]>
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 | |
71 | * The global cache-chain is protected by the semaphore 'cache_chain_sem'. | |
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 | * | |
78 | */ | |
79 | ||
80 | #include <linux/config.h> | |
81 | #include <linux/slab.h> | |
82 | #include <linux/mm.h> | |
83 | #include <linux/swap.h> | |
84 | #include <linux/cache.h> | |
85 | #include <linux/interrupt.h> | |
86 | #include <linux/init.h> | |
87 | #include <linux/compiler.h> | |
88 | #include <linux/seq_file.h> | |
89 | #include <linux/notifier.h> | |
90 | #include <linux/kallsyms.h> | |
91 | #include <linux/cpu.h> | |
92 | #include <linux/sysctl.h> | |
93 | #include <linux/module.h> | |
94 | #include <linux/rcupdate.h> | |
543537bd | 95 | #include <linux/string.h> |
1da177e4 LT |
96 | |
97 | #include <asm/uaccess.h> | |
98 | #include <asm/cacheflush.h> | |
99 | #include <asm/tlbflush.h> | |
100 | #include <asm/page.h> | |
101 | ||
102 | /* | |
103 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
104 | * SLAB_RED_ZONE & SLAB_POISON. | |
105 | * 0 for faster, smaller code (especially in the critical paths). | |
106 | * | |
107 | * STATS - 1 to collect stats for /proc/slabinfo. | |
108 | * 0 for faster, smaller code (especially in the critical paths). | |
109 | * | |
110 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
111 | */ | |
112 | ||
113 | #ifdef CONFIG_DEBUG_SLAB | |
114 | #define DEBUG 1 | |
115 | #define STATS 1 | |
116 | #define FORCED_DEBUG 1 | |
117 | #else | |
118 | #define DEBUG 0 | |
119 | #define STATS 0 | |
120 | #define FORCED_DEBUG 0 | |
121 | #endif | |
122 | ||
123 | ||
124 | /* Shouldn't this be in a header file somewhere? */ | |
125 | #define BYTES_PER_WORD sizeof(void *) | |
126 | ||
127 | #ifndef cache_line_size | |
128 | #define cache_line_size() L1_CACHE_BYTES | |
129 | #endif | |
130 | ||
131 | #ifndef ARCH_KMALLOC_MINALIGN | |
132 | /* | |
133 | * Enforce a minimum alignment for the kmalloc caches. | |
134 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
135 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
136 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
137 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
138 | * Note that this flag disables some debug features. | |
139 | */ | |
140 | #define ARCH_KMALLOC_MINALIGN 0 | |
141 | #endif | |
142 | ||
143 | #ifndef ARCH_SLAB_MINALIGN | |
144 | /* | |
145 | * Enforce a minimum alignment for all caches. | |
146 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
147 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
148 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
149 | * some debug features. | |
150 | */ | |
151 | #define ARCH_SLAB_MINALIGN 0 | |
152 | #endif | |
153 | ||
154 | #ifndef ARCH_KMALLOC_FLAGS | |
155 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
156 | #endif | |
157 | ||
158 | /* Legal flag mask for kmem_cache_create(). */ | |
159 | #if DEBUG | |
160 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
161 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
162 | SLAB_NO_REAP | SLAB_CACHE_DMA | \ | |
163 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ | |
164 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
165 | SLAB_DESTROY_BY_RCU) | |
166 | #else | |
167 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \ | |
168 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ | |
169 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
170 | SLAB_DESTROY_BY_RCU) | |
171 | #endif | |
172 | ||
173 | /* | |
174 | * kmem_bufctl_t: | |
175 | * | |
176 | * Bufctl's are used for linking objs within a slab | |
177 | * linked offsets. | |
178 | * | |
179 | * This implementation relies on "struct page" for locating the cache & | |
180 | * slab an object belongs to. | |
181 | * This allows the bufctl structure to be small (one int), but limits | |
182 | * the number of objects a slab (not a cache) can contain when off-slab | |
183 | * bufctls are used. The limit is the size of the largest general cache | |
184 | * that does not use off-slab slabs. | |
185 | * For 32bit archs with 4 kB pages, is this 56. | |
186 | * This is not serious, as it is only for large objects, when it is unwise | |
187 | * to have too many per slab. | |
188 | * Note: This limit can be raised by introducing a general cache whose size | |
189 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
190 | */ | |
191 | ||
192 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) | |
193 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
194 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2) | |
195 | ||
196 | /* Max number of objs-per-slab for caches which use off-slab slabs. | |
197 | * Needed to avoid a possible looping condition in cache_grow(). | |
198 | */ | |
199 | static unsigned long offslab_limit; | |
200 | ||
201 | /* | |
202 | * struct slab | |
203 | * | |
204 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
205 | * for a slab, or allocated from an general cache. | |
206 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
207 | */ | |
208 | struct slab { | |
209 | struct list_head list; | |
210 | unsigned long colouroff; | |
211 | void *s_mem; /* including colour offset */ | |
212 | unsigned int inuse; /* num of objs active in slab */ | |
213 | kmem_bufctl_t free; | |
214 | }; | |
215 | ||
216 | /* | |
217 | * struct slab_rcu | |
218 | * | |
219 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
220 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
221 | * we need to approach a kernel structure obliquely, from its address | |
222 | * obtained without the usual locking. We can lock the structure to | |
223 | * stabilize it and check it's still at the given address, only if we | |
224 | * can be sure that the memory has not been meanwhile reused for some | |
225 | * other kind of object (which our subsystem's lock might corrupt). | |
226 | * | |
227 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
228 | * taking the spinlock within the structure expected at that address. | |
229 | * | |
230 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
231 | */ | |
232 | struct slab_rcu { | |
233 | struct rcu_head head; | |
234 | kmem_cache_t *cachep; | |
235 | void *addr; | |
236 | }; | |
237 | ||
238 | /* | |
239 | * struct array_cache | |
240 | * | |
241 | * Per cpu structures | |
242 | * Purpose: | |
243 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
244 | * - reduce the number of linked list operations | |
245 | * - reduce spinlock operations | |
246 | * | |
247 | * The limit is stored in the per-cpu structure to reduce the data cache | |
248 | * footprint. | |
249 | * | |
250 | */ | |
251 | struct array_cache { | |
252 | unsigned int avail; | |
253 | unsigned int limit; | |
254 | unsigned int batchcount; | |
255 | unsigned int touched; | |
256 | }; | |
257 | ||
258 | /* bootstrap: The caches do not work without cpuarrays anymore, | |
259 | * but the cpuarrays are allocated from the generic caches... | |
260 | */ | |
261 | #define BOOT_CPUCACHE_ENTRIES 1 | |
262 | struct arraycache_init { | |
263 | struct array_cache cache; | |
264 | void * entries[BOOT_CPUCACHE_ENTRIES]; | |
265 | }; | |
266 | ||
267 | /* | |
268 | * The slab lists of all objects. | |
269 | * Hopefully reduce the internal fragmentation | |
270 | * NUMA: The spinlock could be moved from the kmem_cache_t | |
271 | * into this structure, too. Figure out what causes | |
272 | * fewer cross-node spinlock operations. | |
273 | */ | |
274 | struct kmem_list3 { | |
275 | struct list_head slabs_partial; /* partial list first, better asm code */ | |
276 | struct list_head slabs_full; | |
277 | struct list_head slabs_free; | |
278 | unsigned long free_objects; | |
279 | int free_touched; | |
280 | unsigned long next_reap; | |
281 | struct array_cache *shared; | |
282 | }; | |
283 | ||
284 | #define LIST3_INIT(parent) \ | |
285 | { \ | |
286 | .slabs_full = LIST_HEAD_INIT(parent.slabs_full), \ | |
287 | .slabs_partial = LIST_HEAD_INIT(parent.slabs_partial), \ | |
288 | .slabs_free = LIST_HEAD_INIT(parent.slabs_free) \ | |
289 | } | |
290 | #define list3_data(cachep) \ | |
291 | (&(cachep)->lists) | |
292 | ||
293 | /* NUMA: per-node */ | |
294 | #define list3_data_ptr(cachep, ptr) \ | |
295 | list3_data(cachep) | |
296 | ||
297 | /* | |
298 | * kmem_cache_t | |
299 | * | |
300 | * manages a cache. | |
301 | */ | |
302 | ||
303 | struct kmem_cache_s { | |
304 | /* 1) per-cpu data, touched during every alloc/free */ | |
305 | struct array_cache *array[NR_CPUS]; | |
306 | unsigned int batchcount; | |
307 | unsigned int limit; | |
308 | /* 2) touched by every alloc & free from the backend */ | |
309 | struct kmem_list3 lists; | |
310 | /* NUMA: kmem_3list_t *nodelists[MAX_NUMNODES] */ | |
311 | unsigned int objsize; | |
312 | unsigned int flags; /* constant flags */ | |
313 | unsigned int num; /* # of objs per slab */ | |
314 | unsigned int free_limit; /* upper limit of objects in the lists */ | |
315 | spinlock_t spinlock; | |
316 | ||
317 | /* 3) cache_grow/shrink */ | |
318 | /* order of pgs per slab (2^n) */ | |
319 | unsigned int gfporder; | |
320 | ||
321 | /* force GFP flags, e.g. GFP_DMA */ | |
322 | unsigned int gfpflags; | |
323 | ||
324 | size_t colour; /* cache colouring range */ | |
325 | unsigned int colour_off; /* colour offset */ | |
326 | unsigned int colour_next; /* cache colouring */ | |
327 | kmem_cache_t *slabp_cache; | |
328 | unsigned int slab_size; | |
329 | unsigned int dflags; /* dynamic flags */ | |
330 | ||
331 | /* constructor func */ | |
332 | void (*ctor)(void *, kmem_cache_t *, unsigned long); | |
333 | ||
334 | /* de-constructor func */ | |
335 | void (*dtor)(void *, kmem_cache_t *, unsigned long); | |
336 | ||
337 | /* 4) cache creation/removal */ | |
338 | const char *name; | |
339 | struct list_head next; | |
340 | ||
341 | /* 5) statistics */ | |
342 | #if STATS | |
343 | unsigned long num_active; | |
344 | unsigned long num_allocations; | |
345 | unsigned long high_mark; | |
346 | unsigned long grown; | |
347 | unsigned long reaped; | |
348 | unsigned long errors; | |
349 | unsigned long max_freeable; | |
350 | unsigned long node_allocs; | |
351 | atomic_t allochit; | |
352 | atomic_t allocmiss; | |
353 | atomic_t freehit; | |
354 | atomic_t freemiss; | |
355 | #endif | |
356 | #if DEBUG | |
357 | int dbghead; | |
358 | int reallen; | |
359 | #endif | |
360 | }; | |
361 | ||
362 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
363 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
364 | ||
365 | #define BATCHREFILL_LIMIT 16 | |
366 | /* Optimization question: fewer reaps means less | |
367 | * probability for unnessary cpucache drain/refill cycles. | |
368 | * | |
369 | * OTHO the cpuarrays can contain lots of objects, | |
370 | * which could lock up otherwise freeable slabs. | |
371 | */ | |
372 | #define REAPTIMEOUT_CPUC (2*HZ) | |
373 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
374 | ||
375 | #if STATS | |
376 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
377 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
378 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
379 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
380 | #define STATS_INC_REAPED(x) ((x)->reaped++) | |
381 | #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \ | |
382 | (x)->high_mark = (x)->num_active; \ | |
383 | } while (0) | |
384 | #define STATS_INC_ERR(x) ((x)->errors++) | |
385 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
386 | #define STATS_SET_FREEABLE(x, i) \ | |
387 | do { if ((x)->max_freeable < i) \ | |
388 | (x)->max_freeable = i; \ | |
389 | } while (0) | |
390 | ||
391 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) | |
392 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
393 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
394 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
395 | #else | |
396 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
397 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
398 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
399 | #define STATS_INC_GROWN(x) do { } while (0) | |
400 | #define STATS_INC_REAPED(x) do { } while (0) | |
401 | #define STATS_SET_HIGH(x) do { } while (0) | |
402 | #define STATS_INC_ERR(x) do { } while (0) | |
403 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
404 | #define STATS_SET_FREEABLE(x, i) \ | |
405 | do { } while (0) | |
406 | ||
407 | #define STATS_INC_ALLOCHIT(x) do { } while (0) | |
408 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
409 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
410 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
411 | #endif | |
412 | ||
413 | #if DEBUG | |
414 | /* Magic nums for obj red zoning. | |
415 | * Placed in the first word before and the first word after an obj. | |
416 | */ | |
417 | #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ | |
418 | #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ | |
419 | ||
420 | /* ...and for poisoning */ | |
421 | #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ | |
422 | #define POISON_FREE 0x6b /* for use-after-free poisoning */ | |
423 | #define POISON_END 0xa5 /* end-byte of poisoning */ | |
424 | ||
425 | /* memory layout of objects: | |
426 | * 0 : objp | |
427 | * 0 .. cachep->dbghead - BYTES_PER_WORD - 1: padding. This ensures that | |
428 | * the end of an object is aligned with the end of the real | |
429 | * allocation. Catches writes behind the end of the allocation. | |
430 | * cachep->dbghead - BYTES_PER_WORD .. cachep->dbghead - 1: | |
431 | * redzone word. | |
432 | * cachep->dbghead: The real object. | |
433 | * cachep->objsize - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
434 | * cachep->objsize - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long] | |
435 | */ | |
436 | static int obj_dbghead(kmem_cache_t *cachep) | |
437 | { | |
438 | return cachep->dbghead; | |
439 | } | |
440 | ||
441 | static int obj_reallen(kmem_cache_t *cachep) | |
442 | { | |
443 | return cachep->reallen; | |
444 | } | |
445 | ||
446 | static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp) | |
447 | { | |
448 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
449 | return (unsigned long*) (objp+obj_dbghead(cachep)-BYTES_PER_WORD); | |
450 | } | |
451 | ||
452 | static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp) | |
453 | { | |
454 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
455 | if (cachep->flags & SLAB_STORE_USER) | |
456 | return (unsigned long*) (objp+cachep->objsize-2*BYTES_PER_WORD); | |
457 | return (unsigned long*) (objp+cachep->objsize-BYTES_PER_WORD); | |
458 | } | |
459 | ||
460 | static void **dbg_userword(kmem_cache_t *cachep, void *objp) | |
461 | { | |
462 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
463 | return (void**)(objp+cachep->objsize-BYTES_PER_WORD); | |
464 | } | |
465 | ||
466 | #else | |
467 | ||
468 | #define obj_dbghead(x) 0 | |
469 | #define obj_reallen(cachep) (cachep->objsize) | |
470 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
471 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
472 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
473 | ||
474 | #endif | |
475 | ||
476 | /* | |
477 | * Maximum size of an obj (in 2^order pages) | |
478 | * and absolute limit for the gfp order. | |
479 | */ | |
480 | #if defined(CONFIG_LARGE_ALLOCS) | |
481 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
482 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
483 | #elif defined(CONFIG_MMU) | |
484 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
485 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
486 | #else | |
487 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
488 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
489 | #endif | |
490 | ||
491 | /* | |
492 | * Do not go above this order unless 0 objects fit into the slab. | |
493 | */ | |
494 | #define BREAK_GFP_ORDER_HI 1 | |
495 | #define BREAK_GFP_ORDER_LO 0 | |
496 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
497 | ||
498 | /* Macros for storing/retrieving the cachep and or slab from the | |
499 | * global 'mem_map'. These are used to find the slab an obj belongs to. | |
500 | * With kfree(), these are used to find the cache which an obj belongs to. | |
501 | */ | |
502 | #define SET_PAGE_CACHE(pg,x) ((pg)->lru.next = (struct list_head *)(x)) | |
503 | #define GET_PAGE_CACHE(pg) ((kmem_cache_t *)(pg)->lru.next) | |
504 | #define SET_PAGE_SLAB(pg,x) ((pg)->lru.prev = (struct list_head *)(x)) | |
505 | #define GET_PAGE_SLAB(pg) ((struct slab *)(pg)->lru.prev) | |
506 | ||
507 | /* These are the default caches for kmalloc. Custom caches can have other sizes. */ | |
508 | struct cache_sizes malloc_sizes[] = { | |
509 | #define CACHE(x) { .cs_size = (x) }, | |
510 | #include <linux/kmalloc_sizes.h> | |
511 | CACHE(ULONG_MAX) | |
512 | #undef CACHE | |
513 | }; | |
514 | EXPORT_SYMBOL(malloc_sizes); | |
515 | ||
516 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
517 | struct cache_names { | |
518 | char *name; | |
519 | char *name_dma; | |
520 | }; | |
521 | ||
522 | static struct cache_names __initdata cache_names[] = { | |
523 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
524 | #include <linux/kmalloc_sizes.h> | |
525 | { NULL, } | |
526 | #undef CACHE | |
527 | }; | |
528 | ||
529 | static struct arraycache_init initarray_cache __initdata = | |
530 | { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; | |
531 | static struct arraycache_init initarray_generic = | |
532 | { { 0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; | |
533 | ||
534 | /* internal cache of cache description objs */ | |
535 | static kmem_cache_t cache_cache = { | |
536 | .lists = LIST3_INIT(cache_cache.lists), | |
537 | .batchcount = 1, | |
538 | .limit = BOOT_CPUCACHE_ENTRIES, | |
539 | .objsize = sizeof(kmem_cache_t), | |
540 | .flags = SLAB_NO_REAP, | |
541 | .spinlock = SPIN_LOCK_UNLOCKED, | |
542 | .name = "kmem_cache", | |
543 | #if DEBUG | |
544 | .reallen = sizeof(kmem_cache_t), | |
545 | #endif | |
546 | }; | |
547 | ||
548 | /* Guard access to the cache-chain. */ | |
549 | static struct semaphore cache_chain_sem; | |
550 | static struct list_head cache_chain; | |
551 | ||
552 | /* | |
553 | * vm_enough_memory() looks at this to determine how many | |
554 | * slab-allocated pages are possibly freeable under pressure | |
555 | * | |
556 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
557 | */ | |
558 | atomic_t slab_reclaim_pages; | |
559 | EXPORT_SYMBOL(slab_reclaim_pages); | |
560 | ||
561 | /* | |
562 | * chicken and egg problem: delay the per-cpu array allocation | |
563 | * until the general caches are up. | |
564 | */ | |
565 | static enum { | |
566 | NONE, | |
567 | PARTIAL, | |
568 | FULL | |
569 | } g_cpucache_up; | |
570 | ||
571 | static DEFINE_PER_CPU(struct work_struct, reap_work); | |
572 | ||
573 | static void free_block(kmem_cache_t* cachep, void** objpp, int len); | |
574 | static void enable_cpucache (kmem_cache_t *cachep); | |
575 | static void cache_reap (void *unused); | |
576 | ||
577 | static inline void **ac_entry(struct array_cache *ac) | |
578 | { | |
579 | return (void**)(ac+1); | |
580 | } | |
581 | ||
582 | static inline struct array_cache *ac_data(kmem_cache_t *cachep) | |
583 | { | |
584 | return cachep->array[smp_processor_id()]; | |
585 | } | |
586 | ||
97e2bde4 | 587 | static inline kmem_cache_t *__find_general_cachep(size_t size, int gfpflags) |
1da177e4 LT |
588 | { |
589 | struct cache_sizes *csizep = malloc_sizes; | |
590 | ||
591 | #if DEBUG | |
592 | /* This happens if someone tries to call | |
593 | * kmem_cache_create(), or __kmalloc(), before | |
594 | * the generic caches are initialized. | |
595 | */ | |
596 | BUG_ON(csizep->cs_cachep == NULL); | |
597 | #endif | |
598 | while (size > csizep->cs_size) | |
599 | csizep++; | |
600 | ||
601 | /* | |
602 | * Really subtile: The last entry with cs->cs_size==ULONG_MAX | |
603 | * has cs_{dma,}cachep==NULL. Thus no special case | |
604 | * for large kmalloc calls required. | |
605 | */ | |
606 | if (unlikely(gfpflags & GFP_DMA)) | |
607 | return csizep->cs_dmacachep; | |
608 | return csizep->cs_cachep; | |
609 | } | |
610 | ||
97e2bde4 MS |
611 | kmem_cache_t *kmem_find_general_cachep(size_t size, int gfpflags) |
612 | { | |
613 | return __find_general_cachep(size, gfpflags); | |
614 | } | |
615 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
616 | ||
1da177e4 LT |
617 | /* Cal the num objs, wastage, and bytes left over for a given slab size. */ |
618 | static void cache_estimate(unsigned long gfporder, size_t size, size_t align, | |
619 | int flags, size_t *left_over, unsigned int *num) | |
620 | { | |
621 | int i; | |
622 | size_t wastage = PAGE_SIZE<<gfporder; | |
623 | size_t extra = 0; | |
624 | size_t base = 0; | |
625 | ||
626 | if (!(flags & CFLGS_OFF_SLAB)) { | |
627 | base = sizeof(struct slab); | |
628 | extra = sizeof(kmem_bufctl_t); | |
629 | } | |
630 | i = 0; | |
631 | while (i*size + ALIGN(base+i*extra, align) <= wastage) | |
632 | i++; | |
633 | if (i > 0) | |
634 | i--; | |
635 | ||
636 | if (i > SLAB_LIMIT) | |
637 | i = SLAB_LIMIT; | |
638 | ||
639 | *num = i; | |
640 | wastage -= i*size; | |
641 | wastage -= ALIGN(base+i*extra, align); | |
642 | *left_over = wastage; | |
643 | } | |
644 | ||
645 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
646 | ||
647 | static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg) | |
648 | { | |
649 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
650 | function, cachep->name, msg); | |
651 | dump_stack(); | |
652 | } | |
653 | ||
654 | /* | |
655 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
656 | * via the workqueue/eventd. | |
657 | * Add the CPU number into the expiration time to minimize the possibility of | |
658 | * the CPUs getting into lockstep and contending for the global cache chain | |
659 | * lock. | |
660 | */ | |
661 | static void __devinit start_cpu_timer(int cpu) | |
662 | { | |
663 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
664 | ||
665 | /* | |
666 | * When this gets called from do_initcalls via cpucache_init(), | |
667 | * init_workqueues() has already run, so keventd will be setup | |
668 | * at that time. | |
669 | */ | |
670 | if (keventd_up() && reap_work->func == NULL) { | |
671 | INIT_WORK(reap_work, cache_reap, NULL); | |
672 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
673 | } | |
674 | } | |
675 | ||
676 | static struct array_cache *alloc_arraycache(int cpu, int entries, | |
677 | int batchcount) | |
678 | { | |
679 | int memsize = sizeof(void*)*entries+sizeof(struct array_cache); | |
680 | struct array_cache *nc = NULL; | |
681 | ||
97e2bde4 | 682 | if (cpu == -1) |
1da177e4 | 683 | nc = kmalloc(memsize, GFP_KERNEL); |
97e2bde4 MS |
684 | else |
685 | nc = kmalloc_node(memsize, GFP_KERNEL, cpu_to_node(cpu)); | |
686 | ||
1da177e4 LT |
687 | if (nc) { |
688 | nc->avail = 0; | |
689 | nc->limit = entries; | |
690 | nc->batchcount = batchcount; | |
691 | nc->touched = 0; | |
692 | } | |
693 | return nc; | |
694 | } | |
695 | ||
696 | static int __devinit cpuup_callback(struct notifier_block *nfb, | |
697 | unsigned long action, void *hcpu) | |
698 | { | |
699 | long cpu = (long)hcpu; | |
700 | kmem_cache_t* cachep; | |
701 | ||
702 | switch (action) { | |
703 | case CPU_UP_PREPARE: | |
704 | down(&cache_chain_sem); | |
705 | list_for_each_entry(cachep, &cache_chain, next) { | |
706 | struct array_cache *nc; | |
707 | ||
708 | nc = alloc_arraycache(cpu, cachep->limit, cachep->batchcount); | |
709 | if (!nc) | |
710 | goto bad; | |
711 | ||
712 | spin_lock_irq(&cachep->spinlock); | |
713 | cachep->array[cpu] = nc; | |
714 | cachep->free_limit = (1+num_online_cpus())*cachep->batchcount | |
715 | + cachep->num; | |
716 | spin_unlock_irq(&cachep->spinlock); | |
717 | ||
718 | } | |
719 | up(&cache_chain_sem); | |
720 | break; | |
721 | case CPU_ONLINE: | |
722 | start_cpu_timer(cpu); | |
723 | break; | |
724 | #ifdef CONFIG_HOTPLUG_CPU | |
725 | case CPU_DEAD: | |
726 | /* fall thru */ | |
727 | case CPU_UP_CANCELED: | |
728 | down(&cache_chain_sem); | |
729 | ||
730 | list_for_each_entry(cachep, &cache_chain, next) { | |
731 | struct array_cache *nc; | |
732 | ||
733 | spin_lock_irq(&cachep->spinlock); | |
734 | /* cpu is dead; no one can alloc from it. */ | |
735 | nc = cachep->array[cpu]; | |
736 | cachep->array[cpu] = NULL; | |
737 | cachep->free_limit -= cachep->batchcount; | |
738 | free_block(cachep, ac_entry(nc), nc->avail); | |
739 | spin_unlock_irq(&cachep->spinlock); | |
740 | kfree(nc); | |
741 | } | |
742 | up(&cache_chain_sem); | |
743 | break; | |
744 | #endif | |
745 | } | |
746 | return NOTIFY_OK; | |
747 | bad: | |
748 | up(&cache_chain_sem); | |
749 | return NOTIFY_BAD; | |
750 | } | |
751 | ||
752 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
753 | ||
754 | /* Initialisation. | |
755 | * Called after the gfp() functions have been enabled, and before smp_init(). | |
756 | */ | |
757 | void __init kmem_cache_init(void) | |
758 | { | |
759 | size_t left_over; | |
760 | struct cache_sizes *sizes; | |
761 | struct cache_names *names; | |
762 | ||
763 | /* | |
764 | * Fragmentation resistance on low memory - only use bigger | |
765 | * page orders on machines with more than 32MB of memory. | |
766 | */ | |
767 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
768 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
769 | ||
770 | ||
771 | /* Bootstrap is tricky, because several objects are allocated | |
772 | * from caches that do not exist yet: | |
773 | * 1) initialize the cache_cache cache: it contains the kmem_cache_t | |
774 | * structures of all caches, except cache_cache itself: cache_cache | |
775 | * is statically allocated. | |
776 | * Initially an __init data area is used for the head array, it's | |
777 | * replaced with a kmalloc allocated array at the end of the bootstrap. | |
778 | * 2) Create the first kmalloc cache. | |
779 | * The kmem_cache_t for the new cache is allocated normally. An __init | |
780 | * data area is used for the head array. | |
781 | * 3) Create the remaining kmalloc caches, with minimally sized head arrays. | |
782 | * 4) Replace the __init data head arrays for cache_cache and the first | |
783 | * kmalloc cache with kmalloc allocated arrays. | |
784 | * 5) Resize the head arrays of the kmalloc caches to their final sizes. | |
785 | */ | |
786 | ||
787 | /* 1) create the cache_cache */ | |
788 | init_MUTEX(&cache_chain_sem); | |
789 | INIT_LIST_HEAD(&cache_chain); | |
790 | list_add(&cache_cache.next, &cache_chain); | |
791 | cache_cache.colour_off = cache_line_size(); | |
792 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
793 | ||
794 | cache_cache.objsize = ALIGN(cache_cache.objsize, cache_line_size()); | |
795 | ||
796 | cache_estimate(0, cache_cache.objsize, cache_line_size(), 0, | |
797 | &left_over, &cache_cache.num); | |
798 | if (!cache_cache.num) | |
799 | BUG(); | |
800 | ||
801 | cache_cache.colour = left_over/cache_cache.colour_off; | |
802 | cache_cache.colour_next = 0; | |
803 | cache_cache.slab_size = ALIGN(cache_cache.num*sizeof(kmem_bufctl_t) + | |
804 | sizeof(struct slab), cache_line_size()); | |
805 | ||
806 | /* 2+3) create the kmalloc caches */ | |
807 | sizes = malloc_sizes; | |
808 | names = cache_names; | |
809 | ||
810 | while (sizes->cs_size != ULONG_MAX) { | |
811 | /* For performance, all the general caches are L1 aligned. | |
812 | * This should be particularly beneficial on SMP boxes, as it | |
813 | * eliminates "false sharing". | |
814 | * Note for systems short on memory removing the alignment will | |
815 | * allow tighter packing of the smaller caches. */ | |
816 | sizes->cs_cachep = kmem_cache_create(names->name, | |
817 | sizes->cs_size, ARCH_KMALLOC_MINALIGN, | |
818 | (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, NULL); | |
819 | ||
820 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
821 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
822 | offslab_limit = sizes->cs_size-sizeof(struct slab); | |
823 | offslab_limit /= sizeof(kmem_bufctl_t); | |
824 | } | |
825 | ||
826 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
827 | sizes->cs_size, ARCH_KMALLOC_MINALIGN, | |
828 | (ARCH_KMALLOC_FLAGS | SLAB_CACHE_DMA | SLAB_PANIC), | |
829 | NULL, NULL); | |
830 | ||
831 | sizes++; | |
832 | names++; | |
833 | } | |
834 | /* 4) Replace the bootstrap head arrays */ | |
835 | { | |
836 | void * ptr; | |
837 | ||
838 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
839 | local_irq_disable(); | |
840 | BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache); | |
841 | memcpy(ptr, ac_data(&cache_cache), sizeof(struct arraycache_init)); | |
842 | cache_cache.array[smp_processor_id()] = ptr; | |
843 | local_irq_enable(); | |
844 | ||
845 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); | |
846 | local_irq_disable(); | |
847 | BUG_ON(ac_data(malloc_sizes[0].cs_cachep) != &initarray_generic.cache); | |
848 | memcpy(ptr, ac_data(malloc_sizes[0].cs_cachep), | |
849 | sizeof(struct arraycache_init)); | |
850 | malloc_sizes[0].cs_cachep->array[smp_processor_id()] = ptr; | |
851 | local_irq_enable(); | |
852 | } | |
853 | ||
854 | /* 5) resize the head arrays to their final sizes */ | |
855 | { | |
856 | kmem_cache_t *cachep; | |
857 | down(&cache_chain_sem); | |
858 | list_for_each_entry(cachep, &cache_chain, next) | |
859 | enable_cpucache(cachep); | |
860 | up(&cache_chain_sem); | |
861 | } | |
862 | ||
863 | /* Done! */ | |
864 | g_cpucache_up = FULL; | |
865 | ||
866 | /* Register a cpu startup notifier callback | |
867 | * that initializes ac_data for all new cpus | |
868 | */ | |
869 | register_cpu_notifier(&cpucache_notifier); | |
870 | ||
871 | ||
872 | /* The reap timers are started later, with a module init call: | |
873 | * That part of the kernel is not yet operational. | |
874 | */ | |
875 | } | |
876 | ||
877 | static int __init cpucache_init(void) | |
878 | { | |
879 | int cpu; | |
880 | ||
881 | /* | |
882 | * Register the timers that return unneeded | |
883 | * pages to gfp. | |
884 | */ | |
885 | for (cpu = 0; cpu < NR_CPUS; cpu++) { | |
886 | if (cpu_online(cpu)) | |
887 | start_cpu_timer(cpu); | |
888 | } | |
889 | ||
890 | return 0; | |
891 | } | |
892 | ||
893 | __initcall(cpucache_init); | |
894 | ||
895 | /* | |
896 | * Interface to system's page allocator. No need to hold the cache-lock. | |
897 | * | |
898 | * If we requested dmaable memory, we will get it. Even if we | |
899 | * did not request dmaable memory, we might get it, but that | |
900 | * would be relatively rare and ignorable. | |
901 | */ | |
902 | static void *kmem_getpages(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid) | |
903 | { | |
904 | struct page *page; | |
905 | void *addr; | |
906 | int i; | |
907 | ||
908 | flags |= cachep->gfpflags; | |
909 | if (likely(nodeid == -1)) { | |
910 | page = alloc_pages(flags, cachep->gfporder); | |
911 | } else { | |
912 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); | |
913 | } | |
914 | if (!page) | |
915 | return NULL; | |
916 | addr = page_address(page); | |
917 | ||
918 | i = (1 << cachep->gfporder); | |
919 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
920 | atomic_add(i, &slab_reclaim_pages); | |
921 | add_page_state(nr_slab, i); | |
922 | while (i--) { | |
923 | SetPageSlab(page); | |
924 | page++; | |
925 | } | |
926 | return addr; | |
927 | } | |
928 | ||
929 | /* | |
930 | * Interface to system's page release. | |
931 | */ | |
932 | static void kmem_freepages(kmem_cache_t *cachep, void *addr) | |
933 | { | |
934 | unsigned long i = (1<<cachep->gfporder); | |
935 | struct page *page = virt_to_page(addr); | |
936 | const unsigned long nr_freed = i; | |
937 | ||
938 | while (i--) { | |
939 | if (!TestClearPageSlab(page)) | |
940 | BUG(); | |
941 | page++; | |
942 | } | |
943 | sub_page_state(nr_slab, nr_freed); | |
944 | if (current->reclaim_state) | |
945 | current->reclaim_state->reclaimed_slab += nr_freed; | |
946 | free_pages((unsigned long)addr, cachep->gfporder); | |
947 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
948 | atomic_sub(1<<cachep->gfporder, &slab_reclaim_pages); | |
949 | } | |
950 | ||
951 | static void kmem_rcu_free(struct rcu_head *head) | |
952 | { | |
953 | struct slab_rcu *slab_rcu = (struct slab_rcu *) head; | |
954 | kmem_cache_t *cachep = slab_rcu->cachep; | |
955 | ||
956 | kmem_freepages(cachep, slab_rcu->addr); | |
957 | if (OFF_SLAB(cachep)) | |
958 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
959 | } | |
960 | ||
961 | #if DEBUG | |
962 | ||
963 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
964 | static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, | |
965 | unsigned long caller) | |
966 | { | |
967 | int size = obj_reallen(cachep); | |
968 | ||
969 | addr = (unsigned long *)&((char*)addr)[obj_dbghead(cachep)]; | |
970 | ||
971 | if (size < 5*sizeof(unsigned long)) | |
972 | return; | |
973 | ||
974 | *addr++=0x12345678; | |
975 | *addr++=caller; | |
976 | *addr++=smp_processor_id(); | |
977 | size -= 3*sizeof(unsigned long); | |
978 | { | |
979 | unsigned long *sptr = &caller; | |
980 | unsigned long svalue; | |
981 | ||
982 | while (!kstack_end(sptr)) { | |
983 | svalue = *sptr++; | |
984 | if (kernel_text_address(svalue)) { | |
985 | *addr++=svalue; | |
986 | size -= sizeof(unsigned long); | |
987 | if (size <= sizeof(unsigned long)) | |
988 | break; | |
989 | } | |
990 | } | |
991 | ||
992 | } | |
993 | *addr++=0x87654321; | |
994 | } | |
995 | #endif | |
996 | ||
997 | static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val) | |
998 | { | |
999 | int size = obj_reallen(cachep); | |
1000 | addr = &((char*)addr)[obj_dbghead(cachep)]; | |
1001 | ||
1002 | memset(addr, val, size); | |
1003 | *(unsigned char *)(addr+size-1) = POISON_END; | |
1004 | } | |
1005 | ||
1006 | static void dump_line(char *data, int offset, int limit) | |
1007 | { | |
1008 | int i; | |
1009 | printk(KERN_ERR "%03x:", offset); | |
1010 | for (i=0;i<limit;i++) { | |
1011 | printk(" %02x", (unsigned char)data[offset+i]); | |
1012 | } | |
1013 | printk("\n"); | |
1014 | } | |
1015 | #endif | |
1016 | ||
1017 | #if DEBUG | |
1018 | ||
1019 | static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines) | |
1020 | { | |
1021 | int i, size; | |
1022 | char *realobj; | |
1023 | ||
1024 | if (cachep->flags & SLAB_RED_ZONE) { | |
1025 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
1026 | *dbg_redzone1(cachep, objp), | |
1027 | *dbg_redzone2(cachep, objp)); | |
1028 | } | |
1029 | ||
1030 | if (cachep->flags & SLAB_STORE_USER) { | |
1031 | printk(KERN_ERR "Last user: [<%p>]", | |
1032 | *dbg_userword(cachep, objp)); | |
1033 | print_symbol("(%s)", | |
1034 | (unsigned long)*dbg_userword(cachep, objp)); | |
1035 | printk("\n"); | |
1036 | } | |
1037 | realobj = (char*)objp+obj_dbghead(cachep); | |
1038 | size = obj_reallen(cachep); | |
1039 | for (i=0; i<size && lines;i+=16, lines--) { | |
1040 | int limit; | |
1041 | limit = 16; | |
1042 | if (i+limit > size) | |
1043 | limit = size-i; | |
1044 | dump_line(realobj, i, limit); | |
1045 | } | |
1046 | } | |
1047 | ||
1048 | static void check_poison_obj(kmem_cache_t *cachep, void *objp) | |
1049 | { | |
1050 | char *realobj; | |
1051 | int size, i; | |
1052 | int lines = 0; | |
1053 | ||
1054 | realobj = (char*)objp+obj_dbghead(cachep); | |
1055 | size = obj_reallen(cachep); | |
1056 | ||
1057 | for (i=0;i<size;i++) { | |
1058 | char exp = POISON_FREE; | |
1059 | if (i == size-1) | |
1060 | exp = POISON_END; | |
1061 | if (realobj[i] != exp) { | |
1062 | int limit; | |
1063 | /* Mismatch ! */ | |
1064 | /* Print header */ | |
1065 | if (lines == 0) { | |
1066 | printk(KERN_ERR "Slab corruption: start=%p, len=%d\n", | |
1067 | realobj, size); | |
1068 | print_objinfo(cachep, objp, 0); | |
1069 | } | |
1070 | /* Hexdump the affected line */ | |
1071 | i = (i/16)*16; | |
1072 | limit = 16; | |
1073 | if (i+limit > size) | |
1074 | limit = size-i; | |
1075 | dump_line(realobj, i, limit); | |
1076 | i += 16; | |
1077 | lines++; | |
1078 | /* Limit to 5 lines */ | |
1079 | if (lines > 5) | |
1080 | break; | |
1081 | } | |
1082 | } | |
1083 | if (lines != 0) { | |
1084 | /* Print some data about the neighboring objects, if they | |
1085 | * exist: | |
1086 | */ | |
1087 | struct slab *slabp = GET_PAGE_SLAB(virt_to_page(objp)); | |
1088 | int objnr; | |
1089 | ||
1090 | objnr = (objp-slabp->s_mem)/cachep->objsize; | |
1091 | if (objnr) { | |
1092 | objp = slabp->s_mem+(objnr-1)*cachep->objsize; | |
1093 | realobj = (char*)objp+obj_dbghead(cachep); | |
1094 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", | |
1095 | realobj, size); | |
1096 | print_objinfo(cachep, objp, 2); | |
1097 | } | |
1098 | if (objnr+1 < cachep->num) { | |
1099 | objp = slabp->s_mem+(objnr+1)*cachep->objsize; | |
1100 | realobj = (char*)objp+obj_dbghead(cachep); | |
1101 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", | |
1102 | realobj, size); | |
1103 | print_objinfo(cachep, objp, 2); | |
1104 | } | |
1105 | } | |
1106 | } | |
1107 | #endif | |
1108 | ||
1109 | /* Destroy all the objs in a slab, and release the mem back to the system. | |
1110 | * Before calling the slab must have been unlinked from the cache. | |
1111 | * The cache-lock is not held/needed. | |
1112 | */ | |
1113 | static void slab_destroy (kmem_cache_t *cachep, struct slab *slabp) | |
1114 | { | |
1115 | void *addr = slabp->s_mem - slabp->colouroff; | |
1116 | ||
1117 | #if DEBUG | |
1118 | int i; | |
1119 | for (i = 0; i < cachep->num; i++) { | |
1120 | void *objp = slabp->s_mem + cachep->objsize * i; | |
1121 | ||
1122 | if (cachep->flags & SLAB_POISON) { | |
1123 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
1124 | if ((cachep->objsize%PAGE_SIZE)==0 && OFF_SLAB(cachep)) | |
1125 | kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE,1); | |
1126 | else | |
1127 | check_poison_obj(cachep, objp); | |
1128 | #else | |
1129 | check_poison_obj(cachep, objp); | |
1130 | #endif | |
1131 | } | |
1132 | if (cachep->flags & SLAB_RED_ZONE) { | |
1133 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1134 | slab_error(cachep, "start of a freed object " | |
1135 | "was overwritten"); | |
1136 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
1137 | slab_error(cachep, "end of a freed object " | |
1138 | "was overwritten"); | |
1139 | } | |
1140 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
1141 | (cachep->dtor)(objp+obj_dbghead(cachep), cachep, 0); | |
1142 | } | |
1143 | #else | |
1144 | if (cachep->dtor) { | |
1145 | int i; | |
1146 | for (i = 0; i < cachep->num; i++) { | |
1147 | void* objp = slabp->s_mem+cachep->objsize*i; | |
1148 | (cachep->dtor)(objp, cachep, 0); | |
1149 | } | |
1150 | } | |
1151 | #endif | |
1152 | ||
1153 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { | |
1154 | struct slab_rcu *slab_rcu; | |
1155 | ||
1156 | slab_rcu = (struct slab_rcu *) slabp; | |
1157 | slab_rcu->cachep = cachep; | |
1158 | slab_rcu->addr = addr; | |
1159 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1160 | } else { | |
1161 | kmem_freepages(cachep, addr); | |
1162 | if (OFF_SLAB(cachep)) | |
1163 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1164 | } | |
1165 | } | |
1166 | ||
1167 | /** | |
1168 | * kmem_cache_create - Create a cache. | |
1169 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1170 | * @size: The size of objects to be created in this cache. | |
1171 | * @align: The required alignment for the objects. | |
1172 | * @flags: SLAB flags | |
1173 | * @ctor: A constructor for the objects. | |
1174 | * @dtor: A destructor for the objects. | |
1175 | * | |
1176 | * Returns a ptr to the cache on success, NULL on failure. | |
1177 | * Cannot be called within a int, but can be interrupted. | |
1178 | * The @ctor is run when new pages are allocated by the cache | |
1179 | * and the @dtor is run before the pages are handed back. | |
1180 | * | |
1181 | * @name must be valid until the cache is destroyed. This implies that | |
1182 | * the module calling this has to destroy the cache before getting | |
1183 | * unloaded. | |
1184 | * | |
1185 | * The flags are | |
1186 | * | |
1187 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1188 | * to catch references to uninitialised memory. | |
1189 | * | |
1190 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1191 | * for buffer overruns. | |
1192 | * | |
1193 | * %SLAB_NO_REAP - Don't automatically reap this cache when we're under | |
1194 | * memory pressure. | |
1195 | * | |
1196 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
1197 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1198 | * as davem. | |
1199 | */ | |
1200 | kmem_cache_t * | |
1201 | kmem_cache_create (const char *name, size_t size, size_t align, | |
1202 | unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long), | |
1203 | void (*dtor)(void*, kmem_cache_t *, unsigned long)) | |
1204 | { | |
1205 | size_t left_over, slab_size, ralign; | |
1206 | kmem_cache_t *cachep = NULL; | |
1207 | ||
1208 | /* | |
1209 | * Sanity checks... these are all serious usage bugs. | |
1210 | */ | |
1211 | if ((!name) || | |
1212 | in_interrupt() || | |
1213 | (size < BYTES_PER_WORD) || | |
1214 | (size > (1<<MAX_OBJ_ORDER)*PAGE_SIZE) || | |
1215 | (dtor && !ctor)) { | |
1216 | printk(KERN_ERR "%s: Early error in slab %s\n", | |
1217 | __FUNCTION__, name); | |
1218 | BUG(); | |
1219 | } | |
1220 | ||
1221 | #if DEBUG | |
1222 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1223 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1224 | /* No constructor, but inital state check requested */ | |
1225 | printk(KERN_ERR "%s: No con, but init state check " | |
1226 | "requested - %s\n", __FUNCTION__, name); | |
1227 | flags &= ~SLAB_DEBUG_INITIAL; | |
1228 | } | |
1229 | ||
1230 | #if FORCED_DEBUG | |
1231 | /* | |
1232 | * Enable redzoning and last user accounting, except for caches with | |
1233 | * large objects, if the increased size would increase the object size | |
1234 | * above the next power of two: caches with object sizes just above a | |
1235 | * power of two have a significant amount of internal fragmentation. | |
1236 | */ | |
1237 | if ((size < 4096 || fls(size-1) == fls(size-1+3*BYTES_PER_WORD))) | |
1238 | flags |= SLAB_RED_ZONE|SLAB_STORE_USER; | |
1239 | if (!(flags & SLAB_DESTROY_BY_RCU)) | |
1240 | flags |= SLAB_POISON; | |
1241 | #endif | |
1242 | if (flags & SLAB_DESTROY_BY_RCU) | |
1243 | BUG_ON(flags & SLAB_POISON); | |
1244 | #endif | |
1245 | if (flags & SLAB_DESTROY_BY_RCU) | |
1246 | BUG_ON(dtor); | |
1247 | ||
1248 | /* | |
1249 | * Always checks flags, a caller might be expecting debug | |
1250 | * support which isn't available. | |
1251 | */ | |
1252 | if (flags & ~CREATE_MASK) | |
1253 | BUG(); | |
1254 | ||
1255 | /* Check that size is in terms of words. This is needed to avoid | |
1256 | * unaligned accesses for some archs when redzoning is used, and makes | |
1257 | * sure any on-slab bufctl's are also correctly aligned. | |
1258 | */ | |
1259 | if (size & (BYTES_PER_WORD-1)) { | |
1260 | size += (BYTES_PER_WORD-1); | |
1261 | size &= ~(BYTES_PER_WORD-1); | |
1262 | } | |
1263 | ||
1264 | /* calculate out the final buffer alignment: */ | |
1265 | /* 1) arch recommendation: can be overridden for debug */ | |
1266 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1267 | /* Default alignment: as specified by the arch code. | |
1268 | * Except if an object is really small, then squeeze multiple | |
1269 | * objects into one cacheline. | |
1270 | */ | |
1271 | ralign = cache_line_size(); | |
1272 | while (size <= ralign/2) | |
1273 | ralign /= 2; | |
1274 | } else { | |
1275 | ralign = BYTES_PER_WORD; | |
1276 | } | |
1277 | /* 2) arch mandated alignment: disables debug if necessary */ | |
1278 | if (ralign < ARCH_SLAB_MINALIGN) { | |
1279 | ralign = ARCH_SLAB_MINALIGN; | |
1280 | if (ralign > BYTES_PER_WORD) | |
1281 | flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER); | |
1282 | } | |
1283 | /* 3) caller mandated alignment: disables debug if necessary */ | |
1284 | if (ralign < align) { | |
1285 | ralign = align; | |
1286 | if (ralign > BYTES_PER_WORD) | |
1287 | flags &= ~(SLAB_RED_ZONE|SLAB_STORE_USER); | |
1288 | } | |
1289 | /* 4) Store it. Note that the debug code below can reduce | |
1290 | * the alignment to BYTES_PER_WORD. | |
1291 | */ | |
1292 | align = ralign; | |
1293 | ||
1294 | /* Get cache's description obj. */ | |
1295 | cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL); | |
1296 | if (!cachep) | |
1297 | goto opps; | |
1298 | memset(cachep, 0, sizeof(kmem_cache_t)); | |
1299 | ||
1300 | #if DEBUG | |
1301 | cachep->reallen = size; | |
1302 | ||
1303 | if (flags & SLAB_RED_ZONE) { | |
1304 | /* redzoning only works with word aligned caches */ | |
1305 | align = BYTES_PER_WORD; | |
1306 | ||
1307 | /* add space for red zone words */ | |
1308 | cachep->dbghead += BYTES_PER_WORD; | |
1309 | size += 2*BYTES_PER_WORD; | |
1310 | } | |
1311 | if (flags & SLAB_STORE_USER) { | |
1312 | /* user store requires word alignment and | |
1313 | * one word storage behind the end of the real | |
1314 | * object. | |
1315 | */ | |
1316 | align = BYTES_PER_WORD; | |
1317 | size += BYTES_PER_WORD; | |
1318 | } | |
1319 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
1320 | if (size > 128 && cachep->reallen > cache_line_size() && size < PAGE_SIZE) { | |
1321 | cachep->dbghead += PAGE_SIZE - size; | |
1322 | size = PAGE_SIZE; | |
1323 | } | |
1324 | #endif | |
1325 | #endif | |
1326 | ||
1327 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
1328 | if (size >= (PAGE_SIZE>>3)) | |
1329 | /* | |
1330 | * Size is large, assume best to place the slab management obj | |
1331 | * off-slab (should allow better packing of objs). | |
1332 | */ | |
1333 | flags |= CFLGS_OFF_SLAB; | |
1334 | ||
1335 | size = ALIGN(size, align); | |
1336 | ||
1337 | if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) { | |
1338 | /* | |
1339 | * A VFS-reclaimable slab tends to have most allocations | |
1340 | * as GFP_NOFS and we really don't want to have to be allocating | |
1341 | * higher-order pages when we are unable to shrink dcache. | |
1342 | */ | |
1343 | cachep->gfporder = 0; | |
1344 | cache_estimate(cachep->gfporder, size, align, flags, | |
1345 | &left_over, &cachep->num); | |
1346 | } else { | |
1347 | /* | |
1348 | * Calculate size (in pages) of slabs, and the num of objs per | |
1349 | * slab. This could be made much more intelligent. For now, | |
1350 | * try to avoid using high page-orders for slabs. When the | |
1351 | * gfp() funcs are more friendly towards high-order requests, | |
1352 | * this should be changed. | |
1353 | */ | |
1354 | do { | |
1355 | unsigned int break_flag = 0; | |
1356 | cal_wastage: | |
1357 | cache_estimate(cachep->gfporder, size, align, flags, | |
1358 | &left_over, &cachep->num); | |
1359 | if (break_flag) | |
1360 | break; | |
1361 | if (cachep->gfporder >= MAX_GFP_ORDER) | |
1362 | break; | |
1363 | if (!cachep->num) | |
1364 | goto next; | |
1365 | if (flags & CFLGS_OFF_SLAB && | |
1366 | cachep->num > offslab_limit) { | |
1367 | /* This num of objs will cause problems. */ | |
1368 | cachep->gfporder--; | |
1369 | break_flag++; | |
1370 | goto cal_wastage; | |
1371 | } | |
1372 | ||
1373 | /* | |
1374 | * Large num of objs is good, but v. large slabs are | |
1375 | * currently bad for the gfp()s. | |
1376 | */ | |
1377 | if (cachep->gfporder >= slab_break_gfp_order) | |
1378 | break; | |
1379 | ||
1380 | if ((left_over*8) <= (PAGE_SIZE<<cachep->gfporder)) | |
1381 | break; /* Acceptable internal fragmentation. */ | |
1382 | next: | |
1383 | cachep->gfporder++; | |
1384 | } while (1); | |
1385 | } | |
1386 | ||
1387 | if (!cachep->num) { | |
1388 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
1389 | kmem_cache_free(&cache_cache, cachep); | |
1390 | cachep = NULL; | |
1391 | goto opps; | |
1392 | } | |
1393 | slab_size = ALIGN(cachep->num*sizeof(kmem_bufctl_t) | |
1394 | + sizeof(struct slab), align); | |
1395 | ||
1396 | /* | |
1397 | * If the slab has been placed off-slab, and we have enough space then | |
1398 | * move it on-slab. This is at the expense of any extra colouring. | |
1399 | */ | |
1400 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
1401 | flags &= ~CFLGS_OFF_SLAB; | |
1402 | left_over -= slab_size; | |
1403 | } | |
1404 | ||
1405 | if (flags & CFLGS_OFF_SLAB) { | |
1406 | /* really off slab. No need for manual alignment */ | |
1407 | slab_size = cachep->num*sizeof(kmem_bufctl_t)+sizeof(struct slab); | |
1408 | } | |
1409 | ||
1410 | cachep->colour_off = cache_line_size(); | |
1411 | /* Offset must be a multiple of the alignment. */ | |
1412 | if (cachep->colour_off < align) | |
1413 | cachep->colour_off = align; | |
1414 | cachep->colour = left_over/cachep->colour_off; | |
1415 | cachep->slab_size = slab_size; | |
1416 | cachep->flags = flags; | |
1417 | cachep->gfpflags = 0; | |
1418 | if (flags & SLAB_CACHE_DMA) | |
1419 | cachep->gfpflags |= GFP_DMA; | |
1420 | spin_lock_init(&cachep->spinlock); | |
1421 | cachep->objsize = size; | |
1422 | /* NUMA */ | |
1423 | INIT_LIST_HEAD(&cachep->lists.slabs_full); | |
1424 | INIT_LIST_HEAD(&cachep->lists.slabs_partial); | |
1425 | INIT_LIST_HEAD(&cachep->lists.slabs_free); | |
1426 | ||
1427 | if (flags & CFLGS_OFF_SLAB) | |
1428 | cachep->slabp_cache = kmem_find_general_cachep(slab_size,0); | |
1429 | cachep->ctor = ctor; | |
1430 | cachep->dtor = dtor; | |
1431 | cachep->name = name; | |
1432 | ||
1433 | /* Don't let CPUs to come and go */ | |
1434 | lock_cpu_hotplug(); | |
1435 | ||
1436 | if (g_cpucache_up == FULL) { | |
1437 | enable_cpucache(cachep); | |
1438 | } else { | |
1439 | if (g_cpucache_up == NONE) { | |
1440 | /* Note: the first kmem_cache_create must create | |
1441 | * the cache that's used by kmalloc(24), otherwise | |
1442 | * the creation of further caches will BUG(). | |
1443 | */ | |
1444 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
1445 | g_cpucache_up = PARTIAL; | |
1446 | } else { | |
1447 | cachep->array[smp_processor_id()] = kmalloc(sizeof(struct arraycache_init),GFP_KERNEL); | |
1448 | } | |
1449 | BUG_ON(!ac_data(cachep)); | |
1450 | ac_data(cachep)->avail = 0; | |
1451 | ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1452 | ac_data(cachep)->batchcount = 1; | |
1453 | ac_data(cachep)->touched = 0; | |
1454 | cachep->batchcount = 1; | |
1455 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
1456 | cachep->free_limit = (1+num_online_cpus())*cachep->batchcount | |
1457 | + cachep->num; | |
1458 | } | |
1459 | ||
1460 | cachep->lists.next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1461 | ((unsigned long)cachep)%REAPTIMEOUT_LIST3; | |
1462 | ||
1463 | /* Need the semaphore to access the chain. */ | |
1464 | down(&cache_chain_sem); | |
1465 | { | |
1466 | struct list_head *p; | |
1467 | mm_segment_t old_fs; | |
1468 | ||
1469 | old_fs = get_fs(); | |
1470 | set_fs(KERNEL_DS); | |
1471 | list_for_each(p, &cache_chain) { | |
1472 | kmem_cache_t *pc = list_entry(p, kmem_cache_t, next); | |
1473 | char tmp; | |
1474 | /* This happens when the module gets unloaded and doesn't | |
1475 | destroy its slab cache and noone else reuses the vmalloc | |
1476 | area of the module. Print a warning. */ | |
1477 | if (__get_user(tmp,pc->name)) { | |
1478 | printk("SLAB: cache with size %d has lost its name\n", | |
1479 | pc->objsize); | |
1480 | continue; | |
1481 | } | |
1482 | if (!strcmp(pc->name,name)) { | |
1483 | printk("kmem_cache_create: duplicate cache %s\n",name); | |
1484 | up(&cache_chain_sem); | |
1485 | unlock_cpu_hotplug(); | |
1486 | BUG(); | |
1487 | } | |
1488 | } | |
1489 | set_fs(old_fs); | |
1490 | } | |
1491 | ||
1492 | /* cache setup completed, link it into the list */ | |
1493 | list_add(&cachep->next, &cache_chain); | |
1494 | up(&cache_chain_sem); | |
1495 | unlock_cpu_hotplug(); | |
1496 | opps: | |
1497 | if (!cachep && (flags & SLAB_PANIC)) | |
1498 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
1499 | name); | |
1500 | return cachep; | |
1501 | } | |
1502 | EXPORT_SYMBOL(kmem_cache_create); | |
1503 | ||
1504 | #if DEBUG | |
1505 | static void check_irq_off(void) | |
1506 | { | |
1507 | BUG_ON(!irqs_disabled()); | |
1508 | } | |
1509 | ||
1510 | static void check_irq_on(void) | |
1511 | { | |
1512 | BUG_ON(irqs_disabled()); | |
1513 | } | |
1514 | ||
1515 | static void check_spinlock_acquired(kmem_cache_t *cachep) | |
1516 | { | |
1517 | #ifdef CONFIG_SMP | |
1518 | check_irq_off(); | |
1519 | BUG_ON(spin_trylock(&cachep->spinlock)); | |
1520 | #endif | |
1521 | } | |
1522 | #else | |
1523 | #define check_irq_off() do { } while(0) | |
1524 | #define check_irq_on() do { } while(0) | |
1525 | #define check_spinlock_acquired(x) do { } while(0) | |
1526 | #endif | |
1527 | ||
1528 | /* | |
1529 | * Waits for all CPUs to execute func(). | |
1530 | */ | |
1531 | static void smp_call_function_all_cpus(void (*func) (void *arg), void *arg) | |
1532 | { | |
1533 | check_irq_on(); | |
1534 | preempt_disable(); | |
1535 | ||
1536 | local_irq_disable(); | |
1537 | func(arg); | |
1538 | local_irq_enable(); | |
1539 | ||
1540 | if (smp_call_function(func, arg, 1, 1)) | |
1541 | BUG(); | |
1542 | ||
1543 | preempt_enable(); | |
1544 | } | |
1545 | ||
1546 | static void drain_array_locked(kmem_cache_t* cachep, | |
1547 | struct array_cache *ac, int force); | |
1548 | ||
1549 | static void do_drain(void *arg) | |
1550 | { | |
1551 | kmem_cache_t *cachep = (kmem_cache_t*)arg; | |
1552 | struct array_cache *ac; | |
1553 | ||
1554 | check_irq_off(); | |
1555 | ac = ac_data(cachep); | |
1556 | spin_lock(&cachep->spinlock); | |
1557 | free_block(cachep, &ac_entry(ac)[0], ac->avail); | |
1558 | spin_unlock(&cachep->spinlock); | |
1559 | ac->avail = 0; | |
1560 | } | |
1561 | ||
1562 | static void drain_cpu_caches(kmem_cache_t *cachep) | |
1563 | { | |
1564 | smp_call_function_all_cpus(do_drain, cachep); | |
1565 | check_irq_on(); | |
1566 | spin_lock_irq(&cachep->spinlock); | |
1567 | if (cachep->lists.shared) | |
1568 | drain_array_locked(cachep, cachep->lists.shared, 1); | |
1569 | spin_unlock_irq(&cachep->spinlock); | |
1570 | } | |
1571 | ||
1572 | ||
1573 | /* NUMA shrink all list3s */ | |
1574 | static int __cache_shrink(kmem_cache_t *cachep) | |
1575 | { | |
1576 | struct slab *slabp; | |
1577 | int ret; | |
1578 | ||
1579 | drain_cpu_caches(cachep); | |
1580 | ||
1581 | check_irq_on(); | |
1582 | spin_lock_irq(&cachep->spinlock); | |
1583 | ||
1584 | for(;;) { | |
1585 | struct list_head *p; | |
1586 | ||
1587 | p = cachep->lists.slabs_free.prev; | |
1588 | if (p == &cachep->lists.slabs_free) | |
1589 | break; | |
1590 | ||
1591 | slabp = list_entry(cachep->lists.slabs_free.prev, struct slab, list); | |
1592 | #if DEBUG | |
1593 | if (slabp->inuse) | |
1594 | BUG(); | |
1595 | #endif | |
1596 | list_del(&slabp->list); | |
1597 | ||
1598 | cachep->lists.free_objects -= cachep->num; | |
1599 | spin_unlock_irq(&cachep->spinlock); | |
1600 | slab_destroy(cachep, slabp); | |
1601 | spin_lock_irq(&cachep->spinlock); | |
1602 | } | |
1603 | ret = !list_empty(&cachep->lists.slabs_full) || | |
1604 | !list_empty(&cachep->lists.slabs_partial); | |
1605 | spin_unlock_irq(&cachep->spinlock); | |
1606 | return ret; | |
1607 | } | |
1608 | ||
1609 | /** | |
1610 | * kmem_cache_shrink - Shrink a cache. | |
1611 | * @cachep: The cache to shrink. | |
1612 | * | |
1613 | * Releases as many slabs as possible for a cache. | |
1614 | * To help debugging, a zero exit status indicates all slabs were released. | |
1615 | */ | |
1616 | int kmem_cache_shrink(kmem_cache_t *cachep) | |
1617 | { | |
1618 | if (!cachep || in_interrupt()) | |
1619 | BUG(); | |
1620 | ||
1621 | return __cache_shrink(cachep); | |
1622 | } | |
1623 | EXPORT_SYMBOL(kmem_cache_shrink); | |
1624 | ||
1625 | /** | |
1626 | * kmem_cache_destroy - delete a cache | |
1627 | * @cachep: the cache to destroy | |
1628 | * | |
1629 | * Remove a kmem_cache_t object from the slab cache. | |
1630 | * Returns 0 on success. | |
1631 | * | |
1632 | * It is expected this function will be called by a module when it is | |
1633 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
1634 | * cache being allocated each time a module is loaded and unloaded, if the | |
1635 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
1636 | * | |
1637 | * The cache must be empty before calling this function. | |
1638 | * | |
1639 | * The caller must guarantee that noone will allocate memory from the cache | |
1640 | * during the kmem_cache_destroy(). | |
1641 | */ | |
1642 | int kmem_cache_destroy(kmem_cache_t * cachep) | |
1643 | { | |
1644 | int i; | |
1645 | ||
1646 | if (!cachep || in_interrupt()) | |
1647 | BUG(); | |
1648 | ||
1649 | /* Don't let CPUs to come and go */ | |
1650 | lock_cpu_hotplug(); | |
1651 | ||
1652 | /* Find the cache in the chain of caches. */ | |
1653 | down(&cache_chain_sem); | |
1654 | /* | |
1655 | * the chain is never empty, cache_cache is never destroyed | |
1656 | */ | |
1657 | list_del(&cachep->next); | |
1658 | up(&cache_chain_sem); | |
1659 | ||
1660 | if (__cache_shrink(cachep)) { | |
1661 | slab_error(cachep, "Can't free all objects"); | |
1662 | down(&cache_chain_sem); | |
1663 | list_add(&cachep->next,&cache_chain); | |
1664 | up(&cache_chain_sem); | |
1665 | unlock_cpu_hotplug(); | |
1666 | return 1; | |
1667 | } | |
1668 | ||
1669 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 1670 | synchronize_rcu(); |
1da177e4 LT |
1671 | |
1672 | /* no cpu_online check required here since we clear the percpu | |
1673 | * array on cpu offline and set this to NULL. | |
1674 | */ | |
1675 | for (i = 0; i < NR_CPUS; i++) | |
1676 | kfree(cachep->array[i]); | |
1677 | ||
1678 | /* NUMA: free the list3 structures */ | |
1679 | kfree(cachep->lists.shared); | |
1680 | cachep->lists.shared = NULL; | |
1681 | kmem_cache_free(&cache_cache, cachep); | |
1682 | ||
1683 | unlock_cpu_hotplug(); | |
1684 | ||
1685 | return 0; | |
1686 | } | |
1687 | EXPORT_SYMBOL(kmem_cache_destroy); | |
1688 | ||
1689 | /* Get the memory for a slab management obj. */ | |
1690 | static struct slab* alloc_slabmgmt(kmem_cache_t *cachep, | |
1691 | void *objp, int colour_off, unsigned int __nocast local_flags) | |
1692 | { | |
1693 | struct slab *slabp; | |
1694 | ||
1695 | if (OFF_SLAB(cachep)) { | |
1696 | /* Slab management obj is off-slab. */ | |
1697 | slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags); | |
1698 | if (!slabp) | |
1699 | return NULL; | |
1700 | } else { | |
1701 | slabp = objp+colour_off; | |
1702 | colour_off += cachep->slab_size; | |
1703 | } | |
1704 | slabp->inuse = 0; | |
1705 | slabp->colouroff = colour_off; | |
1706 | slabp->s_mem = objp+colour_off; | |
1707 | ||
1708 | return slabp; | |
1709 | } | |
1710 | ||
1711 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
1712 | { | |
1713 | return (kmem_bufctl_t *)(slabp+1); | |
1714 | } | |
1715 | ||
1716 | static void cache_init_objs(kmem_cache_t *cachep, | |
1717 | struct slab *slabp, unsigned long ctor_flags) | |
1718 | { | |
1719 | int i; | |
1720 | ||
1721 | for (i = 0; i < cachep->num; i++) { | |
1722 | void* objp = slabp->s_mem+cachep->objsize*i; | |
1723 | #if DEBUG | |
1724 | /* need to poison the objs? */ | |
1725 | if (cachep->flags & SLAB_POISON) | |
1726 | poison_obj(cachep, objp, POISON_FREE); | |
1727 | if (cachep->flags & SLAB_STORE_USER) | |
1728 | *dbg_userword(cachep, objp) = NULL; | |
1729 | ||
1730 | if (cachep->flags & SLAB_RED_ZONE) { | |
1731 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
1732 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
1733 | } | |
1734 | /* | |
1735 | * Constructors are not allowed to allocate memory from | |
1736 | * the same cache which they are a constructor for. | |
1737 | * Otherwise, deadlock. They must also be threaded. | |
1738 | */ | |
1739 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
1740 | cachep->ctor(objp+obj_dbghead(cachep), cachep, ctor_flags); | |
1741 | ||
1742 | if (cachep->flags & SLAB_RED_ZONE) { | |
1743 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
1744 | slab_error(cachep, "constructor overwrote the" | |
1745 | " end of an object"); | |
1746 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1747 | slab_error(cachep, "constructor overwrote the" | |
1748 | " start of an object"); | |
1749 | } | |
1750 | if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
1751 | kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0); | |
1752 | #else | |
1753 | if (cachep->ctor) | |
1754 | cachep->ctor(objp, cachep, ctor_flags); | |
1755 | #endif | |
1756 | slab_bufctl(slabp)[i] = i+1; | |
1757 | } | |
1758 | slab_bufctl(slabp)[i-1] = BUFCTL_END; | |
1759 | slabp->free = 0; | |
1760 | } | |
1761 | ||
1762 | static void kmem_flagcheck(kmem_cache_t *cachep, unsigned int flags) | |
1763 | { | |
1764 | if (flags & SLAB_DMA) { | |
1765 | if (!(cachep->gfpflags & GFP_DMA)) | |
1766 | BUG(); | |
1767 | } else { | |
1768 | if (cachep->gfpflags & GFP_DMA) | |
1769 | BUG(); | |
1770 | } | |
1771 | } | |
1772 | ||
1773 | static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp) | |
1774 | { | |
1775 | int i; | |
1776 | struct page *page; | |
1777 | ||
1778 | /* Nasty!!!!!! I hope this is OK. */ | |
1779 | i = 1 << cachep->gfporder; | |
1780 | page = virt_to_page(objp); | |
1781 | do { | |
1782 | SET_PAGE_CACHE(page, cachep); | |
1783 | SET_PAGE_SLAB(page, slabp); | |
1784 | page++; | |
1785 | } while (--i); | |
1786 | } | |
1787 | ||
1788 | /* | |
1789 | * Grow (by 1) the number of slabs within a cache. This is called by | |
1790 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
1791 | */ | |
1792 | static int cache_grow(kmem_cache_t *cachep, unsigned int __nocast flags, int nodeid) | |
1793 | { | |
1794 | struct slab *slabp; | |
1795 | void *objp; | |
1796 | size_t offset; | |
1797 | unsigned int local_flags; | |
1798 | unsigned long ctor_flags; | |
1799 | ||
1800 | /* Be lazy and only check for valid flags here, | |
1801 | * keeping it out of the critical path in kmem_cache_alloc(). | |
1802 | */ | |
1803 | if (flags & ~(SLAB_DMA|SLAB_LEVEL_MASK|SLAB_NO_GROW)) | |
1804 | BUG(); | |
1805 | if (flags & SLAB_NO_GROW) | |
1806 | return 0; | |
1807 | ||
1808 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
1809 | local_flags = (flags & SLAB_LEVEL_MASK); | |
1810 | if (!(local_flags & __GFP_WAIT)) | |
1811 | /* | |
1812 | * Not allowed to sleep. Need to tell a constructor about | |
1813 | * this - it might need to know... | |
1814 | */ | |
1815 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
1816 | ||
1817 | /* About to mess with non-constant members - lock. */ | |
1818 | check_irq_off(); | |
1819 | spin_lock(&cachep->spinlock); | |
1820 | ||
1821 | /* Get colour for the slab, and cal the next value. */ | |
1822 | offset = cachep->colour_next; | |
1823 | cachep->colour_next++; | |
1824 | if (cachep->colour_next >= cachep->colour) | |
1825 | cachep->colour_next = 0; | |
1826 | offset *= cachep->colour_off; | |
1827 | ||
1828 | spin_unlock(&cachep->spinlock); | |
1829 | ||
1830 | if (local_flags & __GFP_WAIT) | |
1831 | local_irq_enable(); | |
1832 | ||
1833 | /* | |
1834 | * The test for missing atomic flag is performed here, rather than | |
1835 | * the more obvious place, simply to reduce the critical path length | |
1836 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
1837 | * will eventually be caught here (where it matters). | |
1838 | */ | |
1839 | kmem_flagcheck(cachep, flags); | |
1840 | ||
1841 | ||
1842 | /* Get mem for the objs. */ | |
1843 | if (!(objp = kmem_getpages(cachep, flags, nodeid))) | |
1844 | goto failed; | |
1845 | ||
1846 | /* Get slab management. */ | |
1847 | if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags))) | |
1848 | goto opps1; | |
1849 | ||
1850 | set_slab_attr(cachep, slabp, objp); | |
1851 | ||
1852 | cache_init_objs(cachep, slabp, ctor_flags); | |
1853 | ||
1854 | if (local_flags & __GFP_WAIT) | |
1855 | local_irq_disable(); | |
1856 | check_irq_off(); | |
1857 | spin_lock(&cachep->spinlock); | |
1858 | ||
1859 | /* Make slab active. */ | |
1860 | list_add_tail(&slabp->list, &(list3_data(cachep)->slabs_free)); | |
1861 | STATS_INC_GROWN(cachep); | |
1862 | list3_data(cachep)->free_objects += cachep->num; | |
1863 | spin_unlock(&cachep->spinlock); | |
1864 | return 1; | |
1865 | opps1: | |
1866 | kmem_freepages(cachep, objp); | |
1867 | failed: | |
1868 | if (local_flags & __GFP_WAIT) | |
1869 | local_irq_disable(); | |
1870 | return 0; | |
1871 | } | |
1872 | ||
1873 | #if DEBUG | |
1874 | ||
1875 | /* | |
1876 | * Perform extra freeing checks: | |
1877 | * - detect bad pointers. | |
1878 | * - POISON/RED_ZONE checking | |
1879 | * - destructor calls, for caches with POISON+dtor | |
1880 | */ | |
1881 | static void kfree_debugcheck(const void *objp) | |
1882 | { | |
1883 | struct page *page; | |
1884 | ||
1885 | if (!virt_addr_valid(objp)) { | |
1886 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
1887 | (unsigned long)objp); | |
1888 | BUG(); | |
1889 | } | |
1890 | page = virt_to_page(objp); | |
1891 | if (!PageSlab(page)) { | |
1892 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", (unsigned long)objp); | |
1893 | BUG(); | |
1894 | } | |
1895 | } | |
1896 | ||
1897 | static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp, | |
1898 | void *caller) | |
1899 | { | |
1900 | struct page *page; | |
1901 | unsigned int objnr; | |
1902 | struct slab *slabp; | |
1903 | ||
1904 | objp -= obj_dbghead(cachep); | |
1905 | kfree_debugcheck(objp); | |
1906 | page = virt_to_page(objp); | |
1907 | ||
1908 | if (GET_PAGE_CACHE(page) != cachep) { | |
1909 | printk(KERN_ERR "mismatch in kmem_cache_free: expected cache %p, got %p\n", | |
1910 | GET_PAGE_CACHE(page),cachep); | |
1911 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); | |
1912 | printk(KERN_ERR "%p is %s.\n", GET_PAGE_CACHE(page), GET_PAGE_CACHE(page)->name); | |
1913 | WARN_ON(1); | |
1914 | } | |
1915 | slabp = GET_PAGE_SLAB(page); | |
1916 | ||
1917 | if (cachep->flags & SLAB_RED_ZONE) { | |
1918 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE || *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
1919 | slab_error(cachep, "double free, or memory outside" | |
1920 | " object was overwritten"); | |
1921 | printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
1922 | objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp)); | |
1923 | } | |
1924 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
1925 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
1926 | } | |
1927 | if (cachep->flags & SLAB_STORE_USER) | |
1928 | *dbg_userword(cachep, objp) = caller; | |
1929 | ||
1930 | objnr = (objp-slabp->s_mem)/cachep->objsize; | |
1931 | ||
1932 | BUG_ON(objnr >= cachep->num); | |
1933 | BUG_ON(objp != slabp->s_mem + objnr*cachep->objsize); | |
1934 | ||
1935 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
1936 | /* Need to call the slab's constructor so the | |
1937 | * caller can perform a verify of its state (debugging). | |
1938 | * Called without the cache-lock held. | |
1939 | */ | |
1940 | cachep->ctor(objp+obj_dbghead(cachep), | |
1941 | cachep, SLAB_CTOR_CONSTRUCTOR|SLAB_CTOR_VERIFY); | |
1942 | } | |
1943 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
1944 | /* we want to cache poison the object, | |
1945 | * call the destruction callback | |
1946 | */ | |
1947 | cachep->dtor(objp+obj_dbghead(cachep), cachep, 0); | |
1948 | } | |
1949 | if (cachep->flags & SLAB_POISON) { | |
1950 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
1951 | if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) { | |
1952 | store_stackinfo(cachep, objp, (unsigned long)caller); | |
1953 | kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 0); | |
1954 | } else { | |
1955 | poison_obj(cachep, objp, POISON_FREE); | |
1956 | } | |
1957 | #else | |
1958 | poison_obj(cachep, objp, POISON_FREE); | |
1959 | #endif | |
1960 | } | |
1961 | return objp; | |
1962 | } | |
1963 | ||
1964 | static void check_slabp(kmem_cache_t *cachep, struct slab *slabp) | |
1965 | { | |
1966 | kmem_bufctl_t i; | |
1967 | int entries = 0; | |
1968 | ||
1969 | check_spinlock_acquired(cachep); | |
1970 | /* Check slab's freelist to see if this obj is there. */ | |
1971 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
1972 | entries++; | |
1973 | if (entries > cachep->num || i >= cachep->num) | |
1974 | goto bad; | |
1975 | } | |
1976 | if (entries != cachep->num - slabp->inuse) { | |
1977 | bad: | |
1978 | printk(KERN_ERR "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
1979 | cachep->name, cachep->num, slabp, slabp->inuse); | |
1980 | for (i=0;i<sizeof(slabp)+cachep->num*sizeof(kmem_bufctl_t);i++) { | |
1981 | if ((i%16)==0) | |
1982 | printk("\n%03x:", i); | |
1983 | printk(" %02x", ((unsigned char*)slabp)[i]); | |
1984 | } | |
1985 | printk("\n"); | |
1986 | BUG(); | |
1987 | } | |
1988 | } | |
1989 | #else | |
1990 | #define kfree_debugcheck(x) do { } while(0) | |
1991 | #define cache_free_debugcheck(x,objp,z) (objp) | |
1992 | #define check_slabp(x,y) do { } while(0) | |
1993 | #endif | |
1994 | ||
1995 | static void *cache_alloc_refill(kmem_cache_t *cachep, unsigned int __nocast flags) | |
1996 | { | |
1997 | int batchcount; | |
1998 | struct kmem_list3 *l3; | |
1999 | struct array_cache *ac; | |
2000 | ||
2001 | check_irq_off(); | |
2002 | ac = ac_data(cachep); | |
2003 | retry: | |
2004 | batchcount = ac->batchcount; | |
2005 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
2006 | /* if there was little recent activity on this | |
2007 | * cache, then perform only a partial refill. | |
2008 | * Otherwise we could generate refill bouncing. | |
2009 | */ | |
2010 | batchcount = BATCHREFILL_LIMIT; | |
2011 | } | |
2012 | l3 = list3_data(cachep); | |
2013 | ||
2014 | BUG_ON(ac->avail > 0); | |
2015 | spin_lock(&cachep->spinlock); | |
2016 | if (l3->shared) { | |
2017 | struct array_cache *shared_array = l3->shared; | |
2018 | if (shared_array->avail) { | |
2019 | if (batchcount > shared_array->avail) | |
2020 | batchcount = shared_array->avail; | |
2021 | shared_array->avail -= batchcount; | |
2022 | ac->avail = batchcount; | |
2023 | memcpy(ac_entry(ac), &ac_entry(shared_array)[shared_array->avail], | |
2024 | sizeof(void*)*batchcount); | |
2025 | shared_array->touched = 1; | |
2026 | goto alloc_done; | |
2027 | } | |
2028 | } | |
2029 | while (batchcount > 0) { | |
2030 | struct list_head *entry; | |
2031 | struct slab *slabp; | |
2032 | /* Get slab alloc is to come from. */ | |
2033 | entry = l3->slabs_partial.next; | |
2034 | if (entry == &l3->slabs_partial) { | |
2035 | l3->free_touched = 1; | |
2036 | entry = l3->slabs_free.next; | |
2037 | if (entry == &l3->slabs_free) | |
2038 | goto must_grow; | |
2039 | } | |
2040 | ||
2041 | slabp = list_entry(entry, struct slab, list); | |
2042 | check_slabp(cachep, slabp); | |
2043 | check_spinlock_acquired(cachep); | |
2044 | while (slabp->inuse < cachep->num && batchcount--) { | |
2045 | kmem_bufctl_t next; | |
2046 | STATS_INC_ALLOCED(cachep); | |
2047 | STATS_INC_ACTIVE(cachep); | |
2048 | STATS_SET_HIGH(cachep); | |
2049 | ||
2050 | /* get obj pointer */ | |
2051 | ac_entry(ac)[ac->avail++] = slabp->s_mem + slabp->free*cachep->objsize; | |
2052 | ||
2053 | slabp->inuse++; | |
2054 | next = slab_bufctl(slabp)[slabp->free]; | |
2055 | #if DEBUG | |
2056 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2057 | #endif | |
2058 | slabp->free = next; | |
2059 | } | |
2060 | check_slabp(cachep, slabp); | |
2061 | ||
2062 | /* move slabp to correct slabp list: */ | |
2063 | list_del(&slabp->list); | |
2064 | if (slabp->free == BUFCTL_END) | |
2065 | list_add(&slabp->list, &l3->slabs_full); | |
2066 | else | |
2067 | list_add(&slabp->list, &l3->slabs_partial); | |
2068 | } | |
2069 | ||
2070 | must_grow: | |
2071 | l3->free_objects -= ac->avail; | |
2072 | alloc_done: | |
2073 | spin_unlock(&cachep->spinlock); | |
2074 | ||
2075 | if (unlikely(!ac->avail)) { | |
2076 | int x; | |
2077 | x = cache_grow(cachep, flags, -1); | |
2078 | ||
2079 | // cache_grow can reenable interrupts, then ac could change. | |
2080 | ac = ac_data(cachep); | |
2081 | if (!x && ac->avail == 0) // no objects in sight? abort | |
2082 | return NULL; | |
2083 | ||
2084 | if (!ac->avail) // objects refilled by interrupt? | |
2085 | goto retry; | |
2086 | } | |
2087 | ac->touched = 1; | |
2088 | return ac_entry(ac)[--ac->avail]; | |
2089 | } | |
2090 | ||
2091 | static inline void | |
2092 | cache_alloc_debugcheck_before(kmem_cache_t *cachep, unsigned int __nocast flags) | |
2093 | { | |
2094 | might_sleep_if(flags & __GFP_WAIT); | |
2095 | #if DEBUG | |
2096 | kmem_flagcheck(cachep, flags); | |
2097 | #endif | |
2098 | } | |
2099 | ||
2100 | #if DEBUG | |
2101 | static void * | |
2102 | cache_alloc_debugcheck_after(kmem_cache_t *cachep, | |
2103 | unsigned long flags, void *objp, void *caller) | |
2104 | { | |
2105 | if (!objp) | |
2106 | return objp; | |
2107 | if (cachep->flags & SLAB_POISON) { | |
2108 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
2109 | if ((cachep->objsize % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) | |
2110 | kernel_map_pages(virt_to_page(objp), cachep->objsize/PAGE_SIZE, 1); | |
2111 | else | |
2112 | check_poison_obj(cachep, objp); | |
2113 | #else | |
2114 | check_poison_obj(cachep, objp); | |
2115 | #endif | |
2116 | poison_obj(cachep, objp, POISON_INUSE); | |
2117 | } | |
2118 | if (cachep->flags & SLAB_STORE_USER) | |
2119 | *dbg_userword(cachep, objp) = caller; | |
2120 | ||
2121 | if (cachep->flags & SLAB_RED_ZONE) { | |
2122 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2123 | slab_error(cachep, "double free, or memory outside" | |
2124 | " object was overwritten"); | |
2125 | printk(KERN_ERR "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2126 | objp, *dbg_redzone1(cachep, objp), *dbg_redzone2(cachep, objp)); | |
2127 | } | |
2128 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2129 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2130 | } | |
2131 | objp += obj_dbghead(cachep); | |
2132 | if (cachep->ctor && cachep->flags & SLAB_POISON) { | |
2133 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2134 | ||
2135 | if (!(flags & __GFP_WAIT)) | |
2136 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2137 | ||
2138 | cachep->ctor(objp, cachep, ctor_flags); | |
2139 | } | |
2140 | return objp; | |
2141 | } | |
2142 | #else | |
2143 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2144 | #endif | |
2145 | ||
2146 | ||
2147 | static inline void *__cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags) | |
2148 | { | |
2149 | unsigned long save_flags; | |
2150 | void* objp; | |
2151 | struct array_cache *ac; | |
2152 | ||
2153 | cache_alloc_debugcheck_before(cachep, flags); | |
2154 | ||
2155 | local_irq_save(save_flags); | |
2156 | ac = ac_data(cachep); | |
2157 | if (likely(ac->avail)) { | |
2158 | STATS_INC_ALLOCHIT(cachep); | |
2159 | ac->touched = 1; | |
2160 | objp = ac_entry(ac)[--ac->avail]; | |
2161 | } else { | |
2162 | STATS_INC_ALLOCMISS(cachep); | |
2163 | objp = cache_alloc_refill(cachep, flags); | |
2164 | } | |
2165 | local_irq_restore(save_flags); | |
2166 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, __builtin_return_address(0)); | |
2167 | return objp; | |
2168 | } | |
2169 | ||
2170 | /* | |
2171 | * NUMA: different approach needed if the spinlock is moved into | |
2172 | * the l3 structure | |
2173 | */ | |
2174 | ||
2175 | static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects) | |
2176 | { | |
2177 | int i; | |
2178 | ||
2179 | check_spinlock_acquired(cachep); | |
2180 | ||
2181 | /* NUMA: move add into loop */ | |
2182 | cachep->lists.free_objects += nr_objects; | |
2183 | ||
2184 | for (i = 0; i < nr_objects; i++) { | |
2185 | void *objp = objpp[i]; | |
2186 | struct slab *slabp; | |
2187 | unsigned int objnr; | |
2188 | ||
2189 | slabp = GET_PAGE_SLAB(virt_to_page(objp)); | |
2190 | list_del(&slabp->list); | |
2191 | objnr = (objp - slabp->s_mem) / cachep->objsize; | |
2192 | check_slabp(cachep, slabp); | |
2193 | #if DEBUG | |
2194 | if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) { | |
2195 | printk(KERN_ERR "slab: double free detected in cache '%s', objp %p.\n", | |
2196 | cachep->name, objp); | |
2197 | BUG(); | |
2198 | } | |
2199 | #endif | |
2200 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2201 | slabp->free = objnr; | |
2202 | STATS_DEC_ACTIVE(cachep); | |
2203 | slabp->inuse--; | |
2204 | check_slabp(cachep, slabp); | |
2205 | ||
2206 | /* fixup slab chains */ | |
2207 | if (slabp->inuse == 0) { | |
2208 | if (cachep->lists.free_objects > cachep->free_limit) { | |
2209 | cachep->lists.free_objects -= cachep->num; | |
2210 | slab_destroy(cachep, slabp); | |
2211 | } else { | |
2212 | list_add(&slabp->list, | |
2213 | &list3_data_ptr(cachep, objp)->slabs_free); | |
2214 | } | |
2215 | } else { | |
2216 | /* Unconditionally move a slab to the end of the | |
2217 | * partial list on free - maximum time for the | |
2218 | * other objects to be freed, too. | |
2219 | */ | |
2220 | list_add_tail(&slabp->list, | |
2221 | &list3_data_ptr(cachep, objp)->slabs_partial); | |
2222 | } | |
2223 | } | |
2224 | } | |
2225 | ||
2226 | static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac) | |
2227 | { | |
2228 | int batchcount; | |
2229 | ||
2230 | batchcount = ac->batchcount; | |
2231 | #if DEBUG | |
2232 | BUG_ON(!batchcount || batchcount > ac->avail); | |
2233 | #endif | |
2234 | check_irq_off(); | |
2235 | spin_lock(&cachep->spinlock); | |
2236 | if (cachep->lists.shared) { | |
2237 | struct array_cache *shared_array = cachep->lists.shared; | |
2238 | int max = shared_array->limit-shared_array->avail; | |
2239 | if (max) { | |
2240 | if (batchcount > max) | |
2241 | batchcount = max; | |
2242 | memcpy(&ac_entry(shared_array)[shared_array->avail], | |
2243 | &ac_entry(ac)[0], | |
2244 | sizeof(void*)*batchcount); | |
2245 | shared_array->avail += batchcount; | |
2246 | goto free_done; | |
2247 | } | |
2248 | } | |
2249 | ||
2250 | free_block(cachep, &ac_entry(ac)[0], batchcount); | |
2251 | free_done: | |
2252 | #if STATS | |
2253 | { | |
2254 | int i = 0; | |
2255 | struct list_head *p; | |
2256 | ||
2257 | p = list3_data(cachep)->slabs_free.next; | |
2258 | while (p != &(list3_data(cachep)->slabs_free)) { | |
2259 | struct slab *slabp; | |
2260 | ||
2261 | slabp = list_entry(p, struct slab, list); | |
2262 | BUG_ON(slabp->inuse); | |
2263 | ||
2264 | i++; | |
2265 | p = p->next; | |
2266 | } | |
2267 | STATS_SET_FREEABLE(cachep, i); | |
2268 | } | |
2269 | #endif | |
2270 | spin_unlock(&cachep->spinlock); | |
2271 | ac->avail -= batchcount; | |
2272 | memmove(&ac_entry(ac)[0], &ac_entry(ac)[batchcount], | |
2273 | sizeof(void*)*ac->avail); | |
2274 | } | |
2275 | ||
2276 | /* | |
2277 | * __cache_free | |
2278 | * Release an obj back to its cache. If the obj has a constructed | |
2279 | * state, it must be in this state _before_ it is released. | |
2280 | * | |
2281 | * Called with disabled ints. | |
2282 | */ | |
2283 | static inline void __cache_free(kmem_cache_t *cachep, void *objp) | |
2284 | { | |
2285 | struct array_cache *ac = ac_data(cachep); | |
2286 | ||
2287 | check_irq_off(); | |
2288 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
2289 | ||
2290 | if (likely(ac->avail < ac->limit)) { | |
2291 | STATS_INC_FREEHIT(cachep); | |
2292 | ac_entry(ac)[ac->avail++] = objp; | |
2293 | return; | |
2294 | } else { | |
2295 | STATS_INC_FREEMISS(cachep); | |
2296 | cache_flusharray(cachep, ac); | |
2297 | ac_entry(ac)[ac->avail++] = objp; | |
2298 | } | |
2299 | } | |
2300 | ||
2301 | /** | |
2302 | * kmem_cache_alloc - Allocate an object | |
2303 | * @cachep: The cache to allocate from. | |
2304 | * @flags: See kmalloc(). | |
2305 | * | |
2306 | * Allocate an object from this cache. The flags are only relevant | |
2307 | * if the cache has no available objects. | |
2308 | */ | |
2309 | void *kmem_cache_alloc(kmem_cache_t *cachep, unsigned int __nocast flags) | |
2310 | { | |
2311 | return __cache_alloc(cachep, flags); | |
2312 | } | |
2313 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2314 | ||
2315 | /** | |
2316 | * kmem_ptr_validate - check if an untrusted pointer might | |
2317 | * be a slab entry. | |
2318 | * @cachep: the cache we're checking against | |
2319 | * @ptr: pointer to validate | |
2320 | * | |
2321 | * This verifies that the untrusted pointer looks sane: | |
2322 | * it is _not_ a guarantee that the pointer is actually | |
2323 | * part of the slab cache in question, but it at least | |
2324 | * validates that the pointer can be dereferenced and | |
2325 | * looks half-way sane. | |
2326 | * | |
2327 | * Currently only used for dentry validation. | |
2328 | */ | |
2329 | int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr) | |
2330 | { | |
2331 | unsigned long addr = (unsigned long) ptr; | |
2332 | unsigned long min_addr = PAGE_OFFSET; | |
2333 | unsigned long align_mask = BYTES_PER_WORD-1; | |
2334 | unsigned long size = cachep->objsize; | |
2335 | struct page *page; | |
2336 | ||
2337 | if (unlikely(addr < min_addr)) | |
2338 | goto out; | |
2339 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
2340 | goto out; | |
2341 | if (unlikely(addr & align_mask)) | |
2342 | goto out; | |
2343 | if (unlikely(!kern_addr_valid(addr))) | |
2344 | goto out; | |
2345 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
2346 | goto out; | |
2347 | page = virt_to_page(ptr); | |
2348 | if (unlikely(!PageSlab(page))) | |
2349 | goto out; | |
2350 | if (unlikely(GET_PAGE_CACHE(page) != cachep)) | |
2351 | goto out; | |
2352 | return 1; | |
2353 | out: | |
2354 | return 0; | |
2355 | } | |
2356 | ||
2357 | #ifdef CONFIG_NUMA | |
2358 | /** | |
2359 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
2360 | * @cachep: The cache to allocate from. | |
2361 | * @flags: See kmalloc(). | |
2362 | * @nodeid: node number of the target node. | |
2363 | * | |
2364 | * Identical to kmem_cache_alloc, except that this function is slow | |
2365 | * and can sleep. And it will allocate memory on the given node, which | |
2366 | * can improve the performance for cpu bound structures. | |
2367 | */ | |
97e2bde4 | 2368 | void *kmem_cache_alloc_node(kmem_cache_t *cachep, int flags, int nodeid) |
1da177e4 LT |
2369 | { |
2370 | int loop; | |
2371 | void *objp; | |
2372 | struct slab *slabp; | |
2373 | kmem_bufctl_t next; | |
2374 | ||
2375 | for (loop = 0;;loop++) { | |
2376 | struct list_head *q; | |
2377 | ||
2378 | objp = NULL; | |
2379 | check_irq_on(); | |
2380 | spin_lock_irq(&cachep->spinlock); | |
2381 | /* walk through all partial and empty slab and find one | |
2382 | * from the right node */ | |
2383 | list_for_each(q,&cachep->lists.slabs_partial) { | |
2384 | slabp = list_entry(q, struct slab, list); | |
2385 | ||
2386 | if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid || | |
2387 | loop > 2) | |
2388 | goto got_slabp; | |
2389 | } | |
2390 | list_for_each(q, &cachep->lists.slabs_free) { | |
2391 | slabp = list_entry(q, struct slab, list); | |
2392 | ||
2393 | if (page_to_nid(virt_to_page(slabp->s_mem)) == nodeid || | |
2394 | loop > 2) | |
2395 | goto got_slabp; | |
2396 | } | |
2397 | spin_unlock_irq(&cachep->spinlock); | |
2398 | ||
2399 | local_irq_disable(); | |
97e2bde4 | 2400 | if (!cache_grow(cachep, flags, nodeid)) { |
1da177e4 LT |
2401 | local_irq_enable(); |
2402 | return NULL; | |
2403 | } | |
2404 | local_irq_enable(); | |
2405 | } | |
2406 | got_slabp: | |
2407 | /* found one: allocate object */ | |
2408 | check_slabp(cachep, slabp); | |
2409 | check_spinlock_acquired(cachep); | |
2410 | ||
2411 | STATS_INC_ALLOCED(cachep); | |
2412 | STATS_INC_ACTIVE(cachep); | |
2413 | STATS_SET_HIGH(cachep); | |
2414 | STATS_INC_NODEALLOCS(cachep); | |
2415 | ||
2416 | objp = slabp->s_mem + slabp->free*cachep->objsize; | |
2417 | ||
2418 | slabp->inuse++; | |
2419 | next = slab_bufctl(slabp)[slabp->free]; | |
2420 | #if DEBUG | |
2421 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2422 | #endif | |
2423 | slabp->free = next; | |
2424 | check_slabp(cachep, slabp); | |
2425 | ||
2426 | /* move slabp to correct slabp list: */ | |
2427 | list_del(&slabp->list); | |
2428 | if (slabp->free == BUFCTL_END) | |
2429 | list_add(&slabp->list, &cachep->lists.slabs_full); | |
2430 | else | |
2431 | list_add(&slabp->list, &cachep->lists.slabs_partial); | |
2432 | ||
2433 | list3_data(cachep)->free_objects--; | |
2434 | spin_unlock_irq(&cachep->spinlock); | |
2435 | ||
2436 | objp = cache_alloc_debugcheck_after(cachep, GFP_KERNEL, objp, | |
2437 | __builtin_return_address(0)); | |
2438 | return objp; | |
2439 | } | |
2440 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
2441 | ||
97e2bde4 MS |
2442 | void *kmalloc_node(size_t size, int flags, int node) |
2443 | { | |
2444 | kmem_cache_t *cachep; | |
2445 | ||
2446 | cachep = kmem_find_general_cachep(size, flags); | |
2447 | if (unlikely(cachep == NULL)) | |
2448 | return NULL; | |
2449 | return kmem_cache_alloc_node(cachep, flags, node); | |
2450 | } | |
2451 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
2452 | #endif |
2453 | ||
2454 | /** | |
2455 | * kmalloc - allocate memory | |
2456 | * @size: how many bytes of memory are required. | |
2457 | * @flags: the type of memory to allocate. | |
2458 | * | |
2459 | * kmalloc is the normal method of allocating memory | |
2460 | * in the kernel. | |
2461 | * | |
2462 | * The @flags argument may be one of: | |
2463 | * | |
2464 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
2465 | * | |
2466 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
2467 | * | |
2468 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
2469 | * | |
2470 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
2471 | * must be suitable for DMA. This can mean different things on different | |
2472 | * platforms. For example, on i386, it means that the memory must come | |
2473 | * from the first 16MB. | |
2474 | */ | |
2475 | void *__kmalloc(size_t size, unsigned int __nocast flags) | |
2476 | { | |
2477 | kmem_cache_t *cachep; | |
2478 | ||
97e2bde4 MS |
2479 | /* If you want to save a few bytes .text space: replace |
2480 | * __ with kmem_. | |
2481 | * Then kmalloc uses the uninlined functions instead of the inline | |
2482 | * functions. | |
2483 | */ | |
2484 | cachep = __find_general_cachep(size, flags); | |
1da177e4 LT |
2485 | if (unlikely(cachep == NULL)) |
2486 | return NULL; | |
2487 | return __cache_alloc(cachep, flags); | |
2488 | } | |
2489 | EXPORT_SYMBOL(__kmalloc); | |
2490 | ||
2491 | #ifdef CONFIG_SMP | |
2492 | /** | |
2493 | * __alloc_percpu - allocate one copy of the object for every present | |
2494 | * cpu in the system, zeroing them. | |
2495 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
2496 | * | |
2497 | * @size: how many bytes of memory are required. | |
2498 | * @align: the alignment, which can't be greater than SMP_CACHE_BYTES. | |
2499 | */ | |
2500 | void *__alloc_percpu(size_t size, size_t align) | |
2501 | { | |
2502 | int i; | |
2503 | struct percpu_data *pdata = kmalloc(sizeof (*pdata), GFP_KERNEL); | |
2504 | ||
2505 | if (!pdata) | |
2506 | return NULL; | |
2507 | ||
2508 | for (i = 0; i < NR_CPUS; i++) { | |
2509 | if (!cpu_possible(i)) | |
2510 | continue; | |
97e2bde4 MS |
2511 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, |
2512 | cpu_to_node(i)); | |
1da177e4 LT |
2513 | |
2514 | if (!pdata->ptrs[i]) | |
2515 | goto unwind_oom; | |
2516 | memset(pdata->ptrs[i], 0, size); | |
2517 | } | |
2518 | ||
2519 | /* Catch derefs w/o wrappers */ | |
2520 | return (void *) (~(unsigned long) pdata); | |
2521 | ||
2522 | unwind_oom: | |
2523 | while (--i >= 0) { | |
2524 | if (!cpu_possible(i)) | |
2525 | continue; | |
2526 | kfree(pdata->ptrs[i]); | |
2527 | } | |
2528 | kfree(pdata); | |
2529 | return NULL; | |
2530 | } | |
2531 | EXPORT_SYMBOL(__alloc_percpu); | |
2532 | #endif | |
2533 | ||
2534 | /** | |
2535 | * kmem_cache_free - Deallocate an object | |
2536 | * @cachep: The cache the allocation was from. | |
2537 | * @objp: The previously allocated object. | |
2538 | * | |
2539 | * Free an object which was previously allocated from this | |
2540 | * cache. | |
2541 | */ | |
2542 | void kmem_cache_free(kmem_cache_t *cachep, void *objp) | |
2543 | { | |
2544 | unsigned long flags; | |
2545 | ||
2546 | local_irq_save(flags); | |
2547 | __cache_free(cachep, objp); | |
2548 | local_irq_restore(flags); | |
2549 | } | |
2550 | EXPORT_SYMBOL(kmem_cache_free); | |
2551 | ||
2552 | /** | |
2553 | * kcalloc - allocate memory for an array. The memory is set to zero. | |
2554 | * @n: number of elements. | |
2555 | * @size: element size. | |
2556 | * @flags: the type of memory to allocate. | |
2557 | */ | |
2558 | void *kcalloc(size_t n, size_t size, unsigned int __nocast flags) | |
2559 | { | |
2560 | void *ret = NULL; | |
2561 | ||
2562 | if (n != 0 && size > INT_MAX / n) | |
2563 | return ret; | |
2564 | ||
2565 | ret = kmalloc(n * size, flags); | |
2566 | if (ret) | |
2567 | memset(ret, 0, n * size); | |
2568 | return ret; | |
2569 | } | |
2570 | EXPORT_SYMBOL(kcalloc); | |
2571 | ||
2572 | /** | |
2573 | * kfree - free previously allocated memory | |
2574 | * @objp: pointer returned by kmalloc. | |
2575 | * | |
2576 | * Don't free memory not originally allocated by kmalloc() | |
2577 | * or you will run into trouble. | |
2578 | */ | |
2579 | void kfree(const void *objp) | |
2580 | { | |
2581 | kmem_cache_t *c; | |
2582 | unsigned long flags; | |
2583 | ||
2584 | if (unlikely(!objp)) | |
2585 | return; | |
2586 | local_irq_save(flags); | |
2587 | kfree_debugcheck(objp); | |
2588 | c = GET_PAGE_CACHE(virt_to_page(objp)); | |
2589 | __cache_free(c, (void*)objp); | |
2590 | local_irq_restore(flags); | |
2591 | } | |
2592 | EXPORT_SYMBOL(kfree); | |
2593 | ||
2594 | #ifdef CONFIG_SMP | |
2595 | /** | |
2596 | * free_percpu - free previously allocated percpu memory | |
2597 | * @objp: pointer returned by alloc_percpu. | |
2598 | * | |
2599 | * Don't free memory not originally allocated by alloc_percpu() | |
2600 | * The complemented objp is to check for that. | |
2601 | */ | |
2602 | void | |
2603 | free_percpu(const void *objp) | |
2604 | { | |
2605 | int i; | |
2606 | struct percpu_data *p = (struct percpu_data *) (~(unsigned long) objp); | |
2607 | ||
2608 | for (i = 0; i < NR_CPUS; i++) { | |
2609 | if (!cpu_possible(i)) | |
2610 | continue; | |
2611 | kfree(p->ptrs[i]); | |
2612 | } | |
2613 | kfree(p); | |
2614 | } | |
2615 | EXPORT_SYMBOL(free_percpu); | |
2616 | #endif | |
2617 | ||
2618 | unsigned int kmem_cache_size(kmem_cache_t *cachep) | |
2619 | { | |
2620 | return obj_reallen(cachep); | |
2621 | } | |
2622 | EXPORT_SYMBOL(kmem_cache_size); | |
2623 | ||
1944972d ACM |
2624 | const char *kmem_cache_name(kmem_cache_t *cachep) |
2625 | { | |
2626 | return cachep->name; | |
2627 | } | |
2628 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
2629 | ||
1da177e4 LT |
2630 | struct ccupdate_struct { |
2631 | kmem_cache_t *cachep; | |
2632 | struct array_cache *new[NR_CPUS]; | |
2633 | }; | |
2634 | ||
2635 | static void do_ccupdate_local(void *info) | |
2636 | { | |
2637 | struct ccupdate_struct *new = (struct ccupdate_struct *)info; | |
2638 | struct array_cache *old; | |
2639 | ||
2640 | check_irq_off(); | |
2641 | old = ac_data(new->cachep); | |
2642 | ||
2643 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; | |
2644 | new->new[smp_processor_id()] = old; | |
2645 | } | |
2646 | ||
2647 | ||
2648 | static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount, | |
2649 | int shared) | |
2650 | { | |
2651 | struct ccupdate_struct new; | |
2652 | struct array_cache *new_shared; | |
2653 | int i; | |
2654 | ||
2655 | memset(&new.new,0,sizeof(new.new)); | |
2656 | for (i = 0; i < NR_CPUS; i++) { | |
2657 | if (cpu_online(i)) { | |
2658 | new.new[i] = alloc_arraycache(i, limit, batchcount); | |
2659 | if (!new.new[i]) { | |
2660 | for (i--; i >= 0; i--) kfree(new.new[i]); | |
2661 | return -ENOMEM; | |
2662 | } | |
2663 | } else { | |
2664 | new.new[i] = NULL; | |
2665 | } | |
2666 | } | |
2667 | new.cachep = cachep; | |
2668 | ||
2669 | smp_call_function_all_cpus(do_ccupdate_local, (void *)&new); | |
2670 | ||
2671 | check_irq_on(); | |
2672 | spin_lock_irq(&cachep->spinlock); | |
2673 | cachep->batchcount = batchcount; | |
2674 | cachep->limit = limit; | |
2675 | cachep->free_limit = (1+num_online_cpus())*cachep->batchcount + cachep->num; | |
2676 | spin_unlock_irq(&cachep->spinlock); | |
2677 | ||
2678 | for (i = 0; i < NR_CPUS; i++) { | |
2679 | struct array_cache *ccold = new.new[i]; | |
2680 | if (!ccold) | |
2681 | continue; | |
2682 | spin_lock_irq(&cachep->spinlock); | |
2683 | free_block(cachep, ac_entry(ccold), ccold->avail); | |
2684 | spin_unlock_irq(&cachep->spinlock); | |
2685 | kfree(ccold); | |
2686 | } | |
2687 | new_shared = alloc_arraycache(-1, batchcount*shared, 0xbaadf00d); | |
2688 | if (new_shared) { | |
2689 | struct array_cache *old; | |
2690 | ||
2691 | spin_lock_irq(&cachep->spinlock); | |
2692 | old = cachep->lists.shared; | |
2693 | cachep->lists.shared = new_shared; | |
2694 | if (old) | |
2695 | free_block(cachep, ac_entry(old), old->avail); | |
2696 | spin_unlock_irq(&cachep->spinlock); | |
2697 | kfree(old); | |
2698 | } | |
2699 | ||
2700 | return 0; | |
2701 | } | |
2702 | ||
2703 | ||
2704 | static void enable_cpucache(kmem_cache_t *cachep) | |
2705 | { | |
2706 | int err; | |
2707 | int limit, shared; | |
2708 | ||
2709 | /* The head array serves three purposes: | |
2710 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
2711 | * - reduce the number of spinlock operations. | |
2712 | * - reduce the number of linked list operations on the slab and | |
2713 | * bufctl chains: array operations are cheaper. | |
2714 | * The numbers are guessed, we should auto-tune as described by | |
2715 | * Bonwick. | |
2716 | */ | |
2717 | if (cachep->objsize > 131072) | |
2718 | limit = 1; | |
2719 | else if (cachep->objsize > PAGE_SIZE) | |
2720 | limit = 8; | |
2721 | else if (cachep->objsize > 1024) | |
2722 | limit = 24; | |
2723 | else if (cachep->objsize > 256) | |
2724 | limit = 54; | |
2725 | else | |
2726 | limit = 120; | |
2727 | ||
2728 | /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound | |
2729 | * allocation behaviour: Most allocs on one cpu, most free operations | |
2730 | * on another cpu. For these cases, an efficient object passing between | |
2731 | * cpus is necessary. This is provided by a shared array. The array | |
2732 | * replaces Bonwick's magazine layer. | |
2733 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
2734 | * to a larger limit. Thus disabled by default. | |
2735 | */ | |
2736 | shared = 0; | |
2737 | #ifdef CONFIG_SMP | |
2738 | if (cachep->objsize <= PAGE_SIZE) | |
2739 | shared = 8; | |
2740 | #endif | |
2741 | ||
2742 | #if DEBUG | |
2743 | /* With debugging enabled, large batchcount lead to excessively | |
2744 | * long periods with disabled local interrupts. Limit the | |
2745 | * batchcount | |
2746 | */ | |
2747 | if (limit > 32) | |
2748 | limit = 32; | |
2749 | #endif | |
2750 | err = do_tune_cpucache(cachep, limit, (limit+1)/2, shared); | |
2751 | if (err) | |
2752 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
2753 | cachep->name, -err); | |
2754 | } | |
2755 | ||
2756 | static void drain_array_locked(kmem_cache_t *cachep, | |
2757 | struct array_cache *ac, int force) | |
2758 | { | |
2759 | int tofree; | |
2760 | ||
2761 | check_spinlock_acquired(cachep); | |
2762 | if (ac->touched && !force) { | |
2763 | ac->touched = 0; | |
2764 | } else if (ac->avail) { | |
2765 | tofree = force ? ac->avail : (ac->limit+4)/5; | |
2766 | if (tofree > ac->avail) { | |
2767 | tofree = (ac->avail+1)/2; | |
2768 | } | |
2769 | free_block(cachep, ac_entry(ac), tofree); | |
2770 | ac->avail -= tofree; | |
2771 | memmove(&ac_entry(ac)[0], &ac_entry(ac)[tofree], | |
2772 | sizeof(void*)*ac->avail); | |
2773 | } | |
2774 | } | |
2775 | ||
2776 | /** | |
2777 | * cache_reap - Reclaim memory from caches. | |
2778 | * | |
2779 | * Called from workqueue/eventd every few seconds. | |
2780 | * Purpose: | |
2781 | * - clear the per-cpu caches for this CPU. | |
2782 | * - return freeable pages to the main free memory pool. | |
2783 | * | |
2784 | * If we cannot acquire the cache chain semaphore then just give up - we'll | |
2785 | * try again on the next iteration. | |
2786 | */ | |
2787 | static void cache_reap(void *unused) | |
2788 | { | |
2789 | struct list_head *walk; | |
2790 | ||
2791 | if (down_trylock(&cache_chain_sem)) { | |
2792 | /* Give up. Setup the next iteration. */ | |
2793 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id()); | |
2794 | return; | |
2795 | } | |
2796 | ||
2797 | list_for_each(walk, &cache_chain) { | |
2798 | kmem_cache_t *searchp; | |
2799 | struct list_head* p; | |
2800 | int tofree; | |
2801 | struct slab *slabp; | |
2802 | ||
2803 | searchp = list_entry(walk, kmem_cache_t, next); | |
2804 | ||
2805 | if (searchp->flags & SLAB_NO_REAP) | |
2806 | goto next; | |
2807 | ||
2808 | check_irq_on(); | |
2809 | ||
2810 | spin_lock_irq(&searchp->spinlock); | |
2811 | ||
2812 | drain_array_locked(searchp, ac_data(searchp), 0); | |
2813 | ||
2814 | if(time_after(searchp->lists.next_reap, jiffies)) | |
2815 | goto next_unlock; | |
2816 | ||
2817 | searchp->lists.next_reap = jiffies + REAPTIMEOUT_LIST3; | |
2818 | ||
2819 | if (searchp->lists.shared) | |
2820 | drain_array_locked(searchp, searchp->lists.shared, 0); | |
2821 | ||
2822 | if (searchp->lists.free_touched) { | |
2823 | searchp->lists.free_touched = 0; | |
2824 | goto next_unlock; | |
2825 | } | |
2826 | ||
2827 | tofree = (searchp->free_limit+5*searchp->num-1)/(5*searchp->num); | |
2828 | do { | |
2829 | p = list3_data(searchp)->slabs_free.next; | |
2830 | if (p == &(list3_data(searchp)->slabs_free)) | |
2831 | break; | |
2832 | ||
2833 | slabp = list_entry(p, struct slab, list); | |
2834 | BUG_ON(slabp->inuse); | |
2835 | list_del(&slabp->list); | |
2836 | STATS_INC_REAPED(searchp); | |
2837 | ||
2838 | /* Safe to drop the lock. The slab is no longer | |
2839 | * linked to the cache. | |
2840 | * searchp cannot disappear, we hold | |
2841 | * cache_chain_lock | |
2842 | */ | |
2843 | searchp->lists.free_objects -= searchp->num; | |
2844 | spin_unlock_irq(&searchp->spinlock); | |
2845 | slab_destroy(searchp, slabp); | |
2846 | spin_lock_irq(&searchp->spinlock); | |
2847 | } while(--tofree > 0); | |
2848 | next_unlock: | |
2849 | spin_unlock_irq(&searchp->spinlock); | |
2850 | next: | |
2851 | cond_resched(); | |
2852 | } | |
2853 | check_irq_on(); | |
2854 | up(&cache_chain_sem); | |
4ae7c039 | 2855 | drain_remote_pages(); |
1da177e4 LT |
2856 | /* Setup the next iteration */ |
2857 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC + smp_processor_id()); | |
2858 | } | |
2859 | ||
2860 | #ifdef CONFIG_PROC_FS | |
2861 | ||
2862 | static void *s_start(struct seq_file *m, loff_t *pos) | |
2863 | { | |
2864 | loff_t n = *pos; | |
2865 | struct list_head *p; | |
2866 | ||
2867 | down(&cache_chain_sem); | |
2868 | if (!n) { | |
2869 | /* | |
2870 | * Output format version, so at least we can change it | |
2871 | * without _too_ many complaints. | |
2872 | */ | |
2873 | #if STATS | |
2874 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); | |
2875 | #else | |
2876 | seq_puts(m, "slabinfo - version: 2.1\n"); | |
2877 | #endif | |
2878 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>"); | |
2879 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
2880 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
2881 | #if STATS | |
2882 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped>" | |
2883 | " <error> <maxfreeable> <freelimit> <nodeallocs>"); | |
2884 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
2885 | #endif | |
2886 | seq_putc(m, '\n'); | |
2887 | } | |
2888 | p = cache_chain.next; | |
2889 | while (n--) { | |
2890 | p = p->next; | |
2891 | if (p == &cache_chain) | |
2892 | return NULL; | |
2893 | } | |
2894 | return list_entry(p, kmem_cache_t, next); | |
2895 | } | |
2896 | ||
2897 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
2898 | { | |
2899 | kmem_cache_t *cachep = p; | |
2900 | ++*pos; | |
2901 | return cachep->next.next == &cache_chain ? NULL | |
2902 | : list_entry(cachep->next.next, kmem_cache_t, next); | |
2903 | } | |
2904 | ||
2905 | static void s_stop(struct seq_file *m, void *p) | |
2906 | { | |
2907 | up(&cache_chain_sem); | |
2908 | } | |
2909 | ||
2910 | static int s_show(struct seq_file *m, void *p) | |
2911 | { | |
2912 | kmem_cache_t *cachep = p; | |
2913 | struct list_head *q; | |
2914 | struct slab *slabp; | |
2915 | unsigned long active_objs; | |
2916 | unsigned long num_objs; | |
2917 | unsigned long active_slabs = 0; | |
2918 | unsigned long num_slabs; | |
2919 | const char *name; | |
2920 | char *error = NULL; | |
2921 | ||
2922 | check_irq_on(); | |
2923 | spin_lock_irq(&cachep->spinlock); | |
2924 | active_objs = 0; | |
2925 | num_slabs = 0; | |
2926 | list_for_each(q,&cachep->lists.slabs_full) { | |
2927 | slabp = list_entry(q, struct slab, list); | |
2928 | if (slabp->inuse != cachep->num && !error) | |
2929 | error = "slabs_full accounting error"; | |
2930 | active_objs += cachep->num; | |
2931 | active_slabs++; | |
2932 | } | |
2933 | list_for_each(q,&cachep->lists.slabs_partial) { | |
2934 | slabp = list_entry(q, struct slab, list); | |
2935 | if (slabp->inuse == cachep->num && !error) | |
2936 | error = "slabs_partial inuse accounting error"; | |
2937 | if (!slabp->inuse && !error) | |
2938 | error = "slabs_partial/inuse accounting error"; | |
2939 | active_objs += slabp->inuse; | |
2940 | active_slabs++; | |
2941 | } | |
2942 | list_for_each(q,&cachep->lists.slabs_free) { | |
2943 | slabp = list_entry(q, struct slab, list); | |
2944 | if (slabp->inuse && !error) | |
2945 | error = "slabs_free/inuse accounting error"; | |
2946 | num_slabs++; | |
2947 | } | |
2948 | num_slabs+=active_slabs; | |
2949 | num_objs = num_slabs*cachep->num; | |
2950 | if (num_objs - active_objs != cachep->lists.free_objects && !error) | |
2951 | error = "free_objects accounting error"; | |
2952 | ||
2953 | name = cachep->name; | |
2954 | if (error) | |
2955 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
2956 | ||
2957 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
2958 | name, active_objs, num_objs, cachep->objsize, | |
2959 | cachep->num, (1<<cachep->gfporder)); | |
2960 | seq_printf(m, " : tunables %4u %4u %4u", | |
2961 | cachep->limit, cachep->batchcount, | |
2962 | cachep->lists.shared->limit/cachep->batchcount); | |
2963 | seq_printf(m, " : slabdata %6lu %6lu %6u", | |
2964 | active_slabs, num_slabs, cachep->lists.shared->avail); | |
2965 | #if STATS | |
2966 | { /* list3 stats */ | |
2967 | unsigned long high = cachep->high_mark; | |
2968 | unsigned long allocs = cachep->num_allocations; | |
2969 | unsigned long grown = cachep->grown; | |
2970 | unsigned long reaped = cachep->reaped; | |
2971 | unsigned long errors = cachep->errors; | |
2972 | unsigned long max_freeable = cachep->max_freeable; | |
2973 | unsigned long free_limit = cachep->free_limit; | |
2974 | unsigned long node_allocs = cachep->node_allocs; | |
2975 | ||
2976 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu", | |
2977 | allocs, high, grown, reaped, errors, | |
2978 | max_freeable, free_limit, node_allocs); | |
2979 | } | |
2980 | /* cpu stats */ | |
2981 | { | |
2982 | unsigned long allochit = atomic_read(&cachep->allochit); | |
2983 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
2984 | unsigned long freehit = atomic_read(&cachep->freehit); | |
2985 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
2986 | ||
2987 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
2988 | allochit, allocmiss, freehit, freemiss); | |
2989 | } | |
2990 | #endif | |
2991 | seq_putc(m, '\n'); | |
2992 | spin_unlock_irq(&cachep->spinlock); | |
2993 | return 0; | |
2994 | } | |
2995 | ||
2996 | /* | |
2997 | * slabinfo_op - iterator that generates /proc/slabinfo | |
2998 | * | |
2999 | * Output layout: | |
3000 | * cache-name | |
3001 | * num-active-objs | |
3002 | * total-objs | |
3003 | * object size | |
3004 | * num-active-slabs | |
3005 | * total-slabs | |
3006 | * num-pages-per-slab | |
3007 | * + further values on SMP and with statistics enabled | |
3008 | */ | |
3009 | ||
3010 | struct seq_operations slabinfo_op = { | |
3011 | .start = s_start, | |
3012 | .next = s_next, | |
3013 | .stop = s_stop, | |
3014 | .show = s_show, | |
3015 | }; | |
3016 | ||
3017 | #define MAX_SLABINFO_WRITE 128 | |
3018 | /** | |
3019 | * slabinfo_write - Tuning for the slab allocator | |
3020 | * @file: unused | |
3021 | * @buffer: user buffer | |
3022 | * @count: data length | |
3023 | * @ppos: unused | |
3024 | */ | |
3025 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, | |
3026 | size_t count, loff_t *ppos) | |
3027 | { | |
3028 | char kbuf[MAX_SLABINFO_WRITE+1], *tmp; | |
3029 | int limit, batchcount, shared, res; | |
3030 | struct list_head *p; | |
3031 | ||
3032 | if (count > MAX_SLABINFO_WRITE) | |
3033 | return -EINVAL; | |
3034 | if (copy_from_user(&kbuf, buffer, count)) | |
3035 | return -EFAULT; | |
3036 | kbuf[MAX_SLABINFO_WRITE] = '\0'; | |
3037 | ||
3038 | tmp = strchr(kbuf, ' '); | |
3039 | if (!tmp) | |
3040 | return -EINVAL; | |
3041 | *tmp = '\0'; | |
3042 | tmp++; | |
3043 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3044 | return -EINVAL; | |
3045 | ||
3046 | /* Find the cache in the chain of caches. */ | |
3047 | down(&cache_chain_sem); | |
3048 | res = -EINVAL; | |
3049 | list_for_each(p,&cache_chain) { | |
3050 | kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next); | |
3051 | ||
3052 | if (!strcmp(cachep->name, kbuf)) { | |
3053 | if (limit < 1 || | |
3054 | batchcount < 1 || | |
3055 | batchcount > limit || | |
3056 | shared < 0) { | |
3057 | res = -EINVAL; | |
3058 | } else { | |
3059 | res = do_tune_cpucache(cachep, limit, batchcount, shared); | |
3060 | } | |
3061 | break; | |
3062 | } | |
3063 | } | |
3064 | up(&cache_chain_sem); | |
3065 | if (res >= 0) | |
3066 | res = count; | |
3067 | return res; | |
3068 | } | |
3069 | #endif | |
3070 | ||
3071 | unsigned int ksize(const void *objp) | |
3072 | { | |
3073 | kmem_cache_t *c; | |
3074 | unsigned long flags; | |
3075 | unsigned int size = 0; | |
3076 | ||
3077 | if (likely(objp != NULL)) { | |
3078 | local_irq_save(flags); | |
3079 | c = GET_PAGE_CACHE(virt_to_page(objp)); | |
3080 | size = kmem_cache_size(c); | |
3081 | local_irq_restore(flags); | |
3082 | } | |
3083 | ||
3084 | return size; | |
3085 | } | |
543537bd PM |
3086 | |
3087 | ||
3088 | /* | |
3089 | * kstrdup - allocate space for and copy an existing string | |
3090 | * | |
3091 | * @s: the string to duplicate | |
3092 | * @gfp: the GFP mask used in the kmalloc() call when allocating memory | |
3093 | */ | |
3094 | char *kstrdup(const char *s, int gfp) | |
3095 | { | |
3096 | size_t len; | |
3097 | char *buf; | |
3098 | ||
3099 | if (!s) | |
3100 | return NULL; | |
3101 | ||
3102 | len = strlen(s) + 1; | |
3103 | buf = kmalloc(len, gfp); | |
3104 | if (buf) | |
3105 | memcpy(buf, s, len); | |
3106 | return buf; | |
3107 | } | |
3108 | EXPORT_SYMBOL(kstrdup); |