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 | |
18004c5d | 71 | * The global cache-chain is protected by the mutex 'slab_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> | |
105 | #include <linux/rcupdate.h> | |
543537bd | 106 | #include <linux/string.h> |
138ae663 | 107 | #include <linux/uaccess.h> |
e498be7d | 108 | #include <linux/nodemask.h> |
d5cff635 | 109 | #include <linux/kmemleak.h> |
dc85da15 | 110 | #include <linux/mempolicy.h> |
fc0abb14 | 111 | #include <linux/mutex.h> |
8a8b6502 | 112 | #include <linux/fault-inject.h> |
e7eebaf6 | 113 | #include <linux/rtmutex.h> |
6a2d7a95 | 114 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 115 | #include <linux/debugobjects.h> |
c175eea4 | 116 | #include <linux/kmemcheck.h> |
8f9f8d9e | 117 | #include <linux/memory.h> |
268bb0ce | 118 | #include <linux/prefetch.h> |
1da177e4 | 119 | |
381760ea MG |
120 | #include <net/sock.h> |
121 | ||
1da177e4 LT |
122 | #include <asm/cacheflush.h> |
123 | #include <asm/tlbflush.h> | |
124 | #include <asm/page.h> | |
125 | ||
4dee6b64 SR |
126 | #include <trace/events/kmem.h> |
127 | ||
072bb0aa MG |
128 | #include "internal.h" |
129 | ||
b9ce5ef4 GC |
130 | #include "slab.h" |
131 | ||
1da177e4 | 132 | /* |
50953fe9 | 133 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
134 | * 0 for faster, smaller code (especially in the critical paths). |
135 | * | |
136 | * STATS - 1 to collect stats for /proc/slabinfo. | |
137 | * 0 for faster, smaller code (especially in the critical paths). | |
138 | * | |
139 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
140 | */ | |
141 | ||
142 | #ifdef CONFIG_DEBUG_SLAB | |
143 | #define DEBUG 1 | |
144 | #define STATS 1 | |
145 | #define FORCED_DEBUG 1 | |
146 | #else | |
147 | #define DEBUG 0 | |
148 | #define STATS 0 | |
149 | #define FORCED_DEBUG 0 | |
150 | #endif | |
151 | ||
1da177e4 LT |
152 | /* Shouldn't this be in a header file somewhere? */ |
153 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 154 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 155 | |
1da177e4 LT |
156 | #ifndef ARCH_KMALLOC_FLAGS |
157 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
158 | #endif | |
159 | ||
072bb0aa MG |
160 | /* |
161 | * true if a page was allocated from pfmemalloc reserves for network-based | |
162 | * swap | |
163 | */ | |
164 | static bool pfmemalloc_active __read_mostly; | |
165 | ||
1da177e4 LT |
166 | /* |
167 | * struct array_cache | |
168 | * | |
1da177e4 LT |
169 | * Purpose: |
170 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
171 | * - reduce the number of linked list operations | |
172 | * - reduce spinlock operations | |
173 | * | |
174 | * The limit is stored in the per-cpu structure to reduce the data cache | |
175 | * footprint. | |
176 | * | |
177 | */ | |
178 | struct array_cache { | |
179 | unsigned int avail; | |
180 | unsigned int limit; | |
181 | unsigned int batchcount; | |
182 | unsigned int touched; | |
e498be7d | 183 | spinlock_t lock; |
bda5b655 | 184 | void *entry[]; /* |
a737b3e2 AM |
185 | * Must have this definition in here for the proper |
186 | * alignment of array_cache. Also simplifies accessing | |
187 | * the entries. | |
072bb0aa MG |
188 | * |
189 | * Entries should not be directly dereferenced as | |
190 | * entries belonging to slabs marked pfmemalloc will | |
191 | * have the lower bits set SLAB_OBJ_PFMEMALLOC | |
a737b3e2 | 192 | */ |
1da177e4 LT |
193 | }; |
194 | ||
072bb0aa MG |
195 | #define SLAB_OBJ_PFMEMALLOC 1 |
196 | static inline bool is_obj_pfmemalloc(void *objp) | |
197 | { | |
198 | return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC; | |
199 | } | |
200 | ||
201 | static inline void set_obj_pfmemalloc(void **objp) | |
202 | { | |
203 | *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC); | |
204 | return; | |
205 | } | |
206 | ||
207 | static inline void clear_obj_pfmemalloc(void **objp) | |
208 | { | |
209 | *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC); | |
210 | } | |
211 | ||
a737b3e2 AM |
212 | /* |
213 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
214 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
215 | */ |
216 | #define BOOT_CPUCACHE_ENTRIES 1 | |
217 | struct arraycache_init { | |
218 | struct array_cache cache; | |
b28a02de | 219 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
220 | }; |
221 | ||
e498be7d CL |
222 | /* |
223 | * Need this for bootstrapping a per node allocator. | |
224 | */ | |
556a169d | 225 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
ce8eb6c4 | 226 | static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS]; |
e498be7d | 227 | #define CACHE_CACHE 0 |
556a169d | 228 | #define SIZE_AC MAX_NUMNODES |
ce8eb6c4 | 229 | #define SIZE_NODE (2 * MAX_NUMNODES) |
e498be7d | 230 | |
ed11d9eb | 231 | static int drain_freelist(struct kmem_cache *cache, |
ce8eb6c4 | 232 | struct kmem_cache_node *n, int tofree); |
ed11d9eb CL |
233 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, |
234 | int node); | |
83b519e8 | 235 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 236 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 237 | |
e0a42726 IM |
238 | static int slab_early_init = 1; |
239 | ||
e3366016 | 240 | #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init)) |
ce8eb6c4 | 241 | #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node)) |
1da177e4 | 242 | |
ce8eb6c4 | 243 | static void kmem_cache_node_init(struct kmem_cache_node *parent) |
e498be7d CL |
244 | { |
245 | INIT_LIST_HEAD(&parent->slabs_full); | |
246 | INIT_LIST_HEAD(&parent->slabs_partial); | |
247 | INIT_LIST_HEAD(&parent->slabs_free); | |
248 | parent->shared = NULL; | |
249 | parent->alien = NULL; | |
2e1217cf | 250 | parent->colour_next = 0; |
e498be7d CL |
251 | spin_lock_init(&parent->list_lock); |
252 | parent->free_objects = 0; | |
253 | parent->free_touched = 0; | |
254 | } | |
255 | ||
a737b3e2 AM |
256 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
257 | do { \ | |
258 | INIT_LIST_HEAD(listp); \ | |
6a67368c | 259 | list_splice(&(cachep->node[nodeid]->slab), listp); \ |
e498be7d CL |
260 | } while (0) |
261 | ||
a737b3e2 AM |
262 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
263 | do { \ | |
e498be7d CL |
264 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
265 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
266 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
267 | } while (0) | |
1da177e4 | 268 | |
1da177e4 LT |
269 | #define CFLGS_OFF_SLAB (0x80000000UL) |
270 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
271 | ||
272 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
273 | /* |
274 | * Optimization question: fewer reaps means less probability for unnessary | |
275 | * cpucache drain/refill cycles. | |
1da177e4 | 276 | * |
dc6f3f27 | 277 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
278 | * which could lock up otherwise freeable slabs. |
279 | */ | |
280 | #define REAPTIMEOUT_CPUC (2*HZ) | |
281 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
282 | ||
283 | #if STATS | |
284 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
285 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
286 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
287 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 288 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
289 | #define STATS_SET_HIGH(x) \ |
290 | do { \ | |
291 | if ((x)->num_active > (x)->high_mark) \ | |
292 | (x)->high_mark = (x)->num_active; \ | |
293 | } while (0) | |
1da177e4 LT |
294 | #define STATS_INC_ERR(x) ((x)->errors++) |
295 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 296 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 297 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
298 | #define STATS_SET_FREEABLE(x, i) \ |
299 | do { \ | |
300 | if ((x)->max_freeable < i) \ | |
301 | (x)->max_freeable = i; \ | |
302 | } while (0) | |
1da177e4 LT |
303 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
304 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
305 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
306 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
307 | #else | |
308 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
309 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
310 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
311 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 312 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
313 | #define STATS_SET_HIGH(x) do { } while (0) |
314 | #define STATS_INC_ERR(x) do { } while (0) | |
315 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 316 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 317 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 318 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
319 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
320 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
321 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
322 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
323 | #endif | |
324 | ||
325 | #if DEBUG | |
1da177e4 | 326 | |
a737b3e2 AM |
327 | /* |
328 | * memory layout of objects: | |
1da177e4 | 329 | * 0 : objp |
3dafccf2 | 330 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
331 | * the end of an object is aligned with the end of the real |
332 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 333 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 334 | * redzone word. |
3dafccf2 | 335 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
336 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
337 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 338 | * [BYTES_PER_WORD long] |
1da177e4 | 339 | */ |
343e0d7a | 340 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 341 | { |
3dafccf2 | 342 | return cachep->obj_offset; |
1da177e4 LT |
343 | } |
344 | ||
b46b8f19 | 345 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
346 | { |
347 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
348 | return (unsigned long long*) (objp + obj_offset(cachep) - |
349 | sizeof(unsigned long long)); | |
1da177e4 LT |
350 | } |
351 | ||
b46b8f19 | 352 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
353 | { |
354 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
355 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 356 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 357 | sizeof(unsigned long long) - |
87a927c7 | 358 | REDZONE_ALIGN); |
3b0efdfa | 359 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 360 | sizeof(unsigned long long)); |
1da177e4 LT |
361 | } |
362 | ||
343e0d7a | 363 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
364 | { |
365 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 366 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
367 | } |
368 | ||
369 | #else | |
370 | ||
3dafccf2 | 371 | #define obj_offset(x) 0 |
b46b8f19 DW |
372 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
373 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
374 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
375 | ||
376 | #endif | |
377 | ||
1da177e4 | 378 | /* |
3df1cccd DR |
379 | * Do not go above this order unless 0 objects fit into the slab or |
380 | * overridden on the command line. | |
1da177e4 | 381 | */ |
543585cc DR |
382 | #define SLAB_MAX_ORDER_HI 1 |
383 | #define SLAB_MAX_ORDER_LO 0 | |
384 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 385 | static bool slab_max_order_set __initdata; |
1da177e4 | 386 | |
6ed5eb22 PE |
387 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
388 | { | |
b49af68f | 389 | struct page *page = virt_to_head_page(obj); |
35026088 | 390 | return page->slab_cache; |
6ed5eb22 PE |
391 | } |
392 | ||
8456a648 | 393 | static inline void *index_to_obj(struct kmem_cache *cache, struct page *page, |
8fea4e96 PE |
394 | unsigned int idx) |
395 | { | |
8456a648 | 396 | return page->s_mem + cache->size * idx; |
8fea4e96 PE |
397 | } |
398 | ||
6a2d7a95 | 399 | /* |
3b0efdfa CL |
400 | * We want to avoid an expensive divide : (offset / cache->size) |
401 | * Using the fact that size is a constant for a particular cache, | |
402 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
403 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
404 | */ | |
405 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
8456a648 | 406 | const struct page *page, void *obj) |
8fea4e96 | 407 | { |
8456a648 | 408 | u32 offset = (obj - page->s_mem); |
6a2d7a95 | 409 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); |
8fea4e96 PE |
410 | } |
411 | ||
1da177e4 | 412 | static struct arraycache_init initarray_generic = |
b28a02de | 413 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
414 | |
415 | /* internal cache of cache description objs */ | |
9b030cb8 | 416 | static struct kmem_cache kmem_cache_boot = { |
b28a02de PE |
417 | .batchcount = 1, |
418 | .limit = BOOT_CPUCACHE_ENTRIES, | |
419 | .shared = 1, | |
3b0efdfa | 420 | .size = sizeof(struct kmem_cache), |
b28a02de | 421 | .name = "kmem_cache", |
1da177e4 LT |
422 | }; |
423 | ||
056c6241 RT |
424 | #define BAD_ALIEN_MAGIC 0x01020304ul |
425 | ||
f1aaee53 AV |
426 | #ifdef CONFIG_LOCKDEP |
427 | ||
428 | /* | |
429 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
430 | * for other slabs "off slab". | |
431 | * The locking for this is tricky in that it nests within the locks | |
432 | * of all other slabs in a few places; to deal with this special | |
433 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
434 | * |
435 | * We set lock class for alien array caches which are up during init. | |
436 | * The lock annotation will be lost if all cpus of a node goes down and | |
437 | * then comes back up during hotplug | |
f1aaee53 | 438 | */ |
056c6241 RT |
439 | static struct lock_class_key on_slab_l3_key; |
440 | static struct lock_class_key on_slab_alc_key; | |
441 | ||
83835b3d PZ |
442 | static struct lock_class_key debugobj_l3_key; |
443 | static struct lock_class_key debugobj_alc_key; | |
444 | ||
445 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
446 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
447 | int q) | |
448 | { | |
449 | struct array_cache **alc; | |
ce8eb6c4 | 450 | struct kmem_cache_node *n; |
83835b3d PZ |
451 | int r; |
452 | ||
ce8eb6c4 CL |
453 | n = cachep->node[q]; |
454 | if (!n) | |
83835b3d PZ |
455 | return; |
456 | ||
ce8eb6c4 CL |
457 | lockdep_set_class(&n->list_lock, l3_key); |
458 | alc = n->alien; | |
83835b3d PZ |
459 | /* |
460 | * FIXME: This check for BAD_ALIEN_MAGIC | |
461 | * should go away when common slab code is taught to | |
462 | * work even without alien caches. | |
463 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
464 | * for alloc_alien_cache, | |
465 | */ | |
466 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
467 | return; | |
468 | for_each_node(r) { | |
469 | if (alc[r]) | |
470 | lockdep_set_class(&alc[r]->lock, alc_key); | |
471 | } | |
472 | } | |
473 | ||
474 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
475 | { | |
476 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
477 | } | |
478 | ||
479 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
480 | { | |
481 | int node; | |
482 | ||
483 | for_each_online_node(node) | |
484 | slab_set_debugobj_lock_classes_node(cachep, node); | |
485 | } | |
486 | ||
ce79ddc8 | 487 | static void init_node_lock_keys(int q) |
f1aaee53 | 488 | { |
e3366016 | 489 | int i; |
056c6241 | 490 | |
97d06609 | 491 | if (slab_state < UP) |
ce79ddc8 PE |
492 | return; |
493 | ||
0f8f8094 | 494 | for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) { |
ce8eb6c4 | 495 | struct kmem_cache_node *n; |
e3366016 CL |
496 | struct kmem_cache *cache = kmalloc_caches[i]; |
497 | ||
498 | if (!cache) | |
499 | continue; | |
ce79ddc8 | 500 | |
ce8eb6c4 CL |
501 | n = cache->node[q]; |
502 | if (!n || OFF_SLAB(cache)) | |
00afa758 | 503 | continue; |
83835b3d | 504 | |
e3366016 | 505 | slab_set_lock_classes(cache, &on_slab_l3_key, |
83835b3d | 506 | &on_slab_alc_key, q); |
f1aaee53 AV |
507 | } |
508 | } | |
ce79ddc8 | 509 | |
6ccfb5bc GC |
510 | static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q) |
511 | { | |
6a67368c | 512 | if (!cachep->node[q]) |
6ccfb5bc GC |
513 | return; |
514 | ||
515 | slab_set_lock_classes(cachep, &on_slab_l3_key, | |
516 | &on_slab_alc_key, q); | |
517 | } | |
518 | ||
519 | static inline void on_slab_lock_classes(struct kmem_cache *cachep) | |
520 | { | |
521 | int node; | |
522 | ||
523 | VM_BUG_ON(OFF_SLAB(cachep)); | |
524 | for_each_node(node) | |
525 | on_slab_lock_classes_node(cachep, node); | |
526 | } | |
527 | ||
ce79ddc8 PE |
528 | static inline void init_lock_keys(void) |
529 | { | |
530 | int node; | |
531 | ||
532 | for_each_node(node) | |
533 | init_node_lock_keys(node); | |
534 | } | |
f1aaee53 | 535 | #else |
ce79ddc8 PE |
536 | static void init_node_lock_keys(int q) |
537 | { | |
538 | } | |
539 | ||
056c6241 | 540 | static inline void init_lock_keys(void) |
f1aaee53 AV |
541 | { |
542 | } | |
83835b3d | 543 | |
6ccfb5bc GC |
544 | static inline void on_slab_lock_classes(struct kmem_cache *cachep) |
545 | { | |
546 | } | |
547 | ||
548 | static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node) | |
549 | { | |
550 | } | |
551 | ||
83835b3d PZ |
552 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) |
553 | { | |
554 | } | |
555 | ||
556 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
557 | { | |
558 | } | |
f1aaee53 AV |
559 | #endif |
560 | ||
1871e52c | 561 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 562 | |
343e0d7a | 563 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
564 | { |
565 | return cachep->array[smp_processor_id()]; | |
566 | } | |
567 | ||
fbaccacf | 568 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 569 | { |
8456a648 | 570 | return ALIGN(nr_objs * sizeof(unsigned int), align); |
fbaccacf | 571 | } |
1da177e4 | 572 | |
a737b3e2 AM |
573 | /* |
574 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
575 | */ | |
fbaccacf SR |
576 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
577 | size_t align, int flags, size_t *left_over, | |
578 | unsigned int *num) | |
579 | { | |
580 | int nr_objs; | |
581 | size_t mgmt_size; | |
582 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 583 | |
fbaccacf SR |
584 | /* |
585 | * The slab management structure can be either off the slab or | |
586 | * on it. For the latter case, the memory allocated for a | |
587 | * slab is used for: | |
588 | * | |
16025177 | 589 | * - One unsigned int for each object |
fbaccacf SR |
590 | * - Padding to respect alignment of @align |
591 | * - @buffer_size bytes for each object | |
592 | * | |
593 | * If the slab management structure is off the slab, then the | |
594 | * alignment will already be calculated into the size. Because | |
595 | * the slabs are all pages aligned, the objects will be at the | |
596 | * correct alignment when allocated. | |
597 | */ | |
598 | if (flags & CFLGS_OFF_SLAB) { | |
599 | mgmt_size = 0; | |
600 | nr_objs = slab_size / buffer_size; | |
601 | ||
fbaccacf SR |
602 | } else { |
603 | /* | |
604 | * Ignore padding for the initial guess. The padding | |
605 | * is at most @align-1 bytes, and @buffer_size is at | |
606 | * least @align. In the worst case, this result will | |
607 | * be one greater than the number of objects that fit | |
608 | * into the memory allocation when taking the padding | |
609 | * into account. | |
610 | */ | |
8456a648 | 611 | nr_objs = (slab_size) / (buffer_size + sizeof(unsigned int)); |
fbaccacf SR |
612 | |
613 | /* | |
614 | * This calculated number will be either the right | |
615 | * amount, or one greater than what we want. | |
616 | */ | |
617 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
618 | > slab_size) | |
619 | nr_objs--; | |
620 | ||
fbaccacf SR |
621 | mgmt_size = slab_mgmt_size(nr_objs, align); |
622 | } | |
623 | *num = nr_objs; | |
624 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
625 | } |
626 | ||
f28510d3 | 627 | #if DEBUG |
d40cee24 | 628 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 629 | |
a737b3e2 AM |
630 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
631 | char *msg) | |
1da177e4 LT |
632 | { |
633 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 634 | function, cachep->name, msg); |
1da177e4 | 635 | dump_stack(); |
373d4d09 | 636 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
1da177e4 | 637 | } |
f28510d3 | 638 | #endif |
1da177e4 | 639 | |
3395ee05 PM |
640 | /* |
641 | * By default on NUMA we use alien caches to stage the freeing of | |
642 | * objects allocated from other nodes. This causes massive memory | |
643 | * inefficiencies when using fake NUMA setup to split memory into a | |
644 | * large number of small nodes, so it can be disabled on the command | |
645 | * line | |
646 | */ | |
647 | ||
648 | static int use_alien_caches __read_mostly = 1; | |
649 | static int __init noaliencache_setup(char *s) | |
650 | { | |
651 | use_alien_caches = 0; | |
652 | return 1; | |
653 | } | |
654 | __setup("noaliencache", noaliencache_setup); | |
655 | ||
3df1cccd DR |
656 | static int __init slab_max_order_setup(char *str) |
657 | { | |
658 | get_option(&str, &slab_max_order); | |
659 | slab_max_order = slab_max_order < 0 ? 0 : | |
660 | min(slab_max_order, MAX_ORDER - 1); | |
661 | slab_max_order_set = true; | |
662 | ||
663 | return 1; | |
664 | } | |
665 | __setup("slab_max_order=", slab_max_order_setup); | |
666 | ||
8fce4d8e CL |
667 | #ifdef CONFIG_NUMA |
668 | /* | |
669 | * Special reaping functions for NUMA systems called from cache_reap(). | |
670 | * These take care of doing round robin flushing of alien caches (containing | |
671 | * objects freed on different nodes from which they were allocated) and the | |
672 | * flushing of remote pcps by calling drain_node_pages. | |
673 | */ | |
1871e52c | 674 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
675 | |
676 | static void init_reap_node(int cpu) | |
677 | { | |
678 | int node; | |
679 | ||
7d6e6d09 | 680 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 681 | if (node == MAX_NUMNODES) |
442295c9 | 682 | node = first_node(node_online_map); |
8fce4d8e | 683 | |
1871e52c | 684 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
685 | } |
686 | ||
687 | static void next_reap_node(void) | |
688 | { | |
909ea964 | 689 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 690 | |
8fce4d8e CL |
691 | node = next_node(node, node_online_map); |
692 | if (unlikely(node >= MAX_NUMNODES)) | |
693 | node = first_node(node_online_map); | |
909ea964 | 694 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
695 | } |
696 | ||
697 | #else | |
698 | #define init_reap_node(cpu) do { } while (0) | |
699 | #define next_reap_node(void) do { } while (0) | |
700 | #endif | |
701 | ||
1da177e4 LT |
702 | /* |
703 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
704 | * via the workqueue/eventd. | |
705 | * Add the CPU number into the expiration time to minimize the possibility of | |
706 | * the CPUs getting into lockstep and contending for the global cache chain | |
707 | * lock. | |
708 | */ | |
0db0628d | 709 | static void start_cpu_timer(int cpu) |
1da177e4 | 710 | { |
1871e52c | 711 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
712 | |
713 | /* | |
714 | * When this gets called from do_initcalls via cpucache_init(), | |
715 | * init_workqueues() has already run, so keventd will be setup | |
716 | * at that time. | |
717 | */ | |
52bad64d | 718 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 719 | init_reap_node(cpu); |
203b42f7 | 720 | INIT_DEFERRABLE_WORK(reap_work, cache_reap); |
2b284214 AV |
721 | schedule_delayed_work_on(cpu, reap_work, |
722 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
723 | } |
724 | } | |
725 | ||
e498be7d | 726 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 727 | int batchcount, gfp_t gfp) |
1da177e4 | 728 | { |
b28a02de | 729 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
730 | struct array_cache *nc = NULL; |
731 | ||
83b519e8 | 732 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
733 | /* |
734 | * The array_cache structures contain pointers to free object. | |
25985edc | 735 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
736 | * cache the pointers are not cleared and they could be counted as |
737 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
738 | * not scan such objects. | |
739 | */ | |
740 | kmemleak_no_scan(nc); | |
1da177e4 LT |
741 | if (nc) { |
742 | nc->avail = 0; | |
743 | nc->limit = entries; | |
744 | nc->batchcount = batchcount; | |
745 | nc->touched = 0; | |
e498be7d | 746 | spin_lock_init(&nc->lock); |
1da177e4 LT |
747 | } |
748 | return nc; | |
749 | } | |
750 | ||
8456a648 | 751 | static inline bool is_slab_pfmemalloc(struct page *page) |
072bb0aa | 752 | { |
8456a648 | 753 | struct page *mem_page = virt_to_page(page->s_mem); |
072bb0aa | 754 | |
8456a648 | 755 | return PageSlabPfmemalloc(mem_page); |
072bb0aa MG |
756 | } |
757 | ||
758 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
759 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
760 | struct array_cache *ac) | |
761 | { | |
ce8eb6c4 | 762 | struct kmem_cache_node *n = cachep->node[numa_mem_id()]; |
8456a648 | 763 | struct page *page; |
072bb0aa MG |
764 | unsigned long flags; |
765 | ||
766 | if (!pfmemalloc_active) | |
767 | return; | |
768 | ||
ce8eb6c4 | 769 | spin_lock_irqsave(&n->list_lock, flags); |
8456a648 JK |
770 | list_for_each_entry(page, &n->slabs_full, lru) |
771 | if (is_slab_pfmemalloc(page)) | |
072bb0aa MG |
772 | goto out; |
773 | ||
8456a648 JK |
774 | list_for_each_entry(page, &n->slabs_partial, lru) |
775 | if (is_slab_pfmemalloc(page)) | |
072bb0aa MG |
776 | goto out; |
777 | ||
8456a648 JK |
778 | list_for_each_entry(page, &n->slabs_free, lru) |
779 | if (is_slab_pfmemalloc(page)) | |
072bb0aa MG |
780 | goto out; |
781 | ||
782 | pfmemalloc_active = false; | |
783 | out: | |
ce8eb6c4 | 784 | spin_unlock_irqrestore(&n->list_lock, flags); |
072bb0aa MG |
785 | } |
786 | ||
381760ea | 787 | static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, |
072bb0aa MG |
788 | gfp_t flags, bool force_refill) |
789 | { | |
790 | int i; | |
791 | void *objp = ac->entry[--ac->avail]; | |
792 | ||
793 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
794 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
ce8eb6c4 | 795 | struct kmem_cache_node *n; |
072bb0aa MG |
796 | |
797 | if (gfp_pfmemalloc_allowed(flags)) { | |
798 | clear_obj_pfmemalloc(&objp); | |
799 | return objp; | |
800 | } | |
801 | ||
802 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
d014dc2e | 803 | for (i = 0; i < ac->avail; i++) { |
072bb0aa MG |
804 | /* If a !PFMEMALLOC object is found, swap them */ |
805 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
806 | objp = ac->entry[i]; | |
807 | ac->entry[i] = ac->entry[ac->avail]; | |
808 | ac->entry[ac->avail] = objp; | |
809 | return objp; | |
810 | } | |
811 | } | |
812 | ||
813 | /* | |
814 | * If there are empty slabs on the slabs_free list and we are | |
815 | * being forced to refill the cache, mark this one !pfmemalloc. | |
816 | */ | |
ce8eb6c4 CL |
817 | n = cachep->node[numa_mem_id()]; |
818 | if (!list_empty(&n->slabs_free) && force_refill) { | |
8456a648 JK |
819 | struct page *page = virt_to_head_page(objp); |
820 | ClearPageSlabPfmemalloc(virt_to_head_page(page->s_mem)); | |
072bb0aa MG |
821 | clear_obj_pfmemalloc(&objp); |
822 | recheck_pfmemalloc_active(cachep, ac); | |
823 | return objp; | |
824 | } | |
825 | ||
826 | /* No !PFMEMALLOC objects available */ | |
827 | ac->avail++; | |
828 | objp = NULL; | |
829 | } | |
830 | ||
831 | return objp; | |
832 | } | |
833 | ||
381760ea MG |
834 | static inline void *ac_get_obj(struct kmem_cache *cachep, |
835 | struct array_cache *ac, gfp_t flags, bool force_refill) | |
836 | { | |
837 | void *objp; | |
838 | ||
839 | if (unlikely(sk_memalloc_socks())) | |
840 | objp = __ac_get_obj(cachep, ac, flags, force_refill); | |
841 | else | |
842 | objp = ac->entry[--ac->avail]; | |
843 | ||
844 | return objp; | |
845 | } | |
846 | ||
847 | static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
072bb0aa MG |
848 | void *objp) |
849 | { | |
850 | if (unlikely(pfmemalloc_active)) { | |
851 | /* Some pfmemalloc slabs exist, check if this is one */ | |
8456a648 JK |
852 | struct page *page = virt_to_head_page(objp); |
853 | struct page *mem_page = virt_to_head_page(page->s_mem); | |
854 | if (PageSlabPfmemalloc(mem_page)) | |
072bb0aa MG |
855 | set_obj_pfmemalloc(&objp); |
856 | } | |
857 | ||
381760ea MG |
858 | return objp; |
859 | } | |
860 | ||
861 | static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
862 | void *objp) | |
863 | { | |
864 | if (unlikely(sk_memalloc_socks())) | |
865 | objp = __ac_put_obj(cachep, ac, objp); | |
866 | ||
072bb0aa MG |
867 | ac->entry[ac->avail++] = objp; |
868 | } | |
869 | ||
3ded175a CL |
870 | /* |
871 | * Transfer objects in one arraycache to another. | |
872 | * Locking must be handled by the caller. | |
873 | * | |
874 | * Return the number of entries transferred. | |
875 | */ | |
876 | static int transfer_objects(struct array_cache *to, | |
877 | struct array_cache *from, unsigned int max) | |
878 | { | |
879 | /* Figure out how many entries to transfer */ | |
732eacc0 | 880 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
881 | |
882 | if (!nr) | |
883 | return 0; | |
884 | ||
885 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
886 | sizeof(void *) *nr); | |
887 | ||
888 | from->avail -= nr; | |
889 | to->avail += nr; | |
3ded175a CL |
890 | return nr; |
891 | } | |
892 | ||
765c4507 CL |
893 | #ifndef CONFIG_NUMA |
894 | ||
895 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
ce8eb6c4 | 896 | #define reap_alien(cachep, n) do { } while (0) |
765c4507 | 897 | |
83b519e8 | 898 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
899 | { |
900 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
901 | } | |
902 | ||
903 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
904 | { | |
905 | } | |
906 | ||
907 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
908 | { | |
909 | return 0; | |
910 | } | |
911 | ||
912 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
913 | gfp_t flags) | |
914 | { | |
915 | return NULL; | |
916 | } | |
917 | ||
8b98c169 | 918 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
919 | gfp_t flags, int nodeid) |
920 | { | |
921 | return NULL; | |
922 | } | |
923 | ||
924 | #else /* CONFIG_NUMA */ | |
925 | ||
8b98c169 | 926 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 927 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 928 | |
83b519e8 | 929 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
930 | { |
931 | struct array_cache **ac_ptr; | |
8ef82866 | 932 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
933 | int i; |
934 | ||
935 | if (limit > 1) | |
936 | limit = 12; | |
f3186a9c | 937 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
938 | if (ac_ptr) { |
939 | for_each_node(i) { | |
f3186a9c | 940 | if (i == node || !node_online(i)) |
e498be7d | 941 | continue; |
83b519e8 | 942 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 943 | if (!ac_ptr[i]) { |
cc550def | 944 | for (i--; i >= 0; i--) |
e498be7d CL |
945 | kfree(ac_ptr[i]); |
946 | kfree(ac_ptr); | |
947 | return NULL; | |
948 | } | |
949 | } | |
950 | } | |
951 | return ac_ptr; | |
952 | } | |
953 | ||
5295a74c | 954 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
955 | { |
956 | int i; | |
957 | ||
958 | if (!ac_ptr) | |
959 | return; | |
e498be7d | 960 | for_each_node(i) |
b28a02de | 961 | kfree(ac_ptr[i]); |
e498be7d CL |
962 | kfree(ac_ptr); |
963 | } | |
964 | ||
343e0d7a | 965 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 966 | struct array_cache *ac, int node) |
e498be7d | 967 | { |
ce8eb6c4 | 968 | struct kmem_cache_node *n = cachep->node[node]; |
e498be7d CL |
969 | |
970 | if (ac->avail) { | |
ce8eb6c4 | 971 | spin_lock(&n->list_lock); |
e00946fe CL |
972 | /* |
973 | * Stuff objects into the remote nodes shared array first. | |
974 | * That way we could avoid the overhead of putting the objects | |
975 | * into the free lists and getting them back later. | |
976 | */ | |
ce8eb6c4 CL |
977 | if (n->shared) |
978 | transfer_objects(n->shared, ac, ac->limit); | |
e00946fe | 979 | |
ff69416e | 980 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d | 981 | ac->avail = 0; |
ce8eb6c4 | 982 | spin_unlock(&n->list_lock); |
e498be7d CL |
983 | } |
984 | } | |
985 | ||
8fce4d8e CL |
986 | /* |
987 | * Called from cache_reap() to regularly drain alien caches round robin. | |
988 | */ | |
ce8eb6c4 | 989 | static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n) |
8fce4d8e | 990 | { |
909ea964 | 991 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 992 | |
ce8eb6c4 CL |
993 | if (n->alien) { |
994 | struct array_cache *ac = n->alien[node]; | |
e00946fe CL |
995 | |
996 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
997 | __drain_alien_cache(cachep, ac, node); |
998 | spin_unlock_irq(&ac->lock); | |
999 | } | |
1000 | } | |
1001 | } | |
1002 | ||
a737b3e2 AM |
1003 | static void drain_alien_cache(struct kmem_cache *cachep, |
1004 | struct array_cache **alien) | |
e498be7d | 1005 | { |
b28a02de | 1006 | int i = 0; |
e498be7d CL |
1007 | struct array_cache *ac; |
1008 | unsigned long flags; | |
1009 | ||
1010 | for_each_online_node(i) { | |
4484ebf1 | 1011 | ac = alien[i]; |
e498be7d CL |
1012 | if (ac) { |
1013 | spin_lock_irqsave(&ac->lock, flags); | |
1014 | __drain_alien_cache(cachep, ac, i); | |
1015 | spin_unlock_irqrestore(&ac->lock, flags); | |
1016 | } | |
1017 | } | |
1018 | } | |
729bd0b7 | 1019 | |
873623df | 1020 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 | 1021 | { |
1ea991b0 | 1022 | int nodeid = page_to_nid(virt_to_page(objp)); |
ce8eb6c4 | 1023 | struct kmem_cache_node *n; |
729bd0b7 | 1024 | struct array_cache *alien = NULL; |
1ca4cb24 PE |
1025 | int node; |
1026 | ||
7d6e6d09 | 1027 | node = numa_mem_id(); |
729bd0b7 PE |
1028 | |
1029 | /* | |
1030 | * Make sure we are not freeing a object from another node to the array | |
1031 | * cache on this cpu. | |
1032 | */ | |
1ea991b0 | 1033 | if (likely(nodeid == node)) |
729bd0b7 PE |
1034 | return 0; |
1035 | ||
ce8eb6c4 | 1036 | n = cachep->node[node]; |
729bd0b7 | 1037 | STATS_INC_NODEFREES(cachep); |
ce8eb6c4 CL |
1038 | if (n->alien && n->alien[nodeid]) { |
1039 | alien = n->alien[nodeid]; | |
873623df | 1040 | spin_lock(&alien->lock); |
729bd0b7 PE |
1041 | if (unlikely(alien->avail == alien->limit)) { |
1042 | STATS_INC_ACOVERFLOW(cachep); | |
1043 | __drain_alien_cache(cachep, alien, nodeid); | |
1044 | } | |
072bb0aa | 1045 | ac_put_obj(cachep, alien, objp); |
729bd0b7 PE |
1046 | spin_unlock(&alien->lock); |
1047 | } else { | |
6a67368c | 1048 | spin_lock(&(cachep->node[nodeid])->list_lock); |
729bd0b7 | 1049 | free_block(cachep, &objp, 1, nodeid); |
6a67368c | 1050 | spin_unlock(&(cachep->node[nodeid])->list_lock); |
729bd0b7 PE |
1051 | } |
1052 | return 1; | |
1053 | } | |
e498be7d CL |
1054 | #endif |
1055 | ||
8f9f8d9e | 1056 | /* |
6a67368c | 1057 | * Allocates and initializes node for a node on each slab cache, used for |
ce8eb6c4 | 1058 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node |
8f9f8d9e | 1059 | * will be allocated off-node since memory is not yet online for the new node. |
6a67368c | 1060 | * When hotplugging memory or a cpu, existing node are not replaced if |
8f9f8d9e DR |
1061 | * already in use. |
1062 | * | |
18004c5d | 1063 | * Must hold slab_mutex. |
8f9f8d9e | 1064 | */ |
6a67368c | 1065 | static int init_cache_node_node(int node) |
8f9f8d9e DR |
1066 | { |
1067 | struct kmem_cache *cachep; | |
ce8eb6c4 | 1068 | struct kmem_cache_node *n; |
6744f087 | 1069 | const int memsize = sizeof(struct kmem_cache_node); |
8f9f8d9e | 1070 | |
18004c5d | 1071 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1072 | /* |
1073 | * Set up the size64 kmemlist for cpu before we can | |
1074 | * begin anything. Make sure some other cpu on this | |
1075 | * node has not already allocated this | |
1076 | */ | |
6a67368c | 1077 | if (!cachep->node[node]) { |
ce8eb6c4 CL |
1078 | n = kmalloc_node(memsize, GFP_KERNEL, node); |
1079 | if (!n) | |
8f9f8d9e | 1080 | return -ENOMEM; |
ce8eb6c4 CL |
1081 | kmem_cache_node_init(n); |
1082 | n->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
8f9f8d9e DR |
1083 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
1084 | ||
1085 | /* | |
1086 | * The l3s don't come and go as CPUs come and | |
18004c5d | 1087 | * go. slab_mutex is sufficient |
8f9f8d9e DR |
1088 | * protection here. |
1089 | */ | |
ce8eb6c4 | 1090 | cachep->node[node] = n; |
8f9f8d9e DR |
1091 | } |
1092 | ||
6a67368c CL |
1093 | spin_lock_irq(&cachep->node[node]->list_lock); |
1094 | cachep->node[node]->free_limit = | |
8f9f8d9e DR |
1095 | (1 + nr_cpus_node(node)) * |
1096 | cachep->batchcount + cachep->num; | |
6a67368c | 1097 | spin_unlock_irq(&cachep->node[node]->list_lock); |
8f9f8d9e DR |
1098 | } |
1099 | return 0; | |
1100 | } | |
1101 | ||
0fa8103b WL |
1102 | static inline int slabs_tofree(struct kmem_cache *cachep, |
1103 | struct kmem_cache_node *n) | |
1104 | { | |
1105 | return (n->free_objects + cachep->num - 1) / cachep->num; | |
1106 | } | |
1107 | ||
0db0628d | 1108 | static void cpuup_canceled(long cpu) |
fbf1e473 AM |
1109 | { |
1110 | struct kmem_cache *cachep; | |
ce8eb6c4 | 1111 | struct kmem_cache_node *n = NULL; |
7d6e6d09 | 1112 | int node = cpu_to_mem(cpu); |
a70f7302 | 1113 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1114 | |
18004c5d | 1115 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1116 | struct array_cache *nc; |
1117 | struct array_cache *shared; | |
1118 | struct array_cache **alien; | |
fbf1e473 | 1119 | |
fbf1e473 AM |
1120 | /* cpu is dead; no one can alloc from it. */ |
1121 | nc = cachep->array[cpu]; | |
1122 | cachep->array[cpu] = NULL; | |
ce8eb6c4 | 1123 | n = cachep->node[node]; |
fbf1e473 | 1124 | |
ce8eb6c4 | 1125 | if (!n) |
fbf1e473 AM |
1126 | goto free_array_cache; |
1127 | ||
ce8eb6c4 | 1128 | spin_lock_irq(&n->list_lock); |
fbf1e473 | 1129 | |
ce8eb6c4 CL |
1130 | /* Free limit for this kmem_cache_node */ |
1131 | n->free_limit -= cachep->batchcount; | |
fbf1e473 AM |
1132 | if (nc) |
1133 | free_block(cachep, nc->entry, nc->avail, node); | |
1134 | ||
58463c1f | 1135 | if (!cpumask_empty(mask)) { |
ce8eb6c4 | 1136 | spin_unlock_irq(&n->list_lock); |
fbf1e473 AM |
1137 | goto free_array_cache; |
1138 | } | |
1139 | ||
ce8eb6c4 | 1140 | shared = n->shared; |
fbf1e473 AM |
1141 | if (shared) { |
1142 | free_block(cachep, shared->entry, | |
1143 | shared->avail, node); | |
ce8eb6c4 | 1144 | n->shared = NULL; |
fbf1e473 AM |
1145 | } |
1146 | ||
ce8eb6c4 CL |
1147 | alien = n->alien; |
1148 | n->alien = NULL; | |
fbf1e473 | 1149 | |
ce8eb6c4 | 1150 | spin_unlock_irq(&n->list_lock); |
fbf1e473 AM |
1151 | |
1152 | kfree(shared); | |
1153 | if (alien) { | |
1154 | drain_alien_cache(cachep, alien); | |
1155 | free_alien_cache(alien); | |
1156 | } | |
1157 | free_array_cache: | |
1158 | kfree(nc); | |
1159 | } | |
1160 | /* | |
1161 | * In the previous loop, all the objects were freed to | |
1162 | * the respective cache's slabs, now we can go ahead and | |
1163 | * shrink each nodelist to its limit. | |
1164 | */ | |
18004c5d | 1165 | list_for_each_entry(cachep, &slab_caches, list) { |
ce8eb6c4 CL |
1166 | n = cachep->node[node]; |
1167 | if (!n) | |
fbf1e473 | 1168 | continue; |
0fa8103b | 1169 | drain_freelist(cachep, n, slabs_tofree(cachep, n)); |
fbf1e473 AM |
1170 | } |
1171 | } | |
1172 | ||
0db0628d | 1173 | static int cpuup_prepare(long cpu) |
1da177e4 | 1174 | { |
343e0d7a | 1175 | struct kmem_cache *cachep; |
ce8eb6c4 | 1176 | struct kmem_cache_node *n = NULL; |
7d6e6d09 | 1177 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1178 | int err; |
1da177e4 | 1179 | |
fbf1e473 AM |
1180 | /* |
1181 | * We need to do this right in the beginning since | |
1182 | * alloc_arraycache's are going to use this list. | |
1183 | * kmalloc_node allows us to add the slab to the right | |
ce8eb6c4 | 1184 | * kmem_cache_node and not this cpu's kmem_cache_node |
fbf1e473 | 1185 | */ |
6a67368c | 1186 | err = init_cache_node_node(node); |
8f9f8d9e DR |
1187 | if (err < 0) |
1188 | goto bad; | |
fbf1e473 AM |
1189 | |
1190 | /* | |
1191 | * Now we can go ahead with allocating the shared arrays and | |
1192 | * array caches | |
1193 | */ | |
18004c5d | 1194 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1195 | struct array_cache *nc; |
1196 | struct array_cache *shared = NULL; | |
1197 | struct array_cache **alien = NULL; | |
1198 | ||
1199 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1200 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1201 | if (!nc) |
1202 | goto bad; | |
1203 | if (cachep->shared) { | |
1204 | shared = alloc_arraycache(node, | |
1205 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1206 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1207 | if (!shared) { |
1208 | kfree(nc); | |
1da177e4 | 1209 | goto bad; |
12d00f6a | 1210 | } |
fbf1e473 AM |
1211 | } |
1212 | if (use_alien_caches) { | |
83b519e8 | 1213 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1214 | if (!alien) { |
1215 | kfree(shared); | |
1216 | kfree(nc); | |
fbf1e473 | 1217 | goto bad; |
12d00f6a | 1218 | } |
fbf1e473 AM |
1219 | } |
1220 | cachep->array[cpu] = nc; | |
ce8eb6c4 CL |
1221 | n = cachep->node[node]; |
1222 | BUG_ON(!n); | |
fbf1e473 | 1223 | |
ce8eb6c4 CL |
1224 | spin_lock_irq(&n->list_lock); |
1225 | if (!n->shared) { | |
fbf1e473 AM |
1226 | /* |
1227 | * We are serialised from CPU_DEAD or | |
1228 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1229 | */ | |
ce8eb6c4 | 1230 | n->shared = shared; |
fbf1e473 AM |
1231 | shared = NULL; |
1232 | } | |
4484ebf1 | 1233 | #ifdef CONFIG_NUMA |
ce8eb6c4 CL |
1234 | if (!n->alien) { |
1235 | n->alien = alien; | |
fbf1e473 | 1236 | alien = NULL; |
1da177e4 | 1237 | } |
fbf1e473 | 1238 | #endif |
ce8eb6c4 | 1239 | spin_unlock_irq(&n->list_lock); |
fbf1e473 AM |
1240 | kfree(shared); |
1241 | free_alien_cache(alien); | |
83835b3d PZ |
1242 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1243 | slab_set_debugobj_lock_classes_node(cachep, node); | |
6ccfb5bc GC |
1244 | else if (!OFF_SLAB(cachep) && |
1245 | !(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
1246 | on_slab_lock_classes_node(cachep, node); | |
fbf1e473 | 1247 | } |
ce79ddc8 PE |
1248 | init_node_lock_keys(node); |
1249 | ||
fbf1e473 AM |
1250 | return 0; |
1251 | bad: | |
12d00f6a | 1252 | cpuup_canceled(cpu); |
fbf1e473 AM |
1253 | return -ENOMEM; |
1254 | } | |
1255 | ||
0db0628d | 1256 | static int cpuup_callback(struct notifier_block *nfb, |
fbf1e473 AM |
1257 | unsigned long action, void *hcpu) |
1258 | { | |
1259 | long cpu = (long)hcpu; | |
1260 | int err = 0; | |
1261 | ||
1262 | switch (action) { | |
fbf1e473 AM |
1263 | case CPU_UP_PREPARE: |
1264 | case CPU_UP_PREPARE_FROZEN: | |
18004c5d | 1265 | mutex_lock(&slab_mutex); |
fbf1e473 | 1266 | err = cpuup_prepare(cpu); |
18004c5d | 1267 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
1268 | break; |
1269 | case CPU_ONLINE: | |
8bb78442 | 1270 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1271 | start_cpu_timer(cpu); |
1272 | break; | |
1273 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1274 | case CPU_DOWN_PREPARE: |
8bb78442 | 1275 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 | 1276 | /* |
18004c5d | 1277 | * Shutdown cache reaper. Note that the slab_mutex is |
5830c590 CL |
1278 | * held so that if cache_reap() is invoked it cannot do |
1279 | * anything expensive but will only modify reap_work | |
1280 | * and reschedule the timer. | |
1281 | */ | |
afe2c511 | 1282 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1283 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1284 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1285 | break; |
1286 | case CPU_DOWN_FAILED: | |
8bb78442 | 1287 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1288 | start_cpu_timer(cpu); |
1289 | break; | |
1da177e4 | 1290 | case CPU_DEAD: |
8bb78442 | 1291 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1292 | /* |
1293 | * Even if all the cpus of a node are down, we don't free the | |
ce8eb6c4 | 1294 | * kmem_cache_node of any cache. This to avoid a race between |
4484ebf1 | 1295 | * cpu_down, and a kmalloc allocation from another cpu for |
ce8eb6c4 | 1296 | * memory from the node of the cpu going down. The node |
4484ebf1 RT |
1297 | * structure is usually allocated from kmem_cache_create() and |
1298 | * gets destroyed at kmem_cache_destroy(). | |
1299 | */ | |
183ff22b | 1300 | /* fall through */ |
8f5be20b | 1301 | #endif |
1da177e4 | 1302 | case CPU_UP_CANCELED: |
8bb78442 | 1303 | case CPU_UP_CANCELED_FROZEN: |
18004c5d | 1304 | mutex_lock(&slab_mutex); |
fbf1e473 | 1305 | cpuup_canceled(cpu); |
18004c5d | 1306 | mutex_unlock(&slab_mutex); |
1da177e4 | 1307 | break; |
1da177e4 | 1308 | } |
eac40680 | 1309 | return notifier_from_errno(err); |
1da177e4 LT |
1310 | } |
1311 | ||
0db0628d | 1312 | static struct notifier_block cpucache_notifier = { |
74b85f37 CS |
1313 | &cpuup_callback, NULL, 0 |
1314 | }; | |
1da177e4 | 1315 | |
8f9f8d9e DR |
1316 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1317 | /* | |
1318 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1319 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1320 | * removed. | |
1321 | * | |
18004c5d | 1322 | * Must hold slab_mutex. |
8f9f8d9e | 1323 | */ |
6a67368c | 1324 | static int __meminit drain_cache_node_node(int node) |
8f9f8d9e DR |
1325 | { |
1326 | struct kmem_cache *cachep; | |
1327 | int ret = 0; | |
1328 | ||
18004c5d | 1329 | list_for_each_entry(cachep, &slab_caches, list) { |
ce8eb6c4 | 1330 | struct kmem_cache_node *n; |
8f9f8d9e | 1331 | |
ce8eb6c4 CL |
1332 | n = cachep->node[node]; |
1333 | if (!n) | |
8f9f8d9e DR |
1334 | continue; |
1335 | ||
0fa8103b | 1336 | drain_freelist(cachep, n, slabs_tofree(cachep, n)); |
8f9f8d9e | 1337 | |
ce8eb6c4 CL |
1338 | if (!list_empty(&n->slabs_full) || |
1339 | !list_empty(&n->slabs_partial)) { | |
8f9f8d9e DR |
1340 | ret = -EBUSY; |
1341 | break; | |
1342 | } | |
1343 | } | |
1344 | return ret; | |
1345 | } | |
1346 | ||
1347 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1348 | unsigned long action, void *arg) | |
1349 | { | |
1350 | struct memory_notify *mnb = arg; | |
1351 | int ret = 0; | |
1352 | int nid; | |
1353 | ||
1354 | nid = mnb->status_change_nid; | |
1355 | if (nid < 0) | |
1356 | goto out; | |
1357 | ||
1358 | switch (action) { | |
1359 | case MEM_GOING_ONLINE: | |
18004c5d | 1360 | mutex_lock(&slab_mutex); |
6a67368c | 1361 | ret = init_cache_node_node(nid); |
18004c5d | 1362 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1363 | break; |
1364 | case MEM_GOING_OFFLINE: | |
18004c5d | 1365 | mutex_lock(&slab_mutex); |
6a67368c | 1366 | ret = drain_cache_node_node(nid); |
18004c5d | 1367 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1368 | break; |
1369 | case MEM_ONLINE: | |
1370 | case MEM_OFFLINE: | |
1371 | case MEM_CANCEL_ONLINE: | |
1372 | case MEM_CANCEL_OFFLINE: | |
1373 | break; | |
1374 | } | |
1375 | out: | |
5fda1bd5 | 1376 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1377 | } |
1378 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1379 | ||
e498be7d | 1380 | /* |
ce8eb6c4 | 1381 | * swap the static kmem_cache_node with kmalloced memory |
e498be7d | 1382 | */ |
6744f087 | 1383 | static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list, |
8f9f8d9e | 1384 | int nodeid) |
e498be7d | 1385 | { |
6744f087 | 1386 | struct kmem_cache_node *ptr; |
e498be7d | 1387 | |
6744f087 | 1388 | ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid); |
e498be7d CL |
1389 | BUG_ON(!ptr); |
1390 | ||
6744f087 | 1391 | memcpy(ptr, list, sizeof(struct kmem_cache_node)); |
2b2d5493 IM |
1392 | /* |
1393 | * Do not assume that spinlocks can be initialized via memcpy: | |
1394 | */ | |
1395 | spin_lock_init(&ptr->list_lock); | |
1396 | ||
e498be7d | 1397 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
6a67368c | 1398 | cachep->node[nodeid] = ptr; |
e498be7d CL |
1399 | } |
1400 | ||
556a169d | 1401 | /* |
ce8eb6c4 CL |
1402 | * For setting up all the kmem_cache_node for cache whose buffer_size is same as |
1403 | * size of kmem_cache_node. | |
556a169d | 1404 | */ |
ce8eb6c4 | 1405 | static void __init set_up_node(struct kmem_cache *cachep, int index) |
556a169d PE |
1406 | { |
1407 | int node; | |
1408 | ||
1409 | for_each_online_node(node) { | |
ce8eb6c4 | 1410 | cachep->node[node] = &init_kmem_cache_node[index + node]; |
6a67368c | 1411 | cachep->node[node]->next_reap = jiffies + |
556a169d PE |
1412 | REAPTIMEOUT_LIST3 + |
1413 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1414 | } | |
1415 | } | |
1416 | ||
3c583465 CL |
1417 | /* |
1418 | * The memory after the last cpu cache pointer is used for the | |
6a67368c | 1419 | * the node pointer. |
3c583465 | 1420 | */ |
6a67368c | 1421 | static void setup_node_pointer(struct kmem_cache *cachep) |
3c583465 | 1422 | { |
6a67368c | 1423 | cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids]; |
3c583465 CL |
1424 | } |
1425 | ||
a737b3e2 AM |
1426 | /* |
1427 | * Initialisation. Called after the page allocator have been initialised and | |
1428 | * before smp_init(). | |
1da177e4 LT |
1429 | */ |
1430 | void __init kmem_cache_init(void) | |
1431 | { | |
e498be7d CL |
1432 | int i; |
1433 | ||
68126702 JK |
1434 | BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) < |
1435 | sizeof(struct rcu_head)); | |
9b030cb8 | 1436 | kmem_cache = &kmem_cache_boot; |
6a67368c | 1437 | setup_node_pointer(kmem_cache); |
9b030cb8 | 1438 | |
b6e68bc1 | 1439 | if (num_possible_nodes() == 1) |
62918a03 SS |
1440 | use_alien_caches = 0; |
1441 | ||
3c583465 | 1442 | for (i = 0; i < NUM_INIT_LISTS; i++) |
ce8eb6c4 | 1443 | kmem_cache_node_init(&init_kmem_cache_node[i]); |
3c583465 | 1444 | |
ce8eb6c4 | 1445 | set_up_node(kmem_cache, CACHE_CACHE); |
1da177e4 LT |
1446 | |
1447 | /* | |
1448 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1449 | * page orders on machines with more than 32MB of memory if |
1450 | * not overridden on the command line. | |
1da177e4 | 1451 | */ |
3df1cccd | 1452 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1453 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1454 | |
1da177e4 LT |
1455 | /* Bootstrap is tricky, because several objects are allocated |
1456 | * from caches that do not exist yet: | |
9b030cb8 CL |
1457 | * 1) initialize the kmem_cache cache: it contains the struct |
1458 | * kmem_cache structures of all caches, except kmem_cache itself: | |
1459 | * kmem_cache is statically allocated. | |
e498be7d | 1460 | * Initially an __init data area is used for the head array and the |
ce8eb6c4 | 1461 | * kmem_cache_node structures, it's replaced with a kmalloc allocated |
e498be7d | 1462 | * array at the end of the bootstrap. |
1da177e4 | 1463 | * 2) Create the first kmalloc cache. |
343e0d7a | 1464 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1465 | * An __init data area is used for the head array. |
1466 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1467 | * head arrays. | |
9b030cb8 | 1468 | * 4) Replace the __init data head arrays for kmem_cache and the first |
1da177e4 | 1469 | * kmalloc cache with kmalloc allocated arrays. |
ce8eb6c4 | 1470 | * 5) Replace the __init data for kmem_cache_node for kmem_cache and |
e498be7d CL |
1471 | * the other cache's with kmalloc allocated memory. |
1472 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1473 | */ |
1474 | ||
9b030cb8 | 1475 | /* 1) create the kmem_cache */ |
1da177e4 | 1476 | |
8da3430d | 1477 | /* |
b56efcf0 | 1478 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1479 | */ |
2f9baa9f CL |
1480 | create_boot_cache(kmem_cache, "kmem_cache", |
1481 | offsetof(struct kmem_cache, array[nr_cpu_ids]) + | |
6744f087 | 1482 | nr_node_ids * sizeof(struct kmem_cache_node *), |
2f9baa9f CL |
1483 | SLAB_HWCACHE_ALIGN); |
1484 | list_add(&kmem_cache->list, &slab_caches); | |
1da177e4 LT |
1485 | |
1486 | /* 2+3) create the kmalloc caches */ | |
1da177e4 | 1487 | |
a737b3e2 AM |
1488 | /* |
1489 | * Initialize the caches that provide memory for the array cache and the | |
ce8eb6c4 | 1490 | * kmem_cache_node structures first. Without this, further allocations will |
a737b3e2 | 1491 | * bug. |
e498be7d CL |
1492 | */ |
1493 | ||
e3366016 CL |
1494 | kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac", |
1495 | kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS); | |
45530c44 | 1496 | |
ce8eb6c4 CL |
1497 | if (INDEX_AC != INDEX_NODE) |
1498 | kmalloc_caches[INDEX_NODE] = | |
1499 | create_kmalloc_cache("kmalloc-node", | |
1500 | kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS); | |
e498be7d | 1501 | |
e0a42726 IM |
1502 | slab_early_init = 0; |
1503 | ||
1da177e4 LT |
1504 | /* 4) Replace the bootstrap head arrays */ |
1505 | { | |
2b2d5493 | 1506 | struct array_cache *ptr; |
e498be7d | 1507 | |
83b519e8 | 1508 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1509 | |
9b030cb8 | 1510 | memcpy(ptr, cpu_cache_get(kmem_cache), |
b28a02de | 1511 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1512 | /* |
1513 | * Do not assume that spinlocks can be initialized via memcpy: | |
1514 | */ | |
1515 | spin_lock_init(&ptr->lock); | |
1516 | ||
9b030cb8 | 1517 | kmem_cache->array[smp_processor_id()] = ptr; |
e498be7d | 1518 | |
83b519e8 | 1519 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1520 | |
e3366016 | 1521 | BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC]) |
b28a02de | 1522 | != &initarray_generic.cache); |
e3366016 | 1523 | memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]), |
b28a02de | 1524 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1525 | /* |
1526 | * Do not assume that spinlocks can be initialized via memcpy: | |
1527 | */ | |
1528 | spin_lock_init(&ptr->lock); | |
1529 | ||
e3366016 | 1530 | kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr; |
1da177e4 | 1531 | } |
ce8eb6c4 | 1532 | /* 5) Replace the bootstrap kmem_cache_node */ |
e498be7d | 1533 | { |
1ca4cb24 PE |
1534 | int nid; |
1535 | ||
9c09a95c | 1536 | for_each_online_node(nid) { |
ce8eb6c4 | 1537 | init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid); |
556a169d | 1538 | |
e3366016 | 1539 | init_list(kmalloc_caches[INDEX_AC], |
ce8eb6c4 | 1540 | &init_kmem_cache_node[SIZE_AC + nid], nid); |
e498be7d | 1541 | |
ce8eb6c4 CL |
1542 | if (INDEX_AC != INDEX_NODE) { |
1543 | init_list(kmalloc_caches[INDEX_NODE], | |
1544 | &init_kmem_cache_node[SIZE_NODE + nid], nid); | |
e498be7d CL |
1545 | } |
1546 | } | |
1547 | } | |
1da177e4 | 1548 | |
f97d5f63 | 1549 | create_kmalloc_caches(ARCH_KMALLOC_FLAGS); |
8429db5c PE |
1550 | } |
1551 | ||
1552 | void __init kmem_cache_init_late(void) | |
1553 | { | |
1554 | struct kmem_cache *cachep; | |
1555 | ||
97d06609 | 1556 | slab_state = UP; |
52cef189 | 1557 | |
8429db5c | 1558 | /* 6) resize the head arrays to their final sizes */ |
18004c5d CL |
1559 | mutex_lock(&slab_mutex); |
1560 | list_for_each_entry(cachep, &slab_caches, list) | |
8429db5c PE |
1561 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1562 | BUG(); | |
18004c5d | 1563 | mutex_unlock(&slab_mutex); |
056c6241 | 1564 | |
947ca185 MW |
1565 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1566 | init_lock_keys(); | |
1567 | ||
97d06609 CL |
1568 | /* Done! */ |
1569 | slab_state = FULL; | |
1570 | ||
a737b3e2 AM |
1571 | /* |
1572 | * Register a cpu startup notifier callback that initializes | |
1573 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1574 | */ |
1575 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1576 | |
8f9f8d9e DR |
1577 | #ifdef CONFIG_NUMA |
1578 | /* | |
1579 | * Register a memory hotplug callback that initializes and frees | |
6a67368c | 1580 | * node. |
8f9f8d9e DR |
1581 | */ |
1582 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1583 | #endif | |
1584 | ||
a737b3e2 AM |
1585 | /* |
1586 | * The reap timers are started later, with a module init call: That part | |
1587 | * of the kernel is not yet operational. | |
1da177e4 LT |
1588 | */ |
1589 | } | |
1590 | ||
1591 | static int __init cpucache_init(void) | |
1592 | { | |
1593 | int cpu; | |
1594 | ||
a737b3e2 AM |
1595 | /* |
1596 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1597 | */ |
e498be7d | 1598 | for_each_online_cpu(cpu) |
a737b3e2 | 1599 | start_cpu_timer(cpu); |
a164f896 GC |
1600 | |
1601 | /* Done! */ | |
97d06609 | 1602 | slab_state = FULL; |
1da177e4 LT |
1603 | return 0; |
1604 | } | |
1da177e4 LT |
1605 | __initcall(cpucache_init); |
1606 | ||
8bdec192 RA |
1607 | static noinline void |
1608 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1609 | { | |
ce8eb6c4 | 1610 | struct kmem_cache_node *n; |
8456a648 | 1611 | struct page *page; |
8bdec192 RA |
1612 | unsigned long flags; |
1613 | int node; | |
1614 | ||
1615 | printk(KERN_WARNING | |
1616 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1617 | nodeid, gfpflags); | |
1618 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1619 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1620 | |
1621 | for_each_online_node(node) { | |
1622 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1623 | unsigned long active_slabs = 0, num_slabs = 0; | |
1624 | ||
ce8eb6c4 CL |
1625 | n = cachep->node[node]; |
1626 | if (!n) | |
8bdec192 RA |
1627 | continue; |
1628 | ||
ce8eb6c4 | 1629 | spin_lock_irqsave(&n->list_lock, flags); |
8456a648 | 1630 | list_for_each_entry(page, &n->slabs_full, lru) { |
8bdec192 RA |
1631 | active_objs += cachep->num; |
1632 | active_slabs++; | |
1633 | } | |
8456a648 JK |
1634 | list_for_each_entry(page, &n->slabs_partial, lru) { |
1635 | active_objs += page->active; | |
8bdec192 RA |
1636 | active_slabs++; |
1637 | } | |
8456a648 | 1638 | list_for_each_entry(page, &n->slabs_free, lru) |
8bdec192 RA |
1639 | num_slabs++; |
1640 | ||
ce8eb6c4 CL |
1641 | free_objects += n->free_objects; |
1642 | spin_unlock_irqrestore(&n->list_lock, flags); | |
8bdec192 RA |
1643 | |
1644 | num_slabs += active_slabs; | |
1645 | num_objs = num_slabs * cachep->num; | |
1646 | printk(KERN_WARNING | |
1647 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1648 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1649 | free_objects); | |
1650 | } | |
1651 | } | |
1652 | ||
1da177e4 LT |
1653 | /* |
1654 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1655 | * | |
1656 | * If we requested dmaable memory, we will get it. Even if we | |
1657 | * did not request dmaable memory, we might get it, but that | |
1658 | * would be relatively rare and ignorable. | |
1659 | */ | |
0c3aa83e JK |
1660 | static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, |
1661 | int nodeid) | |
1da177e4 LT |
1662 | { |
1663 | struct page *page; | |
e1b6aa6f | 1664 | int nr_pages; |
765c4507 | 1665 | |
a618e89f | 1666 | flags |= cachep->allocflags; |
e12ba74d MG |
1667 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1668 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1669 | |
517d0869 | 1670 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1671 | if (!page) { |
1672 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1673 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1674 | return NULL; |
8bdec192 | 1675 | } |
1da177e4 | 1676 | |
b37f1dd0 | 1677 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
072bb0aa MG |
1678 | if (unlikely(page->pfmemalloc)) |
1679 | pfmemalloc_active = true; | |
1680 | ||
e1b6aa6f | 1681 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1682 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1683 | add_zone_page_state(page_zone(page), |
1684 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1685 | else | |
1686 | add_zone_page_state(page_zone(page), | |
1687 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
a57a4988 JK |
1688 | __SetPageSlab(page); |
1689 | if (page->pfmemalloc) | |
1690 | SetPageSlabPfmemalloc(page); | |
1f458cbf | 1691 | memcg_bind_pages(cachep, cachep->gfporder); |
072bb0aa | 1692 | |
b1eeab67 VN |
1693 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1694 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1695 | ||
1696 | if (cachep->ctor) | |
1697 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1698 | else | |
1699 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1700 | } | |
c175eea4 | 1701 | |
0c3aa83e | 1702 | return page; |
1da177e4 LT |
1703 | } |
1704 | ||
1705 | /* | |
1706 | * Interface to system's page release. | |
1707 | */ | |
0c3aa83e | 1708 | static void kmem_freepages(struct kmem_cache *cachep, struct page *page) |
1da177e4 | 1709 | { |
a57a4988 | 1710 | const unsigned long nr_freed = (1 << cachep->gfporder); |
1da177e4 | 1711 | |
b1eeab67 | 1712 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1713 | |
972d1a7b CL |
1714 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1715 | sub_zone_page_state(page_zone(page), | |
1716 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1717 | else | |
1718 | sub_zone_page_state(page_zone(page), | |
1719 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
73293c2f | 1720 | |
a57a4988 | 1721 | BUG_ON(!PageSlab(page)); |
73293c2f | 1722 | __ClearPageSlabPfmemalloc(page); |
a57a4988 | 1723 | __ClearPageSlab(page); |
8456a648 JK |
1724 | page_mapcount_reset(page); |
1725 | page->mapping = NULL; | |
1f458cbf GC |
1726 | |
1727 | memcg_release_pages(cachep, cachep->gfporder); | |
1da177e4 LT |
1728 | if (current->reclaim_state) |
1729 | current->reclaim_state->reclaimed_slab += nr_freed; | |
0c3aa83e | 1730 | __free_memcg_kmem_pages(page, cachep->gfporder); |
1da177e4 LT |
1731 | } |
1732 | ||
1733 | static void kmem_rcu_free(struct rcu_head *head) | |
1734 | { | |
68126702 JK |
1735 | struct kmem_cache *cachep; |
1736 | struct page *page; | |
1da177e4 | 1737 | |
68126702 JK |
1738 | page = container_of(head, struct page, rcu_head); |
1739 | cachep = page->slab_cache; | |
1740 | ||
1741 | kmem_freepages(cachep, page); | |
1da177e4 LT |
1742 | } |
1743 | ||
1744 | #if DEBUG | |
1745 | ||
1746 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1747 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1748 | unsigned long caller) |
1da177e4 | 1749 | { |
8c138bc0 | 1750 | int size = cachep->object_size; |
1da177e4 | 1751 | |
3dafccf2 | 1752 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1753 | |
b28a02de | 1754 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1755 | return; |
1756 | ||
b28a02de PE |
1757 | *addr++ = 0x12345678; |
1758 | *addr++ = caller; | |
1759 | *addr++ = smp_processor_id(); | |
1760 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1761 | { |
1762 | unsigned long *sptr = &caller; | |
1763 | unsigned long svalue; | |
1764 | ||
1765 | while (!kstack_end(sptr)) { | |
1766 | svalue = *sptr++; | |
1767 | if (kernel_text_address(svalue)) { | |
b28a02de | 1768 | *addr++ = svalue; |
1da177e4 LT |
1769 | size -= sizeof(unsigned long); |
1770 | if (size <= sizeof(unsigned long)) | |
1771 | break; | |
1772 | } | |
1773 | } | |
1774 | ||
1775 | } | |
b28a02de | 1776 | *addr++ = 0x87654321; |
1da177e4 LT |
1777 | } |
1778 | #endif | |
1779 | ||
343e0d7a | 1780 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1781 | { |
8c138bc0 | 1782 | int size = cachep->object_size; |
3dafccf2 | 1783 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1784 | |
1785 | memset(addr, val, size); | |
b28a02de | 1786 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1787 | } |
1788 | ||
1789 | static void dump_line(char *data, int offset, int limit) | |
1790 | { | |
1791 | int i; | |
aa83aa40 DJ |
1792 | unsigned char error = 0; |
1793 | int bad_count = 0; | |
1794 | ||
fdde6abb | 1795 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
1796 | for (i = 0; i < limit; i++) { |
1797 | if (data[offset + i] != POISON_FREE) { | |
1798 | error = data[offset + i]; | |
1799 | bad_count++; | |
1800 | } | |
aa83aa40 | 1801 | } |
fdde6abb SAS |
1802 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1803 | &data[offset], limit, 1); | |
aa83aa40 DJ |
1804 | |
1805 | if (bad_count == 1) { | |
1806 | error ^= POISON_FREE; | |
1807 | if (!(error & (error - 1))) { | |
1808 | printk(KERN_ERR "Single bit error detected. Probably " | |
1809 | "bad RAM.\n"); | |
1810 | #ifdef CONFIG_X86 | |
1811 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1812 | "test tool.\n"); | |
1813 | #else | |
1814 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1815 | #endif | |
1816 | } | |
1817 | } | |
1da177e4 LT |
1818 | } |
1819 | #endif | |
1820 | ||
1821 | #if DEBUG | |
1822 | ||
343e0d7a | 1823 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1824 | { |
1825 | int i, size; | |
1826 | char *realobj; | |
1827 | ||
1828 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1829 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1830 | *dbg_redzone1(cachep, objp), |
1831 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1832 | } |
1833 | ||
1834 | if (cachep->flags & SLAB_STORE_USER) { | |
071361d3 JP |
1835 | printk(KERN_ERR "Last user: [<%p>](%pSR)\n", |
1836 | *dbg_userword(cachep, objp), | |
1837 | *dbg_userword(cachep, objp)); | |
1da177e4 | 1838 | } |
3dafccf2 | 1839 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 1840 | size = cachep->object_size; |
b28a02de | 1841 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1842 | int limit; |
1843 | limit = 16; | |
b28a02de PE |
1844 | if (i + limit > size) |
1845 | limit = size - i; | |
1da177e4 LT |
1846 | dump_line(realobj, i, limit); |
1847 | } | |
1848 | } | |
1849 | ||
343e0d7a | 1850 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1851 | { |
1852 | char *realobj; | |
1853 | int size, i; | |
1854 | int lines = 0; | |
1855 | ||
3dafccf2 | 1856 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 1857 | size = cachep->object_size; |
1da177e4 | 1858 | |
b28a02de | 1859 | for (i = 0; i < size; i++) { |
1da177e4 | 1860 | char exp = POISON_FREE; |
b28a02de | 1861 | if (i == size - 1) |
1da177e4 LT |
1862 | exp = POISON_END; |
1863 | if (realobj[i] != exp) { | |
1864 | int limit; | |
1865 | /* Mismatch ! */ | |
1866 | /* Print header */ | |
1867 | if (lines == 0) { | |
b28a02de | 1868 | printk(KERN_ERR |
face37f5 DJ |
1869 | "Slab corruption (%s): %s start=%p, len=%d\n", |
1870 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
1871 | print_objinfo(cachep, objp, 0); |
1872 | } | |
1873 | /* Hexdump the affected line */ | |
b28a02de | 1874 | i = (i / 16) * 16; |
1da177e4 | 1875 | limit = 16; |
b28a02de PE |
1876 | if (i + limit > size) |
1877 | limit = size - i; | |
1da177e4 LT |
1878 | dump_line(realobj, i, limit); |
1879 | i += 16; | |
1880 | lines++; | |
1881 | /* Limit to 5 lines */ | |
1882 | if (lines > 5) | |
1883 | break; | |
1884 | } | |
1885 | } | |
1886 | if (lines != 0) { | |
1887 | /* Print some data about the neighboring objects, if they | |
1888 | * exist: | |
1889 | */ | |
8456a648 | 1890 | struct page *page = virt_to_head_page(objp); |
8fea4e96 | 1891 | unsigned int objnr; |
1da177e4 | 1892 | |
8456a648 | 1893 | objnr = obj_to_index(cachep, page, objp); |
1da177e4 | 1894 | if (objnr) { |
8456a648 | 1895 | objp = index_to_obj(cachep, page, objnr - 1); |
3dafccf2 | 1896 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1897 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1898 | realobj, size); |
1da177e4 LT |
1899 | print_objinfo(cachep, objp, 2); |
1900 | } | |
b28a02de | 1901 | if (objnr + 1 < cachep->num) { |
8456a648 | 1902 | objp = index_to_obj(cachep, page, objnr + 1); |
3dafccf2 | 1903 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1904 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1905 | realobj, size); |
1da177e4 LT |
1906 | print_objinfo(cachep, objp, 2); |
1907 | } | |
1908 | } | |
1909 | } | |
1910 | #endif | |
1911 | ||
12dd36fa | 1912 | #if DEBUG |
8456a648 JK |
1913 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, |
1914 | struct page *page) | |
1da177e4 | 1915 | { |
1da177e4 LT |
1916 | int i; |
1917 | for (i = 0; i < cachep->num; i++) { | |
8456a648 | 1918 | void *objp = index_to_obj(cachep, page, i); |
1da177e4 LT |
1919 | |
1920 | if (cachep->flags & SLAB_POISON) { | |
1921 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 1922 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 1923 | OFF_SLAB(cachep)) |
b28a02de | 1924 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 1925 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
1926 | else |
1927 | check_poison_obj(cachep, objp); | |
1928 | #else | |
1929 | check_poison_obj(cachep, objp); | |
1930 | #endif | |
1931 | } | |
1932 | if (cachep->flags & SLAB_RED_ZONE) { | |
1933 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1934 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1935 | "was overwritten"); |
1da177e4 LT |
1936 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1937 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1938 | "was overwritten"); |
1da177e4 | 1939 | } |
1da177e4 | 1940 | } |
12dd36fa | 1941 | } |
1da177e4 | 1942 | #else |
8456a648 JK |
1943 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, |
1944 | struct page *page) | |
12dd36fa | 1945 | { |
12dd36fa | 1946 | } |
1da177e4 LT |
1947 | #endif |
1948 | ||
911851e6 RD |
1949 | /** |
1950 | * slab_destroy - destroy and release all objects in a slab | |
1951 | * @cachep: cache pointer being destroyed | |
1952 | * @slabp: slab pointer being destroyed | |
1953 | * | |
12dd36fa | 1954 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1955 | * Before calling the slab must have been unlinked from the cache. The |
1956 | * cache-lock is not held/needed. | |
12dd36fa | 1957 | */ |
8456a648 | 1958 | static void slab_destroy(struct kmem_cache *cachep, struct page *page) |
12dd36fa | 1959 | { |
8456a648 | 1960 | struct freelist *freelist; |
12dd36fa | 1961 | |
8456a648 JK |
1962 | freelist = page->freelist; |
1963 | slab_destroy_debugcheck(cachep, page); | |
1da177e4 | 1964 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
68126702 JK |
1965 | struct rcu_head *head; |
1966 | ||
1967 | /* | |
1968 | * RCU free overloads the RCU head over the LRU. | |
1969 | * slab_page has been overloeaded over the LRU, | |
1970 | * however it is not used from now on so that | |
1971 | * we can use it safely. | |
1972 | */ | |
1973 | head = (void *)&page->rcu_head; | |
1974 | call_rcu(head, kmem_rcu_free); | |
1da177e4 | 1975 | |
1da177e4 | 1976 | } else { |
0c3aa83e | 1977 | kmem_freepages(cachep, page); |
1da177e4 | 1978 | } |
68126702 JK |
1979 | |
1980 | /* | |
8456a648 | 1981 | * From now on, we don't use freelist |
68126702 JK |
1982 | * although actual page can be freed in rcu context |
1983 | */ | |
1984 | if (OFF_SLAB(cachep)) | |
8456a648 | 1985 | kmem_cache_free(cachep->freelist_cache, freelist); |
1da177e4 LT |
1986 | } |
1987 | ||
4d268eba | 1988 | /** |
a70773dd RD |
1989 | * calculate_slab_order - calculate size (page order) of slabs |
1990 | * @cachep: pointer to the cache that is being created | |
1991 | * @size: size of objects to be created in this cache. | |
1992 | * @align: required alignment for the objects. | |
1993 | * @flags: slab allocation flags | |
1994 | * | |
1995 | * Also calculates the number of objects per slab. | |
4d268eba PE |
1996 | * |
1997 | * This could be made much more intelligent. For now, try to avoid using | |
1998 | * high order pages for slabs. When the gfp() functions are more friendly | |
1999 | * towards high-order requests, this should be changed. | |
2000 | */ | |
a737b3e2 | 2001 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2002 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2003 | { |
b1ab41c4 | 2004 | unsigned long offslab_limit; |
4d268eba | 2005 | size_t left_over = 0; |
9888e6fa | 2006 | int gfporder; |
4d268eba | 2007 | |
0aa817f0 | 2008 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2009 | unsigned int num; |
2010 | size_t remainder; | |
2011 | ||
9888e6fa | 2012 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2013 | if (!num) |
2014 | continue; | |
9888e6fa | 2015 | |
b1ab41c4 IM |
2016 | if (flags & CFLGS_OFF_SLAB) { |
2017 | /* | |
2018 | * Max number of objs-per-slab for caches which | |
2019 | * use off-slab slabs. Needed to avoid a possible | |
2020 | * looping condition in cache_grow(). | |
2021 | */ | |
8456a648 | 2022 | offslab_limit = size; |
16025177 | 2023 | offslab_limit /= sizeof(unsigned int); |
b1ab41c4 IM |
2024 | |
2025 | if (num > offslab_limit) | |
2026 | break; | |
2027 | } | |
4d268eba | 2028 | |
9888e6fa | 2029 | /* Found something acceptable - save it away */ |
4d268eba | 2030 | cachep->num = num; |
9888e6fa | 2031 | cachep->gfporder = gfporder; |
4d268eba PE |
2032 | left_over = remainder; |
2033 | ||
f78bb8ad LT |
2034 | /* |
2035 | * A VFS-reclaimable slab tends to have most allocations | |
2036 | * as GFP_NOFS and we really don't want to have to be allocating | |
2037 | * higher-order pages when we are unable to shrink dcache. | |
2038 | */ | |
2039 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2040 | break; | |
2041 | ||
4d268eba PE |
2042 | /* |
2043 | * Large number of objects is good, but very large slabs are | |
2044 | * currently bad for the gfp()s. | |
2045 | */ | |
543585cc | 2046 | if (gfporder >= slab_max_order) |
4d268eba PE |
2047 | break; |
2048 | ||
9888e6fa LT |
2049 | /* |
2050 | * Acceptable internal fragmentation? | |
2051 | */ | |
a737b3e2 | 2052 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2053 | break; |
2054 | } | |
2055 | return left_over; | |
2056 | } | |
2057 | ||
83b519e8 | 2058 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2059 | { |
97d06609 | 2060 | if (slab_state >= FULL) |
83b519e8 | 2061 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2062 | |
97d06609 | 2063 | if (slab_state == DOWN) { |
f30cf7d1 | 2064 | /* |
2f9baa9f | 2065 | * Note: Creation of first cache (kmem_cache). |
ce8eb6c4 | 2066 | * The setup_node is taken care |
2f9baa9f CL |
2067 | * of by the caller of __kmem_cache_create |
2068 | */ | |
2069 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2070 | slab_state = PARTIAL; | |
2071 | } else if (slab_state == PARTIAL) { | |
2072 | /* | |
2073 | * Note: the second kmem_cache_create must create the cache | |
f30cf7d1 PE |
2074 | * that's used by kmalloc(24), otherwise the creation of |
2075 | * further caches will BUG(). | |
2076 | */ | |
2077 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2078 | ||
2079 | /* | |
ce8eb6c4 CL |
2080 | * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is |
2081 | * the second cache, then we need to set up all its node/, | |
f30cf7d1 PE |
2082 | * otherwise the creation of further caches will BUG(). |
2083 | */ | |
ce8eb6c4 CL |
2084 | set_up_node(cachep, SIZE_AC); |
2085 | if (INDEX_AC == INDEX_NODE) | |
2086 | slab_state = PARTIAL_NODE; | |
f30cf7d1 | 2087 | else |
97d06609 | 2088 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 | 2089 | } else { |
2f9baa9f | 2090 | /* Remaining boot caches */ |
f30cf7d1 | 2091 | cachep->array[smp_processor_id()] = |
83b519e8 | 2092 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2093 | |
97d06609 | 2094 | if (slab_state == PARTIAL_ARRAYCACHE) { |
ce8eb6c4 CL |
2095 | set_up_node(cachep, SIZE_NODE); |
2096 | slab_state = PARTIAL_NODE; | |
f30cf7d1 PE |
2097 | } else { |
2098 | int node; | |
556a169d | 2099 | for_each_online_node(node) { |
6a67368c | 2100 | cachep->node[node] = |
6744f087 | 2101 | kmalloc_node(sizeof(struct kmem_cache_node), |
eb91f1d0 | 2102 | gfp, node); |
6a67368c | 2103 | BUG_ON(!cachep->node[node]); |
ce8eb6c4 | 2104 | kmem_cache_node_init(cachep->node[node]); |
f30cf7d1 PE |
2105 | } |
2106 | } | |
2107 | } | |
6a67368c | 2108 | cachep->node[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2109 | jiffies + REAPTIMEOUT_LIST3 + |
2110 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2111 | ||
2112 | cpu_cache_get(cachep)->avail = 0; | |
2113 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2114 | cpu_cache_get(cachep)->batchcount = 1; | |
2115 | cpu_cache_get(cachep)->touched = 0; | |
2116 | cachep->batchcount = 1; | |
2117 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2118 | return 0; |
f30cf7d1 PE |
2119 | } |
2120 | ||
1da177e4 | 2121 | /** |
039363f3 | 2122 | * __kmem_cache_create - Create a cache. |
a755b76a | 2123 | * @cachep: cache management descriptor |
1da177e4 | 2124 | * @flags: SLAB flags |
1da177e4 LT |
2125 | * |
2126 | * Returns a ptr to the cache on success, NULL on failure. | |
2127 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2128 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 | 2129 | * |
1da177e4 LT |
2130 | * The flags are |
2131 | * | |
2132 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2133 | * to catch references to uninitialised memory. | |
2134 | * | |
2135 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2136 | * for buffer overruns. | |
2137 | * | |
1da177e4 LT |
2138 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2139 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2140 | * as davem. | |
2141 | */ | |
278b1bb1 | 2142 | int |
8a13a4cc | 2143 | __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) |
1da177e4 | 2144 | { |
8456a648 | 2145 | size_t left_over, freelist_size, ralign; |
83b519e8 | 2146 | gfp_t gfp; |
278b1bb1 | 2147 | int err; |
8a13a4cc | 2148 | size_t size = cachep->size; |
1da177e4 | 2149 | |
1da177e4 | 2150 | #if DEBUG |
1da177e4 LT |
2151 | #if FORCED_DEBUG |
2152 | /* | |
2153 | * Enable redzoning and last user accounting, except for caches with | |
2154 | * large objects, if the increased size would increase the object size | |
2155 | * above the next power of two: caches with object sizes just above a | |
2156 | * power of two have a significant amount of internal fragmentation. | |
2157 | */ | |
87a927c7 DW |
2158 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2159 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2160 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2161 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2162 | flags |= SLAB_POISON; | |
2163 | #endif | |
2164 | if (flags & SLAB_DESTROY_BY_RCU) | |
2165 | BUG_ON(flags & SLAB_POISON); | |
2166 | #endif | |
1da177e4 | 2167 | |
a737b3e2 AM |
2168 | /* |
2169 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2170 | * unaligned accesses for some archs when redzoning is used, and makes |
2171 | * sure any on-slab bufctl's are also correctly aligned. | |
2172 | */ | |
b28a02de PE |
2173 | if (size & (BYTES_PER_WORD - 1)) { |
2174 | size += (BYTES_PER_WORD - 1); | |
2175 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2176 | } |
2177 | ||
ca5f9703 | 2178 | /* |
87a927c7 DW |
2179 | * Redzoning and user store require word alignment or possibly larger. |
2180 | * Note this will be overridden by architecture or caller mandated | |
2181 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2182 | */ |
87a927c7 DW |
2183 | if (flags & SLAB_STORE_USER) |
2184 | ralign = BYTES_PER_WORD; | |
2185 | ||
2186 | if (flags & SLAB_RED_ZONE) { | |
2187 | ralign = REDZONE_ALIGN; | |
2188 | /* If redzoning, ensure that the second redzone is suitably | |
2189 | * aligned, by adjusting the object size accordingly. */ | |
2190 | size += REDZONE_ALIGN - 1; | |
2191 | size &= ~(REDZONE_ALIGN - 1); | |
2192 | } | |
ca5f9703 | 2193 | |
a44b56d3 | 2194 | /* 3) caller mandated alignment */ |
8a13a4cc CL |
2195 | if (ralign < cachep->align) { |
2196 | ralign = cachep->align; | |
1da177e4 | 2197 | } |
3ff84a7f PE |
2198 | /* disable debug if necessary */ |
2199 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2200 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2201 | /* |
ca5f9703 | 2202 | * 4) Store it. |
1da177e4 | 2203 | */ |
8a13a4cc | 2204 | cachep->align = ralign; |
1da177e4 | 2205 | |
83b519e8 PE |
2206 | if (slab_is_available()) |
2207 | gfp = GFP_KERNEL; | |
2208 | else | |
2209 | gfp = GFP_NOWAIT; | |
2210 | ||
6a67368c | 2211 | setup_node_pointer(cachep); |
1da177e4 | 2212 | #if DEBUG |
1da177e4 | 2213 | |
ca5f9703 PE |
2214 | /* |
2215 | * Both debugging options require word-alignment which is calculated | |
2216 | * into align above. | |
2217 | */ | |
1da177e4 | 2218 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2219 | /* add space for red zone words */ |
3ff84a7f PE |
2220 | cachep->obj_offset += sizeof(unsigned long long); |
2221 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2222 | } |
2223 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2224 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2225 | * the real object. But if the second red zone needs to be |
2226 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2227 | */ |
87a927c7 DW |
2228 | if (flags & SLAB_RED_ZONE) |
2229 | size += REDZONE_ALIGN; | |
2230 | else | |
2231 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2232 | } |
2233 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
ce8eb6c4 | 2234 | if (size >= kmalloc_size(INDEX_NODE + 1) |
608da7e3 TH |
2235 | && cachep->object_size > cache_line_size() |
2236 | && ALIGN(size, cachep->align) < PAGE_SIZE) { | |
2237 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align); | |
1da177e4 LT |
2238 | size = PAGE_SIZE; |
2239 | } | |
2240 | #endif | |
2241 | #endif | |
2242 | ||
e0a42726 IM |
2243 | /* |
2244 | * Determine if the slab management is 'on' or 'off' slab. | |
2245 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2246 | * it too early on. Always use on-slab management when |
2247 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2248 | */ |
e7cb55b9 CM |
2249 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2250 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2251 | /* |
2252 | * Size is large, assume best to place the slab management obj | |
2253 | * off-slab (should allow better packing of objs). | |
2254 | */ | |
2255 | flags |= CFLGS_OFF_SLAB; | |
2256 | ||
8a13a4cc | 2257 | size = ALIGN(size, cachep->align); |
1da177e4 | 2258 | |
8a13a4cc | 2259 | left_over = calculate_slab_order(cachep, size, cachep->align, flags); |
1da177e4 | 2260 | |
8a13a4cc | 2261 | if (!cachep->num) |
278b1bb1 | 2262 | return -E2BIG; |
1da177e4 | 2263 | |
8456a648 JK |
2264 | freelist_size = |
2265 | ALIGN(cachep->num * sizeof(unsigned int), cachep->align); | |
1da177e4 LT |
2266 | |
2267 | /* | |
2268 | * If the slab has been placed off-slab, and we have enough space then | |
2269 | * move it on-slab. This is at the expense of any extra colouring. | |
2270 | */ | |
8456a648 | 2271 | if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) { |
1da177e4 | 2272 | flags &= ~CFLGS_OFF_SLAB; |
8456a648 | 2273 | left_over -= freelist_size; |
1da177e4 LT |
2274 | } |
2275 | ||
2276 | if (flags & CFLGS_OFF_SLAB) { | |
2277 | /* really off slab. No need for manual alignment */ | |
8456a648 | 2278 | freelist_size = cachep->num * sizeof(unsigned int); |
67461365 RL |
2279 | |
2280 | #ifdef CONFIG_PAGE_POISONING | |
2281 | /* If we're going to use the generic kernel_map_pages() | |
2282 | * poisoning, then it's going to smash the contents of | |
2283 | * the redzone and userword anyhow, so switch them off. | |
2284 | */ | |
2285 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2286 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2287 | #endif | |
1da177e4 LT |
2288 | } |
2289 | ||
2290 | cachep->colour_off = cache_line_size(); | |
2291 | /* Offset must be a multiple of the alignment. */ | |
8a13a4cc CL |
2292 | if (cachep->colour_off < cachep->align) |
2293 | cachep->colour_off = cachep->align; | |
b28a02de | 2294 | cachep->colour = left_over / cachep->colour_off; |
8456a648 | 2295 | cachep->freelist_size = freelist_size; |
1da177e4 | 2296 | cachep->flags = flags; |
a57a4988 | 2297 | cachep->allocflags = __GFP_COMP; |
4b51d669 | 2298 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2299 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2300 | cachep->size = size; |
6a2d7a95 | 2301 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2302 | |
e5ac9c5a | 2303 | if (flags & CFLGS_OFF_SLAB) { |
8456a648 | 2304 | cachep->freelist_cache = kmalloc_slab(freelist_size, 0u); |
e5ac9c5a RT |
2305 | /* |
2306 | * This is a possibility for one of the malloc_sizes caches. | |
2307 | * But since we go off slab only for object size greater than | |
2308 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2309 | * this should not happen at all. | |
2310 | * But leave a BUG_ON for some lucky dude. | |
2311 | */ | |
8456a648 | 2312 | BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache)); |
e5ac9c5a | 2313 | } |
1da177e4 | 2314 | |
278b1bb1 CL |
2315 | err = setup_cpu_cache(cachep, gfp); |
2316 | if (err) { | |
12c3667f | 2317 | __kmem_cache_shutdown(cachep); |
278b1bb1 | 2318 | return err; |
2ed3a4ef | 2319 | } |
1da177e4 | 2320 | |
83835b3d PZ |
2321 | if (flags & SLAB_DEBUG_OBJECTS) { |
2322 | /* | |
2323 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2324 | * debug_object_activate()->kmem_cache_alloc(). | |
2325 | */ | |
2326 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2327 | ||
2328 | slab_set_debugobj_lock_classes(cachep); | |
6ccfb5bc GC |
2329 | } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU)) |
2330 | on_slab_lock_classes(cachep); | |
83835b3d | 2331 | |
278b1bb1 | 2332 | return 0; |
1da177e4 | 2333 | } |
1da177e4 LT |
2334 | |
2335 | #if DEBUG | |
2336 | static void check_irq_off(void) | |
2337 | { | |
2338 | BUG_ON(!irqs_disabled()); | |
2339 | } | |
2340 | ||
2341 | static void check_irq_on(void) | |
2342 | { | |
2343 | BUG_ON(irqs_disabled()); | |
2344 | } | |
2345 | ||
343e0d7a | 2346 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2347 | { |
2348 | #ifdef CONFIG_SMP | |
2349 | check_irq_off(); | |
6a67368c | 2350 | assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock); |
1da177e4 LT |
2351 | #endif |
2352 | } | |
e498be7d | 2353 | |
343e0d7a | 2354 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2355 | { |
2356 | #ifdef CONFIG_SMP | |
2357 | check_irq_off(); | |
6a67368c | 2358 | assert_spin_locked(&cachep->node[node]->list_lock); |
e498be7d CL |
2359 | #endif |
2360 | } | |
2361 | ||
1da177e4 LT |
2362 | #else |
2363 | #define check_irq_off() do { } while(0) | |
2364 | #define check_irq_on() do { } while(0) | |
2365 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2366 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2367 | #endif |
2368 | ||
ce8eb6c4 | 2369 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n, |
aab2207c CL |
2370 | struct array_cache *ac, |
2371 | int force, int node); | |
2372 | ||
1da177e4 LT |
2373 | static void do_drain(void *arg) |
2374 | { | |
a737b3e2 | 2375 | struct kmem_cache *cachep = arg; |
1da177e4 | 2376 | struct array_cache *ac; |
7d6e6d09 | 2377 | int node = numa_mem_id(); |
1da177e4 LT |
2378 | |
2379 | check_irq_off(); | |
9a2dba4b | 2380 | ac = cpu_cache_get(cachep); |
6a67368c | 2381 | spin_lock(&cachep->node[node]->list_lock); |
ff69416e | 2382 | free_block(cachep, ac->entry, ac->avail, node); |
6a67368c | 2383 | spin_unlock(&cachep->node[node]->list_lock); |
1da177e4 LT |
2384 | ac->avail = 0; |
2385 | } | |
2386 | ||
343e0d7a | 2387 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2388 | { |
ce8eb6c4 | 2389 | struct kmem_cache_node *n; |
e498be7d CL |
2390 | int node; |
2391 | ||
15c8b6c1 | 2392 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2393 | check_irq_on(); |
b28a02de | 2394 | for_each_online_node(node) { |
ce8eb6c4 CL |
2395 | n = cachep->node[node]; |
2396 | if (n && n->alien) | |
2397 | drain_alien_cache(cachep, n->alien); | |
a4523a8b RD |
2398 | } |
2399 | ||
2400 | for_each_online_node(node) { | |
ce8eb6c4 CL |
2401 | n = cachep->node[node]; |
2402 | if (n) | |
2403 | drain_array(cachep, n, n->shared, 1, node); | |
e498be7d | 2404 | } |
1da177e4 LT |
2405 | } |
2406 | ||
ed11d9eb CL |
2407 | /* |
2408 | * Remove slabs from the list of free slabs. | |
2409 | * Specify the number of slabs to drain in tofree. | |
2410 | * | |
2411 | * Returns the actual number of slabs released. | |
2412 | */ | |
2413 | static int drain_freelist(struct kmem_cache *cache, | |
ce8eb6c4 | 2414 | struct kmem_cache_node *n, int tofree) |
1da177e4 | 2415 | { |
ed11d9eb CL |
2416 | struct list_head *p; |
2417 | int nr_freed; | |
8456a648 | 2418 | struct page *page; |
1da177e4 | 2419 | |
ed11d9eb | 2420 | nr_freed = 0; |
ce8eb6c4 | 2421 | while (nr_freed < tofree && !list_empty(&n->slabs_free)) { |
1da177e4 | 2422 | |
ce8eb6c4 CL |
2423 | spin_lock_irq(&n->list_lock); |
2424 | p = n->slabs_free.prev; | |
2425 | if (p == &n->slabs_free) { | |
2426 | spin_unlock_irq(&n->list_lock); | |
ed11d9eb CL |
2427 | goto out; |
2428 | } | |
1da177e4 | 2429 | |
8456a648 | 2430 | page = list_entry(p, struct page, lru); |
1da177e4 | 2431 | #if DEBUG |
8456a648 | 2432 | BUG_ON(page->active); |
1da177e4 | 2433 | #endif |
8456a648 | 2434 | list_del(&page->lru); |
ed11d9eb CL |
2435 | /* |
2436 | * Safe to drop the lock. The slab is no longer linked | |
2437 | * to the cache. | |
2438 | */ | |
ce8eb6c4 CL |
2439 | n->free_objects -= cache->num; |
2440 | spin_unlock_irq(&n->list_lock); | |
8456a648 | 2441 | slab_destroy(cache, page); |
ed11d9eb | 2442 | nr_freed++; |
1da177e4 | 2443 | } |
ed11d9eb CL |
2444 | out: |
2445 | return nr_freed; | |
1da177e4 LT |
2446 | } |
2447 | ||
18004c5d | 2448 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2449 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2450 | { |
2451 | int ret = 0, i = 0; | |
ce8eb6c4 | 2452 | struct kmem_cache_node *n; |
e498be7d CL |
2453 | |
2454 | drain_cpu_caches(cachep); | |
2455 | ||
2456 | check_irq_on(); | |
2457 | for_each_online_node(i) { | |
ce8eb6c4 CL |
2458 | n = cachep->node[i]; |
2459 | if (!n) | |
ed11d9eb CL |
2460 | continue; |
2461 | ||
0fa8103b | 2462 | drain_freelist(cachep, n, slabs_tofree(cachep, n)); |
ed11d9eb | 2463 | |
ce8eb6c4 CL |
2464 | ret += !list_empty(&n->slabs_full) || |
2465 | !list_empty(&n->slabs_partial); | |
e498be7d CL |
2466 | } |
2467 | return (ret ? 1 : 0); | |
2468 | } | |
2469 | ||
1da177e4 LT |
2470 | /** |
2471 | * kmem_cache_shrink - Shrink a cache. | |
2472 | * @cachep: The cache to shrink. | |
2473 | * | |
2474 | * Releases as many slabs as possible for a cache. | |
2475 | * To help debugging, a zero exit status indicates all slabs were released. | |
2476 | */ | |
343e0d7a | 2477 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2478 | { |
8f5be20b | 2479 | int ret; |
40094fa6 | 2480 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2481 | |
95402b38 | 2482 | get_online_cpus(); |
18004c5d | 2483 | mutex_lock(&slab_mutex); |
8f5be20b | 2484 | ret = __cache_shrink(cachep); |
18004c5d | 2485 | mutex_unlock(&slab_mutex); |
95402b38 | 2486 | put_online_cpus(); |
8f5be20b | 2487 | return ret; |
1da177e4 LT |
2488 | } |
2489 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2490 | ||
945cf2b6 | 2491 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
1da177e4 | 2492 | { |
12c3667f | 2493 | int i; |
ce8eb6c4 | 2494 | struct kmem_cache_node *n; |
12c3667f | 2495 | int rc = __cache_shrink(cachep); |
1da177e4 | 2496 | |
12c3667f CL |
2497 | if (rc) |
2498 | return rc; | |
1da177e4 | 2499 | |
12c3667f CL |
2500 | for_each_online_cpu(i) |
2501 | kfree(cachep->array[i]); | |
1da177e4 | 2502 | |
ce8eb6c4 | 2503 | /* NUMA: free the node structures */ |
12c3667f | 2504 | for_each_online_node(i) { |
ce8eb6c4 CL |
2505 | n = cachep->node[i]; |
2506 | if (n) { | |
2507 | kfree(n->shared); | |
2508 | free_alien_cache(n->alien); | |
2509 | kfree(n); | |
12c3667f CL |
2510 | } |
2511 | } | |
2512 | return 0; | |
1da177e4 | 2513 | } |
1da177e4 | 2514 | |
e5ac9c5a RT |
2515 | /* |
2516 | * Get the memory for a slab management obj. | |
2517 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2518 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2519 | * come from the same cache which is getting created because, | |
2520 | * when we are searching for an appropriate cache for these | |
2521 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2522 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2523 | * kmem_find_general_cachep till the initialization is complete. | |
8456a648 | 2524 | * Hence we cannot have freelist_cache same as the original cache. |
e5ac9c5a | 2525 | */ |
8456a648 | 2526 | static struct freelist *alloc_slabmgmt(struct kmem_cache *cachep, |
0c3aa83e JK |
2527 | struct page *page, int colour_off, |
2528 | gfp_t local_flags, int nodeid) | |
1da177e4 | 2529 | { |
8456a648 | 2530 | struct freelist *freelist; |
0c3aa83e | 2531 | void *addr = page_address(page); |
b28a02de | 2532 | |
1da177e4 LT |
2533 | if (OFF_SLAB(cachep)) { |
2534 | /* Slab management obj is off-slab. */ | |
8456a648 | 2535 | freelist = kmem_cache_alloc_node(cachep->freelist_cache, |
8759ec50 | 2536 | local_flags, nodeid); |
d5cff635 CM |
2537 | /* |
2538 | * If the first object in the slab is leaked (it's allocated | |
2539 | * but no one has a reference to it), we want to make sure | |
2540 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2541 | * to the object. Otherwise we will not report the leak. | |
2542 | */ | |
8456a648 | 2543 | kmemleak_scan_area(&page->lru, sizeof(struct list_head), |
c017b4be | 2544 | local_flags); |
8456a648 | 2545 | if (!freelist) |
1da177e4 LT |
2546 | return NULL; |
2547 | } else { | |
8456a648 JK |
2548 | freelist = addr + colour_off; |
2549 | colour_off += cachep->freelist_size; | |
1da177e4 | 2550 | } |
8456a648 JK |
2551 | page->active = 0; |
2552 | page->s_mem = addr + colour_off; | |
2553 | return freelist; | |
1da177e4 LT |
2554 | } |
2555 | ||
8456a648 | 2556 | static inline unsigned int *slab_bufctl(struct page *page) |
1da177e4 | 2557 | { |
8456a648 | 2558 | return (unsigned int *)(page->freelist); |
1da177e4 LT |
2559 | } |
2560 | ||
343e0d7a | 2561 | static void cache_init_objs(struct kmem_cache *cachep, |
8456a648 | 2562 | struct page *page) |
1da177e4 LT |
2563 | { |
2564 | int i; | |
2565 | ||
2566 | for (i = 0; i < cachep->num; i++) { | |
8456a648 | 2567 | void *objp = index_to_obj(cachep, page, i); |
1da177e4 LT |
2568 | #if DEBUG |
2569 | /* need to poison the objs? */ | |
2570 | if (cachep->flags & SLAB_POISON) | |
2571 | poison_obj(cachep, objp, POISON_FREE); | |
2572 | if (cachep->flags & SLAB_STORE_USER) | |
2573 | *dbg_userword(cachep, objp) = NULL; | |
2574 | ||
2575 | if (cachep->flags & SLAB_RED_ZONE) { | |
2576 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2577 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2578 | } | |
2579 | /* | |
a737b3e2 AM |
2580 | * Constructors are not allowed to allocate memory from the same |
2581 | * cache which they are a constructor for. Otherwise, deadlock. | |
2582 | * They must also be threaded. | |
1da177e4 LT |
2583 | */ |
2584 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2585 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2586 | |
2587 | if (cachep->flags & SLAB_RED_ZONE) { | |
2588 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2589 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2590 | " end of an object"); |
1da177e4 LT |
2591 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2592 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2593 | " start of an object"); |
1da177e4 | 2594 | } |
3b0efdfa | 2595 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2596 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2597 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2598 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2599 | #else |
2600 | if (cachep->ctor) | |
51cc5068 | 2601 | cachep->ctor(objp); |
1da177e4 | 2602 | #endif |
8456a648 | 2603 | slab_bufctl(page)[i] = i; |
1da177e4 | 2604 | } |
1da177e4 LT |
2605 | } |
2606 | ||
343e0d7a | 2607 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2608 | { |
4b51d669 CL |
2609 | if (CONFIG_ZONE_DMA_FLAG) { |
2610 | if (flags & GFP_DMA) | |
a618e89f | 2611 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2612 | else |
a618e89f | 2613 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2614 | } |
1da177e4 LT |
2615 | } |
2616 | ||
8456a648 | 2617 | static void *slab_get_obj(struct kmem_cache *cachep, struct page *page, |
a737b3e2 | 2618 | int nodeid) |
78d382d7 | 2619 | { |
b1cb0982 | 2620 | void *objp; |
78d382d7 | 2621 | |
8456a648 JK |
2622 | objp = index_to_obj(cachep, page, slab_bufctl(page)[page->active]); |
2623 | page->active++; | |
78d382d7 | 2624 | #if DEBUG |
1ea991b0 | 2625 | WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid); |
78d382d7 | 2626 | #endif |
78d382d7 MD |
2627 | |
2628 | return objp; | |
2629 | } | |
2630 | ||
8456a648 | 2631 | static void slab_put_obj(struct kmem_cache *cachep, struct page *page, |
a737b3e2 | 2632 | void *objp, int nodeid) |
78d382d7 | 2633 | { |
8456a648 | 2634 | unsigned int objnr = obj_to_index(cachep, page, objp); |
78d382d7 | 2635 | #if DEBUG |
16025177 | 2636 | unsigned int i; |
b1cb0982 | 2637 | |
78d382d7 | 2638 | /* Verify that the slab belongs to the intended node */ |
1ea991b0 | 2639 | WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid); |
78d382d7 | 2640 | |
b1cb0982 | 2641 | /* Verify double free bug */ |
8456a648 JK |
2642 | for (i = page->active; i < cachep->num; i++) { |
2643 | if (slab_bufctl(page)[i] == objnr) { | |
b1cb0982 JK |
2644 | printk(KERN_ERR "slab: double free detected in cache " |
2645 | "'%s', objp %p\n", cachep->name, objp); | |
2646 | BUG(); | |
2647 | } | |
78d382d7 MD |
2648 | } |
2649 | #endif | |
8456a648 JK |
2650 | page->active--; |
2651 | slab_bufctl(page)[page->active] = objnr; | |
78d382d7 MD |
2652 | } |
2653 | ||
4776874f PE |
2654 | /* |
2655 | * Map pages beginning at addr to the given cache and slab. This is required | |
2656 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2657 | * virtual address for kfree, ksize, and slab debugging. |
4776874f | 2658 | */ |
8456a648 JK |
2659 | static void slab_map_pages(struct kmem_cache *cache, struct page *page, |
2660 | struct freelist *freelist) | |
1da177e4 | 2661 | { |
a57a4988 | 2662 | page->slab_cache = cache; |
8456a648 | 2663 | page->freelist = freelist; |
1da177e4 LT |
2664 | } |
2665 | ||
2666 | /* | |
2667 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2668 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2669 | */ | |
3c517a61 | 2670 | static int cache_grow(struct kmem_cache *cachep, |
0c3aa83e | 2671 | gfp_t flags, int nodeid, struct page *page) |
1da177e4 | 2672 | { |
8456a648 | 2673 | struct freelist *freelist; |
b28a02de PE |
2674 | size_t offset; |
2675 | gfp_t local_flags; | |
ce8eb6c4 | 2676 | struct kmem_cache_node *n; |
1da177e4 | 2677 | |
a737b3e2 AM |
2678 | /* |
2679 | * Be lazy and only check for valid flags here, keeping it out of the | |
2680 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2681 | */ |
6cb06229 CL |
2682 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2683 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2684 | |
ce8eb6c4 | 2685 | /* Take the node list lock to change the colour_next on this node */ |
1da177e4 | 2686 | check_irq_off(); |
ce8eb6c4 CL |
2687 | n = cachep->node[nodeid]; |
2688 | spin_lock(&n->list_lock); | |
1da177e4 LT |
2689 | |
2690 | /* Get colour for the slab, and cal the next value. */ | |
ce8eb6c4 CL |
2691 | offset = n->colour_next; |
2692 | n->colour_next++; | |
2693 | if (n->colour_next >= cachep->colour) | |
2694 | n->colour_next = 0; | |
2695 | spin_unlock(&n->list_lock); | |
1da177e4 | 2696 | |
2e1217cf | 2697 | offset *= cachep->colour_off; |
1da177e4 LT |
2698 | |
2699 | if (local_flags & __GFP_WAIT) | |
2700 | local_irq_enable(); | |
2701 | ||
2702 | /* | |
2703 | * The test for missing atomic flag is performed here, rather than | |
2704 | * the more obvious place, simply to reduce the critical path length | |
2705 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2706 | * will eventually be caught here (where it matters). | |
2707 | */ | |
2708 | kmem_flagcheck(cachep, flags); | |
2709 | ||
a737b3e2 AM |
2710 | /* |
2711 | * Get mem for the objs. Attempt to allocate a physical page from | |
2712 | * 'nodeid'. | |
e498be7d | 2713 | */ |
0c3aa83e JK |
2714 | if (!page) |
2715 | page = kmem_getpages(cachep, local_flags, nodeid); | |
2716 | if (!page) | |
1da177e4 LT |
2717 | goto failed; |
2718 | ||
2719 | /* Get slab management. */ | |
8456a648 | 2720 | freelist = alloc_slabmgmt(cachep, page, offset, |
6cb06229 | 2721 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
8456a648 | 2722 | if (!freelist) |
1da177e4 LT |
2723 | goto opps1; |
2724 | ||
8456a648 | 2725 | slab_map_pages(cachep, page, freelist); |
1da177e4 | 2726 | |
8456a648 | 2727 | cache_init_objs(cachep, page); |
1da177e4 LT |
2728 | |
2729 | if (local_flags & __GFP_WAIT) | |
2730 | local_irq_disable(); | |
2731 | check_irq_off(); | |
ce8eb6c4 | 2732 | spin_lock(&n->list_lock); |
1da177e4 LT |
2733 | |
2734 | /* Make slab active. */ | |
8456a648 | 2735 | list_add_tail(&page->lru, &(n->slabs_free)); |
1da177e4 | 2736 | STATS_INC_GROWN(cachep); |
ce8eb6c4 CL |
2737 | n->free_objects += cachep->num; |
2738 | spin_unlock(&n->list_lock); | |
1da177e4 | 2739 | return 1; |
a737b3e2 | 2740 | opps1: |
0c3aa83e | 2741 | kmem_freepages(cachep, page); |
a737b3e2 | 2742 | failed: |
1da177e4 LT |
2743 | if (local_flags & __GFP_WAIT) |
2744 | local_irq_disable(); | |
2745 | return 0; | |
2746 | } | |
2747 | ||
2748 | #if DEBUG | |
2749 | ||
2750 | /* | |
2751 | * Perform extra freeing checks: | |
2752 | * - detect bad pointers. | |
2753 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2754 | */ |
2755 | static void kfree_debugcheck(const void *objp) | |
2756 | { | |
1da177e4 LT |
2757 | if (!virt_addr_valid(objp)) { |
2758 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2759 | (unsigned long)objp); |
2760 | BUG(); | |
1da177e4 | 2761 | } |
1da177e4 LT |
2762 | } |
2763 | ||
58ce1fd5 PE |
2764 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2765 | { | |
b46b8f19 | 2766 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2767 | |
2768 | redzone1 = *dbg_redzone1(cache, obj); | |
2769 | redzone2 = *dbg_redzone2(cache, obj); | |
2770 | ||
2771 | /* | |
2772 | * Redzone is ok. | |
2773 | */ | |
2774 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2775 | return; | |
2776 | ||
2777 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2778 | slab_error(cache, "double free detected"); | |
2779 | else | |
2780 | slab_error(cache, "memory outside object was overwritten"); | |
2781 | ||
b46b8f19 | 2782 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2783 | obj, redzone1, redzone2); |
2784 | } | |
2785 | ||
343e0d7a | 2786 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 2787 | unsigned long caller) |
1da177e4 | 2788 | { |
1da177e4 | 2789 | unsigned int objnr; |
8456a648 | 2790 | struct page *page; |
1da177e4 | 2791 | |
80cbd911 MW |
2792 | BUG_ON(virt_to_cache(objp) != cachep); |
2793 | ||
3dafccf2 | 2794 | objp -= obj_offset(cachep); |
1da177e4 | 2795 | kfree_debugcheck(objp); |
8456a648 | 2796 | page = virt_to_head_page(objp); |
1da177e4 LT |
2797 | |
2798 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2799 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2800 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2801 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2802 | } | |
2803 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 2804 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 | 2805 | |
8456a648 | 2806 | objnr = obj_to_index(cachep, page, objp); |
1da177e4 LT |
2807 | |
2808 | BUG_ON(objnr >= cachep->num); | |
8456a648 | 2809 | BUG_ON(objp != index_to_obj(cachep, page, objnr)); |
1da177e4 | 2810 | |
1da177e4 LT |
2811 | if (cachep->flags & SLAB_POISON) { |
2812 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2813 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
7c0cb9c6 | 2814 | store_stackinfo(cachep, objp, caller); |
b28a02de | 2815 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2816 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2817 | } else { |
2818 | poison_obj(cachep, objp, POISON_FREE); | |
2819 | } | |
2820 | #else | |
2821 | poison_obj(cachep, objp, POISON_FREE); | |
2822 | #endif | |
2823 | } | |
2824 | return objp; | |
2825 | } | |
2826 | ||
1da177e4 LT |
2827 | #else |
2828 | #define kfree_debugcheck(x) do { } while(0) | |
2829 | #define cache_free_debugcheck(x,objp,z) (objp) | |
1da177e4 LT |
2830 | #endif |
2831 | ||
072bb0aa MG |
2832 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
2833 | bool force_refill) | |
1da177e4 LT |
2834 | { |
2835 | int batchcount; | |
ce8eb6c4 | 2836 | struct kmem_cache_node *n; |
1da177e4 | 2837 | struct array_cache *ac; |
1ca4cb24 PE |
2838 | int node; |
2839 | ||
1da177e4 | 2840 | check_irq_off(); |
7d6e6d09 | 2841 | node = numa_mem_id(); |
072bb0aa MG |
2842 | if (unlikely(force_refill)) |
2843 | goto force_grow; | |
2844 | retry: | |
9a2dba4b | 2845 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
2846 | batchcount = ac->batchcount; |
2847 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
2848 | /* |
2849 | * If there was little recent activity on this cache, then | |
2850 | * perform only a partial refill. Otherwise we could generate | |
2851 | * refill bouncing. | |
1da177e4 LT |
2852 | */ |
2853 | batchcount = BATCHREFILL_LIMIT; | |
2854 | } | |
ce8eb6c4 | 2855 | n = cachep->node[node]; |
e498be7d | 2856 | |
ce8eb6c4 CL |
2857 | BUG_ON(ac->avail > 0 || !n); |
2858 | spin_lock(&n->list_lock); | |
1da177e4 | 2859 | |
3ded175a | 2860 | /* See if we can refill from the shared array */ |
ce8eb6c4 CL |
2861 | if (n->shared && transfer_objects(ac, n->shared, batchcount)) { |
2862 | n->shared->touched = 1; | |
3ded175a | 2863 | goto alloc_done; |
44b57f1c | 2864 | } |
3ded175a | 2865 | |
1da177e4 LT |
2866 | while (batchcount > 0) { |
2867 | struct list_head *entry; | |
8456a648 | 2868 | struct page *page; |
1da177e4 | 2869 | /* Get slab alloc is to come from. */ |
ce8eb6c4 CL |
2870 | entry = n->slabs_partial.next; |
2871 | if (entry == &n->slabs_partial) { | |
2872 | n->free_touched = 1; | |
2873 | entry = n->slabs_free.next; | |
2874 | if (entry == &n->slabs_free) | |
1da177e4 LT |
2875 | goto must_grow; |
2876 | } | |
2877 | ||
8456a648 | 2878 | page = list_entry(entry, struct page, lru); |
1da177e4 | 2879 | check_spinlock_acquired(cachep); |
714b8171 PE |
2880 | |
2881 | /* | |
2882 | * The slab was either on partial or free list so | |
2883 | * there must be at least one object available for | |
2884 | * allocation. | |
2885 | */ | |
8456a648 | 2886 | BUG_ON(page->active >= cachep->num); |
714b8171 | 2887 | |
8456a648 | 2888 | while (page->active < cachep->num && batchcount--) { |
1da177e4 LT |
2889 | STATS_INC_ALLOCED(cachep); |
2890 | STATS_INC_ACTIVE(cachep); | |
2891 | STATS_SET_HIGH(cachep); | |
2892 | ||
8456a648 | 2893 | ac_put_obj(cachep, ac, slab_get_obj(cachep, page, |
072bb0aa | 2894 | node)); |
1da177e4 | 2895 | } |
1da177e4 LT |
2896 | |
2897 | /* move slabp to correct slabp list: */ | |
8456a648 JK |
2898 | list_del(&page->lru); |
2899 | if (page->active == cachep->num) | |
2900 | list_add(&page->list, &n->slabs_full); | |
1da177e4 | 2901 | else |
8456a648 | 2902 | list_add(&page->list, &n->slabs_partial); |
1da177e4 LT |
2903 | } |
2904 | ||
a737b3e2 | 2905 | must_grow: |
ce8eb6c4 | 2906 | n->free_objects -= ac->avail; |
a737b3e2 | 2907 | alloc_done: |
ce8eb6c4 | 2908 | spin_unlock(&n->list_lock); |
1da177e4 LT |
2909 | |
2910 | if (unlikely(!ac->avail)) { | |
2911 | int x; | |
072bb0aa | 2912 | force_grow: |
3c517a61 | 2913 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 2914 | |
a737b3e2 | 2915 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 2916 | ac = cpu_cache_get(cachep); |
51cd8e6f | 2917 | node = numa_mem_id(); |
072bb0aa MG |
2918 | |
2919 | /* no objects in sight? abort */ | |
2920 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
2921 | return NULL; |
2922 | ||
a737b3e2 | 2923 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
2924 | goto retry; |
2925 | } | |
2926 | ac->touched = 1; | |
072bb0aa MG |
2927 | |
2928 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
2929 | } |
2930 | ||
a737b3e2 AM |
2931 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
2932 | gfp_t flags) | |
1da177e4 LT |
2933 | { |
2934 | might_sleep_if(flags & __GFP_WAIT); | |
2935 | #if DEBUG | |
2936 | kmem_flagcheck(cachep, flags); | |
2937 | #endif | |
2938 | } | |
2939 | ||
2940 | #if DEBUG | |
a737b3e2 | 2941 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
7c0cb9c6 | 2942 | gfp_t flags, void *objp, unsigned long caller) |
1da177e4 | 2943 | { |
b28a02de | 2944 | if (!objp) |
1da177e4 | 2945 | return objp; |
b28a02de | 2946 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 2947 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 2948 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 2949 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2950 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2951 | else |
2952 | check_poison_obj(cachep, objp); | |
2953 | #else | |
2954 | check_poison_obj(cachep, objp); | |
2955 | #endif | |
2956 | poison_obj(cachep, objp, POISON_INUSE); | |
2957 | } | |
2958 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 2959 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 LT |
2960 | |
2961 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
2962 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
2963 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
2964 | slab_error(cachep, "double free, or memory outside" | |
2965 | " object was overwritten"); | |
b28a02de | 2966 | printk(KERN_ERR |
b46b8f19 | 2967 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
2968 | objp, *dbg_redzone1(cachep, objp), |
2969 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2970 | } |
2971 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
2972 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
2973 | } | |
3dafccf2 | 2974 | objp += obj_offset(cachep); |
4f104934 | 2975 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 2976 | cachep->ctor(objp); |
7ea466f2 TH |
2977 | if (ARCH_SLAB_MINALIGN && |
2978 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 2979 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 2980 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 2981 | } |
1da177e4 LT |
2982 | return objp; |
2983 | } | |
2984 | #else | |
2985 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
2986 | #endif | |
2987 | ||
773ff60e | 2988 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 | 2989 | { |
9b030cb8 | 2990 | if (cachep == kmem_cache) |
773ff60e | 2991 | return false; |
8a8b6502 | 2992 | |
8c138bc0 | 2993 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
2994 | } |
2995 | ||
343e0d7a | 2996 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2997 | { |
b28a02de | 2998 | void *objp; |
1da177e4 | 2999 | struct array_cache *ac; |
072bb0aa | 3000 | bool force_refill = false; |
1da177e4 | 3001 | |
5c382300 | 3002 | check_irq_off(); |
8a8b6502 | 3003 | |
9a2dba4b | 3004 | ac = cpu_cache_get(cachep); |
1da177e4 | 3005 | if (likely(ac->avail)) { |
1da177e4 | 3006 | ac->touched = 1; |
072bb0aa MG |
3007 | objp = ac_get_obj(cachep, ac, flags, false); |
3008 | ||
ddbf2e83 | 3009 | /* |
072bb0aa MG |
3010 | * Allow for the possibility all avail objects are not allowed |
3011 | * by the current flags | |
ddbf2e83 | 3012 | */ |
072bb0aa MG |
3013 | if (objp) { |
3014 | STATS_INC_ALLOCHIT(cachep); | |
3015 | goto out; | |
3016 | } | |
3017 | force_refill = true; | |
1da177e4 | 3018 | } |
072bb0aa MG |
3019 | |
3020 | STATS_INC_ALLOCMISS(cachep); | |
3021 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3022 | /* | |
3023 | * the 'ac' may be updated by cache_alloc_refill(), | |
3024 | * and kmemleak_erase() requires its correct value. | |
3025 | */ | |
3026 | ac = cpu_cache_get(cachep); | |
3027 | ||
3028 | out: | |
d5cff635 CM |
3029 | /* |
3030 | * To avoid a false negative, if an object that is in one of the | |
3031 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3032 | * treat the array pointers as a reference to the object. | |
3033 | */ | |
f3d8b53a O |
3034 | if (objp) |
3035 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3036 | return objp; |
3037 | } | |
3038 | ||
e498be7d | 3039 | #ifdef CONFIG_NUMA |
c61afb18 | 3040 | /* |
b2455396 | 3041 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3042 | * |
3043 | * If we are in_interrupt, then process context, including cpusets and | |
3044 | * mempolicy, may not apply and should not be used for allocation policy. | |
3045 | */ | |
3046 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3047 | { | |
3048 | int nid_alloc, nid_here; | |
3049 | ||
765c4507 | 3050 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3051 | return NULL; |
7d6e6d09 | 3052 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3053 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3054 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3055 | else if (current->mempolicy) |
e7b691b0 | 3056 | nid_alloc = slab_node(); |
c61afb18 | 3057 | if (nid_alloc != nid_here) |
8b98c169 | 3058 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3059 | return NULL; |
3060 | } | |
3061 | ||
765c4507 CL |
3062 | /* |
3063 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 | 3064 | * certain node and fall back is permitted. First we scan all the |
6a67368c | 3065 | * available node for available objects. If that fails then we |
3c517a61 CL |
3066 | * perform an allocation without specifying a node. This allows the page |
3067 | * allocator to do its reclaim / fallback magic. We then insert the | |
3068 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3069 | */ |
8c8cc2c1 | 3070 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3071 | { |
8c8cc2c1 PE |
3072 | struct zonelist *zonelist; |
3073 | gfp_t local_flags; | |
dd1a239f | 3074 | struct zoneref *z; |
54a6eb5c MG |
3075 | struct zone *zone; |
3076 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3077 | void *obj = NULL; |
3c517a61 | 3078 | int nid; |
cc9a6c87 | 3079 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3080 | |
3081 | if (flags & __GFP_THISNODE) | |
3082 | return NULL; | |
3083 | ||
6cb06229 | 3084 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3085 | |
cc9a6c87 MG |
3086 | retry_cpuset: |
3087 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3088 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3089 | |
3c517a61 CL |
3090 | retry: |
3091 | /* | |
3092 | * Look through allowed nodes for objects available | |
3093 | * from existing per node queues. | |
3094 | */ | |
54a6eb5c MG |
3095 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3096 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3097 | |
54a6eb5c | 3098 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
6a67368c CL |
3099 | cache->node[nid] && |
3100 | cache->node[nid]->free_objects) { | |
3c517a61 CL |
3101 | obj = ____cache_alloc_node(cache, |
3102 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3103 | if (obj) |
3104 | break; | |
3105 | } | |
3c517a61 CL |
3106 | } |
3107 | ||
cfce6604 | 3108 | if (!obj) { |
3c517a61 CL |
3109 | /* |
3110 | * This allocation will be performed within the constraints | |
3111 | * of the current cpuset / memory policy requirements. | |
3112 | * We may trigger various forms of reclaim on the allowed | |
3113 | * set and go into memory reserves if necessary. | |
3114 | */ | |
0c3aa83e JK |
3115 | struct page *page; |
3116 | ||
dd47ea75 CL |
3117 | if (local_flags & __GFP_WAIT) |
3118 | local_irq_enable(); | |
3119 | kmem_flagcheck(cache, flags); | |
0c3aa83e | 3120 | page = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3121 | if (local_flags & __GFP_WAIT) |
3122 | local_irq_disable(); | |
0c3aa83e | 3123 | if (page) { |
3c517a61 CL |
3124 | /* |
3125 | * Insert into the appropriate per node queues | |
3126 | */ | |
0c3aa83e JK |
3127 | nid = page_to_nid(page); |
3128 | if (cache_grow(cache, flags, nid, page)) { | |
3c517a61 CL |
3129 | obj = ____cache_alloc_node(cache, |
3130 | flags | GFP_THISNODE, nid); | |
3131 | if (!obj) | |
3132 | /* | |
3133 | * Another processor may allocate the | |
3134 | * objects in the slab since we are | |
3135 | * not holding any locks. | |
3136 | */ | |
3137 | goto retry; | |
3138 | } else { | |
b6a60451 | 3139 | /* cache_grow already freed obj */ |
3c517a61 CL |
3140 | obj = NULL; |
3141 | } | |
3142 | } | |
aedb0eb1 | 3143 | } |
cc9a6c87 MG |
3144 | |
3145 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3146 | goto retry_cpuset; | |
765c4507 CL |
3147 | return obj; |
3148 | } | |
3149 | ||
e498be7d CL |
3150 | /* |
3151 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3152 | */ |
8b98c169 | 3153 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3154 | int nodeid) |
e498be7d CL |
3155 | { |
3156 | struct list_head *entry; | |
8456a648 | 3157 | struct page *page; |
ce8eb6c4 | 3158 | struct kmem_cache_node *n; |
b28a02de | 3159 | void *obj; |
b28a02de PE |
3160 | int x; |
3161 | ||
14e50c6a | 3162 | VM_BUG_ON(nodeid > num_online_nodes()); |
ce8eb6c4 CL |
3163 | n = cachep->node[nodeid]; |
3164 | BUG_ON(!n); | |
b28a02de | 3165 | |
a737b3e2 | 3166 | retry: |
ca3b9b91 | 3167 | check_irq_off(); |
ce8eb6c4 CL |
3168 | spin_lock(&n->list_lock); |
3169 | entry = n->slabs_partial.next; | |
3170 | if (entry == &n->slabs_partial) { | |
3171 | n->free_touched = 1; | |
3172 | entry = n->slabs_free.next; | |
3173 | if (entry == &n->slabs_free) | |
b28a02de PE |
3174 | goto must_grow; |
3175 | } | |
3176 | ||
8456a648 | 3177 | page = list_entry(entry, struct page, lru); |
b28a02de | 3178 | check_spinlock_acquired_node(cachep, nodeid); |
b28a02de PE |
3179 | |
3180 | STATS_INC_NODEALLOCS(cachep); | |
3181 | STATS_INC_ACTIVE(cachep); | |
3182 | STATS_SET_HIGH(cachep); | |
3183 | ||
8456a648 | 3184 | BUG_ON(page->active == cachep->num); |
b28a02de | 3185 | |
8456a648 | 3186 | obj = slab_get_obj(cachep, page, nodeid); |
ce8eb6c4 | 3187 | n->free_objects--; |
b28a02de | 3188 | /* move slabp to correct slabp list: */ |
8456a648 | 3189 | list_del(&page->lru); |
b28a02de | 3190 | |
8456a648 JK |
3191 | if (page->active == cachep->num) |
3192 | list_add(&page->lru, &n->slabs_full); | |
a737b3e2 | 3193 | else |
8456a648 | 3194 | list_add(&page->lru, &n->slabs_partial); |
e498be7d | 3195 | |
ce8eb6c4 | 3196 | spin_unlock(&n->list_lock); |
b28a02de | 3197 | goto done; |
e498be7d | 3198 | |
a737b3e2 | 3199 | must_grow: |
ce8eb6c4 | 3200 | spin_unlock(&n->list_lock); |
3c517a61 | 3201 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3202 | if (x) |
3203 | goto retry; | |
1da177e4 | 3204 | |
8c8cc2c1 | 3205 | return fallback_alloc(cachep, flags); |
e498be7d | 3206 | |
a737b3e2 | 3207 | done: |
b28a02de | 3208 | return obj; |
e498be7d | 3209 | } |
8c8cc2c1 | 3210 | |
8c8cc2c1 | 3211 | static __always_inline void * |
48356303 | 3212 | slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, |
7c0cb9c6 | 3213 | unsigned long caller) |
8c8cc2c1 PE |
3214 | { |
3215 | unsigned long save_flags; | |
3216 | void *ptr; | |
7d6e6d09 | 3217 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3218 | |
dcce284a | 3219 | flags &= gfp_allowed_mask; |
7e85ee0c | 3220 | |
cf40bd16 NP |
3221 | lockdep_trace_alloc(flags); |
3222 | ||
773ff60e | 3223 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3224 | return NULL; |
3225 | ||
d79923fa GC |
3226 | cachep = memcg_kmem_get_cache(cachep, flags); |
3227 | ||
8c8cc2c1 PE |
3228 | cache_alloc_debugcheck_before(cachep, flags); |
3229 | local_irq_save(save_flags); | |
3230 | ||
eacbbae3 | 3231 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3232 | nodeid = slab_node; |
8c8cc2c1 | 3233 | |
6a67368c | 3234 | if (unlikely(!cachep->node[nodeid])) { |
8c8cc2c1 PE |
3235 | /* Node not bootstrapped yet */ |
3236 | ptr = fallback_alloc(cachep, flags); | |
3237 | goto out; | |
3238 | } | |
3239 | ||
7d6e6d09 | 3240 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3241 | /* |
3242 | * Use the locally cached objects if possible. | |
3243 | * However ____cache_alloc does not allow fallback | |
3244 | * to other nodes. It may fail while we still have | |
3245 | * objects on other nodes available. | |
3246 | */ | |
3247 | ptr = ____cache_alloc(cachep, flags); | |
3248 | if (ptr) | |
3249 | goto out; | |
3250 | } | |
3251 | /* ___cache_alloc_node can fall back to other nodes */ | |
3252 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3253 | out: | |
3254 | local_irq_restore(save_flags); | |
3255 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3256 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3257 | flags); |
8c8cc2c1 | 3258 | |
c175eea4 | 3259 | if (likely(ptr)) |
8c138bc0 | 3260 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3261 | |
d07dbea4 | 3262 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3263 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3264 | |
8c8cc2c1 PE |
3265 | return ptr; |
3266 | } | |
3267 | ||
3268 | static __always_inline void * | |
3269 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3270 | { | |
3271 | void *objp; | |
3272 | ||
3273 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3274 | objp = alternate_node_alloc(cache, flags); | |
3275 | if (objp) | |
3276 | goto out; | |
3277 | } | |
3278 | objp = ____cache_alloc(cache, flags); | |
3279 | ||
3280 | /* | |
3281 | * We may just have run out of memory on the local node. | |
3282 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3283 | */ | |
7d6e6d09 LS |
3284 | if (!objp) |
3285 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3286 | |
3287 | out: | |
3288 | return objp; | |
3289 | } | |
3290 | #else | |
3291 | ||
3292 | static __always_inline void * | |
3293 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3294 | { | |
3295 | return ____cache_alloc(cachep, flags); | |
3296 | } | |
3297 | ||
3298 | #endif /* CONFIG_NUMA */ | |
3299 | ||
3300 | static __always_inline void * | |
48356303 | 3301 | slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) |
8c8cc2c1 PE |
3302 | { |
3303 | unsigned long save_flags; | |
3304 | void *objp; | |
3305 | ||
dcce284a | 3306 | flags &= gfp_allowed_mask; |
7e85ee0c | 3307 | |
cf40bd16 NP |
3308 | lockdep_trace_alloc(flags); |
3309 | ||
773ff60e | 3310 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3311 | return NULL; |
3312 | ||
d79923fa GC |
3313 | cachep = memcg_kmem_get_cache(cachep, flags); |
3314 | ||
8c8cc2c1 PE |
3315 | cache_alloc_debugcheck_before(cachep, flags); |
3316 | local_irq_save(save_flags); | |
3317 | objp = __do_cache_alloc(cachep, flags); | |
3318 | local_irq_restore(save_flags); | |
3319 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3320 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3321 | flags); |
8c8cc2c1 PE |
3322 | prefetchw(objp); |
3323 | ||
c175eea4 | 3324 | if (likely(objp)) |
8c138bc0 | 3325 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3326 | |
d07dbea4 | 3327 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3328 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3329 | |
8c8cc2c1 PE |
3330 | return objp; |
3331 | } | |
e498be7d CL |
3332 | |
3333 | /* | |
3334 | * Caller needs to acquire correct kmem_list's list_lock | |
3335 | */ | |
343e0d7a | 3336 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3337 | int node) |
1da177e4 LT |
3338 | { |
3339 | int i; | |
ce8eb6c4 | 3340 | struct kmem_cache_node *n; |
1da177e4 LT |
3341 | |
3342 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3343 | void *objp; |
8456a648 | 3344 | struct page *page; |
1da177e4 | 3345 | |
072bb0aa MG |
3346 | clear_obj_pfmemalloc(&objpp[i]); |
3347 | objp = objpp[i]; | |
3348 | ||
8456a648 | 3349 | page = virt_to_head_page(objp); |
ce8eb6c4 | 3350 | n = cachep->node[node]; |
8456a648 | 3351 | list_del(&page->lru); |
ff69416e | 3352 | check_spinlock_acquired_node(cachep, node); |
8456a648 | 3353 | slab_put_obj(cachep, page, objp, node); |
1da177e4 | 3354 | STATS_DEC_ACTIVE(cachep); |
ce8eb6c4 | 3355 | n->free_objects++; |
1da177e4 LT |
3356 | |
3357 | /* fixup slab chains */ | |
8456a648 | 3358 | if (page->active == 0) { |
ce8eb6c4 CL |
3359 | if (n->free_objects > n->free_limit) { |
3360 | n->free_objects -= cachep->num; | |
e5ac9c5a RT |
3361 | /* No need to drop any previously held |
3362 | * lock here, even if we have a off-slab slab | |
3363 | * descriptor it is guaranteed to come from | |
3364 | * a different cache, refer to comments before | |
3365 | * alloc_slabmgmt. | |
3366 | */ | |
8456a648 | 3367 | slab_destroy(cachep, page); |
1da177e4 | 3368 | } else { |
8456a648 | 3369 | list_add(&page->lru, &n->slabs_free); |
1da177e4 LT |
3370 | } |
3371 | } else { | |
3372 | /* Unconditionally move a slab to the end of the | |
3373 | * partial list on free - maximum time for the | |
3374 | * other objects to be freed, too. | |
3375 | */ | |
8456a648 | 3376 | list_add_tail(&page->lru, &n->slabs_partial); |
1da177e4 LT |
3377 | } |
3378 | } | |
3379 | } | |
3380 | ||
343e0d7a | 3381 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3382 | { |
3383 | int batchcount; | |
ce8eb6c4 | 3384 | struct kmem_cache_node *n; |
7d6e6d09 | 3385 | int node = numa_mem_id(); |
1da177e4 LT |
3386 | |
3387 | batchcount = ac->batchcount; | |
3388 | #if DEBUG | |
3389 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3390 | #endif | |
3391 | check_irq_off(); | |
ce8eb6c4 CL |
3392 | n = cachep->node[node]; |
3393 | spin_lock(&n->list_lock); | |
3394 | if (n->shared) { | |
3395 | struct array_cache *shared_array = n->shared; | |
b28a02de | 3396 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3397 | if (max) { |
3398 | if (batchcount > max) | |
3399 | batchcount = max; | |
e498be7d | 3400 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3401 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3402 | shared_array->avail += batchcount; |
3403 | goto free_done; | |
3404 | } | |
3405 | } | |
3406 | ||
ff69416e | 3407 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3408 | free_done: |
1da177e4 LT |
3409 | #if STATS |
3410 | { | |
3411 | int i = 0; | |
3412 | struct list_head *p; | |
3413 | ||
ce8eb6c4 CL |
3414 | p = n->slabs_free.next; |
3415 | while (p != &(n->slabs_free)) { | |
8456a648 | 3416 | struct page *page; |
1da177e4 | 3417 | |
8456a648 JK |
3418 | page = list_entry(p, struct page, lru); |
3419 | BUG_ON(page->active); | |
1da177e4 LT |
3420 | |
3421 | i++; | |
3422 | p = p->next; | |
3423 | } | |
3424 | STATS_SET_FREEABLE(cachep, i); | |
3425 | } | |
3426 | #endif | |
ce8eb6c4 | 3427 | spin_unlock(&n->list_lock); |
1da177e4 | 3428 | ac->avail -= batchcount; |
a737b3e2 | 3429 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3430 | } |
3431 | ||
3432 | /* | |
a737b3e2 AM |
3433 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3434 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3435 | */ |
a947eb95 | 3436 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 3437 | unsigned long caller) |
1da177e4 | 3438 | { |
9a2dba4b | 3439 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3440 | |
3441 | check_irq_off(); | |
d5cff635 | 3442 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3443 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3444 | |
8c138bc0 | 3445 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3446 | |
1807a1aa SS |
3447 | /* |
3448 | * Skip calling cache_free_alien() when the platform is not numa. | |
3449 | * This will avoid cache misses that happen while accessing slabp (which | |
3450 | * is per page memory reference) to get nodeid. Instead use a global | |
3451 | * variable to skip the call, which is mostly likely to be present in | |
3452 | * the cache. | |
3453 | */ | |
b6e68bc1 | 3454 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3455 | return; |
3456 | ||
1da177e4 LT |
3457 | if (likely(ac->avail < ac->limit)) { |
3458 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3459 | } else { |
3460 | STATS_INC_FREEMISS(cachep); | |
3461 | cache_flusharray(cachep, ac); | |
1da177e4 | 3462 | } |
42c8c99c | 3463 | |
072bb0aa | 3464 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3465 | } |
3466 | ||
3467 | /** | |
3468 | * kmem_cache_alloc - Allocate an object | |
3469 | * @cachep: The cache to allocate from. | |
3470 | * @flags: See kmalloc(). | |
3471 | * | |
3472 | * Allocate an object from this cache. The flags are only relevant | |
3473 | * if the cache has no available objects. | |
3474 | */ | |
343e0d7a | 3475 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3476 | { |
48356303 | 3477 | void *ret = slab_alloc(cachep, flags, _RET_IP_); |
36555751 | 3478 | |
ca2b84cb | 3479 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3480 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3481 | |
3482 | return ret; | |
1da177e4 LT |
3483 | } |
3484 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3485 | ||
0f24f128 | 3486 | #ifdef CONFIG_TRACING |
85beb586 | 3487 | void * |
4052147c | 3488 | kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) |
36555751 | 3489 | { |
85beb586 SR |
3490 | void *ret; |
3491 | ||
48356303 | 3492 | ret = slab_alloc(cachep, flags, _RET_IP_); |
85beb586 SR |
3493 | |
3494 | trace_kmalloc(_RET_IP_, ret, | |
ff4fcd01 | 3495 | size, cachep->size, flags); |
85beb586 | 3496 | return ret; |
36555751 | 3497 | } |
85beb586 | 3498 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3499 | #endif |
3500 | ||
1da177e4 | 3501 | #ifdef CONFIG_NUMA |
d0d04b78 ZL |
3502 | /** |
3503 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3504 | * @cachep: The cache to allocate from. | |
3505 | * @flags: See kmalloc(). | |
3506 | * @nodeid: node number of the target node. | |
3507 | * | |
3508 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3509 | * node, which can improve the performance for cpu bound structures. | |
3510 | * | |
3511 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3512 | */ | |
8b98c169 CH |
3513 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3514 | { | |
48356303 | 3515 | void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
36555751 | 3516 | |
ca2b84cb | 3517 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3518 | cachep->object_size, cachep->size, |
ca2b84cb | 3519 | flags, nodeid); |
36555751 EGM |
3520 | |
3521 | return ret; | |
8b98c169 | 3522 | } |
1da177e4 LT |
3523 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3524 | ||
0f24f128 | 3525 | #ifdef CONFIG_TRACING |
4052147c | 3526 | void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, |
85beb586 | 3527 | gfp_t flags, |
4052147c EG |
3528 | int nodeid, |
3529 | size_t size) | |
36555751 | 3530 | { |
85beb586 SR |
3531 | void *ret; |
3532 | ||
592f4145 | 3533 | ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
7c0cb9c6 | 3534 | |
85beb586 | 3535 | trace_kmalloc_node(_RET_IP_, ret, |
ff4fcd01 | 3536 | size, cachep->size, |
85beb586 SR |
3537 | flags, nodeid); |
3538 | return ret; | |
36555751 | 3539 | } |
85beb586 | 3540 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3541 | #endif |
3542 | ||
8b98c169 | 3543 | static __always_inline void * |
7c0cb9c6 | 3544 | __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) |
97e2bde4 | 3545 | { |
343e0d7a | 3546 | struct kmem_cache *cachep; |
97e2bde4 | 3547 | |
2c59dd65 | 3548 | cachep = kmalloc_slab(size, flags); |
6cb8f913 CL |
3549 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3550 | return cachep; | |
4052147c | 3551 | return kmem_cache_alloc_node_trace(cachep, flags, node, size); |
97e2bde4 | 3552 | } |
8b98c169 | 3553 | |
0bb38a5c | 3554 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3555 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3556 | { | |
7c0cb9c6 | 3557 | return __do_kmalloc_node(size, flags, node, _RET_IP_); |
8b98c169 | 3558 | } |
dbe5e69d | 3559 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3560 | |
3561 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3562 | int node, unsigned long caller) |
8b98c169 | 3563 | { |
7c0cb9c6 | 3564 | return __do_kmalloc_node(size, flags, node, caller); |
8b98c169 CH |
3565 | } |
3566 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3567 | #else | |
3568 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3569 | { | |
7c0cb9c6 | 3570 | return __do_kmalloc_node(size, flags, node, 0); |
8b98c169 CH |
3571 | } |
3572 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3573 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3574 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3575 | |
3576 | /** | |
800590f5 | 3577 | * __do_kmalloc - allocate memory |
1da177e4 | 3578 | * @size: how many bytes of memory are required. |
800590f5 | 3579 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3580 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3581 | */ |
7fd6b141 | 3582 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
7c0cb9c6 | 3583 | unsigned long caller) |
1da177e4 | 3584 | { |
343e0d7a | 3585 | struct kmem_cache *cachep; |
36555751 | 3586 | void *ret; |
1da177e4 | 3587 | |
97e2bde4 MS |
3588 | /* If you want to save a few bytes .text space: replace |
3589 | * __ with kmem_. | |
3590 | * Then kmalloc uses the uninlined functions instead of the inline | |
3591 | * functions. | |
3592 | */ | |
2c59dd65 | 3593 | cachep = kmalloc_slab(size, flags); |
a5c96d8a LT |
3594 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3595 | return cachep; | |
48356303 | 3596 | ret = slab_alloc(cachep, flags, caller); |
36555751 | 3597 | |
7c0cb9c6 | 3598 | trace_kmalloc(caller, ret, |
3b0efdfa | 3599 | size, cachep->size, flags); |
36555751 EGM |
3600 | |
3601 | return ret; | |
7fd6b141 PE |
3602 | } |
3603 | ||
7fd6b141 | 3604 | |
0bb38a5c | 3605 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3606 | void *__kmalloc(size_t size, gfp_t flags) |
3607 | { | |
7c0cb9c6 | 3608 | return __do_kmalloc(size, flags, _RET_IP_); |
1da177e4 LT |
3609 | } |
3610 | EXPORT_SYMBOL(__kmalloc); | |
3611 | ||
ce71e27c | 3612 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3613 | { |
7c0cb9c6 | 3614 | return __do_kmalloc(size, flags, caller); |
7fd6b141 PE |
3615 | } |
3616 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3617 | |
3618 | #else | |
3619 | void *__kmalloc(size_t size, gfp_t flags) | |
3620 | { | |
7c0cb9c6 | 3621 | return __do_kmalloc(size, flags, 0); |
1d2c8eea CH |
3622 | } |
3623 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3624 | #endif |
3625 | ||
1da177e4 LT |
3626 | /** |
3627 | * kmem_cache_free - Deallocate an object | |
3628 | * @cachep: The cache the allocation was from. | |
3629 | * @objp: The previously allocated object. | |
3630 | * | |
3631 | * Free an object which was previously allocated from this | |
3632 | * cache. | |
3633 | */ | |
343e0d7a | 3634 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3635 | { |
3636 | unsigned long flags; | |
b9ce5ef4 GC |
3637 | cachep = cache_from_obj(cachep, objp); |
3638 | if (!cachep) | |
3639 | return; | |
1da177e4 LT |
3640 | |
3641 | local_irq_save(flags); | |
d97d476b | 3642 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3643 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3644 | debug_check_no_obj_freed(objp, cachep->object_size); |
7c0cb9c6 | 3645 | __cache_free(cachep, objp, _RET_IP_); |
1da177e4 | 3646 | local_irq_restore(flags); |
36555751 | 3647 | |
ca2b84cb | 3648 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3649 | } |
3650 | EXPORT_SYMBOL(kmem_cache_free); | |
3651 | ||
1da177e4 LT |
3652 | /** |
3653 | * kfree - free previously allocated memory | |
3654 | * @objp: pointer returned by kmalloc. | |
3655 | * | |
80e93eff PE |
3656 | * If @objp is NULL, no operation is performed. |
3657 | * | |
1da177e4 LT |
3658 | * Don't free memory not originally allocated by kmalloc() |
3659 | * or you will run into trouble. | |
3660 | */ | |
3661 | void kfree(const void *objp) | |
3662 | { | |
343e0d7a | 3663 | struct kmem_cache *c; |
1da177e4 LT |
3664 | unsigned long flags; |
3665 | ||
2121db74 PE |
3666 | trace_kfree(_RET_IP_, objp); |
3667 | ||
6cb8f913 | 3668 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3669 | return; |
3670 | local_irq_save(flags); | |
3671 | kfree_debugcheck(objp); | |
6ed5eb22 | 3672 | c = virt_to_cache(objp); |
8c138bc0 CL |
3673 | debug_check_no_locks_freed(objp, c->object_size); |
3674 | ||
3675 | debug_check_no_obj_freed(objp, c->object_size); | |
7c0cb9c6 | 3676 | __cache_free(c, (void *)objp, _RET_IP_); |
1da177e4 LT |
3677 | local_irq_restore(flags); |
3678 | } | |
3679 | EXPORT_SYMBOL(kfree); | |
3680 | ||
e498be7d | 3681 | /* |
ce8eb6c4 | 3682 | * This initializes kmem_cache_node or resizes various caches for all nodes. |
e498be7d | 3683 | */ |
83b519e8 | 3684 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3685 | { |
3686 | int node; | |
ce8eb6c4 | 3687 | struct kmem_cache_node *n; |
cafeb02e | 3688 | struct array_cache *new_shared; |
3395ee05 | 3689 | struct array_cache **new_alien = NULL; |
e498be7d | 3690 | |
9c09a95c | 3691 | for_each_online_node(node) { |
cafeb02e | 3692 | |
3395ee05 | 3693 | if (use_alien_caches) { |
83b519e8 | 3694 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3695 | if (!new_alien) |
3696 | goto fail; | |
3697 | } | |
cafeb02e | 3698 | |
63109846 ED |
3699 | new_shared = NULL; |
3700 | if (cachep->shared) { | |
3701 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3702 | cachep->shared*cachep->batchcount, |
83b519e8 | 3703 | 0xbaadf00d, gfp); |
63109846 ED |
3704 | if (!new_shared) { |
3705 | free_alien_cache(new_alien); | |
3706 | goto fail; | |
3707 | } | |
0718dc2a | 3708 | } |
cafeb02e | 3709 | |
ce8eb6c4 CL |
3710 | n = cachep->node[node]; |
3711 | if (n) { | |
3712 | struct array_cache *shared = n->shared; | |
cafeb02e | 3713 | |
ce8eb6c4 | 3714 | spin_lock_irq(&n->list_lock); |
e498be7d | 3715 | |
cafeb02e | 3716 | if (shared) |
0718dc2a CL |
3717 | free_block(cachep, shared->entry, |
3718 | shared->avail, node); | |
e498be7d | 3719 | |
ce8eb6c4 CL |
3720 | n->shared = new_shared; |
3721 | if (!n->alien) { | |
3722 | n->alien = new_alien; | |
e498be7d CL |
3723 | new_alien = NULL; |
3724 | } | |
ce8eb6c4 | 3725 | n->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3726 | cachep->batchcount + cachep->num; |
ce8eb6c4 | 3727 | spin_unlock_irq(&n->list_lock); |
cafeb02e | 3728 | kfree(shared); |
e498be7d CL |
3729 | free_alien_cache(new_alien); |
3730 | continue; | |
3731 | } | |
ce8eb6c4 CL |
3732 | n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node); |
3733 | if (!n) { | |
0718dc2a CL |
3734 | free_alien_cache(new_alien); |
3735 | kfree(new_shared); | |
e498be7d | 3736 | goto fail; |
0718dc2a | 3737 | } |
e498be7d | 3738 | |
ce8eb6c4 CL |
3739 | kmem_cache_node_init(n); |
3740 | n->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3741 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
ce8eb6c4 CL |
3742 | n->shared = new_shared; |
3743 | n->alien = new_alien; | |
3744 | n->free_limit = (1 + nr_cpus_node(node)) * | |
a737b3e2 | 3745 | cachep->batchcount + cachep->num; |
ce8eb6c4 | 3746 | cachep->node[node] = n; |
e498be7d | 3747 | } |
cafeb02e | 3748 | return 0; |
0718dc2a | 3749 | |
a737b3e2 | 3750 | fail: |
3b0efdfa | 3751 | if (!cachep->list.next) { |
0718dc2a CL |
3752 | /* Cache is not active yet. Roll back what we did */ |
3753 | node--; | |
3754 | while (node >= 0) { | |
6a67368c | 3755 | if (cachep->node[node]) { |
ce8eb6c4 | 3756 | n = cachep->node[node]; |
0718dc2a | 3757 | |
ce8eb6c4 CL |
3758 | kfree(n->shared); |
3759 | free_alien_cache(n->alien); | |
3760 | kfree(n); | |
6a67368c | 3761 | cachep->node[node] = NULL; |
0718dc2a CL |
3762 | } |
3763 | node--; | |
3764 | } | |
3765 | } | |
cafeb02e | 3766 | return -ENOMEM; |
e498be7d CL |
3767 | } |
3768 | ||
1da177e4 | 3769 | struct ccupdate_struct { |
343e0d7a | 3770 | struct kmem_cache *cachep; |
acfe7d74 | 3771 | struct array_cache *new[0]; |
1da177e4 LT |
3772 | }; |
3773 | ||
3774 | static void do_ccupdate_local(void *info) | |
3775 | { | |
a737b3e2 | 3776 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3777 | struct array_cache *old; |
3778 | ||
3779 | check_irq_off(); | |
9a2dba4b | 3780 | old = cpu_cache_get(new->cachep); |
e498be7d | 3781 | |
1da177e4 LT |
3782 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3783 | new->new[smp_processor_id()] = old; | |
3784 | } | |
3785 | ||
18004c5d | 3786 | /* Always called with the slab_mutex held */ |
943a451a | 3787 | static int __do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 3788 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 3789 | { |
d2e7b7d0 | 3790 | struct ccupdate_struct *new; |
2ed3a4ef | 3791 | int i; |
1da177e4 | 3792 | |
acfe7d74 ED |
3793 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
3794 | gfp); | |
d2e7b7d0 SS |
3795 | if (!new) |
3796 | return -ENOMEM; | |
3797 | ||
e498be7d | 3798 | for_each_online_cpu(i) { |
7d6e6d09 | 3799 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 3800 | batchcount, gfp); |
d2e7b7d0 | 3801 | if (!new->new[i]) { |
b28a02de | 3802 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3803 | kfree(new->new[i]); |
3804 | kfree(new); | |
e498be7d | 3805 | return -ENOMEM; |
1da177e4 LT |
3806 | } |
3807 | } | |
d2e7b7d0 | 3808 | new->cachep = cachep; |
1da177e4 | 3809 | |
15c8b6c1 | 3810 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 3811 | |
1da177e4 | 3812 | check_irq_on(); |
1da177e4 LT |
3813 | cachep->batchcount = batchcount; |
3814 | cachep->limit = limit; | |
e498be7d | 3815 | cachep->shared = shared; |
1da177e4 | 3816 | |
e498be7d | 3817 | for_each_online_cpu(i) { |
d2e7b7d0 | 3818 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3819 | if (!ccold) |
3820 | continue; | |
6a67368c | 3821 | spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock); |
7d6e6d09 | 3822 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); |
6a67368c | 3823 | spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock); |
1da177e4 LT |
3824 | kfree(ccold); |
3825 | } | |
d2e7b7d0 | 3826 | kfree(new); |
83b519e8 | 3827 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
3828 | } |
3829 | ||
943a451a GC |
3830 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3831 | int batchcount, int shared, gfp_t gfp) | |
3832 | { | |
3833 | int ret; | |
3834 | struct kmem_cache *c = NULL; | |
3835 | int i = 0; | |
3836 | ||
3837 | ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp); | |
3838 | ||
3839 | if (slab_state < FULL) | |
3840 | return ret; | |
3841 | ||
3842 | if ((ret < 0) || !is_root_cache(cachep)) | |
3843 | return ret; | |
3844 | ||
ebe945c2 | 3845 | VM_BUG_ON(!mutex_is_locked(&slab_mutex)); |
943a451a GC |
3846 | for_each_memcg_cache_index(i) { |
3847 | c = cache_from_memcg(cachep, i); | |
3848 | if (c) | |
3849 | /* return value determined by the parent cache only */ | |
3850 | __do_tune_cpucache(c, limit, batchcount, shared, gfp); | |
3851 | } | |
3852 | ||
3853 | return ret; | |
3854 | } | |
3855 | ||
18004c5d | 3856 | /* Called with slab_mutex held always */ |
83b519e8 | 3857 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
3858 | { |
3859 | int err; | |
943a451a GC |
3860 | int limit = 0; |
3861 | int shared = 0; | |
3862 | int batchcount = 0; | |
3863 | ||
3864 | if (!is_root_cache(cachep)) { | |
3865 | struct kmem_cache *root = memcg_root_cache(cachep); | |
3866 | limit = root->limit; | |
3867 | shared = root->shared; | |
3868 | batchcount = root->batchcount; | |
3869 | } | |
1da177e4 | 3870 | |
943a451a GC |
3871 | if (limit && shared && batchcount) |
3872 | goto skip_setup; | |
a737b3e2 AM |
3873 | /* |
3874 | * The head array serves three purposes: | |
1da177e4 LT |
3875 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3876 | * - reduce the number of spinlock operations. | |
a737b3e2 | 3877 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
3878 | * bufctl chains: array operations are cheaper. |
3879 | * The numbers are guessed, we should auto-tune as described by | |
3880 | * Bonwick. | |
3881 | */ | |
3b0efdfa | 3882 | if (cachep->size > 131072) |
1da177e4 | 3883 | limit = 1; |
3b0efdfa | 3884 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 3885 | limit = 8; |
3b0efdfa | 3886 | else if (cachep->size > 1024) |
1da177e4 | 3887 | limit = 24; |
3b0efdfa | 3888 | else if (cachep->size > 256) |
1da177e4 LT |
3889 | limit = 54; |
3890 | else | |
3891 | limit = 120; | |
3892 | ||
a737b3e2 AM |
3893 | /* |
3894 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
3895 | * allocation behaviour: Most allocs on one cpu, most free operations |
3896 | * on another cpu. For these cases, an efficient object passing between | |
3897 | * cpus is necessary. This is provided by a shared array. The array | |
3898 | * replaces Bonwick's magazine layer. | |
3899 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
3900 | * to a larger limit. Thus disabled by default. | |
3901 | */ | |
3902 | shared = 0; | |
3b0efdfa | 3903 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 3904 | shared = 8; |
1da177e4 LT |
3905 | |
3906 | #if DEBUG | |
a737b3e2 AM |
3907 | /* |
3908 | * With debugging enabled, large batchcount lead to excessively long | |
3909 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
3910 | */ |
3911 | if (limit > 32) | |
3912 | limit = 32; | |
3913 | #endif | |
943a451a GC |
3914 | batchcount = (limit + 1) / 2; |
3915 | skip_setup: | |
3916 | err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp); | |
1da177e4 LT |
3917 | if (err) |
3918 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 3919 | cachep->name, -err); |
2ed3a4ef | 3920 | return err; |
1da177e4 LT |
3921 | } |
3922 | ||
1b55253a | 3923 | /* |
ce8eb6c4 CL |
3924 | * Drain an array if it contains any elements taking the node lock only if |
3925 | * necessary. Note that the node listlock also protects the array_cache | |
b18e7e65 | 3926 | * if drain_array() is used on the shared array. |
1b55253a | 3927 | */ |
ce8eb6c4 | 3928 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n, |
1b55253a | 3929 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
3930 | { |
3931 | int tofree; | |
3932 | ||
1b55253a CL |
3933 | if (!ac || !ac->avail) |
3934 | return; | |
1da177e4 LT |
3935 | if (ac->touched && !force) { |
3936 | ac->touched = 0; | |
b18e7e65 | 3937 | } else { |
ce8eb6c4 | 3938 | spin_lock_irq(&n->list_lock); |
b18e7e65 CL |
3939 | if (ac->avail) { |
3940 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
3941 | if (tofree > ac->avail) | |
3942 | tofree = (ac->avail + 1) / 2; | |
3943 | free_block(cachep, ac->entry, tofree, node); | |
3944 | ac->avail -= tofree; | |
3945 | memmove(ac->entry, &(ac->entry[tofree]), | |
3946 | sizeof(void *) * ac->avail); | |
3947 | } | |
ce8eb6c4 | 3948 | spin_unlock_irq(&n->list_lock); |
1da177e4 LT |
3949 | } |
3950 | } | |
3951 | ||
3952 | /** | |
3953 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 3954 | * @w: work descriptor |
1da177e4 LT |
3955 | * |
3956 | * Called from workqueue/eventd every few seconds. | |
3957 | * Purpose: | |
3958 | * - clear the per-cpu caches for this CPU. | |
3959 | * - return freeable pages to the main free memory pool. | |
3960 | * | |
a737b3e2 AM |
3961 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
3962 | * again on the next iteration. | |
1da177e4 | 3963 | */ |
7c5cae36 | 3964 | static void cache_reap(struct work_struct *w) |
1da177e4 | 3965 | { |
7a7c381d | 3966 | struct kmem_cache *searchp; |
ce8eb6c4 | 3967 | struct kmem_cache_node *n; |
7d6e6d09 | 3968 | int node = numa_mem_id(); |
bf6aede7 | 3969 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 3970 | |
18004c5d | 3971 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 3972 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 3973 | goto out; |
1da177e4 | 3974 | |
18004c5d | 3975 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
3976 | check_irq_on(); |
3977 | ||
35386e3b | 3978 | /* |
ce8eb6c4 | 3979 | * We only take the node lock if absolutely necessary and we |
35386e3b CL |
3980 | * have established with reasonable certainty that |
3981 | * we can do some work if the lock was obtained. | |
3982 | */ | |
ce8eb6c4 | 3983 | n = searchp->node[node]; |
35386e3b | 3984 | |
ce8eb6c4 | 3985 | reap_alien(searchp, n); |
1da177e4 | 3986 | |
ce8eb6c4 | 3987 | drain_array(searchp, n, cpu_cache_get(searchp), 0, node); |
1da177e4 | 3988 | |
35386e3b CL |
3989 | /* |
3990 | * These are racy checks but it does not matter | |
3991 | * if we skip one check or scan twice. | |
3992 | */ | |
ce8eb6c4 | 3993 | if (time_after(n->next_reap, jiffies)) |
35386e3b | 3994 | goto next; |
1da177e4 | 3995 | |
ce8eb6c4 | 3996 | n->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 3997 | |
ce8eb6c4 | 3998 | drain_array(searchp, n, n->shared, 0, node); |
1da177e4 | 3999 | |
ce8eb6c4 CL |
4000 | if (n->free_touched) |
4001 | n->free_touched = 0; | |
ed11d9eb CL |
4002 | else { |
4003 | int freed; | |
1da177e4 | 4004 | |
ce8eb6c4 | 4005 | freed = drain_freelist(searchp, n, (n->free_limit + |
ed11d9eb CL |
4006 | 5 * searchp->num - 1) / (5 * searchp->num)); |
4007 | STATS_ADD_REAPED(searchp, freed); | |
4008 | } | |
35386e3b | 4009 | next: |
1da177e4 LT |
4010 | cond_resched(); |
4011 | } | |
4012 | check_irq_on(); | |
18004c5d | 4013 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4014 | next_reap_node(); |
7c5cae36 | 4015 | out: |
a737b3e2 | 4016 | /* Set up the next iteration */ |
7c5cae36 | 4017 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4018 | } |
4019 | ||
158a9624 | 4020 | #ifdef CONFIG_SLABINFO |
0d7561c6 | 4021 | void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) |
1da177e4 | 4022 | { |
8456a648 | 4023 | struct page *page; |
b28a02de PE |
4024 | unsigned long active_objs; |
4025 | unsigned long num_objs; | |
4026 | unsigned long active_slabs = 0; | |
4027 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4028 | const char *name; |
1da177e4 | 4029 | char *error = NULL; |
e498be7d | 4030 | int node; |
ce8eb6c4 | 4031 | struct kmem_cache_node *n; |
1da177e4 | 4032 | |
1da177e4 LT |
4033 | active_objs = 0; |
4034 | num_slabs = 0; | |
e498be7d | 4035 | for_each_online_node(node) { |
ce8eb6c4 CL |
4036 | n = cachep->node[node]; |
4037 | if (!n) | |
e498be7d CL |
4038 | continue; |
4039 | ||
ca3b9b91 | 4040 | check_irq_on(); |
ce8eb6c4 | 4041 | spin_lock_irq(&n->list_lock); |
e498be7d | 4042 | |
8456a648 JK |
4043 | list_for_each_entry(page, &n->slabs_full, lru) { |
4044 | if (page->active != cachep->num && !error) | |
e498be7d CL |
4045 | error = "slabs_full accounting error"; |
4046 | active_objs += cachep->num; | |
4047 | active_slabs++; | |
4048 | } | |
8456a648 JK |
4049 | list_for_each_entry(page, &n->slabs_partial, lru) { |
4050 | if (page->active == cachep->num && !error) | |
106a74e1 | 4051 | error = "slabs_partial accounting error"; |
8456a648 | 4052 | if (!page->active && !error) |
106a74e1 | 4053 | error = "slabs_partial accounting error"; |
8456a648 | 4054 | active_objs += page->active; |
e498be7d CL |
4055 | active_slabs++; |
4056 | } | |
8456a648 JK |
4057 | list_for_each_entry(page, &n->slabs_free, lru) { |
4058 | if (page->active && !error) | |
106a74e1 | 4059 | error = "slabs_free accounting error"; |
e498be7d CL |
4060 | num_slabs++; |
4061 | } | |
ce8eb6c4 CL |
4062 | free_objects += n->free_objects; |
4063 | if (n->shared) | |
4064 | shared_avail += n->shared->avail; | |
e498be7d | 4065 | |
ce8eb6c4 | 4066 | spin_unlock_irq(&n->list_lock); |
1da177e4 | 4067 | } |
b28a02de PE |
4068 | num_slabs += active_slabs; |
4069 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4070 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4071 | error = "free_objects accounting error"; |
4072 | ||
b28a02de | 4073 | name = cachep->name; |
1da177e4 LT |
4074 | if (error) |
4075 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4076 | ||
0d7561c6 GC |
4077 | sinfo->active_objs = active_objs; |
4078 | sinfo->num_objs = num_objs; | |
4079 | sinfo->active_slabs = active_slabs; | |
4080 | sinfo->num_slabs = num_slabs; | |
4081 | sinfo->shared_avail = shared_avail; | |
4082 | sinfo->limit = cachep->limit; | |
4083 | sinfo->batchcount = cachep->batchcount; | |
4084 | sinfo->shared = cachep->shared; | |
4085 | sinfo->objects_per_slab = cachep->num; | |
4086 | sinfo->cache_order = cachep->gfporder; | |
4087 | } | |
4088 | ||
4089 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) | |
4090 | { | |
1da177e4 | 4091 | #if STATS |
ce8eb6c4 | 4092 | { /* node stats */ |
1da177e4 LT |
4093 | unsigned long high = cachep->high_mark; |
4094 | unsigned long allocs = cachep->num_allocations; | |
4095 | unsigned long grown = cachep->grown; | |
4096 | unsigned long reaped = cachep->reaped; | |
4097 | unsigned long errors = cachep->errors; | |
4098 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4099 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4100 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4101 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4102 | |
e92dd4fd JP |
4103 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4104 | "%4lu %4lu %4lu %4lu %4lu", | |
4105 | allocs, high, grown, | |
4106 | reaped, errors, max_freeable, node_allocs, | |
4107 | node_frees, overflows); | |
1da177e4 LT |
4108 | } |
4109 | /* cpu stats */ | |
4110 | { | |
4111 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4112 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4113 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4114 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4115 | ||
4116 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4117 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4118 | } |
4119 | #endif | |
1da177e4 LT |
4120 | } |
4121 | ||
1da177e4 LT |
4122 | #define MAX_SLABINFO_WRITE 128 |
4123 | /** | |
4124 | * slabinfo_write - Tuning for the slab allocator | |
4125 | * @file: unused | |
4126 | * @buffer: user buffer | |
4127 | * @count: data length | |
4128 | * @ppos: unused | |
4129 | */ | |
b7454ad3 | 4130 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4131 | size_t count, loff_t *ppos) |
1da177e4 | 4132 | { |
b28a02de | 4133 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4134 | int limit, batchcount, shared, res; |
7a7c381d | 4135 | struct kmem_cache *cachep; |
b28a02de | 4136 | |
1da177e4 LT |
4137 | if (count > MAX_SLABINFO_WRITE) |
4138 | return -EINVAL; | |
4139 | if (copy_from_user(&kbuf, buffer, count)) | |
4140 | return -EFAULT; | |
b28a02de | 4141 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4142 | |
4143 | tmp = strchr(kbuf, ' '); | |
4144 | if (!tmp) | |
4145 | return -EINVAL; | |
4146 | *tmp = '\0'; | |
4147 | tmp++; | |
4148 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4149 | return -EINVAL; | |
4150 | ||
4151 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4152 | mutex_lock(&slab_mutex); |
1da177e4 | 4153 | res = -EINVAL; |
18004c5d | 4154 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4155 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4156 | if (limit < 1 || batchcount < 1 || |
4157 | batchcount > limit || shared < 0) { | |
e498be7d | 4158 | res = 0; |
1da177e4 | 4159 | } else { |
e498be7d | 4160 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4161 | batchcount, shared, |
4162 | GFP_KERNEL); | |
1da177e4 LT |
4163 | } |
4164 | break; | |
4165 | } | |
4166 | } | |
18004c5d | 4167 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4168 | if (res >= 0) |
4169 | res = count; | |
4170 | return res; | |
4171 | } | |
871751e2 AV |
4172 | |
4173 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4174 | ||
4175 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4176 | { | |
18004c5d CL |
4177 | mutex_lock(&slab_mutex); |
4178 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4179 | } |
4180 | ||
4181 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4182 | { | |
4183 | unsigned long *p; | |
4184 | int l; | |
4185 | if (!v) | |
4186 | return 1; | |
4187 | l = n[1]; | |
4188 | p = n + 2; | |
4189 | while (l) { | |
4190 | int i = l/2; | |
4191 | unsigned long *q = p + 2 * i; | |
4192 | if (*q == v) { | |
4193 | q[1]++; | |
4194 | return 1; | |
4195 | } | |
4196 | if (*q > v) { | |
4197 | l = i; | |
4198 | } else { | |
4199 | p = q + 2; | |
4200 | l -= i + 1; | |
4201 | } | |
4202 | } | |
4203 | if (++n[1] == n[0]) | |
4204 | return 0; | |
4205 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4206 | p[0] = v; | |
4207 | p[1] = 1; | |
4208 | return 1; | |
4209 | } | |
4210 | ||
8456a648 JK |
4211 | static void handle_slab(unsigned long *n, struct kmem_cache *c, |
4212 | struct page *page) | |
871751e2 AV |
4213 | { |
4214 | void *p; | |
b1cb0982 JK |
4215 | int i, j; |
4216 | ||
871751e2 AV |
4217 | if (n[0] == n[1]) |
4218 | return; | |
8456a648 | 4219 | for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) { |
b1cb0982 JK |
4220 | bool active = true; |
4221 | ||
8456a648 | 4222 | for (j = page->active; j < c->num; j++) { |
b1cb0982 | 4223 | /* Skip freed item */ |
8456a648 | 4224 | if (slab_bufctl(page)[j] == i) { |
b1cb0982 JK |
4225 | active = false; |
4226 | break; | |
4227 | } | |
4228 | } | |
4229 | if (!active) | |
871751e2 | 4230 | continue; |
b1cb0982 | 4231 | |
871751e2 AV |
4232 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) |
4233 | return; | |
4234 | } | |
4235 | } | |
4236 | ||
4237 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4238 | { | |
4239 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4240 | unsigned long offset, size; |
9281acea | 4241 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4242 | |
a5c43dae | 4243 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4244 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4245 | if (modname[0]) |
871751e2 AV |
4246 | seq_printf(m, " [%s]", modname); |
4247 | return; | |
4248 | } | |
4249 | #endif | |
4250 | seq_printf(m, "%p", (void *)address); | |
4251 | } | |
4252 | ||
4253 | static int leaks_show(struct seq_file *m, void *p) | |
4254 | { | |
0672aa7c | 4255 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
8456a648 | 4256 | struct page *page; |
ce8eb6c4 | 4257 | struct kmem_cache_node *n; |
871751e2 | 4258 | const char *name; |
db845067 | 4259 | unsigned long *x = m->private; |
871751e2 AV |
4260 | int node; |
4261 | int i; | |
4262 | ||
4263 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4264 | return 0; | |
4265 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4266 | return 0; | |
4267 | ||
4268 | /* OK, we can do it */ | |
4269 | ||
db845067 | 4270 | x[1] = 0; |
871751e2 AV |
4271 | |
4272 | for_each_online_node(node) { | |
ce8eb6c4 CL |
4273 | n = cachep->node[node]; |
4274 | if (!n) | |
871751e2 AV |
4275 | continue; |
4276 | ||
4277 | check_irq_on(); | |
ce8eb6c4 | 4278 | spin_lock_irq(&n->list_lock); |
871751e2 | 4279 | |
8456a648 JK |
4280 | list_for_each_entry(page, &n->slabs_full, lru) |
4281 | handle_slab(x, cachep, page); | |
4282 | list_for_each_entry(page, &n->slabs_partial, lru) | |
4283 | handle_slab(x, cachep, page); | |
ce8eb6c4 | 4284 | spin_unlock_irq(&n->list_lock); |
871751e2 AV |
4285 | } |
4286 | name = cachep->name; | |
db845067 | 4287 | if (x[0] == x[1]) { |
871751e2 | 4288 | /* Increase the buffer size */ |
18004c5d | 4289 | mutex_unlock(&slab_mutex); |
db845067 | 4290 | m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
871751e2 AV |
4291 | if (!m->private) { |
4292 | /* Too bad, we are really out */ | |
db845067 | 4293 | m->private = x; |
18004c5d | 4294 | mutex_lock(&slab_mutex); |
871751e2 AV |
4295 | return -ENOMEM; |
4296 | } | |
db845067 CL |
4297 | *(unsigned long *)m->private = x[0] * 2; |
4298 | kfree(x); | |
18004c5d | 4299 | mutex_lock(&slab_mutex); |
871751e2 AV |
4300 | /* Now make sure this entry will be retried */ |
4301 | m->count = m->size; | |
4302 | return 0; | |
4303 | } | |
db845067 CL |
4304 | for (i = 0; i < x[1]; i++) { |
4305 | seq_printf(m, "%s: %lu ", name, x[2*i+3]); | |
4306 | show_symbol(m, x[2*i+2]); | |
871751e2 AV |
4307 | seq_putc(m, '\n'); |
4308 | } | |
d2e7b7d0 | 4309 | |
871751e2 AV |
4310 | return 0; |
4311 | } | |
4312 | ||
a0ec95a8 | 4313 | static const struct seq_operations slabstats_op = { |
871751e2 | 4314 | .start = leaks_start, |
276a2439 WL |
4315 | .next = slab_next, |
4316 | .stop = slab_stop, | |
871751e2 AV |
4317 | .show = leaks_show, |
4318 | }; | |
a0ec95a8 AD |
4319 | |
4320 | static int slabstats_open(struct inode *inode, struct file *file) | |
4321 | { | |
4322 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4323 | int ret = -ENOMEM; | |
4324 | if (n) { | |
4325 | ret = seq_open(file, &slabstats_op); | |
4326 | if (!ret) { | |
4327 | struct seq_file *m = file->private_data; | |
4328 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4329 | m->private = n; | |
4330 | n = NULL; | |
4331 | } | |
4332 | kfree(n); | |
4333 | } | |
4334 | return ret; | |
4335 | } | |
4336 | ||
4337 | static const struct file_operations proc_slabstats_operations = { | |
4338 | .open = slabstats_open, | |
4339 | .read = seq_read, | |
4340 | .llseek = seq_lseek, | |
4341 | .release = seq_release_private, | |
4342 | }; | |
4343 | #endif | |
4344 | ||
4345 | static int __init slab_proc_init(void) | |
4346 | { | |
4347 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4348 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4349 | #endif |
a0ec95a8 AD |
4350 | return 0; |
4351 | } | |
4352 | module_init(slab_proc_init); | |
1da177e4 LT |
4353 | #endif |
4354 | ||
00e145b6 MS |
4355 | /** |
4356 | * ksize - get the actual amount of memory allocated for a given object | |
4357 | * @objp: Pointer to the object | |
4358 | * | |
4359 | * kmalloc may internally round up allocations and return more memory | |
4360 | * than requested. ksize() can be used to determine the actual amount of | |
4361 | * memory allocated. The caller may use this additional memory, even though | |
4362 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4363 | * The caller must guarantee that objp points to a valid object previously | |
4364 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4365 | * must not be freed during the duration of the call. | |
4366 | */ | |
fd76bab2 | 4367 | size_t ksize(const void *objp) |
1da177e4 | 4368 | { |
ef8b4520 CL |
4369 | BUG_ON(!objp); |
4370 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4371 | return 0; |
1da177e4 | 4372 | |
8c138bc0 | 4373 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4374 | } |
b1aabecd | 4375 | EXPORT_SYMBOL(ksize); |