713 | void *caller) | |
1da177e4 | 3714 | { |
343e0d7a | 3715 | struct kmem_cache *cachep; |
1da177e4 | 3716 | |
97e2bde4 MS |
3717 | /* If you want to save a few bytes .text space: replace |
3718 | * __ with kmem_. | |
3719 | * Then kmalloc uses the uninlined functions instead of the inline | |
3720 | * functions. | |
3721 | */ | |
3722 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3723 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3724 | return cachep; | |
7fd6b141 PE |
3725 | return __cache_alloc(cachep, flags, caller); |
3726 | } | |
3727 | ||
7fd6b141 | 3728 | |
1d2c8eea | 3729 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3730 | void *__kmalloc(size_t size, gfp_t flags) |
3731 | { | |
871751e2 | 3732 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3733 | } |
3734 | EXPORT_SYMBOL(__kmalloc); | |
3735 | ||
7fd6b141 PE |
3736 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3737 | { | |
3738 | return __do_kmalloc(size, flags, caller); | |
3739 | } | |
3740 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3741 | |
3742 | #else | |
3743 | void *__kmalloc(size_t size, gfp_t flags) | |
3744 | { | |
3745 | return __do_kmalloc(size, flags, NULL); | |
3746 | } | |
3747 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3748 | #endif |
3749 | ||
1da177e4 LT |
3750 | /** |
3751 | * kmem_cache_free - Deallocate an object | |
3752 | * @cachep: The cache the allocation was from. | |
3753 | * @objp: The previously allocated object. | |
3754 | * | |
3755 | * Free an object which was previously allocated from this | |
3756 | * cache. | |
3757 | */ | |
343e0d7a | 3758 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3759 | { |
3760 | unsigned long flags; | |
3761 | ||
ddc2e812 PE |
3762 | BUG_ON(virt_to_cache(objp) != cachep); |
3763 | ||
1da177e4 | 3764 | local_irq_save(flags); |
898552c9 | 3765 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
873623df | 3766 | __cache_free(cachep, objp); |
1da177e4 LT |
3767 | local_irq_restore(flags); |
3768 | } | |
3769 | EXPORT_SYMBOL(kmem_cache_free); | |
3770 | ||
1da177e4 LT |
3771 | /** |
3772 | * kfree - free previously allocated memory | |
3773 | * @objp: pointer returned by kmalloc. | |
3774 | * | |
80e93eff PE |
3775 | * If @objp is NULL, no operation is performed. |
3776 | * | |
1da177e4 LT |
3777 | * Don't free memory not originally allocated by kmalloc() |
3778 | * or you will run into trouble. | |
3779 | */ | |
3780 | void kfree(const void *objp) | |
3781 | { | |
343e0d7a | 3782 | struct kmem_cache *c; |
1da177e4 LT |
3783 | unsigned long flags; |
3784 | ||
6cb8f913 | 3785 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3786 | return; |
3787 | local_irq_save(flags); | |
3788 | kfree_debugcheck(objp); | |
6ed5eb22 | 3789 | c = virt_to_cache(objp); |
f9b8404c | 3790 | debug_check_no_locks_freed(objp, obj_size(c)); |
873623df | 3791 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3792 | local_irq_restore(flags); |
3793 | } | |
3794 | EXPORT_SYMBOL(kfree); | |
3795 | ||
343e0d7a | 3796 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3797 | { |
3dafccf2 | 3798 | return obj_size(cachep); |
1da177e4 LT |
3799 | } |
3800 | EXPORT_SYMBOL(kmem_cache_size); | |
3801 | ||
343e0d7a | 3802 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3803 | { |
3804 | return cachep->name; | |
3805 | } | |
3806 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3807 | ||
e498be7d | 3808 | /* |
183ff22b | 3809 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3810 | */ |
343e0d7a | 3811 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3812 | { |
3813 | int node; | |
3814 | struct kmem_list3 *l3; | |
cafeb02e | 3815 | struct array_cache *new_shared; |
3395ee05 | 3816 | struct array_cache **new_alien = NULL; |
e498be7d | 3817 | |
04231b30 | 3818 | for_each_node_state(node, N_NORMAL_MEMORY) { |
cafeb02e | 3819 | |
3395ee05 PM |
3820 | if (use_alien_caches) { |
3821 | new_alien = alloc_alien_cache(node, cachep->limit); | |
3822 | if (!new_alien) | |
3823 | goto fail; | |
3824 | } | |
cafeb02e | 3825 | |
63109846 ED |
3826 | new_shared = NULL; |
3827 | if (cachep->shared) { | |
3828 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3829 | cachep->shared*cachep->batchcount, |
a737b3e2 | 3830 | 0xbaadf00d); |
63109846 ED |
3831 | if (!new_shared) { |
3832 | free_alien_cache(new_alien); | |
3833 | goto fail; | |
3834 | } | |
0718dc2a | 3835 | } |
cafeb02e | 3836 | |
a737b3e2 AM |
3837 | l3 = cachep->nodelists[node]; |
3838 | if (l3) { | |
cafeb02e CL |
3839 | struct array_cache *shared = l3->shared; |
3840 | ||
e498be7d CL |
3841 | spin_lock_irq(&l3->list_lock); |
3842 | ||
cafeb02e | 3843 | if (shared) |
0718dc2a CL |
3844 | free_block(cachep, shared->entry, |
3845 | shared->avail, node); | |
e498be7d | 3846 | |
cafeb02e CL |
3847 | l3->shared = new_shared; |
3848 | if (!l3->alien) { | |
e498be7d CL |
3849 | l3->alien = new_alien; |
3850 | new_alien = NULL; | |
3851 | } | |
b28a02de | 3852 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3853 | cachep->batchcount + cachep->num; |
e498be7d | 3854 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3855 | kfree(shared); |
e498be7d CL |
3856 | free_alien_cache(new_alien); |
3857 | continue; | |
3858 | } | |
a737b3e2 | 3859 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3860 | if (!l3) { |
3861 | free_alien_cache(new_alien); | |
3862 | kfree(new_shared); | |
e498be7d | 3863 | goto fail; |
0718dc2a | 3864 | } |
e498be7d CL |
3865 | |
3866 | kmem_list3_init(l3); | |
3867 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3868 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3869 | l3->shared = new_shared; |
e498be7d | 3870 | l3->alien = new_alien; |
b28a02de | 3871 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3872 | cachep->batchcount + cachep->num; |
e498be7d CL |
3873 | cachep->nodelists[node] = l3; |
3874 | } | |
cafeb02e | 3875 | return 0; |
0718dc2a | 3876 | |
a737b3e2 | 3877 | fail: |
0718dc2a CL |
3878 | if (!cachep->next.next) { |
3879 | /* Cache is not active yet. Roll back what we did */ | |
3880 | node--; | |
3881 | while (node >= 0) { | |
3882 | if (cachep->nodelists[node]) { | |
3883 | l3 = cachep->nodelists[node]; | |
3884 | ||
3885 | kfree(l3->shared); | |
3886 | free_alien_cache(l3->alien); | |
3887 | kfree(l3); | |
3888 | cachep->nodelists[node] = NULL; | |
3889 | } | |
3890 | node--; | |
3891 | } | |
3892 | } | |
cafeb02e | 3893 | return -ENOMEM; |
e498be7d CL |
3894 | } |
3895 | ||
1da177e4 | 3896 | struct ccupdate_struct { |
343e0d7a | 3897 | struct kmem_cache *cachep; |
1da177e4 LT |
3898 | struct array_cache *new[NR_CPUS]; |
3899 | }; | |
3900 | ||
3901 | static void do_ccupdate_local(void *info) | |
3902 | { | |
a737b3e2 | 3903 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3904 | struct array_cache *old; |
3905 | ||
3906 | check_irq_off(); | |
9a2dba4b | 3907 | old = cpu_cache_get(new->cachep); |
e498be7d | 3908 | |
1da177e4 LT |
3909 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3910 | new->new[smp_processor_id()] = old; | |
3911 | } | |
3912 | ||
b5d8ca7c | 3913 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3914 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3915 | int batchcount, int shared) | |
1da177e4 | 3916 | { |
d2e7b7d0 | 3917 | struct ccupdate_struct *new; |
2ed3a4ef | 3918 | int i; |
1da177e4 | 3919 | |
d2e7b7d0 SS |
3920 | new = kzalloc(sizeof(*new), GFP_KERNEL); |
3921 | if (!new) | |
3922 | return -ENOMEM; | |
3923 | ||
e498be7d | 3924 | for_each_online_cpu(i) { |
d2e7b7d0 | 3925 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
a737b3e2 | 3926 | batchcount); |
d2e7b7d0 | 3927 | if (!new->new[i]) { |
b28a02de | 3928 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3929 | kfree(new->new[i]); |
3930 | kfree(new); | |
e498be7d | 3931 | return -ENOMEM; |
1da177e4 LT |
3932 | } |
3933 | } | |
d2e7b7d0 | 3934 | new->cachep = cachep; |
1da177e4 | 3935 | |
d2e7b7d0 | 3936 | on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); |
e498be7d | 3937 | |
1da177e4 | 3938 | check_irq_on(); |
1da177e4 LT |
3939 | cachep->batchcount = batchcount; |
3940 | cachep->limit = limit; | |
e498be7d | 3941 | cachep->shared = shared; |
1da177e4 | 3942 | |
e498be7d | 3943 | for_each_online_cpu(i) { |
d2e7b7d0 | 3944 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3945 | if (!ccold) |
3946 | continue; | |
e498be7d | 3947 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3948 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3949 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3950 | kfree(ccold); |
3951 | } | |
d2e7b7d0 | 3952 | kfree(new); |
2ed3a4ef | 3953 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3954 | } |
3955 | ||
b5d8ca7c | 3956 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3957 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3958 | { |
3959 | int err; | |
3960 | int limit, shared; | |
3961 | ||
a737b3e2 AM |
3962 | /* |
3963 | * The head array serves three purposes: | |
1da177e4 LT |
3964 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3965 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3966 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3967 | * bufctl chains: array operations are cheaper. |
3968 | * The numbers are guessed, we should auto-tune as described by | |
3969 | * Bonwick. | |
3970 | */ | |
3dafccf2 | 3971 | if (cachep->buffer_size > 131072) |
1da177e4 | 3972 | limit = 1; |
3dafccf2 | 3973 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3974 | limit = 8; |
3dafccf2 | 3975 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3976 | limit = 24; |
3dafccf2 | 3977 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3978 | limit = 54; |
3979 | else | |
3980 | limit = 120; | |
3981 | ||
a737b3e2 AM |
3982 | /* |
3983 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3984 | * allocation behaviour: Most allocs on one cpu, most free operations |
3985 | * on another cpu. For these cases, an efficient object passing between | |
3986 | * cpus is necessary. This is provided by a shared array. The array | |
3987 | * replaces Bonwick's magazine layer. | |
3988 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3989 | * to a larger limit. Thus disabled by default. | |
3990 | */ | |
3991 | shared = 0; | |
364fbb29 | 3992 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 3993 | shared = 8; |
1da177e4 LT |
3994 | |
3995 | #if DEBUG | |
a737b3e2 AM |
3996 | /* |
3997 | * With debugging enabled, large batchcount lead to excessively long | |
3998 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3999 | */ |
4000 | if (limit > 32) | |
4001 | limit = 32; | |
4002 | #endif | |
b28a02de | 4003 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
4004 | if (err) |
4005 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4006 | cachep->name, -err); |
2ed3a4ef | 4007 | return err; |
1da177e4 LT |
4008 | } |
4009 | ||
1b55253a CL |
4010 | /* |
4011 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4012 | * necessary. Note that the l3 listlock also protects the array_cache |
4013 | * if drain_array() is used on the shared array. | |
1b55253a CL |
4014 | */ |
4015 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4016 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
4017 | { |
4018 | int tofree; | |
4019 | ||
1b55253a CL |
4020 | if (!ac || !ac->avail) |
4021 | return; | |
1da177e4 LT |
4022 | if (ac->touched && !force) { |
4023 | ac->touched = 0; | |
b18e7e65 | 4024 | } else { |
1b55253a | 4025 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4026 | if (ac->avail) { |
4027 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4028 | if (tofree > ac->avail) | |
4029 | tofree = (ac->avail + 1) / 2; | |
4030 | free_block(cachep, ac->entry, tofree, node); | |
4031 | ac->avail -= tofree; | |
4032 | memmove(ac->entry, &(ac->entry[tofree]), | |
4033 | sizeof(void *) * ac->avail); | |
4034 | } | |
1b55253a | 4035 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4036 | } |
4037 | } | |
4038 | ||
4039 | /** | |
4040 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4041 | * @w: work descriptor |
1da177e4 LT |
4042 | * |
4043 | * Called from workqueue/eventd every few seconds. | |
4044 | * Purpose: | |
4045 | * - clear the per-cpu caches for this CPU. | |
4046 | * - return freeable pages to the main free memory pool. | |
4047 | * | |
a737b3e2 AM |
4048 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4049 | * again on the next iteration. | |
1da177e4 | 4050 | */ |
7c5cae36 | 4051 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4052 | { |
7a7c381d | 4053 | struct kmem_cache *searchp; |
e498be7d | 4054 | struct kmem_list3 *l3; |
aab2207c | 4055 | int node = numa_node_id(); |
7c5cae36 CL |
4056 | struct delayed_work *work = |
4057 | container_of(w, struct delayed_work, work); | |
1da177e4 | 4058 | |
7c5cae36 | 4059 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4060 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4061 | goto out; |
1da177e4 | 4062 | |
7a7c381d | 4063 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4064 | check_irq_on(); |
4065 | ||
35386e3b CL |
4066 | /* |
4067 | * We only take the l3 lock if absolutely necessary and we | |
4068 | * have established with reasonable certainty that | |
4069 | * we can do some work if the lock was obtained. | |
4070 | */ | |
aab2207c | 4071 | l3 = searchp->nodelists[node]; |
35386e3b | 4072 | |
8fce4d8e | 4073 | reap_alien(searchp, l3); |
1da177e4 | 4074 | |
aab2207c | 4075 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4076 | |
35386e3b CL |
4077 | /* |
4078 | * These are racy checks but it does not matter | |
4079 | * if we skip one check or scan twice. | |
4080 | */ | |
e498be7d | 4081 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4082 | goto next; |
1da177e4 | 4083 | |
e498be7d | 4084 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4085 | |
aab2207c | 4086 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4087 | |
ed11d9eb | 4088 | if (l3->free_touched) |
e498be7d | 4089 | l3->free_touched = 0; |
ed11d9eb CL |
4090 | else { |
4091 | int freed; | |
1da177e4 | 4092 | |
ed11d9eb CL |
4093 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4094 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4095 | STATS_ADD_REAPED(searchp, freed); | |
4096 | } | |
35386e3b | 4097 | next: |
1da177e4 LT |
4098 | cond_resched(); |
4099 | } | |
4100 | check_irq_on(); | |
fc0abb14 | 4101 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4102 | next_reap_node(); |
7c5cae36 | 4103 | out: |
a737b3e2 | 4104 | /* Set up the next iteration */ |
7c5cae36 | 4105 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4106 | } |
4107 | ||
4108 | #ifdef CONFIG_PROC_FS | |
4109 | ||
85289f98 | 4110 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4111 | { |
85289f98 PE |
4112 | /* |
4113 | * Output format version, so at least we can change it | |
4114 | * without _too_ many complaints. | |
4115 | */ | |
1da177e4 | 4116 | #if STATS |
85289f98 | 4117 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4118 | #else |
85289f98 | 4119 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4120 | #endif |
85289f98 PE |
4121 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4122 | "<objperslab> <pagesperslab>"); | |
4123 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4124 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4125 | #if STATS |
85289f98 | 4126 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4127 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4128 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4129 | #endif |
85289f98 PE |
4130 | seq_putc(m, '\n'); |
4131 | } | |
4132 | ||
4133 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4134 | { | |
4135 | loff_t n = *pos; | |
85289f98 | 4136 | |
fc0abb14 | 4137 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4138 | if (!n) |
4139 | print_slabinfo_header(m); | |
b92151ba PE |
4140 | |
4141 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4142 | } |
4143 | ||
4144 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4145 | { | |
b92151ba | 4146 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4147 | } |
4148 | ||
4149 | static void s_stop(struct seq_file *m, void *p) | |
4150 | { | |
fc0abb14 | 4151 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4152 | } |
4153 | ||
4154 | static int s_show(struct seq_file *m, void *p) | |
4155 | { | |
b92151ba | 4156 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
b28a02de PE |
4157 | struct slab *slabp; |
4158 | unsigned long active_objs; | |
4159 | unsigned long num_objs; | |
4160 | unsigned long active_slabs = 0; | |
4161 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4162 | const char *name; |
1da177e4 | 4163 | char *error = NULL; |
e498be7d CL |
4164 | int node; |
4165 | struct kmem_list3 *l3; | |
1da177e4 | 4166 | |
1da177e4 LT |
4167 | active_objs = 0; |
4168 | num_slabs = 0; | |
e498be7d CL |
4169 | for_each_online_node(node) { |
4170 | l3 = cachep->nodelists[node]; | |
4171 | if (!l3) | |
4172 | continue; | |
4173 | ||
ca3b9b91 RT |
4174 | check_irq_on(); |
4175 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4176 | |
7a7c381d | 4177 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4178 | if (slabp->inuse != cachep->num && !error) |
4179 | error = "slabs_full accounting error"; | |
4180 | active_objs += cachep->num; | |
4181 | active_slabs++; | |
4182 | } | |
7a7c381d | 4183 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4184 | if (slabp->inuse == cachep->num && !error) |
4185 | error = "slabs_partial inuse accounting error"; | |
4186 | if (!slabp->inuse && !error) | |
4187 | error = "slabs_partial/inuse accounting error"; | |
4188 | active_objs += slabp->inuse; | |
4189 | active_slabs++; | |
4190 | } | |
7a7c381d | 4191 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4192 | if (slabp->inuse && !error) |
4193 | error = "slabs_free/inuse accounting error"; | |
4194 | num_slabs++; | |
4195 | } | |
4196 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4197 | if (l3->shared) |
4198 | shared_avail += l3->shared->avail; | |
e498be7d | 4199 | |
ca3b9b91 | 4200 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4201 | } |
b28a02de PE |
4202 | num_slabs += active_slabs; |
4203 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4204 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4205 | error = "free_objects accounting error"; |
4206 | ||
b28a02de | 4207 | name = cachep->name; |
1da177e4 LT |
4208 | if (error) |
4209 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4210 | ||
4211 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4212 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4213 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4214 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4215 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4216 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4217 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4218 | #if STATS |
b28a02de | 4219 | { /* list3 stats */ |
1da177e4 LT |
4220 | unsigned long high = cachep->high_mark; |
4221 | unsigned long allocs = cachep->num_allocations; | |
4222 | unsigned long grown = cachep->grown; | |
4223 | unsigned long reaped = cachep->reaped; | |
4224 | unsigned long errors = cachep->errors; | |
4225 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4226 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4227 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4228 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4229 | |
e498be7d | 4230 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4231 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4232 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4233 | node_frees, overflows); |
1da177e4 LT |
4234 | } |
4235 | /* cpu stats */ | |
4236 | { | |
4237 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4238 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4239 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4240 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4241 | ||
4242 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4243 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4244 | } |
4245 | #endif | |
4246 | seq_putc(m, '\n'); | |
1da177e4 LT |
4247 | return 0; |
4248 | } | |
4249 | ||
4250 | /* | |
4251 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4252 | * | |
4253 | * Output layout: | |
4254 | * cache-name | |
4255 | * num-active-objs | |
4256 | * total-objs | |
4257 | * object size | |
4258 | * num-active-slabs | |
4259 | * total-slabs | |
4260 | * num-pages-per-slab | |
4261 | * + further values on SMP and with statistics enabled | |
4262 | */ | |
4263 | ||
15ad7cdc | 4264 | const struct seq_operations slabinfo_op = { |
b28a02de PE |
4265 | .start = s_start, |
4266 | .next = s_next, | |
4267 | .stop = s_stop, | |
4268 | .show = s_show, | |
1da177e4 LT |
4269 | }; |
4270 | ||
4271 | #define MAX_SLABINFO_WRITE 128 | |
4272 | /** | |
4273 | * slabinfo_write - Tuning for the slab allocator | |
4274 | * @file: unused | |
4275 | * @buffer: user buffer | |
4276 | * @count: data length | |
4277 | * @ppos: unused | |
4278 | */ | |
b28a02de PE |
4279 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4280 | size_t count, loff_t *ppos) | |
1da177e4 | 4281 | { |
b28a02de | 4282 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4283 | int limit, batchcount, shared, res; |
7a7c381d | 4284 | struct kmem_cache *cachep; |
b28a02de | 4285 | |
1da177e4 LT |
4286 | if (count > MAX_SLABINFO_WRITE) |
4287 | return -EINVAL; | |
4288 | if (copy_from_user(&kbuf, buffer, count)) | |
4289 | return -EFAULT; | |
b28a02de | 4290 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4291 | |
4292 | tmp = strchr(kbuf, ' '); | |
4293 | if (!tmp) | |
4294 | return -EINVAL; | |
4295 | *tmp = '\0'; | |
4296 | tmp++; | |
4297 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4298 | return -EINVAL; | |
4299 | ||
4300 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4301 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4302 | res = -EINVAL; |
7a7c381d | 4303 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4304 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4305 | if (limit < 1 || batchcount < 1 || |
4306 | batchcount > limit || shared < 0) { | |
e498be7d | 4307 | res = 0; |
1da177e4 | 4308 | } else { |
e498be7d | 4309 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4310 | batchcount, shared); |
1da177e4 LT |
4311 | } |
4312 | break; | |
4313 | } | |
4314 | } | |
fc0abb14 | 4315 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4316 | if (res >= 0) |
4317 | res = count; | |
4318 | return res; | |
4319 | } | |
871751e2 AV |
4320 | |
4321 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4322 | ||
4323 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4324 | { | |
871751e2 | 4325 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4326 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4327 | } |
4328 | ||
4329 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4330 | { | |
4331 | unsigned long *p; | |
4332 | int l; | |
4333 | if (!v) | |
4334 | return 1; | |
4335 | l = n[1]; | |
4336 | p = n + 2; | |
4337 | while (l) { | |
4338 | int i = l/2; | |
4339 | unsigned long *q = p + 2 * i; | |
4340 | if (*q == v) { | |
4341 | q[1]++; | |
4342 | return 1; | |
4343 | } | |
4344 | if (*q > v) { | |
4345 | l = i; | |
4346 | } else { | |
4347 | p = q + 2; | |
4348 | l -= i + 1; | |
4349 | } | |
4350 | } | |
4351 | if (++n[1] == n[0]) | |
4352 | return 0; | |
4353 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4354 | p[0] = v; | |
4355 | p[1] = 1; | |
4356 | return 1; | |
4357 | } | |
4358 | ||
4359 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4360 | { | |
4361 | void *p; | |
4362 | int i; | |
4363 | if (n[0] == n[1]) | |
4364 | return; | |
4365 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4366 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4367 | continue; | |
4368 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4369 | return; | |
4370 | } | |
4371 | } | |
4372 | ||
4373 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4374 | { | |
4375 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4376 | unsigned long offset, size; |
9281acea | 4377 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4378 | |
a5c43dae | 4379 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4380 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4381 | if (modname[0]) |
871751e2 AV |
4382 | seq_printf(m, " [%s]", modname); |
4383 | return; | |
4384 | } | |
4385 | #endif | |
4386 | seq_printf(m, "%p", (void *)address); | |
4387 | } | |
4388 | ||
4389 | static int leaks_show(struct seq_file *m, void *p) | |
4390 | { | |
b92151ba | 4391 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4392 | struct slab *slabp; |
4393 | struct kmem_list3 *l3; | |
4394 | const char *name; | |
4395 | unsigned long *n = m->private; | |
4396 | int node; | |
4397 | int i; | |
4398 | ||
4399 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4400 | return 0; | |
4401 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4402 | return 0; | |
4403 | ||
4404 | /* OK, we can do it */ | |
4405 | ||
4406 | n[1] = 0; | |
4407 | ||
4408 | for_each_online_node(node) { | |
4409 | l3 = cachep->nodelists[node]; | |
4410 | if (!l3) | |
4411 | continue; | |
4412 | ||
4413 | check_irq_on(); | |
4414 | spin_lock_irq(&l3->list_lock); | |
4415 | ||
7a7c381d | 4416 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4417 | handle_slab(n, cachep, slabp); |
7a7c381d | 4418 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4419 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4420 | spin_unlock_irq(&l3->list_lock); |
4421 | } | |
4422 | name = cachep->name; | |
4423 | if (n[0] == n[1]) { | |
4424 | /* Increase the buffer size */ | |
4425 | mutex_unlock(&cache_chain_mutex); | |
4426 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4427 | if (!m->private) { | |
4428 | /* Too bad, we are really out */ | |
4429 | m->private = n; | |
4430 | mutex_lock(&cache_chain_mutex); | |
4431 | return -ENOMEM; | |
4432 | } | |
4433 | *(unsigned long *)m->private = n[0] * 2; | |
4434 | kfree(n); | |
4435 | mutex_lock(&cache_chain_mutex); | |
4436 | /* Now make sure this entry will be retried */ | |
4437 | m->count = m->size; | |
4438 | return 0; | |
4439 | } | |
4440 | for (i = 0; i < n[1]; i++) { | |
4441 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4442 | show_symbol(m, n[2*i+2]); | |
4443 | seq_putc(m, '\n'); | |
4444 | } | |
d2e7b7d0 | 4445 | |
871751e2 AV |
4446 | return 0; |
4447 | } | |
4448 | ||
15ad7cdc | 4449 | const struct seq_operations slabstats_op = { |
871751e2 AV |
4450 | .start = leaks_start, |
4451 | .next = s_next, | |
4452 | .stop = s_stop, | |
4453 | .show = leaks_show, | |
4454 | }; | |
4455 | #endif | |
1da177e4 LT |
4456 | #endif |
4457 | ||
00e145b6 MS |
4458 | /** |
4459 | * ksize - get the actual amount of memory allocated for a given object | |
4460 | * @objp: Pointer to the object | |
4461 | * | |
4462 | * kmalloc may internally round up allocations and return more memory | |
4463 | * than requested. ksize() can be used to determine the actual amount of | |
4464 | * memory allocated. The caller may use this additional memory, even though | |
4465 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4466 | * The caller must guarantee that objp points to a valid object previously | |
4467 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4468 | * must not be freed during the duration of the call. | |
4469 | */ | |
fd76bab2 | 4470 | size_t ksize(const void *objp) |
1da177e4 | 4471 | { |
ef8b4520 CL |
4472 | BUG_ON(!objp); |
4473 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4474 | return 0; |
1da177e4 | 4475 | |
6ed5eb22 | 4476 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4477 | } |