]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
29 | * slabs and you must pass objects with the same intializations to | |
30 | * kmem_cache_free. | |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
54 | * it's changed with a smp_call_function(). |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
fc0abb14 | 71 | * The global cache-chain is protected by the mutex 'cache_chain_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
89 | #include <linux/config.h> | |
90 | #include <linux/slab.h> | |
91 | #include <linux/mm.h> | |
c9cf5528 | 92 | #include <linux/poison.h> |
1da177e4 LT |
93 | #include <linux/swap.h> |
94 | #include <linux/cache.h> | |
95 | #include <linux/interrupt.h> | |
96 | #include <linux/init.h> | |
97 | #include <linux/compiler.h> | |
101a5001 | 98 | #include <linux/cpuset.h> |
1da177e4 LT |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> | |
101 | #include <linux/kallsyms.h> | |
102 | #include <linux/cpu.h> | |
103 | #include <linux/sysctl.h> | |
104 | #include <linux/module.h> | |
105 | #include <linux/rcupdate.h> | |
543537bd | 106 | #include <linux/string.h> |
e498be7d | 107 | #include <linux/nodemask.h> |
dc85da15 | 108 | #include <linux/mempolicy.h> |
fc0abb14 | 109 | #include <linux/mutex.h> |
e7eebaf6 | 110 | #include <linux/rtmutex.h> |
1da177e4 LT |
111 | |
112 | #include <asm/uaccess.h> | |
113 | #include <asm/cacheflush.h> | |
114 | #include <asm/tlbflush.h> | |
115 | #include <asm/page.h> | |
116 | ||
117 | /* | |
118 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
119 | * SLAB_RED_ZONE & SLAB_POISON. | |
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 | |
151 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
152 | * Note that this flag disables some debug features. | |
153 | */ | |
154 | #define ARCH_KMALLOC_MINALIGN 0 | |
155 | #endif | |
156 | ||
157 | #ifndef ARCH_SLAB_MINALIGN | |
158 | /* | |
159 | * Enforce a minimum alignment for all caches. | |
160 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
161 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
162 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
163 | * some debug features. | |
164 | */ | |
165 | #define ARCH_SLAB_MINALIGN 0 | |
166 | #endif | |
167 | ||
168 | #ifndef ARCH_KMALLOC_FLAGS | |
169 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
170 | #endif | |
171 | ||
172 | /* Legal flag mask for kmem_cache_create(). */ | |
173 | #if DEBUG | |
174 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
175 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
ac2b898c | 176 | SLAB_CACHE_DMA | \ |
1da177e4 LT |
177 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ |
178 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 179 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 | 180 | #else |
ac2b898c | 181 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
1da177e4 LT |
182 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ |
183 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
101a5001 | 184 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD) |
1da177e4 LT |
185 | #endif |
186 | ||
187 | /* | |
188 | * kmem_bufctl_t: | |
189 | * | |
190 | * Bufctl's are used for linking objs within a slab | |
191 | * linked offsets. | |
192 | * | |
193 | * This implementation relies on "struct page" for locating the cache & | |
194 | * slab an object belongs to. | |
195 | * This allows the bufctl structure to be small (one int), but limits | |
196 | * the number of objects a slab (not a cache) can contain when off-slab | |
197 | * bufctls are used. The limit is the size of the largest general cache | |
198 | * that does not use off-slab slabs. | |
199 | * For 32bit archs with 4 kB pages, is this 56. | |
200 | * This is not serious, as it is only for large objects, when it is unwise | |
201 | * to have too many per slab. | |
202 | * Note: This limit can be raised by introducing a general cache whose size | |
203 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
204 | */ | |
205 | ||
fa5b08d5 | 206 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
207 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
208 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
209 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
210 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 211 | |
1da177e4 LT |
212 | /* |
213 | * struct slab | |
214 | * | |
215 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
216 | * for a slab, or allocated from an general cache. | |
217 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
218 | */ | |
219 | struct slab { | |
b28a02de PE |
220 | struct list_head list; |
221 | unsigned long colouroff; | |
222 | void *s_mem; /* including colour offset */ | |
223 | unsigned int inuse; /* num of objs active in slab */ | |
224 | kmem_bufctl_t free; | |
225 | unsigned short nodeid; | |
1da177e4 LT |
226 | }; |
227 | ||
228 | /* | |
229 | * struct slab_rcu | |
230 | * | |
231 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
232 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
233 | * we need to approach a kernel structure obliquely, from its address | |
234 | * obtained without the usual locking. We can lock the structure to | |
235 | * stabilize it and check it's still at the given address, only if we | |
236 | * can be sure that the memory has not been meanwhile reused for some | |
237 | * other kind of object (which our subsystem's lock might corrupt). | |
238 | * | |
239 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
240 | * taking the spinlock within the structure expected at that address. | |
241 | * | |
242 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
243 | */ | |
244 | struct slab_rcu { | |
b28a02de | 245 | struct rcu_head head; |
343e0d7a | 246 | struct kmem_cache *cachep; |
b28a02de | 247 | void *addr; |
1da177e4 LT |
248 | }; |
249 | ||
250 | /* | |
251 | * struct array_cache | |
252 | * | |
1da177e4 LT |
253 | * Purpose: |
254 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
255 | * - reduce the number of linked list operations | |
256 | * - reduce spinlock operations | |
257 | * | |
258 | * The limit is stored in the per-cpu structure to reduce the data cache | |
259 | * footprint. | |
260 | * | |
261 | */ | |
262 | struct array_cache { | |
263 | unsigned int avail; | |
264 | unsigned int limit; | |
265 | unsigned int batchcount; | |
266 | unsigned int touched; | |
e498be7d | 267 | spinlock_t lock; |
a737b3e2 AM |
268 | void *entry[0]; /* |
269 | * Must have this definition in here for the proper | |
270 | * alignment of array_cache. Also simplifies accessing | |
271 | * the entries. | |
272 | * [0] is for gcc 2.95. It should really be []. | |
273 | */ | |
1da177e4 LT |
274 | }; |
275 | ||
a737b3e2 AM |
276 | /* |
277 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
278 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
279 | */ |
280 | #define BOOT_CPUCACHE_ENTRIES 1 | |
281 | struct arraycache_init { | |
282 | struct array_cache cache; | |
b28a02de | 283 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
284 | }; |
285 | ||
286 | /* | |
e498be7d | 287 | * The slab lists for all objects. |
1da177e4 LT |
288 | */ |
289 | struct kmem_list3 { | |
b28a02de PE |
290 | struct list_head slabs_partial; /* partial list first, better asm code */ |
291 | struct list_head slabs_full; | |
292 | struct list_head slabs_free; | |
293 | unsigned long free_objects; | |
b28a02de | 294 | unsigned int free_limit; |
2e1217cf | 295 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
296 | spinlock_t list_lock; |
297 | struct array_cache *shared; /* shared per node */ | |
298 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
299 | unsigned long next_reap; /* updated without locking */ |
300 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
301 | }; |
302 | ||
e498be7d CL |
303 | /* |
304 | * Need this for bootstrapping a per node allocator. | |
305 | */ | |
306 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
307 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
308 | #define CACHE_CACHE 0 | |
309 | #define SIZE_AC 1 | |
310 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
311 | ||
ed11d9eb CL |
312 | static int drain_freelist(struct kmem_cache *cache, |
313 | struct kmem_list3 *l3, int tofree); | |
314 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
315 | int node); | |
2ed3a4ef | 316 | static int enable_cpucache(struct kmem_cache *cachep); |
ed11d9eb CL |
317 | static void cache_reap(void *unused); |
318 | ||
e498be7d | 319 | /* |
a737b3e2 AM |
320 | * This function must be completely optimized away if a constant is passed to |
321 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 322 | */ |
7243cc05 | 323 | static __always_inline int index_of(const size_t size) |
e498be7d | 324 | { |
5ec8a847 SR |
325 | extern void __bad_size(void); |
326 | ||
e498be7d CL |
327 | if (__builtin_constant_p(size)) { |
328 | int i = 0; | |
329 | ||
330 | #define CACHE(x) \ | |
331 | if (size <=x) \ | |
332 | return i; \ | |
333 | else \ | |
334 | i++; | |
335 | #include "linux/kmalloc_sizes.h" | |
336 | #undef CACHE | |
5ec8a847 | 337 | __bad_size(); |
7243cc05 | 338 | } else |
5ec8a847 | 339 | __bad_size(); |
e498be7d CL |
340 | return 0; |
341 | } | |
342 | ||
e0a42726 IM |
343 | static int slab_early_init = 1; |
344 | ||
e498be7d CL |
345 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
346 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 347 | |
5295a74c | 348 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
349 | { |
350 | INIT_LIST_HEAD(&parent->slabs_full); | |
351 | INIT_LIST_HEAD(&parent->slabs_partial); | |
352 | INIT_LIST_HEAD(&parent->slabs_free); | |
353 | parent->shared = NULL; | |
354 | parent->alien = NULL; | |
2e1217cf | 355 | parent->colour_next = 0; |
e498be7d CL |
356 | spin_lock_init(&parent->list_lock); |
357 | parent->free_objects = 0; | |
358 | parent->free_touched = 0; | |
359 | } | |
360 | ||
a737b3e2 AM |
361 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
362 | do { \ | |
363 | INIT_LIST_HEAD(listp); \ | |
364 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
365 | } while (0) |
366 | ||
a737b3e2 AM |
367 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
368 | do { \ | |
e498be7d CL |
369 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
370 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
371 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
372 | } while (0) | |
1da177e4 LT |
373 | |
374 | /* | |
343e0d7a | 375 | * struct kmem_cache |
1da177e4 LT |
376 | * |
377 | * manages a cache. | |
378 | */ | |
b28a02de | 379 | |
2109a2d1 | 380 | struct kmem_cache { |
1da177e4 | 381 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de | 382 | struct array_cache *array[NR_CPUS]; |
b5d8ca7c | 383 | /* 2) Cache tunables. Protected by cache_chain_mutex */ |
b28a02de PE |
384 | unsigned int batchcount; |
385 | unsigned int limit; | |
386 | unsigned int shared; | |
b5d8ca7c | 387 | |
3dafccf2 | 388 | unsigned int buffer_size; |
b5d8ca7c | 389 | /* 3) touched by every alloc & free from the backend */ |
b28a02de | 390 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
b5d8ca7c | 391 | |
a737b3e2 AM |
392 | unsigned int flags; /* constant flags */ |
393 | unsigned int num; /* # of objs per slab */ | |
1da177e4 | 394 | |
b5d8ca7c | 395 | /* 4) cache_grow/shrink */ |
1da177e4 | 396 | /* order of pgs per slab (2^n) */ |
b28a02de | 397 | unsigned int gfporder; |
1da177e4 LT |
398 | |
399 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 400 | gfp_t gfpflags; |
1da177e4 | 401 | |
a737b3e2 | 402 | size_t colour; /* cache colouring range */ |
b28a02de | 403 | unsigned int colour_off; /* colour offset */ |
343e0d7a | 404 | struct kmem_cache *slabp_cache; |
b28a02de | 405 | unsigned int slab_size; |
a737b3e2 | 406 | unsigned int dflags; /* dynamic flags */ |
1da177e4 LT |
407 | |
408 | /* constructor func */ | |
343e0d7a | 409 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
410 | |
411 | /* de-constructor func */ | |
343e0d7a | 412 | void (*dtor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 | 413 | |
b5d8ca7c | 414 | /* 5) cache creation/removal */ |
b28a02de PE |
415 | const char *name; |
416 | struct list_head next; | |
1da177e4 | 417 | |
b5d8ca7c | 418 | /* 6) statistics */ |
1da177e4 | 419 | #if STATS |
b28a02de PE |
420 | unsigned long num_active; |
421 | unsigned long num_allocations; | |
422 | unsigned long high_mark; | |
423 | unsigned long grown; | |
424 | unsigned long reaped; | |
425 | unsigned long errors; | |
426 | unsigned long max_freeable; | |
427 | unsigned long node_allocs; | |
428 | unsigned long node_frees; | |
fb7faf33 | 429 | unsigned long node_overflow; |
b28a02de PE |
430 | atomic_t allochit; |
431 | atomic_t allocmiss; | |
432 | atomic_t freehit; | |
433 | atomic_t freemiss; | |
1da177e4 LT |
434 | #endif |
435 | #if DEBUG | |
3dafccf2 MS |
436 | /* |
437 | * If debugging is enabled, then the allocator can add additional | |
438 | * fields and/or padding to every object. buffer_size contains the total | |
439 | * object size including these internal fields, the following two | |
440 | * variables contain the offset to the user object and its size. | |
441 | */ | |
442 | int obj_offset; | |
443 | int obj_size; | |
1da177e4 LT |
444 | #endif |
445 | }; | |
446 | ||
447 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
448 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
449 | ||
450 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
451 | /* |
452 | * Optimization question: fewer reaps means less probability for unnessary | |
453 | * cpucache drain/refill cycles. | |
1da177e4 | 454 | * |
dc6f3f27 | 455 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
456 | * which could lock up otherwise freeable slabs. |
457 | */ | |
458 | #define REAPTIMEOUT_CPUC (2*HZ) | |
459 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
460 | ||
461 | #if STATS | |
462 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
463 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
464 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
465 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 466 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
467 | #define STATS_SET_HIGH(x) \ |
468 | do { \ | |
469 | if ((x)->num_active > (x)->high_mark) \ | |
470 | (x)->high_mark = (x)->num_active; \ | |
471 | } while (0) | |
1da177e4 LT |
472 | #define STATS_INC_ERR(x) ((x)->errors++) |
473 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 474 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 475 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
476 | #define STATS_SET_FREEABLE(x, i) \ |
477 | do { \ | |
478 | if ((x)->max_freeable < i) \ | |
479 | (x)->max_freeable = i; \ | |
480 | } while (0) | |
1da177e4 LT |
481 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
482 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
483 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
484 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
485 | #else | |
486 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
487 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
488 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
489 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 490 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
491 | #define STATS_SET_HIGH(x) do { } while (0) |
492 | #define STATS_INC_ERR(x) do { } while (0) | |
493 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 494 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 495 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 496 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
497 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
498 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
499 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
500 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
501 | #endif | |
502 | ||
503 | #if DEBUG | |
1da177e4 | 504 | |
a737b3e2 AM |
505 | /* |
506 | * memory layout of objects: | |
1da177e4 | 507 | * 0 : objp |
3dafccf2 | 508 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
509 | * the end of an object is aligned with the end of the real |
510 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 511 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 512 | * redzone word. |
3dafccf2 MS |
513 | * cachep->obj_offset: The real object. |
514 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
515 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
516 | * [BYTES_PER_WORD long] | |
1da177e4 | 517 | */ |
343e0d7a | 518 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 519 | { |
3dafccf2 | 520 | return cachep->obj_offset; |
1da177e4 LT |
521 | } |
522 | ||
343e0d7a | 523 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 524 | { |
3dafccf2 | 525 | return cachep->obj_size; |
1da177e4 LT |
526 | } |
527 | ||
343e0d7a | 528 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
529 | { |
530 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 531 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
532 | } |
533 | ||
343e0d7a | 534 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
535 | { |
536 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
537 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 538 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 539 | 2 * BYTES_PER_WORD); |
3dafccf2 | 540 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
541 | } |
542 | ||
343e0d7a | 543 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
544 | { |
545 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 546 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
547 | } |
548 | ||
549 | #else | |
550 | ||
3dafccf2 MS |
551 | #define obj_offset(x) 0 |
552 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
553 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
554 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
555 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
556 | ||
557 | #endif | |
558 | ||
559 | /* | |
a737b3e2 AM |
560 | * Maximum size of an obj (in 2^order pages) and absolute limit for the gfp |
561 | * order. | |
1da177e4 LT |
562 | */ |
563 | #if defined(CONFIG_LARGE_ALLOCS) | |
564 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
565 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
566 | #elif defined(CONFIG_MMU) | |
567 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
568 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
569 | #else | |
570 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
571 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
572 | #endif | |
573 | ||
574 | /* | |
575 | * Do not go above this order unless 0 objects fit into the slab. | |
576 | */ | |
577 | #define BREAK_GFP_ORDER_HI 1 | |
578 | #define BREAK_GFP_ORDER_LO 0 | |
579 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
580 | ||
a737b3e2 AM |
581 | /* |
582 | * Functions for storing/retrieving the cachep and or slab from the page | |
583 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
584 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 585 | */ |
065d41cb PE |
586 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
587 | { | |
588 | page->lru.next = (struct list_head *)cache; | |
589 | } | |
590 | ||
591 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
592 | { | |
84097518 NP |
593 | if (unlikely(PageCompound(page))) |
594 | page = (struct page *)page_private(page); | |
ddc2e812 | 595 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
596 | return (struct kmem_cache *)page->lru.next; |
597 | } | |
598 | ||
599 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
600 | { | |
601 | page->lru.prev = (struct list_head *)slab; | |
602 | } | |
603 | ||
604 | static inline struct slab *page_get_slab(struct page *page) | |
605 | { | |
84097518 NP |
606 | if (unlikely(PageCompound(page))) |
607 | page = (struct page *)page_private(page); | |
ddc2e812 | 608 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
609 | return (struct slab *)page->lru.prev; |
610 | } | |
1da177e4 | 611 | |
6ed5eb22 PE |
612 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
613 | { | |
614 | struct page *page = virt_to_page(obj); | |
615 | return page_get_cache(page); | |
616 | } | |
617 | ||
618 | static inline struct slab *virt_to_slab(const void *obj) | |
619 | { | |
620 | struct page *page = virt_to_page(obj); | |
621 | return page_get_slab(page); | |
622 | } | |
623 | ||
8fea4e96 PE |
624 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
625 | unsigned int idx) | |
626 | { | |
627 | return slab->s_mem + cache->buffer_size * idx; | |
628 | } | |
629 | ||
630 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
631 | struct slab *slab, void *obj) | |
632 | { | |
633 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
634 | } | |
635 | ||
a737b3e2 AM |
636 | /* |
637 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
638 | */ | |
1da177e4 LT |
639 | struct cache_sizes malloc_sizes[] = { |
640 | #define CACHE(x) { .cs_size = (x) }, | |
641 | #include <linux/kmalloc_sizes.h> | |
642 | CACHE(ULONG_MAX) | |
643 | #undef CACHE | |
644 | }; | |
645 | EXPORT_SYMBOL(malloc_sizes); | |
646 | ||
647 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
648 | struct cache_names { | |
649 | char *name; | |
650 | char *name_dma; | |
651 | }; | |
652 | ||
653 | static struct cache_names __initdata cache_names[] = { | |
654 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
655 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 656 | {NULL,} |
1da177e4 LT |
657 | #undef CACHE |
658 | }; | |
659 | ||
660 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 661 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 662 | static struct arraycache_init initarray_generic = |
b28a02de | 663 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
664 | |
665 | /* internal cache of cache description objs */ | |
343e0d7a | 666 | static struct kmem_cache cache_cache = { |
b28a02de PE |
667 | .batchcount = 1, |
668 | .limit = BOOT_CPUCACHE_ENTRIES, | |
669 | .shared = 1, | |
343e0d7a | 670 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 671 | .name = "kmem_cache", |
1da177e4 | 672 | #if DEBUG |
343e0d7a | 673 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
674 | #endif |
675 | }; | |
676 | ||
056c6241 RT |
677 | #define BAD_ALIEN_MAGIC 0x01020304ul |
678 | ||
f1aaee53 AV |
679 | #ifdef CONFIG_LOCKDEP |
680 | ||
681 | /* | |
682 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
683 | * for other slabs "off slab". | |
684 | * The locking for this is tricky in that it nests within the locks | |
685 | * of all other slabs in a few places; to deal with this special | |
686 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
687 | * |
688 | * We set lock class for alien array caches which are up during init. | |
689 | * The lock annotation will be lost if all cpus of a node goes down and | |
690 | * then comes back up during hotplug | |
f1aaee53 | 691 | */ |
056c6241 RT |
692 | static struct lock_class_key on_slab_l3_key; |
693 | static struct lock_class_key on_slab_alc_key; | |
694 | ||
695 | static inline void init_lock_keys(void) | |
f1aaee53 | 696 | |
f1aaee53 AV |
697 | { |
698 | int q; | |
056c6241 RT |
699 | struct cache_sizes *s = malloc_sizes; |
700 | ||
701 | while (s->cs_size != ULONG_MAX) { | |
702 | for_each_node(q) { | |
703 | struct array_cache **alc; | |
704 | int r; | |
705 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
706 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
707 | continue; | |
708 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
709 | alc = l3->alien; | |
710 | /* | |
711 | * FIXME: This check for BAD_ALIEN_MAGIC | |
712 | * should go away when common slab code is taught to | |
713 | * work even without alien caches. | |
714 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
715 | * for alloc_alien_cache, | |
716 | */ | |
717 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
718 | continue; | |
719 | for_each_node(r) { | |
720 | if (alc[r]) | |
721 | lockdep_set_class(&alc[r]->lock, | |
722 | &on_slab_alc_key); | |
723 | } | |
724 | } | |
725 | s++; | |
f1aaee53 AV |
726 | } |
727 | } | |
f1aaee53 | 728 | #else |
056c6241 | 729 | static inline void init_lock_keys(void) |
f1aaee53 AV |
730 | { |
731 | } | |
732 | #endif | |
733 | ||
1da177e4 | 734 | /* Guard access to the cache-chain. */ |
fc0abb14 | 735 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
736 | static struct list_head cache_chain; |
737 | ||
1da177e4 LT |
738 | /* |
739 | * chicken and egg problem: delay the per-cpu array allocation | |
740 | * until the general caches are up. | |
741 | */ | |
742 | static enum { | |
743 | NONE, | |
e498be7d CL |
744 | PARTIAL_AC, |
745 | PARTIAL_L3, | |
1da177e4 LT |
746 | FULL |
747 | } g_cpucache_up; | |
748 | ||
39d24e64 MK |
749 | /* |
750 | * used by boot code to determine if it can use slab based allocator | |
751 | */ | |
752 | int slab_is_available(void) | |
753 | { | |
754 | return g_cpucache_up == FULL; | |
755 | } | |
756 | ||
1da177e4 LT |
757 | static DEFINE_PER_CPU(struct work_struct, reap_work); |
758 | ||
343e0d7a | 759 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
760 | { |
761 | return cachep->array[smp_processor_id()]; | |
762 | } | |
763 | ||
a737b3e2 AM |
764 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
765 | gfp_t gfpflags) | |
1da177e4 LT |
766 | { |
767 | struct cache_sizes *csizep = malloc_sizes; | |
768 | ||
769 | #if DEBUG | |
770 | /* This happens if someone tries to call | |
b28a02de PE |
771 | * kmem_cache_create(), or __kmalloc(), before |
772 | * the generic caches are initialized. | |
773 | */ | |
c7e43c78 | 774 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
775 | #endif |
776 | while (size > csizep->cs_size) | |
777 | csizep++; | |
778 | ||
779 | /* | |
0abf40c1 | 780 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
781 | * has cs_{dma,}cachep==NULL. Thus no special case |
782 | * for large kmalloc calls required. | |
783 | */ | |
784 | if (unlikely(gfpflags & GFP_DMA)) | |
785 | return csizep->cs_dmacachep; | |
786 | return csizep->cs_cachep; | |
787 | } | |
788 | ||
b221385b | 789 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
790 | { |
791 | return __find_general_cachep(size, gfpflags); | |
792 | } | |
97e2bde4 | 793 | |
fbaccacf | 794 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 795 | { |
fbaccacf SR |
796 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
797 | } | |
1da177e4 | 798 | |
a737b3e2 AM |
799 | /* |
800 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
801 | */ | |
fbaccacf SR |
802 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
803 | size_t align, int flags, size_t *left_over, | |
804 | unsigned int *num) | |
805 | { | |
806 | int nr_objs; | |
807 | size_t mgmt_size; | |
808 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 809 | |
fbaccacf SR |
810 | /* |
811 | * The slab management structure can be either off the slab or | |
812 | * on it. For the latter case, the memory allocated for a | |
813 | * slab is used for: | |
814 | * | |
815 | * - The struct slab | |
816 | * - One kmem_bufctl_t for each object | |
817 | * - Padding to respect alignment of @align | |
818 | * - @buffer_size bytes for each object | |
819 | * | |
820 | * If the slab management structure is off the slab, then the | |
821 | * alignment will already be calculated into the size. Because | |
822 | * the slabs are all pages aligned, the objects will be at the | |
823 | * correct alignment when allocated. | |
824 | */ | |
825 | if (flags & CFLGS_OFF_SLAB) { | |
826 | mgmt_size = 0; | |
827 | nr_objs = slab_size / buffer_size; | |
828 | ||
829 | if (nr_objs > SLAB_LIMIT) | |
830 | nr_objs = SLAB_LIMIT; | |
831 | } else { | |
832 | /* | |
833 | * Ignore padding for the initial guess. The padding | |
834 | * is at most @align-1 bytes, and @buffer_size is at | |
835 | * least @align. In the worst case, this result will | |
836 | * be one greater than the number of objects that fit | |
837 | * into the memory allocation when taking the padding | |
838 | * into account. | |
839 | */ | |
840 | nr_objs = (slab_size - sizeof(struct slab)) / | |
841 | (buffer_size + sizeof(kmem_bufctl_t)); | |
842 | ||
843 | /* | |
844 | * This calculated number will be either the right | |
845 | * amount, or one greater than what we want. | |
846 | */ | |
847 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
848 | > slab_size) | |
849 | nr_objs--; | |
850 | ||
851 | if (nr_objs > SLAB_LIMIT) | |
852 | nr_objs = SLAB_LIMIT; | |
853 | ||
854 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
855 | } | |
856 | *num = nr_objs; | |
857 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
858 | } |
859 | ||
860 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
861 | ||
a737b3e2 AM |
862 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
863 | char *msg) | |
1da177e4 LT |
864 | { |
865 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 866 | function, cachep->name, msg); |
1da177e4 LT |
867 | dump_stack(); |
868 | } | |
869 | ||
8fce4d8e CL |
870 | #ifdef CONFIG_NUMA |
871 | /* | |
872 | * Special reaping functions for NUMA systems called from cache_reap(). | |
873 | * These take care of doing round robin flushing of alien caches (containing | |
874 | * objects freed on different nodes from which they were allocated) and the | |
875 | * flushing of remote pcps by calling drain_node_pages. | |
876 | */ | |
877 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
878 | ||
879 | static void init_reap_node(int cpu) | |
880 | { | |
881 | int node; | |
882 | ||
883 | node = next_node(cpu_to_node(cpu), node_online_map); | |
884 | if (node == MAX_NUMNODES) | |
442295c9 | 885 | node = first_node(node_online_map); |
8fce4d8e CL |
886 | |
887 | __get_cpu_var(reap_node) = node; | |
888 | } | |
889 | ||
890 | static void next_reap_node(void) | |
891 | { | |
892 | int node = __get_cpu_var(reap_node); | |
893 | ||
894 | /* | |
895 | * Also drain per cpu pages on remote zones | |
896 | */ | |
897 | if (node != numa_node_id()) | |
898 | drain_node_pages(node); | |
899 | ||
900 | node = next_node(node, node_online_map); | |
901 | if (unlikely(node >= MAX_NUMNODES)) | |
902 | node = first_node(node_online_map); | |
903 | __get_cpu_var(reap_node) = node; | |
904 | } | |
905 | ||
906 | #else | |
907 | #define init_reap_node(cpu) do { } while (0) | |
908 | #define next_reap_node(void) do { } while (0) | |
909 | #endif | |
910 | ||
1da177e4 LT |
911 | /* |
912 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
913 | * via the workqueue/eventd. | |
914 | * Add the CPU number into the expiration time to minimize the possibility of | |
915 | * the CPUs getting into lockstep and contending for the global cache chain | |
916 | * lock. | |
917 | */ | |
918 | static void __devinit start_cpu_timer(int cpu) | |
919 | { | |
920 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
921 | ||
922 | /* | |
923 | * When this gets called from do_initcalls via cpucache_init(), | |
924 | * init_workqueues() has already run, so keventd will be setup | |
925 | * at that time. | |
926 | */ | |
927 | if (keventd_up() && reap_work->func == NULL) { | |
8fce4d8e | 928 | init_reap_node(cpu); |
1da177e4 LT |
929 | INIT_WORK(reap_work, cache_reap, NULL); |
930 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
931 | } | |
932 | } | |
933 | ||
e498be7d | 934 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 935 | int batchcount) |
1da177e4 | 936 | { |
b28a02de | 937 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
938 | struct array_cache *nc = NULL; |
939 | ||
e498be7d | 940 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
941 | if (nc) { |
942 | nc->avail = 0; | |
943 | nc->limit = entries; | |
944 | nc->batchcount = batchcount; | |
945 | nc->touched = 0; | |
e498be7d | 946 | spin_lock_init(&nc->lock); |
1da177e4 LT |
947 | } |
948 | return nc; | |
949 | } | |
950 | ||
3ded175a CL |
951 | /* |
952 | * Transfer objects in one arraycache to another. | |
953 | * Locking must be handled by the caller. | |
954 | * | |
955 | * Return the number of entries transferred. | |
956 | */ | |
957 | static int transfer_objects(struct array_cache *to, | |
958 | struct array_cache *from, unsigned int max) | |
959 | { | |
960 | /* Figure out how many entries to transfer */ | |
961 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
962 | ||
963 | if (!nr) | |
964 | return 0; | |
965 | ||
966 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
967 | sizeof(void *) *nr); | |
968 | ||
969 | from->avail -= nr; | |
970 | to->avail += nr; | |
971 | to->touched = 1; | |
972 | return nr; | |
973 | } | |
974 | ||
e498be7d | 975 | #ifdef CONFIG_NUMA |
343e0d7a | 976 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 977 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 978 | |
5295a74c | 979 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
980 | { |
981 | struct array_cache **ac_ptr; | |
b28a02de | 982 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
983 | int i; |
984 | ||
985 | if (limit > 1) | |
986 | limit = 12; | |
987 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
988 | if (ac_ptr) { | |
989 | for_each_node(i) { | |
990 | if (i == node || !node_online(i)) { | |
991 | ac_ptr[i] = NULL; | |
992 | continue; | |
993 | } | |
994 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
995 | if (!ac_ptr[i]) { | |
b28a02de | 996 | for (i--; i <= 0; i--) |
e498be7d CL |
997 | kfree(ac_ptr[i]); |
998 | kfree(ac_ptr); | |
999 | return NULL; | |
1000 | } | |
1001 | } | |
1002 | } | |
1003 | return ac_ptr; | |
1004 | } | |
1005 | ||
5295a74c | 1006 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1007 | { |
1008 | int i; | |
1009 | ||
1010 | if (!ac_ptr) | |
1011 | return; | |
e498be7d | 1012 | for_each_node(i) |
b28a02de | 1013 | kfree(ac_ptr[i]); |
e498be7d CL |
1014 | kfree(ac_ptr); |
1015 | } | |
1016 | ||
343e0d7a | 1017 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1018 | struct array_cache *ac, int node) |
e498be7d CL |
1019 | { |
1020 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1021 | ||
1022 | if (ac->avail) { | |
1023 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1024 | /* |
1025 | * Stuff objects into the remote nodes shared array first. | |
1026 | * That way we could avoid the overhead of putting the objects | |
1027 | * into the free lists and getting them back later. | |
1028 | */ | |
693f7d36 JS |
1029 | if (rl3->shared) |
1030 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1031 | |
ff69416e | 1032 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1033 | ac->avail = 0; |
1034 | spin_unlock(&rl3->list_lock); | |
1035 | } | |
1036 | } | |
1037 | ||
8fce4d8e CL |
1038 | /* |
1039 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1040 | */ | |
1041 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1042 | { | |
1043 | int node = __get_cpu_var(reap_node); | |
1044 | ||
1045 | if (l3->alien) { | |
1046 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1047 | |
1048 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1049 | __drain_alien_cache(cachep, ac, node); |
1050 | spin_unlock_irq(&ac->lock); | |
1051 | } | |
1052 | } | |
1053 | } | |
1054 | ||
a737b3e2 AM |
1055 | static void drain_alien_cache(struct kmem_cache *cachep, |
1056 | struct array_cache **alien) | |
e498be7d | 1057 | { |
b28a02de | 1058 | int i = 0; |
e498be7d CL |
1059 | struct array_cache *ac; |
1060 | unsigned long flags; | |
1061 | ||
1062 | for_each_online_node(i) { | |
4484ebf1 | 1063 | ac = alien[i]; |
e498be7d CL |
1064 | if (ac) { |
1065 | spin_lock_irqsave(&ac->lock, flags); | |
1066 | __drain_alien_cache(cachep, ac, i); | |
1067 | spin_unlock_irqrestore(&ac->lock, flags); | |
1068 | } | |
1069 | } | |
1070 | } | |
729bd0b7 | 1071 | |
873623df | 1072 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1073 | { |
1074 | struct slab *slabp = virt_to_slab(objp); | |
1075 | int nodeid = slabp->nodeid; | |
1076 | struct kmem_list3 *l3; | |
1077 | struct array_cache *alien = NULL; | |
1078 | ||
1079 | /* | |
1080 | * Make sure we are not freeing a object from another node to the array | |
1081 | * cache on this cpu. | |
1082 | */ | |
1083 | if (likely(slabp->nodeid == numa_node_id())) | |
1084 | return 0; | |
1085 | ||
1086 | l3 = cachep->nodelists[numa_node_id()]; | |
1087 | STATS_INC_NODEFREES(cachep); | |
1088 | if (l3->alien && l3->alien[nodeid]) { | |
1089 | alien = l3->alien[nodeid]; | |
873623df | 1090 | spin_lock(&alien->lock); |
729bd0b7 PE |
1091 | if (unlikely(alien->avail == alien->limit)) { |
1092 | STATS_INC_ACOVERFLOW(cachep); | |
1093 | __drain_alien_cache(cachep, alien, nodeid); | |
1094 | } | |
1095 | alien->entry[alien->avail++] = objp; | |
1096 | spin_unlock(&alien->lock); | |
1097 | } else { | |
1098 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1099 | free_block(cachep, &objp, 1, nodeid); | |
1100 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1101 | } | |
1102 | return 1; | |
1103 | } | |
1104 | ||
e498be7d | 1105 | #else |
7a21ef6f | 1106 | |
4484ebf1 | 1107 | #define drain_alien_cache(cachep, alien) do { } while (0) |
8fce4d8e | 1108 | #define reap_alien(cachep, l3) do { } while (0) |
4484ebf1 | 1109 | |
7a21ef6f LT |
1110 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
1111 | { | |
056c6241 | 1112 | return (struct array_cache **)BAD_ALIEN_MAGIC; |
7a21ef6f LT |
1113 | } |
1114 | ||
4484ebf1 RT |
1115 | static inline void free_alien_cache(struct array_cache **ac_ptr) |
1116 | { | |
1117 | } | |
7a21ef6f | 1118 | |
873623df | 1119 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1120 | { |
1121 | return 0; | |
1122 | } | |
1123 | ||
e498be7d CL |
1124 | #endif |
1125 | ||
8c78f307 | 1126 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 1127 | unsigned long action, void *hcpu) |
1da177e4 LT |
1128 | { |
1129 | long cpu = (long)hcpu; | |
343e0d7a | 1130 | struct kmem_cache *cachep; |
e498be7d CL |
1131 | struct kmem_list3 *l3 = NULL; |
1132 | int node = cpu_to_node(cpu); | |
1133 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
1134 | |
1135 | switch (action) { | |
1136 | case CPU_UP_PREPARE: | |
fc0abb14 | 1137 | mutex_lock(&cache_chain_mutex); |
a737b3e2 AM |
1138 | /* |
1139 | * We need to do this right in the beginning since | |
e498be7d CL |
1140 | * alloc_arraycache's are going to use this list. |
1141 | * kmalloc_node allows us to add the slab to the right | |
1142 | * kmem_list3 and not this cpu's kmem_list3 | |
1143 | */ | |
1144 | ||
1da177e4 | 1145 | list_for_each_entry(cachep, &cache_chain, next) { |
a737b3e2 AM |
1146 | /* |
1147 | * Set up the size64 kmemlist for cpu before we can | |
e498be7d CL |
1148 | * begin anything. Make sure some other cpu on this |
1149 | * node has not already allocated this | |
1150 | */ | |
1151 | if (!cachep->nodelists[node]) { | |
a737b3e2 AM |
1152 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1153 | if (!l3) | |
e498be7d CL |
1154 | goto bad; |
1155 | kmem_list3_init(l3); | |
1156 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1157 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1158 | |
4484ebf1 RT |
1159 | /* |
1160 | * The l3s don't come and go as CPUs come and | |
1161 | * go. cache_chain_mutex is sufficient | |
1162 | * protection here. | |
1163 | */ | |
e498be7d CL |
1164 | cachep->nodelists[node] = l3; |
1165 | } | |
1da177e4 | 1166 | |
e498be7d CL |
1167 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1168 | cachep->nodelists[node]->free_limit = | |
a737b3e2 AM |
1169 | (1 + nr_cpus_node(node)) * |
1170 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1171 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1172 | } | |
1173 | ||
a737b3e2 AM |
1174 | /* |
1175 | * Now we can go ahead with allocating the shared arrays and | |
1176 | * array caches | |
1177 | */ | |
e498be7d | 1178 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1179 | struct array_cache *nc; |
4484ebf1 RT |
1180 | struct array_cache *shared; |
1181 | struct array_cache **alien; | |
cd105df4 | 1182 | |
e498be7d | 1183 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1184 | cachep->batchcount); |
1da177e4 LT |
1185 | if (!nc) |
1186 | goto bad; | |
4484ebf1 RT |
1187 | shared = alloc_arraycache(node, |
1188 | cachep->shared * cachep->batchcount, | |
1189 | 0xbaadf00d); | |
1190 | if (!shared) | |
1191 | goto bad; | |
7a21ef6f | 1192 | |
4484ebf1 RT |
1193 | alien = alloc_alien_cache(node, cachep->limit); |
1194 | if (!alien) | |
1195 | goto bad; | |
1da177e4 | 1196 | cachep->array[cpu] = nc; |
e498be7d CL |
1197 | l3 = cachep->nodelists[node]; |
1198 | BUG_ON(!l3); | |
e498be7d | 1199 | |
4484ebf1 RT |
1200 | spin_lock_irq(&l3->list_lock); |
1201 | if (!l3->shared) { | |
1202 | /* | |
1203 | * We are serialised from CPU_DEAD or | |
1204 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1205 | */ | |
1206 | l3->shared = shared; | |
1207 | shared = NULL; | |
e498be7d | 1208 | } |
4484ebf1 RT |
1209 | #ifdef CONFIG_NUMA |
1210 | if (!l3->alien) { | |
1211 | l3->alien = alien; | |
1212 | alien = NULL; | |
1213 | } | |
1214 | #endif | |
1215 | spin_unlock_irq(&l3->list_lock); | |
4484ebf1 RT |
1216 | kfree(shared); |
1217 | free_alien_cache(alien); | |
1da177e4 | 1218 | } |
fc0abb14 | 1219 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1220 | break; |
1221 | case CPU_ONLINE: | |
1222 | start_cpu_timer(cpu); | |
1223 | break; | |
1224 | #ifdef CONFIG_HOTPLUG_CPU | |
1225 | case CPU_DEAD: | |
4484ebf1 RT |
1226 | /* |
1227 | * Even if all the cpus of a node are down, we don't free the | |
1228 | * kmem_list3 of any cache. This to avoid a race between | |
1229 | * cpu_down, and a kmalloc allocation from another cpu for | |
1230 | * memory from the node of the cpu going down. The list3 | |
1231 | * structure is usually allocated from kmem_cache_create() and | |
1232 | * gets destroyed at kmem_cache_destroy(). | |
1233 | */ | |
1da177e4 LT |
1234 | /* fall thru */ |
1235 | case CPU_UP_CANCELED: | |
fc0abb14 | 1236 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1237 | list_for_each_entry(cachep, &cache_chain, next) { |
1238 | struct array_cache *nc; | |
4484ebf1 RT |
1239 | struct array_cache *shared; |
1240 | struct array_cache **alien; | |
e498be7d | 1241 | cpumask_t mask; |
1da177e4 | 1242 | |
e498be7d | 1243 | mask = node_to_cpumask(node); |
1da177e4 LT |
1244 | /* cpu is dead; no one can alloc from it. */ |
1245 | nc = cachep->array[cpu]; | |
1246 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1247 | l3 = cachep->nodelists[node]; |
1248 | ||
1249 | if (!l3) | |
4484ebf1 | 1250 | goto free_array_cache; |
e498be7d | 1251 | |
ca3b9b91 | 1252 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1253 | |
1254 | /* Free limit for this kmem_list3 */ | |
1255 | l3->free_limit -= cachep->batchcount; | |
1256 | if (nc) | |
ff69416e | 1257 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1258 | |
1259 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1260 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1261 | goto free_array_cache; |
b28a02de | 1262 | } |
e498be7d | 1263 | |
4484ebf1 RT |
1264 | shared = l3->shared; |
1265 | if (shared) { | |
e498be7d | 1266 | free_block(cachep, l3->shared->entry, |
b28a02de | 1267 | l3->shared->avail, node); |
e498be7d CL |
1268 | l3->shared = NULL; |
1269 | } | |
e498be7d | 1270 | |
4484ebf1 RT |
1271 | alien = l3->alien; |
1272 | l3->alien = NULL; | |
1273 | ||
1274 | spin_unlock_irq(&l3->list_lock); | |
1275 | ||
1276 | kfree(shared); | |
1277 | if (alien) { | |
1278 | drain_alien_cache(cachep, alien); | |
1279 | free_alien_cache(alien); | |
e498be7d | 1280 | } |
4484ebf1 | 1281 | free_array_cache: |
1da177e4 LT |
1282 | kfree(nc); |
1283 | } | |
4484ebf1 RT |
1284 | /* |
1285 | * In the previous loop, all the objects were freed to | |
1286 | * the respective cache's slabs, now we can go ahead and | |
1287 | * shrink each nodelist to its limit. | |
1288 | */ | |
1289 | list_for_each_entry(cachep, &cache_chain, next) { | |
1290 | l3 = cachep->nodelists[node]; | |
1291 | if (!l3) | |
1292 | continue; | |
ed11d9eb | 1293 | drain_freelist(cachep, l3, l3->free_objects); |
4484ebf1 | 1294 | } |
fc0abb14 | 1295 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1296 | break; |
1297 | #endif | |
1298 | } | |
1299 | return NOTIFY_OK; | |
a737b3e2 | 1300 | bad: |
fc0abb14 | 1301 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1302 | return NOTIFY_BAD; |
1303 | } | |
1304 | ||
74b85f37 CS |
1305 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1306 | &cpuup_callback, NULL, 0 | |
1307 | }; | |
1da177e4 | 1308 | |
e498be7d CL |
1309 | /* |
1310 | * swap the static kmem_list3 with kmalloced memory | |
1311 | */ | |
a737b3e2 AM |
1312 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1313 | int nodeid) | |
e498be7d CL |
1314 | { |
1315 | struct kmem_list3 *ptr; | |
1316 | ||
1317 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1318 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1319 | BUG_ON(!ptr); | |
1320 | ||
1321 | local_irq_disable(); | |
1322 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
2b2d5493 IM |
1323 | /* |
1324 | * Do not assume that spinlocks can be initialized via memcpy: | |
1325 | */ | |
1326 | spin_lock_init(&ptr->list_lock); | |
1327 | ||
e498be7d CL |
1328 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1329 | cachep->nodelists[nodeid] = ptr; | |
1330 | local_irq_enable(); | |
1331 | } | |
1332 | ||
a737b3e2 AM |
1333 | /* |
1334 | * Initialisation. Called after the page allocator have been initialised and | |
1335 | * before smp_init(). | |
1da177e4 LT |
1336 | */ |
1337 | void __init kmem_cache_init(void) | |
1338 | { | |
1339 | size_t left_over; | |
1340 | struct cache_sizes *sizes; | |
1341 | struct cache_names *names; | |
e498be7d | 1342 | int i; |
07ed76b2 | 1343 | int order; |
e498be7d CL |
1344 | |
1345 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1346 | kmem_list3_init(&initkmem_list3[i]); | |
1347 | if (i < MAX_NUMNODES) | |
1348 | cache_cache.nodelists[i] = NULL; | |
1349 | } | |
1da177e4 LT |
1350 | |
1351 | /* | |
1352 | * Fragmentation resistance on low memory - only use bigger | |
1353 | * page orders on machines with more than 32MB of memory. | |
1354 | */ | |
1355 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1356 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1357 | ||
1da177e4 LT |
1358 | /* Bootstrap is tricky, because several objects are allocated |
1359 | * from caches that do not exist yet: | |
a737b3e2 AM |
1360 | * 1) initialize the cache_cache cache: it contains the struct |
1361 | * kmem_cache structures of all caches, except cache_cache itself: | |
1362 | * cache_cache is statically allocated. | |
e498be7d CL |
1363 | * Initially an __init data area is used for the head array and the |
1364 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1365 | * array at the end of the bootstrap. | |
1da177e4 | 1366 | * 2) Create the first kmalloc cache. |
343e0d7a | 1367 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1368 | * An __init data area is used for the head array. |
1369 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1370 | * head arrays. | |
1da177e4 LT |
1371 | * 4) Replace the __init data head arrays for cache_cache and the first |
1372 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1373 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1374 | * the other cache's with kmalloc allocated memory. | |
1375 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1376 | */ |
1377 | ||
1378 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1379 | INIT_LIST_HEAD(&cache_chain); |
1380 | list_add(&cache_cache.next, &cache_chain); | |
1381 | cache_cache.colour_off = cache_line_size(); | |
1382 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1383 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1384 | |
a737b3e2 AM |
1385 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1386 | cache_line_size()); | |
1da177e4 | 1387 | |
07ed76b2 JS |
1388 | for (order = 0; order < MAX_ORDER; order++) { |
1389 | cache_estimate(order, cache_cache.buffer_size, | |
1390 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1391 | if (cache_cache.num) | |
1392 | break; | |
1393 | } | |
40094fa6 | 1394 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1395 | cache_cache.gfporder = order; |
b28a02de | 1396 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1397 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1398 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1399 | |
1400 | /* 2+3) create the kmalloc caches */ | |
1401 | sizes = malloc_sizes; | |
1402 | names = cache_names; | |
1403 | ||
a737b3e2 AM |
1404 | /* |
1405 | * Initialize the caches that provide memory for the array cache and the | |
1406 | * kmem_list3 structures first. Without this, further allocations will | |
1407 | * bug. | |
e498be7d CL |
1408 | */ |
1409 | ||
1410 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1411 | sizes[INDEX_AC].cs_size, |
1412 | ARCH_KMALLOC_MINALIGN, | |
1413 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1414 | NULL, NULL); | |
e498be7d | 1415 | |
a737b3e2 | 1416 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1417 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1418 | kmem_cache_create(names[INDEX_L3].name, |
1419 | sizes[INDEX_L3].cs_size, | |
1420 | ARCH_KMALLOC_MINALIGN, | |
1421 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1422 | NULL, NULL); | |
1423 | } | |
e498be7d | 1424 | |
e0a42726 IM |
1425 | slab_early_init = 0; |
1426 | ||
1da177e4 | 1427 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1428 | /* |
1429 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1430 | * This should be particularly beneficial on SMP boxes, as it |
1431 | * eliminates "false sharing". | |
1432 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1433 | * allow tighter packing of the smaller caches. |
1434 | */ | |
a737b3e2 | 1435 | if (!sizes->cs_cachep) { |
e498be7d | 1436 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1437 | sizes->cs_size, |
1438 | ARCH_KMALLOC_MINALIGN, | |
1439 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1440 | NULL, NULL); | |
1441 | } | |
1da177e4 | 1442 | |
1da177e4 | 1443 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, |
a737b3e2 AM |
1444 | sizes->cs_size, |
1445 | ARCH_KMALLOC_MINALIGN, | |
1446 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1447 | SLAB_PANIC, | |
1448 | NULL, NULL); | |
1da177e4 LT |
1449 | sizes++; |
1450 | names++; | |
1451 | } | |
1452 | /* 4) Replace the bootstrap head arrays */ | |
1453 | { | |
2b2d5493 | 1454 | struct array_cache *ptr; |
e498be7d | 1455 | |
1da177e4 | 1456 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1457 | |
1da177e4 | 1458 | local_irq_disable(); |
9a2dba4b PE |
1459 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1460 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1461 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1462 | /* |
1463 | * Do not assume that spinlocks can be initialized via memcpy: | |
1464 | */ | |
1465 | spin_lock_init(&ptr->lock); | |
1466 | ||
1da177e4 LT |
1467 | cache_cache.array[smp_processor_id()] = ptr; |
1468 | local_irq_enable(); | |
e498be7d | 1469 | |
1da177e4 | 1470 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1471 | |
1da177e4 | 1472 | local_irq_disable(); |
9a2dba4b | 1473 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1474 | != &initarray_generic.cache); |
9a2dba4b | 1475 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1476 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1477 | /* |
1478 | * Do not assume that spinlocks can be initialized via memcpy: | |
1479 | */ | |
1480 | spin_lock_init(&ptr->lock); | |
1481 | ||
e498be7d | 1482 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1483 | ptr; |
1da177e4 LT |
1484 | local_irq_enable(); |
1485 | } | |
e498be7d CL |
1486 | /* 5) Replace the bootstrap kmem_list3's */ |
1487 | { | |
1488 | int node; | |
1489 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1490 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1491 | numa_node_id()); |
e498be7d CL |
1492 | |
1493 | for_each_online_node(node) { | |
1494 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1495 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1496 | |
1497 | if (INDEX_AC != INDEX_L3) { | |
1498 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1499 | &initkmem_list3[SIZE_L3 + node], |
1500 | node); | |
e498be7d CL |
1501 | } |
1502 | } | |
1503 | } | |
1da177e4 | 1504 | |
e498be7d | 1505 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1506 | { |
343e0d7a | 1507 | struct kmem_cache *cachep; |
fc0abb14 | 1508 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1509 | list_for_each_entry(cachep, &cache_chain, next) |
2ed3a4ef CL |
1510 | if (enable_cpucache(cachep)) |
1511 | BUG(); | |
fc0abb14 | 1512 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1513 | } |
1514 | ||
056c6241 RT |
1515 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1516 | init_lock_keys(); | |
1517 | ||
1518 | ||
1da177e4 LT |
1519 | /* Done! */ |
1520 | g_cpucache_up = FULL; | |
1521 | ||
a737b3e2 AM |
1522 | /* |
1523 | * Register a cpu startup notifier callback that initializes | |
1524 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1525 | */ |
1526 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1527 | |
a737b3e2 AM |
1528 | /* |
1529 | * The reap timers are started later, with a module init call: That part | |
1530 | * of the kernel is not yet operational. | |
1da177e4 LT |
1531 | */ |
1532 | } | |
1533 | ||
1534 | static int __init cpucache_init(void) | |
1535 | { | |
1536 | int cpu; | |
1537 | ||
a737b3e2 AM |
1538 | /* |
1539 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1540 | */ |
e498be7d | 1541 | for_each_online_cpu(cpu) |
a737b3e2 | 1542 | start_cpu_timer(cpu); |
1da177e4 LT |
1543 | return 0; |
1544 | } | |
1da177e4 LT |
1545 | __initcall(cpucache_init); |
1546 | ||
1547 | /* | |
1548 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1549 | * | |
1550 | * If we requested dmaable memory, we will get it. Even if we | |
1551 | * did not request dmaable memory, we might get it, but that | |
1552 | * would be relatively rare and ignorable. | |
1553 | */ | |
343e0d7a | 1554 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1555 | { |
1556 | struct page *page; | |
e1b6aa6f | 1557 | int nr_pages; |
1da177e4 LT |
1558 | int i; |
1559 | ||
d6fef9da | 1560 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1561 | /* |
1562 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1563 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1564 | */ |
e1b6aa6f | 1565 | flags |= __GFP_COMP; |
d6fef9da | 1566 | #endif |
e1b6aa6f CH |
1567 | flags |= cachep->gfpflags; |
1568 | ||
1569 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
1da177e4 LT |
1570 | if (!page) |
1571 | return NULL; | |
1da177e4 | 1572 | |
e1b6aa6f | 1573 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1574 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1575 | add_zone_page_state(page_zone(page), |
1576 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1577 | else | |
1578 | add_zone_page_state(page_zone(page), | |
1579 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1580 | for (i = 0; i < nr_pages; i++) |
1581 | __SetPageSlab(page + i); | |
1582 | return page_address(page); | |
1da177e4 LT |
1583 | } |
1584 | ||
1585 | /* | |
1586 | * Interface to system's page release. | |
1587 | */ | |
343e0d7a | 1588 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1589 | { |
b28a02de | 1590 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1591 | struct page *page = virt_to_page(addr); |
1592 | const unsigned long nr_freed = i; | |
1593 | ||
972d1a7b CL |
1594 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1595 | sub_zone_page_state(page_zone(page), | |
1596 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1597 | else | |
1598 | sub_zone_page_state(page_zone(page), | |
1599 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1600 | while (i--) { |
f205b2fe NP |
1601 | BUG_ON(!PageSlab(page)); |
1602 | __ClearPageSlab(page); | |
1da177e4 LT |
1603 | page++; |
1604 | } | |
1da177e4 LT |
1605 | if (current->reclaim_state) |
1606 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1607 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1608 | } |
1609 | ||
1610 | static void kmem_rcu_free(struct rcu_head *head) | |
1611 | { | |
b28a02de | 1612 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1613 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1614 | |
1615 | kmem_freepages(cachep, slab_rcu->addr); | |
1616 | if (OFF_SLAB(cachep)) | |
1617 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1618 | } | |
1619 | ||
1620 | #if DEBUG | |
1621 | ||
1622 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1623 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1624 | unsigned long caller) |
1da177e4 | 1625 | { |
3dafccf2 | 1626 | int size = obj_size(cachep); |
1da177e4 | 1627 | |
3dafccf2 | 1628 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1629 | |
b28a02de | 1630 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1631 | return; |
1632 | ||
b28a02de PE |
1633 | *addr++ = 0x12345678; |
1634 | *addr++ = caller; | |
1635 | *addr++ = smp_processor_id(); | |
1636 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1637 | { |
1638 | unsigned long *sptr = &caller; | |
1639 | unsigned long svalue; | |
1640 | ||
1641 | while (!kstack_end(sptr)) { | |
1642 | svalue = *sptr++; | |
1643 | if (kernel_text_address(svalue)) { | |
b28a02de | 1644 | *addr++ = svalue; |
1da177e4 LT |
1645 | size -= sizeof(unsigned long); |
1646 | if (size <= sizeof(unsigned long)) | |
1647 | break; | |
1648 | } | |
1649 | } | |
1650 | ||
1651 | } | |
b28a02de | 1652 | *addr++ = 0x87654321; |
1da177e4 LT |
1653 | } |
1654 | #endif | |
1655 | ||
343e0d7a | 1656 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1657 | { |
3dafccf2 MS |
1658 | int size = obj_size(cachep); |
1659 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1660 | |
1661 | memset(addr, val, size); | |
b28a02de | 1662 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1663 | } |
1664 | ||
1665 | static void dump_line(char *data, int offset, int limit) | |
1666 | { | |
1667 | int i; | |
1668 | printk(KERN_ERR "%03x:", offset); | |
a737b3e2 | 1669 | for (i = 0; i < limit; i++) |
b28a02de | 1670 | printk(" %02x", (unsigned char)data[offset + i]); |
1da177e4 LT |
1671 | printk("\n"); |
1672 | } | |
1673 | #endif | |
1674 | ||
1675 | #if DEBUG | |
1676 | ||
343e0d7a | 1677 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1678 | { |
1679 | int i, size; | |
1680 | char *realobj; | |
1681 | ||
1682 | if (cachep->flags & SLAB_RED_ZONE) { | |
1683 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
a737b3e2 AM |
1684 | *dbg_redzone1(cachep, objp), |
1685 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1686 | } |
1687 | ||
1688 | if (cachep->flags & SLAB_STORE_USER) { | |
1689 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1690 | *dbg_userword(cachep, objp)); |
1da177e4 | 1691 | print_symbol("(%s)", |
a737b3e2 | 1692 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1693 | printk("\n"); |
1694 | } | |
3dafccf2 MS |
1695 | realobj = (char *)objp + obj_offset(cachep); |
1696 | size = obj_size(cachep); | |
b28a02de | 1697 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1698 | int limit; |
1699 | limit = 16; | |
b28a02de PE |
1700 | if (i + limit > size) |
1701 | limit = size - i; | |
1da177e4 LT |
1702 | dump_line(realobj, i, limit); |
1703 | } | |
1704 | } | |
1705 | ||
343e0d7a | 1706 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1707 | { |
1708 | char *realobj; | |
1709 | int size, i; | |
1710 | int lines = 0; | |
1711 | ||
3dafccf2 MS |
1712 | realobj = (char *)objp + obj_offset(cachep); |
1713 | size = obj_size(cachep); | |
1da177e4 | 1714 | |
b28a02de | 1715 | for (i = 0; i < size; i++) { |
1da177e4 | 1716 | char exp = POISON_FREE; |
b28a02de | 1717 | if (i == size - 1) |
1da177e4 LT |
1718 | exp = POISON_END; |
1719 | if (realobj[i] != exp) { | |
1720 | int limit; | |
1721 | /* Mismatch ! */ | |
1722 | /* Print header */ | |
1723 | if (lines == 0) { | |
b28a02de | 1724 | printk(KERN_ERR |
a737b3e2 AM |
1725 | "Slab corruption: start=%p, len=%d\n", |
1726 | realobj, size); | |
1da177e4 LT |
1727 | print_objinfo(cachep, objp, 0); |
1728 | } | |
1729 | /* Hexdump the affected line */ | |
b28a02de | 1730 | i = (i / 16) * 16; |
1da177e4 | 1731 | limit = 16; |
b28a02de PE |
1732 | if (i + limit > size) |
1733 | limit = size - i; | |
1da177e4 LT |
1734 | dump_line(realobj, i, limit); |
1735 | i += 16; | |
1736 | lines++; | |
1737 | /* Limit to 5 lines */ | |
1738 | if (lines > 5) | |
1739 | break; | |
1740 | } | |
1741 | } | |
1742 | if (lines != 0) { | |
1743 | /* Print some data about the neighboring objects, if they | |
1744 | * exist: | |
1745 | */ | |
6ed5eb22 | 1746 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1747 | unsigned int objnr; |
1da177e4 | 1748 | |
8fea4e96 | 1749 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1750 | if (objnr) { |
8fea4e96 | 1751 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1752 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1753 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1754 | realobj, size); |
1da177e4 LT |
1755 | print_objinfo(cachep, objp, 2); |
1756 | } | |
b28a02de | 1757 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1758 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1759 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1760 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1761 | realobj, size); |
1da177e4 LT |
1762 | print_objinfo(cachep, objp, 2); |
1763 | } | |
1764 | } | |
1765 | } | |
1766 | #endif | |
1767 | ||
12dd36fa MD |
1768 | #if DEBUG |
1769 | /** | |
911851e6 RD |
1770 | * slab_destroy_objs - destroy a slab and its objects |
1771 | * @cachep: cache pointer being destroyed | |
1772 | * @slabp: slab pointer being destroyed | |
1773 | * | |
1774 | * Call the registered destructor for each object in a slab that is being | |
1775 | * destroyed. | |
1da177e4 | 1776 | */ |
343e0d7a | 1777 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1778 | { |
1da177e4 LT |
1779 | int i; |
1780 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1781 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1782 | |
1783 | if (cachep->flags & SLAB_POISON) { | |
1784 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1785 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1786 | OFF_SLAB(cachep)) | |
b28a02de | 1787 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1788 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1789 | else |
1790 | check_poison_obj(cachep, objp); | |
1791 | #else | |
1792 | check_poison_obj(cachep, objp); | |
1793 | #endif | |
1794 | } | |
1795 | if (cachep->flags & SLAB_RED_ZONE) { | |
1796 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1797 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1798 | "was overwritten"); |
1da177e4 LT |
1799 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1800 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1801 | "was overwritten"); |
1da177e4 LT |
1802 | } |
1803 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1804 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1805 | } |
12dd36fa | 1806 | } |
1da177e4 | 1807 | #else |
343e0d7a | 1808 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1809 | { |
1da177e4 LT |
1810 | if (cachep->dtor) { |
1811 | int i; | |
1812 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1813 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1814 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1815 | } |
1816 | } | |
12dd36fa | 1817 | } |
1da177e4 LT |
1818 | #endif |
1819 | ||
911851e6 RD |
1820 | /** |
1821 | * slab_destroy - destroy and release all objects in a slab | |
1822 | * @cachep: cache pointer being destroyed | |
1823 | * @slabp: slab pointer being destroyed | |
1824 | * | |
12dd36fa | 1825 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1826 | * Before calling the slab must have been unlinked from the cache. The |
1827 | * cache-lock is not held/needed. | |
12dd36fa | 1828 | */ |
343e0d7a | 1829 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1830 | { |
1831 | void *addr = slabp->s_mem - slabp->colouroff; | |
1832 | ||
1833 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1834 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1835 | struct slab_rcu *slab_rcu; | |
1836 | ||
b28a02de | 1837 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1838 | slab_rcu->cachep = cachep; |
1839 | slab_rcu->addr = addr; | |
1840 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1841 | } else { | |
1842 | kmem_freepages(cachep, addr); | |
873623df IM |
1843 | if (OFF_SLAB(cachep)) |
1844 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1845 | } |
1846 | } | |
1847 | ||
a737b3e2 AM |
1848 | /* |
1849 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1850 | * size of kmem_list3. | |
1851 | */ | |
343e0d7a | 1852 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1853 | { |
1854 | int node; | |
1855 | ||
1856 | for_each_online_node(node) { | |
b28a02de | 1857 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1858 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1859 | REAPTIMEOUT_LIST3 + |
1860 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1861 | } |
1862 | } | |
1863 | ||
117f6eb1 CL |
1864 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1865 | { | |
1866 | int i; | |
1867 | struct kmem_list3 *l3; | |
1868 | ||
1869 | for_each_online_cpu(i) | |
1870 | kfree(cachep->array[i]); | |
1871 | ||
1872 | /* NUMA: free the list3 structures */ | |
1873 | for_each_online_node(i) { | |
1874 | l3 = cachep->nodelists[i]; | |
1875 | if (l3) { | |
1876 | kfree(l3->shared); | |
1877 | free_alien_cache(l3->alien); | |
1878 | kfree(l3); | |
1879 | } | |
1880 | } | |
1881 | kmem_cache_free(&cache_cache, cachep); | |
1882 | } | |
1883 | ||
1884 | ||
4d268eba | 1885 | /** |
a70773dd RD |
1886 | * calculate_slab_order - calculate size (page order) of slabs |
1887 | * @cachep: pointer to the cache that is being created | |
1888 | * @size: size of objects to be created in this cache. | |
1889 | * @align: required alignment for the objects. | |
1890 | * @flags: slab allocation flags | |
1891 | * | |
1892 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1893 | * |
1894 | * This could be made much more intelligent. For now, try to avoid using | |
1895 | * high order pages for slabs. When the gfp() functions are more friendly | |
1896 | * towards high-order requests, this should be changed. | |
1897 | */ | |
a737b3e2 | 1898 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1899 | size_t size, size_t align, unsigned long flags) |
4d268eba | 1900 | { |
b1ab41c4 | 1901 | unsigned long offslab_limit; |
4d268eba | 1902 | size_t left_over = 0; |
9888e6fa | 1903 | int gfporder; |
4d268eba | 1904 | |
a737b3e2 | 1905 | for (gfporder = 0; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1906 | unsigned int num; |
1907 | size_t remainder; | |
1908 | ||
9888e6fa | 1909 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1910 | if (!num) |
1911 | continue; | |
9888e6fa | 1912 | |
b1ab41c4 IM |
1913 | if (flags & CFLGS_OFF_SLAB) { |
1914 | /* | |
1915 | * Max number of objs-per-slab for caches which | |
1916 | * use off-slab slabs. Needed to avoid a possible | |
1917 | * looping condition in cache_grow(). | |
1918 | */ | |
1919 | offslab_limit = size - sizeof(struct slab); | |
1920 | offslab_limit /= sizeof(kmem_bufctl_t); | |
1921 | ||
1922 | if (num > offslab_limit) | |
1923 | break; | |
1924 | } | |
4d268eba | 1925 | |
9888e6fa | 1926 | /* Found something acceptable - save it away */ |
4d268eba | 1927 | cachep->num = num; |
9888e6fa | 1928 | cachep->gfporder = gfporder; |
4d268eba PE |
1929 | left_over = remainder; |
1930 | ||
f78bb8ad LT |
1931 | /* |
1932 | * A VFS-reclaimable slab tends to have most allocations | |
1933 | * as GFP_NOFS and we really don't want to have to be allocating | |
1934 | * higher-order pages when we are unable to shrink dcache. | |
1935 | */ | |
1936 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1937 | break; | |
1938 | ||
4d268eba PE |
1939 | /* |
1940 | * Large number of objects is good, but very large slabs are | |
1941 | * currently bad for the gfp()s. | |
1942 | */ | |
9888e6fa | 1943 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1944 | break; |
1945 | ||
9888e6fa LT |
1946 | /* |
1947 | * Acceptable internal fragmentation? | |
1948 | */ | |
a737b3e2 | 1949 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
1950 | break; |
1951 | } | |
1952 | return left_over; | |
1953 | } | |
1954 | ||
2ed3a4ef | 1955 | static int setup_cpu_cache(struct kmem_cache *cachep) |
f30cf7d1 | 1956 | { |
2ed3a4ef CL |
1957 | if (g_cpucache_up == FULL) |
1958 | return enable_cpucache(cachep); | |
1959 | ||
f30cf7d1 PE |
1960 | if (g_cpucache_up == NONE) { |
1961 | /* | |
1962 | * Note: the first kmem_cache_create must create the cache | |
1963 | * that's used by kmalloc(24), otherwise the creation of | |
1964 | * further caches will BUG(). | |
1965 | */ | |
1966 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
1967 | ||
1968 | /* | |
1969 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
1970 | * the first cache, then we need to set up all its list3s, | |
1971 | * otherwise the creation of further caches will BUG(). | |
1972 | */ | |
1973 | set_up_list3s(cachep, SIZE_AC); | |
1974 | if (INDEX_AC == INDEX_L3) | |
1975 | g_cpucache_up = PARTIAL_L3; | |
1976 | else | |
1977 | g_cpucache_up = PARTIAL_AC; | |
1978 | } else { | |
1979 | cachep->array[smp_processor_id()] = | |
1980 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
1981 | ||
1982 | if (g_cpucache_up == PARTIAL_AC) { | |
1983 | set_up_list3s(cachep, SIZE_L3); | |
1984 | g_cpucache_up = PARTIAL_L3; | |
1985 | } else { | |
1986 | int node; | |
1987 | for_each_online_node(node) { | |
1988 | cachep->nodelists[node] = | |
1989 | kmalloc_node(sizeof(struct kmem_list3), | |
1990 | GFP_KERNEL, node); | |
1991 | BUG_ON(!cachep->nodelists[node]); | |
1992 | kmem_list3_init(cachep->nodelists[node]); | |
1993 | } | |
1994 | } | |
1995 | } | |
1996 | cachep->nodelists[numa_node_id()]->next_reap = | |
1997 | jiffies + REAPTIMEOUT_LIST3 + | |
1998 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1999 | ||
2000 | cpu_cache_get(cachep)->avail = 0; | |
2001 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2002 | cpu_cache_get(cachep)->batchcount = 1; | |
2003 | cpu_cache_get(cachep)->touched = 0; | |
2004 | cachep->batchcount = 1; | |
2005 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2006 | return 0; |
f30cf7d1 PE |
2007 | } |
2008 | ||
1da177e4 LT |
2009 | /** |
2010 | * kmem_cache_create - Create a cache. | |
2011 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2012 | * @size: The size of objects to be created in this cache. | |
2013 | * @align: The required alignment for the objects. | |
2014 | * @flags: SLAB flags | |
2015 | * @ctor: A constructor for the objects. | |
2016 | * @dtor: A destructor for the objects. | |
2017 | * | |
2018 | * Returns a ptr to the cache on success, NULL on failure. | |
2019 | * Cannot be called within a int, but can be interrupted. | |
2020 | * The @ctor is run when new pages are allocated by the cache | |
2021 | * and the @dtor is run before the pages are handed back. | |
2022 | * | |
2023 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2024 | * the module calling this has to destroy the cache before getting unloaded. |
2025 | * | |
1da177e4 LT |
2026 | * The flags are |
2027 | * | |
2028 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2029 | * to catch references to uninitialised memory. | |
2030 | * | |
2031 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2032 | * for buffer overruns. | |
2033 | * | |
1da177e4 LT |
2034 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2035 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2036 | * as davem. | |
2037 | */ | |
343e0d7a | 2038 | struct kmem_cache * |
1da177e4 | 2039 | kmem_cache_create (const char *name, size_t size, size_t align, |
a737b3e2 AM |
2040 | unsigned long flags, |
2041 | void (*ctor)(void*, struct kmem_cache *, unsigned long), | |
343e0d7a | 2042 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) |
1da177e4 LT |
2043 | { |
2044 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2045 | struct kmem_cache *cachep = NULL, *pc; |
1da177e4 LT |
2046 | |
2047 | /* | |
2048 | * Sanity checks... these are all serious usage bugs. | |
2049 | */ | |
a737b3e2 | 2050 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
b28a02de | 2051 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { |
a737b3e2 AM |
2052 | printk(KERN_ERR "%s: Early error in slab %s\n", __FUNCTION__, |
2053 | name); | |
b28a02de PE |
2054 | BUG(); |
2055 | } | |
1da177e4 | 2056 | |
f0188f47 RT |
2057 | /* |
2058 | * Prevent CPUs from coming and going. | |
2059 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
2060 | */ | |
2061 | lock_cpu_hotplug(); | |
2062 | ||
fc0abb14 | 2063 | mutex_lock(&cache_chain_mutex); |
4f12bb4f | 2064 | |
7a7c381d | 2065 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2066 | mm_segment_t old_fs = get_fs(); |
2067 | char tmp; | |
2068 | int res; | |
2069 | ||
2070 | /* | |
2071 | * This happens when the module gets unloaded and doesn't | |
2072 | * destroy its slab cache and no-one else reuses the vmalloc | |
2073 | * area of the module. Print a warning. | |
2074 | */ | |
2075 | set_fs(KERNEL_DS); | |
2076 | res = __get_user(tmp, pc->name); | |
2077 | set_fs(old_fs); | |
2078 | if (res) { | |
2079 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2080 | pc->buffer_size); |
4f12bb4f AM |
2081 | continue; |
2082 | } | |
2083 | ||
b28a02de | 2084 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
2085 | printk("kmem_cache_create: duplicate cache %s\n", name); |
2086 | dump_stack(); | |
2087 | goto oops; | |
2088 | } | |
2089 | } | |
2090 | ||
1da177e4 LT |
2091 | #if DEBUG |
2092 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
2093 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
2094 | /* No constructor, but inital state check requested */ | |
2095 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 2096 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
2097 | flags &= ~SLAB_DEBUG_INITIAL; |
2098 | } | |
1da177e4 LT |
2099 | #if FORCED_DEBUG |
2100 | /* | |
2101 | * Enable redzoning and last user accounting, except for caches with | |
2102 | * large objects, if the increased size would increase the object size | |
2103 | * above the next power of two: caches with object sizes just above a | |
2104 | * power of two have a significant amount of internal fragmentation. | |
2105 | */ | |
a737b3e2 | 2106 | if (size < 4096 || fls(size - 1) == fls(size-1 + 3 * BYTES_PER_WORD)) |
b28a02de | 2107 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2108 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2109 | flags |= SLAB_POISON; | |
2110 | #endif | |
2111 | if (flags & SLAB_DESTROY_BY_RCU) | |
2112 | BUG_ON(flags & SLAB_POISON); | |
2113 | #endif | |
2114 | if (flags & SLAB_DESTROY_BY_RCU) | |
2115 | BUG_ON(dtor); | |
2116 | ||
2117 | /* | |
a737b3e2 AM |
2118 | * Always checks flags, a caller might be expecting debug support which |
2119 | * isn't available. | |
1da177e4 | 2120 | */ |
40094fa6 | 2121 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2122 | |
a737b3e2 AM |
2123 | /* |
2124 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2125 | * unaligned accesses for some archs when redzoning is used, and makes |
2126 | * sure any on-slab bufctl's are also correctly aligned. | |
2127 | */ | |
b28a02de PE |
2128 | if (size & (BYTES_PER_WORD - 1)) { |
2129 | size += (BYTES_PER_WORD - 1); | |
2130 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2131 | } |
2132 | ||
a737b3e2 AM |
2133 | /* calculate the final buffer alignment: */ |
2134 | ||
1da177e4 LT |
2135 | /* 1) arch recommendation: can be overridden for debug */ |
2136 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2137 | /* |
2138 | * Default alignment: as specified by the arch code. Except if | |
2139 | * an object is really small, then squeeze multiple objects into | |
2140 | * one cacheline. | |
1da177e4 LT |
2141 | */ |
2142 | ralign = cache_line_size(); | |
b28a02de | 2143 | while (size <= ralign / 2) |
1da177e4 LT |
2144 | ralign /= 2; |
2145 | } else { | |
2146 | ralign = BYTES_PER_WORD; | |
2147 | } | |
ca5f9703 PE |
2148 | |
2149 | /* | |
2150 | * Redzoning and user store require word alignment. Note this will be | |
2151 | * overridden by architecture or caller mandated alignment if either | |
2152 | * is greater than BYTES_PER_WORD. | |
2153 | */ | |
2154 | if (flags & SLAB_RED_ZONE || flags & SLAB_STORE_USER) | |
2155 | ralign = BYTES_PER_WORD; | |
2156 | ||
1da177e4 LT |
2157 | /* 2) arch mandated alignment: disables debug if necessary */ |
2158 | if (ralign < ARCH_SLAB_MINALIGN) { | |
2159 | ralign = ARCH_SLAB_MINALIGN; | |
2160 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2161 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
2162 | } |
2163 | /* 3) caller mandated alignment: disables debug if necessary */ | |
2164 | if (ralign < align) { | |
2165 | ralign = align; | |
2166 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 2167 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 | 2168 | } |
a737b3e2 | 2169 | /* |
ca5f9703 | 2170 | * 4) Store it. |
1da177e4 LT |
2171 | */ |
2172 | align = ralign; | |
2173 | ||
2174 | /* Get cache's description obj. */ | |
c5e3b83e | 2175 | cachep = kmem_cache_zalloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 2176 | if (!cachep) |
4f12bb4f | 2177 | goto oops; |
1da177e4 LT |
2178 | |
2179 | #if DEBUG | |
3dafccf2 | 2180 | cachep->obj_size = size; |
1da177e4 | 2181 | |
ca5f9703 PE |
2182 | /* |
2183 | * Both debugging options require word-alignment which is calculated | |
2184 | * into align above. | |
2185 | */ | |
1da177e4 | 2186 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2187 | /* add space for red zone words */ |
3dafccf2 | 2188 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 2189 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
2190 | } |
2191 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 PE |
2192 | /* user store requires one word storage behind the end of |
2193 | * the real object. | |
1da177e4 | 2194 | */ |
1da177e4 LT |
2195 | size += BYTES_PER_WORD; |
2196 | } | |
2197 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2198 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2199 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2200 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2201 | size = PAGE_SIZE; |
2202 | } | |
2203 | #endif | |
2204 | #endif | |
2205 | ||
e0a42726 IM |
2206 | /* |
2207 | * Determine if the slab management is 'on' or 'off' slab. | |
2208 | * (bootstrapping cannot cope with offslab caches so don't do | |
2209 | * it too early on.) | |
2210 | */ | |
2211 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2212 | /* |
2213 | * Size is large, assume best to place the slab management obj | |
2214 | * off-slab (should allow better packing of objs). | |
2215 | */ | |
2216 | flags |= CFLGS_OFF_SLAB; | |
2217 | ||
2218 | size = ALIGN(size, align); | |
2219 | ||
f78bb8ad | 2220 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2221 | |
2222 | if (!cachep->num) { | |
2223 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
2224 | kmem_cache_free(&cache_cache, cachep); | |
2225 | cachep = NULL; | |
4f12bb4f | 2226 | goto oops; |
1da177e4 | 2227 | } |
b28a02de PE |
2228 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2229 | + sizeof(struct slab), align); | |
1da177e4 LT |
2230 | |
2231 | /* | |
2232 | * If the slab has been placed off-slab, and we have enough space then | |
2233 | * move it on-slab. This is at the expense of any extra colouring. | |
2234 | */ | |
2235 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2236 | flags &= ~CFLGS_OFF_SLAB; | |
2237 | left_over -= slab_size; | |
2238 | } | |
2239 | ||
2240 | if (flags & CFLGS_OFF_SLAB) { | |
2241 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2242 | slab_size = |
2243 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2244 | } |
2245 | ||
2246 | cachep->colour_off = cache_line_size(); | |
2247 | /* Offset must be a multiple of the alignment. */ | |
2248 | if (cachep->colour_off < align) | |
2249 | cachep->colour_off = align; | |
b28a02de | 2250 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2251 | cachep->slab_size = slab_size; |
2252 | cachep->flags = flags; | |
2253 | cachep->gfpflags = 0; | |
2254 | if (flags & SLAB_CACHE_DMA) | |
2255 | cachep->gfpflags |= GFP_DMA; | |
3dafccf2 | 2256 | cachep->buffer_size = size; |
1da177e4 | 2257 | |
e5ac9c5a | 2258 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2259 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2260 | /* |
2261 | * This is a possibility for one of the malloc_sizes caches. | |
2262 | * But since we go off slab only for object size greater than | |
2263 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2264 | * this should not happen at all. | |
2265 | * But leave a BUG_ON for some lucky dude. | |
2266 | */ | |
2267 | BUG_ON(!cachep->slabp_cache); | |
2268 | } | |
1da177e4 LT |
2269 | cachep->ctor = ctor; |
2270 | cachep->dtor = dtor; | |
2271 | cachep->name = name; | |
2272 | ||
2ed3a4ef CL |
2273 | if (setup_cpu_cache(cachep)) { |
2274 | __kmem_cache_destroy(cachep); | |
2275 | cachep = NULL; | |
2276 | goto oops; | |
2277 | } | |
1da177e4 | 2278 | |
1da177e4 LT |
2279 | /* cache setup completed, link it into the list */ |
2280 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2281 | oops: |
1da177e4 LT |
2282 | if (!cachep && (flags & SLAB_PANIC)) |
2283 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2284 | name); |
fc0abb14 | 2285 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 2286 | unlock_cpu_hotplug(); |
1da177e4 LT |
2287 | return cachep; |
2288 | } | |
2289 | EXPORT_SYMBOL(kmem_cache_create); | |
2290 | ||
2291 | #if DEBUG | |
2292 | static void check_irq_off(void) | |
2293 | { | |
2294 | BUG_ON(!irqs_disabled()); | |
2295 | } | |
2296 | ||
2297 | static void check_irq_on(void) | |
2298 | { | |
2299 | BUG_ON(irqs_disabled()); | |
2300 | } | |
2301 | ||
343e0d7a | 2302 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2303 | { |
2304 | #ifdef CONFIG_SMP | |
2305 | check_irq_off(); | |
e498be7d | 2306 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2307 | #endif |
2308 | } | |
e498be7d | 2309 | |
343e0d7a | 2310 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2311 | { |
2312 | #ifdef CONFIG_SMP | |
2313 | check_irq_off(); | |
2314 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2315 | #endif | |
2316 | } | |
2317 | ||
1da177e4 LT |
2318 | #else |
2319 | #define check_irq_off() do { } while(0) | |
2320 | #define check_irq_on() do { } while(0) | |
2321 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2322 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2323 | #endif |
2324 | ||
aab2207c CL |
2325 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2326 | struct array_cache *ac, | |
2327 | int force, int node); | |
2328 | ||
1da177e4 LT |
2329 | static void do_drain(void *arg) |
2330 | { | |
a737b3e2 | 2331 | struct kmem_cache *cachep = arg; |
1da177e4 | 2332 | struct array_cache *ac; |
ff69416e | 2333 | int node = numa_node_id(); |
1da177e4 LT |
2334 | |
2335 | check_irq_off(); | |
9a2dba4b | 2336 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2337 | spin_lock(&cachep->nodelists[node]->list_lock); |
2338 | free_block(cachep, ac->entry, ac->avail, node); | |
2339 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2340 | ac->avail = 0; |
2341 | } | |
2342 | ||
343e0d7a | 2343 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2344 | { |
e498be7d CL |
2345 | struct kmem_list3 *l3; |
2346 | int node; | |
2347 | ||
a07fa394 | 2348 | on_each_cpu(do_drain, cachep, 1, 1); |
1da177e4 | 2349 | check_irq_on(); |
b28a02de | 2350 | for_each_online_node(node) { |
e498be7d | 2351 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2352 | if (l3 && l3->alien) |
2353 | drain_alien_cache(cachep, l3->alien); | |
2354 | } | |
2355 | ||
2356 | for_each_online_node(node) { | |
2357 | l3 = cachep->nodelists[node]; | |
2358 | if (l3) | |
aab2207c | 2359 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2360 | } |
1da177e4 LT |
2361 | } |
2362 | ||
ed11d9eb CL |
2363 | /* |
2364 | * Remove slabs from the list of free slabs. | |
2365 | * Specify the number of slabs to drain in tofree. | |
2366 | * | |
2367 | * Returns the actual number of slabs released. | |
2368 | */ | |
2369 | static int drain_freelist(struct kmem_cache *cache, | |
2370 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2371 | { |
ed11d9eb CL |
2372 | struct list_head *p; |
2373 | int nr_freed; | |
1da177e4 | 2374 | struct slab *slabp; |
1da177e4 | 2375 | |
ed11d9eb CL |
2376 | nr_freed = 0; |
2377 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2378 | |
ed11d9eb | 2379 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2380 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2381 | if (p == &l3->slabs_free) { |
2382 | spin_unlock_irq(&l3->list_lock); | |
2383 | goto out; | |
2384 | } | |
1da177e4 | 2385 | |
ed11d9eb | 2386 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2387 | #if DEBUG |
40094fa6 | 2388 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2389 | #endif |
2390 | list_del(&slabp->list); | |
ed11d9eb CL |
2391 | /* |
2392 | * Safe to drop the lock. The slab is no longer linked | |
2393 | * to the cache. | |
2394 | */ | |
2395 | l3->free_objects -= cache->num; | |
e498be7d | 2396 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2397 | slab_destroy(cache, slabp); |
2398 | nr_freed++; | |
1da177e4 | 2399 | } |
ed11d9eb CL |
2400 | out: |
2401 | return nr_freed; | |
1da177e4 LT |
2402 | } |
2403 | ||
343e0d7a | 2404 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2405 | { |
2406 | int ret = 0, i = 0; | |
2407 | struct kmem_list3 *l3; | |
2408 | ||
2409 | drain_cpu_caches(cachep); | |
2410 | ||
2411 | check_irq_on(); | |
2412 | for_each_online_node(i) { | |
2413 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2414 | if (!l3) |
2415 | continue; | |
2416 | ||
2417 | drain_freelist(cachep, l3, l3->free_objects); | |
2418 | ||
2419 | ret += !list_empty(&l3->slabs_full) || | |
2420 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2421 | } |
2422 | return (ret ? 1 : 0); | |
2423 | } | |
2424 | ||
1da177e4 LT |
2425 | /** |
2426 | * kmem_cache_shrink - Shrink a cache. | |
2427 | * @cachep: The cache to shrink. | |
2428 | * | |
2429 | * Releases as many slabs as possible for a cache. | |
2430 | * To help debugging, a zero exit status indicates all slabs were released. | |
2431 | */ | |
343e0d7a | 2432 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2433 | { |
40094fa6 | 2434 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2435 | |
2436 | return __cache_shrink(cachep); | |
2437 | } | |
2438 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2439 | ||
2440 | /** | |
2441 | * kmem_cache_destroy - delete a cache | |
2442 | * @cachep: the cache to destroy | |
2443 | * | |
343e0d7a | 2444 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2445 | * |
2446 | * It is expected this function will be called by a module when it is | |
2447 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2448 | * cache being allocated each time a module is loaded and unloaded, if the | |
2449 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2450 | * | |
2451 | * The cache must be empty before calling this function. | |
2452 | * | |
2453 | * The caller must guarantee that noone will allocate memory from the cache | |
2454 | * during the kmem_cache_destroy(). | |
2455 | */ | |
133d205a | 2456 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2457 | { |
40094fa6 | 2458 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 LT |
2459 | |
2460 | /* Don't let CPUs to come and go */ | |
2461 | lock_cpu_hotplug(); | |
2462 | ||
2463 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2464 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2465 | /* |
2466 | * the chain is never empty, cache_cache is never destroyed | |
2467 | */ | |
2468 | list_del(&cachep->next); | |
fc0abb14 | 2469 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2470 | |
2471 | if (__cache_shrink(cachep)) { | |
2472 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2473 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2474 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2475 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 2476 | unlock_cpu_hotplug(); |
133d205a | 2477 | return; |
1da177e4 LT |
2478 | } |
2479 | ||
2480 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2481 | synchronize_rcu(); |
1da177e4 | 2482 | |
117f6eb1 | 2483 | __kmem_cache_destroy(cachep); |
1da177e4 | 2484 | unlock_cpu_hotplug(); |
1da177e4 LT |
2485 | } |
2486 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2487 | ||
e5ac9c5a RT |
2488 | /* |
2489 | * Get the memory for a slab management obj. | |
2490 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2491 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2492 | * come from the same cache which is getting created because, | |
2493 | * when we are searching for an appropriate cache for these | |
2494 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2495 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2496 | * kmem_find_general_cachep till the initialization is complete. | |
2497 | * Hence we cannot have slabp_cache same as the original cache. | |
2498 | */ | |
343e0d7a | 2499 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2500 | int colour_off, gfp_t local_flags, |
2501 | int nodeid) | |
1da177e4 LT |
2502 | { |
2503 | struct slab *slabp; | |
b28a02de | 2504 | |
1da177e4 LT |
2505 | if (OFF_SLAB(cachep)) { |
2506 | /* Slab management obj is off-slab. */ | |
5b74ada7 RT |
2507 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
2508 | local_flags, nodeid); | |
1da177e4 LT |
2509 | if (!slabp) |
2510 | return NULL; | |
2511 | } else { | |
b28a02de | 2512 | slabp = objp + colour_off; |
1da177e4 LT |
2513 | colour_off += cachep->slab_size; |
2514 | } | |
2515 | slabp->inuse = 0; | |
2516 | slabp->colouroff = colour_off; | |
b28a02de | 2517 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2518 | slabp->nodeid = nodeid; |
1da177e4 LT |
2519 | return slabp; |
2520 | } | |
2521 | ||
2522 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2523 | { | |
b28a02de | 2524 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2525 | } |
2526 | ||
343e0d7a | 2527 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2528 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2529 | { |
2530 | int i; | |
2531 | ||
2532 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2533 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2534 | #if DEBUG |
2535 | /* need to poison the objs? */ | |
2536 | if (cachep->flags & SLAB_POISON) | |
2537 | poison_obj(cachep, objp, POISON_FREE); | |
2538 | if (cachep->flags & SLAB_STORE_USER) | |
2539 | *dbg_userword(cachep, objp) = NULL; | |
2540 | ||
2541 | if (cachep->flags & SLAB_RED_ZONE) { | |
2542 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2543 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2544 | } | |
2545 | /* | |
a737b3e2 AM |
2546 | * Constructors are not allowed to allocate memory from the same |
2547 | * cache which they are a constructor for. Otherwise, deadlock. | |
2548 | * They must also be threaded. | |
1da177e4 LT |
2549 | */ |
2550 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2551 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2552 | ctor_flags); |
1da177e4 LT |
2553 | |
2554 | if (cachep->flags & SLAB_RED_ZONE) { | |
2555 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2556 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2557 | " end of an object"); |
1da177e4 LT |
2558 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2559 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2560 | " start of an object"); |
1da177e4 | 2561 | } |
a737b3e2 AM |
2562 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2563 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2564 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2565 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2566 | #else |
2567 | if (cachep->ctor) | |
2568 | cachep->ctor(objp, cachep, ctor_flags); | |
2569 | #endif | |
b28a02de | 2570 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2571 | } |
b28a02de | 2572 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2573 | slabp->free = 0; |
2574 | } | |
2575 | ||
343e0d7a | 2576 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2577 | { |
a737b3e2 AM |
2578 | if (flags & SLAB_DMA) |
2579 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2580 | else | |
2581 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
1da177e4 LT |
2582 | } |
2583 | ||
a737b3e2 AM |
2584 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2585 | int nodeid) | |
78d382d7 | 2586 | { |
8fea4e96 | 2587 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2588 | kmem_bufctl_t next; |
2589 | ||
2590 | slabp->inuse++; | |
2591 | next = slab_bufctl(slabp)[slabp->free]; | |
2592 | #if DEBUG | |
2593 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2594 | WARN_ON(slabp->nodeid != nodeid); | |
2595 | #endif | |
2596 | slabp->free = next; | |
2597 | ||
2598 | return objp; | |
2599 | } | |
2600 | ||
a737b3e2 AM |
2601 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2602 | void *objp, int nodeid) | |
78d382d7 | 2603 | { |
8fea4e96 | 2604 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2605 | |
2606 | #if DEBUG | |
2607 | /* Verify that the slab belongs to the intended node */ | |
2608 | WARN_ON(slabp->nodeid != nodeid); | |
2609 | ||
871751e2 | 2610 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2611 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2612 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2613 | BUG(); |
2614 | } | |
2615 | #endif | |
2616 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2617 | slabp->free = objnr; | |
2618 | slabp->inuse--; | |
2619 | } | |
2620 | ||
4776874f PE |
2621 | /* |
2622 | * Map pages beginning at addr to the given cache and slab. This is required | |
2623 | * for the slab allocator to be able to lookup the cache and slab of a | |
2624 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2625 | */ | |
2626 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2627 | void *addr) | |
1da177e4 | 2628 | { |
4776874f | 2629 | int nr_pages; |
1da177e4 LT |
2630 | struct page *page; |
2631 | ||
4776874f | 2632 | page = virt_to_page(addr); |
84097518 | 2633 | |
4776874f | 2634 | nr_pages = 1; |
84097518 | 2635 | if (likely(!PageCompound(page))) |
4776874f PE |
2636 | nr_pages <<= cache->gfporder; |
2637 | ||
1da177e4 | 2638 | do { |
4776874f PE |
2639 | page_set_cache(page, cache); |
2640 | page_set_slab(page, slab); | |
1da177e4 | 2641 | page++; |
4776874f | 2642 | } while (--nr_pages); |
1da177e4 LT |
2643 | } |
2644 | ||
2645 | /* | |
2646 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2647 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2648 | */ | |
343e0d7a | 2649 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2650 | { |
b28a02de PE |
2651 | struct slab *slabp; |
2652 | void *objp; | |
2653 | size_t offset; | |
2654 | gfp_t local_flags; | |
2655 | unsigned long ctor_flags; | |
e498be7d | 2656 | struct kmem_list3 *l3; |
1da177e4 | 2657 | |
a737b3e2 AM |
2658 | /* |
2659 | * Be lazy and only check for valid flags here, keeping it out of the | |
2660 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2661 | */ |
40094fa6 | 2662 | BUG_ON(flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)); |
1da177e4 LT |
2663 | if (flags & SLAB_NO_GROW) |
2664 | return 0; | |
2665 | ||
2666 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2667 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2668 | if (!(local_flags & __GFP_WAIT)) | |
2669 | /* | |
2670 | * Not allowed to sleep. Need to tell a constructor about | |
2671 | * this - it might need to know... | |
2672 | */ | |
2673 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2674 | ||
2e1217cf | 2675 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2676 | check_irq_off(); |
2e1217cf RT |
2677 | l3 = cachep->nodelists[nodeid]; |
2678 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2679 | |
2680 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2681 | offset = l3->colour_next; |
2682 | l3->colour_next++; | |
2683 | if (l3->colour_next >= cachep->colour) | |
2684 | l3->colour_next = 0; | |
2685 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2686 | |
2e1217cf | 2687 | offset *= cachep->colour_off; |
1da177e4 LT |
2688 | |
2689 | if (local_flags & __GFP_WAIT) | |
2690 | local_irq_enable(); | |
2691 | ||
2692 | /* | |
2693 | * The test for missing atomic flag is performed here, rather than | |
2694 | * the more obvious place, simply to reduce the critical path length | |
2695 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2696 | * will eventually be caught here (where it matters). | |
2697 | */ | |
2698 | kmem_flagcheck(cachep, flags); | |
2699 | ||
a737b3e2 AM |
2700 | /* |
2701 | * Get mem for the objs. Attempt to allocate a physical page from | |
2702 | * 'nodeid'. | |
e498be7d | 2703 | */ |
a737b3e2 AM |
2704 | objp = kmem_getpages(cachep, flags, nodeid); |
2705 | if (!objp) | |
1da177e4 LT |
2706 | goto failed; |
2707 | ||
2708 | /* Get slab management. */ | |
5b74ada7 | 2709 | slabp = alloc_slabmgmt(cachep, objp, offset, local_flags, nodeid); |
a737b3e2 | 2710 | if (!slabp) |
1da177e4 LT |
2711 | goto opps1; |
2712 | ||
e498be7d | 2713 | slabp->nodeid = nodeid; |
4776874f | 2714 | slab_map_pages(cachep, slabp, objp); |
1da177e4 LT |
2715 | |
2716 | cache_init_objs(cachep, slabp, ctor_flags); | |
2717 | ||
2718 | if (local_flags & __GFP_WAIT) | |
2719 | local_irq_disable(); | |
2720 | check_irq_off(); | |
e498be7d | 2721 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2722 | |
2723 | /* Make slab active. */ | |
e498be7d | 2724 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2725 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2726 | l3->free_objects += cachep->num; |
2727 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2728 | return 1; |
a737b3e2 | 2729 | opps1: |
1da177e4 | 2730 | kmem_freepages(cachep, objp); |
a737b3e2 | 2731 | failed: |
1da177e4 LT |
2732 | if (local_flags & __GFP_WAIT) |
2733 | local_irq_disable(); | |
2734 | return 0; | |
2735 | } | |
2736 | ||
2737 | #if DEBUG | |
2738 | ||
2739 | /* | |
2740 | * Perform extra freeing checks: | |
2741 | * - detect bad pointers. | |
2742 | * - POISON/RED_ZONE checking | |
2743 | * - destructor calls, for caches with POISON+dtor | |
2744 | */ | |
2745 | static void kfree_debugcheck(const void *objp) | |
2746 | { | |
2747 | struct page *page; | |
2748 | ||
2749 | if (!virt_addr_valid(objp)) { | |
2750 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2751 | (unsigned long)objp); |
2752 | BUG(); | |
1da177e4 LT |
2753 | } |
2754 | page = virt_to_page(objp); | |
2755 | if (!PageSlab(page)) { | |
b28a02de PE |
2756 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2757 | (unsigned long)objp); | |
1da177e4 LT |
2758 | BUG(); |
2759 | } | |
2760 | } | |
2761 | ||
58ce1fd5 PE |
2762 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2763 | { | |
2764 | unsigned long redzone1, redzone2; | |
2765 | ||
2766 | redzone1 = *dbg_redzone1(cache, obj); | |
2767 | redzone2 = *dbg_redzone2(cache, obj); | |
2768 | ||
2769 | /* | |
2770 | * Redzone is ok. | |
2771 | */ | |
2772 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2773 | return; | |
2774 | ||
2775 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2776 | slab_error(cache, "double free detected"); | |
2777 | else | |
2778 | slab_error(cache, "memory outside object was overwritten"); | |
2779 | ||
2780 | printk(KERN_ERR "%p: redzone 1:0x%lx, redzone 2:0x%lx.\n", | |
2781 | obj, redzone1, redzone2); | |
2782 | } | |
2783 | ||
343e0d7a | 2784 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2785 | void *caller) |
1da177e4 LT |
2786 | { |
2787 | struct page *page; | |
2788 | unsigned int objnr; | |
2789 | struct slab *slabp; | |
2790 | ||
3dafccf2 | 2791 | objp -= obj_offset(cachep); |
1da177e4 LT |
2792 | kfree_debugcheck(objp); |
2793 | page = virt_to_page(objp); | |
2794 | ||
065d41cb | 2795 | slabp = page_get_slab(page); |
1da177e4 LT |
2796 | |
2797 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2798 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2799 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2800 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2801 | } | |
2802 | if (cachep->flags & SLAB_STORE_USER) | |
2803 | *dbg_userword(cachep, objp) = caller; | |
2804 | ||
8fea4e96 | 2805 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2806 | |
2807 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2808 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2809 | |
2810 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
a737b3e2 AM |
2811 | /* |
2812 | * Need to call the slab's constructor so the caller can | |
2813 | * perform a verify of its state (debugging). Called without | |
2814 | * the cache-lock held. | |
1da177e4 | 2815 | */ |
3dafccf2 | 2816 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2817 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2818 | } |
2819 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2820 | /* we want to cache poison the object, | |
2821 | * call the destruction callback | |
2822 | */ | |
3dafccf2 | 2823 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 2824 | } |
871751e2 AV |
2825 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2826 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2827 | #endif | |
1da177e4 LT |
2828 | if (cachep->flags & SLAB_POISON) { |
2829 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2830 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2831 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2832 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2833 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2834 | } else { |
2835 | poison_obj(cachep, objp, POISON_FREE); | |
2836 | } | |
2837 | #else | |
2838 | poison_obj(cachep, objp, POISON_FREE); | |
2839 | #endif | |
2840 | } | |
2841 | return objp; | |
2842 | } | |
2843 | ||
343e0d7a | 2844 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2845 | { |
2846 | kmem_bufctl_t i; | |
2847 | int entries = 0; | |
b28a02de | 2848 | |
1da177e4 LT |
2849 | /* Check slab's freelist to see if this obj is there. */ |
2850 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2851 | entries++; | |
2852 | if (entries > cachep->num || i >= cachep->num) | |
2853 | goto bad; | |
2854 | } | |
2855 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2856 | bad: |
2857 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2858 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2859 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2860 | for (i = 0; |
264132bc | 2861 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2862 | i++) { |
a737b3e2 | 2863 | if (i % 16 == 0) |
1da177e4 | 2864 | printk("\n%03x:", i); |
b28a02de | 2865 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2866 | } |
2867 | printk("\n"); | |
2868 | BUG(); | |
2869 | } | |
2870 | } | |
2871 | #else | |
2872 | #define kfree_debugcheck(x) do { } while(0) | |
2873 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2874 | #define check_slabp(x,y) do { } while(0) | |
2875 | #endif | |
2876 | ||
343e0d7a | 2877 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2878 | { |
2879 | int batchcount; | |
2880 | struct kmem_list3 *l3; | |
2881 | struct array_cache *ac; | |
2882 | ||
2883 | check_irq_off(); | |
9a2dba4b | 2884 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2885 | retry: |
1da177e4 LT |
2886 | batchcount = ac->batchcount; |
2887 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2888 | /* |
2889 | * If there was little recent activity on this cache, then | |
2890 | * perform only a partial refill. Otherwise we could generate | |
2891 | * refill bouncing. | |
1da177e4 LT |
2892 | */ |
2893 | batchcount = BATCHREFILL_LIMIT; | |
2894 | } | |
e498be7d CL |
2895 | l3 = cachep->nodelists[numa_node_id()]; |
2896 | ||
2897 | BUG_ON(ac->avail > 0 || !l3); | |
2898 | spin_lock(&l3->list_lock); | |
1da177e4 | 2899 | |
3ded175a CL |
2900 | /* See if we can refill from the shared array */ |
2901 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2902 | goto alloc_done; | |
2903 | ||
1da177e4 LT |
2904 | while (batchcount > 0) { |
2905 | struct list_head *entry; | |
2906 | struct slab *slabp; | |
2907 | /* Get slab alloc is to come from. */ | |
2908 | entry = l3->slabs_partial.next; | |
2909 | if (entry == &l3->slabs_partial) { | |
2910 | l3->free_touched = 1; | |
2911 | entry = l3->slabs_free.next; | |
2912 | if (entry == &l3->slabs_free) | |
2913 | goto must_grow; | |
2914 | } | |
2915 | ||
2916 | slabp = list_entry(entry, struct slab, list); | |
2917 | check_slabp(cachep, slabp); | |
2918 | check_spinlock_acquired(cachep); | |
2919 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2920 | STATS_INC_ALLOCED(cachep); |
2921 | STATS_INC_ACTIVE(cachep); | |
2922 | STATS_SET_HIGH(cachep); | |
2923 | ||
78d382d7 MD |
2924 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2925 | numa_node_id()); | |
1da177e4 LT |
2926 | } |
2927 | check_slabp(cachep, slabp); | |
2928 | ||
2929 | /* move slabp to correct slabp list: */ | |
2930 | list_del(&slabp->list); | |
2931 | if (slabp->free == BUFCTL_END) | |
2932 | list_add(&slabp->list, &l3->slabs_full); | |
2933 | else | |
2934 | list_add(&slabp->list, &l3->slabs_partial); | |
2935 | } | |
2936 | ||
a737b3e2 | 2937 | must_grow: |
1da177e4 | 2938 | l3->free_objects -= ac->avail; |
a737b3e2 | 2939 | alloc_done: |
e498be7d | 2940 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2941 | |
2942 | if (unlikely(!ac->avail)) { | |
2943 | int x; | |
e498be7d CL |
2944 | x = cache_grow(cachep, flags, numa_node_id()); |
2945 | ||
a737b3e2 | 2946 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 2947 | ac = cpu_cache_get(cachep); |
a737b3e2 | 2948 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
2949 | return NULL; |
2950 | ||
a737b3e2 | 2951 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
2952 | goto retry; |
2953 | } | |
2954 | ac->touched = 1; | |
e498be7d | 2955 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2956 | } |
2957 | ||
a737b3e2 AM |
2958 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
2959 | gfp_t flags) | |
1da177e4 LT |
2960 | { |
2961 | might_sleep_if(flags & __GFP_WAIT); | |
2962 | #if DEBUG | |
2963 | kmem_flagcheck(cachep, flags); | |
2964 | #endif | |
2965 | } | |
2966 | ||
2967 | #if DEBUG | |
a737b3e2 AM |
2968 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
2969 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 2970 | { |
b28a02de | 2971 | if (!objp) |
1da177e4 | 2972 | return objp; |
b28a02de | 2973 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2974 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2975 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2976 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2977 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2978 | else |
2979 | check_poison_obj(cachep, objp); | |
2980 | #else | |
2981 | check_poison_obj(cachep, objp); | |
2982 | #endif | |
2983 | poison_obj(cachep, objp, POISON_INUSE); | |
2984 | } | |
2985 | if (cachep->flags & SLAB_STORE_USER) | |
2986 | *dbg_userword(cachep, objp) = caller; | |
2987 | ||
2988 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2989 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
2990 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2991 | slab_error(cachep, "double free, or memory outside" | |
2992 | " object was overwritten"); | |
b28a02de | 2993 | printk(KERN_ERR |
a737b3e2 AM |
2994 | "%p: redzone 1:0x%lx, redzone 2:0x%lx\n", |
2995 | objp, *dbg_redzone1(cachep, objp), | |
2996 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2997 | } |
2998 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2999 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3000 | } | |
871751e2 AV |
3001 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3002 | { | |
3003 | struct slab *slabp; | |
3004 | unsigned objnr; | |
3005 | ||
3006 | slabp = page_get_slab(virt_to_page(objp)); | |
3007 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; | |
3008 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3009 | } | |
3010 | #endif | |
3dafccf2 | 3011 | objp += obj_offset(cachep); |
1da177e4 | 3012 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 3013 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
3014 | |
3015 | if (!(flags & __GFP_WAIT)) | |
3016 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
3017 | ||
3018 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 3019 | } |
1da177e4 LT |
3020 | return objp; |
3021 | } | |
3022 | #else | |
3023 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3024 | #endif | |
3025 | ||
343e0d7a | 3026 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3027 | { |
b28a02de | 3028 | void *objp; |
1da177e4 LT |
3029 | struct array_cache *ac; |
3030 | ||
dc85da15 | 3031 | #ifdef CONFIG_NUMA |
b2455396 | 3032 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { |
c61afb18 PJ |
3033 | objp = alternate_node_alloc(cachep, flags); |
3034 | if (objp != NULL) | |
3035 | return objp; | |
dc85da15 CL |
3036 | } |
3037 | #endif | |
3038 | ||
5c382300 | 3039 | check_irq_off(); |
9a2dba4b | 3040 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3041 | if (likely(ac->avail)) { |
3042 | STATS_INC_ALLOCHIT(cachep); | |
3043 | ac->touched = 1; | |
e498be7d | 3044 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3045 | } else { |
3046 | STATS_INC_ALLOCMISS(cachep); | |
3047 | objp = cache_alloc_refill(cachep, flags); | |
3048 | } | |
5c382300 AK |
3049 | return objp; |
3050 | } | |
3051 | ||
a737b3e2 AM |
3052 | static __always_inline void *__cache_alloc(struct kmem_cache *cachep, |
3053 | gfp_t flags, void *caller) | |
5c382300 AK |
3054 | { |
3055 | unsigned long save_flags; | |
b28a02de | 3056 | void *objp; |
5c382300 AK |
3057 | |
3058 | cache_alloc_debugcheck_before(cachep, flags); | |
3059 | ||
3060 | local_irq_save(save_flags); | |
3061 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 3062 | local_irq_restore(save_flags); |
34342e86 | 3063 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 3064 | caller); |
34342e86 | 3065 | prefetchw(objp); |
1da177e4 LT |
3066 | return objp; |
3067 | } | |
3068 | ||
e498be7d | 3069 | #ifdef CONFIG_NUMA |
c61afb18 | 3070 | /* |
b2455396 | 3071 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3072 | * |
3073 | * If we are in_interrupt, then process context, including cpusets and | |
3074 | * mempolicy, may not apply and should not be used for allocation policy. | |
3075 | */ | |
3076 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3077 | { | |
3078 | int nid_alloc, nid_here; | |
3079 | ||
3080 | if (in_interrupt()) | |
3081 | return NULL; | |
3082 | nid_alloc = nid_here = numa_node_id(); | |
3083 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3084 | nid_alloc = cpuset_mem_spread_node(); | |
3085 | else if (current->mempolicy) | |
3086 | nid_alloc = slab_node(current->mempolicy); | |
3087 | if (nid_alloc != nid_here) | |
3088 | return __cache_alloc_node(cachep, flags, nid_alloc); | |
3089 | return NULL; | |
3090 | } | |
3091 | ||
e498be7d CL |
3092 | /* |
3093 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3094 | */ |
a737b3e2 AM |
3095 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
3096 | int nodeid) | |
e498be7d CL |
3097 | { |
3098 | struct list_head *entry; | |
b28a02de PE |
3099 | struct slab *slabp; |
3100 | struct kmem_list3 *l3; | |
3101 | void *obj; | |
b28a02de PE |
3102 | int x; |
3103 | ||
3104 | l3 = cachep->nodelists[nodeid]; | |
3105 | BUG_ON(!l3); | |
3106 | ||
a737b3e2 | 3107 | retry: |
ca3b9b91 | 3108 | check_irq_off(); |
b28a02de PE |
3109 | spin_lock(&l3->list_lock); |
3110 | entry = l3->slabs_partial.next; | |
3111 | if (entry == &l3->slabs_partial) { | |
3112 | l3->free_touched = 1; | |
3113 | entry = l3->slabs_free.next; | |
3114 | if (entry == &l3->slabs_free) | |
3115 | goto must_grow; | |
3116 | } | |
3117 | ||
3118 | slabp = list_entry(entry, struct slab, list); | |
3119 | check_spinlock_acquired_node(cachep, nodeid); | |
3120 | check_slabp(cachep, slabp); | |
3121 | ||
3122 | STATS_INC_NODEALLOCS(cachep); | |
3123 | STATS_INC_ACTIVE(cachep); | |
3124 | STATS_SET_HIGH(cachep); | |
3125 | ||
3126 | BUG_ON(slabp->inuse == cachep->num); | |
3127 | ||
78d382d7 | 3128 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3129 | check_slabp(cachep, slabp); |
3130 | l3->free_objects--; | |
3131 | /* move slabp to correct slabp list: */ | |
3132 | list_del(&slabp->list); | |
3133 | ||
a737b3e2 | 3134 | if (slabp->free == BUFCTL_END) |
b28a02de | 3135 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3136 | else |
b28a02de | 3137 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3138 | |
b28a02de PE |
3139 | spin_unlock(&l3->list_lock); |
3140 | goto done; | |
e498be7d | 3141 | |
a737b3e2 | 3142 | must_grow: |
b28a02de PE |
3143 | spin_unlock(&l3->list_lock); |
3144 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 3145 | |
b28a02de PE |
3146 | if (!x) |
3147 | return NULL; | |
e498be7d | 3148 | |
b28a02de | 3149 | goto retry; |
a737b3e2 | 3150 | done: |
b28a02de | 3151 | return obj; |
e498be7d CL |
3152 | } |
3153 | #endif | |
3154 | ||
3155 | /* | |
3156 | * Caller needs to acquire correct kmem_list's list_lock | |
3157 | */ | |
343e0d7a | 3158 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3159 | int node) |
1da177e4 LT |
3160 | { |
3161 | int i; | |
e498be7d | 3162 | struct kmem_list3 *l3; |
1da177e4 LT |
3163 | |
3164 | for (i = 0; i < nr_objects; i++) { | |
3165 | void *objp = objpp[i]; | |
3166 | struct slab *slabp; | |
1da177e4 | 3167 | |
6ed5eb22 | 3168 | slabp = virt_to_slab(objp); |
ff69416e | 3169 | l3 = cachep->nodelists[node]; |
1da177e4 | 3170 | list_del(&slabp->list); |
ff69416e | 3171 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3172 | check_slabp(cachep, slabp); |
78d382d7 | 3173 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3174 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3175 | l3->free_objects++; |
1da177e4 LT |
3176 | check_slabp(cachep, slabp); |
3177 | ||
3178 | /* fixup slab chains */ | |
3179 | if (slabp->inuse == 0) { | |
e498be7d CL |
3180 | if (l3->free_objects > l3->free_limit) { |
3181 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3182 | /* No need to drop any previously held |
3183 | * lock here, even if we have a off-slab slab | |
3184 | * descriptor it is guaranteed to come from | |
3185 | * a different cache, refer to comments before | |
3186 | * alloc_slabmgmt. | |
3187 | */ | |
1da177e4 LT |
3188 | slab_destroy(cachep, slabp); |
3189 | } else { | |
e498be7d | 3190 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3191 | } |
3192 | } else { | |
3193 | /* Unconditionally move a slab to the end of the | |
3194 | * partial list on free - maximum time for the | |
3195 | * other objects to be freed, too. | |
3196 | */ | |
e498be7d | 3197 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3198 | } |
3199 | } | |
3200 | } | |
3201 | ||
343e0d7a | 3202 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3203 | { |
3204 | int batchcount; | |
e498be7d | 3205 | struct kmem_list3 *l3; |
ff69416e | 3206 | int node = numa_node_id(); |
1da177e4 LT |
3207 | |
3208 | batchcount = ac->batchcount; | |
3209 | #if DEBUG | |
3210 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3211 | #endif | |
3212 | check_irq_off(); | |
ff69416e | 3213 | l3 = cachep->nodelists[node]; |
873623df | 3214 | spin_lock(&l3->list_lock); |
e498be7d CL |
3215 | if (l3->shared) { |
3216 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3217 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3218 | if (max) { |
3219 | if (batchcount > max) | |
3220 | batchcount = max; | |
e498be7d | 3221 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3222 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3223 | shared_array->avail += batchcount; |
3224 | goto free_done; | |
3225 | } | |
3226 | } | |
3227 | ||
ff69416e | 3228 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3229 | free_done: |
1da177e4 LT |
3230 | #if STATS |
3231 | { | |
3232 | int i = 0; | |
3233 | struct list_head *p; | |
3234 | ||
e498be7d CL |
3235 | p = l3->slabs_free.next; |
3236 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3237 | struct slab *slabp; |
3238 | ||
3239 | slabp = list_entry(p, struct slab, list); | |
3240 | BUG_ON(slabp->inuse); | |
3241 | ||
3242 | i++; | |
3243 | p = p->next; | |
3244 | } | |
3245 | STATS_SET_FREEABLE(cachep, i); | |
3246 | } | |
3247 | #endif | |
e498be7d | 3248 | spin_unlock(&l3->list_lock); |
1da177e4 | 3249 | ac->avail -= batchcount; |
a737b3e2 | 3250 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3251 | } |
3252 | ||
3253 | /* | |
a737b3e2 AM |
3254 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3255 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3256 | */ |
873623df | 3257 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3258 | { |
9a2dba4b | 3259 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3260 | |
3261 | check_irq_off(); | |
3262 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3263 | ||
873623df | 3264 | if (cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3265 | return; |
3266 | ||
1da177e4 LT |
3267 | if (likely(ac->avail < ac->limit)) { |
3268 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3269 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3270 | return; |
3271 | } else { | |
3272 | STATS_INC_FREEMISS(cachep); | |
3273 | cache_flusharray(cachep, ac); | |
e498be7d | 3274 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3275 | } |
3276 | } | |
3277 | ||
3278 | /** | |
3279 | * kmem_cache_alloc - Allocate an object | |
3280 | * @cachep: The cache to allocate from. | |
3281 | * @flags: See kmalloc(). | |
3282 | * | |
3283 | * Allocate an object from this cache. The flags are only relevant | |
3284 | * if the cache has no available objects. | |
3285 | */ | |
343e0d7a | 3286 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3287 | { |
7fd6b141 | 3288 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3289 | } |
3290 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3291 | ||
a8c0f9a4 | 3292 | /** |
b8008b2b | 3293 | * kmem_cache_zalloc - Allocate an object. The memory is set to zero. |
a8c0f9a4 PE |
3294 | * @cache: The cache to allocate from. |
3295 | * @flags: See kmalloc(). | |
3296 | * | |
3297 | * Allocate an object from this cache and set the allocated memory to zero. | |
3298 | * The flags are only relevant if the cache has no available objects. | |
3299 | */ | |
3300 | void *kmem_cache_zalloc(struct kmem_cache *cache, gfp_t flags) | |
3301 | { | |
3302 | void *ret = __cache_alloc(cache, flags, __builtin_return_address(0)); | |
3303 | if (ret) | |
3304 | memset(ret, 0, obj_size(cache)); | |
3305 | return ret; | |
3306 | } | |
3307 | EXPORT_SYMBOL(kmem_cache_zalloc); | |
3308 | ||
1da177e4 LT |
3309 | /** |
3310 | * kmem_ptr_validate - check if an untrusted pointer might | |
3311 | * be a slab entry. | |
3312 | * @cachep: the cache we're checking against | |
3313 | * @ptr: pointer to validate | |
3314 | * | |
3315 | * This verifies that the untrusted pointer looks sane: | |
3316 | * it is _not_ a guarantee that the pointer is actually | |
3317 | * part of the slab cache in question, but it at least | |
3318 | * validates that the pointer can be dereferenced and | |
3319 | * looks half-way sane. | |
3320 | * | |
3321 | * Currently only used for dentry validation. | |
3322 | */ | |
343e0d7a | 3323 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3324 | { |
b28a02de | 3325 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3326 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3327 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3328 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3329 | struct page *page; |
3330 | ||
3331 | if (unlikely(addr < min_addr)) | |
3332 | goto out; | |
3333 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3334 | goto out; | |
3335 | if (unlikely(addr & align_mask)) | |
3336 | goto out; | |
3337 | if (unlikely(!kern_addr_valid(addr))) | |
3338 | goto out; | |
3339 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3340 | goto out; | |
3341 | page = virt_to_page(ptr); | |
3342 | if (unlikely(!PageSlab(page))) | |
3343 | goto out; | |
065d41cb | 3344 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3345 | goto out; |
3346 | return 1; | |
a737b3e2 | 3347 | out: |
1da177e4 LT |
3348 | return 0; |
3349 | } | |
3350 | ||
3351 | #ifdef CONFIG_NUMA | |
3352 | /** | |
3353 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3354 | * @cachep: The cache to allocate from. | |
3355 | * @flags: See kmalloc(). | |
3356 | * @nodeid: node number of the target node. | |
3357 | * | |
3358 | * Identical to kmem_cache_alloc, except that this function is slow | |
3359 | * and can sleep. And it will allocate memory on the given node, which | |
3360 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3361 | * New and improved: it will now make sure that the object gets |
3362 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3363 | */ |
343e0d7a | 3364 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3365 | { |
e498be7d CL |
3366 | unsigned long save_flags; |
3367 | void *ptr; | |
1da177e4 | 3368 | |
e498be7d CL |
3369 | cache_alloc_debugcheck_before(cachep, flags); |
3370 | local_irq_save(save_flags); | |
18f820f6 CL |
3371 | |
3372 | if (nodeid == -1 || nodeid == numa_node_id() || | |
a737b3e2 | 3373 | !cachep->nodelists[nodeid]) |
5c382300 AK |
3374 | ptr = ____cache_alloc(cachep, flags); |
3375 | else | |
3376 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3377 | local_irq_restore(save_flags); |
18f820f6 CL |
3378 | |
3379 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3380 | __builtin_return_address(0)); | |
1da177e4 | 3381 | |
e498be7d | 3382 | return ptr; |
1da177e4 LT |
3383 | } |
3384 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3385 | ||
dbe5e69d | 3386 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3387 | { |
343e0d7a | 3388 | struct kmem_cache *cachep; |
97e2bde4 MS |
3389 | |
3390 | cachep = kmem_find_general_cachep(size, flags); | |
3391 | if (unlikely(cachep == NULL)) | |
3392 | return NULL; | |
3393 | return kmem_cache_alloc_node(cachep, flags, node); | |
3394 | } | |
dbe5e69d | 3395 | EXPORT_SYMBOL(__kmalloc_node); |
1da177e4 LT |
3396 | #endif |
3397 | ||
3398 | /** | |
800590f5 | 3399 | * __do_kmalloc - allocate memory |
1da177e4 | 3400 | * @size: how many bytes of memory are required. |
800590f5 | 3401 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3402 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3403 | */ |
7fd6b141 PE |
3404 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3405 | void *caller) | |
1da177e4 | 3406 | { |
343e0d7a | 3407 | struct kmem_cache *cachep; |
1da177e4 | 3408 | |
97e2bde4 MS |
3409 | /* If you want to save a few bytes .text space: replace |
3410 | * __ with kmem_. | |
3411 | * Then kmalloc uses the uninlined functions instead of the inline | |
3412 | * functions. | |
3413 | */ | |
3414 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3415 | if (unlikely(cachep == NULL)) |
3416 | return NULL; | |
7fd6b141 PE |
3417 | return __cache_alloc(cachep, flags, caller); |
3418 | } | |
3419 | ||
7fd6b141 PE |
3420 | |
3421 | void *__kmalloc(size_t size, gfp_t flags) | |
3422 | { | |
871751e2 | 3423 | #ifndef CONFIG_DEBUG_SLAB |
7fd6b141 | 3424 | return __do_kmalloc(size, flags, NULL); |
871751e2 AV |
3425 | #else |
3426 | return __do_kmalloc(size, flags, __builtin_return_address(0)); | |
3427 | #endif | |
1da177e4 LT |
3428 | } |
3429 | EXPORT_SYMBOL(__kmalloc); | |
3430 | ||
871751e2 | 3431 | #ifdef CONFIG_DEBUG_SLAB |
7fd6b141 PE |
3432 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) |
3433 | { | |
3434 | return __do_kmalloc(size, flags, caller); | |
3435 | } | |
3436 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
7fd6b141 PE |
3437 | #endif |
3438 | ||
1da177e4 LT |
3439 | /** |
3440 | * kmem_cache_free - Deallocate an object | |
3441 | * @cachep: The cache the allocation was from. | |
3442 | * @objp: The previously allocated object. | |
3443 | * | |
3444 | * Free an object which was previously allocated from this | |
3445 | * cache. | |
3446 | */ | |
343e0d7a | 3447 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3448 | { |
3449 | unsigned long flags; | |
3450 | ||
ddc2e812 PE |
3451 | BUG_ON(virt_to_cache(objp) != cachep); |
3452 | ||
1da177e4 | 3453 | local_irq_save(flags); |
873623df | 3454 | __cache_free(cachep, objp); |
1da177e4 LT |
3455 | local_irq_restore(flags); |
3456 | } | |
3457 | EXPORT_SYMBOL(kmem_cache_free); | |
3458 | ||
1da177e4 LT |
3459 | /** |
3460 | * kfree - free previously allocated memory | |
3461 | * @objp: pointer returned by kmalloc. | |
3462 | * | |
80e93eff PE |
3463 | * If @objp is NULL, no operation is performed. |
3464 | * | |
1da177e4 LT |
3465 | * Don't free memory not originally allocated by kmalloc() |
3466 | * or you will run into trouble. | |
3467 | */ | |
3468 | void kfree(const void *objp) | |
3469 | { | |
343e0d7a | 3470 | struct kmem_cache *c; |
1da177e4 LT |
3471 | unsigned long flags; |
3472 | ||
3473 | if (unlikely(!objp)) | |
3474 | return; | |
3475 | local_irq_save(flags); | |
3476 | kfree_debugcheck(objp); | |
6ed5eb22 | 3477 | c = virt_to_cache(objp); |
f9b8404c | 3478 | debug_check_no_locks_freed(objp, obj_size(c)); |
873623df | 3479 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3480 | local_irq_restore(flags); |
3481 | } | |
3482 | EXPORT_SYMBOL(kfree); | |
3483 | ||
343e0d7a | 3484 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3485 | { |
3dafccf2 | 3486 | return obj_size(cachep); |
1da177e4 LT |
3487 | } |
3488 | EXPORT_SYMBOL(kmem_cache_size); | |
3489 | ||
343e0d7a | 3490 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3491 | { |
3492 | return cachep->name; | |
3493 | } | |
3494 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3495 | ||
e498be7d | 3496 | /* |
0718dc2a | 3497 | * This initializes kmem_list3 or resizes varioius caches for all nodes. |
e498be7d | 3498 | */ |
343e0d7a | 3499 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3500 | { |
3501 | int node; | |
3502 | struct kmem_list3 *l3; | |
cafeb02e CL |
3503 | struct array_cache *new_shared; |
3504 | struct array_cache **new_alien; | |
e498be7d CL |
3505 | |
3506 | for_each_online_node(node) { | |
cafeb02e | 3507 | |
a737b3e2 AM |
3508 | new_alien = alloc_alien_cache(node, cachep->limit); |
3509 | if (!new_alien) | |
e498be7d | 3510 | goto fail; |
cafeb02e | 3511 | |
0718dc2a CL |
3512 | new_shared = alloc_arraycache(node, |
3513 | cachep->shared*cachep->batchcount, | |
a737b3e2 | 3514 | 0xbaadf00d); |
0718dc2a CL |
3515 | if (!new_shared) { |
3516 | free_alien_cache(new_alien); | |
e498be7d | 3517 | goto fail; |
0718dc2a | 3518 | } |
cafeb02e | 3519 | |
a737b3e2 AM |
3520 | l3 = cachep->nodelists[node]; |
3521 | if (l3) { | |
cafeb02e CL |
3522 | struct array_cache *shared = l3->shared; |
3523 | ||
e498be7d CL |
3524 | spin_lock_irq(&l3->list_lock); |
3525 | ||
cafeb02e | 3526 | if (shared) |
0718dc2a CL |
3527 | free_block(cachep, shared->entry, |
3528 | shared->avail, node); | |
e498be7d | 3529 | |
cafeb02e CL |
3530 | l3->shared = new_shared; |
3531 | if (!l3->alien) { | |
e498be7d CL |
3532 | l3->alien = new_alien; |
3533 | new_alien = NULL; | |
3534 | } | |
b28a02de | 3535 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3536 | cachep->batchcount + cachep->num; |
e498be7d | 3537 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3538 | kfree(shared); |
e498be7d CL |
3539 | free_alien_cache(new_alien); |
3540 | continue; | |
3541 | } | |
a737b3e2 | 3542 | l3 = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, node); |
0718dc2a CL |
3543 | if (!l3) { |
3544 | free_alien_cache(new_alien); | |
3545 | kfree(new_shared); | |
e498be7d | 3546 | goto fail; |
0718dc2a | 3547 | } |
e498be7d CL |
3548 | |
3549 | kmem_list3_init(l3); | |
3550 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3551 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3552 | l3->shared = new_shared; |
e498be7d | 3553 | l3->alien = new_alien; |
b28a02de | 3554 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3555 | cachep->batchcount + cachep->num; |
e498be7d CL |
3556 | cachep->nodelists[node] = l3; |
3557 | } | |
cafeb02e | 3558 | return 0; |
0718dc2a | 3559 | |
a737b3e2 | 3560 | fail: |
0718dc2a CL |
3561 | if (!cachep->next.next) { |
3562 | /* Cache is not active yet. Roll back what we did */ | |
3563 | node--; | |
3564 | while (node >= 0) { | |
3565 | if (cachep->nodelists[node]) { | |
3566 | l3 = cachep->nodelists[node]; | |
3567 | ||
3568 | kfree(l3->shared); | |
3569 | free_alien_cache(l3->alien); | |
3570 | kfree(l3); | |
3571 | cachep->nodelists[node] = NULL; | |
3572 | } | |
3573 | node--; | |
3574 | } | |
3575 | } | |
cafeb02e | 3576 | return -ENOMEM; |
e498be7d CL |
3577 | } |
3578 | ||
1da177e4 | 3579 | struct ccupdate_struct { |
343e0d7a | 3580 | struct kmem_cache *cachep; |
1da177e4 LT |
3581 | struct array_cache *new[NR_CPUS]; |
3582 | }; | |
3583 | ||
3584 | static void do_ccupdate_local(void *info) | |
3585 | { | |
a737b3e2 | 3586 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3587 | struct array_cache *old; |
3588 | ||
3589 | check_irq_off(); | |
9a2dba4b | 3590 | old = cpu_cache_get(new->cachep); |
e498be7d | 3591 | |
1da177e4 LT |
3592 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3593 | new->new[smp_processor_id()] = old; | |
3594 | } | |
3595 | ||
b5d8ca7c | 3596 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 AM |
3597 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3598 | int batchcount, int shared) | |
1da177e4 | 3599 | { |
d2e7b7d0 | 3600 | struct ccupdate_struct *new; |
2ed3a4ef | 3601 | int i; |
1da177e4 | 3602 | |
d2e7b7d0 SS |
3603 | new = kzalloc(sizeof(*new), GFP_KERNEL); |
3604 | if (!new) | |
3605 | return -ENOMEM; | |
3606 | ||
e498be7d | 3607 | for_each_online_cpu(i) { |
d2e7b7d0 | 3608 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
a737b3e2 | 3609 | batchcount); |
d2e7b7d0 | 3610 | if (!new->new[i]) { |
b28a02de | 3611 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3612 | kfree(new->new[i]); |
3613 | kfree(new); | |
e498be7d | 3614 | return -ENOMEM; |
1da177e4 LT |
3615 | } |
3616 | } | |
d2e7b7d0 | 3617 | new->cachep = cachep; |
1da177e4 | 3618 | |
d2e7b7d0 | 3619 | on_each_cpu(do_ccupdate_local, (void *)new, 1, 1); |
e498be7d | 3620 | |
1da177e4 | 3621 | check_irq_on(); |
1da177e4 LT |
3622 | cachep->batchcount = batchcount; |
3623 | cachep->limit = limit; | |
e498be7d | 3624 | cachep->shared = shared; |
1da177e4 | 3625 | |
e498be7d | 3626 | for_each_online_cpu(i) { |
d2e7b7d0 | 3627 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3628 | if (!ccold) |
3629 | continue; | |
e498be7d | 3630 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3631 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3632 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3633 | kfree(ccold); |
3634 | } | |
d2e7b7d0 | 3635 | kfree(new); |
2ed3a4ef | 3636 | return alloc_kmemlist(cachep); |
1da177e4 LT |
3637 | } |
3638 | ||
b5d8ca7c | 3639 | /* Called with cache_chain_mutex held always */ |
2ed3a4ef | 3640 | static int enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3641 | { |
3642 | int err; | |
3643 | int limit, shared; | |
3644 | ||
a737b3e2 AM |
3645 | /* |
3646 | * The head array serves three purposes: | |
1da177e4 LT |
3647 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3648 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3649 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3650 | * bufctl chains: array operations are cheaper. |
3651 | * The numbers are guessed, we should auto-tune as described by | |
3652 | * Bonwick. | |
3653 | */ | |
3dafccf2 | 3654 | if (cachep->buffer_size > 131072) |
1da177e4 | 3655 | limit = 1; |
3dafccf2 | 3656 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3657 | limit = 8; |
3dafccf2 | 3658 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3659 | limit = 24; |
3dafccf2 | 3660 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3661 | limit = 54; |
3662 | else | |
3663 | limit = 120; | |
3664 | ||
a737b3e2 AM |
3665 | /* |
3666 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3667 | * allocation behaviour: Most allocs on one cpu, most free operations |
3668 | * on another cpu. For these cases, an efficient object passing between | |
3669 | * cpus is necessary. This is provided by a shared array. The array | |
3670 | * replaces Bonwick's magazine layer. | |
3671 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3672 | * to a larger limit. Thus disabled by default. | |
3673 | */ | |
3674 | shared = 0; | |
3675 | #ifdef CONFIG_SMP | |
3dafccf2 | 3676 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3677 | shared = 8; |
3678 | #endif | |
3679 | ||
3680 | #if DEBUG | |
a737b3e2 AM |
3681 | /* |
3682 | * With debugging enabled, large batchcount lead to excessively long | |
3683 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3684 | */ |
3685 | if (limit > 32) | |
3686 | limit = 32; | |
3687 | #endif | |
b28a02de | 3688 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3689 | if (err) |
3690 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3691 | cachep->name, -err); |
2ed3a4ef | 3692 | return err; |
1da177e4 LT |
3693 | } |
3694 | ||
1b55253a CL |
3695 | /* |
3696 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3697 | * necessary. Note that the l3 listlock also protects the array_cache |
3698 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3699 | */ |
3700 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
3701 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
3702 | { |
3703 | int tofree; | |
3704 | ||
1b55253a CL |
3705 | if (!ac || !ac->avail) |
3706 | return; | |
1da177e4 LT |
3707 | if (ac->touched && !force) { |
3708 | ac->touched = 0; | |
b18e7e65 | 3709 | } else { |
1b55253a | 3710 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
3711 | if (ac->avail) { |
3712 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3713 | if (tofree > ac->avail) | |
3714 | tofree = (ac->avail + 1) / 2; | |
3715 | free_block(cachep, ac->entry, tofree, node); | |
3716 | ac->avail -= tofree; | |
3717 | memmove(ac->entry, &(ac->entry[tofree]), | |
3718 | sizeof(void *) * ac->avail); | |
3719 | } | |
1b55253a | 3720 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
3721 | } |
3722 | } | |
3723 | ||
3724 | /** | |
3725 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3726 | * @unused: unused parameter |
1da177e4 LT |
3727 | * |
3728 | * Called from workqueue/eventd every few seconds. | |
3729 | * Purpose: | |
3730 | * - clear the per-cpu caches for this CPU. | |
3731 | * - return freeable pages to the main free memory pool. | |
3732 | * | |
a737b3e2 AM |
3733 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3734 | * again on the next iteration. | |
1da177e4 LT |
3735 | */ |
3736 | static void cache_reap(void *unused) | |
3737 | { | |
7a7c381d | 3738 | struct kmem_cache *searchp; |
e498be7d | 3739 | struct kmem_list3 *l3; |
aab2207c | 3740 | int node = numa_node_id(); |
1da177e4 | 3741 | |
fc0abb14 | 3742 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3743 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3744 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3745 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3746 | return; |
3747 | } | |
3748 | ||
7a7c381d | 3749 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
3750 | check_irq_on(); |
3751 | ||
35386e3b CL |
3752 | /* |
3753 | * We only take the l3 lock if absolutely necessary and we | |
3754 | * have established with reasonable certainty that | |
3755 | * we can do some work if the lock was obtained. | |
3756 | */ | |
aab2207c | 3757 | l3 = searchp->nodelists[node]; |
35386e3b | 3758 | |
8fce4d8e | 3759 | reap_alien(searchp, l3); |
1da177e4 | 3760 | |
aab2207c | 3761 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 3762 | |
35386e3b CL |
3763 | /* |
3764 | * These are racy checks but it does not matter | |
3765 | * if we skip one check or scan twice. | |
3766 | */ | |
e498be7d | 3767 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 3768 | goto next; |
1da177e4 | 3769 | |
e498be7d | 3770 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3771 | |
aab2207c | 3772 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 3773 | |
ed11d9eb | 3774 | if (l3->free_touched) |
e498be7d | 3775 | l3->free_touched = 0; |
ed11d9eb CL |
3776 | else { |
3777 | int freed; | |
1da177e4 | 3778 | |
ed11d9eb CL |
3779 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
3780 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
3781 | STATS_ADD_REAPED(searchp, freed); | |
3782 | } | |
35386e3b | 3783 | next: |
1da177e4 LT |
3784 | cond_resched(); |
3785 | } | |
3786 | check_irq_on(); | |
fc0abb14 | 3787 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 3788 | next_reap_node(); |
2244b95a | 3789 | refresh_cpu_vm_stats(smp_processor_id()); |
a737b3e2 | 3790 | /* Set up the next iteration */ |
cd61ef62 | 3791 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3792 | } |
3793 | ||
3794 | #ifdef CONFIG_PROC_FS | |
3795 | ||
85289f98 | 3796 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3797 | { |
85289f98 PE |
3798 | /* |
3799 | * Output format version, so at least we can change it | |
3800 | * without _too_ many complaints. | |
3801 | */ | |
1da177e4 | 3802 | #if STATS |
85289f98 | 3803 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3804 | #else |
85289f98 | 3805 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3806 | #endif |
85289f98 PE |
3807 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3808 | "<objperslab> <pagesperslab>"); | |
3809 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3810 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3811 | #if STATS |
85289f98 | 3812 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 3813 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 3814 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 3815 | #endif |
85289f98 PE |
3816 | seq_putc(m, '\n'); |
3817 | } | |
3818 | ||
3819 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3820 | { | |
3821 | loff_t n = *pos; | |
3822 | struct list_head *p; | |
3823 | ||
fc0abb14 | 3824 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3825 | if (!n) |
3826 | print_slabinfo_header(m); | |
1da177e4 LT |
3827 | p = cache_chain.next; |
3828 | while (n--) { | |
3829 | p = p->next; | |
3830 | if (p == &cache_chain) | |
3831 | return NULL; | |
3832 | } | |
343e0d7a | 3833 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3834 | } |
3835 | ||
3836 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3837 | { | |
343e0d7a | 3838 | struct kmem_cache *cachep = p; |
1da177e4 | 3839 | ++*pos; |
a737b3e2 AM |
3840 | return cachep->next.next == &cache_chain ? |
3841 | NULL : list_entry(cachep->next.next, struct kmem_cache, next); | |
1da177e4 LT |
3842 | } |
3843 | ||
3844 | static void s_stop(struct seq_file *m, void *p) | |
3845 | { | |
fc0abb14 | 3846 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3847 | } |
3848 | ||
3849 | static int s_show(struct seq_file *m, void *p) | |
3850 | { | |
343e0d7a | 3851 | struct kmem_cache *cachep = p; |
b28a02de PE |
3852 | struct slab *slabp; |
3853 | unsigned long active_objs; | |
3854 | unsigned long num_objs; | |
3855 | unsigned long active_slabs = 0; | |
3856 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3857 | const char *name; |
1da177e4 | 3858 | char *error = NULL; |
e498be7d CL |
3859 | int node; |
3860 | struct kmem_list3 *l3; | |
1da177e4 | 3861 | |
1da177e4 LT |
3862 | active_objs = 0; |
3863 | num_slabs = 0; | |
e498be7d CL |
3864 | for_each_online_node(node) { |
3865 | l3 = cachep->nodelists[node]; | |
3866 | if (!l3) | |
3867 | continue; | |
3868 | ||
ca3b9b91 RT |
3869 | check_irq_on(); |
3870 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3871 | |
7a7c381d | 3872 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
3873 | if (slabp->inuse != cachep->num && !error) |
3874 | error = "slabs_full accounting error"; | |
3875 | active_objs += cachep->num; | |
3876 | active_slabs++; | |
3877 | } | |
7a7c381d | 3878 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
3879 | if (slabp->inuse == cachep->num && !error) |
3880 | error = "slabs_partial inuse accounting error"; | |
3881 | if (!slabp->inuse && !error) | |
3882 | error = "slabs_partial/inuse accounting error"; | |
3883 | active_objs += slabp->inuse; | |
3884 | active_slabs++; | |
3885 | } | |
7a7c381d | 3886 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
3887 | if (slabp->inuse && !error) |
3888 | error = "slabs_free/inuse accounting error"; | |
3889 | num_slabs++; | |
3890 | } | |
3891 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3892 | if (l3->shared) |
3893 | shared_avail += l3->shared->avail; | |
e498be7d | 3894 | |
ca3b9b91 | 3895 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3896 | } |
b28a02de PE |
3897 | num_slabs += active_slabs; |
3898 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3899 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3900 | error = "free_objects accounting error"; |
3901 | ||
b28a02de | 3902 | name = cachep->name; |
1da177e4 LT |
3903 | if (error) |
3904 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3905 | ||
3906 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3907 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3908 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3909 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3910 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3911 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3912 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3913 | #if STATS |
b28a02de | 3914 | { /* list3 stats */ |
1da177e4 LT |
3915 | unsigned long high = cachep->high_mark; |
3916 | unsigned long allocs = cachep->num_allocations; | |
3917 | unsigned long grown = cachep->grown; | |
3918 | unsigned long reaped = cachep->reaped; | |
3919 | unsigned long errors = cachep->errors; | |
3920 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3921 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3922 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 3923 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 3924 | |
e498be7d | 3925 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 3926 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 3927 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 3928 | node_frees, overflows); |
1da177e4 LT |
3929 | } |
3930 | /* cpu stats */ | |
3931 | { | |
3932 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3933 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3934 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3935 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3936 | ||
3937 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3938 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3939 | } |
3940 | #endif | |
3941 | seq_putc(m, '\n'); | |
1da177e4 LT |
3942 | return 0; |
3943 | } | |
3944 | ||
3945 | /* | |
3946 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3947 | * | |
3948 | * Output layout: | |
3949 | * cache-name | |
3950 | * num-active-objs | |
3951 | * total-objs | |
3952 | * object size | |
3953 | * num-active-slabs | |
3954 | * total-slabs | |
3955 | * num-pages-per-slab | |
3956 | * + further values on SMP and with statistics enabled | |
3957 | */ | |
3958 | ||
3959 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3960 | .start = s_start, |
3961 | .next = s_next, | |
3962 | .stop = s_stop, | |
3963 | .show = s_show, | |
1da177e4 LT |
3964 | }; |
3965 | ||
3966 | #define MAX_SLABINFO_WRITE 128 | |
3967 | /** | |
3968 | * slabinfo_write - Tuning for the slab allocator | |
3969 | * @file: unused | |
3970 | * @buffer: user buffer | |
3971 | * @count: data length | |
3972 | * @ppos: unused | |
3973 | */ | |
b28a02de PE |
3974 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3975 | size_t count, loff_t *ppos) | |
1da177e4 | 3976 | { |
b28a02de | 3977 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 3978 | int limit, batchcount, shared, res; |
7a7c381d | 3979 | struct kmem_cache *cachep; |
b28a02de | 3980 | |
1da177e4 LT |
3981 | if (count > MAX_SLABINFO_WRITE) |
3982 | return -EINVAL; | |
3983 | if (copy_from_user(&kbuf, buffer, count)) | |
3984 | return -EFAULT; | |
b28a02de | 3985 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3986 | |
3987 | tmp = strchr(kbuf, ' '); | |
3988 | if (!tmp) | |
3989 | return -EINVAL; | |
3990 | *tmp = '\0'; | |
3991 | tmp++; | |
3992 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3993 | return -EINVAL; | |
3994 | ||
3995 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3996 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3997 | res = -EINVAL; |
7a7c381d | 3998 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 3999 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4000 | if (limit < 1 || batchcount < 1 || |
4001 | batchcount > limit || shared < 0) { | |
e498be7d | 4002 | res = 0; |
1da177e4 | 4003 | } else { |
e498be7d | 4004 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 4005 | batchcount, shared); |
1da177e4 LT |
4006 | } |
4007 | break; | |
4008 | } | |
4009 | } | |
fc0abb14 | 4010 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4011 | if (res >= 0) |
4012 | res = count; | |
4013 | return res; | |
4014 | } | |
871751e2 AV |
4015 | |
4016 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4017 | ||
4018 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4019 | { | |
4020 | loff_t n = *pos; | |
4021 | struct list_head *p; | |
4022 | ||
4023 | mutex_lock(&cache_chain_mutex); | |
4024 | p = cache_chain.next; | |
4025 | while (n--) { | |
4026 | p = p->next; | |
4027 | if (p == &cache_chain) | |
4028 | return NULL; | |
4029 | } | |
4030 | return list_entry(p, struct kmem_cache, next); | |
4031 | } | |
4032 | ||
4033 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4034 | { | |
4035 | unsigned long *p; | |
4036 | int l; | |
4037 | if (!v) | |
4038 | return 1; | |
4039 | l = n[1]; | |
4040 | p = n + 2; | |
4041 | while (l) { | |
4042 | int i = l/2; | |
4043 | unsigned long *q = p + 2 * i; | |
4044 | if (*q == v) { | |
4045 | q[1]++; | |
4046 | return 1; | |
4047 | } | |
4048 | if (*q > v) { | |
4049 | l = i; | |
4050 | } else { | |
4051 | p = q + 2; | |
4052 | l -= i + 1; | |
4053 | } | |
4054 | } | |
4055 | if (++n[1] == n[0]) | |
4056 | return 0; | |
4057 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4058 | p[0] = v; | |
4059 | p[1] = 1; | |
4060 | return 1; | |
4061 | } | |
4062 | ||
4063 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4064 | { | |
4065 | void *p; | |
4066 | int i; | |
4067 | if (n[0] == n[1]) | |
4068 | return; | |
4069 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4070 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4071 | continue; | |
4072 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4073 | return; | |
4074 | } | |
4075 | } | |
4076 | ||
4077 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4078 | { | |
4079 | #ifdef CONFIG_KALLSYMS | |
4080 | char *modname; | |
4081 | const char *name; | |
4082 | unsigned long offset, size; | |
4083 | char namebuf[KSYM_NAME_LEN+1]; | |
4084 | ||
4085 | name = kallsyms_lookup(address, &size, &offset, &modname, namebuf); | |
4086 | ||
4087 | if (name) { | |
4088 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4089 | if (modname) | |
4090 | seq_printf(m, " [%s]", modname); | |
4091 | return; | |
4092 | } | |
4093 | #endif | |
4094 | seq_printf(m, "%p", (void *)address); | |
4095 | } | |
4096 | ||
4097 | static int leaks_show(struct seq_file *m, void *p) | |
4098 | { | |
4099 | struct kmem_cache *cachep = p; | |
871751e2 AV |
4100 | struct slab *slabp; |
4101 | struct kmem_list3 *l3; | |
4102 | const char *name; | |
4103 | unsigned long *n = m->private; | |
4104 | int node; | |
4105 | int i; | |
4106 | ||
4107 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4108 | return 0; | |
4109 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4110 | return 0; | |
4111 | ||
4112 | /* OK, we can do it */ | |
4113 | ||
4114 | n[1] = 0; | |
4115 | ||
4116 | for_each_online_node(node) { | |
4117 | l3 = cachep->nodelists[node]; | |
4118 | if (!l3) | |
4119 | continue; | |
4120 | ||
4121 | check_irq_on(); | |
4122 | spin_lock_irq(&l3->list_lock); | |
4123 | ||
7a7c381d | 4124 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4125 | handle_slab(n, cachep, slabp); |
7a7c381d | 4126 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4127 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4128 | spin_unlock_irq(&l3->list_lock); |
4129 | } | |
4130 | name = cachep->name; | |
4131 | if (n[0] == n[1]) { | |
4132 | /* Increase the buffer size */ | |
4133 | mutex_unlock(&cache_chain_mutex); | |
4134 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4135 | if (!m->private) { | |
4136 | /* Too bad, we are really out */ | |
4137 | m->private = n; | |
4138 | mutex_lock(&cache_chain_mutex); | |
4139 | return -ENOMEM; | |
4140 | } | |
4141 | *(unsigned long *)m->private = n[0] * 2; | |
4142 | kfree(n); | |
4143 | mutex_lock(&cache_chain_mutex); | |
4144 | /* Now make sure this entry will be retried */ | |
4145 | m->count = m->size; | |
4146 | return 0; | |
4147 | } | |
4148 | for (i = 0; i < n[1]; i++) { | |
4149 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4150 | show_symbol(m, n[2*i+2]); | |
4151 | seq_putc(m, '\n'); | |
4152 | } | |
d2e7b7d0 | 4153 | |
871751e2 AV |
4154 | return 0; |
4155 | } | |
4156 | ||
4157 | struct seq_operations slabstats_op = { | |
4158 | .start = leaks_start, | |
4159 | .next = s_next, | |
4160 | .stop = s_stop, | |
4161 | .show = leaks_show, | |
4162 | }; | |
4163 | #endif | |
1da177e4 LT |
4164 | #endif |
4165 | ||
00e145b6 MS |
4166 | /** |
4167 | * ksize - get the actual amount of memory allocated for a given object | |
4168 | * @objp: Pointer to the object | |
4169 | * | |
4170 | * kmalloc may internally round up allocations and return more memory | |
4171 | * than requested. ksize() can be used to determine the actual amount of | |
4172 | * memory allocated. The caller may use this additional memory, even though | |
4173 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4174 | * The caller must guarantee that objp points to a valid object previously | |
4175 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4176 | * must not be freed during the duration of the call. | |
4177 | */ | |
1da177e4 LT |
4178 | unsigned int ksize(const void *objp) |
4179 | { | |
00e145b6 MS |
4180 | if (unlikely(objp == NULL)) |
4181 | return 0; | |
1da177e4 | 4182 | |
6ed5eb22 | 4183 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4184 | } |