]>
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 | * | |
53 | * The c_cpuarray may not be read with enabled local interrupts - | |
54 | * it's changed with a smp_call_function(). | |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
58 | * Several members in kmem_cache_t and struct slab never change, they | |
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> | |
92 | #include <linux/swap.h> | |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
97 | #include <linux/seq_file.h> | |
98 | #include <linux/notifier.h> | |
99 | #include <linux/kallsyms.h> | |
100 | #include <linux/cpu.h> | |
101 | #include <linux/sysctl.h> | |
102 | #include <linux/module.h> | |
103 | #include <linux/rcupdate.h> | |
543537bd | 104 | #include <linux/string.h> |
e498be7d | 105 | #include <linux/nodemask.h> |
dc85da15 | 106 | #include <linux/mempolicy.h> |
fc0abb14 | 107 | #include <linux/mutex.h> |
1da177e4 LT |
108 | |
109 | #include <asm/uaccess.h> | |
110 | #include <asm/cacheflush.h> | |
111 | #include <asm/tlbflush.h> | |
112 | #include <asm/page.h> | |
113 | ||
114 | /* | |
115 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
116 | * SLAB_RED_ZONE & SLAB_POISON. | |
117 | * 0 for faster, smaller code (especially in the critical paths). | |
118 | * | |
119 | * STATS - 1 to collect stats for /proc/slabinfo. | |
120 | * 0 for faster, smaller code (especially in the critical paths). | |
121 | * | |
122 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
123 | */ | |
124 | ||
125 | #ifdef CONFIG_DEBUG_SLAB | |
126 | #define DEBUG 1 | |
127 | #define STATS 1 | |
128 | #define FORCED_DEBUG 1 | |
129 | #else | |
130 | #define DEBUG 0 | |
131 | #define STATS 0 | |
132 | #define FORCED_DEBUG 0 | |
133 | #endif | |
134 | ||
1da177e4 LT |
135 | /* Shouldn't this be in a header file somewhere? */ |
136 | #define BYTES_PER_WORD sizeof(void *) | |
137 | ||
138 | #ifndef cache_line_size | |
139 | #define cache_line_size() L1_CACHE_BYTES | |
140 | #endif | |
141 | ||
142 | #ifndef ARCH_KMALLOC_MINALIGN | |
143 | /* | |
144 | * Enforce a minimum alignment for the kmalloc caches. | |
145 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
146 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
147 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
148 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
149 | * Note that this flag disables some debug features. | |
150 | */ | |
151 | #define ARCH_KMALLOC_MINALIGN 0 | |
152 | #endif | |
153 | ||
154 | #ifndef ARCH_SLAB_MINALIGN | |
155 | /* | |
156 | * Enforce a minimum alignment for all caches. | |
157 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
158 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
159 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
160 | * some debug features. | |
161 | */ | |
162 | #define ARCH_SLAB_MINALIGN 0 | |
163 | #endif | |
164 | ||
165 | #ifndef ARCH_KMALLOC_FLAGS | |
166 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
167 | #endif | |
168 | ||
169 | /* Legal flag mask for kmem_cache_create(). */ | |
170 | #if DEBUG | |
171 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
172 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
173 | SLAB_NO_REAP | SLAB_CACHE_DMA | \ | |
174 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ | |
175 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
176 | SLAB_DESTROY_BY_RCU) | |
177 | #else | |
178 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \ | |
179 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ | |
180 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
181 | SLAB_DESTROY_BY_RCU) | |
182 | #endif | |
183 | ||
184 | /* | |
185 | * kmem_bufctl_t: | |
186 | * | |
187 | * Bufctl's are used for linking objs within a slab | |
188 | * linked offsets. | |
189 | * | |
190 | * This implementation relies on "struct page" for locating the cache & | |
191 | * slab an object belongs to. | |
192 | * This allows the bufctl structure to be small (one int), but limits | |
193 | * the number of objects a slab (not a cache) can contain when off-slab | |
194 | * bufctls are used. The limit is the size of the largest general cache | |
195 | * that does not use off-slab slabs. | |
196 | * For 32bit archs with 4 kB pages, is this 56. | |
197 | * This is not serious, as it is only for large objects, when it is unwise | |
198 | * to have too many per slab. | |
199 | * Note: This limit can be raised by introducing a general cache whose size | |
200 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
201 | */ | |
202 | ||
fa5b08d5 | 203 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
204 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
205 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
206 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2) | |
207 | ||
208 | /* Max number of objs-per-slab for caches which use off-slab slabs. | |
209 | * Needed to avoid a possible looping condition in cache_grow(). | |
210 | */ | |
211 | static unsigned long offslab_limit; | |
212 | ||
213 | /* | |
214 | * struct slab | |
215 | * | |
216 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
217 | * for a slab, or allocated from an general cache. | |
218 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
219 | */ | |
220 | struct slab { | |
b28a02de PE |
221 | struct list_head list; |
222 | unsigned long colouroff; | |
223 | void *s_mem; /* including colour offset */ | |
224 | unsigned int inuse; /* num of objs active in slab */ | |
225 | kmem_bufctl_t free; | |
226 | unsigned short nodeid; | |
1da177e4 LT |
227 | }; |
228 | ||
229 | /* | |
230 | * struct slab_rcu | |
231 | * | |
232 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
233 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
234 | * we need to approach a kernel structure obliquely, from its address | |
235 | * obtained without the usual locking. We can lock the structure to | |
236 | * stabilize it and check it's still at the given address, only if we | |
237 | * can be sure that the memory has not been meanwhile reused for some | |
238 | * other kind of object (which our subsystem's lock might corrupt). | |
239 | * | |
240 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
241 | * taking the spinlock within the structure expected at that address. | |
242 | * | |
243 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
244 | */ | |
245 | struct slab_rcu { | |
b28a02de PE |
246 | struct rcu_head head; |
247 | kmem_cache_t *cachep; | |
248 | void *addr; | |
1da177e4 LT |
249 | }; |
250 | ||
251 | /* | |
252 | * struct array_cache | |
253 | * | |
1da177e4 LT |
254 | * Purpose: |
255 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
256 | * - reduce the number of linked list operations | |
257 | * - reduce spinlock operations | |
258 | * | |
259 | * The limit is stored in the per-cpu structure to reduce the data cache | |
260 | * footprint. | |
261 | * | |
262 | */ | |
263 | struct array_cache { | |
264 | unsigned int avail; | |
265 | unsigned int limit; | |
266 | unsigned int batchcount; | |
267 | unsigned int touched; | |
e498be7d CL |
268 | spinlock_t lock; |
269 | void *entry[0]; /* | |
270 | * Must have this definition in here for the proper | |
271 | * alignment of array_cache. Also simplifies accessing | |
272 | * the entries. | |
273 | * [0] is for gcc 2.95. It should really be []. | |
274 | */ | |
1da177e4 LT |
275 | }; |
276 | ||
277 | /* bootstrap: The caches do not work without cpuarrays anymore, | |
278 | * but the cpuarrays are allocated from the generic caches... | |
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; | |
294 | unsigned long next_reap; | |
295 | int free_touched; | |
296 | unsigned int free_limit; | |
297 | spinlock_t list_lock; | |
298 | struct array_cache *shared; /* shared per node */ | |
299 | struct array_cache **alien; /* on other nodes */ | |
1da177e4 LT |
300 | }; |
301 | ||
e498be7d CL |
302 | /* |
303 | * Need this for bootstrapping a per node allocator. | |
304 | */ | |
305 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
306 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
307 | #define CACHE_CACHE 0 | |
308 | #define SIZE_AC 1 | |
309 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
310 | ||
311 | /* | |
7243cc05 | 312 | * This function must be completely optimized away if |
e498be7d CL |
313 | * a constant is passed to it. Mostly the same as |
314 | * what is in linux/slab.h except it returns an | |
315 | * index. | |
316 | */ | |
7243cc05 | 317 | static __always_inline int index_of(const size_t size) |
e498be7d | 318 | { |
5ec8a847 SR |
319 | extern void __bad_size(void); |
320 | ||
e498be7d CL |
321 | if (__builtin_constant_p(size)) { |
322 | int i = 0; | |
323 | ||
324 | #define CACHE(x) \ | |
325 | if (size <=x) \ | |
326 | return i; \ | |
327 | else \ | |
328 | i++; | |
329 | #include "linux/kmalloc_sizes.h" | |
330 | #undef CACHE | |
5ec8a847 | 331 | __bad_size(); |
7243cc05 | 332 | } else |
5ec8a847 | 333 | __bad_size(); |
e498be7d CL |
334 | return 0; |
335 | } | |
336 | ||
337 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) | |
338 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 339 | |
e498be7d CL |
340 | static inline void kmem_list3_init(struct kmem_list3 *parent) |
341 | { | |
342 | INIT_LIST_HEAD(&parent->slabs_full); | |
343 | INIT_LIST_HEAD(&parent->slabs_partial); | |
344 | INIT_LIST_HEAD(&parent->slabs_free); | |
345 | parent->shared = NULL; | |
346 | parent->alien = NULL; | |
347 | spin_lock_init(&parent->list_lock); | |
348 | parent->free_objects = 0; | |
349 | parent->free_touched = 0; | |
350 | } | |
351 | ||
352 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ | |
353 | do { \ | |
354 | INIT_LIST_HEAD(listp); \ | |
355 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
356 | } while (0) | |
357 | ||
358 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ | |
359 | do { \ | |
360 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ | |
361 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
362 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
363 | } while (0) | |
1da177e4 LT |
364 | |
365 | /* | |
366 | * kmem_cache_t | |
367 | * | |
368 | * manages a cache. | |
369 | */ | |
b28a02de | 370 | |
2109a2d1 | 371 | struct kmem_cache { |
1da177e4 | 372 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de PE |
373 | struct array_cache *array[NR_CPUS]; |
374 | unsigned int batchcount; | |
375 | unsigned int limit; | |
376 | unsigned int shared; | |
3dafccf2 | 377 | unsigned int buffer_size; |
e498be7d | 378 | /* 2) touched by every alloc & free from the backend */ |
b28a02de PE |
379 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
380 | unsigned int flags; /* constant flags */ | |
381 | unsigned int num; /* # of objs per slab */ | |
382 | spinlock_t spinlock; | |
1da177e4 LT |
383 | |
384 | /* 3) cache_grow/shrink */ | |
385 | /* order of pgs per slab (2^n) */ | |
b28a02de | 386 | unsigned int gfporder; |
1da177e4 LT |
387 | |
388 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 389 | gfp_t gfpflags; |
1da177e4 | 390 | |
b28a02de PE |
391 | size_t colour; /* cache colouring range */ |
392 | unsigned int colour_off; /* colour offset */ | |
393 | unsigned int colour_next; /* cache colouring */ | |
394 | kmem_cache_t *slabp_cache; | |
395 | unsigned int slab_size; | |
396 | unsigned int dflags; /* dynamic flags */ | |
1da177e4 LT |
397 | |
398 | /* constructor func */ | |
b28a02de | 399 | void (*ctor) (void *, kmem_cache_t *, unsigned long); |
1da177e4 LT |
400 | |
401 | /* de-constructor func */ | |
b28a02de | 402 | void (*dtor) (void *, kmem_cache_t *, unsigned long); |
1da177e4 LT |
403 | |
404 | /* 4) cache creation/removal */ | |
b28a02de PE |
405 | const char *name; |
406 | struct list_head next; | |
1da177e4 LT |
407 | |
408 | /* 5) statistics */ | |
409 | #if STATS | |
b28a02de PE |
410 | unsigned long num_active; |
411 | unsigned long num_allocations; | |
412 | unsigned long high_mark; | |
413 | unsigned long grown; | |
414 | unsigned long reaped; | |
415 | unsigned long errors; | |
416 | unsigned long max_freeable; | |
417 | unsigned long node_allocs; | |
418 | unsigned long node_frees; | |
419 | atomic_t allochit; | |
420 | atomic_t allocmiss; | |
421 | atomic_t freehit; | |
422 | atomic_t freemiss; | |
1da177e4 LT |
423 | #endif |
424 | #if DEBUG | |
3dafccf2 MS |
425 | /* |
426 | * If debugging is enabled, then the allocator can add additional | |
427 | * fields and/or padding to every object. buffer_size contains the total | |
428 | * object size including these internal fields, the following two | |
429 | * variables contain the offset to the user object and its size. | |
430 | */ | |
431 | int obj_offset; | |
432 | int obj_size; | |
1da177e4 LT |
433 | #endif |
434 | }; | |
435 | ||
436 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
437 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
438 | ||
439 | #define BATCHREFILL_LIMIT 16 | |
440 | /* Optimization question: fewer reaps means less | |
441 | * probability for unnessary cpucache drain/refill cycles. | |
442 | * | |
dc6f3f27 | 443 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
444 | * which could lock up otherwise freeable slabs. |
445 | */ | |
446 | #define REAPTIMEOUT_CPUC (2*HZ) | |
447 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
448 | ||
449 | #if STATS | |
450 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
451 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
452 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
453 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
454 | #define STATS_INC_REAPED(x) ((x)->reaped++) | |
455 | #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \ | |
456 | (x)->high_mark = (x)->num_active; \ | |
457 | } while (0) | |
458 | #define STATS_INC_ERR(x) ((x)->errors++) | |
459 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 460 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
1da177e4 LT |
461 | #define STATS_SET_FREEABLE(x, i) \ |
462 | do { if ((x)->max_freeable < i) \ | |
463 | (x)->max_freeable = i; \ | |
464 | } while (0) | |
465 | ||
466 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) | |
467 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
468 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
469 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
470 | #else | |
471 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
472 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
473 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
474 | #define STATS_INC_GROWN(x) do { } while (0) | |
475 | #define STATS_INC_REAPED(x) do { } while (0) | |
476 | #define STATS_SET_HIGH(x) do { } while (0) | |
477 | #define STATS_INC_ERR(x) do { } while (0) | |
478 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 479 | #define STATS_INC_NODEFREES(x) do { } while (0) |
1da177e4 LT |
480 | #define STATS_SET_FREEABLE(x, i) \ |
481 | do { } while (0) | |
482 | ||
483 | #define STATS_INC_ALLOCHIT(x) do { } while (0) | |
484 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
485 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
486 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
487 | #endif | |
488 | ||
489 | #if DEBUG | |
490 | /* Magic nums for obj red zoning. | |
491 | * Placed in the first word before and the first word after an obj. | |
492 | */ | |
493 | #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ | |
494 | #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ | |
495 | ||
496 | /* ...and for poisoning */ | |
497 | #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ | |
498 | #define POISON_FREE 0x6b /* for use-after-free poisoning */ | |
499 | #define POISON_END 0xa5 /* end-byte of poisoning */ | |
500 | ||
501 | /* memory layout of objects: | |
502 | * 0 : objp | |
3dafccf2 | 503 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
504 | * the end of an object is aligned with the end of the real |
505 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 506 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 507 | * redzone word. |
3dafccf2 MS |
508 | * cachep->obj_offset: The real object. |
509 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
510 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long] | |
1da177e4 | 511 | */ |
3dafccf2 | 512 | static int obj_offset(kmem_cache_t *cachep) |
1da177e4 | 513 | { |
3dafccf2 | 514 | return cachep->obj_offset; |
1da177e4 LT |
515 | } |
516 | ||
3dafccf2 | 517 | static int obj_size(kmem_cache_t *cachep) |
1da177e4 | 518 | { |
3dafccf2 | 519 | return cachep->obj_size; |
1da177e4 LT |
520 | } |
521 | ||
522 | static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp) | |
523 | { | |
524 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 525 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
526 | } |
527 | ||
528 | static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp) | |
529 | { | |
530 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
531 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 532 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 533 | 2 * BYTES_PER_WORD); |
3dafccf2 | 534 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
535 | } |
536 | ||
537 | static void **dbg_userword(kmem_cache_t *cachep, void *objp) | |
538 | { | |
539 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 540 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
541 | } |
542 | ||
543 | #else | |
544 | ||
3dafccf2 MS |
545 | #define obj_offset(x) 0 |
546 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
547 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
548 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
549 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
550 | ||
551 | #endif | |
552 | ||
553 | /* | |
554 | * Maximum size of an obj (in 2^order pages) | |
555 | * and absolute limit for the gfp order. | |
556 | */ | |
557 | #if defined(CONFIG_LARGE_ALLOCS) | |
558 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
559 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
560 | #elif defined(CONFIG_MMU) | |
561 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
562 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
563 | #else | |
564 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
565 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
566 | #endif | |
567 | ||
568 | /* | |
569 | * Do not go above this order unless 0 objects fit into the slab. | |
570 | */ | |
571 | #define BREAK_GFP_ORDER_HI 1 | |
572 | #define BREAK_GFP_ORDER_LO 0 | |
573 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
574 | ||
065d41cb | 575 | /* Functions for storing/retrieving the cachep and or slab from the |
1da177e4 LT |
576 | * global 'mem_map'. These are used to find the slab an obj belongs to. |
577 | * With kfree(), these are used to find the cache which an obj belongs to. | |
578 | */ | |
065d41cb PE |
579 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
580 | { | |
581 | page->lru.next = (struct list_head *)cache; | |
582 | } | |
583 | ||
584 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
585 | { | |
586 | return (struct kmem_cache *)page->lru.next; | |
587 | } | |
588 | ||
589 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
590 | { | |
591 | page->lru.prev = (struct list_head *)slab; | |
592 | } | |
593 | ||
594 | static inline struct slab *page_get_slab(struct page *page) | |
595 | { | |
596 | return (struct slab *)page->lru.prev; | |
597 | } | |
1da177e4 LT |
598 | |
599 | /* These are the default caches for kmalloc. Custom caches can have other sizes. */ | |
600 | struct cache_sizes malloc_sizes[] = { | |
601 | #define CACHE(x) { .cs_size = (x) }, | |
602 | #include <linux/kmalloc_sizes.h> | |
603 | CACHE(ULONG_MAX) | |
604 | #undef CACHE | |
605 | }; | |
606 | EXPORT_SYMBOL(malloc_sizes); | |
607 | ||
608 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
609 | struct cache_names { | |
610 | char *name; | |
611 | char *name_dma; | |
612 | }; | |
613 | ||
614 | static struct cache_names __initdata cache_names[] = { | |
615 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
616 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 617 | {NULL,} |
1da177e4 LT |
618 | #undef CACHE |
619 | }; | |
620 | ||
621 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 622 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 623 | static struct arraycache_init initarray_generic = |
b28a02de | 624 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
625 | |
626 | /* internal cache of cache description objs */ | |
627 | static kmem_cache_t cache_cache = { | |
b28a02de PE |
628 | .batchcount = 1, |
629 | .limit = BOOT_CPUCACHE_ENTRIES, | |
630 | .shared = 1, | |
3dafccf2 | 631 | .buffer_size = sizeof(kmem_cache_t), |
b28a02de PE |
632 | .flags = SLAB_NO_REAP, |
633 | .spinlock = SPIN_LOCK_UNLOCKED, | |
634 | .name = "kmem_cache", | |
1da177e4 | 635 | #if DEBUG |
3dafccf2 | 636 | .obj_size = sizeof(kmem_cache_t), |
1da177e4 LT |
637 | #endif |
638 | }; | |
639 | ||
640 | /* Guard access to the cache-chain. */ | |
fc0abb14 | 641 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
642 | static struct list_head cache_chain; |
643 | ||
644 | /* | |
645 | * vm_enough_memory() looks at this to determine how many | |
646 | * slab-allocated pages are possibly freeable under pressure | |
647 | * | |
648 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
649 | */ | |
650 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
651 | |
652 | /* | |
653 | * chicken and egg problem: delay the per-cpu array allocation | |
654 | * until the general caches are up. | |
655 | */ | |
656 | static enum { | |
657 | NONE, | |
e498be7d CL |
658 | PARTIAL_AC, |
659 | PARTIAL_L3, | |
1da177e4 LT |
660 | FULL |
661 | } g_cpucache_up; | |
662 | ||
663 | static DEFINE_PER_CPU(struct work_struct, reap_work); | |
664 | ||
b28a02de PE |
665 | static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node); |
666 | static void enable_cpucache(kmem_cache_t *cachep); | |
667 | static void cache_reap(void *unused); | |
e498be7d | 668 | static int __node_shrink(kmem_cache_t *cachep, int node); |
1da177e4 LT |
669 | |
670 | static inline struct array_cache *ac_data(kmem_cache_t *cachep) | |
671 | { | |
672 | return cachep->array[smp_processor_id()]; | |
673 | } | |
674 | ||
dd0fc66f | 675 | static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags) |
1da177e4 LT |
676 | { |
677 | struct cache_sizes *csizep = malloc_sizes; | |
678 | ||
679 | #if DEBUG | |
680 | /* This happens if someone tries to call | |
b28a02de PE |
681 | * kmem_cache_create(), or __kmalloc(), before |
682 | * the generic caches are initialized. | |
683 | */ | |
c7e43c78 | 684 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
685 | #endif |
686 | while (size > csizep->cs_size) | |
687 | csizep++; | |
688 | ||
689 | /* | |
0abf40c1 | 690 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
691 | * has cs_{dma,}cachep==NULL. Thus no special case |
692 | * for large kmalloc calls required. | |
693 | */ | |
694 | if (unlikely(gfpflags & GFP_DMA)) | |
695 | return csizep->cs_dmacachep; | |
696 | return csizep->cs_cachep; | |
697 | } | |
698 | ||
dd0fc66f | 699 | kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
700 | { |
701 | return __find_general_cachep(size, gfpflags); | |
702 | } | |
703 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
704 | ||
fbaccacf | 705 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 706 | { |
fbaccacf SR |
707 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
708 | } | |
1da177e4 | 709 | |
fbaccacf SR |
710 | /* Calculate the number of objects and left-over bytes for a given |
711 | buffer size. */ | |
712 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, | |
713 | size_t align, int flags, size_t *left_over, | |
714 | unsigned int *num) | |
715 | { | |
716 | int nr_objs; | |
717 | size_t mgmt_size; | |
718 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 719 | |
fbaccacf SR |
720 | /* |
721 | * The slab management structure can be either off the slab or | |
722 | * on it. For the latter case, the memory allocated for a | |
723 | * slab is used for: | |
724 | * | |
725 | * - The struct slab | |
726 | * - One kmem_bufctl_t for each object | |
727 | * - Padding to respect alignment of @align | |
728 | * - @buffer_size bytes for each object | |
729 | * | |
730 | * If the slab management structure is off the slab, then the | |
731 | * alignment will already be calculated into the size. Because | |
732 | * the slabs are all pages aligned, the objects will be at the | |
733 | * correct alignment when allocated. | |
734 | */ | |
735 | if (flags & CFLGS_OFF_SLAB) { | |
736 | mgmt_size = 0; | |
737 | nr_objs = slab_size / buffer_size; | |
738 | ||
739 | if (nr_objs > SLAB_LIMIT) | |
740 | nr_objs = SLAB_LIMIT; | |
741 | } else { | |
742 | /* | |
743 | * Ignore padding for the initial guess. The padding | |
744 | * is at most @align-1 bytes, and @buffer_size is at | |
745 | * least @align. In the worst case, this result will | |
746 | * be one greater than the number of objects that fit | |
747 | * into the memory allocation when taking the padding | |
748 | * into account. | |
749 | */ | |
750 | nr_objs = (slab_size - sizeof(struct slab)) / | |
751 | (buffer_size + sizeof(kmem_bufctl_t)); | |
752 | ||
753 | /* | |
754 | * This calculated number will be either the right | |
755 | * amount, or one greater than what we want. | |
756 | */ | |
757 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
758 | > slab_size) | |
759 | nr_objs--; | |
760 | ||
761 | if (nr_objs > SLAB_LIMIT) | |
762 | nr_objs = SLAB_LIMIT; | |
763 | ||
764 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
765 | } | |
766 | *num = nr_objs; | |
767 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
768 | } |
769 | ||
770 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
771 | ||
772 | static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg) | |
773 | { | |
774 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 775 | function, cachep->name, msg); |
1da177e4 LT |
776 | dump_stack(); |
777 | } | |
778 | ||
779 | /* | |
780 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
781 | * via the workqueue/eventd. | |
782 | * Add the CPU number into the expiration time to minimize the possibility of | |
783 | * the CPUs getting into lockstep and contending for the global cache chain | |
784 | * lock. | |
785 | */ | |
786 | static void __devinit start_cpu_timer(int cpu) | |
787 | { | |
788 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
789 | ||
790 | /* | |
791 | * When this gets called from do_initcalls via cpucache_init(), | |
792 | * init_workqueues() has already run, so keventd will be setup | |
793 | * at that time. | |
794 | */ | |
795 | if (keventd_up() && reap_work->func == NULL) { | |
796 | INIT_WORK(reap_work, cache_reap, NULL); | |
797 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
798 | } | |
799 | } | |
800 | ||
e498be7d | 801 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 802 | int batchcount) |
1da177e4 | 803 | { |
b28a02de | 804 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
805 | struct array_cache *nc = NULL; |
806 | ||
e498be7d | 807 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
808 | if (nc) { |
809 | nc->avail = 0; | |
810 | nc->limit = entries; | |
811 | nc->batchcount = batchcount; | |
812 | nc->touched = 0; | |
e498be7d | 813 | spin_lock_init(&nc->lock); |
1da177e4 LT |
814 | } |
815 | return nc; | |
816 | } | |
817 | ||
e498be7d | 818 | #ifdef CONFIG_NUMA |
dc85da15 CL |
819 | static void *__cache_alloc_node(kmem_cache_t *, gfp_t, int); |
820 | ||
e498be7d CL |
821 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
822 | { | |
823 | struct array_cache **ac_ptr; | |
b28a02de | 824 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
825 | int i; |
826 | ||
827 | if (limit > 1) | |
828 | limit = 12; | |
829 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
830 | if (ac_ptr) { | |
831 | for_each_node(i) { | |
832 | if (i == node || !node_online(i)) { | |
833 | ac_ptr[i] = NULL; | |
834 | continue; | |
835 | } | |
836 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
837 | if (!ac_ptr[i]) { | |
b28a02de | 838 | for (i--; i <= 0; i--) |
e498be7d CL |
839 | kfree(ac_ptr[i]); |
840 | kfree(ac_ptr); | |
841 | return NULL; | |
842 | } | |
843 | } | |
844 | } | |
845 | return ac_ptr; | |
846 | } | |
847 | ||
848 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
849 | { | |
850 | int i; | |
851 | ||
852 | if (!ac_ptr) | |
853 | return; | |
854 | ||
855 | for_each_node(i) | |
b28a02de | 856 | kfree(ac_ptr[i]); |
e498be7d CL |
857 | |
858 | kfree(ac_ptr); | |
859 | } | |
860 | ||
b28a02de PE |
861 | static inline void __drain_alien_cache(kmem_cache_t *cachep, |
862 | struct array_cache *ac, int node) | |
e498be7d CL |
863 | { |
864 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
865 | ||
866 | if (ac->avail) { | |
867 | spin_lock(&rl3->list_lock); | |
ff69416e | 868 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
869 | ac->avail = 0; |
870 | spin_unlock(&rl3->list_lock); | |
871 | } | |
872 | } | |
873 | ||
874 | static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3) | |
875 | { | |
b28a02de | 876 | int i = 0; |
e498be7d CL |
877 | struct array_cache *ac; |
878 | unsigned long flags; | |
879 | ||
880 | for_each_online_node(i) { | |
881 | ac = l3->alien[i]; | |
882 | if (ac) { | |
883 | spin_lock_irqsave(&ac->lock, flags); | |
884 | __drain_alien_cache(cachep, ac, i); | |
885 | spin_unlock_irqrestore(&ac->lock, flags); | |
886 | } | |
887 | } | |
888 | } | |
889 | #else | |
890 | #define alloc_alien_cache(node, limit) do { } while (0) | |
891 | #define free_alien_cache(ac_ptr) do { } while (0) | |
892 | #define drain_alien_cache(cachep, l3) do { } while (0) | |
893 | #endif | |
894 | ||
1da177e4 | 895 | static int __devinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 896 | unsigned long action, void *hcpu) |
1da177e4 LT |
897 | { |
898 | long cpu = (long)hcpu; | |
b28a02de | 899 | kmem_cache_t *cachep; |
e498be7d CL |
900 | struct kmem_list3 *l3 = NULL; |
901 | int node = cpu_to_node(cpu); | |
902 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
903 | |
904 | switch (action) { | |
905 | case CPU_UP_PREPARE: | |
fc0abb14 | 906 | mutex_lock(&cache_chain_mutex); |
e498be7d CL |
907 | /* we need to do this right in the beginning since |
908 | * alloc_arraycache's are going to use this list. | |
909 | * kmalloc_node allows us to add the slab to the right | |
910 | * kmem_list3 and not this cpu's kmem_list3 | |
911 | */ | |
912 | ||
1da177e4 | 913 | list_for_each_entry(cachep, &cache_chain, next) { |
e498be7d CL |
914 | /* setup the size64 kmemlist for cpu before we can |
915 | * begin anything. Make sure some other cpu on this | |
916 | * node has not already allocated this | |
917 | */ | |
918 | if (!cachep->nodelists[node]) { | |
919 | if (!(l3 = kmalloc_node(memsize, | |
b28a02de | 920 | GFP_KERNEL, node))) |
e498be7d CL |
921 | goto bad; |
922 | kmem_list3_init(l3); | |
923 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 924 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
925 | |
926 | cachep->nodelists[node] = l3; | |
927 | } | |
1da177e4 | 928 | |
e498be7d CL |
929 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
930 | cachep->nodelists[node]->free_limit = | |
b28a02de PE |
931 | (1 + nr_cpus_node(node)) * |
932 | cachep->batchcount + cachep->num; | |
e498be7d CL |
933 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
934 | } | |
935 | ||
936 | /* Now we can go ahead with allocating the shared array's | |
b28a02de | 937 | & array cache's */ |
e498be7d | 938 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 TK |
939 | struct array_cache *nc; |
940 | ||
e498be7d | 941 | nc = alloc_arraycache(node, cachep->limit, |
b28a02de | 942 | cachep->batchcount); |
1da177e4 LT |
943 | if (!nc) |
944 | goto bad; | |
1da177e4 | 945 | cachep->array[cpu] = nc; |
1da177e4 | 946 | |
e498be7d CL |
947 | l3 = cachep->nodelists[node]; |
948 | BUG_ON(!l3); | |
949 | if (!l3->shared) { | |
950 | if (!(nc = alloc_arraycache(node, | |
b28a02de PE |
951 | cachep->shared * |
952 | cachep->batchcount, | |
953 | 0xbaadf00d))) | |
954 | goto bad; | |
e498be7d CL |
955 | |
956 | /* we are serialised from CPU_DEAD or | |
b28a02de | 957 | CPU_UP_CANCELLED by the cpucontrol lock */ |
e498be7d CL |
958 | l3->shared = nc; |
959 | } | |
1da177e4 | 960 | } |
fc0abb14 | 961 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
962 | break; |
963 | case CPU_ONLINE: | |
964 | start_cpu_timer(cpu); | |
965 | break; | |
966 | #ifdef CONFIG_HOTPLUG_CPU | |
967 | case CPU_DEAD: | |
968 | /* fall thru */ | |
969 | case CPU_UP_CANCELED: | |
fc0abb14 | 970 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
971 | |
972 | list_for_each_entry(cachep, &cache_chain, next) { | |
973 | struct array_cache *nc; | |
e498be7d | 974 | cpumask_t mask; |
1da177e4 | 975 | |
e498be7d | 976 | mask = node_to_cpumask(node); |
1da177e4 LT |
977 | spin_lock_irq(&cachep->spinlock); |
978 | /* cpu is dead; no one can alloc from it. */ | |
979 | nc = cachep->array[cpu]; | |
980 | cachep->array[cpu] = NULL; | |
e498be7d CL |
981 | l3 = cachep->nodelists[node]; |
982 | ||
983 | if (!l3) | |
984 | goto unlock_cache; | |
985 | ||
986 | spin_lock(&l3->list_lock); | |
987 | ||
988 | /* Free limit for this kmem_list3 */ | |
989 | l3->free_limit -= cachep->batchcount; | |
990 | if (nc) | |
ff69416e | 991 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
992 | |
993 | if (!cpus_empty(mask)) { | |
b28a02de PE |
994 | spin_unlock(&l3->list_lock); |
995 | goto unlock_cache; | |
996 | } | |
e498be7d CL |
997 | |
998 | if (l3->shared) { | |
999 | free_block(cachep, l3->shared->entry, | |
b28a02de | 1000 | l3->shared->avail, node); |
e498be7d CL |
1001 | kfree(l3->shared); |
1002 | l3->shared = NULL; | |
1003 | } | |
1004 | if (l3->alien) { | |
1005 | drain_alien_cache(cachep, l3); | |
1006 | free_alien_cache(l3->alien); | |
1007 | l3->alien = NULL; | |
1008 | } | |
1009 | ||
1010 | /* free slabs belonging to this node */ | |
1011 | if (__node_shrink(cachep, node)) { | |
1012 | cachep->nodelists[node] = NULL; | |
1013 | spin_unlock(&l3->list_lock); | |
1014 | kfree(l3); | |
1015 | } else { | |
1016 | spin_unlock(&l3->list_lock); | |
1017 | } | |
b28a02de | 1018 | unlock_cache: |
1da177e4 LT |
1019 | spin_unlock_irq(&cachep->spinlock); |
1020 | kfree(nc); | |
1021 | } | |
fc0abb14 | 1022 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1023 | break; |
1024 | #endif | |
1025 | } | |
1026 | return NOTIFY_OK; | |
b28a02de | 1027 | bad: |
fc0abb14 | 1028 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1029 | return NOTIFY_BAD; |
1030 | } | |
1031 | ||
1032 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
1033 | ||
e498be7d CL |
1034 | /* |
1035 | * swap the static kmem_list3 with kmalloced memory | |
1036 | */ | |
b28a02de | 1037 | static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid) |
e498be7d CL |
1038 | { |
1039 | struct kmem_list3 *ptr; | |
1040 | ||
1041 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1042 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1043 | BUG_ON(!ptr); | |
1044 | ||
1045 | local_irq_disable(); | |
1046 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1047 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1048 | cachep->nodelists[nodeid] = ptr; | |
1049 | local_irq_enable(); | |
1050 | } | |
1051 | ||
1da177e4 LT |
1052 | /* Initialisation. |
1053 | * Called after the gfp() functions have been enabled, and before smp_init(). | |
1054 | */ | |
1055 | void __init kmem_cache_init(void) | |
1056 | { | |
1057 | size_t left_over; | |
1058 | struct cache_sizes *sizes; | |
1059 | struct cache_names *names; | |
e498be7d CL |
1060 | int i; |
1061 | ||
1062 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1063 | kmem_list3_init(&initkmem_list3[i]); | |
1064 | if (i < MAX_NUMNODES) | |
1065 | cache_cache.nodelists[i] = NULL; | |
1066 | } | |
1da177e4 LT |
1067 | |
1068 | /* | |
1069 | * Fragmentation resistance on low memory - only use bigger | |
1070 | * page orders on machines with more than 32MB of memory. | |
1071 | */ | |
1072 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1073 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1074 | ||
1da177e4 LT |
1075 | /* Bootstrap is tricky, because several objects are allocated |
1076 | * from caches that do not exist yet: | |
1077 | * 1) initialize the cache_cache cache: it contains the kmem_cache_t | |
1078 | * structures of all caches, except cache_cache itself: cache_cache | |
1079 | * is statically allocated. | |
e498be7d CL |
1080 | * Initially an __init data area is used for the head array and the |
1081 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1082 | * array at the end of the bootstrap. | |
1da177e4 | 1083 | * 2) Create the first kmalloc cache. |
e498be7d CL |
1084 | * The kmem_cache_t for the new cache is allocated normally. |
1085 | * An __init data area is used for the head array. | |
1086 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1087 | * head arrays. | |
1da177e4 LT |
1088 | * 4) Replace the __init data head arrays for cache_cache and the first |
1089 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1090 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1091 | * the other cache's with kmalloc allocated memory. | |
1092 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1093 | */ |
1094 | ||
1095 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1096 | INIT_LIST_HEAD(&cache_chain); |
1097 | list_add(&cache_cache.next, &cache_chain); | |
1098 | cache_cache.colour_off = cache_line_size(); | |
1099 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1100 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1101 | |
3dafccf2 | 1102 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size()); |
1da177e4 | 1103 | |
3dafccf2 | 1104 | cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0, |
b28a02de | 1105 | &left_over, &cache_cache.num); |
1da177e4 LT |
1106 | if (!cache_cache.num) |
1107 | BUG(); | |
1108 | ||
b28a02de | 1109 | cache_cache.colour = left_over / cache_cache.colour_off; |
1da177e4 | 1110 | cache_cache.colour_next = 0; |
b28a02de PE |
1111 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1112 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1113 | |
1114 | /* 2+3) create the kmalloc caches */ | |
1115 | sizes = malloc_sizes; | |
1116 | names = cache_names; | |
1117 | ||
e498be7d CL |
1118 | /* Initialize the caches that provide memory for the array cache |
1119 | * and the kmem_list3 structures first. | |
1120 | * Without this, further allocations will bug | |
1121 | */ | |
1122 | ||
1123 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
b28a02de PE |
1124 | sizes[INDEX_AC].cs_size, |
1125 | ARCH_KMALLOC_MINALIGN, | |
1126 | (ARCH_KMALLOC_FLAGS | | |
1127 | SLAB_PANIC), NULL, NULL); | |
e498be7d CL |
1128 | |
1129 | if (INDEX_AC != INDEX_L3) | |
1130 | sizes[INDEX_L3].cs_cachep = | |
b28a02de PE |
1131 | kmem_cache_create(names[INDEX_L3].name, |
1132 | sizes[INDEX_L3].cs_size, | |
1133 | ARCH_KMALLOC_MINALIGN, | |
1134 | (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, | |
1135 | NULL); | |
e498be7d | 1136 | |
1da177e4 | 1137 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1138 | /* |
1139 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1140 | * This should be particularly beneficial on SMP boxes, as it |
1141 | * eliminates "false sharing". | |
1142 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1143 | * allow tighter packing of the smaller caches. |
1144 | */ | |
b28a02de | 1145 | if (!sizes->cs_cachep) |
e498be7d | 1146 | sizes->cs_cachep = kmem_cache_create(names->name, |
b28a02de PE |
1147 | sizes->cs_size, |
1148 | ARCH_KMALLOC_MINALIGN, | |
1149 | (ARCH_KMALLOC_FLAGS | |
1150 | | SLAB_PANIC), | |
1151 | NULL, NULL); | |
1da177e4 LT |
1152 | |
1153 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
1154 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
b28a02de | 1155 | offslab_limit = sizes->cs_size - sizeof(struct slab); |
1da177e4 LT |
1156 | offslab_limit /= sizeof(kmem_bufctl_t); |
1157 | } | |
1158 | ||
1159 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
b28a02de PE |
1160 | sizes->cs_size, |
1161 | ARCH_KMALLOC_MINALIGN, | |
1162 | (ARCH_KMALLOC_FLAGS | | |
1163 | SLAB_CACHE_DMA | | |
1164 | SLAB_PANIC), NULL, | |
1165 | NULL); | |
1da177e4 LT |
1166 | |
1167 | sizes++; | |
1168 | names++; | |
1169 | } | |
1170 | /* 4) Replace the bootstrap head arrays */ | |
1171 | { | |
b28a02de | 1172 | void *ptr; |
e498be7d | 1173 | |
1da177e4 | 1174 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1175 | |
1da177e4 LT |
1176 | local_irq_disable(); |
1177 | BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache); | |
e498be7d | 1178 | memcpy(ptr, ac_data(&cache_cache), |
b28a02de | 1179 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1180 | cache_cache.array[smp_processor_id()] = ptr; |
1181 | local_irq_enable(); | |
e498be7d | 1182 | |
1da177e4 | 1183 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1184 | |
1da177e4 | 1185 | local_irq_disable(); |
e498be7d | 1186 | BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1187 | != &initarray_generic.cache); |
e498be7d | 1188 | memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1189 | sizeof(struct arraycache_init)); |
e498be7d | 1190 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1191 | ptr; |
1da177e4 LT |
1192 | local_irq_enable(); |
1193 | } | |
e498be7d CL |
1194 | /* 5) Replace the bootstrap kmem_list3's */ |
1195 | { | |
1196 | int node; | |
1197 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1198 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1199 | numa_node_id()); |
e498be7d CL |
1200 | |
1201 | for_each_online_node(node) { | |
1202 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1203 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1204 | |
1205 | if (INDEX_AC != INDEX_L3) { | |
1206 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1207 | &initkmem_list3[SIZE_L3 + node], |
1208 | node); | |
e498be7d CL |
1209 | } |
1210 | } | |
1211 | } | |
1da177e4 | 1212 | |
e498be7d | 1213 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 LT |
1214 | { |
1215 | kmem_cache_t *cachep; | |
fc0abb14 | 1216 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1217 | list_for_each_entry(cachep, &cache_chain, next) |
b28a02de | 1218 | enable_cpucache(cachep); |
fc0abb14 | 1219 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1220 | } |
1221 | ||
1222 | /* Done! */ | |
1223 | g_cpucache_up = FULL; | |
1224 | ||
1225 | /* Register a cpu startup notifier callback | |
1226 | * that initializes ac_data for all new cpus | |
1227 | */ | |
1228 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 LT |
1229 | |
1230 | /* The reap timers are started later, with a module init call: | |
1231 | * That part of the kernel is not yet operational. | |
1232 | */ | |
1233 | } | |
1234 | ||
1235 | static int __init cpucache_init(void) | |
1236 | { | |
1237 | int cpu; | |
1238 | ||
1239 | /* | |
1240 | * Register the timers that return unneeded | |
1241 | * pages to gfp. | |
1242 | */ | |
e498be7d | 1243 | for_each_online_cpu(cpu) |
b28a02de | 1244 | start_cpu_timer(cpu); |
1da177e4 LT |
1245 | |
1246 | return 0; | |
1247 | } | |
1248 | ||
1249 | __initcall(cpucache_init); | |
1250 | ||
1251 | /* | |
1252 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1253 | * | |
1254 | * If we requested dmaable memory, we will get it. Even if we | |
1255 | * did not request dmaable memory, we might get it, but that | |
1256 | * would be relatively rare and ignorable. | |
1257 | */ | |
dd0fc66f | 1258 | static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1259 | { |
1260 | struct page *page; | |
1261 | void *addr; | |
1262 | int i; | |
1263 | ||
1264 | flags |= cachep->gfpflags; | |
50c85a19 | 1265 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); |
1da177e4 LT |
1266 | if (!page) |
1267 | return NULL; | |
1268 | addr = page_address(page); | |
1269 | ||
1270 | i = (1 << cachep->gfporder); | |
1271 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1272 | atomic_add(i, &slab_reclaim_pages); | |
1273 | add_page_state(nr_slab, i); | |
1274 | while (i--) { | |
1275 | SetPageSlab(page); | |
1276 | page++; | |
1277 | } | |
1278 | return addr; | |
1279 | } | |
1280 | ||
1281 | /* | |
1282 | * Interface to system's page release. | |
1283 | */ | |
1284 | static void kmem_freepages(kmem_cache_t *cachep, void *addr) | |
1285 | { | |
b28a02de | 1286 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1287 | struct page *page = virt_to_page(addr); |
1288 | const unsigned long nr_freed = i; | |
1289 | ||
1290 | while (i--) { | |
1291 | if (!TestClearPageSlab(page)) | |
1292 | BUG(); | |
1293 | page++; | |
1294 | } | |
1295 | sub_page_state(nr_slab, nr_freed); | |
1296 | if (current->reclaim_state) | |
1297 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1298 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1299 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1300 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1301 | } |
1302 | ||
1303 | static void kmem_rcu_free(struct rcu_head *head) | |
1304 | { | |
b28a02de | 1305 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
1da177e4 LT |
1306 | kmem_cache_t *cachep = slab_rcu->cachep; |
1307 | ||
1308 | kmem_freepages(cachep, slab_rcu->addr); | |
1309 | if (OFF_SLAB(cachep)) | |
1310 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1311 | } | |
1312 | ||
1313 | #if DEBUG | |
1314 | ||
1315 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
1316 | static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, | |
b28a02de | 1317 | unsigned long caller) |
1da177e4 | 1318 | { |
3dafccf2 | 1319 | int size = obj_size(cachep); |
1da177e4 | 1320 | |
3dafccf2 | 1321 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1322 | |
b28a02de | 1323 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1324 | return; |
1325 | ||
b28a02de PE |
1326 | *addr++ = 0x12345678; |
1327 | *addr++ = caller; | |
1328 | *addr++ = smp_processor_id(); | |
1329 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1330 | { |
1331 | unsigned long *sptr = &caller; | |
1332 | unsigned long svalue; | |
1333 | ||
1334 | while (!kstack_end(sptr)) { | |
1335 | svalue = *sptr++; | |
1336 | if (kernel_text_address(svalue)) { | |
b28a02de | 1337 | *addr++ = svalue; |
1da177e4 LT |
1338 | size -= sizeof(unsigned long); |
1339 | if (size <= sizeof(unsigned long)) | |
1340 | break; | |
1341 | } | |
1342 | } | |
1343 | ||
1344 | } | |
b28a02de | 1345 | *addr++ = 0x87654321; |
1da177e4 LT |
1346 | } |
1347 | #endif | |
1348 | ||
1349 | static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val) | |
1350 | { | |
3dafccf2 MS |
1351 | int size = obj_size(cachep); |
1352 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1353 | |
1354 | memset(addr, val, size); | |
b28a02de | 1355 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1356 | } |
1357 | ||
1358 | static void dump_line(char *data, int offset, int limit) | |
1359 | { | |
1360 | int i; | |
1361 | printk(KERN_ERR "%03x:", offset); | |
b28a02de PE |
1362 | for (i = 0; i < limit; i++) { |
1363 | printk(" %02x", (unsigned char)data[offset + i]); | |
1da177e4 LT |
1364 | } |
1365 | printk("\n"); | |
1366 | } | |
1367 | #endif | |
1368 | ||
1369 | #if DEBUG | |
1370 | ||
1371 | static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines) | |
1372 | { | |
1373 | int i, size; | |
1374 | char *realobj; | |
1375 | ||
1376 | if (cachep->flags & SLAB_RED_ZONE) { | |
1377 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
b28a02de PE |
1378 | *dbg_redzone1(cachep, objp), |
1379 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1380 | } |
1381 | ||
1382 | if (cachep->flags & SLAB_STORE_USER) { | |
1383 | printk(KERN_ERR "Last user: [<%p>]", | |
b28a02de | 1384 | *dbg_userword(cachep, objp)); |
1da177e4 | 1385 | print_symbol("(%s)", |
b28a02de | 1386 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1387 | printk("\n"); |
1388 | } | |
3dafccf2 MS |
1389 | realobj = (char *)objp + obj_offset(cachep); |
1390 | size = obj_size(cachep); | |
b28a02de | 1391 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1392 | int limit; |
1393 | limit = 16; | |
b28a02de PE |
1394 | if (i + limit > size) |
1395 | limit = size - i; | |
1da177e4 LT |
1396 | dump_line(realobj, i, limit); |
1397 | } | |
1398 | } | |
1399 | ||
1400 | static void check_poison_obj(kmem_cache_t *cachep, void *objp) | |
1401 | { | |
1402 | char *realobj; | |
1403 | int size, i; | |
1404 | int lines = 0; | |
1405 | ||
3dafccf2 MS |
1406 | realobj = (char *)objp + obj_offset(cachep); |
1407 | size = obj_size(cachep); | |
1da177e4 | 1408 | |
b28a02de | 1409 | for (i = 0; i < size; i++) { |
1da177e4 | 1410 | char exp = POISON_FREE; |
b28a02de | 1411 | if (i == size - 1) |
1da177e4 LT |
1412 | exp = POISON_END; |
1413 | if (realobj[i] != exp) { | |
1414 | int limit; | |
1415 | /* Mismatch ! */ | |
1416 | /* Print header */ | |
1417 | if (lines == 0) { | |
b28a02de PE |
1418 | printk(KERN_ERR |
1419 | "Slab corruption: start=%p, len=%d\n", | |
1420 | realobj, size); | |
1da177e4 LT |
1421 | print_objinfo(cachep, objp, 0); |
1422 | } | |
1423 | /* Hexdump the affected line */ | |
b28a02de | 1424 | i = (i / 16) * 16; |
1da177e4 | 1425 | limit = 16; |
b28a02de PE |
1426 | if (i + limit > size) |
1427 | limit = size - i; | |
1da177e4 LT |
1428 | dump_line(realobj, i, limit); |
1429 | i += 16; | |
1430 | lines++; | |
1431 | /* Limit to 5 lines */ | |
1432 | if (lines > 5) | |
1433 | break; | |
1434 | } | |
1435 | } | |
1436 | if (lines != 0) { | |
1437 | /* Print some data about the neighboring objects, if they | |
1438 | * exist: | |
1439 | */ | |
065d41cb | 1440 | struct slab *slabp = page_get_slab(virt_to_page(objp)); |
1da177e4 LT |
1441 | int objnr; |
1442 | ||
3dafccf2 | 1443 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
1da177e4 | 1444 | if (objnr) { |
3dafccf2 MS |
1445 | objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size; |
1446 | realobj = (char *)objp + obj_offset(cachep); | |
1da177e4 | 1447 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1448 | realobj, size); |
1da177e4 LT |
1449 | print_objinfo(cachep, objp, 2); |
1450 | } | |
b28a02de | 1451 | if (objnr + 1 < cachep->num) { |
3dafccf2 MS |
1452 | objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size; |
1453 | realobj = (char *)objp + obj_offset(cachep); | |
1da177e4 | 1454 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1455 | realobj, size); |
1da177e4 LT |
1456 | print_objinfo(cachep, objp, 2); |
1457 | } | |
1458 | } | |
1459 | } | |
1460 | #endif | |
1461 | ||
1462 | /* Destroy all the objs in a slab, and release the mem back to the system. | |
1463 | * Before calling the slab must have been unlinked from the cache. | |
1464 | * The cache-lock is not held/needed. | |
1465 | */ | |
b28a02de | 1466 | static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp) |
1da177e4 LT |
1467 | { |
1468 | void *addr = slabp->s_mem - slabp->colouroff; | |
1469 | ||
1470 | #if DEBUG | |
1471 | int i; | |
1472 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 1473 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
1da177e4 LT |
1474 | |
1475 | if (cachep->flags & SLAB_POISON) { | |
1476 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 1477 | if ((cachep->buffer_size % PAGE_SIZE) == 0 |
b28a02de PE |
1478 | && OFF_SLAB(cachep)) |
1479 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 1480 | cachep->buffer_size / PAGE_SIZE, |
b28a02de | 1481 | 1); |
1da177e4 LT |
1482 | else |
1483 | check_poison_obj(cachep, objp); | |
1484 | #else | |
1485 | check_poison_obj(cachep, objp); | |
1486 | #endif | |
1487 | } | |
1488 | if (cachep->flags & SLAB_RED_ZONE) { | |
1489 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1490 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1491 | "was overwritten"); |
1da177e4 LT |
1492 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1493 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1494 | "was overwritten"); |
1da177e4 LT |
1495 | } |
1496 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1497 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 LT |
1498 | } |
1499 | #else | |
1500 | if (cachep->dtor) { | |
1501 | int i; | |
1502 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 1503 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
b28a02de | 1504 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1505 | } |
1506 | } | |
1507 | #endif | |
1508 | ||
1509 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { | |
1510 | struct slab_rcu *slab_rcu; | |
1511 | ||
b28a02de | 1512 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1513 | slab_rcu->cachep = cachep; |
1514 | slab_rcu->addr = addr; | |
1515 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1516 | } else { | |
1517 | kmem_freepages(cachep, addr); | |
1518 | if (OFF_SLAB(cachep)) | |
1519 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1520 | } | |
1521 | } | |
1522 | ||
3dafccf2 | 1523 | /* For setting up all the kmem_list3s for cache whose buffer_size is same |
e498be7d CL |
1524 | as size of kmem_list3. */ |
1525 | static inline void set_up_list3s(kmem_cache_t *cachep, int index) | |
1526 | { | |
1527 | int node; | |
1528 | ||
1529 | for_each_online_node(node) { | |
b28a02de | 1530 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1531 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1532 | REAPTIMEOUT_LIST3 + |
1533 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1534 | } |
1535 | } | |
1536 | ||
4d268eba PE |
1537 | /** |
1538 | * calculate_slab_order - calculate size (page order) of slabs and the number | |
1539 | * of objects per slab. | |
1540 | * | |
1541 | * This could be made much more intelligent. For now, try to avoid using | |
1542 | * high order pages for slabs. When the gfp() functions are more friendly | |
1543 | * towards high-order requests, this should be changed. | |
1544 | */ | |
1545 | static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size, | |
1546 | size_t align, gfp_t flags) | |
1547 | { | |
1548 | size_t left_over = 0; | |
1549 | ||
b28a02de | 1550 | for (;; cachep->gfporder++) { |
4d268eba PE |
1551 | unsigned int num; |
1552 | size_t remainder; | |
1553 | ||
1554 | if (cachep->gfporder > MAX_GFP_ORDER) { | |
1555 | cachep->num = 0; | |
1556 | break; | |
1557 | } | |
1558 | ||
1559 | cache_estimate(cachep->gfporder, size, align, flags, | |
1560 | &remainder, &num); | |
1561 | if (!num) | |
1562 | continue; | |
1563 | /* More than offslab_limit objects will cause problems */ | |
1564 | if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit) | |
1565 | break; | |
1566 | ||
1567 | cachep->num = num; | |
1568 | left_over = remainder; | |
1569 | ||
1570 | /* | |
1571 | * Large number of objects is good, but very large slabs are | |
1572 | * currently bad for the gfp()s. | |
1573 | */ | |
1574 | if (cachep->gfporder >= slab_break_gfp_order) | |
1575 | break; | |
1576 | ||
1577 | if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder)) | |
1578 | /* Acceptable internal fragmentation */ | |
1579 | break; | |
1580 | } | |
1581 | return left_over; | |
1582 | } | |
1583 | ||
1da177e4 LT |
1584 | /** |
1585 | * kmem_cache_create - Create a cache. | |
1586 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1587 | * @size: The size of objects to be created in this cache. | |
1588 | * @align: The required alignment for the objects. | |
1589 | * @flags: SLAB flags | |
1590 | * @ctor: A constructor for the objects. | |
1591 | * @dtor: A destructor for the objects. | |
1592 | * | |
1593 | * Returns a ptr to the cache on success, NULL on failure. | |
1594 | * Cannot be called within a int, but can be interrupted. | |
1595 | * The @ctor is run when new pages are allocated by the cache | |
1596 | * and the @dtor is run before the pages are handed back. | |
1597 | * | |
1598 | * @name must be valid until the cache is destroyed. This implies that | |
1599 | * the module calling this has to destroy the cache before getting | |
1600 | * unloaded. | |
1601 | * | |
1602 | * The flags are | |
1603 | * | |
1604 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1605 | * to catch references to uninitialised memory. | |
1606 | * | |
1607 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1608 | * for buffer overruns. | |
1609 | * | |
1610 | * %SLAB_NO_REAP - Don't automatically reap this cache when we're under | |
1611 | * memory pressure. | |
1612 | * | |
1613 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
1614 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1615 | * as davem. | |
1616 | */ | |
1617 | kmem_cache_t * | |
1618 | kmem_cache_create (const char *name, size_t size, size_t align, | |
1619 | unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long), | |
1620 | void (*dtor)(void*, kmem_cache_t *, unsigned long)) | |
1621 | { | |
1622 | size_t left_over, slab_size, ralign; | |
1623 | kmem_cache_t *cachep = NULL; | |
4f12bb4f | 1624 | struct list_head *p; |
1da177e4 LT |
1625 | |
1626 | /* | |
1627 | * Sanity checks... these are all serious usage bugs. | |
1628 | */ | |
1629 | if ((!name) || | |
b28a02de PE |
1630 | in_interrupt() || |
1631 | (size < BYTES_PER_WORD) || | |
1632 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { | |
1633 | printk(KERN_ERR "%s: Early error in slab %s\n", | |
1634 | __FUNCTION__, name); | |
1635 | BUG(); | |
1636 | } | |
1da177e4 | 1637 | |
fc0abb14 | 1638 | mutex_lock(&cache_chain_mutex); |
4f12bb4f AM |
1639 | |
1640 | list_for_each(p, &cache_chain) { | |
1641 | kmem_cache_t *pc = list_entry(p, kmem_cache_t, next); | |
1642 | mm_segment_t old_fs = get_fs(); | |
1643 | char tmp; | |
1644 | int res; | |
1645 | ||
1646 | /* | |
1647 | * This happens when the module gets unloaded and doesn't | |
1648 | * destroy its slab cache and no-one else reuses the vmalloc | |
1649 | * area of the module. Print a warning. | |
1650 | */ | |
1651 | set_fs(KERNEL_DS); | |
1652 | res = __get_user(tmp, pc->name); | |
1653 | set_fs(old_fs); | |
1654 | if (res) { | |
1655 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 1656 | pc->buffer_size); |
4f12bb4f AM |
1657 | continue; |
1658 | } | |
1659 | ||
b28a02de | 1660 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1661 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1662 | dump_stack(); | |
1663 | goto oops; | |
1664 | } | |
1665 | } | |
1666 | ||
1da177e4 LT |
1667 | #if DEBUG |
1668 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1669 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1670 | /* No constructor, but inital state check requested */ | |
1671 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 1672 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
1673 | flags &= ~SLAB_DEBUG_INITIAL; |
1674 | } | |
1da177e4 LT |
1675 | #if FORCED_DEBUG |
1676 | /* | |
1677 | * Enable redzoning and last user accounting, except for caches with | |
1678 | * large objects, if the increased size would increase the object size | |
1679 | * above the next power of two: caches with object sizes just above a | |
1680 | * power of two have a significant amount of internal fragmentation. | |
1681 | */ | |
b28a02de PE |
1682 | if ((size < 4096 |
1683 | || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD))) | |
1684 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; | |
1da177e4 LT |
1685 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
1686 | flags |= SLAB_POISON; | |
1687 | #endif | |
1688 | if (flags & SLAB_DESTROY_BY_RCU) | |
1689 | BUG_ON(flags & SLAB_POISON); | |
1690 | #endif | |
1691 | if (flags & SLAB_DESTROY_BY_RCU) | |
1692 | BUG_ON(dtor); | |
1693 | ||
1694 | /* | |
1695 | * Always checks flags, a caller might be expecting debug | |
1696 | * support which isn't available. | |
1697 | */ | |
1698 | if (flags & ~CREATE_MASK) | |
1699 | BUG(); | |
1700 | ||
1701 | /* Check that size is in terms of words. This is needed to avoid | |
1702 | * unaligned accesses for some archs when redzoning is used, and makes | |
1703 | * sure any on-slab bufctl's are also correctly aligned. | |
1704 | */ | |
b28a02de PE |
1705 | if (size & (BYTES_PER_WORD - 1)) { |
1706 | size += (BYTES_PER_WORD - 1); | |
1707 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
1708 | } |
1709 | ||
1710 | /* calculate out the final buffer alignment: */ | |
1711 | /* 1) arch recommendation: can be overridden for debug */ | |
1712 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1713 | /* Default alignment: as specified by the arch code. | |
1714 | * Except if an object is really small, then squeeze multiple | |
1715 | * objects into one cacheline. | |
1716 | */ | |
1717 | ralign = cache_line_size(); | |
b28a02de | 1718 | while (size <= ralign / 2) |
1da177e4 LT |
1719 | ralign /= 2; |
1720 | } else { | |
1721 | ralign = BYTES_PER_WORD; | |
1722 | } | |
1723 | /* 2) arch mandated alignment: disables debug if necessary */ | |
1724 | if (ralign < ARCH_SLAB_MINALIGN) { | |
1725 | ralign = ARCH_SLAB_MINALIGN; | |
1726 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1727 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1728 | } |
1729 | /* 3) caller mandated alignment: disables debug if necessary */ | |
1730 | if (ralign < align) { | |
1731 | ralign = align; | |
1732 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1733 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1734 | } |
1735 | /* 4) Store it. Note that the debug code below can reduce | |
1736 | * the alignment to BYTES_PER_WORD. | |
1737 | */ | |
1738 | align = ralign; | |
1739 | ||
1740 | /* Get cache's description obj. */ | |
1741 | cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL); | |
1742 | if (!cachep) | |
4f12bb4f | 1743 | goto oops; |
1da177e4 LT |
1744 | memset(cachep, 0, sizeof(kmem_cache_t)); |
1745 | ||
1746 | #if DEBUG | |
3dafccf2 | 1747 | cachep->obj_size = size; |
1da177e4 LT |
1748 | |
1749 | if (flags & SLAB_RED_ZONE) { | |
1750 | /* redzoning only works with word aligned caches */ | |
1751 | align = BYTES_PER_WORD; | |
1752 | ||
1753 | /* add space for red zone words */ | |
3dafccf2 | 1754 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 1755 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
1756 | } |
1757 | if (flags & SLAB_STORE_USER) { | |
1758 | /* user store requires word alignment and | |
1759 | * one word storage behind the end of the real | |
1760 | * object. | |
1761 | */ | |
1762 | align = BYTES_PER_WORD; | |
1763 | size += BYTES_PER_WORD; | |
1764 | } | |
1765 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 1766 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
1767 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
1768 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
1769 | size = PAGE_SIZE; |
1770 | } | |
1771 | #endif | |
1772 | #endif | |
1773 | ||
1774 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 1775 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
1776 | /* |
1777 | * Size is large, assume best to place the slab management obj | |
1778 | * off-slab (should allow better packing of objs). | |
1779 | */ | |
1780 | flags |= CFLGS_OFF_SLAB; | |
1781 | ||
1782 | size = ALIGN(size, align); | |
1783 | ||
1784 | if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) { | |
1785 | /* | |
1786 | * A VFS-reclaimable slab tends to have most allocations | |
1787 | * as GFP_NOFS and we really don't want to have to be allocating | |
1788 | * higher-order pages when we are unable to shrink dcache. | |
1789 | */ | |
1790 | cachep->gfporder = 0; | |
1791 | cache_estimate(cachep->gfporder, size, align, flags, | |
b28a02de | 1792 | &left_over, &cachep->num); |
4d268eba PE |
1793 | } else |
1794 | left_over = calculate_slab_order(cachep, size, align, flags); | |
1da177e4 LT |
1795 | |
1796 | if (!cachep->num) { | |
1797 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
1798 | kmem_cache_free(&cache_cache, cachep); | |
1799 | cachep = NULL; | |
4f12bb4f | 1800 | goto oops; |
1da177e4 | 1801 | } |
b28a02de PE |
1802 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
1803 | + sizeof(struct slab), align); | |
1da177e4 LT |
1804 | |
1805 | /* | |
1806 | * If the slab has been placed off-slab, and we have enough space then | |
1807 | * move it on-slab. This is at the expense of any extra colouring. | |
1808 | */ | |
1809 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
1810 | flags &= ~CFLGS_OFF_SLAB; | |
1811 | left_over -= slab_size; | |
1812 | } | |
1813 | ||
1814 | if (flags & CFLGS_OFF_SLAB) { | |
1815 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
1816 | slab_size = |
1817 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
1818 | } |
1819 | ||
1820 | cachep->colour_off = cache_line_size(); | |
1821 | /* Offset must be a multiple of the alignment. */ | |
1822 | if (cachep->colour_off < align) | |
1823 | cachep->colour_off = align; | |
b28a02de | 1824 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
1825 | cachep->slab_size = slab_size; |
1826 | cachep->flags = flags; | |
1827 | cachep->gfpflags = 0; | |
1828 | if (flags & SLAB_CACHE_DMA) | |
1829 | cachep->gfpflags |= GFP_DMA; | |
1830 | spin_lock_init(&cachep->spinlock); | |
3dafccf2 | 1831 | cachep->buffer_size = size; |
1da177e4 LT |
1832 | |
1833 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 1834 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
1835 | cachep->ctor = ctor; |
1836 | cachep->dtor = dtor; | |
1837 | cachep->name = name; | |
1838 | ||
1839 | /* Don't let CPUs to come and go */ | |
1840 | lock_cpu_hotplug(); | |
1841 | ||
1842 | if (g_cpucache_up == FULL) { | |
1843 | enable_cpucache(cachep); | |
1844 | } else { | |
1845 | if (g_cpucache_up == NONE) { | |
1846 | /* Note: the first kmem_cache_create must create | |
1847 | * the cache that's used by kmalloc(24), otherwise | |
1848 | * the creation of further caches will BUG(). | |
1849 | */ | |
e498be7d | 1850 | cachep->array[smp_processor_id()] = |
b28a02de | 1851 | &initarray_generic.cache; |
e498be7d CL |
1852 | |
1853 | /* If the cache that's used by | |
1854 | * kmalloc(sizeof(kmem_list3)) is the first cache, | |
1855 | * then we need to set up all its list3s, otherwise | |
1856 | * the creation of further caches will BUG(). | |
1857 | */ | |
1858 | set_up_list3s(cachep, SIZE_AC); | |
1859 | if (INDEX_AC == INDEX_L3) | |
1860 | g_cpucache_up = PARTIAL_L3; | |
1861 | else | |
1862 | g_cpucache_up = PARTIAL_AC; | |
1da177e4 | 1863 | } else { |
e498be7d | 1864 | cachep->array[smp_processor_id()] = |
b28a02de | 1865 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d CL |
1866 | |
1867 | if (g_cpucache_up == PARTIAL_AC) { | |
1868 | set_up_list3s(cachep, SIZE_L3); | |
1869 | g_cpucache_up = PARTIAL_L3; | |
1870 | } else { | |
1871 | int node; | |
1872 | for_each_online_node(node) { | |
1873 | ||
1874 | cachep->nodelists[node] = | |
b28a02de PE |
1875 | kmalloc_node(sizeof |
1876 | (struct kmem_list3), | |
1877 | GFP_KERNEL, node); | |
e498be7d | 1878 | BUG_ON(!cachep->nodelists[node]); |
b28a02de PE |
1879 | kmem_list3_init(cachep-> |
1880 | nodelists[node]); | |
e498be7d CL |
1881 | } |
1882 | } | |
1da177e4 | 1883 | } |
e498be7d | 1884 | cachep->nodelists[numa_node_id()]->next_reap = |
b28a02de PE |
1885 | jiffies + REAPTIMEOUT_LIST3 + |
1886 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1887 | |
1da177e4 LT |
1888 | BUG_ON(!ac_data(cachep)); |
1889 | ac_data(cachep)->avail = 0; | |
1890 | ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1891 | ac_data(cachep)->batchcount = 1; | |
1892 | ac_data(cachep)->touched = 0; | |
1893 | cachep->batchcount = 1; | |
1894 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
b28a02de | 1895 | } |
1da177e4 | 1896 | |
1da177e4 LT |
1897 | /* cache setup completed, link it into the list */ |
1898 | list_add(&cachep->next, &cache_chain); | |
1da177e4 | 1899 | unlock_cpu_hotplug(); |
b28a02de | 1900 | oops: |
1da177e4 LT |
1901 | if (!cachep && (flags & SLAB_PANIC)) |
1902 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 1903 | name); |
fc0abb14 | 1904 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1905 | return cachep; |
1906 | } | |
1907 | EXPORT_SYMBOL(kmem_cache_create); | |
1908 | ||
1909 | #if DEBUG | |
1910 | static void check_irq_off(void) | |
1911 | { | |
1912 | BUG_ON(!irqs_disabled()); | |
1913 | } | |
1914 | ||
1915 | static void check_irq_on(void) | |
1916 | { | |
1917 | BUG_ON(irqs_disabled()); | |
1918 | } | |
1919 | ||
1920 | static void check_spinlock_acquired(kmem_cache_t *cachep) | |
1921 | { | |
1922 | #ifdef CONFIG_SMP | |
1923 | check_irq_off(); | |
e498be7d | 1924 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
1925 | #endif |
1926 | } | |
e498be7d CL |
1927 | |
1928 | static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node) | |
1929 | { | |
1930 | #ifdef CONFIG_SMP | |
1931 | check_irq_off(); | |
1932 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
1933 | #endif | |
1934 | } | |
1935 | ||
1da177e4 LT |
1936 | #else |
1937 | #define check_irq_off() do { } while(0) | |
1938 | #define check_irq_on() do { } while(0) | |
1939 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 1940 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
1941 | #endif |
1942 | ||
1943 | /* | |
1944 | * Waits for all CPUs to execute func(). | |
1945 | */ | |
b28a02de | 1946 | static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg) |
1da177e4 LT |
1947 | { |
1948 | check_irq_on(); | |
1949 | preempt_disable(); | |
1950 | ||
1951 | local_irq_disable(); | |
1952 | func(arg); | |
1953 | local_irq_enable(); | |
1954 | ||
1955 | if (smp_call_function(func, arg, 1, 1)) | |
1956 | BUG(); | |
1957 | ||
1958 | preempt_enable(); | |
1959 | } | |
1960 | ||
b28a02de PE |
1961 | static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac, |
1962 | int force, int node); | |
1da177e4 LT |
1963 | |
1964 | static void do_drain(void *arg) | |
1965 | { | |
b28a02de | 1966 | kmem_cache_t *cachep = (kmem_cache_t *) arg; |
1da177e4 | 1967 | struct array_cache *ac; |
ff69416e | 1968 | int node = numa_node_id(); |
1da177e4 LT |
1969 | |
1970 | check_irq_off(); | |
1971 | ac = ac_data(cachep); | |
ff69416e CL |
1972 | spin_lock(&cachep->nodelists[node]->list_lock); |
1973 | free_block(cachep, ac->entry, ac->avail, node); | |
1974 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
1975 | ac->avail = 0; |
1976 | } | |
1977 | ||
1978 | static void drain_cpu_caches(kmem_cache_t *cachep) | |
1979 | { | |
e498be7d CL |
1980 | struct kmem_list3 *l3; |
1981 | int node; | |
1982 | ||
1da177e4 LT |
1983 | smp_call_function_all_cpus(do_drain, cachep); |
1984 | check_irq_on(); | |
1985 | spin_lock_irq(&cachep->spinlock); | |
b28a02de | 1986 | for_each_online_node(node) { |
e498be7d CL |
1987 | l3 = cachep->nodelists[node]; |
1988 | if (l3) { | |
1989 | spin_lock(&l3->list_lock); | |
1990 | drain_array_locked(cachep, l3->shared, 1, node); | |
1991 | spin_unlock(&l3->list_lock); | |
1992 | if (l3->alien) | |
1993 | drain_alien_cache(cachep, l3); | |
1994 | } | |
1995 | } | |
1da177e4 LT |
1996 | spin_unlock_irq(&cachep->spinlock); |
1997 | } | |
1998 | ||
e498be7d | 1999 | static int __node_shrink(kmem_cache_t *cachep, int node) |
1da177e4 LT |
2000 | { |
2001 | struct slab *slabp; | |
e498be7d | 2002 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
2003 | int ret; |
2004 | ||
e498be7d | 2005 | for (;;) { |
1da177e4 LT |
2006 | struct list_head *p; |
2007 | ||
e498be7d CL |
2008 | p = l3->slabs_free.prev; |
2009 | if (p == &l3->slabs_free) | |
1da177e4 LT |
2010 | break; |
2011 | ||
e498be7d | 2012 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 LT |
2013 | #if DEBUG |
2014 | if (slabp->inuse) | |
2015 | BUG(); | |
2016 | #endif | |
2017 | list_del(&slabp->list); | |
2018 | ||
e498be7d CL |
2019 | l3->free_objects -= cachep->num; |
2020 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 2021 | slab_destroy(cachep, slabp); |
e498be7d | 2022 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 2023 | } |
b28a02de | 2024 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
2025 | return ret; |
2026 | } | |
2027 | ||
e498be7d CL |
2028 | static int __cache_shrink(kmem_cache_t *cachep) |
2029 | { | |
2030 | int ret = 0, i = 0; | |
2031 | struct kmem_list3 *l3; | |
2032 | ||
2033 | drain_cpu_caches(cachep); | |
2034 | ||
2035 | check_irq_on(); | |
2036 | for_each_online_node(i) { | |
2037 | l3 = cachep->nodelists[i]; | |
2038 | if (l3) { | |
2039 | spin_lock_irq(&l3->list_lock); | |
2040 | ret += __node_shrink(cachep, i); | |
2041 | spin_unlock_irq(&l3->list_lock); | |
2042 | } | |
2043 | } | |
2044 | return (ret ? 1 : 0); | |
2045 | } | |
2046 | ||
1da177e4 LT |
2047 | /** |
2048 | * kmem_cache_shrink - Shrink a cache. | |
2049 | * @cachep: The cache to shrink. | |
2050 | * | |
2051 | * Releases as many slabs as possible for a cache. | |
2052 | * To help debugging, a zero exit status indicates all slabs were released. | |
2053 | */ | |
2054 | int kmem_cache_shrink(kmem_cache_t *cachep) | |
2055 | { | |
2056 | if (!cachep || in_interrupt()) | |
2057 | BUG(); | |
2058 | ||
2059 | return __cache_shrink(cachep); | |
2060 | } | |
2061 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2062 | ||
2063 | /** | |
2064 | * kmem_cache_destroy - delete a cache | |
2065 | * @cachep: the cache to destroy | |
2066 | * | |
2067 | * Remove a kmem_cache_t object from the slab cache. | |
2068 | * Returns 0 on success. | |
2069 | * | |
2070 | * It is expected this function will be called by a module when it is | |
2071 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2072 | * cache being allocated each time a module is loaded and unloaded, if the | |
2073 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2074 | * | |
2075 | * The cache must be empty before calling this function. | |
2076 | * | |
2077 | * The caller must guarantee that noone will allocate memory from the cache | |
2078 | * during the kmem_cache_destroy(). | |
2079 | */ | |
b28a02de | 2080 | int kmem_cache_destroy(kmem_cache_t *cachep) |
1da177e4 LT |
2081 | { |
2082 | int i; | |
e498be7d | 2083 | struct kmem_list3 *l3; |
1da177e4 LT |
2084 | |
2085 | if (!cachep || in_interrupt()) | |
2086 | BUG(); | |
2087 | ||
2088 | /* Don't let CPUs to come and go */ | |
2089 | lock_cpu_hotplug(); | |
2090 | ||
2091 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2092 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2093 | /* |
2094 | * the chain is never empty, cache_cache is never destroyed | |
2095 | */ | |
2096 | list_del(&cachep->next); | |
fc0abb14 | 2097 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2098 | |
2099 | if (__cache_shrink(cachep)) { | |
2100 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2101 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2102 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2103 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2104 | unlock_cpu_hotplug(); |
2105 | return 1; | |
2106 | } | |
2107 | ||
2108 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2109 | synchronize_rcu(); |
1da177e4 | 2110 | |
e498be7d | 2111 | for_each_online_cpu(i) |
b28a02de | 2112 | kfree(cachep->array[i]); |
1da177e4 LT |
2113 | |
2114 | /* NUMA: free the list3 structures */ | |
e498be7d CL |
2115 | for_each_online_node(i) { |
2116 | if ((l3 = cachep->nodelists[i])) { | |
2117 | kfree(l3->shared); | |
2118 | free_alien_cache(l3->alien); | |
2119 | kfree(l3); | |
2120 | } | |
2121 | } | |
1da177e4 LT |
2122 | kmem_cache_free(&cache_cache, cachep); |
2123 | ||
2124 | unlock_cpu_hotplug(); | |
2125 | ||
2126 | return 0; | |
2127 | } | |
2128 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2129 | ||
2130 | /* Get the memory for a slab management obj. */ | |
b28a02de PE |
2131 | static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp, |
2132 | int colour_off, gfp_t local_flags) | |
1da177e4 LT |
2133 | { |
2134 | struct slab *slabp; | |
b28a02de | 2135 | |
1da177e4 LT |
2136 | if (OFF_SLAB(cachep)) { |
2137 | /* Slab management obj is off-slab. */ | |
2138 | slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags); | |
2139 | if (!slabp) | |
2140 | return NULL; | |
2141 | } else { | |
b28a02de | 2142 | slabp = objp + colour_off; |
1da177e4 LT |
2143 | colour_off += cachep->slab_size; |
2144 | } | |
2145 | slabp->inuse = 0; | |
2146 | slabp->colouroff = colour_off; | |
b28a02de | 2147 | slabp->s_mem = objp + colour_off; |
1da177e4 LT |
2148 | |
2149 | return slabp; | |
2150 | } | |
2151 | ||
2152 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2153 | { | |
b28a02de | 2154 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2155 | } |
2156 | ||
2157 | static void cache_init_objs(kmem_cache_t *cachep, | |
b28a02de | 2158 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2159 | { |
2160 | int i; | |
2161 | ||
2162 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 2163 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
1da177e4 LT |
2164 | #if DEBUG |
2165 | /* need to poison the objs? */ | |
2166 | if (cachep->flags & SLAB_POISON) | |
2167 | poison_obj(cachep, objp, POISON_FREE); | |
2168 | if (cachep->flags & SLAB_STORE_USER) | |
2169 | *dbg_userword(cachep, objp) = NULL; | |
2170 | ||
2171 | if (cachep->flags & SLAB_RED_ZONE) { | |
2172 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2173 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2174 | } | |
2175 | /* | |
2176 | * Constructors are not allowed to allocate memory from | |
2177 | * the same cache which they are a constructor for. | |
2178 | * Otherwise, deadlock. They must also be threaded. | |
2179 | */ | |
2180 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2181 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2182 | ctor_flags); |
1da177e4 LT |
2183 | |
2184 | if (cachep->flags & SLAB_RED_ZONE) { | |
2185 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2186 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2187 | " end of an object"); |
1da177e4 LT |
2188 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2189 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2190 | " start of an object"); |
1da177e4 | 2191 | } |
3dafccf2 | 2192 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep) |
b28a02de PE |
2193 | && cachep->flags & SLAB_POISON) |
2194 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 2195 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2196 | #else |
2197 | if (cachep->ctor) | |
2198 | cachep->ctor(objp, cachep, ctor_flags); | |
2199 | #endif | |
b28a02de | 2200 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2201 | } |
b28a02de | 2202 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2203 | slabp->free = 0; |
2204 | } | |
2205 | ||
6daa0e28 | 2206 | static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2207 | { |
2208 | if (flags & SLAB_DMA) { | |
2209 | if (!(cachep->gfpflags & GFP_DMA)) | |
2210 | BUG(); | |
2211 | } else { | |
2212 | if (cachep->gfpflags & GFP_DMA) | |
2213 | BUG(); | |
2214 | } | |
2215 | } | |
2216 | ||
2217 | static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp) | |
2218 | { | |
2219 | int i; | |
2220 | struct page *page; | |
2221 | ||
2222 | /* Nasty!!!!!! I hope this is OK. */ | |
2223 | i = 1 << cachep->gfporder; | |
2224 | page = virt_to_page(objp); | |
2225 | do { | |
065d41cb PE |
2226 | page_set_cache(page, cachep); |
2227 | page_set_slab(page, slabp); | |
1da177e4 LT |
2228 | page++; |
2229 | } while (--i); | |
2230 | } | |
2231 | ||
2232 | /* | |
2233 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2234 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2235 | */ | |
dd0fc66f | 2236 | static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2237 | { |
b28a02de PE |
2238 | struct slab *slabp; |
2239 | void *objp; | |
2240 | size_t offset; | |
2241 | gfp_t local_flags; | |
2242 | unsigned long ctor_flags; | |
e498be7d | 2243 | struct kmem_list3 *l3; |
1da177e4 LT |
2244 | |
2245 | /* Be lazy and only check for valid flags here, | |
b28a02de | 2246 | * keeping it out of the critical path in kmem_cache_alloc(). |
1da177e4 | 2247 | */ |
b28a02de | 2248 | if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)) |
1da177e4 LT |
2249 | BUG(); |
2250 | if (flags & SLAB_NO_GROW) | |
2251 | return 0; | |
2252 | ||
2253 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2254 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2255 | if (!(local_flags & __GFP_WAIT)) | |
2256 | /* | |
2257 | * Not allowed to sleep. Need to tell a constructor about | |
2258 | * this - it might need to know... | |
2259 | */ | |
2260 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2261 | ||
2262 | /* About to mess with non-constant members - lock. */ | |
2263 | check_irq_off(); | |
2264 | spin_lock(&cachep->spinlock); | |
2265 | ||
2266 | /* Get colour for the slab, and cal the next value. */ | |
2267 | offset = cachep->colour_next; | |
2268 | cachep->colour_next++; | |
2269 | if (cachep->colour_next >= cachep->colour) | |
2270 | cachep->colour_next = 0; | |
2271 | offset *= cachep->colour_off; | |
2272 | ||
2273 | spin_unlock(&cachep->spinlock); | |
2274 | ||
e498be7d | 2275 | check_irq_off(); |
1da177e4 LT |
2276 | if (local_flags & __GFP_WAIT) |
2277 | local_irq_enable(); | |
2278 | ||
2279 | /* | |
2280 | * The test for missing atomic flag is performed here, rather than | |
2281 | * the more obvious place, simply to reduce the critical path length | |
2282 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2283 | * will eventually be caught here (where it matters). | |
2284 | */ | |
2285 | kmem_flagcheck(cachep, flags); | |
2286 | ||
e498be7d CL |
2287 | /* Get mem for the objs. |
2288 | * Attempt to allocate a physical page from 'nodeid', | |
2289 | */ | |
1da177e4 LT |
2290 | if (!(objp = kmem_getpages(cachep, flags, nodeid))) |
2291 | goto failed; | |
2292 | ||
2293 | /* Get slab management. */ | |
2294 | if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags))) | |
2295 | goto opps1; | |
2296 | ||
e498be7d | 2297 | slabp->nodeid = nodeid; |
1da177e4 LT |
2298 | set_slab_attr(cachep, slabp, objp); |
2299 | ||
2300 | cache_init_objs(cachep, slabp, ctor_flags); | |
2301 | ||
2302 | if (local_flags & __GFP_WAIT) | |
2303 | local_irq_disable(); | |
2304 | check_irq_off(); | |
e498be7d CL |
2305 | l3 = cachep->nodelists[nodeid]; |
2306 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2307 | |
2308 | /* Make slab active. */ | |
e498be7d | 2309 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2310 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2311 | l3->free_objects += cachep->num; |
2312 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2313 | return 1; |
b28a02de | 2314 | opps1: |
1da177e4 | 2315 | kmem_freepages(cachep, objp); |
b28a02de | 2316 | failed: |
1da177e4 LT |
2317 | if (local_flags & __GFP_WAIT) |
2318 | local_irq_disable(); | |
2319 | return 0; | |
2320 | } | |
2321 | ||
2322 | #if DEBUG | |
2323 | ||
2324 | /* | |
2325 | * Perform extra freeing checks: | |
2326 | * - detect bad pointers. | |
2327 | * - POISON/RED_ZONE checking | |
2328 | * - destructor calls, for caches with POISON+dtor | |
2329 | */ | |
2330 | static void kfree_debugcheck(const void *objp) | |
2331 | { | |
2332 | struct page *page; | |
2333 | ||
2334 | if (!virt_addr_valid(objp)) { | |
2335 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2336 | (unsigned long)objp); |
2337 | BUG(); | |
1da177e4 LT |
2338 | } |
2339 | page = virt_to_page(objp); | |
2340 | if (!PageSlab(page)) { | |
b28a02de PE |
2341 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2342 | (unsigned long)objp); | |
1da177e4 LT |
2343 | BUG(); |
2344 | } | |
2345 | } | |
2346 | ||
2347 | static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp, | |
b28a02de | 2348 | void *caller) |
1da177e4 LT |
2349 | { |
2350 | struct page *page; | |
2351 | unsigned int objnr; | |
2352 | struct slab *slabp; | |
2353 | ||
3dafccf2 | 2354 | objp -= obj_offset(cachep); |
1da177e4 LT |
2355 | kfree_debugcheck(objp); |
2356 | page = virt_to_page(objp); | |
2357 | ||
065d41cb | 2358 | if (page_get_cache(page) != cachep) { |
b28a02de PE |
2359 | printk(KERN_ERR |
2360 | "mismatch in kmem_cache_free: expected cache %p, got %p\n", | |
2361 | page_get_cache(page), cachep); | |
1da177e4 | 2362 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2363 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2364 | page_get_cache(page)->name); | |
1da177e4 LT |
2365 | WARN_ON(1); |
2366 | } | |
065d41cb | 2367 | slabp = page_get_slab(page); |
1da177e4 LT |
2368 | |
2369 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2370 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE |
2371 | || *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2372 | slab_error(cachep, | |
2373 | "double free, or memory outside" | |
2374 | " object was overwritten"); | |
2375 | printk(KERN_ERR | |
2376 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2377 | objp, *dbg_redzone1(cachep, objp), | |
2378 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2379 | } |
2380 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2381 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2382 | } | |
2383 | if (cachep->flags & SLAB_STORE_USER) | |
2384 | *dbg_userword(cachep, objp) = caller; | |
2385 | ||
3dafccf2 | 2386 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
1da177e4 LT |
2387 | |
2388 | BUG_ON(objnr >= cachep->num); | |
3dafccf2 | 2389 | BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size); |
1da177e4 LT |
2390 | |
2391 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
2392 | /* Need to call the slab's constructor so the | |
2393 | * caller can perform a verify of its state (debugging). | |
2394 | * Called without the cache-lock held. | |
2395 | */ | |
3dafccf2 | 2396 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2397 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2398 | } |
2399 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2400 | /* we want to cache poison the object, | |
2401 | * call the destruction callback | |
2402 | */ | |
3dafccf2 | 2403 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 LT |
2404 | } |
2405 | if (cachep->flags & SLAB_POISON) { | |
2406 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 2407 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) { |
1da177e4 | 2408 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2409 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2410 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2411 | } else { |
2412 | poison_obj(cachep, objp, POISON_FREE); | |
2413 | } | |
2414 | #else | |
2415 | poison_obj(cachep, objp, POISON_FREE); | |
2416 | #endif | |
2417 | } | |
2418 | return objp; | |
2419 | } | |
2420 | ||
2421 | static void check_slabp(kmem_cache_t *cachep, struct slab *slabp) | |
2422 | { | |
2423 | kmem_bufctl_t i; | |
2424 | int entries = 0; | |
b28a02de | 2425 | |
1da177e4 LT |
2426 | /* Check slab's freelist to see if this obj is there. */ |
2427 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2428 | entries++; | |
2429 | if (entries > cachep->num || i >= cachep->num) | |
2430 | goto bad; | |
2431 | } | |
2432 | if (entries != cachep->num - slabp->inuse) { | |
b28a02de PE |
2433 | bad: |
2434 | printk(KERN_ERR | |
2435 | "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2436 | cachep->name, cachep->num, slabp, slabp->inuse); | |
2437 | for (i = 0; | |
2438 | i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t); | |
2439 | i++) { | |
2440 | if ((i % 16) == 0) | |
1da177e4 | 2441 | printk("\n%03x:", i); |
b28a02de | 2442 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2443 | } |
2444 | printk("\n"); | |
2445 | BUG(); | |
2446 | } | |
2447 | } | |
2448 | #else | |
2449 | #define kfree_debugcheck(x) do { } while(0) | |
2450 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2451 | #define check_slabp(x,y) do { } while(0) | |
2452 | #endif | |
2453 | ||
dd0fc66f | 2454 | static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2455 | { |
2456 | int batchcount; | |
2457 | struct kmem_list3 *l3; | |
2458 | struct array_cache *ac; | |
2459 | ||
2460 | check_irq_off(); | |
2461 | ac = ac_data(cachep); | |
b28a02de | 2462 | retry: |
1da177e4 LT |
2463 | batchcount = ac->batchcount; |
2464 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
2465 | /* if there was little recent activity on this | |
2466 | * cache, then perform only a partial refill. | |
2467 | * Otherwise we could generate refill bouncing. | |
2468 | */ | |
2469 | batchcount = BATCHREFILL_LIMIT; | |
2470 | } | |
e498be7d CL |
2471 | l3 = cachep->nodelists[numa_node_id()]; |
2472 | ||
2473 | BUG_ON(ac->avail > 0 || !l3); | |
2474 | spin_lock(&l3->list_lock); | |
1da177e4 | 2475 | |
1da177e4 LT |
2476 | if (l3->shared) { |
2477 | struct array_cache *shared_array = l3->shared; | |
2478 | if (shared_array->avail) { | |
2479 | if (batchcount > shared_array->avail) | |
2480 | batchcount = shared_array->avail; | |
2481 | shared_array->avail -= batchcount; | |
2482 | ac->avail = batchcount; | |
e498be7d | 2483 | memcpy(ac->entry, |
b28a02de PE |
2484 | &(shared_array->entry[shared_array->avail]), |
2485 | sizeof(void *) * batchcount); | |
1da177e4 LT |
2486 | shared_array->touched = 1; |
2487 | goto alloc_done; | |
2488 | } | |
2489 | } | |
2490 | while (batchcount > 0) { | |
2491 | struct list_head *entry; | |
2492 | struct slab *slabp; | |
2493 | /* Get slab alloc is to come from. */ | |
2494 | entry = l3->slabs_partial.next; | |
2495 | if (entry == &l3->slabs_partial) { | |
2496 | l3->free_touched = 1; | |
2497 | entry = l3->slabs_free.next; | |
2498 | if (entry == &l3->slabs_free) | |
2499 | goto must_grow; | |
2500 | } | |
2501 | ||
2502 | slabp = list_entry(entry, struct slab, list); | |
2503 | check_slabp(cachep, slabp); | |
2504 | check_spinlock_acquired(cachep); | |
2505 | while (slabp->inuse < cachep->num && batchcount--) { | |
2506 | kmem_bufctl_t next; | |
2507 | STATS_INC_ALLOCED(cachep); | |
2508 | STATS_INC_ACTIVE(cachep); | |
2509 | STATS_SET_HIGH(cachep); | |
2510 | ||
2511 | /* get obj pointer */ | |
e498be7d | 2512 | ac->entry[ac->avail++] = slabp->s_mem + |
3dafccf2 | 2513 | slabp->free * cachep->buffer_size; |
1da177e4 LT |
2514 | |
2515 | slabp->inuse++; | |
2516 | next = slab_bufctl(slabp)[slabp->free]; | |
2517 | #if DEBUG | |
2518 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
09ad4bbc | 2519 | WARN_ON(numa_node_id() != slabp->nodeid); |
1da177e4 | 2520 | #endif |
b28a02de | 2521 | slabp->free = next; |
1da177e4 LT |
2522 | } |
2523 | check_slabp(cachep, slabp); | |
2524 | ||
2525 | /* move slabp to correct slabp list: */ | |
2526 | list_del(&slabp->list); | |
2527 | if (slabp->free == BUFCTL_END) | |
2528 | list_add(&slabp->list, &l3->slabs_full); | |
2529 | else | |
2530 | list_add(&slabp->list, &l3->slabs_partial); | |
2531 | } | |
2532 | ||
b28a02de | 2533 | must_grow: |
1da177e4 | 2534 | l3->free_objects -= ac->avail; |
b28a02de | 2535 | alloc_done: |
e498be7d | 2536 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2537 | |
2538 | if (unlikely(!ac->avail)) { | |
2539 | int x; | |
e498be7d CL |
2540 | x = cache_grow(cachep, flags, numa_node_id()); |
2541 | ||
1da177e4 LT |
2542 | // cache_grow can reenable interrupts, then ac could change. |
2543 | ac = ac_data(cachep); | |
2544 | if (!x && ac->avail == 0) // no objects in sight? abort | |
2545 | return NULL; | |
2546 | ||
b28a02de | 2547 | if (!ac->avail) // objects refilled by interrupt? |
1da177e4 LT |
2548 | goto retry; |
2549 | } | |
2550 | ac->touched = 1; | |
e498be7d | 2551 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2552 | } |
2553 | ||
2554 | static inline void | |
dd0fc66f | 2555 | cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2556 | { |
2557 | might_sleep_if(flags & __GFP_WAIT); | |
2558 | #if DEBUG | |
2559 | kmem_flagcheck(cachep, flags); | |
2560 | #endif | |
2561 | } | |
2562 | ||
2563 | #if DEBUG | |
b28a02de PE |
2564 | static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags, |
2565 | void *objp, void *caller) | |
1da177e4 | 2566 | { |
b28a02de | 2567 | if (!objp) |
1da177e4 | 2568 | return objp; |
b28a02de | 2569 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2570 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2571 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2572 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2573 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2574 | else |
2575 | check_poison_obj(cachep, objp); | |
2576 | #else | |
2577 | check_poison_obj(cachep, objp); | |
2578 | #endif | |
2579 | poison_obj(cachep, objp, POISON_INUSE); | |
2580 | } | |
2581 | if (cachep->flags & SLAB_STORE_USER) | |
2582 | *dbg_userword(cachep, objp) = caller; | |
2583 | ||
2584 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2585 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE |
2586 | || *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2587 | slab_error(cachep, | |
2588 | "double free, or memory outside" | |
2589 | " object was overwritten"); | |
2590 | printk(KERN_ERR | |
2591 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2592 | objp, *dbg_redzone1(cachep, objp), | |
2593 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2594 | } |
2595 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2596 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2597 | } | |
3dafccf2 | 2598 | objp += obj_offset(cachep); |
1da177e4 | 2599 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2600 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2601 | |
2602 | if (!(flags & __GFP_WAIT)) | |
2603 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2604 | ||
2605 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2606 | } |
1da177e4 LT |
2607 | return objp; |
2608 | } | |
2609 | #else | |
2610 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2611 | #endif | |
2612 | ||
dd0fc66f | 2613 | static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 | 2614 | { |
b28a02de | 2615 | void *objp; |
1da177e4 LT |
2616 | struct array_cache *ac; |
2617 | ||
dc85da15 | 2618 | #ifdef CONFIG_NUMA |
86c562a9 | 2619 | if (unlikely(current->mempolicy && !in_interrupt())) { |
dc85da15 CL |
2620 | int nid = slab_node(current->mempolicy); |
2621 | ||
2622 | if (nid != numa_node_id()) | |
2623 | return __cache_alloc_node(cachep, flags, nid); | |
2624 | } | |
2625 | #endif | |
2626 | ||
5c382300 | 2627 | check_irq_off(); |
1da177e4 LT |
2628 | ac = ac_data(cachep); |
2629 | if (likely(ac->avail)) { | |
2630 | STATS_INC_ALLOCHIT(cachep); | |
2631 | ac->touched = 1; | |
e498be7d | 2632 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2633 | } else { |
2634 | STATS_INC_ALLOCMISS(cachep); | |
2635 | objp = cache_alloc_refill(cachep, flags); | |
2636 | } | |
5c382300 AK |
2637 | return objp; |
2638 | } | |
2639 | ||
dd0fc66f | 2640 | static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
5c382300 AK |
2641 | { |
2642 | unsigned long save_flags; | |
b28a02de | 2643 | void *objp; |
5c382300 AK |
2644 | |
2645 | cache_alloc_debugcheck_before(cachep, flags); | |
2646 | ||
2647 | local_irq_save(save_flags); | |
2648 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2649 | local_irq_restore(save_flags); |
34342e86 | 2650 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
b28a02de | 2651 | __builtin_return_address(0)); |
34342e86 | 2652 | prefetchw(objp); |
1da177e4 LT |
2653 | return objp; |
2654 | } | |
2655 | ||
e498be7d CL |
2656 | #ifdef CONFIG_NUMA |
2657 | /* | |
2658 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2659 | */ |
6daa0e28 | 2660 | static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
e498be7d CL |
2661 | { |
2662 | struct list_head *entry; | |
b28a02de PE |
2663 | struct slab *slabp; |
2664 | struct kmem_list3 *l3; | |
2665 | void *obj; | |
2666 | kmem_bufctl_t next; | |
2667 | int x; | |
2668 | ||
2669 | l3 = cachep->nodelists[nodeid]; | |
2670 | BUG_ON(!l3); | |
2671 | ||
2672 | retry: | |
2673 | spin_lock(&l3->list_lock); | |
2674 | entry = l3->slabs_partial.next; | |
2675 | if (entry == &l3->slabs_partial) { | |
2676 | l3->free_touched = 1; | |
2677 | entry = l3->slabs_free.next; | |
2678 | if (entry == &l3->slabs_free) | |
2679 | goto must_grow; | |
2680 | } | |
2681 | ||
2682 | slabp = list_entry(entry, struct slab, list); | |
2683 | check_spinlock_acquired_node(cachep, nodeid); | |
2684 | check_slabp(cachep, slabp); | |
2685 | ||
2686 | STATS_INC_NODEALLOCS(cachep); | |
2687 | STATS_INC_ACTIVE(cachep); | |
2688 | STATS_SET_HIGH(cachep); | |
2689 | ||
2690 | BUG_ON(slabp->inuse == cachep->num); | |
2691 | ||
2692 | /* get obj pointer */ | |
3dafccf2 | 2693 | obj = slabp->s_mem + slabp->free * cachep->buffer_size; |
b28a02de PE |
2694 | slabp->inuse++; |
2695 | next = slab_bufctl(slabp)[slabp->free]; | |
e498be7d | 2696 | #if DEBUG |
b28a02de | 2697 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; |
e498be7d | 2698 | #endif |
b28a02de PE |
2699 | slabp->free = next; |
2700 | check_slabp(cachep, slabp); | |
2701 | l3->free_objects--; | |
2702 | /* move slabp to correct slabp list: */ | |
2703 | list_del(&slabp->list); | |
2704 | ||
2705 | if (slabp->free == BUFCTL_END) { | |
2706 | list_add(&slabp->list, &l3->slabs_full); | |
2707 | } else { | |
2708 | list_add(&slabp->list, &l3->slabs_partial); | |
2709 | } | |
e498be7d | 2710 | |
b28a02de PE |
2711 | spin_unlock(&l3->list_lock); |
2712 | goto done; | |
e498be7d | 2713 | |
b28a02de PE |
2714 | must_grow: |
2715 | spin_unlock(&l3->list_lock); | |
2716 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 2717 | |
b28a02de PE |
2718 | if (!x) |
2719 | return NULL; | |
e498be7d | 2720 | |
b28a02de PE |
2721 | goto retry; |
2722 | done: | |
2723 | return obj; | |
e498be7d CL |
2724 | } |
2725 | #endif | |
2726 | ||
2727 | /* | |
2728 | * Caller needs to acquire correct kmem_list's list_lock | |
2729 | */ | |
b28a02de PE |
2730 | static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects, |
2731 | int node) | |
1da177e4 LT |
2732 | { |
2733 | int i; | |
e498be7d | 2734 | struct kmem_list3 *l3; |
1da177e4 LT |
2735 | |
2736 | for (i = 0; i < nr_objects; i++) { | |
2737 | void *objp = objpp[i]; | |
2738 | struct slab *slabp; | |
2739 | unsigned int objnr; | |
2740 | ||
065d41cb | 2741 | slabp = page_get_slab(virt_to_page(objp)); |
ff69416e | 2742 | l3 = cachep->nodelists[node]; |
1da177e4 | 2743 | list_del(&slabp->list); |
3dafccf2 | 2744 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
ff69416e | 2745 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 2746 | check_slabp(cachep, slabp); |
e498be7d | 2747 | |
1da177e4 | 2748 | #if DEBUG |
09ad4bbc CL |
2749 | /* Verify that the slab belongs to the intended node */ |
2750 | WARN_ON(slabp->nodeid != node); | |
2751 | ||
1da177e4 | 2752 | if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) { |
e498be7d | 2753 | printk(KERN_ERR "slab: double free detected in cache " |
b28a02de | 2754 | "'%s', objp %p\n", cachep->name, objp); |
1da177e4 LT |
2755 | BUG(); |
2756 | } | |
2757 | #endif | |
2758 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2759 | slabp->free = objnr; | |
2760 | STATS_DEC_ACTIVE(cachep); | |
2761 | slabp->inuse--; | |
e498be7d | 2762 | l3->free_objects++; |
1da177e4 LT |
2763 | check_slabp(cachep, slabp); |
2764 | ||
2765 | /* fixup slab chains */ | |
2766 | if (slabp->inuse == 0) { | |
e498be7d CL |
2767 | if (l3->free_objects > l3->free_limit) { |
2768 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
2769 | slab_destroy(cachep, slabp); |
2770 | } else { | |
e498be7d | 2771 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
2772 | } |
2773 | } else { | |
2774 | /* Unconditionally move a slab to the end of the | |
2775 | * partial list on free - maximum time for the | |
2776 | * other objects to be freed, too. | |
2777 | */ | |
e498be7d | 2778 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
2779 | } |
2780 | } | |
2781 | } | |
2782 | ||
2783 | static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac) | |
2784 | { | |
2785 | int batchcount; | |
e498be7d | 2786 | struct kmem_list3 *l3; |
ff69416e | 2787 | int node = numa_node_id(); |
1da177e4 LT |
2788 | |
2789 | batchcount = ac->batchcount; | |
2790 | #if DEBUG | |
2791 | BUG_ON(!batchcount || batchcount > ac->avail); | |
2792 | #endif | |
2793 | check_irq_off(); | |
ff69416e | 2794 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2795 | spin_lock(&l3->list_lock); |
2796 | if (l3->shared) { | |
2797 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 2798 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
2799 | if (max) { |
2800 | if (batchcount > max) | |
2801 | batchcount = max; | |
e498be7d | 2802 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 2803 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
2804 | shared_array->avail += batchcount; |
2805 | goto free_done; | |
2806 | } | |
2807 | } | |
2808 | ||
ff69416e | 2809 | free_block(cachep, ac->entry, batchcount, node); |
b28a02de | 2810 | free_done: |
1da177e4 LT |
2811 | #if STATS |
2812 | { | |
2813 | int i = 0; | |
2814 | struct list_head *p; | |
2815 | ||
e498be7d CL |
2816 | p = l3->slabs_free.next; |
2817 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
2818 | struct slab *slabp; |
2819 | ||
2820 | slabp = list_entry(p, struct slab, list); | |
2821 | BUG_ON(slabp->inuse); | |
2822 | ||
2823 | i++; | |
2824 | p = p->next; | |
2825 | } | |
2826 | STATS_SET_FREEABLE(cachep, i); | |
2827 | } | |
2828 | #endif | |
e498be7d | 2829 | spin_unlock(&l3->list_lock); |
1da177e4 | 2830 | ac->avail -= batchcount; |
e498be7d | 2831 | memmove(ac->entry, &(ac->entry[batchcount]), |
b28a02de | 2832 | sizeof(void *) * ac->avail); |
1da177e4 LT |
2833 | } |
2834 | ||
2835 | /* | |
2836 | * __cache_free | |
2837 | * Release an obj back to its cache. If the obj has a constructed | |
2838 | * state, it must be in this state _before_ it is released. | |
2839 | * | |
2840 | * Called with disabled ints. | |
2841 | */ | |
2842 | static inline void __cache_free(kmem_cache_t *cachep, void *objp) | |
2843 | { | |
2844 | struct array_cache *ac = ac_data(cachep); | |
2845 | ||
2846 | check_irq_off(); | |
2847 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
2848 | ||
e498be7d CL |
2849 | /* Make sure we are not freeing a object from another |
2850 | * node to the array cache on this cpu. | |
2851 | */ | |
2852 | #ifdef CONFIG_NUMA | |
2853 | { | |
2854 | struct slab *slabp; | |
065d41cb | 2855 | slabp = page_get_slab(virt_to_page(objp)); |
e498be7d CL |
2856 | if (unlikely(slabp->nodeid != numa_node_id())) { |
2857 | struct array_cache *alien = NULL; | |
2858 | int nodeid = slabp->nodeid; | |
b28a02de PE |
2859 | struct kmem_list3 *l3 = |
2860 | cachep->nodelists[numa_node_id()]; | |
e498be7d CL |
2861 | |
2862 | STATS_INC_NODEFREES(cachep); | |
2863 | if (l3->alien && l3->alien[nodeid]) { | |
2864 | alien = l3->alien[nodeid]; | |
2865 | spin_lock(&alien->lock); | |
2866 | if (unlikely(alien->avail == alien->limit)) | |
2867 | __drain_alien_cache(cachep, | |
b28a02de | 2868 | alien, nodeid); |
e498be7d CL |
2869 | alien->entry[alien->avail++] = objp; |
2870 | spin_unlock(&alien->lock); | |
2871 | } else { | |
2872 | spin_lock(&(cachep->nodelists[nodeid])-> | |
b28a02de | 2873 | list_lock); |
ff69416e | 2874 | free_block(cachep, &objp, 1, nodeid); |
e498be7d | 2875 | spin_unlock(&(cachep->nodelists[nodeid])-> |
b28a02de | 2876 | list_lock); |
e498be7d CL |
2877 | } |
2878 | return; | |
2879 | } | |
2880 | } | |
2881 | #endif | |
1da177e4 LT |
2882 | if (likely(ac->avail < ac->limit)) { |
2883 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 2884 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2885 | return; |
2886 | } else { | |
2887 | STATS_INC_FREEMISS(cachep); | |
2888 | cache_flusharray(cachep, ac); | |
e498be7d | 2889 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2890 | } |
2891 | } | |
2892 | ||
2893 | /** | |
2894 | * kmem_cache_alloc - Allocate an object | |
2895 | * @cachep: The cache to allocate from. | |
2896 | * @flags: See kmalloc(). | |
2897 | * | |
2898 | * Allocate an object from this cache. The flags are only relevant | |
2899 | * if the cache has no available objects. | |
2900 | */ | |
dd0fc66f | 2901 | void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2902 | { |
2903 | return __cache_alloc(cachep, flags); | |
2904 | } | |
2905 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2906 | ||
2907 | /** | |
2908 | * kmem_ptr_validate - check if an untrusted pointer might | |
2909 | * be a slab entry. | |
2910 | * @cachep: the cache we're checking against | |
2911 | * @ptr: pointer to validate | |
2912 | * | |
2913 | * This verifies that the untrusted pointer looks sane: | |
2914 | * it is _not_ a guarantee that the pointer is actually | |
2915 | * part of the slab cache in question, but it at least | |
2916 | * validates that the pointer can be dereferenced and | |
2917 | * looks half-way sane. | |
2918 | * | |
2919 | * Currently only used for dentry validation. | |
2920 | */ | |
2921 | int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr) | |
2922 | { | |
b28a02de | 2923 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 2924 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 2925 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 2926 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
2927 | struct page *page; |
2928 | ||
2929 | if (unlikely(addr < min_addr)) | |
2930 | goto out; | |
2931 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
2932 | goto out; | |
2933 | if (unlikely(addr & align_mask)) | |
2934 | goto out; | |
2935 | if (unlikely(!kern_addr_valid(addr))) | |
2936 | goto out; | |
2937 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
2938 | goto out; | |
2939 | page = virt_to_page(ptr); | |
2940 | if (unlikely(!PageSlab(page))) | |
2941 | goto out; | |
065d41cb | 2942 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
2943 | goto out; |
2944 | return 1; | |
b28a02de | 2945 | out: |
1da177e4 LT |
2946 | return 0; |
2947 | } | |
2948 | ||
2949 | #ifdef CONFIG_NUMA | |
2950 | /** | |
2951 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
2952 | * @cachep: The cache to allocate from. | |
2953 | * @flags: See kmalloc(). | |
2954 | * @nodeid: node number of the target node. | |
2955 | * | |
2956 | * Identical to kmem_cache_alloc, except that this function is slow | |
2957 | * and can sleep. And it will allocate memory on the given node, which | |
2958 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
2959 | * New and improved: it will now make sure that the object gets |
2960 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 2961 | */ |
dd0fc66f | 2962 | void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2963 | { |
e498be7d CL |
2964 | unsigned long save_flags; |
2965 | void *ptr; | |
1da177e4 | 2966 | |
e498be7d CL |
2967 | cache_alloc_debugcheck_before(cachep, flags); |
2968 | local_irq_save(save_flags); | |
18f820f6 CL |
2969 | |
2970 | if (nodeid == -1 || nodeid == numa_node_id() || | |
2971 | !cachep->nodelists[nodeid]) | |
5c382300 AK |
2972 | ptr = ____cache_alloc(cachep, flags); |
2973 | else | |
2974 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 2975 | local_irq_restore(save_flags); |
18f820f6 CL |
2976 | |
2977 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
2978 | __builtin_return_address(0)); | |
1da177e4 | 2979 | |
e498be7d | 2980 | return ptr; |
1da177e4 LT |
2981 | } |
2982 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
2983 | ||
dd0fc66f | 2984 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 MS |
2985 | { |
2986 | kmem_cache_t *cachep; | |
2987 | ||
2988 | cachep = kmem_find_general_cachep(size, flags); | |
2989 | if (unlikely(cachep == NULL)) | |
2990 | return NULL; | |
2991 | return kmem_cache_alloc_node(cachep, flags, node); | |
2992 | } | |
2993 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
2994 | #endif |
2995 | ||
2996 | /** | |
2997 | * kmalloc - allocate memory | |
2998 | * @size: how many bytes of memory are required. | |
2999 | * @flags: the type of memory to allocate. | |
3000 | * | |
3001 | * kmalloc is the normal method of allocating memory | |
3002 | * in the kernel. | |
3003 | * | |
3004 | * The @flags argument may be one of: | |
3005 | * | |
3006 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
3007 | * | |
3008 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
3009 | * | |
3010 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
3011 | * | |
3012 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
3013 | * must be suitable for DMA. This can mean different things on different | |
3014 | * platforms. For example, on i386, it means that the memory must come | |
3015 | * from the first 16MB. | |
3016 | */ | |
dd0fc66f | 3017 | void *__kmalloc(size_t size, gfp_t flags) |
1da177e4 LT |
3018 | { |
3019 | kmem_cache_t *cachep; | |
3020 | ||
97e2bde4 MS |
3021 | /* If you want to save a few bytes .text space: replace |
3022 | * __ with kmem_. | |
3023 | * Then kmalloc uses the uninlined functions instead of the inline | |
3024 | * functions. | |
3025 | */ | |
3026 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3027 | if (unlikely(cachep == NULL)) |
3028 | return NULL; | |
1da177e4 LT |
3029 | return __cache_alloc(cachep, flags); |
3030 | } | |
3031 | EXPORT_SYMBOL(__kmalloc); | |
3032 | ||
3033 | #ifdef CONFIG_SMP | |
3034 | /** | |
3035 | * __alloc_percpu - allocate one copy of the object for every present | |
3036 | * cpu in the system, zeroing them. | |
3037 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
3038 | * | |
3039 | * @size: how many bytes of memory are required. | |
1da177e4 | 3040 | */ |
f9f75005 | 3041 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
3042 | { |
3043 | int i; | |
b28a02de | 3044 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3045 | |
3046 | if (!pdata) | |
3047 | return NULL; | |
3048 | ||
e498be7d CL |
3049 | /* |
3050 | * Cannot use for_each_online_cpu since a cpu may come online | |
3051 | * and we have no way of figuring out how to fix the array | |
3052 | * that we have allocated then.... | |
3053 | */ | |
3054 | for_each_cpu(i) { | |
3055 | int node = cpu_to_node(i); | |
3056 | ||
3057 | if (node_online(node)) | |
3058 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3059 | else | |
3060 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3061 | |
3062 | if (!pdata->ptrs[i]) | |
3063 | goto unwind_oom; | |
3064 | memset(pdata->ptrs[i], 0, size); | |
3065 | } | |
3066 | ||
3067 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3068 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3069 | |
b28a02de | 3070 | unwind_oom: |
1da177e4 LT |
3071 | while (--i >= 0) { |
3072 | if (!cpu_possible(i)) | |
3073 | continue; | |
3074 | kfree(pdata->ptrs[i]); | |
3075 | } | |
3076 | kfree(pdata); | |
3077 | return NULL; | |
3078 | } | |
3079 | EXPORT_SYMBOL(__alloc_percpu); | |
3080 | #endif | |
3081 | ||
3082 | /** | |
3083 | * kmem_cache_free - Deallocate an object | |
3084 | * @cachep: The cache the allocation was from. | |
3085 | * @objp: The previously allocated object. | |
3086 | * | |
3087 | * Free an object which was previously allocated from this | |
3088 | * cache. | |
3089 | */ | |
3090 | void kmem_cache_free(kmem_cache_t *cachep, void *objp) | |
3091 | { | |
3092 | unsigned long flags; | |
3093 | ||
3094 | local_irq_save(flags); | |
3095 | __cache_free(cachep, objp); | |
3096 | local_irq_restore(flags); | |
3097 | } | |
3098 | EXPORT_SYMBOL(kmem_cache_free); | |
3099 | ||
1da177e4 LT |
3100 | /** |
3101 | * kfree - free previously allocated memory | |
3102 | * @objp: pointer returned by kmalloc. | |
3103 | * | |
80e93eff PE |
3104 | * If @objp is NULL, no operation is performed. |
3105 | * | |
1da177e4 LT |
3106 | * Don't free memory not originally allocated by kmalloc() |
3107 | * or you will run into trouble. | |
3108 | */ | |
3109 | void kfree(const void *objp) | |
3110 | { | |
3111 | kmem_cache_t *c; | |
3112 | unsigned long flags; | |
3113 | ||
3114 | if (unlikely(!objp)) | |
3115 | return; | |
3116 | local_irq_save(flags); | |
3117 | kfree_debugcheck(objp); | |
065d41cb | 3118 | c = page_get_cache(virt_to_page(objp)); |
3dafccf2 | 3119 | mutex_debug_check_no_locks_freed(objp, obj_size(c)); |
b28a02de | 3120 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3121 | local_irq_restore(flags); |
3122 | } | |
3123 | EXPORT_SYMBOL(kfree); | |
3124 | ||
3125 | #ifdef CONFIG_SMP | |
3126 | /** | |
3127 | * free_percpu - free previously allocated percpu memory | |
3128 | * @objp: pointer returned by alloc_percpu. | |
3129 | * | |
3130 | * Don't free memory not originally allocated by alloc_percpu() | |
3131 | * The complemented objp is to check for that. | |
3132 | */ | |
b28a02de | 3133 | void free_percpu(const void *objp) |
1da177e4 LT |
3134 | { |
3135 | int i; | |
b28a02de | 3136 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3137 | |
e498be7d CL |
3138 | /* |
3139 | * We allocate for all cpus so we cannot use for online cpu here. | |
3140 | */ | |
3141 | for_each_cpu(i) | |
b28a02de | 3142 | kfree(p->ptrs[i]); |
1da177e4 LT |
3143 | kfree(p); |
3144 | } | |
3145 | EXPORT_SYMBOL(free_percpu); | |
3146 | #endif | |
3147 | ||
3148 | unsigned int kmem_cache_size(kmem_cache_t *cachep) | |
3149 | { | |
3dafccf2 | 3150 | return obj_size(cachep); |
1da177e4 LT |
3151 | } |
3152 | EXPORT_SYMBOL(kmem_cache_size); | |
3153 | ||
1944972d ACM |
3154 | const char *kmem_cache_name(kmem_cache_t *cachep) |
3155 | { | |
3156 | return cachep->name; | |
3157 | } | |
3158 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3159 | ||
e498be7d CL |
3160 | /* |
3161 | * This initializes kmem_list3 for all nodes. | |
3162 | */ | |
3163 | static int alloc_kmemlist(kmem_cache_t *cachep) | |
3164 | { | |
3165 | int node; | |
3166 | struct kmem_list3 *l3; | |
3167 | int err = 0; | |
3168 | ||
3169 | for_each_online_node(node) { | |
3170 | struct array_cache *nc = NULL, *new; | |
3171 | struct array_cache **new_alien = NULL; | |
3172 | #ifdef CONFIG_NUMA | |
3173 | if (!(new_alien = alloc_alien_cache(node, cachep->limit))) | |
3174 | goto fail; | |
3175 | #endif | |
b28a02de PE |
3176 | if (!(new = alloc_arraycache(node, (cachep->shared * |
3177 | cachep->batchcount), | |
3178 | 0xbaadf00d))) | |
e498be7d CL |
3179 | goto fail; |
3180 | if ((l3 = cachep->nodelists[node])) { | |
3181 | ||
3182 | spin_lock_irq(&l3->list_lock); | |
3183 | ||
3184 | if ((nc = cachep->nodelists[node]->shared)) | |
b28a02de | 3185 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
3186 | |
3187 | l3->shared = new; | |
3188 | if (!cachep->nodelists[node]->alien) { | |
3189 | l3->alien = new_alien; | |
3190 | new_alien = NULL; | |
3191 | } | |
b28a02de PE |
3192 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3193 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3194 | spin_unlock_irq(&l3->list_lock); |
3195 | kfree(nc); | |
3196 | free_alien_cache(new_alien); | |
3197 | continue; | |
3198 | } | |
3199 | if (!(l3 = kmalloc_node(sizeof(struct kmem_list3), | |
b28a02de | 3200 | GFP_KERNEL, node))) |
e498be7d CL |
3201 | goto fail; |
3202 | ||
3203 | kmem_list3_init(l3); | |
3204 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 3205 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
3206 | l3->shared = new; |
3207 | l3->alien = new_alien; | |
b28a02de PE |
3208 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3209 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3210 | cachep->nodelists[node] = l3; |
3211 | } | |
3212 | return err; | |
b28a02de | 3213 | fail: |
e498be7d CL |
3214 | err = -ENOMEM; |
3215 | return err; | |
3216 | } | |
3217 | ||
1da177e4 LT |
3218 | struct ccupdate_struct { |
3219 | kmem_cache_t *cachep; | |
3220 | struct array_cache *new[NR_CPUS]; | |
3221 | }; | |
3222 | ||
3223 | static void do_ccupdate_local(void *info) | |
3224 | { | |
3225 | struct ccupdate_struct *new = (struct ccupdate_struct *)info; | |
3226 | struct array_cache *old; | |
3227 | ||
3228 | check_irq_off(); | |
3229 | old = ac_data(new->cachep); | |
e498be7d | 3230 | |
1da177e4 LT |
3231 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3232 | new->new[smp_processor_id()] = old; | |
3233 | } | |
3234 | ||
1da177e4 | 3235 | static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount, |
b28a02de | 3236 | int shared) |
1da177e4 LT |
3237 | { |
3238 | struct ccupdate_struct new; | |
e498be7d | 3239 | int i, err; |
1da177e4 | 3240 | |
b28a02de | 3241 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3242 | for_each_online_cpu(i) { |
b28a02de PE |
3243 | new.new[i] = |
3244 | alloc_arraycache(cpu_to_node(i), limit, batchcount); | |
e498be7d | 3245 | if (!new.new[i]) { |
b28a02de PE |
3246 | for (i--; i >= 0; i--) |
3247 | kfree(new.new[i]); | |
e498be7d | 3248 | return -ENOMEM; |
1da177e4 LT |
3249 | } |
3250 | } | |
3251 | new.cachep = cachep; | |
3252 | ||
3253 | smp_call_function_all_cpus(do_ccupdate_local, (void *)&new); | |
e498be7d | 3254 | |
1da177e4 LT |
3255 | check_irq_on(); |
3256 | spin_lock_irq(&cachep->spinlock); | |
3257 | cachep->batchcount = batchcount; | |
3258 | cachep->limit = limit; | |
e498be7d | 3259 | cachep->shared = shared; |
1da177e4 LT |
3260 | spin_unlock_irq(&cachep->spinlock); |
3261 | ||
e498be7d | 3262 | for_each_online_cpu(i) { |
1da177e4 LT |
3263 | struct array_cache *ccold = new.new[i]; |
3264 | if (!ccold) | |
3265 | continue; | |
e498be7d | 3266 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3267 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3268 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3269 | kfree(ccold); |
3270 | } | |
1da177e4 | 3271 | |
e498be7d CL |
3272 | err = alloc_kmemlist(cachep); |
3273 | if (err) { | |
3274 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3275 | cachep->name, -err); |
e498be7d | 3276 | BUG(); |
1da177e4 | 3277 | } |
1da177e4 LT |
3278 | return 0; |
3279 | } | |
3280 | ||
1da177e4 LT |
3281 | static void enable_cpucache(kmem_cache_t *cachep) |
3282 | { | |
3283 | int err; | |
3284 | int limit, shared; | |
3285 | ||
3286 | /* The head array serves three purposes: | |
3287 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
3288 | * - reduce the number of spinlock operations. | |
3289 | * - reduce the number of linked list operations on the slab and | |
3290 | * bufctl chains: array operations are cheaper. | |
3291 | * The numbers are guessed, we should auto-tune as described by | |
3292 | * Bonwick. | |
3293 | */ | |
3dafccf2 | 3294 | if (cachep->buffer_size > 131072) |
1da177e4 | 3295 | limit = 1; |
3dafccf2 | 3296 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3297 | limit = 8; |
3dafccf2 | 3298 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3299 | limit = 24; |
3dafccf2 | 3300 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3301 | limit = 54; |
3302 | else | |
3303 | limit = 120; | |
3304 | ||
3305 | /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound | |
3306 | * allocation behaviour: Most allocs on one cpu, most free operations | |
3307 | * on another cpu. For these cases, an efficient object passing between | |
3308 | * cpus is necessary. This is provided by a shared array. The array | |
3309 | * replaces Bonwick's magazine layer. | |
3310 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3311 | * to a larger limit. Thus disabled by default. | |
3312 | */ | |
3313 | shared = 0; | |
3314 | #ifdef CONFIG_SMP | |
3dafccf2 | 3315 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3316 | shared = 8; |
3317 | #endif | |
3318 | ||
3319 | #if DEBUG | |
3320 | /* With debugging enabled, large batchcount lead to excessively | |
3321 | * long periods with disabled local interrupts. Limit the | |
3322 | * batchcount | |
3323 | */ | |
3324 | if (limit > 32) | |
3325 | limit = 32; | |
3326 | #endif | |
b28a02de | 3327 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3328 | if (err) |
3329 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3330 | cachep->name, -err); |
1da177e4 LT |
3331 | } |
3332 | ||
b28a02de PE |
3333 | static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac, |
3334 | int force, int node) | |
1da177e4 LT |
3335 | { |
3336 | int tofree; | |
3337 | ||
e498be7d | 3338 | check_spinlock_acquired_node(cachep, node); |
1da177e4 LT |
3339 | if (ac->touched && !force) { |
3340 | ac->touched = 0; | |
3341 | } else if (ac->avail) { | |
b28a02de | 3342 | tofree = force ? ac->avail : (ac->limit + 4) / 5; |
1da177e4 | 3343 | if (tofree > ac->avail) { |
b28a02de | 3344 | tofree = (ac->avail + 1) / 2; |
1da177e4 | 3345 | } |
ff69416e | 3346 | free_block(cachep, ac->entry, tofree, node); |
1da177e4 | 3347 | ac->avail -= tofree; |
e498be7d | 3348 | memmove(ac->entry, &(ac->entry[tofree]), |
b28a02de | 3349 | sizeof(void *) * ac->avail); |
1da177e4 LT |
3350 | } |
3351 | } | |
3352 | ||
3353 | /** | |
3354 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3355 | * @unused: unused parameter |
1da177e4 LT |
3356 | * |
3357 | * Called from workqueue/eventd every few seconds. | |
3358 | * Purpose: | |
3359 | * - clear the per-cpu caches for this CPU. | |
3360 | * - return freeable pages to the main free memory pool. | |
3361 | * | |
fc0abb14 | 3362 | * If we cannot acquire the cache chain mutex then just give up - we'll |
1da177e4 LT |
3363 | * try again on the next iteration. |
3364 | */ | |
3365 | static void cache_reap(void *unused) | |
3366 | { | |
3367 | struct list_head *walk; | |
e498be7d | 3368 | struct kmem_list3 *l3; |
1da177e4 | 3369 | |
fc0abb14 | 3370 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3371 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3372 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3373 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3374 | return; |
3375 | } | |
3376 | ||
3377 | list_for_each(walk, &cache_chain) { | |
3378 | kmem_cache_t *searchp; | |
b28a02de | 3379 | struct list_head *p; |
1da177e4 LT |
3380 | int tofree; |
3381 | struct slab *slabp; | |
3382 | ||
3383 | searchp = list_entry(walk, kmem_cache_t, next); | |
3384 | ||
3385 | if (searchp->flags & SLAB_NO_REAP) | |
3386 | goto next; | |
3387 | ||
3388 | check_irq_on(); | |
3389 | ||
e498be7d CL |
3390 | l3 = searchp->nodelists[numa_node_id()]; |
3391 | if (l3->alien) | |
3392 | drain_alien_cache(searchp, l3); | |
3393 | spin_lock_irq(&l3->list_lock); | |
1da177e4 | 3394 | |
e498be7d | 3395 | drain_array_locked(searchp, ac_data(searchp), 0, |
b28a02de | 3396 | numa_node_id()); |
1da177e4 | 3397 | |
e498be7d | 3398 | if (time_after(l3->next_reap, jiffies)) |
1da177e4 LT |
3399 | goto next_unlock; |
3400 | ||
e498be7d | 3401 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3402 | |
e498be7d CL |
3403 | if (l3->shared) |
3404 | drain_array_locked(searchp, l3->shared, 0, | |
b28a02de | 3405 | numa_node_id()); |
1da177e4 | 3406 | |
e498be7d CL |
3407 | if (l3->free_touched) { |
3408 | l3->free_touched = 0; | |
1da177e4 LT |
3409 | goto next_unlock; |
3410 | } | |
3411 | ||
b28a02de PE |
3412 | tofree = |
3413 | (l3->free_limit + 5 * searchp->num - | |
3414 | 1) / (5 * searchp->num); | |
1da177e4 | 3415 | do { |
e498be7d CL |
3416 | p = l3->slabs_free.next; |
3417 | if (p == &(l3->slabs_free)) | |
1da177e4 LT |
3418 | break; |
3419 | ||
3420 | slabp = list_entry(p, struct slab, list); | |
3421 | BUG_ON(slabp->inuse); | |
3422 | list_del(&slabp->list); | |
3423 | STATS_INC_REAPED(searchp); | |
3424 | ||
3425 | /* Safe to drop the lock. The slab is no longer | |
3426 | * linked to the cache. | |
3427 | * searchp cannot disappear, we hold | |
3428 | * cache_chain_lock | |
3429 | */ | |
e498be7d CL |
3430 | l3->free_objects -= searchp->num; |
3431 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3432 | slab_destroy(searchp, slabp); |
e498be7d | 3433 | spin_lock_irq(&l3->list_lock); |
b28a02de PE |
3434 | } while (--tofree > 0); |
3435 | next_unlock: | |
e498be7d | 3436 | spin_unlock_irq(&l3->list_lock); |
b28a02de | 3437 | next: |
1da177e4 LT |
3438 | cond_resched(); |
3439 | } | |
3440 | check_irq_on(); | |
fc0abb14 | 3441 | mutex_unlock(&cache_chain_mutex); |
4ae7c039 | 3442 | drain_remote_pages(); |
1da177e4 | 3443 | /* Setup the next iteration */ |
cd61ef62 | 3444 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3445 | } |
3446 | ||
3447 | #ifdef CONFIG_PROC_FS | |
3448 | ||
85289f98 | 3449 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3450 | { |
85289f98 PE |
3451 | /* |
3452 | * Output format version, so at least we can change it | |
3453 | * without _too_ many complaints. | |
3454 | */ | |
1da177e4 | 3455 | #if STATS |
85289f98 | 3456 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3457 | #else |
85289f98 | 3458 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3459 | #endif |
85289f98 PE |
3460 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3461 | "<objperslab> <pagesperslab>"); | |
3462 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3463 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3464 | #if STATS |
85289f98 PE |
3465 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
3466 | "<error> <maxfreeable> <nodeallocs> <remotefrees>"); | |
3467 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
1da177e4 | 3468 | #endif |
85289f98 PE |
3469 | seq_putc(m, '\n'); |
3470 | } | |
3471 | ||
3472 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3473 | { | |
3474 | loff_t n = *pos; | |
3475 | struct list_head *p; | |
3476 | ||
fc0abb14 | 3477 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3478 | if (!n) |
3479 | print_slabinfo_header(m); | |
1da177e4 LT |
3480 | p = cache_chain.next; |
3481 | while (n--) { | |
3482 | p = p->next; | |
3483 | if (p == &cache_chain) | |
3484 | return NULL; | |
3485 | } | |
3486 | return list_entry(p, kmem_cache_t, next); | |
3487 | } | |
3488 | ||
3489 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3490 | { | |
3491 | kmem_cache_t *cachep = p; | |
3492 | ++*pos; | |
3493 | return cachep->next.next == &cache_chain ? NULL | |
b28a02de | 3494 | : list_entry(cachep->next.next, kmem_cache_t, next); |
1da177e4 LT |
3495 | } |
3496 | ||
3497 | static void s_stop(struct seq_file *m, void *p) | |
3498 | { | |
fc0abb14 | 3499 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3500 | } |
3501 | ||
3502 | static int s_show(struct seq_file *m, void *p) | |
3503 | { | |
3504 | kmem_cache_t *cachep = p; | |
3505 | struct list_head *q; | |
b28a02de PE |
3506 | struct slab *slabp; |
3507 | unsigned long active_objs; | |
3508 | unsigned long num_objs; | |
3509 | unsigned long active_slabs = 0; | |
3510 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3511 | const char *name; |
1da177e4 | 3512 | char *error = NULL; |
e498be7d CL |
3513 | int node; |
3514 | struct kmem_list3 *l3; | |
1da177e4 LT |
3515 | |
3516 | check_irq_on(); | |
3517 | spin_lock_irq(&cachep->spinlock); | |
3518 | active_objs = 0; | |
3519 | num_slabs = 0; | |
e498be7d CL |
3520 | for_each_online_node(node) { |
3521 | l3 = cachep->nodelists[node]; | |
3522 | if (!l3) | |
3523 | continue; | |
3524 | ||
3525 | spin_lock(&l3->list_lock); | |
3526 | ||
b28a02de | 3527 | list_for_each(q, &l3->slabs_full) { |
e498be7d CL |
3528 | slabp = list_entry(q, struct slab, list); |
3529 | if (slabp->inuse != cachep->num && !error) | |
3530 | error = "slabs_full accounting error"; | |
3531 | active_objs += cachep->num; | |
3532 | active_slabs++; | |
3533 | } | |
b28a02de | 3534 | list_for_each(q, &l3->slabs_partial) { |
e498be7d CL |
3535 | slabp = list_entry(q, struct slab, list); |
3536 | if (slabp->inuse == cachep->num && !error) | |
3537 | error = "slabs_partial inuse accounting error"; | |
3538 | if (!slabp->inuse && !error) | |
3539 | error = "slabs_partial/inuse accounting error"; | |
3540 | active_objs += slabp->inuse; | |
3541 | active_slabs++; | |
3542 | } | |
b28a02de | 3543 | list_for_each(q, &l3->slabs_free) { |
e498be7d CL |
3544 | slabp = list_entry(q, struct slab, list); |
3545 | if (slabp->inuse && !error) | |
3546 | error = "slabs_free/inuse accounting error"; | |
3547 | num_slabs++; | |
3548 | } | |
3549 | free_objects += l3->free_objects; | |
3550 | shared_avail += l3->shared->avail; | |
3551 | ||
3552 | spin_unlock(&l3->list_lock); | |
1da177e4 | 3553 | } |
b28a02de PE |
3554 | num_slabs += active_slabs; |
3555 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3556 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3557 | error = "free_objects accounting error"; |
3558 | ||
b28a02de | 3559 | name = cachep->name; |
1da177e4 LT |
3560 | if (error) |
3561 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3562 | ||
3563 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3564 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3565 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3566 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3567 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3568 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3569 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3570 | #if STATS |
b28a02de | 3571 | { /* list3 stats */ |
1da177e4 LT |
3572 | unsigned long high = cachep->high_mark; |
3573 | unsigned long allocs = cachep->num_allocations; | |
3574 | unsigned long grown = cachep->grown; | |
3575 | unsigned long reaped = cachep->reaped; | |
3576 | unsigned long errors = cachep->errors; | |
3577 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3578 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3579 | unsigned long node_frees = cachep->node_frees; |
1da177e4 | 3580 | |
e498be7d | 3581 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
b28a02de | 3582 | %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees); |
1da177e4 LT |
3583 | } |
3584 | /* cpu stats */ | |
3585 | { | |
3586 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3587 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3588 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3589 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3590 | ||
3591 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3592 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3593 | } |
3594 | #endif | |
3595 | seq_putc(m, '\n'); | |
3596 | spin_unlock_irq(&cachep->spinlock); | |
3597 | return 0; | |
3598 | } | |
3599 | ||
3600 | /* | |
3601 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3602 | * | |
3603 | * Output layout: | |
3604 | * cache-name | |
3605 | * num-active-objs | |
3606 | * total-objs | |
3607 | * object size | |
3608 | * num-active-slabs | |
3609 | * total-slabs | |
3610 | * num-pages-per-slab | |
3611 | * + further values on SMP and with statistics enabled | |
3612 | */ | |
3613 | ||
3614 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3615 | .start = s_start, |
3616 | .next = s_next, | |
3617 | .stop = s_stop, | |
3618 | .show = s_show, | |
1da177e4 LT |
3619 | }; |
3620 | ||
3621 | #define MAX_SLABINFO_WRITE 128 | |
3622 | /** | |
3623 | * slabinfo_write - Tuning for the slab allocator | |
3624 | * @file: unused | |
3625 | * @buffer: user buffer | |
3626 | * @count: data length | |
3627 | * @ppos: unused | |
3628 | */ | |
b28a02de PE |
3629 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3630 | size_t count, loff_t *ppos) | |
1da177e4 | 3631 | { |
b28a02de | 3632 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 LT |
3633 | int limit, batchcount, shared, res; |
3634 | struct list_head *p; | |
b28a02de | 3635 | |
1da177e4 LT |
3636 | if (count > MAX_SLABINFO_WRITE) |
3637 | return -EINVAL; | |
3638 | if (copy_from_user(&kbuf, buffer, count)) | |
3639 | return -EFAULT; | |
b28a02de | 3640 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3641 | |
3642 | tmp = strchr(kbuf, ' '); | |
3643 | if (!tmp) | |
3644 | return -EINVAL; | |
3645 | *tmp = '\0'; | |
3646 | tmp++; | |
3647 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3648 | return -EINVAL; | |
3649 | ||
3650 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3651 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3652 | res = -EINVAL; |
b28a02de | 3653 | list_for_each(p, &cache_chain) { |
1da177e4 LT |
3654 | kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next); |
3655 | ||
3656 | if (!strcmp(cachep->name, kbuf)) { | |
3657 | if (limit < 1 || | |
3658 | batchcount < 1 || | |
b28a02de | 3659 | batchcount > limit || shared < 0) { |
e498be7d | 3660 | res = 0; |
1da177e4 | 3661 | } else { |
e498be7d | 3662 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3663 | batchcount, shared); |
1da177e4 LT |
3664 | } |
3665 | break; | |
3666 | } | |
3667 | } | |
fc0abb14 | 3668 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3669 | if (res >= 0) |
3670 | res = count; | |
3671 | return res; | |
3672 | } | |
3673 | #endif | |
3674 | ||
00e145b6 MS |
3675 | /** |
3676 | * ksize - get the actual amount of memory allocated for a given object | |
3677 | * @objp: Pointer to the object | |
3678 | * | |
3679 | * kmalloc may internally round up allocations and return more memory | |
3680 | * than requested. ksize() can be used to determine the actual amount of | |
3681 | * memory allocated. The caller may use this additional memory, even though | |
3682 | * a smaller amount of memory was initially specified with the kmalloc call. | |
3683 | * The caller must guarantee that objp points to a valid object previously | |
3684 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
3685 | * must not be freed during the duration of the call. | |
3686 | */ | |
1da177e4 LT |
3687 | unsigned int ksize(const void *objp) |
3688 | { | |
00e145b6 MS |
3689 | if (unlikely(objp == NULL)) |
3690 | return 0; | |
1da177e4 | 3691 | |
3dafccf2 | 3692 | return obj_size(page_get_cache(virt_to_page(objp))); |
1da177e4 | 3693 | } |