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