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. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
fc0abb14 | 71 | * The global cache-chain is protected by the mutex 'cache_chain_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
89 | #include <linux/config.h> | |
90 | #include <linux/slab.h> | |
91 | #include <linux/mm.h> | |
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 | 246 | struct rcu_head head; |
343e0d7a | 247 | struct kmem_cache *cachep; |
b28a02de | 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; | |
2e1217cf | 297 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
298 | spinlock_t list_lock; |
299 | struct array_cache *shared; /* shared per node */ | |
300 | struct array_cache **alien; /* on other nodes */ | |
1da177e4 LT |
301 | }; |
302 | ||
e498be7d CL |
303 | /* |
304 | * Need this for bootstrapping a per node allocator. | |
305 | */ | |
306 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES + 1) | |
307 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
308 | #define CACHE_CACHE 0 | |
309 | #define SIZE_AC 1 | |
310 | #define SIZE_L3 (1 + MAX_NUMNODES) | |
311 | ||
312 | /* | |
7243cc05 | 313 | * This function must be completely optimized away if |
e498be7d CL |
314 | * a constant is passed to it. Mostly the same as |
315 | * what is in linux/slab.h except it returns an | |
316 | * index. | |
317 | */ | |
7243cc05 | 318 | static __always_inline int index_of(const size_t size) |
e498be7d | 319 | { |
5ec8a847 SR |
320 | extern void __bad_size(void); |
321 | ||
e498be7d CL |
322 | if (__builtin_constant_p(size)) { |
323 | int i = 0; | |
324 | ||
325 | #define CACHE(x) \ | |
326 | if (size <=x) \ | |
327 | return i; \ | |
328 | else \ | |
329 | i++; | |
330 | #include "linux/kmalloc_sizes.h" | |
331 | #undef CACHE | |
5ec8a847 | 332 | __bad_size(); |
7243cc05 | 333 | } else |
5ec8a847 | 334 | __bad_size(); |
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 | |
5295a74c | 341 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
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; | |
2e1217cf | 348 | parent->colour_next = 0; |
e498be7d CL |
349 | spin_lock_init(&parent->list_lock); |
350 | parent->free_objects = 0; | |
351 | parent->free_touched = 0; | |
352 | } | |
353 | ||
354 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ | |
355 | do { \ | |
356 | INIT_LIST_HEAD(listp); \ | |
357 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
358 | } while (0) | |
359 | ||
360 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ | |
361 | do { \ | |
362 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ | |
363 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
364 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
365 | } while (0) | |
1da177e4 LT |
366 | |
367 | /* | |
343e0d7a | 368 | * struct kmem_cache |
1da177e4 LT |
369 | * |
370 | * manages a cache. | |
371 | */ | |
b28a02de | 372 | |
2109a2d1 | 373 | struct kmem_cache { |
1da177e4 | 374 | /* 1) per-cpu data, touched during every alloc/free */ |
b28a02de PE |
375 | struct array_cache *array[NR_CPUS]; |
376 | unsigned int batchcount; | |
377 | unsigned int limit; | |
378 | unsigned int shared; | |
3dafccf2 | 379 | unsigned int buffer_size; |
e498be7d | 380 | /* 2) touched by every alloc & free from the backend */ |
b28a02de PE |
381 | struct kmem_list3 *nodelists[MAX_NUMNODES]; |
382 | unsigned int flags; /* constant flags */ | |
383 | unsigned int num; /* # of objs per slab */ | |
384 | spinlock_t spinlock; | |
1da177e4 LT |
385 | |
386 | /* 3) cache_grow/shrink */ | |
387 | /* order of pgs per slab (2^n) */ | |
b28a02de | 388 | unsigned int gfporder; |
1da177e4 LT |
389 | |
390 | /* force GFP flags, e.g. GFP_DMA */ | |
b28a02de | 391 | gfp_t gfpflags; |
1da177e4 | 392 | |
b28a02de PE |
393 | size_t colour; /* cache colouring range */ |
394 | unsigned int colour_off; /* colour offset */ | |
343e0d7a | 395 | struct kmem_cache *slabp_cache; |
b28a02de PE |
396 | unsigned int slab_size; |
397 | unsigned int dflags; /* dynamic flags */ | |
1da177e4 LT |
398 | |
399 | /* constructor func */ | |
343e0d7a | 400 | void (*ctor) (void *, struct kmem_cache *, unsigned long); |
1da177e4 LT |
401 | |
402 | /* de-constructor func */ | |
343e0d7a | 403 | void (*dtor) (void *, struct kmem_cache *, 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 | */ |
343e0d7a | 513 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 514 | { |
3dafccf2 | 515 | return cachep->obj_offset; |
1da177e4 LT |
516 | } |
517 | ||
343e0d7a | 518 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 519 | { |
3dafccf2 | 520 | return cachep->obj_size; |
1da177e4 LT |
521 | } |
522 | ||
343e0d7a | 523 | static unsigned long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
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 | ||
343e0d7a | 529 | static unsigned long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
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 | ||
343e0d7a | 538 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
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 | 599 | |
6ed5eb22 PE |
600 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
601 | { | |
602 | struct page *page = virt_to_page(obj); | |
603 | return page_get_cache(page); | |
604 | } | |
605 | ||
606 | static inline struct slab *virt_to_slab(const void *obj) | |
607 | { | |
608 | struct page *page = virt_to_page(obj); | |
609 | return page_get_slab(page); | |
610 | } | |
611 | ||
1da177e4 LT |
612 | /* These are the default caches for kmalloc. Custom caches can have other sizes. */ |
613 | struct cache_sizes malloc_sizes[] = { | |
614 | #define CACHE(x) { .cs_size = (x) }, | |
615 | #include <linux/kmalloc_sizes.h> | |
616 | CACHE(ULONG_MAX) | |
617 | #undef CACHE | |
618 | }; | |
619 | EXPORT_SYMBOL(malloc_sizes); | |
620 | ||
621 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
622 | struct cache_names { | |
623 | char *name; | |
624 | char *name_dma; | |
625 | }; | |
626 | ||
627 | static struct cache_names __initdata cache_names[] = { | |
628 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
629 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 630 | {NULL,} |
1da177e4 LT |
631 | #undef CACHE |
632 | }; | |
633 | ||
634 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 635 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 636 | static struct arraycache_init initarray_generic = |
b28a02de | 637 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
638 | |
639 | /* internal cache of cache description objs */ | |
343e0d7a | 640 | static struct kmem_cache cache_cache = { |
b28a02de PE |
641 | .batchcount = 1, |
642 | .limit = BOOT_CPUCACHE_ENTRIES, | |
643 | .shared = 1, | |
343e0d7a | 644 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de PE |
645 | .flags = SLAB_NO_REAP, |
646 | .spinlock = SPIN_LOCK_UNLOCKED, | |
647 | .name = "kmem_cache", | |
1da177e4 | 648 | #if DEBUG |
343e0d7a | 649 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
650 | #endif |
651 | }; | |
652 | ||
653 | /* Guard access to the cache-chain. */ | |
fc0abb14 | 654 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
655 | static struct list_head cache_chain; |
656 | ||
657 | /* | |
658 | * vm_enough_memory() looks at this to determine how many | |
659 | * slab-allocated pages are possibly freeable under pressure | |
660 | * | |
661 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
662 | */ | |
663 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
664 | |
665 | /* | |
666 | * chicken and egg problem: delay the per-cpu array allocation | |
667 | * until the general caches are up. | |
668 | */ | |
669 | static enum { | |
670 | NONE, | |
e498be7d CL |
671 | PARTIAL_AC, |
672 | PARTIAL_L3, | |
1da177e4 LT |
673 | FULL |
674 | } g_cpucache_up; | |
675 | ||
676 | static DEFINE_PER_CPU(struct work_struct, reap_work); | |
677 | ||
343e0d7a PE |
678 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node); |
679 | static void enable_cpucache(struct kmem_cache *cachep); | |
b28a02de | 680 | static void cache_reap(void *unused); |
343e0d7a | 681 | static int __node_shrink(struct kmem_cache *cachep, int node); |
1da177e4 | 682 | |
343e0d7a | 683 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
684 | { |
685 | return cachep->array[smp_processor_id()]; | |
686 | } | |
687 | ||
343e0d7a | 688 | static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags) |
1da177e4 LT |
689 | { |
690 | struct cache_sizes *csizep = malloc_sizes; | |
691 | ||
692 | #if DEBUG | |
693 | /* This happens if someone tries to call | |
b28a02de PE |
694 | * kmem_cache_create(), or __kmalloc(), before |
695 | * the generic caches are initialized. | |
696 | */ | |
c7e43c78 | 697 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
698 | #endif |
699 | while (size > csizep->cs_size) | |
700 | csizep++; | |
701 | ||
702 | /* | |
0abf40c1 | 703 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
704 | * has cs_{dma,}cachep==NULL. Thus no special case |
705 | * for large kmalloc calls required. | |
706 | */ | |
707 | if (unlikely(gfpflags & GFP_DMA)) | |
708 | return csizep->cs_dmacachep; | |
709 | return csizep->cs_cachep; | |
710 | } | |
711 | ||
343e0d7a | 712 | struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
713 | { |
714 | return __find_general_cachep(size, gfpflags); | |
715 | } | |
716 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
717 | ||
fbaccacf | 718 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 719 | { |
fbaccacf SR |
720 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
721 | } | |
1da177e4 | 722 | |
fbaccacf SR |
723 | /* Calculate the number of objects and left-over bytes for a given |
724 | buffer size. */ | |
725 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, | |
726 | size_t align, int flags, size_t *left_over, | |
727 | unsigned int *num) | |
728 | { | |
729 | int nr_objs; | |
730 | size_t mgmt_size; | |
731 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 732 | |
fbaccacf SR |
733 | /* |
734 | * The slab management structure can be either off the slab or | |
735 | * on it. For the latter case, the memory allocated for a | |
736 | * slab is used for: | |
737 | * | |
738 | * - The struct slab | |
739 | * - One kmem_bufctl_t for each object | |
740 | * - Padding to respect alignment of @align | |
741 | * - @buffer_size bytes for each object | |
742 | * | |
743 | * If the slab management structure is off the slab, then the | |
744 | * alignment will already be calculated into the size. Because | |
745 | * the slabs are all pages aligned, the objects will be at the | |
746 | * correct alignment when allocated. | |
747 | */ | |
748 | if (flags & CFLGS_OFF_SLAB) { | |
749 | mgmt_size = 0; | |
750 | nr_objs = slab_size / buffer_size; | |
751 | ||
752 | if (nr_objs > SLAB_LIMIT) | |
753 | nr_objs = SLAB_LIMIT; | |
754 | } else { | |
755 | /* | |
756 | * Ignore padding for the initial guess. The padding | |
757 | * is at most @align-1 bytes, and @buffer_size is at | |
758 | * least @align. In the worst case, this result will | |
759 | * be one greater than the number of objects that fit | |
760 | * into the memory allocation when taking the padding | |
761 | * into account. | |
762 | */ | |
763 | nr_objs = (slab_size - sizeof(struct slab)) / | |
764 | (buffer_size + sizeof(kmem_bufctl_t)); | |
765 | ||
766 | /* | |
767 | * This calculated number will be either the right | |
768 | * amount, or one greater than what we want. | |
769 | */ | |
770 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
771 | > slab_size) | |
772 | nr_objs--; | |
773 | ||
774 | if (nr_objs > SLAB_LIMIT) | |
775 | nr_objs = SLAB_LIMIT; | |
776 | ||
777 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
778 | } | |
779 | *num = nr_objs; | |
780 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
781 | } |
782 | ||
783 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
784 | ||
343e0d7a | 785 | static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg) |
1da177e4 LT |
786 | { |
787 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 788 | function, cachep->name, msg); |
1da177e4 LT |
789 | dump_stack(); |
790 | } | |
791 | ||
792 | /* | |
793 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
794 | * via the workqueue/eventd. | |
795 | * Add the CPU number into the expiration time to minimize the possibility of | |
796 | * the CPUs getting into lockstep and contending for the global cache chain | |
797 | * lock. | |
798 | */ | |
799 | static void __devinit start_cpu_timer(int cpu) | |
800 | { | |
801 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
802 | ||
803 | /* | |
804 | * When this gets called from do_initcalls via cpucache_init(), | |
805 | * init_workqueues() has already run, so keventd will be setup | |
806 | * at that time. | |
807 | */ | |
808 | if (keventd_up() && reap_work->func == NULL) { | |
809 | INIT_WORK(reap_work, cache_reap, NULL); | |
810 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
811 | } | |
812 | } | |
813 | ||
e498be7d | 814 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 815 | int batchcount) |
1da177e4 | 816 | { |
b28a02de | 817 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
818 | struct array_cache *nc = NULL; |
819 | ||
e498be7d | 820 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
821 | if (nc) { |
822 | nc->avail = 0; | |
823 | nc->limit = entries; | |
824 | nc->batchcount = batchcount; | |
825 | nc->touched = 0; | |
e498be7d | 826 | spin_lock_init(&nc->lock); |
1da177e4 LT |
827 | } |
828 | return nc; | |
829 | } | |
830 | ||
e498be7d | 831 | #ifdef CONFIG_NUMA |
343e0d7a | 832 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
dc85da15 | 833 | |
5295a74c | 834 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
835 | { |
836 | struct array_cache **ac_ptr; | |
b28a02de | 837 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
838 | int i; |
839 | ||
840 | if (limit > 1) | |
841 | limit = 12; | |
842 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
843 | if (ac_ptr) { | |
844 | for_each_node(i) { | |
845 | if (i == node || !node_online(i)) { | |
846 | ac_ptr[i] = NULL; | |
847 | continue; | |
848 | } | |
849 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
850 | if (!ac_ptr[i]) { | |
b28a02de | 851 | for (i--; i <= 0; i--) |
e498be7d CL |
852 | kfree(ac_ptr[i]); |
853 | kfree(ac_ptr); | |
854 | return NULL; | |
855 | } | |
856 | } | |
857 | } | |
858 | return ac_ptr; | |
859 | } | |
860 | ||
5295a74c | 861 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
862 | { |
863 | int i; | |
864 | ||
865 | if (!ac_ptr) | |
866 | return; | |
867 | ||
868 | for_each_node(i) | |
b28a02de | 869 | kfree(ac_ptr[i]); |
e498be7d CL |
870 | |
871 | kfree(ac_ptr); | |
872 | } | |
873 | ||
343e0d7a | 874 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 875 | struct array_cache *ac, int node) |
e498be7d CL |
876 | { |
877 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
878 | ||
879 | if (ac->avail) { | |
880 | spin_lock(&rl3->list_lock); | |
ff69416e | 881 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
882 | ac->avail = 0; |
883 | spin_unlock(&rl3->list_lock); | |
884 | } | |
885 | } | |
886 | ||
4484ebf1 | 887 | static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien) |
e498be7d | 888 | { |
b28a02de | 889 | int i = 0; |
e498be7d CL |
890 | struct array_cache *ac; |
891 | unsigned long flags; | |
892 | ||
893 | for_each_online_node(i) { | |
4484ebf1 | 894 | ac = alien[i]; |
e498be7d CL |
895 | if (ac) { |
896 | spin_lock_irqsave(&ac->lock, flags); | |
897 | __drain_alien_cache(cachep, ac, i); | |
898 | spin_unlock_irqrestore(&ac->lock, flags); | |
899 | } | |
900 | } | |
901 | } | |
902 | #else | |
7a21ef6f | 903 | |
4484ebf1 RT |
904 | #define drain_alien_cache(cachep, alien) do { } while (0) |
905 | ||
7a21ef6f LT |
906 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
907 | { | |
908 | return (struct array_cache **) 0x01020304ul; | |
909 | } | |
910 | ||
4484ebf1 RT |
911 | static inline void free_alien_cache(struct array_cache **ac_ptr) |
912 | { | |
913 | } | |
7a21ef6f | 914 | |
e498be7d CL |
915 | #endif |
916 | ||
1da177e4 | 917 | static int __devinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 918 | unsigned long action, void *hcpu) |
1da177e4 LT |
919 | { |
920 | long cpu = (long)hcpu; | |
343e0d7a | 921 | struct kmem_cache *cachep; |
e498be7d CL |
922 | struct kmem_list3 *l3 = NULL; |
923 | int node = cpu_to_node(cpu); | |
924 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
925 | |
926 | switch (action) { | |
927 | case CPU_UP_PREPARE: | |
fc0abb14 | 928 | mutex_lock(&cache_chain_mutex); |
e498be7d CL |
929 | /* we need to do this right in the beginning since |
930 | * alloc_arraycache's are going to use this list. | |
931 | * kmalloc_node allows us to add the slab to the right | |
932 | * kmem_list3 and not this cpu's kmem_list3 | |
933 | */ | |
934 | ||
1da177e4 | 935 | list_for_each_entry(cachep, &cache_chain, next) { |
e498be7d CL |
936 | /* setup the size64 kmemlist for cpu before we can |
937 | * begin anything. Make sure some other cpu on this | |
938 | * node has not already allocated this | |
939 | */ | |
940 | if (!cachep->nodelists[node]) { | |
941 | if (!(l3 = kmalloc_node(memsize, | |
b28a02de | 942 | GFP_KERNEL, node))) |
e498be7d CL |
943 | goto bad; |
944 | kmem_list3_init(l3); | |
945 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 946 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 947 | |
4484ebf1 RT |
948 | /* |
949 | * The l3s don't come and go as CPUs come and | |
950 | * go. cache_chain_mutex is sufficient | |
951 | * protection here. | |
952 | */ | |
e498be7d CL |
953 | cachep->nodelists[node] = l3; |
954 | } | |
1da177e4 | 955 | |
e498be7d CL |
956 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
957 | cachep->nodelists[node]->free_limit = | |
b28a02de PE |
958 | (1 + nr_cpus_node(node)) * |
959 | cachep->batchcount + cachep->num; | |
e498be7d CL |
960 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
961 | } | |
962 | ||
963 | /* Now we can go ahead with allocating the shared array's | |
b28a02de | 964 | & array cache's */ |
e498be7d | 965 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 966 | struct array_cache *nc; |
4484ebf1 RT |
967 | struct array_cache *shared; |
968 | struct array_cache **alien; | |
cd105df4 | 969 | |
e498be7d | 970 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 971 | cachep->batchcount); |
1da177e4 LT |
972 | if (!nc) |
973 | goto bad; | |
4484ebf1 RT |
974 | shared = alloc_arraycache(node, |
975 | cachep->shared * cachep->batchcount, | |
976 | 0xbaadf00d); | |
977 | if (!shared) | |
978 | goto bad; | |
7a21ef6f | 979 | |
4484ebf1 RT |
980 | alien = alloc_alien_cache(node, cachep->limit); |
981 | if (!alien) | |
982 | goto bad; | |
1da177e4 | 983 | cachep->array[cpu] = nc; |
1da177e4 | 984 | |
e498be7d CL |
985 | l3 = cachep->nodelists[node]; |
986 | BUG_ON(!l3); | |
e498be7d | 987 | |
4484ebf1 RT |
988 | spin_lock_irq(&l3->list_lock); |
989 | if (!l3->shared) { | |
990 | /* | |
991 | * We are serialised from CPU_DEAD or | |
992 | * CPU_UP_CANCELLED by the cpucontrol lock | |
993 | */ | |
994 | l3->shared = shared; | |
995 | shared = NULL; | |
e498be7d | 996 | } |
4484ebf1 RT |
997 | #ifdef CONFIG_NUMA |
998 | if (!l3->alien) { | |
999 | l3->alien = alien; | |
1000 | alien = NULL; | |
1001 | } | |
1002 | #endif | |
1003 | spin_unlock_irq(&l3->list_lock); | |
1004 | ||
1005 | kfree(shared); | |
1006 | free_alien_cache(alien); | |
1da177e4 | 1007 | } |
fc0abb14 | 1008 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1009 | break; |
1010 | case CPU_ONLINE: | |
1011 | start_cpu_timer(cpu); | |
1012 | break; | |
1013 | #ifdef CONFIG_HOTPLUG_CPU | |
1014 | case CPU_DEAD: | |
4484ebf1 RT |
1015 | /* |
1016 | * Even if all the cpus of a node are down, we don't free the | |
1017 | * kmem_list3 of any cache. This to avoid a race between | |
1018 | * cpu_down, and a kmalloc allocation from another cpu for | |
1019 | * memory from the node of the cpu going down. The list3 | |
1020 | * structure is usually allocated from kmem_cache_create() and | |
1021 | * gets destroyed at kmem_cache_destroy(). | |
1022 | */ | |
1da177e4 LT |
1023 | /* fall thru */ |
1024 | case CPU_UP_CANCELED: | |
fc0abb14 | 1025 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1026 | |
1027 | list_for_each_entry(cachep, &cache_chain, next) { | |
1028 | struct array_cache *nc; | |
4484ebf1 RT |
1029 | struct array_cache *shared; |
1030 | struct array_cache **alien; | |
e498be7d | 1031 | cpumask_t mask; |
1da177e4 | 1032 | |
e498be7d | 1033 | mask = node_to_cpumask(node); |
1da177e4 LT |
1034 | /* cpu is dead; no one can alloc from it. */ |
1035 | nc = cachep->array[cpu]; | |
1036 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1037 | l3 = cachep->nodelists[node]; |
1038 | ||
1039 | if (!l3) | |
4484ebf1 | 1040 | goto free_array_cache; |
e498be7d | 1041 | |
ca3b9b91 | 1042 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1043 | |
1044 | /* Free limit for this kmem_list3 */ | |
1045 | l3->free_limit -= cachep->batchcount; | |
1046 | if (nc) | |
ff69416e | 1047 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1048 | |
1049 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1050 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1051 | goto free_array_cache; |
b28a02de | 1052 | } |
e498be7d | 1053 | |
4484ebf1 RT |
1054 | shared = l3->shared; |
1055 | if (shared) { | |
e498be7d | 1056 | free_block(cachep, l3->shared->entry, |
b28a02de | 1057 | l3->shared->avail, node); |
e498be7d CL |
1058 | l3->shared = NULL; |
1059 | } | |
e498be7d | 1060 | |
4484ebf1 RT |
1061 | alien = l3->alien; |
1062 | l3->alien = NULL; | |
1063 | ||
1064 | spin_unlock_irq(&l3->list_lock); | |
1065 | ||
1066 | kfree(shared); | |
1067 | if (alien) { | |
1068 | drain_alien_cache(cachep, alien); | |
1069 | free_alien_cache(alien); | |
e498be7d | 1070 | } |
4484ebf1 | 1071 | free_array_cache: |
1da177e4 LT |
1072 | kfree(nc); |
1073 | } | |
4484ebf1 RT |
1074 | /* |
1075 | * In the previous loop, all the objects were freed to | |
1076 | * the respective cache's slabs, now we can go ahead and | |
1077 | * shrink each nodelist to its limit. | |
1078 | */ | |
1079 | list_for_each_entry(cachep, &cache_chain, next) { | |
1080 | l3 = cachep->nodelists[node]; | |
1081 | if (!l3) | |
1082 | continue; | |
1083 | spin_lock_irq(&l3->list_lock); | |
1084 | /* free slabs belonging to this node */ | |
1085 | __node_shrink(cachep, node); | |
1086 | spin_unlock_irq(&l3->list_lock); | |
1087 | } | |
fc0abb14 | 1088 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1089 | break; |
1090 | #endif | |
1091 | } | |
1092 | return NOTIFY_OK; | |
b28a02de | 1093 | bad: |
fc0abb14 | 1094 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1095 | return NOTIFY_BAD; |
1096 | } | |
1097 | ||
1098 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
1099 | ||
e498be7d CL |
1100 | /* |
1101 | * swap the static kmem_list3 with kmalloced memory | |
1102 | */ | |
343e0d7a | 1103 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid) |
e498be7d CL |
1104 | { |
1105 | struct kmem_list3 *ptr; | |
1106 | ||
1107 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1108 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1109 | BUG_ON(!ptr); | |
1110 | ||
1111 | local_irq_disable(); | |
1112 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1113 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1114 | cachep->nodelists[nodeid] = ptr; | |
1115 | local_irq_enable(); | |
1116 | } | |
1117 | ||
1da177e4 LT |
1118 | /* Initialisation. |
1119 | * Called after the gfp() functions have been enabled, and before smp_init(). | |
1120 | */ | |
1121 | void __init kmem_cache_init(void) | |
1122 | { | |
1123 | size_t left_over; | |
1124 | struct cache_sizes *sizes; | |
1125 | struct cache_names *names; | |
e498be7d CL |
1126 | int i; |
1127 | ||
1128 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1129 | kmem_list3_init(&initkmem_list3[i]); | |
1130 | if (i < MAX_NUMNODES) | |
1131 | cache_cache.nodelists[i] = NULL; | |
1132 | } | |
1da177e4 LT |
1133 | |
1134 | /* | |
1135 | * Fragmentation resistance on low memory - only use bigger | |
1136 | * page orders on machines with more than 32MB of memory. | |
1137 | */ | |
1138 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1139 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1140 | ||
1da177e4 LT |
1141 | /* Bootstrap is tricky, because several objects are allocated |
1142 | * from caches that do not exist yet: | |
343e0d7a | 1143 | * 1) initialize the cache_cache cache: it contains the struct kmem_cache |
1da177e4 LT |
1144 | * structures of all caches, except cache_cache itself: cache_cache |
1145 | * is statically allocated. | |
e498be7d CL |
1146 | * Initially an __init data area is used for the head array and the |
1147 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1148 | * array at the end of the bootstrap. | |
1da177e4 | 1149 | * 2) Create the first kmalloc cache. |
343e0d7a | 1150 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1151 | * An __init data area is used for the head array. |
1152 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1153 | * head arrays. | |
1da177e4 LT |
1154 | * 4) Replace the __init data head arrays for cache_cache and the first |
1155 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1156 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1157 | * the other cache's with kmalloc allocated memory. | |
1158 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1159 | */ |
1160 | ||
1161 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1162 | INIT_LIST_HEAD(&cache_chain); |
1163 | list_add(&cache_cache.next, &cache_chain); | |
1164 | cache_cache.colour_off = cache_line_size(); | |
1165 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1166 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1167 | |
3dafccf2 | 1168 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size()); |
1da177e4 | 1169 | |
3dafccf2 | 1170 | cache_estimate(0, cache_cache.buffer_size, cache_line_size(), 0, |
b28a02de | 1171 | &left_over, &cache_cache.num); |
1da177e4 LT |
1172 | if (!cache_cache.num) |
1173 | BUG(); | |
1174 | ||
b28a02de | 1175 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1176 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1177 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1178 | |
1179 | /* 2+3) create the kmalloc caches */ | |
1180 | sizes = malloc_sizes; | |
1181 | names = cache_names; | |
1182 | ||
e498be7d CL |
1183 | /* Initialize the caches that provide memory for the array cache |
1184 | * and the kmem_list3 structures first. | |
1185 | * Without this, further allocations will bug | |
1186 | */ | |
1187 | ||
1188 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
b28a02de PE |
1189 | sizes[INDEX_AC].cs_size, |
1190 | ARCH_KMALLOC_MINALIGN, | |
1191 | (ARCH_KMALLOC_FLAGS | | |
1192 | SLAB_PANIC), NULL, NULL); | |
e498be7d CL |
1193 | |
1194 | if (INDEX_AC != INDEX_L3) | |
1195 | sizes[INDEX_L3].cs_cachep = | |
b28a02de PE |
1196 | kmem_cache_create(names[INDEX_L3].name, |
1197 | sizes[INDEX_L3].cs_size, | |
1198 | ARCH_KMALLOC_MINALIGN, | |
1199 | (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, | |
1200 | NULL); | |
e498be7d | 1201 | |
1da177e4 | 1202 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1203 | /* |
1204 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1205 | * This should be particularly beneficial on SMP boxes, as it |
1206 | * eliminates "false sharing". | |
1207 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1208 | * allow tighter packing of the smaller caches. |
1209 | */ | |
b28a02de | 1210 | if (!sizes->cs_cachep) |
e498be7d | 1211 | sizes->cs_cachep = kmem_cache_create(names->name, |
b28a02de PE |
1212 | sizes->cs_size, |
1213 | ARCH_KMALLOC_MINALIGN, | |
1214 | (ARCH_KMALLOC_FLAGS | |
1215 | | SLAB_PANIC), | |
1216 | NULL, NULL); | |
1da177e4 LT |
1217 | |
1218 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
1219 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
b28a02de | 1220 | offslab_limit = sizes->cs_size - sizeof(struct slab); |
1da177e4 LT |
1221 | offslab_limit /= sizeof(kmem_bufctl_t); |
1222 | } | |
1223 | ||
1224 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
b28a02de PE |
1225 | sizes->cs_size, |
1226 | ARCH_KMALLOC_MINALIGN, | |
1227 | (ARCH_KMALLOC_FLAGS | | |
1228 | SLAB_CACHE_DMA | | |
1229 | SLAB_PANIC), NULL, | |
1230 | NULL); | |
1da177e4 LT |
1231 | |
1232 | sizes++; | |
1233 | names++; | |
1234 | } | |
1235 | /* 4) Replace the bootstrap head arrays */ | |
1236 | { | |
b28a02de | 1237 | void *ptr; |
e498be7d | 1238 | |
1da177e4 | 1239 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1240 | |
1da177e4 | 1241 | local_irq_disable(); |
9a2dba4b PE |
1242 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1243 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1244 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1245 | cache_cache.array[smp_processor_id()] = ptr; |
1246 | local_irq_enable(); | |
e498be7d | 1247 | |
1da177e4 | 1248 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1249 | |
1da177e4 | 1250 | local_irq_disable(); |
9a2dba4b | 1251 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1252 | != &initarray_generic.cache); |
9a2dba4b | 1253 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1254 | sizeof(struct arraycache_init)); |
e498be7d | 1255 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1256 | ptr; |
1da177e4 LT |
1257 | local_irq_enable(); |
1258 | } | |
e498be7d CL |
1259 | /* 5) Replace the bootstrap kmem_list3's */ |
1260 | { | |
1261 | int node; | |
1262 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1263 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1264 | numa_node_id()); |
e498be7d CL |
1265 | |
1266 | for_each_online_node(node) { | |
1267 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1268 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1269 | |
1270 | if (INDEX_AC != INDEX_L3) { | |
1271 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1272 | &initkmem_list3[SIZE_L3 + node], |
1273 | node); | |
e498be7d CL |
1274 | } |
1275 | } | |
1276 | } | |
1da177e4 | 1277 | |
e498be7d | 1278 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1279 | { |
343e0d7a | 1280 | struct kmem_cache *cachep; |
fc0abb14 | 1281 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1282 | list_for_each_entry(cachep, &cache_chain, next) |
b28a02de | 1283 | enable_cpucache(cachep); |
fc0abb14 | 1284 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1285 | } |
1286 | ||
1287 | /* Done! */ | |
1288 | g_cpucache_up = FULL; | |
1289 | ||
1290 | /* Register a cpu startup notifier callback | |
9a2dba4b | 1291 | * that initializes cpu_cache_get for all new cpus |
1da177e4 LT |
1292 | */ |
1293 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 LT |
1294 | |
1295 | /* The reap timers are started later, with a module init call: | |
1296 | * That part of the kernel is not yet operational. | |
1297 | */ | |
1298 | } | |
1299 | ||
1300 | static int __init cpucache_init(void) | |
1301 | { | |
1302 | int cpu; | |
1303 | ||
1304 | /* | |
1305 | * Register the timers that return unneeded | |
1306 | * pages to gfp. | |
1307 | */ | |
e498be7d | 1308 | for_each_online_cpu(cpu) |
b28a02de | 1309 | start_cpu_timer(cpu); |
1da177e4 LT |
1310 | |
1311 | return 0; | |
1312 | } | |
1313 | ||
1314 | __initcall(cpucache_init); | |
1315 | ||
1316 | /* | |
1317 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1318 | * | |
1319 | * If we requested dmaable memory, we will get it. Even if we | |
1320 | * did not request dmaable memory, we might get it, but that | |
1321 | * would be relatively rare and ignorable. | |
1322 | */ | |
343e0d7a | 1323 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1324 | { |
1325 | struct page *page; | |
1326 | void *addr; | |
1327 | int i; | |
1328 | ||
1329 | flags |= cachep->gfpflags; | |
50c85a19 | 1330 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); |
1da177e4 LT |
1331 | if (!page) |
1332 | return NULL; | |
1333 | addr = page_address(page); | |
1334 | ||
1335 | i = (1 << cachep->gfporder); | |
1336 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1337 | atomic_add(i, &slab_reclaim_pages); | |
1338 | add_page_state(nr_slab, i); | |
1339 | while (i--) { | |
1340 | SetPageSlab(page); | |
1341 | page++; | |
1342 | } | |
1343 | return addr; | |
1344 | } | |
1345 | ||
1346 | /* | |
1347 | * Interface to system's page release. | |
1348 | */ | |
343e0d7a | 1349 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1350 | { |
b28a02de | 1351 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1352 | struct page *page = virt_to_page(addr); |
1353 | const unsigned long nr_freed = i; | |
1354 | ||
1355 | while (i--) { | |
1356 | if (!TestClearPageSlab(page)) | |
1357 | BUG(); | |
1358 | page++; | |
1359 | } | |
1360 | sub_page_state(nr_slab, nr_freed); | |
1361 | if (current->reclaim_state) | |
1362 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1363 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1364 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1365 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1366 | } |
1367 | ||
1368 | static void kmem_rcu_free(struct rcu_head *head) | |
1369 | { | |
b28a02de | 1370 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1371 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1372 | |
1373 | kmem_freepages(cachep, slab_rcu->addr); | |
1374 | if (OFF_SLAB(cachep)) | |
1375 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1376 | } | |
1377 | ||
1378 | #if DEBUG | |
1379 | ||
1380 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1381 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1382 | unsigned long caller) |
1da177e4 | 1383 | { |
3dafccf2 | 1384 | int size = obj_size(cachep); |
1da177e4 | 1385 | |
3dafccf2 | 1386 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1387 | |
b28a02de | 1388 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1389 | return; |
1390 | ||
b28a02de PE |
1391 | *addr++ = 0x12345678; |
1392 | *addr++ = caller; | |
1393 | *addr++ = smp_processor_id(); | |
1394 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1395 | { |
1396 | unsigned long *sptr = &caller; | |
1397 | unsigned long svalue; | |
1398 | ||
1399 | while (!kstack_end(sptr)) { | |
1400 | svalue = *sptr++; | |
1401 | if (kernel_text_address(svalue)) { | |
b28a02de | 1402 | *addr++ = svalue; |
1da177e4 LT |
1403 | size -= sizeof(unsigned long); |
1404 | if (size <= sizeof(unsigned long)) | |
1405 | break; | |
1406 | } | |
1407 | } | |
1408 | ||
1409 | } | |
b28a02de | 1410 | *addr++ = 0x87654321; |
1da177e4 LT |
1411 | } |
1412 | #endif | |
1413 | ||
343e0d7a | 1414 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1415 | { |
3dafccf2 MS |
1416 | int size = obj_size(cachep); |
1417 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1418 | |
1419 | memset(addr, val, size); | |
b28a02de | 1420 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1421 | } |
1422 | ||
1423 | static void dump_line(char *data, int offset, int limit) | |
1424 | { | |
1425 | int i; | |
1426 | printk(KERN_ERR "%03x:", offset); | |
b28a02de PE |
1427 | for (i = 0; i < limit; i++) { |
1428 | printk(" %02x", (unsigned char)data[offset + i]); | |
1da177e4 LT |
1429 | } |
1430 | printk("\n"); | |
1431 | } | |
1432 | #endif | |
1433 | ||
1434 | #if DEBUG | |
1435 | ||
343e0d7a | 1436 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1437 | { |
1438 | int i, size; | |
1439 | char *realobj; | |
1440 | ||
1441 | if (cachep->flags & SLAB_RED_ZONE) { | |
1442 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
b28a02de PE |
1443 | *dbg_redzone1(cachep, objp), |
1444 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1445 | } |
1446 | ||
1447 | if (cachep->flags & SLAB_STORE_USER) { | |
1448 | printk(KERN_ERR "Last user: [<%p>]", | |
b28a02de | 1449 | *dbg_userword(cachep, objp)); |
1da177e4 | 1450 | print_symbol("(%s)", |
b28a02de | 1451 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1452 | printk("\n"); |
1453 | } | |
3dafccf2 MS |
1454 | realobj = (char *)objp + obj_offset(cachep); |
1455 | size = obj_size(cachep); | |
b28a02de | 1456 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1457 | int limit; |
1458 | limit = 16; | |
b28a02de PE |
1459 | if (i + limit > size) |
1460 | limit = size - i; | |
1da177e4 LT |
1461 | dump_line(realobj, i, limit); |
1462 | } | |
1463 | } | |
1464 | ||
343e0d7a | 1465 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1466 | { |
1467 | char *realobj; | |
1468 | int size, i; | |
1469 | int lines = 0; | |
1470 | ||
3dafccf2 MS |
1471 | realobj = (char *)objp + obj_offset(cachep); |
1472 | size = obj_size(cachep); | |
1da177e4 | 1473 | |
b28a02de | 1474 | for (i = 0; i < size; i++) { |
1da177e4 | 1475 | char exp = POISON_FREE; |
b28a02de | 1476 | if (i == size - 1) |
1da177e4 LT |
1477 | exp = POISON_END; |
1478 | if (realobj[i] != exp) { | |
1479 | int limit; | |
1480 | /* Mismatch ! */ | |
1481 | /* Print header */ | |
1482 | if (lines == 0) { | |
b28a02de PE |
1483 | printk(KERN_ERR |
1484 | "Slab corruption: start=%p, len=%d\n", | |
1485 | realobj, size); | |
1da177e4 LT |
1486 | print_objinfo(cachep, objp, 0); |
1487 | } | |
1488 | /* Hexdump the affected line */ | |
b28a02de | 1489 | i = (i / 16) * 16; |
1da177e4 | 1490 | limit = 16; |
b28a02de PE |
1491 | if (i + limit > size) |
1492 | limit = size - i; | |
1da177e4 LT |
1493 | dump_line(realobj, i, limit); |
1494 | i += 16; | |
1495 | lines++; | |
1496 | /* Limit to 5 lines */ | |
1497 | if (lines > 5) | |
1498 | break; | |
1499 | } | |
1500 | } | |
1501 | if (lines != 0) { | |
1502 | /* Print some data about the neighboring objects, if they | |
1503 | * exist: | |
1504 | */ | |
6ed5eb22 | 1505 | struct slab *slabp = virt_to_slab(objp); |
1da177e4 LT |
1506 | int objnr; |
1507 | ||
3dafccf2 | 1508 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
1da177e4 | 1509 | if (objnr) { |
3dafccf2 MS |
1510 | objp = slabp->s_mem + (objnr - 1) * cachep->buffer_size; |
1511 | realobj = (char *)objp + obj_offset(cachep); | |
1da177e4 | 1512 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1513 | realobj, size); |
1da177e4 LT |
1514 | print_objinfo(cachep, objp, 2); |
1515 | } | |
b28a02de | 1516 | if (objnr + 1 < cachep->num) { |
3dafccf2 MS |
1517 | objp = slabp->s_mem + (objnr + 1) * cachep->buffer_size; |
1518 | realobj = (char *)objp + obj_offset(cachep); | |
1da177e4 | 1519 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1520 | realobj, size); |
1da177e4 LT |
1521 | print_objinfo(cachep, objp, 2); |
1522 | } | |
1523 | } | |
1524 | } | |
1525 | #endif | |
1526 | ||
12dd36fa MD |
1527 | #if DEBUG |
1528 | /** | |
1529 | * slab_destroy_objs - call the registered destructor for each object in | |
1530 | * a slab that is to be destroyed. | |
1da177e4 | 1531 | */ |
343e0d7a | 1532 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1533 | { |
1da177e4 LT |
1534 | int i; |
1535 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 1536 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
1da177e4 LT |
1537 | |
1538 | if (cachep->flags & SLAB_POISON) { | |
1539 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 1540 | if ((cachep->buffer_size % PAGE_SIZE) == 0 |
b28a02de PE |
1541 | && OFF_SLAB(cachep)) |
1542 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 1543 | cachep->buffer_size / PAGE_SIZE, |
b28a02de | 1544 | 1); |
1da177e4 LT |
1545 | else |
1546 | check_poison_obj(cachep, objp); | |
1547 | #else | |
1548 | check_poison_obj(cachep, objp); | |
1549 | #endif | |
1550 | } | |
1551 | if (cachep->flags & SLAB_RED_ZONE) { | |
1552 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1553 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1554 | "was overwritten"); |
1da177e4 LT |
1555 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1556 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1557 | "was overwritten"); |
1da177e4 LT |
1558 | } |
1559 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1560 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1561 | } |
12dd36fa | 1562 | } |
1da177e4 | 1563 | #else |
343e0d7a | 1564 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1565 | { |
1da177e4 LT |
1566 | if (cachep->dtor) { |
1567 | int i; | |
1568 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 1569 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
b28a02de | 1570 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1571 | } |
1572 | } | |
12dd36fa | 1573 | } |
1da177e4 LT |
1574 | #endif |
1575 | ||
12dd36fa MD |
1576 | /** |
1577 | * Destroy all the objs in a slab, and release the mem back to the system. | |
1578 | * Before calling the slab must have been unlinked from the cache. | |
1579 | * The cache-lock is not held/needed. | |
1580 | */ | |
343e0d7a | 1581 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1582 | { |
1583 | void *addr = slabp->s_mem - slabp->colouroff; | |
1584 | ||
1585 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1586 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1587 | struct slab_rcu *slab_rcu; | |
1588 | ||
b28a02de | 1589 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1590 | slab_rcu->cachep = cachep; |
1591 | slab_rcu->addr = addr; | |
1592 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1593 | } else { | |
1594 | kmem_freepages(cachep, addr); | |
1595 | if (OFF_SLAB(cachep)) | |
1596 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1597 | } | |
1598 | } | |
1599 | ||
3dafccf2 | 1600 | /* For setting up all the kmem_list3s for cache whose buffer_size is same |
e498be7d | 1601 | as size of kmem_list3. */ |
343e0d7a | 1602 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1603 | { |
1604 | int node; | |
1605 | ||
1606 | for_each_online_node(node) { | |
b28a02de | 1607 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1608 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1609 | REAPTIMEOUT_LIST3 + |
1610 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1611 | } |
1612 | } | |
1613 | ||
4d268eba | 1614 | /** |
a70773dd RD |
1615 | * calculate_slab_order - calculate size (page order) of slabs |
1616 | * @cachep: pointer to the cache that is being created | |
1617 | * @size: size of objects to be created in this cache. | |
1618 | * @align: required alignment for the objects. | |
1619 | * @flags: slab allocation flags | |
1620 | * | |
1621 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1622 | * |
1623 | * This could be made much more intelligent. For now, try to avoid using | |
1624 | * high order pages for slabs. When the gfp() functions are more friendly | |
1625 | * towards high-order requests, this should be changed. | |
1626 | */ | |
ee13d785 RD |
1627 | static inline size_t calculate_slab_order(struct kmem_cache *cachep, |
1628 | size_t size, size_t align, unsigned long flags) | |
4d268eba PE |
1629 | { |
1630 | size_t left_over = 0; | |
9888e6fa | 1631 | int gfporder; |
4d268eba | 1632 | |
9888e6fa | 1633 | for (gfporder = 0 ; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1634 | unsigned int num; |
1635 | size_t remainder; | |
1636 | ||
9888e6fa | 1637 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1638 | if (!num) |
1639 | continue; | |
9888e6fa | 1640 | |
4d268eba | 1641 | /* More than offslab_limit objects will cause problems */ |
9888e6fa | 1642 | if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit) |
4d268eba PE |
1643 | break; |
1644 | ||
9888e6fa | 1645 | /* Found something acceptable - save it away */ |
4d268eba | 1646 | cachep->num = num; |
9888e6fa | 1647 | cachep->gfporder = gfporder; |
4d268eba PE |
1648 | left_over = remainder; |
1649 | ||
f78bb8ad LT |
1650 | /* |
1651 | * A VFS-reclaimable slab tends to have most allocations | |
1652 | * as GFP_NOFS and we really don't want to have to be allocating | |
1653 | * higher-order pages when we are unable to shrink dcache. | |
1654 | */ | |
1655 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1656 | break; | |
1657 | ||
4d268eba PE |
1658 | /* |
1659 | * Large number of objects is good, but very large slabs are | |
1660 | * currently bad for the gfp()s. | |
1661 | */ | |
9888e6fa | 1662 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1663 | break; |
1664 | ||
9888e6fa LT |
1665 | /* |
1666 | * Acceptable internal fragmentation? | |
1667 | */ | |
1668 | if ((left_over * 8) <= (PAGE_SIZE << gfporder)) | |
4d268eba PE |
1669 | break; |
1670 | } | |
1671 | return left_over; | |
1672 | } | |
1673 | ||
1da177e4 LT |
1674 | /** |
1675 | * kmem_cache_create - Create a cache. | |
1676 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1677 | * @size: The size of objects to be created in this cache. | |
1678 | * @align: The required alignment for the objects. | |
1679 | * @flags: SLAB flags | |
1680 | * @ctor: A constructor for the objects. | |
1681 | * @dtor: A destructor for the objects. | |
1682 | * | |
1683 | * Returns a ptr to the cache on success, NULL on failure. | |
1684 | * Cannot be called within a int, but can be interrupted. | |
1685 | * The @ctor is run when new pages are allocated by the cache | |
1686 | * and the @dtor is run before the pages are handed back. | |
1687 | * | |
1688 | * @name must be valid until the cache is destroyed. This implies that | |
1689 | * the module calling this has to destroy the cache before getting | |
1690 | * unloaded. | |
1691 | * | |
1692 | * The flags are | |
1693 | * | |
1694 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1695 | * to catch references to uninitialised memory. | |
1696 | * | |
1697 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1698 | * for buffer overruns. | |
1699 | * | |
1700 | * %SLAB_NO_REAP - Don't automatically reap this cache when we're under | |
1701 | * memory pressure. | |
1702 | * | |
1703 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
1704 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1705 | * as davem. | |
1706 | */ | |
343e0d7a | 1707 | struct kmem_cache * |
1da177e4 | 1708 | kmem_cache_create (const char *name, size_t size, size_t align, |
343e0d7a PE |
1709 | unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long), |
1710 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) | |
1da177e4 LT |
1711 | { |
1712 | size_t left_over, slab_size, ralign; | |
343e0d7a | 1713 | struct kmem_cache *cachep = NULL; |
4f12bb4f | 1714 | struct list_head *p; |
1da177e4 LT |
1715 | |
1716 | /* | |
1717 | * Sanity checks... these are all serious usage bugs. | |
1718 | */ | |
1719 | if ((!name) || | |
b28a02de PE |
1720 | in_interrupt() || |
1721 | (size < BYTES_PER_WORD) || | |
1722 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { | |
1723 | printk(KERN_ERR "%s: Early error in slab %s\n", | |
1724 | __FUNCTION__, name); | |
1725 | BUG(); | |
1726 | } | |
1da177e4 | 1727 | |
f0188f47 RT |
1728 | /* |
1729 | * Prevent CPUs from coming and going. | |
1730 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
1731 | */ | |
1732 | lock_cpu_hotplug(); | |
1733 | ||
fc0abb14 | 1734 | mutex_lock(&cache_chain_mutex); |
4f12bb4f AM |
1735 | |
1736 | list_for_each(p, &cache_chain) { | |
343e0d7a | 1737 | struct kmem_cache *pc = list_entry(p, struct kmem_cache, next); |
4f12bb4f AM |
1738 | mm_segment_t old_fs = get_fs(); |
1739 | char tmp; | |
1740 | int res; | |
1741 | ||
1742 | /* | |
1743 | * This happens when the module gets unloaded and doesn't | |
1744 | * destroy its slab cache and no-one else reuses the vmalloc | |
1745 | * area of the module. Print a warning. | |
1746 | */ | |
1747 | set_fs(KERNEL_DS); | |
1748 | res = __get_user(tmp, pc->name); | |
1749 | set_fs(old_fs); | |
1750 | if (res) { | |
1751 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 1752 | pc->buffer_size); |
4f12bb4f AM |
1753 | continue; |
1754 | } | |
1755 | ||
b28a02de | 1756 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1757 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1758 | dump_stack(); | |
1759 | goto oops; | |
1760 | } | |
1761 | } | |
1762 | ||
1da177e4 LT |
1763 | #if DEBUG |
1764 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1765 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1766 | /* No constructor, but inital state check requested */ | |
1767 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 1768 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
1769 | flags &= ~SLAB_DEBUG_INITIAL; |
1770 | } | |
1da177e4 LT |
1771 | #if FORCED_DEBUG |
1772 | /* | |
1773 | * Enable redzoning and last user accounting, except for caches with | |
1774 | * large objects, if the increased size would increase the object size | |
1775 | * above the next power of two: caches with object sizes just above a | |
1776 | * power of two have a significant amount of internal fragmentation. | |
1777 | */ | |
b28a02de PE |
1778 | if ((size < 4096 |
1779 | || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD))) | |
1780 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; | |
1da177e4 LT |
1781 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
1782 | flags |= SLAB_POISON; | |
1783 | #endif | |
1784 | if (flags & SLAB_DESTROY_BY_RCU) | |
1785 | BUG_ON(flags & SLAB_POISON); | |
1786 | #endif | |
1787 | if (flags & SLAB_DESTROY_BY_RCU) | |
1788 | BUG_ON(dtor); | |
1789 | ||
1790 | /* | |
1791 | * Always checks flags, a caller might be expecting debug | |
1792 | * support which isn't available. | |
1793 | */ | |
1794 | if (flags & ~CREATE_MASK) | |
1795 | BUG(); | |
1796 | ||
1797 | /* Check that size is in terms of words. This is needed to avoid | |
1798 | * unaligned accesses for some archs when redzoning is used, and makes | |
1799 | * sure any on-slab bufctl's are also correctly aligned. | |
1800 | */ | |
b28a02de PE |
1801 | if (size & (BYTES_PER_WORD - 1)) { |
1802 | size += (BYTES_PER_WORD - 1); | |
1803 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
1804 | } |
1805 | ||
1806 | /* calculate out the final buffer alignment: */ | |
1807 | /* 1) arch recommendation: can be overridden for debug */ | |
1808 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1809 | /* Default alignment: as specified by the arch code. | |
1810 | * Except if an object is really small, then squeeze multiple | |
1811 | * objects into one cacheline. | |
1812 | */ | |
1813 | ralign = cache_line_size(); | |
b28a02de | 1814 | while (size <= ralign / 2) |
1da177e4 LT |
1815 | ralign /= 2; |
1816 | } else { | |
1817 | ralign = BYTES_PER_WORD; | |
1818 | } | |
1819 | /* 2) arch mandated alignment: disables debug if necessary */ | |
1820 | if (ralign < ARCH_SLAB_MINALIGN) { | |
1821 | ralign = ARCH_SLAB_MINALIGN; | |
1822 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1823 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1824 | } |
1825 | /* 3) caller mandated alignment: disables debug if necessary */ | |
1826 | if (ralign < align) { | |
1827 | ralign = align; | |
1828 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1829 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1830 | } |
1831 | /* 4) Store it. Note that the debug code below can reduce | |
1832 | * the alignment to BYTES_PER_WORD. | |
1833 | */ | |
1834 | align = ralign; | |
1835 | ||
1836 | /* Get cache's description obj. */ | |
343e0d7a | 1837 | cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 1838 | if (!cachep) |
4f12bb4f | 1839 | goto oops; |
343e0d7a | 1840 | memset(cachep, 0, sizeof(struct kmem_cache)); |
1da177e4 LT |
1841 | |
1842 | #if DEBUG | |
3dafccf2 | 1843 | cachep->obj_size = size; |
1da177e4 LT |
1844 | |
1845 | if (flags & SLAB_RED_ZONE) { | |
1846 | /* redzoning only works with word aligned caches */ | |
1847 | align = BYTES_PER_WORD; | |
1848 | ||
1849 | /* add space for red zone words */ | |
3dafccf2 | 1850 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 1851 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
1852 | } |
1853 | if (flags & SLAB_STORE_USER) { | |
1854 | /* user store requires word alignment and | |
1855 | * one word storage behind the end of the real | |
1856 | * object. | |
1857 | */ | |
1858 | align = BYTES_PER_WORD; | |
1859 | size += BYTES_PER_WORD; | |
1860 | } | |
1861 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 1862 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
1863 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
1864 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
1865 | size = PAGE_SIZE; |
1866 | } | |
1867 | #endif | |
1868 | #endif | |
1869 | ||
1870 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 1871 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
1872 | /* |
1873 | * Size is large, assume best to place the slab management obj | |
1874 | * off-slab (should allow better packing of objs). | |
1875 | */ | |
1876 | flags |= CFLGS_OFF_SLAB; | |
1877 | ||
1878 | size = ALIGN(size, align); | |
1879 | ||
f78bb8ad | 1880 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
1881 | |
1882 | if (!cachep->num) { | |
1883 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
1884 | kmem_cache_free(&cache_cache, cachep); | |
1885 | cachep = NULL; | |
4f12bb4f | 1886 | goto oops; |
1da177e4 | 1887 | } |
b28a02de PE |
1888 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
1889 | + sizeof(struct slab), align); | |
1da177e4 LT |
1890 | |
1891 | /* | |
1892 | * If the slab has been placed off-slab, and we have enough space then | |
1893 | * move it on-slab. This is at the expense of any extra colouring. | |
1894 | */ | |
1895 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
1896 | flags &= ~CFLGS_OFF_SLAB; | |
1897 | left_over -= slab_size; | |
1898 | } | |
1899 | ||
1900 | if (flags & CFLGS_OFF_SLAB) { | |
1901 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
1902 | slab_size = |
1903 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
1904 | } |
1905 | ||
1906 | cachep->colour_off = cache_line_size(); | |
1907 | /* Offset must be a multiple of the alignment. */ | |
1908 | if (cachep->colour_off < align) | |
1909 | cachep->colour_off = align; | |
b28a02de | 1910 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
1911 | cachep->slab_size = slab_size; |
1912 | cachep->flags = flags; | |
1913 | cachep->gfpflags = 0; | |
1914 | if (flags & SLAB_CACHE_DMA) | |
1915 | cachep->gfpflags |= GFP_DMA; | |
1916 | spin_lock_init(&cachep->spinlock); | |
3dafccf2 | 1917 | cachep->buffer_size = size; |
1da177e4 LT |
1918 | |
1919 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 1920 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
1921 | cachep->ctor = ctor; |
1922 | cachep->dtor = dtor; | |
1923 | cachep->name = name; | |
1924 | ||
1da177e4 LT |
1925 | |
1926 | if (g_cpucache_up == FULL) { | |
1927 | enable_cpucache(cachep); | |
1928 | } else { | |
1929 | if (g_cpucache_up == NONE) { | |
1930 | /* Note: the first kmem_cache_create must create | |
1931 | * the cache that's used by kmalloc(24), otherwise | |
1932 | * the creation of further caches will BUG(). | |
1933 | */ | |
e498be7d | 1934 | cachep->array[smp_processor_id()] = |
b28a02de | 1935 | &initarray_generic.cache; |
e498be7d CL |
1936 | |
1937 | /* If the cache that's used by | |
1938 | * kmalloc(sizeof(kmem_list3)) is the first cache, | |
1939 | * then we need to set up all its list3s, otherwise | |
1940 | * the creation of further caches will BUG(). | |
1941 | */ | |
1942 | set_up_list3s(cachep, SIZE_AC); | |
1943 | if (INDEX_AC == INDEX_L3) | |
1944 | g_cpucache_up = PARTIAL_L3; | |
1945 | else | |
1946 | g_cpucache_up = PARTIAL_AC; | |
1da177e4 | 1947 | } else { |
e498be7d | 1948 | cachep->array[smp_processor_id()] = |
b28a02de | 1949 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d CL |
1950 | |
1951 | if (g_cpucache_up == PARTIAL_AC) { | |
1952 | set_up_list3s(cachep, SIZE_L3); | |
1953 | g_cpucache_up = PARTIAL_L3; | |
1954 | } else { | |
1955 | int node; | |
1956 | for_each_online_node(node) { | |
1957 | ||
1958 | cachep->nodelists[node] = | |
b28a02de PE |
1959 | kmalloc_node(sizeof |
1960 | (struct kmem_list3), | |
1961 | GFP_KERNEL, node); | |
e498be7d | 1962 | BUG_ON(!cachep->nodelists[node]); |
b28a02de PE |
1963 | kmem_list3_init(cachep-> |
1964 | nodelists[node]); | |
e498be7d CL |
1965 | } |
1966 | } | |
1da177e4 | 1967 | } |
e498be7d | 1968 | cachep->nodelists[numa_node_id()]->next_reap = |
b28a02de PE |
1969 | jiffies + REAPTIMEOUT_LIST3 + |
1970 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1971 | |
9a2dba4b PE |
1972 | BUG_ON(!cpu_cache_get(cachep)); |
1973 | cpu_cache_get(cachep)->avail = 0; | |
1974 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
1975 | cpu_cache_get(cachep)->batchcount = 1; | |
1976 | cpu_cache_get(cachep)->touched = 0; | |
1da177e4 LT |
1977 | cachep->batchcount = 1; |
1978 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
b28a02de | 1979 | } |
1da177e4 | 1980 | |
1da177e4 LT |
1981 | /* cache setup completed, link it into the list */ |
1982 | list_add(&cachep->next, &cache_chain); | |
b28a02de | 1983 | oops: |
1da177e4 LT |
1984 | if (!cachep && (flags & SLAB_PANIC)) |
1985 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 1986 | name); |
fc0abb14 | 1987 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 1988 | unlock_cpu_hotplug(); |
1da177e4 LT |
1989 | return cachep; |
1990 | } | |
1991 | EXPORT_SYMBOL(kmem_cache_create); | |
1992 | ||
1993 | #if DEBUG | |
1994 | static void check_irq_off(void) | |
1995 | { | |
1996 | BUG_ON(!irqs_disabled()); | |
1997 | } | |
1998 | ||
1999 | static void check_irq_on(void) | |
2000 | { | |
2001 | BUG_ON(irqs_disabled()); | |
2002 | } | |
2003 | ||
343e0d7a | 2004 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2005 | { |
2006 | #ifdef CONFIG_SMP | |
2007 | check_irq_off(); | |
e498be7d | 2008 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2009 | #endif |
2010 | } | |
e498be7d | 2011 | |
343e0d7a | 2012 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2013 | { |
2014 | #ifdef CONFIG_SMP | |
2015 | check_irq_off(); | |
2016 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2017 | #endif | |
2018 | } | |
2019 | ||
1da177e4 LT |
2020 | #else |
2021 | #define check_irq_off() do { } while(0) | |
2022 | #define check_irq_on() do { } while(0) | |
2023 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2024 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2025 | #endif |
2026 | ||
2027 | /* | |
2028 | * Waits for all CPUs to execute func(). | |
2029 | */ | |
b28a02de | 2030 | static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg) |
1da177e4 LT |
2031 | { |
2032 | check_irq_on(); | |
2033 | preempt_disable(); | |
2034 | ||
2035 | local_irq_disable(); | |
2036 | func(arg); | |
2037 | local_irq_enable(); | |
2038 | ||
2039 | if (smp_call_function(func, arg, 1, 1)) | |
2040 | BUG(); | |
2041 | ||
2042 | preempt_enable(); | |
2043 | } | |
2044 | ||
343e0d7a | 2045 | static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, |
b28a02de | 2046 | int force, int node); |
1da177e4 LT |
2047 | |
2048 | static void do_drain(void *arg) | |
2049 | { | |
343e0d7a | 2050 | struct kmem_cache *cachep = (struct kmem_cache *) arg; |
1da177e4 | 2051 | struct array_cache *ac; |
ff69416e | 2052 | int node = numa_node_id(); |
1da177e4 LT |
2053 | |
2054 | check_irq_off(); | |
9a2dba4b | 2055 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2056 | spin_lock(&cachep->nodelists[node]->list_lock); |
2057 | free_block(cachep, ac->entry, ac->avail, node); | |
2058 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2059 | ac->avail = 0; |
2060 | } | |
2061 | ||
343e0d7a | 2062 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2063 | { |
e498be7d CL |
2064 | struct kmem_list3 *l3; |
2065 | int node; | |
2066 | ||
1da177e4 LT |
2067 | smp_call_function_all_cpus(do_drain, cachep); |
2068 | check_irq_on(); | |
b28a02de | 2069 | for_each_online_node(node) { |
e498be7d CL |
2070 | l3 = cachep->nodelists[node]; |
2071 | if (l3) { | |
ca3b9b91 | 2072 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2073 | drain_array_locked(cachep, l3->shared, 1, node); |
ca3b9b91 | 2074 | spin_unlock_irq(&l3->list_lock); |
e498be7d | 2075 | if (l3->alien) |
4484ebf1 | 2076 | drain_alien_cache(cachep, l3->alien); |
e498be7d CL |
2077 | } |
2078 | } | |
1da177e4 LT |
2079 | } |
2080 | ||
343e0d7a | 2081 | static int __node_shrink(struct kmem_cache *cachep, int node) |
1da177e4 LT |
2082 | { |
2083 | struct slab *slabp; | |
e498be7d | 2084 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
2085 | int ret; |
2086 | ||
e498be7d | 2087 | for (;;) { |
1da177e4 LT |
2088 | struct list_head *p; |
2089 | ||
e498be7d CL |
2090 | p = l3->slabs_free.prev; |
2091 | if (p == &l3->slabs_free) | |
1da177e4 LT |
2092 | break; |
2093 | ||
e498be7d | 2094 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 LT |
2095 | #if DEBUG |
2096 | if (slabp->inuse) | |
2097 | BUG(); | |
2098 | #endif | |
2099 | list_del(&slabp->list); | |
2100 | ||
e498be7d CL |
2101 | l3->free_objects -= cachep->num; |
2102 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 2103 | slab_destroy(cachep, slabp); |
e498be7d | 2104 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 2105 | } |
b28a02de | 2106 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
2107 | return ret; |
2108 | } | |
2109 | ||
343e0d7a | 2110 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2111 | { |
2112 | int ret = 0, i = 0; | |
2113 | struct kmem_list3 *l3; | |
2114 | ||
2115 | drain_cpu_caches(cachep); | |
2116 | ||
2117 | check_irq_on(); | |
2118 | for_each_online_node(i) { | |
2119 | l3 = cachep->nodelists[i]; | |
2120 | if (l3) { | |
2121 | spin_lock_irq(&l3->list_lock); | |
2122 | ret += __node_shrink(cachep, i); | |
2123 | spin_unlock_irq(&l3->list_lock); | |
2124 | } | |
2125 | } | |
2126 | return (ret ? 1 : 0); | |
2127 | } | |
2128 | ||
1da177e4 LT |
2129 | /** |
2130 | * kmem_cache_shrink - Shrink a cache. | |
2131 | * @cachep: The cache to shrink. | |
2132 | * | |
2133 | * Releases as many slabs as possible for a cache. | |
2134 | * To help debugging, a zero exit status indicates all slabs were released. | |
2135 | */ | |
343e0d7a | 2136 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 LT |
2137 | { |
2138 | if (!cachep || in_interrupt()) | |
2139 | BUG(); | |
2140 | ||
2141 | return __cache_shrink(cachep); | |
2142 | } | |
2143 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2144 | ||
2145 | /** | |
2146 | * kmem_cache_destroy - delete a cache | |
2147 | * @cachep: the cache to destroy | |
2148 | * | |
343e0d7a | 2149 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2150 | * Returns 0 on success. |
2151 | * | |
2152 | * It is expected this function will be called by a module when it is | |
2153 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2154 | * cache being allocated each time a module is loaded and unloaded, if the | |
2155 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2156 | * | |
2157 | * The cache must be empty before calling this function. | |
2158 | * | |
2159 | * The caller must guarantee that noone will allocate memory from the cache | |
2160 | * during the kmem_cache_destroy(). | |
2161 | */ | |
343e0d7a | 2162 | int kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 LT |
2163 | { |
2164 | int i; | |
e498be7d | 2165 | struct kmem_list3 *l3; |
1da177e4 LT |
2166 | |
2167 | if (!cachep || in_interrupt()) | |
2168 | BUG(); | |
2169 | ||
2170 | /* Don't let CPUs to come and go */ | |
2171 | lock_cpu_hotplug(); | |
2172 | ||
2173 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2174 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2175 | /* |
2176 | * the chain is never empty, cache_cache is never destroyed | |
2177 | */ | |
2178 | list_del(&cachep->next); | |
fc0abb14 | 2179 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2180 | |
2181 | if (__cache_shrink(cachep)) { | |
2182 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2183 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2184 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2185 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2186 | unlock_cpu_hotplug(); |
2187 | return 1; | |
2188 | } | |
2189 | ||
2190 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2191 | synchronize_rcu(); |
1da177e4 | 2192 | |
e498be7d | 2193 | for_each_online_cpu(i) |
b28a02de | 2194 | kfree(cachep->array[i]); |
1da177e4 LT |
2195 | |
2196 | /* NUMA: free the list3 structures */ | |
e498be7d CL |
2197 | for_each_online_node(i) { |
2198 | if ((l3 = cachep->nodelists[i])) { | |
2199 | kfree(l3->shared); | |
2200 | free_alien_cache(l3->alien); | |
2201 | kfree(l3); | |
2202 | } | |
2203 | } | |
1da177e4 LT |
2204 | kmem_cache_free(&cache_cache, cachep); |
2205 | ||
2206 | unlock_cpu_hotplug(); | |
2207 | ||
2208 | return 0; | |
2209 | } | |
2210 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2211 | ||
2212 | /* Get the memory for a slab management obj. */ | |
343e0d7a | 2213 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
b28a02de | 2214 | int colour_off, gfp_t local_flags) |
1da177e4 LT |
2215 | { |
2216 | struct slab *slabp; | |
b28a02de | 2217 | |
1da177e4 LT |
2218 | if (OFF_SLAB(cachep)) { |
2219 | /* Slab management obj is off-slab. */ | |
2220 | slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags); | |
2221 | if (!slabp) | |
2222 | return NULL; | |
2223 | } else { | |
b28a02de | 2224 | slabp = objp + colour_off; |
1da177e4 LT |
2225 | colour_off += cachep->slab_size; |
2226 | } | |
2227 | slabp->inuse = 0; | |
2228 | slabp->colouroff = colour_off; | |
b28a02de | 2229 | slabp->s_mem = objp + colour_off; |
1da177e4 LT |
2230 | |
2231 | return slabp; | |
2232 | } | |
2233 | ||
2234 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2235 | { | |
b28a02de | 2236 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2237 | } |
2238 | ||
343e0d7a | 2239 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2240 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2241 | { |
2242 | int i; | |
2243 | ||
2244 | for (i = 0; i < cachep->num; i++) { | |
3dafccf2 | 2245 | void *objp = slabp->s_mem + cachep->buffer_size * i; |
1da177e4 LT |
2246 | #if DEBUG |
2247 | /* need to poison the objs? */ | |
2248 | if (cachep->flags & SLAB_POISON) | |
2249 | poison_obj(cachep, objp, POISON_FREE); | |
2250 | if (cachep->flags & SLAB_STORE_USER) | |
2251 | *dbg_userword(cachep, objp) = NULL; | |
2252 | ||
2253 | if (cachep->flags & SLAB_RED_ZONE) { | |
2254 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2255 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2256 | } | |
2257 | /* | |
2258 | * Constructors are not allowed to allocate memory from | |
2259 | * the same cache which they are a constructor for. | |
2260 | * Otherwise, deadlock. They must also be threaded. | |
2261 | */ | |
2262 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2263 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2264 | ctor_flags); |
1da177e4 LT |
2265 | |
2266 | if (cachep->flags & SLAB_RED_ZONE) { | |
2267 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2268 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2269 | " end of an object"); |
1da177e4 LT |
2270 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2271 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2272 | " start of an object"); |
1da177e4 | 2273 | } |
3dafccf2 | 2274 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep) |
b28a02de PE |
2275 | && cachep->flags & SLAB_POISON) |
2276 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 2277 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2278 | #else |
2279 | if (cachep->ctor) | |
2280 | cachep->ctor(objp, cachep, ctor_flags); | |
2281 | #endif | |
b28a02de | 2282 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2283 | } |
b28a02de | 2284 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2285 | slabp->free = 0; |
2286 | } | |
2287 | ||
343e0d7a | 2288 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2289 | { |
2290 | if (flags & SLAB_DMA) { | |
2291 | if (!(cachep->gfpflags & GFP_DMA)) | |
2292 | BUG(); | |
2293 | } else { | |
2294 | if (cachep->gfpflags & GFP_DMA) | |
2295 | BUG(); | |
2296 | } | |
2297 | } | |
2298 | ||
343e0d7a | 2299 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid) |
78d382d7 MD |
2300 | { |
2301 | void *objp = slabp->s_mem + (slabp->free * cachep->buffer_size); | |
2302 | kmem_bufctl_t next; | |
2303 | ||
2304 | slabp->inuse++; | |
2305 | next = slab_bufctl(slabp)[slabp->free]; | |
2306 | #if DEBUG | |
2307 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2308 | WARN_ON(slabp->nodeid != nodeid); | |
2309 | #endif | |
2310 | slabp->free = next; | |
2311 | ||
2312 | return objp; | |
2313 | } | |
2314 | ||
343e0d7a | 2315 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp, |
78d382d7 MD |
2316 | int nodeid) |
2317 | { | |
2318 | unsigned int objnr = (unsigned)(objp-slabp->s_mem) / cachep->buffer_size; | |
2319 | ||
2320 | #if DEBUG | |
2321 | /* Verify that the slab belongs to the intended node */ | |
2322 | WARN_ON(slabp->nodeid != nodeid); | |
2323 | ||
2324 | if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) { | |
2325 | printk(KERN_ERR "slab: double free detected in cache " | |
2326 | "'%s', objp %p\n", cachep->name, objp); | |
2327 | BUG(); | |
2328 | } | |
2329 | #endif | |
2330 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2331 | slabp->free = objnr; | |
2332 | slabp->inuse--; | |
2333 | } | |
2334 | ||
343e0d7a | 2335 | static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp) |
1da177e4 LT |
2336 | { |
2337 | int i; | |
2338 | struct page *page; | |
2339 | ||
2340 | /* Nasty!!!!!! I hope this is OK. */ | |
2341 | i = 1 << cachep->gfporder; | |
2342 | page = virt_to_page(objp); | |
2343 | do { | |
065d41cb PE |
2344 | page_set_cache(page, cachep); |
2345 | page_set_slab(page, slabp); | |
1da177e4 LT |
2346 | page++; |
2347 | } while (--i); | |
2348 | } | |
2349 | ||
2350 | /* | |
2351 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2352 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2353 | */ | |
343e0d7a | 2354 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2355 | { |
b28a02de PE |
2356 | struct slab *slabp; |
2357 | void *objp; | |
2358 | size_t offset; | |
2359 | gfp_t local_flags; | |
2360 | unsigned long ctor_flags; | |
e498be7d | 2361 | struct kmem_list3 *l3; |
1da177e4 LT |
2362 | |
2363 | /* Be lazy and only check for valid flags here, | |
b28a02de | 2364 | * keeping it out of the critical path in kmem_cache_alloc(). |
1da177e4 | 2365 | */ |
b28a02de | 2366 | if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)) |
1da177e4 LT |
2367 | BUG(); |
2368 | if (flags & SLAB_NO_GROW) | |
2369 | return 0; | |
2370 | ||
2371 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2372 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2373 | if (!(local_flags & __GFP_WAIT)) | |
2374 | /* | |
2375 | * Not allowed to sleep. Need to tell a constructor about | |
2376 | * this - it might need to know... | |
2377 | */ | |
2378 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2379 | ||
2e1217cf | 2380 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2381 | check_irq_off(); |
2e1217cf RT |
2382 | l3 = cachep->nodelists[nodeid]; |
2383 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2384 | |
2385 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2386 | offset = l3->colour_next; |
2387 | l3->colour_next++; | |
2388 | if (l3->colour_next >= cachep->colour) | |
2389 | l3->colour_next = 0; | |
2390 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2391 | |
2e1217cf | 2392 | offset *= cachep->colour_off; |
1da177e4 LT |
2393 | |
2394 | if (local_flags & __GFP_WAIT) | |
2395 | local_irq_enable(); | |
2396 | ||
2397 | /* | |
2398 | * The test for missing atomic flag is performed here, rather than | |
2399 | * the more obvious place, simply to reduce the critical path length | |
2400 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2401 | * will eventually be caught here (where it matters). | |
2402 | */ | |
2403 | kmem_flagcheck(cachep, flags); | |
2404 | ||
e498be7d CL |
2405 | /* Get mem for the objs. |
2406 | * Attempt to allocate a physical page from 'nodeid', | |
2407 | */ | |
1da177e4 LT |
2408 | if (!(objp = kmem_getpages(cachep, flags, nodeid))) |
2409 | goto failed; | |
2410 | ||
2411 | /* Get slab management. */ | |
2412 | if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags))) | |
2413 | goto opps1; | |
2414 | ||
e498be7d | 2415 | slabp->nodeid = nodeid; |
1da177e4 LT |
2416 | set_slab_attr(cachep, slabp, objp); |
2417 | ||
2418 | cache_init_objs(cachep, slabp, ctor_flags); | |
2419 | ||
2420 | if (local_flags & __GFP_WAIT) | |
2421 | local_irq_disable(); | |
2422 | check_irq_off(); | |
e498be7d | 2423 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2424 | |
2425 | /* Make slab active. */ | |
e498be7d | 2426 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2427 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2428 | l3->free_objects += cachep->num; |
2429 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2430 | return 1; |
b28a02de | 2431 | opps1: |
1da177e4 | 2432 | kmem_freepages(cachep, objp); |
b28a02de | 2433 | failed: |
1da177e4 LT |
2434 | if (local_flags & __GFP_WAIT) |
2435 | local_irq_disable(); | |
2436 | return 0; | |
2437 | } | |
2438 | ||
2439 | #if DEBUG | |
2440 | ||
2441 | /* | |
2442 | * Perform extra freeing checks: | |
2443 | * - detect bad pointers. | |
2444 | * - POISON/RED_ZONE checking | |
2445 | * - destructor calls, for caches with POISON+dtor | |
2446 | */ | |
2447 | static void kfree_debugcheck(const void *objp) | |
2448 | { | |
2449 | struct page *page; | |
2450 | ||
2451 | if (!virt_addr_valid(objp)) { | |
2452 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2453 | (unsigned long)objp); |
2454 | BUG(); | |
1da177e4 LT |
2455 | } |
2456 | page = virt_to_page(objp); | |
2457 | if (!PageSlab(page)) { | |
b28a02de PE |
2458 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2459 | (unsigned long)objp); | |
1da177e4 LT |
2460 | BUG(); |
2461 | } | |
2462 | } | |
2463 | ||
343e0d7a | 2464 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2465 | void *caller) |
1da177e4 LT |
2466 | { |
2467 | struct page *page; | |
2468 | unsigned int objnr; | |
2469 | struct slab *slabp; | |
2470 | ||
3dafccf2 | 2471 | objp -= obj_offset(cachep); |
1da177e4 LT |
2472 | kfree_debugcheck(objp); |
2473 | page = virt_to_page(objp); | |
2474 | ||
065d41cb | 2475 | if (page_get_cache(page) != cachep) { |
b28a02de PE |
2476 | printk(KERN_ERR |
2477 | "mismatch in kmem_cache_free: expected cache %p, got %p\n", | |
2478 | page_get_cache(page), cachep); | |
1da177e4 | 2479 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2480 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2481 | page_get_cache(page)->name); | |
1da177e4 LT |
2482 | WARN_ON(1); |
2483 | } | |
065d41cb | 2484 | slabp = page_get_slab(page); |
1da177e4 LT |
2485 | |
2486 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2487 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE |
2488 | || *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2489 | slab_error(cachep, | |
2490 | "double free, or memory outside" | |
2491 | " object was overwritten"); | |
2492 | printk(KERN_ERR | |
2493 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2494 | objp, *dbg_redzone1(cachep, objp), | |
2495 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2496 | } |
2497 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2498 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2499 | } | |
2500 | if (cachep->flags & SLAB_STORE_USER) | |
2501 | *dbg_userword(cachep, objp) = caller; | |
2502 | ||
3dafccf2 | 2503 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
1da177e4 LT |
2504 | |
2505 | BUG_ON(objnr >= cachep->num); | |
3dafccf2 | 2506 | BUG_ON(objp != slabp->s_mem + objnr * cachep->buffer_size); |
1da177e4 LT |
2507 | |
2508 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
2509 | /* Need to call the slab's constructor so the | |
2510 | * caller can perform a verify of its state (debugging). | |
2511 | * Called without the cache-lock held. | |
2512 | */ | |
3dafccf2 | 2513 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2514 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2515 | } |
2516 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2517 | /* we want to cache poison the object, | |
2518 | * call the destruction callback | |
2519 | */ | |
3dafccf2 | 2520 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 LT |
2521 | } |
2522 | if (cachep->flags & SLAB_POISON) { | |
2523 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 2524 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) { |
1da177e4 | 2525 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2526 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2527 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2528 | } else { |
2529 | poison_obj(cachep, objp, POISON_FREE); | |
2530 | } | |
2531 | #else | |
2532 | poison_obj(cachep, objp, POISON_FREE); | |
2533 | #endif | |
2534 | } | |
2535 | return objp; | |
2536 | } | |
2537 | ||
343e0d7a | 2538 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2539 | { |
2540 | kmem_bufctl_t i; | |
2541 | int entries = 0; | |
b28a02de | 2542 | |
1da177e4 LT |
2543 | /* Check slab's freelist to see if this obj is there. */ |
2544 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2545 | entries++; | |
2546 | if (entries > cachep->num || i >= cachep->num) | |
2547 | goto bad; | |
2548 | } | |
2549 | if (entries != cachep->num - slabp->inuse) { | |
b28a02de PE |
2550 | bad: |
2551 | printk(KERN_ERR | |
2552 | "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2553 | cachep->name, cachep->num, slabp, slabp->inuse); | |
2554 | for (i = 0; | |
264132bc | 2555 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de PE |
2556 | i++) { |
2557 | if ((i % 16) == 0) | |
1da177e4 | 2558 | printk("\n%03x:", i); |
b28a02de | 2559 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2560 | } |
2561 | printk("\n"); | |
2562 | BUG(); | |
2563 | } | |
2564 | } | |
2565 | #else | |
2566 | #define kfree_debugcheck(x) do { } while(0) | |
2567 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2568 | #define check_slabp(x,y) do { } while(0) | |
2569 | #endif | |
2570 | ||
343e0d7a | 2571 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2572 | { |
2573 | int batchcount; | |
2574 | struct kmem_list3 *l3; | |
2575 | struct array_cache *ac; | |
2576 | ||
2577 | check_irq_off(); | |
9a2dba4b | 2578 | ac = cpu_cache_get(cachep); |
b28a02de | 2579 | retry: |
1da177e4 LT |
2580 | batchcount = ac->batchcount; |
2581 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
2582 | /* if there was little recent activity on this | |
2583 | * cache, then perform only a partial refill. | |
2584 | * Otherwise we could generate refill bouncing. | |
2585 | */ | |
2586 | batchcount = BATCHREFILL_LIMIT; | |
2587 | } | |
e498be7d CL |
2588 | l3 = cachep->nodelists[numa_node_id()]; |
2589 | ||
2590 | BUG_ON(ac->avail > 0 || !l3); | |
2591 | spin_lock(&l3->list_lock); | |
1da177e4 | 2592 | |
1da177e4 LT |
2593 | if (l3->shared) { |
2594 | struct array_cache *shared_array = l3->shared; | |
2595 | if (shared_array->avail) { | |
2596 | if (batchcount > shared_array->avail) | |
2597 | batchcount = shared_array->avail; | |
2598 | shared_array->avail -= batchcount; | |
2599 | ac->avail = batchcount; | |
e498be7d | 2600 | memcpy(ac->entry, |
b28a02de PE |
2601 | &(shared_array->entry[shared_array->avail]), |
2602 | sizeof(void *) * batchcount); | |
1da177e4 LT |
2603 | shared_array->touched = 1; |
2604 | goto alloc_done; | |
2605 | } | |
2606 | } | |
2607 | while (batchcount > 0) { | |
2608 | struct list_head *entry; | |
2609 | struct slab *slabp; | |
2610 | /* Get slab alloc is to come from. */ | |
2611 | entry = l3->slabs_partial.next; | |
2612 | if (entry == &l3->slabs_partial) { | |
2613 | l3->free_touched = 1; | |
2614 | entry = l3->slabs_free.next; | |
2615 | if (entry == &l3->slabs_free) | |
2616 | goto must_grow; | |
2617 | } | |
2618 | ||
2619 | slabp = list_entry(entry, struct slab, list); | |
2620 | check_slabp(cachep, slabp); | |
2621 | check_spinlock_acquired(cachep); | |
2622 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2623 | STATS_INC_ALLOCED(cachep); |
2624 | STATS_INC_ACTIVE(cachep); | |
2625 | STATS_SET_HIGH(cachep); | |
2626 | ||
78d382d7 MD |
2627 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2628 | numa_node_id()); | |
1da177e4 LT |
2629 | } |
2630 | check_slabp(cachep, slabp); | |
2631 | ||
2632 | /* move slabp to correct slabp list: */ | |
2633 | list_del(&slabp->list); | |
2634 | if (slabp->free == BUFCTL_END) | |
2635 | list_add(&slabp->list, &l3->slabs_full); | |
2636 | else | |
2637 | list_add(&slabp->list, &l3->slabs_partial); | |
2638 | } | |
2639 | ||
b28a02de | 2640 | must_grow: |
1da177e4 | 2641 | l3->free_objects -= ac->avail; |
b28a02de | 2642 | alloc_done: |
e498be7d | 2643 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2644 | |
2645 | if (unlikely(!ac->avail)) { | |
2646 | int x; | |
e498be7d CL |
2647 | x = cache_grow(cachep, flags, numa_node_id()); |
2648 | ||
1da177e4 | 2649 | // cache_grow can reenable interrupts, then ac could change. |
9a2dba4b | 2650 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2651 | if (!x && ac->avail == 0) // no objects in sight? abort |
2652 | return NULL; | |
2653 | ||
b28a02de | 2654 | if (!ac->avail) // objects refilled by interrupt? |
1da177e4 LT |
2655 | goto retry; |
2656 | } | |
2657 | ac->touched = 1; | |
e498be7d | 2658 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2659 | } |
2660 | ||
2661 | static inline void | |
343e0d7a | 2662 | cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2663 | { |
2664 | might_sleep_if(flags & __GFP_WAIT); | |
2665 | #if DEBUG | |
2666 | kmem_flagcheck(cachep, flags); | |
2667 | #endif | |
2668 | } | |
2669 | ||
2670 | #if DEBUG | |
343e0d7a | 2671 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags, |
b28a02de | 2672 | void *objp, void *caller) |
1da177e4 | 2673 | { |
b28a02de | 2674 | if (!objp) |
1da177e4 | 2675 | return objp; |
b28a02de | 2676 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2677 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2678 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2679 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2680 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2681 | else |
2682 | check_poison_obj(cachep, objp); | |
2683 | #else | |
2684 | check_poison_obj(cachep, objp); | |
2685 | #endif | |
2686 | poison_obj(cachep, objp, POISON_INUSE); | |
2687 | } | |
2688 | if (cachep->flags & SLAB_STORE_USER) | |
2689 | *dbg_userword(cachep, objp) = caller; | |
2690 | ||
2691 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2692 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE |
2693 | || *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2694 | slab_error(cachep, | |
2695 | "double free, or memory outside" | |
2696 | " object was overwritten"); | |
2697 | printk(KERN_ERR | |
2698 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2699 | objp, *dbg_redzone1(cachep, objp), | |
2700 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2701 | } |
2702 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2703 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2704 | } | |
3dafccf2 | 2705 | objp += obj_offset(cachep); |
1da177e4 | 2706 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2707 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2708 | |
2709 | if (!(flags & __GFP_WAIT)) | |
2710 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2711 | ||
2712 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2713 | } |
1da177e4 LT |
2714 | return objp; |
2715 | } | |
2716 | #else | |
2717 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2718 | #endif | |
2719 | ||
343e0d7a | 2720 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2721 | { |
b28a02de | 2722 | void *objp; |
1da177e4 LT |
2723 | struct array_cache *ac; |
2724 | ||
dc85da15 | 2725 | #ifdef CONFIG_NUMA |
86c562a9 | 2726 | if (unlikely(current->mempolicy && !in_interrupt())) { |
dc85da15 CL |
2727 | int nid = slab_node(current->mempolicy); |
2728 | ||
2729 | if (nid != numa_node_id()) | |
2730 | return __cache_alloc_node(cachep, flags, nid); | |
2731 | } | |
2732 | #endif | |
2733 | ||
5c382300 | 2734 | check_irq_off(); |
9a2dba4b | 2735 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2736 | if (likely(ac->avail)) { |
2737 | STATS_INC_ALLOCHIT(cachep); | |
2738 | ac->touched = 1; | |
e498be7d | 2739 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2740 | } else { |
2741 | STATS_INC_ALLOCMISS(cachep); | |
2742 | objp = cache_alloc_refill(cachep, flags); | |
2743 | } | |
5c382300 AK |
2744 | return objp; |
2745 | } | |
2746 | ||
7fd6b141 PE |
2747 | static __always_inline void * |
2748 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
5c382300 AK |
2749 | { |
2750 | unsigned long save_flags; | |
b28a02de | 2751 | void *objp; |
5c382300 AK |
2752 | |
2753 | cache_alloc_debugcheck_before(cachep, flags); | |
2754 | ||
2755 | local_irq_save(save_flags); | |
2756 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2757 | local_irq_restore(save_flags); |
34342e86 | 2758 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 2759 | caller); |
34342e86 | 2760 | prefetchw(objp); |
1da177e4 LT |
2761 | return objp; |
2762 | } | |
2763 | ||
e498be7d CL |
2764 | #ifdef CONFIG_NUMA |
2765 | /* | |
2766 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2767 | */ |
343e0d7a | 2768 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
e498be7d CL |
2769 | { |
2770 | struct list_head *entry; | |
b28a02de PE |
2771 | struct slab *slabp; |
2772 | struct kmem_list3 *l3; | |
2773 | void *obj; | |
b28a02de PE |
2774 | int x; |
2775 | ||
2776 | l3 = cachep->nodelists[nodeid]; | |
2777 | BUG_ON(!l3); | |
2778 | ||
2779 | retry: | |
ca3b9b91 | 2780 | check_irq_off(); |
b28a02de PE |
2781 | spin_lock(&l3->list_lock); |
2782 | entry = l3->slabs_partial.next; | |
2783 | if (entry == &l3->slabs_partial) { | |
2784 | l3->free_touched = 1; | |
2785 | entry = l3->slabs_free.next; | |
2786 | if (entry == &l3->slabs_free) | |
2787 | goto must_grow; | |
2788 | } | |
2789 | ||
2790 | slabp = list_entry(entry, struct slab, list); | |
2791 | check_spinlock_acquired_node(cachep, nodeid); | |
2792 | check_slabp(cachep, slabp); | |
2793 | ||
2794 | STATS_INC_NODEALLOCS(cachep); | |
2795 | STATS_INC_ACTIVE(cachep); | |
2796 | STATS_SET_HIGH(cachep); | |
2797 | ||
2798 | BUG_ON(slabp->inuse == cachep->num); | |
2799 | ||
78d382d7 | 2800 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
2801 | check_slabp(cachep, slabp); |
2802 | l3->free_objects--; | |
2803 | /* move slabp to correct slabp list: */ | |
2804 | list_del(&slabp->list); | |
2805 | ||
2806 | if (slabp->free == BUFCTL_END) { | |
2807 | list_add(&slabp->list, &l3->slabs_full); | |
2808 | } else { | |
2809 | list_add(&slabp->list, &l3->slabs_partial); | |
2810 | } | |
e498be7d | 2811 | |
b28a02de PE |
2812 | spin_unlock(&l3->list_lock); |
2813 | goto done; | |
e498be7d | 2814 | |
b28a02de PE |
2815 | must_grow: |
2816 | spin_unlock(&l3->list_lock); | |
2817 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 2818 | |
b28a02de PE |
2819 | if (!x) |
2820 | return NULL; | |
e498be7d | 2821 | |
b28a02de PE |
2822 | goto retry; |
2823 | done: | |
2824 | return obj; | |
e498be7d CL |
2825 | } |
2826 | #endif | |
2827 | ||
2828 | /* | |
2829 | * Caller needs to acquire correct kmem_list's list_lock | |
2830 | */ | |
343e0d7a | 2831 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 2832 | int node) |
1da177e4 LT |
2833 | { |
2834 | int i; | |
e498be7d | 2835 | struct kmem_list3 *l3; |
1da177e4 LT |
2836 | |
2837 | for (i = 0; i < nr_objects; i++) { | |
2838 | void *objp = objpp[i]; | |
2839 | struct slab *slabp; | |
1da177e4 | 2840 | |
6ed5eb22 | 2841 | slabp = virt_to_slab(objp); |
ff69416e | 2842 | l3 = cachep->nodelists[node]; |
1da177e4 | 2843 | list_del(&slabp->list); |
ff69416e | 2844 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 2845 | check_slabp(cachep, slabp); |
78d382d7 | 2846 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 2847 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 2848 | l3->free_objects++; |
1da177e4 LT |
2849 | check_slabp(cachep, slabp); |
2850 | ||
2851 | /* fixup slab chains */ | |
2852 | if (slabp->inuse == 0) { | |
e498be7d CL |
2853 | if (l3->free_objects > l3->free_limit) { |
2854 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
2855 | slab_destroy(cachep, slabp); |
2856 | } else { | |
e498be7d | 2857 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
2858 | } |
2859 | } else { | |
2860 | /* Unconditionally move a slab to the end of the | |
2861 | * partial list on free - maximum time for the | |
2862 | * other objects to be freed, too. | |
2863 | */ | |
e498be7d | 2864 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
2865 | } |
2866 | } | |
2867 | } | |
2868 | ||
343e0d7a | 2869 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
2870 | { |
2871 | int batchcount; | |
e498be7d | 2872 | struct kmem_list3 *l3; |
ff69416e | 2873 | int node = numa_node_id(); |
1da177e4 LT |
2874 | |
2875 | batchcount = ac->batchcount; | |
2876 | #if DEBUG | |
2877 | BUG_ON(!batchcount || batchcount > ac->avail); | |
2878 | #endif | |
2879 | check_irq_off(); | |
ff69416e | 2880 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2881 | spin_lock(&l3->list_lock); |
2882 | if (l3->shared) { | |
2883 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 2884 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
2885 | if (max) { |
2886 | if (batchcount > max) | |
2887 | batchcount = max; | |
e498be7d | 2888 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 2889 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
2890 | shared_array->avail += batchcount; |
2891 | goto free_done; | |
2892 | } | |
2893 | } | |
2894 | ||
ff69416e | 2895 | free_block(cachep, ac->entry, batchcount, node); |
b28a02de | 2896 | free_done: |
1da177e4 LT |
2897 | #if STATS |
2898 | { | |
2899 | int i = 0; | |
2900 | struct list_head *p; | |
2901 | ||
e498be7d CL |
2902 | p = l3->slabs_free.next; |
2903 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
2904 | struct slab *slabp; |
2905 | ||
2906 | slabp = list_entry(p, struct slab, list); | |
2907 | BUG_ON(slabp->inuse); | |
2908 | ||
2909 | i++; | |
2910 | p = p->next; | |
2911 | } | |
2912 | STATS_SET_FREEABLE(cachep, i); | |
2913 | } | |
2914 | #endif | |
e498be7d | 2915 | spin_unlock(&l3->list_lock); |
1da177e4 | 2916 | ac->avail -= batchcount; |
e498be7d | 2917 | memmove(ac->entry, &(ac->entry[batchcount]), |
b28a02de | 2918 | sizeof(void *) * ac->avail); |
1da177e4 LT |
2919 | } |
2920 | ||
2921 | /* | |
2922 | * __cache_free | |
2923 | * Release an obj back to its cache. If the obj has a constructed | |
2924 | * state, it must be in this state _before_ it is released. | |
2925 | * | |
2926 | * Called with disabled ints. | |
2927 | */ | |
343e0d7a | 2928 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 2929 | { |
9a2dba4b | 2930 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
2931 | |
2932 | check_irq_off(); | |
2933 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
2934 | ||
e498be7d CL |
2935 | /* Make sure we are not freeing a object from another |
2936 | * node to the array cache on this cpu. | |
2937 | */ | |
2938 | #ifdef CONFIG_NUMA | |
2939 | { | |
2940 | struct slab *slabp; | |
6ed5eb22 | 2941 | slabp = virt_to_slab(objp); |
e498be7d CL |
2942 | if (unlikely(slabp->nodeid != numa_node_id())) { |
2943 | struct array_cache *alien = NULL; | |
2944 | int nodeid = slabp->nodeid; | |
b28a02de PE |
2945 | struct kmem_list3 *l3 = |
2946 | cachep->nodelists[numa_node_id()]; | |
e498be7d CL |
2947 | |
2948 | STATS_INC_NODEFREES(cachep); | |
2949 | if (l3->alien && l3->alien[nodeid]) { | |
2950 | alien = l3->alien[nodeid]; | |
2951 | spin_lock(&alien->lock); | |
2952 | if (unlikely(alien->avail == alien->limit)) | |
2953 | __drain_alien_cache(cachep, | |
b28a02de | 2954 | alien, nodeid); |
e498be7d CL |
2955 | alien->entry[alien->avail++] = objp; |
2956 | spin_unlock(&alien->lock); | |
2957 | } else { | |
2958 | spin_lock(&(cachep->nodelists[nodeid])-> | |
b28a02de | 2959 | list_lock); |
ff69416e | 2960 | free_block(cachep, &objp, 1, nodeid); |
e498be7d | 2961 | spin_unlock(&(cachep->nodelists[nodeid])-> |
b28a02de | 2962 | list_lock); |
e498be7d CL |
2963 | } |
2964 | return; | |
2965 | } | |
2966 | } | |
2967 | #endif | |
1da177e4 LT |
2968 | if (likely(ac->avail < ac->limit)) { |
2969 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 2970 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2971 | return; |
2972 | } else { | |
2973 | STATS_INC_FREEMISS(cachep); | |
2974 | cache_flusharray(cachep, ac); | |
e498be7d | 2975 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
2976 | } |
2977 | } | |
2978 | ||
2979 | /** | |
2980 | * kmem_cache_alloc - Allocate an object | |
2981 | * @cachep: The cache to allocate from. | |
2982 | * @flags: See kmalloc(). | |
2983 | * | |
2984 | * Allocate an object from this cache. The flags are only relevant | |
2985 | * if the cache has no available objects. | |
2986 | */ | |
343e0d7a | 2987 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2988 | { |
7fd6b141 | 2989 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
2990 | } |
2991 | EXPORT_SYMBOL(kmem_cache_alloc); | |
2992 | ||
2993 | /** | |
2994 | * kmem_ptr_validate - check if an untrusted pointer might | |
2995 | * be a slab entry. | |
2996 | * @cachep: the cache we're checking against | |
2997 | * @ptr: pointer to validate | |
2998 | * | |
2999 | * This verifies that the untrusted pointer looks sane: | |
3000 | * it is _not_ a guarantee that the pointer is actually | |
3001 | * part of the slab cache in question, but it at least | |
3002 | * validates that the pointer can be dereferenced and | |
3003 | * looks half-way sane. | |
3004 | * | |
3005 | * Currently only used for dentry validation. | |
3006 | */ | |
343e0d7a | 3007 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3008 | { |
b28a02de | 3009 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3010 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3011 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3012 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3013 | struct page *page; |
3014 | ||
3015 | if (unlikely(addr < min_addr)) | |
3016 | goto out; | |
3017 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3018 | goto out; | |
3019 | if (unlikely(addr & align_mask)) | |
3020 | goto out; | |
3021 | if (unlikely(!kern_addr_valid(addr))) | |
3022 | goto out; | |
3023 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3024 | goto out; | |
3025 | page = virt_to_page(ptr); | |
3026 | if (unlikely(!PageSlab(page))) | |
3027 | goto out; | |
065d41cb | 3028 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3029 | goto out; |
3030 | return 1; | |
b28a02de | 3031 | out: |
1da177e4 LT |
3032 | return 0; |
3033 | } | |
3034 | ||
3035 | #ifdef CONFIG_NUMA | |
3036 | /** | |
3037 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3038 | * @cachep: The cache to allocate from. | |
3039 | * @flags: See kmalloc(). | |
3040 | * @nodeid: node number of the target node. | |
3041 | * | |
3042 | * Identical to kmem_cache_alloc, except that this function is slow | |
3043 | * and can sleep. And it will allocate memory on the given node, which | |
3044 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3045 | * New and improved: it will now make sure that the object gets |
3046 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3047 | */ |
343e0d7a | 3048 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3049 | { |
e498be7d CL |
3050 | unsigned long save_flags; |
3051 | void *ptr; | |
1da177e4 | 3052 | |
e498be7d CL |
3053 | cache_alloc_debugcheck_before(cachep, flags); |
3054 | local_irq_save(save_flags); | |
18f820f6 CL |
3055 | |
3056 | if (nodeid == -1 || nodeid == numa_node_id() || | |
3057 | !cachep->nodelists[nodeid]) | |
5c382300 AK |
3058 | ptr = ____cache_alloc(cachep, flags); |
3059 | else | |
3060 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3061 | local_irq_restore(save_flags); |
18f820f6 CL |
3062 | |
3063 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3064 | __builtin_return_address(0)); | |
1da177e4 | 3065 | |
e498be7d | 3066 | return ptr; |
1da177e4 LT |
3067 | } |
3068 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3069 | ||
dd0fc66f | 3070 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3071 | { |
343e0d7a | 3072 | struct kmem_cache *cachep; |
97e2bde4 MS |
3073 | |
3074 | cachep = kmem_find_general_cachep(size, flags); | |
3075 | if (unlikely(cachep == NULL)) | |
3076 | return NULL; | |
3077 | return kmem_cache_alloc_node(cachep, flags, node); | |
3078 | } | |
3079 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
3080 | #endif |
3081 | ||
3082 | /** | |
3083 | * kmalloc - allocate memory | |
3084 | * @size: how many bytes of memory are required. | |
3085 | * @flags: the type of memory to allocate. | |
3086 | * | |
3087 | * kmalloc is the normal method of allocating memory | |
3088 | * in the kernel. | |
3089 | * | |
3090 | * The @flags argument may be one of: | |
3091 | * | |
3092 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
3093 | * | |
3094 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
3095 | * | |
3096 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
3097 | * | |
3098 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
3099 | * must be suitable for DMA. This can mean different things on different | |
3100 | * platforms. For example, on i386, it means that the memory must come | |
3101 | * from the first 16MB. | |
3102 | */ | |
7fd6b141 PE |
3103 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3104 | void *caller) | |
1da177e4 | 3105 | { |
343e0d7a | 3106 | struct kmem_cache *cachep; |
1da177e4 | 3107 | |
97e2bde4 MS |
3108 | /* If you want to save a few bytes .text space: replace |
3109 | * __ with kmem_. | |
3110 | * Then kmalloc uses the uninlined functions instead of the inline | |
3111 | * functions. | |
3112 | */ | |
3113 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3114 | if (unlikely(cachep == NULL)) |
3115 | return NULL; | |
7fd6b141 PE |
3116 | return __cache_alloc(cachep, flags, caller); |
3117 | } | |
3118 | ||
3119 | #ifndef CONFIG_DEBUG_SLAB | |
3120 | ||
3121 | void *__kmalloc(size_t size, gfp_t flags) | |
3122 | { | |
3123 | return __do_kmalloc(size, flags, NULL); | |
1da177e4 LT |
3124 | } |
3125 | EXPORT_SYMBOL(__kmalloc); | |
3126 | ||
7fd6b141 PE |
3127 | #else |
3128 | ||
3129 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) | |
3130 | { | |
3131 | return __do_kmalloc(size, flags, caller); | |
3132 | } | |
3133 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
3134 | ||
3135 | #endif | |
3136 | ||
1da177e4 LT |
3137 | #ifdef CONFIG_SMP |
3138 | /** | |
3139 | * __alloc_percpu - allocate one copy of the object for every present | |
3140 | * cpu in the system, zeroing them. | |
3141 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
3142 | * | |
3143 | * @size: how many bytes of memory are required. | |
1da177e4 | 3144 | */ |
f9f75005 | 3145 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
3146 | { |
3147 | int i; | |
b28a02de | 3148 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3149 | |
3150 | if (!pdata) | |
3151 | return NULL; | |
3152 | ||
e498be7d CL |
3153 | /* |
3154 | * Cannot use for_each_online_cpu since a cpu may come online | |
3155 | * and we have no way of figuring out how to fix the array | |
3156 | * that we have allocated then.... | |
3157 | */ | |
3158 | for_each_cpu(i) { | |
3159 | int node = cpu_to_node(i); | |
3160 | ||
3161 | if (node_online(node)) | |
3162 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3163 | else | |
3164 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3165 | |
3166 | if (!pdata->ptrs[i]) | |
3167 | goto unwind_oom; | |
3168 | memset(pdata->ptrs[i], 0, size); | |
3169 | } | |
3170 | ||
3171 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3172 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3173 | |
b28a02de | 3174 | unwind_oom: |
1da177e4 LT |
3175 | while (--i >= 0) { |
3176 | if (!cpu_possible(i)) | |
3177 | continue; | |
3178 | kfree(pdata->ptrs[i]); | |
3179 | } | |
3180 | kfree(pdata); | |
3181 | return NULL; | |
3182 | } | |
3183 | EXPORT_SYMBOL(__alloc_percpu); | |
3184 | #endif | |
3185 | ||
3186 | /** | |
3187 | * kmem_cache_free - Deallocate an object | |
3188 | * @cachep: The cache the allocation was from. | |
3189 | * @objp: The previously allocated object. | |
3190 | * | |
3191 | * Free an object which was previously allocated from this | |
3192 | * cache. | |
3193 | */ | |
343e0d7a | 3194 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3195 | { |
3196 | unsigned long flags; | |
3197 | ||
3198 | local_irq_save(flags); | |
3199 | __cache_free(cachep, objp); | |
3200 | local_irq_restore(flags); | |
3201 | } | |
3202 | EXPORT_SYMBOL(kmem_cache_free); | |
3203 | ||
1da177e4 LT |
3204 | /** |
3205 | * kfree - free previously allocated memory | |
3206 | * @objp: pointer returned by kmalloc. | |
3207 | * | |
80e93eff PE |
3208 | * If @objp is NULL, no operation is performed. |
3209 | * | |
1da177e4 LT |
3210 | * Don't free memory not originally allocated by kmalloc() |
3211 | * or you will run into trouble. | |
3212 | */ | |
3213 | void kfree(const void *objp) | |
3214 | { | |
343e0d7a | 3215 | struct kmem_cache *c; |
1da177e4 LT |
3216 | unsigned long flags; |
3217 | ||
3218 | if (unlikely(!objp)) | |
3219 | return; | |
3220 | local_irq_save(flags); | |
3221 | kfree_debugcheck(objp); | |
6ed5eb22 | 3222 | c = virt_to_cache(objp); |
3dafccf2 | 3223 | mutex_debug_check_no_locks_freed(objp, obj_size(c)); |
b28a02de | 3224 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3225 | local_irq_restore(flags); |
3226 | } | |
3227 | EXPORT_SYMBOL(kfree); | |
3228 | ||
3229 | #ifdef CONFIG_SMP | |
3230 | /** | |
3231 | * free_percpu - free previously allocated percpu memory | |
3232 | * @objp: pointer returned by alloc_percpu. | |
3233 | * | |
3234 | * Don't free memory not originally allocated by alloc_percpu() | |
3235 | * The complemented objp is to check for that. | |
3236 | */ | |
b28a02de | 3237 | void free_percpu(const void *objp) |
1da177e4 LT |
3238 | { |
3239 | int i; | |
b28a02de | 3240 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3241 | |
e498be7d CL |
3242 | /* |
3243 | * We allocate for all cpus so we cannot use for online cpu here. | |
3244 | */ | |
3245 | for_each_cpu(i) | |
b28a02de | 3246 | kfree(p->ptrs[i]); |
1da177e4 LT |
3247 | kfree(p); |
3248 | } | |
3249 | EXPORT_SYMBOL(free_percpu); | |
3250 | #endif | |
3251 | ||
343e0d7a | 3252 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3253 | { |
3dafccf2 | 3254 | return obj_size(cachep); |
1da177e4 LT |
3255 | } |
3256 | EXPORT_SYMBOL(kmem_cache_size); | |
3257 | ||
343e0d7a | 3258 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3259 | { |
3260 | return cachep->name; | |
3261 | } | |
3262 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3263 | ||
e498be7d CL |
3264 | /* |
3265 | * This initializes kmem_list3 for all nodes. | |
3266 | */ | |
343e0d7a | 3267 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3268 | { |
3269 | int node; | |
3270 | struct kmem_list3 *l3; | |
3271 | int err = 0; | |
3272 | ||
3273 | for_each_online_node(node) { | |
3274 | struct array_cache *nc = NULL, *new; | |
3275 | struct array_cache **new_alien = NULL; | |
3276 | #ifdef CONFIG_NUMA | |
3277 | if (!(new_alien = alloc_alien_cache(node, cachep->limit))) | |
3278 | goto fail; | |
3279 | #endif | |
b28a02de PE |
3280 | if (!(new = alloc_arraycache(node, (cachep->shared * |
3281 | cachep->batchcount), | |
3282 | 0xbaadf00d))) | |
e498be7d CL |
3283 | goto fail; |
3284 | if ((l3 = cachep->nodelists[node])) { | |
3285 | ||
3286 | spin_lock_irq(&l3->list_lock); | |
3287 | ||
3288 | if ((nc = cachep->nodelists[node]->shared)) | |
b28a02de | 3289 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
3290 | |
3291 | l3->shared = new; | |
3292 | if (!cachep->nodelists[node]->alien) { | |
3293 | l3->alien = new_alien; | |
3294 | new_alien = NULL; | |
3295 | } | |
b28a02de PE |
3296 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3297 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3298 | spin_unlock_irq(&l3->list_lock); |
3299 | kfree(nc); | |
3300 | free_alien_cache(new_alien); | |
3301 | continue; | |
3302 | } | |
3303 | if (!(l3 = kmalloc_node(sizeof(struct kmem_list3), | |
b28a02de | 3304 | GFP_KERNEL, node))) |
e498be7d CL |
3305 | goto fail; |
3306 | ||
3307 | kmem_list3_init(l3); | |
3308 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 3309 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
3310 | l3->shared = new; |
3311 | l3->alien = new_alien; | |
b28a02de PE |
3312 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3313 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3314 | cachep->nodelists[node] = l3; |
3315 | } | |
3316 | return err; | |
b28a02de | 3317 | fail: |
e498be7d CL |
3318 | err = -ENOMEM; |
3319 | return err; | |
3320 | } | |
3321 | ||
1da177e4 | 3322 | struct ccupdate_struct { |
343e0d7a | 3323 | struct kmem_cache *cachep; |
1da177e4 LT |
3324 | struct array_cache *new[NR_CPUS]; |
3325 | }; | |
3326 | ||
3327 | static void do_ccupdate_local(void *info) | |
3328 | { | |
3329 | struct ccupdate_struct *new = (struct ccupdate_struct *)info; | |
3330 | struct array_cache *old; | |
3331 | ||
3332 | check_irq_off(); | |
9a2dba4b | 3333 | old = cpu_cache_get(new->cachep); |
e498be7d | 3334 | |
1da177e4 LT |
3335 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3336 | new->new[smp_processor_id()] = old; | |
3337 | } | |
3338 | ||
343e0d7a | 3339 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount, |
b28a02de | 3340 | int shared) |
1da177e4 LT |
3341 | { |
3342 | struct ccupdate_struct new; | |
e498be7d | 3343 | int i, err; |
1da177e4 | 3344 | |
b28a02de | 3345 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3346 | for_each_online_cpu(i) { |
b28a02de PE |
3347 | new.new[i] = |
3348 | alloc_arraycache(cpu_to_node(i), limit, batchcount); | |
e498be7d | 3349 | if (!new.new[i]) { |
b28a02de PE |
3350 | for (i--; i >= 0; i--) |
3351 | kfree(new.new[i]); | |
e498be7d | 3352 | return -ENOMEM; |
1da177e4 LT |
3353 | } |
3354 | } | |
3355 | new.cachep = cachep; | |
3356 | ||
3357 | smp_call_function_all_cpus(do_ccupdate_local, (void *)&new); | |
e498be7d | 3358 | |
1da177e4 | 3359 | check_irq_on(); |
ca3b9b91 | 3360 | spin_lock(&cachep->spinlock); |
1da177e4 LT |
3361 | cachep->batchcount = batchcount; |
3362 | cachep->limit = limit; | |
e498be7d | 3363 | cachep->shared = shared; |
ca3b9b91 | 3364 | spin_unlock(&cachep->spinlock); |
1da177e4 | 3365 | |
e498be7d | 3366 | for_each_online_cpu(i) { |
1da177e4 LT |
3367 | struct array_cache *ccold = new.new[i]; |
3368 | if (!ccold) | |
3369 | continue; | |
e498be7d | 3370 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3371 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3372 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3373 | kfree(ccold); |
3374 | } | |
1da177e4 | 3375 | |
e498be7d CL |
3376 | err = alloc_kmemlist(cachep); |
3377 | if (err) { | |
3378 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3379 | cachep->name, -err); |
e498be7d | 3380 | BUG(); |
1da177e4 | 3381 | } |
1da177e4 LT |
3382 | return 0; |
3383 | } | |
3384 | ||
343e0d7a | 3385 | static void enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3386 | { |
3387 | int err; | |
3388 | int limit, shared; | |
3389 | ||
3390 | /* The head array serves three purposes: | |
3391 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
3392 | * - reduce the number of spinlock operations. | |
3393 | * - reduce the number of linked list operations on the slab and | |
3394 | * bufctl chains: array operations are cheaper. | |
3395 | * The numbers are guessed, we should auto-tune as described by | |
3396 | * Bonwick. | |
3397 | */ | |
3dafccf2 | 3398 | if (cachep->buffer_size > 131072) |
1da177e4 | 3399 | limit = 1; |
3dafccf2 | 3400 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3401 | limit = 8; |
3dafccf2 | 3402 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3403 | limit = 24; |
3dafccf2 | 3404 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3405 | limit = 54; |
3406 | else | |
3407 | limit = 120; | |
3408 | ||
3409 | /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound | |
3410 | * allocation behaviour: Most allocs on one cpu, most free operations | |
3411 | * on another cpu. For these cases, an efficient object passing between | |
3412 | * cpus is necessary. This is provided by a shared array. The array | |
3413 | * replaces Bonwick's magazine layer. | |
3414 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3415 | * to a larger limit. Thus disabled by default. | |
3416 | */ | |
3417 | shared = 0; | |
3418 | #ifdef CONFIG_SMP | |
3dafccf2 | 3419 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3420 | shared = 8; |
3421 | #endif | |
3422 | ||
3423 | #if DEBUG | |
3424 | /* With debugging enabled, large batchcount lead to excessively | |
3425 | * long periods with disabled local interrupts. Limit the | |
3426 | * batchcount | |
3427 | */ | |
3428 | if (limit > 32) | |
3429 | limit = 32; | |
3430 | #endif | |
b28a02de | 3431 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3432 | if (err) |
3433 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3434 | cachep->name, -err); |
1da177e4 LT |
3435 | } |
3436 | ||
343e0d7a | 3437 | static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, |
b28a02de | 3438 | int force, int node) |
1da177e4 LT |
3439 | { |
3440 | int tofree; | |
3441 | ||
e498be7d | 3442 | check_spinlock_acquired_node(cachep, node); |
1da177e4 LT |
3443 | if (ac->touched && !force) { |
3444 | ac->touched = 0; | |
3445 | } else if (ac->avail) { | |
b28a02de | 3446 | tofree = force ? ac->avail : (ac->limit + 4) / 5; |
1da177e4 | 3447 | if (tofree > ac->avail) { |
b28a02de | 3448 | tofree = (ac->avail + 1) / 2; |
1da177e4 | 3449 | } |
ff69416e | 3450 | free_block(cachep, ac->entry, tofree, node); |
1da177e4 | 3451 | ac->avail -= tofree; |
e498be7d | 3452 | memmove(ac->entry, &(ac->entry[tofree]), |
b28a02de | 3453 | sizeof(void *) * ac->avail); |
1da177e4 LT |
3454 | } |
3455 | } | |
3456 | ||
3457 | /** | |
3458 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3459 | * @unused: unused parameter |
1da177e4 LT |
3460 | * |
3461 | * Called from workqueue/eventd every few seconds. | |
3462 | * Purpose: | |
3463 | * - clear the per-cpu caches for this CPU. | |
3464 | * - return freeable pages to the main free memory pool. | |
3465 | * | |
fc0abb14 | 3466 | * If we cannot acquire the cache chain mutex then just give up - we'll |
1da177e4 LT |
3467 | * try again on the next iteration. |
3468 | */ | |
3469 | static void cache_reap(void *unused) | |
3470 | { | |
3471 | struct list_head *walk; | |
e498be7d | 3472 | struct kmem_list3 *l3; |
1da177e4 | 3473 | |
fc0abb14 | 3474 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3475 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3476 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3477 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3478 | return; |
3479 | } | |
3480 | ||
3481 | list_for_each(walk, &cache_chain) { | |
343e0d7a | 3482 | struct kmem_cache *searchp; |
b28a02de | 3483 | struct list_head *p; |
1da177e4 LT |
3484 | int tofree; |
3485 | struct slab *slabp; | |
3486 | ||
343e0d7a | 3487 | searchp = list_entry(walk, struct kmem_cache, next); |
1da177e4 LT |
3488 | |
3489 | if (searchp->flags & SLAB_NO_REAP) | |
3490 | goto next; | |
3491 | ||
3492 | check_irq_on(); | |
3493 | ||
e498be7d CL |
3494 | l3 = searchp->nodelists[numa_node_id()]; |
3495 | if (l3->alien) | |
4484ebf1 | 3496 | drain_alien_cache(searchp, l3->alien); |
e498be7d | 3497 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 3498 | |
9a2dba4b | 3499 | drain_array_locked(searchp, cpu_cache_get(searchp), 0, |
b28a02de | 3500 | numa_node_id()); |
1da177e4 | 3501 | |
e498be7d | 3502 | if (time_after(l3->next_reap, jiffies)) |
1da177e4 LT |
3503 | goto next_unlock; |
3504 | ||
e498be7d | 3505 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3506 | |
e498be7d CL |
3507 | if (l3->shared) |
3508 | drain_array_locked(searchp, l3->shared, 0, | |
b28a02de | 3509 | numa_node_id()); |
1da177e4 | 3510 | |
e498be7d CL |
3511 | if (l3->free_touched) { |
3512 | l3->free_touched = 0; | |
1da177e4 LT |
3513 | goto next_unlock; |
3514 | } | |
3515 | ||
b28a02de PE |
3516 | tofree = |
3517 | (l3->free_limit + 5 * searchp->num - | |
3518 | 1) / (5 * searchp->num); | |
1da177e4 | 3519 | do { |
e498be7d CL |
3520 | p = l3->slabs_free.next; |
3521 | if (p == &(l3->slabs_free)) | |
1da177e4 LT |
3522 | break; |
3523 | ||
3524 | slabp = list_entry(p, struct slab, list); | |
3525 | BUG_ON(slabp->inuse); | |
3526 | list_del(&slabp->list); | |
3527 | STATS_INC_REAPED(searchp); | |
3528 | ||
3529 | /* Safe to drop the lock. The slab is no longer | |
3530 | * linked to the cache. | |
3531 | * searchp cannot disappear, we hold | |
3532 | * cache_chain_lock | |
3533 | */ | |
e498be7d CL |
3534 | l3->free_objects -= searchp->num; |
3535 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3536 | slab_destroy(searchp, slabp); |
e498be7d | 3537 | spin_lock_irq(&l3->list_lock); |
b28a02de PE |
3538 | } while (--tofree > 0); |
3539 | next_unlock: | |
e498be7d | 3540 | spin_unlock_irq(&l3->list_lock); |
b28a02de | 3541 | next: |
1da177e4 LT |
3542 | cond_resched(); |
3543 | } | |
3544 | check_irq_on(); | |
fc0abb14 | 3545 | mutex_unlock(&cache_chain_mutex); |
4ae7c039 | 3546 | drain_remote_pages(); |
1da177e4 | 3547 | /* Setup the next iteration */ |
cd61ef62 | 3548 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3549 | } |
3550 | ||
3551 | #ifdef CONFIG_PROC_FS | |
3552 | ||
85289f98 | 3553 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3554 | { |
85289f98 PE |
3555 | /* |
3556 | * Output format version, so at least we can change it | |
3557 | * without _too_ many complaints. | |
3558 | */ | |
1da177e4 | 3559 | #if STATS |
85289f98 | 3560 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3561 | #else |
85289f98 | 3562 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3563 | #endif |
85289f98 PE |
3564 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3565 | "<objperslab> <pagesperslab>"); | |
3566 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3567 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3568 | #if STATS |
85289f98 PE |
3569 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
3570 | "<error> <maxfreeable> <nodeallocs> <remotefrees>"); | |
3571 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
1da177e4 | 3572 | #endif |
85289f98 PE |
3573 | seq_putc(m, '\n'); |
3574 | } | |
3575 | ||
3576 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3577 | { | |
3578 | loff_t n = *pos; | |
3579 | struct list_head *p; | |
3580 | ||
fc0abb14 | 3581 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3582 | if (!n) |
3583 | print_slabinfo_header(m); | |
1da177e4 LT |
3584 | p = cache_chain.next; |
3585 | while (n--) { | |
3586 | p = p->next; | |
3587 | if (p == &cache_chain) | |
3588 | return NULL; | |
3589 | } | |
343e0d7a | 3590 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3591 | } |
3592 | ||
3593 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3594 | { | |
343e0d7a | 3595 | struct kmem_cache *cachep = p; |
1da177e4 LT |
3596 | ++*pos; |
3597 | return cachep->next.next == &cache_chain ? NULL | |
343e0d7a | 3598 | : list_entry(cachep->next.next, struct kmem_cache, next); |
1da177e4 LT |
3599 | } |
3600 | ||
3601 | static void s_stop(struct seq_file *m, void *p) | |
3602 | { | |
fc0abb14 | 3603 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3604 | } |
3605 | ||
3606 | static int s_show(struct seq_file *m, void *p) | |
3607 | { | |
343e0d7a | 3608 | struct kmem_cache *cachep = p; |
1da177e4 | 3609 | struct list_head *q; |
b28a02de PE |
3610 | struct slab *slabp; |
3611 | unsigned long active_objs; | |
3612 | unsigned long num_objs; | |
3613 | unsigned long active_slabs = 0; | |
3614 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3615 | const char *name; |
1da177e4 | 3616 | char *error = NULL; |
e498be7d CL |
3617 | int node; |
3618 | struct kmem_list3 *l3; | |
1da177e4 | 3619 | |
ca3b9b91 | 3620 | spin_lock(&cachep->spinlock); |
1da177e4 LT |
3621 | active_objs = 0; |
3622 | num_slabs = 0; | |
e498be7d CL |
3623 | for_each_online_node(node) { |
3624 | l3 = cachep->nodelists[node]; | |
3625 | if (!l3) | |
3626 | continue; | |
3627 | ||
ca3b9b91 RT |
3628 | check_irq_on(); |
3629 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3630 | |
b28a02de | 3631 | list_for_each(q, &l3->slabs_full) { |
e498be7d CL |
3632 | slabp = list_entry(q, struct slab, list); |
3633 | if (slabp->inuse != cachep->num && !error) | |
3634 | error = "slabs_full accounting error"; | |
3635 | active_objs += cachep->num; | |
3636 | active_slabs++; | |
3637 | } | |
b28a02de | 3638 | list_for_each(q, &l3->slabs_partial) { |
e498be7d CL |
3639 | slabp = list_entry(q, struct slab, list); |
3640 | if (slabp->inuse == cachep->num && !error) | |
3641 | error = "slabs_partial inuse accounting error"; | |
3642 | if (!slabp->inuse && !error) | |
3643 | error = "slabs_partial/inuse accounting error"; | |
3644 | active_objs += slabp->inuse; | |
3645 | active_slabs++; | |
3646 | } | |
b28a02de | 3647 | list_for_each(q, &l3->slabs_free) { |
e498be7d CL |
3648 | slabp = list_entry(q, struct slab, list); |
3649 | if (slabp->inuse && !error) | |
3650 | error = "slabs_free/inuse accounting error"; | |
3651 | num_slabs++; | |
3652 | } | |
3653 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3654 | if (l3->shared) |
3655 | shared_avail += l3->shared->avail; | |
e498be7d | 3656 | |
ca3b9b91 | 3657 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3658 | } |
b28a02de PE |
3659 | num_slabs += active_slabs; |
3660 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3661 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3662 | error = "free_objects accounting error"; |
3663 | ||
b28a02de | 3664 | name = cachep->name; |
1da177e4 LT |
3665 | if (error) |
3666 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3667 | ||
3668 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3669 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3670 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3671 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3672 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3673 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3674 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3675 | #if STATS |
b28a02de | 3676 | { /* list3 stats */ |
1da177e4 LT |
3677 | unsigned long high = cachep->high_mark; |
3678 | unsigned long allocs = cachep->num_allocations; | |
3679 | unsigned long grown = cachep->grown; | |
3680 | unsigned long reaped = cachep->reaped; | |
3681 | unsigned long errors = cachep->errors; | |
3682 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3683 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3684 | unsigned long node_frees = cachep->node_frees; |
1da177e4 | 3685 | |
e498be7d | 3686 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
b28a02de | 3687 | %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees); |
1da177e4 LT |
3688 | } |
3689 | /* cpu stats */ | |
3690 | { | |
3691 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3692 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3693 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3694 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3695 | ||
3696 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3697 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3698 | } |
3699 | #endif | |
3700 | seq_putc(m, '\n'); | |
ca3b9b91 | 3701 | spin_unlock(&cachep->spinlock); |
1da177e4 LT |
3702 | return 0; |
3703 | } | |
3704 | ||
3705 | /* | |
3706 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3707 | * | |
3708 | * Output layout: | |
3709 | * cache-name | |
3710 | * num-active-objs | |
3711 | * total-objs | |
3712 | * object size | |
3713 | * num-active-slabs | |
3714 | * total-slabs | |
3715 | * num-pages-per-slab | |
3716 | * + further values on SMP and with statistics enabled | |
3717 | */ | |
3718 | ||
3719 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3720 | .start = s_start, |
3721 | .next = s_next, | |
3722 | .stop = s_stop, | |
3723 | .show = s_show, | |
1da177e4 LT |
3724 | }; |
3725 | ||
3726 | #define MAX_SLABINFO_WRITE 128 | |
3727 | /** | |
3728 | * slabinfo_write - Tuning for the slab allocator | |
3729 | * @file: unused | |
3730 | * @buffer: user buffer | |
3731 | * @count: data length | |
3732 | * @ppos: unused | |
3733 | */ | |
b28a02de PE |
3734 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3735 | size_t count, loff_t *ppos) | |
1da177e4 | 3736 | { |
b28a02de | 3737 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 LT |
3738 | int limit, batchcount, shared, res; |
3739 | struct list_head *p; | |
b28a02de | 3740 | |
1da177e4 LT |
3741 | if (count > MAX_SLABINFO_WRITE) |
3742 | return -EINVAL; | |
3743 | if (copy_from_user(&kbuf, buffer, count)) | |
3744 | return -EFAULT; | |
b28a02de | 3745 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3746 | |
3747 | tmp = strchr(kbuf, ' '); | |
3748 | if (!tmp) | |
3749 | return -EINVAL; | |
3750 | *tmp = '\0'; | |
3751 | tmp++; | |
3752 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3753 | return -EINVAL; | |
3754 | ||
3755 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3756 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3757 | res = -EINVAL; |
b28a02de | 3758 | list_for_each(p, &cache_chain) { |
343e0d7a PE |
3759 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, |
3760 | next); | |
1da177e4 LT |
3761 | |
3762 | if (!strcmp(cachep->name, kbuf)) { | |
3763 | if (limit < 1 || | |
3764 | batchcount < 1 || | |
b28a02de | 3765 | batchcount > limit || shared < 0) { |
e498be7d | 3766 | res = 0; |
1da177e4 | 3767 | } else { |
e498be7d | 3768 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3769 | batchcount, shared); |
1da177e4 LT |
3770 | } |
3771 | break; | |
3772 | } | |
3773 | } | |
fc0abb14 | 3774 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3775 | if (res >= 0) |
3776 | res = count; | |
3777 | return res; | |
3778 | } | |
3779 | #endif | |
3780 | ||
00e145b6 MS |
3781 | /** |
3782 | * ksize - get the actual amount of memory allocated for a given object | |
3783 | * @objp: Pointer to the object | |
3784 | * | |
3785 | * kmalloc may internally round up allocations and return more memory | |
3786 | * than requested. ksize() can be used to determine the actual amount of | |
3787 | * memory allocated. The caller may use this additional memory, even though | |
3788 | * a smaller amount of memory was initially specified with the kmalloc call. | |
3789 | * The caller must guarantee that objp points to a valid object previously | |
3790 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
3791 | * must not be freed during the duration of the call. | |
3792 | */ | |
1da177e4 LT |
3793 | unsigned int ksize(const void *objp) |
3794 | { | |
00e145b6 MS |
3795 | if (unlikely(objp == NULL)) |
3796 | return 0; | |
1da177e4 | 3797 | |
6ed5eb22 | 3798 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 3799 | } |