]>
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 | ||
8fea4e96 PE |
612 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
613 | unsigned int idx) | |
614 | { | |
615 | return slab->s_mem + cache->buffer_size * idx; | |
616 | } | |
617 | ||
618 | static inline unsigned int obj_to_index(struct kmem_cache *cache, | |
619 | struct slab *slab, void *obj) | |
620 | { | |
621 | return (unsigned)(obj - slab->s_mem) / cache->buffer_size; | |
622 | } | |
623 | ||
1da177e4 LT |
624 | /* These are the default caches for kmalloc. Custom caches can have other sizes. */ |
625 | struct cache_sizes malloc_sizes[] = { | |
626 | #define CACHE(x) { .cs_size = (x) }, | |
627 | #include <linux/kmalloc_sizes.h> | |
628 | CACHE(ULONG_MAX) | |
629 | #undef CACHE | |
630 | }; | |
631 | EXPORT_SYMBOL(malloc_sizes); | |
632 | ||
633 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
634 | struct cache_names { | |
635 | char *name; | |
636 | char *name_dma; | |
637 | }; | |
638 | ||
639 | static struct cache_names __initdata cache_names[] = { | |
640 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
641 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 642 | {NULL,} |
1da177e4 LT |
643 | #undef CACHE |
644 | }; | |
645 | ||
646 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 647 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 648 | static struct arraycache_init initarray_generic = |
b28a02de | 649 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
650 | |
651 | /* internal cache of cache description objs */ | |
343e0d7a | 652 | static struct kmem_cache cache_cache = { |
b28a02de PE |
653 | .batchcount = 1, |
654 | .limit = BOOT_CPUCACHE_ENTRIES, | |
655 | .shared = 1, | |
343e0d7a | 656 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de PE |
657 | .flags = SLAB_NO_REAP, |
658 | .spinlock = SPIN_LOCK_UNLOCKED, | |
659 | .name = "kmem_cache", | |
1da177e4 | 660 | #if DEBUG |
343e0d7a | 661 | .obj_size = sizeof(struct kmem_cache), |
1da177e4 LT |
662 | #endif |
663 | }; | |
664 | ||
665 | /* Guard access to the cache-chain. */ | |
fc0abb14 | 666 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
667 | static struct list_head cache_chain; |
668 | ||
669 | /* | |
670 | * vm_enough_memory() looks at this to determine how many | |
671 | * slab-allocated pages are possibly freeable under pressure | |
672 | * | |
673 | * SLAB_RECLAIM_ACCOUNT turns this on per-slab | |
674 | */ | |
675 | atomic_t slab_reclaim_pages; | |
1da177e4 LT |
676 | |
677 | /* | |
678 | * chicken and egg problem: delay the per-cpu array allocation | |
679 | * until the general caches are up. | |
680 | */ | |
681 | static enum { | |
682 | NONE, | |
e498be7d CL |
683 | PARTIAL_AC, |
684 | PARTIAL_L3, | |
1da177e4 LT |
685 | FULL |
686 | } g_cpucache_up; | |
687 | ||
688 | static DEFINE_PER_CPU(struct work_struct, reap_work); | |
689 | ||
343e0d7a PE |
690 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, int node); |
691 | static void enable_cpucache(struct kmem_cache *cachep); | |
b28a02de | 692 | static void cache_reap(void *unused); |
343e0d7a | 693 | static int __node_shrink(struct kmem_cache *cachep, int node); |
1da177e4 | 694 | |
343e0d7a | 695 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
696 | { |
697 | return cachep->array[smp_processor_id()]; | |
698 | } | |
699 | ||
343e0d7a | 700 | static inline struct kmem_cache *__find_general_cachep(size_t size, gfp_t gfpflags) |
1da177e4 LT |
701 | { |
702 | struct cache_sizes *csizep = malloc_sizes; | |
703 | ||
704 | #if DEBUG | |
705 | /* This happens if someone tries to call | |
b28a02de PE |
706 | * kmem_cache_create(), or __kmalloc(), before |
707 | * the generic caches are initialized. | |
708 | */ | |
c7e43c78 | 709 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 LT |
710 | #endif |
711 | while (size > csizep->cs_size) | |
712 | csizep++; | |
713 | ||
714 | /* | |
0abf40c1 | 715 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
716 | * has cs_{dma,}cachep==NULL. Thus no special case |
717 | * for large kmalloc calls required. | |
718 | */ | |
719 | if (unlikely(gfpflags & GFP_DMA)) | |
720 | return csizep->cs_dmacachep; | |
721 | return csizep->cs_cachep; | |
722 | } | |
723 | ||
343e0d7a | 724 | struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
725 | { |
726 | return __find_general_cachep(size, gfpflags); | |
727 | } | |
728 | EXPORT_SYMBOL(kmem_find_general_cachep); | |
729 | ||
fbaccacf | 730 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 731 | { |
fbaccacf SR |
732 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
733 | } | |
1da177e4 | 734 | |
fbaccacf SR |
735 | /* Calculate the number of objects and left-over bytes for a given |
736 | buffer size. */ | |
737 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, | |
738 | size_t align, int flags, size_t *left_over, | |
739 | unsigned int *num) | |
740 | { | |
741 | int nr_objs; | |
742 | size_t mgmt_size; | |
743 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 744 | |
fbaccacf SR |
745 | /* |
746 | * The slab management structure can be either off the slab or | |
747 | * on it. For the latter case, the memory allocated for a | |
748 | * slab is used for: | |
749 | * | |
750 | * - The struct slab | |
751 | * - One kmem_bufctl_t for each object | |
752 | * - Padding to respect alignment of @align | |
753 | * - @buffer_size bytes for each object | |
754 | * | |
755 | * If the slab management structure is off the slab, then the | |
756 | * alignment will already be calculated into the size. Because | |
757 | * the slabs are all pages aligned, the objects will be at the | |
758 | * correct alignment when allocated. | |
759 | */ | |
760 | if (flags & CFLGS_OFF_SLAB) { | |
761 | mgmt_size = 0; | |
762 | nr_objs = slab_size / buffer_size; | |
763 | ||
764 | if (nr_objs > SLAB_LIMIT) | |
765 | nr_objs = SLAB_LIMIT; | |
766 | } else { | |
767 | /* | |
768 | * Ignore padding for the initial guess. The padding | |
769 | * is at most @align-1 bytes, and @buffer_size is at | |
770 | * least @align. In the worst case, this result will | |
771 | * be one greater than the number of objects that fit | |
772 | * into the memory allocation when taking the padding | |
773 | * into account. | |
774 | */ | |
775 | nr_objs = (slab_size - sizeof(struct slab)) / | |
776 | (buffer_size + sizeof(kmem_bufctl_t)); | |
777 | ||
778 | /* | |
779 | * This calculated number will be either the right | |
780 | * amount, or one greater than what we want. | |
781 | */ | |
782 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
783 | > slab_size) | |
784 | nr_objs--; | |
785 | ||
786 | if (nr_objs > SLAB_LIMIT) | |
787 | nr_objs = SLAB_LIMIT; | |
788 | ||
789 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
790 | } | |
791 | *num = nr_objs; | |
792 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
793 | } |
794 | ||
795 | #define slab_error(cachep, msg) __slab_error(__FUNCTION__, cachep, msg) | |
796 | ||
343e0d7a | 797 | static void __slab_error(const char *function, struct kmem_cache *cachep, char *msg) |
1da177e4 LT |
798 | { |
799 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 800 | function, cachep->name, msg); |
1da177e4 LT |
801 | dump_stack(); |
802 | } | |
803 | ||
8fce4d8e CL |
804 | #ifdef CONFIG_NUMA |
805 | /* | |
806 | * Special reaping functions for NUMA systems called from cache_reap(). | |
807 | * These take care of doing round robin flushing of alien caches (containing | |
808 | * objects freed on different nodes from which they were allocated) and the | |
809 | * flushing of remote pcps by calling drain_node_pages. | |
810 | */ | |
811 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
812 | ||
813 | static void init_reap_node(int cpu) | |
814 | { | |
815 | int node; | |
816 | ||
817 | node = next_node(cpu_to_node(cpu), node_online_map); | |
818 | if (node == MAX_NUMNODES) | |
819 | node = 0; | |
820 | ||
821 | __get_cpu_var(reap_node) = node; | |
822 | } | |
823 | ||
824 | static void next_reap_node(void) | |
825 | { | |
826 | int node = __get_cpu_var(reap_node); | |
827 | ||
828 | /* | |
829 | * Also drain per cpu pages on remote zones | |
830 | */ | |
831 | if (node != numa_node_id()) | |
832 | drain_node_pages(node); | |
833 | ||
834 | node = next_node(node, node_online_map); | |
835 | if (unlikely(node >= MAX_NUMNODES)) | |
836 | node = first_node(node_online_map); | |
837 | __get_cpu_var(reap_node) = node; | |
838 | } | |
839 | ||
840 | #else | |
841 | #define init_reap_node(cpu) do { } while (0) | |
842 | #define next_reap_node(void) do { } while (0) | |
843 | #endif | |
844 | ||
1da177e4 LT |
845 | /* |
846 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
847 | * via the workqueue/eventd. | |
848 | * Add the CPU number into the expiration time to minimize the possibility of | |
849 | * the CPUs getting into lockstep and contending for the global cache chain | |
850 | * lock. | |
851 | */ | |
852 | static void __devinit start_cpu_timer(int cpu) | |
853 | { | |
854 | struct work_struct *reap_work = &per_cpu(reap_work, cpu); | |
855 | ||
856 | /* | |
857 | * When this gets called from do_initcalls via cpucache_init(), | |
858 | * init_workqueues() has already run, so keventd will be setup | |
859 | * at that time. | |
860 | */ | |
861 | if (keventd_up() && reap_work->func == NULL) { | |
8fce4d8e | 862 | init_reap_node(cpu); |
1da177e4 LT |
863 | INIT_WORK(reap_work, cache_reap, NULL); |
864 | schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu); | |
865 | } | |
866 | } | |
867 | ||
e498be7d | 868 | static struct array_cache *alloc_arraycache(int node, int entries, |
b28a02de | 869 | int batchcount) |
1da177e4 | 870 | { |
b28a02de | 871 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
872 | struct array_cache *nc = NULL; |
873 | ||
e498be7d | 874 | nc = kmalloc_node(memsize, GFP_KERNEL, node); |
1da177e4 LT |
875 | if (nc) { |
876 | nc->avail = 0; | |
877 | nc->limit = entries; | |
878 | nc->batchcount = batchcount; | |
879 | nc->touched = 0; | |
e498be7d | 880 | spin_lock_init(&nc->lock); |
1da177e4 LT |
881 | } |
882 | return nc; | |
883 | } | |
884 | ||
e498be7d | 885 | #ifdef CONFIG_NUMA |
343e0d7a | 886 | static void *__cache_alloc_node(struct kmem_cache *, gfp_t, int); |
dc85da15 | 887 | |
5295a74c | 888 | static struct array_cache **alloc_alien_cache(int node, int limit) |
e498be7d CL |
889 | { |
890 | struct array_cache **ac_ptr; | |
b28a02de | 891 | int memsize = sizeof(void *) * MAX_NUMNODES; |
e498be7d CL |
892 | int i; |
893 | ||
894 | if (limit > 1) | |
895 | limit = 12; | |
896 | ac_ptr = kmalloc_node(memsize, GFP_KERNEL, node); | |
897 | if (ac_ptr) { | |
898 | for_each_node(i) { | |
899 | if (i == node || !node_online(i)) { | |
900 | ac_ptr[i] = NULL; | |
901 | continue; | |
902 | } | |
903 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d); | |
904 | if (!ac_ptr[i]) { | |
b28a02de | 905 | for (i--; i <= 0; i--) |
e498be7d CL |
906 | kfree(ac_ptr[i]); |
907 | kfree(ac_ptr); | |
908 | return NULL; | |
909 | } | |
910 | } | |
911 | } | |
912 | return ac_ptr; | |
913 | } | |
914 | ||
5295a74c | 915 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
916 | { |
917 | int i; | |
918 | ||
919 | if (!ac_ptr) | |
920 | return; | |
921 | ||
922 | for_each_node(i) | |
b28a02de | 923 | kfree(ac_ptr[i]); |
e498be7d CL |
924 | |
925 | kfree(ac_ptr); | |
926 | } | |
927 | ||
343e0d7a | 928 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 929 | struct array_cache *ac, int node) |
e498be7d CL |
930 | { |
931 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
932 | ||
933 | if (ac->avail) { | |
934 | spin_lock(&rl3->list_lock); | |
ff69416e | 935 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
936 | ac->avail = 0; |
937 | spin_unlock(&rl3->list_lock); | |
938 | } | |
939 | } | |
940 | ||
8fce4d8e CL |
941 | /* |
942 | * Called from cache_reap() to regularly drain alien caches round robin. | |
943 | */ | |
944 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
945 | { | |
946 | int node = __get_cpu_var(reap_node); | |
947 | ||
948 | if (l3->alien) { | |
949 | struct array_cache *ac = l3->alien[node]; | |
950 | if (ac && ac->avail) { | |
951 | spin_lock_irq(&ac->lock); | |
952 | __drain_alien_cache(cachep, ac, node); | |
953 | spin_unlock_irq(&ac->lock); | |
954 | } | |
955 | } | |
956 | } | |
957 | ||
4484ebf1 | 958 | static void drain_alien_cache(struct kmem_cache *cachep, struct array_cache **alien) |
e498be7d | 959 | { |
b28a02de | 960 | int i = 0; |
e498be7d CL |
961 | struct array_cache *ac; |
962 | unsigned long flags; | |
963 | ||
964 | for_each_online_node(i) { | |
4484ebf1 | 965 | ac = alien[i]; |
e498be7d CL |
966 | if (ac) { |
967 | spin_lock_irqsave(&ac->lock, flags); | |
968 | __drain_alien_cache(cachep, ac, i); | |
969 | spin_unlock_irqrestore(&ac->lock, flags); | |
970 | } | |
971 | } | |
972 | } | |
973 | #else | |
7a21ef6f | 974 | |
4484ebf1 | 975 | #define drain_alien_cache(cachep, alien) do { } while (0) |
8fce4d8e | 976 | #define reap_alien(cachep, l3) do { } while (0) |
4484ebf1 | 977 | |
7a21ef6f LT |
978 | static inline struct array_cache **alloc_alien_cache(int node, int limit) |
979 | { | |
980 | return (struct array_cache **) 0x01020304ul; | |
981 | } | |
982 | ||
4484ebf1 RT |
983 | static inline void free_alien_cache(struct array_cache **ac_ptr) |
984 | { | |
985 | } | |
7a21ef6f | 986 | |
e498be7d CL |
987 | #endif |
988 | ||
1da177e4 | 989 | static int __devinit cpuup_callback(struct notifier_block *nfb, |
b28a02de | 990 | unsigned long action, void *hcpu) |
1da177e4 LT |
991 | { |
992 | long cpu = (long)hcpu; | |
343e0d7a | 993 | struct kmem_cache *cachep; |
e498be7d CL |
994 | struct kmem_list3 *l3 = NULL; |
995 | int node = cpu_to_node(cpu); | |
996 | int memsize = sizeof(struct kmem_list3); | |
1da177e4 LT |
997 | |
998 | switch (action) { | |
999 | case CPU_UP_PREPARE: | |
fc0abb14 | 1000 | mutex_lock(&cache_chain_mutex); |
e498be7d CL |
1001 | /* we need to do this right in the beginning since |
1002 | * alloc_arraycache's are going to use this list. | |
1003 | * kmalloc_node allows us to add the slab to the right | |
1004 | * kmem_list3 and not this cpu's kmem_list3 | |
1005 | */ | |
1006 | ||
1da177e4 | 1007 | list_for_each_entry(cachep, &cache_chain, next) { |
e498be7d CL |
1008 | /* setup the size64 kmemlist for cpu before we can |
1009 | * begin anything. Make sure some other cpu on this | |
1010 | * node has not already allocated this | |
1011 | */ | |
1012 | if (!cachep->nodelists[node]) { | |
1013 | if (!(l3 = kmalloc_node(memsize, | |
b28a02de | 1014 | GFP_KERNEL, node))) |
e498be7d CL |
1015 | goto bad; |
1016 | kmem_list3_init(l3); | |
1017 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 1018 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d | 1019 | |
4484ebf1 RT |
1020 | /* |
1021 | * The l3s don't come and go as CPUs come and | |
1022 | * go. cache_chain_mutex is sufficient | |
1023 | * protection here. | |
1024 | */ | |
e498be7d CL |
1025 | cachep->nodelists[node] = l3; |
1026 | } | |
1da177e4 | 1027 | |
e498be7d CL |
1028 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1029 | cachep->nodelists[node]->free_limit = | |
b28a02de PE |
1030 | (1 + nr_cpus_node(node)) * |
1031 | cachep->batchcount + cachep->num; | |
e498be7d CL |
1032 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); |
1033 | } | |
1034 | ||
1035 | /* Now we can go ahead with allocating the shared array's | |
b28a02de | 1036 | & array cache's */ |
e498be7d | 1037 | list_for_each_entry(cachep, &cache_chain, next) { |
cd105df4 | 1038 | struct array_cache *nc; |
4484ebf1 RT |
1039 | struct array_cache *shared; |
1040 | struct array_cache **alien; | |
cd105df4 | 1041 | |
e498be7d | 1042 | nc = alloc_arraycache(node, cachep->limit, |
4484ebf1 | 1043 | cachep->batchcount); |
1da177e4 LT |
1044 | if (!nc) |
1045 | goto bad; | |
4484ebf1 RT |
1046 | shared = alloc_arraycache(node, |
1047 | cachep->shared * cachep->batchcount, | |
1048 | 0xbaadf00d); | |
1049 | if (!shared) | |
1050 | goto bad; | |
7a21ef6f | 1051 | |
4484ebf1 RT |
1052 | alien = alloc_alien_cache(node, cachep->limit); |
1053 | if (!alien) | |
1054 | goto bad; | |
1da177e4 | 1055 | cachep->array[cpu] = nc; |
1da177e4 | 1056 | |
e498be7d CL |
1057 | l3 = cachep->nodelists[node]; |
1058 | BUG_ON(!l3); | |
e498be7d | 1059 | |
4484ebf1 RT |
1060 | spin_lock_irq(&l3->list_lock); |
1061 | if (!l3->shared) { | |
1062 | /* | |
1063 | * We are serialised from CPU_DEAD or | |
1064 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1065 | */ | |
1066 | l3->shared = shared; | |
1067 | shared = NULL; | |
e498be7d | 1068 | } |
4484ebf1 RT |
1069 | #ifdef CONFIG_NUMA |
1070 | if (!l3->alien) { | |
1071 | l3->alien = alien; | |
1072 | alien = NULL; | |
1073 | } | |
1074 | #endif | |
1075 | spin_unlock_irq(&l3->list_lock); | |
1076 | ||
1077 | kfree(shared); | |
1078 | free_alien_cache(alien); | |
1da177e4 | 1079 | } |
fc0abb14 | 1080 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1081 | break; |
1082 | case CPU_ONLINE: | |
1083 | start_cpu_timer(cpu); | |
1084 | break; | |
1085 | #ifdef CONFIG_HOTPLUG_CPU | |
1086 | case CPU_DEAD: | |
4484ebf1 RT |
1087 | /* |
1088 | * Even if all the cpus of a node are down, we don't free the | |
1089 | * kmem_list3 of any cache. This to avoid a race between | |
1090 | * cpu_down, and a kmalloc allocation from another cpu for | |
1091 | * memory from the node of the cpu going down. The list3 | |
1092 | * structure is usually allocated from kmem_cache_create() and | |
1093 | * gets destroyed at kmem_cache_destroy(). | |
1094 | */ | |
1da177e4 LT |
1095 | /* fall thru */ |
1096 | case CPU_UP_CANCELED: | |
fc0abb14 | 1097 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
1098 | |
1099 | list_for_each_entry(cachep, &cache_chain, next) { | |
1100 | struct array_cache *nc; | |
4484ebf1 RT |
1101 | struct array_cache *shared; |
1102 | struct array_cache **alien; | |
e498be7d | 1103 | cpumask_t mask; |
1da177e4 | 1104 | |
e498be7d | 1105 | mask = node_to_cpumask(node); |
1da177e4 LT |
1106 | /* cpu is dead; no one can alloc from it. */ |
1107 | nc = cachep->array[cpu]; | |
1108 | cachep->array[cpu] = NULL; | |
e498be7d CL |
1109 | l3 = cachep->nodelists[node]; |
1110 | ||
1111 | if (!l3) | |
4484ebf1 | 1112 | goto free_array_cache; |
e498be7d | 1113 | |
ca3b9b91 | 1114 | spin_lock_irq(&l3->list_lock); |
e498be7d CL |
1115 | |
1116 | /* Free limit for this kmem_list3 */ | |
1117 | l3->free_limit -= cachep->batchcount; | |
1118 | if (nc) | |
ff69416e | 1119 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
1120 | |
1121 | if (!cpus_empty(mask)) { | |
ca3b9b91 | 1122 | spin_unlock_irq(&l3->list_lock); |
4484ebf1 | 1123 | goto free_array_cache; |
b28a02de | 1124 | } |
e498be7d | 1125 | |
4484ebf1 RT |
1126 | shared = l3->shared; |
1127 | if (shared) { | |
e498be7d | 1128 | free_block(cachep, l3->shared->entry, |
b28a02de | 1129 | l3->shared->avail, node); |
e498be7d CL |
1130 | l3->shared = NULL; |
1131 | } | |
e498be7d | 1132 | |
4484ebf1 RT |
1133 | alien = l3->alien; |
1134 | l3->alien = NULL; | |
1135 | ||
1136 | spin_unlock_irq(&l3->list_lock); | |
1137 | ||
1138 | kfree(shared); | |
1139 | if (alien) { | |
1140 | drain_alien_cache(cachep, alien); | |
1141 | free_alien_cache(alien); | |
e498be7d | 1142 | } |
4484ebf1 | 1143 | free_array_cache: |
1da177e4 LT |
1144 | kfree(nc); |
1145 | } | |
4484ebf1 RT |
1146 | /* |
1147 | * In the previous loop, all the objects were freed to | |
1148 | * the respective cache's slabs, now we can go ahead and | |
1149 | * shrink each nodelist to its limit. | |
1150 | */ | |
1151 | list_for_each_entry(cachep, &cache_chain, next) { | |
1152 | l3 = cachep->nodelists[node]; | |
1153 | if (!l3) | |
1154 | continue; | |
1155 | spin_lock_irq(&l3->list_lock); | |
1156 | /* free slabs belonging to this node */ | |
1157 | __node_shrink(cachep, node); | |
1158 | spin_unlock_irq(&l3->list_lock); | |
1159 | } | |
fc0abb14 | 1160 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1161 | break; |
1162 | #endif | |
1163 | } | |
1164 | return NOTIFY_OK; | |
b28a02de | 1165 | bad: |
fc0abb14 | 1166 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1167 | return NOTIFY_BAD; |
1168 | } | |
1169 | ||
1170 | static struct notifier_block cpucache_notifier = { &cpuup_callback, NULL, 0 }; | |
1171 | ||
e498be7d CL |
1172 | /* |
1173 | * swap the static kmem_list3 with kmalloced memory | |
1174 | */ | |
343e0d7a | 1175 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, int nodeid) |
e498be7d CL |
1176 | { |
1177 | struct kmem_list3 *ptr; | |
1178 | ||
1179 | BUG_ON(cachep->nodelists[nodeid] != list); | |
1180 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_KERNEL, nodeid); | |
1181 | BUG_ON(!ptr); | |
1182 | ||
1183 | local_irq_disable(); | |
1184 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1185 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1186 | cachep->nodelists[nodeid] = ptr; | |
1187 | local_irq_enable(); | |
1188 | } | |
1189 | ||
1da177e4 LT |
1190 | /* Initialisation. |
1191 | * Called after the gfp() functions have been enabled, and before smp_init(). | |
1192 | */ | |
1193 | void __init kmem_cache_init(void) | |
1194 | { | |
1195 | size_t left_over; | |
1196 | struct cache_sizes *sizes; | |
1197 | struct cache_names *names; | |
e498be7d | 1198 | int i; |
07ed76b2 | 1199 | int order; |
e498be7d CL |
1200 | |
1201 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1202 | kmem_list3_init(&initkmem_list3[i]); | |
1203 | if (i < MAX_NUMNODES) | |
1204 | cache_cache.nodelists[i] = NULL; | |
1205 | } | |
1da177e4 LT |
1206 | |
1207 | /* | |
1208 | * Fragmentation resistance on low memory - only use bigger | |
1209 | * page orders on machines with more than 32MB of memory. | |
1210 | */ | |
1211 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1212 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1213 | ||
1da177e4 LT |
1214 | /* Bootstrap is tricky, because several objects are allocated |
1215 | * from caches that do not exist yet: | |
343e0d7a | 1216 | * 1) initialize the cache_cache cache: it contains the struct kmem_cache |
1da177e4 LT |
1217 | * structures of all caches, except cache_cache itself: cache_cache |
1218 | * is statically allocated. | |
e498be7d CL |
1219 | * Initially an __init data area is used for the head array and the |
1220 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1221 | * array at the end of the bootstrap. | |
1da177e4 | 1222 | * 2) Create the first kmalloc cache. |
343e0d7a | 1223 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1224 | * An __init data area is used for the head array. |
1225 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1226 | * head arrays. | |
1da177e4 LT |
1227 | * 4) Replace the __init data head arrays for cache_cache and the first |
1228 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1229 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1230 | * the other cache's with kmalloc allocated memory. | |
1231 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1232 | */ |
1233 | ||
1234 | /* 1) create the cache_cache */ | |
1da177e4 LT |
1235 | INIT_LIST_HEAD(&cache_chain); |
1236 | list_add(&cache_cache.next, &cache_chain); | |
1237 | cache_cache.colour_off = cache_line_size(); | |
1238 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
e498be7d | 1239 | cache_cache.nodelists[numa_node_id()] = &initkmem_list3[CACHE_CACHE]; |
1da177e4 | 1240 | |
3dafccf2 | 1241 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, cache_line_size()); |
1da177e4 | 1242 | |
07ed76b2 JS |
1243 | for (order = 0; order < MAX_ORDER; order++) { |
1244 | cache_estimate(order, cache_cache.buffer_size, | |
1245 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1246 | if (cache_cache.num) | |
1247 | break; | |
1248 | } | |
1da177e4 LT |
1249 | if (!cache_cache.num) |
1250 | BUG(); | |
07ed76b2 | 1251 | cache_cache.gfporder = order; |
b28a02de | 1252 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1253 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1254 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1255 | |
1256 | /* 2+3) create the kmalloc caches */ | |
1257 | sizes = malloc_sizes; | |
1258 | names = cache_names; | |
1259 | ||
e498be7d CL |
1260 | /* Initialize the caches that provide memory for the array cache |
1261 | * and the kmem_list3 structures first. | |
1262 | * Without this, further allocations will bug | |
1263 | */ | |
1264 | ||
1265 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
b28a02de PE |
1266 | sizes[INDEX_AC].cs_size, |
1267 | ARCH_KMALLOC_MINALIGN, | |
1268 | (ARCH_KMALLOC_FLAGS | | |
1269 | SLAB_PANIC), NULL, NULL); | |
e498be7d CL |
1270 | |
1271 | if (INDEX_AC != INDEX_L3) | |
1272 | sizes[INDEX_L3].cs_cachep = | |
b28a02de PE |
1273 | kmem_cache_create(names[INDEX_L3].name, |
1274 | sizes[INDEX_L3].cs_size, | |
1275 | ARCH_KMALLOC_MINALIGN, | |
1276 | (ARCH_KMALLOC_FLAGS | SLAB_PANIC), NULL, | |
1277 | NULL); | |
e498be7d | 1278 | |
1da177e4 | 1279 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1280 | /* |
1281 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1282 | * This should be particularly beneficial on SMP boxes, as it |
1283 | * eliminates "false sharing". | |
1284 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1285 | * allow tighter packing of the smaller caches. |
1286 | */ | |
b28a02de | 1287 | if (!sizes->cs_cachep) |
e498be7d | 1288 | sizes->cs_cachep = kmem_cache_create(names->name, |
b28a02de PE |
1289 | sizes->cs_size, |
1290 | ARCH_KMALLOC_MINALIGN, | |
1291 | (ARCH_KMALLOC_FLAGS | |
1292 | | SLAB_PANIC), | |
1293 | NULL, NULL); | |
1da177e4 LT |
1294 | |
1295 | /* Inc off-slab bufctl limit until the ceiling is hit. */ | |
1296 | if (!(OFF_SLAB(sizes->cs_cachep))) { | |
b28a02de | 1297 | offslab_limit = sizes->cs_size - sizeof(struct slab); |
1da177e4 LT |
1298 | offslab_limit /= sizeof(kmem_bufctl_t); |
1299 | } | |
1300 | ||
1301 | sizes->cs_dmacachep = kmem_cache_create(names->name_dma, | |
b28a02de PE |
1302 | sizes->cs_size, |
1303 | ARCH_KMALLOC_MINALIGN, | |
1304 | (ARCH_KMALLOC_FLAGS | | |
1305 | SLAB_CACHE_DMA | | |
1306 | SLAB_PANIC), NULL, | |
1307 | NULL); | |
1da177e4 LT |
1308 | |
1309 | sizes++; | |
1310 | names++; | |
1311 | } | |
1312 | /* 4) Replace the bootstrap head arrays */ | |
1313 | { | |
b28a02de | 1314 | void *ptr; |
e498be7d | 1315 | |
1da177e4 | 1316 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1317 | |
1da177e4 | 1318 | local_irq_disable(); |
9a2dba4b PE |
1319 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1320 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1321 | sizeof(struct arraycache_init)); |
1da177e4 LT |
1322 | cache_cache.array[smp_processor_id()] = ptr; |
1323 | local_irq_enable(); | |
e498be7d | 1324 | |
1da177e4 | 1325 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d | 1326 | |
1da177e4 | 1327 | local_irq_disable(); |
9a2dba4b | 1328 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1329 | != &initarray_generic.cache); |
9a2dba4b | 1330 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1331 | sizeof(struct arraycache_init)); |
e498be7d | 1332 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1333 | ptr; |
1da177e4 LT |
1334 | local_irq_enable(); |
1335 | } | |
e498be7d CL |
1336 | /* 5) Replace the bootstrap kmem_list3's */ |
1337 | { | |
1338 | int node; | |
1339 | /* Replace the static kmem_list3 structures for the boot cpu */ | |
1340 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE], | |
b28a02de | 1341 | numa_node_id()); |
e498be7d CL |
1342 | |
1343 | for_each_online_node(node) { | |
1344 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
b28a02de | 1345 | &initkmem_list3[SIZE_AC + node], node); |
e498be7d CL |
1346 | |
1347 | if (INDEX_AC != INDEX_L3) { | |
1348 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
b28a02de PE |
1349 | &initkmem_list3[SIZE_L3 + node], |
1350 | node); | |
e498be7d CL |
1351 | } |
1352 | } | |
1353 | } | |
1da177e4 | 1354 | |
e498be7d | 1355 | /* 6) resize the head arrays to their final sizes */ |
1da177e4 | 1356 | { |
343e0d7a | 1357 | struct kmem_cache *cachep; |
fc0abb14 | 1358 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 1359 | list_for_each_entry(cachep, &cache_chain, next) |
b28a02de | 1360 | enable_cpucache(cachep); |
fc0abb14 | 1361 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1362 | } |
1363 | ||
1364 | /* Done! */ | |
1365 | g_cpucache_up = FULL; | |
1366 | ||
1367 | /* Register a cpu startup notifier callback | |
9a2dba4b | 1368 | * that initializes cpu_cache_get for all new cpus |
1da177e4 LT |
1369 | */ |
1370 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 LT |
1371 | |
1372 | /* The reap timers are started later, with a module init call: | |
1373 | * That part of the kernel is not yet operational. | |
1374 | */ | |
1375 | } | |
1376 | ||
1377 | static int __init cpucache_init(void) | |
1378 | { | |
1379 | int cpu; | |
1380 | ||
1381 | /* | |
1382 | * Register the timers that return unneeded | |
1383 | * pages to gfp. | |
1384 | */ | |
e498be7d | 1385 | for_each_online_cpu(cpu) |
b28a02de | 1386 | start_cpu_timer(cpu); |
1da177e4 LT |
1387 | |
1388 | return 0; | |
1389 | } | |
1390 | ||
1391 | __initcall(cpucache_init); | |
1392 | ||
1393 | /* | |
1394 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1395 | * | |
1396 | * If we requested dmaable memory, we will get it. Even if we | |
1397 | * did not request dmaable memory, we might get it, but that | |
1398 | * would be relatively rare and ignorable. | |
1399 | */ | |
343e0d7a | 1400 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1401 | { |
1402 | struct page *page; | |
1403 | void *addr; | |
1404 | int i; | |
1405 | ||
1406 | flags |= cachep->gfpflags; | |
50c85a19 | 1407 | page = alloc_pages_node(nodeid, flags, cachep->gfporder); |
1da177e4 LT |
1408 | if (!page) |
1409 | return NULL; | |
1410 | addr = page_address(page); | |
1411 | ||
1412 | i = (1 << cachep->gfporder); | |
1413 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1414 | atomic_add(i, &slab_reclaim_pages); | |
1415 | add_page_state(nr_slab, i); | |
1416 | while (i--) { | |
f205b2fe | 1417 | __SetPageSlab(page); |
1da177e4 LT |
1418 | page++; |
1419 | } | |
1420 | return addr; | |
1421 | } | |
1422 | ||
1423 | /* | |
1424 | * Interface to system's page release. | |
1425 | */ | |
343e0d7a | 1426 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1427 | { |
b28a02de | 1428 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1429 | struct page *page = virt_to_page(addr); |
1430 | const unsigned long nr_freed = i; | |
1431 | ||
1432 | while (i--) { | |
f205b2fe NP |
1433 | BUG_ON(!PageSlab(page)); |
1434 | __ClearPageSlab(page); | |
1da177e4 LT |
1435 | page++; |
1436 | } | |
1437 | sub_page_state(nr_slab, nr_freed); | |
1438 | if (current->reclaim_state) | |
1439 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1440 | free_pages((unsigned long)addr, cachep->gfporder); | |
b28a02de PE |
1441 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1442 | atomic_sub(1 << cachep->gfporder, &slab_reclaim_pages); | |
1da177e4 LT |
1443 | } |
1444 | ||
1445 | static void kmem_rcu_free(struct rcu_head *head) | |
1446 | { | |
b28a02de | 1447 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1448 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1449 | |
1450 | kmem_freepages(cachep, slab_rcu->addr); | |
1451 | if (OFF_SLAB(cachep)) | |
1452 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1453 | } | |
1454 | ||
1455 | #if DEBUG | |
1456 | ||
1457 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1458 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1459 | unsigned long caller) |
1da177e4 | 1460 | { |
3dafccf2 | 1461 | int size = obj_size(cachep); |
1da177e4 | 1462 | |
3dafccf2 | 1463 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1464 | |
b28a02de | 1465 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1466 | return; |
1467 | ||
b28a02de PE |
1468 | *addr++ = 0x12345678; |
1469 | *addr++ = caller; | |
1470 | *addr++ = smp_processor_id(); | |
1471 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1472 | { |
1473 | unsigned long *sptr = &caller; | |
1474 | unsigned long svalue; | |
1475 | ||
1476 | while (!kstack_end(sptr)) { | |
1477 | svalue = *sptr++; | |
1478 | if (kernel_text_address(svalue)) { | |
b28a02de | 1479 | *addr++ = svalue; |
1da177e4 LT |
1480 | size -= sizeof(unsigned long); |
1481 | if (size <= sizeof(unsigned long)) | |
1482 | break; | |
1483 | } | |
1484 | } | |
1485 | ||
1486 | } | |
b28a02de | 1487 | *addr++ = 0x87654321; |
1da177e4 LT |
1488 | } |
1489 | #endif | |
1490 | ||
343e0d7a | 1491 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1492 | { |
3dafccf2 MS |
1493 | int size = obj_size(cachep); |
1494 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1495 | |
1496 | memset(addr, val, size); | |
b28a02de | 1497 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1498 | } |
1499 | ||
1500 | static void dump_line(char *data, int offset, int limit) | |
1501 | { | |
1502 | int i; | |
1503 | printk(KERN_ERR "%03x:", offset); | |
b28a02de PE |
1504 | for (i = 0; i < limit; i++) { |
1505 | printk(" %02x", (unsigned char)data[offset + i]); | |
1da177e4 LT |
1506 | } |
1507 | printk("\n"); | |
1508 | } | |
1509 | #endif | |
1510 | ||
1511 | #if DEBUG | |
1512 | ||
343e0d7a | 1513 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1514 | { |
1515 | int i, size; | |
1516 | char *realobj; | |
1517 | ||
1518 | if (cachep->flags & SLAB_RED_ZONE) { | |
1519 | printk(KERN_ERR "Redzone: 0x%lx/0x%lx.\n", | |
b28a02de PE |
1520 | *dbg_redzone1(cachep, objp), |
1521 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1522 | } |
1523 | ||
1524 | if (cachep->flags & SLAB_STORE_USER) { | |
1525 | printk(KERN_ERR "Last user: [<%p>]", | |
b28a02de | 1526 | *dbg_userword(cachep, objp)); |
1da177e4 | 1527 | print_symbol("(%s)", |
b28a02de | 1528 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1529 | printk("\n"); |
1530 | } | |
3dafccf2 MS |
1531 | realobj = (char *)objp + obj_offset(cachep); |
1532 | size = obj_size(cachep); | |
b28a02de | 1533 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1534 | int limit; |
1535 | limit = 16; | |
b28a02de PE |
1536 | if (i + limit > size) |
1537 | limit = size - i; | |
1da177e4 LT |
1538 | dump_line(realobj, i, limit); |
1539 | } | |
1540 | } | |
1541 | ||
343e0d7a | 1542 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1543 | { |
1544 | char *realobj; | |
1545 | int size, i; | |
1546 | int lines = 0; | |
1547 | ||
3dafccf2 MS |
1548 | realobj = (char *)objp + obj_offset(cachep); |
1549 | size = obj_size(cachep); | |
1da177e4 | 1550 | |
b28a02de | 1551 | for (i = 0; i < size; i++) { |
1da177e4 | 1552 | char exp = POISON_FREE; |
b28a02de | 1553 | if (i == size - 1) |
1da177e4 LT |
1554 | exp = POISON_END; |
1555 | if (realobj[i] != exp) { | |
1556 | int limit; | |
1557 | /* Mismatch ! */ | |
1558 | /* Print header */ | |
1559 | if (lines == 0) { | |
b28a02de PE |
1560 | printk(KERN_ERR |
1561 | "Slab corruption: start=%p, len=%d\n", | |
1562 | realobj, size); | |
1da177e4 LT |
1563 | print_objinfo(cachep, objp, 0); |
1564 | } | |
1565 | /* Hexdump the affected line */ | |
b28a02de | 1566 | i = (i / 16) * 16; |
1da177e4 | 1567 | limit = 16; |
b28a02de PE |
1568 | if (i + limit > size) |
1569 | limit = size - i; | |
1da177e4 LT |
1570 | dump_line(realobj, i, limit); |
1571 | i += 16; | |
1572 | lines++; | |
1573 | /* Limit to 5 lines */ | |
1574 | if (lines > 5) | |
1575 | break; | |
1576 | } | |
1577 | } | |
1578 | if (lines != 0) { | |
1579 | /* Print some data about the neighboring objects, if they | |
1580 | * exist: | |
1581 | */ | |
6ed5eb22 | 1582 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1583 | unsigned int objnr; |
1da177e4 | 1584 | |
8fea4e96 | 1585 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1586 | if (objnr) { |
8fea4e96 | 1587 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1588 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1589 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1590 | realobj, size); |
1da177e4 LT |
1591 | print_objinfo(cachep, objp, 2); |
1592 | } | |
b28a02de | 1593 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1594 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1595 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1596 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1597 | realobj, size); |
1da177e4 LT |
1598 | print_objinfo(cachep, objp, 2); |
1599 | } | |
1600 | } | |
1601 | } | |
1602 | #endif | |
1603 | ||
12dd36fa MD |
1604 | #if DEBUG |
1605 | /** | |
1606 | * slab_destroy_objs - call the registered destructor for each object in | |
1607 | * a slab that is to be destroyed. | |
1da177e4 | 1608 | */ |
343e0d7a | 1609 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1610 | { |
1da177e4 LT |
1611 | int i; |
1612 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1613 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1614 | |
1615 | if (cachep->flags & SLAB_POISON) { | |
1616 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 1617 | if ((cachep->buffer_size % PAGE_SIZE) == 0 |
b28a02de PE |
1618 | && OFF_SLAB(cachep)) |
1619 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 1620 | cachep->buffer_size / PAGE_SIZE, |
b28a02de | 1621 | 1); |
1da177e4 LT |
1622 | else |
1623 | check_poison_obj(cachep, objp); | |
1624 | #else | |
1625 | check_poison_obj(cachep, objp); | |
1626 | #endif | |
1627 | } | |
1628 | if (cachep->flags & SLAB_RED_ZONE) { | |
1629 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1630 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1631 | "was overwritten"); |
1da177e4 LT |
1632 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1633 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1634 | "was overwritten"); |
1da177e4 LT |
1635 | } |
1636 | if (cachep->dtor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 1637 | (cachep->dtor) (objp + obj_offset(cachep), cachep, 0); |
1da177e4 | 1638 | } |
12dd36fa | 1639 | } |
1da177e4 | 1640 | #else |
343e0d7a | 1641 | static void slab_destroy_objs(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1642 | { |
1da177e4 LT |
1643 | if (cachep->dtor) { |
1644 | int i; | |
1645 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1646 | void *objp = index_to_obj(cachep, slabp, i); |
b28a02de | 1647 | (cachep->dtor) (objp, cachep, 0); |
1da177e4 LT |
1648 | } |
1649 | } | |
12dd36fa | 1650 | } |
1da177e4 LT |
1651 | #endif |
1652 | ||
12dd36fa MD |
1653 | /** |
1654 | * Destroy all the objs in a slab, and release the mem back to the system. | |
1655 | * Before calling the slab must have been unlinked from the cache. | |
1656 | * The cache-lock is not held/needed. | |
1657 | */ | |
343e0d7a | 1658 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1659 | { |
1660 | void *addr = slabp->s_mem - slabp->colouroff; | |
1661 | ||
1662 | slab_destroy_objs(cachep, slabp); | |
1da177e4 LT |
1663 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1664 | struct slab_rcu *slab_rcu; | |
1665 | ||
b28a02de | 1666 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1667 | slab_rcu->cachep = cachep; |
1668 | slab_rcu->addr = addr; | |
1669 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1670 | } else { | |
1671 | kmem_freepages(cachep, addr); | |
1672 | if (OFF_SLAB(cachep)) | |
1673 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1674 | } | |
1675 | } | |
1676 | ||
3dafccf2 | 1677 | /* For setting up all the kmem_list3s for cache whose buffer_size is same |
e498be7d | 1678 | as size of kmem_list3. */ |
343e0d7a | 1679 | static void set_up_list3s(struct kmem_cache *cachep, int index) |
e498be7d CL |
1680 | { |
1681 | int node; | |
1682 | ||
1683 | for_each_online_node(node) { | |
b28a02de | 1684 | cachep->nodelists[node] = &initkmem_list3[index + node]; |
e498be7d | 1685 | cachep->nodelists[node]->next_reap = jiffies + |
b28a02de PE |
1686 | REAPTIMEOUT_LIST3 + |
1687 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d CL |
1688 | } |
1689 | } | |
1690 | ||
4d268eba | 1691 | /** |
a70773dd RD |
1692 | * calculate_slab_order - calculate size (page order) of slabs |
1693 | * @cachep: pointer to the cache that is being created | |
1694 | * @size: size of objects to be created in this cache. | |
1695 | * @align: required alignment for the objects. | |
1696 | * @flags: slab allocation flags | |
1697 | * | |
1698 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1699 | * |
1700 | * This could be made much more intelligent. For now, try to avoid using | |
1701 | * high order pages for slabs. When the gfp() functions are more friendly | |
1702 | * towards high-order requests, this should be changed. | |
1703 | */ | |
ee13d785 RD |
1704 | static inline size_t calculate_slab_order(struct kmem_cache *cachep, |
1705 | size_t size, size_t align, unsigned long flags) | |
4d268eba PE |
1706 | { |
1707 | size_t left_over = 0; | |
9888e6fa | 1708 | int gfporder; |
4d268eba | 1709 | |
9888e6fa | 1710 | for (gfporder = 0 ; gfporder <= MAX_GFP_ORDER; gfporder++) { |
4d268eba PE |
1711 | unsigned int num; |
1712 | size_t remainder; | |
1713 | ||
9888e6fa | 1714 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1715 | if (!num) |
1716 | continue; | |
9888e6fa | 1717 | |
4d268eba | 1718 | /* More than offslab_limit objects will cause problems */ |
9888e6fa | 1719 | if ((flags & CFLGS_OFF_SLAB) && num > offslab_limit) |
4d268eba PE |
1720 | break; |
1721 | ||
9888e6fa | 1722 | /* Found something acceptable - save it away */ |
4d268eba | 1723 | cachep->num = num; |
9888e6fa | 1724 | cachep->gfporder = gfporder; |
4d268eba PE |
1725 | left_over = remainder; |
1726 | ||
f78bb8ad LT |
1727 | /* |
1728 | * A VFS-reclaimable slab tends to have most allocations | |
1729 | * as GFP_NOFS and we really don't want to have to be allocating | |
1730 | * higher-order pages when we are unable to shrink dcache. | |
1731 | */ | |
1732 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1733 | break; | |
1734 | ||
4d268eba PE |
1735 | /* |
1736 | * Large number of objects is good, but very large slabs are | |
1737 | * currently bad for the gfp()s. | |
1738 | */ | |
9888e6fa | 1739 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
1740 | break; |
1741 | ||
9888e6fa LT |
1742 | /* |
1743 | * Acceptable internal fragmentation? | |
1744 | */ | |
1745 | if ((left_over * 8) <= (PAGE_SIZE << gfporder)) | |
4d268eba PE |
1746 | break; |
1747 | } | |
1748 | return left_over; | |
1749 | } | |
1750 | ||
1da177e4 LT |
1751 | /** |
1752 | * kmem_cache_create - Create a cache. | |
1753 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
1754 | * @size: The size of objects to be created in this cache. | |
1755 | * @align: The required alignment for the objects. | |
1756 | * @flags: SLAB flags | |
1757 | * @ctor: A constructor for the objects. | |
1758 | * @dtor: A destructor for the objects. | |
1759 | * | |
1760 | * Returns a ptr to the cache on success, NULL on failure. | |
1761 | * Cannot be called within a int, but can be interrupted. | |
1762 | * The @ctor is run when new pages are allocated by the cache | |
1763 | * and the @dtor is run before the pages are handed back. | |
1764 | * | |
1765 | * @name must be valid until the cache is destroyed. This implies that | |
1766 | * the module calling this has to destroy the cache before getting | |
1767 | * unloaded. | |
1768 | * | |
1769 | * The flags are | |
1770 | * | |
1771 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
1772 | * to catch references to uninitialised memory. | |
1773 | * | |
1774 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
1775 | * for buffer overruns. | |
1776 | * | |
1777 | * %SLAB_NO_REAP - Don't automatically reap this cache when we're under | |
1778 | * memory pressure. | |
1779 | * | |
1780 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
1781 | * cacheline. This can be beneficial if you're counting cycles as closely | |
1782 | * as davem. | |
1783 | */ | |
343e0d7a | 1784 | struct kmem_cache * |
1da177e4 | 1785 | kmem_cache_create (const char *name, size_t size, size_t align, |
343e0d7a PE |
1786 | unsigned long flags, void (*ctor)(void*, struct kmem_cache *, unsigned long), |
1787 | void (*dtor)(void*, struct kmem_cache *, unsigned long)) | |
1da177e4 LT |
1788 | { |
1789 | size_t left_over, slab_size, ralign; | |
343e0d7a | 1790 | struct kmem_cache *cachep = NULL; |
4f12bb4f | 1791 | struct list_head *p; |
1da177e4 LT |
1792 | |
1793 | /* | |
1794 | * Sanity checks... these are all serious usage bugs. | |
1795 | */ | |
1796 | if ((!name) || | |
b28a02de PE |
1797 | in_interrupt() || |
1798 | (size < BYTES_PER_WORD) || | |
1799 | (size > (1 << MAX_OBJ_ORDER) * PAGE_SIZE) || (dtor && !ctor)) { | |
1800 | printk(KERN_ERR "%s: Early error in slab %s\n", | |
1801 | __FUNCTION__, name); | |
1802 | BUG(); | |
1803 | } | |
1da177e4 | 1804 | |
f0188f47 RT |
1805 | /* |
1806 | * Prevent CPUs from coming and going. | |
1807 | * lock_cpu_hotplug() nests outside cache_chain_mutex | |
1808 | */ | |
1809 | lock_cpu_hotplug(); | |
1810 | ||
fc0abb14 | 1811 | mutex_lock(&cache_chain_mutex); |
4f12bb4f AM |
1812 | |
1813 | list_for_each(p, &cache_chain) { | |
343e0d7a | 1814 | struct kmem_cache *pc = list_entry(p, struct kmem_cache, next); |
4f12bb4f AM |
1815 | mm_segment_t old_fs = get_fs(); |
1816 | char tmp; | |
1817 | int res; | |
1818 | ||
1819 | /* | |
1820 | * This happens when the module gets unloaded and doesn't | |
1821 | * destroy its slab cache and no-one else reuses the vmalloc | |
1822 | * area of the module. Print a warning. | |
1823 | */ | |
1824 | set_fs(KERNEL_DS); | |
1825 | res = __get_user(tmp, pc->name); | |
1826 | set_fs(old_fs); | |
1827 | if (res) { | |
1828 | printk("SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 1829 | pc->buffer_size); |
4f12bb4f AM |
1830 | continue; |
1831 | } | |
1832 | ||
b28a02de | 1833 | if (!strcmp(pc->name, name)) { |
4f12bb4f AM |
1834 | printk("kmem_cache_create: duplicate cache %s\n", name); |
1835 | dump_stack(); | |
1836 | goto oops; | |
1837 | } | |
1838 | } | |
1839 | ||
1da177e4 LT |
1840 | #if DEBUG |
1841 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1842 | if ((flags & SLAB_DEBUG_INITIAL) && !ctor) { | |
1843 | /* No constructor, but inital state check requested */ | |
1844 | printk(KERN_ERR "%s: No con, but init state check " | |
b28a02de | 1845 | "requested - %s\n", __FUNCTION__, name); |
1da177e4 LT |
1846 | flags &= ~SLAB_DEBUG_INITIAL; |
1847 | } | |
1da177e4 LT |
1848 | #if FORCED_DEBUG |
1849 | /* | |
1850 | * Enable redzoning and last user accounting, except for caches with | |
1851 | * large objects, if the increased size would increase the object size | |
1852 | * above the next power of two: caches with object sizes just above a | |
1853 | * power of two have a significant amount of internal fragmentation. | |
1854 | */ | |
b28a02de PE |
1855 | if ((size < 4096 |
1856 | || fls(size - 1) == fls(size - 1 + 3 * BYTES_PER_WORD))) | |
1857 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; | |
1da177e4 LT |
1858 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
1859 | flags |= SLAB_POISON; | |
1860 | #endif | |
1861 | if (flags & SLAB_DESTROY_BY_RCU) | |
1862 | BUG_ON(flags & SLAB_POISON); | |
1863 | #endif | |
1864 | if (flags & SLAB_DESTROY_BY_RCU) | |
1865 | BUG_ON(dtor); | |
1866 | ||
1867 | /* | |
1868 | * Always checks flags, a caller might be expecting debug | |
1869 | * support which isn't available. | |
1870 | */ | |
1871 | if (flags & ~CREATE_MASK) | |
1872 | BUG(); | |
1873 | ||
1874 | /* Check that size is in terms of words. This is needed to avoid | |
1875 | * unaligned accesses for some archs when redzoning is used, and makes | |
1876 | * sure any on-slab bufctl's are also correctly aligned. | |
1877 | */ | |
b28a02de PE |
1878 | if (size & (BYTES_PER_WORD - 1)) { |
1879 | size += (BYTES_PER_WORD - 1); | |
1880 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
1881 | } |
1882 | ||
1883 | /* calculate out the final buffer alignment: */ | |
1884 | /* 1) arch recommendation: can be overridden for debug */ | |
1885 | if (flags & SLAB_HWCACHE_ALIGN) { | |
1886 | /* Default alignment: as specified by the arch code. | |
1887 | * Except if an object is really small, then squeeze multiple | |
1888 | * objects into one cacheline. | |
1889 | */ | |
1890 | ralign = cache_line_size(); | |
b28a02de | 1891 | while (size <= ralign / 2) |
1da177e4 LT |
1892 | ralign /= 2; |
1893 | } else { | |
1894 | ralign = BYTES_PER_WORD; | |
1895 | } | |
1896 | /* 2) arch mandated alignment: disables debug if necessary */ | |
1897 | if (ralign < ARCH_SLAB_MINALIGN) { | |
1898 | ralign = ARCH_SLAB_MINALIGN; | |
1899 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1900 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1901 | } |
1902 | /* 3) caller mandated alignment: disables debug if necessary */ | |
1903 | if (ralign < align) { | |
1904 | ralign = align; | |
1905 | if (ralign > BYTES_PER_WORD) | |
b28a02de | 1906 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
1da177e4 LT |
1907 | } |
1908 | /* 4) Store it. Note that the debug code below can reduce | |
1909 | * the alignment to BYTES_PER_WORD. | |
1910 | */ | |
1911 | align = ralign; | |
1912 | ||
1913 | /* Get cache's description obj. */ | |
343e0d7a | 1914 | cachep = kmem_cache_alloc(&cache_cache, SLAB_KERNEL); |
1da177e4 | 1915 | if (!cachep) |
4f12bb4f | 1916 | goto oops; |
343e0d7a | 1917 | memset(cachep, 0, sizeof(struct kmem_cache)); |
1da177e4 LT |
1918 | |
1919 | #if DEBUG | |
3dafccf2 | 1920 | cachep->obj_size = size; |
1da177e4 LT |
1921 | |
1922 | if (flags & SLAB_RED_ZONE) { | |
1923 | /* redzoning only works with word aligned caches */ | |
1924 | align = BYTES_PER_WORD; | |
1925 | ||
1926 | /* add space for red zone words */ | |
3dafccf2 | 1927 | cachep->obj_offset += BYTES_PER_WORD; |
b28a02de | 1928 | size += 2 * BYTES_PER_WORD; |
1da177e4 LT |
1929 | } |
1930 | if (flags & SLAB_STORE_USER) { | |
1931 | /* user store requires word alignment and | |
1932 | * one word storage behind the end of the real | |
1933 | * object. | |
1934 | */ | |
1935 | align = BYTES_PER_WORD; | |
1936 | size += BYTES_PER_WORD; | |
1937 | } | |
1938 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 1939 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
1940 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
1941 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
1942 | size = PAGE_SIZE; |
1943 | } | |
1944 | #endif | |
1945 | #endif | |
1946 | ||
1947 | /* Determine if the slab management is 'on' or 'off' slab. */ | |
b28a02de | 1948 | if (size >= (PAGE_SIZE >> 3)) |
1da177e4 LT |
1949 | /* |
1950 | * Size is large, assume best to place the slab management obj | |
1951 | * off-slab (should allow better packing of objs). | |
1952 | */ | |
1953 | flags |= CFLGS_OFF_SLAB; | |
1954 | ||
1955 | size = ALIGN(size, align); | |
1956 | ||
f78bb8ad | 1957 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
1958 | |
1959 | if (!cachep->num) { | |
1960 | printk("kmem_cache_create: couldn't create cache %s.\n", name); | |
1961 | kmem_cache_free(&cache_cache, cachep); | |
1962 | cachep = NULL; | |
4f12bb4f | 1963 | goto oops; |
1da177e4 | 1964 | } |
b28a02de PE |
1965 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
1966 | + sizeof(struct slab), align); | |
1da177e4 LT |
1967 | |
1968 | /* | |
1969 | * If the slab has been placed off-slab, and we have enough space then | |
1970 | * move it on-slab. This is at the expense of any extra colouring. | |
1971 | */ | |
1972 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
1973 | flags &= ~CFLGS_OFF_SLAB; | |
1974 | left_over -= slab_size; | |
1975 | } | |
1976 | ||
1977 | if (flags & CFLGS_OFF_SLAB) { | |
1978 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
1979 | slab_size = |
1980 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
1981 | } |
1982 | ||
1983 | cachep->colour_off = cache_line_size(); | |
1984 | /* Offset must be a multiple of the alignment. */ | |
1985 | if (cachep->colour_off < align) | |
1986 | cachep->colour_off = align; | |
b28a02de | 1987 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
1988 | cachep->slab_size = slab_size; |
1989 | cachep->flags = flags; | |
1990 | cachep->gfpflags = 0; | |
1991 | if (flags & SLAB_CACHE_DMA) | |
1992 | cachep->gfpflags |= GFP_DMA; | |
1993 | spin_lock_init(&cachep->spinlock); | |
3dafccf2 | 1994 | cachep->buffer_size = size; |
1da177e4 LT |
1995 | |
1996 | if (flags & CFLGS_OFF_SLAB) | |
b2d55073 | 1997 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
1da177e4 LT |
1998 | cachep->ctor = ctor; |
1999 | cachep->dtor = dtor; | |
2000 | cachep->name = name; | |
2001 | ||
1da177e4 LT |
2002 | |
2003 | if (g_cpucache_up == FULL) { | |
2004 | enable_cpucache(cachep); | |
2005 | } else { | |
2006 | if (g_cpucache_up == NONE) { | |
2007 | /* Note: the first kmem_cache_create must create | |
2008 | * the cache that's used by kmalloc(24), otherwise | |
2009 | * the creation of further caches will BUG(). | |
2010 | */ | |
e498be7d | 2011 | cachep->array[smp_processor_id()] = |
b28a02de | 2012 | &initarray_generic.cache; |
e498be7d CL |
2013 | |
2014 | /* If the cache that's used by | |
2015 | * kmalloc(sizeof(kmem_list3)) is the first cache, | |
2016 | * then we need to set up all its list3s, otherwise | |
2017 | * the creation of further caches will BUG(). | |
2018 | */ | |
2019 | set_up_list3s(cachep, SIZE_AC); | |
2020 | if (INDEX_AC == INDEX_L3) | |
2021 | g_cpucache_up = PARTIAL_L3; | |
2022 | else | |
2023 | g_cpucache_up = PARTIAL_AC; | |
1da177e4 | 2024 | } else { |
e498be7d | 2025 | cachep->array[smp_processor_id()] = |
b28a02de | 2026 | kmalloc(sizeof(struct arraycache_init), GFP_KERNEL); |
e498be7d CL |
2027 | |
2028 | if (g_cpucache_up == PARTIAL_AC) { | |
2029 | set_up_list3s(cachep, SIZE_L3); | |
2030 | g_cpucache_up = PARTIAL_L3; | |
2031 | } else { | |
2032 | int node; | |
2033 | for_each_online_node(node) { | |
2034 | ||
2035 | cachep->nodelists[node] = | |
b28a02de PE |
2036 | kmalloc_node(sizeof |
2037 | (struct kmem_list3), | |
2038 | GFP_KERNEL, node); | |
e498be7d | 2039 | BUG_ON(!cachep->nodelists[node]); |
b28a02de PE |
2040 | kmem_list3_init(cachep-> |
2041 | nodelists[node]); | |
e498be7d CL |
2042 | } |
2043 | } | |
1da177e4 | 2044 | } |
e498be7d | 2045 | cachep->nodelists[numa_node_id()]->next_reap = |
b28a02de PE |
2046 | jiffies + REAPTIMEOUT_LIST3 + |
2047 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 2048 | |
9a2dba4b PE |
2049 | BUG_ON(!cpu_cache_get(cachep)); |
2050 | cpu_cache_get(cachep)->avail = 0; | |
2051 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2052 | cpu_cache_get(cachep)->batchcount = 1; | |
2053 | cpu_cache_get(cachep)->touched = 0; | |
1da177e4 LT |
2054 | cachep->batchcount = 1; |
2055 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
b28a02de | 2056 | } |
1da177e4 | 2057 | |
1da177e4 LT |
2058 | /* cache setup completed, link it into the list */ |
2059 | list_add(&cachep->next, &cache_chain); | |
b28a02de | 2060 | oops: |
1da177e4 LT |
2061 | if (!cachep && (flags & SLAB_PANIC)) |
2062 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2063 | name); |
fc0abb14 | 2064 | mutex_unlock(&cache_chain_mutex); |
f0188f47 | 2065 | unlock_cpu_hotplug(); |
1da177e4 LT |
2066 | return cachep; |
2067 | } | |
2068 | EXPORT_SYMBOL(kmem_cache_create); | |
2069 | ||
2070 | #if DEBUG | |
2071 | static void check_irq_off(void) | |
2072 | { | |
2073 | BUG_ON(!irqs_disabled()); | |
2074 | } | |
2075 | ||
2076 | static void check_irq_on(void) | |
2077 | { | |
2078 | BUG_ON(irqs_disabled()); | |
2079 | } | |
2080 | ||
343e0d7a | 2081 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2082 | { |
2083 | #ifdef CONFIG_SMP | |
2084 | check_irq_off(); | |
e498be7d | 2085 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2086 | #endif |
2087 | } | |
e498be7d | 2088 | |
343e0d7a | 2089 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2090 | { |
2091 | #ifdef CONFIG_SMP | |
2092 | check_irq_off(); | |
2093 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2094 | #endif | |
2095 | } | |
2096 | ||
1da177e4 LT |
2097 | #else |
2098 | #define check_irq_off() do { } while(0) | |
2099 | #define check_irq_on() do { } while(0) | |
2100 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2101 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2102 | #endif |
2103 | ||
2104 | /* | |
2105 | * Waits for all CPUs to execute func(). | |
2106 | */ | |
b28a02de | 2107 | static void smp_call_function_all_cpus(void (*func)(void *arg), void *arg) |
1da177e4 LT |
2108 | { |
2109 | check_irq_on(); | |
2110 | preempt_disable(); | |
2111 | ||
2112 | local_irq_disable(); | |
2113 | func(arg); | |
2114 | local_irq_enable(); | |
2115 | ||
2116 | if (smp_call_function(func, arg, 1, 1)) | |
2117 | BUG(); | |
2118 | ||
2119 | preempt_enable(); | |
2120 | } | |
2121 | ||
343e0d7a | 2122 | static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, |
b28a02de | 2123 | int force, int node); |
1da177e4 LT |
2124 | |
2125 | static void do_drain(void *arg) | |
2126 | { | |
343e0d7a | 2127 | struct kmem_cache *cachep = (struct kmem_cache *) arg; |
1da177e4 | 2128 | struct array_cache *ac; |
ff69416e | 2129 | int node = numa_node_id(); |
1da177e4 LT |
2130 | |
2131 | check_irq_off(); | |
9a2dba4b | 2132 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2133 | spin_lock(&cachep->nodelists[node]->list_lock); |
2134 | free_block(cachep, ac->entry, ac->avail, node); | |
2135 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2136 | ac->avail = 0; |
2137 | } | |
2138 | ||
343e0d7a | 2139 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2140 | { |
e498be7d CL |
2141 | struct kmem_list3 *l3; |
2142 | int node; | |
2143 | ||
1da177e4 LT |
2144 | smp_call_function_all_cpus(do_drain, cachep); |
2145 | check_irq_on(); | |
b28a02de | 2146 | for_each_online_node(node) { |
e498be7d CL |
2147 | l3 = cachep->nodelists[node]; |
2148 | if (l3) { | |
ca3b9b91 | 2149 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2150 | drain_array_locked(cachep, l3->shared, 1, node); |
ca3b9b91 | 2151 | spin_unlock_irq(&l3->list_lock); |
e498be7d | 2152 | if (l3->alien) |
4484ebf1 | 2153 | drain_alien_cache(cachep, l3->alien); |
e498be7d CL |
2154 | } |
2155 | } | |
1da177e4 LT |
2156 | } |
2157 | ||
343e0d7a | 2158 | static int __node_shrink(struct kmem_cache *cachep, int node) |
1da177e4 LT |
2159 | { |
2160 | struct slab *slabp; | |
e498be7d | 2161 | struct kmem_list3 *l3 = cachep->nodelists[node]; |
1da177e4 LT |
2162 | int ret; |
2163 | ||
e498be7d | 2164 | for (;;) { |
1da177e4 LT |
2165 | struct list_head *p; |
2166 | ||
e498be7d CL |
2167 | p = l3->slabs_free.prev; |
2168 | if (p == &l3->slabs_free) | |
1da177e4 LT |
2169 | break; |
2170 | ||
e498be7d | 2171 | slabp = list_entry(l3->slabs_free.prev, struct slab, list); |
1da177e4 LT |
2172 | #if DEBUG |
2173 | if (slabp->inuse) | |
2174 | BUG(); | |
2175 | #endif | |
2176 | list_del(&slabp->list); | |
2177 | ||
e498be7d CL |
2178 | l3->free_objects -= cachep->num; |
2179 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 2180 | slab_destroy(cachep, slabp); |
e498be7d | 2181 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 2182 | } |
b28a02de | 2183 | ret = !list_empty(&l3->slabs_full) || !list_empty(&l3->slabs_partial); |
1da177e4 LT |
2184 | return ret; |
2185 | } | |
2186 | ||
343e0d7a | 2187 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2188 | { |
2189 | int ret = 0, i = 0; | |
2190 | struct kmem_list3 *l3; | |
2191 | ||
2192 | drain_cpu_caches(cachep); | |
2193 | ||
2194 | check_irq_on(); | |
2195 | for_each_online_node(i) { | |
2196 | l3 = cachep->nodelists[i]; | |
2197 | if (l3) { | |
2198 | spin_lock_irq(&l3->list_lock); | |
2199 | ret += __node_shrink(cachep, i); | |
2200 | spin_unlock_irq(&l3->list_lock); | |
2201 | } | |
2202 | } | |
2203 | return (ret ? 1 : 0); | |
2204 | } | |
2205 | ||
1da177e4 LT |
2206 | /** |
2207 | * kmem_cache_shrink - Shrink a cache. | |
2208 | * @cachep: The cache to shrink. | |
2209 | * | |
2210 | * Releases as many slabs as possible for a cache. | |
2211 | * To help debugging, a zero exit status indicates all slabs were released. | |
2212 | */ | |
343e0d7a | 2213 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 LT |
2214 | { |
2215 | if (!cachep || in_interrupt()) | |
2216 | BUG(); | |
2217 | ||
2218 | return __cache_shrink(cachep); | |
2219 | } | |
2220 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2221 | ||
2222 | /** | |
2223 | * kmem_cache_destroy - delete a cache | |
2224 | * @cachep: the cache to destroy | |
2225 | * | |
343e0d7a | 2226 | * Remove a struct kmem_cache object from the slab cache. |
1da177e4 LT |
2227 | * Returns 0 on success. |
2228 | * | |
2229 | * It is expected this function will be called by a module when it is | |
2230 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2231 | * cache being allocated each time a module is loaded and unloaded, if the | |
2232 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2233 | * | |
2234 | * The cache must be empty before calling this function. | |
2235 | * | |
2236 | * The caller must guarantee that noone will allocate memory from the cache | |
2237 | * during the kmem_cache_destroy(). | |
2238 | */ | |
343e0d7a | 2239 | int kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 LT |
2240 | { |
2241 | int i; | |
e498be7d | 2242 | struct kmem_list3 *l3; |
1da177e4 LT |
2243 | |
2244 | if (!cachep || in_interrupt()) | |
2245 | BUG(); | |
2246 | ||
2247 | /* Don't let CPUs to come and go */ | |
2248 | lock_cpu_hotplug(); | |
2249 | ||
2250 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 2251 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2252 | /* |
2253 | * the chain is never empty, cache_cache is never destroyed | |
2254 | */ | |
2255 | list_del(&cachep->next); | |
fc0abb14 | 2256 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2257 | |
2258 | if (__cache_shrink(cachep)) { | |
2259 | slab_error(cachep, "Can't free all objects"); | |
fc0abb14 | 2260 | mutex_lock(&cache_chain_mutex); |
b28a02de | 2261 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2262 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
2263 | unlock_cpu_hotplug(); |
2264 | return 1; | |
2265 | } | |
2266 | ||
2267 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2268 | synchronize_rcu(); |
1da177e4 | 2269 | |
e498be7d | 2270 | for_each_online_cpu(i) |
b28a02de | 2271 | kfree(cachep->array[i]); |
1da177e4 LT |
2272 | |
2273 | /* NUMA: free the list3 structures */ | |
e498be7d CL |
2274 | for_each_online_node(i) { |
2275 | if ((l3 = cachep->nodelists[i])) { | |
2276 | kfree(l3->shared); | |
2277 | free_alien_cache(l3->alien); | |
2278 | kfree(l3); | |
2279 | } | |
2280 | } | |
1da177e4 LT |
2281 | kmem_cache_free(&cache_cache, cachep); |
2282 | ||
2283 | unlock_cpu_hotplug(); | |
2284 | ||
2285 | return 0; | |
2286 | } | |
2287 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2288 | ||
2289 | /* Get the memory for a slab management obj. */ | |
343e0d7a | 2290 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
b28a02de | 2291 | int colour_off, gfp_t local_flags) |
1da177e4 LT |
2292 | { |
2293 | struct slab *slabp; | |
b28a02de | 2294 | |
1da177e4 LT |
2295 | if (OFF_SLAB(cachep)) { |
2296 | /* Slab management obj is off-slab. */ | |
2297 | slabp = kmem_cache_alloc(cachep->slabp_cache, local_flags); | |
2298 | if (!slabp) | |
2299 | return NULL; | |
2300 | } else { | |
b28a02de | 2301 | slabp = objp + colour_off; |
1da177e4 LT |
2302 | colour_off += cachep->slab_size; |
2303 | } | |
2304 | slabp->inuse = 0; | |
2305 | slabp->colouroff = colour_off; | |
b28a02de | 2306 | slabp->s_mem = objp + colour_off; |
1da177e4 LT |
2307 | |
2308 | return slabp; | |
2309 | } | |
2310 | ||
2311 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2312 | { | |
b28a02de | 2313 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2314 | } |
2315 | ||
343e0d7a | 2316 | static void cache_init_objs(struct kmem_cache *cachep, |
b28a02de | 2317 | struct slab *slabp, unsigned long ctor_flags) |
1da177e4 LT |
2318 | { |
2319 | int i; | |
2320 | ||
2321 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2322 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2323 | #if DEBUG |
2324 | /* need to poison the objs? */ | |
2325 | if (cachep->flags & SLAB_POISON) | |
2326 | poison_obj(cachep, objp, POISON_FREE); | |
2327 | if (cachep->flags & SLAB_STORE_USER) | |
2328 | *dbg_userword(cachep, objp) = NULL; | |
2329 | ||
2330 | if (cachep->flags & SLAB_RED_ZONE) { | |
2331 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2332 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2333 | } | |
2334 | /* | |
2335 | * Constructors are not allowed to allocate memory from | |
2336 | * the same cache which they are a constructor for. | |
2337 | * Otherwise, deadlock. They must also be threaded. | |
2338 | */ | |
2339 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
3dafccf2 | 2340 | cachep->ctor(objp + obj_offset(cachep), cachep, |
b28a02de | 2341 | ctor_flags); |
1da177e4 LT |
2342 | |
2343 | if (cachep->flags & SLAB_RED_ZONE) { | |
2344 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2345 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2346 | " end of an object"); |
1da177e4 LT |
2347 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2348 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2349 | " start of an object"); |
1da177e4 | 2350 | } |
3dafccf2 | 2351 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep) |
b28a02de PE |
2352 | && cachep->flags & SLAB_POISON) |
2353 | kernel_map_pages(virt_to_page(objp), | |
3dafccf2 | 2354 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2355 | #else |
2356 | if (cachep->ctor) | |
2357 | cachep->ctor(objp, cachep, ctor_flags); | |
2358 | #endif | |
b28a02de | 2359 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2360 | } |
b28a02de | 2361 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2362 | slabp->free = 0; |
2363 | } | |
2364 | ||
343e0d7a | 2365 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2366 | { |
2367 | if (flags & SLAB_DMA) { | |
2368 | if (!(cachep->gfpflags & GFP_DMA)) | |
2369 | BUG(); | |
2370 | } else { | |
2371 | if (cachep->gfpflags & GFP_DMA) | |
2372 | BUG(); | |
2373 | } | |
2374 | } | |
2375 | ||
343e0d7a | 2376 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, int nodeid) |
78d382d7 | 2377 | { |
8fea4e96 | 2378 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2379 | kmem_bufctl_t next; |
2380 | ||
2381 | slabp->inuse++; | |
2382 | next = slab_bufctl(slabp)[slabp->free]; | |
2383 | #if DEBUG | |
2384 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2385 | WARN_ON(slabp->nodeid != nodeid); | |
2386 | #endif | |
2387 | slabp->free = next; | |
2388 | ||
2389 | return objp; | |
2390 | } | |
2391 | ||
343e0d7a | 2392 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, void *objp, |
78d382d7 MD |
2393 | int nodeid) |
2394 | { | |
8fea4e96 | 2395 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2396 | |
2397 | #if DEBUG | |
2398 | /* Verify that the slab belongs to the intended node */ | |
2399 | WARN_ON(slabp->nodeid != nodeid); | |
2400 | ||
2401 | if (slab_bufctl(slabp)[objnr] != BUFCTL_FREE) { | |
2402 | printk(KERN_ERR "slab: double free detected in cache " | |
2403 | "'%s', objp %p\n", cachep->name, objp); | |
2404 | BUG(); | |
2405 | } | |
2406 | #endif | |
2407 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2408 | slabp->free = objnr; | |
2409 | slabp->inuse--; | |
2410 | } | |
2411 | ||
343e0d7a | 2412 | static void set_slab_attr(struct kmem_cache *cachep, struct slab *slabp, void *objp) |
1da177e4 LT |
2413 | { |
2414 | int i; | |
2415 | struct page *page; | |
2416 | ||
2417 | /* Nasty!!!!!! I hope this is OK. */ | |
2418 | i = 1 << cachep->gfporder; | |
2419 | page = virt_to_page(objp); | |
2420 | do { | |
065d41cb PE |
2421 | page_set_cache(page, cachep); |
2422 | page_set_slab(page, slabp); | |
1da177e4 LT |
2423 | page++; |
2424 | } while (--i); | |
2425 | } | |
2426 | ||
2427 | /* | |
2428 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2429 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2430 | */ | |
343e0d7a | 2431 | static int cache_grow(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 2432 | { |
b28a02de PE |
2433 | struct slab *slabp; |
2434 | void *objp; | |
2435 | size_t offset; | |
2436 | gfp_t local_flags; | |
2437 | unsigned long ctor_flags; | |
e498be7d | 2438 | struct kmem_list3 *l3; |
1da177e4 LT |
2439 | |
2440 | /* Be lazy and only check for valid flags here, | |
b28a02de | 2441 | * keeping it out of the critical path in kmem_cache_alloc(). |
1da177e4 | 2442 | */ |
b28a02de | 2443 | if (flags & ~(SLAB_DMA | SLAB_LEVEL_MASK | SLAB_NO_GROW)) |
1da177e4 LT |
2444 | BUG(); |
2445 | if (flags & SLAB_NO_GROW) | |
2446 | return 0; | |
2447 | ||
2448 | ctor_flags = SLAB_CTOR_CONSTRUCTOR; | |
2449 | local_flags = (flags & SLAB_LEVEL_MASK); | |
2450 | if (!(local_flags & __GFP_WAIT)) | |
2451 | /* | |
2452 | * Not allowed to sleep. Need to tell a constructor about | |
2453 | * this - it might need to know... | |
2454 | */ | |
2455 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2456 | ||
2e1217cf | 2457 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2458 | check_irq_off(); |
2e1217cf RT |
2459 | l3 = cachep->nodelists[nodeid]; |
2460 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2461 | |
2462 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2463 | offset = l3->colour_next; |
2464 | l3->colour_next++; | |
2465 | if (l3->colour_next >= cachep->colour) | |
2466 | l3->colour_next = 0; | |
2467 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2468 | |
2e1217cf | 2469 | offset *= cachep->colour_off; |
1da177e4 LT |
2470 | |
2471 | if (local_flags & __GFP_WAIT) | |
2472 | local_irq_enable(); | |
2473 | ||
2474 | /* | |
2475 | * The test for missing atomic flag is performed here, rather than | |
2476 | * the more obvious place, simply to reduce the critical path length | |
2477 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2478 | * will eventually be caught here (where it matters). | |
2479 | */ | |
2480 | kmem_flagcheck(cachep, flags); | |
2481 | ||
e498be7d CL |
2482 | /* Get mem for the objs. |
2483 | * Attempt to allocate a physical page from 'nodeid', | |
2484 | */ | |
1da177e4 LT |
2485 | if (!(objp = kmem_getpages(cachep, flags, nodeid))) |
2486 | goto failed; | |
2487 | ||
2488 | /* Get slab management. */ | |
2489 | if (!(slabp = alloc_slabmgmt(cachep, objp, offset, local_flags))) | |
2490 | goto opps1; | |
2491 | ||
e498be7d | 2492 | slabp->nodeid = nodeid; |
1da177e4 LT |
2493 | set_slab_attr(cachep, slabp, objp); |
2494 | ||
2495 | cache_init_objs(cachep, slabp, ctor_flags); | |
2496 | ||
2497 | if (local_flags & __GFP_WAIT) | |
2498 | local_irq_disable(); | |
2499 | check_irq_off(); | |
e498be7d | 2500 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2501 | |
2502 | /* Make slab active. */ | |
e498be7d | 2503 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2504 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2505 | l3->free_objects += cachep->num; |
2506 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2507 | return 1; |
b28a02de | 2508 | opps1: |
1da177e4 | 2509 | kmem_freepages(cachep, objp); |
b28a02de | 2510 | failed: |
1da177e4 LT |
2511 | if (local_flags & __GFP_WAIT) |
2512 | local_irq_disable(); | |
2513 | return 0; | |
2514 | } | |
2515 | ||
2516 | #if DEBUG | |
2517 | ||
2518 | /* | |
2519 | * Perform extra freeing checks: | |
2520 | * - detect bad pointers. | |
2521 | * - POISON/RED_ZONE checking | |
2522 | * - destructor calls, for caches with POISON+dtor | |
2523 | */ | |
2524 | static void kfree_debugcheck(const void *objp) | |
2525 | { | |
2526 | struct page *page; | |
2527 | ||
2528 | if (!virt_addr_valid(objp)) { | |
2529 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2530 | (unsigned long)objp); |
2531 | BUG(); | |
1da177e4 LT |
2532 | } |
2533 | page = virt_to_page(objp); | |
2534 | if (!PageSlab(page)) { | |
b28a02de PE |
2535 | printk(KERN_ERR "kfree_debugcheck: bad ptr %lxh.\n", |
2536 | (unsigned long)objp); | |
1da177e4 LT |
2537 | BUG(); |
2538 | } | |
2539 | } | |
2540 | ||
343e0d7a | 2541 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2542 | void *caller) |
1da177e4 LT |
2543 | { |
2544 | struct page *page; | |
2545 | unsigned int objnr; | |
2546 | struct slab *slabp; | |
2547 | ||
3dafccf2 | 2548 | objp -= obj_offset(cachep); |
1da177e4 LT |
2549 | kfree_debugcheck(objp); |
2550 | page = virt_to_page(objp); | |
2551 | ||
065d41cb | 2552 | if (page_get_cache(page) != cachep) { |
b28a02de PE |
2553 | printk(KERN_ERR |
2554 | "mismatch in kmem_cache_free: expected cache %p, got %p\n", | |
2555 | page_get_cache(page), cachep); | |
1da177e4 | 2556 | printk(KERN_ERR "%p is %s.\n", cachep, cachep->name); |
b28a02de PE |
2557 | printk(KERN_ERR "%p is %s.\n", page_get_cache(page), |
2558 | page_get_cache(page)->name); | |
1da177e4 LT |
2559 | WARN_ON(1); |
2560 | } | |
065d41cb | 2561 | slabp = page_get_slab(page); |
1da177e4 LT |
2562 | |
2563 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2564 | if (*dbg_redzone1(cachep, objp) != RED_ACTIVE |
2565 | || *dbg_redzone2(cachep, objp) != RED_ACTIVE) { | |
2566 | slab_error(cachep, | |
2567 | "double free, or memory outside" | |
2568 | " object was overwritten"); | |
2569 | printk(KERN_ERR | |
2570 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2571 | objp, *dbg_redzone1(cachep, objp), | |
2572 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2573 | } |
2574 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2575 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2576 | } | |
2577 | if (cachep->flags & SLAB_STORE_USER) | |
2578 | *dbg_userword(cachep, objp) = caller; | |
2579 | ||
8fea4e96 | 2580 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2581 | |
2582 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2583 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 LT |
2584 | |
2585 | if (cachep->flags & SLAB_DEBUG_INITIAL) { | |
2586 | /* Need to call the slab's constructor so the | |
2587 | * caller can perform a verify of its state (debugging). | |
2588 | * Called without the cache-lock held. | |
2589 | */ | |
3dafccf2 | 2590 | cachep->ctor(objp + obj_offset(cachep), |
b28a02de | 2591 | cachep, SLAB_CTOR_CONSTRUCTOR | SLAB_CTOR_VERIFY); |
1da177e4 LT |
2592 | } |
2593 | if (cachep->flags & SLAB_POISON && cachep->dtor) { | |
2594 | /* we want to cache poison the object, | |
2595 | * call the destruction callback | |
2596 | */ | |
3dafccf2 | 2597 | cachep->dtor(objp + obj_offset(cachep), cachep, 0); |
1da177e4 LT |
2598 | } |
2599 | if (cachep->flags & SLAB_POISON) { | |
2600 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3dafccf2 | 2601 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) { |
1da177e4 | 2602 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2603 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2604 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2605 | } else { |
2606 | poison_obj(cachep, objp, POISON_FREE); | |
2607 | } | |
2608 | #else | |
2609 | poison_obj(cachep, objp, POISON_FREE); | |
2610 | #endif | |
2611 | } | |
2612 | return objp; | |
2613 | } | |
2614 | ||
343e0d7a | 2615 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2616 | { |
2617 | kmem_bufctl_t i; | |
2618 | int entries = 0; | |
b28a02de | 2619 | |
1da177e4 LT |
2620 | /* Check slab's freelist to see if this obj is there. */ |
2621 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2622 | entries++; | |
2623 | if (entries > cachep->num || i >= cachep->num) | |
2624 | goto bad; | |
2625 | } | |
2626 | if (entries != cachep->num - slabp->inuse) { | |
b28a02de PE |
2627 | bad: |
2628 | printk(KERN_ERR | |
2629 | "slab: Internal list corruption detected in cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2630 | cachep->name, cachep->num, slabp, slabp->inuse); | |
2631 | for (i = 0; | |
264132bc | 2632 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de PE |
2633 | i++) { |
2634 | if ((i % 16) == 0) | |
1da177e4 | 2635 | printk("\n%03x:", i); |
b28a02de | 2636 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2637 | } |
2638 | printk("\n"); | |
2639 | BUG(); | |
2640 | } | |
2641 | } | |
2642 | #else | |
2643 | #define kfree_debugcheck(x) do { } while(0) | |
2644 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2645 | #define check_slabp(x,y) do { } while(0) | |
2646 | #endif | |
2647 | ||
343e0d7a | 2648 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2649 | { |
2650 | int batchcount; | |
2651 | struct kmem_list3 *l3; | |
2652 | struct array_cache *ac; | |
2653 | ||
2654 | check_irq_off(); | |
9a2dba4b | 2655 | ac = cpu_cache_get(cachep); |
b28a02de | 2656 | retry: |
1da177e4 LT |
2657 | batchcount = ac->batchcount; |
2658 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
2659 | /* if there was little recent activity on this | |
2660 | * cache, then perform only a partial refill. | |
2661 | * Otherwise we could generate refill bouncing. | |
2662 | */ | |
2663 | batchcount = BATCHREFILL_LIMIT; | |
2664 | } | |
e498be7d CL |
2665 | l3 = cachep->nodelists[numa_node_id()]; |
2666 | ||
2667 | BUG_ON(ac->avail > 0 || !l3); | |
2668 | spin_lock(&l3->list_lock); | |
1da177e4 | 2669 | |
1da177e4 LT |
2670 | if (l3->shared) { |
2671 | struct array_cache *shared_array = l3->shared; | |
2672 | if (shared_array->avail) { | |
2673 | if (batchcount > shared_array->avail) | |
2674 | batchcount = shared_array->avail; | |
2675 | shared_array->avail -= batchcount; | |
2676 | ac->avail = batchcount; | |
e498be7d | 2677 | memcpy(ac->entry, |
b28a02de PE |
2678 | &(shared_array->entry[shared_array->avail]), |
2679 | sizeof(void *) * batchcount); | |
1da177e4 LT |
2680 | shared_array->touched = 1; |
2681 | goto alloc_done; | |
2682 | } | |
2683 | } | |
2684 | while (batchcount > 0) { | |
2685 | struct list_head *entry; | |
2686 | struct slab *slabp; | |
2687 | /* Get slab alloc is to come from. */ | |
2688 | entry = l3->slabs_partial.next; | |
2689 | if (entry == &l3->slabs_partial) { | |
2690 | l3->free_touched = 1; | |
2691 | entry = l3->slabs_free.next; | |
2692 | if (entry == &l3->slabs_free) | |
2693 | goto must_grow; | |
2694 | } | |
2695 | ||
2696 | slabp = list_entry(entry, struct slab, list); | |
2697 | check_slabp(cachep, slabp); | |
2698 | check_spinlock_acquired(cachep); | |
2699 | while (slabp->inuse < cachep->num && batchcount--) { | |
1da177e4 LT |
2700 | STATS_INC_ALLOCED(cachep); |
2701 | STATS_INC_ACTIVE(cachep); | |
2702 | STATS_SET_HIGH(cachep); | |
2703 | ||
78d382d7 MD |
2704 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
2705 | numa_node_id()); | |
1da177e4 LT |
2706 | } |
2707 | check_slabp(cachep, slabp); | |
2708 | ||
2709 | /* move slabp to correct slabp list: */ | |
2710 | list_del(&slabp->list); | |
2711 | if (slabp->free == BUFCTL_END) | |
2712 | list_add(&slabp->list, &l3->slabs_full); | |
2713 | else | |
2714 | list_add(&slabp->list, &l3->slabs_partial); | |
2715 | } | |
2716 | ||
b28a02de | 2717 | must_grow: |
1da177e4 | 2718 | l3->free_objects -= ac->avail; |
b28a02de | 2719 | alloc_done: |
e498be7d | 2720 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
2721 | |
2722 | if (unlikely(!ac->avail)) { | |
2723 | int x; | |
e498be7d CL |
2724 | x = cache_grow(cachep, flags, numa_node_id()); |
2725 | ||
1da177e4 | 2726 | // cache_grow can reenable interrupts, then ac could change. |
9a2dba4b | 2727 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2728 | if (!x && ac->avail == 0) // no objects in sight? abort |
2729 | return NULL; | |
2730 | ||
b28a02de | 2731 | if (!ac->avail) // objects refilled by interrupt? |
1da177e4 LT |
2732 | goto retry; |
2733 | } | |
2734 | ac->touched = 1; | |
e498be7d | 2735 | return ac->entry[--ac->avail]; |
1da177e4 LT |
2736 | } |
2737 | ||
2738 | static inline void | |
343e0d7a | 2739 | cache_alloc_debugcheck_before(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2740 | { |
2741 | might_sleep_if(flags & __GFP_WAIT); | |
2742 | #if DEBUG | |
2743 | kmem_flagcheck(cachep, flags); | |
2744 | #endif | |
2745 | } | |
2746 | ||
2747 | #if DEBUG | |
343e0d7a | 2748 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, gfp_t flags, |
b28a02de | 2749 | void *objp, void *caller) |
1da177e4 | 2750 | { |
b28a02de | 2751 | if (!objp) |
1da177e4 | 2752 | return objp; |
b28a02de | 2753 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2754 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 2755 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2756 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2757 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
2758 | else |
2759 | check_poison_obj(cachep, objp); | |
2760 | #else | |
2761 | check_poison_obj(cachep, objp); | |
2762 | #endif | |
2763 | poison_obj(cachep, objp, POISON_INUSE); | |
2764 | } | |
2765 | if (cachep->flags & SLAB_STORE_USER) | |
2766 | *dbg_userword(cachep, objp) = caller; | |
2767 | ||
2768 | if (cachep->flags & SLAB_RED_ZONE) { | |
b28a02de PE |
2769 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE |
2770 | || *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2771 | slab_error(cachep, | |
2772 | "double free, or memory outside" | |
2773 | " object was overwritten"); | |
2774 | printk(KERN_ERR | |
2775 | "%p: redzone 1: 0x%lx, redzone 2: 0x%lx.\n", | |
2776 | objp, *dbg_redzone1(cachep, objp), | |
2777 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2778 | } |
2779 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2780 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2781 | } | |
3dafccf2 | 2782 | objp += obj_offset(cachep); |
1da177e4 | 2783 | if (cachep->ctor && cachep->flags & SLAB_POISON) { |
b28a02de | 2784 | unsigned long ctor_flags = SLAB_CTOR_CONSTRUCTOR; |
1da177e4 LT |
2785 | |
2786 | if (!(flags & __GFP_WAIT)) | |
2787 | ctor_flags |= SLAB_CTOR_ATOMIC; | |
2788 | ||
2789 | cachep->ctor(objp, cachep, ctor_flags); | |
b28a02de | 2790 | } |
1da177e4 LT |
2791 | return objp; |
2792 | } | |
2793 | #else | |
2794 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2795 | #endif | |
2796 | ||
343e0d7a | 2797 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2798 | { |
b28a02de | 2799 | void *objp; |
1da177e4 LT |
2800 | struct array_cache *ac; |
2801 | ||
dc85da15 | 2802 | #ifdef CONFIG_NUMA |
86c562a9 | 2803 | if (unlikely(current->mempolicy && !in_interrupt())) { |
dc85da15 CL |
2804 | int nid = slab_node(current->mempolicy); |
2805 | ||
2806 | if (nid != numa_node_id()) | |
2807 | return __cache_alloc_node(cachep, flags, nid); | |
2808 | } | |
2809 | #endif | |
2810 | ||
5c382300 | 2811 | check_irq_off(); |
9a2dba4b | 2812 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2813 | if (likely(ac->avail)) { |
2814 | STATS_INC_ALLOCHIT(cachep); | |
2815 | ac->touched = 1; | |
e498be7d | 2816 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
2817 | } else { |
2818 | STATS_INC_ALLOCMISS(cachep); | |
2819 | objp = cache_alloc_refill(cachep, flags); | |
2820 | } | |
5c382300 AK |
2821 | return objp; |
2822 | } | |
2823 | ||
7fd6b141 PE |
2824 | static __always_inline void * |
2825 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
5c382300 AK |
2826 | { |
2827 | unsigned long save_flags; | |
b28a02de | 2828 | void *objp; |
5c382300 AK |
2829 | |
2830 | cache_alloc_debugcheck_before(cachep, flags); | |
2831 | ||
2832 | local_irq_save(save_flags); | |
2833 | objp = ____cache_alloc(cachep, flags); | |
1da177e4 | 2834 | local_irq_restore(save_flags); |
34342e86 | 2835 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, |
7fd6b141 | 2836 | caller); |
34342e86 | 2837 | prefetchw(objp); |
1da177e4 LT |
2838 | return objp; |
2839 | } | |
2840 | ||
e498be7d CL |
2841 | #ifdef CONFIG_NUMA |
2842 | /* | |
2843 | * A interface to enable slab creation on nodeid | |
1da177e4 | 2844 | */ |
343e0d7a | 2845 | static void *__cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
e498be7d CL |
2846 | { |
2847 | struct list_head *entry; | |
b28a02de PE |
2848 | struct slab *slabp; |
2849 | struct kmem_list3 *l3; | |
2850 | void *obj; | |
b28a02de PE |
2851 | int x; |
2852 | ||
2853 | l3 = cachep->nodelists[nodeid]; | |
2854 | BUG_ON(!l3); | |
2855 | ||
2856 | retry: | |
ca3b9b91 | 2857 | check_irq_off(); |
b28a02de PE |
2858 | spin_lock(&l3->list_lock); |
2859 | entry = l3->slabs_partial.next; | |
2860 | if (entry == &l3->slabs_partial) { | |
2861 | l3->free_touched = 1; | |
2862 | entry = l3->slabs_free.next; | |
2863 | if (entry == &l3->slabs_free) | |
2864 | goto must_grow; | |
2865 | } | |
2866 | ||
2867 | slabp = list_entry(entry, struct slab, list); | |
2868 | check_spinlock_acquired_node(cachep, nodeid); | |
2869 | check_slabp(cachep, slabp); | |
2870 | ||
2871 | STATS_INC_NODEALLOCS(cachep); | |
2872 | STATS_INC_ACTIVE(cachep); | |
2873 | STATS_SET_HIGH(cachep); | |
2874 | ||
2875 | BUG_ON(slabp->inuse == cachep->num); | |
2876 | ||
78d382d7 | 2877 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
2878 | check_slabp(cachep, slabp); |
2879 | l3->free_objects--; | |
2880 | /* move slabp to correct slabp list: */ | |
2881 | list_del(&slabp->list); | |
2882 | ||
2883 | if (slabp->free == BUFCTL_END) { | |
2884 | list_add(&slabp->list, &l3->slabs_full); | |
2885 | } else { | |
2886 | list_add(&slabp->list, &l3->slabs_partial); | |
2887 | } | |
e498be7d | 2888 | |
b28a02de PE |
2889 | spin_unlock(&l3->list_lock); |
2890 | goto done; | |
e498be7d | 2891 | |
b28a02de PE |
2892 | must_grow: |
2893 | spin_unlock(&l3->list_lock); | |
2894 | x = cache_grow(cachep, flags, nodeid); | |
1da177e4 | 2895 | |
b28a02de PE |
2896 | if (!x) |
2897 | return NULL; | |
e498be7d | 2898 | |
b28a02de PE |
2899 | goto retry; |
2900 | done: | |
2901 | return obj; | |
e498be7d CL |
2902 | } |
2903 | #endif | |
2904 | ||
2905 | /* | |
2906 | * Caller needs to acquire correct kmem_list's list_lock | |
2907 | */ | |
343e0d7a | 2908 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 2909 | int node) |
1da177e4 LT |
2910 | { |
2911 | int i; | |
e498be7d | 2912 | struct kmem_list3 *l3; |
1da177e4 LT |
2913 | |
2914 | for (i = 0; i < nr_objects; i++) { | |
2915 | void *objp = objpp[i]; | |
2916 | struct slab *slabp; | |
1da177e4 | 2917 | |
6ed5eb22 | 2918 | slabp = virt_to_slab(objp); |
ff69416e | 2919 | l3 = cachep->nodelists[node]; |
1da177e4 | 2920 | list_del(&slabp->list); |
ff69416e | 2921 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 2922 | check_slabp(cachep, slabp); |
78d382d7 | 2923 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 2924 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 2925 | l3->free_objects++; |
1da177e4 LT |
2926 | check_slabp(cachep, slabp); |
2927 | ||
2928 | /* fixup slab chains */ | |
2929 | if (slabp->inuse == 0) { | |
e498be7d CL |
2930 | if (l3->free_objects > l3->free_limit) { |
2931 | l3->free_objects -= cachep->num; | |
1da177e4 LT |
2932 | slab_destroy(cachep, slabp); |
2933 | } else { | |
e498be7d | 2934 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
2935 | } |
2936 | } else { | |
2937 | /* Unconditionally move a slab to the end of the | |
2938 | * partial list on free - maximum time for the | |
2939 | * other objects to be freed, too. | |
2940 | */ | |
e498be7d | 2941 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
2942 | } |
2943 | } | |
2944 | } | |
2945 | ||
343e0d7a | 2946 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
2947 | { |
2948 | int batchcount; | |
e498be7d | 2949 | struct kmem_list3 *l3; |
ff69416e | 2950 | int node = numa_node_id(); |
1da177e4 LT |
2951 | |
2952 | batchcount = ac->batchcount; | |
2953 | #if DEBUG | |
2954 | BUG_ON(!batchcount || batchcount > ac->avail); | |
2955 | #endif | |
2956 | check_irq_off(); | |
ff69416e | 2957 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2958 | spin_lock(&l3->list_lock); |
2959 | if (l3->shared) { | |
2960 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 2961 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
2962 | if (max) { |
2963 | if (batchcount > max) | |
2964 | batchcount = max; | |
e498be7d | 2965 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 2966 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
2967 | shared_array->avail += batchcount; |
2968 | goto free_done; | |
2969 | } | |
2970 | } | |
2971 | ||
ff69416e | 2972 | free_block(cachep, ac->entry, batchcount, node); |
b28a02de | 2973 | free_done: |
1da177e4 LT |
2974 | #if STATS |
2975 | { | |
2976 | int i = 0; | |
2977 | struct list_head *p; | |
2978 | ||
e498be7d CL |
2979 | p = l3->slabs_free.next; |
2980 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
2981 | struct slab *slabp; |
2982 | ||
2983 | slabp = list_entry(p, struct slab, list); | |
2984 | BUG_ON(slabp->inuse); | |
2985 | ||
2986 | i++; | |
2987 | p = p->next; | |
2988 | } | |
2989 | STATS_SET_FREEABLE(cachep, i); | |
2990 | } | |
2991 | #endif | |
e498be7d | 2992 | spin_unlock(&l3->list_lock); |
1da177e4 | 2993 | ac->avail -= batchcount; |
e498be7d | 2994 | memmove(ac->entry, &(ac->entry[batchcount]), |
b28a02de | 2995 | sizeof(void *) * ac->avail); |
1da177e4 LT |
2996 | } |
2997 | ||
2998 | /* | |
2999 | * __cache_free | |
3000 | * Release an obj back to its cache. If the obj has a constructed | |
3001 | * state, it must be in this state _before_ it is released. | |
3002 | * | |
3003 | * Called with disabled ints. | |
3004 | */ | |
343e0d7a | 3005 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3006 | { |
9a2dba4b | 3007 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3008 | |
3009 | check_irq_off(); | |
3010 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); | |
3011 | ||
e498be7d CL |
3012 | /* Make sure we are not freeing a object from another |
3013 | * node to the array cache on this cpu. | |
3014 | */ | |
3015 | #ifdef CONFIG_NUMA | |
3016 | { | |
3017 | struct slab *slabp; | |
6ed5eb22 | 3018 | slabp = virt_to_slab(objp); |
e498be7d CL |
3019 | if (unlikely(slabp->nodeid != numa_node_id())) { |
3020 | struct array_cache *alien = NULL; | |
3021 | int nodeid = slabp->nodeid; | |
b28a02de PE |
3022 | struct kmem_list3 *l3 = |
3023 | cachep->nodelists[numa_node_id()]; | |
e498be7d CL |
3024 | |
3025 | STATS_INC_NODEFREES(cachep); | |
3026 | if (l3->alien && l3->alien[nodeid]) { | |
3027 | alien = l3->alien[nodeid]; | |
3028 | spin_lock(&alien->lock); | |
3029 | if (unlikely(alien->avail == alien->limit)) | |
3030 | __drain_alien_cache(cachep, | |
b28a02de | 3031 | alien, nodeid); |
e498be7d CL |
3032 | alien->entry[alien->avail++] = objp; |
3033 | spin_unlock(&alien->lock); | |
3034 | } else { | |
3035 | spin_lock(&(cachep->nodelists[nodeid])-> | |
b28a02de | 3036 | list_lock); |
ff69416e | 3037 | free_block(cachep, &objp, 1, nodeid); |
e498be7d | 3038 | spin_unlock(&(cachep->nodelists[nodeid])-> |
b28a02de | 3039 | list_lock); |
e498be7d CL |
3040 | } |
3041 | return; | |
3042 | } | |
3043 | } | |
3044 | #endif | |
1da177e4 LT |
3045 | if (likely(ac->avail < ac->limit)) { |
3046 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3047 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3048 | return; |
3049 | } else { | |
3050 | STATS_INC_FREEMISS(cachep); | |
3051 | cache_flusharray(cachep, ac); | |
e498be7d | 3052 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3053 | } |
3054 | } | |
3055 | ||
3056 | /** | |
3057 | * kmem_cache_alloc - Allocate an object | |
3058 | * @cachep: The cache to allocate from. | |
3059 | * @flags: See kmalloc(). | |
3060 | * | |
3061 | * Allocate an object from this cache. The flags are only relevant | |
3062 | * if the cache has no available objects. | |
3063 | */ | |
343e0d7a | 3064 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3065 | { |
7fd6b141 | 3066 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); |
1da177e4 LT |
3067 | } |
3068 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3069 | ||
3070 | /** | |
3071 | * kmem_ptr_validate - check if an untrusted pointer might | |
3072 | * be a slab entry. | |
3073 | * @cachep: the cache we're checking against | |
3074 | * @ptr: pointer to validate | |
3075 | * | |
3076 | * This verifies that the untrusted pointer looks sane: | |
3077 | * it is _not_ a guarantee that the pointer is actually | |
3078 | * part of the slab cache in question, but it at least | |
3079 | * validates that the pointer can be dereferenced and | |
3080 | * looks half-way sane. | |
3081 | * | |
3082 | * Currently only used for dentry validation. | |
3083 | */ | |
343e0d7a | 3084 | int fastcall kmem_ptr_validate(struct kmem_cache *cachep, void *ptr) |
1da177e4 | 3085 | { |
b28a02de | 3086 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3087 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3088 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3089 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3090 | struct page *page; |
3091 | ||
3092 | if (unlikely(addr < min_addr)) | |
3093 | goto out; | |
3094 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3095 | goto out; | |
3096 | if (unlikely(addr & align_mask)) | |
3097 | goto out; | |
3098 | if (unlikely(!kern_addr_valid(addr))) | |
3099 | goto out; | |
3100 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3101 | goto out; | |
3102 | page = virt_to_page(ptr); | |
3103 | if (unlikely(!PageSlab(page))) | |
3104 | goto out; | |
065d41cb | 3105 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3106 | goto out; |
3107 | return 1; | |
b28a02de | 3108 | out: |
1da177e4 LT |
3109 | return 0; |
3110 | } | |
3111 | ||
3112 | #ifdef CONFIG_NUMA | |
3113 | /** | |
3114 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3115 | * @cachep: The cache to allocate from. | |
3116 | * @flags: See kmalloc(). | |
3117 | * @nodeid: node number of the target node. | |
3118 | * | |
3119 | * Identical to kmem_cache_alloc, except that this function is slow | |
3120 | * and can sleep. And it will allocate memory on the given node, which | |
3121 | * can improve the performance for cpu bound structures. | |
e498be7d CL |
3122 | * New and improved: it will now make sure that the object gets |
3123 | * put on the correct node list so that there is no false sharing. | |
1da177e4 | 3124 | */ |
343e0d7a | 3125 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 | 3126 | { |
e498be7d CL |
3127 | unsigned long save_flags; |
3128 | void *ptr; | |
1da177e4 | 3129 | |
e498be7d CL |
3130 | cache_alloc_debugcheck_before(cachep, flags); |
3131 | local_irq_save(save_flags); | |
18f820f6 CL |
3132 | |
3133 | if (nodeid == -1 || nodeid == numa_node_id() || | |
3134 | !cachep->nodelists[nodeid]) | |
5c382300 AK |
3135 | ptr = ____cache_alloc(cachep, flags); |
3136 | else | |
3137 | ptr = __cache_alloc_node(cachep, flags, nodeid); | |
e498be7d | 3138 | local_irq_restore(save_flags); |
18f820f6 CL |
3139 | |
3140 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, | |
3141 | __builtin_return_address(0)); | |
1da177e4 | 3142 | |
e498be7d | 3143 | return ptr; |
1da177e4 LT |
3144 | } |
3145 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3146 | ||
dd0fc66f | 3147 | void *kmalloc_node(size_t size, gfp_t flags, int node) |
97e2bde4 | 3148 | { |
343e0d7a | 3149 | struct kmem_cache *cachep; |
97e2bde4 MS |
3150 | |
3151 | cachep = kmem_find_general_cachep(size, flags); | |
3152 | if (unlikely(cachep == NULL)) | |
3153 | return NULL; | |
3154 | return kmem_cache_alloc_node(cachep, flags, node); | |
3155 | } | |
3156 | EXPORT_SYMBOL(kmalloc_node); | |
1da177e4 LT |
3157 | #endif |
3158 | ||
3159 | /** | |
3160 | * kmalloc - allocate memory | |
3161 | * @size: how many bytes of memory are required. | |
3162 | * @flags: the type of memory to allocate. | |
3163 | * | |
3164 | * kmalloc is the normal method of allocating memory | |
3165 | * in the kernel. | |
3166 | * | |
3167 | * The @flags argument may be one of: | |
3168 | * | |
3169 | * %GFP_USER - Allocate memory on behalf of user. May sleep. | |
3170 | * | |
3171 | * %GFP_KERNEL - Allocate normal kernel ram. May sleep. | |
3172 | * | |
3173 | * %GFP_ATOMIC - Allocation will not sleep. Use inside interrupt handlers. | |
3174 | * | |
3175 | * Additionally, the %GFP_DMA flag may be set to indicate the memory | |
3176 | * must be suitable for DMA. This can mean different things on different | |
3177 | * platforms. For example, on i386, it means that the memory must come | |
3178 | * from the first 16MB. | |
3179 | */ | |
7fd6b141 PE |
3180 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3181 | void *caller) | |
1da177e4 | 3182 | { |
343e0d7a | 3183 | struct kmem_cache *cachep; |
1da177e4 | 3184 | |
97e2bde4 MS |
3185 | /* If you want to save a few bytes .text space: replace |
3186 | * __ with kmem_. | |
3187 | * Then kmalloc uses the uninlined functions instead of the inline | |
3188 | * functions. | |
3189 | */ | |
3190 | cachep = __find_general_cachep(size, flags); | |
dbdb9045 AM |
3191 | if (unlikely(cachep == NULL)) |
3192 | return NULL; | |
7fd6b141 PE |
3193 | return __cache_alloc(cachep, flags, caller); |
3194 | } | |
3195 | ||
3196 | #ifndef CONFIG_DEBUG_SLAB | |
3197 | ||
3198 | void *__kmalloc(size_t size, gfp_t flags) | |
3199 | { | |
3200 | return __do_kmalloc(size, flags, NULL); | |
1da177e4 LT |
3201 | } |
3202 | EXPORT_SYMBOL(__kmalloc); | |
3203 | ||
7fd6b141 PE |
3204 | #else |
3205 | ||
3206 | void *__kmalloc_track_caller(size_t size, gfp_t flags, void *caller) | |
3207 | { | |
3208 | return __do_kmalloc(size, flags, caller); | |
3209 | } | |
3210 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
3211 | ||
3212 | #endif | |
3213 | ||
1da177e4 LT |
3214 | #ifdef CONFIG_SMP |
3215 | /** | |
3216 | * __alloc_percpu - allocate one copy of the object for every present | |
3217 | * cpu in the system, zeroing them. | |
3218 | * Objects should be dereferenced using the per_cpu_ptr macro only. | |
3219 | * | |
3220 | * @size: how many bytes of memory are required. | |
1da177e4 | 3221 | */ |
f9f75005 | 3222 | void *__alloc_percpu(size_t size) |
1da177e4 LT |
3223 | { |
3224 | int i; | |
b28a02de | 3225 | struct percpu_data *pdata = kmalloc(sizeof(*pdata), GFP_KERNEL); |
1da177e4 LT |
3226 | |
3227 | if (!pdata) | |
3228 | return NULL; | |
3229 | ||
e498be7d CL |
3230 | /* |
3231 | * Cannot use for_each_online_cpu since a cpu may come online | |
3232 | * and we have no way of figuring out how to fix the array | |
3233 | * that we have allocated then.... | |
3234 | */ | |
3235 | for_each_cpu(i) { | |
3236 | int node = cpu_to_node(i); | |
3237 | ||
3238 | if (node_online(node)) | |
3239 | pdata->ptrs[i] = kmalloc_node(size, GFP_KERNEL, node); | |
3240 | else | |
3241 | pdata->ptrs[i] = kmalloc(size, GFP_KERNEL); | |
1da177e4 LT |
3242 | |
3243 | if (!pdata->ptrs[i]) | |
3244 | goto unwind_oom; | |
3245 | memset(pdata->ptrs[i], 0, size); | |
3246 | } | |
3247 | ||
3248 | /* Catch derefs w/o wrappers */ | |
b28a02de | 3249 | return (void *)(~(unsigned long)pdata); |
1da177e4 | 3250 | |
b28a02de | 3251 | unwind_oom: |
1da177e4 LT |
3252 | while (--i >= 0) { |
3253 | if (!cpu_possible(i)) | |
3254 | continue; | |
3255 | kfree(pdata->ptrs[i]); | |
3256 | } | |
3257 | kfree(pdata); | |
3258 | return NULL; | |
3259 | } | |
3260 | EXPORT_SYMBOL(__alloc_percpu); | |
3261 | #endif | |
3262 | ||
3263 | /** | |
3264 | * kmem_cache_free - Deallocate an object | |
3265 | * @cachep: The cache the allocation was from. | |
3266 | * @objp: The previously allocated object. | |
3267 | * | |
3268 | * Free an object which was previously allocated from this | |
3269 | * cache. | |
3270 | */ | |
343e0d7a | 3271 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3272 | { |
3273 | unsigned long flags; | |
3274 | ||
3275 | local_irq_save(flags); | |
3276 | __cache_free(cachep, objp); | |
3277 | local_irq_restore(flags); | |
3278 | } | |
3279 | EXPORT_SYMBOL(kmem_cache_free); | |
3280 | ||
1da177e4 LT |
3281 | /** |
3282 | * kfree - free previously allocated memory | |
3283 | * @objp: pointer returned by kmalloc. | |
3284 | * | |
80e93eff PE |
3285 | * If @objp is NULL, no operation is performed. |
3286 | * | |
1da177e4 LT |
3287 | * Don't free memory not originally allocated by kmalloc() |
3288 | * or you will run into trouble. | |
3289 | */ | |
3290 | void kfree(const void *objp) | |
3291 | { | |
343e0d7a | 3292 | struct kmem_cache *c; |
1da177e4 LT |
3293 | unsigned long flags; |
3294 | ||
3295 | if (unlikely(!objp)) | |
3296 | return; | |
3297 | local_irq_save(flags); | |
3298 | kfree_debugcheck(objp); | |
6ed5eb22 | 3299 | c = virt_to_cache(objp); |
3dafccf2 | 3300 | mutex_debug_check_no_locks_freed(objp, obj_size(c)); |
b28a02de | 3301 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3302 | local_irq_restore(flags); |
3303 | } | |
3304 | EXPORT_SYMBOL(kfree); | |
3305 | ||
3306 | #ifdef CONFIG_SMP | |
3307 | /** | |
3308 | * free_percpu - free previously allocated percpu memory | |
3309 | * @objp: pointer returned by alloc_percpu. | |
3310 | * | |
3311 | * Don't free memory not originally allocated by alloc_percpu() | |
3312 | * The complemented objp is to check for that. | |
3313 | */ | |
b28a02de | 3314 | void free_percpu(const void *objp) |
1da177e4 LT |
3315 | { |
3316 | int i; | |
b28a02de | 3317 | struct percpu_data *p = (struct percpu_data *)(~(unsigned long)objp); |
1da177e4 | 3318 | |
e498be7d CL |
3319 | /* |
3320 | * We allocate for all cpus so we cannot use for online cpu here. | |
3321 | */ | |
3322 | for_each_cpu(i) | |
b28a02de | 3323 | kfree(p->ptrs[i]); |
1da177e4 LT |
3324 | kfree(p); |
3325 | } | |
3326 | EXPORT_SYMBOL(free_percpu); | |
3327 | #endif | |
3328 | ||
343e0d7a | 3329 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3330 | { |
3dafccf2 | 3331 | return obj_size(cachep); |
1da177e4 LT |
3332 | } |
3333 | EXPORT_SYMBOL(kmem_cache_size); | |
3334 | ||
343e0d7a | 3335 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3336 | { |
3337 | return cachep->name; | |
3338 | } | |
3339 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3340 | ||
e498be7d CL |
3341 | /* |
3342 | * This initializes kmem_list3 for all nodes. | |
3343 | */ | |
343e0d7a | 3344 | static int alloc_kmemlist(struct kmem_cache *cachep) |
e498be7d CL |
3345 | { |
3346 | int node; | |
3347 | struct kmem_list3 *l3; | |
3348 | int err = 0; | |
3349 | ||
3350 | for_each_online_node(node) { | |
3351 | struct array_cache *nc = NULL, *new; | |
3352 | struct array_cache **new_alien = NULL; | |
3353 | #ifdef CONFIG_NUMA | |
3354 | if (!(new_alien = alloc_alien_cache(node, cachep->limit))) | |
3355 | goto fail; | |
3356 | #endif | |
b28a02de PE |
3357 | if (!(new = alloc_arraycache(node, (cachep->shared * |
3358 | cachep->batchcount), | |
3359 | 0xbaadf00d))) | |
e498be7d CL |
3360 | goto fail; |
3361 | if ((l3 = cachep->nodelists[node])) { | |
3362 | ||
3363 | spin_lock_irq(&l3->list_lock); | |
3364 | ||
3365 | if ((nc = cachep->nodelists[node]->shared)) | |
b28a02de | 3366 | free_block(cachep, nc->entry, nc->avail, node); |
e498be7d CL |
3367 | |
3368 | l3->shared = new; | |
3369 | if (!cachep->nodelists[node]->alien) { | |
3370 | l3->alien = new_alien; | |
3371 | new_alien = NULL; | |
3372 | } | |
b28a02de PE |
3373 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3374 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3375 | spin_unlock_irq(&l3->list_lock); |
3376 | kfree(nc); | |
3377 | free_alien_cache(new_alien); | |
3378 | continue; | |
3379 | } | |
3380 | if (!(l3 = kmalloc_node(sizeof(struct kmem_list3), | |
b28a02de | 3381 | GFP_KERNEL, node))) |
e498be7d CL |
3382 | goto fail; |
3383 | ||
3384 | kmem_list3_init(l3); | |
3385 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
b28a02de | 3386 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
e498be7d CL |
3387 | l3->shared = new; |
3388 | l3->alien = new_alien; | |
b28a02de PE |
3389 | l3->free_limit = (1 + nr_cpus_node(node)) * |
3390 | cachep->batchcount + cachep->num; | |
e498be7d CL |
3391 | cachep->nodelists[node] = l3; |
3392 | } | |
3393 | return err; | |
b28a02de | 3394 | fail: |
e498be7d CL |
3395 | err = -ENOMEM; |
3396 | return err; | |
3397 | } | |
3398 | ||
1da177e4 | 3399 | struct ccupdate_struct { |
343e0d7a | 3400 | struct kmem_cache *cachep; |
1da177e4 LT |
3401 | struct array_cache *new[NR_CPUS]; |
3402 | }; | |
3403 | ||
3404 | static void do_ccupdate_local(void *info) | |
3405 | { | |
3406 | struct ccupdate_struct *new = (struct ccupdate_struct *)info; | |
3407 | struct array_cache *old; | |
3408 | ||
3409 | check_irq_off(); | |
9a2dba4b | 3410 | old = cpu_cache_get(new->cachep); |
e498be7d | 3411 | |
1da177e4 LT |
3412 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3413 | new->new[smp_processor_id()] = old; | |
3414 | } | |
3415 | ||
343e0d7a | 3416 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, int batchcount, |
b28a02de | 3417 | int shared) |
1da177e4 LT |
3418 | { |
3419 | struct ccupdate_struct new; | |
e498be7d | 3420 | int i, err; |
1da177e4 | 3421 | |
b28a02de | 3422 | memset(&new.new, 0, sizeof(new.new)); |
e498be7d | 3423 | for_each_online_cpu(i) { |
b28a02de PE |
3424 | new.new[i] = |
3425 | alloc_arraycache(cpu_to_node(i), limit, batchcount); | |
e498be7d | 3426 | if (!new.new[i]) { |
b28a02de PE |
3427 | for (i--; i >= 0; i--) |
3428 | kfree(new.new[i]); | |
e498be7d | 3429 | return -ENOMEM; |
1da177e4 LT |
3430 | } |
3431 | } | |
3432 | new.cachep = cachep; | |
3433 | ||
3434 | smp_call_function_all_cpus(do_ccupdate_local, (void *)&new); | |
e498be7d | 3435 | |
1da177e4 | 3436 | check_irq_on(); |
ca3b9b91 | 3437 | spin_lock(&cachep->spinlock); |
1da177e4 LT |
3438 | cachep->batchcount = batchcount; |
3439 | cachep->limit = limit; | |
e498be7d | 3440 | cachep->shared = shared; |
ca3b9b91 | 3441 | spin_unlock(&cachep->spinlock); |
1da177e4 | 3442 | |
e498be7d | 3443 | for_each_online_cpu(i) { |
1da177e4 LT |
3444 | struct array_cache *ccold = new.new[i]; |
3445 | if (!ccold) | |
3446 | continue; | |
e498be7d | 3447 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3448 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3449 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3450 | kfree(ccold); |
3451 | } | |
1da177e4 | 3452 | |
e498be7d CL |
3453 | err = alloc_kmemlist(cachep); |
3454 | if (err) { | |
3455 | printk(KERN_ERR "alloc_kmemlist failed for %s, error %d.\n", | |
b28a02de | 3456 | cachep->name, -err); |
e498be7d | 3457 | BUG(); |
1da177e4 | 3458 | } |
1da177e4 LT |
3459 | return 0; |
3460 | } | |
3461 | ||
343e0d7a | 3462 | static void enable_cpucache(struct kmem_cache *cachep) |
1da177e4 LT |
3463 | { |
3464 | int err; | |
3465 | int limit, shared; | |
3466 | ||
3467 | /* The head array serves three purposes: | |
3468 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
3469 | * - reduce the number of spinlock operations. | |
3470 | * - reduce the number of linked list operations on the slab and | |
3471 | * bufctl chains: array operations are cheaper. | |
3472 | * The numbers are guessed, we should auto-tune as described by | |
3473 | * Bonwick. | |
3474 | */ | |
3dafccf2 | 3475 | if (cachep->buffer_size > 131072) |
1da177e4 | 3476 | limit = 1; |
3dafccf2 | 3477 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3478 | limit = 8; |
3dafccf2 | 3479 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3480 | limit = 24; |
3dafccf2 | 3481 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3482 | limit = 54; |
3483 | else | |
3484 | limit = 120; | |
3485 | ||
3486 | /* Cpu bound tasks (e.g. network routing) can exhibit cpu bound | |
3487 | * allocation behaviour: Most allocs on one cpu, most free operations | |
3488 | * on another cpu. For these cases, an efficient object passing between | |
3489 | * cpus is necessary. This is provided by a shared array. The array | |
3490 | * replaces Bonwick's magazine layer. | |
3491 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3492 | * to a larger limit. Thus disabled by default. | |
3493 | */ | |
3494 | shared = 0; | |
3495 | #ifdef CONFIG_SMP | |
3dafccf2 | 3496 | if (cachep->buffer_size <= PAGE_SIZE) |
1da177e4 LT |
3497 | shared = 8; |
3498 | #endif | |
3499 | ||
3500 | #if DEBUG | |
3501 | /* With debugging enabled, large batchcount lead to excessively | |
3502 | * long periods with disabled local interrupts. Limit the | |
3503 | * batchcount | |
3504 | */ | |
3505 | if (limit > 32) | |
3506 | limit = 32; | |
3507 | #endif | |
b28a02de | 3508 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared); |
1da177e4 LT |
3509 | if (err) |
3510 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3511 | cachep->name, -err); |
1da177e4 LT |
3512 | } |
3513 | ||
343e0d7a | 3514 | static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, |
b28a02de | 3515 | int force, int node) |
1da177e4 LT |
3516 | { |
3517 | int tofree; | |
3518 | ||
e498be7d | 3519 | check_spinlock_acquired_node(cachep, node); |
1da177e4 LT |
3520 | if (ac->touched && !force) { |
3521 | ac->touched = 0; | |
3522 | } else if (ac->avail) { | |
b28a02de | 3523 | tofree = force ? ac->avail : (ac->limit + 4) / 5; |
1da177e4 | 3524 | if (tofree > ac->avail) { |
b28a02de | 3525 | tofree = (ac->avail + 1) / 2; |
1da177e4 | 3526 | } |
ff69416e | 3527 | free_block(cachep, ac->entry, tofree, node); |
1da177e4 | 3528 | ac->avail -= tofree; |
e498be7d | 3529 | memmove(ac->entry, &(ac->entry[tofree]), |
b28a02de | 3530 | sizeof(void *) * ac->avail); |
1da177e4 LT |
3531 | } |
3532 | } | |
3533 | ||
3534 | /** | |
3535 | * cache_reap - Reclaim memory from caches. | |
1e5d5331 | 3536 | * @unused: unused parameter |
1da177e4 LT |
3537 | * |
3538 | * Called from workqueue/eventd every few seconds. | |
3539 | * Purpose: | |
3540 | * - clear the per-cpu caches for this CPU. | |
3541 | * - return freeable pages to the main free memory pool. | |
3542 | * | |
fc0abb14 | 3543 | * If we cannot acquire the cache chain mutex then just give up - we'll |
1da177e4 LT |
3544 | * try again on the next iteration. |
3545 | */ | |
3546 | static void cache_reap(void *unused) | |
3547 | { | |
3548 | struct list_head *walk; | |
e498be7d | 3549 | struct kmem_list3 *l3; |
1da177e4 | 3550 | |
fc0abb14 | 3551 | if (!mutex_trylock(&cache_chain_mutex)) { |
1da177e4 | 3552 | /* Give up. Setup the next iteration. */ |
b28a02de PE |
3553 | schedule_delayed_work(&__get_cpu_var(reap_work), |
3554 | REAPTIMEOUT_CPUC); | |
1da177e4 LT |
3555 | return; |
3556 | } | |
3557 | ||
3558 | list_for_each(walk, &cache_chain) { | |
343e0d7a | 3559 | struct kmem_cache *searchp; |
b28a02de | 3560 | struct list_head *p; |
1da177e4 LT |
3561 | int tofree; |
3562 | struct slab *slabp; | |
3563 | ||
343e0d7a | 3564 | searchp = list_entry(walk, struct kmem_cache, next); |
1da177e4 LT |
3565 | |
3566 | if (searchp->flags & SLAB_NO_REAP) | |
3567 | goto next; | |
3568 | ||
3569 | check_irq_on(); | |
3570 | ||
e498be7d | 3571 | l3 = searchp->nodelists[numa_node_id()]; |
8fce4d8e | 3572 | reap_alien(searchp, l3); |
e498be7d | 3573 | spin_lock_irq(&l3->list_lock); |
1da177e4 | 3574 | |
9a2dba4b | 3575 | drain_array_locked(searchp, cpu_cache_get(searchp), 0, |
b28a02de | 3576 | numa_node_id()); |
1da177e4 | 3577 | |
e498be7d | 3578 | if (time_after(l3->next_reap, jiffies)) |
1da177e4 LT |
3579 | goto next_unlock; |
3580 | ||
e498be7d | 3581 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3582 | |
e498be7d CL |
3583 | if (l3->shared) |
3584 | drain_array_locked(searchp, l3->shared, 0, | |
b28a02de | 3585 | numa_node_id()); |
1da177e4 | 3586 | |
e498be7d CL |
3587 | if (l3->free_touched) { |
3588 | l3->free_touched = 0; | |
1da177e4 LT |
3589 | goto next_unlock; |
3590 | } | |
3591 | ||
b28a02de PE |
3592 | tofree = |
3593 | (l3->free_limit + 5 * searchp->num - | |
3594 | 1) / (5 * searchp->num); | |
1da177e4 | 3595 | do { |
e498be7d CL |
3596 | p = l3->slabs_free.next; |
3597 | if (p == &(l3->slabs_free)) | |
1da177e4 LT |
3598 | break; |
3599 | ||
3600 | slabp = list_entry(p, struct slab, list); | |
3601 | BUG_ON(slabp->inuse); | |
3602 | list_del(&slabp->list); | |
3603 | STATS_INC_REAPED(searchp); | |
3604 | ||
3605 | /* Safe to drop the lock. The slab is no longer | |
3606 | * linked to the cache. | |
3607 | * searchp cannot disappear, we hold | |
3608 | * cache_chain_lock | |
3609 | */ | |
e498be7d CL |
3610 | l3->free_objects -= searchp->num; |
3611 | spin_unlock_irq(&l3->list_lock); | |
1da177e4 | 3612 | slab_destroy(searchp, slabp); |
e498be7d | 3613 | spin_lock_irq(&l3->list_lock); |
b28a02de PE |
3614 | } while (--tofree > 0); |
3615 | next_unlock: | |
e498be7d | 3616 | spin_unlock_irq(&l3->list_lock); |
b28a02de | 3617 | next: |
1da177e4 LT |
3618 | cond_resched(); |
3619 | } | |
3620 | check_irq_on(); | |
fc0abb14 | 3621 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 3622 | next_reap_node(); |
1da177e4 | 3623 | /* Setup the next iteration */ |
cd61ef62 | 3624 | schedule_delayed_work(&__get_cpu_var(reap_work), REAPTIMEOUT_CPUC); |
1da177e4 LT |
3625 | } |
3626 | ||
3627 | #ifdef CONFIG_PROC_FS | |
3628 | ||
85289f98 | 3629 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 3630 | { |
85289f98 PE |
3631 | /* |
3632 | * Output format version, so at least we can change it | |
3633 | * without _too_ many complaints. | |
3634 | */ | |
1da177e4 | 3635 | #if STATS |
85289f98 | 3636 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 3637 | #else |
85289f98 | 3638 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 3639 | #endif |
85289f98 PE |
3640 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
3641 | "<objperslab> <pagesperslab>"); | |
3642 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
3643 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 3644 | #if STATS |
85289f98 PE |
3645 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
3646 | "<error> <maxfreeable> <nodeallocs> <remotefrees>"); | |
3647 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
1da177e4 | 3648 | #endif |
85289f98 PE |
3649 | seq_putc(m, '\n'); |
3650 | } | |
3651 | ||
3652 | static void *s_start(struct seq_file *m, loff_t *pos) | |
3653 | { | |
3654 | loff_t n = *pos; | |
3655 | struct list_head *p; | |
3656 | ||
fc0abb14 | 3657 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
3658 | if (!n) |
3659 | print_slabinfo_header(m); | |
1da177e4 LT |
3660 | p = cache_chain.next; |
3661 | while (n--) { | |
3662 | p = p->next; | |
3663 | if (p == &cache_chain) | |
3664 | return NULL; | |
3665 | } | |
343e0d7a | 3666 | return list_entry(p, struct kmem_cache, next); |
1da177e4 LT |
3667 | } |
3668 | ||
3669 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
3670 | { | |
343e0d7a | 3671 | struct kmem_cache *cachep = p; |
1da177e4 LT |
3672 | ++*pos; |
3673 | return cachep->next.next == &cache_chain ? NULL | |
343e0d7a | 3674 | : list_entry(cachep->next.next, struct kmem_cache, next); |
1da177e4 LT |
3675 | } |
3676 | ||
3677 | static void s_stop(struct seq_file *m, void *p) | |
3678 | { | |
fc0abb14 | 3679 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3680 | } |
3681 | ||
3682 | static int s_show(struct seq_file *m, void *p) | |
3683 | { | |
343e0d7a | 3684 | struct kmem_cache *cachep = p; |
1da177e4 | 3685 | struct list_head *q; |
b28a02de PE |
3686 | struct slab *slabp; |
3687 | unsigned long active_objs; | |
3688 | unsigned long num_objs; | |
3689 | unsigned long active_slabs = 0; | |
3690 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 3691 | const char *name; |
1da177e4 | 3692 | char *error = NULL; |
e498be7d CL |
3693 | int node; |
3694 | struct kmem_list3 *l3; | |
1da177e4 | 3695 | |
ca3b9b91 | 3696 | spin_lock(&cachep->spinlock); |
1da177e4 LT |
3697 | active_objs = 0; |
3698 | num_slabs = 0; | |
e498be7d CL |
3699 | for_each_online_node(node) { |
3700 | l3 = cachep->nodelists[node]; | |
3701 | if (!l3) | |
3702 | continue; | |
3703 | ||
ca3b9b91 RT |
3704 | check_irq_on(); |
3705 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 3706 | |
b28a02de | 3707 | list_for_each(q, &l3->slabs_full) { |
e498be7d CL |
3708 | slabp = list_entry(q, struct slab, list); |
3709 | if (slabp->inuse != cachep->num && !error) | |
3710 | error = "slabs_full accounting error"; | |
3711 | active_objs += cachep->num; | |
3712 | active_slabs++; | |
3713 | } | |
b28a02de | 3714 | list_for_each(q, &l3->slabs_partial) { |
e498be7d CL |
3715 | slabp = list_entry(q, struct slab, list); |
3716 | if (slabp->inuse == cachep->num && !error) | |
3717 | error = "slabs_partial inuse accounting error"; | |
3718 | if (!slabp->inuse && !error) | |
3719 | error = "slabs_partial/inuse accounting error"; | |
3720 | active_objs += slabp->inuse; | |
3721 | active_slabs++; | |
3722 | } | |
b28a02de | 3723 | list_for_each(q, &l3->slabs_free) { |
e498be7d CL |
3724 | slabp = list_entry(q, struct slab, list); |
3725 | if (slabp->inuse && !error) | |
3726 | error = "slabs_free/inuse accounting error"; | |
3727 | num_slabs++; | |
3728 | } | |
3729 | free_objects += l3->free_objects; | |
4484ebf1 RT |
3730 | if (l3->shared) |
3731 | shared_avail += l3->shared->avail; | |
e498be7d | 3732 | |
ca3b9b91 | 3733 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 3734 | } |
b28a02de PE |
3735 | num_slabs += active_slabs; |
3736 | num_objs = num_slabs * cachep->num; | |
e498be7d | 3737 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
3738 | error = "free_objects accounting error"; |
3739 | ||
b28a02de | 3740 | name = cachep->name; |
1da177e4 LT |
3741 | if (error) |
3742 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
3743 | ||
3744 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 3745 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 3746 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 3747 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 3748 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 3749 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 3750 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 3751 | #if STATS |
b28a02de | 3752 | { /* list3 stats */ |
1da177e4 LT |
3753 | unsigned long high = cachep->high_mark; |
3754 | unsigned long allocs = cachep->num_allocations; | |
3755 | unsigned long grown = cachep->grown; | |
3756 | unsigned long reaped = cachep->reaped; | |
3757 | unsigned long errors = cachep->errors; | |
3758 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 3759 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 3760 | unsigned long node_frees = cachep->node_frees; |
1da177e4 | 3761 | |
e498be7d | 3762 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
b28a02de | 3763 | %4lu %4lu %4lu %4lu", allocs, high, grown, reaped, errors, max_freeable, node_allocs, node_frees); |
1da177e4 LT |
3764 | } |
3765 | /* cpu stats */ | |
3766 | { | |
3767 | unsigned long allochit = atomic_read(&cachep->allochit); | |
3768 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
3769 | unsigned long freehit = atomic_read(&cachep->freehit); | |
3770 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
3771 | ||
3772 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 3773 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
3774 | } |
3775 | #endif | |
3776 | seq_putc(m, '\n'); | |
ca3b9b91 | 3777 | spin_unlock(&cachep->spinlock); |
1da177e4 LT |
3778 | return 0; |
3779 | } | |
3780 | ||
3781 | /* | |
3782 | * slabinfo_op - iterator that generates /proc/slabinfo | |
3783 | * | |
3784 | * Output layout: | |
3785 | * cache-name | |
3786 | * num-active-objs | |
3787 | * total-objs | |
3788 | * object size | |
3789 | * num-active-slabs | |
3790 | * total-slabs | |
3791 | * num-pages-per-slab | |
3792 | * + further values on SMP and with statistics enabled | |
3793 | */ | |
3794 | ||
3795 | struct seq_operations slabinfo_op = { | |
b28a02de PE |
3796 | .start = s_start, |
3797 | .next = s_next, | |
3798 | .stop = s_stop, | |
3799 | .show = s_show, | |
1da177e4 LT |
3800 | }; |
3801 | ||
3802 | #define MAX_SLABINFO_WRITE 128 | |
3803 | /** | |
3804 | * slabinfo_write - Tuning for the slab allocator | |
3805 | * @file: unused | |
3806 | * @buffer: user buffer | |
3807 | * @count: data length | |
3808 | * @ppos: unused | |
3809 | */ | |
b28a02de PE |
3810 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
3811 | size_t count, loff_t *ppos) | |
1da177e4 | 3812 | { |
b28a02de | 3813 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 LT |
3814 | int limit, batchcount, shared, res; |
3815 | struct list_head *p; | |
b28a02de | 3816 | |
1da177e4 LT |
3817 | if (count > MAX_SLABINFO_WRITE) |
3818 | return -EINVAL; | |
3819 | if (copy_from_user(&kbuf, buffer, count)) | |
3820 | return -EFAULT; | |
b28a02de | 3821 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
3822 | |
3823 | tmp = strchr(kbuf, ' '); | |
3824 | if (!tmp) | |
3825 | return -EINVAL; | |
3826 | *tmp = '\0'; | |
3827 | tmp++; | |
3828 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
3829 | return -EINVAL; | |
3830 | ||
3831 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 3832 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 3833 | res = -EINVAL; |
b28a02de | 3834 | list_for_each(p, &cache_chain) { |
343e0d7a PE |
3835 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, |
3836 | next); | |
1da177e4 LT |
3837 | |
3838 | if (!strcmp(cachep->name, kbuf)) { | |
3839 | if (limit < 1 || | |
3840 | batchcount < 1 || | |
b28a02de | 3841 | batchcount > limit || shared < 0) { |
e498be7d | 3842 | res = 0; |
1da177e4 | 3843 | } else { |
e498be7d | 3844 | res = do_tune_cpucache(cachep, limit, |
b28a02de | 3845 | batchcount, shared); |
1da177e4 LT |
3846 | } |
3847 | break; | |
3848 | } | |
3849 | } | |
fc0abb14 | 3850 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
3851 | if (res >= 0) |
3852 | res = count; | |
3853 | return res; | |
3854 | } | |
3855 | #endif | |
3856 | ||
00e145b6 MS |
3857 | /** |
3858 | * ksize - get the actual amount of memory allocated for a given object | |
3859 | * @objp: Pointer to the object | |
3860 | * | |
3861 | * kmalloc may internally round up allocations and return more memory | |
3862 | * than requested. ksize() can be used to determine the actual amount of | |
3863 | * memory allocated. The caller may use this additional memory, even though | |
3864 | * a smaller amount of memory was initially specified with the kmalloc call. | |
3865 | * The caller must guarantee that objp points to a valid object previously | |
3866 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
3867 | * must not be freed during the duration of the call. | |
3868 | */ | |
1da177e4 LT |
3869 | unsigned int ksize(const void *objp) |
3870 | { | |
00e145b6 MS |
3871 | if (unlikely(objp == NULL)) |
3872 | return 0; | |
1da177e4 | 3873 | |
6ed5eb22 | 3874 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 3875 | } |