Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
29 | * slabs and you must pass objects with the same intializations to | |
30 | * kmem_cache_free. | |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
53 | * The c_cpuarray may not be read with enabled local interrupts - | |
54 | * it's changed with a smp_call_function(). | |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
58 | * Several members in kmem_cache_t and struct slab never change, they | |
59 | * are accessed without any locking. | |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
fc0abb14 | 71 | * The global cache-chain is protected by the mutex 'cache_chain_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
89 | #include <linux/config.h> | |
90 | #include <linux/slab.h> | |
91 | #include <linux/mm.h> | |
92 | #include <linux/swap.h> | |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
97 | #include <linux/seq_file.h> | |
98 | #include <linux/notifier.h> | |
99 | #include <linux/kallsyms.h> | |
100 | #include <linux/cpu.h> | |
101 | #include <linux/sysctl.h> | |
102 | #include <linux/module.h> | |
103 | #include <linux/rcupdate.h> | |
543537bd | 104 | #include <linux/string.h> |
e498be7d | 105 | #include <linux/nodemask.h> |
dc85da15 | 106 | #include <linux/mempolicy.h> |
fc0abb14 | 107 | #include <linux/mutex.h> |
1da177e4 LT |
108 | |
109 | #include <asm/uaccess.h> | |
110 | #include <asm/cacheflush.h> | |
111 | #include <asm/tlbflush.h> | |
112 | #include <asm/page.h> | |
113 | ||
114 | /* | |
115 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_DEBUG_INITIAL, | |
116 | * SLAB_RED_ZONE & SLAB_POISON. | |
117 | * 0 for faster, smaller code (especially in the critical paths). | |
118 | * | |
119 | * STATS - 1 to collect stats for /proc/slabinfo. | |
120 | * 0 for faster, smaller code (especially in the critical paths). | |
121 | * | |
122 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
123 | */ | |
124 | ||
125 | #ifdef CONFIG_DEBUG_SLAB | |
126 | #define DEBUG 1 | |
127 | #define STATS 1 | |
128 | #define FORCED_DEBUG 1 | |
129 | #else | |
130 | #define DEBUG 0 | |
131 | #define STATS 0 | |
132 | #define FORCED_DEBUG 0 | |
133 | #endif | |
134 | ||
1da177e4 LT |
135 | /* Shouldn't this be in a header file somewhere? */ |
136 | #define BYTES_PER_WORD sizeof(void *) | |
137 | ||
138 | #ifndef cache_line_size | |
139 | #define cache_line_size() L1_CACHE_BYTES | |
140 | #endif | |
141 | ||
142 | #ifndef ARCH_KMALLOC_MINALIGN | |
143 | /* | |
144 | * Enforce a minimum alignment for the kmalloc caches. | |
145 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
146 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
147 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
148 | * alignment larger than BYTES_PER_WORD. ARCH_KMALLOC_MINALIGN allows that. | |
149 | * Note that this flag disables some debug features. | |
150 | */ | |
151 | #define ARCH_KMALLOC_MINALIGN 0 | |
152 | #endif | |
153 | ||
154 | #ifndef ARCH_SLAB_MINALIGN | |
155 | /* | |
156 | * Enforce a minimum alignment for all caches. | |
157 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
158 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
159 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
160 | * some debug features. | |
161 | */ | |
162 | #define ARCH_SLAB_MINALIGN 0 | |
163 | #endif | |
164 | ||
165 | #ifndef ARCH_KMALLOC_FLAGS | |
166 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
167 | #endif | |
168 | ||
169 | /* Legal flag mask for kmem_cache_create(). */ | |
170 | #if DEBUG | |
171 | # define CREATE_MASK (SLAB_DEBUG_INITIAL | SLAB_RED_ZONE | \ | |
172 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
173 | SLAB_NO_REAP | SLAB_CACHE_DMA | \ | |
174 | SLAB_MUST_HWCACHE_ALIGN | SLAB_STORE_USER | \ | |
175 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
176 | SLAB_DESTROY_BY_RCU) | |
177 | #else | |
178 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | SLAB_NO_REAP | \ | |
179 | SLAB_CACHE_DMA | SLAB_MUST_HWCACHE_ALIGN | \ | |
180 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
181 | SLAB_DESTROY_BY_RCU) | |
182 | #endif | |
183 | ||
184 | /* | |
185 | * kmem_bufctl_t: | |
186 | * | |
187 | * Bufctl's are used for linking objs within a slab | |
188 | * linked offsets. | |
189 | * | |
190 | * This implementation relies on "struct page" for locating the cache & | |
191 | * slab an object belongs to. | |
192 | * This allows the bufctl structure to be small (one int), but limits | |
193 | * the number of objects a slab (not a cache) can contain when off-slab | |
194 | * bufctls are used. The limit is the size of the largest general cache | |
195 | * that does not use off-slab slabs. | |
196 | * For 32bit archs with 4 kB pages, is this 56. | |
197 | * This is not serious, as it is only for large objects, when it is unwise | |
198 | * to have too many per slab. | |
199 | * Note: This limit can be raised by introducing a general cache whose size | |
200 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
201 | */ | |
202 | ||
fa5b08d5 | 203 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
204 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
205 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
206 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-2) | |
207 | ||
208 | /* Max number of objs-per-slab for caches which use off-slab slabs. | |
209 | * Needed to avoid a possible looping condition in cache_grow(). | |
210 | */ | |
211 | static unsigned long offslab_limit; | |
212 | ||
213 | /* | |
214 | * struct slab | |
215 | * | |
216 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
217 | * for a slab, or allocated from an general cache. | |
218 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
219 | */ | |
220 | struct slab { | |
b28a02de PE |
221 | struct list_head list; |
222 | unsigned long colouroff; | |
223 | void *s_mem; /* including colour offset */ | |
224 | unsigned int inuse; /* num of objs active in slab */ | |
225 | kmem_bufctl_t free; | |
226 | unsigned short nodeid; | |
1da177e4 LT |
227 | }; |
228 | ||
229 | /* | |
230 | * struct slab_rcu | |
231 | * | |
232 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
233 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
234 | * we need to approach a kernel structure obliquely, from its address | |
235 | * obtained without the usual locking. We can lock the structure to | |
236 | * stabilize it and check it's still at the given address, only if we | |
237 | * can be sure that the memory has not been meanwhile reused for some | |
238 | * other kind of object (which our subsystem's lock might corrupt). | |
239 | * | |
240 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
241 | * taking the spinlock within the structure expected at that address. | |
242 | * | |
243 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
244 | */ | |
245 | struct slab_rcu { | |
b28a02de PE |
246 | struct rcu_head head; |
247 | kmem_cache_t *cachep; | |
248 | void *addr; | |
1da177e4 LT |
249 | }; |
250 | ||
251 | /* | |
252 | * struct array_cache | |
253 | * | |
1da177e4 LT |
254 | * Purpose: |
255 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
256 | * - reduce the number of linked list operations | |
257 | * - reduce spinlock operations | |
258 | * | |
259 | * The limit is stored in the per-cpu structure to reduce the data cache | |
260 | * footprint. | |
261 | * | |
262 | */ | |
263 | struct array_cache { | |
264 | unsigned int avail; | |
265 | unsigned int limit; | |
266 | unsigned int batchcount; | |
267 | unsigned int touched; | |
e498be7d CL |
268 | spinlock_t lock; |
269 | void *entry[0]; /* | |
270 | * Must have this definition in here for the proper | |
271 | * alignment of array_cache. Also simplifies accessing | |
272 | * the entries. | |
273 | * [0] is for gcc 2.95. It should really be []. | |
274 | */ | |
1da177e4 LT |
275 | }; |
276 | ||
277 | /* bootstrap: The caches do not work without cpuarrays anymore, | |
278 | * but the cpuarrays are allocated from the generic caches... | |
279 | */ | |
280 | #define BOOT_CPUCACHE_ENTRIES 1 | |
281 | struct arraycache_init { | |
282 | struct array_cache cache; | |
b28a02de | 283 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
284 | }; |
285 | ||
286 | /* | |
e498be7d | 287 | * The slab lists for all objects. |
1da177e4 LT |
288 | */ |
289 | struct kmem_list3 { | |
b28a02de PE |
290 | struct list_head slabs_partial; /* partial list first, better asm code */ |
291 | struct list_head slabs_full; | |
292 | struct list_head slabs_free; | |
293 | unsigned long free_objects; | |
294 | unsigned long next_reap; | |
295 | int free_touched; | |
296 | unsigned int free_limit; | |
297 | spinlock_t list_lock; | |
298 | struct array_cache *shared; /* shared per node */ | |
299 | struct array_cache **alien; /* on other nodes */ | |
1da177e4 LT |
300 | }; |
301 | ||
e498be7d CL |
302 | /* |
303 | * Need this for bootstrapping a per node allocator. | |
304 | */ | |
305 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
306 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
307 | #define CACHE_CACHE 0 | |
308 | #define SIZE_AC 1 | |
309 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
310 | ||
311 | /* | |
7243cc05 | 312 | * This function must be completely optimized away if |
e498be7d CL |
313 | * a constant is passed to it. Mostly the same as |
314 | * what is in linux/slab.h except it returns an | |
315 | * index. | |
316 | */ | |
7243cc05 | 317 | static __always_inline int index_of(const size_t size) |
e498be7d | 318 | { |
5ec8a847 SR |
319 | extern void __bad_size(void); |
320 | ||
e498be7d CL |
321 | if (__builtin_constant_p(size)) { |
322 | int i = 0; | |
323 | ||
324 | #define CACHE(x) \ | |
325 | if (size <=x) \ | |
326 | return i; \ | |
327 | else \ | |
328 | i++; | |
329 | #include "linux/kmalloc_sizes.h" | |
330 | #undef CACHE | |
5ec8a847 | 331 | __bad_size(); |
7243cc05 | 332 | } else |
5ec8a847 | 333 | __bad_size(); |
e498be7d CL |
334 | return 0; |
335 | } | |
336 | ||
337 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) | |
338 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 339 | |
e498be7d CL |
340 | static inline void kmem_list3_init(struct kmem_list3 *parent) |
341 | { | |
342 | INIT_LIST_HEAD(&parent->slabs_full); | |
343 | INIT_LIST_HEAD(&parent->slabs_partial); | |
344 | INIT_LIST_HEAD(&parent->slabs_free); | |
345 | parent->shared = NULL; | |
346 | parent->alien = NULL; | |
347 | spin_lock_init(&parent->list_lock); | |
348 | parent->free_objects = 0; | |
349 | parent->free_touched = 0; | |
350 | } | |
351 | ||
352 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ | |
353 | do { \ | |
354 | INIT_LIST_HEAD(listp); \ | |
355 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
356 | } while (0) | |
357 | ||
358 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ | |
359 | do { \ | |
360 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ | |
361 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
362 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
363 | } while (0) | |
1da177e4 LT |
364 | |
365 | /* | |
366 | * kmem_cache_t | |
367 | * | |
368 | * manages a cache. | |
369 | */ | |
b28a02de | 370 | |
2109a2d1 | 371 | struct kmem_cache { |
1da177e4 | 372 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de PE |
373 | struct array_cache *array[NR_CPUS]; |
374 | unsigned int batchcount; | |
375 | unsigned int limit; | |
376 | unsigned int shared; | |
3dafccf2 | 377 | unsigned int buffer_size; |
e498be7d | 378 | /* 2) touched by every alloc & free from the backend */ |
b28a02de PE |
379 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
380 | unsigned int flags; /* constant flags */ | |
381 | unsigned int num; /* # of objs per slab */ | |
382 | spinlock_t spinlock; | |
1da177e4 LT |
383 | |
384 | /* 3) cache_grow/shrink */ | |
385 | /* order of pgs per slab (2^n) */ | |
b28a02de | 386 | unsigned int gfporder; |
1da177e4 LT |
387 | |
388 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 389 | gfp_t gfpflags; |
1da177e4 | 390 | |
b28a02de PE |
391 | size_t colour; /* cache colouring range */ |
392 | unsigned int colour_off; /* colour offset */ | |
393 | unsigned int colour_next; /* cache colouring */ | |
394 | kmem_cache_t *slabp_cache; | |
395 | unsigned int slab_size; | |
396 | unsigned int dflags; /* dynamic flags */ | |
1da177e4 LT |
397 | |
398 | /* constructor func */ | |
b28a02de | 399 | void (*ctor) (void *, kmem_cache_t *, unsigned long); |
1da177e4 LT |
400 | |
401 | /* de-constructor func */ | |
b28a02de | 402 | void (*dtor) (void *, kmem_cache_t *, unsigned long); |
1da177e4 LT |
403 | |
404 | /* 4) cache creation/removal */ | |
b28a02de PE |
405 | const char *name; |
406 | struct list_head next; | |
1da177e4 LT |
407 | |
408 | /* 5) statistics */ | |
409 | #if STATS | |
b28a02de PE |
410 | unsigned long num_active; |
411 | unsigned long num_allocations; | |
412 | unsigned long high_mark; | |
413 | unsigned long grown; | |
414 | unsigned long reaped; | |
415 | unsigned long errors; | |
416 | unsigned long max_freeable; | |
417 | unsigned long node_allocs; | |
418 | unsigned long node_frees; | |
419 | atomic_t allochit; | |
420 | atomic_t allocmiss; | |
421 | atomic_t freehit; | |
422 | atomic_t freemiss; | |
1da177e4 LT |
423 | #endif |
424 | #if DEBUG | |
3dafccf2 MS |
425 | /* |
426 | * If debugging is enabled, then the allocator can add additional | |
427 | * fields and/or padding to every object. buffer_size contains the total | |
428 | * object size including these internal fields, the following two | |
429 | * variables contain the offset to the user object and its size. | |
430 | */ | |
431 | int obj_offset; | |
432 | int obj_size; | |
1da177e4 LT |
433 | #endif |
434 | }; | |
435 | ||
436 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
437 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
438 | ||
439 | #define BATCHREFILL_LIMIT 16 | |
440 | /* Optimization question: fewer reaps means less | |
441 | * probability for unnessary cpucache drain/refill cycles. | |
442 | * | |
dc6f3f27 | 443 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
444 | * which could lock up otherwise freeable slabs. |
445 | */ | |
446 | #define REAPTIMEOUT_CPUC (2*HZ) | |
447 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
448 | ||
449 | #if STATS | |
450 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
451 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
452 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
453 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
454 | #define STATS_INC_REAPED(x) ((x)->reaped++) | |
455 | #define STATS_SET_HIGH(x) do { if ((x)->num_active > (x)->high_mark) \ | |
456 | (x)->high_mark = (x)->num_active; \ | |
457 | } while (0) | |
458 | #define STATS_INC_ERR(x) ((x)->errors++) | |
459 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 460 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
1da177e4 LT |
461 | #define STATS_SET_FREEABLE(x, i) \ |
462 | do { if ((x)->max_freeable < i) \ | |
463 | (x)->max_freeable = i; \ | |
464 | } while (0) | |
465 | ||
466 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) | |
467 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
468 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
469 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
470 | #else | |
471 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
472 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
473 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
474 | #define STATS_INC_GROWN(x) do { } while (0) | |
475 | #define STATS_INC_REAPED(x) do { } while (0) | |
476 | #define STATS_SET_HIGH(x) do { } while (0) | |
477 | #define STATS_INC_ERR(x) do { } while (0) | |
478 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 479 | #define STATS_INC_NODEFREES(x) do { } while (0) |
1da177e4 LT |
480 | #define STATS_SET_FREEABLE(x, i) \ |
481 | do { } while (0) | |
482 | ||
483 | #define STATS_INC_ALLOCHIT(x) do { } while (0) | |
484 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
485 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
486 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
487 | #endif | |
488 | ||
489 | #if DEBUG | |
490 | /* Magic nums for obj red zoning. | |
491 | * Placed in the first word before and the first word after an obj. | |
492 | */ | |
493 | #define RED_INACTIVE 0x5A2CF071UL /* when obj is inactive */ | |
494 | #define RED_ACTIVE 0x170FC2A5UL /* when obj is active */ | |
495 | ||
496 | /* ...and for poisoning */ | |
497 | #define POISON_INUSE 0x5a /* for use-uninitialised poisoning */ | |
498 | #define POISON_FREE 0x6b /* for use-after-free poisoning */ | |
499 | #define POISON_END 0xa5 /* end-byte of poisoning */ | |
500 | ||
501 | /* memory layout of objects: | |
502 | * 0 : objp | |
3dafccf2 | 503 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
504 | * the end of an object is aligned with the end of the real |
505 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 506 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 507 | * redzone word. |
3dafccf2 MS |
508 | * cachep->obj_offset: The real object. |
509 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
510 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address [BYTES_PER_WORD long] | |
1da177e4 | 511 | */ |
3dafccf2 | 512 | static int obj_offset(kmem_cache_t *cachep) |
1da177e4 | 513 | { |
3dafccf2 | 514 | return cachep->obj_offset; |
1da177e4 LT |
515 | } |
516 | ||
3dafccf2 | 517 | static int obj_size(kmem_cache_t *cachep) |
1da177e4 | 518 | { |
3dafccf2 | 519 | return cachep->obj_size; |
1da177e4 LT |
520 | } |
521 | ||
522 | static unsigned long *dbg_redzone1(kmem_cache_t *cachep, void *objp) | |
523 | { | |
524 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
3dafccf2 | 525 | return (unsigned long*) (objp+obj_offset(cachep)-BYTES_PER_WORD); |
1da177e4 LT |
526 | } |
527 | ||
528 | static unsigned long *dbg_redzone2(kmem_cache_t *cachep, void *objp) | |
529 | { | |
530 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
531 | if (cachep->flags & SLAB_STORE_USER) | |
3dafccf2 | 532 | return (unsigned long *)(objp + cachep->buffer_size - |
b28a02de | 533 | 2 * BYTES_PER_WORD); |
3dafccf2 | 534 | return (unsigned long *)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
535 | } |
536 | ||
537 | static void **dbg_userword(kmem_cache_t *cachep, void *objp) | |
538 | { | |
539 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 540 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
541 | } |
542 | ||
543 | #else | |
544 | ||
3dafccf2 MS |
545 | #define obj_offset(x) 0 |
546 | #define obj_size(cachep) (cachep->buffer_size) | |
1da177e4 LT |
547 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long *)NULL;}) |
548 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long *)NULL;}) | |
549 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
550 | ||
551 | #endif | |
552 | ||
553 | /* | |
554 | * Maximum size of an obj (in 2^order pages) | |
555 | * and absolute limit for the gfp order. | |
556 | */ | |
557 | #if defined(CONFIG_LARGE_ALLOCS) | |
558 | #define MAX_OBJ_ORDER 13 /* up to 32Mb */ | |
559 | #define MAX_GFP_ORDER 13 /* up to 32Mb */ | |
560 | #elif defined(CONFIG_MMU) | |
561 | #define MAX_OBJ_ORDER 5 /* 32 pages */ | |
562 | #define MAX_GFP_ORDER 5 /* 32 pages */ | |
563 | #else | |
564 | #define MAX_OBJ_ORDER 8 /* up to 1Mb */ | |
565 | #define MAX_GFP_ORDER 8 /* up to 1Mb */ | |
566 | #endif | |
567 | ||
568 | /* | |
569 | * Do not go above this order unless 0 objects fit into the slab. | |
570 | */ | |
571 | #define BREAK_GFP_ORDER_HI 1 | |
572 | #define BREAK_GFP_ORDER_LO 0 | |
573 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
574 | ||
065d41cb | 575 | /* Functions for storing/retrieving the cachep and or slab from the |
1da177e4 LT |
576 | * global 'mem_map'. These are used to find the slab an obj belongs to. |
577 | * With kfree(), these are used to find the cache which an obj belongs to. | |
578 | */ | |
065d41cb PE |
579 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
580 | { | |
581 | page->lru.next = (struct list_head *)cache; | |
582 | } | |
583 | ||
584 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
585 | { | |
586 | return (struct kmem_cache *)page->lru.next; | |
587 | } | |
588 | ||
589 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
590 | { | |
591 | page->lru.prev = (struct list_head *)slab; | |
592 | } | |
593 | ||
594 | static inline struct slab *page_get_slab(struct page *page) | |
595 | { | |
596 | return (struct slab *)page->lru.prev; | |
597 | } | |
1da177e4 LT |
598 | |
599 | /* These are the default caches for kmalloc. Custom caches can have other sizes. */ | |
600 | struct cache_sizes malloc_sizes[] = { | |
601 | #define CACHE(x) { .cs_size = (x) }, | |
602 | #include <linux/kmalloc_sizes.h> | |
603 | CACHE(ULONG_MAX) | |
604 | #undef CACHE | |
605 | }; | |
606 | EXPORT_SYMBOL(malloc_sizes); | |
607 | ||
608 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
609 | struct cache_names { | |
610 | char *name; | |
611 | char *name_dma; | |
612 | }; | |
613 | ||
614 | static struct cache_names __initdata cache_names[] = { | |
615 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
616 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 617 | {NULL,} |
1da177e4 LT |
618 | #undef CACHE |
619 | }; | |
620 | ||
621 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 622 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 623 | static struct arraycache_init initarray_generic = |
b28a02de | 624 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
625 | |
626 | /* internal cache of cache description objs */ | |
627 | static kmem_cache_t cache_cache = { | |
b28a02de PE |
628 | .batchcount = 1, |
629 | .limit = BOOT_CPUCACHE_ENTRIES, | |
630 | .shared = 1, | |
3dafccf2 | 631 | .buffer_size = sizeof(kmem_cache_t), |
b28a02de PE |
632 | .flags = SLAB_NO_REAP, |
633 | .spinlock = SPIN_LOCK_UNLOCKED, | |
634 | .name = "kmem_cache", | |
1da177e4 | 635 | #if DEBUG |
3dafccf2 | 636 | .obj_size = sizeof(kmem_cache_t), |
1da177e4 LT |
637 | #endif |
638 | }; | |
639 | ||
640 | /* Guard access to the cache-chain. */ | |
fc0abb14 | 641 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
642 | static struct list_head cache_chain; |
643 | ||
644 | /* | |
645 | * vm_enough_memory() looks at this to determine how many | |
646 | * slab-allocated pages are possibly freeable under pressure | |
647 | * | |
648 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
649 | */ | |
650 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
651 | |
652 | /* | |
653 | * chicken and egg problem: delay the per-cpu array allocation | |
654 | * until the general caches are up. | |
655 | */ | |
656 | static enum { | |
657 | NONE, | |
e498be7d CL |
658 | PARTIAL_AC, |
659 | PARTIAL_L3, | |
1da177e4 LT |
660 | FULL |
661 | } g_cpucache_up; | |
662 | ||
663 | static DEFINE_PER_CPU(struct work_struct, reap_work); | |
664 | ||
b28a02de PE |
665 | static void free_block(kmem_cache_t *cachep, void **objpp, int len, int node); |
666 | static void enable_cpucache(kmem_cache_t *cachep); | |
667 | static void cache_reap(void *unused); | |
e498be7d | 668 | static int __node_shrink(kmem_cache_t *cachep, int node); |
1da177e4 LT |
669 | |
670 | static inline struct array_cache *ac_data(kmem_cache_t *cachep) | |
671 | { | |
672 | return cachep->array[smp_processor_id()]; | |
673 | } | |
674 | ||
dd0fc66f | 675 | static inline kmem_cache_t *__find_general_cachep(size_t size, gfp_t gfpflags) |
1da177e4 LT |
676 | { |
677 | struct cache_sizes *csizep = malloc_sizes; | |
678 | ||
679 | #if DEBUG | |
680 | /* This happens if someone tries to call | |
b28a02de PE |
681 | * kmem_cache_create(), or __kmalloc(), before |
682 | * the generic caches are initialized. | |
683 | */ | |
c7e43c78 | 684 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
685 | #endif |
686 | while (size > csizep->cs_size) | |
687 | csizep++; | |
688 | ||
689 | /* | |
0abf40c1 | 690 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
691 | * has cs_{dma,}cachep==NULL. Thus no special case |
692 | * for large kmalloc calls required. | |
693 | */ | |
694 | if (unlikely(gfpflags & GFP_DMA)) | |
695 | return csizep->cs_dmacachep; | |
696 | return csizep->cs_cachep; | |
697 | } | |
698 | ||
dd0fc66f | 699 | kmem_cache_t *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
700 | { |
701 | return __find_general_cachep(size, gfpflags); | |
702 | } | |
703 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
704 | ||
1da177e4 LT |
705 | /* Cal the num objs, wastage, and bytes left over for a given slab size. */ |
706 | static void cache_estimate(unsigned long gfporder, size_t size, size_t align, | |
b28a02de | 707 | int flags, size_t *left_over, unsigned int *num) |
1da177e4 LT |
708 | { |
709 | int i; | |
b28a02de | 710 | size_t wastage = PAGE_SIZE << gfporder; |
1da177e4 LT |
711 | size_t extra = 0; |
712 | size_t base = 0; | |
713 | ||
714 | if (!(flags & CFLGS_OFF_SLAB)) { | |
715 | base = sizeof(struct slab); | |
716 | extra = sizeof(kmem_bufctl_t); | |
717 | } | |
718 | i = 0; | |
b28a02de | 719 | while (i * size + ALIGN(base + i * extra, align) <= wastage) |
1da177e4 LT |
720 | i++; |
721 | if (i > 0) | |
722 | i--; | |
723 | ||
724 | if (i > SLAB_LIMIT) | |
725 | i = SLAB_LIMIT; | |
726 | ||
727 | *num = i; | |
b28a02de PE |
728 | wastage -= i * size; |
729 | wastage -= ALIGN(base + i * extra, align); | |
1da177e4 LT |
730 | *left_over = wastage; |
731 | } | |
732 | ||
733 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
734 | ||
735 | static void __slab_error(const char *function, kmem_cache_t *cachep, char *msg) | |
736 | { | |
737 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 738 | function, cachep->name, msg); |
1da177e4 LT |
739 | dump_stack(); |
740 | } | |
741 | ||
742 | /* | |
743 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
744 | * via the workqueue/eventd. | |
745 | * Add the CPU number into the expiration time to minimize the possibility of | |
746 | * the CPUs getting into lockstep and contending for the global cache chain | |
747 | * lock. | |
748 | */ | |
749 | static void __devinit start_cpu_timer(int cpu) | |
750 | { | |
751 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
752 | ||
753 | /* | |
754 | * When this gets called from do_initcalls via cpucache_init(), | |
755 | * init_workqueues() has already run, so keventd will be setup | |
756 | * at that time. | |
757 | */ | |
758 | if (keventd_up() && reap_work->func == NULL) { | |
759 | INIT_WORK(reap_work, cache_reap, NULL); | |
760 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
761 | } | |
762 | } | |
763 | ||
e498be7d | 764 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 765 | int batchcount) |
1da177e4 | 766 | { |
b28a02de | 767 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
768 | struct array_cache *nc = NULL; |
769 | ||
e498be7d | 770 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
771 | if (nc) { |
772 | nc->avail = 0; | |
773 | nc->limit = entries; | |
774 | nc->batchcount = batchcount; | |
775 | nc->touched = 0; | |
e498be7d | 776 | spin_lock_init(&nc->lock); |
1da177e4 LT |
777 | } |
778 | return nc; | |
779 | } | |
780 | ||
e498be7d | 781 | #ifdef CONFIG_NUMA |
dc85da15 CL |
782 | static void *__cache_alloc_node(kmem_cache_t *, gfp_t, int); |
783 | ||
e498be7d CL |
784 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
785 | { | |
786 | struct array_cache **ac_ptr; | |
b28a02de | 787 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
788 | int i; |
789 | ||
790 | if (limit > 1) | |
791 | limit = 12; | |
792 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
793 | if (ac_ptr) { | |
794 | for_each_node(i) { | |
795 | if (i == node || !node_online(i)) { | |
796 | ac_ptr[i] = NULL; | |
797 | continue; | |
798 | } | |
799 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
800 | if (!ac_ptr[i]) { | |
b28a02de | 801 | for (i--; i <= 0; i--) |
e498be7d CL |
802 | kfree(ac_ptr[i]); |
803 | kfree(ac_ptr); | |
804 | return NULL; | |
805 | } | |
806 | } | |
807 | } | |
808 | return ac_ptr; | |
809 | } | |
810 | ||
811 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
812 | { | |
813 | int i; | |
814 | ||
815 | if (!ac_ptr) | |
816 | return; | |
817 | ||
818 | for_each_node(i) | |
b28a02de | 819 | kfree(ac_ptr[i]); |
e498be7d CL |
820 | |
821 | kfree(ac_ptr); | |
822 | } | |
823 | ||
b28a02de PE |
824 | static inline void __drain_alien_cache(kmem_cache_t *cachep, |
825 | struct array_cache *ac, int node) | |
e498be7d CL |
826 | { |
827 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
828 | ||
829 | if (ac->avail) { | |
830 | spin_lock(&rl3->list_lock); | |
ff69416e | 831 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
832 | ac->avail = 0; |
833 | spin_unlock(&rl3->list_lock); | |
834 | } | |
835 | } | |
836 | ||
837 | static void drain_alien_cache(kmem_cache_t *cachep, struct kmem_list3 *l3) | |
838 | { | |
b28a02de | 839 | int i = 0; |
e498be7d CL |
840 | struct array_cache *ac; |
841 | unsigned long flags; | |
842 | ||
843 | for_each_online_node(i) { | |
844 | ac = l3->alien[i]; | |
845 | if (ac) { | |
846 | spin_lock_irqsave(&ac->lock, flags); | |
847 | __drain_alien_cache(cachep, ac, i); | |
848 | spin_unlock_irqrestore(&ac->lock, flags); | |
849 | } | |
850 | } | |
851 | } | |
852 | #else | |
853 | #define alloc_alien_cache(node, limit) do { } while (0) | |
854 | #define free_alien_cache(ac_ptr) do { } while (0) | |
855 | #define drain_alien_cache(cachep, l3) do { } while (0) | |
856 | #endif | |
857 | ||
1da177e4 | 858 | static int __devinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 859 | unsigned long action, void *hcpu) |
1da177e4 LT |
860 | { |
861 | long cpu = (long)hcpu; | |
b28a02de | 862 | kmem_cache_t *cachep; |
e498be7d CL |
863 | struct kmem_list3 *l3 = NULL; |
864 | int node = cpu_to_node(cpu); | |
865 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
866 | |
867 | switch (action) { | |
868 | case CPU_UP_PREPARE: | |
fc0abb14 | 869 | mutex_lock(&cache_chain_mutex); |
e498be7d CL |
870 | /* we need to do this right in the beginning since |
871 | * alloc_arraycache's are going to use this list. | |
872 | * kmalloc_node allows us to add the slab to the right | |
873 | * kmem_list3 and not this cpu's kmem_list3 | |
874 | */ | |
875 | ||
1da177e4 | 876 | list_for_each_entry(cachep, &cache_chain, next) { |
e498be7d CL |
877 | /* setup the size64 kmemlist for cpu before we can |
878 | * begin anything. Make sure some other cpu on this | |
879 | * node has not already allocated this | |
880 | */ | |
881 | if (!cachep->nodelists[node]) { | |
882 | if (!(l3 = kmalloc_node(memsize, | |
b28a02de | 883 | GFP_KERNEL, node))) |
e498be7d CL |
884 | goto bad; |
885 | kmem_list3_init(l3); | |
886 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 887 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
888 | |
889 | cachep->nodelists[node] = l3; | |
890 | } | |
1da177e4 | 891 | |
e498be7d CL |
892 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
893 | cachep->nodelists[node]->free_limit = | |
b28a02de PE |
894 | (1 + nr_cpus_node(node)) * |
895 | cachep->batchcount + cachep->num; | |
e498be7d CL |
896 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
897 | } | |
898 | ||
899 | /* Now we can go ahead with allocating the shared array's | |
b28a02de | 900 | & array cache's */ |
e498be7d | 901 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 TK |
902 | struct array_cache *nc; |
903 | ||
e498be7d | 904 | nc = alloc_arraycache(node, cachep->limit, |
b28a02de | 905 | cachep->batchcount); |
1da177e4 LT |
906 | if (!nc) |
907 | goto bad; | |
1da177e4 | 908 | cachep->array[cpu] = nc; |
1da177e4 | 909 | |
e498be7d CL |
910 | l3 = cachep->nodelists[node]; |
911 | BUG_ON(!l3); | |
912 | if (!l3->shared) { | |
913 | if (!(nc = alloc_arraycache(node, | |
b28a02de PE |
914 | cachep->shared * |
915 | cachep->batchcount, | |
916 | 0xbaadf00d))) | |
917 | goto bad; | |
e498be7d CL |
918 | |
919 | /* we are serialised from CPU_DEAD or | |
b28a02de | 920 | CPU_UP_CANCELLED by the cpucontrol lock */ |
e498be7d CL |
921 | l3->shared = nc; |
922 | } | |
1da177e4 | 923 | } |
fc0abb14 | 924 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
925 | break; |
926 | case CPU_ONLINE: | |
927 | start_cpu_timer(cpu); | |
928 | break; | |
929 | #ifdef CONFIG_HOTPLUG_CPU | |
930 | case CPU_DEAD: | |
931 | /* fall thru */ | |
932 | case CPU_UP_CANCELED: | |
fc0abb14 | 933 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
934 | |
935 | list_for_each_entry(cachep, &cache_chain, next) { | |
936 | struct array_cache *nc; | |
e498be7d | 937 | cpumask_t mask; |
1da177e4 | 938 | |
e498be7d | 939 | mask = node_to_cpumask(node); |
1da177e4 LT |
940 | spin_lock_irq(&cachep->spinlock); |
941 | /* cpu is dead; no one can alloc from it. */ | |
942 | nc = cachep->array[cpu]; | |
943 | cachep->array[cpu] = NULL; | |
e498be7d CL |
944 | l3 = cachep->nodelists[node]; |
945 | ||
946 | if (!l3) | |
947 | goto unlock_cache; | |
948 | ||
949 | spin_lock(&l3->list_lock); | |
950 | ||
951 | /* Free limit for this kmem_list3 */ | |
952 | l3->free_limit -= cachep->batchcount; | |
953 | if (nc) | |
ff69416e | 954 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
955 | |
956 | if (!cpus_empty(mask)) { | |
b28a02de PE |
957 | spin_unlock(&l3->list_lock); |
958 | goto unlock_cache; | |
959 | } | |
e498be7d CL |
960 | |
961 | if (l3->shared) { | |
962 | free_block(cachep, l3->shared->entry, | |
b28a02de | 963 | l3->shared->avail, node); |
e498be7d CL |
964 | kfree(l3->shared); |
965 | l3->shared = NULL; | |
966 | } | |
967 | if (l3->alien) { | |
968 | drain_alien_cache(cachep, l3); | |
969 | free_alien_cache(l3->alien); | |
970 | l3->alien = NULL; | |
971 | } | |
972 | ||
973 | /* free slabs belonging to this node */ | |
974 | if (__node_shrink(cachep, node)) { | |
975 | cachep->nodelists[node] = NULL; | |
976 | spin_unlock(&l3->list_lock); | |
977 | kfree(l3); | |
978 | } else { | |
979 | spin_unlock(&l3->list_lock); | |
980 | } | |
b28a02de | 981 | unlock_cache: |
1da177e4 LT |
982 | spin_unlock_irq(&cachep->spinlock); |
983 | kfree(nc); | |
984 | } | |
fc0abb14 | 985 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
986 | break; |
987 | #endif | |
988 | } | |
989 | return NOTIFY_OK; | |
b28a02de | 990 | bad: |
fc0abb14 | 991 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
992 | return NOTIFY_BAD; |
993 | } | |
994 | ||
995 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
996 | ||
e498be7d CL |
997 | /* |
998 | * swap the static kmem_list3 with kmalloced memory | |
999 | */ | |
b28a02de | 1000 | static void init_list(kmem_cache_t *cachep, struct kmem_list3 *list, int nodeid) |
e498be7d CL |
1001 | { |
1002 | struct kmem_list3 *ptr; | |
1003 | ||
1004 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1005 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1006 | BUG_ON(!ptr); | |
1007 | ||
1008 | local_irq_disable(); | |
1009 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1010 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1011 | cachep->nodelists[nodeid] = ptr; | |
1012 | local_irq_enable(); | |
1013 | } | |
1014 | ||
1da177e4 LT |
1015 | /* Initialisation. |
1016 | * Called after the gfp() functions have been enabled, and before smp_init(). | |
1017 | */ | |
1018 | void __init kmem_cache_init(void) | |
1019 | { | |
1020 | size_t left_over; | |
1021 | struct cache_sizes *sizes; | |
1022 | struct cache_names *names; | |
e498be7d CL |
1023 | int i; |
1024 | ||
1025 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1026 | kmem_list3_init(&initkmem_list3[i]); | |
1027 | if (i < MAX_NUMNODES) | |
1028 | cache_cache.nodelists[i] = NULL; | |
1029 | } | |
1da177e4 LT |
1030 | |
1031 | /* | |
1032 | * Fragmentation resistance on low memory - only use bigger | |
1033 | * page orders on machines with more than 32MB of memory. | |
1034 | */ | |
1035 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1036 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1037 | ||
1da177e4 LT |
1038 | /* Bootstrap is tricky, because several objects are allocated |
1039 | * from caches that do not exist yet: | |
1040 | * 1) initialize the cache_cache cache: it contains the kmem_cache_t | |
1041 | * structures of all caches, except cache_cache itself: cache_cache | |
1042 | * is statically allocated. | |
e498be7d CL |
1043 | * Initially an __init data area is used for the head array and the |
1044 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1045 | * array at the end of the bootstrap. | |
1da177e4 | 1046 | * 2) Create the first kmalloc cache. |
e498be7d CL |
1047 | * The kmem_cache_t for the new cache is allocated normally. |
1048 | * An __init data area is used for the head array. | |
1049 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1050 | * head arrays. | |
1da177e4 LT |
1051 | * 4) Replace the __init data head arrays for cache_cache and the first |
1052 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1053 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1054 | * the other cache's with kmalloc allocated memory. | |
1055 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1056 | */ |
1057 | ||
1058 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1059 | INIT_LIST_HEAD(&cache_chain); |
1060 | list_add(&cache_cache.next, &cache_chain); | |
1061 | cache_cache.colour_off = cache_line_size(); | |
1062 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1063 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1064 | |
3dafccf2 | 1065 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size()); |
1da177e4 | 1066 | |
3dafccf2 | 1067 | cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0, |
b28a02de | 1068 | &left_over, &cache_cache.num); |
1da177e4 LT |
1069 | if (!cache_cache.num) |
1070 | BUG(); | |
1071 | ||
b28a02de | 1072 | cache_cache.colour = left_over / cache_cache.colour_off; |
1da177e4 | 1073 | cache_cache.colour_next = 0; |
b28a02de PE |
1074 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1075 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1076 | |
1077 | /* 2+3) create the kmalloc caches */ | |
1078 | sizes = malloc_sizes; | |
1079 | names = cache_names; | |
1080 | ||
e498be7d CL |
1081 | /* Initialize the caches that provide memory for the array cache |
1082 | * and the kmem_list3 structures first. | |
1083 | * Without this, further allocations will bug | |
1084 | */ | |
1085 | ||
1086 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
b28a02de PE |
1087 | sizes[INDEX_AC].cs_size, |
1088 | ARCH_KMALLOC_MINALIGN, | |
1089 | (ARCH_KMALLOC_FLAGS | | |
1090 | SLAB_PANIC), NULL, NULL); | |
e498be7d CL |
1091 | |
1092 | if (INDEX_AC != INDEX_L3) | |
1093 | sizes[INDEX_L3].cs_cachep = | |
b28a02de PE |
1094 | kmem_cache_create(names[INDEX_L3].name, |
1095 | sizes[INDEX_L3].cs_size, | |
1096 | ARCH_KMALLOC_MINALIGN, | |
1097 | (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, | |
1098 | NULL); | |
e498be7d | 1099 | |
1da177e4 | 1100 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1101 | /* |
1102 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1103 | * This should be particularly beneficial on SMP boxes, as it |
1104 | * eliminates "false sharing". | |
1105 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1106 | * allow tighter packing of the smaller caches. |
1107 | */ | |
b28a02de | 1108 | if (!sizes->cs_cachep) |
e498be7d | 1109 | sizes->cs_cachep = kmem_cache_create(names->name, |
b28a02de PE |
1110 | sizes->cs_size, |
1111 | ARCH_KMALLOC_MINALIGN, | |
1112 | (ARCH_KMALLOC_FLAGS | |
1113 | | SLAB_PANIC), | |
1114 | NULL, NULL); | |
1da177e4 LT |
1115 | |
1116 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
1117 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
b28a02de | 1118 | offslab_limit = sizes->cs_size - sizeof(struct slab); |
1da177e4 LT |
1119 | offslab_limit /= sizeof(kmem_bufctl_t); |
1120 | } | |
1121 | ||
1122 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
b28a02de PE |
1123 | sizes->cs_size, |
1124 | ARCH_KMALLOC_MINALIGN, | |
1125 | (ARCH_KMALLOC_FLAGS | | |
1126 | SLAB_CACHE_DMA | | |
1127 | SLAB_PANIC), NULL, | |
1128 | NULL); | |
1da177e4 LT |
1129 | |
1130 | sizes++; | |
1131 | names++; | |
1132 | } | |
1133 | /* 4) Replace the bootstrap head arrays */ | |
1134 | { | |
b28a02de | 1135 | void *ptr; |
e498be7d | 1136 | |
1da177e4 | 1137 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1138 | |
1da177e4 LT |
1139 | local_irq_disable(); |
1140 | BUG_ON(ac_data(&cache_cache) != &initarray_cache.cache); | |
e498be7d | 1141 | memcpy(ptr, ac_data(&cache_cache), |
b28a02de | 1142 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1143 | cache_cache.array[smp_processor_id()] = ptr; |
1144 | local_irq_enable(); | |
e498be7d | 1145 | |
1da177e4 | 1146 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1147 | |
1da177e4 | 1148 | local_irq_disable(); |
e498be7d | 1149 | BUG_ON(ac_data(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1150 | != &initarray_generic.cache); |
e498be7d | 1151 | memcpy(ptr, ac_data(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1152 | sizeof(struct arraycache_init)); |
e498be7d | 1153 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1154 | ptr; |
1da177e4 LT |
1155 | local_irq_enable(); |
1156 | } | |
e498be7d CL |
1157 | /* 5) Replace the bootstrap kmem_list3's */ |
1158 | { | |
1159 | int node; | |
1160 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1161 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1162 | numa_node_id()); |
e498be7d CL |
1163 | |
1164 | for_each_online_node(node) { | |
1165 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1166 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1167 | |
1168 | if (INDEX_AC != INDEX_L3) { | |
1169 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1170 | &initkmem_list3[SIZE_L3 + node], |
1171 | node); | |
e498be7d CL |
1172 | } |
1173 | } | |
1174 | } | |
1da177e4 | 1175 | |
e498be7d | 1176 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 LT |
1177 | { |
1178 | kmem_cache_t *cachep; | |
fc0abb14 | 1179 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1180 | list_for_each_entry(cachep, &cache_chain, next) |
b28a02de | 1181 | enable_cpucache(cachep); |
fc0abb14 | 1182 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1183 | } |
1184 | ||
1185 | /* Done! */ | |
1186 | g_cpucache_up = FULL; | |
1187 | ||
1188 | /* Register a cpu startup notifier callback | |
1189 | * that initializes ac_data for all new cpus | |
1190 | */ | |
1191 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 LT |
1192 | |
1193 | /* The reap timers are started later, with a module init call: | |
1194 | * That part of the kernel is not yet operational. | |
1195 | */ | |
1196 | } | |
1197 | ||
1198 | static int __init cpucache_init(void) | |
1199 | { | |
1200 | int cpu; | |
1201 | ||
1202 | /* | |
1203 | * Register the timers that return unneeded | |
1204 | * pages to gfp. | |
1205 | */ | |
e498be7d | 1206 | for_each_online_cpu(cpu) |
b28a02de | 1207 | start_cpu_timer(cpu); |
1da177e4 LT |
1208 | |
1209 | return 0; | |
1210 | } | |
1211 | ||
1212 | __initcall(cpucache_init); | |
1213 | ||
1214 | /* | |
1215 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1216 | * | |
1217 | * If we requested dmaable memory, we will get it. Even if we | |
1218 | * did not request dmaable memory, we might get it, but that | |
1219 | * would be relatively rare and ignorable. | |
1220 | */ | |
dd0fc66f | 1221 | static void *kmem_getpages(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1222 | { |
1223 | struct page *page; | |
1224 | void *addr; | |
1225 | int i; | |
1226 | ||
1227 | flags |= cachep->gfpflags; | |
50c85a19 | 1228 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); |
1da177e4 LT |
1229 | if (!page) |
1230 | return NULL; | |
1231 | addr = page_address(page); | |
1232 | ||
1233 | i = (1 << cachep->gfporder); | |
1234 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1235 | atomic_add(i, &slab_reclaim_pages); | |
1236 | add_page_state(nr_slab, i); | |
1237 | while (i--) { | |
1238 | SetPageSlab(page); | |
1239 | page++; | |
1240 | } | |
1241 | return addr; | |
1242 | } | |
1243 | ||
1244 | /* | |
1245 | * Interface to system's page release. | |
1246 | */ | |
1247 | static void kmem_freepages(kmem_cache_t *cachep, void *addr) | |
1248 | { | |
b28a02de | 1249 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1250 | struct page *page = virt_to_page(addr); |
1251 | const unsigned long nr_freed = i; | |
1252 | ||
1253 | while (i--) { | |
1254 | if (!TestClearPageSlab(page)) | |
1255 | BUG(); | |
1256 | page++; | |
1257 | } | |
1258 | sub_page_state(nr_slab, nr_freed); | |
1259 | if (current->reclaim_state) | |
1260 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1261 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1262 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1263 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1264 | } |
1265 | ||
1266 | static void kmem_rcu_free(struct rcu_head *head) | |
1267 | { | |
b28a02de | 1268 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
1da177e4 LT |
1269 | kmem_cache_t *cachep = slab_rcu->cachep; |
1270 | ||
1271 | kmem_freepages(cachep, slab_rcu->addr); | |
1272 | if (OFF_SLAB(cachep)) | |
1273 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1274 | } | |
1275 | ||
1276 | #if DEBUG | |
1277 | ||
1278 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
1279 | static void store_stackinfo(kmem_cache_t *cachep, unsigned long *addr, | |
b28a02de | 1280 | unsigned long caller) |
1da177e4 | 1281 | { |
3dafccf2 | 1282 | int size = obj_size(cachep); |
1da177e4 | 1283 | |
3dafccf2 | 1284 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1285 | |
b28a02de | 1286 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1287 | return; |
1288 | ||
b28a02de PE |
1289 | *addr++ = 0x12345678; |
1290 | *addr++ = caller; | |
1291 | *addr++ = smp_processor_id(); | |
1292 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1293 | { |
1294 | unsigned long *sptr = &caller; | |
1295 | unsigned long svalue; | |
1296 | ||
1297 | while (!kstack_end(sptr)) { | |
1298 | svalue = *sptr++; | |
1299 | if (kernel_text_address(svalue)) { | |
b28a02de | 1300 | *addr++ = svalue; |
1da177e4 LT |
1301 | size -= sizeof(unsigned long); |
1302 | if (size <= sizeof(unsigned long)) | |
1303 | break; | |
1304 | } | |
1305 | } | |
1306 | ||
1307 | } | |
b28a02de | 1308 | *addr++ = 0x87654321; |
1da177e4 LT |
1309 | } |
1310 | #endif | |
1311 | ||
1312 | static void poison_obj(kmem_cache_t *cachep, void *addr, unsigned char val) | |
1313 | { | |
3dafccf2 MS |
1314 | int size = obj_size(cachep); |
1315 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1316 | |
1317 | memset(addr, val, size); | |
b28a02de | 1318 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1319 | } |
1320 | ||
1321 | static void dump_line(char *data, int offset, int limit) | |
1322 | { | |
1323 | int i; | |
1324 | printk(KERN_ERR "%03x:", offset); | |
b28a02de PE |
1325 | for (i = 0; i < limit; i++) { |
1326 | printk(" %02x", (unsigned char)data[offset + i]); | |
1da177e4 LT |
1327 | } |
1328 | printk("\n"); | |
1329 | } | |
1330 | #endif | |
1331 | ||
1332 | #if DEBUG | |
1333 | ||
1334 | static void print_objinfo(kmem_cache_t *cachep, void *objp, int lines) | |
1335 | { | |
1336 | int i, size; | |
1337 | char *realobj; | |
1338 | ||
1339 | if (cachep->flags & SLAB_RED_ZONE) { | |
1340 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
b28a02de PE |
1341 | *dbg_redzone1(cachep, objp), |
1342 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1343 | } |
1344 | ||
1345 | if (cachep->flags & SLAB_STORE_USER) { | |
1346 | printk(KERN_ERR "Last user: [<%p>]", | |
b28a02de | 1347 | *dbg_userword(cachep, objp)); |
1da177e4 | 1348 | print_symbol("(%s)", |
b28a02de | 1349 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1350 | printk("\n"); |
1351 | } | |
3dafccf2 MS |
1352 | realobj = (char *)objp + obj_offset(cachep); |
1353 | size = obj_size(cachep); | |
b28a02de | 1354 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1355 | int limit; |
1356 | limit = 16; | |
b28a02de PE |
1357 | if (i + limit > size) |
1358 | limit = size - i; | |
1da177e4 LT |
1359 | dump_line(realobj, i, limit); |
1360 | } | |
1361 | } | |
1362 | ||
1363 | static void check_poison_obj(kmem_cache_t *cachep, void *objp) | |
1364 | { | |
1365 | char *realobj; | |
1366 | int size, i; | |
1367 | int lines = 0; | |
1368 | ||
3dafccf2 MS |
1369 | realobj = (char *)objp + obj_offset(cachep); |
1370 | size = obj_size(cachep); | |
1da177e4 | 1371 | |
b28a02de | 1372 | for (i = 0; i < size; i++) { |
1da177e4 | 1373 | char exp = POISON_FREE; |
b28a02de | 1374 | if (i == size - 1) |
1da177e4 LT |
1375 | exp = POISON_END; |
1376 | if (realobj[i] != exp) { | |
1377 | int limit; | |
1378 | /* Mismatch ! */ | |
1379 | /* Print header */ | |
1380 | if (lines == 0) { | |
b28a02de PE |
1381 | printk(KERN_ERR |
1382 | "Slab corruption: start=%p, len=%d\n", | |
1383 | realobj, size); | |
1da177e4 LT |
1384 | print_objinfo(cachep, objp, 0); |
1385 | } | |
1386 | /* Hexdump the affected line */ | |
b28a02de | 1387 | i = (i / 16) * 16; |
1da177e4 | 1388 | limit = 16; |
b28a02de PE |
1389 | if (i + limit > size) |
1390 | limit = size - i; | |
1da177e4 LT |
1391 | dump_line(realobj, i, limit); |
1392 | i += 16; | |
1393 | lines++; | |
1394 | /* Limit to 5 lines */ | |
1395 | if (lines > 5) | |
1396 | break; | |
1397 | } | |
1398 | } | |
1399 | if (lines != 0) { | |
1400 | /* Print some data about the neighboring objects, if they | |
1401 | * exist: | |
1402 | */ | |
065d41cb | 1403 | struct slab *slabp = page_get_slab(virt_to_page(objp)); |
1da177e4 LT |
1404 | int objnr; |
1405 | ||
3dafccf2 | 1406 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
1da177e4 | 1407 | if (objnr) { |
3dafccf2 MS |
1408 | objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size; |
1409 | realobj = (char *)objp + obj_offset(cachep); | |
1da177e4 | 1410 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1411 | realobj, size); |
1da177e4 LT |
1412 | print_objinfo(cachep, objp, 2); |
1413 | } | |
b28a02de | 1414 | if (objnr + 1 < cachep->num) { |
3dafccf2 MS |
1415 | objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size; |
1416 | realobj = (char *)objp + obj_offset(cachep); | |
1da177e4 | 1417 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1418 | realobj, size); |
1da177e4 LT |
1419 | print_objinfo(cachep, objp, 2); |
1420 | } | |
1421 | } | |
1422 | } | |
1423 | #endif | |
1424 | ||
1425 | /* Destroy all the objs in a slab, and release the mem back to the system. | |
1426 | * Before calling the slab must have been unlinked from the cache. | |
1427 | * The cache-lock is not held/needed. | |
1428 | */ | |
b28a02de | 1429 | static void slab_destroy(kmem_cache_t *cachep, struct slab *slabp) |
1da177e4 LT |
1430 | { |
1431 | void *addr = slabp->s_mem - slabp->colouroff; | |
1432 | ||
1433 | #if DEBUG | |
1434 | int i; | |
1435 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 1436 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
1da177e4 LT |
1437 | |
1438 | if (cachep->flags & SLAB_POISON) { | |
1439 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 1440 | if ((cachep->buffer_size % PAGE_SIZE) == 0 |
b28a02de PE |
1441 | && OFF_SLAB(cachep)) |
1442 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 1443 | cachep->buffer_size / PAGE_SIZE, |
b28a02de | 1444 | 1); |
1da177e4 LT |
1445 | else |
1446 | check_poison_obj(cachep, objp); | |
1447 | #else | |
1448 | check_poison_obj(cachep, objp); | |
1449 | #endif | |
1450 | } | |
1451 | if (cachep->flags & SLAB_RED_ZONE) { | |
1452 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1453 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1454 | "was overwritten"); |
1da177e4 LT |
1455 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1456 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1457 | "was overwritten"); |
1da177e4 LT |
1458 | } |
1459 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1460 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 LT |
1461 | } |
1462 | #else | |
1463 | if (cachep->dtor) { | |
1464 | int i; | |
1465 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 1466 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
b28a02de | 1467 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1468 | } |
1469 | } | |
1470 | #endif | |
1471 | ||
1472 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { | |
1473 | struct slab_rcu *slab_rcu; | |
1474 | ||
b28a02de | 1475 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1476 | slab_rcu->cachep = cachep; |
1477 | slab_rcu->addr = addr; | |
1478 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1479 | } else { | |
1480 | kmem_freepages(cachep, addr); | |
1481 | if (OFF_SLAB(cachep)) | |
1482 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1483 | } | |
1484 | } | |
1485 | ||
3dafccf2 | 1486 | /* For setting up all the kmem_list3s for cache whose buffer_size is same |
e498be7d CL |
1487 | as size of kmem_list3. */ |
1488 | static inline void set_up_list3s(kmem_cache_t *cachep, int index) | |
1489 | { | |
1490 | int node; | |
1491 | ||
1492 | for_each_online_node(node) { | |
b28a02de | 1493 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1494 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1495 | REAPTIMEOUT_LIST3 + |
1496 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1497 | } |
1498 | } | |
1499 | ||
4d268eba PE |
1500 | /** |
1501 | * calculate_slab_order - calculate size (page order) of slabs and the number | |
1502 | * of objects per slab. | |
1503 | * | |
1504 | * This could be made much more intelligent. For now, try to avoid using | |
1505 | * high order pages for slabs. When the gfp() functions are more friendly | |
1506 | * towards high-order requests, this should be changed. | |
1507 | */ | |
1508 | static inline size_t calculate_slab_order(kmem_cache_t *cachep, size_t size, | |
1509 | size_t align, gfp_t flags) | |
1510 | { | |
1511 | size_t left_over = 0; | |
1512 | ||
b28a02de | 1513 | for (;; cachep->gfporder++) { |
4d268eba PE |
1514 | unsigned int num; |
1515 | size_t remainder; | |
1516 | ||
1517 | if (cachep->gfporder > MAX_GFP_ORDER) { | |
1518 | cachep->num = 0; | |
1519 | break; | |
1520 | } | |
1521 | ||
1522 | cache_estimate(cachep->gfporder, size, align, flags, | |
1523 | &remainder, &num); | |
1524 | if (!num) | |
1525 | continue; | |
1526 | /* More than offslab_limit objects will cause problems */ | |
1527 | if (flags & CFLGS_OFF_SLAB && cachep->num > offslab_limit) | |
1528 | break; | |
1529 | ||
1530 | cachep->num = num; | |
1531 | left_over = remainder; | |
1532 | ||
1533 | /* | |
1534 | * Large number of objects is good, but very large slabs are | |
1535 | * currently bad for the gfp()s. | |
1536 | */ | |
1537 | if (cachep->gfporder >= slab_break_gfp_order) | |
1538 | break; | |
1539 | ||
1540 | if ((left_over * 8) <= (PAGE_SIZE << cachep->gfporder)) | |
1541 | /* Acceptable internal fragmentation */ | |
1542 | break; | |
1543 | } | |
1544 | return left_over; | |
1545 | } | |
1546 | ||
1da177e4 LT |
1547 | /** |
1548 | * kmem_cache_create - Create a cache. | |
1549 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1550 | * @size: The size of objects to be created in this cache. | |
1551 | * @align: The required alignment for the objects. | |
1552 | * @flags: SLAB flags | |
1553 | * @ctor: A constructor for the objects. | |
1554 | * @dtor: A destructor for the objects. | |
1555 | * | |
1556 | * Returns a ptr to the cache on success, NULL on failure. | |
1557 | * Cannot be called within a int, but can be interrupted. | |
1558 | * The @ctor is run when new pages are allocated by the cache | |
1559 | * and the @dtor is run before the pages are handed back. | |
1560 | * | |
1561 | * @name must be valid until the cache is destroyed. This implies that | |
1562 | * the module calling this has to destroy the cache before getting | |
1563 | * unloaded. | |
1564 | * | |
1565 | * The flags are | |
1566 | * | |
1567 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1568 | * to catch references to uninitialised memory. | |
1569 | * | |
1570 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1571 | * for buffer overruns. | |
1572 | * | |
1573 | * %SLAB_NO_REAP - Don't automatically reap this cache when we're under | |
1574 | * memory pressure. | |
1575 | * | |
1576 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
1577 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1578 | * as davem. | |
1579 | */ | |
1580 | kmem_cache_t * | |
1581 | kmem_cache_create (const char *name, size_t size, size_t align, | |
1582 | unsigned long flags, void (*ctor)(void*, kmem_cache_t *, unsigned long), | |
1583 | void (*dtor)(void*, kmem_cache_t *, unsigned long)) | |
1584 | { | |
1585 | size_t left_over, slab_size, ralign; | |
1586 | kmem_cache_t *cachep = NULL; | |
4f12bb4f | 1587 | struct list_head *p; |
1da177e4 LT |
1588 | |
1589 | /* | |
1590 | * Sanity checks... these are all serious usage bugs. | |
1591 | */ | |
1592 | if ((!name) || | |
b28a02de PE |
1593 | in_interrupt() || |
1594 | (size < BYTES_PER_WORD) || | |
1595 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { | |
1596 | printk(KERN_ERR "%s: Early error in slab %s\n", | |
1597 | __FUNCTION__, name); | |
1598 | BUG(); | |
1599 | } | |
1da177e4 | 1600 | |
fc0abb14 | 1601 | mutex_lock(&cache_chain_mutex); |
4f12bb4f AM |
1602 | |
1603 | list_for_each(p, &cache_chain) { | |
1604 | kmem_cache_t *pc = list_entry(p, kmem_cache_t, next); | |
1605 | mm_segment_t old_fs = get_fs(); | |
1606 | char tmp; | |
1607 | int res; | |
1608 | ||
1609 | /* | |
1610 | * This happens when the module gets unloaded and doesn't | |
1611 | * destroy its slab cache and no-one else reuses the vmalloc | |
1612 | * area of the module. Print a warning. | |
1613 | */ | |
1614 | set_fs(KERNEL_DS); | |
1615 | res = __get_user(tmp, pc->name); | |
1616 | set_fs(old_fs); | |
1617 | if (res) { | |
1618 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 1619 | pc->buffer_size); |
4f12bb4f AM |
1620 | continue; |
1621 | } | |
1622 | ||
b28a02de | 1623 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1624 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1625 | dump_stack(); | |
1626 | goto oops; | |
1627 | } | |
1628 | } | |
1629 | ||
1da177e4 LT |
1630 | #if DEBUG |
1631 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1632 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1633 | /* No constructor, but inital state check requested */ | |
1634 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 1635 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
1636 | flags &= ~SLAB_DEBUG_INITIAL; |
1637 | } | |
1da177e4 LT |
1638 | #if FORCED_DEBUG |
1639 | /* | |
1640 | * Enable redzoning and last user accounting, except for caches with | |
1641 | * large objects, if the increased size would increase the object size | |
1642 | * above the next power of two: caches with object sizes just above a | |
1643 | * power of two have a significant amount of internal fragmentation. | |
1644 | */ | |
b28a02de PE |
1645 | if ((size < 4096 |
1646 | || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD))) | |
1647 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; | |
1da177e4 LT |
1648 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
1649 | flags |= SLAB_POISON; | |
1650 | #endif | |
1651 | if (flags & SLAB_DESTROY_BY_RCU) | |
1652 | BUG_ON(flags & SLAB_POISON); | |
1653 | #endif | |
1654 | if (flags & SLAB_DESTROY_BY_RCU) | |
1655 | BUG_ON(dtor); | |
1656 | ||
1657 | /* | |
1658 | * Always checks flags, a caller might be expecting debug | |
1659 | * support which isn't available. | |
1660 | */ | |
1661 | if (flags & ~CREATE_MASK) | |
1662 | BUG(); | |
1663 | ||
1664 | /* Check that size is in terms of words. This is needed to avoid | |
1665 | * unaligned accesses for some archs when redzoning is used, and makes | |
1666 | * sure any on-slab bufctl's are also correctly aligned. | |
1667 | */ | |
b28a02de PE |
1668 | if (size & (BYTES_PER_WORD - 1)) { |
1669 | size += (BYTES_PER_WORD - 1); | |
1670 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
1671 | } |
1672 | ||
1673 | /* calculate out the final buffer alignment: */ | |
1674 | /* 1) arch recommendation: can be overridden for debug */ | |
1675 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1676 | /* Default alignment: as specified by the arch code. | |
1677 | * Except if an object is really small, then squeeze multiple | |
1678 | * objects into one cacheline. | |
1679 | */ | |
1680 | ralign = cache_line_size(); | |
b28a02de | 1681 | while (size <= ralign / 2) |
1da177e4 LT |
1682 | ralign /= 2; |
1683 | } else { | |
1684 | ralign = BYTES_PER_WORD; | |
1685 | } | |
1686 | /* 2) arch mandated alignment: disables debug if necessary */ | |
1687 | if (ralign < ARCH_SLAB_MINALIGN) { | |
1688 | ralign = ARCH_SLAB_MINALIGN; | |
1689 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1690 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1691 | } |
1692 | /* 3) caller mandated alignment: disables debug if necessary */ | |
1693 | if (ralign < align) { | |
1694 | ralign = align; | |
1695 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1696 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1697 | } |
1698 | /* 4) Store it. Note that the debug code below can reduce | |
1699 | * the alignment to BYTES_PER_WORD. | |
1700 | */ | |
1701 | align = ralign; | |
1702 | ||
1703 | /* Get cache's description obj. */ | |
1704 | cachep = (kmem_cache_t *) kmem_cache_alloc(&cache_cache, SLAB_KERNEL); | |
1705 | if (!cachep) | |
4f12bb4f | 1706 | goto oops; |
1da177e4 LT |
1707 | memset(cachep, 0, sizeof(kmem_cache_t)); |
1708 | ||
1709 | #if DEBUG | |
3dafccf2 | 1710 | cachep->obj_size = size; |
1da177e4 LT |
1711 | |
1712 | if (flags & SLAB_RED_ZONE) { | |
1713 | /* redzoning only works with word aligned caches */ | |
1714 | align = BYTES_PER_WORD; | |
1715 | ||
1716 | /* add space for red zone words */ | |
3dafccf2 | 1717 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 1718 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
1719 | } |
1720 | if (flags & SLAB_STORE_USER) { | |
1721 | /* user store requires word alignment and | |
1722 | * one word storage behind the end of the real | |
1723 | * object. | |
1724 | */ | |
1725 | align = BYTES_PER_WORD; | |
1726 | size += BYTES_PER_WORD; | |
1727 | } | |
1728 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 1729 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
1730 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
1731 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
1732 | size = PAGE_SIZE; |
1733 | } | |
1734 | #endif | |
1735 | #endif | |
1736 | ||
1737 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 1738 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
1739 | /* |
1740 | * Size is large, assume best to place the slab management obj | |
1741 | * off-slab (should allow better packing of objs). | |
1742 | */ | |
1743 | flags |= CFLGS_OFF_SLAB; | |
1744 | ||
1745 | size = ALIGN(size, align); | |
1746 | ||
1747 | if ((flags & SLAB_RECLAIM_ACCOUNT) && size <= PAGE_SIZE) { | |
1748 | /* | |
1749 | * A VFS-reclaimable slab tends to have most allocations | |
1750 | * as GFP_NOFS and we really don't want to have to be allocating | |
1751 | * higher-order pages when we are unable to shrink dcache. | |
1752 | */ | |
1753 | cachep->gfporder = 0; | |
1754 | cache_estimate(cachep->gfporder, size, align, flags, | |
b28a02de | 1755 | &left_over, &cachep->num); |
4d268eba PE |
1756 | } else |
1757 | left_over = calculate_slab_order(cachep, size, align, flags); | |
1da177e4 LT |
1758 | |
1759 | if (!cachep->num) { | |
1760 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
1761 | kmem_cache_free(&cache_cache, cachep); | |
1762 | cachep = NULL; | |
4f12bb4f | 1763 | goto oops; |
1da177e4 | 1764 | } |
b28a02de PE |
1765 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
1766 | + sizeof(struct slab), align); | |
1da177e4 LT |
1767 | |
1768 | /* | |
1769 | * If the slab has been placed off-slab, and we have enough space then | |
1770 | * move it on-slab. This is at the expense of any extra colouring. | |
1771 | */ | |
1772 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
1773 | flags &= ~CFLGS_OFF_SLAB; | |
1774 | left_over -= slab_size; | |
1775 | } | |
1776 | ||
1777 | if (flags & CFLGS_OFF_SLAB) { | |
1778 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
1779 | slab_size = |
1780 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
1781 | } |
1782 | ||
1783 | cachep->colour_off = cache_line_size(); | |
1784 | /* Offset must be a multiple of the alignment. */ | |
1785 | if (cachep->colour_off < align) | |
1786 | cachep->colour_off = align; | |
b28a02de | 1787 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
1788 | cachep->slab_size = slab_size; |
1789 | cachep->flags = flags; | |
1790 | cachep->gfpflags = 0; | |
1791 | if (flags & SLAB_CACHE_DMA) | |
1792 | cachep->gfpflags |= GFP_DMA; | |
1793 | spin_lock_init(&cachep->spinlock); | |
3dafccf2 | 1794 | cachep->buffer_size = size; |
1da177e4 LT |
1795 | |
1796 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 1797 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
1798 | cachep->ctor = ctor; |
1799 | cachep->dtor = dtor; | |
1800 | cachep->name = name; | |
1801 | ||
1802 | /* Don't let CPUs to come and go */ | |
1803 | lock_cpu_hotplug(); | |
1804 | ||
1805 | if (g_cpucache_up == FULL) { | |
1806 | enable_cpucache(cachep); | |
1807 | } else { | |
1808 | if (g_cpucache_up == NONE) { | |
1809 | /* Note: the first kmem_cache_create must create | |
1810 | * the cache that's used by kmalloc(24), otherwise | |
1811 | * the creation of further caches will BUG(). | |
1812 | */ | |
e498be7d | 1813 | cachep->array[smp_processor_id()] = |
b28a02de | 1814 | &initarray_generic.cache; |
e498be7d CL |
1815 | |
1816 | /* If the cache that's used by | |
1817 | * kmalloc(sizeof(kmem_list3)) is the first cache, | |
1818 | * then we need to set up all its list3s, otherwise | |
1819 | * the creation of further caches will BUG(). | |
1820 | */ | |
1821 | set_up_list3s(cachep, SIZE_AC); | |
1822 | if (INDEX_AC == INDEX_L3) | |
1823 | g_cpucache_up = PARTIAL_L3; | |
1824 | else | |
1825 | g_cpucache_up = PARTIAL_AC; | |
1da177e4 | 1826 | } else { |
e498be7d | 1827 | cachep->array[smp_processor_id()] = |
b28a02de | 1828 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d CL |
1829 | |
1830 | if (g_cpucache_up == PARTIAL_AC) { | |
1831 | set_up_list3s(cachep, SIZE_L3); | |
1832 | g_cpucache_up = PARTIAL_L3; | |
1833 | } else { | |
1834 | int node; | |
1835 | for_each_online_node(node) { | |
1836 | ||
1837 | cachep->nodelists[node] = | |
b28a02de PE |
1838 | kmalloc_node(sizeof |
1839 | (struct kmem_list3), | |
1840 | GFP_KERNEL, node); | |
e498be7d | 1841 | BUG_ON(!cachep->nodelists[node]); |
b28a02de PE |
1842 | kmem_list3_init(cachep-> |
1843 | nodelists[node]); | |
e498be7d CL |
1844 | } |
1845 | } | |
1da177e4 | 1846 | } |
e498be7d | 1847 | cachep->nodelists[numa_node_id()]->next_reap = |
b28a02de PE |
1848 | jiffies + REAPTIMEOUT_LIST3 + |
1849 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1850 | |
1da177e4 LT |
1851 | BUG_ON(!ac_data(cachep)); |
1852 | ac_data(cachep)->avail = 0; | |
1853 | ac_data(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1854 | ac_data(cachep)->batchcount = 1; | |
1855 | ac_data(cachep)->touched = 0; | |
1856 | cachep->batchcount = 1; | |
1857 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
b28a02de | 1858 | } |
1da177e4 | 1859 | |
1da177e4 LT |
1860 | /* cache setup completed, link it into the list */ |
1861 | list_add(&cachep->next, &cache_chain); | |
1da177e4 | 1862 | unlock_cpu_hotplug(); |
b28a02de | 1863 | oops: |
1da177e4 LT |
1864 | if (!cachep && (flags & SLAB_PANIC)) |
1865 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 1866 | name); |
fc0abb14 | 1867 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1868 | return cachep; |
1869 | } | |
1870 | EXPORT_SYMBOL(kmem_cache_create); | |
1871 | ||
1872 | #if DEBUG | |
1873 | static void check_irq_off(void) | |
1874 | { | |
1875 | BUG_ON(!irqs_disabled()); | |
1876 | } | |
1877 | ||
1878 | static void check_irq_on(void) | |
1879 | { | |
1880 | BUG_ON(irqs_disabled()); | |
1881 | } | |
1882 | ||
1883 | static void check_spinlock_acquired(kmem_cache_t *cachep) | |
1884 | { | |
1885 | #ifdef CONFIG_SMP | |
1886 | check_irq_off(); | |
e498be7d | 1887 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
1888 | #endif |
1889 | } | |
e498be7d CL |
1890 | |
1891 | static inline void check_spinlock_acquired_node(kmem_cache_t *cachep, int node) | |
1892 | { | |
1893 | #ifdef CONFIG_SMP | |
1894 | check_irq_off(); | |
1895 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
1896 | #endif | |
1897 | } | |
1898 | ||
1da177e4 LT |
1899 | #else |
1900 | #define check_irq_off() do { } while(0) | |
1901 | #define check_irq_on() do { } while(0) | |
1902 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 1903 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
1904 | #endif |
1905 | ||
1906 | /* | |
1907 | * Waits for all CPUs to execute func(). | |
1908 | */ | |
b28a02de | 1909 | static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg) |
1da177e4 LT |
1910 | { |
1911 | check_irq_on(); | |
1912 | preempt_disable(); | |
1913 | ||
1914 | local_irq_disable(); | |
1915 | func(arg); | |
1916 | local_irq_enable(); | |
1917 | ||
1918 | if (smp_call_function(func, arg, 1, 1)) | |
1919 | BUG(); | |
1920 | ||
1921 | preempt_enable(); | |
1922 | } | |
1923 | ||
b28a02de PE |
1924 | static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac, |
1925 | int force, int node); | |
1da177e4 LT |
1926 | |
1927 | static void do_drain(void *arg) | |
1928 | { | |
b28a02de | 1929 | kmem_cache_t *cachep = (kmem_cache_t *) arg; |
1da177e4 | 1930 | struct array_cache *ac; |
ff69416e | 1931 | int node = numa_node_id(); |
1da177e4 LT |
1932 | |
1933 | check_irq_off(); | |
1934 | ac = ac_data(cachep); | |
ff69416e CL |
1935 | spin_lock(&cachep->nodelists[node]->list_lock); |
1936 | free_block(cachep, ac->entry, ac->avail, node); | |
1937 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
1938 | ac->avail = 0; |
1939 | } | |
1940 | ||
1941 | static void drain_cpu_caches(kmem_cache_t *cachep) | |
1942 | { | |
e498be7d CL |
1943 | struct kmem_list3 *l3; |
1944 | int node; | |
1945 | ||
1da177e4 LT |
1946 | smp_call_function_all_cpus(do_drain, cachep); |
1947 | check_irq_on(); | |
1948 | spin_lock_irq(&cachep->spinlock); | |
b28a02de | 1949 | for_each_online_node(node) { |
e498be7d CL |
1950 | l3 = cachep->nodelists[node]; |
1951 | if (l3) { | |
1952 | spin_lock(&l3->list_lock); | |
1953 | drain_array_locked(cachep, l3->shared, 1, node); | |
1954 | spin_unlock(&l3->list_lock); | |
1955 | if (l3->alien) | |
1956 | drain_alien_cache(cachep, l3); | |
1957 | } | |
1958 | } | |
1da177e4 LT |
1959 | spin_unlock_irq(&cachep->spinlock); |
1960 | } | |
1961 | ||
e498be7d | 1962 | static int __node_shrink(kmem_cache_t *cachep, int node) |
1da177e4 LT |
1963 | { |
1964 | struct slab *slabp; | |
e498be7d | 1965 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
1966 | int ret; |
1967 | ||
e498be7d | 1968 | for (;;) { |
1da177e4 LT |
1969 | struct list_head *p; |
1970 | ||
e498be7d CL |
1971 | p = l3->slabs_free.prev; |
1972 | if (p == &l3->slabs_free) | |
1da177e4 LT |
1973 | break; |
1974 | ||
e498be7d | 1975 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 LT |
1976 | #if DEBUG |
1977 | if (slabp->inuse) | |
1978 | BUG(); | |
1979 | #endif | |
1980 | list_del(&slabp->list); | |
1981 | ||
e498be7d CL |
1982 | l3->free_objects -= cachep->num; |
1983 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 1984 | slab_destroy(cachep, slabp); |
e498be7d | 1985 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 1986 | } |
b28a02de | 1987 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
1988 | return ret; |
1989 | } | |
1990 | ||
e498be7d CL |
1991 | static int __cache_shrink(kmem_cache_t *cachep) |
1992 | { | |
1993 | int ret = 0, i = 0; | |
1994 | struct kmem_list3 *l3; | |
1995 | ||
1996 | drain_cpu_caches(cachep); | |
1997 | ||
1998 | check_irq_on(); | |
1999 | for_each_online_node(i) { | |
2000 | l3 = cachep->nodelists[i]; | |
2001 | if (l3) { | |
2002 | spin_lock_irq(&l3->list_lock); | |
2003 | ret += __node_shrink(cachep, i); | |
2004 | spin_unlock_irq(&l3->list_lock); | |
2005 | } | |
2006 | } | |
2007 | return (ret ? 1 : 0); | |
2008 | } | |
2009 | ||
1da177e4 LT |
2010 | /** |
2011 | * kmem_cache_shrink - Shrink a cache. | |
2012 | * @cachep: The cache to shrink. | |
2013 | * | |
2014 | * Releases as many slabs as possible for a cache. | |
2015 | * To help debugging, a zero exit status indicates all slabs were released. | |
2016 | */ | |
2017 | int kmem_cache_shrink(kmem_cache_t *cachep) | |
2018 | { | |
2019 | if (!cachep || in_interrupt()) | |
2020 | BUG(); | |
2021 | ||
2022 | return __cache_shrink(cachep); | |
2023 | } | |
2024 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2025 | ||
2026 | /** | |
2027 | * kmem_cache_destroy - delete a cache | |
2028 | * @cachep: the cache to destroy | |
2029 | * | |
2030 | * Remove a kmem_cache_t object from the slab cache. | |
2031 | * Returns 0 on success. | |
2032 | * | |
2033 | * It is expected this function will be called by a module when it is | |
2034 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2035 | * cache being allocated each time a module is loaded and unloaded, if the | |
2036 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2037 | * | |
2038 | * The cache must be empty before calling this function. | |
2039 | * | |
2040 | * The caller must guarantee that noone will allocate memory from the cache | |
2041 | * during the kmem_cache_destroy(). | |
2042 | */ | |
b28a02de | 2043 | int kmem_cache_destroy(kmem_cache_t *cachep) |
1da177e4 LT |
2044 | { |
2045 | int i; | |
e498be7d | 2046 | struct kmem_list3 *l3; |
1da177e4 LT |
2047 | |
2048 | if (!cachep || in_interrupt()) | |
2049 | BUG(); | |
2050 | ||
2051 | /* Don't let CPUs to come and go */ | |
2052 | lock_cpu_hotplug(); | |
2053 | ||
2054 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2055 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2056 | /* |
2057 | * the chain is never empty, cache_cache is never destroyed | |
2058 | */ | |
2059 | list_del(&cachep->next); | |
fc0abb14 | 2060 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2061 | |
2062 | if (__cache_shrink(cachep)) { | |
2063 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2064 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2065 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2066 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2067 | unlock_cpu_hotplug(); |
2068 | return 1; | |
2069 | } | |
2070 | ||
2071 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2072 | synchronize_rcu(); |
1da177e4 | 2073 | |
e498be7d | 2074 | for_each_online_cpu(i) |
b28a02de | 2075 | kfree(cachep->array[i]); |
1da177e4 LT |
2076 | |
2077 | /* NUMA: free the list3 structures */ | |
e498be7d CL |
2078 | for_each_online_node(i) { |
2079 | if ((l3 = cachep->nodelists[i])) { | |
2080 | kfree(l3->shared); | |
2081 | free_alien_cache(l3->alien); | |
2082 | kfree(l3); | |
2083 | } | |
2084 | } | |
1da177e4 LT |
2085 | kmem_cache_free(&cache_cache, cachep); |
2086 | ||
2087 | unlock_cpu_hotplug(); | |
2088 | ||
2089 | return 0; | |
2090 | } | |
2091 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2092 | ||
2093 | /* Get the memory for a slab management obj. */ | |
b28a02de PE |
2094 | static struct slab *alloc_slabmgmt(kmem_cache_t *cachep, void *objp, |
2095 | int colour_off, gfp_t local_flags) | |
1da177e4 LT |
2096 | { |
2097 | struct slab *slabp; | |
b28a02de | 2098 | |
1da177e4 LT |
2099 | if (OFF_SLAB(cachep)) { |
2100 | /* Slab management obj is off-slab. */ | |
2101 | slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags); | |
2102 | if (!slabp) | |
2103 | return NULL; | |
2104 | } else { | |
b28a02de | 2105 | slabp = objp + colour_off; |
1da177e4 LT |
2106 | colour_off += cachep->slab_size; |
2107 | } | |
2108 | slabp->inuse = 0; | |
2109 | slabp->colouroff = colour_off; | |
b28a02de | 2110 | slabp->s_mem = objp + colour_off; |
1da177e4 LT |
2111 | |
2112 | return slabp; | |
2113 | } | |
2114 | ||
2115 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2116 | { | |
b28a02de | 2117 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2118 | } |
2119 | ||
2120 | static void cache_init_objs(kmem_cache_t *cachep, | |
b28a02de | 2121 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2122 | { |
2123 | int i; | |
2124 | ||
2125 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 2126 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
1da177e4 LT |
2127 | #if DEBUG |
2128 | /* need to poison the objs? */ | |
2129 | if (cachep->flags & SLAB_POISON) | |
2130 | poison_obj(cachep, objp, POISON_FREE); | |
2131 | if (cachep->flags & SLAB_STORE_USER) | |
2132 | *dbg_userword(cachep, objp) = NULL; | |
2133 | ||
2134 | if (cachep->flags & SLAB_RED_ZONE) { | |
2135 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2136 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2137 | } | |
2138 | /* | |
2139 | * Constructors are not allowed to allocate memory from | |
2140 | * the same cache which they are a constructor for. | |
2141 | * Otherwise, deadlock. They must also be threaded. | |
2142 | */ | |
2143 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2144 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2145 | ctor_flags); |
1da177e4 LT |
2146 | |
2147 | if (cachep->flags & SLAB_RED_ZONE) { | |
2148 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2149 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2150 | " end of an object"); |
1da177e4 LT |
2151 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2152 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2153 | " start of an object"); |
1da177e4 | 2154 | } |
3dafccf2 | 2155 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep) |
b28a02de PE |
2156 | && cachep->flags & SLAB_POISON) |
2157 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 2158 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2159 | #else |
2160 | if (cachep->ctor) | |
2161 | cachep->ctor(objp, cachep, ctor_flags); | |
2162 | #endif | |
b28a02de | 2163 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2164 | } |
b28a02de | 2165 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2166 | slabp->free = 0; |
2167 | } | |
2168 | ||
6daa0e28 | 2169 | static void kmem_flagcheck(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2170 | { |
2171 | if (flags & SLAB_DMA) { | |
2172 | if (!(cachep->gfpflags & GFP_DMA)) | |
2173 | BUG(); | |
2174 | } else { | |
2175 | if (cachep->gfpflags & GFP_DMA) | |
2176 | BUG(); | |
2177 | } | |
2178 | } | |
2179 | ||
2180 | static void set_slab_attr(kmem_cache_t *cachep, struct slab *slabp, void *objp) | |
2181 | { | |
2182 | int i; | |
2183 | struct page *page; | |
2184 | ||
2185 | /* Nasty!!!!!! I hope this is OK. */ | |
2186 | i = 1 << cachep->gfporder; | |
2187 | page = virt_to_page(objp); | |
2188 | do { | |
065d41cb PE |
2189 | page_set_cache(page, cachep); |
2190 | page_set_slab(page, slabp); | |
1da177e4 LT |
2191 | page++; |
2192 | } while (--i); | |
2193 | } | |
2194 | ||
2195 | /* | |
2196 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2197 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2198 | */ | |
dd0fc66f | 2199 | static int cache_grow(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2200 | { |
b28a02de PE |
2201 | struct slab *slabp; |
2202 | void *objp; | |
2203 | size_t offset; | |
2204 | gfp_t local_flags; | |
2205 | unsigned long ctor_flags; | |
e498be7d | 2206 | struct kmem_list3 *l3; |
1da177e4 LT |
2207 | |
2208 | /* Be lazy and only check for valid flags here, | |
b28a02de | 2209 | * keeping it out of the critical path in kmem_cache_alloc(). |
1da177e4 | 2210 | */ |
b28a02de | 2211 | if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)) |
1da177e4 LT |
2212 | BUG(); |
2213 | if (flags & SLAB_NO_GROW) | |
2214 | return 0; | |
2215 | ||
2216 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2217 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2218 | if (!(local_flags & __GFP_WAIT)) | |
2219 | /* | |
2220 | * Not allowed to sleep. Need to tell a constructor about | |
2221 | * this - it might need to know... | |
2222 | */ | |
2223 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2224 | ||
2225 | /* About to mess with non-constant members - lock. */ | |
2226 | check_irq_off(); | |
2227 | spin_lock(&cachep->spinlock); | |
2228 | ||
2229 | /* Get colour for the slab, and cal the next value. */ | |
2230 | offset = cachep->colour_next; | |
2231 | cachep->colour_next++; | |
2232 | if (cachep->colour_next >= cachep->colour) | |
2233 | cachep->colour_next = 0; | |
2234 | offset *= cachep->colour_off; | |
2235 | ||
2236 | spin_unlock(&cachep->spinlock); | |
2237 | ||
e498be7d | 2238 | check_irq_off(); |
1da177e4 LT |
2239 | if (local_flags & __GFP_WAIT) |
2240 | local_irq_enable(); | |
2241 | ||
2242 | /* | |
2243 | * The test for missing atomic flag is performed here, rather than | |
2244 | * the more obvious place, simply to reduce the critical path length | |
2245 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2246 | * will eventually be caught here (where it matters). | |
2247 | */ | |
2248 | kmem_flagcheck(cachep, flags); | |
2249 | ||
e498be7d CL |
2250 | /* Get mem for the objs. |
2251 | * Attempt to allocate a physical page from 'nodeid', | |
2252 | */ | |
1da177e4 LT |
2253 | if (!(objp = kmem_getpages(cachep, flags, nodeid))) |
2254 | goto failed; | |
2255 | ||
2256 | /* Get slab management. */ | |
2257 | if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags))) | |
2258 | goto opps1; | |
2259 | ||
e498be7d | 2260 | slabp->nodeid = nodeid; |
1da177e4 LT |
2261 | set_slab_attr(cachep, slabp, objp); |
2262 | ||
2263 | cache_init_objs(cachep, slabp, ctor_flags); | |
2264 | ||
2265 | if (local_flags & __GFP_WAIT) | |
2266 | local_irq_disable(); | |
2267 | check_irq_off(); | |
e498be7d CL |
2268 | l3 = cachep->nodelists[nodeid]; |
2269 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2270 | |
2271 | /* Make slab active. */ | |
e498be7d | 2272 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2273 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2274 | l3->free_objects += cachep->num; |
2275 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2276 | return 1; |
b28a02de | 2277 | opps1: |
1da177e4 | 2278 | kmem_freepages(cachep, objp); |
b28a02de | 2279 | failed: |
1da177e4 LT |
2280 | if (local_flags & __GFP_WAIT) |
2281 | local_irq_disable(); | |
2282 | return 0; | |
2283 | } | |
2284 | ||
2285 | #if DEBUG | |
2286 | ||
2287 | /* | |
2288 | * Perform extra freeing checks: | |
2289 | * - detect bad pointers. | |
2290 | * - POISON/RED_ZONE checking | |
2291 | * - destructor calls, for caches with POISON+dtor | |
2292 | */ | |
2293 | static void kfree_debugcheck(const void *objp) | |
2294 | { | |
2295 | struct page *page; | |
2296 | ||
2297 | if (!virt_addr_valid(objp)) { | |
2298 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2299 | (unsigned long)objp); |
2300 | BUG(); | |
1da177e4 LT |
2301 | } |
2302 | page = virt_to_page(objp); | |
2303 | if (!PageSlab(page)) { | |
b28a02de PE |
2304 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2305 | (unsigned long)objp); | |
1da177e4 LT |
2306 | BUG(); |
2307 | } | |
2308 | } | |
2309 | ||
2310 | static void *cache_free_debugcheck(kmem_cache_t *cachep, void *objp, | |
b28a02de | 2311 | void *caller) |
1da177e4 LT |
2312 | { |
2313 | struct page *page; | |
2314 | unsigned int objnr; | |
2315 | struct slab *slabp; | |
2316 | ||
3dafccf2 | 2317 | objp -= obj_offset(cachep); |
1da177e4 LT |
2318 | kfree_debugcheck(objp); |
2319 | page = virt_to_page(objp); | |
2320 | ||
065d41cb | 2321 | if (page_get_cache(page) != cachep) { |
b28a02de PE |
2322 | printk(KERN_ERR |
2323 | "mismatch in kmem_cache_free: expected cache %p, got %p\n", | |
2324 | page_get_cache(page), cachep); | |
1da177e4 | 2325 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2326 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2327 | page_get_cache(page)->name); | |
1da177e4 LT |
2328 | WARN_ON(1); |
2329 | } | |
065d41cb | 2330 | slabp = page_get_slab(page); |
1da177e4 LT |
2331 | |
2332 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2333 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE |
2334 | || *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2335 | slab_error(cachep, | |
2336 | "double free, or memory outside" | |
2337 | " object was overwritten"); | |
2338 | printk(KERN_ERR | |
2339 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2340 | objp, *dbg_redzone1(cachep, objp), | |
2341 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2342 | } |
2343 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2344 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2345 | } | |
2346 | if (cachep->flags & SLAB_STORE_USER) | |
2347 | *dbg_userword(cachep, objp) = caller; | |
2348 | ||
3dafccf2 | 2349 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
1da177e4 LT |
2350 | |
2351 | BUG_ON(objnr >= cachep->num); | |
3dafccf2 | 2352 | BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size); |
1da177e4 LT |
2353 | |
2354 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
2355 | /* Need to call the slab's constructor so the | |
2356 | * caller can perform a verify of its state (debugging). | |
2357 | * Called without the cache-lock held. | |
2358 | */ | |
3dafccf2 | 2359 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2360 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2361 | } |
2362 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2363 | /* we want to cache poison the object, | |
2364 | * call the destruction callback | |
2365 | */ | |
3dafccf2 | 2366 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 LT |
2367 | } |
2368 | if (cachep->flags & SLAB_POISON) { | |
2369 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 2370 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) { |
1da177e4 | 2371 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2372 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2373 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2374 | } else { |
2375 | poison_obj(cachep, objp, POISON_FREE); | |
2376 | } | |
2377 | #else | |
2378 | poison_obj(cachep, objp, POISON_FREE); | |
2379 | #endif | |
2380 | } | |
2381 | return objp; | |
2382 | } | |
2383 | ||
2384 | static void check_slabp(kmem_cache_t *cachep, struct slab *slabp) | |
2385 | { | |
2386 | kmem_bufctl_t i; | |
2387 | int entries = 0; | |
b28a02de | 2388 | |
1da177e4 LT |
2389 | /* Check slab's freelist to see if this obj is there. */ |
2390 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2391 | entries++; | |
2392 | if (entries > cachep->num || i >= cachep->num) | |
2393 | goto bad; | |
2394 | } | |
2395 | if (entries != cachep->num - slabp->inuse) { | |
b28a02de PE |
2396 | bad: |
2397 | printk(KERN_ERR | |
2398 | "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2399 | cachep->name, cachep->num, slabp, slabp->inuse); | |
2400 | for (i = 0; | |
2401 | i < sizeof(slabp) + cachep->num * sizeof(kmem_bufctl_t); | |
2402 | i++) { | |
2403 | if ((i % 16) == 0) | |
1da177e4 | 2404 | printk("\n%03x:", i); |
b28a02de | 2405 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2406 | } |
2407 | printk("\n"); | |
2408 | BUG(); | |
2409 | } | |
2410 | } | |
2411 | #else | |
2412 | #define kfree_debugcheck(x) do { } while(0) | |
2413 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2414 | #define check_slabp(x,y) do { } while(0) | |
2415 | #endif | |
2416 | ||
dd0fc66f | 2417 | static void *cache_alloc_refill(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2418 | { |
2419 | int batchcount; | |
2420 | struct kmem_list3 *l3; | |
2421 | struct array_cache *ac; | |
2422 | ||
2423 | check_irq_off(); | |
2424 | ac = ac_data(cachep); | |
b28a02de | 2425 | retry: |
1da177e4 LT |
2426 | batchcount = ac->batchcount; |
2427 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
2428 | /* if there was little recent activity on this | |
2429 | * cache, then perform only a partial refill. | |
2430 | * Otherwise we could generate refill bouncing. | |
2431 | */ | |
2432 | batchcount = BATCHREFILL_LIMIT; | |
2433 | } | |
e498be7d CL |
2434 | l3 = cachep->nodelists[numa_node_id()]; |
2435 | ||
2436 | BUG_ON(ac->avail > 0 || !l3); | |
2437 | spin_lock(&l3->list_lock); | |
1da177e4 | 2438 | |
1da177e4 LT |
2439 | if (l3->shared) { |
2440 | struct array_cache *shared_array = l3->shared; | |
2441 | if (shared_array->avail) { | |
2442 | if (batchcount > shared_array->avail) | |
2443 | batchcount = shared_array->avail; | |
2444 | shared_array->avail -= batchcount; | |
2445 | ac->avail = batchcount; | |
e498be7d | 2446 | memcpy(ac->entry, |
b28a02de PE |
2447 | &(shared_array->entry[shared_array->avail]), |
2448 | sizeof(void *) * batchcount); | |
1da177e4 LT |
2449 | shared_array->touched = 1; |
2450 | goto alloc_done; | |
2451 | } | |
2452 | } | |
2453 | while (batchcount > 0) { | |
2454 | struct list_head *entry; | |
2455 | struct slab *slabp; | |
2456 | /* Get slab alloc is to come from. */ | |
2457 | entry = l3->slabs_partial.next; | |
2458 | if (entry == &l3->slabs_partial) { | |
2459 | l3->free_touched = 1; | |
2460 | entry = l3->slabs_free.next; | |
2461 | if (entry == &l3->slabs_free) | |
2462 | goto must_grow; | |
2463 | } | |
2464 | ||
2465 | slabp = list_entry(entry, struct slab, list); | |
2466 | check_slabp(cachep, slabp); | |
2467 | check_spinlock_acquired(cachep); | |
2468 | while (slabp->inuse < cachep->num && batchcount--) { | |
2469 | kmem_bufctl_t next; | |
2470 | STATS_INC_ALLOCED(cachep); | |
2471 | STATS_INC_ACTIVE(cachep); | |
2472 | STATS_SET_HIGH(cachep); | |
2473 | ||
2474 | /* get obj pointer */ | |
e498be7d | 2475 | ac->entry[ac->avail++] = slabp->s_mem + |
3dafccf2 | 2476 | slabp->free * cachep->buffer_size; |
1da177e4 LT |
2477 | |
2478 | slabp->inuse++; | |
2479 | next = slab_bufctl(slabp)[slabp->free]; | |
2480 | #if DEBUG | |
2481 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
09ad4bbc | 2482 | WARN_ON(numa_node_id() != slabp->nodeid); |
1da177e4 | 2483 | #endif |
b28a02de | 2484 | slabp->free = next; |
1da177e4 LT |
2485 | } |
2486 | check_slabp(cachep, slabp); | |
2487 | ||
2488 | /* move slabp to correct slabp list: */ | |
2489 | list_del(&slabp->list); | |
2490 | if (slabp->free == BUFCTL_END) | |
2491 | list_add(&slabp->list, &l3->slabs_full); | |
2492 | else | |
2493 | list_add(&slabp->list, &l3->slabs_partial); | |
2494 | } | |
2495 | ||
b28a02de | 2496 | must_grow: |
1da177e4 | 2497 | l3->free_objects -= ac->avail; |
b28a02de | 2498 | alloc_done: |
e498be7d | 2499 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2500 | |
2501 | if (unlikely(!ac->avail)) { | |
2502 | int x; | |
e498be7d CL |
2503 | x = cache_grow(cachep, flags, numa_node_id()); |
2504 | ||
1da177e4 LT |
2505 | // cache_grow can reenable interrupts, then ac could change. |
2506 | ac = ac_data(cachep); | |
2507 | if (!x && ac->avail == 0) // no objects in sight? abort | |
2508 | return NULL; | |
2509 | ||
b28a02de | 2510 | if (!ac->avail) // objects refilled by interrupt? |
1da177e4 LT |
2511 | goto retry; |
2512 | } | |
2513 | ac->touched = 1; | |
e498be7d | 2514 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2515 | } |
2516 | ||
2517 | static inline void | |
dd0fc66f | 2518 | cache_alloc_debugcheck_before(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2519 | { |
2520 | might_sleep_if(flags & __GFP_WAIT); | |
2521 | #if DEBUG | |
2522 | kmem_flagcheck(cachep, flags); | |
2523 | #endif | |
2524 | } | |
2525 | ||
2526 | #if DEBUG | |
b28a02de PE |
2527 | static void *cache_alloc_debugcheck_after(kmem_cache_t *cachep, gfp_t flags, |
2528 | void *objp, void *caller) | |
1da177e4 | 2529 | { |
b28a02de | 2530 | if (!objp) |
1da177e4 | 2531 | return objp; |
b28a02de | 2532 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2533 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2534 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2535 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2536 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2537 | else |
2538 | check_poison_obj(cachep, objp); | |
2539 | #else | |
2540 | check_poison_obj(cachep, objp); | |
2541 | #endif | |
2542 | poison_obj(cachep, objp, POISON_INUSE); | |
2543 | } | |
2544 | if (cachep->flags & SLAB_STORE_USER) | |
2545 | *dbg_userword(cachep, objp) = caller; | |
2546 | ||
2547 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2548 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE |
2549 | || *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2550 | slab_error(cachep, | |
2551 | "double free, or memory outside" | |
2552 | " object was overwritten"); | |
2553 | printk(KERN_ERR | |
2554 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2555 | objp, *dbg_redzone1(cachep, objp), | |
2556 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2557 | } |
2558 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2559 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2560 | } | |
3dafccf2 | 2561 | objp += obj_offset(cachep); |
1da177e4 | 2562 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2563 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2564 | |
2565 | if (!(flags & __GFP_WAIT)) | |
2566 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2567 | ||
2568 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2569 | } |
1da177e4 LT |
2570 | return objp; |
2571 | } | |
2572 | #else | |
2573 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2574 | #endif | |
2575 | ||
dd0fc66f | 2576 | static inline void *____cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 | 2577 | { |
b28a02de | 2578 | void *objp; |
1da177e4 LT |
2579 | struct array_cache *ac; |
2580 | ||
dc85da15 | 2581 | #ifdef CONFIG_NUMA |
86c562a9 | 2582 | if (unlikely(current->mempolicy && !in_interrupt())) { |
dc85da15 CL |
2583 | int nid = slab_node(current->mempolicy); |
2584 | ||
2585 | if (nid != numa_node_id()) | |
2586 | return __cache_alloc_node(cachep, flags, nid); | |
2587 | } | |
2588 | #endif | |
2589 | ||
5c382300 | 2590 | check_irq_off(); |
1da177e4 LT |
2591 | ac = ac_data(cachep); |
2592 | if (likely(ac->avail)) { | |
2593 | STATS_INC_ALLOCHIT(cachep); | |
2594 | ac->touched = 1; | |
e498be7d | 2595 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2596 | } else { |
2597 | STATS_INC_ALLOCMISS(cachep); | |
2598 | objp = cache_alloc_refill(cachep, flags); | |
2599 | } | |
5c382300 AK |
2600 | return objp; |
2601 | } | |
2602 | ||
dd0fc66f | 2603 | static inline void *__cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
5c382300 AK |
2604 | { |
2605 | unsigned long save_flags; | |
b28a02de | 2606 | void *objp; |
5c382300 AK |
2607 | |
2608 | cache_alloc_debugcheck_before(cachep, flags); | |
2609 | ||
2610 | local_irq_save(save_flags); | |
2611 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2612 | local_irq_restore(save_flags); |
34342e86 | 2613 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
b28a02de | 2614 | __builtin_return_address(0)); |
34342e86 | 2615 | prefetchw(objp); |
1da177e4 LT |
2616 | return objp; |
2617 | } | |
2618 | ||
e498be7d CL |
2619 | #ifdef CONFIG_NUMA |
2620 | /* | |
2621 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2622 | */ |
6daa0e28 | 2623 | static void *__cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
e498be7d CL |
2624 | { |
2625 | struct list_head *entry; | |
b28a02de PE |
2626 | struct slab *slabp; |
2627 | struct kmem_list3 *l3; | |
2628 | void *obj; | |
2629 | kmem_bufctl_t next; | |
2630 | int x; | |
2631 | ||
2632 | l3 = cachep->nodelists[nodeid]; | |
2633 | BUG_ON(!l3); | |
2634 | ||
2635 | retry: | |
2636 | spin_lock(&l3->list_lock); | |
2637 | entry = l3->slabs_partial.next; | |
2638 | if (entry == &l3->slabs_partial) { | |
2639 | l3->free_touched = 1; | |
2640 | entry = l3->slabs_free.next; | |
2641 | if (entry == &l3->slabs_free) | |
2642 | goto must_grow; | |
2643 | } | |
2644 | ||
2645 | slabp = list_entry(entry, struct slab, list); | |
2646 | check_spinlock_acquired_node(cachep, nodeid); | |
2647 | check_slabp(cachep, slabp); | |
2648 | ||
2649 | STATS_INC_NODEALLOCS(cachep); | |
2650 | STATS_INC_ACTIVE(cachep); | |
2651 | STATS_SET_HIGH(cachep); | |
2652 | ||
2653 | BUG_ON(slabp->inuse == cachep->num); | |
2654 | ||
2655 | /* get obj pointer */ | |
3dafccf2 | 2656 | obj = slabp->s_mem + slabp->free * cachep->buffer_size; |
b28a02de PE |
2657 | slabp->inuse++; |
2658 | next = slab_bufctl(slabp)[slabp->free]; | |
e498be7d | 2659 | #if DEBUG |
b28a02de | 2660 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; |
e498be7d | 2661 | #endif |
b28a02de PE |
2662 | slabp->free = next; |
2663 | check_slabp(cachep, slabp); | |
2664 | l3->free_objects--; | |
2665 | /* move slabp to correct slabp list: */ | |
2666 | list_del(&slabp->list); | |
2667 | ||
2668 | if (slabp->free == BUFCTL_END) { | |
2669 | list_add(&slabp->list, &l3->slabs_full); | |
2670 | } else { | |
2671 | list_add(&slabp->list, &l3->slabs_partial); | |
2672 | } | |
e498be7d | 2673 | |
b28a02de PE |
2674 | spin_unlock(&l3->list_lock); |
2675 | goto done; | |
e498be7d | 2676 | |
b28a02de PE |
2677 | must_grow: |
2678 | spin_unlock(&l3->list_lock); | |
2679 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 2680 | |
b28a02de PE |
2681 | if (!x) |
2682 | return NULL; | |
e498be7d | 2683 | |
b28a02de PE |
2684 | goto retry; |
2685 | done: | |
2686 | return obj; | |
e498be7d CL |
2687 | } |
2688 | #endif | |
2689 | ||
2690 | /* | |
2691 | * Caller needs to acquire correct kmem_list's list_lock | |
2692 | */ | |
b28a02de PE |
2693 | static void free_block(kmem_cache_t *cachep, void **objpp, int nr_objects, |
2694 | int node) | |
1da177e4 LT |
2695 | { |
2696 | int i; | |
e498be7d | 2697 | struct kmem_list3 *l3; |
1da177e4 LT |
2698 | |
2699 | for (i = 0; i < nr_objects; i++) { | |
2700 | void *objp = objpp[i]; | |
2701 | struct slab *slabp; | |
2702 | unsigned int objnr; | |
2703 | ||
065d41cb | 2704 | slabp = page_get_slab(virt_to_page(objp)); |
ff69416e | 2705 | l3 = cachep->nodelists[node]; |
1da177e4 | 2706 | list_del(&slabp->list); |
3dafccf2 | 2707 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
ff69416e | 2708 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 2709 | check_slabp(cachep, slabp); |
e498be7d | 2710 | |
1da177e4 | 2711 | #if DEBUG |
09ad4bbc CL |
2712 | /* Verify that the slab belongs to the intended node */ |
2713 | WARN_ON(slabp->nodeid != node); | |
2714 | ||
1da177e4 | 2715 | if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) { |
e498be7d | 2716 | printk(KERN_ERR "slab: double free detected in cache " |
b28a02de | 2717 | "'%s', objp %p\n", cachep->name, objp); |
1da177e4 LT |
2718 | BUG(); |
2719 | } | |
2720 | #endif | |
2721 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2722 | slabp->free = objnr; | |
2723 | STATS_DEC_ACTIVE(cachep); | |
2724 | slabp->inuse--; | |
e498be7d | 2725 | l3->free_objects++; |
1da177e4 LT |
2726 | check_slabp(cachep, slabp); |
2727 | ||
2728 | /* fixup slab chains */ | |
2729 | if (slabp->inuse == 0) { | |
e498be7d CL |
2730 | if (l3->free_objects > l3->free_limit) { |
2731 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
2732 | slab_destroy(cachep, slabp); |
2733 | } else { | |
e498be7d | 2734 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
2735 | } |
2736 | } else { | |
2737 | /* Unconditionally move a slab to the end of the | |
2738 | * partial list on free - maximum time for the | |
2739 | * other objects to be freed, too. | |
2740 | */ | |
e498be7d | 2741 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
2742 | } |
2743 | } | |
2744 | } | |
2745 | ||
2746 | static void cache_flusharray(kmem_cache_t *cachep, struct array_cache *ac) | |
2747 | { | |
2748 | int batchcount; | |
e498be7d | 2749 | struct kmem_list3 *l3; |
ff69416e | 2750 | int node = numa_node_id(); |
1da177e4 LT |
2751 | |
2752 | batchcount = ac->batchcount; | |
2753 | #if DEBUG | |
2754 | BUG_ON(!batchcount || batchcount > ac->avail); | |
2755 | #endif | |
2756 | check_irq_off(); | |
ff69416e | 2757 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2758 | spin_lock(&l3->list_lock); |
2759 | if (l3->shared) { | |
2760 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 2761 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
2762 | if (max) { |
2763 | if (batchcount > max) | |
2764 | batchcount = max; | |
e498be7d | 2765 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 2766 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
2767 | shared_array->avail += batchcount; |
2768 | goto free_done; | |
2769 | } | |
2770 | } | |
2771 | ||
ff69416e | 2772 | free_block(cachep, ac->entry, batchcount, node); |
b28a02de | 2773 | free_done: |
1da177e4 LT |
2774 | #if STATS |
2775 | { | |
2776 | int i = 0; | |
2777 | struct list_head *p; | |
2778 | ||
e498be7d CL |
2779 | p = l3->slabs_free.next; |
2780 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
2781 | struct slab *slabp; |
2782 | ||
2783 | slabp = list_entry(p, struct slab, list); | |
2784 | BUG_ON(slabp->inuse); | |
2785 | ||
2786 | i++; | |
2787 | p = p->next; | |
2788 | } | |
2789 | STATS_SET_FREEABLE(cachep, i); | |
2790 | } | |
2791 | #endif | |
e498be7d | 2792 | spin_unlock(&l3->list_lock); |
1da177e4 | 2793 | ac->avail -= batchcount; |
e498be7d | 2794 | memmove(ac->entry, &(ac->entry[batchcount]), |
b28a02de | 2795 | sizeof(void *) * ac->avail); |
1da177e4 LT |
2796 | } |
2797 | ||
2798 | /* | |
2799 | * __cache_free | |
2800 | * Release an obj back to its cache. If the obj has a constructed | |
2801 | * state, it must be in this state _before_ it is released. | |
2802 | * | |
2803 | * Called with disabled ints. | |
2804 | */ | |
2805 | static inline void __cache_free(kmem_cache_t *cachep, void *objp) | |
2806 | { | |
2807 | struct array_cache *ac = ac_data(cachep); | |
2808 | ||
2809 | check_irq_off(); | |
2810 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
2811 | ||
e498be7d CL |
2812 | /* Make sure we are not freeing a object from another |
2813 | * node to the array cache on this cpu. | |
2814 | */ | |
2815 | #ifdef CONFIG_NUMA | |
2816 | { | |
2817 | struct slab *slabp; | |
065d41cb | 2818 | slabp = page_get_slab(virt_to_page(objp)); |
e498be7d CL |
2819 | if (unlikely(slabp->nodeid != numa_node_id())) { |
2820 | struct array_cache *alien = NULL; | |
2821 | int nodeid = slabp->nodeid; | |
b28a02de PE |
2822 | struct kmem_list3 *l3 = |
2823 | cachep->nodelists[numa_node_id()]; | |
e498be7d CL |
2824 | |
2825 | STATS_INC_NODEFREES(cachep); | |
2826 | if (l3->alien && l3->alien[nodeid]) { | |
2827 | alien = l3->alien[nodeid]; | |
2828 | spin_lock(&alien->lock); | |
2829 | if (unlikely(alien->avail == alien->limit)) | |
2830 | __drain_alien_cache(cachep, | |
b28a02de | 2831 | alien, nodeid); |
e498be7d CL |
2832 | alien->entry[alien->avail++] = objp; |
2833 | spin_unlock(&alien->lock); | |
2834 | } else { | |
2835 | spin_lock(&(cachep->nodelists[nodeid])-> | |
b28a02de | 2836 | list_lock); |
ff69416e | 2837 | free_block(cachep, &objp, 1, nodeid); |
e498be7d | 2838 | spin_unlock(&(cachep->nodelists[nodeid])-> |
b28a02de | 2839 | list_lock); |
e498be7d CL |
2840 | } |
2841 | return; | |
2842 | } | |
2843 | } | |
2844 | #endif | |
1da177e4 LT |
2845 | if (likely(ac->avail < ac->limit)) { |
2846 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 2847 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2848 | return; |
2849 | } else { | |
2850 | STATS_INC_FREEMISS(cachep); | |
2851 | cache_flusharray(cachep, ac); | |
e498be7d | 2852 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2853 | } |
2854 | } | |
2855 | ||
2856 | /** | |
2857 | * kmem_cache_alloc - Allocate an object | |
2858 | * @cachep: The cache to allocate from. | |
2859 | * @flags: See kmalloc(). | |
2860 | * | |
2861 | * Allocate an object from this cache. The flags are only relevant | |
2862 | * if the cache has no available objects. | |
2863 | */ | |
dd0fc66f | 2864 | void *kmem_cache_alloc(kmem_cache_t *cachep, gfp_t flags) |
1da177e4 LT |
2865 | { |
2866 | return __cache_alloc(cachep, flags); | |
2867 | } | |
2868 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2869 | ||
2870 | /** | |
2871 | * kmem_ptr_validate - check if an untrusted pointer might | |
2872 | * be a slab entry. | |
2873 | * @cachep: the cache we're checking against | |
2874 | * @ptr: pointer to validate | |
2875 | * | |
2876 | * This verifies that the untrusted pointer looks sane: | |
2877 | * it is _not_ a guarantee that the pointer is actually | |
2878 | * part of the slab cache in question, but it at least | |
2879 | * validates that the pointer can be dereferenced and | |
2880 | * looks half-way sane. | |
2881 | * | |
2882 | * Currently only used for dentry validation. | |
2883 | */ | |
2884 | int fastcall kmem_ptr_validate(kmem_cache_t *cachep, void *ptr) | |
2885 | { | |
b28a02de | 2886 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 2887 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 2888 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 2889 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
2890 | struct page *page; |
2891 | ||
2892 | if (unlikely(addr < min_addr)) | |
2893 | goto out; | |
2894 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
2895 | goto out; | |
2896 | if (unlikely(addr & align_mask)) | |
2897 | goto out; | |
2898 | if (unlikely(!kern_addr_valid(addr))) | |
2899 | goto out; | |
2900 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
2901 | goto out; | |
2902 | page = virt_to_page(ptr); | |
2903 | if (unlikely(!PageSlab(page))) | |
2904 | goto out; | |
065d41cb | 2905 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
2906 | goto out; |
2907 | return 1; | |
b28a02de | 2908 | out: |
1da177e4 LT |
2909 | return 0; |
2910 | } | |
2911 | ||
2912 | #ifdef CONFIG_NUMA | |
2913 | /** | |
2914 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
2915 | * @cachep: The cache to allocate from. | |
2916 | * @flags: See kmalloc(). | |
2917 | * @nodeid: node number of the target node. | |
2918 | * | |
2919 | * Identical to kmem_cache_alloc, except that this function is slow | |
2920 | * and can sleep. And it will allocate memory on the given node, which | |
2921 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
2922 | * New and improved: it will now make sure that the object gets |
2923 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 2924 | */ |
dd0fc66f | 2925 | void *kmem_cache_alloc_node(kmem_cache_t *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2926 | { |
e498be7d CL |
2927 | unsigned long save_flags; |
2928 | void *ptr; | |
1da177e4 | 2929 | |
e498be7d CL |
2930 | cache_alloc_debugcheck_before(cachep, flags); |
2931 | local_irq_save(save_flags); | |
18f820f6 CL |
2932 | |
2933 | if (nodeid == -1 || nodeid == numa_node_id() || | |
2934 | !cachep->nodelists[nodeid]) | |
5c382300 AK |
2935 | ptr = ____cache_alloc(cachep, flags); |
2936 | else | |
2937 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 2938 | local_irq_restore(save_flags); |
18f820f6 CL |
2939 | |
2940 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
2941 | __builtin_return_address(0)); | |
1da177e4 | 2942 | |
e498be7d | 2943 | return ptr; |
1da177e4 LT |
2944 | } |
2945 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
2946 | ||
dd0fc66f | 2947 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 MS |
2948 | { |
2949 | kmem_cache_t *cachep; | |
2950 | ||
2951 | cachep = kmem_find_general_cachep(size, flags); | |
2952 | if (unlikely(cachep == NULL)) | |
2953 | return NULL; | |
2954 | return kmem_cache_alloc_node(cachep, flags, node); | |
2955 | } | |
2956 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
2957 | #endif |
2958 | ||
2959 | /** | |
2960 | * kmalloc - allocate memory | |
2961 | * @size: how many bytes of memory are required. | |
2962 | * @flags: the type of memory to allocate. | |
2963 | * | |
2964 | * kmalloc is the normal method of allocating memory | |
2965 | * in the kernel. | |
2966 | * | |
2967 | * The @flags argument may be one of: | |
2968 | * | |
2969 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
2970 | * | |
2971 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
2972 | * | |
2973 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
2974 | * | |
2975 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
2976 | * must be suitable for DMA. This can mean different things on different | |
2977 | * platforms. For example, on i386, it means that the memory must come | |
2978 | * from the first 16MB. | |
2979 | */ | |
dd0fc66f | 2980 | void *__kmalloc(size_t size, gfp_t flags) |
1da177e4 LT |
2981 | { |
2982 | kmem_cache_t *cachep; | |
2983 | ||
97e2bde4 MS |
2984 | /* If you want to save a few bytes .text space: replace |
2985 | * __ with kmem_. | |
2986 | * Then kmalloc uses the uninlined functions instead of the inline | |
2987 | * functions. | |
2988 | */ | |
2989 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
2990 | if (unlikely(cachep == NULL)) |
2991 | return NULL; | |
1da177e4 LT |
2992 | return __cache_alloc(cachep, flags); |
2993 | } | |
2994 | EXPORT_SYMBOL(__kmalloc); | |
2995 | ||
2996 | #ifdef CONFIG_SMP | |
2997 | /** | |
2998 | * __alloc_percpu - allocate one copy of the object for every present | |
2999 | * cpu in the system, zeroing them. | |
3000 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
3001 | * | |
3002 | * @size: how many bytes of memory are required. | |
1da177e4 | 3003 | */ |
f9f75005 | 3004 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
3005 | { |
3006 | int i; | |
b28a02de | 3007 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3008 | |
3009 | if (!pdata) | |
3010 | return NULL; | |
3011 | ||
e498be7d CL |
3012 | /* |
3013 | * Cannot use for_each_online_cpu since a cpu may come online | |
3014 | * and we have no way of figuring out how to fix the array | |
3015 | * that we have allocated then.... | |
3016 | */ | |
3017 | for_each_cpu(i) { | |
3018 | int node = cpu_to_node(i); | |
3019 | ||
3020 | if (node_online(node)) | |
3021 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3022 | else | |
3023 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3024 | |
3025 | if (!pdata->ptrs[i]) | |
3026 | goto unwind_oom; | |
3027 | memset(pdata->ptrs[i], 0, size); | |
3028 | } | |
3029 | ||
3030 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3031 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3032 | |
b28a02de | 3033 | unwind_oom: |
1da177e4 LT |
3034 | while (--i >= 0) { |
3035 | if (!cpu_possible(i)) | |
3036 | continue; | |
3037 | kfree(pdata->ptrs[i]); | |
3038 | } | |
3039 | kfree(pdata); | |
3040 | return NULL; | |
3041 | } | |
3042 | EXPORT_SYMBOL(__alloc_percpu); | |
3043 | #endif | |
3044 | ||
3045 | /** | |
3046 | * kmem_cache_free - Deallocate an object | |
3047 | * @cachep: The cache the allocation was from. | |
3048 | * @objp: The previously allocated object. | |
3049 | * | |
3050 | * Free an object which was previously allocated from this | |
3051 | * cache. | |
3052 | */ | |
3053 | void kmem_cache_free(kmem_cache_t *cachep, void *objp) | |
3054 | { | |
3055 | unsigned long flags; | |
3056 | ||
3057 | local_irq_save(flags); | |
3058 | __cache_free(cachep, objp); | |
3059 | local_irq_restore(flags); | |
3060 | } | |
3061 | EXPORT_SYMBOL(kmem_cache_free); | |
3062 | ||
1da177e4 LT |
3063 | /** |
3064 | * kfree - free previously allocated memory | |
3065 | * @objp: pointer returned by kmalloc. | |
3066 | * | |
80e93eff PE |
3067 | * If @objp is NULL, no operation is performed. |
3068 | * | |
1da177e4 LT |
3069 | * Don't free memory not originally allocated by kmalloc() |
3070 | * or you will run into trouble. | |
3071 | */ | |
3072 | void kfree(const void *objp) | |
3073 | { | |
3074 | kmem_cache_t *c; | |
3075 | unsigned long flags; | |
3076 | ||
3077 | if (unlikely(!objp)) | |
3078 | return; | |
3079 | local_irq_save(flags); | |
3080 | kfree_debugcheck(objp); | |
065d41cb | 3081 | c = page_get_cache(virt_to_page(objp)); |
3dafccf2 | 3082 | mutex_debug_check_no_locks_freed(objp, obj_size(c)); |
b28a02de | 3083 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3084 | local_irq_restore(flags); |
3085 | } | |
3086 | EXPORT_SYMBOL(kfree); | |
3087 | ||
3088 | #ifdef CONFIG_SMP | |
3089 | /** | |
3090 | * free_percpu - free previously allocated percpu memory | |
3091 | * @objp: pointer returned by alloc_percpu. | |
3092 | * | |
3093 | * Don't free memory not originally allocated by alloc_percpu() | |
3094 | * The complemented objp is to check for that. | |
3095 | */ | |
b28a02de | 3096 | void free_percpu(const void *objp) |
1da177e4 LT |
3097 | { |
3098 | int i; | |
b28a02de | 3099 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3100 | |
e498be7d CL |
3101 | /* |
3102 | * We allocate for all cpus so we cannot use for online cpu here. | |
3103 | */ | |
3104 | for_each_cpu(i) | |
b28a02de | 3105 | kfree(p->ptrs[i]); |
1da177e4 LT |
3106 | kfree(p); |
3107 | } | |
3108 | EXPORT_SYMBOL(free_percpu); | |
3109 | #endif | |
3110 | ||
3111 | unsigned int kmem_cache_size(kmem_cache_t *cachep) | |
3112 | { | |
3dafccf2 | 3113 | return obj_size(cachep); |
1da177e4 LT |
3114 | } |
3115 | EXPORT_SYMBOL(kmem_cache_size); | |
3116 | ||
1944972d ACM |
3117 | const char *kmem_cache_name(kmem_cache_t *cachep) |
3118 | { | |
3119 | return cachep->name; | |
3120 | } | |
3121 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3122 | ||
e498be7d CL |
3123 | /* |
3124 | * This initializes kmem_list3 for all nodes. | |
3125 | */ | |
3126 | static int alloc_kmemlist(kmem_cache_t *cachep) | |
3127 | { | |
3128 | int node; | |
3129 | struct kmem_list3 *l3; | |
3130 | int err = 0; | |
3131 | ||
3132 | for_each_online_node(node) { | |
3133 | struct array_cache *nc = NULL, *new; | |
3134 | struct array_cache **new_alien = NULL; | |
3135 | #ifdef CONFIG_NUMA | |
3136 | if (!(new_alien = alloc_alien_cache(node, cachep->limit))) | |
3137 | goto fail; | |
3138 | #endif | |
b28a02de PE |
3139 | if (!(new = alloc_arraycache(node, (cachep->shared * |
3140 | cachep->batchcount), | |
3141 | 0xbaadf00d))) | |
e498be7d CL |
3142 | goto fail; |
3143 | if ((l3 = cachep->nodelists[node])) { | |
3144 | ||
3145 | spin_lock_irq(&l3->list_lock); | |
3146 | ||
3147 | if ((nc = cachep->nodelists[node]->shared)) | |
b28a02de | 3148 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
3149 | |
3150 | l3->shared = new; | |
3151 | if (!cachep->nodelists[node]->alien) { | |
3152 | l3->alien = new_alien; | |
3153 | new_alien = NULL; | |
3154 | } | |
b28a02de PE |
3155 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3156 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3157 | spin_unlock_irq(&l3->list_lock); |
3158 | kfree(nc); | |
3159 | free_alien_cache(new_alien); | |
3160 | continue; | |
3161 | } | |
3162 | if (!(l3 = kmalloc_node(sizeof(struct kmem_list3), | |
b28a02de | 3163 | GFP_KERNEL, node))) |
e498be7d CL |
3164 | goto fail; |
3165 | ||
3166 | kmem_list3_init(l3); | |
3167 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 3168 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
3169 | l3->shared = new; |
3170 | l3->alien = new_alien; | |
b28a02de PE |
3171 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3172 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3173 | cachep->nodelists[node] = l3; |
3174 | } | |
3175 | return err; | |
b28a02de | 3176 | fail: |
e498be7d CL |
3177 | err = -ENOMEM; |
3178 | return err; | |
3179 | } | |
3180 | ||
1da177e4 LT |
3181 | struct ccupdate_struct { |
3182 | kmem_cache_t *cachep; | |
3183 | struct array_cache *new[NR_CPUS]; | |
3184 | }; | |
3185 | ||
3186 | static void do_ccupdate_local(void *info) | |
3187 | { | |
3188 | struct ccupdate_struct *new = (struct ccupdate_struct *)info; | |
3189 | struct array_cache *old; | |
3190 | ||
3191 | check_irq_off(); | |
3192 | old = ac_data(new->cachep); | |
e498be7d | 3193 | |
1da177e4 LT |
3194 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3195 | new->new[smp_processor_id()] = old; | |
3196 | } | |
3197 | ||
1da177e4 | 3198 | static int do_tune_cpucache(kmem_cache_t *cachep, int limit, int batchcount, |
b28a02de | 3199 | int shared) |
1da177e4 LT |
3200 | { |
3201 | struct ccupdate_struct new; | |
e498be7d | 3202 | int i, err; |
1da177e4 | 3203 | |
b28a02de | 3204 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3205 | for_each_online_cpu(i) { |
b28a02de PE |
3206 | new.new[i] = |
3207 | alloc_arraycache(cpu_to_node(i), limit, batchcount); | |
e498be7d | 3208 | if (!new.new[i]) { |
b28a02de PE |
3209 | for (i--; i >= 0; i--) |
3210 | kfree(new.new[i]); | |
e498be7d | 3211 | return -ENOMEM; |
1da177e4 LT |
3212 | } |
3213 | } | |
3214 | new.cachep = cachep; | |
3215 | ||
3216 | smp_call_function_all_cpus(do_ccupdate_local, (void *)&new); | |
e498be7d | 3217 | |
1da177e4 LT |
3218 | check_irq_on(); |
3219 | spin_lock_irq(&cachep->spinlock); | |
3220 | cachep->batchcount = batchcount; | |
3221 | cachep->limit = limit; | |
e498be7d | 3222 | cachep->shared = shared; |
1da177e4 LT |
3223 | spin_unlock_irq(&cachep->spinlock); |
3224 | ||
e498be7d | 3225 | for_each_online_cpu(i) { |
1da177e4 LT |
3226 | struct array_cache *ccold = new.new[i]; |
3227 | if (!ccold) | |
3228 | continue; | |
e498be7d | 3229 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3230 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3231 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3232 | kfree(ccold); |
3233 | } | |
1da177e4 | 3234 | |
e498be7d CL |
3235 | err = alloc_kmemlist(cachep); |
3236 | if (err) { | |
3237 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3238 | cachep->name, -err); |
e498be7d | 3239 | BUG(); |
1da177e4 | 3240 | } |
1da177e4 LT |
3241 | return 0; |
3242 | } | |
3243 | ||
1da177e4 LT |
3244 | static void enable_cpucache(kmem_cache_t *cachep) |
3245 | { | |
3246 | int err; | |
3247 | int limit, shared; | |
3248 | ||
3249 | /* The head array serves three purposes: | |
3250 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
3251 | * - reduce the number of spinlock operations. | |
3252 | * - reduce the number of linked list operations on the slab and | |
3253 | * bufctl chains: array operations are cheaper. | |
3254 | * The numbers are guessed, we should auto-tune as described by | |
3255 | * Bonwick. | |
3256 | */ | |
3dafccf2 | 3257 | if (cachep->buffer_size > 131072) |
1da177e4 | 3258 | limit = 1; |
3dafccf2 | 3259 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3260 | limit = 8; |
3dafccf2 | 3261 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3262 | limit = 24; |
3dafccf2 | 3263 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3264 | limit = 54; |
3265 | else | |
3266 | limit = 120; | |
3267 | ||
3268 | /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound | |
3269 | * allocation behaviour: Most allocs on one cpu, most free operations | |
3270 | * on another cpu. For these cases, an efficient object passing between | |
3271 | * cpus is necessary. This is provided by a shared array. The array | |
3272 | * replaces Bonwick's magazine layer. | |
3273 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3274 | * to a larger limit. Thus disabled by default. | |
3275 | */ | |
3276 | shared = 0; | |
3277 | #ifdef CONFIG_SMP | |
3dafccf2 | 3278 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3279 | shared = 8; |
3280 | #endif | |
3281 | ||
3282 | #if DEBUG | |
3283 | /* With debugging enabled, large batchcount lead to excessively | |
3284 | * long periods with disabled local interrupts. Limit the | |
3285 | * batchcount | |
3286 | */ | |
3287 | if (limit > 32) | |
3288 | limit = 32; | |
3289 | #endif | |
b28a02de | 3290 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3291 | if (err) |
3292 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3293 | cachep->name, -err); |
1da177e4 LT |
3294 | } |
3295 | ||
b28a02de PE |
3296 | static void drain_array_locked(kmem_cache_t *cachep, struct array_cache *ac, |
3297 | int force, int node) | |
1da177e4 LT |
3298 | { |
3299 | int tofree; | |
3300 | ||
e498be7d | 3301 | check_spinlock_acquired_node(cachep, node); |
1da177e4 LT |
3302 | if (ac->touched && !force) { |
3303 | ac->touched = 0; | |
3304 | } else if (ac->avail) { | |
b28a02de | 3305 | tofree = force ? ac->avail : (ac->limit + 4) / 5; |
1da177e4 | 3306 | if (tofree > ac->avail) { |
b28a02de | 3307 | tofree = (ac->avail + 1) / 2; |
1da177e4 | 3308 | } |
ff69416e | 3309 | free_block(cachep, ac->entry, tofree, node); |
1da177e4 | 3310 | ac->avail -= tofree; |
e498be7d | 3311 | memmove(ac->entry, &(ac->entry[tofree]), |
b28a02de | 3312 | sizeof(void *) * ac->avail); |
1da177e4 LT |
3313 | } |
3314 | } | |
3315 | ||
3316 | /** | |
3317 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3318 | * @unused: unused parameter |
1da177e4 LT |
3319 | * |
3320 | * Called from workqueue/eventd every few seconds. | |
3321 | * Purpose: | |
3322 | * - clear the per-cpu caches for this CPU. | |
3323 | * - return freeable pages to the main free memory pool. | |
3324 | * | |
fc0abb14 | 3325 | * If we cannot acquire the cache chain mutex then just give up - we'll |
1da177e4 LT |
3326 | * try again on the next iteration. |
3327 | */ | |
3328 | static void cache_reap(void *unused) | |
3329 | { | |
3330 | struct list_head *walk; | |
e498be7d | 3331 | struct kmem_list3 *l3; |
1da177e4 | 3332 | |
fc0abb14 | 3333 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3334 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3335 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3336 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3337 | return; |
3338 | } | |
3339 | ||
3340 | list_for_each(walk, &cache_chain) { | |
3341 | kmem_cache_t *searchp; | |
b28a02de | 3342 | struct list_head *p; |
1da177e4 LT |
3343 | int tofree; |
3344 | struct slab *slabp; | |
3345 | ||
3346 | searchp = list_entry(walk, kmem_cache_t, next); | |
3347 | ||
3348 | if (searchp->flags & SLAB_NO_REAP) | |
3349 | goto next; | |
3350 | ||
3351 | check_irq_on(); | |
3352 | ||
e498be7d CL |
3353 | l3 = searchp->nodelists[numa_node_id()]; |
3354 | if (l3->alien) | |
3355 | drain_alien_cache(searchp, l3); | |
3356 | spin_lock_irq(&l3->list_lock); | |
1da177e4 | 3357 | |
e498be7d | 3358 | drain_array_locked(searchp, ac_data(searchp), 0, |
b28a02de | 3359 | numa_node_id()); |
1da177e4 | 3360 | |
e498be7d | 3361 | if (time_after(l3->next_reap, jiffies)) |
1da177e4 LT |
3362 | goto next_unlock; |
3363 | ||
e498be7d | 3364 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3365 | |
e498be7d CL |
3366 | if (l3->shared) |
3367 | drain_array_locked(searchp, l3->shared, 0, | |
b28a02de | 3368 | numa_node_id()); |
1da177e4 | 3369 | |
e498be7d CL |
3370 | if (l3->free_touched) { |
3371 | l3->free_touched = 0; | |
1da177e4 LT |
3372 | goto next_unlock; |
3373 | } | |
3374 | ||
b28a02de PE |
3375 | tofree = |
3376 | (l3->free_limit + 5 * searchp->num - | |
3377 | 1) / (5 * searchp->num); | |
1da177e4 | 3378 | do { |
e498be7d CL |
3379 | p = l3->slabs_free.next; |
3380 | if (p == &(l3->slabs_free)) | |
1da177e4 LT |
3381 | break; |
3382 | ||
3383 | slabp = list_entry(p, struct slab, list); | |
3384 | BUG_ON(slabp->inuse); | |
3385 | list_del(&slabp->list); | |
3386 | STATS_INC_REAPED(searchp); | |
3387 | ||
3388 | /* Safe to drop the lock. The slab is no longer | |
3389 | * linked to the cache. | |
3390 | * searchp cannot disappear, we hold | |
3391 | * cache_chain_lock | |
3392 | */ | |
e498be7d CL |
3393 | l3->free_objects -= searchp->num; |
3394 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3395 | slab_destroy(searchp, slabp); |
e498be7d | 3396 | spin_lock_irq(&l3->list_lock); |
b28a02de PE |
3397 | } while (--tofree > 0); |
3398 | next_unlock: | |
e498be7d | 3399 | spin_unlock_irq(&l3->list_lock); |
b28a02de | 3400 | next: |
1da177e4 LT |
3401 | cond_resched(); |
3402 | } | |
3403 | check_irq_on(); | |
fc0abb14 | 3404 | mutex_unlock(&cache_chain_mutex); |
4ae7c039 | 3405 | drain_remote_pages(); |
1da177e4 | 3406 | /* Setup the next iteration */ |
cd61ef62 | 3407 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3408 | } |
3409 | ||
3410 | #ifdef CONFIG_PROC_FS | |
3411 | ||
85289f98 | 3412 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3413 | { |
85289f98 PE |
3414 | /* |
3415 | * Output format version, so at least we can change it | |
3416 | * without _too_ many complaints. | |
3417 | */ | |
1da177e4 | 3418 | #if STATS |
85289f98 | 3419 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3420 | #else |
85289f98 | 3421 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3422 | #endif |
85289f98 PE |
3423 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3424 | "<objperslab> <pagesperslab>"); | |
3425 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3426 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3427 | #if STATS |
85289f98 PE |
3428 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
3429 | "<error> <maxfreeable> <nodeallocs> <remotefrees>"); | |
3430 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
1da177e4 | 3431 | #endif |
85289f98 PE |
3432 | seq_putc(m, '\n'); |
3433 | } | |
3434 | ||
3435 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3436 | { | |
3437 | loff_t n = *pos; | |
3438 | struct list_head *p; | |
3439 | ||
fc0abb14 | 3440 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3441 | if (!n) |
3442 | print_slabinfo_header(m); | |
1da177e4 LT |
3443 | p = cache_chain.next; |
3444 | while (n--) { | |
3445 | p = p->next; | |
3446 | if (p == &cache_chain) | |
3447 | return NULL; | |
3448 | } | |
3449 | return list_entry(p, kmem_cache_t, next); | |
3450 | } | |
3451 | ||
3452 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3453 | { | |
3454 | kmem_cache_t *cachep = p; | |
3455 | ++*pos; | |
3456 | return cachep->next.next == &cache_chain ? NULL | |
b28a02de | 3457 | : list_entry(cachep->next.next, kmem_cache_t, next); |
1da177e4 LT |
3458 | } |
3459 | ||
3460 | static void s_stop(struct seq_file *m, void *p) | |
3461 | { | |
fc0abb14 | 3462 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3463 | } |
3464 | ||
3465 | static int s_show(struct seq_file *m, void *p) | |
3466 | { | |
3467 | kmem_cache_t *cachep = p; | |
3468 | struct list_head *q; | |
b28a02de PE |
3469 | struct slab *slabp; |
3470 | unsigned long active_objs; | |
3471 | unsigned long num_objs; | |
3472 | unsigned long active_slabs = 0; | |
3473 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3474 | const char *name; |
1da177e4 | 3475 | char *error = NULL; |
e498be7d CL |
3476 | int node; |
3477 | struct kmem_list3 *l3; | |
1da177e4 LT |
3478 | |
3479 | check_irq_on(); | |
3480 | spin_lock_irq(&cachep->spinlock); | |
3481 | active_objs = 0; | |
3482 | num_slabs = 0; | |
e498be7d CL |
3483 | for_each_online_node(node) { |
3484 | l3 = cachep->nodelists[node]; | |
3485 | if (!l3) | |
3486 | continue; | |
3487 | ||
3488 | spin_lock(&l3->list_lock); | |
3489 | ||
b28a02de | 3490 | list_for_each(q, &l3->slabs_full) { |
e498be7d CL |
3491 | slabp = list_entry(q, struct slab, list); |
3492 | if (slabp->inuse != cachep->num && !error) | |
3493 | error = "slabs_full accounting error"; | |
3494 | active_objs += cachep->num; | |
3495 | active_slabs++; | |
3496 | } | |
b28a02de | 3497 | list_for_each(q, &l3->slabs_partial) { |
e498be7d CL |
3498 | slabp = list_entry(q, struct slab, list); |
3499 | if (slabp->inuse == cachep->num && !error) | |
3500 | error = "slabs_partial inuse accounting error"; | |
3501 | if (!slabp->inuse && !error) | |
3502 | error = "slabs_partial/inuse accounting error"; | |
3503 | active_objs += slabp->inuse; | |
3504 | active_slabs++; | |
3505 | } | |
b28a02de | 3506 | list_for_each(q, &l3->slabs_free) { |
e498be7d CL |
3507 | slabp = list_entry(q, struct slab, list); |
3508 | if (slabp->inuse && !error) | |
3509 | error = "slabs_free/inuse accounting error"; | |
3510 | num_slabs++; | |
3511 | } | |
3512 | free_objects += l3->free_objects; | |
3513 | shared_avail += l3->shared->avail; | |
3514 | ||
3515 | spin_unlock(&l3->list_lock); | |
1da177e4 | 3516 | } |
b28a02de PE |
3517 | num_slabs += active_slabs; |
3518 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3519 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3520 | error = "free_objects accounting error"; |
3521 | ||
b28a02de | 3522 | name = cachep->name; |
1da177e4 LT |
3523 | if (error) |
3524 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3525 | ||
3526 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3527 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3528 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3529 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3530 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3531 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3532 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3533 | #if STATS |
b28a02de | 3534 | { /* list3 stats */ |
1da177e4 LT |
3535 | unsigned long high = cachep->high_mark; |
3536 | unsigned long allocs = cachep->num_allocations; | |
3537 | unsigned long grown = cachep->grown; | |
3538 | unsigned long reaped = cachep->reaped; | |
3539 | unsigned long errors = cachep->errors; | |
3540 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3541 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3542 | unsigned long node_frees = cachep->node_frees; |
1da177e4 | 3543 | |
e498be7d | 3544 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
b28a02de | 3545 | %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees); |
1da177e4 LT |
3546 | } |
3547 | /* cpu stats */ | |
3548 | { | |
3549 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3550 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3551 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3552 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3553 | ||
3554 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3555 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3556 | } |
3557 | #endif | |
3558 | seq_putc(m, '\n'); | |
3559 | spin_unlock_irq(&cachep->spinlock); | |
3560 | return 0; | |
3561 | } | |
3562 | ||
3563 | /* | |
3564 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3565 | * | |
3566 | * Output layout: | |
3567 | * cache-name | |
3568 | * num-active-objs | |
3569 | * total-objs | |
3570 | * object size | |
3571 | * num-active-slabs | |
3572 | * total-slabs | |
3573 | * num-pages-per-slab | |
3574 | * + further values on SMP and with statistics enabled | |
3575 | */ | |
3576 | ||
3577 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3578 | .start = s_start, |
3579 | .next = s_next, | |
3580 | .stop = s_stop, | |
3581 | .show = s_show, | |
1da177e4 LT |
3582 | }; |
3583 | ||
3584 | #define MAX_SLABINFO_WRITE 128 | |
3585 | /** | |
3586 | * slabinfo_write - Tuning for the slab allocator | |
3587 | * @file: unused | |
3588 | * @buffer: user buffer | |
3589 | * @count: data length | |
3590 | * @ppos: unused | |
3591 | */ | |
b28a02de PE |
3592 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3593 | size_t count, loff_t *ppos) | |
1da177e4 | 3594 | { |
b28a02de | 3595 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 LT |
3596 | int limit, batchcount, shared, res; |
3597 | struct list_head *p; | |
b28a02de | 3598 | |
1da177e4 LT |
3599 | if (count > MAX_SLABINFO_WRITE) |
3600 | return -EINVAL; | |
3601 | if (copy_from_user(&kbuf, buffer, count)) | |
3602 | return -EFAULT; | |
b28a02de | 3603 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3604 | |
3605 | tmp = strchr(kbuf, ' '); | |
3606 | if (!tmp) | |
3607 | return -EINVAL; | |
3608 | *tmp = '\0'; | |
3609 | tmp++; | |
3610 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3611 | return -EINVAL; | |
3612 | ||
3613 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3614 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3615 | res = -EINVAL; |
b28a02de | 3616 | list_for_each(p, &cache_chain) { |
1da177e4 LT |
3617 | kmem_cache_t *cachep = list_entry(p, kmem_cache_t, next); |
3618 | ||
3619 | if (!strcmp(cachep->name, kbuf)) { | |
3620 | if (limit < 1 || | |
3621 | batchcount < 1 || | |
b28a02de | 3622 | batchcount > limit || shared < 0) { |
e498be7d | 3623 | res = 0; |
1da177e4 | 3624 | } else { |
e498be7d | 3625 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3626 | batchcount, shared); |
1da177e4 LT |
3627 | } |
3628 | break; | |
3629 | } | |
3630 | } | |
fc0abb14 | 3631 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3632 | if (res >= 0) |
3633 | res = count; | |
3634 | return res; | |
3635 | } | |
3636 | #endif | |
3637 | ||
00e145b6 MS |
3638 | /** |
3639 | * ksize - get the actual amount of memory allocated for a given object | |
3640 | * @objp: Pointer to the object | |
3641 | * | |
3642 | * kmalloc may internally round up allocations and return more memory | |
3643 | * than requested. ksize() can be used to determine the actual amount of | |
3644 | * memory allocated. The caller may use this additional memory, even though | |
3645 | * a smaller amount of memory was initially specified with the kmalloc call. | |
3646 | * The caller must guarantee that objp points to a valid object previously | |
3647 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
3648 | * must not be freed during the duration of the call. | |
3649 | */ | |
1da177e4 LT |
3650 | unsigned int ksize(const void *objp) |
3651 | { | |
00e145b6 MS |
3652 | if (unlikely(objp == NULL)) |
3653 | return 0; | |
1da177e4 | 3654 | |
3dafccf2 | 3655 | return obj_size(page_get_cache(virt_to_page(objp))); |
1da177e4 | 3656 | } |