]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
29 | * slabs and you must pass objects with the same intializations to | |
30 | * kmem_cache_free. | |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
54 | * it's changed with a smp_call_function(). |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
fc0abb14 | 71 | * The global cache-chain is protected by the mutex 'cache_chain_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
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 *) | |
140 | ||
141 | #ifndef cache_line_size | |
142 | #define cache_line_size() L1_CACHE_BYTES | |
143 | #endif | |
144 | ||
145 | #ifndef ARCH_KMALLOC_MINALIGN | |
146 | /* | |
147 | * Enforce a minimum alignment for the kmalloc caches. | |
148 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
149 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
150 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
b46b8f19 DW |
151 | * alignment larger than the alignment of a 64-bit integer. |
152 | * ARCH_KMALLOC_MINALIGN allows that. | |
153 | * Note that increasing this value may disable some debug features. | |
1da177e4 | 154 | */ |
b46b8f19 | 155 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
1da177e4 LT |
156 | #endif |
157 | ||
158 | #ifndef ARCH_SLAB_MINALIGN | |
159 | /* | |
160 | * Enforce a minimum alignment for all caches. | |
161 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
162 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
163 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
164 | * some debug features. | |
165 | */ | |
166 | #define ARCH_SLAB_MINALIGN 0 | |
167 | #endif | |
168 | ||
169 | #ifndef ARCH_KMALLOC_FLAGS | |
170 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
171 | #endif | |
172 | ||
173 | /* Legal flag mask for kmem_cache_create(). */ | |
174 | #if DEBUG | |
50953fe9 | 175 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 176 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 177 | SLAB_CACHE_DMA | \ |
5af60839 | 178 | SLAB_STORE_USER | \ |
1da177e4 | 179 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
101a5001 | 180 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 181 | #else |
ac2b898c | 182 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 183 | SLAB_CACHE_DMA | \ |
1da177e4 | 184 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
101a5001 | 185 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 LT |
186 | #endif |
187 | ||
188 | /* | |
189 | * kmem_bufctl_t: | |
190 | * | |
191 | * Bufctl's are used for linking objs within a slab | |
192 | * linked offsets. | |
193 | * | |
194 | * This implementation relies on "struct page" for locating the cache & | |
195 | * slab an object belongs to. | |
196 | * This allows the bufctl structure to be small (one int), but limits | |
197 | * the number of objects a slab (not a cache) can contain when off-slab | |
198 | * bufctls are used. The limit is the size of the largest general cache | |
199 | * that does not use off-slab slabs. | |
200 | * For 32bit archs with 4 kB pages, is this 56. | |
201 | * This is not serious, as it is only for large objects, when it is unwise | |
202 | * to have too many per slab. | |
203 | * Note: This limit can be raised by introducing a general cache whose size | |
204 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
205 | */ | |
206 | ||
fa5b08d5 | 207 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
208 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
209 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
210 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
211 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 212 | |
1da177e4 LT |
213 | /* |
214 | * struct slab | |
215 | * | |
216 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
217 | * for a slab, or allocated from an general cache. | |
218 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
219 | */ | |
220 | struct slab { | |
b28a02de PE |
221 | struct list_head list; |
222 | unsigned long colouroff; | |
223 | void *s_mem; /* including colour offset */ | |
224 | unsigned int inuse; /* num of objs active in slab */ | |
225 | kmem_bufctl_t free; | |
226 | unsigned short nodeid; | |
1da177e4 LT |
227 | }; |
228 | ||
229 | /* | |
230 | * struct slab_rcu | |
231 | * | |
232 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
233 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
234 | * we need to approach a kernel structure obliquely, from its address | |
235 | * obtained without the usual locking. We can lock the structure to | |
236 | * stabilize it and check it's still at the given address, only if we | |
237 | * can be sure that the memory has not been meanwhile reused for some | |
238 | * other kind of object (which our subsystem's lock might corrupt). | |
239 | * | |
240 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
241 | * taking the spinlock within the structure expected at that address. | |
242 | * | |
243 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
244 | */ | |
245 | struct slab_rcu { | |
b28a02de | 246 | struct rcu_head head; |
343e0d7a | 247 | struct kmem_cache *cachep; |
b28a02de | 248 | void *addr; |
1da177e4 LT |
249 | }; |
250 | ||
251 | /* | |
252 | * struct array_cache | |
253 | * | |
1da177e4 LT |
254 | * Purpose: |
255 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
256 | * - reduce the number of linked list operations | |
257 | * - reduce spinlock operations | |
258 | * | |
259 | * The limit is stored in the per-cpu structure to reduce the data cache | |
260 | * footprint. | |
261 | * | |
262 | */ | |
263 | struct array_cache { | |
264 | unsigned int avail; | |
265 | unsigned int limit; | |
266 | unsigned int batchcount; | |
267 | unsigned int touched; | |
e498be7d | 268 | spinlock_t lock; |
a737b3e2 AM |
269 | void *entry[0]; /* |
270 | * Must have this definition in here for the proper | |
271 | * alignment of array_cache. Also simplifies accessing | |
272 | * the entries. | |
273 | * [0] is for gcc 2.95. It should really be []. | |
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 */ | |
343e0d7a | 410 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
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) - | |
550 | BYTES_PER_WORD); | |
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 | ||
571 | /* | |
a737b3e2 AM |
572 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
573 | * order. | |
1da177e4 LT |
574 | */ |
575 | #if defined(CONFIG_LARGE_ALLOCS) | |
576 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
577 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
578 | #elif defined(CONFIG_MMU) | |
579 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
580 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
581 | #else | |
582 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
583 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
584 | #endif | |
585 | ||
586 | /* | |
587 | * Do not go above this order unless 0 objects fit into the slab. | |
588 | */ | |
589 | #define BREAK_GFP_ORDER_HI 1 | |
590 | #define BREAK_GFP_ORDER_LO 0 | |
591 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
592 | ||
a737b3e2 AM |
593 | /* |
594 | * Functions for storing/retrieving the cachep and or slab from the page | |
595 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
596 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 597 | */ |
065d41cb PE |
598 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
599 | { | |
600 | page->lru.next = (struct list_head *)cache; | |
601 | } | |
602 | ||
603 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
604 | { | |
d85f3385 | 605 | page = compound_head(page); |
ddc2e812 | 606 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
607 | return (struct kmem_cache *)page->lru.next; |
608 | } | |
609 | ||
610 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
611 | { | |
612 | page->lru.prev = (struct list_head *)slab; | |
613 | } | |
614 | ||
615 | static inline struct slab *page_get_slab(struct page *page) | |
616 | { | |
ddc2e812 | 617 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
618 | return (struct slab *)page->lru.prev; |
619 | } | |
1da177e4 | 620 | |
6ed5eb22 PE |
621 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
622 | { | |
b49af68f | 623 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
624 | return page_get_cache(page); |
625 | } | |
626 | ||
627 | static inline struct slab *virt_to_slab(const void *obj) | |
628 | { | |
b49af68f | 629 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
630 | return page_get_slab(page); |
631 | } | |
632 | ||
8fea4e96 PE |
633 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
634 | unsigned int idx) | |
635 | { | |
636 | return slab->s_mem + cache->buffer_size * idx; | |
637 | } | |
638 | ||
6a2d7a95 ED |
639 | /* |
640 | * We want to avoid an expensive divide : (offset / cache->buffer_size) | |
641 | * Using the fact that buffer_size is a constant for a particular cache, | |
642 | * we can replace (offset / cache->buffer_size) by | |
643 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
644 | */ | |
645 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
646 | const struct slab *slab, void *obj) | |
8fea4e96 | 647 | { |
6a2d7a95 ED |
648 | u32 offset = (obj - slab->s_mem); |
649 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
650 | } |
651 | ||
a737b3e2 AM |
652 | /* |
653 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
654 | */ | |
1da177e4 LT |
655 | struct cache_sizes malloc_sizes[] = { |
656 | #define CACHE(x) { .cs_size = (x) }, | |
657 | #include <linux/kmalloc_sizes.h> | |
658 | CACHE(ULONG_MAX) | |
659 | #undef CACHE | |
660 | }; | |
661 | EXPORT_SYMBOL(malloc_sizes); | |
662 | ||
663 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
664 | struct cache_names { | |
665 | char *name; | |
666 | char *name_dma; | |
667 | }; | |
668 | ||
669 | static struct cache_names __initdata cache_names[] = { | |
670 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
671 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 672 | {NULL,} |
1da177e4 LT |
673 | #undef CACHE |
674 | }; | |
675 | ||
676 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 677 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 678 | static struct arraycache_init initarray_generic = |
b28a02de | 679 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
680 | |
681 | /* internal cache of cache description objs */ | |
343e0d7a | 682 | static struct kmem_cache cache_cache = { |
b28a02de PE |
683 | .batchcount = 1, |
684 | .limit = BOOT_CPUCACHE_ENTRIES, | |
685 | .shared = 1, | |
343e0d7a | 686 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 687 | .name = "kmem_cache", |
1da177e4 LT |
688 | }; |
689 | ||
056c6241 RT |
690 | #define BAD_ALIEN_MAGIC 0x01020304ul |
691 | ||
f1aaee53 AV |
692 | #ifdef CONFIG_LOCKDEP |
693 | ||
694 | /* | |
695 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
696 | * for other slabs "off slab". | |
697 | * The locking for this is tricky in that it nests within the locks | |
698 | * of all other slabs in a few places; to deal with this special | |
699 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
700 | * |
701 | * We set lock class for alien array caches which are up during init. | |
702 | * The lock annotation will be lost if all cpus of a node goes down and | |
703 | * then comes back up during hotplug | |
f1aaee53 | 704 | */ |
056c6241 RT |
705 | static struct lock_class_key on_slab_l3_key; |
706 | static struct lock_class_key on_slab_alc_key; | |
707 | ||
708 | static inline void init_lock_keys(void) | |
f1aaee53 | 709 | |
f1aaee53 AV |
710 | { |
711 | int q; | |
056c6241 RT |
712 | struct cache_sizes *s = malloc_sizes; |
713 | ||
714 | while (s->cs_size != ULONG_MAX) { | |
715 | for_each_node(q) { | |
716 | struct array_cache **alc; | |
717 | int r; | |
718 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
719 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
720 | continue; | |
721 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
722 | alc = l3->alien; | |
723 | /* | |
724 | * FIXME: This check for BAD_ALIEN_MAGIC | |
725 | * should go away when common slab code is taught to | |
726 | * work even without alien caches. | |
727 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
728 | * for alloc_alien_cache, | |
729 | */ | |
730 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
731 | continue; | |
732 | for_each_node(r) { | |
733 | if (alc[r]) | |
734 | lockdep_set_class(&alc[r]->lock, | |
735 | &on_slab_alc_key); | |
736 | } | |
737 | } | |
738 | s++; | |
f1aaee53 AV |
739 | } |
740 | } | |
f1aaee53 | 741 | #else |
056c6241 | 742 | static inline void init_lock_keys(void) |
f1aaee53 AV |
743 | { |
744 | } | |
745 | #endif | |
746 | ||
8f5be20b RT |
747 | /* |
748 | * 1. Guard access to the cache-chain. | |
749 | * 2. Protect sanity of cpu_online_map against cpu hotplug events | |
750 | */ | |
fc0abb14 | 751 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
752 | static struct list_head cache_chain; |
753 | ||
1da177e4 LT |
754 | /* |
755 | * chicken and egg problem: delay the per-cpu array allocation | |
756 | * until the general caches are up. | |
757 | */ | |
758 | static enum { | |
759 | NONE, | |
e498be7d CL |
760 | PARTIAL_AC, |
761 | PARTIAL_L3, | |
1da177e4 LT |
762 | FULL |
763 | } g_cpucache_up; | |
764 | ||
39d24e64 MK |
765 | /* |
766 | * used by boot code to determine if it can use slab based allocator | |
767 | */ | |
768 | int slab_is_available(void) | |
769 | { | |
770 | return g_cpucache_up == FULL; | |
771 | } | |
772 | ||
52bad64d | 773 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 774 | |
343e0d7a | 775 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
776 | { |
777 | return cachep->array[smp_processor_id()]; | |
778 | } | |
779 | ||
a737b3e2 AM |
780 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
781 | gfp_t gfpflags) | |
1da177e4 LT |
782 | { |
783 | struct cache_sizes *csizep = malloc_sizes; | |
784 | ||
785 | #if DEBUG | |
786 | /* This happens if someone tries to call | |
b28a02de PE |
787 | * kmem_cache_create(), or __kmalloc(), before |
788 | * the generic caches are initialized. | |
789 | */ | |
c7e43c78 | 790 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 791 | #endif |
0b44f7a5 | 792 | WARN_ON_ONCE(size == 0); |
1da177e4 LT |
793 | while (size > csizep->cs_size) |
794 | csizep++; | |
795 | ||
796 | /* | |
0abf40c1 | 797 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
798 | * has cs_{dma,}cachep==NULL. Thus no special case |
799 | * for large kmalloc calls required. | |
800 | */ | |
4b51d669 | 801 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
802 | if (unlikely(gfpflags & GFP_DMA)) |
803 | return csizep->cs_dmacachep; | |
4b51d669 | 804 | #endif |
1da177e4 LT |
805 | return csizep->cs_cachep; |
806 | } | |
807 | ||
b221385b | 808 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
809 | { |
810 | return __find_general_cachep(size, gfpflags); | |
811 | } | |
97e2bde4 | 812 | |
fbaccacf | 813 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 814 | { |
fbaccacf SR |
815 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
816 | } | |
1da177e4 | 817 | |
a737b3e2 AM |
818 | /* |
819 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
820 | */ | |
fbaccacf SR |
821 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
822 | size_t align, int flags, size_t *left_over, | |
823 | unsigned int *num) | |
824 | { | |
825 | int nr_objs; | |
826 | size_t mgmt_size; | |
827 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 828 | |
fbaccacf SR |
829 | /* |
830 | * The slab management structure can be either off the slab or | |
831 | * on it. For the latter case, the memory allocated for a | |
832 | * slab is used for: | |
833 | * | |
834 | * - The struct slab | |
835 | * - One kmem_bufctl_t for each object | |
836 | * - Padding to respect alignment of @align | |
837 | * - @buffer_size bytes for each object | |
838 | * | |
839 | * If the slab management structure is off the slab, then the | |
840 | * alignment will already be calculated into the size. Because | |
841 | * the slabs are all pages aligned, the objects will be at the | |
842 | * correct alignment when allocated. | |
843 | */ | |
844 | if (flags & CFLGS_OFF_SLAB) { | |
845 | mgmt_size = 0; | |
846 | nr_objs = slab_size / buffer_size; | |
847 | ||
848 | if (nr_objs > SLAB_LIMIT) | |
849 | nr_objs = SLAB_LIMIT; | |
850 | } else { | |
851 | /* | |
852 | * Ignore padding for the initial guess. The padding | |
853 | * is at most @align-1 bytes, and @buffer_size is at | |
854 | * least @align. In the worst case, this result will | |
855 | * be one greater than the number of objects that fit | |
856 | * into the memory allocation when taking the padding | |
857 | * into account. | |
858 | */ | |
859 | nr_objs = (slab_size - sizeof(struct slab)) / | |
860 | (buffer_size + sizeof(kmem_bufctl_t)); | |
861 | ||
862 | /* | |
863 | * This calculated number will be either the right | |
864 | * amount, or one greater than what we want. | |
865 | */ | |
866 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
867 | > slab_size) | |
868 | nr_objs--; | |
869 | ||
870 | if (nr_objs > SLAB_LIMIT) | |
871 | nr_objs = SLAB_LIMIT; | |
872 | ||
873 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
874 | } | |
875 | *num = nr_objs; | |
876 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
877 | } |
878 | ||
879 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
880 | ||
a737b3e2 AM |
881 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
882 | char *msg) | |
1da177e4 LT |
883 | { |
884 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 885 | function, cachep->name, msg); |
1da177e4 LT |
886 | dump_stack(); |
887 | } | |
888 | ||
3395ee05 PM |
889 | /* |
890 | * By default on NUMA we use alien caches to stage the freeing of | |
891 | * objects allocated from other nodes. This causes massive memory | |
892 | * inefficiencies when using fake NUMA setup to split memory into a | |
893 | * large number of small nodes, so it can be disabled on the command | |
894 | * line | |
895 | */ | |
896 | ||
897 | static int use_alien_caches __read_mostly = 1; | |
898 | static int __init noaliencache_setup(char *s) | |
899 | { | |
900 | use_alien_caches = 0; | |
901 | return 1; | |
902 | } | |
903 | __setup("noaliencache", noaliencache_setup); | |
904 | ||
8fce4d8e CL |
905 | #ifdef CONFIG_NUMA |
906 | /* | |
907 | * Special reaping functions for NUMA systems called from cache_reap(). | |
908 | * These take care of doing round robin flushing of alien caches (containing | |
909 | * objects freed on different nodes from which they were allocated) and the | |
910 | * flushing of remote pcps by calling drain_node_pages. | |
911 | */ | |
912 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
913 | ||
914 | static void init_reap_node(int cpu) | |
915 | { | |
916 | int node; | |
917 | ||
918 | node = next_node(cpu_to_node(cpu), node_online_map); | |
919 | if (node == MAX_NUMNODES) | |
442295c9 | 920 | node = first_node(node_online_map); |
8fce4d8e | 921 | |
7f6b8876 | 922 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
923 | } |
924 | ||
925 | static void next_reap_node(void) | |
926 | { | |
927 | int node = __get_cpu_var(reap_node); | |
928 | ||
8fce4d8e CL |
929 | node = next_node(node, node_online_map); |
930 | if (unlikely(node >= MAX_NUMNODES)) | |
931 | node = first_node(node_online_map); | |
932 | __get_cpu_var(reap_node) = node; | |
933 | } | |
934 | ||
935 | #else | |
936 | #define init_reap_node(cpu) do { } while (0) | |
937 | #define next_reap_node(void) do { } while (0) | |
938 | #endif | |
939 | ||
1da177e4 LT |
940 | /* |
941 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
942 | * via the workqueue/eventd. | |
943 | * Add the CPU number into the expiration time to minimize the possibility of | |
944 | * the CPUs getting into lockstep and contending for the global cache chain | |
945 | * lock. | |
946 | */ | |
947 | static void __devinit start_cpu_timer(int cpu) | |
948 | { | |
52bad64d | 949 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
950 | |
951 | /* | |
952 | * When this gets called from do_initcalls via cpucache_init(), | |
953 | * init_workqueues() has already run, so keventd will be setup | |
954 | * at that time. | |
955 | */ | |
52bad64d | 956 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 957 | init_reap_node(cpu); |
65f27f38 | 958 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
959 | schedule_delayed_work_on(cpu, reap_work, |
960 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
961 | } |
962 | } | |
963 | ||
e498be7d | 964 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 965 | int batchcount) |
1da177e4 | 966 | { |
b28a02de | 967 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
968 | struct array_cache *nc = NULL; |
969 | ||
e498be7d | 970 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
971 | if (nc) { |
972 | nc->avail = 0; | |
973 | nc->limit = entries; | |
974 | nc->batchcount = batchcount; | |
975 | nc->touched = 0; | |
e498be7d | 976 | spin_lock_init(&nc->lock); |
1da177e4 LT |
977 | } |
978 | return nc; | |
979 | } | |
980 | ||
3ded175a CL |
981 | /* |
982 | * Transfer objects in one arraycache to another. | |
983 | * Locking must be handled by the caller. | |
984 | * | |
985 | * Return the number of entries transferred. | |
986 | */ | |
987 | static int transfer_objects(struct array_cache *to, | |
988 | struct array_cache *from, unsigned int max) | |
989 | { | |
990 | /* Figure out how many entries to transfer */ | |
991 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
992 | ||
993 | if (!nr) | |
994 | return 0; | |
995 | ||
996 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
997 | sizeof(void *) *nr); | |
998 | ||
999 | from->avail -= nr; | |
1000 | to->avail += nr; | |
1001 | to->touched = 1; | |
1002 | return nr; | |
1003 | } | |
1004 | ||
765c4507 CL |
1005 | #ifndef CONFIG_NUMA |
1006 | ||
1007 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1008 | #define reap_alien(cachep, l3) do { } while (0) | |
1009 | ||
1010 | static inline struct array_cache **alloc_alien_cache(int node, int limit) | |
1011 | { | |
1012 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1013 | } | |
1014 | ||
1015 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1016 | { | |
1017 | } | |
1018 | ||
1019 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1020 | { | |
1021 | return 0; | |
1022 | } | |
1023 | ||
1024 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1025 | gfp_t flags) | |
1026 | { | |
1027 | return NULL; | |
1028 | } | |
1029 | ||
8b98c169 | 1030 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1031 | gfp_t flags, int nodeid) |
1032 | { | |
1033 | return NULL; | |
1034 | } | |
1035 | ||
1036 | #else /* CONFIG_NUMA */ | |
1037 | ||
8b98c169 | 1038 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1039 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1040 | |
5295a74c | 1041 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
1042 | { |
1043 | struct array_cache **ac_ptr; | |
8ef82866 | 1044 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1045 | int i; |
1046 | ||
1047 | if (limit > 1) | |
1048 | limit = 12; | |
1049 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
1050 | if (ac_ptr) { | |
1051 | for_each_node(i) { | |
1052 | if (i == node || !node_online(i)) { | |
1053 | ac_ptr[i] = NULL; | |
1054 | continue; | |
1055 | } | |
1056 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
1057 | if (!ac_ptr[i]) { | |
b28a02de | 1058 | for (i--; i <= 0; i--) |
e498be7d CL |
1059 | kfree(ac_ptr[i]); |
1060 | kfree(ac_ptr); | |
1061 | return NULL; | |
1062 | } | |
1063 | } | |
1064 | } | |
1065 | return ac_ptr; | |
1066 | } | |
1067 | ||
5295a74c | 1068 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1069 | { |
1070 | int i; | |
1071 | ||
1072 | if (!ac_ptr) | |
1073 | return; | |
e498be7d | 1074 | for_each_node(i) |
b28a02de | 1075 | kfree(ac_ptr[i]); |
e498be7d CL |
1076 | kfree(ac_ptr); |
1077 | } | |
1078 | ||
343e0d7a | 1079 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1080 | struct array_cache *ac, int node) |
e498be7d CL |
1081 | { |
1082 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1083 | ||
1084 | if (ac->avail) { | |
1085 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1086 | /* |
1087 | * Stuff objects into the remote nodes shared array first. | |
1088 | * That way we could avoid the overhead of putting the objects | |
1089 | * into the free lists and getting them back later. | |
1090 | */ | |
693f7d36 JS |
1091 | if (rl3->shared) |
1092 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1093 | |
ff69416e | 1094 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1095 | ac->avail = 0; |
1096 | spin_unlock(&rl3->list_lock); | |
1097 | } | |
1098 | } | |
1099 | ||
8fce4d8e CL |
1100 | /* |
1101 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1102 | */ | |
1103 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1104 | { | |
1105 | int node = __get_cpu_var(reap_node); | |
1106 | ||
1107 | if (l3->alien) { | |
1108 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1109 | |
1110 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1111 | __drain_alien_cache(cachep, ac, node); |
1112 | spin_unlock_irq(&ac->lock); | |
1113 | } | |
1114 | } | |
1115 | } | |
1116 | ||
a737b3e2 AM |
1117 | static void drain_alien_cache(struct kmem_cache *cachep, |
1118 | struct array_cache **alien) | |
e498be7d | 1119 | { |
b28a02de | 1120 | int i = 0; |
e498be7d CL |
1121 | struct array_cache *ac; |
1122 | unsigned long flags; | |
1123 | ||
1124 | for_each_online_node(i) { | |
4484ebf1 | 1125 | ac = alien[i]; |
e498be7d CL |
1126 | if (ac) { |
1127 | spin_lock_irqsave(&ac->lock, flags); | |
1128 | __drain_alien_cache(cachep, ac, i); | |
1129 | spin_unlock_irqrestore(&ac->lock, flags); | |
1130 | } | |
1131 | } | |
1132 | } | |
729bd0b7 | 1133 | |
873623df | 1134 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1135 | { |
1136 | struct slab *slabp = virt_to_slab(objp); | |
1137 | int nodeid = slabp->nodeid; | |
1138 | struct kmem_list3 *l3; | |
1139 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1140 | int node; |
1141 | ||
1142 | node = numa_node_id(); | |
729bd0b7 PE |
1143 | |
1144 | /* | |
1145 | * Make sure we are not freeing a object from another node to the array | |
1146 | * cache on this cpu. | |
1147 | */ | |
62918a03 | 1148 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1149 | return 0; |
1150 | ||
1ca4cb24 | 1151 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1152 | STATS_INC_NODEFREES(cachep); |
1153 | if (l3->alien && l3->alien[nodeid]) { | |
1154 | alien = l3->alien[nodeid]; | |
873623df | 1155 | spin_lock(&alien->lock); |
729bd0b7 PE |
1156 | if (unlikely(alien->avail == alien->limit)) { |
1157 | STATS_INC_ACOVERFLOW(cachep); | |
1158 | __drain_alien_cache(cachep, alien, nodeid); | |
1159 | } | |
1160 | alien->entry[alien->avail++] = objp; | |
1161 | spin_unlock(&alien->lock); | |
1162 | } else { | |
1163 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1164 | free_block(cachep, &objp, 1, nodeid); | |
1165 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1166 | } | |
1167 | return 1; | |
1168 | } | |
e498be7d CL |
1169 | #endif |
1170 | ||
8c78f307 | 1171 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1172 | unsigned long action, void *hcpu) |
1da177e4 LT |
1173 | { |
1174 | long cpu = (long)hcpu; | |
343e0d7a | 1175 | struct kmem_cache *cachep; |
e498be7d CL |
1176 | struct kmem_list3 *l3 = NULL; |
1177 | int node = cpu_to_node(cpu); | |
1178 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1179 | |
1180 | switch (action) { | |
38c3bd96 | 1181 | case CPU_LOCK_ACQUIRE: |
fc0abb14 | 1182 | mutex_lock(&cache_chain_mutex); |
38c3bd96 HC |
1183 | break; |
1184 | case CPU_UP_PREPARE: | |
8bb78442 | 1185 | case CPU_UP_PREPARE_FROZEN: |
a737b3e2 AM |
1186 | /* |
1187 | * We need to do this right in the beginning since | |
e498be7d CL |
1188 | * alloc_arraycache's are going to use this list. |
1189 | * kmalloc_node allows us to add the slab to the right | |
1190 | * kmem_list3 and not this cpu's kmem_list3 | |
1191 | */ | |
1192 | ||
1da177e4 | 1193 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1194 | /* |
1195 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1196 | * begin anything. Make sure some other cpu on this |
1197 | * node has not already allocated this | |
1198 | */ | |
1199 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1200 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1201 | if (!l3) | |
e498be7d CL |
1202 | goto bad; |
1203 | kmem_list3_init(l3); | |
1204 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1205 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1206 | |
4484ebf1 RT |
1207 | /* |
1208 | * The l3s don't come and go as CPUs come and | |
1209 | * go. cache_chain_mutex is sufficient | |
1210 | * protection here. | |
1211 | */ | |
e498be7d CL |
1212 | cachep->nodelists[node] = l3; |
1213 | } | |
1da177e4 | 1214 | |
e498be7d CL |
1215 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1216 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1217 | (1 + nr_cpus_node(node)) * |
1218 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1219 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1220 | } | |
1221 | ||
a737b3e2 AM |
1222 | /* |
1223 | * Now we can go ahead with allocating the shared arrays and | |
1224 | * array caches | |
1225 | */ | |
e498be7d | 1226 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1227 | struct array_cache *nc; |
63109846 | 1228 | struct array_cache *shared = NULL; |
3395ee05 | 1229 | struct array_cache **alien = NULL; |
cd105df4 | 1230 | |
e498be7d | 1231 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1232 | cachep->batchcount); |
1da177e4 LT |
1233 | if (!nc) |
1234 | goto bad; | |
63109846 ED |
1235 | if (cachep->shared) { |
1236 | shared = alloc_arraycache(node, | |
4484ebf1 RT |
1237 | cachep->shared * cachep->batchcount, |
1238 | 0xbaadf00d); | |
63109846 ED |
1239 | if (!shared) |
1240 | goto bad; | |
1241 | } | |
3395ee05 PM |
1242 | if (use_alien_caches) { |
1243 | alien = alloc_alien_cache(node, cachep->limit); | |
1244 | if (!alien) | |
1245 | goto bad; | |
1246 | } | |
1da177e4 | 1247 | cachep->array[cpu] = nc; |
e498be7d CL |
1248 | l3 = cachep->nodelists[node]; |
1249 | BUG_ON(!l3); | |
e498be7d | 1250 | |
4484ebf1 RT |
1251 | spin_lock_irq(&l3->list_lock); |
1252 | if (!l3->shared) { | |
1253 | /* | |
1254 | * We are serialised from CPU_DEAD or | |
1255 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1256 | */ | |
1257 | l3->shared = shared; | |
1258 | shared = NULL; | |
e498be7d | 1259 | } |
4484ebf1 RT |
1260 | #ifdef CONFIG_NUMA |
1261 | if (!l3->alien) { | |
1262 | l3->alien = alien; | |
1263 | alien = NULL; | |
1264 | } | |
1265 | #endif | |
1266 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1267 | kfree(shared); |
1268 | free_alien_cache(alien); | |
1da177e4 | 1269 | } |
1da177e4 LT |
1270 | break; |
1271 | case CPU_ONLINE: | |
8bb78442 | 1272 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1273 | start_cpu_timer(cpu); |
1274 | break; | |
1275 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1276 | case CPU_DOWN_PREPARE: |
8bb78442 | 1277 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1278 | /* |
1279 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1280 | * held so that if cache_reap() is invoked it cannot do | |
1281 | * anything expensive but will only modify reap_work | |
1282 | * and reschedule the timer. | |
1283 | */ | |
1284 | cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); | |
1285 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1286 | per_cpu(reap_work, cpu).work.func = NULL; | |
1287 | break; | |
1288 | case CPU_DOWN_FAILED: | |
8bb78442 | 1289 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1290 | start_cpu_timer(cpu); |
1291 | break; | |
1da177e4 | 1292 | case CPU_DEAD: |
8bb78442 | 1293 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1294 | /* |
1295 | * Even if all the cpus of a node are down, we don't free the | |
1296 | * kmem_list3 of any cache. This to avoid a race between | |
1297 | * cpu_down, and a kmalloc allocation from another cpu for | |
1298 | * memory from the node of the cpu going down. The list3 | |
1299 | * structure is usually allocated from kmem_cache_create() and | |
1300 | * gets destroyed at kmem_cache_destroy(). | |
1301 | */ | |
1da177e4 | 1302 | /* fall thru */ |
8f5be20b | 1303 | #endif |
1da177e4 | 1304 | case CPU_UP_CANCELED: |
8bb78442 | 1305 | case CPU_UP_CANCELED_FROZEN: |
1da177e4 LT |
1306 | list_for_each_entry(cachep, &cache_chain, next) { |
1307 | struct array_cache *nc; | |
4484ebf1 RT |
1308 | struct array_cache *shared; |
1309 | struct array_cache **alien; | |
e498be7d | 1310 | cpumask_t mask; |
1da177e4 | 1311 | |
e498be7d | 1312 | mask = node_to_cpumask(node); |
1da177e4 LT |
1313 | /* cpu is dead; no one can alloc from it. */ |
1314 | nc = cachep->array[cpu]; | |
1315 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1316 | l3 = cachep->nodelists[node]; |
1317 | ||
1318 | if (!l3) | |
4484ebf1 | 1319 | goto free_array_cache; |
e498be7d | 1320 | |
ca3b9b91 | 1321 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1322 | |
1323 | /* Free limit for this kmem_list3 */ | |
1324 | l3->free_limit -= cachep->batchcount; | |
1325 | if (nc) | |
ff69416e | 1326 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1327 | |
1328 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1329 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1330 | goto free_array_cache; |
b28a02de | 1331 | } |
e498be7d | 1332 | |
4484ebf1 RT |
1333 | shared = l3->shared; |
1334 | if (shared) { | |
63109846 ED |
1335 | free_block(cachep, shared->entry, |
1336 | shared->avail, node); | |
e498be7d CL |
1337 | l3->shared = NULL; |
1338 | } | |
e498be7d | 1339 | |
4484ebf1 RT |
1340 | alien = l3->alien; |
1341 | l3->alien = NULL; | |
1342 | ||
1343 | spin_unlock_irq(&l3->list_lock); | |
1344 | ||
1345 | kfree(shared); | |
1346 | if (alien) { | |
1347 | drain_alien_cache(cachep, alien); | |
1348 | free_alien_cache(alien); | |
e498be7d | 1349 | } |
4484ebf1 | 1350 | free_array_cache: |
1da177e4 LT |
1351 | kfree(nc); |
1352 | } | |
4484ebf1 RT |
1353 | /* |
1354 | * In the previous loop, all the objects were freed to | |
1355 | * the respective cache's slabs, now we can go ahead and | |
1356 | * shrink each nodelist to its limit. | |
1357 | */ | |
1358 | list_for_each_entry(cachep, &cache_chain, next) { | |
1359 | l3 = cachep->nodelists[node]; | |
1360 | if (!l3) | |
1361 | continue; | |
ed11d9eb | 1362 | drain_freelist(cachep, l3, l3->free_objects); |
4484ebf1 | 1363 | } |
38c3bd96 HC |
1364 | break; |
1365 | case CPU_LOCK_RELEASE: | |
fc0abb14 | 1366 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1367 | break; |
1da177e4 LT |
1368 | } |
1369 | return NOTIFY_OK; | |
a737b3e2 | 1370 | bad: |
1da177e4 LT |
1371 | return NOTIFY_BAD; |
1372 | } | |
1373 | ||
74b85f37 CS |
1374 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1375 | &cpuup_callback, NULL, 0 | |
1376 | }; | |
1da177e4 | 1377 | |
e498be7d CL |
1378 | /* |
1379 | * swap the static kmem_list3 with kmalloced memory | |
1380 | */ | |
a737b3e2 AM |
1381 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1382 | int nodeid) | |
e498be7d CL |
1383 | { |
1384 | struct kmem_list3 *ptr; | |
1385 | ||
e498be7d CL |
1386 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); |
1387 | BUG_ON(!ptr); | |
1388 | ||
1389 | local_irq_disable(); | |
1390 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1391 | /* |
1392 | * Do not assume that spinlocks can be initialized via memcpy: | |
1393 | */ | |
1394 | spin_lock_init(&ptr->list_lock); | |
1395 | ||
e498be7d CL |
1396 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1397 | cachep->nodelists[nodeid] = ptr; | |
1398 | local_irq_enable(); | |
1399 | } | |
1400 | ||
a737b3e2 AM |
1401 | /* |
1402 | * Initialisation. Called after the page allocator have been initialised and | |
1403 | * before smp_init(). | |
1da177e4 LT |
1404 | */ |
1405 | void __init kmem_cache_init(void) | |
1406 | { | |
1407 | size_t left_over; | |
1408 | struct cache_sizes *sizes; | |
1409 | struct cache_names *names; | |
e498be7d | 1410 | int i; |
07ed76b2 | 1411 | int order; |
1ca4cb24 | 1412 | int node; |
e498be7d | 1413 | |
62918a03 SS |
1414 | if (num_possible_nodes() == 1) |
1415 | use_alien_caches = 0; | |
1416 | ||
e498be7d CL |
1417 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1418 | kmem_list3_init(&initkmem_list3[i]); | |
1419 | if (i < MAX_NUMNODES) | |
1420 | cache_cache.nodelists[i] = NULL; | |
1421 | } | |
1da177e4 LT |
1422 | |
1423 | /* | |
1424 | * Fragmentation resistance on low memory - only use bigger | |
1425 | * page orders on machines with more than 32MB of memory. | |
1426 | */ | |
1427 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1428 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1429 | ||
1da177e4 LT |
1430 | /* Bootstrap is tricky, because several objects are allocated |
1431 | * from caches that do not exist yet: | |
a737b3e2 AM |
1432 | * 1) initialize the cache_cache cache: it contains the struct |
1433 | * kmem_cache structures of all caches, except cache_cache itself: | |
1434 | * cache_cache is statically allocated. | |
e498be7d CL |
1435 | * Initially an __init data area is used for the head array and the |
1436 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1437 | * array at the end of the bootstrap. | |
1da177e4 | 1438 | * 2) Create the first kmalloc cache. |
343e0d7a | 1439 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1440 | * An __init data area is used for the head array. |
1441 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1442 | * head arrays. | |
1da177e4 LT |
1443 | * 4) Replace the __init data head arrays for cache_cache and the first |
1444 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1445 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1446 | * the other cache's with kmalloc allocated memory. | |
1447 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1448 | */ |
1449 | ||
1ca4cb24 PE |
1450 | node = numa_node_id(); |
1451 | ||
1da177e4 | 1452 | /* 1) create the cache_cache */ |
1da177e4 LT |
1453 | INIT_LIST_HEAD(&cache_chain); |
1454 | list_add(&cache_cache.next, &cache_chain); | |
1455 | cache_cache.colour_off = cache_line_size(); | |
1456 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
1ca4cb24 | 1457 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1458 | |
8da3430d ED |
1459 | /* |
1460 | * struct kmem_cache size depends on nr_node_ids, which | |
1461 | * can be less than MAX_NUMNODES. | |
1462 | */ | |
1463 | cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + | |
1464 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1465 | #if DEBUG | |
1466 | cache_cache.obj_size = cache_cache.buffer_size; | |
1467 | #endif | |
a737b3e2 AM |
1468 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1469 | cache_line_size()); | |
6a2d7a95 ED |
1470 | cache_cache.reciprocal_buffer_size = |
1471 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1472 | |
07ed76b2 JS |
1473 | for (order = 0; order < MAX_ORDER; order++) { |
1474 | cache_estimate(order, cache_cache.buffer_size, | |
1475 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1476 | if (cache_cache.num) | |
1477 | break; | |
1478 | } | |
40094fa6 | 1479 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1480 | cache_cache.gfporder = order; |
b28a02de | 1481 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1482 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1483 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1484 | |
1485 | /* 2+3) create the kmalloc caches */ | |
1486 | sizes = malloc_sizes; | |
1487 | names = cache_names; | |
1488 | ||
a737b3e2 AM |
1489 | /* |
1490 | * Initialize the caches that provide memory for the array cache and the | |
1491 | * kmem_list3 structures first. Without this, further allocations will | |
1492 | * bug. | |
e498be7d CL |
1493 | */ |
1494 | ||
1495 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1496 | sizes[INDEX_AC].cs_size, |
1497 | ARCH_KMALLOC_MINALIGN, | |
1498 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1499 | NULL, NULL); | |
e498be7d | 1500 | |
a737b3e2 | 1501 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1502 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1503 | kmem_cache_create(names[INDEX_L3].name, |
1504 | sizes[INDEX_L3].cs_size, | |
1505 | ARCH_KMALLOC_MINALIGN, | |
1506 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1507 | NULL, NULL); | |
1508 | } | |
e498be7d | 1509 | |
e0a42726 IM |
1510 | slab_early_init = 0; |
1511 | ||
1da177e4 | 1512 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1513 | /* |
1514 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1515 | * This should be particularly beneficial on SMP boxes, as it |
1516 | * eliminates "false sharing". | |
1517 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1518 | * allow tighter packing of the smaller caches. |
1519 | */ | |
a737b3e2 | 1520 | if (!sizes->cs_cachep) { |
e498be7d | 1521 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1522 | sizes->cs_size, |
1523 | ARCH_KMALLOC_MINALIGN, | |
1524 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1525 | NULL, NULL); | |
1526 | } | |
4b51d669 CL |
1527 | #ifdef CONFIG_ZONE_DMA |
1528 | sizes->cs_dmacachep = kmem_cache_create( | |
1529 | names->name_dma, | |
a737b3e2 AM |
1530 | sizes->cs_size, |
1531 | ARCH_KMALLOC_MINALIGN, | |
1532 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1533 | SLAB_PANIC, | |
1534 | NULL, NULL); | |
4b51d669 | 1535 | #endif |
1da177e4 LT |
1536 | sizes++; |
1537 | names++; | |
1538 | } | |
1539 | /* 4) Replace the bootstrap head arrays */ | |
1540 | { | |
2b2d5493 | 1541 | struct array_cache *ptr; |
e498be7d | 1542 | |
1da177e4 | 1543 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1544 | |
1da177e4 | 1545 | local_irq_disable(); |
9a2dba4b PE |
1546 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1547 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1548 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1549 | /* |
1550 | * Do not assume that spinlocks can be initialized via memcpy: | |
1551 | */ | |
1552 | spin_lock_init(&ptr->lock); | |
1553 | ||
1da177e4 LT |
1554 | cache_cache.array[smp_processor_id()] = ptr; |
1555 | local_irq_enable(); | |
e498be7d | 1556 | |
1da177e4 | 1557 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1558 | |
1da177e4 | 1559 | local_irq_disable(); |
9a2dba4b | 1560 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1561 | != &initarray_generic.cache); |
9a2dba4b | 1562 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1563 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1564 | /* |
1565 | * Do not assume that spinlocks can be initialized via memcpy: | |
1566 | */ | |
1567 | spin_lock_init(&ptr->lock); | |
1568 | ||
e498be7d | 1569 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1570 | ptr; |
1da177e4 LT |
1571 | local_irq_enable(); |
1572 | } | |
e498be7d CL |
1573 | /* 5) Replace the bootstrap kmem_list3's */ |
1574 | { | |
1ca4cb24 PE |
1575 | int nid; |
1576 | ||
e498be7d | 1577 | /* Replace the static kmem_list3 structures for the boot cpu */ |
1ca4cb24 | 1578 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], node); |
e498be7d | 1579 | |
1ca4cb24 | 1580 | for_each_online_node(nid) { |
e498be7d | 1581 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1582 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1583 | |
1584 | if (INDEX_AC != INDEX_L3) { | |
1585 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1586 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1587 | } |
1588 | } | |
1589 | } | |
1da177e4 | 1590 | |
e498be7d | 1591 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1592 | { |
343e0d7a | 1593 | struct kmem_cache *cachep; |
fc0abb14 | 1594 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1595 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1596 | if (enable_cpucache(cachep)) |
1597 | BUG(); | |
fc0abb14 | 1598 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1599 | } |
1600 | ||
056c6241 RT |
1601 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1602 | init_lock_keys(); | |
1603 | ||
1604 | ||
1da177e4 LT |
1605 | /* Done! */ |
1606 | g_cpucache_up = FULL; | |
1607 | ||
a737b3e2 AM |
1608 | /* |
1609 | * Register a cpu startup notifier callback that initializes | |
1610 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1611 | */ |
1612 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1613 | |
a737b3e2 AM |
1614 | /* |
1615 | * The reap timers are started later, with a module init call: That part | |
1616 | * of the kernel is not yet operational. | |
1da177e4 LT |
1617 | */ |
1618 | } | |
1619 | ||
1620 | static int __init cpucache_init(void) | |
1621 | { | |
1622 | int cpu; | |
1623 | ||
a737b3e2 AM |
1624 | /* |
1625 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1626 | */ |
e498be7d | 1627 | for_each_online_cpu(cpu) |
a737b3e2 | 1628 | start_cpu_timer(cpu); |
1da177e4 LT |
1629 | return 0; |
1630 | } | |
1da177e4 LT |
1631 | __initcall(cpucache_init); |
1632 | ||
1633 | /* | |
1634 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1635 | * | |
1636 | * If we requested dmaable memory, we will get it. Even if we | |
1637 | * did not request dmaable memory, we might get it, but that | |
1638 | * would be relatively rare and ignorable. | |
1639 | */ | |
343e0d7a | 1640 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1641 | { |
1642 | struct page *page; | |
e1b6aa6f | 1643 | int nr_pages; |
1da177e4 LT |
1644 | int i; |
1645 | ||
d6fef9da | 1646 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1647 | /* |
1648 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1649 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1650 | */ |
e1b6aa6f | 1651 | flags |= __GFP_COMP; |
d6fef9da | 1652 | #endif |
765c4507 | 1653 | |
3c517a61 | 1654 | flags |= cachep->gfpflags; |
e1b6aa6f CH |
1655 | |
1656 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1657 | if (!page) |
1658 | return NULL; | |
1da177e4 | 1659 | |
e1b6aa6f | 1660 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1661 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1662 | add_zone_page_state(page_zone(page), |
1663 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1664 | else | |
1665 | add_zone_page_state(page_zone(page), | |
1666 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1667 | for (i = 0; i < nr_pages; i++) |
1668 | __SetPageSlab(page + i); | |
1669 | return page_address(page); | |
1da177e4 LT |
1670 | } |
1671 | ||
1672 | /* | |
1673 | * Interface to system's page release. | |
1674 | */ | |
343e0d7a | 1675 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1676 | { |
b28a02de | 1677 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1678 | struct page *page = virt_to_page(addr); |
1679 | const unsigned long nr_freed = i; | |
1680 | ||
972d1a7b CL |
1681 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1682 | sub_zone_page_state(page_zone(page), | |
1683 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1684 | else | |
1685 | sub_zone_page_state(page_zone(page), | |
1686 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1687 | while (i--) { |
f205b2fe NP |
1688 | BUG_ON(!PageSlab(page)); |
1689 | __ClearPageSlab(page); | |
1da177e4 LT |
1690 | page++; |
1691 | } | |
1da177e4 LT |
1692 | if (current->reclaim_state) |
1693 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1694 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1695 | } |
1696 | ||
1697 | static void kmem_rcu_free(struct rcu_head *head) | |
1698 | { | |
b28a02de | 1699 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1700 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1701 | |
1702 | kmem_freepages(cachep, slab_rcu->addr); | |
1703 | if (OFF_SLAB(cachep)) | |
1704 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1705 | } | |
1706 | ||
1707 | #if DEBUG | |
1708 | ||
1709 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1710 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1711 | unsigned long caller) |
1da177e4 | 1712 | { |
3dafccf2 | 1713 | int size = obj_size(cachep); |
1da177e4 | 1714 | |
3dafccf2 | 1715 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1716 | |
b28a02de | 1717 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1718 | return; |
1719 | ||
b28a02de PE |
1720 | *addr++ = 0x12345678; |
1721 | *addr++ = caller; | |
1722 | *addr++ = smp_processor_id(); | |
1723 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1724 | { |
1725 | unsigned long *sptr = &caller; | |
1726 | unsigned long svalue; | |
1727 | ||
1728 | while (!kstack_end(sptr)) { | |
1729 | svalue = *sptr++; | |
1730 | if (kernel_text_address(svalue)) { | |
b28a02de | 1731 | *addr++ = svalue; |
1da177e4 LT |
1732 | size -= sizeof(unsigned long); |
1733 | if (size <= sizeof(unsigned long)) | |
1734 | break; | |
1735 | } | |
1736 | } | |
1737 | ||
1738 | } | |
b28a02de | 1739 | *addr++ = 0x87654321; |
1da177e4 LT |
1740 | } |
1741 | #endif | |
1742 | ||
343e0d7a | 1743 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1744 | { |
3dafccf2 MS |
1745 | int size = obj_size(cachep); |
1746 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1747 | |
1748 | memset(addr, val, size); | |
b28a02de | 1749 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1750 | } |
1751 | ||
1752 | static void dump_line(char *data, int offset, int limit) | |
1753 | { | |
1754 | int i; | |
aa83aa40 DJ |
1755 | unsigned char error = 0; |
1756 | int bad_count = 0; | |
1757 | ||
1da177e4 | 1758 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1759 | for (i = 0; i < limit; i++) { |
1760 | if (data[offset + i] != POISON_FREE) { | |
1761 | error = data[offset + i]; | |
1762 | bad_count++; | |
1763 | } | |
b28a02de | 1764 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1765 | } |
1da177e4 | 1766 | printk("\n"); |
aa83aa40 DJ |
1767 | |
1768 | if (bad_count == 1) { | |
1769 | error ^= POISON_FREE; | |
1770 | if (!(error & (error - 1))) { | |
1771 | printk(KERN_ERR "Single bit error detected. Probably " | |
1772 | "bad RAM.\n"); | |
1773 | #ifdef CONFIG_X86 | |
1774 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1775 | "test tool.\n"); | |
1776 | #else | |
1777 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1778 | #endif | |
1779 | } | |
1780 | } | |
1da177e4 LT |
1781 | } |
1782 | #endif | |
1783 | ||
1784 | #if DEBUG | |
1785 | ||
343e0d7a | 1786 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1787 | { |
1788 | int i, size; | |
1789 | char *realobj; | |
1790 | ||
1791 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1792 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1793 | *dbg_redzone1(cachep, objp), |
1794 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1795 | } |
1796 | ||
1797 | if (cachep->flags & SLAB_STORE_USER) { | |
1798 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1799 | *dbg_userword(cachep, objp)); |
1da177e4 | 1800 | print_symbol("(%s)", |
a737b3e2 | 1801 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1802 | printk("\n"); |
1803 | } | |
3dafccf2 MS |
1804 | realobj = (char *)objp + obj_offset(cachep); |
1805 | size = obj_size(cachep); | |
b28a02de | 1806 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1807 | int limit; |
1808 | limit = 16; | |
b28a02de PE |
1809 | if (i + limit > size) |
1810 | limit = size - i; | |
1da177e4 LT |
1811 | dump_line(realobj, i, limit); |
1812 | } | |
1813 | } | |
1814 | ||
343e0d7a | 1815 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1816 | { |
1817 | char *realobj; | |
1818 | int size, i; | |
1819 | int lines = 0; | |
1820 | ||
3dafccf2 MS |
1821 | realobj = (char *)objp + obj_offset(cachep); |
1822 | size = obj_size(cachep); | |
1da177e4 | 1823 | |
b28a02de | 1824 | for (i = 0; i < size; i++) { |
1da177e4 | 1825 | char exp = POISON_FREE; |
b28a02de | 1826 | if (i == size - 1) |
1da177e4 LT |
1827 | exp = POISON_END; |
1828 | if (realobj[i] != exp) { | |
1829 | int limit; | |
1830 | /* Mismatch ! */ | |
1831 | /* Print header */ | |
1832 | if (lines == 0) { | |
b28a02de | 1833 | printk(KERN_ERR |
e94a40c5 DH |
1834 | "Slab corruption: %s start=%p, len=%d\n", |
1835 | cachep->name, realobj, size); | |
1da177e4 LT |
1836 | print_objinfo(cachep, objp, 0); |
1837 | } | |
1838 | /* Hexdump the affected line */ | |
b28a02de | 1839 | i = (i / 16) * 16; |
1da177e4 | 1840 | limit = 16; |
b28a02de PE |
1841 | if (i + limit > size) |
1842 | limit = size - i; | |
1da177e4 LT |
1843 | dump_line(realobj, i, limit); |
1844 | i += 16; | |
1845 | lines++; | |
1846 | /* Limit to 5 lines */ | |
1847 | if (lines > 5) | |
1848 | break; | |
1849 | } | |
1850 | } | |
1851 | if (lines != 0) { | |
1852 | /* Print some data about the neighboring objects, if they | |
1853 | * exist: | |
1854 | */ | |
6ed5eb22 | 1855 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1856 | unsigned int objnr; |
1da177e4 | 1857 | |
8fea4e96 | 1858 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1859 | if (objnr) { |
8fea4e96 | 1860 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1861 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1862 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1863 | realobj, size); |
1da177e4 LT |
1864 | print_objinfo(cachep, objp, 2); |
1865 | } | |
b28a02de | 1866 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1867 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1868 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1869 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1870 | realobj, size); |
1da177e4 LT |
1871 | print_objinfo(cachep, objp, 2); |
1872 | } | |
1873 | } | |
1874 | } | |
1875 | #endif | |
1876 | ||
12dd36fa MD |
1877 | #if DEBUG |
1878 | /** | |
911851e6 RD |
1879 | * slab_destroy_objs - destroy a slab and its objects |
1880 | * @cachep: cache pointer being destroyed | |
1881 | * @slabp: slab pointer being destroyed | |
1882 | * | |
1883 | * Call the registered destructor for each object in a slab that is being | |
1884 | * destroyed. | |
1da177e4 | 1885 | */ |
343e0d7a | 1886 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1887 | { |
1da177e4 LT |
1888 | int i; |
1889 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1890 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1891 | |
1892 | if (cachep->flags & SLAB_POISON) { | |
1893 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1894 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1895 | OFF_SLAB(cachep)) | |
b28a02de | 1896 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1897 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1898 | else |
1899 | check_poison_obj(cachep, objp); | |
1900 | #else | |
1901 | check_poison_obj(cachep, objp); | |
1902 | #endif | |
1903 | } | |
1904 | if (cachep->flags & SLAB_RED_ZONE) { | |
1905 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1906 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1907 | "was overwritten"); |
1da177e4 LT |
1908 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1909 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1910 | "was overwritten"); |
1da177e4 | 1911 | } |
1da177e4 | 1912 | } |
12dd36fa | 1913 | } |
1da177e4 | 1914 | #else |
343e0d7a | 1915 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1916 | { |
12dd36fa | 1917 | } |
1da177e4 LT |
1918 | #endif |
1919 | ||
911851e6 RD |
1920 | /** |
1921 | * slab_destroy - destroy and release all objects in a slab | |
1922 | * @cachep: cache pointer being destroyed | |
1923 | * @slabp: slab pointer being destroyed | |
1924 | * | |
12dd36fa | 1925 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1926 | * Before calling the slab must have been unlinked from the cache. The |
1927 | * cache-lock is not held/needed. | |
12dd36fa | 1928 | */ |
343e0d7a | 1929 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1930 | { |
1931 | void *addr = slabp->s_mem - slabp->colouroff; | |
1932 | ||
1933 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1934 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1935 | struct slab_rcu *slab_rcu; | |
1936 | ||
b28a02de | 1937 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1938 | slab_rcu->cachep = cachep; |
1939 | slab_rcu->addr = addr; | |
1940 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1941 | } else { | |
1942 | kmem_freepages(cachep, addr); | |
873623df IM |
1943 | if (OFF_SLAB(cachep)) |
1944 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1945 | } |
1946 | } | |
1947 | ||
a737b3e2 AM |
1948 | /* |
1949 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1950 | * size of kmem_list3. | |
1951 | */ | |
a3a02be7 | 1952 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1953 | { |
1954 | int node; | |
1955 | ||
1956 | for_each_online_node(node) { | |
b28a02de | 1957 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1958 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1959 | REAPTIMEOUT_LIST3 + |
1960 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1961 | } |
1962 | } | |
1963 | ||
117f6eb1 CL |
1964 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1965 | { | |
1966 | int i; | |
1967 | struct kmem_list3 *l3; | |
1968 | ||
1969 | for_each_online_cpu(i) | |
1970 | kfree(cachep->array[i]); | |
1971 | ||
1972 | /* NUMA: free the list3 structures */ | |
1973 | for_each_online_node(i) { | |
1974 | l3 = cachep->nodelists[i]; | |
1975 | if (l3) { | |
1976 | kfree(l3->shared); | |
1977 | free_alien_cache(l3->alien); | |
1978 | kfree(l3); | |
1979 | } | |
1980 | } | |
1981 | kmem_cache_free(&cache_cache, cachep); | |
1982 | } | |
1983 | ||
1984 | ||
4d268eba | 1985 | /** |
a70773dd RD |
1986 | * calculate_slab_order - calculate size (page order) of slabs |
1987 | * @cachep: pointer to the cache that is being created | |
1988 | * @size: size of objects to be created in this cache. | |
1989 | * @align: required alignment for the objects. | |
1990 | * @flags: slab allocation flags | |
1991 | * | |
1992 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1993 | * |
1994 | * This could be made much more intelligent. For now, try to avoid using | |
1995 | * high order pages for slabs. When the gfp() functions are more friendly | |
1996 | * towards high-order requests, this should be changed. | |
1997 | */ | |
a737b3e2 | 1998 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1999 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2000 | { |
b1ab41c4 | 2001 | unsigned long offslab_limit; |
4d268eba | 2002 | size_t left_over = 0; |
9888e6fa | 2003 | int gfporder; |
4d268eba | 2004 | |
a737b3e2 | 2005 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
2006 | unsigned int num; |
2007 | size_t remainder; | |
2008 | ||
9888e6fa | 2009 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2010 | if (!num) |
2011 | continue; | |
9888e6fa | 2012 | |
b1ab41c4 IM |
2013 | if (flags & CFLGS_OFF_SLAB) { |
2014 | /* | |
2015 | * Max number of objs-per-slab for caches which | |
2016 | * use off-slab slabs. Needed to avoid a possible | |
2017 | * looping condition in cache_grow(). | |
2018 | */ | |
2019 | offslab_limit = size - sizeof(struct slab); | |
2020 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2021 | ||
2022 | if (num > offslab_limit) | |
2023 | break; | |
2024 | } | |
4d268eba | 2025 | |
9888e6fa | 2026 | /* Found something acceptable - save it away */ |
4d268eba | 2027 | cachep->num = num; |
9888e6fa | 2028 | cachep->gfporder = gfporder; |
4d268eba PE |
2029 | left_over = remainder; |
2030 | ||
f78bb8ad LT |
2031 | /* |
2032 | * A VFS-reclaimable slab tends to have most allocations | |
2033 | * as GFP_NOFS and we really don't want to have to be allocating | |
2034 | * higher-order pages when we are unable to shrink dcache. | |
2035 | */ | |
2036 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2037 | break; | |
2038 | ||
4d268eba PE |
2039 | /* |
2040 | * Large number of objects is good, but very large slabs are | |
2041 | * currently bad for the gfp()s. | |
2042 | */ | |
9888e6fa | 2043 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2044 | break; |
2045 | ||
9888e6fa LT |
2046 | /* |
2047 | * Acceptable internal fragmentation? | |
2048 | */ | |
a737b3e2 | 2049 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2050 | break; |
2051 | } | |
2052 | return left_over; | |
2053 | } | |
2054 | ||
2ed3a4ef | 2055 | static int setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 2056 | { |
2ed3a4ef CL |
2057 | if (g_cpucache_up == FULL) |
2058 | return enable_cpucache(cachep); | |
2059 | ||
f30cf7d1 PE |
2060 | if (g_cpucache_up == NONE) { |
2061 | /* | |
2062 | * Note: the first kmem_cache_create must create the cache | |
2063 | * that's used by kmalloc(24), otherwise the creation of | |
2064 | * further caches will BUG(). | |
2065 | */ | |
2066 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2067 | ||
2068 | /* | |
2069 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2070 | * the first cache, then we need to set up all its list3s, | |
2071 | * otherwise the creation of further caches will BUG(). | |
2072 | */ | |
2073 | set_up_list3s(cachep, SIZE_AC); | |
2074 | if (INDEX_AC == INDEX_L3) | |
2075 | g_cpucache_up = PARTIAL_L3; | |
2076 | else | |
2077 | g_cpucache_up = PARTIAL_AC; | |
2078 | } else { | |
2079 | cachep->array[smp_processor_id()] = | |
2080 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
2081 | ||
2082 | if (g_cpucache_up == PARTIAL_AC) { | |
2083 | set_up_list3s(cachep, SIZE_L3); | |
2084 | g_cpucache_up = PARTIAL_L3; | |
2085 | } else { | |
2086 | int node; | |
2087 | for_each_online_node(node) { | |
2088 | cachep->nodelists[node] = | |
2089 | kmalloc_node(sizeof(struct kmem_list3), | |
2090 | GFP_KERNEL, node); | |
2091 | BUG_ON(!cachep->nodelists[node]); | |
2092 | kmem_list3_init(cachep->nodelists[node]); | |
2093 | } | |
2094 | } | |
2095 | } | |
2096 | cachep->nodelists[numa_node_id()]->next_reap = | |
2097 | jiffies + REAPTIMEOUT_LIST3 + | |
2098 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2099 | ||
2100 | cpu_cache_get(cachep)->avail = 0; | |
2101 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2102 | cpu_cache_get(cachep)->batchcount = 1; | |
2103 | cpu_cache_get(cachep)->touched = 0; | |
2104 | cachep->batchcount = 1; | |
2105 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2106 | return 0; |
f30cf7d1 PE |
2107 | } |
2108 | ||
1da177e4 LT |
2109 | /** |
2110 | * kmem_cache_create - Create a cache. | |
2111 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2112 | * @size: The size of objects to be created in this cache. | |
2113 | * @align: The required alignment for the objects. | |
2114 | * @flags: SLAB flags | |
2115 | * @ctor: A constructor for the objects. | |
c59def9f | 2116 | * @dtor: A destructor for the objects (not implemented anymore). |
1da177e4 LT |
2117 | * |
2118 | * Returns a ptr to the cache on success, NULL on failure. | |
2119 | * Cannot be called within a int, but can be interrupted. | |
2120 | * The @ctor is run when new pages are allocated by the cache | |
2121 | * and the @dtor is run before the pages are handed back. | |
2122 | * | |
2123 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2124 | * the module calling this has to destroy the cache before getting unloaded. |
2125 | * | |
1da177e4 LT |
2126 | * The flags are |
2127 | * | |
2128 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2129 | * to catch references to uninitialised memory. | |
2130 | * | |
2131 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2132 | * for buffer overruns. | |
2133 | * | |
1da177e4 LT |
2134 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2135 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2136 | * as davem. | |
2137 | */ | |
343e0d7a | 2138 | struct kmem_cache * |
1da177e4 | 2139 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
2140 | unsigned long flags, |
2141 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 2142 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
2143 | { |
2144 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2145 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2146 | |
2147 | /* | |
2148 | * Sanity checks... these are all serious usage bugs. | |
2149 | */ | |
a737b3e2 | 2150 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
c59def9f | 2151 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || dtor) { |
a737b3e2 AM |
2152 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
2153 | name); | |
b28a02de PE |
2154 | BUG(); |
2155 | } | |
1da177e4 | 2156 | |
f0188f47 | 2157 | /* |
8f5be20b RT |
2158 | * We use cache_chain_mutex to ensure a consistent view of |
2159 | * cpu_online_map as well. Please see cpuup_callback | |
f0188f47 | 2160 | */ |
fc0abb14 | 2161 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2162 | |
7a7c381d | 2163 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2164 | char tmp; |
2165 | int res; | |
2166 | ||
2167 | /* | |
2168 | * This happens when the module gets unloaded and doesn't | |
2169 | * destroy its slab cache and no-one else reuses the vmalloc | |
2170 | * area of the module. Print a warning. | |
2171 | */ | |
138ae663 | 2172 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2173 | if (res) { |
b4169525 | 2174 | printk(KERN_ERR |
2175 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2176 | pc->buffer_size); |
4f12bb4f AM |
2177 | continue; |
2178 | } | |
2179 | ||
b28a02de | 2180 | if (!strcmp(pc->name, name)) { |
b4169525 | 2181 | printk(KERN_ERR |
2182 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2183 | dump_stack(); |
2184 | goto oops; | |
2185 | } | |
2186 | } | |
2187 | ||
1da177e4 LT |
2188 | #if DEBUG |
2189 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2190 | #if FORCED_DEBUG |
2191 | /* | |
2192 | * Enable redzoning and last user accounting, except for caches with | |
2193 | * large objects, if the increased size would increase the object size | |
2194 | * above the next power of two: caches with object sizes just above a | |
2195 | * power of two have a significant amount of internal fragmentation. | |
2196 | */ | |
a737b3e2 | 2197 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 2198 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2199 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2200 | flags |= SLAB_POISON; | |
2201 | #endif | |
2202 | if (flags & SLAB_DESTROY_BY_RCU) | |
2203 | BUG_ON(flags & SLAB_POISON); | |
2204 | #endif | |
1da177e4 | 2205 | /* |
a737b3e2 AM |
2206 | * Always checks flags, a caller might be expecting debug support which |
2207 | * isn't available. | |
1da177e4 | 2208 | */ |
40094fa6 | 2209 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2210 | |
a737b3e2 AM |
2211 | /* |
2212 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2213 | * unaligned accesses for some archs when redzoning is used, and makes |
2214 | * sure any on-slab bufctl's are also correctly aligned. | |
2215 | */ | |
b28a02de PE |
2216 | if (size & (BYTES_PER_WORD - 1)) { |
2217 | size += (BYTES_PER_WORD - 1); | |
2218 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2219 | } |
2220 | ||
a737b3e2 AM |
2221 | /* calculate the final buffer alignment: */ |
2222 | ||
1da177e4 LT |
2223 | /* 1) arch recommendation: can be overridden for debug */ |
2224 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2225 | /* |
2226 | * Default alignment: as specified by the arch code. Except if | |
2227 | * an object is really small, then squeeze multiple objects into | |
2228 | * one cacheline. | |
1da177e4 LT |
2229 | */ |
2230 | ralign = cache_line_size(); | |
b28a02de | 2231 | while (size <= ralign / 2) |
1da177e4 LT |
2232 | ralign /= 2; |
2233 | } else { | |
2234 | ralign = BYTES_PER_WORD; | |
2235 | } | |
ca5f9703 PE |
2236 | |
2237 | /* | |
2238 | * Redzoning and user store require word alignment. Note this will be | |
2239 | * overridden by architecture or caller mandated alignment if either | |
2240 | * is greater than BYTES_PER_WORD. | |
2241 | */ | |
2242 | if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) | |
b46b8f19 | 2243 | ralign = __alignof__(unsigned long long); |
ca5f9703 | 2244 | |
a44b56d3 | 2245 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2246 | if (ralign < ARCH_SLAB_MINALIGN) { |
2247 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2248 | } |
a44b56d3 | 2249 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2250 | if (ralign < align) { |
2251 | ralign = align; | |
1da177e4 | 2252 | } |
a44b56d3 | 2253 | /* disable debug if necessary */ |
b46b8f19 | 2254 | if (ralign > __alignof__(unsigned long long)) |
a44b56d3 | 2255 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2256 | /* |
ca5f9703 | 2257 | * 4) Store it. |
1da177e4 LT |
2258 | */ |
2259 | align = ralign; | |
2260 | ||
2261 | /* Get cache's description obj. */ | |
e94b1766 | 2262 | cachep = kmem_cache_zalloc(&cache_cache, GFP_KERNEL); |
1da177e4 | 2263 | if (!cachep) |
4f12bb4f | 2264 | goto oops; |
1da177e4 LT |
2265 | |
2266 | #if DEBUG | |
3dafccf2 | 2267 | cachep->obj_size = size; |
1da177e4 | 2268 | |
ca5f9703 PE |
2269 | /* |
2270 | * Both debugging options require word-alignment which is calculated | |
2271 | * into align above. | |
2272 | */ | |
1da177e4 | 2273 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2274 | /* add space for red zone words */ |
b46b8f19 DW |
2275 | cachep->obj_offset += sizeof(unsigned long long); |
2276 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2277 | } |
2278 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 PE |
2279 | /* user store requires one word storage behind the end of |
2280 | * the real object. | |
1da177e4 | 2281 | */ |
1da177e4 LT |
2282 | size += BYTES_PER_WORD; |
2283 | } | |
2284 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2285 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2286 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2287 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2288 | size = PAGE_SIZE; |
2289 | } | |
2290 | #endif | |
2291 | #endif | |
2292 | ||
e0a42726 IM |
2293 | /* |
2294 | * Determine if the slab management is 'on' or 'off' slab. | |
2295 | * (bootstrapping cannot cope with offslab caches so don't do | |
2296 | * it too early on.) | |
2297 | */ | |
2298 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2299 | /* |
2300 | * Size is large, assume best to place the slab management obj | |
2301 | * off-slab (should allow better packing of objs). | |
2302 | */ | |
2303 | flags |= CFLGS_OFF_SLAB; | |
2304 | ||
2305 | size = ALIGN(size, align); | |
2306 | ||
f78bb8ad | 2307 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2308 | |
2309 | if (!cachep->num) { | |
b4169525 | 2310 | printk(KERN_ERR |
2311 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2312 | kmem_cache_free(&cache_cache, cachep); |
2313 | cachep = NULL; | |
4f12bb4f | 2314 | goto oops; |
1da177e4 | 2315 | } |
b28a02de PE |
2316 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2317 | + sizeof(struct slab), align); | |
1da177e4 LT |
2318 | |
2319 | /* | |
2320 | * If the slab has been placed off-slab, and we have enough space then | |
2321 | * move it on-slab. This is at the expense of any extra colouring. | |
2322 | */ | |
2323 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2324 | flags &= ~CFLGS_OFF_SLAB; | |
2325 | left_over -= slab_size; | |
2326 | } | |
2327 | ||
2328 | if (flags & CFLGS_OFF_SLAB) { | |
2329 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2330 | slab_size = |
2331 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2332 | } |
2333 | ||
2334 | cachep->colour_off = cache_line_size(); | |
2335 | /* Offset must be a multiple of the alignment. */ | |
2336 | if (cachep->colour_off < align) | |
2337 | cachep->colour_off = align; | |
b28a02de | 2338 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2339 | cachep->slab_size = slab_size; |
2340 | cachep->flags = flags; | |
2341 | cachep->gfpflags = 0; | |
4b51d669 | 2342 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2343 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2344 | cachep->buffer_size = size; |
6a2d7a95 | 2345 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2346 | |
e5ac9c5a | 2347 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2348 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2349 | /* |
2350 | * This is a possibility for one of the malloc_sizes caches. | |
2351 | * But since we go off slab only for object size greater than | |
2352 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2353 | * this should not happen at all. | |
2354 | * But leave a BUG_ON for some lucky dude. | |
2355 | */ | |
2356 | BUG_ON(!cachep->slabp_cache); | |
2357 | } | |
1da177e4 | 2358 | cachep->ctor = ctor; |
1da177e4 LT |
2359 | cachep->name = name; |
2360 | ||
2ed3a4ef CL |
2361 | if (setup_cpu_cache(cachep)) { |
2362 | __kmem_cache_destroy(cachep); | |
2363 | cachep = NULL; | |
2364 | goto oops; | |
2365 | } | |
1da177e4 | 2366 | |
1da177e4 LT |
2367 | /* cache setup completed, link it into the list */ |
2368 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2369 | oops: |
1da177e4 LT |
2370 | if (!cachep && (flags & SLAB_PANIC)) |
2371 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2372 | name); |
fc0abb14 | 2373 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2374 | return cachep; |
2375 | } | |
2376 | EXPORT_SYMBOL(kmem_cache_create); | |
2377 | ||
2378 | #if DEBUG | |
2379 | static void check_irq_off(void) | |
2380 | { | |
2381 | BUG_ON(!irqs_disabled()); | |
2382 | } | |
2383 | ||
2384 | static void check_irq_on(void) | |
2385 | { | |
2386 | BUG_ON(irqs_disabled()); | |
2387 | } | |
2388 | ||
343e0d7a | 2389 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2390 | { |
2391 | #ifdef CONFIG_SMP | |
2392 | check_irq_off(); | |
e498be7d | 2393 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2394 | #endif |
2395 | } | |
e498be7d | 2396 | |
343e0d7a | 2397 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2398 | { |
2399 | #ifdef CONFIG_SMP | |
2400 | check_irq_off(); | |
2401 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2402 | #endif | |
2403 | } | |
2404 | ||
1da177e4 LT |
2405 | #else |
2406 | #define check_irq_off() do { } while(0) | |
2407 | #define check_irq_on() do { } while(0) | |
2408 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2409 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2410 | #endif |
2411 | ||
aab2207c CL |
2412 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2413 | struct array_cache *ac, | |
2414 | int force, int node); | |
2415 | ||
1da177e4 LT |
2416 | static void do_drain(void *arg) |
2417 | { | |
a737b3e2 | 2418 | struct kmem_cache *cachep = arg; |
1da177e4 | 2419 | struct array_cache *ac; |
ff69416e | 2420 | int node = numa_node_id(); |
1da177e4 LT |
2421 | |
2422 | check_irq_off(); | |
9a2dba4b | 2423 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2424 | spin_lock(&cachep->nodelists[node]->list_lock); |
2425 | free_block(cachep, ac->entry, ac->avail, node); | |
2426 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2427 | ac->avail = 0; |
2428 | } | |
2429 | ||
343e0d7a | 2430 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2431 | { |
e498be7d CL |
2432 | struct kmem_list3 *l3; |
2433 | int node; | |
2434 | ||
a07fa394 | 2435 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2436 | check_irq_on(); |
b28a02de | 2437 | for_each_online_node(node) { |
e498be7d | 2438 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2439 | if (l3 && l3->alien) |
2440 | drain_alien_cache(cachep, l3->alien); | |
2441 | } | |
2442 | ||
2443 | for_each_online_node(node) { | |
2444 | l3 = cachep->nodelists[node]; | |
2445 | if (l3) | |
aab2207c | 2446 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2447 | } |
1da177e4 LT |
2448 | } |
2449 | ||
ed11d9eb CL |
2450 | /* |
2451 | * Remove slabs from the list of free slabs. | |
2452 | * Specify the number of slabs to drain in tofree. | |
2453 | * | |
2454 | * Returns the actual number of slabs released. | |
2455 | */ | |
2456 | static int drain_freelist(struct kmem_cache *cache, | |
2457 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2458 | { |
ed11d9eb CL |
2459 | struct list_head *p; |
2460 | int nr_freed; | |
1da177e4 | 2461 | struct slab *slabp; |
1da177e4 | 2462 | |
ed11d9eb CL |
2463 | nr_freed = 0; |
2464 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2465 | |
ed11d9eb | 2466 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2467 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2468 | if (p == &l3->slabs_free) { |
2469 | spin_unlock_irq(&l3->list_lock); | |
2470 | goto out; | |
2471 | } | |
1da177e4 | 2472 | |
ed11d9eb | 2473 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2474 | #if DEBUG |
40094fa6 | 2475 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2476 | #endif |
2477 | list_del(&slabp->list); | |
ed11d9eb CL |
2478 | /* |
2479 | * Safe to drop the lock. The slab is no longer linked | |
2480 | * to the cache. | |
2481 | */ | |
2482 | l3->free_objects -= cache->num; | |
e498be7d | 2483 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2484 | slab_destroy(cache, slabp); |
2485 | nr_freed++; | |
1da177e4 | 2486 | } |
ed11d9eb CL |
2487 | out: |
2488 | return nr_freed; | |
1da177e4 LT |
2489 | } |
2490 | ||
8f5be20b | 2491 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2492 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2493 | { |
2494 | int ret = 0, i = 0; | |
2495 | struct kmem_list3 *l3; | |
2496 | ||
2497 | drain_cpu_caches(cachep); | |
2498 | ||
2499 | check_irq_on(); | |
2500 | for_each_online_node(i) { | |
2501 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2502 | if (!l3) |
2503 | continue; | |
2504 | ||
2505 | drain_freelist(cachep, l3, l3->free_objects); | |
2506 | ||
2507 | ret += !list_empty(&l3->slabs_full) || | |
2508 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2509 | } |
2510 | return (ret ? 1 : 0); | |
2511 | } | |
2512 | ||
1da177e4 LT |
2513 | /** |
2514 | * kmem_cache_shrink - Shrink a cache. | |
2515 | * @cachep: The cache to shrink. | |
2516 | * | |
2517 | * Releases as many slabs as possible for a cache. | |
2518 | * To help debugging, a zero exit status indicates all slabs were released. | |
2519 | */ | |
343e0d7a | 2520 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2521 | { |
8f5be20b | 2522 | int ret; |
40094fa6 | 2523 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2524 | |
8f5be20b RT |
2525 | mutex_lock(&cache_chain_mutex); |
2526 | ret = __cache_shrink(cachep); | |
2527 | mutex_unlock(&cache_chain_mutex); | |
2528 | return ret; | |
1da177e4 LT |
2529 | } |
2530 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2531 | ||
2532 | /** | |
2533 | * kmem_cache_destroy - delete a cache | |
2534 | * @cachep: the cache to destroy | |
2535 | * | |
72fd4a35 | 2536 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2537 | * |
2538 | * It is expected this function will be called by a module when it is | |
2539 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2540 | * cache being allocated each time a module is loaded and unloaded, if the | |
2541 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2542 | * | |
2543 | * The cache must be empty before calling this function. | |
2544 | * | |
2545 | * The caller must guarantee that noone will allocate memory from the cache | |
2546 | * during the kmem_cache_destroy(). | |
2547 | */ | |
133d205a | 2548 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2549 | { |
40094fa6 | 2550 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2551 | |
1da177e4 | 2552 | /* Find the cache in the chain of caches. */ |
fc0abb14 | 2553 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2554 | /* |
2555 | * the chain is never empty, cache_cache is never destroyed | |
2556 | */ | |
2557 | list_del(&cachep->next); | |
1da177e4 LT |
2558 | if (__cache_shrink(cachep)) { |
2559 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2560 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2561 | mutex_unlock(&cache_chain_mutex); |
133d205a | 2562 | return; |
1da177e4 LT |
2563 | } |
2564 | ||
2565 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2566 | synchronize_rcu(); |
1da177e4 | 2567 | |
117f6eb1 | 2568 | __kmem_cache_destroy(cachep); |
8f5be20b | 2569 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2570 | } |
2571 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2572 | ||
e5ac9c5a RT |
2573 | /* |
2574 | * Get the memory for a slab management obj. | |
2575 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2576 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2577 | * come from the same cache which is getting created because, | |
2578 | * when we are searching for an appropriate cache for these | |
2579 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2580 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2581 | * kmem_find_general_cachep till the initialization is complete. | |
2582 | * Hence we cannot have slabp_cache same as the original cache. | |
2583 | */ | |
343e0d7a | 2584 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2585 | int colour_off, gfp_t local_flags, |
2586 | int nodeid) | |
1da177e4 LT |
2587 | { |
2588 | struct slab *slabp; | |
b28a02de | 2589 | |
1da177e4 LT |
2590 | if (OFF_SLAB(cachep)) { |
2591 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2592 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
3c517a61 | 2593 | local_flags & ~GFP_THISNODE, nodeid); |
1da177e4 LT |
2594 | if (!slabp) |
2595 | return NULL; | |
2596 | } else { | |
b28a02de | 2597 | slabp = objp + colour_off; |
1da177e4 LT |
2598 | colour_off += cachep->slab_size; |
2599 | } | |
2600 | slabp->inuse = 0; | |
2601 | slabp->colouroff = colour_off; | |
b28a02de | 2602 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2603 | slabp->nodeid = nodeid; |
1da177e4 LT |
2604 | return slabp; |
2605 | } | |
2606 | ||
2607 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2608 | { | |
b28a02de | 2609 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2610 | } |
2611 | ||
343e0d7a | 2612 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2613 | struct slab *slabp) |
1da177e4 LT |
2614 | { |
2615 | int i; | |
2616 | ||
2617 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2618 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2619 | #if DEBUG |
2620 | /* need to poison the objs? */ | |
2621 | if (cachep->flags & SLAB_POISON) | |
2622 | poison_obj(cachep, objp, POISON_FREE); | |
2623 | if (cachep->flags & SLAB_STORE_USER) | |
2624 | *dbg_userword(cachep, objp) = NULL; | |
2625 | ||
2626 | if (cachep->flags & SLAB_RED_ZONE) { | |
2627 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2628 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2629 | } | |
2630 | /* | |
a737b3e2 AM |
2631 | * Constructors are not allowed to allocate memory from the same |
2632 | * cache which they are a constructor for. Otherwise, deadlock. | |
2633 | * They must also be threaded. | |
1da177e4 LT |
2634 | */ |
2635 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2636 | cachep->ctor(objp + obj_offset(cachep), cachep, |
a35afb83 | 2637 | 0); |
1da177e4 LT |
2638 | |
2639 | if (cachep->flags & SLAB_RED_ZONE) { | |
2640 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2641 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2642 | " end of an object"); |
1da177e4 LT |
2643 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2644 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2645 | " start of an object"); |
1da177e4 | 2646 | } |
a737b3e2 AM |
2647 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2648 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2649 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2650 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2651 | #else |
2652 | if (cachep->ctor) | |
a35afb83 | 2653 | cachep->ctor(objp, cachep, 0); |
1da177e4 | 2654 | #endif |
b28a02de | 2655 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2656 | } |
b28a02de | 2657 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2658 | slabp->free = 0; |
2659 | } | |
2660 | ||
343e0d7a | 2661 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2662 | { |
4b51d669 CL |
2663 | if (CONFIG_ZONE_DMA_FLAG) { |
2664 | if (flags & GFP_DMA) | |
2665 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2666 | else | |
2667 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2668 | } | |
1da177e4 LT |
2669 | } |
2670 | ||
a737b3e2 AM |
2671 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2672 | int nodeid) | |
78d382d7 | 2673 | { |
8fea4e96 | 2674 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2675 | kmem_bufctl_t next; |
2676 | ||
2677 | slabp->inuse++; | |
2678 | next = slab_bufctl(slabp)[slabp->free]; | |
2679 | #if DEBUG | |
2680 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2681 | WARN_ON(slabp->nodeid != nodeid); | |
2682 | #endif | |
2683 | slabp->free = next; | |
2684 | ||
2685 | return objp; | |
2686 | } | |
2687 | ||
a737b3e2 AM |
2688 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2689 | void *objp, int nodeid) | |
78d382d7 | 2690 | { |
8fea4e96 | 2691 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2692 | |
2693 | #if DEBUG | |
2694 | /* Verify that the slab belongs to the intended node */ | |
2695 | WARN_ON(slabp->nodeid != nodeid); | |
2696 | ||
871751e2 | 2697 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2698 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2699 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2700 | BUG(); |
2701 | } | |
2702 | #endif | |
2703 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2704 | slabp->free = objnr; | |
2705 | slabp->inuse--; | |
2706 | } | |
2707 | ||
4776874f PE |
2708 | /* |
2709 | * Map pages beginning at addr to the given cache and slab. This is required | |
2710 | * for the slab allocator to be able to lookup the cache and slab of a | |
2711 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2712 | */ | |
2713 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2714 | void *addr) | |
1da177e4 | 2715 | { |
4776874f | 2716 | int nr_pages; |
1da177e4 LT |
2717 | struct page *page; |
2718 | ||
4776874f | 2719 | page = virt_to_page(addr); |
84097518 | 2720 | |
4776874f | 2721 | nr_pages = 1; |
84097518 | 2722 | if (likely(!PageCompound(page))) |
4776874f PE |
2723 | nr_pages <<= cache->gfporder; |
2724 | ||
1da177e4 | 2725 | do { |
4776874f PE |
2726 | page_set_cache(page, cache); |
2727 | page_set_slab(page, slab); | |
1da177e4 | 2728 | page++; |
4776874f | 2729 | } while (--nr_pages); |
1da177e4 LT |
2730 | } |
2731 | ||
2732 | /* | |
2733 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2734 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2735 | */ | |
3c517a61 CL |
2736 | static int cache_grow(struct kmem_cache *cachep, |
2737 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2738 | { |
b28a02de | 2739 | struct slab *slabp; |
b28a02de PE |
2740 | size_t offset; |
2741 | gfp_t local_flags; | |
e498be7d | 2742 | struct kmem_list3 *l3; |
1da177e4 | 2743 | |
a737b3e2 AM |
2744 | /* |
2745 | * Be lazy and only check for valid flags here, keeping it out of the | |
2746 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2747 | */ |
cfce6604 | 2748 | BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK)); |
1da177e4 | 2749 | |
a06d72c1 | 2750 | local_flags = (flags & GFP_LEVEL_MASK); |
2e1217cf | 2751 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2752 | check_irq_off(); |
2e1217cf RT |
2753 | l3 = cachep->nodelists[nodeid]; |
2754 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2755 | |
2756 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2757 | offset = l3->colour_next; |
2758 | l3->colour_next++; | |
2759 | if (l3->colour_next >= cachep->colour) | |
2760 | l3->colour_next = 0; | |
2761 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2762 | |
2e1217cf | 2763 | offset *= cachep->colour_off; |
1da177e4 LT |
2764 | |
2765 | if (local_flags & __GFP_WAIT) | |
2766 | local_irq_enable(); | |
2767 | ||
2768 | /* | |
2769 | * The test for missing atomic flag is performed here, rather than | |
2770 | * the more obvious place, simply to reduce the critical path length | |
2771 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2772 | * will eventually be caught here (where it matters). | |
2773 | */ | |
2774 | kmem_flagcheck(cachep, flags); | |
2775 | ||
a737b3e2 AM |
2776 | /* |
2777 | * Get mem for the objs. Attempt to allocate a physical page from | |
2778 | * 'nodeid'. | |
e498be7d | 2779 | */ |
3c517a61 CL |
2780 | if (!objp) |
2781 | objp = kmem_getpages(cachep, flags, nodeid); | |
a737b3e2 | 2782 | if (!objp) |
1da177e4 LT |
2783 | goto failed; |
2784 | ||
2785 | /* Get slab management. */ | |
3c517a61 CL |
2786 | slabp = alloc_slabmgmt(cachep, objp, offset, |
2787 | local_flags & ~GFP_THISNODE, nodeid); | |
a737b3e2 | 2788 | if (!slabp) |
1da177e4 LT |
2789 | goto opps1; |
2790 | ||
e498be7d | 2791 | slabp->nodeid = nodeid; |
4776874f | 2792 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2793 | |
a35afb83 | 2794 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2795 | |
2796 | if (local_flags & __GFP_WAIT) | |
2797 | local_irq_disable(); | |
2798 | check_irq_off(); | |
e498be7d | 2799 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2800 | |
2801 | /* Make slab active. */ | |
e498be7d | 2802 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2803 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2804 | l3->free_objects += cachep->num; |
2805 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2806 | return 1; |
a737b3e2 | 2807 | opps1: |
1da177e4 | 2808 | kmem_freepages(cachep, objp); |
a737b3e2 | 2809 | failed: |
1da177e4 LT |
2810 | if (local_flags & __GFP_WAIT) |
2811 | local_irq_disable(); | |
2812 | return 0; | |
2813 | } | |
2814 | ||
2815 | #if DEBUG | |
2816 | ||
2817 | /* | |
2818 | * Perform extra freeing checks: | |
2819 | * - detect bad pointers. | |
2820 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2821 | */ |
2822 | static void kfree_debugcheck(const void *objp) | |
2823 | { | |
1da177e4 LT |
2824 | if (!virt_addr_valid(objp)) { |
2825 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2826 | (unsigned long)objp); |
2827 | BUG(); | |
1da177e4 | 2828 | } |
1da177e4 LT |
2829 | } |
2830 | ||
58ce1fd5 PE |
2831 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2832 | { | |
b46b8f19 | 2833 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2834 | |
2835 | redzone1 = *dbg_redzone1(cache, obj); | |
2836 | redzone2 = *dbg_redzone2(cache, obj); | |
2837 | ||
2838 | /* | |
2839 | * Redzone is ok. | |
2840 | */ | |
2841 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2842 | return; | |
2843 | ||
2844 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2845 | slab_error(cache, "double free detected"); | |
2846 | else | |
2847 | slab_error(cache, "memory outside object was overwritten"); | |
2848 | ||
b46b8f19 | 2849 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2850 | obj, redzone1, redzone2); |
2851 | } | |
2852 | ||
343e0d7a | 2853 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2854 | void *caller) |
1da177e4 LT |
2855 | { |
2856 | struct page *page; | |
2857 | unsigned int objnr; | |
2858 | struct slab *slabp; | |
2859 | ||
3dafccf2 | 2860 | objp -= obj_offset(cachep); |
1da177e4 | 2861 | kfree_debugcheck(objp); |
b49af68f | 2862 | page = virt_to_head_page(objp); |
1da177e4 | 2863 | |
065d41cb | 2864 | slabp = page_get_slab(page); |
1da177e4 LT |
2865 | |
2866 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2867 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2868 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2869 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2870 | } | |
2871 | if (cachep->flags & SLAB_STORE_USER) | |
2872 | *dbg_userword(cachep, objp) = caller; | |
2873 | ||
8fea4e96 | 2874 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2875 | |
2876 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2877 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2878 | |
871751e2 AV |
2879 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2880 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2881 | #endif | |
1da177e4 LT |
2882 | if (cachep->flags & SLAB_POISON) { |
2883 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2884 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2885 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2886 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2887 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2888 | } else { |
2889 | poison_obj(cachep, objp, POISON_FREE); | |
2890 | } | |
2891 | #else | |
2892 | poison_obj(cachep, objp, POISON_FREE); | |
2893 | #endif | |
2894 | } | |
2895 | return objp; | |
2896 | } | |
2897 | ||
343e0d7a | 2898 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2899 | { |
2900 | kmem_bufctl_t i; | |
2901 | int entries = 0; | |
b28a02de | 2902 | |
1da177e4 LT |
2903 | /* Check slab's freelist to see if this obj is there. */ |
2904 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2905 | entries++; | |
2906 | if (entries > cachep->num || i >= cachep->num) | |
2907 | goto bad; | |
2908 | } | |
2909 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2910 | bad: |
2911 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2912 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2913 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2914 | for (i = 0; |
264132bc | 2915 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2916 | i++) { |
a737b3e2 | 2917 | if (i % 16 == 0) |
1da177e4 | 2918 | printk("\n%03x:", i); |
b28a02de | 2919 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2920 | } |
2921 | printk("\n"); | |
2922 | BUG(); | |
2923 | } | |
2924 | } | |
2925 | #else | |
2926 | #define kfree_debugcheck(x) do { } while(0) | |
2927 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2928 | #define check_slabp(x,y) do { } while(0) | |
2929 | #endif | |
2930 | ||
343e0d7a | 2931 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2932 | { |
2933 | int batchcount; | |
2934 | struct kmem_list3 *l3; | |
2935 | struct array_cache *ac; | |
1ca4cb24 PE |
2936 | int node; |
2937 | ||
2938 | node = numa_node_id(); | |
1da177e4 LT |
2939 | |
2940 | check_irq_off(); | |
9a2dba4b | 2941 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2942 | retry: |
1da177e4 LT |
2943 | batchcount = ac->batchcount; |
2944 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2945 | /* |
2946 | * If there was little recent activity on this cache, then | |
2947 | * perform only a partial refill. Otherwise we could generate | |
2948 | * refill bouncing. | |
1da177e4 LT |
2949 | */ |
2950 | batchcount = BATCHREFILL_LIMIT; | |
2951 | } | |
1ca4cb24 | 2952 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2953 | |
2954 | BUG_ON(ac->avail > 0 || !l3); | |
2955 | spin_lock(&l3->list_lock); | |
1da177e4 | 2956 | |
3ded175a CL |
2957 | /* See if we can refill from the shared array */ |
2958 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2959 | goto alloc_done; | |
2960 | ||
1da177e4 LT |
2961 | while (batchcount > 0) { |
2962 | struct list_head *entry; | |
2963 | struct slab *slabp; | |
2964 | /* Get slab alloc is to come from. */ | |
2965 | entry = l3->slabs_partial.next; | |
2966 | if (entry == &l3->slabs_partial) { | |
2967 | l3->free_touched = 1; | |
2968 | entry = l3->slabs_free.next; | |
2969 | if (entry == &l3->slabs_free) | |
2970 | goto must_grow; | |
2971 | } | |
2972 | ||
2973 | slabp = list_entry(entry, struct slab, list); | |
2974 | check_slabp(cachep, slabp); | |
2975 | check_spinlock_acquired(cachep); | |
714b8171 PE |
2976 | |
2977 | /* | |
2978 | * The slab was either on partial or free list so | |
2979 | * there must be at least one object available for | |
2980 | * allocation. | |
2981 | */ | |
2982 | BUG_ON(slabp->inuse < 0 || slabp->inuse >= cachep->num); | |
2983 | ||
1da177e4 | 2984 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
2985 | STATS_INC_ALLOCED(cachep); |
2986 | STATS_INC_ACTIVE(cachep); | |
2987 | STATS_SET_HIGH(cachep); | |
2988 | ||
78d382d7 | 2989 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 2990 | node); |
1da177e4 LT |
2991 | } |
2992 | check_slabp(cachep, slabp); | |
2993 | ||
2994 | /* move slabp to correct slabp list: */ | |
2995 | list_del(&slabp->list); | |
2996 | if (slabp->free == BUFCTL_END) | |
2997 | list_add(&slabp->list, &l3->slabs_full); | |
2998 | else | |
2999 | list_add(&slabp->list, &l3->slabs_partial); | |
3000 | } | |
3001 | ||
a737b3e2 | 3002 | must_grow: |
1da177e4 | 3003 | l3->free_objects -= ac->avail; |
a737b3e2 | 3004 | alloc_done: |
e498be7d | 3005 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3006 | |
3007 | if (unlikely(!ac->avail)) { | |
3008 | int x; | |
3c517a61 | 3009 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3010 | |
a737b3e2 | 3011 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3012 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3013 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3014 | return NULL; |
3015 | ||
a737b3e2 | 3016 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3017 | goto retry; |
3018 | } | |
3019 | ac->touched = 1; | |
e498be7d | 3020 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3021 | } |
3022 | ||
a737b3e2 AM |
3023 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3024 | gfp_t flags) | |
1da177e4 LT |
3025 | { |
3026 | might_sleep_if(flags & __GFP_WAIT); | |
3027 | #if DEBUG | |
3028 | kmem_flagcheck(cachep, flags); | |
3029 | #endif | |
3030 | } | |
3031 | ||
3032 | #if DEBUG | |
a737b3e2 AM |
3033 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3034 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3035 | { |
b28a02de | 3036 | if (!objp) |
1da177e4 | 3037 | return objp; |
b28a02de | 3038 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3039 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3040 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3041 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3042 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3043 | else |
3044 | check_poison_obj(cachep, objp); | |
3045 | #else | |
3046 | check_poison_obj(cachep, objp); | |
3047 | #endif | |
3048 | poison_obj(cachep, objp, POISON_INUSE); | |
3049 | } | |
3050 | if (cachep->flags & SLAB_STORE_USER) | |
3051 | *dbg_userword(cachep, objp) = caller; | |
3052 | ||
3053 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3054 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3055 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3056 | slab_error(cachep, "double free, or memory outside" | |
3057 | " object was overwritten"); | |
b28a02de | 3058 | printk(KERN_ERR |
b46b8f19 | 3059 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3060 | objp, *dbg_redzone1(cachep, objp), |
3061 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3062 | } |
3063 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3064 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3065 | } | |
871751e2 AV |
3066 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3067 | { | |
3068 | struct slab *slabp; | |
3069 | unsigned objnr; | |
3070 | ||
b49af68f | 3071 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3072 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3073 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3074 | } | |
3075 | #endif | |
3dafccf2 | 3076 | objp += obj_offset(cachep); |
4f104934 | 3077 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
a35afb83 | 3078 | cachep->ctor(objp, cachep, 0); |
a44b56d3 KH |
3079 | #if ARCH_SLAB_MINALIGN |
3080 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3081 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3082 | objp, ARCH_SLAB_MINALIGN); | |
3083 | } | |
3084 | #endif | |
1da177e4 LT |
3085 | return objp; |
3086 | } | |
3087 | #else | |
3088 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3089 | #endif | |
3090 | ||
8a8b6502 AM |
3091 | #ifdef CONFIG_FAILSLAB |
3092 | ||
3093 | static struct failslab_attr { | |
3094 | ||
3095 | struct fault_attr attr; | |
3096 | ||
3097 | u32 ignore_gfp_wait; | |
3098 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3099 | struct dentry *ignore_gfp_wait_file; | |
3100 | #endif | |
3101 | ||
3102 | } failslab = { | |
3103 | .attr = FAULT_ATTR_INITIALIZER, | |
6b1b60f4 | 3104 | .ignore_gfp_wait = 1, |
8a8b6502 AM |
3105 | }; |
3106 | ||
3107 | static int __init setup_failslab(char *str) | |
3108 | { | |
3109 | return setup_fault_attr(&failslab.attr, str); | |
3110 | } | |
3111 | __setup("failslab=", setup_failslab); | |
3112 | ||
3113 | static int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3114 | { | |
3115 | if (cachep == &cache_cache) | |
3116 | return 0; | |
3117 | if (flags & __GFP_NOFAIL) | |
3118 | return 0; | |
3119 | if (failslab.ignore_gfp_wait && (flags & __GFP_WAIT)) | |
3120 | return 0; | |
3121 | ||
3122 | return should_fail(&failslab.attr, obj_size(cachep)); | |
3123 | } | |
3124 | ||
3125 | #ifdef CONFIG_FAULT_INJECTION_DEBUG_FS | |
3126 | ||
3127 | static int __init failslab_debugfs(void) | |
3128 | { | |
3129 | mode_t mode = S_IFREG | S_IRUSR | S_IWUSR; | |
3130 | struct dentry *dir; | |
3131 | int err; | |
3132 | ||
824ebef1 | 3133 | err = init_fault_attr_dentries(&failslab.attr, "failslab"); |
8a8b6502 AM |
3134 | if (err) |
3135 | return err; | |
3136 | dir = failslab.attr.dentries.dir; | |
3137 | ||
3138 | failslab.ignore_gfp_wait_file = | |
3139 | debugfs_create_bool("ignore-gfp-wait", mode, dir, | |
3140 | &failslab.ignore_gfp_wait); | |
3141 | ||
3142 | if (!failslab.ignore_gfp_wait_file) { | |
3143 | err = -ENOMEM; | |
3144 | debugfs_remove(failslab.ignore_gfp_wait_file); | |
3145 | cleanup_fault_attr_dentries(&failslab.attr); | |
3146 | } | |
3147 | ||
3148 | return err; | |
3149 | } | |
3150 | ||
3151 | late_initcall(failslab_debugfs); | |
3152 | ||
3153 | #endif /* CONFIG_FAULT_INJECTION_DEBUG_FS */ | |
3154 | ||
3155 | #else /* CONFIG_FAILSLAB */ | |
3156 | ||
3157 | static inline int should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3158 | { | |
3159 | return 0; | |
3160 | } | |
3161 | ||
3162 | #endif /* CONFIG_FAILSLAB */ | |
3163 | ||
343e0d7a | 3164 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3165 | { |
b28a02de | 3166 | void *objp; |
1da177e4 LT |
3167 | struct array_cache *ac; |
3168 | ||
5c382300 | 3169 | check_irq_off(); |
8a8b6502 | 3170 | |
9a2dba4b | 3171 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3172 | if (likely(ac->avail)) { |
3173 | STATS_INC_ALLOCHIT(cachep); | |
3174 | ac->touched = 1; | |
e498be7d | 3175 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3176 | } else { |
3177 | STATS_INC_ALLOCMISS(cachep); | |
3178 | objp = cache_alloc_refill(cachep, flags); | |
3179 | } | |
5c382300 AK |
3180 | return objp; |
3181 | } | |
3182 | ||
e498be7d | 3183 | #ifdef CONFIG_NUMA |
c61afb18 | 3184 | /* |
b2455396 | 3185 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3186 | * |
3187 | * If we are in_interrupt, then process context, including cpusets and | |
3188 | * mempolicy, may not apply and should not be used for allocation policy. | |
3189 | */ | |
3190 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3191 | { | |
3192 | int nid_alloc, nid_here; | |
3193 | ||
765c4507 | 3194 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3195 | return NULL; |
3196 | nid_alloc = nid_here = numa_node_id(); | |
3197 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3198 | nid_alloc = cpuset_mem_spread_node(); | |
3199 | else if (current->mempolicy) | |
3200 | nid_alloc = slab_node(current->mempolicy); | |
3201 | if (nid_alloc != nid_here) | |
8b98c169 | 3202 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3203 | return NULL; |
3204 | } | |
3205 | ||
765c4507 CL |
3206 | /* |
3207 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3208 | * certain node and fall back is permitted. First we scan all the |
3209 | * available nodelists for available objects. If that fails then we | |
3210 | * perform an allocation without specifying a node. This allows the page | |
3211 | * allocator to do its reclaim / fallback magic. We then insert the | |
3212 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3213 | */ |
8c8cc2c1 | 3214 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3215 | { |
8c8cc2c1 PE |
3216 | struct zonelist *zonelist; |
3217 | gfp_t local_flags; | |
765c4507 CL |
3218 | struct zone **z; |
3219 | void *obj = NULL; | |
3c517a61 | 3220 | int nid; |
8c8cc2c1 PE |
3221 | |
3222 | if (flags & __GFP_THISNODE) | |
3223 | return NULL; | |
3224 | ||
3225 | zonelist = &NODE_DATA(slab_node(current->mempolicy)) | |
3226 | ->node_zonelists[gfp_zone(flags)]; | |
3227 | local_flags = (flags & GFP_LEVEL_MASK); | |
765c4507 | 3228 | |
3c517a61 CL |
3229 | retry: |
3230 | /* | |
3231 | * Look through allowed nodes for objects available | |
3232 | * from existing per node queues. | |
3233 | */ | |
aedb0eb1 | 3234 | for (z = zonelist->zones; *z && !obj; z++) { |
3c517a61 | 3235 | nid = zone_to_nid(*z); |
aedb0eb1 | 3236 | |
02a0e53d | 3237 | if (cpuset_zone_allowed_hardwall(*z, flags) && |
3c517a61 CL |
3238 | cache->nodelists[nid] && |
3239 | cache->nodelists[nid]->free_objects) | |
3240 | obj = ____cache_alloc_node(cache, | |
3241 | flags | GFP_THISNODE, nid); | |
3242 | } | |
3243 | ||
cfce6604 | 3244 | if (!obj) { |
3c517a61 CL |
3245 | /* |
3246 | * This allocation will be performed within the constraints | |
3247 | * of the current cpuset / memory policy requirements. | |
3248 | * We may trigger various forms of reclaim on the allowed | |
3249 | * set and go into memory reserves if necessary. | |
3250 | */ | |
dd47ea75 CL |
3251 | if (local_flags & __GFP_WAIT) |
3252 | local_irq_enable(); | |
3253 | kmem_flagcheck(cache, flags); | |
3c517a61 | 3254 | obj = kmem_getpages(cache, flags, -1); |
dd47ea75 CL |
3255 | if (local_flags & __GFP_WAIT) |
3256 | local_irq_disable(); | |
3c517a61 CL |
3257 | if (obj) { |
3258 | /* | |
3259 | * Insert into the appropriate per node queues | |
3260 | */ | |
3261 | nid = page_to_nid(virt_to_page(obj)); | |
3262 | if (cache_grow(cache, flags, nid, obj)) { | |
3263 | obj = ____cache_alloc_node(cache, | |
3264 | flags | GFP_THISNODE, nid); | |
3265 | if (!obj) | |
3266 | /* | |
3267 | * Another processor may allocate the | |
3268 | * objects in the slab since we are | |
3269 | * not holding any locks. | |
3270 | */ | |
3271 | goto retry; | |
3272 | } else { | |
b6a60451 | 3273 | /* cache_grow already freed obj */ |
3c517a61 CL |
3274 | obj = NULL; |
3275 | } | |
3276 | } | |
aedb0eb1 | 3277 | } |
765c4507 CL |
3278 | return obj; |
3279 | } | |
3280 | ||
e498be7d CL |
3281 | /* |
3282 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3283 | */ |
8b98c169 | 3284 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3285 | int nodeid) |
e498be7d CL |
3286 | { |
3287 | struct list_head *entry; | |
b28a02de PE |
3288 | struct slab *slabp; |
3289 | struct kmem_list3 *l3; | |
3290 | void *obj; | |
b28a02de PE |
3291 | int x; |
3292 | ||
3293 | l3 = cachep->nodelists[nodeid]; | |
3294 | BUG_ON(!l3); | |
3295 | ||
a737b3e2 | 3296 | retry: |
ca3b9b91 | 3297 | check_irq_off(); |
b28a02de PE |
3298 | spin_lock(&l3->list_lock); |
3299 | entry = l3->slabs_partial.next; | |
3300 | if (entry == &l3->slabs_partial) { | |
3301 | l3->free_touched = 1; | |
3302 | entry = l3->slabs_free.next; | |
3303 | if (entry == &l3->slabs_free) | |
3304 | goto must_grow; | |
3305 | } | |
3306 | ||
3307 | slabp = list_entry(entry, struct slab, list); | |
3308 | check_spinlock_acquired_node(cachep, nodeid); | |
3309 | check_slabp(cachep, slabp); | |
3310 | ||
3311 | STATS_INC_NODEALLOCS(cachep); | |
3312 | STATS_INC_ACTIVE(cachep); | |
3313 | STATS_SET_HIGH(cachep); | |
3314 | ||
3315 | BUG_ON(slabp->inuse == cachep->num); | |
3316 | ||
78d382d7 | 3317 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3318 | check_slabp(cachep, slabp); |
3319 | l3->free_objects--; | |
3320 | /* move slabp to correct slabp list: */ | |
3321 | list_del(&slabp->list); | |
3322 | ||
a737b3e2 | 3323 | if (slabp->free == BUFCTL_END) |
b28a02de | 3324 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3325 | else |
b28a02de | 3326 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3327 | |
b28a02de PE |
3328 | spin_unlock(&l3->list_lock); |
3329 | goto done; | |
e498be7d | 3330 | |
a737b3e2 | 3331 | must_grow: |
b28a02de | 3332 | spin_unlock(&l3->list_lock); |
3c517a61 | 3333 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3334 | if (x) |
3335 | goto retry; | |
1da177e4 | 3336 | |
8c8cc2c1 | 3337 | return fallback_alloc(cachep, flags); |
e498be7d | 3338 | |
a737b3e2 | 3339 | done: |
b28a02de | 3340 | return obj; |
e498be7d | 3341 | } |
8c8cc2c1 PE |
3342 | |
3343 | /** | |
3344 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3345 | * @cachep: The cache to allocate from. | |
3346 | * @flags: See kmalloc(). | |
3347 | * @nodeid: node number of the target node. | |
3348 | * @caller: return address of caller, used for debug information | |
3349 | * | |
3350 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3351 | * node, which can improve the performance for cpu bound structures. | |
3352 | * | |
3353 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3354 | */ | |
3355 | static __always_inline void * | |
3356 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3357 | void *caller) | |
3358 | { | |
3359 | unsigned long save_flags; | |
3360 | void *ptr; | |
3361 | ||
824ebef1 AM |
3362 | if (should_failslab(cachep, flags)) |
3363 | return NULL; | |
3364 | ||
8c8cc2c1 PE |
3365 | cache_alloc_debugcheck_before(cachep, flags); |
3366 | local_irq_save(save_flags); | |
3367 | ||
3368 | if (unlikely(nodeid == -1)) | |
3369 | nodeid = numa_node_id(); | |
3370 | ||
3371 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3372 | /* Node not bootstrapped yet */ | |
3373 | ptr = fallback_alloc(cachep, flags); | |
3374 | goto out; | |
3375 | } | |
3376 | ||
3377 | if (nodeid == numa_node_id()) { | |
3378 | /* | |
3379 | * Use the locally cached objects if possible. | |
3380 | * However ____cache_alloc does not allow fallback | |
3381 | * to other nodes. It may fail while we still have | |
3382 | * objects on other nodes available. | |
3383 | */ | |
3384 | ptr = ____cache_alloc(cachep, flags); | |
3385 | if (ptr) | |
3386 | goto out; | |
3387 | } | |
3388 | /* ___cache_alloc_node can fall back to other nodes */ | |
3389 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3390 | out: | |
3391 | local_irq_restore(save_flags); | |
3392 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
3393 | ||
3394 | return ptr; | |
3395 | } | |
3396 | ||
3397 | static __always_inline void * | |
3398 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3399 | { | |
3400 | void *objp; | |
3401 | ||
3402 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3403 | objp = alternate_node_alloc(cache, flags); | |
3404 | if (objp) | |
3405 | goto out; | |
3406 | } | |
3407 | objp = ____cache_alloc(cache, flags); | |
3408 | ||
3409 | /* | |
3410 | * We may just have run out of memory on the local node. | |
3411 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3412 | */ | |
3413 | if (!objp) | |
3414 | objp = ____cache_alloc_node(cache, flags, numa_node_id()); | |
3415 | ||
3416 | out: | |
3417 | return objp; | |
3418 | } | |
3419 | #else | |
3420 | ||
3421 | static __always_inline void * | |
3422 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3423 | { | |
3424 | return ____cache_alloc(cachep, flags); | |
3425 | } | |
3426 | ||
3427 | #endif /* CONFIG_NUMA */ | |
3428 | ||
3429 | static __always_inline void * | |
3430 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3431 | { | |
3432 | unsigned long save_flags; | |
3433 | void *objp; | |
3434 | ||
824ebef1 AM |
3435 | if (should_failslab(cachep, flags)) |
3436 | return NULL; | |
3437 | ||
8c8cc2c1 PE |
3438 | cache_alloc_debugcheck_before(cachep, flags); |
3439 | local_irq_save(save_flags); | |
3440 | objp = __do_cache_alloc(cachep, flags); | |
3441 | local_irq_restore(save_flags); | |
3442 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
3443 | prefetchw(objp); | |
3444 | ||
3445 | return objp; | |
3446 | } | |
e498be7d CL |
3447 | |
3448 | /* | |
3449 | * Caller needs to acquire correct kmem_list's list_lock | |
3450 | */ | |
343e0d7a | 3451 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3452 | int node) |
1da177e4 LT |
3453 | { |
3454 | int i; | |
e498be7d | 3455 | struct kmem_list3 *l3; |
1da177e4 LT |
3456 | |
3457 | for (i = 0; i < nr_objects; i++) { | |
3458 | void *objp = objpp[i]; | |
3459 | struct slab *slabp; | |
1da177e4 | 3460 | |
6ed5eb22 | 3461 | slabp = virt_to_slab(objp); |
ff69416e | 3462 | l3 = cachep->nodelists[node]; |
1da177e4 | 3463 | list_del(&slabp->list); |
ff69416e | 3464 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3465 | check_slabp(cachep, slabp); |
78d382d7 | 3466 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3467 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3468 | l3->free_objects++; |
1da177e4 LT |
3469 | check_slabp(cachep, slabp); |
3470 | ||
3471 | /* fixup slab chains */ | |
3472 | if (slabp->inuse == 0) { | |
e498be7d CL |
3473 | if (l3->free_objects > l3->free_limit) { |
3474 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3475 | /* No need to drop any previously held |
3476 | * lock here, even if we have a off-slab slab | |
3477 | * descriptor it is guaranteed to come from | |
3478 | * a different cache, refer to comments before | |
3479 | * alloc_slabmgmt. | |
3480 | */ | |
1da177e4 LT |
3481 | slab_destroy(cachep, slabp); |
3482 | } else { | |
e498be7d | 3483 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3484 | } |
3485 | } else { | |
3486 | /* Unconditionally move a slab to the end of the | |
3487 | * partial list on free - maximum time for the | |
3488 | * other objects to be freed, too. | |
3489 | */ | |
e498be7d | 3490 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3491 | } |
3492 | } | |
3493 | } | |
3494 | ||
343e0d7a | 3495 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3496 | { |
3497 | int batchcount; | |
e498be7d | 3498 | struct kmem_list3 *l3; |
ff69416e | 3499 | int node = numa_node_id(); |
1da177e4 LT |
3500 | |
3501 | batchcount = ac->batchcount; | |
3502 | #if DEBUG | |
3503 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3504 | #endif | |
3505 | check_irq_off(); | |
ff69416e | 3506 | l3 = cachep->nodelists[node]; |
873623df | 3507 | spin_lock(&l3->list_lock); |
e498be7d CL |
3508 | if (l3->shared) { |
3509 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3510 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3511 | if (max) { |
3512 | if (batchcount > max) | |
3513 | batchcount = max; | |
e498be7d | 3514 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3515 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3516 | shared_array->avail += batchcount; |
3517 | goto free_done; | |
3518 | } | |
3519 | } | |
3520 | ||
ff69416e | 3521 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3522 | free_done: |
1da177e4 LT |
3523 | #if STATS |
3524 | { | |
3525 | int i = 0; | |
3526 | struct list_head *p; | |
3527 | ||
e498be7d CL |
3528 | p = l3->slabs_free.next; |
3529 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3530 | struct slab *slabp; |
3531 | ||
3532 | slabp = list_entry(p, struct slab, list); | |
3533 | BUG_ON(slabp->inuse); | |
3534 | ||
3535 | i++; | |
3536 | p = p->next; | |
3537 | } | |
3538 | STATS_SET_FREEABLE(cachep, i); | |
3539 | } | |
3540 | #endif | |
e498be7d | 3541 | spin_unlock(&l3->list_lock); |
1da177e4 | 3542 | ac->avail -= batchcount; |
a737b3e2 | 3543 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3544 | } |
3545 | ||
3546 | /* | |
a737b3e2 AM |
3547 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3548 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3549 | */ |
873623df | 3550 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3551 | { |
9a2dba4b | 3552 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3553 | |
3554 | check_irq_off(); | |
3555 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3556 | ||
62918a03 | 3557 | if (use_alien_caches && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3558 | return; |
3559 | ||
1da177e4 LT |
3560 | if (likely(ac->avail < ac->limit)) { |
3561 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3562 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3563 | return; |
3564 | } else { | |
3565 | STATS_INC_FREEMISS(cachep); | |
3566 | cache_flusharray(cachep, ac); | |
e498be7d | 3567 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3568 | } |
3569 | } | |
3570 | ||
3571 | /** | |
3572 | * kmem_cache_alloc - Allocate an object | |
3573 | * @cachep: The cache to allocate from. | |
3574 | * @flags: See kmalloc(). | |
3575 | * | |
3576 | * Allocate an object from this cache. The flags are only relevant | |
3577 | * if the cache has no available objects. | |
3578 | */ | |
343e0d7a | 3579 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3580 | { |
7fd6b141 | 3581 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3582 | } |
3583 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3584 | ||
a8c0f9a4 | 3585 | /** |
b8008b2b | 3586 | * kmem_cache_zalloc - Allocate an object. The memory is set to zero. |
a8c0f9a4 PE |
3587 | * @cache: The cache to allocate from. |
3588 | * @flags: See kmalloc(). | |
3589 | * | |
3590 | * Allocate an object from this cache and set the allocated memory to zero. | |
3591 | * The flags are only relevant if the cache has no available objects. | |
3592 | */ | |
3593 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3594 | { | |
3595 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3596 | if (ret) | |
3597 | memset(ret, 0, obj_size(cache)); | |
3598 | return ret; | |
3599 | } | |
3600 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3601 | ||
1da177e4 LT |
3602 | /** |
3603 | * kmem_ptr_validate - check if an untrusted pointer might | |
3604 | * be a slab entry. | |
3605 | * @cachep: the cache we're checking against | |
3606 | * @ptr: pointer to validate | |
3607 | * | |
3608 | * This verifies that the untrusted pointer looks sane: | |
3609 | * it is _not_ a guarantee that the pointer is actually | |
3610 | * part of the slab cache in question, but it at least | |
3611 | * validates that the pointer can be dereferenced and | |
3612 | * looks half-way sane. | |
3613 | * | |
3614 | * Currently only used for dentry validation. | |
3615 | */ | |
b7f869a2 | 3616 | int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) |
1da177e4 | 3617 | { |
b28a02de | 3618 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3619 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3620 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3621 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3622 | struct page *page; |
3623 | ||
3624 | if (unlikely(addr < min_addr)) | |
3625 | goto out; | |
3626 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3627 | goto out; | |
3628 | if (unlikely(addr & align_mask)) | |
3629 | goto out; | |
3630 | if (unlikely(!kern_addr_valid(addr))) | |
3631 | goto out; | |
3632 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3633 | goto out; | |
3634 | page = virt_to_page(ptr); | |
3635 | if (unlikely(!PageSlab(page))) | |
3636 | goto out; | |
065d41cb | 3637 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3638 | goto out; |
3639 | return 1; | |
a737b3e2 | 3640 | out: |
1da177e4 LT |
3641 | return 0; |
3642 | } | |
3643 | ||
3644 | #ifdef CONFIG_NUMA | |
8b98c169 CH |
3645 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3646 | { | |
3647 | return __cache_alloc_node(cachep, flags, nodeid, | |
3648 | __builtin_return_address(0)); | |
3649 | } | |
1da177e4 LT |
3650 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3651 | ||
8b98c169 CH |
3652 | static __always_inline void * |
3653 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3654 | { |
343e0d7a | 3655 | struct kmem_cache *cachep; |
97e2bde4 MS |
3656 | |
3657 | cachep = kmem_find_general_cachep(size, flags); | |
3658 | if (unlikely(cachep == NULL)) | |
3659 | return NULL; | |
3660 | return kmem_cache_alloc_node(cachep, flags, node); | |
3661 | } | |
8b98c169 CH |
3662 | |
3663 | #ifdef CONFIG_DEBUG_SLAB | |
3664 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3665 | { | |
3666 | return __do_kmalloc_node(size, flags, node, | |
3667 | __builtin_return_address(0)); | |
3668 | } | |
dbe5e69d | 3669 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3670 | |
3671 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
3672 | int node, void *caller) | |
3673 | { | |
3674 | return __do_kmalloc_node(size, flags, node, caller); | |
3675 | } | |
3676 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3677 | #else | |
3678 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3679 | { | |
3680 | return __do_kmalloc_node(size, flags, node, NULL); | |
3681 | } | |
3682 | EXPORT_SYMBOL(__kmalloc_node); | |
3683 | #endif /* CONFIG_DEBUG_SLAB */ | |
3684 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3685 | |
3686 | /** | |
800590f5 | 3687 | * __do_kmalloc - allocate memory |
1da177e4 | 3688 | * @size: how many bytes of memory are required. |
800590f5 | 3689 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3690 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3691 | */ |
7fd6b141 PE |
3692 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3693 | void *caller) | |
1da177e4 | 3694 | { |
343e0d7a | 3695 | struct kmem_cache *cachep; |
1da177e4 | 3696 | |
97e2bde4 MS |
3697 | /* If you want to save a few bytes .text space: replace |
3698 | * __ with kmem_. | |
3699 | * Then kmalloc uses the uninlined functions instead of the inline | |
3700 | * functions. | |
3701 | */ | |
3702 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3703 | if (unlikely(cachep == NULL)) |
3704 | return NULL; | |
7fd6b141 PE |
3705 | return __cache_alloc(cachep, flags, caller); |
3706 | } | |
3707 | ||
7fd6b141 | 3708 | |
1d2c8eea | 3709 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3710 | void *__kmalloc(size_t size, gfp_t flags) |
3711 | { | |
871751e2 | 3712 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3713 | } |
3714 | EXPORT_SYMBOL(__kmalloc); | |
3715 | ||
7fd6b141 PE |
3716 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3717 | { | |
3718 | return __do_kmalloc(size, flags, caller); | |
3719 | } | |
3720 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3721 | |
3722 | #else | |
3723 | void *__kmalloc(size_t size, gfp_t flags) | |
3724 | { | |
3725 | return __do_kmalloc(size, flags, NULL); | |
3726 | } | |
3727 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3728 | #endif |
3729 | ||
fd76bab2 PE |
3730 | /** |
3731 | * krealloc - reallocate memory. The contents will remain unchanged. | |
fd76bab2 PE |
3732 | * @p: object to reallocate memory for. |
3733 | * @new_size: how many bytes of memory are required. | |
3734 | * @flags: the type of memory to allocate. | |
3735 | * | |
3736 | * The contents of the object pointed to are preserved up to the | |
3737 | * lesser of the new and old sizes. If @p is %NULL, krealloc() | |
3738 | * behaves exactly like kmalloc(). If @size is 0 and @p is not a | |
3739 | * %NULL pointer, the object pointed to is freed. | |
3740 | */ | |
3741 | void *krealloc(const void *p, size_t new_size, gfp_t flags) | |
3742 | { | |
3743 | struct kmem_cache *cache, *new_cache; | |
3744 | void *ret; | |
3745 | ||
3746 | if (unlikely(!p)) | |
3747 | return kmalloc_track_caller(new_size, flags); | |
3748 | ||
3749 | if (unlikely(!new_size)) { | |
3750 | kfree(p); | |
3751 | return NULL; | |
3752 | } | |
3753 | ||
3754 | cache = virt_to_cache(p); | |
3755 | new_cache = __find_general_cachep(new_size, flags); | |
3756 | ||
3757 | /* | |
3758 | * If new size fits in the current cache, bail out. | |
3759 | */ | |
3760 | if (likely(cache == new_cache)) | |
3761 | return (void *)p; | |
3762 | ||
3763 | /* | |
3764 | * We are on the slow-path here so do not use __cache_alloc | |
3765 | * because it bloats kernel text. | |
3766 | */ | |
3767 | ret = kmalloc_track_caller(new_size, flags); | |
3768 | if (ret) { | |
3769 | memcpy(ret, p, min(new_size, ksize(p))); | |
3770 | kfree(p); | |
3771 | } | |
3772 | return ret; | |
3773 | } | |
3774 | EXPORT_SYMBOL(krealloc); | |
3775 | ||
1da177e4 LT |
3776 | /** |
3777 | * kmem_cache_free - Deallocate an object | |
3778 | * @cachep: The cache the allocation was from. | |
3779 | * @objp: The previously allocated object. | |
3780 | * | |
3781 | * Free an object which was previously allocated from this | |
3782 | * cache. | |
3783 | */ | |
343e0d7a | 3784 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3785 | { |
3786 | unsigned long flags; | |
3787 | ||
ddc2e812 PE |
3788 | BUG_ON(virt_to_cache(objp) != cachep); |
3789 | ||
1da177e4 | 3790 | local_irq_save(flags); |
898552c9 | 3791 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
873623df | 3792 | __cache_free(cachep, objp); |
1da177e4 LT |
3793 | local_irq_restore(flags); |
3794 | } | |
3795 | EXPORT_SYMBOL(kmem_cache_free); | |
3796 | ||
1da177e4 LT |
3797 | /** |
3798 | * kfree - free previously allocated memory | |
3799 | * @objp: pointer returned by kmalloc. | |
3800 | * | |
80e93eff PE |
3801 | * If @objp is NULL, no operation is performed. |
3802 | * | |
1da177e4 LT |
3803 | * Don't free memory not originally allocated by kmalloc() |
3804 | * or you will run into trouble. | |
3805 | */ | |
3806 | void kfree(const void *objp) | |
3807 | { | |
343e0d7a | 3808 | struct kmem_cache *c; |
1da177e4 LT |
3809 | unsigned long flags; |
3810 | ||
3811 | if (unlikely(!objp)) | |
3812 | return; | |
3813 | local_irq_save(flags); | |
3814 | kfree_debugcheck(objp); | |
6ed5eb22 | 3815 | c = virt_to_cache(objp); |
f9b8404c | 3816 | debug_check_no_locks_freed(objp, obj_size(c)); |
873623df | 3817 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3818 | local_irq_restore(flags); |
3819 | } | |
3820 | EXPORT_SYMBOL(kfree); | |
3821 | ||
343e0d7a | 3822 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3823 | { |
3dafccf2 | 3824 | return obj_size(cachep); |
1da177e4 LT |
3825 | } |
3826 | EXPORT_SYMBOL(kmem_cache_size); | |
3827 | ||
343e0d7a | 3828 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3829 | { |
3830 | return cachep->name; | |
3831 | } | |
3832 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3833 | ||
e498be7d | 3834 | /* |
0718dc2a | 3835 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3836 | */ |
343e0d7a | 3837 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3838 | { |
3839 | int node; | |
3840 | struct kmem_list3 *l3; | |
cafeb02e | 3841 | struct array_cache *new_shared; |
3395ee05 | 3842 | struct array_cache **new_alien = NULL; |
e498be7d CL |
3843 | |
3844 | for_each_online_node(node) { | |
cafeb02e | 3845 | |
3395ee05 PM |
3846 | if (use_alien_caches) { |
3847 | new_alien = alloc_alien_cache(node, cachep->limit); | |
3848 | if (!new_alien) | |
3849 | goto fail; | |
3850 | } | |
cafeb02e | 3851 | |
63109846 ED |
3852 | new_shared = NULL; |
3853 | if (cachep->shared) { | |
3854 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3855 | cachep->shared*cachep->batchcount, |
a737b3e2 | 3856 | 0xbaadf00d); |
63109846 ED |
3857 | if (!new_shared) { |
3858 | free_alien_cache(new_alien); | |
3859 | goto fail; | |
3860 | } | |
0718dc2a | 3861 | } |
cafeb02e | 3862 | |
a737b3e2 AM |
3863 | l3 = cachep->nodelists[node]; |
3864 | if (l3) { | |
cafeb02e CL |
3865 | struct array_cache *shared = l3->shared; |
3866 | ||
e498be7d CL |
3867 | spin_lock_irq(&l3->list_lock); |
3868 | ||
cafeb02e | 3869 | if (shared) |
0718dc2a CL |
3870 | free_block(cachep, shared->entry, |
3871 | shared->avail, node); | |
e498be7d | 3872 | |
cafeb02e CL |
3873 | l3->shared = new_shared; |
3874 | if (!l3->alien) { | |
e498be7d CL |
3875 | l3->alien = new_alien; |
3876 | new_alien = NULL; | |
3877 | } | |
b28a02de | 3878 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3879 | cachep->batchcount + cachep->num; |
e498be7d | 3880 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3881 | kfree(shared); |
e498be7d CL |
3882 | free_alien_cache(new_alien); |
3883 | continue; | |
3884 | } | |
a737b3e2 | 3885 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3886 | if (!l3) { |
3887 | free_alien_cache(new_alien); | |
3888 | kfree(new_shared); | |
e498be7d | 3889 | goto fail; |
0718dc2a | 3890 | } |
e498be7d CL |
3891 | |
3892 | kmem_list3_init(l3); | |
3893 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3894 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3895 | l3->shared = new_shared; |
e498be7d | 3896 | l3->alien = new_alien; |
b28a02de | 3897 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3898 | cachep->batchcount + cachep->num; |
e498be7d CL |
3899 | cachep->nodelists[node] = l3; |
3900 | } | |
cafeb02e | 3901 | return 0; |
0718dc2a | 3902 | |
a737b3e2 | 3903 | fail: |
0718dc2a CL |
3904 | if (!cachep->next.next) { |
3905 | /* Cache is not active yet. Roll back what we did */ | |
3906 | node--; | |
3907 | while (node >= 0) { | |
3908 | if (cachep->nodelists[node]) { | |
3909 | l3 = cachep->nodelists[node]; | |
3910 | ||
3911 | kfree(l3->shared); | |
3912 | free_alien_cache(l3->alien); | |
3913 | kfree(l3); | |
3914 | cachep->nodelists[node] = NULL; | |
3915 | } | |
3916 | node--; | |
3917 | } | |
3918 | } | |
cafeb02e | 3919 | return -ENOMEM; |
e498be7d CL |
3920 | } |
3921 | ||
1da177e4 | 3922 | struct ccupdate_struct { |
343e0d7a | 3923 | struct kmem_cache *cachep; |
1da177e4 LT |
3924 | struct array_cache *new[NR_CPUS]; |
3925 | }; | |
3926 | ||
3927 | static void do_ccupdate_local(void *info) | |
3928 | { | |
a737b3e2 | 3929 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3930 | struct array_cache *old; |
3931 | ||
3932 | check_irq_off(); | |
9a2dba4b | 3933 | old = cpu_cache_get(new->cachep); |
e498be7d | 3934 | |
1da177e4 LT |
3935 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3936 | new->new[smp_processor_id()] = old; | |
3937 | } | |
3938 | ||
b5d8ca7c | 3939 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3940 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3941 | int batchcount, int shared) | |
1da177e4 | 3942 | { |
d2e7b7d0 | 3943 | struct ccupdate_struct *new; |
2ed3a4ef | 3944 | int i; |
1da177e4 | 3945 | |
d2e7b7d0 SS |
3946 | new = kzalloc(sizeof(*new), GFP_KERNEL); |
3947 | if (!new) | |
3948 | return -ENOMEM; | |
3949 | ||
e498be7d | 3950 | for_each_online_cpu(i) { |
d2e7b7d0 | 3951 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
a737b3e2 | 3952 | batchcount); |
d2e7b7d0 | 3953 | if (!new->new[i]) { |
b28a02de | 3954 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3955 | kfree(new->new[i]); |
3956 | kfree(new); | |
e498be7d | 3957 | return -ENOMEM; |
1da177e4 LT |
3958 | } |
3959 | } | |
d2e7b7d0 | 3960 | new->cachep = cachep; |
1da177e4 | 3961 | |
d2e7b7d0 | 3962 | on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); |
e498be7d | 3963 | |
1da177e4 | 3964 | check_irq_on(); |
1da177e4 LT |
3965 | cachep->batchcount = batchcount; |
3966 | cachep->limit = limit; | |
e498be7d | 3967 | cachep->shared = shared; |
1da177e4 | 3968 | |
e498be7d | 3969 | for_each_online_cpu(i) { |
d2e7b7d0 | 3970 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3971 | if (!ccold) |
3972 | continue; | |
e498be7d | 3973 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3974 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3975 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3976 | kfree(ccold); |
3977 | } | |
d2e7b7d0 | 3978 | kfree(new); |
2ed3a4ef | 3979 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3980 | } |
3981 | ||
b5d8ca7c | 3982 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3983 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3984 | { |
3985 | int err; | |
3986 | int limit, shared; | |
3987 | ||
a737b3e2 AM |
3988 | /* |
3989 | * The head array serves three purposes: | |
1da177e4 LT |
3990 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3991 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3992 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3993 | * bufctl chains: array operations are cheaper. |
3994 | * The numbers are guessed, we should auto-tune as described by | |
3995 | * Bonwick. | |
3996 | */ | |
3dafccf2 | 3997 | if (cachep->buffer_size > 131072) |
1da177e4 | 3998 | limit = 1; |
3dafccf2 | 3999 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 4000 | limit = 8; |
3dafccf2 | 4001 | else if (cachep->buffer_size > 1024) |
1da177e4 | 4002 | limit = 24; |
3dafccf2 | 4003 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
4004 | limit = 54; |
4005 | else | |
4006 | limit = 120; | |
4007 | ||
a737b3e2 AM |
4008 | /* |
4009 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4010 | * allocation behaviour: Most allocs on one cpu, most free operations |
4011 | * on another cpu. For these cases, an efficient object passing between | |
4012 | * cpus is necessary. This is provided by a shared array. The array | |
4013 | * replaces Bonwick's magazine layer. | |
4014 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4015 | * to a larger limit. Thus disabled by default. | |
4016 | */ | |
4017 | shared = 0; | |
364fbb29 | 4018 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4019 | shared = 8; |
1da177e4 LT |
4020 | |
4021 | #if DEBUG | |
a737b3e2 AM |
4022 | /* |
4023 | * With debugging enabled, large batchcount lead to excessively long | |
4024 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4025 | */ |
4026 | if (limit > 32) | |
4027 | limit = 32; | |
4028 | #endif | |
b28a02de | 4029 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
4030 | if (err) |
4031 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4032 | cachep->name, -err); |
2ed3a4ef | 4033 | return err; |
1da177e4 LT |
4034 | } |
4035 | ||
1b55253a CL |
4036 | /* |
4037 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4038 | * necessary. Note that the l3 listlock also protects the array_cache |
4039 | * if drain_array() is used on the shared array. | |
1b55253a CL |
4040 | */ |
4041 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4042 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
4043 | { |
4044 | int tofree; | |
4045 | ||
1b55253a CL |
4046 | if (!ac || !ac->avail) |
4047 | return; | |
1da177e4 LT |
4048 | if (ac->touched && !force) { |
4049 | ac->touched = 0; | |
b18e7e65 | 4050 | } else { |
1b55253a | 4051 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4052 | if (ac->avail) { |
4053 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4054 | if (tofree > ac->avail) | |
4055 | tofree = (ac->avail + 1) / 2; | |
4056 | free_block(cachep, ac->entry, tofree, node); | |
4057 | ac->avail -= tofree; | |
4058 | memmove(ac->entry, &(ac->entry[tofree]), | |
4059 | sizeof(void *) * ac->avail); | |
4060 | } | |
1b55253a | 4061 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4062 | } |
4063 | } | |
4064 | ||
4065 | /** | |
4066 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4067 | * @w: work descriptor |
1da177e4 LT |
4068 | * |
4069 | * Called from workqueue/eventd every few seconds. | |
4070 | * Purpose: | |
4071 | * - clear the per-cpu caches for this CPU. | |
4072 | * - return freeable pages to the main free memory pool. | |
4073 | * | |
a737b3e2 AM |
4074 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4075 | * again on the next iteration. | |
1da177e4 | 4076 | */ |
7c5cae36 | 4077 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4078 | { |
7a7c381d | 4079 | struct kmem_cache *searchp; |
e498be7d | 4080 | struct kmem_list3 *l3; |
aab2207c | 4081 | int node = numa_node_id(); |
7c5cae36 CL |
4082 | struct delayed_work *work = |
4083 | container_of(w, struct delayed_work, work); | |
1da177e4 | 4084 | |
7c5cae36 | 4085 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4086 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4087 | goto out; |
1da177e4 | 4088 | |
7a7c381d | 4089 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4090 | check_irq_on(); |
4091 | ||
35386e3b CL |
4092 | /* |
4093 | * We only take the l3 lock if absolutely necessary and we | |
4094 | * have established with reasonable certainty that | |
4095 | * we can do some work if the lock was obtained. | |
4096 | */ | |
aab2207c | 4097 | l3 = searchp->nodelists[node]; |
35386e3b | 4098 | |
8fce4d8e | 4099 | reap_alien(searchp, l3); |
1da177e4 | 4100 | |
aab2207c | 4101 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4102 | |
35386e3b CL |
4103 | /* |
4104 | * These are racy checks but it does not matter | |
4105 | * if we skip one check or scan twice. | |
4106 | */ | |
e498be7d | 4107 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4108 | goto next; |
1da177e4 | 4109 | |
e498be7d | 4110 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4111 | |
aab2207c | 4112 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4113 | |
ed11d9eb | 4114 | if (l3->free_touched) |
e498be7d | 4115 | l3->free_touched = 0; |
ed11d9eb CL |
4116 | else { |
4117 | int freed; | |
1da177e4 | 4118 | |
ed11d9eb CL |
4119 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4120 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4121 | STATS_ADD_REAPED(searchp, freed); | |
4122 | } | |
35386e3b | 4123 | next: |
1da177e4 LT |
4124 | cond_resched(); |
4125 | } | |
4126 | check_irq_on(); | |
fc0abb14 | 4127 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4128 | next_reap_node(); |
7c5cae36 | 4129 | out: |
a737b3e2 | 4130 | /* Set up the next iteration */ |
7c5cae36 | 4131 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4132 | } |
4133 | ||
4134 | #ifdef CONFIG_PROC_FS | |
4135 | ||
85289f98 | 4136 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4137 | { |
85289f98 PE |
4138 | /* |
4139 | * Output format version, so at least we can change it | |
4140 | * without _too_ many complaints. | |
4141 | */ | |
1da177e4 | 4142 | #if STATS |
85289f98 | 4143 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4144 | #else |
85289f98 | 4145 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4146 | #endif |
85289f98 PE |
4147 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4148 | "<objperslab> <pagesperslab>"); | |
4149 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4150 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4151 | #if STATS |
85289f98 | 4152 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4153 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4154 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4155 | #endif |
85289f98 PE |
4156 | seq_putc(m, '\n'); |
4157 | } | |
4158 | ||
4159 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4160 | { | |
4161 | loff_t n = *pos; | |
4162 | struct list_head *p; | |
4163 | ||
fc0abb14 | 4164 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4165 | if (!n) |
4166 | print_slabinfo_header(m); | |
1da177e4 LT |
4167 | p = cache_chain.next; |
4168 | while (n--) { | |
4169 | p = p->next; | |
4170 | if (p == &cache_chain) | |
4171 | return NULL; | |
4172 | } | |
343e0d7a | 4173 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
4174 | } |
4175 | ||
4176 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4177 | { | |
343e0d7a | 4178 | struct kmem_cache *cachep = p; |
1da177e4 | 4179 | ++*pos; |
a737b3e2 AM |
4180 | return cachep->next.next == &cache_chain ? |
4181 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
4182 | } |
4183 | ||
4184 | static void s_stop(struct seq_file *m, void *p) | |
4185 | { | |
fc0abb14 | 4186 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4187 | } |
4188 | ||
4189 | static int s_show(struct seq_file *m, void *p) | |
4190 | { | |
343e0d7a | 4191 | struct kmem_cache *cachep = p; |
b28a02de PE |
4192 | struct slab *slabp; |
4193 | unsigned long active_objs; | |
4194 | unsigned long num_objs; | |
4195 | unsigned long active_slabs = 0; | |
4196 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4197 | const char *name; |
1da177e4 | 4198 | char *error = NULL; |
e498be7d CL |
4199 | int node; |
4200 | struct kmem_list3 *l3; | |
1da177e4 | 4201 | |
1da177e4 LT |
4202 | active_objs = 0; |
4203 | num_slabs = 0; | |
e498be7d CL |
4204 | for_each_online_node(node) { |
4205 | l3 = cachep->nodelists[node]; | |
4206 | if (!l3) | |
4207 | continue; | |
4208 | ||
ca3b9b91 RT |
4209 | check_irq_on(); |
4210 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4211 | |
7a7c381d | 4212 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4213 | if (slabp->inuse != cachep->num && !error) |
4214 | error = "slabs_full accounting error"; | |
4215 | active_objs += cachep->num; | |
4216 | active_slabs++; | |
4217 | } | |
7a7c381d | 4218 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4219 | if (slabp->inuse == cachep->num && !error) |
4220 | error = "slabs_partial inuse accounting error"; | |
4221 | if (!slabp->inuse && !error) | |
4222 | error = "slabs_partial/inuse accounting error"; | |
4223 | active_objs += slabp->inuse; | |
4224 | active_slabs++; | |
4225 | } | |
7a7c381d | 4226 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4227 | if (slabp->inuse && !error) |
4228 | error = "slabs_free/inuse accounting error"; | |
4229 | num_slabs++; | |
4230 | } | |
4231 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4232 | if (l3->shared) |
4233 | shared_avail += l3->shared->avail; | |
e498be7d | 4234 | |
ca3b9b91 | 4235 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4236 | } |
b28a02de PE |
4237 | num_slabs += active_slabs; |
4238 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4239 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4240 | error = "free_objects accounting error"; |
4241 | ||
b28a02de | 4242 | name = cachep->name; |
1da177e4 LT |
4243 | if (error) |
4244 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4245 | ||
4246 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4247 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4248 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4249 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4250 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4251 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4252 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4253 | #if STATS |
b28a02de | 4254 | { /* list3 stats */ |
1da177e4 LT |
4255 | unsigned long high = cachep->high_mark; |
4256 | unsigned long allocs = cachep->num_allocations; | |
4257 | unsigned long grown = cachep->grown; | |
4258 | unsigned long reaped = cachep->reaped; | |
4259 | unsigned long errors = cachep->errors; | |
4260 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4261 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4262 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4263 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4264 | |
e498be7d | 4265 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4266 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4267 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4268 | node_frees, overflows); |
1da177e4 LT |
4269 | } |
4270 | /* cpu stats */ | |
4271 | { | |
4272 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4273 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4274 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4275 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4276 | ||
4277 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4278 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4279 | } |
4280 | #endif | |
4281 | seq_putc(m, '\n'); | |
1da177e4 LT |
4282 | return 0; |
4283 | } | |
4284 | ||
4285 | /* | |
4286 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4287 | * | |
4288 | * Output layout: | |
4289 | * cache-name | |
4290 | * num-active-objs | |
4291 | * total-objs | |
4292 | * object size | |
4293 | * num-active-slabs | |
4294 | * total-slabs | |
4295 | * num-pages-per-slab | |
4296 | * + further values on SMP and with statistics enabled | |
4297 | */ | |
4298 | ||
15ad7cdc | 4299 | const struct seq_operations slabinfo_op = { |
b28a02de PE |
4300 | .start = s_start, |
4301 | .next = s_next, | |
4302 | .stop = s_stop, | |
4303 | .show = s_show, | |
1da177e4 LT |
4304 | }; |
4305 | ||
4306 | #define MAX_SLABINFO_WRITE 128 | |
4307 | /** | |
4308 | * slabinfo_write - Tuning for the slab allocator | |
4309 | * @file: unused | |
4310 | * @buffer: user buffer | |
4311 | * @count: data length | |
4312 | * @ppos: unused | |
4313 | */ | |
b28a02de PE |
4314 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4315 | size_t count, loff_t *ppos) | |
1da177e4 | 4316 | { |
b28a02de | 4317 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4318 | int limit, batchcount, shared, res; |
7a7c381d | 4319 | struct kmem_cache *cachep; |
b28a02de | 4320 | |
1da177e4 LT |
4321 | if (count > MAX_SLABINFO_WRITE) |
4322 | return -EINVAL; | |
4323 | if (copy_from_user(&kbuf, buffer, count)) | |
4324 | return -EFAULT; | |
b28a02de | 4325 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4326 | |
4327 | tmp = strchr(kbuf, ' '); | |
4328 | if (!tmp) | |
4329 | return -EINVAL; | |
4330 | *tmp = '\0'; | |
4331 | tmp++; | |
4332 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4333 | return -EINVAL; | |
4334 | ||
4335 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4336 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4337 | res = -EINVAL; |
7a7c381d | 4338 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4339 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4340 | if (limit < 1 || batchcount < 1 || |
4341 | batchcount > limit || shared < 0) { | |
e498be7d | 4342 | res = 0; |
1da177e4 | 4343 | } else { |
e498be7d | 4344 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4345 | batchcount, shared); |
1da177e4 LT |
4346 | } |
4347 | break; | |
4348 | } | |
4349 | } | |
fc0abb14 | 4350 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4351 | if (res >= 0) |
4352 | res = count; | |
4353 | return res; | |
4354 | } | |
871751e2 AV |
4355 | |
4356 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4357 | ||
4358 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4359 | { | |
4360 | loff_t n = *pos; | |
4361 | struct list_head *p; | |
4362 | ||
4363 | mutex_lock(&cache_chain_mutex); | |
4364 | p = cache_chain.next; | |
4365 | while (n--) { | |
4366 | p = p->next; | |
4367 | if (p == &cache_chain) | |
4368 | return NULL; | |
4369 | } | |
4370 | return list_entry(p, struct kmem_cache, next); | |
4371 | } | |
4372 | ||
4373 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4374 | { | |
4375 | unsigned long *p; | |
4376 | int l; | |
4377 | if (!v) | |
4378 | return 1; | |
4379 | l = n[1]; | |
4380 | p = n + 2; | |
4381 | while (l) { | |
4382 | int i = l/2; | |
4383 | unsigned long *q = p + 2 * i; | |
4384 | if (*q == v) { | |
4385 | q[1]++; | |
4386 | return 1; | |
4387 | } | |
4388 | if (*q > v) { | |
4389 | l = i; | |
4390 | } else { | |
4391 | p = q + 2; | |
4392 | l -= i + 1; | |
4393 | } | |
4394 | } | |
4395 | if (++n[1] == n[0]) | |
4396 | return 0; | |
4397 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4398 | p[0] = v; | |
4399 | p[1] = 1; | |
4400 | return 1; | |
4401 | } | |
4402 | ||
4403 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4404 | { | |
4405 | void *p; | |
4406 | int i; | |
4407 | if (n[0] == n[1]) | |
4408 | return; | |
4409 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4410 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4411 | continue; | |
4412 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4413 | return; | |
4414 | } | |
4415 | } | |
4416 | ||
4417 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4418 | { | |
4419 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4420 | unsigned long offset, size; |
a5c43dae | 4421 | char modname[MODULE_NAME_LEN + 1], name[KSYM_NAME_LEN + 1]; |
871751e2 | 4422 | |
a5c43dae | 4423 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4424 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4425 | if (modname[0]) |
871751e2 AV |
4426 | seq_printf(m, " [%s]", modname); |
4427 | return; | |
4428 | } | |
4429 | #endif | |
4430 | seq_printf(m, "%p", (void *)address); | |
4431 | } | |
4432 | ||
4433 | static int leaks_show(struct seq_file *m, void *p) | |
4434 | { | |
4435 | struct kmem_cache *cachep = p; | |
871751e2 AV |
4436 | struct slab *slabp; |
4437 | struct kmem_list3 *l3; | |
4438 | const char *name; | |
4439 | unsigned long *n = m->private; | |
4440 | int node; | |
4441 | int i; | |
4442 | ||
4443 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4444 | return 0; | |
4445 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4446 | return 0; | |
4447 | ||
4448 | /* OK, we can do it */ | |
4449 | ||
4450 | n[1] = 0; | |
4451 | ||
4452 | for_each_online_node(node) { | |
4453 | l3 = cachep->nodelists[node]; | |
4454 | if (!l3) | |
4455 | continue; | |
4456 | ||
4457 | check_irq_on(); | |
4458 | spin_lock_irq(&l3->list_lock); | |
4459 | ||
7a7c381d | 4460 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4461 | handle_slab(n, cachep, slabp); |
7a7c381d | 4462 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4463 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4464 | spin_unlock_irq(&l3->list_lock); |
4465 | } | |
4466 | name = cachep->name; | |
4467 | if (n[0] == n[1]) { | |
4468 | /* Increase the buffer size */ | |
4469 | mutex_unlock(&cache_chain_mutex); | |
4470 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4471 | if (!m->private) { | |
4472 | /* Too bad, we are really out */ | |
4473 | m->private = n; | |
4474 | mutex_lock(&cache_chain_mutex); | |
4475 | return -ENOMEM; | |
4476 | } | |
4477 | *(unsigned long *)m->private = n[0] * 2; | |
4478 | kfree(n); | |
4479 | mutex_lock(&cache_chain_mutex); | |
4480 | /* Now make sure this entry will be retried */ | |
4481 | m->count = m->size; | |
4482 | return 0; | |
4483 | } | |
4484 | for (i = 0; i < n[1]; i++) { | |
4485 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4486 | show_symbol(m, n[2*i+2]); | |
4487 | seq_putc(m, '\n'); | |
4488 | } | |
d2e7b7d0 | 4489 | |
871751e2 AV |
4490 | return 0; |
4491 | } | |
4492 | ||
15ad7cdc | 4493 | const struct seq_operations slabstats_op = { |
871751e2 AV |
4494 | .start = leaks_start, |
4495 | .next = s_next, | |
4496 | .stop = s_stop, | |
4497 | .show = leaks_show, | |
4498 | }; | |
4499 | #endif | |
1da177e4 LT |
4500 | #endif |
4501 | ||
00e145b6 MS |
4502 | /** |
4503 | * ksize - get the actual amount of memory allocated for a given object | |
4504 | * @objp: Pointer to the object | |
4505 | * | |
4506 | * kmalloc may internally round up allocations and return more memory | |
4507 | * than requested. ksize() can be used to determine the actual amount of | |
4508 | * memory allocated. The caller may use this additional memory, even though | |
4509 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4510 | * The caller must guarantee that objp points to a valid object previously | |
4511 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4512 | * must not be freed during the duration of the call. | |
4513 | */ | |
fd76bab2 | 4514 | size_t ksize(const void *objp) |
1da177e4 | 4515 | { |
00e145b6 MS |
4516 | if (unlikely(objp == NULL)) |
4517 | return 0; | |
1da177e4 | 4518 | |
6ed5eb22 | 4519 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4520 | } |