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 | |
183ff22b | 29 | * slabs and you must pass objects with the same initializations to |
1da177e4 LT |
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 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
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 | ||
1da177e4 LT |
89 | #include <linux/slab.h> |
90 | #include <linux/mm.h> | |
c9cf5528 | 91 | #include <linux/poison.h> |
1da177e4 LT |
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> | |
101a5001 | 97 | #include <linux/cpuset.h> |
a0ec95a8 | 98 | #include <linux/proc_fs.h> |
1da177e4 LT |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> | |
101 | #include <linux/kallsyms.h> | |
102 | #include <linux/cpu.h> | |
103 | #include <linux/sysctl.h> | |
104 | #include <linux/module.h> | |
02af61bb | 105 | #include <linux/kmemtrace.h> |
1da177e4 | 106 | #include <linux/rcupdate.h> |
543537bd | 107 | #include <linux/string.h> |
138ae663 | 108 | #include <linux/uaccess.h> |
e498be7d | 109 | #include <linux/nodemask.h> |
d5cff635 | 110 | #include <linux/kmemleak.h> |
dc85da15 | 111 | #include <linux/mempolicy.h> |
fc0abb14 | 112 | #include <linux/mutex.h> |
8a8b6502 | 113 | #include <linux/fault-inject.h> |
e7eebaf6 | 114 | #include <linux/rtmutex.h> |
6a2d7a95 | 115 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 116 | #include <linux/debugobjects.h> |
c175eea4 | 117 | #include <linux/kmemcheck.h> |
1da177e4 | 118 | |
1da177e4 LT |
119 | #include <asm/cacheflush.h> |
120 | #include <asm/tlbflush.h> | |
121 | #include <asm/page.h> | |
122 | ||
123 | /* | |
50953fe9 | 124 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
125 | * 0 for faster, smaller code (especially in the critical paths). |
126 | * | |
127 | * STATS - 1 to collect stats for /proc/slabinfo. | |
128 | * 0 for faster, smaller code (especially in the critical paths). | |
129 | * | |
130 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
131 | */ | |
132 | ||
133 | #ifdef CONFIG_DEBUG_SLAB | |
134 | #define DEBUG 1 | |
135 | #define STATS 1 | |
136 | #define FORCED_DEBUG 1 | |
137 | #else | |
138 | #define DEBUG 0 | |
139 | #define STATS 0 | |
140 | #define FORCED_DEBUG 0 | |
141 | #endif | |
142 | ||
1da177e4 LT |
143 | /* Shouldn't this be in a header file somewhere? */ |
144 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 145 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 146 | |
1da177e4 LT |
147 | #ifndef ARCH_KMALLOC_MINALIGN |
148 | /* | |
149 | * Enforce a minimum alignment for the kmalloc caches. | |
150 | * Usually, the kmalloc caches are cache_line_size() aligned, except when | |
151 | * DEBUG and FORCED_DEBUG are enabled, then they are BYTES_PER_WORD aligned. | |
152 | * Some archs want to perform DMA into kmalloc caches and need a guaranteed | |
b46b8f19 DW |
153 | * alignment larger than the alignment of a 64-bit integer. |
154 | * ARCH_KMALLOC_MINALIGN allows that. | |
155 | * Note that increasing this value may disable some debug features. | |
1da177e4 | 156 | */ |
b46b8f19 | 157 | #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long) |
1da177e4 LT |
158 | #endif |
159 | ||
160 | #ifndef ARCH_SLAB_MINALIGN | |
161 | /* | |
162 | * Enforce a minimum alignment for all caches. | |
163 | * Intended for archs that get misalignment faults even for BYTES_PER_WORD | |
164 | * aligned buffers. Includes ARCH_KMALLOC_MINALIGN. | |
165 | * If possible: Do not enable this flag for CONFIG_DEBUG_SLAB, it disables | |
166 | * some debug features. | |
167 | */ | |
168 | #define ARCH_SLAB_MINALIGN 0 | |
169 | #endif | |
170 | ||
171 | #ifndef ARCH_KMALLOC_FLAGS | |
172 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
173 | #endif | |
174 | ||
175 | /* Legal flag mask for kmem_cache_create(). */ | |
176 | #if DEBUG | |
50953fe9 | 177 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 178 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 179 | SLAB_CACHE_DMA | \ |
5af60839 | 180 | SLAB_STORE_USER | \ |
1da177e4 | 181 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 182 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 183 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 | 184 | #else |
ac2b898c | 185 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 186 | SLAB_CACHE_DMA | \ |
1da177e4 | 187 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 188 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 189 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 LT |
190 | #endif |
191 | ||
192 | /* | |
193 | * kmem_bufctl_t: | |
194 | * | |
195 | * Bufctl's are used for linking objs within a slab | |
196 | * linked offsets. | |
197 | * | |
198 | * This implementation relies on "struct page" for locating the cache & | |
199 | * slab an object belongs to. | |
200 | * This allows the bufctl structure to be small (one int), but limits | |
201 | * the number of objects a slab (not a cache) can contain when off-slab | |
202 | * bufctls are used. The limit is the size of the largest general cache | |
203 | * that does not use off-slab slabs. | |
204 | * For 32bit archs with 4 kB pages, is this 56. | |
205 | * This is not serious, as it is only for large objects, when it is unwise | |
206 | * to have too many per slab. | |
207 | * Note: This limit can be raised by introducing a general cache whose size | |
208 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
209 | */ | |
210 | ||
fa5b08d5 | 211 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
212 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
213 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
214 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
215 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 216 | |
1da177e4 LT |
217 | /* |
218 | * struct slab | |
219 | * | |
220 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
221 | * for a slab, or allocated from an general cache. | |
222 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
223 | */ | |
224 | struct slab { | |
b28a02de PE |
225 | struct list_head list; |
226 | unsigned long colouroff; | |
227 | void *s_mem; /* including colour offset */ | |
228 | unsigned int inuse; /* num of objs active in slab */ | |
229 | kmem_bufctl_t free; | |
230 | unsigned short nodeid; | |
1da177e4 LT |
231 | }; |
232 | ||
233 | /* | |
234 | * struct slab_rcu | |
235 | * | |
236 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
237 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
238 | * we need to approach a kernel structure obliquely, from its address | |
239 | * obtained without the usual locking. We can lock the structure to | |
240 | * stabilize it and check it's still at the given address, only if we | |
241 | * can be sure that the memory has not been meanwhile reused for some | |
242 | * other kind of object (which our subsystem's lock might corrupt). | |
243 | * | |
244 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
245 | * taking the spinlock within the structure expected at that address. | |
246 | * | |
247 | * We assume struct slab_rcu can overlay struct slab when destroying. | |
248 | */ | |
249 | struct slab_rcu { | |
b28a02de | 250 | struct rcu_head head; |
343e0d7a | 251 | struct kmem_cache *cachep; |
b28a02de | 252 | void *addr; |
1da177e4 LT |
253 | }; |
254 | ||
255 | /* | |
256 | * struct array_cache | |
257 | * | |
1da177e4 LT |
258 | * Purpose: |
259 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
260 | * - reduce the number of linked list operations | |
261 | * - reduce spinlock operations | |
262 | * | |
263 | * The limit is stored in the per-cpu structure to reduce the data cache | |
264 | * footprint. | |
265 | * | |
266 | */ | |
267 | struct array_cache { | |
268 | unsigned int avail; | |
269 | unsigned int limit; | |
270 | unsigned int batchcount; | |
271 | unsigned int touched; | |
e498be7d | 272 | spinlock_t lock; |
bda5b655 | 273 | void *entry[]; /* |
a737b3e2 AM |
274 | * Must have this definition in here for the proper |
275 | * alignment of array_cache. Also simplifies accessing | |
276 | * the entries. | |
a737b3e2 | 277 | */ |
1da177e4 LT |
278 | }; |
279 | ||
a737b3e2 AM |
280 | /* |
281 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
282 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
283 | */ |
284 | #define BOOT_CPUCACHE_ENTRIES 1 | |
285 | struct arraycache_init { | |
286 | struct array_cache cache; | |
b28a02de | 287 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
288 | }; |
289 | ||
290 | /* | |
e498be7d | 291 | * The slab lists for all objects. |
1da177e4 LT |
292 | */ |
293 | struct kmem_list3 { | |
b28a02de PE |
294 | struct list_head slabs_partial; /* partial list first, better asm code */ |
295 | struct list_head slabs_full; | |
296 | struct list_head slabs_free; | |
297 | unsigned long free_objects; | |
b28a02de | 298 | unsigned int free_limit; |
2e1217cf | 299 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
300 | spinlock_t list_lock; |
301 | struct array_cache *shared; /* shared per node */ | |
302 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
303 | unsigned long next_reap; /* updated without locking */ |
304 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
305 | }; |
306 | ||
7e85ee0c PE |
307 | /* |
308 | * The slab allocator is initialized with interrupts disabled. Therefore, make | |
309 | * sure early boot allocations don't accidentally enable interrupts. | |
310 | */ | |
311 | static gfp_t slab_gfp_mask __read_mostly = SLAB_GFP_BOOT_MASK; | |
312 | ||
e498be7d CL |
313 | /* |
314 | * Need this for bootstrapping a per node allocator. | |
315 | */ | |
556a169d | 316 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
e498be7d CL |
317 | struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
318 | #define CACHE_CACHE 0 | |
556a169d PE |
319 | #define SIZE_AC MAX_NUMNODES |
320 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 321 | |
ed11d9eb CL |
322 | static int drain_freelist(struct kmem_cache *cache, |
323 | struct kmem_list3 *l3, int tofree); | |
324 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
325 | int node); | |
83b519e8 | 326 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 327 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 328 | |
e498be7d | 329 | /* |
a737b3e2 AM |
330 | * This function must be completely optimized away if a constant is passed to |
331 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 332 | */ |
7243cc05 | 333 | static __always_inline int index_of(const size_t size) |
e498be7d | 334 | { |
5ec8a847 SR |
335 | extern void __bad_size(void); |
336 | ||
e498be7d CL |
337 | if (__builtin_constant_p(size)) { |
338 | int i = 0; | |
339 | ||
340 | #define CACHE(x) \ | |
341 | if (size <=x) \ | |
342 | return i; \ | |
343 | else \ | |
344 | i++; | |
1c61fc40 | 345 | #include <linux/kmalloc_sizes.h> |
e498be7d | 346 | #undef CACHE |
5ec8a847 | 347 | __bad_size(); |
7243cc05 | 348 | } else |
5ec8a847 | 349 | __bad_size(); |
e498be7d CL |
350 | return 0; |
351 | } | |
352 | ||
e0a42726 IM |
353 | static int slab_early_init = 1; |
354 | ||
e498be7d CL |
355 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
356 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 357 | |
5295a74c | 358 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
359 | { |
360 | INIT_LIST_HEAD(&parent->slabs_full); | |
361 | INIT_LIST_HEAD(&parent->slabs_partial); | |
362 | INIT_LIST_HEAD(&parent->slabs_free); | |
363 | parent->shared = NULL; | |
364 | parent->alien = NULL; | |
2e1217cf | 365 | parent->colour_next = 0; |
e498be7d CL |
366 | spin_lock_init(&parent->list_lock); |
367 | parent->free_objects = 0; | |
368 | parent->free_touched = 0; | |
369 | } | |
370 | ||
a737b3e2 AM |
371 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
372 | do { \ | |
373 | INIT_LIST_HEAD(listp); \ | |
374 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
375 | } while (0) |
376 | ||
a737b3e2 AM |
377 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
378 | do { \ | |
e498be7d CL |
379 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
380 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
381 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
382 | } while (0) | |
1da177e4 | 383 | |
1da177e4 LT |
384 | #define CFLGS_OFF_SLAB (0x80000000UL) |
385 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
386 | ||
387 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
388 | /* |
389 | * Optimization question: fewer reaps means less probability for unnessary | |
390 | * cpucache drain/refill cycles. | |
1da177e4 | 391 | * |
dc6f3f27 | 392 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
393 | * which could lock up otherwise freeable slabs. |
394 | */ | |
395 | #define REAPTIMEOUT_CPUC (2*HZ) | |
396 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
397 | ||
398 | #if STATS | |
399 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
400 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
401 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
402 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 403 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
404 | #define STATS_SET_HIGH(x) \ |
405 | do { \ | |
406 | if ((x)->num_active > (x)->high_mark) \ | |
407 | (x)->high_mark = (x)->num_active; \ | |
408 | } while (0) | |
1da177e4 LT |
409 | #define STATS_INC_ERR(x) ((x)->errors++) |
410 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 411 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 412 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
413 | #define STATS_SET_FREEABLE(x, i) \ |
414 | do { \ | |
415 | if ((x)->max_freeable < i) \ | |
416 | (x)->max_freeable = i; \ | |
417 | } while (0) | |
1da177e4 LT |
418 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
419 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
420 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
421 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
422 | #else | |
423 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
424 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
425 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
426 | #define STATS_INC_GROWN(x) do { } while (0) | |
ed11d9eb | 427 | #define STATS_ADD_REAPED(x,y) do { } while (0) |
1da177e4 LT |
428 | #define STATS_SET_HIGH(x) do { } while (0) |
429 | #define STATS_INC_ERR(x) do { } while (0) | |
430 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 431 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 432 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 433 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
434 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
435 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
436 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
437 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
438 | #endif | |
439 | ||
440 | #if DEBUG | |
1da177e4 | 441 | |
a737b3e2 AM |
442 | /* |
443 | * memory layout of objects: | |
1da177e4 | 444 | * 0 : objp |
3dafccf2 | 445 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
446 | * the end of an object is aligned with the end of the real |
447 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 448 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 449 | * redzone word. |
3dafccf2 MS |
450 | * cachep->obj_offset: The real object. |
451 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
452 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
453 | * [BYTES_PER_WORD long] | |
1da177e4 | 454 | */ |
343e0d7a | 455 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 456 | { |
3dafccf2 | 457 | return cachep->obj_offset; |
1da177e4 LT |
458 | } |
459 | ||
343e0d7a | 460 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 461 | { |
3dafccf2 | 462 | return cachep->obj_size; |
1da177e4 LT |
463 | } |
464 | ||
b46b8f19 | 465 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
466 | { |
467 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
468 | return (unsigned long long*) (objp + obj_offset(cachep) - |
469 | sizeof(unsigned long long)); | |
1da177e4 LT |
470 | } |
471 | ||
b46b8f19 | 472 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
473 | { |
474 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
475 | if (cachep->flags & SLAB_STORE_USER) | |
b46b8f19 DW |
476 | return (unsigned long long *)(objp + cachep->buffer_size - |
477 | sizeof(unsigned long long) - | |
87a927c7 | 478 | REDZONE_ALIGN); |
b46b8f19 DW |
479 | return (unsigned long long *) (objp + cachep->buffer_size - |
480 | sizeof(unsigned long long)); | |
1da177e4 LT |
481 | } |
482 | ||
343e0d7a | 483 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
484 | { |
485 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 486 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
487 | } |
488 | ||
489 | #else | |
490 | ||
3dafccf2 MS |
491 | #define obj_offset(x) 0 |
492 | #define obj_size(cachep) (cachep->buffer_size) | |
b46b8f19 DW |
493 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
494 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
495 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
496 | ||
497 | #endif | |
498 | ||
36555751 EGM |
499 | #ifdef CONFIG_KMEMTRACE |
500 | size_t slab_buffer_size(struct kmem_cache *cachep) | |
501 | { | |
502 | return cachep->buffer_size; | |
503 | } | |
504 | EXPORT_SYMBOL(slab_buffer_size); | |
505 | #endif | |
506 | ||
1da177e4 LT |
507 | /* |
508 | * Do not go above this order unless 0 objects fit into the slab. | |
509 | */ | |
510 | #define BREAK_GFP_ORDER_HI 1 | |
511 | #define BREAK_GFP_ORDER_LO 0 | |
512 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
513 | ||
a737b3e2 AM |
514 | /* |
515 | * Functions for storing/retrieving the cachep and or slab from the page | |
516 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
517 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 518 | */ |
065d41cb PE |
519 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
520 | { | |
521 | page->lru.next = (struct list_head *)cache; | |
522 | } | |
523 | ||
524 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
525 | { | |
d85f3385 | 526 | page = compound_head(page); |
ddc2e812 | 527 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
528 | return (struct kmem_cache *)page->lru.next; |
529 | } | |
530 | ||
531 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
532 | { | |
533 | page->lru.prev = (struct list_head *)slab; | |
534 | } | |
535 | ||
536 | static inline struct slab *page_get_slab(struct page *page) | |
537 | { | |
ddc2e812 | 538 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
539 | return (struct slab *)page->lru.prev; |
540 | } | |
1da177e4 | 541 | |
6ed5eb22 PE |
542 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
543 | { | |
b49af68f | 544 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
545 | return page_get_cache(page); |
546 | } | |
547 | ||
548 | static inline struct slab *virt_to_slab(const void *obj) | |
549 | { | |
b49af68f | 550 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
551 | return page_get_slab(page); |
552 | } | |
553 | ||
8fea4e96 PE |
554 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
555 | unsigned int idx) | |
556 | { | |
557 | return slab->s_mem + cache->buffer_size * idx; | |
558 | } | |
559 | ||
6a2d7a95 ED |
560 | /* |
561 | * We want to avoid an expensive divide : (offset / cache->buffer_size) | |
562 | * Using the fact that buffer_size is a constant for a particular cache, | |
563 | * we can replace (offset / cache->buffer_size) by | |
564 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
565 | */ | |
566 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
567 | const struct slab *slab, void *obj) | |
8fea4e96 | 568 | { |
6a2d7a95 ED |
569 | u32 offset = (obj - slab->s_mem); |
570 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
571 | } |
572 | ||
a737b3e2 AM |
573 | /* |
574 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
575 | */ | |
1da177e4 LT |
576 | struct cache_sizes malloc_sizes[] = { |
577 | #define CACHE(x) { .cs_size = (x) }, | |
578 | #include <linux/kmalloc_sizes.h> | |
579 | CACHE(ULONG_MAX) | |
580 | #undef CACHE | |
581 | }; | |
582 | EXPORT_SYMBOL(malloc_sizes); | |
583 | ||
584 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
585 | struct cache_names { | |
586 | char *name; | |
587 | char *name_dma; | |
588 | }; | |
589 | ||
590 | static struct cache_names __initdata cache_names[] = { | |
591 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
592 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 593 | {NULL,} |
1da177e4 LT |
594 | #undef CACHE |
595 | }; | |
596 | ||
597 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 598 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 599 | static struct arraycache_init initarray_generic = |
b28a02de | 600 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
601 | |
602 | /* internal cache of cache description objs */ | |
343e0d7a | 603 | static struct kmem_cache cache_cache = { |
b28a02de PE |
604 | .batchcount = 1, |
605 | .limit = BOOT_CPUCACHE_ENTRIES, | |
606 | .shared = 1, | |
343e0d7a | 607 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 608 | .name = "kmem_cache", |
1da177e4 LT |
609 | }; |
610 | ||
056c6241 RT |
611 | #define BAD_ALIEN_MAGIC 0x01020304ul |
612 | ||
f1aaee53 AV |
613 | #ifdef CONFIG_LOCKDEP |
614 | ||
615 | /* | |
616 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
617 | * for other slabs "off slab". | |
618 | * The locking for this is tricky in that it nests within the locks | |
619 | * of all other slabs in a few places; to deal with this special | |
620 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
621 | * |
622 | * We set lock class for alien array caches which are up during init. | |
623 | * The lock annotation will be lost if all cpus of a node goes down and | |
624 | * then comes back up during hotplug | |
f1aaee53 | 625 | */ |
056c6241 RT |
626 | static struct lock_class_key on_slab_l3_key; |
627 | static struct lock_class_key on_slab_alc_key; | |
628 | ||
629 | static inline void init_lock_keys(void) | |
f1aaee53 | 630 | |
f1aaee53 AV |
631 | { |
632 | int q; | |
056c6241 RT |
633 | struct cache_sizes *s = malloc_sizes; |
634 | ||
635 | while (s->cs_size != ULONG_MAX) { | |
636 | for_each_node(q) { | |
637 | struct array_cache **alc; | |
638 | int r; | |
639 | struct kmem_list3 *l3 = s->cs_cachep->nodelists[q]; | |
640 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
641 | continue; | |
642 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); | |
643 | alc = l3->alien; | |
644 | /* | |
645 | * FIXME: This check for BAD_ALIEN_MAGIC | |
646 | * should go away when common slab code is taught to | |
647 | * work even without alien caches. | |
648 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
649 | * for alloc_alien_cache, | |
650 | */ | |
651 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
652 | continue; | |
653 | for_each_node(r) { | |
654 | if (alc[r]) | |
655 | lockdep_set_class(&alc[r]->lock, | |
656 | &on_slab_alc_key); | |
657 | } | |
658 | } | |
659 | s++; | |
f1aaee53 AV |
660 | } |
661 | } | |
f1aaee53 | 662 | #else |
056c6241 | 663 | static inline void init_lock_keys(void) |
f1aaee53 AV |
664 | { |
665 | } | |
666 | #endif | |
667 | ||
8f5be20b | 668 | /* |
95402b38 | 669 | * Guard access to the cache-chain. |
8f5be20b | 670 | */ |
fc0abb14 | 671 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
672 | static struct list_head cache_chain; |
673 | ||
1da177e4 LT |
674 | /* |
675 | * chicken and egg problem: delay the per-cpu array allocation | |
676 | * until the general caches are up. | |
677 | */ | |
678 | static enum { | |
679 | NONE, | |
e498be7d CL |
680 | PARTIAL_AC, |
681 | PARTIAL_L3, | |
8429db5c | 682 | EARLY, |
1da177e4 LT |
683 | FULL |
684 | } g_cpucache_up; | |
685 | ||
39d24e64 MK |
686 | /* |
687 | * used by boot code to determine if it can use slab based allocator | |
688 | */ | |
689 | int slab_is_available(void) | |
690 | { | |
8429db5c | 691 | return g_cpucache_up >= EARLY; |
39d24e64 MK |
692 | } |
693 | ||
52bad64d | 694 | static DEFINE_PER_CPU(struct delayed_work, reap_work); |
1da177e4 | 695 | |
343e0d7a | 696 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
697 | { |
698 | return cachep->array[smp_processor_id()]; | |
699 | } | |
700 | ||
a737b3e2 AM |
701 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
702 | gfp_t gfpflags) | |
1da177e4 LT |
703 | { |
704 | struct cache_sizes *csizep = malloc_sizes; | |
705 | ||
706 | #if DEBUG | |
707 | /* This happens if someone tries to call | |
b28a02de PE |
708 | * kmem_cache_create(), or __kmalloc(), before |
709 | * the generic caches are initialized. | |
710 | */ | |
c7e43c78 | 711 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 712 | #endif |
6cb8f913 CL |
713 | if (!size) |
714 | return ZERO_SIZE_PTR; | |
715 | ||
1da177e4 LT |
716 | while (size > csizep->cs_size) |
717 | csizep++; | |
718 | ||
719 | /* | |
0abf40c1 | 720 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
721 | * has cs_{dma,}cachep==NULL. Thus no special case |
722 | * for large kmalloc calls required. | |
723 | */ | |
4b51d669 | 724 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
725 | if (unlikely(gfpflags & GFP_DMA)) |
726 | return csizep->cs_dmacachep; | |
4b51d669 | 727 | #endif |
1da177e4 LT |
728 | return csizep->cs_cachep; |
729 | } | |
730 | ||
b221385b | 731 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
732 | { |
733 | return __find_general_cachep(size, gfpflags); | |
734 | } | |
97e2bde4 | 735 | |
fbaccacf | 736 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 737 | { |
fbaccacf SR |
738 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
739 | } | |
1da177e4 | 740 | |
a737b3e2 AM |
741 | /* |
742 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
743 | */ | |
fbaccacf SR |
744 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
745 | size_t align, int flags, size_t *left_over, | |
746 | unsigned int *num) | |
747 | { | |
748 | int nr_objs; | |
749 | size_t mgmt_size; | |
750 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 751 | |
fbaccacf SR |
752 | /* |
753 | * The slab management structure can be either off the slab or | |
754 | * on it. For the latter case, the memory allocated for a | |
755 | * slab is used for: | |
756 | * | |
757 | * - The struct slab | |
758 | * - One kmem_bufctl_t for each object | |
759 | * - Padding to respect alignment of @align | |
760 | * - @buffer_size bytes for each object | |
761 | * | |
762 | * If the slab management structure is off the slab, then the | |
763 | * alignment will already be calculated into the size. Because | |
764 | * the slabs are all pages aligned, the objects will be at the | |
765 | * correct alignment when allocated. | |
766 | */ | |
767 | if (flags & CFLGS_OFF_SLAB) { | |
768 | mgmt_size = 0; | |
769 | nr_objs = slab_size / buffer_size; | |
770 | ||
771 | if (nr_objs > SLAB_LIMIT) | |
772 | nr_objs = SLAB_LIMIT; | |
773 | } else { | |
774 | /* | |
775 | * Ignore padding for the initial guess. The padding | |
776 | * is at most @align-1 bytes, and @buffer_size is at | |
777 | * least @align. In the worst case, this result will | |
778 | * be one greater than the number of objects that fit | |
779 | * into the memory allocation when taking the padding | |
780 | * into account. | |
781 | */ | |
782 | nr_objs = (slab_size - sizeof(struct slab)) / | |
783 | (buffer_size + sizeof(kmem_bufctl_t)); | |
784 | ||
785 | /* | |
786 | * This calculated number will be either the right | |
787 | * amount, or one greater than what we want. | |
788 | */ | |
789 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
790 | > slab_size) | |
791 | nr_objs--; | |
792 | ||
793 | if (nr_objs > SLAB_LIMIT) | |
794 | nr_objs = SLAB_LIMIT; | |
795 | ||
796 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
797 | } | |
798 | *num = nr_objs; | |
799 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
800 | } |
801 | ||
d40cee24 | 802 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 803 | |
a737b3e2 AM |
804 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
805 | char *msg) | |
1da177e4 LT |
806 | { |
807 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 808 | function, cachep->name, msg); |
1da177e4 LT |
809 | dump_stack(); |
810 | } | |
811 | ||
3395ee05 PM |
812 | /* |
813 | * By default on NUMA we use alien caches to stage the freeing of | |
814 | * objects allocated from other nodes. This causes massive memory | |
815 | * inefficiencies when using fake NUMA setup to split memory into a | |
816 | * large number of small nodes, so it can be disabled on the command | |
817 | * line | |
818 | */ | |
819 | ||
820 | static int use_alien_caches __read_mostly = 1; | |
1807a1aa | 821 | static int numa_platform __read_mostly = 1; |
3395ee05 PM |
822 | static int __init noaliencache_setup(char *s) |
823 | { | |
824 | use_alien_caches = 0; | |
825 | return 1; | |
826 | } | |
827 | __setup("noaliencache", noaliencache_setup); | |
828 | ||
8fce4d8e CL |
829 | #ifdef CONFIG_NUMA |
830 | /* | |
831 | * Special reaping functions for NUMA systems called from cache_reap(). | |
832 | * These take care of doing round robin flushing of alien caches (containing | |
833 | * objects freed on different nodes from which they were allocated) and the | |
834 | * flushing of remote pcps by calling drain_node_pages. | |
835 | */ | |
836 | static DEFINE_PER_CPU(unsigned long, reap_node); | |
837 | ||
838 | static void init_reap_node(int cpu) | |
839 | { | |
840 | int node; | |
841 | ||
842 | node = next_node(cpu_to_node(cpu), node_online_map); | |
843 | if (node == MAX_NUMNODES) | |
442295c9 | 844 | node = first_node(node_online_map); |
8fce4d8e | 845 | |
7f6b8876 | 846 | per_cpu(reap_node, cpu) = node; |
8fce4d8e CL |
847 | } |
848 | ||
849 | static void next_reap_node(void) | |
850 | { | |
851 | int node = __get_cpu_var(reap_node); | |
852 | ||
8fce4d8e CL |
853 | node = next_node(node, node_online_map); |
854 | if (unlikely(node >= MAX_NUMNODES)) | |
855 | node = first_node(node_online_map); | |
856 | __get_cpu_var(reap_node) = node; | |
857 | } | |
858 | ||
859 | #else | |
860 | #define init_reap_node(cpu) do { } while (0) | |
861 | #define next_reap_node(void) do { } while (0) | |
862 | #endif | |
863 | ||
1da177e4 LT |
864 | /* |
865 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
866 | * via the workqueue/eventd. | |
867 | * Add the CPU number into the expiration time to minimize the possibility of | |
868 | * the CPUs getting into lockstep and contending for the global cache chain | |
869 | * lock. | |
870 | */ | |
897e679b | 871 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 872 | { |
52bad64d | 873 | struct delayed_work *reap_work = &per_cpu(reap_work, cpu); |
1da177e4 LT |
874 | |
875 | /* | |
876 | * When this gets called from do_initcalls via cpucache_init(), | |
877 | * init_workqueues() has already run, so keventd will be setup | |
878 | * at that time. | |
879 | */ | |
52bad64d | 880 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 881 | init_reap_node(cpu); |
65f27f38 | 882 | INIT_DELAYED_WORK(reap_work, cache_reap); |
2b284214 AV |
883 | schedule_delayed_work_on(cpu, reap_work, |
884 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
885 | } |
886 | } | |
887 | ||
e498be7d | 888 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 889 | int batchcount, gfp_t gfp) |
1da177e4 | 890 | { |
b28a02de | 891 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
892 | struct array_cache *nc = NULL; |
893 | ||
83b519e8 | 894 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
895 | /* |
896 | * The array_cache structures contain pointers to free object. | |
897 | * However, when such objects are allocated or transfered to another | |
898 | * cache the pointers are not cleared and they could be counted as | |
899 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
900 | * not scan such objects. | |
901 | */ | |
902 | kmemleak_no_scan(nc); | |
1da177e4 LT |
903 | if (nc) { |
904 | nc->avail = 0; | |
905 | nc->limit = entries; | |
906 | nc->batchcount = batchcount; | |
907 | nc->touched = 0; | |
e498be7d | 908 | spin_lock_init(&nc->lock); |
1da177e4 LT |
909 | } |
910 | return nc; | |
911 | } | |
912 | ||
3ded175a CL |
913 | /* |
914 | * Transfer objects in one arraycache to another. | |
915 | * Locking must be handled by the caller. | |
916 | * | |
917 | * Return the number of entries transferred. | |
918 | */ | |
919 | static int transfer_objects(struct array_cache *to, | |
920 | struct array_cache *from, unsigned int max) | |
921 | { | |
922 | /* Figure out how many entries to transfer */ | |
923 | int nr = min(min(from->avail, max), to->limit - to->avail); | |
924 | ||
925 | if (!nr) | |
926 | return 0; | |
927 | ||
928 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
929 | sizeof(void *) *nr); | |
930 | ||
931 | from->avail -= nr; | |
932 | to->avail += nr; | |
933 | to->touched = 1; | |
934 | return nr; | |
935 | } | |
936 | ||
765c4507 CL |
937 | #ifndef CONFIG_NUMA |
938 | ||
939 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
940 | #define reap_alien(cachep, l3) do { } while (0) | |
941 | ||
83b519e8 | 942 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
943 | { |
944 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
945 | } | |
946 | ||
947 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
948 | { | |
949 | } | |
950 | ||
951 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
952 | { | |
953 | return 0; | |
954 | } | |
955 | ||
956 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
957 | gfp_t flags) | |
958 | { | |
959 | return NULL; | |
960 | } | |
961 | ||
8b98c169 | 962 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
963 | gfp_t flags, int nodeid) |
964 | { | |
965 | return NULL; | |
966 | } | |
967 | ||
968 | #else /* CONFIG_NUMA */ | |
969 | ||
8b98c169 | 970 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 971 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 972 | |
83b519e8 | 973 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
974 | { |
975 | struct array_cache **ac_ptr; | |
8ef82866 | 976 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
977 | int i; |
978 | ||
979 | if (limit > 1) | |
980 | limit = 12; | |
83b519e8 | 981 | ac_ptr = kmalloc_node(memsize, gfp, node); |
e498be7d CL |
982 | if (ac_ptr) { |
983 | for_each_node(i) { | |
984 | if (i == node || !node_online(i)) { | |
985 | ac_ptr[i] = NULL; | |
986 | continue; | |
987 | } | |
83b519e8 | 988 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 989 | if (!ac_ptr[i]) { |
cc550def | 990 | for (i--; i >= 0; i--) |
e498be7d CL |
991 | kfree(ac_ptr[i]); |
992 | kfree(ac_ptr); | |
993 | return NULL; | |
994 | } | |
995 | } | |
996 | } | |
997 | return ac_ptr; | |
998 | } | |
999 | ||
5295a74c | 1000 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1001 | { |
1002 | int i; | |
1003 | ||
1004 | if (!ac_ptr) | |
1005 | return; | |
e498be7d | 1006 | for_each_node(i) |
b28a02de | 1007 | kfree(ac_ptr[i]); |
e498be7d CL |
1008 | kfree(ac_ptr); |
1009 | } | |
1010 | ||
343e0d7a | 1011 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1012 | struct array_cache *ac, int node) |
e498be7d CL |
1013 | { |
1014 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1015 | ||
1016 | if (ac->avail) { | |
1017 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1018 | /* |
1019 | * Stuff objects into the remote nodes shared array first. | |
1020 | * That way we could avoid the overhead of putting the objects | |
1021 | * into the free lists and getting them back later. | |
1022 | */ | |
693f7d36 | 1023 | if (rl3->shared) |
1024 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1025 | |
ff69416e | 1026 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1027 | ac->avail = 0; |
1028 | spin_unlock(&rl3->list_lock); | |
1029 | } | |
1030 | } | |
1031 | ||
8fce4d8e CL |
1032 | /* |
1033 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1034 | */ | |
1035 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1036 | { | |
1037 | int node = __get_cpu_var(reap_node); | |
1038 | ||
1039 | if (l3->alien) { | |
1040 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1041 | |
1042 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1043 | __drain_alien_cache(cachep, ac, node); |
1044 | spin_unlock_irq(&ac->lock); | |
1045 | } | |
1046 | } | |
1047 | } | |
1048 | ||
a737b3e2 AM |
1049 | static void drain_alien_cache(struct kmem_cache *cachep, |
1050 | struct array_cache **alien) | |
e498be7d | 1051 | { |
b28a02de | 1052 | int i = 0; |
e498be7d CL |
1053 | struct array_cache *ac; |
1054 | unsigned long flags; | |
1055 | ||
1056 | for_each_online_node(i) { | |
4484ebf1 | 1057 | ac = alien[i]; |
e498be7d CL |
1058 | if (ac) { |
1059 | spin_lock_irqsave(&ac->lock, flags); | |
1060 | __drain_alien_cache(cachep, ac, i); | |
1061 | spin_unlock_irqrestore(&ac->lock, flags); | |
1062 | } | |
1063 | } | |
1064 | } | |
729bd0b7 | 1065 | |
873623df | 1066 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1067 | { |
1068 | struct slab *slabp = virt_to_slab(objp); | |
1069 | int nodeid = slabp->nodeid; | |
1070 | struct kmem_list3 *l3; | |
1071 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1072 | int node; |
1073 | ||
1074 | node = numa_node_id(); | |
729bd0b7 PE |
1075 | |
1076 | /* | |
1077 | * Make sure we are not freeing a object from another node to the array | |
1078 | * cache on this cpu. | |
1079 | */ | |
62918a03 | 1080 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1081 | return 0; |
1082 | ||
1ca4cb24 | 1083 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1084 | STATS_INC_NODEFREES(cachep); |
1085 | if (l3->alien && l3->alien[nodeid]) { | |
1086 | alien = l3->alien[nodeid]; | |
873623df | 1087 | spin_lock(&alien->lock); |
729bd0b7 PE |
1088 | if (unlikely(alien->avail == alien->limit)) { |
1089 | STATS_INC_ACOVERFLOW(cachep); | |
1090 | __drain_alien_cache(cachep, alien, nodeid); | |
1091 | } | |
1092 | alien->entry[alien->avail++] = objp; | |
1093 | spin_unlock(&alien->lock); | |
1094 | } else { | |
1095 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1096 | free_block(cachep, &objp, 1, nodeid); | |
1097 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1098 | } | |
1099 | return 1; | |
1100 | } | |
e498be7d CL |
1101 | #endif |
1102 | ||
fbf1e473 AM |
1103 | static void __cpuinit cpuup_canceled(long cpu) |
1104 | { | |
1105 | struct kmem_cache *cachep; | |
1106 | struct kmem_list3 *l3 = NULL; | |
1107 | int node = cpu_to_node(cpu); | |
a70f7302 | 1108 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 AM |
1109 | |
1110 | list_for_each_entry(cachep, &cache_chain, next) { | |
1111 | struct array_cache *nc; | |
1112 | struct array_cache *shared; | |
1113 | struct array_cache **alien; | |
fbf1e473 | 1114 | |
fbf1e473 AM |
1115 | /* cpu is dead; no one can alloc from it. */ |
1116 | nc = cachep->array[cpu]; | |
1117 | cachep->array[cpu] = NULL; | |
1118 | l3 = cachep->nodelists[node]; | |
1119 | ||
1120 | if (!l3) | |
1121 | goto free_array_cache; | |
1122 | ||
1123 | spin_lock_irq(&l3->list_lock); | |
1124 | ||
1125 | /* Free limit for this kmem_list3 */ | |
1126 | l3->free_limit -= cachep->batchcount; | |
1127 | if (nc) | |
1128 | free_block(cachep, nc->entry, nc->avail, node); | |
1129 | ||
c5f59f08 | 1130 | if (!cpus_empty(*mask)) { |
fbf1e473 AM |
1131 | spin_unlock_irq(&l3->list_lock); |
1132 | goto free_array_cache; | |
1133 | } | |
1134 | ||
1135 | shared = l3->shared; | |
1136 | if (shared) { | |
1137 | free_block(cachep, shared->entry, | |
1138 | shared->avail, node); | |
1139 | l3->shared = NULL; | |
1140 | } | |
1141 | ||
1142 | alien = l3->alien; | |
1143 | l3->alien = NULL; | |
1144 | ||
1145 | spin_unlock_irq(&l3->list_lock); | |
1146 | ||
1147 | kfree(shared); | |
1148 | if (alien) { | |
1149 | drain_alien_cache(cachep, alien); | |
1150 | free_alien_cache(alien); | |
1151 | } | |
1152 | free_array_cache: | |
1153 | kfree(nc); | |
1154 | } | |
1155 | /* | |
1156 | * In the previous loop, all the objects were freed to | |
1157 | * the respective cache's slabs, now we can go ahead and | |
1158 | * shrink each nodelist to its limit. | |
1159 | */ | |
1160 | list_for_each_entry(cachep, &cache_chain, next) { | |
1161 | l3 = cachep->nodelists[node]; | |
1162 | if (!l3) | |
1163 | continue; | |
1164 | drain_freelist(cachep, l3, l3->free_objects); | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1169 | { |
343e0d7a | 1170 | struct kmem_cache *cachep; |
e498be7d CL |
1171 | struct kmem_list3 *l3 = NULL; |
1172 | int node = cpu_to_node(cpu); | |
ea02e3dd | 1173 | const int memsize = sizeof(struct kmem_list3); |
1da177e4 | 1174 | |
fbf1e473 AM |
1175 | /* |
1176 | * We need to do this right in the beginning since | |
1177 | * alloc_arraycache's are going to use this list. | |
1178 | * kmalloc_node allows us to add the slab to the right | |
1179 | * kmem_list3 and not this cpu's kmem_list3 | |
1180 | */ | |
1181 | ||
1182 | list_for_each_entry(cachep, &cache_chain, next) { | |
a737b3e2 | 1183 | /* |
fbf1e473 AM |
1184 | * Set up the size64 kmemlist for cpu before we can |
1185 | * begin anything. Make sure some other cpu on this | |
1186 | * node has not already allocated this | |
e498be7d | 1187 | */ |
fbf1e473 AM |
1188 | if (!cachep->nodelists[node]) { |
1189 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1190 | if (!l3) | |
1191 | goto bad; | |
1192 | kmem_list3_init(l3); | |
1193 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1194 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
e498be7d | 1195 | |
a737b3e2 | 1196 | /* |
fbf1e473 AM |
1197 | * The l3s don't come and go as CPUs come and |
1198 | * go. cache_chain_mutex is sufficient | |
1199 | * protection here. | |
e498be7d | 1200 | */ |
fbf1e473 | 1201 | cachep->nodelists[node] = l3; |
e498be7d CL |
1202 | } |
1203 | ||
fbf1e473 AM |
1204 | spin_lock_irq(&cachep->nodelists[node]->list_lock); |
1205 | cachep->nodelists[node]->free_limit = | |
1206 | (1 + nr_cpus_node(node)) * | |
1207 | cachep->batchcount + cachep->num; | |
1208 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1209 | } | |
1210 | ||
1211 | /* | |
1212 | * Now we can go ahead with allocating the shared arrays and | |
1213 | * array caches | |
1214 | */ | |
1215 | list_for_each_entry(cachep, &cache_chain, next) { | |
1216 | struct array_cache *nc; | |
1217 | struct array_cache *shared = NULL; | |
1218 | struct array_cache **alien = NULL; | |
1219 | ||
1220 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1221 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1222 | if (!nc) |
1223 | goto bad; | |
1224 | if (cachep->shared) { | |
1225 | shared = alloc_arraycache(node, | |
1226 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1227 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1228 | if (!shared) { |
1229 | kfree(nc); | |
1da177e4 | 1230 | goto bad; |
12d00f6a | 1231 | } |
fbf1e473 AM |
1232 | } |
1233 | if (use_alien_caches) { | |
83b519e8 | 1234 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1235 | if (!alien) { |
1236 | kfree(shared); | |
1237 | kfree(nc); | |
fbf1e473 | 1238 | goto bad; |
12d00f6a | 1239 | } |
fbf1e473 AM |
1240 | } |
1241 | cachep->array[cpu] = nc; | |
1242 | l3 = cachep->nodelists[node]; | |
1243 | BUG_ON(!l3); | |
1244 | ||
1245 | spin_lock_irq(&l3->list_lock); | |
1246 | if (!l3->shared) { | |
1247 | /* | |
1248 | * We are serialised from CPU_DEAD or | |
1249 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1250 | */ | |
1251 | l3->shared = shared; | |
1252 | shared = NULL; | |
1253 | } | |
4484ebf1 | 1254 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1255 | if (!l3->alien) { |
1256 | l3->alien = alien; | |
1257 | alien = NULL; | |
1da177e4 | 1258 | } |
fbf1e473 AM |
1259 | #endif |
1260 | spin_unlock_irq(&l3->list_lock); | |
1261 | kfree(shared); | |
1262 | free_alien_cache(alien); | |
1263 | } | |
1264 | return 0; | |
1265 | bad: | |
12d00f6a | 1266 | cpuup_canceled(cpu); |
fbf1e473 AM |
1267 | return -ENOMEM; |
1268 | } | |
1269 | ||
1270 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1271 | unsigned long action, void *hcpu) | |
1272 | { | |
1273 | long cpu = (long)hcpu; | |
1274 | int err = 0; | |
1275 | ||
1276 | switch (action) { | |
fbf1e473 AM |
1277 | case CPU_UP_PREPARE: |
1278 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1279 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1280 | err = cpuup_prepare(cpu); |
95402b38 | 1281 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1282 | break; |
1283 | case CPU_ONLINE: | |
8bb78442 | 1284 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1285 | start_cpu_timer(cpu); |
1286 | break; | |
1287 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1288 | case CPU_DOWN_PREPARE: |
8bb78442 | 1289 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1290 | /* |
1291 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1292 | * held so that if cache_reap() is invoked it cannot do | |
1293 | * anything expensive but will only modify reap_work | |
1294 | * and reschedule the timer. | |
1295 | */ | |
1296 | cancel_rearming_delayed_work(&per_cpu(reap_work, cpu)); | |
1297 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1298 | per_cpu(reap_work, cpu).work.func = NULL; | |
1299 | break; | |
1300 | case CPU_DOWN_FAILED: | |
8bb78442 | 1301 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1302 | start_cpu_timer(cpu); |
1303 | break; | |
1da177e4 | 1304 | case CPU_DEAD: |
8bb78442 | 1305 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1306 | /* |
1307 | * Even if all the cpus of a node are down, we don't free the | |
1308 | * kmem_list3 of any cache. This to avoid a race between | |
1309 | * cpu_down, and a kmalloc allocation from another cpu for | |
1310 | * memory from the node of the cpu going down. The list3 | |
1311 | * structure is usually allocated from kmem_cache_create() and | |
1312 | * gets destroyed at kmem_cache_destroy(). | |
1313 | */ | |
183ff22b | 1314 | /* fall through */ |
8f5be20b | 1315 | #endif |
1da177e4 | 1316 | case CPU_UP_CANCELED: |
8bb78442 | 1317 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1318 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1319 | cpuup_canceled(cpu); |
fc0abb14 | 1320 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1321 | break; |
1da177e4 | 1322 | } |
fbf1e473 | 1323 | return err ? NOTIFY_BAD : NOTIFY_OK; |
1da177e4 LT |
1324 | } |
1325 | ||
74b85f37 CS |
1326 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1327 | &cpuup_callback, NULL, 0 | |
1328 | }; | |
1da177e4 | 1329 | |
e498be7d CL |
1330 | /* |
1331 | * swap the static kmem_list3 with kmalloced memory | |
1332 | */ | |
a737b3e2 AM |
1333 | static void init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1334 | int nodeid) | |
e498be7d CL |
1335 | { |
1336 | struct kmem_list3 *ptr; | |
1337 | ||
83b519e8 | 1338 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1339 | BUG_ON(!ptr); |
1340 | ||
e498be7d | 1341 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1342 | /* |
1343 | * Do not assume that spinlocks can be initialized via memcpy: | |
1344 | */ | |
1345 | spin_lock_init(&ptr->list_lock); | |
1346 | ||
e498be7d CL |
1347 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1348 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1349 | } |
1350 | ||
556a169d PE |
1351 | /* |
1352 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1353 | * size of kmem_list3. | |
1354 | */ | |
1355 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1356 | { | |
1357 | int node; | |
1358 | ||
1359 | for_each_online_node(node) { | |
1360 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1361 | cachep->nodelists[node]->next_reap = jiffies + | |
1362 | REAPTIMEOUT_LIST3 + | |
1363 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1364 | } | |
1365 | } | |
1366 | ||
a737b3e2 AM |
1367 | /* |
1368 | * Initialisation. Called after the page allocator have been initialised and | |
1369 | * before smp_init(). | |
1da177e4 LT |
1370 | */ |
1371 | void __init kmem_cache_init(void) | |
1372 | { | |
1373 | size_t left_over; | |
1374 | struct cache_sizes *sizes; | |
1375 | struct cache_names *names; | |
e498be7d | 1376 | int i; |
07ed76b2 | 1377 | int order; |
1ca4cb24 | 1378 | int node; |
e498be7d | 1379 | |
1807a1aa | 1380 | if (num_possible_nodes() == 1) { |
62918a03 | 1381 | use_alien_caches = 0; |
1807a1aa SS |
1382 | numa_platform = 0; |
1383 | } | |
62918a03 | 1384 | |
e498be7d CL |
1385 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1386 | kmem_list3_init(&initkmem_list3[i]); | |
1387 | if (i < MAX_NUMNODES) | |
1388 | cache_cache.nodelists[i] = NULL; | |
1389 | } | |
556a169d | 1390 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1391 | |
1392 | /* | |
1393 | * Fragmentation resistance on low memory - only use bigger | |
1394 | * page orders on machines with more than 32MB of memory. | |
1395 | */ | |
1396 | if (num_physpages > (32 << 20) >> PAGE_SHIFT) | |
1397 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; | |
1398 | ||
1da177e4 LT |
1399 | /* Bootstrap is tricky, because several objects are allocated |
1400 | * from caches that do not exist yet: | |
a737b3e2 AM |
1401 | * 1) initialize the cache_cache cache: it contains the struct |
1402 | * kmem_cache structures of all caches, except cache_cache itself: | |
1403 | * cache_cache is statically allocated. | |
e498be7d CL |
1404 | * Initially an __init data area is used for the head array and the |
1405 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1406 | * array at the end of the bootstrap. | |
1da177e4 | 1407 | * 2) Create the first kmalloc cache. |
343e0d7a | 1408 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1409 | * An __init data area is used for the head array. |
1410 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1411 | * head arrays. | |
1da177e4 LT |
1412 | * 4) Replace the __init data head arrays for cache_cache and the first |
1413 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1414 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1415 | * the other cache's with kmalloc allocated memory. | |
1416 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1417 | */ |
1418 | ||
1ca4cb24 PE |
1419 | node = numa_node_id(); |
1420 | ||
1da177e4 | 1421 | /* 1) create the cache_cache */ |
1da177e4 LT |
1422 | INIT_LIST_HEAD(&cache_chain); |
1423 | list_add(&cache_cache.next, &cache_chain); | |
1424 | cache_cache.colour_off = cache_line_size(); | |
1425 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1426 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1427 | |
8da3430d ED |
1428 | /* |
1429 | * struct kmem_cache size depends on nr_node_ids, which | |
1430 | * can be less than MAX_NUMNODES. | |
1431 | */ | |
1432 | cache_cache.buffer_size = offsetof(struct kmem_cache, nodelists) + | |
1433 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1434 | #if DEBUG | |
1435 | cache_cache.obj_size = cache_cache.buffer_size; | |
1436 | #endif | |
a737b3e2 AM |
1437 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1438 | cache_line_size()); | |
6a2d7a95 ED |
1439 | cache_cache.reciprocal_buffer_size = |
1440 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1441 | |
07ed76b2 JS |
1442 | for (order = 0; order < MAX_ORDER; order++) { |
1443 | cache_estimate(order, cache_cache.buffer_size, | |
1444 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1445 | if (cache_cache.num) | |
1446 | break; | |
1447 | } | |
40094fa6 | 1448 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1449 | cache_cache.gfporder = order; |
b28a02de | 1450 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1451 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1452 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1453 | |
1454 | /* 2+3) create the kmalloc caches */ | |
1455 | sizes = malloc_sizes; | |
1456 | names = cache_names; | |
1457 | ||
a737b3e2 AM |
1458 | /* |
1459 | * Initialize the caches that provide memory for the array cache and the | |
1460 | * kmem_list3 structures first. Without this, further allocations will | |
1461 | * bug. | |
e498be7d CL |
1462 | */ |
1463 | ||
1464 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1465 | sizes[INDEX_AC].cs_size, |
1466 | ARCH_KMALLOC_MINALIGN, | |
1467 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1468 | NULL); |
e498be7d | 1469 | |
a737b3e2 | 1470 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1471 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1472 | kmem_cache_create(names[INDEX_L3].name, |
1473 | sizes[INDEX_L3].cs_size, | |
1474 | ARCH_KMALLOC_MINALIGN, | |
1475 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1476 | NULL); |
a737b3e2 | 1477 | } |
e498be7d | 1478 | |
e0a42726 IM |
1479 | slab_early_init = 0; |
1480 | ||
1da177e4 | 1481 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1482 | /* |
1483 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1484 | * This should be particularly beneficial on SMP boxes, as it |
1485 | * eliminates "false sharing". | |
1486 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1487 | * allow tighter packing of the smaller caches. |
1488 | */ | |
a737b3e2 | 1489 | if (!sizes->cs_cachep) { |
e498be7d | 1490 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1491 | sizes->cs_size, |
1492 | ARCH_KMALLOC_MINALIGN, | |
1493 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1494 | NULL); |
a737b3e2 | 1495 | } |
4b51d669 CL |
1496 | #ifdef CONFIG_ZONE_DMA |
1497 | sizes->cs_dmacachep = kmem_cache_create( | |
1498 | names->name_dma, | |
a737b3e2 AM |
1499 | sizes->cs_size, |
1500 | ARCH_KMALLOC_MINALIGN, | |
1501 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1502 | SLAB_PANIC, | |
20c2df83 | 1503 | NULL); |
4b51d669 | 1504 | #endif |
1da177e4 LT |
1505 | sizes++; |
1506 | names++; | |
1507 | } | |
1508 | /* 4) Replace the bootstrap head arrays */ | |
1509 | { | |
2b2d5493 | 1510 | struct array_cache *ptr; |
e498be7d | 1511 | |
83b519e8 | 1512 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1513 | |
9a2dba4b PE |
1514 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1515 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1516 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1517 | /* |
1518 | * Do not assume that spinlocks can be initialized via memcpy: | |
1519 | */ | |
1520 | spin_lock_init(&ptr->lock); | |
1521 | ||
1da177e4 | 1522 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1523 | |
83b519e8 | 1524 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1525 | |
9a2dba4b | 1526 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1527 | != &initarray_generic.cache); |
9a2dba4b | 1528 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1529 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1530 | /* |
1531 | * Do not assume that spinlocks can be initialized via memcpy: | |
1532 | */ | |
1533 | spin_lock_init(&ptr->lock); | |
1534 | ||
e498be7d | 1535 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1536 | ptr; |
1da177e4 | 1537 | } |
e498be7d CL |
1538 | /* 5) Replace the bootstrap kmem_list3's */ |
1539 | { | |
1ca4cb24 PE |
1540 | int nid; |
1541 | ||
9c09a95c | 1542 | for_each_online_node(nid) { |
ec1f5eee | 1543 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1544 | |
e498be7d | 1545 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1546 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1547 | |
1548 | if (INDEX_AC != INDEX_L3) { | |
1549 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1550 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1551 | } |
1552 | } | |
1553 | } | |
1da177e4 | 1554 | |
8429db5c | 1555 | g_cpucache_up = EARLY; |
1da177e4 | 1556 | |
056c6241 RT |
1557 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1558 | init_lock_keys(); | |
8429db5c PE |
1559 | } |
1560 | ||
1561 | void __init kmem_cache_init_late(void) | |
1562 | { | |
1563 | struct kmem_cache *cachep; | |
1564 | ||
1565 | /* | |
1566 | * Interrupts are enabled now so all GFP allocations are safe. | |
1567 | */ | |
1568 | slab_gfp_mask = __GFP_BITS_MASK; | |
056c6241 | 1569 | |
8429db5c PE |
1570 | /* 6) resize the head arrays to their final sizes */ |
1571 | mutex_lock(&cache_chain_mutex); | |
1572 | list_for_each_entry(cachep, &cache_chain, next) | |
1573 | if (enable_cpucache(cachep, GFP_NOWAIT)) | |
1574 | BUG(); | |
1575 | mutex_unlock(&cache_chain_mutex); | |
056c6241 | 1576 | |
1da177e4 LT |
1577 | /* Done! */ |
1578 | g_cpucache_up = FULL; | |
1579 | ||
a737b3e2 AM |
1580 | /* |
1581 | * Register a cpu startup notifier callback that initializes | |
1582 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1583 | */ |
1584 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1585 | |
a737b3e2 AM |
1586 | /* |
1587 | * The reap timers are started later, with a module init call: That part | |
1588 | * of the kernel is not yet operational. | |
1da177e4 LT |
1589 | */ |
1590 | } | |
1591 | ||
1592 | static int __init cpucache_init(void) | |
1593 | { | |
1594 | int cpu; | |
1595 | ||
a737b3e2 AM |
1596 | /* |
1597 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1598 | */ |
e498be7d | 1599 | for_each_online_cpu(cpu) |
a737b3e2 | 1600 | start_cpu_timer(cpu); |
1da177e4 LT |
1601 | return 0; |
1602 | } | |
1da177e4 LT |
1603 | __initcall(cpucache_init); |
1604 | ||
1605 | /* | |
1606 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1607 | * | |
1608 | * If we requested dmaable memory, we will get it. Even if we | |
1609 | * did not request dmaable memory, we might get it, but that | |
1610 | * would be relatively rare and ignorable. | |
1611 | */ | |
343e0d7a | 1612 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1613 | { |
1614 | struct page *page; | |
e1b6aa6f | 1615 | int nr_pages; |
1da177e4 LT |
1616 | int i; |
1617 | ||
d6fef9da | 1618 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1619 | /* |
1620 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1621 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1622 | */ |
e1b6aa6f | 1623 | flags |= __GFP_COMP; |
d6fef9da | 1624 | #endif |
765c4507 | 1625 | |
3c517a61 | 1626 | flags |= cachep->gfpflags; |
e12ba74d MG |
1627 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1628 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1629 | |
b1eeab67 | 1630 | page = alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
1da177e4 LT |
1631 | if (!page) |
1632 | return NULL; | |
1da177e4 | 1633 | |
e1b6aa6f | 1634 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1635 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1636 | add_zone_page_state(page_zone(page), |
1637 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1638 | else | |
1639 | add_zone_page_state(page_zone(page), | |
1640 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1641 | for (i = 0; i < nr_pages; i++) |
1642 | __SetPageSlab(page + i); | |
c175eea4 | 1643 | |
b1eeab67 VN |
1644 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1645 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1646 | ||
1647 | if (cachep->ctor) | |
1648 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1649 | else | |
1650 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1651 | } | |
c175eea4 | 1652 | |
e1b6aa6f | 1653 | return page_address(page); |
1da177e4 LT |
1654 | } |
1655 | ||
1656 | /* | |
1657 | * Interface to system's page release. | |
1658 | */ | |
343e0d7a | 1659 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1660 | { |
b28a02de | 1661 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1662 | struct page *page = virt_to_page(addr); |
1663 | const unsigned long nr_freed = i; | |
1664 | ||
b1eeab67 | 1665 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1666 | |
972d1a7b CL |
1667 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1668 | sub_zone_page_state(page_zone(page), | |
1669 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1670 | else | |
1671 | sub_zone_page_state(page_zone(page), | |
1672 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1673 | while (i--) { |
f205b2fe NP |
1674 | BUG_ON(!PageSlab(page)); |
1675 | __ClearPageSlab(page); | |
1da177e4 LT |
1676 | page++; |
1677 | } | |
1da177e4 LT |
1678 | if (current->reclaim_state) |
1679 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1680 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1681 | } |
1682 | ||
1683 | static void kmem_rcu_free(struct rcu_head *head) | |
1684 | { | |
b28a02de | 1685 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1686 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1687 | |
1688 | kmem_freepages(cachep, slab_rcu->addr); | |
1689 | if (OFF_SLAB(cachep)) | |
1690 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1691 | } | |
1692 | ||
1693 | #if DEBUG | |
1694 | ||
1695 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1696 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1697 | unsigned long caller) |
1da177e4 | 1698 | { |
3dafccf2 | 1699 | int size = obj_size(cachep); |
1da177e4 | 1700 | |
3dafccf2 | 1701 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1702 | |
b28a02de | 1703 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1704 | return; |
1705 | ||
b28a02de PE |
1706 | *addr++ = 0x12345678; |
1707 | *addr++ = caller; | |
1708 | *addr++ = smp_processor_id(); | |
1709 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1710 | { |
1711 | unsigned long *sptr = &caller; | |
1712 | unsigned long svalue; | |
1713 | ||
1714 | while (!kstack_end(sptr)) { | |
1715 | svalue = *sptr++; | |
1716 | if (kernel_text_address(svalue)) { | |
b28a02de | 1717 | *addr++ = svalue; |
1da177e4 LT |
1718 | size -= sizeof(unsigned long); |
1719 | if (size <= sizeof(unsigned long)) | |
1720 | break; | |
1721 | } | |
1722 | } | |
1723 | ||
1724 | } | |
b28a02de | 1725 | *addr++ = 0x87654321; |
1da177e4 LT |
1726 | } |
1727 | #endif | |
1728 | ||
343e0d7a | 1729 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1730 | { |
3dafccf2 MS |
1731 | int size = obj_size(cachep); |
1732 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1733 | |
1734 | memset(addr, val, size); | |
b28a02de | 1735 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1736 | } |
1737 | ||
1738 | static void dump_line(char *data, int offset, int limit) | |
1739 | { | |
1740 | int i; | |
aa83aa40 DJ |
1741 | unsigned char error = 0; |
1742 | int bad_count = 0; | |
1743 | ||
1da177e4 | 1744 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1745 | for (i = 0; i < limit; i++) { |
1746 | if (data[offset + i] != POISON_FREE) { | |
1747 | error = data[offset + i]; | |
1748 | bad_count++; | |
1749 | } | |
b28a02de | 1750 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1751 | } |
1da177e4 | 1752 | printk("\n"); |
aa83aa40 DJ |
1753 | |
1754 | if (bad_count == 1) { | |
1755 | error ^= POISON_FREE; | |
1756 | if (!(error & (error - 1))) { | |
1757 | printk(KERN_ERR "Single bit error detected. Probably " | |
1758 | "bad RAM.\n"); | |
1759 | #ifdef CONFIG_X86 | |
1760 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1761 | "test tool.\n"); | |
1762 | #else | |
1763 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1764 | #endif | |
1765 | } | |
1766 | } | |
1da177e4 LT |
1767 | } |
1768 | #endif | |
1769 | ||
1770 | #if DEBUG | |
1771 | ||
343e0d7a | 1772 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1773 | { |
1774 | int i, size; | |
1775 | char *realobj; | |
1776 | ||
1777 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1778 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1779 | *dbg_redzone1(cachep, objp), |
1780 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1781 | } |
1782 | ||
1783 | if (cachep->flags & SLAB_STORE_USER) { | |
1784 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1785 | *dbg_userword(cachep, objp)); |
1da177e4 | 1786 | print_symbol("(%s)", |
a737b3e2 | 1787 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1788 | printk("\n"); |
1789 | } | |
3dafccf2 MS |
1790 | realobj = (char *)objp + obj_offset(cachep); |
1791 | size = obj_size(cachep); | |
b28a02de | 1792 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1793 | int limit; |
1794 | limit = 16; | |
b28a02de PE |
1795 | if (i + limit > size) |
1796 | limit = size - i; | |
1da177e4 LT |
1797 | dump_line(realobj, i, limit); |
1798 | } | |
1799 | } | |
1800 | ||
343e0d7a | 1801 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1802 | { |
1803 | char *realobj; | |
1804 | int size, i; | |
1805 | int lines = 0; | |
1806 | ||
3dafccf2 MS |
1807 | realobj = (char *)objp + obj_offset(cachep); |
1808 | size = obj_size(cachep); | |
1da177e4 | 1809 | |
b28a02de | 1810 | for (i = 0; i < size; i++) { |
1da177e4 | 1811 | char exp = POISON_FREE; |
b28a02de | 1812 | if (i == size - 1) |
1da177e4 LT |
1813 | exp = POISON_END; |
1814 | if (realobj[i] != exp) { | |
1815 | int limit; | |
1816 | /* Mismatch ! */ | |
1817 | /* Print header */ | |
1818 | if (lines == 0) { | |
b28a02de | 1819 | printk(KERN_ERR |
e94a40c5 DH |
1820 | "Slab corruption: %s start=%p, len=%d\n", |
1821 | cachep->name, realobj, size); | |
1da177e4 LT |
1822 | print_objinfo(cachep, objp, 0); |
1823 | } | |
1824 | /* Hexdump the affected line */ | |
b28a02de | 1825 | i = (i / 16) * 16; |
1da177e4 | 1826 | limit = 16; |
b28a02de PE |
1827 | if (i + limit > size) |
1828 | limit = size - i; | |
1da177e4 LT |
1829 | dump_line(realobj, i, limit); |
1830 | i += 16; | |
1831 | lines++; | |
1832 | /* Limit to 5 lines */ | |
1833 | if (lines > 5) | |
1834 | break; | |
1835 | } | |
1836 | } | |
1837 | if (lines != 0) { | |
1838 | /* Print some data about the neighboring objects, if they | |
1839 | * exist: | |
1840 | */ | |
6ed5eb22 | 1841 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1842 | unsigned int objnr; |
1da177e4 | 1843 | |
8fea4e96 | 1844 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1845 | if (objnr) { |
8fea4e96 | 1846 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1847 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1848 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1849 | realobj, size); |
1da177e4 LT |
1850 | print_objinfo(cachep, objp, 2); |
1851 | } | |
b28a02de | 1852 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1853 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1854 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1855 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1856 | realobj, size); |
1da177e4 LT |
1857 | print_objinfo(cachep, objp, 2); |
1858 | } | |
1859 | } | |
1860 | } | |
1861 | #endif | |
1862 | ||
12dd36fa | 1863 | #if DEBUG |
e79aec29 | 1864 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1865 | { |
1da177e4 LT |
1866 | int i; |
1867 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1868 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1869 | |
1870 | if (cachep->flags & SLAB_POISON) { | |
1871 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1872 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1873 | OFF_SLAB(cachep)) | |
b28a02de | 1874 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1875 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1876 | else |
1877 | check_poison_obj(cachep, objp); | |
1878 | #else | |
1879 | check_poison_obj(cachep, objp); | |
1880 | #endif | |
1881 | } | |
1882 | if (cachep->flags & SLAB_RED_ZONE) { | |
1883 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1884 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1885 | "was overwritten"); |
1da177e4 LT |
1886 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1887 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1888 | "was overwritten"); |
1da177e4 | 1889 | } |
1da177e4 | 1890 | } |
12dd36fa | 1891 | } |
1da177e4 | 1892 | #else |
e79aec29 | 1893 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1894 | { |
12dd36fa | 1895 | } |
1da177e4 LT |
1896 | #endif |
1897 | ||
911851e6 RD |
1898 | /** |
1899 | * slab_destroy - destroy and release all objects in a slab | |
1900 | * @cachep: cache pointer being destroyed | |
1901 | * @slabp: slab pointer being destroyed | |
1902 | * | |
12dd36fa | 1903 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1904 | * Before calling the slab must have been unlinked from the cache. The |
1905 | * cache-lock is not held/needed. | |
12dd36fa | 1906 | */ |
343e0d7a | 1907 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1908 | { |
1909 | void *addr = slabp->s_mem - slabp->colouroff; | |
1910 | ||
e79aec29 | 1911 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
1912 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1913 | struct slab_rcu *slab_rcu; | |
1914 | ||
b28a02de | 1915 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1916 | slab_rcu->cachep = cachep; |
1917 | slab_rcu->addr = addr; | |
1918 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1919 | } else { | |
1920 | kmem_freepages(cachep, addr); | |
873623df IM |
1921 | if (OFF_SLAB(cachep)) |
1922 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1923 | } |
1924 | } | |
1925 | ||
117f6eb1 CL |
1926 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1927 | { | |
1928 | int i; | |
1929 | struct kmem_list3 *l3; | |
1930 | ||
1931 | for_each_online_cpu(i) | |
1932 | kfree(cachep->array[i]); | |
1933 | ||
1934 | /* NUMA: free the list3 structures */ | |
1935 | for_each_online_node(i) { | |
1936 | l3 = cachep->nodelists[i]; | |
1937 | if (l3) { | |
1938 | kfree(l3->shared); | |
1939 | free_alien_cache(l3->alien); | |
1940 | kfree(l3); | |
1941 | } | |
1942 | } | |
1943 | kmem_cache_free(&cache_cache, cachep); | |
1944 | } | |
1945 | ||
1946 | ||
4d268eba | 1947 | /** |
a70773dd RD |
1948 | * calculate_slab_order - calculate size (page order) of slabs |
1949 | * @cachep: pointer to the cache that is being created | |
1950 | * @size: size of objects to be created in this cache. | |
1951 | * @align: required alignment for the objects. | |
1952 | * @flags: slab allocation flags | |
1953 | * | |
1954 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1955 | * |
1956 | * This could be made much more intelligent. For now, try to avoid using | |
1957 | * high order pages for slabs. When the gfp() functions are more friendly | |
1958 | * towards high-order requests, this should be changed. | |
1959 | */ | |
a737b3e2 | 1960 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 1961 | size_t size, size_t align, unsigned long flags) |
4d268eba | 1962 | { |
b1ab41c4 | 1963 | unsigned long offslab_limit; |
4d268eba | 1964 | size_t left_over = 0; |
9888e6fa | 1965 | int gfporder; |
4d268eba | 1966 | |
0aa817f0 | 1967 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
1968 | unsigned int num; |
1969 | size_t remainder; | |
1970 | ||
9888e6fa | 1971 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
1972 | if (!num) |
1973 | continue; | |
9888e6fa | 1974 | |
b1ab41c4 IM |
1975 | if (flags & CFLGS_OFF_SLAB) { |
1976 | /* | |
1977 | * Max number of objs-per-slab for caches which | |
1978 | * use off-slab slabs. Needed to avoid a possible | |
1979 | * looping condition in cache_grow(). | |
1980 | */ | |
1981 | offslab_limit = size - sizeof(struct slab); | |
1982 | offslab_limit /= sizeof(kmem_bufctl_t); | |
1983 | ||
1984 | if (num > offslab_limit) | |
1985 | break; | |
1986 | } | |
4d268eba | 1987 | |
9888e6fa | 1988 | /* Found something acceptable - save it away */ |
4d268eba | 1989 | cachep->num = num; |
9888e6fa | 1990 | cachep->gfporder = gfporder; |
4d268eba PE |
1991 | left_over = remainder; |
1992 | ||
f78bb8ad LT |
1993 | /* |
1994 | * A VFS-reclaimable slab tends to have most allocations | |
1995 | * as GFP_NOFS and we really don't want to have to be allocating | |
1996 | * higher-order pages when we are unable to shrink dcache. | |
1997 | */ | |
1998 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
1999 | break; | |
2000 | ||
4d268eba PE |
2001 | /* |
2002 | * Large number of objects is good, but very large slabs are | |
2003 | * currently bad for the gfp()s. | |
2004 | */ | |
9888e6fa | 2005 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2006 | break; |
2007 | ||
9888e6fa LT |
2008 | /* |
2009 | * Acceptable internal fragmentation? | |
2010 | */ | |
a737b3e2 | 2011 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2012 | break; |
2013 | } | |
2014 | return left_over; | |
2015 | } | |
2016 | ||
83b519e8 | 2017 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2018 | { |
2ed3a4ef | 2019 | if (g_cpucache_up == FULL) |
83b519e8 | 2020 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2021 | |
f30cf7d1 PE |
2022 | if (g_cpucache_up == NONE) { |
2023 | /* | |
2024 | * Note: the first kmem_cache_create must create the cache | |
2025 | * that's used by kmalloc(24), otherwise the creation of | |
2026 | * further caches will BUG(). | |
2027 | */ | |
2028 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2029 | ||
2030 | /* | |
2031 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2032 | * the first cache, then we need to set up all its list3s, | |
2033 | * otherwise the creation of further caches will BUG(). | |
2034 | */ | |
2035 | set_up_list3s(cachep, SIZE_AC); | |
2036 | if (INDEX_AC == INDEX_L3) | |
2037 | g_cpucache_up = PARTIAL_L3; | |
2038 | else | |
2039 | g_cpucache_up = PARTIAL_AC; | |
2040 | } else { | |
2041 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2042 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 PE |
2043 | |
2044 | if (g_cpucache_up == PARTIAL_AC) { | |
2045 | set_up_list3s(cachep, SIZE_L3); | |
2046 | g_cpucache_up = PARTIAL_L3; | |
2047 | } else { | |
2048 | int node; | |
556a169d | 2049 | for_each_online_node(node) { |
f30cf7d1 PE |
2050 | cachep->nodelists[node] = |
2051 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2052 | gfp, node); |
f30cf7d1 PE |
2053 | BUG_ON(!cachep->nodelists[node]); |
2054 | kmem_list3_init(cachep->nodelists[node]); | |
2055 | } | |
2056 | } | |
2057 | } | |
2058 | cachep->nodelists[numa_node_id()]->next_reap = | |
2059 | jiffies + REAPTIMEOUT_LIST3 + | |
2060 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2061 | ||
2062 | cpu_cache_get(cachep)->avail = 0; | |
2063 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2064 | cpu_cache_get(cachep)->batchcount = 1; | |
2065 | cpu_cache_get(cachep)->touched = 0; | |
2066 | cachep->batchcount = 1; | |
2067 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2068 | return 0; |
f30cf7d1 PE |
2069 | } |
2070 | ||
1da177e4 LT |
2071 | /** |
2072 | * kmem_cache_create - Create a cache. | |
2073 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2074 | * @size: The size of objects to be created in this cache. | |
2075 | * @align: The required alignment for the objects. | |
2076 | * @flags: SLAB flags | |
2077 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2078 | * |
2079 | * Returns a ptr to the cache on success, NULL on failure. | |
2080 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2081 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2082 | * |
2083 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 | 2084 | * the module calling this has to destroy the cache before getting unloaded. |
249da166 CM |
2085 | * Note that kmem_cache_name() is not guaranteed to return the same pointer, |
2086 | * therefore applications must manage it themselves. | |
a737b3e2 | 2087 | * |
1da177e4 LT |
2088 | * The flags are |
2089 | * | |
2090 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2091 | * to catch references to uninitialised memory. | |
2092 | * | |
2093 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2094 | * for buffer overruns. | |
2095 | * | |
1da177e4 LT |
2096 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2097 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2098 | * as davem. | |
2099 | */ | |
343e0d7a | 2100 | struct kmem_cache * |
1da177e4 | 2101 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2102 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2103 | { |
2104 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2105 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2106 | gfp_t gfp; |
1da177e4 LT |
2107 | |
2108 | /* | |
2109 | * Sanity checks... these are all serious usage bugs. | |
2110 | */ | |
a737b3e2 | 2111 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2112 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2113 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2114 | name); |
b28a02de PE |
2115 | BUG(); |
2116 | } | |
1da177e4 | 2117 | |
f0188f47 | 2118 | /* |
8f5be20b | 2119 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2120 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2121 | */ |
83b519e8 PE |
2122 | if (slab_is_available()) { |
2123 | get_online_cpus(); | |
2124 | mutex_lock(&cache_chain_mutex); | |
2125 | } | |
4f12bb4f | 2126 | |
7a7c381d | 2127 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2128 | char tmp; |
2129 | int res; | |
2130 | ||
2131 | /* | |
2132 | * This happens when the module gets unloaded and doesn't | |
2133 | * destroy its slab cache and no-one else reuses the vmalloc | |
2134 | * area of the module. Print a warning. | |
2135 | */ | |
138ae663 | 2136 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2137 | if (res) { |
b4169525 | 2138 | printk(KERN_ERR |
2139 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2140 | pc->buffer_size); |
4f12bb4f AM |
2141 | continue; |
2142 | } | |
2143 | ||
b28a02de | 2144 | if (!strcmp(pc->name, name)) { |
b4169525 | 2145 | printk(KERN_ERR |
2146 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2147 | dump_stack(); |
2148 | goto oops; | |
2149 | } | |
2150 | } | |
2151 | ||
1da177e4 LT |
2152 | #if DEBUG |
2153 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2154 | #if FORCED_DEBUG |
2155 | /* | |
2156 | * Enable redzoning and last user accounting, except for caches with | |
2157 | * large objects, if the increased size would increase the object size | |
2158 | * above the next power of two: caches with object sizes just above a | |
2159 | * power of two have a significant amount of internal fragmentation. | |
2160 | */ | |
87a927c7 DW |
2161 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2162 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2163 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2164 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2165 | flags |= SLAB_POISON; | |
2166 | #endif | |
2167 | if (flags & SLAB_DESTROY_BY_RCU) | |
2168 | BUG_ON(flags & SLAB_POISON); | |
2169 | #endif | |
1da177e4 | 2170 | /* |
a737b3e2 AM |
2171 | * Always checks flags, a caller might be expecting debug support which |
2172 | * isn't available. | |
1da177e4 | 2173 | */ |
40094fa6 | 2174 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2175 | |
a737b3e2 AM |
2176 | /* |
2177 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2178 | * unaligned accesses for some archs when redzoning is used, and makes |
2179 | * sure any on-slab bufctl's are also correctly aligned. | |
2180 | */ | |
b28a02de PE |
2181 | if (size & (BYTES_PER_WORD - 1)) { |
2182 | size += (BYTES_PER_WORD - 1); | |
2183 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2184 | } |
2185 | ||
a737b3e2 AM |
2186 | /* calculate the final buffer alignment: */ |
2187 | ||
1da177e4 LT |
2188 | /* 1) arch recommendation: can be overridden for debug */ |
2189 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2190 | /* |
2191 | * Default alignment: as specified by the arch code. Except if | |
2192 | * an object is really small, then squeeze multiple objects into | |
2193 | * one cacheline. | |
1da177e4 LT |
2194 | */ |
2195 | ralign = cache_line_size(); | |
b28a02de | 2196 | while (size <= ralign / 2) |
1da177e4 LT |
2197 | ralign /= 2; |
2198 | } else { | |
2199 | ralign = BYTES_PER_WORD; | |
2200 | } | |
ca5f9703 PE |
2201 | |
2202 | /* | |
87a927c7 DW |
2203 | * Redzoning and user store require word alignment or possibly larger. |
2204 | * Note this will be overridden by architecture or caller mandated | |
2205 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2206 | */ |
87a927c7 DW |
2207 | if (flags & SLAB_STORE_USER) |
2208 | ralign = BYTES_PER_WORD; | |
2209 | ||
2210 | if (flags & SLAB_RED_ZONE) { | |
2211 | ralign = REDZONE_ALIGN; | |
2212 | /* If redzoning, ensure that the second redzone is suitably | |
2213 | * aligned, by adjusting the object size accordingly. */ | |
2214 | size += REDZONE_ALIGN - 1; | |
2215 | size &= ~(REDZONE_ALIGN - 1); | |
2216 | } | |
ca5f9703 | 2217 | |
a44b56d3 | 2218 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2219 | if (ralign < ARCH_SLAB_MINALIGN) { |
2220 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2221 | } |
a44b56d3 | 2222 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2223 | if (ralign < align) { |
2224 | ralign = align; | |
1da177e4 | 2225 | } |
a44b56d3 | 2226 | /* disable debug if necessary */ |
b46b8f19 | 2227 | if (ralign > __alignof__(unsigned long long)) |
a44b56d3 | 2228 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2229 | /* |
ca5f9703 | 2230 | * 4) Store it. |
1da177e4 LT |
2231 | */ |
2232 | align = ralign; | |
2233 | ||
83b519e8 PE |
2234 | if (slab_is_available()) |
2235 | gfp = GFP_KERNEL; | |
2236 | else | |
2237 | gfp = GFP_NOWAIT; | |
2238 | ||
1da177e4 | 2239 | /* Get cache's description obj. */ |
83b519e8 | 2240 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2241 | if (!cachep) |
4f12bb4f | 2242 | goto oops; |
1da177e4 LT |
2243 | |
2244 | #if DEBUG | |
3dafccf2 | 2245 | cachep->obj_size = size; |
1da177e4 | 2246 | |
ca5f9703 PE |
2247 | /* |
2248 | * Both debugging options require word-alignment which is calculated | |
2249 | * into align above. | |
2250 | */ | |
1da177e4 | 2251 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2252 | /* add space for red zone words */ |
b46b8f19 DW |
2253 | cachep->obj_offset += sizeof(unsigned long long); |
2254 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2255 | } |
2256 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2257 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2258 | * the real object. But if the second red zone needs to be |
2259 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2260 | */ |
87a927c7 DW |
2261 | if (flags & SLAB_RED_ZONE) |
2262 | size += REDZONE_ALIGN; | |
2263 | else | |
2264 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2265 | } |
2266 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2267 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3dafccf2 MS |
2268 | && cachep->obj_size > cache_line_size() && size < PAGE_SIZE) { |
2269 | cachep->obj_offset += PAGE_SIZE - size; | |
1da177e4 LT |
2270 | size = PAGE_SIZE; |
2271 | } | |
2272 | #endif | |
2273 | #endif | |
2274 | ||
e0a42726 IM |
2275 | /* |
2276 | * Determine if the slab management is 'on' or 'off' slab. | |
2277 | * (bootstrapping cannot cope with offslab caches so don't do | |
2278 | * it too early on.) | |
2279 | */ | |
2280 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init) | |
1da177e4 LT |
2281 | /* |
2282 | * Size is large, assume best to place the slab management obj | |
2283 | * off-slab (should allow better packing of objs). | |
2284 | */ | |
2285 | flags |= CFLGS_OFF_SLAB; | |
2286 | ||
2287 | size = ALIGN(size, align); | |
2288 | ||
f78bb8ad | 2289 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2290 | |
2291 | if (!cachep->num) { | |
b4169525 | 2292 | printk(KERN_ERR |
2293 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2294 | kmem_cache_free(&cache_cache, cachep); |
2295 | cachep = NULL; | |
4f12bb4f | 2296 | goto oops; |
1da177e4 | 2297 | } |
b28a02de PE |
2298 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2299 | + sizeof(struct slab), align); | |
1da177e4 LT |
2300 | |
2301 | /* | |
2302 | * If the slab has been placed off-slab, and we have enough space then | |
2303 | * move it on-slab. This is at the expense of any extra colouring. | |
2304 | */ | |
2305 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2306 | flags &= ~CFLGS_OFF_SLAB; | |
2307 | left_over -= slab_size; | |
2308 | } | |
2309 | ||
2310 | if (flags & CFLGS_OFF_SLAB) { | |
2311 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2312 | slab_size = |
2313 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
1da177e4 LT |
2314 | } |
2315 | ||
2316 | cachep->colour_off = cache_line_size(); | |
2317 | /* Offset must be a multiple of the alignment. */ | |
2318 | if (cachep->colour_off < align) | |
2319 | cachep->colour_off = align; | |
b28a02de | 2320 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2321 | cachep->slab_size = slab_size; |
2322 | cachep->flags = flags; | |
2323 | cachep->gfpflags = 0; | |
4b51d669 | 2324 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2325 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2326 | cachep->buffer_size = size; |
6a2d7a95 | 2327 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2328 | |
e5ac9c5a | 2329 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2330 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2331 | /* |
2332 | * This is a possibility for one of the malloc_sizes caches. | |
2333 | * But since we go off slab only for object size greater than | |
2334 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2335 | * this should not happen at all. | |
2336 | * But leave a BUG_ON for some lucky dude. | |
2337 | */ | |
6cb8f913 | 2338 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2339 | } |
1da177e4 | 2340 | cachep->ctor = ctor; |
1da177e4 LT |
2341 | cachep->name = name; |
2342 | ||
83b519e8 | 2343 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef CL |
2344 | __kmem_cache_destroy(cachep); |
2345 | cachep = NULL; | |
2346 | goto oops; | |
2347 | } | |
1da177e4 | 2348 | |
1da177e4 LT |
2349 | /* cache setup completed, link it into the list */ |
2350 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2351 | oops: |
1da177e4 LT |
2352 | if (!cachep && (flags & SLAB_PANIC)) |
2353 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2354 | name); |
83b519e8 PE |
2355 | if (slab_is_available()) { |
2356 | mutex_unlock(&cache_chain_mutex); | |
2357 | put_online_cpus(); | |
2358 | } | |
1da177e4 LT |
2359 | return cachep; |
2360 | } | |
2361 | EXPORT_SYMBOL(kmem_cache_create); | |
2362 | ||
2363 | #if DEBUG | |
2364 | static void check_irq_off(void) | |
2365 | { | |
2366 | BUG_ON(!irqs_disabled()); | |
2367 | } | |
2368 | ||
2369 | static void check_irq_on(void) | |
2370 | { | |
2371 | BUG_ON(irqs_disabled()); | |
2372 | } | |
2373 | ||
343e0d7a | 2374 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2375 | { |
2376 | #ifdef CONFIG_SMP | |
2377 | check_irq_off(); | |
e498be7d | 2378 | assert_spin_locked(&cachep->nodelists[numa_node_id()]->list_lock); |
1da177e4 LT |
2379 | #endif |
2380 | } | |
e498be7d | 2381 | |
343e0d7a | 2382 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2383 | { |
2384 | #ifdef CONFIG_SMP | |
2385 | check_irq_off(); | |
2386 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2387 | #endif | |
2388 | } | |
2389 | ||
1da177e4 LT |
2390 | #else |
2391 | #define check_irq_off() do { } while(0) | |
2392 | #define check_irq_on() do { } while(0) | |
2393 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2394 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2395 | #endif |
2396 | ||
aab2207c CL |
2397 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2398 | struct array_cache *ac, | |
2399 | int force, int node); | |
2400 | ||
1da177e4 LT |
2401 | static void do_drain(void *arg) |
2402 | { | |
a737b3e2 | 2403 | struct kmem_cache *cachep = arg; |
1da177e4 | 2404 | struct array_cache *ac; |
ff69416e | 2405 | int node = numa_node_id(); |
1da177e4 LT |
2406 | |
2407 | check_irq_off(); | |
9a2dba4b | 2408 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2409 | spin_lock(&cachep->nodelists[node]->list_lock); |
2410 | free_block(cachep, ac->entry, ac->avail, node); | |
2411 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2412 | ac->avail = 0; |
2413 | } | |
2414 | ||
343e0d7a | 2415 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2416 | { |
e498be7d CL |
2417 | struct kmem_list3 *l3; |
2418 | int node; | |
2419 | ||
15c8b6c1 | 2420 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2421 | check_irq_on(); |
b28a02de | 2422 | for_each_online_node(node) { |
e498be7d | 2423 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2424 | if (l3 && l3->alien) |
2425 | drain_alien_cache(cachep, l3->alien); | |
2426 | } | |
2427 | ||
2428 | for_each_online_node(node) { | |
2429 | l3 = cachep->nodelists[node]; | |
2430 | if (l3) | |
aab2207c | 2431 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2432 | } |
1da177e4 LT |
2433 | } |
2434 | ||
ed11d9eb CL |
2435 | /* |
2436 | * Remove slabs from the list of free slabs. | |
2437 | * Specify the number of slabs to drain in tofree. | |
2438 | * | |
2439 | * Returns the actual number of slabs released. | |
2440 | */ | |
2441 | static int drain_freelist(struct kmem_cache *cache, | |
2442 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2443 | { |
ed11d9eb CL |
2444 | struct list_head *p; |
2445 | int nr_freed; | |
1da177e4 | 2446 | struct slab *slabp; |
1da177e4 | 2447 | |
ed11d9eb CL |
2448 | nr_freed = 0; |
2449 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2450 | |
ed11d9eb | 2451 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2452 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2453 | if (p == &l3->slabs_free) { |
2454 | spin_unlock_irq(&l3->list_lock); | |
2455 | goto out; | |
2456 | } | |
1da177e4 | 2457 | |
ed11d9eb | 2458 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2459 | #if DEBUG |
40094fa6 | 2460 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2461 | #endif |
2462 | list_del(&slabp->list); | |
ed11d9eb CL |
2463 | /* |
2464 | * Safe to drop the lock. The slab is no longer linked | |
2465 | * to the cache. | |
2466 | */ | |
2467 | l3->free_objects -= cache->num; | |
e498be7d | 2468 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2469 | slab_destroy(cache, slabp); |
2470 | nr_freed++; | |
1da177e4 | 2471 | } |
ed11d9eb CL |
2472 | out: |
2473 | return nr_freed; | |
1da177e4 LT |
2474 | } |
2475 | ||
8f5be20b | 2476 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2477 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2478 | { |
2479 | int ret = 0, i = 0; | |
2480 | struct kmem_list3 *l3; | |
2481 | ||
2482 | drain_cpu_caches(cachep); | |
2483 | ||
2484 | check_irq_on(); | |
2485 | for_each_online_node(i) { | |
2486 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2487 | if (!l3) |
2488 | continue; | |
2489 | ||
2490 | drain_freelist(cachep, l3, l3->free_objects); | |
2491 | ||
2492 | ret += !list_empty(&l3->slabs_full) || | |
2493 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2494 | } |
2495 | return (ret ? 1 : 0); | |
2496 | } | |
2497 | ||
1da177e4 LT |
2498 | /** |
2499 | * kmem_cache_shrink - Shrink a cache. | |
2500 | * @cachep: The cache to shrink. | |
2501 | * | |
2502 | * Releases as many slabs as possible for a cache. | |
2503 | * To help debugging, a zero exit status indicates all slabs were released. | |
2504 | */ | |
343e0d7a | 2505 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2506 | { |
8f5be20b | 2507 | int ret; |
40094fa6 | 2508 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2509 | |
95402b38 | 2510 | get_online_cpus(); |
8f5be20b RT |
2511 | mutex_lock(&cache_chain_mutex); |
2512 | ret = __cache_shrink(cachep); | |
2513 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2514 | put_online_cpus(); |
8f5be20b | 2515 | return ret; |
1da177e4 LT |
2516 | } |
2517 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2518 | ||
2519 | /** | |
2520 | * kmem_cache_destroy - delete a cache | |
2521 | * @cachep: the cache to destroy | |
2522 | * | |
72fd4a35 | 2523 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2524 | * |
2525 | * It is expected this function will be called by a module when it is | |
2526 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2527 | * cache being allocated each time a module is loaded and unloaded, if the | |
2528 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2529 | * | |
2530 | * The cache must be empty before calling this function. | |
2531 | * | |
2532 | * The caller must guarantee that noone will allocate memory from the cache | |
2533 | * during the kmem_cache_destroy(). | |
2534 | */ | |
133d205a | 2535 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2536 | { |
40094fa6 | 2537 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2538 | |
1da177e4 | 2539 | /* Find the cache in the chain of caches. */ |
95402b38 | 2540 | get_online_cpus(); |
fc0abb14 | 2541 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2542 | /* |
2543 | * the chain is never empty, cache_cache is never destroyed | |
2544 | */ | |
2545 | list_del(&cachep->next); | |
1da177e4 LT |
2546 | if (__cache_shrink(cachep)) { |
2547 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2548 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2549 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2550 | put_online_cpus(); |
133d205a | 2551 | return; |
1da177e4 LT |
2552 | } |
2553 | ||
2554 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
fbd568a3 | 2555 | synchronize_rcu(); |
1da177e4 | 2556 | |
117f6eb1 | 2557 | __kmem_cache_destroy(cachep); |
8f5be20b | 2558 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2559 | put_online_cpus(); |
1da177e4 LT |
2560 | } |
2561 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2562 | ||
e5ac9c5a RT |
2563 | /* |
2564 | * Get the memory for a slab management obj. | |
2565 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2566 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2567 | * come from the same cache which is getting created because, | |
2568 | * when we are searching for an appropriate cache for these | |
2569 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2570 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2571 | * kmem_find_general_cachep till the initialization is complete. | |
2572 | * Hence we cannot have slabp_cache same as the original cache. | |
2573 | */ | |
343e0d7a | 2574 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2575 | int colour_off, gfp_t local_flags, |
2576 | int nodeid) | |
1da177e4 LT |
2577 | { |
2578 | struct slab *slabp; | |
b28a02de | 2579 | |
1da177e4 LT |
2580 | if (OFF_SLAB(cachep)) { |
2581 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2582 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2583 | local_flags, nodeid); |
d5cff635 CM |
2584 | /* |
2585 | * If the first object in the slab is leaked (it's allocated | |
2586 | * but no one has a reference to it), we want to make sure | |
2587 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2588 | * to the object. Otherwise we will not report the leak. | |
2589 | */ | |
2590 | kmemleak_scan_area(slabp, offsetof(struct slab, list), | |
2591 | sizeof(struct list_head), local_flags); | |
1da177e4 LT |
2592 | if (!slabp) |
2593 | return NULL; | |
2594 | } else { | |
b28a02de | 2595 | slabp = objp + colour_off; |
1da177e4 LT |
2596 | colour_off += cachep->slab_size; |
2597 | } | |
2598 | slabp->inuse = 0; | |
2599 | slabp->colouroff = colour_off; | |
b28a02de | 2600 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2601 | slabp->nodeid = nodeid; |
e51bfd0a | 2602 | slabp->free = 0; |
1da177e4 LT |
2603 | return slabp; |
2604 | } | |
2605 | ||
2606 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2607 | { | |
b28a02de | 2608 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2609 | } |
2610 | ||
343e0d7a | 2611 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2612 | struct slab *slabp) |
1da177e4 LT |
2613 | { |
2614 | int i; | |
2615 | ||
2616 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2617 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2618 | #if DEBUG |
2619 | /* need to poison the objs? */ | |
2620 | if (cachep->flags & SLAB_POISON) | |
2621 | poison_obj(cachep, objp, POISON_FREE); | |
2622 | if (cachep->flags & SLAB_STORE_USER) | |
2623 | *dbg_userword(cachep, objp) = NULL; | |
2624 | ||
2625 | if (cachep->flags & SLAB_RED_ZONE) { | |
2626 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2627 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2628 | } | |
2629 | /* | |
a737b3e2 AM |
2630 | * Constructors are not allowed to allocate memory from the same |
2631 | * cache which they are a constructor for. Otherwise, deadlock. | |
2632 | * They must also be threaded. | |
1da177e4 LT |
2633 | */ |
2634 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2635 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2636 | |
2637 | if (cachep->flags & SLAB_RED_ZONE) { | |
2638 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2639 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2640 | " end of an object"); |
1da177e4 LT |
2641 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2642 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2643 | " start of an object"); |
1da177e4 | 2644 | } |
a737b3e2 AM |
2645 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2646 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2647 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2648 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2649 | #else |
2650 | if (cachep->ctor) | |
51cc5068 | 2651 | cachep->ctor(objp); |
1da177e4 | 2652 | #endif |
b28a02de | 2653 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2654 | } |
b28a02de | 2655 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2656 | } |
2657 | ||
343e0d7a | 2658 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2659 | { |
4b51d669 CL |
2660 | if (CONFIG_ZONE_DMA_FLAG) { |
2661 | if (flags & GFP_DMA) | |
2662 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2663 | else | |
2664 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2665 | } | |
1da177e4 LT |
2666 | } |
2667 | ||
a737b3e2 AM |
2668 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2669 | int nodeid) | |
78d382d7 | 2670 | { |
8fea4e96 | 2671 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2672 | kmem_bufctl_t next; |
2673 | ||
2674 | slabp->inuse++; | |
2675 | next = slab_bufctl(slabp)[slabp->free]; | |
2676 | #if DEBUG | |
2677 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2678 | WARN_ON(slabp->nodeid != nodeid); | |
2679 | #endif | |
2680 | slabp->free = next; | |
2681 | ||
2682 | return objp; | |
2683 | } | |
2684 | ||
a737b3e2 AM |
2685 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2686 | void *objp, int nodeid) | |
78d382d7 | 2687 | { |
8fea4e96 | 2688 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2689 | |
2690 | #if DEBUG | |
2691 | /* Verify that the slab belongs to the intended node */ | |
2692 | WARN_ON(slabp->nodeid != nodeid); | |
2693 | ||
871751e2 | 2694 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2695 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2696 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2697 | BUG(); |
2698 | } | |
2699 | #endif | |
2700 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2701 | slabp->free = objnr; | |
2702 | slabp->inuse--; | |
2703 | } | |
2704 | ||
4776874f PE |
2705 | /* |
2706 | * Map pages beginning at addr to the given cache and slab. This is required | |
2707 | * for the slab allocator to be able to lookup the cache and slab of a | |
2708 | * virtual address for kfree, ksize, kmem_ptr_validate, and slab debugging. | |
2709 | */ | |
2710 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2711 | void *addr) | |
1da177e4 | 2712 | { |
4776874f | 2713 | int nr_pages; |
1da177e4 LT |
2714 | struct page *page; |
2715 | ||
4776874f | 2716 | page = virt_to_page(addr); |
84097518 | 2717 | |
4776874f | 2718 | nr_pages = 1; |
84097518 | 2719 | if (likely(!PageCompound(page))) |
4776874f PE |
2720 | nr_pages <<= cache->gfporder; |
2721 | ||
1da177e4 | 2722 | do { |
4776874f PE |
2723 | page_set_cache(page, cache); |
2724 | page_set_slab(page, slab); | |
1da177e4 | 2725 | page++; |
4776874f | 2726 | } while (--nr_pages); |
1da177e4 LT |
2727 | } |
2728 | ||
2729 | /* | |
2730 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2731 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2732 | */ | |
3c517a61 CL |
2733 | static int cache_grow(struct kmem_cache *cachep, |
2734 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2735 | { |
b28a02de | 2736 | struct slab *slabp; |
b28a02de PE |
2737 | size_t offset; |
2738 | gfp_t local_flags; | |
e498be7d | 2739 | struct kmem_list3 *l3; |
1da177e4 | 2740 | |
a737b3e2 AM |
2741 | /* |
2742 | * Be lazy and only check for valid flags here, keeping it out of the | |
2743 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2744 | */ |
6cb06229 CL |
2745 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2746 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2747 | |
2e1217cf | 2748 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2749 | check_irq_off(); |
2e1217cf RT |
2750 | l3 = cachep->nodelists[nodeid]; |
2751 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2752 | |
2753 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2754 | offset = l3->colour_next; |
2755 | l3->colour_next++; | |
2756 | if (l3->colour_next >= cachep->colour) | |
2757 | l3->colour_next = 0; | |
2758 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2759 | |
2e1217cf | 2760 | offset *= cachep->colour_off; |
1da177e4 LT |
2761 | |
2762 | if (local_flags & __GFP_WAIT) | |
2763 | local_irq_enable(); | |
2764 | ||
2765 | /* | |
2766 | * The test for missing atomic flag is performed here, rather than | |
2767 | * the more obvious place, simply to reduce the critical path length | |
2768 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2769 | * will eventually be caught here (where it matters). | |
2770 | */ | |
2771 | kmem_flagcheck(cachep, flags); | |
2772 | ||
a737b3e2 AM |
2773 | /* |
2774 | * Get mem for the objs. Attempt to allocate a physical page from | |
2775 | * 'nodeid'. | |
e498be7d | 2776 | */ |
3c517a61 | 2777 | if (!objp) |
b8c1c5da | 2778 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2779 | if (!objp) |
1da177e4 LT |
2780 | goto failed; |
2781 | ||
2782 | /* Get slab management. */ | |
3c517a61 | 2783 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2784 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2785 | if (!slabp) |
1da177e4 LT |
2786 | goto opps1; |
2787 | ||
4776874f | 2788 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2789 | |
a35afb83 | 2790 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2791 | |
2792 | if (local_flags & __GFP_WAIT) | |
2793 | local_irq_disable(); | |
2794 | check_irq_off(); | |
e498be7d | 2795 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2796 | |
2797 | /* Make slab active. */ | |
e498be7d | 2798 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2799 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2800 | l3->free_objects += cachep->num; |
2801 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2802 | return 1; |
a737b3e2 | 2803 | opps1: |
1da177e4 | 2804 | kmem_freepages(cachep, objp); |
a737b3e2 | 2805 | failed: |
1da177e4 LT |
2806 | if (local_flags & __GFP_WAIT) |
2807 | local_irq_disable(); | |
2808 | return 0; | |
2809 | } | |
2810 | ||
2811 | #if DEBUG | |
2812 | ||
2813 | /* | |
2814 | * Perform extra freeing checks: | |
2815 | * - detect bad pointers. | |
2816 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2817 | */ |
2818 | static void kfree_debugcheck(const void *objp) | |
2819 | { | |
1da177e4 LT |
2820 | if (!virt_addr_valid(objp)) { |
2821 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2822 | (unsigned long)objp); |
2823 | BUG(); | |
1da177e4 | 2824 | } |
1da177e4 LT |
2825 | } |
2826 | ||
58ce1fd5 PE |
2827 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2828 | { | |
b46b8f19 | 2829 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2830 | |
2831 | redzone1 = *dbg_redzone1(cache, obj); | |
2832 | redzone2 = *dbg_redzone2(cache, obj); | |
2833 | ||
2834 | /* | |
2835 | * Redzone is ok. | |
2836 | */ | |
2837 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2838 | return; | |
2839 | ||
2840 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2841 | slab_error(cache, "double free detected"); | |
2842 | else | |
2843 | slab_error(cache, "memory outside object was overwritten"); | |
2844 | ||
b46b8f19 | 2845 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2846 | obj, redzone1, redzone2); |
2847 | } | |
2848 | ||
343e0d7a | 2849 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2850 | void *caller) |
1da177e4 LT |
2851 | { |
2852 | struct page *page; | |
2853 | unsigned int objnr; | |
2854 | struct slab *slabp; | |
2855 | ||
80cbd911 MW |
2856 | BUG_ON(virt_to_cache(objp) != cachep); |
2857 | ||
3dafccf2 | 2858 | objp -= obj_offset(cachep); |
1da177e4 | 2859 | kfree_debugcheck(objp); |
b49af68f | 2860 | page = virt_to_head_page(objp); |
1da177e4 | 2861 | |
065d41cb | 2862 | slabp = page_get_slab(page); |
1da177e4 LT |
2863 | |
2864 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2865 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2866 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2867 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2868 | } | |
2869 | if (cachep->flags & SLAB_STORE_USER) | |
2870 | *dbg_userword(cachep, objp) = caller; | |
2871 | ||
8fea4e96 | 2872 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2873 | |
2874 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2875 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2876 | |
871751e2 AV |
2877 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2878 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2879 | #endif | |
1da177e4 LT |
2880 | if (cachep->flags & SLAB_POISON) { |
2881 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2882 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2883 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2884 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2885 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2886 | } else { |
2887 | poison_obj(cachep, objp, POISON_FREE); | |
2888 | } | |
2889 | #else | |
2890 | poison_obj(cachep, objp, POISON_FREE); | |
2891 | #endif | |
2892 | } | |
2893 | return objp; | |
2894 | } | |
2895 | ||
343e0d7a | 2896 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2897 | { |
2898 | kmem_bufctl_t i; | |
2899 | int entries = 0; | |
b28a02de | 2900 | |
1da177e4 LT |
2901 | /* Check slab's freelist to see if this obj is there. */ |
2902 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2903 | entries++; | |
2904 | if (entries > cachep->num || i >= cachep->num) | |
2905 | goto bad; | |
2906 | } | |
2907 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2908 | bad: |
2909 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2910 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2911 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2912 | for (i = 0; |
264132bc | 2913 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2914 | i++) { |
a737b3e2 | 2915 | if (i % 16 == 0) |
1da177e4 | 2916 | printk("\n%03x:", i); |
b28a02de | 2917 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2918 | } |
2919 | printk("\n"); | |
2920 | BUG(); | |
2921 | } | |
2922 | } | |
2923 | #else | |
2924 | #define kfree_debugcheck(x) do { } while(0) | |
2925 | #define cache_free_debugcheck(x,objp,z) (objp) | |
2926 | #define check_slabp(x,y) do { } while(0) | |
2927 | #endif | |
2928 | ||
343e0d7a | 2929 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
2930 | { |
2931 | int batchcount; | |
2932 | struct kmem_list3 *l3; | |
2933 | struct array_cache *ac; | |
1ca4cb24 PE |
2934 | int node; |
2935 | ||
6d2144d3 | 2936 | retry: |
1da177e4 | 2937 | check_irq_off(); |
6d2144d3 | 2938 | node = numa_node_id(); |
9a2dba4b | 2939 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2940 | batchcount = ac->batchcount; |
2941 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2942 | /* |
2943 | * If there was little recent activity on this cache, then | |
2944 | * perform only a partial refill. Otherwise we could generate | |
2945 | * refill bouncing. | |
1da177e4 LT |
2946 | */ |
2947 | batchcount = BATCHREFILL_LIMIT; | |
2948 | } | |
1ca4cb24 | 2949 | l3 = cachep->nodelists[node]; |
e498be7d CL |
2950 | |
2951 | BUG_ON(ac->avail > 0 || !l3); | |
2952 | spin_lock(&l3->list_lock); | |
1da177e4 | 2953 | |
3ded175a CL |
2954 | /* See if we can refill from the shared array */ |
2955 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) | |
2956 | goto alloc_done; | |
2957 | ||
1da177e4 LT |
2958 | while (batchcount > 0) { |
2959 | struct list_head *entry; | |
2960 | struct slab *slabp; | |
2961 | /* Get slab alloc is to come from. */ | |
2962 | entry = l3->slabs_partial.next; | |
2963 | if (entry == &l3->slabs_partial) { | |
2964 | l3->free_touched = 1; | |
2965 | entry = l3->slabs_free.next; | |
2966 | if (entry == &l3->slabs_free) | |
2967 | goto must_grow; | |
2968 | } | |
2969 | ||
2970 | slabp = list_entry(entry, struct slab, list); | |
2971 | check_slabp(cachep, slabp); | |
2972 | check_spinlock_acquired(cachep); | |
714b8171 PE |
2973 | |
2974 | /* | |
2975 | * The slab was either on partial or free list so | |
2976 | * there must be at least one object available for | |
2977 | * allocation. | |
2978 | */ | |
249b9f33 | 2979 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 2980 | |
1da177e4 | 2981 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
2982 | STATS_INC_ALLOCED(cachep); |
2983 | STATS_INC_ACTIVE(cachep); | |
2984 | STATS_SET_HIGH(cachep); | |
2985 | ||
78d382d7 | 2986 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 2987 | node); |
1da177e4 LT |
2988 | } |
2989 | check_slabp(cachep, slabp); | |
2990 | ||
2991 | /* move slabp to correct slabp list: */ | |
2992 | list_del(&slabp->list); | |
2993 | if (slabp->free == BUFCTL_END) | |
2994 | list_add(&slabp->list, &l3->slabs_full); | |
2995 | else | |
2996 | list_add(&slabp->list, &l3->slabs_partial); | |
2997 | } | |
2998 | ||
a737b3e2 | 2999 | must_grow: |
1da177e4 | 3000 | l3->free_objects -= ac->avail; |
a737b3e2 | 3001 | alloc_done: |
e498be7d | 3002 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3003 | |
3004 | if (unlikely(!ac->avail)) { | |
3005 | int x; | |
3c517a61 | 3006 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3007 | |
a737b3e2 | 3008 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3009 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3010 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3011 | return NULL; |
3012 | ||
a737b3e2 | 3013 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3014 | goto retry; |
3015 | } | |
3016 | ac->touched = 1; | |
e498be7d | 3017 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3018 | } |
3019 | ||
a737b3e2 AM |
3020 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3021 | gfp_t flags) | |
1da177e4 LT |
3022 | { |
3023 | might_sleep_if(flags & __GFP_WAIT); | |
3024 | #if DEBUG | |
3025 | kmem_flagcheck(cachep, flags); | |
3026 | #endif | |
3027 | } | |
3028 | ||
3029 | #if DEBUG | |
a737b3e2 AM |
3030 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3031 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3032 | { |
b28a02de | 3033 | if (!objp) |
1da177e4 | 3034 | return objp; |
b28a02de | 3035 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3036 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3037 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3038 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3039 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3040 | else |
3041 | check_poison_obj(cachep, objp); | |
3042 | #else | |
3043 | check_poison_obj(cachep, objp); | |
3044 | #endif | |
3045 | poison_obj(cachep, objp, POISON_INUSE); | |
3046 | } | |
3047 | if (cachep->flags & SLAB_STORE_USER) | |
3048 | *dbg_userword(cachep, objp) = caller; | |
3049 | ||
3050 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3051 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3052 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3053 | slab_error(cachep, "double free, or memory outside" | |
3054 | " object was overwritten"); | |
b28a02de | 3055 | printk(KERN_ERR |
b46b8f19 | 3056 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3057 | objp, *dbg_redzone1(cachep, objp), |
3058 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3059 | } |
3060 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3061 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3062 | } | |
871751e2 AV |
3063 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3064 | { | |
3065 | struct slab *slabp; | |
3066 | unsigned objnr; | |
3067 | ||
b49af68f | 3068 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3069 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3070 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3071 | } | |
3072 | #endif | |
3dafccf2 | 3073 | objp += obj_offset(cachep); |
4f104934 | 3074 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3075 | cachep->ctor(objp); |
a44b56d3 KH |
3076 | #if ARCH_SLAB_MINALIGN |
3077 | if ((u32)objp & (ARCH_SLAB_MINALIGN-1)) { | |
3078 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3079 | objp, ARCH_SLAB_MINALIGN); | |
3080 | } | |
3081 | #endif | |
1da177e4 LT |
3082 | return objp; |
3083 | } | |
3084 | #else | |
3085 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3086 | #endif | |
3087 | ||
773ff60e | 3088 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3089 | { |
3090 | if (cachep == &cache_cache) | |
773ff60e | 3091 | return false; |
8a8b6502 | 3092 | |
773ff60e | 3093 | return should_failslab(obj_size(cachep), flags); |
8a8b6502 AM |
3094 | } |
3095 | ||
343e0d7a | 3096 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3097 | { |
b28a02de | 3098 | void *objp; |
1da177e4 LT |
3099 | struct array_cache *ac; |
3100 | ||
5c382300 | 3101 | check_irq_off(); |
8a8b6502 | 3102 | |
9a2dba4b | 3103 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3104 | if (likely(ac->avail)) { |
3105 | STATS_INC_ALLOCHIT(cachep); | |
3106 | ac->touched = 1; | |
e498be7d | 3107 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3108 | } else { |
3109 | STATS_INC_ALLOCMISS(cachep); | |
3110 | objp = cache_alloc_refill(cachep, flags); | |
3111 | } | |
d5cff635 CM |
3112 | /* |
3113 | * To avoid a false negative, if an object that is in one of the | |
3114 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3115 | * treat the array pointers as a reference to the object. | |
3116 | */ | |
3117 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3118 | return objp; |
3119 | } | |
3120 | ||
e498be7d | 3121 | #ifdef CONFIG_NUMA |
c61afb18 | 3122 | /* |
b2455396 | 3123 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3124 | * |
3125 | * If we are in_interrupt, then process context, including cpusets and | |
3126 | * mempolicy, may not apply and should not be used for allocation policy. | |
3127 | */ | |
3128 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3129 | { | |
3130 | int nid_alloc, nid_here; | |
3131 | ||
765c4507 | 3132 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 PJ |
3133 | return NULL; |
3134 | nid_alloc = nid_here = numa_node_id(); | |
3135 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3136 | nid_alloc = cpuset_mem_spread_node(); | |
3137 | else if (current->mempolicy) | |
3138 | nid_alloc = slab_node(current->mempolicy); | |
3139 | if (nid_alloc != nid_here) | |
8b98c169 | 3140 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3141 | return NULL; |
3142 | } | |
3143 | ||
765c4507 CL |
3144 | /* |
3145 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3146 | * certain node and fall back is permitted. First we scan all the |
3147 | * available nodelists for available objects. If that fails then we | |
3148 | * perform an allocation without specifying a node. This allows the page | |
3149 | * allocator to do its reclaim / fallback magic. We then insert the | |
3150 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3151 | */ |
8c8cc2c1 | 3152 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3153 | { |
8c8cc2c1 PE |
3154 | struct zonelist *zonelist; |
3155 | gfp_t local_flags; | |
dd1a239f | 3156 | struct zoneref *z; |
54a6eb5c MG |
3157 | struct zone *zone; |
3158 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3159 | void *obj = NULL; |
3c517a61 | 3160 | int nid; |
8c8cc2c1 PE |
3161 | |
3162 | if (flags & __GFP_THISNODE) | |
3163 | return NULL; | |
3164 | ||
0e88460d | 3165 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
6cb06229 | 3166 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3167 | |
3c517a61 CL |
3168 | retry: |
3169 | /* | |
3170 | * Look through allowed nodes for objects available | |
3171 | * from existing per node queues. | |
3172 | */ | |
54a6eb5c MG |
3173 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3174 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3175 | |
54a6eb5c | 3176 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3177 | cache->nodelists[nid] && |
481c5346 | 3178 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3179 | obj = ____cache_alloc_node(cache, |
3180 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3181 | if (obj) |
3182 | break; | |
3183 | } | |
3c517a61 CL |
3184 | } |
3185 | ||
cfce6604 | 3186 | if (!obj) { |
3c517a61 CL |
3187 | /* |
3188 | * This allocation will be performed within the constraints | |
3189 | * of the current cpuset / memory policy requirements. | |
3190 | * We may trigger various forms of reclaim on the allowed | |
3191 | * set and go into memory reserves if necessary. | |
3192 | */ | |
dd47ea75 CL |
3193 | if (local_flags & __GFP_WAIT) |
3194 | local_irq_enable(); | |
3195 | kmem_flagcheck(cache, flags); | |
9ac33b2b | 3196 | obj = kmem_getpages(cache, local_flags, -1); |
dd47ea75 CL |
3197 | if (local_flags & __GFP_WAIT) |
3198 | local_irq_disable(); | |
3c517a61 CL |
3199 | if (obj) { |
3200 | /* | |
3201 | * Insert into the appropriate per node queues | |
3202 | */ | |
3203 | nid = page_to_nid(virt_to_page(obj)); | |
3204 | if (cache_grow(cache, flags, nid, obj)) { | |
3205 | obj = ____cache_alloc_node(cache, | |
3206 | flags | GFP_THISNODE, nid); | |
3207 | if (!obj) | |
3208 | /* | |
3209 | * Another processor may allocate the | |
3210 | * objects in the slab since we are | |
3211 | * not holding any locks. | |
3212 | */ | |
3213 | goto retry; | |
3214 | } else { | |
b6a60451 | 3215 | /* cache_grow already freed obj */ |
3c517a61 CL |
3216 | obj = NULL; |
3217 | } | |
3218 | } | |
aedb0eb1 | 3219 | } |
765c4507 CL |
3220 | return obj; |
3221 | } | |
3222 | ||
e498be7d CL |
3223 | /* |
3224 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3225 | */ |
8b98c169 | 3226 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3227 | int nodeid) |
e498be7d CL |
3228 | { |
3229 | struct list_head *entry; | |
b28a02de PE |
3230 | struct slab *slabp; |
3231 | struct kmem_list3 *l3; | |
3232 | void *obj; | |
b28a02de PE |
3233 | int x; |
3234 | ||
3235 | l3 = cachep->nodelists[nodeid]; | |
3236 | BUG_ON(!l3); | |
3237 | ||
a737b3e2 | 3238 | retry: |
ca3b9b91 | 3239 | check_irq_off(); |
b28a02de PE |
3240 | spin_lock(&l3->list_lock); |
3241 | entry = l3->slabs_partial.next; | |
3242 | if (entry == &l3->slabs_partial) { | |
3243 | l3->free_touched = 1; | |
3244 | entry = l3->slabs_free.next; | |
3245 | if (entry == &l3->slabs_free) | |
3246 | goto must_grow; | |
3247 | } | |
3248 | ||
3249 | slabp = list_entry(entry, struct slab, list); | |
3250 | check_spinlock_acquired_node(cachep, nodeid); | |
3251 | check_slabp(cachep, slabp); | |
3252 | ||
3253 | STATS_INC_NODEALLOCS(cachep); | |
3254 | STATS_INC_ACTIVE(cachep); | |
3255 | STATS_SET_HIGH(cachep); | |
3256 | ||
3257 | BUG_ON(slabp->inuse == cachep->num); | |
3258 | ||
78d382d7 | 3259 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3260 | check_slabp(cachep, slabp); |
3261 | l3->free_objects--; | |
3262 | /* move slabp to correct slabp list: */ | |
3263 | list_del(&slabp->list); | |
3264 | ||
a737b3e2 | 3265 | if (slabp->free == BUFCTL_END) |
b28a02de | 3266 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3267 | else |
b28a02de | 3268 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3269 | |
b28a02de PE |
3270 | spin_unlock(&l3->list_lock); |
3271 | goto done; | |
e498be7d | 3272 | |
a737b3e2 | 3273 | must_grow: |
b28a02de | 3274 | spin_unlock(&l3->list_lock); |
3c517a61 | 3275 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3276 | if (x) |
3277 | goto retry; | |
1da177e4 | 3278 | |
8c8cc2c1 | 3279 | return fallback_alloc(cachep, flags); |
e498be7d | 3280 | |
a737b3e2 | 3281 | done: |
b28a02de | 3282 | return obj; |
e498be7d | 3283 | } |
8c8cc2c1 PE |
3284 | |
3285 | /** | |
3286 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3287 | * @cachep: The cache to allocate from. | |
3288 | * @flags: See kmalloc(). | |
3289 | * @nodeid: node number of the target node. | |
3290 | * @caller: return address of caller, used for debug information | |
3291 | * | |
3292 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3293 | * node, which can improve the performance for cpu bound structures. | |
3294 | * | |
3295 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3296 | */ | |
3297 | static __always_inline void * | |
3298 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3299 | void *caller) | |
3300 | { | |
3301 | unsigned long save_flags; | |
3302 | void *ptr; | |
3303 | ||
7e85ee0c PE |
3304 | flags &= slab_gfp_mask; |
3305 | ||
cf40bd16 NP |
3306 | lockdep_trace_alloc(flags); |
3307 | ||
773ff60e | 3308 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3309 | return NULL; |
3310 | ||
8c8cc2c1 PE |
3311 | cache_alloc_debugcheck_before(cachep, flags); |
3312 | local_irq_save(save_flags); | |
3313 | ||
3314 | if (unlikely(nodeid == -1)) | |
3315 | nodeid = numa_node_id(); | |
3316 | ||
3317 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3318 | /* Node not bootstrapped yet */ | |
3319 | ptr = fallback_alloc(cachep, flags); | |
3320 | goto out; | |
3321 | } | |
3322 | ||
3323 | if (nodeid == numa_node_id()) { | |
3324 | /* | |
3325 | * Use the locally cached objects if possible. | |
3326 | * However ____cache_alloc does not allow fallback | |
3327 | * to other nodes. It may fail while we still have | |
3328 | * objects on other nodes available. | |
3329 | */ | |
3330 | ptr = ____cache_alloc(cachep, flags); | |
3331 | if (ptr) | |
3332 | goto out; | |
3333 | } | |
3334 | /* ___cache_alloc_node can fall back to other nodes */ | |
3335 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3336 | out: | |
3337 | local_irq_restore(save_flags); | |
3338 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
d5cff635 CM |
3339 | kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags, |
3340 | flags); | |
8c8cc2c1 | 3341 | |
c175eea4 PE |
3342 | if (likely(ptr)) |
3343 | kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep)); | |
3344 | ||
d07dbea4 CL |
3345 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3346 | memset(ptr, 0, obj_size(cachep)); | |
3347 | ||
8c8cc2c1 PE |
3348 | return ptr; |
3349 | } | |
3350 | ||
3351 | static __always_inline void * | |
3352 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3353 | { | |
3354 | void *objp; | |
3355 | ||
3356 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3357 | objp = alternate_node_alloc(cache, flags); | |
3358 | if (objp) | |
3359 | goto out; | |
3360 | } | |
3361 | objp = ____cache_alloc(cache, flags); | |
3362 | ||
3363 | /* | |
3364 | * We may just have run out of memory on the local node. | |
3365 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3366 | */ | |
3367 | if (!objp) | |
3368 | objp = ____cache_alloc_node(cache, flags, numa_node_id()); | |
3369 | ||
3370 | out: | |
3371 | return objp; | |
3372 | } | |
3373 | #else | |
3374 | ||
3375 | static __always_inline void * | |
3376 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3377 | { | |
3378 | return ____cache_alloc(cachep, flags); | |
3379 | } | |
3380 | ||
3381 | #endif /* CONFIG_NUMA */ | |
3382 | ||
3383 | static __always_inline void * | |
3384 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3385 | { | |
3386 | unsigned long save_flags; | |
3387 | void *objp; | |
3388 | ||
7e85ee0c PE |
3389 | flags &= slab_gfp_mask; |
3390 | ||
cf40bd16 NP |
3391 | lockdep_trace_alloc(flags); |
3392 | ||
773ff60e | 3393 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3394 | return NULL; |
3395 | ||
8c8cc2c1 PE |
3396 | cache_alloc_debugcheck_before(cachep, flags); |
3397 | local_irq_save(save_flags); | |
3398 | objp = __do_cache_alloc(cachep, flags); | |
3399 | local_irq_restore(save_flags); | |
3400 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
d5cff635 CM |
3401 | kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags, |
3402 | flags); | |
8c8cc2c1 PE |
3403 | prefetchw(objp); |
3404 | ||
c175eea4 PE |
3405 | if (likely(objp)) |
3406 | kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep)); | |
3407 | ||
d07dbea4 CL |
3408 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3409 | memset(objp, 0, obj_size(cachep)); | |
3410 | ||
8c8cc2c1 PE |
3411 | return objp; |
3412 | } | |
e498be7d CL |
3413 | |
3414 | /* | |
3415 | * Caller needs to acquire correct kmem_list's list_lock | |
3416 | */ | |
343e0d7a | 3417 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3418 | int node) |
1da177e4 LT |
3419 | { |
3420 | int i; | |
e498be7d | 3421 | struct kmem_list3 *l3; |
1da177e4 LT |
3422 | |
3423 | for (i = 0; i < nr_objects; i++) { | |
3424 | void *objp = objpp[i]; | |
3425 | struct slab *slabp; | |
1da177e4 | 3426 | |
6ed5eb22 | 3427 | slabp = virt_to_slab(objp); |
ff69416e | 3428 | l3 = cachep->nodelists[node]; |
1da177e4 | 3429 | list_del(&slabp->list); |
ff69416e | 3430 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3431 | check_slabp(cachep, slabp); |
78d382d7 | 3432 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3433 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3434 | l3->free_objects++; |
1da177e4 LT |
3435 | check_slabp(cachep, slabp); |
3436 | ||
3437 | /* fixup slab chains */ | |
3438 | if (slabp->inuse == 0) { | |
e498be7d CL |
3439 | if (l3->free_objects > l3->free_limit) { |
3440 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3441 | /* No need to drop any previously held |
3442 | * lock here, even if we have a off-slab slab | |
3443 | * descriptor it is guaranteed to come from | |
3444 | * a different cache, refer to comments before | |
3445 | * alloc_slabmgmt. | |
3446 | */ | |
1da177e4 LT |
3447 | slab_destroy(cachep, slabp); |
3448 | } else { | |
e498be7d | 3449 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3450 | } |
3451 | } else { | |
3452 | /* Unconditionally move a slab to the end of the | |
3453 | * partial list on free - maximum time for the | |
3454 | * other objects to be freed, too. | |
3455 | */ | |
e498be7d | 3456 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3457 | } |
3458 | } | |
3459 | } | |
3460 | ||
343e0d7a | 3461 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3462 | { |
3463 | int batchcount; | |
e498be7d | 3464 | struct kmem_list3 *l3; |
ff69416e | 3465 | int node = numa_node_id(); |
1da177e4 LT |
3466 | |
3467 | batchcount = ac->batchcount; | |
3468 | #if DEBUG | |
3469 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3470 | #endif | |
3471 | check_irq_off(); | |
ff69416e | 3472 | l3 = cachep->nodelists[node]; |
873623df | 3473 | spin_lock(&l3->list_lock); |
e498be7d CL |
3474 | if (l3->shared) { |
3475 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3476 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3477 | if (max) { |
3478 | if (batchcount > max) | |
3479 | batchcount = max; | |
e498be7d | 3480 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3481 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3482 | shared_array->avail += batchcount; |
3483 | goto free_done; | |
3484 | } | |
3485 | } | |
3486 | ||
ff69416e | 3487 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3488 | free_done: |
1da177e4 LT |
3489 | #if STATS |
3490 | { | |
3491 | int i = 0; | |
3492 | struct list_head *p; | |
3493 | ||
e498be7d CL |
3494 | p = l3->slabs_free.next; |
3495 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3496 | struct slab *slabp; |
3497 | ||
3498 | slabp = list_entry(p, struct slab, list); | |
3499 | BUG_ON(slabp->inuse); | |
3500 | ||
3501 | i++; | |
3502 | p = p->next; | |
3503 | } | |
3504 | STATS_SET_FREEABLE(cachep, i); | |
3505 | } | |
3506 | #endif | |
e498be7d | 3507 | spin_unlock(&l3->list_lock); |
1da177e4 | 3508 | ac->avail -= batchcount; |
a737b3e2 | 3509 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3510 | } |
3511 | ||
3512 | /* | |
a737b3e2 AM |
3513 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3514 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3515 | */ |
873623df | 3516 | static inline void __cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 | 3517 | { |
9a2dba4b | 3518 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3519 | |
3520 | check_irq_off(); | |
d5cff635 | 3521 | kmemleak_free_recursive(objp, cachep->flags); |
1da177e4 LT |
3522 | objp = cache_free_debugcheck(cachep, objp, __builtin_return_address(0)); |
3523 | ||
c175eea4 PE |
3524 | kmemcheck_slab_free(cachep, objp, obj_size(cachep)); |
3525 | ||
1807a1aa SS |
3526 | /* |
3527 | * Skip calling cache_free_alien() when the platform is not numa. | |
3528 | * This will avoid cache misses that happen while accessing slabp (which | |
3529 | * is per page memory reference) to get nodeid. Instead use a global | |
3530 | * variable to skip the call, which is mostly likely to be present in | |
3531 | * the cache. | |
3532 | */ | |
3533 | if (numa_platform && cache_free_alien(cachep, objp)) | |
729bd0b7 PE |
3534 | return; |
3535 | ||
1da177e4 LT |
3536 | if (likely(ac->avail < ac->limit)) { |
3537 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3538 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3539 | return; |
3540 | } else { | |
3541 | STATS_INC_FREEMISS(cachep); | |
3542 | cache_flusharray(cachep, ac); | |
e498be7d | 3543 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3544 | } |
3545 | } | |
3546 | ||
3547 | /** | |
3548 | * kmem_cache_alloc - Allocate an object | |
3549 | * @cachep: The cache to allocate from. | |
3550 | * @flags: See kmalloc(). | |
3551 | * | |
3552 | * Allocate an object from this cache. The flags are only relevant | |
3553 | * if the cache has no available objects. | |
3554 | */ | |
343e0d7a | 3555 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3556 | { |
36555751 EGM |
3557 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3558 | ||
ca2b84cb EGM |
3559 | trace_kmem_cache_alloc(_RET_IP_, ret, |
3560 | obj_size(cachep), cachep->buffer_size, flags); | |
36555751 EGM |
3561 | |
3562 | return ret; | |
1da177e4 LT |
3563 | } |
3564 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3565 | ||
36555751 EGM |
3566 | #ifdef CONFIG_KMEMTRACE |
3567 | void *kmem_cache_alloc_notrace(struct kmem_cache *cachep, gfp_t flags) | |
3568 | { | |
3569 | return __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3570 | } | |
3571 | EXPORT_SYMBOL(kmem_cache_alloc_notrace); | |
3572 | #endif | |
3573 | ||
1da177e4 | 3574 | /** |
7682486b | 3575 | * kmem_ptr_validate - check if an untrusted pointer might be a slab entry. |
1da177e4 LT |
3576 | * @cachep: the cache we're checking against |
3577 | * @ptr: pointer to validate | |
3578 | * | |
7682486b | 3579 | * This verifies that the untrusted pointer looks sane; |
1da177e4 LT |
3580 | * it is _not_ a guarantee that the pointer is actually |
3581 | * part of the slab cache in question, but it at least | |
3582 | * validates that the pointer can be dereferenced and | |
3583 | * looks half-way sane. | |
3584 | * | |
3585 | * Currently only used for dentry validation. | |
3586 | */ | |
b7f869a2 | 3587 | int kmem_ptr_validate(struct kmem_cache *cachep, const void *ptr) |
1da177e4 | 3588 | { |
b28a02de | 3589 | unsigned long addr = (unsigned long)ptr; |
1da177e4 | 3590 | unsigned long min_addr = PAGE_OFFSET; |
b28a02de | 3591 | unsigned long align_mask = BYTES_PER_WORD - 1; |
3dafccf2 | 3592 | unsigned long size = cachep->buffer_size; |
1da177e4 LT |
3593 | struct page *page; |
3594 | ||
3595 | if (unlikely(addr < min_addr)) | |
3596 | goto out; | |
3597 | if (unlikely(addr > (unsigned long)high_memory - size)) | |
3598 | goto out; | |
3599 | if (unlikely(addr & align_mask)) | |
3600 | goto out; | |
3601 | if (unlikely(!kern_addr_valid(addr))) | |
3602 | goto out; | |
3603 | if (unlikely(!kern_addr_valid(addr + size - 1))) | |
3604 | goto out; | |
3605 | page = virt_to_page(ptr); | |
3606 | if (unlikely(!PageSlab(page))) | |
3607 | goto out; | |
065d41cb | 3608 | if (unlikely(page_get_cache(page) != cachep)) |
1da177e4 LT |
3609 | goto out; |
3610 | return 1; | |
a737b3e2 | 3611 | out: |
1da177e4 LT |
3612 | return 0; |
3613 | } | |
3614 | ||
3615 | #ifdef CONFIG_NUMA | |
8b98c169 CH |
3616 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3617 | { | |
36555751 EGM |
3618 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3619 | __builtin_return_address(0)); | |
3620 | ||
ca2b84cb EGM |
3621 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
3622 | obj_size(cachep), cachep->buffer_size, | |
3623 | flags, nodeid); | |
36555751 EGM |
3624 | |
3625 | return ret; | |
8b98c169 | 3626 | } |
1da177e4 LT |
3627 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3628 | ||
36555751 EGM |
3629 | #ifdef CONFIG_KMEMTRACE |
3630 | void *kmem_cache_alloc_node_notrace(struct kmem_cache *cachep, | |
3631 | gfp_t flags, | |
3632 | int nodeid) | |
3633 | { | |
3634 | return __cache_alloc_node(cachep, flags, nodeid, | |
3635 | __builtin_return_address(0)); | |
3636 | } | |
3637 | EXPORT_SYMBOL(kmem_cache_alloc_node_notrace); | |
3638 | #endif | |
3639 | ||
8b98c169 CH |
3640 | static __always_inline void * |
3641 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3642 | { |
343e0d7a | 3643 | struct kmem_cache *cachep; |
36555751 | 3644 | void *ret; |
97e2bde4 MS |
3645 | |
3646 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3647 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3648 | return cachep; | |
36555751 EGM |
3649 | ret = kmem_cache_alloc_node_notrace(cachep, flags, node); |
3650 | ||
ca2b84cb EGM |
3651 | trace_kmalloc_node((unsigned long) caller, ret, |
3652 | size, cachep->buffer_size, flags, node); | |
36555751 EGM |
3653 | |
3654 | return ret; | |
97e2bde4 | 3655 | } |
8b98c169 | 3656 | |
36555751 | 3657 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE) |
8b98c169 CH |
3658 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3659 | { | |
3660 | return __do_kmalloc_node(size, flags, node, | |
3661 | __builtin_return_address(0)); | |
3662 | } | |
dbe5e69d | 3663 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3664 | |
3665 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3666 | int node, unsigned long caller) |
8b98c169 | 3667 | { |
ce71e27c | 3668 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3669 | } |
3670 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3671 | #else | |
3672 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3673 | { | |
3674 | return __do_kmalloc_node(size, flags, node, NULL); | |
3675 | } | |
3676 | EXPORT_SYMBOL(__kmalloc_node); | |
3677 | #endif /* CONFIG_DEBUG_SLAB */ | |
3678 | #endif /* CONFIG_NUMA */ | |
1da177e4 LT |
3679 | |
3680 | /** | |
800590f5 | 3681 | * __do_kmalloc - allocate memory |
1da177e4 | 3682 | * @size: how many bytes of memory are required. |
800590f5 | 3683 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3684 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3685 | */ |
7fd6b141 PE |
3686 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3687 | void *caller) | |
1da177e4 | 3688 | { |
343e0d7a | 3689 | struct kmem_cache *cachep; |
36555751 | 3690 | void *ret; |
1da177e4 | 3691 | |
97e2bde4 MS |
3692 | /* If you want to save a few bytes .text space: replace |
3693 | * __ with kmem_. | |
3694 | * Then kmalloc uses the uninlined functions instead of the inline | |
3695 | * functions. | |
3696 | */ | |
3697 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3698 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3699 | return cachep; | |
36555751 EGM |
3700 | ret = __cache_alloc(cachep, flags, caller); |
3701 | ||
ca2b84cb EGM |
3702 | trace_kmalloc((unsigned long) caller, ret, |
3703 | size, cachep->buffer_size, flags); | |
36555751 EGM |
3704 | |
3705 | return ret; | |
7fd6b141 PE |
3706 | } |
3707 | ||
7fd6b141 | 3708 | |
36555751 | 3709 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_KMEMTRACE) |
7fd6b141 PE |
3710 | void *__kmalloc(size_t size, gfp_t flags) |
3711 | { | |
871751e2 | 3712 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3713 | } |
3714 | EXPORT_SYMBOL(__kmalloc); | |
3715 | ||
ce71e27c | 3716 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3717 | { |
ce71e27c | 3718 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3719 | } |
3720 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3721 | |
3722 | #else | |
3723 | void *__kmalloc(size_t size, gfp_t flags) | |
3724 | { | |
3725 | return __do_kmalloc(size, flags, NULL); | |
3726 | } | |
3727 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3728 | #endif |
3729 | ||
1da177e4 LT |
3730 | /** |
3731 | * kmem_cache_free - Deallocate an object | |
3732 | * @cachep: The cache the allocation was from. | |
3733 | * @objp: The previously allocated object. | |
3734 | * | |
3735 | * Free an object which was previously allocated from this | |
3736 | * cache. | |
3737 | */ | |
343e0d7a | 3738 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3739 | { |
3740 | unsigned long flags; | |
3741 | ||
3742 | local_irq_save(flags); | |
898552c9 | 3743 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
3ac7fe5a TG |
3744 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3745 | debug_check_no_obj_freed(objp, obj_size(cachep)); | |
873623df | 3746 | __cache_free(cachep, objp); |
1da177e4 | 3747 | local_irq_restore(flags); |
36555751 | 3748 | |
ca2b84cb | 3749 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3750 | } |
3751 | EXPORT_SYMBOL(kmem_cache_free); | |
3752 | ||
1da177e4 LT |
3753 | /** |
3754 | * kfree - free previously allocated memory | |
3755 | * @objp: pointer returned by kmalloc. | |
3756 | * | |
80e93eff PE |
3757 | * If @objp is NULL, no operation is performed. |
3758 | * | |
1da177e4 LT |
3759 | * Don't free memory not originally allocated by kmalloc() |
3760 | * or you will run into trouble. | |
3761 | */ | |
3762 | void kfree(const void *objp) | |
3763 | { | |
343e0d7a | 3764 | struct kmem_cache *c; |
1da177e4 LT |
3765 | unsigned long flags; |
3766 | ||
2121db74 PE |
3767 | trace_kfree(_RET_IP_, objp); |
3768 | ||
6cb8f913 | 3769 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3770 | return; |
3771 | local_irq_save(flags); | |
3772 | kfree_debugcheck(objp); | |
6ed5eb22 | 3773 | c = virt_to_cache(objp); |
f9b8404c | 3774 | debug_check_no_locks_freed(objp, obj_size(c)); |
3ac7fe5a | 3775 | debug_check_no_obj_freed(objp, obj_size(c)); |
873623df | 3776 | __cache_free(c, (void *)objp); |
1da177e4 LT |
3777 | local_irq_restore(flags); |
3778 | } | |
3779 | EXPORT_SYMBOL(kfree); | |
3780 | ||
343e0d7a | 3781 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3782 | { |
3dafccf2 | 3783 | return obj_size(cachep); |
1da177e4 LT |
3784 | } |
3785 | EXPORT_SYMBOL(kmem_cache_size); | |
3786 | ||
343e0d7a | 3787 | const char *kmem_cache_name(struct kmem_cache *cachep) |
1944972d ACM |
3788 | { |
3789 | return cachep->name; | |
3790 | } | |
3791 | EXPORT_SYMBOL_GPL(kmem_cache_name); | |
3792 | ||
e498be7d | 3793 | /* |
183ff22b | 3794 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3795 | */ |
83b519e8 | 3796 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3797 | { |
3798 | int node; | |
3799 | struct kmem_list3 *l3; | |
cafeb02e | 3800 | struct array_cache *new_shared; |
3395ee05 | 3801 | struct array_cache **new_alien = NULL; |
e498be7d | 3802 | |
9c09a95c | 3803 | for_each_online_node(node) { |
cafeb02e | 3804 | |
3395ee05 | 3805 | if (use_alien_caches) { |
83b519e8 | 3806 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3807 | if (!new_alien) |
3808 | goto fail; | |
3809 | } | |
cafeb02e | 3810 | |
63109846 ED |
3811 | new_shared = NULL; |
3812 | if (cachep->shared) { | |
3813 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3814 | cachep->shared*cachep->batchcount, |
83b519e8 | 3815 | 0xbaadf00d, gfp); |
63109846 ED |
3816 | if (!new_shared) { |
3817 | free_alien_cache(new_alien); | |
3818 | goto fail; | |
3819 | } | |
0718dc2a | 3820 | } |
cafeb02e | 3821 | |
a737b3e2 AM |
3822 | l3 = cachep->nodelists[node]; |
3823 | if (l3) { | |
cafeb02e CL |
3824 | struct array_cache *shared = l3->shared; |
3825 | ||
e498be7d CL |
3826 | spin_lock_irq(&l3->list_lock); |
3827 | ||
cafeb02e | 3828 | if (shared) |
0718dc2a CL |
3829 | free_block(cachep, shared->entry, |
3830 | shared->avail, node); | |
e498be7d | 3831 | |
cafeb02e CL |
3832 | l3->shared = new_shared; |
3833 | if (!l3->alien) { | |
e498be7d CL |
3834 | l3->alien = new_alien; |
3835 | new_alien = NULL; | |
3836 | } | |
b28a02de | 3837 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3838 | cachep->batchcount + cachep->num; |
e498be7d | 3839 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3840 | kfree(shared); |
e498be7d CL |
3841 | free_alien_cache(new_alien); |
3842 | continue; | |
3843 | } | |
83b519e8 | 3844 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3845 | if (!l3) { |
3846 | free_alien_cache(new_alien); | |
3847 | kfree(new_shared); | |
e498be7d | 3848 | goto fail; |
0718dc2a | 3849 | } |
e498be7d CL |
3850 | |
3851 | kmem_list3_init(l3); | |
3852 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3853 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3854 | l3->shared = new_shared; |
e498be7d | 3855 | l3->alien = new_alien; |
b28a02de | 3856 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3857 | cachep->batchcount + cachep->num; |
e498be7d CL |
3858 | cachep->nodelists[node] = l3; |
3859 | } | |
cafeb02e | 3860 | return 0; |
0718dc2a | 3861 | |
a737b3e2 | 3862 | fail: |
0718dc2a CL |
3863 | if (!cachep->next.next) { |
3864 | /* Cache is not active yet. Roll back what we did */ | |
3865 | node--; | |
3866 | while (node >= 0) { | |
3867 | if (cachep->nodelists[node]) { | |
3868 | l3 = cachep->nodelists[node]; | |
3869 | ||
3870 | kfree(l3->shared); | |
3871 | free_alien_cache(l3->alien); | |
3872 | kfree(l3); | |
3873 | cachep->nodelists[node] = NULL; | |
3874 | } | |
3875 | node--; | |
3876 | } | |
3877 | } | |
cafeb02e | 3878 | return -ENOMEM; |
e498be7d CL |
3879 | } |
3880 | ||
1da177e4 | 3881 | struct ccupdate_struct { |
343e0d7a | 3882 | struct kmem_cache *cachep; |
1da177e4 LT |
3883 | struct array_cache *new[NR_CPUS]; |
3884 | }; | |
3885 | ||
3886 | static void do_ccupdate_local(void *info) | |
3887 | { | |
a737b3e2 | 3888 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3889 | struct array_cache *old; |
3890 | ||
3891 | check_irq_off(); | |
9a2dba4b | 3892 | old = cpu_cache_get(new->cachep); |
e498be7d | 3893 | |
1da177e4 LT |
3894 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3895 | new->new[smp_processor_id()] = old; | |
3896 | } | |
3897 | ||
b5d8ca7c | 3898 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 3899 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 3900 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 3901 | { |
d2e7b7d0 | 3902 | struct ccupdate_struct *new; |
2ed3a4ef | 3903 | int i; |
1da177e4 | 3904 | |
83b519e8 | 3905 | new = kzalloc(sizeof(*new), gfp); |
d2e7b7d0 SS |
3906 | if (!new) |
3907 | return -ENOMEM; | |
3908 | ||
e498be7d | 3909 | for_each_online_cpu(i) { |
d2e7b7d0 | 3910 | new->new[i] = alloc_arraycache(cpu_to_node(i), limit, |
83b519e8 | 3911 | batchcount, gfp); |
d2e7b7d0 | 3912 | if (!new->new[i]) { |
b28a02de | 3913 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3914 | kfree(new->new[i]); |
3915 | kfree(new); | |
e498be7d | 3916 | return -ENOMEM; |
1da177e4 LT |
3917 | } |
3918 | } | |
d2e7b7d0 | 3919 | new->cachep = cachep; |
1da177e4 | 3920 | |
15c8b6c1 | 3921 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 3922 | |
1da177e4 | 3923 | check_irq_on(); |
1da177e4 LT |
3924 | cachep->batchcount = batchcount; |
3925 | cachep->limit = limit; | |
e498be7d | 3926 | cachep->shared = shared; |
1da177e4 | 3927 | |
e498be7d | 3928 | for_each_online_cpu(i) { |
d2e7b7d0 | 3929 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3930 | if (!ccold) |
3931 | continue; | |
e498be7d | 3932 | spin_lock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
ff69416e | 3933 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_node(i)); |
e498be7d | 3934 | spin_unlock_irq(&cachep->nodelists[cpu_to_node(i)]->list_lock); |
1da177e4 LT |
3935 | kfree(ccold); |
3936 | } | |
d2e7b7d0 | 3937 | kfree(new); |
83b519e8 | 3938 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
3939 | } |
3940 | ||
b5d8ca7c | 3941 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 3942 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
3943 | { |
3944 | int err; | |
3945 | int limit, shared; | |
3946 | ||
a737b3e2 AM |
3947 | /* |
3948 | * The head array serves three purposes: | |
1da177e4 LT |
3949 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3950 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3951 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3952 | * bufctl chains: array operations are cheaper. |
3953 | * The numbers are guessed, we should auto-tune as described by | |
3954 | * Bonwick. | |
3955 | */ | |
3dafccf2 | 3956 | if (cachep->buffer_size > 131072) |
1da177e4 | 3957 | limit = 1; |
3dafccf2 | 3958 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 3959 | limit = 8; |
3dafccf2 | 3960 | else if (cachep->buffer_size > 1024) |
1da177e4 | 3961 | limit = 24; |
3dafccf2 | 3962 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
3963 | limit = 54; |
3964 | else | |
3965 | limit = 120; | |
3966 | ||
a737b3e2 AM |
3967 | /* |
3968 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3969 | * allocation behaviour: Most allocs on one cpu, most free operations |
3970 | * on another cpu. For these cases, an efficient object passing between | |
3971 | * cpus is necessary. This is provided by a shared array. The array | |
3972 | * replaces Bonwick's magazine layer. | |
3973 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3974 | * to a larger limit. Thus disabled by default. | |
3975 | */ | |
3976 | shared = 0; | |
364fbb29 | 3977 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 3978 | shared = 8; |
1da177e4 LT |
3979 | |
3980 | #if DEBUG | |
a737b3e2 AM |
3981 | /* |
3982 | * With debugging enabled, large batchcount lead to excessively long | |
3983 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3984 | */ |
3985 | if (limit > 32) | |
3986 | limit = 32; | |
3987 | #endif | |
83b519e8 | 3988 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
3989 | if (err) |
3990 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3991 | cachep->name, -err); |
2ed3a4ef | 3992 | return err; |
1da177e4 LT |
3993 | } |
3994 | ||
1b55253a CL |
3995 | /* |
3996 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
3997 | * necessary. Note that the l3 listlock also protects the array_cache |
3998 | * if drain_array() is used on the shared array. | |
1b55253a CL |
3999 | */ |
4000 | void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4001 | struct array_cache *ac, int force, int node) | |
1da177e4 LT |
4002 | { |
4003 | int tofree; | |
4004 | ||
1b55253a CL |
4005 | if (!ac || !ac->avail) |
4006 | return; | |
1da177e4 LT |
4007 | if (ac->touched && !force) { |
4008 | ac->touched = 0; | |
b18e7e65 | 4009 | } else { |
1b55253a | 4010 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4011 | if (ac->avail) { |
4012 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4013 | if (tofree > ac->avail) | |
4014 | tofree = (ac->avail + 1) / 2; | |
4015 | free_block(cachep, ac->entry, tofree, node); | |
4016 | ac->avail -= tofree; | |
4017 | memmove(ac->entry, &(ac->entry[tofree]), | |
4018 | sizeof(void *) * ac->avail); | |
4019 | } | |
1b55253a | 4020 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4021 | } |
4022 | } | |
4023 | ||
4024 | /** | |
4025 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4026 | * @w: work descriptor |
1da177e4 LT |
4027 | * |
4028 | * Called from workqueue/eventd every few seconds. | |
4029 | * Purpose: | |
4030 | * - clear the per-cpu caches for this CPU. | |
4031 | * - return freeable pages to the main free memory pool. | |
4032 | * | |
a737b3e2 AM |
4033 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4034 | * again on the next iteration. | |
1da177e4 | 4035 | */ |
7c5cae36 | 4036 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4037 | { |
7a7c381d | 4038 | struct kmem_cache *searchp; |
e498be7d | 4039 | struct kmem_list3 *l3; |
aab2207c | 4040 | int node = numa_node_id(); |
bf6aede7 | 4041 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4042 | |
7c5cae36 | 4043 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4044 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4045 | goto out; |
1da177e4 | 4046 | |
7a7c381d | 4047 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4048 | check_irq_on(); |
4049 | ||
35386e3b CL |
4050 | /* |
4051 | * We only take the l3 lock if absolutely necessary and we | |
4052 | * have established with reasonable certainty that | |
4053 | * we can do some work if the lock was obtained. | |
4054 | */ | |
aab2207c | 4055 | l3 = searchp->nodelists[node]; |
35386e3b | 4056 | |
8fce4d8e | 4057 | reap_alien(searchp, l3); |
1da177e4 | 4058 | |
aab2207c | 4059 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4060 | |
35386e3b CL |
4061 | /* |
4062 | * These are racy checks but it does not matter | |
4063 | * if we skip one check or scan twice. | |
4064 | */ | |
e498be7d | 4065 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4066 | goto next; |
1da177e4 | 4067 | |
e498be7d | 4068 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4069 | |
aab2207c | 4070 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4071 | |
ed11d9eb | 4072 | if (l3->free_touched) |
e498be7d | 4073 | l3->free_touched = 0; |
ed11d9eb CL |
4074 | else { |
4075 | int freed; | |
1da177e4 | 4076 | |
ed11d9eb CL |
4077 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4078 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4079 | STATS_ADD_REAPED(searchp, freed); | |
4080 | } | |
35386e3b | 4081 | next: |
1da177e4 LT |
4082 | cond_resched(); |
4083 | } | |
4084 | check_irq_on(); | |
fc0abb14 | 4085 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4086 | next_reap_node(); |
7c5cae36 | 4087 | out: |
a737b3e2 | 4088 | /* Set up the next iteration */ |
7c5cae36 | 4089 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4090 | } |
4091 | ||
158a9624 | 4092 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4093 | |
85289f98 | 4094 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4095 | { |
85289f98 PE |
4096 | /* |
4097 | * Output format version, so at least we can change it | |
4098 | * without _too_ many complaints. | |
4099 | */ | |
1da177e4 | 4100 | #if STATS |
85289f98 | 4101 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4102 | #else |
85289f98 | 4103 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4104 | #endif |
85289f98 PE |
4105 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4106 | "<objperslab> <pagesperslab>"); | |
4107 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4108 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4109 | #if STATS |
85289f98 | 4110 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4111 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4112 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4113 | #endif |
85289f98 PE |
4114 | seq_putc(m, '\n'); |
4115 | } | |
4116 | ||
4117 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4118 | { | |
4119 | loff_t n = *pos; | |
85289f98 | 4120 | |
fc0abb14 | 4121 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4122 | if (!n) |
4123 | print_slabinfo_header(m); | |
b92151ba PE |
4124 | |
4125 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4126 | } |
4127 | ||
4128 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4129 | { | |
b92151ba | 4130 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4131 | } |
4132 | ||
4133 | static void s_stop(struct seq_file *m, void *p) | |
4134 | { | |
fc0abb14 | 4135 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4136 | } |
4137 | ||
4138 | static int s_show(struct seq_file *m, void *p) | |
4139 | { | |
b92151ba | 4140 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
b28a02de PE |
4141 | struct slab *slabp; |
4142 | unsigned long active_objs; | |
4143 | unsigned long num_objs; | |
4144 | unsigned long active_slabs = 0; | |
4145 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4146 | const char *name; |
1da177e4 | 4147 | char *error = NULL; |
e498be7d CL |
4148 | int node; |
4149 | struct kmem_list3 *l3; | |
1da177e4 | 4150 | |
1da177e4 LT |
4151 | active_objs = 0; |
4152 | num_slabs = 0; | |
e498be7d CL |
4153 | for_each_online_node(node) { |
4154 | l3 = cachep->nodelists[node]; | |
4155 | if (!l3) | |
4156 | continue; | |
4157 | ||
ca3b9b91 RT |
4158 | check_irq_on(); |
4159 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4160 | |
7a7c381d | 4161 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4162 | if (slabp->inuse != cachep->num && !error) |
4163 | error = "slabs_full accounting error"; | |
4164 | active_objs += cachep->num; | |
4165 | active_slabs++; | |
4166 | } | |
7a7c381d | 4167 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4168 | if (slabp->inuse == cachep->num && !error) |
4169 | error = "slabs_partial inuse accounting error"; | |
4170 | if (!slabp->inuse && !error) | |
4171 | error = "slabs_partial/inuse accounting error"; | |
4172 | active_objs += slabp->inuse; | |
4173 | active_slabs++; | |
4174 | } | |
7a7c381d | 4175 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4176 | if (slabp->inuse && !error) |
4177 | error = "slabs_free/inuse accounting error"; | |
4178 | num_slabs++; | |
4179 | } | |
4180 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4181 | if (l3->shared) |
4182 | shared_avail += l3->shared->avail; | |
e498be7d | 4183 | |
ca3b9b91 | 4184 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4185 | } |
b28a02de PE |
4186 | num_slabs += active_slabs; |
4187 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4188 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4189 | error = "free_objects accounting error"; |
4190 | ||
b28a02de | 4191 | name = cachep->name; |
1da177e4 LT |
4192 | if (error) |
4193 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4194 | ||
4195 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4196 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4197 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4198 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4199 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4200 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4201 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4202 | #if STATS |
b28a02de | 4203 | { /* list3 stats */ |
1da177e4 LT |
4204 | unsigned long high = cachep->high_mark; |
4205 | unsigned long allocs = cachep->num_allocations; | |
4206 | unsigned long grown = cachep->grown; | |
4207 | unsigned long reaped = cachep->reaped; | |
4208 | unsigned long errors = cachep->errors; | |
4209 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4210 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4211 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4212 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4213 | |
e498be7d | 4214 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu \ |
fb7faf33 | 4215 | %4lu %4lu %4lu %4lu %4lu", allocs, high, grown, |
a737b3e2 | 4216 | reaped, errors, max_freeable, node_allocs, |
fb7faf33 | 4217 | node_frees, overflows); |
1da177e4 LT |
4218 | } |
4219 | /* cpu stats */ | |
4220 | { | |
4221 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4222 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4223 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4224 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4225 | ||
4226 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4227 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4228 | } |
4229 | #endif | |
4230 | seq_putc(m, '\n'); | |
1da177e4 LT |
4231 | return 0; |
4232 | } | |
4233 | ||
4234 | /* | |
4235 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4236 | * | |
4237 | * Output layout: | |
4238 | * cache-name | |
4239 | * num-active-objs | |
4240 | * total-objs | |
4241 | * object size | |
4242 | * num-active-slabs | |
4243 | * total-slabs | |
4244 | * num-pages-per-slab | |
4245 | * + further values on SMP and with statistics enabled | |
4246 | */ | |
4247 | ||
7b3c3a50 | 4248 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4249 | .start = s_start, |
4250 | .next = s_next, | |
4251 | .stop = s_stop, | |
4252 | .show = s_show, | |
1da177e4 LT |
4253 | }; |
4254 | ||
4255 | #define MAX_SLABINFO_WRITE 128 | |
4256 | /** | |
4257 | * slabinfo_write - Tuning for the slab allocator | |
4258 | * @file: unused | |
4259 | * @buffer: user buffer | |
4260 | * @count: data length | |
4261 | * @ppos: unused | |
4262 | */ | |
b28a02de PE |
4263 | ssize_t slabinfo_write(struct file *file, const char __user * buffer, |
4264 | size_t count, loff_t *ppos) | |
1da177e4 | 4265 | { |
b28a02de | 4266 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4267 | int limit, batchcount, shared, res; |
7a7c381d | 4268 | struct kmem_cache *cachep; |
b28a02de | 4269 | |
1da177e4 LT |
4270 | if (count > MAX_SLABINFO_WRITE) |
4271 | return -EINVAL; | |
4272 | if (copy_from_user(&kbuf, buffer, count)) | |
4273 | return -EFAULT; | |
b28a02de | 4274 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4275 | |
4276 | tmp = strchr(kbuf, ' '); | |
4277 | if (!tmp) | |
4278 | return -EINVAL; | |
4279 | *tmp = '\0'; | |
4280 | tmp++; | |
4281 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4282 | return -EINVAL; | |
4283 | ||
4284 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4285 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4286 | res = -EINVAL; |
7a7c381d | 4287 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4288 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4289 | if (limit < 1 || batchcount < 1 || |
4290 | batchcount > limit || shared < 0) { | |
e498be7d | 4291 | res = 0; |
1da177e4 | 4292 | } else { |
e498be7d | 4293 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4294 | batchcount, shared, |
4295 | GFP_KERNEL); | |
1da177e4 LT |
4296 | } |
4297 | break; | |
4298 | } | |
4299 | } | |
fc0abb14 | 4300 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4301 | if (res >= 0) |
4302 | res = count; | |
4303 | return res; | |
4304 | } | |
871751e2 | 4305 | |
7b3c3a50 AD |
4306 | static int slabinfo_open(struct inode *inode, struct file *file) |
4307 | { | |
4308 | return seq_open(file, &slabinfo_op); | |
4309 | } | |
4310 | ||
4311 | static const struct file_operations proc_slabinfo_operations = { | |
4312 | .open = slabinfo_open, | |
4313 | .read = seq_read, | |
4314 | .write = slabinfo_write, | |
4315 | .llseek = seq_lseek, | |
4316 | .release = seq_release, | |
4317 | }; | |
4318 | ||
871751e2 AV |
4319 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4320 | ||
4321 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4322 | { | |
871751e2 | 4323 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4324 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4325 | } |
4326 | ||
4327 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4328 | { | |
4329 | unsigned long *p; | |
4330 | int l; | |
4331 | if (!v) | |
4332 | return 1; | |
4333 | l = n[1]; | |
4334 | p = n + 2; | |
4335 | while (l) { | |
4336 | int i = l/2; | |
4337 | unsigned long *q = p + 2 * i; | |
4338 | if (*q == v) { | |
4339 | q[1]++; | |
4340 | return 1; | |
4341 | } | |
4342 | if (*q > v) { | |
4343 | l = i; | |
4344 | } else { | |
4345 | p = q + 2; | |
4346 | l -= i + 1; | |
4347 | } | |
4348 | } | |
4349 | if (++n[1] == n[0]) | |
4350 | return 0; | |
4351 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4352 | p[0] = v; | |
4353 | p[1] = 1; | |
4354 | return 1; | |
4355 | } | |
4356 | ||
4357 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4358 | { | |
4359 | void *p; | |
4360 | int i; | |
4361 | if (n[0] == n[1]) | |
4362 | return; | |
4363 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4364 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4365 | continue; | |
4366 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4367 | return; | |
4368 | } | |
4369 | } | |
4370 | ||
4371 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4372 | { | |
4373 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4374 | unsigned long offset, size; |
9281acea | 4375 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4376 | |
a5c43dae | 4377 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4378 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4379 | if (modname[0]) |
871751e2 AV |
4380 | seq_printf(m, " [%s]", modname); |
4381 | return; | |
4382 | } | |
4383 | #endif | |
4384 | seq_printf(m, "%p", (void *)address); | |
4385 | } | |
4386 | ||
4387 | static int leaks_show(struct seq_file *m, void *p) | |
4388 | { | |
b92151ba | 4389 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4390 | struct slab *slabp; |
4391 | struct kmem_list3 *l3; | |
4392 | const char *name; | |
4393 | unsigned long *n = m->private; | |
4394 | int node; | |
4395 | int i; | |
4396 | ||
4397 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4398 | return 0; | |
4399 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4400 | return 0; | |
4401 | ||
4402 | /* OK, we can do it */ | |
4403 | ||
4404 | n[1] = 0; | |
4405 | ||
4406 | for_each_online_node(node) { | |
4407 | l3 = cachep->nodelists[node]; | |
4408 | if (!l3) | |
4409 | continue; | |
4410 | ||
4411 | check_irq_on(); | |
4412 | spin_lock_irq(&l3->list_lock); | |
4413 | ||
7a7c381d | 4414 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4415 | handle_slab(n, cachep, slabp); |
7a7c381d | 4416 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4417 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4418 | spin_unlock_irq(&l3->list_lock); |
4419 | } | |
4420 | name = cachep->name; | |
4421 | if (n[0] == n[1]) { | |
4422 | /* Increase the buffer size */ | |
4423 | mutex_unlock(&cache_chain_mutex); | |
4424 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4425 | if (!m->private) { | |
4426 | /* Too bad, we are really out */ | |
4427 | m->private = n; | |
4428 | mutex_lock(&cache_chain_mutex); | |
4429 | return -ENOMEM; | |
4430 | } | |
4431 | *(unsigned long *)m->private = n[0] * 2; | |
4432 | kfree(n); | |
4433 | mutex_lock(&cache_chain_mutex); | |
4434 | /* Now make sure this entry will be retried */ | |
4435 | m->count = m->size; | |
4436 | return 0; | |
4437 | } | |
4438 | for (i = 0; i < n[1]; i++) { | |
4439 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4440 | show_symbol(m, n[2*i+2]); | |
4441 | seq_putc(m, '\n'); | |
4442 | } | |
d2e7b7d0 | 4443 | |
871751e2 AV |
4444 | return 0; |
4445 | } | |
4446 | ||
a0ec95a8 | 4447 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4448 | .start = leaks_start, |
4449 | .next = s_next, | |
4450 | .stop = s_stop, | |
4451 | .show = leaks_show, | |
4452 | }; | |
a0ec95a8 AD |
4453 | |
4454 | static int slabstats_open(struct inode *inode, struct file *file) | |
4455 | { | |
4456 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4457 | int ret = -ENOMEM; | |
4458 | if (n) { | |
4459 | ret = seq_open(file, &slabstats_op); | |
4460 | if (!ret) { | |
4461 | struct seq_file *m = file->private_data; | |
4462 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4463 | m->private = n; | |
4464 | n = NULL; | |
4465 | } | |
4466 | kfree(n); | |
4467 | } | |
4468 | return ret; | |
4469 | } | |
4470 | ||
4471 | static const struct file_operations proc_slabstats_operations = { | |
4472 | .open = slabstats_open, | |
4473 | .read = seq_read, | |
4474 | .llseek = seq_lseek, | |
4475 | .release = seq_release_private, | |
4476 | }; | |
4477 | #endif | |
4478 | ||
4479 | static int __init slab_proc_init(void) | |
4480 | { | |
7b3c3a50 | 4481 | proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4482 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4483 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4484 | #endif |
a0ec95a8 AD |
4485 | return 0; |
4486 | } | |
4487 | module_init(slab_proc_init); | |
1da177e4 LT |
4488 | #endif |
4489 | ||
00e145b6 MS |
4490 | /** |
4491 | * ksize - get the actual amount of memory allocated for a given object | |
4492 | * @objp: Pointer to the object | |
4493 | * | |
4494 | * kmalloc may internally round up allocations and return more memory | |
4495 | * than requested. ksize() can be used to determine the actual amount of | |
4496 | * memory allocated. The caller may use this additional memory, even though | |
4497 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4498 | * The caller must guarantee that objp points to a valid object previously | |
4499 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4500 | * must not be freed during the duration of the call. | |
4501 | */ | |
fd76bab2 | 4502 | size_t ksize(const void *objp) |
1da177e4 | 4503 | { |
ef8b4520 CL |
4504 | BUG_ON(!objp); |
4505 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4506 | return 0; |
1da177e4 | 4507 | |
6ed5eb22 | 4508 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4509 | } |
b1aabecd | 4510 | EXPORT_SYMBOL(ksize); |