Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
183ff22b | 29 | * slabs and you must pass objects with the same initializations to |
1da177e4 LT |
30 | * kmem_cache_free. |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
54 | * it's changed with a smp_call_function(). |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
fc0abb14 | 71 | * The global cache-chain is protected by the mutex 'cache_chain_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
1da177e4 LT |
89 | #include <linux/slab.h> |
90 | #include <linux/mm.h> | |
c9cf5528 | 91 | #include <linux/poison.h> |
1da177e4 LT |
92 | #include <linux/swap.h> |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
101a5001 | 97 | #include <linux/cpuset.h> |
a0ec95a8 | 98 | #include <linux/proc_fs.h> |
1da177e4 LT |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> | |
101 | #include <linux/kallsyms.h> | |
102 | #include <linux/cpu.h> | |
103 | #include <linux/sysctl.h> | |
104 | #include <linux/module.h> | |
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 | |
1da177e4 LT |
120 | #include <asm/cacheflush.h> |
121 | #include <asm/tlbflush.h> | |
122 | #include <asm/page.h> | |
123 | ||
124 | /* | |
50953fe9 | 125 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
126 | * 0 for faster, smaller code (especially in the critical paths). |
127 | * | |
128 | * STATS - 1 to collect stats for /proc/slabinfo. | |
129 | * 0 for faster, smaller code (especially in the critical paths). | |
130 | * | |
131 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
132 | */ | |
133 | ||
134 | #ifdef CONFIG_DEBUG_SLAB | |
135 | #define DEBUG 1 | |
136 | #define STATS 1 | |
137 | #define FORCED_DEBUG 1 | |
138 | #else | |
139 | #define DEBUG 0 | |
140 | #define STATS 0 | |
141 | #define FORCED_DEBUG 0 | |
142 | #endif | |
143 | ||
1da177e4 LT |
144 | /* Shouldn't this be in a header file somewhere? */ |
145 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 146 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 147 | |
1da177e4 LT |
148 | #ifndef ARCH_KMALLOC_FLAGS |
149 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
150 | #endif | |
151 | ||
152 | /* Legal flag mask for kmem_cache_create(). */ | |
153 | #if DEBUG | |
50953fe9 | 154 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 155 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 156 | SLAB_CACHE_DMA | \ |
5af60839 | 157 | SLAB_STORE_USER | \ |
1da177e4 | 158 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 159 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 160 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 | 161 | #else |
ac2b898c | 162 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 163 | SLAB_CACHE_DMA | \ |
1da177e4 | 164 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 165 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 166 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 LT |
167 | #endif |
168 | ||
169 | /* | |
170 | * kmem_bufctl_t: | |
171 | * | |
172 | * Bufctl's are used for linking objs within a slab | |
173 | * linked offsets. | |
174 | * | |
175 | * This implementation relies on "struct page" for locating the cache & | |
176 | * slab an object belongs to. | |
177 | * This allows the bufctl structure to be small (one int), but limits | |
178 | * the number of objects a slab (not a cache) can contain when off-slab | |
179 | * bufctls are used. The limit is the size of the largest general cache | |
180 | * that does not use off-slab slabs. | |
181 | * For 32bit archs with 4 kB pages, is this 56. | |
182 | * This is not serious, as it is only for large objects, when it is unwise | |
183 | * to have too many per slab. | |
184 | * Note: This limit can be raised by introducing a general cache whose size | |
185 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
186 | */ | |
187 | ||
fa5b08d5 | 188 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
189 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
190 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
191 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
192 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 193 | |
1da177e4 LT |
194 | /* |
195 | * struct slab_rcu | |
196 | * | |
197 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
198 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
199 | * we need to approach a kernel structure obliquely, from its address | |
200 | * obtained without the usual locking. We can lock the structure to | |
201 | * stabilize it and check it's still at the given address, only if we | |
202 | * can be sure that the memory has not been meanwhile reused for some | |
203 | * other kind of object (which our subsystem's lock might corrupt). | |
204 | * | |
205 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
206 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
207 | */ |
208 | struct slab_rcu { | |
b28a02de | 209 | struct rcu_head head; |
343e0d7a | 210 | struct kmem_cache *cachep; |
b28a02de | 211 | void *addr; |
1da177e4 LT |
212 | }; |
213 | ||
5bfe53a7 LJ |
214 | /* |
215 | * struct slab | |
216 | * | |
217 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
218 | * for a slab, or allocated from an general cache. | |
219 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
220 | */ | |
221 | struct slab { | |
222 | union { | |
223 | struct { | |
224 | struct list_head list; | |
225 | unsigned long colouroff; | |
226 | void *s_mem; /* including colour offset */ | |
227 | unsigned int inuse; /* num of objs active in slab */ | |
228 | kmem_bufctl_t free; | |
229 | unsigned short nodeid; | |
230 | }; | |
231 | struct slab_rcu __slab_cover_slab_rcu; | |
232 | }; | |
233 | }; | |
234 | ||
1da177e4 LT |
235 | /* |
236 | * struct array_cache | |
237 | * | |
1da177e4 LT |
238 | * Purpose: |
239 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
240 | * - reduce the number of linked list operations | |
241 | * - reduce spinlock operations | |
242 | * | |
243 | * The limit is stored in the per-cpu structure to reduce the data cache | |
244 | * footprint. | |
245 | * | |
246 | */ | |
247 | struct array_cache { | |
248 | unsigned int avail; | |
249 | unsigned int limit; | |
250 | unsigned int batchcount; | |
251 | unsigned int touched; | |
e498be7d | 252 | spinlock_t lock; |
bda5b655 | 253 | void *entry[]; /* |
a737b3e2 AM |
254 | * Must have this definition in here for the proper |
255 | * alignment of array_cache. Also simplifies accessing | |
256 | * the entries. | |
a737b3e2 | 257 | */ |
1da177e4 LT |
258 | }; |
259 | ||
a737b3e2 AM |
260 | /* |
261 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
262 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
263 | */ |
264 | #define BOOT_CPUCACHE_ENTRIES 1 | |
265 | struct arraycache_init { | |
266 | struct array_cache cache; | |
b28a02de | 267 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
268 | }; |
269 | ||
270 | /* | |
e498be7d | 271 | * The slab lists for all objects. |
1da177e4 LT |
272 | */ |
273 | struct kmem_list3 { | |
b28a02de PE |
274 | struct list_head slabs_partial; /* partial list first, better asm code */ |
275 | struct list_head slabs_full; | |
276 | struct list_head slabs_free; | |
277 | unsigned long free_objects; | |
b28a02de | 278 | unsigned int free_limit; |
2e1217cf | 279 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
280 | spinlock_t list_lock; |
281 | struct array_cache *shared; /* shared per node */ | |
282 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
283 | unsigned long next_reap; /* updated without locking */ |
284 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
285 | }; |
286 | ||
e498be7d CL |
287 | /* |
288 | * Need this for bootstrapping a per node allocator. | |
289 | */ | |
556a169d | 290 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
68a1b195 | 291 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 292 | #define CACHE_CACHE 0 |
556a169d PE |
293 | #define SIZE_AC MAX_NUMNODES |
294 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 295 | |
ed11d9eb CL |
296 | static int drain_freelist(struct kmem_cache *cache, |
297 | struct kmem_list3 *l3, int tofree); | |
298 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
299 | int node); | |
83b519e8 | 300 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 301 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 302 | |
e498be7d | 303 | /* |
a737b3e2 AM |
304 | * This function must be completely optimized away if a constant is passed to |
305 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 306 | */ |
7243cc05 | 307 | static __always_inline int index_of(const size_t size) |
e498be7d | 308 | { |
5ec8a847 SR |
309 | extern void __bad_size(void); |
310 | ||
e498be7d CL |
311 | if (__builtin_constant_p(size)) { |
312 | int i = 0; | |
313 | ||
314 | #define CACHE(x) \ | |
315 | if (size <=x) \ | |
316 | return i; \ | |
317 | else \ | |
318 | i++; | |
1c61fc40 | 319 | #include <linux/kmalloc_sizes.h> |
e498be7d | 320 | #undef CACHE |
5ec8a847 | 321 | __bad_size(); |
7243cc05 | 322 | } else |
5ec8a847 | 323 | __bad_size(); |
e498be7d CL |
324 | return 0; |
325 | } | |
326 | ||
e0a42726 IM |
327 | static int slab_early_init = 1; |
328 | ||
e498be7d CL |
329 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
330 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 331 | |
5295a74c | 332 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
333 | { |
334 | INIT_LIST_HEAD(&parent->slabs_full); | |
335 | INIT_LIST_HEAD(&parent->slabs_partial); | |
336 | INIT_LIST_HEAD(&parent->slabs_free); | |
337 | parent->shared = NULL; | |
338 | parent->alien = NULL; | |
2e1217cf | 339 | parent->colour_next = 0; |
e498be7d CL |
340 | spin_lock_init(&parent->list_lock); |
341 | parent->free_objects = 0; | |
342 | parent->free_touched = 0; | |
343 | } | |
344 | ||
a737b3e2 AM |
345 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
346 | do { \ | |
347 | INIT_LIST_HEAD(listp); \ | |
348 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
349 | } while (0) |
350 | ||
a737b3e2 AM |
351 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
352 | do { \ | |
e498be7d CL |
353 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
354 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
355 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
356 | } while (0) | |
1da177e4 | 357 | |
1da177e4 LT |
358 | #define CFLGS_OFF_SLAB (0x80000000UL) |
359 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
360 | ||
361 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
362 | /* |
363 | * Optimization question: fewer reaps means less probability for unnessary | |
364 | * cpucache drain/refill cycles. | |
1da177e4 | 365 | * |
dc6f3f27 | 366 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
367 | * which could lock up otherwise freeable slabs. |
368 | */ | |
369 | #define REAPTIMEOUT_CPUC (2*HZ) | |
370 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
371 | ||
372 | #if STATS | |
373 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
374 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
375 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
376 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 377 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
378 | #define STATS_SET_HIGH(x) \ |
379 | do { \ | |
380 | if ((x)->num_active > (x)->high_mark) \ | |
381 | (x)->high_mark = (x)->num_active; \ | |
382 | } while (0) | |
1da177e4 LT |
383 | #define STATS_INC_ERR(x) ((x)->errors++) |
384 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 385 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 386 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
387 | #define STATS_SET_FREEABLE(x, i) \ |
388 | do { \ | |
389 | if ((x)->max_freeable < i) \ | |
390 | (x)->max_freeable = i; \ | |
391 | } while (0) | |
1da177e4 LT |
392 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
393 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
394 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
395 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
396 | #else | |
397 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
398 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
399 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
400 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 401 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
402 | #define STATS_SET_HIGH(x) do { } while (0) |
403 | #define STATS_INC_ERR(x) do { } while (0) | |
404 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 405 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 406 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 407 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
408 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
409 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
410 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
411 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
412 | #endif | |
413 | ||
414 | #if DEBUG | |
1da177e4 | 415 | |
a737b3e2 AM |
416 | /* |
417 | * memory layout of objects: | |
1da177e4 | 418 | * 0 : objp |
3dafccf2 | 419 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
420 | * the end of an object is aligned with the end of the real |
421 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 422 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 423 | * redzone word. |
3dafccf2 MS |
424 | * cachep->obj_offset: The real object. |
425 | * cachep->buffer_size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
a737b3e2 AM |
426 | * cachep->buffer_size - 1* BYTES_PER_WORD: last caller address |
427 | * [BYTES_PER_WORD long] | |
1da177e4 | 428 | */ |
343e0d7a | 429 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 430 | { |
3dafccf2 | 431 | return cachep->obj_offset; |
1da177e4 LT |
432 | } |
433 | ||
343e0d7a | 434 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 435 | { |
3dafccf2 | 436 | return cachep->obj_size; |
1da177e4 LT |
437 | } |
438 | ||
b46b8f19 | 439 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
440 | { |
441 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
442 | return (unsigned long long*) (objp + obj_offset(cachep) - |
443 | sizeof(unsigned long long)); | |
1da177e4 LT |
444 | } |
445 | ||
b46b8f19 | 446 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
447 | { |
448 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
449 | if (cachep->flags & SLAB_STORE_USER) | |
b46b8f19 DW |
450 | return (unsigned long long *)(objp + cachep->buffer_size - |
451 | sizeof(unsigned long long) - | |
87a927c7 | 452 | REDZONE_ALIGN); |
b46b8f19 DW |
453 | return (unsigned long long *) (objp + cachep->buffer_size - |
454 | sizeof(unsigned long long)); | |
1da177e4 LT |
455 | } |
456 | ||
343e0d7a | 457 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
458 | { |
459 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3dafccf2 | 460 | return (void **)(objp + cachep->buffer_size - BYTES_PER_WORD); |
1da177e4 LT |
461 | } |
462 | ||
463 | #else | |
464 | ||
3dafccf2 MS |
465 | #define obj_offset(x) 0 |
466 | #define obj_size(cachep) (cachep->buffer_size) | |
b46b8f19 DW |
467 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
468 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
469 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
470 | ||
471 | #endif | |
472 | ||
0f24f128 | 473 | #ifdef CONFIG_TRACING |
36555751 EGM |
474 | size_t slab_buffer_size(struct kmem_cache *cachep) |
475 | { | |
476 | return cachep->buffer_size; | |
477 | } | |
478 | EXPORT_SYMBOL(slab_buffer_size); | |
479 | #endif | |
480 | ||
1da177e4 LT |
481 | /* |
482 | * Do not go above this order unless 0 objects fit into the slab. | |
483 | */ | |
484 | #define BREAK_GFP_ORDER_HI 1 | |
485 | #define BREAK_GFP_ORDER_LO 0 | |
486 | static int slab_break_gfp_order = BREAK_GFP_ORDER_LO; | |
487 | ||
a737b3e2 AM |
488 | /* |
489 | * Functions for storing/retrieving the cachep and or slab from the page | |
490 | * allocator. These are used to find the slab an obj belongs to. With kfree(), | |
491 | * these are used to find the cache which an obj belongs to. | |
1da177e4 | 492 | */ |
065d41cb PE |
493 | static inline void page_set_cache(struct page *page, struct kmem_cache *cache) |
494 | { | |
495 | page->lru.next = (struct list_head *)cache; | |
496 | } | |
497 | ||
498 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
499 | { | |
d85f3385 | 500 | page = compound_head(page); |
ddc2e812 | 501 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
502 | return (struct kmem_cache *)page->lru.next; |
503 | } | |
504 | ||
505 | static inline void page_set_slab(struct page *page, struct slab *slab) | |
506 | { | |
507 | page->lru.prev = (struct list_head *)slab; | |
508 | } | |
509 | ||
510 | static inline struct slab *page_get_slab(struct page *page) | |
511 | { | |
ddc2e812 | 512 | BUG_ON(!PageSlab(page)); |
065d41cb PE |
513 | return (struct slab *)page->lru.prev; |
514 | } | |
1da177e4 | 515 | |
6ed5eb22 PE |
516 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
517 | { | |
b49af68f | 518 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
519 | return page_get_cache(page); |
520 | } | |
521 | ||
522 | static inline struct slab *virt_to_slab(const void *obj) | |
523 | { | |
b49af68f | 524 | struct page *page = virt_to_head_page(obj); |
6ed5eb22 PE |
525 | return page_get_slab(page); |
526 | } | |
527 | ||
8fea4e96 PE |
528 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
529 | unsigned int idx) | |
530 | { | |
531 | return slab->s_mem + cache->buffer_size * idx; | |
532 | } | |
533 | ||
6a2d7a95 ED |
534 | /* |
535 | * We want to avoid an expensive divide : (offset / cache->buffer_size) | |
536 | * Using the fact that buffer_size is a constant for a particular cache, | |
537 | * we can replace (offset / cache->buffer_size) by | |
538 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
539 | */ | |
540 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
541 | const struct slab *slab, void *obj) | |
8fea4e96 | 542 | { |
6a2d7a95 ED |
543 | u32 offset = (obj - slab->s_mem); |
544 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
545 | } |
546 | ||
a737b3e2 AM |
547 | /* |
548 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
549 | */ | |
1da177e4 LT |
550 | struct cache_sizes malloc_sizes[] = { |
551 | #define CACHE(x) { .cs_size = (x) }, | |
552 | #include <linux/kmalloc_sizes.h> | |
553 | CACHE(ULONG_MAX) | |
554 | #undef CACHE | |
555 | }; | |
556 | EXPORT_SYMBOL(malloc_sizes); | |
557 | ||
558 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
559 | struct cache_names { | |
560 | char *name; | |
561 | char *name_dma; | |
562 | }; | |
563 | ||
564 | static struct cache_names __initdata cache_names[] = { | |
565 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
566 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 567 | {NULL,} |
1da177e4 LT |
568 | #undef CACHE |
569 | }; | |
570 | ||
571 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 572 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 573 | static struct arraycache_init initarray_generic = |
b28a02de | 574 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
575 | |
576 | /* internal cache of cache description objs */ | |
b56efcf0 | 577 | static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES]; |
343e0d7a | 578 | static struct kmem_cache cache_cache = { |
b56efcf0 | 579 | .nodelists = cache_cache_nodelists, |
b28a02de PE |
580 | .batchcount = 1, |
581 | .limit = BOOT_CPUCACHE_ENTRIES, | |
582 | .shared = 1, | |
343e0d7a | 583 | .buffer_size = sizeof(struct kmem_cache), |
b28a02de | 584 | .name = "kmem_cache", |
1da177e4 LT |
585 | }; |
586 | ||
056c6241 RT |
587 | #define BAD_ALIEN_MAGIC 0x01020304ul |
588 | ||
ce79ddc8 PE |
589 | /* |
590 | * chicken and egg problem: delay the per-cpu array allocation | |
591 | * until the general caches are up. | |
592 | */ | |
593 | static enum { | |
594 | NONE, | |
595 | PARTIAL_AC, | |
596 | PARTIAL_L3, | |
597 | EARLY, | |
598 | FULL | |
599 | } g_cpucache_up; | |
600 | ||
601 | /* | |
602 | * used by boot code to determine if it can use slab based allocator | |
603 | */ | |
604 | int slab_is_available(void) | |
605 | { | |
606 | return g_cpucache_up >= EARLY; | |
607 | } | |
608 | ||
f1aaee53 AV |
609 | #ifdef CONFIG_LOCKDEP |
610 | ||
611 | /* | |
612 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
613 | * for other slabs "off slab". | |
614 | * The locking for this is tricky in that it nests within the locks | |
615 | * of all other slabs in a few places; to deal with this special | |
616 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
617 | * |
618 | * We set lock class for alien array caches which are up during init. | |
619 | * The lock annotation will be lost if all cpus of a node goes down and | |
620 | * then comes back up during hotplug | |
f1aaee53 | 621 | */ |
056c6241 RT |
622 | static struct lock_class_key on_slab_l3_key; |
623 | static struct lock_class_key on_slab_alc_key; | |
624 | ||
ce79ddc8 | 625 | static void init_node_lock_keys(int q) |
f1aaee53 | 626 | { |
056c6241 RT |
627 | struct cache_sizes *s = malloc_sizes; |
628 | ||
ce79ddc8 PE |
629 | if (g_cpucache_up != FULL) |
630 | return; | |
631 | ||
632 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
633 | struct array_cache **alc; | |
634 | struct kmem_list3 *l3; | |
635 | int r; | |
636 | ||
637 | l3 = s->cs_cachep->nodelists[q]; | |
638 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 639 | continue; |
ce79ddc8 PE |
640 | lockdep_set_class(&l3->list_lock, &on_slab_l3_key); |
641 | alc = l3->alien; | |
642 | /* | |
643 | * FIXME: This check for BAD_ALIEN_MAGIC | |
644 | * should go away when common slab code is taught to | |
645 | * work even without alien caches. | |
646 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
647 | * for alloc_alien_cache, | |
648 | */ | |
649 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
00afa758 | 650 | continue; |
ce79ddc8 PE |
651 | for_each_node(r) { |
652 | if (alc[r]) | |
653 | lockdep_set_class(&alc[r]->lock, | |
654 | &on_slab_alc_key); | |
056c6241 | 655 | } |
f1aaee53 AV |
656 | } |
657 | } | |
ce79ddc8 PE |
658 | |
659 | static inline void init_lock_keys(void) | |
660 | { | |
661 | int node; | |
662 | ||
663 | for_each_node(node) | |
664 | init_node_lock_keys(node); | |
665 | } | |
f1aaee53 | 666 | #else |
ce79ddc8 PE |
667 | static void init_node_lock_keys(int q) |
668 | { | |
669 | } | |
670 | ||
056c6241 | 671 | static inline void init_lock_keys(void) |
f1aaee53 AV |
672 | { |
673 | } | |
674 | #endif | |
675 | ||
8f5be20b | 676 | /* |
95402b38 | 677 | * Guard access to the cache-chain. |
8f5be20b | 678 | */ |
fc0abb14 | 679 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
680 | static struct list_head cache_chain; |
681 | ||
1871e52c | 682 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 683 | |
343e0d7a | 684 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
685 | { |
686 | return cachep->array[smp_processor_id()]; | |
687 | } | |
688 | ||
a737b3e2 AM |
689 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
690 | gfp_t gfpflags) | |
1da177e4 LT |
691 | { |
692 | struct cache_sizes *csizep = malloc_sizes; | |
693 | ||
694 | #if DEBUG | |
695 | /* This happens if someone tries to call | |
b28a02de PE |
696 | * kmem_cache_create(), or __kmalloc(), before |
697 | * the generic caches are initialized. | |
698 | */ | |
c7e43c78 | 699 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 700 | #endif |
6cb8f913 CL |
701 | if (!size) |
702 | return ZERO_SIZE_PTR; | |
703 | ||
1da177e4 LT |
704 | while (size > csizep->cs_size) |
705 | csizep++; | |
706 | ||
707 | /* | |
0abf40c1 | 708 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
709 | * has cs_{dma,}cachep==NULL. Thus no special case |
710 | * for large kmalloc calls required. | |
711 | */ | |
4b51d669 | 712 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
713 | if (unlikely(gfpflags & GFP_DMA)) |
714 | return csizep->cs_dmacachep; | |
4b51d669 | 715 | #endif |
1da177e4 LT |
716 | return csizep->cs_cachep; |
717 | } | |
718 | ||
b221385b | 719 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
720 | { |
721 | return __find_general_cachep(size, gfpflags); | |
722 | } | |
97e2bde4 | 723 | |
fbaccacf | 724 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 725 | { |
fbaccacf SR |
726 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
727 | } | |
1da177e4 | 728 | |
a737b3e2 AM |
729 | /* |
730 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
731 | */ | |
fbaccacf SR |
732 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
733 | size_t align, int flags, size_t *left_over, | |
734 | unsigned int *num) | |
735 | { | |
736 | int nr_objs; | |
737 | size_t mgmt_size; | |
738 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 739 | |
fbaccacf SR |
740 | /* |
741 | * The slab management structure can be either off the slab or | |
742 | * on it. For the latter case, the memory allocated for a | |
743 | * slab is used for: | |
744 | * | |
745 | * - The struct slab | |
746 | * - One kmem_bufctl_t for each object | |
747 | * - Padding to respect alignment of @align | |
748 | * - @buffer_size bytes for each object | |
749 | * | |
750 | * If the slab management structure is off the slab, then the | |
751 | * alignment will already be calculated into the size. Because | |
752 | * the slabs are all pages aligned, the objects will be at the | |
753 | * correct alignment when allocated. | |
754 | */ | |
755 | if (flags & CFLGS_OFF_SLAB) { | |
756 | mgmt_size = 0; | |
757 | nr_objs = slab_size / buffer_size; | |
758 | ||
759 | if (nr_objs > SLAB_LIMIT) | |
760 | nr_objs = SLAB_LIMIT; | |
761 | } else { | |
762 | /* | |
763 | * Ignore padding for the initial guess. The padding | |
764 | * is at most @align-1 bytes, and @buffer_size is at | |
765 | * least @align. In the worst case, this result will | |
766 | * be one greater than the number of objects that fit | |
767 | * into the memory allocation when taking the padding | |
768 | * into account. | |
769 | */ | |
770 | nr_objs = (slab_size - sizeof(struct slab)) / | |
771 | (buffer_size + sizeof(kmem_bufctl_t)); | |
772 | ||
773 | /* | |
774 | * This calculated number will be either the right | |
775 | * amount, or one greater than what we want. | |
776 | */ | |
777 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
778 | > slab_size) | |
779 | nr_objs--; | |
780 | ||
781 | if (nr_objs > SLAB_LIMIT) | |
782 | nr_objs = SLAB_LIMIT; | |
783 | ||
784 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
785 | } | |
786 | *num = nr_objs; | |
787 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
788 | } |
789 | ||
d40cee24 | 790 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 791 | |
a737b3e2 AM |
792 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
793 | char *msg) | |
1da177e4 LT |
794 | { |
795 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 796 | function, cachep->name, msg); |
1da177e4 LT |
797 | dump_stack(); |
798 | } | |
799 | ||
3395ee05 PM |
800 | /* |
801 | * By default on NUMA we use alien caches to stage the freeing of | |
802 | * objects allocated from other nodes. This causes massive memory | |
803 | * inefficiencies when using fake NUMA setup to split memory into a | |
804 | * large number of small nodes, so it can be disabled on the command | |
805 | * line | |
806 | */ | |
807 | ||
808 | static int use_alien_caches __read_mostly = 1; | |
809 | static int __init noaliencache_setup(char *s) | |
810 | { | |
811 | use_alien_caches = 0; | |
812 | return 1; | |
813 | } | |
814 | __setup("noaliencache", noaliencache_setup); | |
815 | ||
8fce4d8e CL |
816 | #ifdef CONFIG_NUMA |
817 | /* | |
818 | * Special reaping functions for NUMA systems called from cache_reap(). | |
819 | * These take care of doing round robin flushing of alien caches (containing | |
820 | * objects freed on different nodes from which they were allocated) and the | |
821 | * flushing of remote pcps by calling drain_node_pages. | |
822 | */ | |
1871e52c | 823 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
824 | |
825 | static void init_reap_node(int cpu) | |
826 | { | |
827 | int node; | |
828 | ||
7d6e6d09 | 829 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 830 | if (node == MAX_NUMNODES) |
442295c9 | 831 | node = first_node(node_online_map); |
8fce4d8e | 832 | |
1871e52c | 833 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
834 | } |
835 | ||
836 | static void next_reap_node(void) | |
837 | { | |
909ea964 | 838 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 839 | |
8fce4d8e CL |
840 | node = next_node(node, node_online_map); |
841 | if (unlikely(node >= MAX_NUMNODES)) | |
842 | node = first_node(node_online_map); | |
909ea964 | 843 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
844 | } |
845 | ||
846 | #else | |
847 | #define init_reap_node(cpu) do { } while (0) | |
848 | #define next_reap_node(void) do { } while (0) | |
849 | #endif | |
850 | ||
1da177e4 LT |
851 | /* |
852 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
853 | * via the workqueue/eventd. | |
854 | * Add the CPU number into the expiration time to minimize the possibility of | |
855 | * the CPUs getting into lockstep and contending for the global cache chain | |
856 | * lock. | |
857 | */ | |
897e679b | 858 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 859 | { |
1871e52c | 860 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
861 | |
862 | /* | |
863 | * When this gets called from do_initcalls via cpucache_init(), | |
864 | * init_workqueues() has already run, so keventd will be setup | |
865 | * at that time. | |
866 | */ | |
52bad64d | 867 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 868 | init_reap_node(cpu); |
78b43536 | 869 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); |
2b284214 AV |
870 | schedule_delayed_work_on(cpu, reap_work, |
871 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
872 | } |
873 | } | |
874 | ||
e498be7d | 875 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 876 | int batchcount, gfp_t gfp) |
1da177e4 | 877 | { |
b28a02de | 878 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
879 | struct array_cache *nc = NULL; |
880 | ||
83b519e8 | 881 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
882 | /* |
883 | * The array_cache structures contain pointers to free object. | |
25985edc | 884 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
885 | * cache the pointers are not cleared and they could be counted as |
886 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
887 | * not scan such objects. | |
888 | */ | |
889 | kmemleak_no_scan(nc); | |
1da177e4 LT |
890 | if (nc) { |
891 | nc->avail = 0; | |
892 | nc->limit = entries; | |
893 | nc->batchcount = batchcount; | |
894 | nc->touched = 0; | |
e498be7d | 895 | spin_lock_init(&nc->lock); |
1da177e4 LT |
896 | } |
897 | return nc; | |
898 | } | |
899 | ||
3ded175a CL |
900 | /* |
901 | * Transfer objects in one arraycache to another. | |
902 | * Locking must be handled by the caller. | |
903 | * | |
904 | * Return the number of entries transferred. | |
905 | */ | |
906 | static int transfer_objects(struct array_cache *to, | |
907 | struct array_cache *from, unsigned int max) | |
908 | { | |
909 | /* Figure out how many entries to transfer */ | |
732eacc0 | 910 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
911 | |
912 | if (!nr) | |
913 | return 0; | |
914 | ||
915 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
916 | sizeof(void *) *nr); | |
917 | ||
918 | from->avail -= nr; | |
919 | to->avail += nr; | |
3ded175a CL |
920 | return nr; |
921 | } | |
922 | ||
765c4507 CL |
923 | #ifndef CONFIG_NUMA |
924 | ||
925 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
926 | #define reap_alien(cachep, l3) do { } while (0) | |
927 | ||
83b519e8 | 928 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
929 | { |
930 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
931 | } | |
932 | ||
933 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
934 | { | |
935 | } | |
936 | ||
937 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
938 | { | |
939 | return 0; | |
940 | } | |
941 | ||
942 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
943 | gfp_t flags) | |
944 | { | |
945 | return NULL; | |
946 | } | |
947 | ||
8b98c169 | 948 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
949 | gfp_t flags, int nodeid) |
950 | { | |
951 | return NULL; | |
952 | } | |
953 | ||
954 | #else /* CONFIG_NUMA */ | |
955 | ||
8b98c169 | 956 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 957 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 958 | |
83b519e8 | 959 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
960 | { |
961 | struct array_cache **ac_ptr; | |
8ef82866 | 962 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
963 | int i; |
964 | ||
965 | if (limit > 1) | |
966 | limit = 12; | |
f3186a9c | 967 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
968 | if (ac_ptr) { |
969 | for_each_node(i) { | |
f3186a9c | 970 | if (i == node || !node_online(i)) |
e498be7d | 971 | continue; |
83b519e8 | 972 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 973 | if (!ac_ptr[i]) { |
cc550def | 974 | for (i--; i >= 0; i--) |
e498be7d CL |
975 | kfree(ac_ptr[i]); |
976 | kfree(ac_ptr); | |
977 | return NULL; | |
978 | } | |
979 | } | |
980 | } | |
981 | return ac_ptr; | |
982 | } | |
983 | ||
5295a74c | 984 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
985 | { |
986 | int i; | |
987 | ||
988 | if (!ac_ptr) | |
989 | return; | |
e498be7d | 990 | for_each_node(i) |
b28a02de | 991 | kfree(ac_ptr[i]); |
e498be7d CL |
992 | kfree(ac_ptr); |
993 | } | |
994 | ||
343e0d7a | 995 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 996 | struct array_cache *ac, int node) |
e498be7d CL |
997 | { |
998 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
999 | ||
1000 | if (ac->avail) { | |
1001 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1002 | /* |
1003 | * Stuff objects into the remote nodes shared array first. | |
1004 | * That way we could avoid the overhead of putting the objects | |
1005 | * into the free lists and getting them back later. | |
1006 | */ | |
693f7d36 | 1007 | if (rl3->shared) |
1008 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1009 | |
ff69416e | 1010 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1011 | ac->avail = 0; |
1012 | spin_unlock(&rl3->list_lock); | |
1013 | } | |
1014 | } | |
1015 | ||
8fce4d8e CL |
1016 | /* |
1017 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1018 | */ | |
1019 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1020 | { | |
909ea964 | 1021 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1022 | |
1023 | if (l3->alien) { | |
1024 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1025 | |
1026 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1027 | __drain_alien_cache(cachep, ac, node); |
1028 | spin_unlock_irq(&ac->lock); | |
1029 | } | |
1030 | } | |
1031 | } | |
1032 | ||
a737b3e2 AM |
1033 | static void drain_alien_cache(struct kmem_cache *cachep, |
1034 | struct array_cache **alien) | |
e498be7d | 1035 | { |
b28a02de | 1036 | int i = 0; |
e498be7d CL |
1037 | struct array_cache *ac; |
1038 | unsigned long flags; | |
1039 | ||
1040 | for_each_online_node(i) { | |
4484ebf1 | 1041 | ac = alien[i]; |
e498be7d CL |
1042 | if (ac) { |
1043 | spin_lock_irqsave(&ac->lock, flags); | |
1044 | __drain_alien_cache(cachep, ac, i); | |
1045 | spin_unlock_irqrestore(&ac->lock, flags); | |
1046 | } | |
1047 | } | |
1048 | } | |
729bd0b7 | 1049 | |
873623df | 1050 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1051 | { |
1052 | struct slab *slabp = virt_to_slab(objp); | |
1053 | int nodeid = slabp->nodeid; | |
1054 | struct kmem_list3 *l3; | |
1055 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1056 | int node; |
1057 | ||
7d6e6d09 | 1058 | node = numa_mem_id(); |
729bd0b7 PE |
1059 | |
1060 | /* | |
1061 | * Make sure we are not freeing a object from another node to the array | |
1062 | * cache on this cpu. | |
1063 | */ | |
62918a03 | 1064 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1065 | return 0; |
1066 | ||
1ca4cb24 | 1067 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1068 | STATS_INC_NODEFREES(cachep); |
1069 | if (l3->alien && l3->alien[nodeid]) { | |
1070 | alien = l3->alien[nodeid]; | |
873623df | 1071 | spin_lock(&alien->lock); |
729bd0b7 PE |
1072 | if (unlikely(alien->avail == alien->limit)) { |
1073 | STATS_INC_ACOVERFLOW(cachep); | |
1074 | __drain_alien_cache(cachep, alien, nodeid); | |
1075 | } | |
1076 | alien->entry[alien->avail++] = objp; | |
1077 | spin_unlock(&alien->lock); | |
1078 | } else { | |
1079 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1080 | free_block(cachep, &objp, 1, nodeid); | |
1081 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1082 | } | |
1083 | return 1; | |
1084 | } | |
e498be7d CL |
1085 | #endif |
1086 | ||
8f9f8d9e DR |
1087 | /* |
1088 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1089 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1090 | * will be allocated off-node since memory is not yet online for the new node. | |
1091 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1092 | * already in use. | |
1093 | * | |
1094 | * Must hold cache_chain_mutex. | |
1095 | */ | |
1096 | static int init_cache_nodelists_node(int node) | |
1097 | { | |
1098 | struct kmem_cache *cachep; | |
1099 | struct kmem_list3 *l3; | |
1100 | const int memsize = sizeof(struct kmem_list3); | |
1101 | ||
1102 | list_for_each_entry(cachep, &cache_chain, next) { | |
1103 | /* | |
1104 | * Set up the size64 kmemlist for cpu before we can | |
1105 | * begin anything. Make sure some other cpu on this | |
1106 | * node has not already allocated this | |
1107 | */ | |
1108 | if (!cachep->nodelists[node]) { | |
1109 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1110 | if (!l3) | |
1111 | return -ENOMEM; | |
1112 | kmem_list3_init(l3); | |
1113 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1114 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1115 | ||
1116 | /* | |
1117 | * The l3s don't come and go as CPUs come and | |
1118 | * go. cache_chain_mutex is sufficient | |
1119 | * protection here. | |
1120 | */ | |
1121 | cachep->nodelists[node] = l3; | |
1122 | } | |
1123 | ||
1124 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1125 | cachep->nodelists[node]->free_limit = | |
1126 | (1 + nr_cpus_node(node)) * | |
1127 | cachep->batchcount + cachep->num; | |
1128 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1129 | } | |
1130 | return 0; | |
1131 | } | |
1132 | ||
fbf1e473 AM |
1133 | static void __cpuinit cpuup_canceled(long cpu) |
1134 | { | |
1135 | struct kmem_cache *cachep; | |
1136 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1137 | int node = cpu_to_mem(cpu); |
a70f7302 | 1138 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 AM |
1139 | |
1140 | list_for_each_entry(cachep, &cache_chain, next) { | |
1141 | struct array_cache *nc; | |
1142 | struct array_cache *shared; | |
1143 | struct array_cache **alien; | |
fbf1e473 | 1144 | |
fbf1e473 AM |
1145 | /* cpu is dead; no one can alloc from it. */ |
1146 | nc = cachep->array[cpu]; | |
1147 | cachep->array[cpu] = NULL; | |
1148 | l3 = cachep->nodelists[node]; | |
1149 | ||
1150 | if (!l3) | |
1151 | goto free_array_cache; | |
1152 | ||
1153 | spin_lock_irq(&l3->list_lock); | |
1154 | ||
1155 | /* Free limit for this kmem_list3 */ | |
1156 | l3->free_limit -= cachep->batchcount; | |
1157 | if (nc) | |
1158 | free_block(cachep, nc->entry, nc->avail, node); | |
1159 | ||
58463c1f | 1160 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1161 | spin_unlock_irq(&l3->list_lock); |
1162 | goto free_array_cache; | |
1163 | } | |
1164 | ||
1165 | shared = l3->shared; | |
1166 | if (shared) { | |
1167 | free_block(cachep, shared->entry, | |
1168 | shared->avail, node); | |
1169 | l3->shared = NULL; | |
1170 | } | |
1171 | ||
1172 | alien = l3->alien; | |
1173 | l3->alien = NULL; | |
1174 | ||
1175 | spin_unlock_irq(&l3->list_lock); | |
1176 | ||
1177 | kfree(shared); | |
1178 | if (alien) { | |
1179 | drain_alien_cache(cachep, alien); | |
1180 | free_alien_cache(alien); | |
1181 | } | |
1182 | free_array_cache: | |
1183 | kfree(nc); | |
1184 | } | |
1185 | /* | |
1186 | * In the previous loop, all the objects were freed to | |
1187 | * the respective cache's slabs, now we can go ahead and | |
1188 | * shrink each nodelist to its limit. | |
1189 | */ | |
1190 | list_for_each_entry(cachep, &cache_chain, next) { | |
1191 | l3 = cachep->nodelists[node]; | |
1192 | if (!l3) | |
1193 | continue; | |
1194 | drain_freelist(cachep, l3, l3->free_objects); | |
1195 | } | |
1196 | } | |
1197 | ||
1198 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1199 | { |
343e0d7a | 1200 | struct kmem_cache *cachep; |
e498be7d | 1201 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1202 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1203 | int err; |
1da177e4 | 1204 | |
fbf1e473 AM |
1205 | /* |
1206 | * We need to do this right in the beginning since | |
1207 | * alloc_arraycache's are going to use this list. | |
1208 | * kmalloc_node allows us to add the slab to the right | |
1209 | * kmem_list3 and not this cpu's kmem_list3 | |
1210 | */ | |
8f9f8d9e DR |
1211 | err = init_cache_nodelists_node(node); |
1212 | if (err < 0) | |
1213 | goto bad; | |
fbf1e473 AM |
1214 | |
1215 | /* | |
1216 | * Now we can go ahead with allocating the shared arrays and | |
1217 | * array caches | |
1218 | */ | |
1219 | list_for_each_entry(cachep, &cache_chain, next) { | |
1220 | struct array_cache *nc; | |
1221 | struct array_cache *shared = NULL; | |
1222 | struct array_cache **alien = NULL; | |
1223 | ||
1224 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1225 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1226 | if (!nc) |
1227 | goto bad; | |
1228 | if (cachep->shared) { | |
1229 | shared = alloc_arraycache(node, | |
1230 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1231 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1232 | if (!shared) { |
1233 | kfree(nc); | |
1da177e4 | 1234 | goto bad; |
12d00f6a | 1235 | } |
fbf1e473 AM |
1236 | } |
1237 | if (use_alien_caches) { | |
83b519e8 | 1238 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1239 | if (!alien) { |
1240 | kfree(shared); | |
1241 | kfree(nc); | |
fbf1e473 | 1242 | goto bad; |
12d00f6a | 1243 | } |
fbf1e473 AM |
1244 | } |
1245 | cachep->array[cpu] = nc; | |
1246 | l3 = cachep->nodelists[node]; | |
1247 | BUG_ON(!l3); | |
1248 | ||
1249 | spin_lock_irq(&l3->list_lock); | |
1250 | if (!l3->shared) { | |
1251 | /* | |
1252 | * We are serialised from CPU_DEAD or | |
1253 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1254 | */ | |
1255 | l3->shared = shared; | |
1256 | shared = NULL; | |
1257 | } | |
4484ebf1 | 1258 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1259 | if (!l3->alien) { |
1260 | l3->alien = alien; | |
1261 | alien = NULL; | |
1da177e4 | 1262 | } |
fbf1e473 AM |
1263 | #endif |
1264 | spin_unlock_irq(&l3->list_lock); | |
1265 | kfree(shared); | |
1266 | free_alien_cache(alien); | |
1267 | } | |
ce79ddc8 PE |
1268 | init_node_lock_keys(node); |
1269 | ||
fbf1e473 AM |
1270 | return 0; |
1271 | bad: | |
12d00f6a | 1272 | cpuup_canceled(cpu); |
fbf1e473 AM |
1273 | return -ENOMEM; |
1274 | } | |
1275 | ||
1276 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1277 | unsigned long action, void *hcpu) | |
1278 | { | |
1279 | long cpu = (long)hcpu; | |
1280 | int err = 0; | |
1281 | ||
1282 | switch (action) { | |
fbf1e473 AM |
1283 | case CPU_UP_PREPARE: |
1284 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1285 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1286 | err = cpuup_prepare(cpu); |
95402b38 | 1287 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1288 | break; |
1289 | case CPU_ONLINE: | |
8bb78442 | 1290 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1291 | start_cpu_timer(cpu); |
1292 | break; | |
1293 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1294 | case CPU_DOWN_PREPARE: |
8bb78442 | 1295 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1296 | /* |
1297 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1298 | * held so that if cache_reap() is invoked it cannot do | |
1299 | * anything expensive but will only modify reap_work | |
1300 | * and reschedule the timer. | |
1301 | */ | |
afe2c511 | 1302 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1303 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1304 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1305 | break; |
1306 | case CPU_DOWN_FAILED: | |
8bb78442 | 1307 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1308 | start_cpu_timer(cpu); |
1309 | break; | |
1da177e4 | 1310 | case CPU_DEAD: |
8bb78442 | 1311 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1312 | /* |
1313 | * Even if all the cpus of a node are down, we don't free the | |
1314 | * kmem_list3 of any cache. This to avoid a race between | |
1315 | * cpu_down, and a kmalloc allocation from another cpu for | |
1316 | * memory from the node of the cpu going down. The list3 | |
1317 | * structure is usually allocated from kmem_cache_create() and | |
1318 | * gets destroyed at kmem_cache_destroy(). | |
1319 | */ | |
183ff22b | 1320 | /* fall through */ |
8f5be20b | 1321 | #endif |
1da177e4 | 1322 | case CPU_UP_CANCELED: |
8bb78442 | 1323 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1324 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1325 | cpuup_canceled(cpu); |
fc0abb14 | 1326 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1327 | break; |
1da177e4 | 1328 | } |
eac40680 | 1329 | return notifier_from_errno(err); |
1da177e4 LT |
1330 | } |
1331 | ||
74b85f37 CS |
1332 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1333 | &cpuup_callback, NULL, 0 | |
1334 | }; | |
1da177e4 | 1335 | |
8f9f8d9e DR |
1336 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1337 | /* | |
1338 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1339 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1340 | * removed. | |
1341 | * | |
1342 | * Must hold cache_chain_mutex. | |
1343 | */ | |
1344 | static int __meminit drain_cache_nodelists_node(int node) | |
1345 | { | |
1346 | struct kmem_cache *cachep; | |
1347 | int ret = 0; | |
1348 | ||
1349 | list_for_each_entry(cachep, &cache_chain, next) { | |
1350 | struct kmem_list3 *l3; | |
1351 | ||
1352 | l3 = cachep->nodelists[node]; | |
1353 | if (!l3) | |
1354 | continue; | |
1355 | ||
1356 | drain_freelist(cachep, l3, l3->free_objects); | |
1357 | ||
1358 | if (!list_empty(&l3->slabs_full) || | |
1359 | !list_empty(&l3->slabs_partial)) { | |
1360 | ret = -EBUSY; | |
1361 | break; | |
1362 | } | |
1363 | } | |
1364 | return ret; | |
1365 | } | |
1366 | ||
1367 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1368 | unsigned long action, void *arg) | |
1369 | { | |
1370 | struct memory_notify *mnb = arg; | |
1371 | int ret = 0; | |
1372 | int nid; | |
1373 | ||
1374 | nid = mnb->status_change_nid; | |
1375 | if (nid < 0) | |
1376 | goto out; | |
1377 | ||
1378 | switch (action) { | |
1379 | case MEM_GOING_ONLINE: | |
1380 | mutex_lock(&cache_chain_mutex); | |
1381 | ret = init_cache_nodelists_node(nid); | |
1382 | mutex_unlock(&cache_chain_mutex); | |
1383 | break; | |
1384 | case MEM_GOING_OFFLINE: | |
1385 | mutex_lock(&cache_chain_mutex); | |
1386 | ret = drain_cache_nodelists_node(nid); | |
1387 | mutex_unlock(&cache_chain_mutex); | |
1388 | break; | |
1389 | case MEM_ONLINE: | |
1390 | case MEM_OFFLINE: | |
1391 | case MEM_CANCEL_ONLINE: | |
1392 | case MEM_CANCEL_OFFLINE: | |
1393 | break; | |
1394 | } | |
1395 | out: | |
5fda1bd5 | 1396 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1397 | } |
1398 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1399 | ||
e498be7d CL |
1400 | /* |
1401 | * swap the static kmem_list3 with kmalloced memory | |
1402 | */ | |
8f9f8d9e DR |
1403 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1404 | int nodeid) | |
e498be7d CL |
1405 | { |
1406 | struct kmem_list3 *ptr; | |
1407 | ||
83b519e8 | 1408 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1409 | BUG_ON(!ptr); |
1410 | ||
e498be7d | 1411 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1412 | /* |
1413 | * Do not assume that spinlocks can be initialized via memcpy: | |
1414 | */ | |
1415 | spin_lock_init(&ptr->list_lock); | |
1416 | ||
e498be7d CL |
1417 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1418 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1419 | } |
1420 | ||
556a169d PE |
1421 | /* |
1422 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1423 | * size of kmem_list3. | |
1424 | */ | |
1425 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1426 | { | |
1427 | int node; | |
1428 | ||
1429 | for_each_online_node(node) { | |
1430 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1431 | cachep->nodelists[node]->next_reap = jiffies + | |
1432 | REAPTIMEOUT_LIST3 + | |
1433 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1434 | } | |
1435 | } | |
1436 | ||
a737b3e2 AM |
1437 | /* |
1438 | * Initialisation. Called after the page allocator have been initialised and | |
1439 | * before smp_init(). | |
1da177e4 LT |
1440 | */ |
1441 | void __init kmem_cache_init(void) | |
1442 | { | |
1443 | size_t left_over; | |
1444 | struct cache_sizes *sizes; | |
1445 | struct cache_names *names; | |
e498be7d | 1446 | int i; |
07ed76b2 | 1447 | int order; |
1ca4cb24 | 1448 | int node; |
e498be7d | 1449 | |
b6e68bc1 | 1450 | if (num_possible_nodes() == 1) |
62918a03 SS |
1451 | use_alien_caches = 0; |
1452 | ||
e498be7d CL |
1453 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1454 | kmem_list3_init(&initkmem_list3[i]); | |
1455 | if (i < MAX_NUMNODES) | |
1456 | cache_cache.nodelists[i] = NULL; | |
1457 | } | |
556a169d | 1458 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1459 | |
1460 | /* | |
1461 | * Fragmentation resistance on low memory - only use bigger | |
1462 | * page orders on machines with more than 32MB of memory. | |
1463 | */ | |
4481374c | 1464 | if (totalram_pages > (32 << 20) >> PAGE_SHIFT) |
1da177e4 LT |
1465 | slab_break_gfp_order = BREAK_GFP_ORDER_HI; |
1466 | ||
1da177e4 LT |
1467 | /* Bootstrap is tricky, because several objects are allocated |
1468 | * from caches that do not exist yet: | |
a737b3e2 AM |
1469 | * 1) initialize the cache_cache cache: it contains the struct |
1470 | * kmem_cache structures of all caches, except cache_cache itself: | |
1471 | * cache_cache is statically allocated. | |
e498be7d CL |
1472 | * Initially an __init data area is used for the head array and the |
1473 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1474 | * array at the end of the bootstrap. | |
1da177e4 | 1475 | * 2) Create the first kmalloc cache. |
343e0d7a | 1476 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1477 | * An __init data area is used for the head array. |
1478 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1479 | * head arrays. | |
1da177e4 LT |
1480 | * 4) Replace the __init data head arrays for cache_cache and the first |
1481 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1482 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1483 | * the other cache's with kmalloc allocated memory. | |
1484 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1485 | */ |
1486 | ||
7d6e6d09 | 1487 | node = numa_mem_id(); |
1ca4cb24 | 1488 | |
1da177e4 | 1489 | /* 1) create the cache_cache */ |
1da177e4 LT |
1490 | INIT_LIST_HEAD(&cache_chain); |
1491 | list_add(&cache_cache.next, &cache_chain); | |
1492 | cache_cache.colour_off = cache_line_size(); | |
1493 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1494 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1495 | |
8da3430d | 1496 | /* |
b56efcf0 | 1497 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1498 | */ |
b56efcf0 ED |
1499 | cache_cache.buffer_size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
1500 | nr_node_ids * sizeof(struct kmem_list3 *); | |
8da3430d ED |
1501 | #if DEBUG |
1502 | cache_cache.obj_size = cache_cache.buffer_size; | |
1503 | #endif | |
a737b3e2 AM |
1504 | cache_cache.buffer_size = ALIGN(cache_cache.buffer_size, |
1505 | cache_line_size()); | |
6a2d7a95 ED |
1506 | cache_cache.reciprocal_buffer_size = |
1507 | reciprocal_value(cache_cache.buffer_size); | |
1da177e4 | 1508 | |
07ed76b2 JS |
1509 | for (order = 0; order < MAX_ORDER; order++) { |
1510 | cache_estimate(order, cache_cache.buffer_size, | |
1511 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1512 | if (cache_cache.num) | |
1513 | break; | |
1514 | } | |
40094fa6 | 1515 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1516 | cache_cache.gfporder = order; |
b28a02de | 1517 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1518 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1519 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1520 | |
1521 | /* 2+3) create the kmalloc caches */ | |
1522 | sizes = malloc_sizes; | |
1523 | names = cache_names; | |
1524 | ||
a737b3e2 AM |
1525 | /* |
1526 | * Initialize the caches that provide memory for the array cache and the | |
1527 | * kmem_list3 structures first. Without this, further allocations will | |
1528 | * bug. | |
e498be7d CL |
1529 | */ |
1530 | ||
1531 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1532 | sizes[INDEX_AC].cs_size, |
1533 | ARCH_KMALLOC_MINALIGN, | |
1534 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1535 | NULL); |
e498be7d | 1536 | |
a737b3e2 | 1537 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1538 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1539 | kmem_cache_create(names[INDEX_L3].name, |
1540 | sizes[INDEX_L3].cs_size, | |
1541 | ARCH_KMALLOC_MINALIGN, | |
1542 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1543 | NULL); |
a737b3e2 | 1544 | } |
e498be7d | 1545 | |
e0a42726 IM |
1546 | slab_early_init = 0; |
1547 | ||
1da177e4 | 1548 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1549 | /* |
1550 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1551 | * This should be particularly beneficial on SMP boxes, as it |
1552 | * eliminates "false sharing". | |
1553 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1554 | * allow tighter packing of the smaller caches. |
1555 | */ | |
a737b3e2 | 1556 | if (!sizes->cs_cachep) { |
e498be7d | 1557 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1558 | sizes->cs_size, |
1559 | ARCH_KMALLOC_MINALIGN, | |
1560 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1561 | NULL); |
a737b3e2 | 1562 | } |
4b51d669 CL |
1563 | #ifdef CONFIG_ZONE_DMA |
1564 | sizes->cs_dmacachep = kmem_cache_create( | |
1565 | names->name_dma, | |
a737b3e2 AM |
1566 | sizes->cs_size, |
1567 | ARCH_KMALLOC_MINALIGN, | |
1568 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1569 | SLAB_PANIC, | |
20c2df83 | 1570 | NULL); |
4b51d669 | 1571 | #endif |
1da177e4 LT |
1572 | sizes++; |
1573 | names++; | |
1574 | } | |
1575 | /* 4) Replace the bootstrap head arrays */ | |
1576 | { | |
2b2d5493 | 1577 | struct array_cache *ptr; |
e498be7d | 1578 | |
83b519e8 | 1579 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1580 | |
9a2dba4b PE |
1581 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1582 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1583 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1584 | /* |
1585 | * Do not assume that spinlocks can be initialized via memcpy: | |
1586 | */ | |
1587 | spin_lock_init(&ptr->lock); | |
1588 | ||
1da177e4 | 1589 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1590 | |
83b519e8 | 1591 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1592 | |
9a2dba4b | 1593 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1594 | != &initarray_generic.cache); |
9a2dba4b | 1595 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1596 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1597 | /* |
1598 | * Do not assume that spinlocks can be initialized via memcpy: | |
1599 | */ | |
1600 | spin_lock_init(&ptr->lock); | |
1601 | ||
e498be7d | 1602 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1603 | ptr; |
1da177e4 | 1604 | } |
e498be7d CL |
1605 | /* 5) Replace the bootstrap kmem_list3's */ |
1606 | { | |
1ca4cb24 PE |
1607 | int nid; |
1608 | ||
9c09a95c | 1609 | for_each_online_node(nid) { |
ec1f5eee | 1610 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1611 | |
e498be7d | 1612 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1613 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1614 | |
1615 | if (INDEX_AC != INDEX_L3) { | |
1616 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1617 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1618 | } |
1619 | } | |
1620 | } | |
1da177e4 | 1621 | |
8429db5c | 1622 | g_cpucache_up = EARLY; |
8429db5c PE |
1623 | } |
1624 | ||
1625 | void __init kmem_cache_init_late(void) | |
1626 | { | |
1627 | struct kmem_cache *cachep; | |
1628 | ||
8429db5c PE |
1629 | /* 6) resize the head arrays to their final sizes */ |
1630 | mutex_lock(&cache_chain_mutex); | |
1631 | list_for_each_entry(cachep, &cache_chain, next) | |
1632 | if (enable_cpucache(cachep, GFP_NOWAIT)) | |
1633 | BUG(); | |
1634 | mutex_unlock(&cache_chain_mutex); | |
056c6241 | 1635 | |
1da177e4 LT |
1636 | /* Done! */ |
1637 | g_cpucache_up = FULL; | |
1638 | ||
ec5a36f9 PE |
1639 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1640 | init_lock_keys(); | |
1641 | ||
a737b3e2 AM |
1642 | /* |
1643 | * Register a cpu startup notifier callback that initializes | |
1644 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1645 | */ |
1646 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1647 | |
8f9f8d9e DR |
1648 | #ifdef CONFIG_NUMA |
1649 | /* | |
1650 | * Register a memory hotplug callback that initializes and frees | |
1651 | * nodelists. | |
1652 | */ | |
1653 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1654 | #endif | |
1655 | ||
a737b3e2 AM |
1656 | /* |
1657 | * The reap timers are started later, with a module init call: That part | |
1658 | * of the kernel is not yet operational. | |
1da177e4 LT |
1659 | */ |
1660 | } | |
1661 | ||
1662 | static int __init cpucache_init(void) | |
1663 | { | |
1664 | int cpu; | |
1665 | ||
a737b3e2 AM |
1666 | /* |
1667 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1668 | */ |
e498be7d | 1669 | for_each_online_cpu(cpu) |
a737b3e2 | 1670 | start_cpu_timer(cpu); |
1da177e4 LT |
1671 | return 0; |
1672 | } | |
1da177e4 LT |
1673 | __initcall(cpucache_init); |
1674 | ||
1675 | /* | |
1676 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1677 | * | |
1678 | * If we requested dmaable memory, we will get it. Even if we | |
1679 | * did not request dmaable memory, we might get it, but that | |
1680 | * would be relatively rare and ignorable. | |
1681 | */ | |
343e0d7a | 1682 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1683 | { |
1684 | struct page *page; | |
e1b6aa6f | 1685 | int nr_pages; |
1da177e4 LT |
1686 | int i; |
1687 | ||
d6fef9da | 1688 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1689 | /* |
1690 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1691 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1692 | */ |
e1b6aa6f | 1693 | flags |= __GFP_COMP; |
d6fef9da | 1694 | #endif |
765c4507 | 1695 | |
3c517a61 | 1696 | flags |= cachep->gfpflags; |
e12ba74d MG |
1697 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1698 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1699 | |
517d0869 | 1700 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
1da177e4 LT |
1701 | if (!page) |
1702 | return NULL; | |
1da177e4 | 1703 | |
e1b6aa6f | 1704 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1705 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1706 | add_zone_page_state(page_zone(page), |
1707 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1708 | else | |
1709 | add_zone_page_state(page_zone(page), | |
1710 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1711 | for (i = 0; i < nr_pages; i++) |
1712 | __SetPageSlab(page + i); | |
c175eea4 | 1713 | |
b1eeab67 VN |
1714 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1715 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1716 | ||
1717 | if (cachep->ctor) | |
1718 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1719 | else | |
1720 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1721 | } | |
c175eea4 | 1722 | |
e1b6aa6f | 1723 | return page_address(page); |
1da177e4 LT |
1724 | } |
1725 | ||
1726 | /* | |
1727 | * Interface to system's page release. | |
1728 | */ | |
343e0d7a | 1729 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1730 | { |
b28a02de | 1731 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1732 | struct page *page = virt_to_page(addr); |
1733 | const unsigned long nr_freed = i; | |
1734 | ||
b1eeab67 | 1735 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1736 | |
972d1a7b CL |
1737 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1738 | sub_zone_page_state(page_zone(page), | |
1739 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1740 | else | |
1741 | sub_zone_page_state(page_zone(page), | |
1742 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1743 | while (i--) { |
f205b2fe NP |
1744 | BUG_ON(!PageSlab(page)); |
1745 | __ClearPageSlab(page); | |
1da177e4 LT |
1746 | page++; |
1747 | } | |
1da177e4 LT |
1748 | if (current->reclaim_state) |
1749 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1750 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1751 | } |
1752 | ||
1753 | static void kmem_rcu_free(struct rcu_head *head) | |
1754 | { | |
b28a02de | 1755 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1756 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1757 | |
1758 | kmem_freepages(cachep, slab_rcu->addr); | |
1759 | if (OFF_SLAB(cachep)) | |
1760 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1761 | } | |
1762 | ||
1763 | #if DEBUG | |
1764 | ||
1765 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1766 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1767 | unsigned long caller) |
1da177e4 | 1768 | { |
3dafccf2 | 1769 | int size = obj_size(cachep); |
1da177e4 | 1770 | |
3dafccf2 | 1771 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1772 | |
b28a02de | 1773 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1774 | return; |
1775 | ||
b28a02de PE |
1776 | *addr++ = 0x12345678; |
1777 | *addr++ = caller; | |
1778 | *addr++ = smp_processor_id(); | |
1779 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1780 | { |
1781 | unsigned long *sptr = &caller; | |
1782 | unsigned long svalue; | |
1783 | ||
1784 | while (!kstack_end(sptr)) { | |
1785 | svalue = *sptr++; | |
1786 | if (kernel_text_address(svalue)) { | |
b28a02de | 1787 | *addr++ = svalue; |
1da177e4 LT |
1788 | size -= sizeof(unsigned long); |
1789 | if (size <= sizeof(unsigned long)) | |
1790 | break; | |
1791 | } | |
1792 | } | |
1793 | ||
1794 | } | |
b28a02de | 1795 | *addr++ = 0x87654321; |
1da177e4 LT |
1796 | } |
1797 | #endif | |
1798 | ||
343e0d7a | 1799 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1800 | { |
3dafccf2 MS |
1801 | int size = obj_size(cachep); |
1802 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1803 | |
1804 | memset(addr, val, size); | |
b28a02de | 1805 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1806 | } |
1807 | ||
1808 | static void dump_line(char *data, int offset, int limit) | |
1809 | { | |
1810 | int i; | |
aa83aa40 DJ |
1811 | unsigned char error = 0; |
1812 | int bad_count = 0; | |
1813 | ||
1da177e4 | 1814 | printk(KERN_ERR "%03x:", offset); |
aa83aa40 DJ |
1815 | for (i = 0; i < limit; i++) { |
1816 | if (data[offset + i] != POISON_FREE) { | |
1817 | error = data[offset + i]; | |
1818 | bad_count++; | |
1819 | } | |
b28a02de | 1820 | printk(" %02x", (unsigned char)data[offset + i]); |
aa83aa40 | 1821 | } |
1da177e4 | 1822 | printk("\n"); |
aa83aa40 DJ |
1823 | |
1824 | if (bad_count == 1) { | |
1825 | error ^= POISON_FREE; | |
1826 | if (!(error & (error - 1))) { | |
1827 | printk(KERN_ERR "Single bit error detected. Probably " | |
1828 | "bad RAM.\n"); | |
1829 | #ifdef CONFIG_X86 | |
1830 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1831 | "test tool.\n"); | |
1832 | #else | |
1833 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1834 | #endif | |
1835 | } | |
1836 | } | |
1da177e4 LT |
1837 | } |
1838 | #endif | |
1839 | ||
1840 | #if DEBUG | |
1841 | ||
343e0d7a | 1842 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1843 | { |
1844 | int i, size; | |
1845 | char *realobj; | |
1846 | ||
1847 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1848 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1849 | *dbg_redzone1(cachep, objp), |
1850 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1851 | } |
1852 | ||
1853 | if (cachep->flags & SLAB_STORE_USER) { | |
1854 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1855 | *dbg_userword(cachep, objp)); |
1da177e4 | 1856 | print_symbol("(%s)", |
a737b3e2 | 1857 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1858 | printk("\n"); |
1859 | } | |
3dafccf2 MS |
1860 | realobj = (char *)objp + obj_offset(cachep); |
1861 | size = obj_size(cachep); | |
b28a02de | 1862 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1863 | int limit; |
1864 | limit = 16; | |
b28a02de PE |
1865 | if (i + limit > size) |
1866 | limit = size - i; | |
1da177e4 LT |
1867 | dump_line(realobj, i, limit); |
1868 | } | |
1869 | } | |
1870 | ||
343e0d7a | 1871 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1872 | { |
1873 | char *realobj; | |
1874 | int size, i; | |
1875 | int lines = 0; | |
1876 | ||
3dafccf2 MS |
1877 | realobj = (char *)objp + obj_offset(cachep); |
1878 | size = obj_size(cachep); | |
1da177e4 | 1879 | |
b28a02de | 1880 | for (i = 0; i < size; i++) { |
1da177e4 | 1881 | char exp = POISON_FREE; |
b28a02de | 1882 | if (i == size - 1) |
1da177e4 LT |
1883 | exp = POISON_END; |
1884 | if (realobj[i] != exp) { | |
1885 | int limit; | |
1886 | /* Mismatch ! */ | |
1887 | /* Print header */ | |
1888 | if (lines == 0) { | |
b28a02de | 1889 | printk(KERN_ERR |
e94a40c5 DH |
1890 | "Slab corruption: %s start=%p, len=%d\n", |
1891 | cachep->name, realobj, size); | |
1da177e4 LT |
1892 | print_objinfo(cachep, objp, 0); |
1893 | } | |
1894 | /* Hexdump the affected line */ | |
b28a02de | 1895 | i = (i / 16) * 16; |
1da177e4 | 1896 | limit = 16; |
b28a02de PE |
1897 | if (i + limit > size) |
1898 | limit = size - i; | |
1da177e4 LT |
1899 | dump_line(realobj, i, limit); |
1900 | i += 16; | |
1901 | lines++; | |
1902 | /* Limit to 5 lines */ | |
1903 | if (lines > 5) | |
1904 | break; | |
1905 | } | |
1906 | } | |
1907 | if (lines != 0) { | |
1908 | /* Print some data about the neighboring objects, if they | |
1909 | * exist: | |
1910 | */ | |
6ed5eb22 | 1911 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1912 | unsigned int objnr; |
1da177e4 | 1913 | |
8fea4e96 | 1914 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1915 | if (objnr) { |
8fea4e96 | 1916 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1917 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1918 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1919 | realobj, size); |
1da177e4 LT |
1920 | print_objinfo(cachep, objp, 2); |
1921 | } | |
b28a02de | 1922 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1923 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1924 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1925 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1926 | realobj, size); |
1da177e4 LT |
1927 | print_objinfo(cachep, objp, 2); |
1928 | } | |
1929 | } | |
1930 | } | |
1931 | #endif | |
1932 | ||
12dd36fa | 1933 | #if DEBUG |
e79aec29 | 1934 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1935 | { |
1da177e4 LT |
1936 | int i; |
1937 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 1938 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
1939 | |
1940 | if (cachep->flags & SLAB_POISON) { | |
1941 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 AM |
1942 | if (cachep->buffer_size % PAGE_SIZE == 0 && |
1943 | OFF_SLAB(cachep)) | |
b28a02de | 1944 | kernel_map_pages(virt_to_page(objp), |
a737b3e2 | 1945 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
1946 | else |
1947 | check_poison_obj(cachep, objp); | |
1948 | #else | |
1949 | check_poison_obj(cachep, objp); | |
1950 | #endif | |
1951 | } | |
1952 | if (cachep->flags & SLAB_RED_ZONE) { | |
1953 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
1954 | slab_error(cachep, "start of a freed object " | |
b28a02de | 1955 | "was overwritten"); |
1da177e4 LT |
1956 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1957 | slab_error(cachep, "end of a freed object " | |
b28a02de | 1958 | "was overwritten"); |
1da177e4 | 1959 | } |
1da177e4 | 1960 | } |
12dd36fa | 1961 | } |
1da177e4 | 1962 | #else |
e79aec29 | 1963 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 1964 | { |
12dd36fa | 1965 | } |
1da177e4 LT |
1966 | #endif |
1967 | ||
911851e6 RD |
1968 | /** |
1969 | * slab_destroy - destroy and release all objects in a slab | |
1970 | * @cachep: cache pointer being destroyed | |
1971 | * @slabp: slab pointer being destroyed | |
1972 | * | |
12dd36fa | 1973 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
1974 | * Before calling the slab must have been unlinked from the cache. The |
1975 | * cache-lock is not held/needed. | |
12dd36fa | 1976 | */ |
343e0d7a | 1977 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
1978 | { |
1979 | void *addr = slabp->s_mem - slabp->colouroff; | |
1980 | ||
e79aec29 | 1981 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
1982 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
1983 | struct slab_rcu *slab_rcu; | |
1984 | ||
b28a02de | 1985 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
1986 | slab_rcu->cachep = cachep; |
1987 | slab_rcu->addr = addr; | |
1988 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
1989 | } else { | |
1990 | kmem_freepages(cachep, addr); | |
873623df IM |
1991 | if (OFF_SLAB(cachep)) |
1992 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
1993 | } |
1994 | } | |
1995 | ||
117f6eb1 CL |
1996 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
1997 | { | |
1998 | int i; | |
1999 | struct kmem_list3 *l3; | |
2000 | ||
2001 | for_each_online_cpu(i) | |
2002 | kfree(cachep->array[i]); | |
2003 | ||
2004 | /* NUMA: free the list3 structures */ | |
2005 | for_each_online_node(i) { | |
2006 | l3 = cachep->nodelists[i]; | |
2007 | if (l3) { | |
2008 | kfree(l3->shared); | |
2009 | free_alien_cache(l3->alien); | |
2010 | kfree(l3); | |
2011 | } | |
2012 | } | |
2013 | kmem_cache_free(&cache_cache, cachep); | |
2014 | } | |
2015 | ||
2016 | ||
4d268eba | 2017 | /** |
a70773dd RD |
2018 | * calculate_slab_order - calculate size (page order) of slabs |
2019 | * @cachep: pointer to the cache that is being created | |
2020 | * @size: size of objects to be created in this cache. | |
2021 | * @align: required alignment for the objects. | |
2022 | * @flags: slab allocation flags | |
2023 | * | |
2024 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2025 | * |
2026 | * This could be made much more intelligent. For now, try to avoid using | |
2027 | * high order pages for slabs. When the gfp() functions are more friendly | |
2028 | * towards high-order requests, this should be changed. | |
2029 | */ | |
a737b3e2 | 2030 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2031 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2032 | { |
b1ab41c4 | 2033 | unsigned long offslab_limit; |
4d268eba | 2034 | size_t left_over = 0; |
9888e6fa | 2035 | int gfporder; |
4d268eba | 2036 | |
0aa817f0 | 2037 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2038 | unsigned int num; |
2039 | size_t remainder; | |
2040 | ||
9888e6fa | 2041 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2042 | if (!num) |
2043 | continue; | |
9888e6fa | 2044 | |
b1ab41c4 IM |
2045 | if (flags & CFLGS_OFF_SLAB) { |
2046 | /* | |
2047 | * Max number of objs-per-slab for caches which | |
2048 | * use off-slab slabs. Needed to avoid a possible | |
2049 | * looping condition in cache_grow(). | |
2050 | */ | |
2051 | offslab_limit = size - sizeof(struct slab); | |
2052 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2053 | ||
2054 | if (num > offslab_limit) | |
2055 | break; | |
2056 | } | |
4d268eba | 2057 | |
9888e6fa | 2058 | /* Found something acceptable - save it away */ |
4d268eba | 2059 | cachep->num = num; |
9888e6fa | 2060 | cachep->gfporder = gfporder; |
4d268eba PE |
2061 | left_over = remainder; |
2062 | ||
f78bb8ad LT |
2063 | /* |
2064 | * A VFS-reclaimable slab tends to have most allocations | |
2065 | * as GFP_NOFS and we really don't want to have to be allocating | |
2066 | * higher-order pages when we are unable to shrink dcache. | |
2067 | */ | |
2068 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2069 | break; | |
2070 | ||
4d268eba PE |
2071 | /* |
2072 | * Large number of objects is good, but very large slabs are | |
2073 | * currently bad for the gfp()s. | |
2074 | */ | |
9888e6fa | 2075 | if (gfporder >= slab_break_gfp_order) |
4d268eba PE |
2076 | break; |
2077 | ||
9888e6fa LT |
2078 | /* |
2079 | * Acceptable internal fragmentation? | |
2080 | */ | |
a737b3e2 | 2081 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2082 | break; |
2083 | } | |
2084 | return left_over; | |
2085 | } | |
2086 | ||
83b519e8 | 2087 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2088 | { |
2ed3a4ef | 2089 | if (g_cpucache_up == FULL) |
83b519e8 | 2090 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2091 | |
f30cf7d1 PE |
2092 | if (g_cpucache_up == NONE) { |
2093 | /* | |
2094 | * Note: the first kmem_cache_create must create the cache | |
2095 | * that's used by kmalloc(24), otherwise the creation of | |
2096 | * further caches will BUG(). | |
2097 | */ | |
2098 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2099 | ||
2100 | /* | |
2101 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2102 | * the first cache, then we need to set up all its list3s, | |
2103 | * otherwise the creation of further caches will BUG(). | |
2104 | */ | |
2105 | set_up_list3s(cachep, SIZE_AC); | |
2106 | if (INDEX_AC == INDEX_L3) | |
2107 | g_cpucache_up = PARTIAL_L3; | |
2108 | else | |
2109 | g_cpucache_up = PARTIAL_AC; | |
2110 | } else { | |
2111 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2112 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 PE |
2113 | |
2114 | if (g_cpucache_up == PARTIAL_AC) { | |
2115 | set_up_list3s(cachep, SIZE_L3); | |
2116 | g_cpucache_up = PARTIAL_L3; | |
2117 | } else { | |
2118 | int node; | |
556a169d | 2119 | for_each_online_node(node) { |
f30cf7d1 PE |
2120 | cachep->nodelists[node] = |
2121 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2122 | gfp, node); |
f30cf7d1 PE |
2123 | BUG_ON(!cachep->nodelists[node]); |
2124 | kmem_list3_init(cachep->nodelists[node]); | |
2125 | } | |
2126 | } | |
2127 | } | |
7d6e6d09 | 2128 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2129 | jiffies + REAPTIMEOUT_LIST3 + |
2130 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2131 | ||
2132 | cpu_cache_get(cachep)->avail = 0; | |
2133 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2134 | cpu_cache_get(cachep)->batchcount = 1; | |
2135 | cpu_cache_get(cachep)->touched = 0; | |
2136 | cachep->batchcount = 1; | |
2137 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2138 | return 0; |
f30cf7d1 PE |
2139 | } |
2140 | ||
1da177e4 LT |
2141 | /** |
2142 | * kmem_cache_create - Create a cache. | |
2143 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2144 | * @size: The size of objects to be created in this cache. | |
2145 | * @align: The required alignment for the objects. | |
2146 | * @flags: SLAB flags | |
2147 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2148 | * |
2149 | * Returns a ptr to the cache on success, NULL on failure. | |
2150 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2151 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2152 | * |
2153 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2154 | * the module calling this has to destroy the cache before getting unloaded. |
2155 | * | |
1da177e4 LT |
2156 | * The flags are |
2157 | * | |
2158 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2159 | * to catch references to uninitialised memory. | |
2160 | * | |
2161 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2162 | * for buffer overruns. | |
2163 | * | |
1da177e4 LT |
2164 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2165 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2166 | * as davem. | |
2167 | */ | |
343e0d7a | 2168 | struct kmem_cache * |
1da177e4 | 2169 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2170 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2171 | { |
2172 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2173 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2174 | gfp_t gfp; |
1da177e4 LT |
2175 | |
2176 | /* | |
2177 | * Sanity checks... these are all serious usage bugs. | |
2178 | */ | |
a737b3e2 | 2179 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2180 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2181 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2182 | name); |
b28a02de PE |
2183 | BUG(); |
2184 | } | |
1da177e4 | 2185 | |
f0188f47 | 2186 | /* |
8f5be20b | 2187 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2188 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2189 | */ |
83b519e8 PE |
2190 | if (slab_is_available()) { |
2191 | get_online_cpus(); | |
2192 | mutex_lock(&cache_chain_mutex); | |
2193 | } | |
4f12bb4f | 2194 | |
7a7c381d | 2195 | list_for_each_entry(pc, &cache_chain, next) { |
4f12bb4f AM |
2196 | char tmp; |
2197 | int res; | |
2198 | ||
2199 | /* | |
2200 | * This happens when the module gets unloaded and doesn't | |
2201 | * destroy its slab cache and no-one else reuses the vmalloc | |
2202 | * area of the module. Print a warning. | |
2203 | */ | |
138ae663 | 2204 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2205 | if (res) { |
b4169525 | 2206 | printk(KERN_ERR |
2207 | "SLAB: cache with size %d has lost its name\n", | |
3dafccf2 | 2208 | pc->buffer_size); |
4f12bb4f AM |
2209 | continue; |
2210 | } | |
2211 | ||
b28a02de | 2212 | if (!strcmp(pc->name, name)) { |
b4169525 | 2213 | printk(KERN_ERR |
2214 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2215 | dump_stack(); |
2216 | goto oops; | |
2217 | } | |
2218 | } | |
2219 | ||
1da177e4 LT |
2220 | #if DEBUG |
2221 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2222 | #if FORCED_DEBUG |
2223 | /* | |
2224 | * Enable redzoning and last user accounting, except for caches with | |
2225 | * large objects, if the increased size would increase the object size | |
2226 | * above the next power of two: caches with object sizes just above a | |
2227 | * power of two have a significant amount of internal fragmentation. | |
2228 | */ | |
87a927c7 DW |
2229 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2230 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2231 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2232 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2233 | flags |= SLAB_POISON; | |
2234 | #endif | |
2235 | if (flags & SLAB_DESTROY_BY_RCU) | |
2236 | BUG_ON(flags & SLAB_POISON); | |
2237 | #endif | |
1da177e4 | 2238 | /* |
a737b3e2 AM |
2239 | * Always checks flags, a caller might be expecting debug support which |
2240 | * isn't available. | |
1da177e4 | 2241 | */ |
40094fa6 | 2242 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2243 | |
a737b3e2 AM |
2244 | /* |
2245 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2246 | * unaligned accesses for some archs when redzoning is used, and makes |
2247 | * sure any on-slab bufctl's are also correctly aligned. | |
2248 | */ | |
b28a02de PE |
2249 | if (size & (BYTES_PER_WORD - 1)) { |
2250 | size += (BYTES_PER_WORD - 1); | |
2251 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2252 | } |
2253 | ||
a737b3e2 AM |
2254 | /* calculate the final buffer alignment: */ |
2255 | ||
1da177e4 LT |
2256 | /* 1) arch recommendation: can be overridden for debug */ |
2257 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2258 | /* |
2259 | * Default alignment: as specified by the arch code. Except if | |
2260 | * an object is really small, then squeeze multiple objects into | |
2261 | * one cacheline. | |
1da177e4 LT |
2262 | */ |
2263 | ralign = cache_line_size(); | |
b28a02de | 2264 | while (size <= ralign / 2) |
1da177e4 LT |
2265 | ralign /= 2; |
2266 | } else { | |
2267 | ralign = BYTES_PER_WORD; | |
2268 | } | |
ca5f9703 PE |
2269 | |
2270 | /* | |
87a927c7 DW |
2271 | * Redzoning and user store require word alignment or possibly larger. |
2272 | * Note this will be overridden by architecture or caller mandated | |
2273 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2274 | */ |
87a927c7 DW |
2275 | if (flags & SLAB_STORE_USER) |
2276 | ralign = BYTES_PER_WORD; | |
2277 | ||
2278 | if (flags & SLAB_RED_ZONE) { | |
2279 | ralign = REDZONE_ALIGN; | |
2280 | /* If redzoning, ensure that the second redzone is suitably | |
2281 | * aligned, by adjusting the object size accordingly. */ | |
2282 | size += REDZONE_ALIGN - 1; | |
2283 | size &= ~(REDZONE_ALIGN - 1); | |
2284 | } | |
ca5f9703 | 2285 | |
a44b56d3 | 2286 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2287 | if (ralign < ARCH_SLAB_MINALIGN) { |
2288 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2289 | } |
a44b56d3 | 2290 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2291 | if (ralign < align) { |
2292 | ralign = align; | |
1da177e4 | 2293 | } |
3ff84a7f PE |
2294 | /* disable debug if necessary */ |
2295 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2296 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2297 | /* |
ca5f9703 | 2298 | * 4) Store it. |
1da177e4 LT |
2299 | */ |
2300 | align = ralign; | |
2301 | ||
83b519e8 PE |
2302 | if (slab_is_available()) |
2303 | gfp = GFP_KERNEL; | |
2304 | else | |
2305 | gfp = GFP_NOWAIT; | |
2306 | ||
1da177e4 | 2307 | /* Get cache's description obj. */ |
83b519e8 | 2308 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2309 | if (!cachep) |
4f12bb4f | 2310 | goto oops; |
1da177e4 | 2311 | |
b56efcf0 | 2312 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
1da177e4 | 2313 | #if DEBUG |
3dafccf2 | 2314 | cachep->obj_size = size; |
1da177e4 | 2315 | |
ca5f9703 PE |
2316 | /* |
2317 | * Both debugging options require word-alignment which is calculated | |
2318 | * into align above. | |
2319 | */ | |
1da177e4 | 2320 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2321 | /* add space for red zone words */ |
3ff84a7f PE |
2322 | cachep->obj_offset += sizeof(unsigned long long); |
2323 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2324 | } |
2325 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2326 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2327 | * the real object. But if the second red zone needs to be |
2328 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2329 | */ |
87a927c7 DW |
2330 | if (flags & SLAB_RED_ZONE) |
2331 | size += REDZONE_ALIGN; | |
2332 | else | |
2333 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2334 | } |
2335 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2336 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
1ab335d8 CO |
2337 | && cachep->obj_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
2338 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); | |
1da177e4 LT |
2339 | size = PAGE_SIZE; |
2340 | } | |
2341 | #endif | |
2342 | #endif | |
2343 | ||
e0a42726 IM |
2344 | /* |
2345 | * Determine if the slab management is 'on' or 'off' slab. | |
2346 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2347 | * it too early on. Always use on-slab management when |
2348 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2349 | */ |
e7cb55b9 CM |
2350 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2351 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2352 | /* |
2353 | * Size is large, assume best to place the slab management obj | |
2354 | * off-slab (should allow better packing of objs). | |
2355 | */ | |
2356 | flags |= CFLGS_OFF_SLAB; | |
2357 | ||
2358 | size = ALIGN(size, align); | |
2359 | ||
f78bb8ad | 2360 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2361 | |
2362 | if (!cachep->num) { | |
b4169525 | 2363 | printk(KERN_ERR |
2364 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2365 | kmem_cache_free(&cache_cache, cachep); |
2366 | cachep = NULL; | |
4f12bb4f | 2367 | goto oops; |
1da177e4 | 2368 | } |
b28a02de PE |
2369 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2370 | + sizeof(struct slab), align); | |
1da177e4 LT |
2371 | |
2372 | /* | |
2373 | * If the slab has been placed off-slab, and we have enough space then | |
2374 | * move it on-slab. This is at the expense of any extra colouring. | |
2375 | */ | |
2376 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2377 | flags &= ~CFLGS_OFF_SLAB; | |
2378 | left_over -= slab_size; | |
2379 | } | |
2380 | ||
2381 | if (flags & CFLGS_OFF_SLAB) { | |
2382 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2383 | slab_size = |
2384 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2385 | |
2386 | #ifdef CONFIG_PAGE_POISONING | |
2387 | /* If we're going to use the generic kernel_map_pages() | |
2388 | * poisoning, then it's going to smash the contents of | |
2389 | * the redzone and userword anyhow, so switch them off. | |
2390 | */ | |
2391 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2392 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2393 | #endif | |
1da177e4 LT |
2394 | } |
2395 | ||
2396 | cachep->colour_off = cache_line_size(); | |
2397 | /* Offset must be a multiple of the alignment. */ | |
2398 | if (cachep->colour_off < align) | |
2399 | cachep->colour_off = align; | |
b28a02de | 2400 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2401 | cachep->slab_size = slab_size; |
2402 | cachep->flags = flags; | |
2403 | cachep->gfpflags = 0; | |
4b51d669 | 2404 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2405 | cachep->gfpflags |= GFP_DMA; |
3dafccf2 | 2406 | cachep->buffer_size = size; |
6a2d7a95 | 2407 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2408 | |
e5ac9c5a | 2409 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2410 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2411 | /* |
2412 | * This is a possibility for one of the malloc_sizes caches. | |
2413 | * But since we go off slab only for object size greater than | |
2414 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2415 | * this should not happen at all. | |
2416 | * But leave a BUG_ON for some lucky dude. | |
2417 | */ | |
6cb8f913 | 2418 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2419 | } |
1da177e4 | 2420 | cachep->ctor = ctor; |
1da177e4 LT |
2421 | cachep->name = name; |
2422 | ||
83b519e8 | 2423 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef CL |
2424 | __kmem_cache_destroy(cachep); |
2425 | cachep = NULL; | |
2426 | goto oops; | |
2427 | } | |
1da177e4 | 2428 | |
1da177e4 LT |
2429 | /* cache setup completed, link it into the list */ |
2430 | list_add(&cachep->next, &cache_chain); | |
a737b3e2 | 2431 | oops: |
1da177e4 LT |
2432 | if (!cachep && (flags & SLAB_PANIC)) |
2433 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2434 | name); |
83b519e8 PE |
2435 | if (slab_is_available()) { |
2436 | mutex_unlock(&cache_chain_mutex); | |
2437 | put_online_cpus(); | |
2438 | } | |
1da177e4 LT |
2439 | return cachep; |
2440 | } | |
2441 | EXPORT_SYMBOL(kmem_cache_create); | |
2442 | ||
2443 | #if DEBUG | |
2444 | static void check_irq_off(void) | |
2445 | { | |
2446 | BUG_ON(!irqs_disabled()); | |
2447 | } | |
2448 | ||
2449 | static void check_irq_on(void) | |
2450 | { | |
2451 | BUG_ON(irqs_disabled()); | |
2452 | } | |
2453 | ||
343e0d7a | 2454 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2455 | { |
2456 | #ifdef CONFIG_SMP | |
2457 | check_irq_off(); | |
7d6e6d09 | 2458 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2459 | #endif |
2460 | } | |
e498be7d | 2461 | |
343e0d7a | 2462 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2463 | { |
2464 | #ifdef CONFIG_SMP | |
2465 | check_irq_off(); | |
2466 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2467 | #endif | |
2468 | } | |
2469 | ||
1da177e4 LT |
2470 | #else |
2471 | #define check_irq_off() do { } while(0) | |
2472 | #define check_irq_on() do { } while(0) | |
2473 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2474 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2475 | #endif |
2476 | ||
aab2207c CL |
2477 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2478 | struct array_cache *ac, | |
2479 | int force, int node); | |
2480 | ||
1da177e4 LT |
2481 | static void do_drain(void *arg) |
2482 | { | |
a737b3e2 | 2483 | struct kmem_cache *cachep = arg; |
1da177e4 | 2484 | struct array_cache *ac; |
7d6e6d09 | 2485 | int node = numa_mem_id(); |
1da177e4 LT |
2486 | |
2487 | check_irq_off(); | |
9a2dba4b | 2488 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2489 | spin_lock(&cachep->nodelists[node]->list_lock); |
2490 | free_block(cachep, ac->entry, ac->avail, node); | |
2491 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2492 | ac->avail = 0; |
2493 | } | |
2494 | ||
343e0d7a | 2495 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2496 | { |
e498be7d CL |
2497 | struct kmem_list3 *l3; |
2498 | int node; | |
2499 | ||
15c8b6c1 | 2500 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2501 | check_irq_on(); |
b28a02de | 2502 | for_each_online_node(node) { |
e498be7d | 2503 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2504 | if (l3 && l3->alien) |
2505 | drain_alien_cache(cachep, l3->alien); | |
2506 | } | |
2507 | ||
2508 | for_each_online_node(node) { | |
2509 | l3 = cachep->nodelists[node]; | |
2510 | if (l3) | |
aab2207c | 2511 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2512 | } |
1da177e4 LT |
2513 | } |
2514 | ||
ed11d9eb CL |
2515 | /* |
2516 | * Remove slabs from the list of free slabs. | |
2517 | * Specify the number of slabs to drain in tofree. | |
2518 | * | |
2519 | * Returns the actual number of slabs released. | |
2520 | */ | |
2521 | static int drain_freelist(struct kmem_cache *cache, | |
2522 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2523 | { |
ed11d9eb CL |
2524 | struct list_head *p; |
2525 | int nr_freed; | |
1da177e4 | 2526 | struct slab *slabp; |
1da177e4 | 2527 | |
ed11d9eb CL |
2528 | nr_freed = 0; |
2529 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2530 | |
ed11d9eb | 2531 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2532 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2533 | if (p == &l3->slabs_free) { |
2534 | spin_unlock_irq(&l3->list_lock); | |
2535 | goto out; | |
2536 | } | |
1da177e4 | 2537 | |
ed11d9eb | 2538 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2539 | #if DEBUG |
40094fa6 | 2540 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2541 | #endif |
2542 | list_del(&slabp->list); | |
ed11d9eb CL |
2543 | /* |
2544 | * Safe to drop the lock. The slab is no longer linked | |
2545 | * to the cache. | |
2546 | */ | |
2547 | l3->free_objects -= cache->num; | |
e498be7d | 2548 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2549 | slab_destroy(cache, slabp); |
2550 | nr_freed++; | |
1da177e4 | 2551 | } |
ed11d9eb CL |
2552 | out: |
2553 | return nr_freed; | |
1da177e4 LT |
2554 | } |
2555 | ||
8f5be20b | 2556 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2557 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2558 | { |
2559 | int ret = 0, i = 0; | |
2560 | struct kmem_list3 *l3; | |
2561 | ||
2562 | drain_cpu_caches(cachep); | |
2563 | ||
2564 | check_irq_on(); | |
2565 | for_each_online_node(i) { | |
2566 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2567 | if (!l3) |
2568 | continue; | |
2569 | ||
2570 | drain_freelist(cachep, l3, l3->free_objects); | |
2571 | ||
2572 | ret += !list_empty(&l3->slabs_full) || | |
2573 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2574 | } |
2575 | return (ret ? 1 : 0); | |
2576 | } | |
2577 | ||
1da177e4 LT |
2578 | /** |
2579 | * kmem_cache_shrink - Shrink a cache. | |
2580 | * @cachep: The cache to shrink. | |
2581 | * | |
2582 | * Releases as many slabs as possible for a cache. | |
2583 | * To help debugging, a zero exit status indicates all slabs were released. | |
2584 | */ | |
343e0d7a | 2585 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2586 | { |
8f5be20b | 2587 | int ret; |
40094fa6 | 2588 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2589 | |
95402b38 | 2590 | get_online_cpus(); |
8f5be20b RT |
2591 | mutex_lock(&cache_chain_mutex); |
2592 | ret = __cache_shrink(cachep); | |
2593 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2594 | put_online_cpus(); |
8f5be20b | 2595 | return ret; |
1da177e4 LT |
2596 | } |
2597 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2598 | ||
2599 | /** | |
2600 | * kmem_cache_destroy - delete a cache | |
2601 | * @cachep: the cache to destroy | |
2602 | * | |
72fd4a35 | 2603 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2604 | * |
2605 | * It is expected this function will be called by a module when it is | |
2606 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2607 | * cache being allocated each time a module is loaded and unloaded, if the | |
2608 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2609 | * | |
2610 | * The cache must be empty before calling this function. | |
2611 | * | |
25985edc | 2612 | * The caller must guarantee that no one will allocate memory from the cache |
1da177e4 LT |
2613 | * during the kmem_cache_destroy(). |
2614 | */ | |
133d205a | 2615 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2616 | { |
40094fa6 | 2617 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2618 | |
1da177e4 | 2619 | /* Find the cache in the chain of caches. */ |
95402b38 | 2620 | get_online_cpus(); |
fc0abb14 | 2621 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2622 | /* |
2623 | * the chain is never empty, cache_cache is never destroyed | |
2624 | */ | |
2625 | list_del(&cachep->next); | |
1da177e4 LT |
2626 | if (__cache_shrink(cachep)) { |
2627 | slab_error(cachep, "Can't free all objects"); | |
b28a02de | 2628 | list_add(&cachep->next, &cache_chain); |
fc0abb14 | 2629 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2630 | put_online_cpus(); |
133d205a | 2631 | return; |
1da177e4 LT |
2632 | } |
2633 | ||
2634 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
7ed9f7e5 | 2635 | rcu_barrier(); |
1da177e4 | 2636 | |
117f6eb1 | 2637 | __kmem_cache_destroy(cachep); |
8f5be20b | 2638 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2639 | put_online_cpus(); |
1da177e4 LT |
2640 | } |
2641 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2642 | ||
e5ac9c5a RT |
2643 | /* |
2644 | * Get the memory for a slab management obj. | |
2645 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2646 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2647 | * come from the same cache which is getting created because, | |
2648 | * when we are searching for an appropriate cache for these | |
2649 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2650 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2651 | * kmem_find_general_cachep till the initialization is complete. | |
2652 | * Hence we cannot have slabp_cache same as the original cache. | |
2653 | */ | |
343e0d7a | 2654 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2655 | int colour_off, gfp_t local_flags, |
2656 | int nodeid) | |
1da177e4 LT |
2657 | { |
2658 | struct slab *slabp; | |
b28a02de | 2659 | |
1da177e4 LT |
2660 | if (OFF_SLAB(cachep)) { |
2661 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2662 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2663 | local_flags, nodeid); |
d5cff635 CM |
2664 | /* |
2665 | * If the first object in the slab is leaked (it's allocated | |
2666 | * but no one has a reference to it), we want to make sure | |
2667 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2668 | * to the object. Otherwise we will not report the leak. | |
2669 | */ | |
c017b4be CM |
2670 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2671 | local_flags); | |
1da177e4 LT |
2672 | if (!slabp) |
2673 | return NULL; | |
2674 | } else { | |
b28a02de | 2675 | slabp = objp + colour_off; |
1da177e4 LT |
2676 | colour_off += cachep->slab_size; |
2677 | } | |
2678 | slabp->inuse = 0; | |
2679 | slabp->colouroff = colour_off; | |
b28a02de | 2680 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2681 | slabp->nodeid = nodeid; |
e51bfd0a | 2682 | slabp->free = 0; |
1da177e4 LT |
2683 | return slabp; |
2684 | } | |
2685 | ||
2686 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2687 | { | |
b28a02de | 2688 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2689 | } |
2690 | ||
343e0d7a | 2691 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2692 | struct slab *slabp) |
1da177e4 LT |
2693 | { |
2694 | int i; | |
2695 | ||
2696 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2697 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2698 | #if DEBUG |
2699 | /* need to poison the objs? */ | |
2700 | if (cachep->flags & SLAB_POISON) | |
2701 | poison_obj(cachep, objp, POISON_FREE); | |
2702 | if (cachep->flags & SLAB_STORE_USER) | |
2703 | *dbg_userword(cachep, objp) = NULL; | |
2704 | ||
2705 | if (cachep->flags & SLAB_RED_ZONE) { | |
2706 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2707 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2708 | } | |
2709 | /* | |
a737b3e2 AM |
2710 | * Constructors are not allowed to allocate memory from the same |
2711 | * cache which they are a constructor for. Otherwise, deadlock. | |
2712 | * They must also be threaded. | |
1da177e4 LT |
2713 | */ |
2714 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2715 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2716 | |
2717 | if (cachep->flags & SLAB_RED_ZONE) { | |
2718 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2719 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2720 | " end of an object"); |
1da177e4 LT |
2721 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2722 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2723 | " start of an object"); |
1da177e4 | 2724 | } |
a737b3e2 AM |
2725 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && |
2726 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
b28a02de | 2727 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2728 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2729 | #else |
2730 | if (cachep->ctor) | |
51cc5068 | 2731 | cachep->ctor(objp); |
1da177e4 | 2732 | #endif |
b28a02de | 2733 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2734 | } |
b28a02de | 2735 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2736 | } |
2737 | ||
343e0d7a | 2738 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2739 | { |
4b51d669 CL |
2740 | if (CONFIG_ZONE_DMA_FLAG) { |
2741 | if (flags & GFP_DMA) | |
2742 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2743 | else | |
2744 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2745 | } | |
1da177e4 LT |
2746 | } |
2747 | ||
a737b3e2 AM |
2748 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2749 | int nodeid) | |
78d382d7 | 2750 | { |
8fea4e96 | 2751 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2752 | kmem_bufctl_t next; |
2753 | ||
2754 | slabp->inuse++; | |
2755 | next = slab_bufctl(slabp)[slabp->free]; | |
2756 | #if DEBUG | |
2757 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2758 | WARN_ON(slabp->nodeid != nodeid); | |
2759 | #endif | |
2760 | slabp->free = next; | |
2761 | ||
2762 | return objp; | |
2763 | } | |
2764 | ||
a737b3e2 AM |
2765 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2766 | void *objp, int nodeid) | |
78d382d7 | 2767 | { |
8fea4e96 | 2768 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2769 | |
2770 | #if DEBUG | |
2771 | /* Verify that the slab belongs to the intended node */ | |
2772 | WARN_ON(slabp->nodeid != nodeid); | |
2773 | ||
871751e2 | 2774 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2775 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2776 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2777 | BUG(); |
2778 | } | |
2779 | #endif | |
2780 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2781 | slabp->free = objnr; | |
2782 | slabp->inuse--; | |
2783 | } | |
2784 | ||
4776874f PE |
2785 | /* |
2786 | * Map pages beginning at addr to the given cache and slab. This is required | |
2787 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2788 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2789 | */ |
2790 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2791 | void *addr) | |
1da177e4 | 2792 | { |
4776874f | 2793 | int nr_pages; |
1da177e4 LT |
2794 | struct page *page; |
2795 | ||
4776874f | 2796 | page = virt_to_page(addr); |
84097518 | 2797 | |
4776874f | 2798 | nr_pages = 1; |
84097518 | 2799 | if (likely(!PageCompound(page))) |
4776874f PE |
2800 | nr_pages <<= cache->gfporder; |
2801 | ||
1da177e4 | 2802 | do { |
4776874f PE |
2803 | page_set_cache(page, cache); |
2804 | page_set_slab(page, slab); | |
1da177e4 | 2805 | page++; |
4776874f | 2806 | } while (--nr_pages); |
1da177e4 LT |
2807 | } |
2808 | ||
2809 | /* | |
2810 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2811 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2812 | */ | |
3c517a61 CL |
2813 | static int cache_grow(struct kmem_cache *cachep, |
2814 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2815 | { |
b28a02de | 2816 | struct slab *slabp; |
b28a02de PE |
2817 | size_t offset; |
2818 | gfp_t local_flags; | |
e498be7d | 2819 | struct kmem_list3 *l3; |
1da177e4 | 2820 | |
a737b3e2 AM |
2821 | /* |
2822 | * Be lazy and only check for valid flags here, keeping it out of the | |
2823 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2824 | */ |
6cb06229 CL |
2825 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2826 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2827 | |
2e1217cf | 2828 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2829 | check_irq_off(); |
2e1217cf RT |
2830 | l3 = cachep->nodelists[nodeid]; |
2831 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2832 | |
2833 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2834 | offset = l3->colour_next; |
2835 | l3->colour_next++; | |
2836 | if (l3->colour_next >= cachep->colour) | |
2837 | l3->colour_next = 0; | |
2838 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2839 | |
2e1217cf | 2840 | offset *= cachep->colour_off; |
1da177e4 LT |
2841 | |
2842 | if (local_flags & __GFP_WAIT) | |
2843 | local_irq_enable(); | |
2844 | ||
2845 | /* | |
2846 | * The test for missing atomic flag is performed here, rather than | |
2847 | * the more obvious place, simply to reduce the critical path length | |
2848 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2849 | * will eventually be caught here (where it matters). | |
2850 | */ | |
2851 | kmem_flagcheck(cachep, flags); | |
2852 | ||
a737b3e2 AM |
2853 | /* |
2854 | * Get mem for the objs. Attempt to allocate a physical page from | |
2855 | * 'nodeid'. | |
e498be7d | 2856 | */ |
3c517a61 | 2857 | if (!objp) |
b8c1c5da | 2858 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2859 | if (!objp) |
1da177e4 LT |
2860 | goto failed; |
2861 | ||
2862 | /* Get slab management. */ | |
3c517a61 | 2863 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2864 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2865 | if (!slabp) |
1da177e4 LT |
2866 | goto opps1; |
2867 | ||
4776874f | 2868 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2869 | |
a35afb83 | 2870 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2871 | |
2872 | if (local_flags & __GFP_WAIT) | |
2873 | local_irq_disable(); | |
2874 | check_irq_off(); | |
e498be7d | 2875 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2876 | |
2877 | /* Make slab active. */ | |
e498be7d | 2878 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2879 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2880 | l3->free_objects += cachep->num; |
2881 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2882 | return 1; |
a737b3e2 | 2883 | opps1: |
1da177e4 | 2884 | kmem_freepages(cachep, objp); |
a737b3e2 | 2885 | failed: |
1da177e4 LT |
2886 | if (local_flags & __GFP_WAIT) |
2887 | local_irq_disable(); | |
2888 | return 0; | |
2889 | } | |
2890 | ||
2891 | #if DEBUG | |
2892 | ||
2893 | /* | |
2894 | * Perform extra freeing checks: | |
2895 | * - detect bad pointers. | |
2896 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2897 | */ |
2898 | static void kfree_debugcheck(const void *objp) | |
2899 | { | |
1da177e4 LT |
2900 | if (!virt_addr_valid(objp)) { |
2901 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2902 | (unsigned long)objp); |
2903 | BUG(); | |
1da177e4 | 2904 | } |
1da177e4 LT |
2905 | } |
2906 | ||
58ce1fd5 PE |
2907 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2908 | { | |
b46b8f19 | 2909 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2910 | |
2911 | redzone1 = *dbg_redzone1(cache, obj); | |
2912 | redzone2 = *dbg_redzone2(cache, obj); | |
2913 | ||
2914 | /* | |
2915 | * Redzone is ok. | |
2916 | */ | |
2917 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2918 | return; | |
2919 | ||
2920 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2921 | slab_error(cache, "double free detected"); | |
2922 | else | |
2923 | slab_error(cache, "memory outside object was overwritten"); | |
2924 | ||
b46b8f19 | 2925 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2926 | obj, redzone1, redzone2); |
2927 | } | |
2928 | ||
343e0d7a | 2929 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2930 | void *caller) |
1da177e4 LT |
2931 | { |
2932 | struct page *page; | |
2933 | unsigned int objnr; | |
2934 | struct slab *slabp; | |
2935 | ||
80cbd911 MW |
2936 | BUG_ON(virt_to_cache(objp) != cachep); |
2937 | ||
3dafccf2 | 2938 | objp -= obj_offset(cachep); |
1da177e4 | 2939 | kfree_debugcheck(objp); |
b49af68f | 2940 | page = virt_to_head_page(objp); |
1da177e4 | 2941 | |
065d41cb | 2942 | slabp = page_get_slab(page); |
1da177e4 LT |
2943 | |
2944 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2945 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2946 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2947 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2948 | } | |
2949 | if (cachep->flags & SLAB_STORE_USER) | |
2950 | *dbg_userword(cachep, objp) = caller; | |
2951 | ||
8fea4e96 | 2952 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2953 | |
2954 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2955 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2956 | |
871751e2 AV |
2957 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
2958 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
2959 | #endif | |
1da177e4 LT |
2960 | if (cachep->flags & SLAB_POISON) { |
2961 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
a737b3e2 | 2962 | if ((cachep->buffer_size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 2963 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 2964 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 2965 | cachep->buffer_size / PAGE_SIZE, 0); |
1da177e4 LT |
2966 | } else { |
2967 | poison_obj(cachep, objp, POISON_FREE); | |
2968 | } | |
2969 | #else | |
2970 | poison_obj(cachep, objp, POISON_FREE); | |
2971 | #endif | |
2972 | } | |
2973 | return objp; | |
2974 | } | |
2975 | ||
343e0d7a | 2976 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
2977 | { |
2978 | kmem_bufctl_t i; | |
2979 | int entries = 0; | |
b28a02de | 2980 | |
1da177e4 LT |
2981 | /* Check slab's freelist to see if this obj is there. */ |
2982 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
2983 | entries++; | |
2984 | if (entries > cachep->num || i >= cachep->num) | |
2985 | goto bad; | |
2986 | } | |
2987 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
2988 | bad: |
2989 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
2990 | "cache '%s'(%d), slabp %p(%d). Hexdump:\n", | |
2991 | cachep->name, cachep->num, slabp, slabp->inuse); | |
b28a02de | 2992 | for (i = 0; |
264132bc | 2993 | i < sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t); |
b28a02de | 2994 | i++) { |
a737b3e2 | 2995 | if (i % 16 == 0) |
1da177e4 | 2996 | printk("\n%03x:", i); |
b28a02de | 2997 | printk(" %02x", ((unsigned char *)slabp)[i]); |
1da177e4 LT |
2998 | } |
2999 | printk("\n"); | |
3000 | BUG(); | |
3001 | } | |
3002 | } | |
3003 | #else | |
3004 | #define kfree_debugcheck(x) do { } while(0) | |
3005 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3006 | #define check_slabp(x,y) do { } while(0) | |
3007 | #endif | |
3008 | ||
343e0d7a | 3009 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
3010 | { |
3011 | int batchcount; | |
3012 | struct kmem_list3 *l3; | |
3013 | struct array_cache *ac; | |
1ca4cb24 PE |
3014 | int node; |
3015 | ||
6d2144d3 | 3016 | retry: |
1da177e4 | 3017 | check_irq_off(); |
7d6e6d09 | 3018 | node = numa_mem_id(); |
9a2dba4b | 3019 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3020 | batchcount = ac->batchcount; |
3021 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3022 | /* |
3023 | * If there was little recent activity on this cache, then | |
3024 | * perform only a partial refill. Otherwise we could generate | |
3025 | * refill bouncing. | |
1da177e4 LT |
3026 | */ |
3027 | batchcount = BATCHREFILL_LIMIT; | |
3028 | } | |
1ca4cb24 | 3029 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3030 | |
3031 | BUG_ON(ac->avail > 0 || !l3); | |
3032 | spin_lock(&l3->list_lock); | |
1da177e4 | 3033 | |
3ded175a | 3034 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3035 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3036 | l3->shared->touched = 1; | |
3ded175a | 3037 | goto alloc_done; |
44b57f1c | 3038 | } |
3ded175a | 3039 | |
1da177e4 LT |
3040 | while (batchcount > 0) { |
3041 | struct list_head *entry; | |
3042 | struct slab *slabp; | |
3043 | /* Get slab alloc is to come from. */ | |
3044 | entry = l3->slabs_partial.next; | |
3045 | if (entry == &l3->slabs_partial) { | |
3046 | l3->free_touched = 1; | |
3047 | entry = l3->slabs_free.next; | |
3048 | if (entry == &l3->slabs_free) | |
3049 | goto must_grow; | |
3050 | } | |
3051 | ||
3052 | slabp = list_entry(entry, struct slab, list); | |
3053 | check_slabp(cachep, slabp); | |
3054 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3055 | |
3056 | /* | |
3057 | * The slab was either on partial or free list so | |
3058 | * there must be at least one object available for | |
3059 | * allocation. | |
3060 | */ | |
249b9f33 | 3061 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3062 | |
1da177e4 | 3063 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3064 | STATS_INC_ALLOCED(cachep); |
3065 | STATS_INC_ACTIVE(cachep); | |
3066 | STATS_SET_HIGH(cachep); | |
3067 | ||
78d382d7 | 3068 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3069 | node); |
1da177e4 LT |
3070 | } |
3071 | check_slabp(cachep, slabp); | |
3072 | ||
3073 | /* move slabp to correct slabp list: */ | |
3074 | list_del(&slabp->list); | |
3075 | if (slabp->free == BUFCTL_END) | |
3076 | list_add(&slabp->list, &l3->slabs_full); | |
3077 | else | |
3078 | list_add(&slabp->list, &l3->slabs_partial); | |
3079 | } | |
3080 | ||
a737b3e2 | 3081 | must_grow: |
1da177e4 | 3082 | l3->free_objects -= ac->avail; |
a737b3e2 | 3083 | alloc_done: |
e498be7d | 3084 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3085 | |
3086 | if (unlikely(!ac->avail)) { | |
3087 | int x; | |
3c517a61 | 3088 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3089 | |
a737b3e2 | 3090 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3091 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3092 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3093 | return NULL; |
3094 | ||
a737b3e2 | 3095 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3096 | goto retry; |
3097 | } | |
3098 | ac->touched = 1; | |
e498be7d | 3099 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3100 | } |
3101 | ||
a737b3e2 AM |
3102 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3103 | gfp_t flags) | |
1da177e4 LT |
3104 | { |
3105 | might_sleep_if(flags & __GFP_WAIT); | |
3106 | #if DEBUG | |
3107 | kmem_flagcheck(cachep, flags); | |
3108 | #endif | |
3109 | } | |
3110 | ||
3111 | #if DEBUG | |
a737b3e2 AM |
3112 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3113 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3114 | { |
b28a02de | 3115 | if (!objp) |
1da177e4 | 3116 | return objp; |
b28a02de | 3117 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3118 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3dafccf2 | 3119 | if ((cachep->buffer_size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3120 | kernel_map_pages(virt_to_page(objp), |
3dafccf2 | 3121 | cachep->buffer_size / PAGE_SIZE, 1); |
1da177e4 LT |
3122 | else |
3123 | check_poison_obj(cachep, objp); | |
3124 | #else | |
3125 | check_poison_obj(cachep, objp); | |
3126 | #endif | |
3127 | poison_obj(cachep, objp, POISON_INUSE); | |
3128 | } | |
3129 | if (cachep->flags & SLAB_STORE_USER) | |
3130 | *dbg_userword(cachep, objp) = caller; | |
3131 | ||
3132 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3133 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3134 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3135 | slab_error(cachep, "double free, or memory outside" | |
3136 | " object was overwritten"); | |
b28a02de | 3137 | printk(KERN_ERR |
b46b8f19 | 3138 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3139 | objp, *dbg_redzone1(cachep, objp), |
3140 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3141 | } |
3142 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3143 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3144 | } | |
871751e2 AV |
3145 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3146 | { | |
3147 | struct slab *slabp; | |
3148 | unsigned objnr; | |
3149 | ||
b49af68f | 3150 | slabp = page_get_slab(virt_to_head_page(objp)); |
871751e2 AV |
3151 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->buffer_size; |
3152 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3153 | } | |
3154 | #endif | |
3dafccf2 | 3155 | objp += obj_offset(cachep); |
4f104934 | 3156 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3157 | cachep->ctor(objp); |
7ea466f2 TH |
3158 | if (ARCH_SLAB_MINALIGN && |
3159 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3160 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3161 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3162 | } |
1da177e4 LT |
3163 | return objp; |
3164 | } | |
3165 | #else | |
3166 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3167 | #endif | |
3168 | ||
773ff60e | 3169 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3170 | { |
3171 | if (cachep == &cache_cache) | |
773ff60e | 3172 | return false; |
8a8b6502 | 3173 | |
4c13dd3b | 3174 | return should_failslab(obj_size(cachep), flags, cachep->flags); |
8a8b6502 AM |
3175 | } |
3176 | ||
343e0d7a | 3177 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3178 | { |
b28a02de | 3179 | void *objp; |
1da177e4 LT |
3180 | struct array_cache *ac; |
3181 | ||
5c382300 | 3182 | check_irq_off(); |
8a8b6502 | 3183 | |
9a2dba4b | 3184 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3185 | if (likely(ac->avail)) { |
3186 | STATS_INC_ALLOCHIT(cachep); | |
3187 | ac->touched = 1; | |
e498be7d | 3188 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3189 | } else { |
3190 | STATS_INC_ALLOCMISS(cachep); | |
3191 | objp = cache_alloc_refill(cachep, flags); | |
ddbf2e83 O |
3192 | /* |
3193 | * the 'ac' may be updated by cache_alloc_refill(), | |
3194 | * and kmemleak_erase() requires its correct value. | |
3195 | */ | |
3196 | ac = cpu_cache_get(cachep); | |
1da177e4 | 3197 | } |
d5cff635 CM |
3198 | /* |
3199 | * To avoid a false negative, if an object that is in one of the | |
3200 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3201 | * treat the array pointers as a reference to the object. | |
3202 | */ | |
f3d8b53a O |
3203 | if (objp) |
3204 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3205 | return objp; |
3206 | } | |
3207 | ||
e498be7d | 3208 | #ifdef CONFIG_NUMA |
c61afb18 | 3209 | /* |
b2455396 | 3210 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3211 | * |
3212 | * If we are in_interrupt, then process context, including cpusets and | |
3213 | * mempolicy, may not apply and should not be used for allocation policy. | |
3214 | */ | |
3215 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3216 | { | |
3217 | int nid_alloc, nid_here; | |
3218 | ||
765c4507 | 3219 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3220 | return NULL; |
7d6e6d09 | 3221 | nid_alloc = nid_here = numa_mem_id(); |
c0ff7453 | 3222 | get_mems_allowed(); |
c61afb18 | 3223 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3224 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 PJ |
3225 | else if (current->mempolicy) |
3226 | nid_alloc = slab_node(current->mempolicy); | |
c0ff7453 | 3227 | put_mems_allowed(); |
c61afb18 | 3228 | if (nid_alloc != nid_here) |
8b98c169 | 3229 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3230 | return NULL; |
3231 | } | |
3232 | ||
765c4507 CL |
3233 | /* |
3234 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3235 | * certain node and fall back is permitted. First we scan all the |
3236 | * available nodelists for available objects. If that fails then we | |
3237 | * perform an allocation without specifying a node. This allows the page | |
3238 | * allocator to do its reclaim / fallback magic. We then insert the | |
3239 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3240 | */ |
8c8cc2c1 | 3241 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3242 | { |
8c8cc2c1 PE |
3243 | struct zonelist *zonelist; |
3244 | gfp_t local_flags; | |
dd1a239f | 3245 | struct zoneref *z; |
54a6eb5c MG |
3246 | struct zone *zone; |
3247 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3248 | void *obj = NULL; |
3c517a61 | 3249 | int nid; |
8c8cc2c1 PE |
3250 | |
3251 | if (flags & __GFP_THISNODE) | |
3252 | return NULL; | |
3253 | ||
c0ff7453 | 3254 | get_mems_allowed(); |
0e88460d | 3255 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); |
6cb06229 | 3256 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3257 | |
3c517a61 CL |
3258 | retry: |
3259 | /* | |
3260 | * Look through allowed nodes for objects available | |
3261 | * from existing per node queues. | |
3262 | */ | |
54a6eb5c MG |
3263 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3264 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3265 | |
54a6eb5c | 3266 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3267 | cache->nodelists[nid] && |
481c5346 | 3268 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3269 | obj = ____cache_alloc_node(cache, |
3270 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3271 | if (obj) |
3272 | break; | |
3273 | } | |
3c517a61 CL |
3274 | } |
3275 | ||
cfce6604 | 3276 | if (!obj) { |
3c517a61 CL |
3277 | /* |
3278 | * This allocation will be performed within the constraints | |
3279 | * of the current cpuset / memory policy requirements. | |
3280 | * We may trigger various forms of reclaim on the allowed | |
3281 | * set and go into memory reserves if necessary. | |
3282 | */ | |
dd47ea75 CL |
3283 | if (local_flags & __GFP_WAIT) |
3284 | local_irq_enable(); | |
3285 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3286 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3287 | if (local_flags & __GFP_WAIT) |
3288 | local_irq_disable(); | |
3c517a61 CL |
3289 | if (obj) { |
3290 | /* | |
3291 | * Insert into the appropriate per node queues | |
3292 | */ | |
3293 | nid = page_to_nid(virt_to_page(obj)); | |
3294 | if (cache_grow(cache, flags, nid, obj)) { | |
3295 | obj = ____cache_alloc_node(cache, | |
3296 | flags | GFP_THISNODE, nid); | |
3297 | if (!obj) | |
3298 | /* | |
3299 | * Another processor may allocate the | |
3300 | * objects in the slab since we are | |
3301 | * not holding any locks. | |
3302 | */ | |
3303 | goto retry; | |
3304 | } else { | |
b6a60451 | 3305 | /* cache_grow already freed obj */ |
3c517a61 CL |
3306 | obj = NULL; |
3307 | } | |
3308 | } | |
aedb0eb1 | 3309 | } |
c0ff7453 | 3310 | put_mems_allowed(); |
765c4507 CL |
3311 | return obj; |
3312 | } | |
3313 | ||
e498be7d CL |
3314 | /* |
3315 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3316 | */ |
8b98c169 | 3317 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3318 | int nodeid) |
e498be7d CL |
3319 | { |
3320 | struct list_head *entry; | |
b28a02de PE |
3321 | struct slab *slabp; |
3322 | struct kmem_list3 *l3; | |
3323 | void *obj; | |
b28a02de PE |
3324 | int x; |
3325 | ||
3326 | l3 = cachep->nodelists[nodeid]; | |
3327 | BUG_ON(!l3); | |
3328 | ||
a737b3e2 | 3329 | retry: |
ca3b9b91 | 3330 | check_irq_off(); |
b28a02de PE |
3331 | spin_lock(&l3->list_lock); |
3332 | entry = l3->slabs_partial.next; | |
3333 | if (entry == &l3->slabs_partial) { | |
3334 | l3->free_touched = 1; | |
3335 | entry = l3->slabs_free.next; | |
3336 | if (entry == &l3->slabs_free) | |
3337 | goto must_grow; | |
3338 | } | |
3339 | ||
3340 | slabp = list_entry(entry, struct slab, list); | |
3341 | check_spinlock_acquired_node(cachep, nodeid); | |
3342 | check_slabp(cachep, slabp); | |
3343 | ||
3344 | STATS_INC_NODEALLOCS(cachep); | |
3345 | STATS_INC_ACTIVE(cachep); | |
3346 | STATS_SET_HIGH(cachep); | |
3347 | ||
3348 | BUG_ON(slabp->inuse == cachep->num); | |
3349 | ||
78d382d7 | 3350 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3351 | check_slabp(cachep, slabp); |
3352 | l3->free_objects--; | |
3353 | /* move slabp to correct slabp list: */ | |
3354 | list_del(&slabp->list); | |
3355 | ||
a737b3e2 | 3356 | if (slabp->free == BUFCTL_END) |
b28a02de | 3357 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3358 | else |
b28a02de | 3359 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3360 | |
b28a02de PE |
3361 | spin_unlock(&l3->list_lock); |
3362 | goto done; | |
e498be7d | 3363 | |
a737b3e2 | 3364 | must_grow: |
b28a02de | 3365 | spin_unlock(&l3->list_lock); |
3c517a61 | 3366 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3367 | if (x) |
3368 | goto retry; | |
1da177e4 | 3369 | |
8c8cc2c1 | 3370 | return fallback_alloc(cachep, flags); |
e498be7d | 3371 | |
a737b3e2 | 3372 | done: |
b28a02de | 3373 | return obj; |
e498be7d | 3374 | } |
8c8cc2c1 PE |
3375 | |
3376 | /** | |
3377 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3378 | * @cachep: The cache to allocate from. | |
3379 | * @flags: See kmalloc(). | |
3380 | * @nodeid: node number of the target node. | |
3381 | * @caller: return address of caller, used for debug information | |
3382 | * | |
3383 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3384 | * node, which can improve the performance for cpu bound structures. | |
3385 | * | |
3386 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3387 | */ | |
3388 | static __always_inline void * | |
3389 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3390 | void *caller) | |
3391 | { | |
3392 | unsigned long save_flags; | |
3393 | void *ptr; | |
7d6e6d09 | 3394 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3395 | |
dcce284a | 3396 | flags &= gfp_allowed_mask; |
7e85ee0c | 3397 | |
cf40bd16 NP |
3398 | lockdep_trace_alloc(flags); |
3399 | ||
773ff60e | 3400 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3401 | return NULL; |
3402 | ||
8c8cc2c1 PE |
3403 | cache_alloc_debugcheck_before(cachep, flags); |
3404 | local_irq_save(save_flags); | |
3405 | ||
8e15b79c | 3406 | if (nodeid == -1) |
7d6e6d09 | 3407 | nodeid = slab_node; |
8c8cc2c1 PE |
3408 | |
3409 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3410 | /* Node not bootstrapped yet */ | |
3411 | ptr = fallback_alloc(cachep, flags); | |
3412 | goto out; | |
3413 | } | |
3414 | ||
7d6e6d09 | 3415 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3416 | /* |
3417 | * Use the locally cached objects if possible. | |
3418 | * However ____cache_alloc does not allow fallback | |
3419 | * to other nodes. It may fail while we still have | |
3420 | * objects on other nodes available. | |
3421 | */ | |
3422 | ptr = ____cache_alloc(cachep, flags); | |
3423 | if (ptr) | |
3424 | goto out; | |
3425 | } | |
3426 | /* ___cache_alloc_node can fall back to other nodes */ | |
3427 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3428 | out: | |
3429 | local_irq_restore(save_flags); | |
3430 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
d5cff635 CM |
3431 | kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags, |
3432 | flags); | |
8c8cc2c1 | 3433 | |
c175eea4 PE |
3434 | if (likely(ptr)) |
3435 | kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep)); | |
3436 | ||
d07dbea4 CL |
3437 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3438 | memset(ptr, 0, obj_size(cachep)); | |
3439 | ||
8c8cc2c1 PE |
3440 | return ptr; |
3441 | } | |
3442 | ||
3443 | static __always_inline void * | |
3444 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3445 | { | |
3446 | void *objp; | |
3447 | ||
3448 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3449 | objp = alternate_node_alloc(cache, flags); | |
3450 | if (objp) | |
3451 | goto out; | |
3452 | } | |
3453 | objp = ____cache_alloc(cache, flags); | |
3454 | ||
3455 | /* | |
3456 | * We may just have run out of memory on the local node. | |
3457 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3458 | */ | |
7d6e6d09 LS |
3459 | if (!objp) |
3460 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3461 | |
3462 | out: | |
3463 | return objp; | |
3464 | } | |
3465 | #else | |
3466 | ||
3467 | static __always_inline void * | |
3468 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3469 | { | |
3470 | return ____cache_alloc(cachep, flags); | |
3471 | } | |
3472 | ||
3473 | #endif /* CONFIG_NUMA */ | |
3474 | ||
3475 | static __always_inline void * | |
3476 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3477 | { | |
3478 | unsigned long save_flags; | |
3479 | void *objp; | |
3480 | ||
dcce284a | 3481 | flags &= gfp_allowed_mask; |
7e85ee0c | 3482 | |
cf40bd16 NP |
3483 | lockdep_trace_alloc(flags); |
3484 | ||
773ff60e | 3485 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3486 | return NULL; |
3487 | ||
8c8cc2c1 PE |
3488 | cache_alloc_debugcheck_before(cachep, flags); |
3489 | local_irq_save(save_flags); | |
3490 | objp = __do_cache_alloc(cachep, flags); | |
3491 | local_irq_restore(save_flags); | |
3492 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
d5cff635 CM |
3493 | kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags, |
3494 | flags); | |
8c8cc2c1 PE |
3495 | prefetchw(objp); |
3496 | ||
c175eea4 PE |
3497 | if (likely(objp)) |
3498 | kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep)); | |
3499 | ||
d07dbea4 CL |
3500 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3501 | memset(objp, 0, obj_size(cachep)); | |
3502 | ||
8c8cc2c1 PE |
3503 | return objp; |
3504 | } | |
e498be7d CL |
3505 | |
3506 | /* | |
3507 | * Caller needs to acquire correct kmem_list's list_lock | |
3508 | */ | |
343e0d7a | 3509 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3510 | int node) |
1da177e4 LT |
3511 | { |
3512 | int i; | |
e498be7d | 3513 | struct kmem_list3 *l3; |
1da177e4 LT |
3514 | |
3515 | for (i = 0; i < nr_objects; i++) { | |
3516 | void *objp = objpp[i]; | |
3517 | struct slab *slabp; | |
1da177e4 | 3518 | |
6ed5eb22 | 3519 | slabp = virt_to_slab(objp); |
ff69416e | 3520 | l3 = cachep->nodelists[node]; |
1da177e4 | 3521 | list_del(&slabp->list); |
ff69416e | 3522 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3523 | check_slabp(cachep, slabp); |
78d382d7 | 3524 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3525 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3526 | l3->free_objects++; |
1da177e4 LT |
3527 | check_slabp(cachep, slabp); |
3528 | ||
3529 | /* fixup slab chains */ | |
3530 | if (slabp->inuse == 0) { | |
e498be7d CL |
3531 | if (l3->free_objects > l3->free_limit) { |
3532 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3533 | /* No need to drop any previously held |
3534 | * lock here, even if we have a off-slab slab | |
3535 | * descriptor it is guaranteed to come from | |
3536 | * a different cache, refer to comments before | |
3537 | * alloc_slabmgmt. | |
3538 | */ | |
1da177e4 LT |
3539 | slab_destroy(cachep, slabp); |
3540 | } else { | |
e498be7d | 3541 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3542 | } |
3543 | } else { | |
3544 | /* Unconditionally move a slab to the end of the | |
3545 | * partial list on free - maximum time for the | |
3546 | * other objects to be freed, too. | |
3547 | */ | |
e498be7d | 3548 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3549 | } |
3550 | } | |
3551 | } | |
3552 | ||
343e0d7a | 3553 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3554 | { |
3555 | int batchcount; | |
e498be7d | 3556 | struct kmem_list3 *l3; |
7d6e6d09 | 3557 | int node = numa_mem_id(); |
1da177e4 LT |
3558 | |
3559 | batchcount = ac->batchcount; | |
3560 | #if DEBUG | |
3561 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3562 | #endif | |
3563 | check_irq_off(); | |
ff69416e | 3564 | l3 = cachep->nodelists[node]; |
873623df | 3565 | spin_lock(&l3->list_lock); |
e498be7d CL |
3566 | if (l3->shared) { |
3567 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3568 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3569 | if (max) { |
3570 | if (batchcount > max) | |
3571 | batchcount = max; | |
e498be7d | 3572 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3573 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3574 | shared_array->avail += batchcount; |
3575 | goto free_done; | |
3576 | } | |
3577 | } | |
3578 | ||
ff69416e | 3579 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3580 | free_done: |
1da177e4 LT |
3581 | #if STATS |
3582 | { | |
3583 | int i = 0; | |
3584 | struct list_head *p; | |
3585 | ||
e498be7d CL |
3586 | p = l3->slabs_free.next; |
3587 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3588 | struct slab *slabp; |
3589 | ||
3590 | slabp = list_entry(p, struct slab, list); | |
3591 | BUG_ON(slabp->inuse); | |
3592 | ||
3593 | i++; | |
3594 | p = p->next; | |
3595 | } | |
3596 | STATS_SET_FREEABLE(cachep, i); | |
3597 | } | |
3598 | #endif | |
e498be7d | 3599 | spin_unlock(&l3->list_lock); |
1da177e4 | 3600 | ac->avail -= batchcount; |
a737b3e2 | 3601 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3602 | } |
3603 | ||
3604 | /* | |
a737b3e2 AM |
3605 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3606 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3607 | */ |
a947eb95 SS |
3608 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3609 | void *caller) | |
1da177e4 | 3610 | { |
9a2dba4b | 3611 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3612 | |
3613 | check_irq_off(); | |
d5cff635 | 3614 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3615 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3616 | |
c175eea4 PE |
3617 | kmemcheck_slab_free(cachep, objp, obj_size(cachep)); |
3618 | ||
1807a1aa SS |
3619 | /* |
3620 | * Skip calling cache_free_alien() when the platform is not numa. | |
3621 | * This will avoid cache misses that happen while accessing slabp (which | |
3622 | * is per page memory reference) to get nodeid. Instead use a global | |
3623 | * variable to skip the call, which is mostly likely to be present in | |
3624 | * the cache. | |
3625 | */ | |
b6e68bc1 | 3626 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3627 | return; |
3628 | ||
1da177e4 LT |
3629 | if (likely(ac->avail < ac->limit)) { |
3630 | STATS_INC_FREEHIT(cachep); | |
e498be7d | 3631 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3632 | return; |
3633 | } else { | |
3634 | STATS_INC_FREEMISS(cachep); | |
3635 | cache_flusharray(cachep, ac); | |
e498be7d | 3636 | ac->entry[ac->avail++] = objp; |
1da177e4 LT |
3637 | } |
3638 | } | |
3639 | ||
3640 | /** | |
3641 | * kmem_cache_alloc - Allocate an object | |
3642 | * @cachep: The cache to allocate from. | |
3643 | * @flags: See kmalloc(). | |
3644 | * | |
3645 | * Allocate an object from this cache. The flags are only relevant | |
3646 | * if the cache has no available objects. | |
3647 | */ | |
343e0d7a | 3648 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3649 | { |
36555751 EGM |
3650 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3651 | ||
ca2b84cb EGM |
3652 | trace_kmem_cache_alloc(_RET_IP_, ret, |
3653 | obj_size(cachep), cachep->buffer_size, flags); | |
36555751 EGM |
3654 | |
3655 | return ret; | |
1da177e4 LT |
3656 | } |
3657 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3658 | ||
0f24f128 | 3659 | #ifdef CONFIG_TRACING |
85beb586 SR |
3660 | void * |
3661 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3662 | { |
85beb586 SR |
3663 | void *ret; |
3664 | ||
3665 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3666 | ||
3667 | trace_kmalloc(_RET_IP_, ret, | |
3668 | size, slab_buffer_size(cachep), flags); | |
3669 | return ret; | |
36555751 | 3670 | } |
85beb586 | 3671 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3672 | #endif |
3673 | ||
1da177e4 | 3674 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3675 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3676 | { | |
36555751 EGM |
3677 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3678 | __builtin_return_address(0)); | |
3679 | ||
ca2b84cb EGM |
3680 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
3681 | obj_size(cachep), cachep->buffer_size, | |
3682 | flags, nodeid); | |
36555751 EGM |
3683 | |
3684 | return ret; | |
8b98c169 | 3685 | } |
1da177e4 LT |
3686 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3687 | ||
0f24f128 | 3688 | #ifdef CONFIG_TRACING |
85beb586 SR |
3689 | void *kmem_cache_alloc_node_trace(size_t size, |
3690 | struct kmem_cache *cachep, | |
3691 | gfp_t flags, | |
3692 | int nodeid) | |
36555751 | 3693 | { |
85beb586 SR |
3694 | void *ret; |
3695 | ||
3696 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3697 | __builtin_return_address(0)); |
85beb586 SR |
3698 | trace_kmalloc_node(_RET_IP_, ret, |
3699 | size, slab_buffer_size(cachep), | |
3700 | flags, nodeid); | |
3701 | return ret; | |
36555751 | 3702 | } |
85beb586 | 3703 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3704 | #endif |
3705 | ||
8b98c169 CH |
3706 | static __always_inline void * |
3707 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3708 | { |
343e0d7a | 3709 | struct kmem_cache *cachep; |
97e2bde4 MS |
3710 | |
3711 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3712 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3713 | return cachep; | |
85beb586 | 3714 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3715 | } |
8b98c169 | 3716 | |
0bb38a5c | 3717 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3718 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3719 | { | |
3720 | return __do_kmalloc_node(size, flags, node, | |
3721 | __builtin_return_address(0)); | |
3722 | } | |
dbe5e69d | 3723 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3724 | |
3725 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3726 | int node, unsigned long caller) |
8b98c169 | 3727 | { |
ce71e27c | 3728 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3729 | } |
3730 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3731 | #else | |
3732 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3733 | { | |
3734 | return __do_kmalloc_node(size, flags, node, NULL); | |
3735 | } | |
3736 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3737 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3738 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3739 | |
3740 | /** | |
800590f5 | 3741 | * __do_kmalloc - allocate memory |
1da177e4 | 3742 | * @size: how many bytes of memory are required. |
800590f5 | 3743 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3744 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3745 | */ |
7fd6b141 PE |
3746 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3747 | void *caller) | |
1da177e4 | 3748 | { |
343e0d7a | 3749 | struct kmem_cache *cachep; |
36555751 | 3750 | void *ret; |
1da177e4 | 3751 | |
97e2bde4 MS |
3752 | /* If you want to save a few bytes .text space: replace |
3753 | * __ with kmem_. | |
3754 | * Then kmalloc uses the uninlined functions instead of the inline | |
3755 | * functions. | |
3756 | */ | |
3757 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3758 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3759 | return cachep; | |
36555751 EGM |
3760 | ret = __cache_alloc(cachep, flags, caller); |
3761 | ||
ca2b84cb EGM |
3762 | trace_kmalloc((unsigned long) caller, ret, |
3763 | size, cachep->buffer_size, flags); | |
36555751 EGM |
3764 | |
3765 | return ret; | |
7fd6b141 PE |
3766 | } |
3767 | ||
7fd6b141 | 3768 | |
0bb38a5c | 3769 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3770 | void *__kmalloc(size_t size, gfp_t flags) |
3771 | { | |
871751e2 | 3772 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3773 | } |
3774 | EXPORT_SYMBOL(__kmalloc); | |
3775 | ||
ce71e27c | 3776 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3777 | { |
ce71e27c | 3778 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3779 | } |
3780 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3781 | |
3782 | #else | |
3783 | void *__kmalloc(size_t size, gfp_t flags) | |
3784 | { | |
3785 | return __do_kmalloc(size, flags, NULL); | |
3786 | } | |
3787 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3788 | #endif |
3789 | ||
1da177e4 LT |
3790 | /** |
3791 | * kmem_cache_free - Deallocate an object | |
3792 | * @cachep: The cache the allocation was from. | |
3793 | * @objp: The previously allocated object. | |
3794 | * | |
3795 | * Free an object which was previously allocated from this | |
3796 | * cache. | |
3797 | */ | |
343e0d7a | 3798 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3799 | { |
3800 | unsigned long flags; | |
3801 | ||
3802 | local_irq_save(flags); | |
898552c9 | 3803 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
3ac7fe5a TG |
3804 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3805 | debug_check_no_obj_freed(objp, obj_size(cachep)); | |
a947eb95 | 3806 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3807 | local_irq_restore(flags); |
36555751 | 3808 | |
ca2b84cb | 3809 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3810 | } |
3811 | EXPORT_SYMBOL(kmem_cache_free); | |
3812 | ||
1da177e4 LT |
3813 | /** |
3814 | * kfree - free previously allocated memory | |
3815 | * @objp: pointer returned by kmalloc. | |
3816 | * | |
80e93eff PE |
3817 | * If @objp is NULL, no operation is performed. |
3818 | * | |
1da177e4 LT |
3819 | * Don't free memory not originally allocated by kmalloc() |
3820 | * or you will run into trouble. | |
3821 | */ | |
3822 | void kfree(const void *objp) | |
3823 | { | |
343e0d7a | 3824 | struct kmem_cache *c; |
1da177e4 LT |
3825 | unsigned long flags; |
3826 | ||
2121db74 PE |
3827 | trace_kfree(_RET_IP_, objp); |
3828 | ||
6cb8f913 | 3829 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3830 | return; |
3831 | local_irq_save(flags); | |
3832 | kfree_debugcheck(objp); | |
6ed5eb22 | 3833 | c = virt_to_cache(objp); |
f9b8404c | 3834 | debug_check_no_locks_freed(objp, obj_size(c)); |
3ac7fe5a | 3835 | debug_check_no_obj_freed(objp, obj_size(c)); |
a947eb95 | 3836 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
3837 | local_irq_restore(flags); |
3838 | } | |
3839 | EXPORT_SYMBOL(kfree); | |
3840 | ||
343e0d7a | 3841 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3842 | { |
3dafccf2 | 3843 | return obj_size(cachep); |
1da177e4 LT |
3844 | } |
3845 | EXPORT_SYMBOL(kmem_cache_size); | |
3846 | ||
e498be7d | 3847 | /* |
183ff22b | 3848 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3849 | */ |
83b519e8 | 3850 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3851 | { |
3852 | int node; | |
3853 | struct kmem_list3 *l3; | |
cafeb02e | 3854 | struct array_cache *new_shared; |
3395ee05 | 3855 | struct array_cache **new_alien = NULL; |
e498be7d | 3856 | |
9c09a95c | 3857 | for_each_online_node(node) { |
cafeb02e | 3858 | |
3395ee05 | 3859 | if (use_alien_caches) { |
83b519e8 | 3860 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3861 | if (!new_alien) |
3862 | goto fail; | |
3863 | } | |
cafeb02e | 3864 | |
63109846 ED |
3865 | new_shared = NULL; |
3866 | if (cachep->shared) { | |
3867 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3868 | cachep->shared*cachep->batchcount, |
83b519e8 | 3869 | 0xbaadf00d, gfp); |
63109846 ED |
3870 | if (!new_shared) { |
3871 | free_alien_cache(new_alien); | |
3872 | goto fail; | |
3873 | } | |
0718dc2a | 3874 | } |
cafeb02e | 3875 | |
a737b3e2 AM |
3876 | l3 = cachep->nodelists[node]; |
3877 | if (l3) { | |
cafeb02e CL |
3878 | struct array_cache *shared = l3->shared; |
3879 | ||
e498be7d CL |
3880 | spin_lock_irq(&l3->list_lock); |
3881 | ||
cafeb02e | 3882 | if (shared) |
0718dc2a CL |
3883 | free_block(cachep, shared->entry, |
3884 | shared->avail, node); | |
e498be7d | 3885 | |
cafeb02e CL |
3886 | l3->shared = new_shared; |
3887 | if (!l3->alien) { | |
e498be7d CL |
3888 | l3->alien = new_alien; |
3889 | new_alien = NULL; | |
3890 | } | |
b28a02de | 3891 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3892 | cachep->batchcount + cachep->num; |
e498be7d | 3893 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3894 | kfree(shared); |
e498be7d CL |
3895 | free_alien_cache(new_alien); |
3896 | continue; | |
3897 | } | |
83b519e8 | 3898 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3899 | if (!l3) { |
3900 | free_alien_cache(new_alien); | |
3901 | kfree(new_shared); | |
e498be7d | 3902 | goto fail; |
0718dc2a | 3903 | } |
e498be7d CL |
3904 | |
3905 | kmem_list3_init(l3); | |
3906 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3907 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3908 | l3->shared = new_shared; |
e498be7d | 3909 | l3->alien = new_alien; |
b28a02de | 3910 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3911 | cachep->batchcount + cachep->num; |
e498be7d CL |
3912 | cachep->nodelists[node] = l3; |
3913 | } | |
cafeb02e | 3914 | return 0; |
0718dc2a | 3915 | |
a737b3e2 | 3916 | fail: |
0718dc2a CL |
3917 | if (!cachep->next.next) { |
3918 | /* Cache is not active yet. Roll back what we did */ | |
3919 | node--; | |
3920 | while (node >= 0) { | |
3921 | if (cachep->nodelists[node]) { | |
3922 | l3 = cachep->nodelists[node]; | |
3923 | ||
3924 | kfree(l3->shared); | |
3925 | free_alien_cache(l3->alien); | |
3926 | kfree(l3); | |
3927 | cachep->nodelists[node] = NULL; | |
3928 | } | |
3929 | node--; | |
3930 | } | |
3931 | } | |
cafeb02e | 3932 | return -ENOMEM; |
e498be7d CL |
3933 | } |
3934 | ||
1da177e4 | 3935 | struct ccupdate_struct { |
343e0d7a | 3936 | struct kmem_cache *cachep; |
acfe7d74 | 3937 | struct array_cache *new[0]; |
1da177e4 LT |
3938 | }; |
3939 | ||
3940 | static void do_ccupdate_local(void *info) | |
3941 | { | |
a737b3e2 | 3942 | struct ccupdate_struct *new = info; |
1da177e4 LT |
3943 | struct array_cache *old; |
3944 | ||
3945 | check_irq_off(); | |
9a2dba4b | 3946 | old = cpu_cache_get(new->cachep); |
e498be7d | 3947 | |
1da177e4 LT |
3948 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
3949 | new->new[smp_processor_id()] = old; | |
3950 | } | |
3951 | ||
b5d8ca7c | 3952 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 3953 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 3954 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 3955 | { |
d2e7b7d0 | 3956 | struct ccupdate_struct *new; |
2ed3a4ef | 3957 | int i; |
1da177e4 | 3958 | |
acfe7d74 ED |
3959 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
3960 | gfp); | |
d2e7b7d0 SS |
3961 | if (!new) |
3962 | return -ENOMEM; | |
3963 | ||
e498be7d | 3964 | for_each_online_cpu(i) { |
7d6e6d09 | 3965 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 3966 | batchcount, gfp); |
d2e7b7d0 | 3967 | if (!new->new[i]) { |
b28a02de | 3968 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
3969 | kfree(new->new[i]); |
3970 | kfree(new); | |
e498be7d | 3971 | return -ENOMEM; |
1da177e4 LT |
3972 | } |
3973 | } | |
d2e7b7d0 | 3974 | new->cachep = cachep; |
1da177e4 | 3975 | |
15c8b6c1 | 3976 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 3977 | |
1da177e4 | 3978 | check_irq_on(); |
1da177e4 LT |
3979 | cachep->batchcount = batchcount; |
3980 | cachep->limit = limit; | |
e498be7d | 3981 | cachep->shared = shared; |
1da177e4 | 3982 | |
e498be7d | 3983 | for_each_online_cpu(i) { |
d2e7b7d0 | 3984 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
3985 | if (!ccold) |
3986 | continue; | |
7d6e6d09 LS |
3987 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
3988 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
3989 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
3990 | kfree(ccold); |
3991 | } | |
d2e7b7d0 | 3992 | kfree(new); |
83b519e8 | 3993 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
3994 | } |
3995 | ||
b5d8ca7c | 3996 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 3997 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
3998 | { |
3999 | int err; | |
4000 | int limit, shared; | |
4001 | ||
a737b3e2 AM |
4002 | /* |
4003 | * The head array serves three purposes: | |
1da177e4 LT |
4004 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4005 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4006 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4007 | * bufctl chains: array operations are cheaper. |
4008 | * The numbers are guessed, we should auto-tune as described by | |
4009 | * Bonwick. | |
4010 | */ | |
3dafccf2 | 4011 | if (cachep->buffer_size > 131072) |
1da177e4 | 4012 | limit = 1; |
3dafccf2 | 4013 | else if (cachep->buffer_size > PAGE_SIZE) |
1da177e4 | 4014 | limit = 8; |
3dafccf2 | 4015 | else if (cachep->buffer_size > 1024) |
1da177e4 | 4016 | limit = 24; |
3dafccf2 | 4017 | else if (cachep->buffer_size > 256) |
1da177e4 LT |
4018 | limit = 54; |
4019 | else | |
4020 | limit = 120; | |
4021 | ||
a737b3e2 AM |
4022 | /* |
4023 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4024 | * allocation behaviour: Most allocs on one cpu, most free operations |
4025 | * on another cpu. For these cases, an efficient object passing between | |
4026 | * cpus is necessary. This is provided by a shared array. The array | |
4027 | * replaces Bonwick's magazine layer. | |
4028 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4029 | * to a larger limit. Thus disabled by default. | |
4030 | */ | |
4031 | shared = 0; | |
364fbb29 | 4032 | if (cachep->buffer_size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4033 | shared = 8; |
1da177e4 LT |
4034 | |
4035 | #if DEBUG | |
a737b3e2 AM |
4036 | /* |
4037 | * With debugging enabled, large batchcount lead to excessively long | |
4038 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4039 | */ |
4040 | if (limit > 32) | |
4041 | limit = 32; | |
4042 | #endif | |
83b519e8 | 4043 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4044 | if (err) |
4045 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4046 | cachep->name, -err); |
2ed3a4ef | 4047 | return err; |
1da177e4 LT |
4048 | } |
4049 | ||
1b55253a CL |
4050 | /* |
4051 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4052 | * necessary. Note that the l3 listlock also protects the array_cache |
4053 | * if drain_array() is used on the shared array. | |
1b55253a | 4054 | */ |
68a1b195 | 4055 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4056 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4057 | { |
4058 | int tofree; | |
4059 | ||
1b55253a CL |
4060 | if (!ac || !ac->avail) |
4061 | return; | |
1da177e4 LT |
4062 | if (ac->touched && !force) { |
4063 | ac->touched = 0; | |
b18e7e65 | 4064 | } else { |
1b55253a | 4065 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4066 | if (ac->avail) { |
4067 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4068 | if (tofree > ac->avail) | |
4069 | tofree = (ac->avail + 1) / 2; | |
4070 | free_block(cachep, ac->entry, tofree, node); | |
4071 | ac->avail -= tofree; | |
4072 | memmove(ac->entry, &(ac->entry[tofree]), | |
4073 | sizeof(void *) * ac->avail); | |
4074 | } | |
1b55253a | 4075 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4076 | } |
4077 | } | |
4078 | ||
4079 | /** | |
4080 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4081 | * @w: work descriptor |
1da177e4 LT |
4082 | * |
4083 | * Called from workqueue/eventd every few seconds. | |
4084 | * Purpose: | |
4085 | * - clear the per-cpu caches for this CPU. | |
4086 | * - return freeable pages to the main free memory pool. | |
4087 | * | |
a737b3e2 AM |
4088 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4089 | * again on the next iteration. | |
1da177e4 | 4090 | */ |
7c5cae36 | 4091 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4092 | { |
7a7c381d | 4093 | struct kmem_cache *searchp; |
e498be7d | 4094 | struct kmem_list3 *l3; |
7d6e6d09 | 4095 | int node = numa_mem_id(); |
bf6aede7 | 4096 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4097 | |
7c5cae36 | 4098 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4099 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4100 | goto out; |
1da177e4 | 4101 | |
7a7c381d | 4102 | list_for_each_entry(searchp, &cache_chain, next) { |
1da177e4 LT |
4103 | check_irq_on(); |
4104 | ||
35386e3b CL |
4105 | /* |
4106 | * We only take the l3 lock if absolutely necessary and we | |
4107 | * have established with reasonable certainty that | |
4108 | * we can do some work if the lock was obtained. | |
4109 | */ | |
aab2207c | 4110 | l3 = searchp->nodelists[node]; |
35386e3b | 4111 | |
8fce4d8e | 4112 | reap_alien(searchp, l3); |
1da177e4 | 4113 | |
aab2207c | 4114 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4115 | |
35386e3b CL |
4116 | /* |
4117 | * These are racy checks but it does not matter | |
4118 | * if we skip one check or scan twice. | |
4119 | */ | |
e498be7d | 4120 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4121 | goto next; |
1da177e4 | 4122 | |
e498be7d | 4123 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4124 | |
aab2207c | 4125 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4126 | |
ed11d9eb | 4127 | if (l3->free_touched) |
e498be7d | 4128 | l3->free_touched = 0; |
ed11d9eb CL |
4129 | else { |
4130 | int freed; | |
1da177e4 | 4131 | |
ed11d9eb CL |
4132 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4133 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4134 | STATS_ADD_REAPED(searchp, freed); | |
4135 | } | |
35386e3b | 4136 | next: |
1da177e4 LT |
4137 | cond_resched(); |
4138 | } | |
4139 | check_irq_on(); | |
fc0abb14 | 4140 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4141 | next_reap_node(); |
7c5cae36 | 4142 | out: |
a737b3e2 | 4143 | /* Set up the next iteration */ |
7c5cae36 | 4144 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4145 | } |
4146 | ||
158a9624 | 4147 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4148 | |
85289f98 | 4149 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4150 | { |
85289f98 PE |
4151 | /* |
4152 | * Output format version, so at least we can change it | |
4153 | * without _too_ many complaints. | |
4154 | */ | |
1da177e4 | 4155 | #if STATS |
85289f98 | 4156 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4157 | #else |
85289f98 | 4158 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4159 | #endif |
85289f98 PE |
4160 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4161 | "<objperslab> <pagesperslab>"); | |
4162 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4163 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4164 | #if STATS |
85289f98 | 4165 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4166 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4167 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4168 | #endif |
85289f98 PE |
4169 | seq_putc(m, '\n'); |
4170 | } | |
4171 | ||
4172 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4173 | { | |
4174 | loff_t n = *pos; | |
85289f98 | 4175 | |
fc0abb14 | 4176 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4177 | if (!n) |
4178 | print_slabinfo_header(m); | |
b92151ba PE |
4179 | |
4180 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4181 | } |
4182 | ||
4183 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4184 | { | |
b92151ba | 4185 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4186 | } |
4187 | ||
4188 | static void s_stop(struct seq_file *m, void *p) | |
4189 | { | |
fc0abb14 | 4190 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4191 | } |
4192 | ||
4193 | static int s_show(struct seq_file *m, void *p) | |
4194 | { | |
b92151ba | 4195 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
b28a02de PE |
4196 | struct slab *slabp; |
4197 | unsigned long active_objs; | |
4198 | unsigned long num_objs; | |
4199 | unsigned long active_slabs = 0; | |
4200 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4201 | const char *name; |
1da177e4 | 4202 | char *error = NULL; |
e498be7d CL |
4203 | int node; |
4204 | struct kmem_list3 *l3; | |
1da177e4 | 4205 | |
1da177e4 LT |
4206 | active_objs = 0; |
4207 | num_slabs = 0; | |
e498be7d CL |
4208 | for_each_online_node(node) { |
4209 | l3 = cachep->nodelists[node]; | |
4210 | if (!l3) | |
4211 | continue; | |
4212 | ||
ca3b9b91 RT |
4213 | check_irq_on(); |
4214 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4215 | |
7a7c381d | 4216 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4217 | if (slabp->inuse != cachep->num && !error) |
4218 | error = "slabs_full accounting error"; | |
4219 | active_objs += cachep->num; | |
4220 | active_slabs++; | |
4221 | } | |
7a7c381d | 4222 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4223 | if (slabp->inuse == cachep->num && !error) |
4224 | error = "slabs_partial inuse accounting error"; | |
4225 | if (!slabp->inuse && !error) | |
4226 | error = "slabs_partial/inuse accounting error"; | |
4227 | active_objs += slabp->inuse; | |
4228 | active_slabs++; | |
4229 | } | |
7a7c381d | 4230 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4231 | if (slabp->inuse && !error) |
4232 | error = "slabs_free/inuse accounting error"; | |
4233 | num_slabs++; | |
4234 | } | |
4235 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4236 | if (l3->shared) |
4237 | shared_avail += l3->shared->avail; | |
e498be7d | 4238 | |
ca3b9b91 | 4239 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4240 | } |
b28a02de PE |
4241 | num_slabs += active_slabs; |
4242 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4243 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4244 | error = "free_objects accounting error"; |
4245 | ||
b28a02de | 4246 | name = cachep->name; |
1da177e4 LT |
4247 | if (error) |
4248 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4249 | ||
4250 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3dafccf2 | 4251 | name, active_objs, num_objs, cachep->buffer_size, |
b28a02de | 4252 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4253 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4254 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4255 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4256 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4257 | #if STATS |
b28a02de | 4258 | { /* list3 stats */ |
1da177e4 LT |
4259 | unsigned long high = cachep->high_mark; |
4260 | unsigned long allocs = cachep->num_allocations; | |
4261 | unsigned long grown = cachep->grown; | |
4262 | unsigned long reaped = cachep->reaped; | |
4263 | unsigned long errors = cachep->errors; | |
4264 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4265 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4266 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4267 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4268 | |
e92dd4fd JP |
4269 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4270 | "%4lu %4lu %4lu %4lu %4lu", | |
4271 | allocs, high, grown, | |
4272 | reaped, errors, max_freeable, node_allocs, | |
4273 | node_frees, overflows); | |
1da177e4 LT |
4274 | } |
4275 | /* cpu stats */ | |
4276 | { | |
4277 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4278 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4279 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4280 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4281 | ||
4282 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4283 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4284 | } |
4285 | #endif | |
4286 | seq_putc(m, '\n'); | |
1da177e4 LT |
4287 | return 0; |
4288 | } | |
4289 | ||
4290 | /* | |
4291 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4292 | * | |
4293 | * Output layout: | |
4294 | * cache-name | |
4295 | * num-active-objs | |
4296 | * total-objs | |
4297 | * object size | |
4298 | * num-active-slabs | |
4299 | * total-slabs | |
4300 | * num-pages-per-slab | |
4301 | * + further values on SMP and with statistics enabled | |
4302 | */ | |
4303 | ||
7b3c3a50 | 4304 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4305 | .start = s_start, |
4306 | .next = s_next, | |
4307 | .stop = s_stop, | |
4308 | .show = s_show, | |
1da177e4 LT |
4309 | }; |
4310 | ||
4311 | #define MAX_SLABINFO_WRITE 128 | |
4312 | /** | |
4313 | * slabinfo_write - Tuning for the slab allocator | |
4314 | * @file: unused | |
4315 | * @buffer: user buffer | |
4316 | * @count: data length | |
4317 | * @ppos: unused | |
4318 | */ | |
68a1b195 | 4319 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4320 | size_t count, loff_t *ppos) |
1da177e4 | 4321 | { |
b28a02de | 4322 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4323 | int limit, batchcount, shared, res; |
7a7c381d | 4324 | struct kmem_cache *cachep; |
b28a02de | 4325 | |
1da177e4 LT |
4326 | if (count > MAX_SLABINFO_WRITE) |
4327 | return -EINVAL; | |
4328 | if (copy_from_user(&kbuf, buffer, count)) | |
4329 | return -EFAULT; | |
b28a02de | 4330 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4331 | |
4332 | tmp = strchr(kbuf, ' '); | |
4333 | if (!tmp) | |
4334 | return -EINVAL; | |
4335 | *tmp = '\0'; | |
4336 | tmp++; | |
4337 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4338 | return -EINVAL; | |
4339 | ||
4340 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4341 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4342 | res = -EINVAL; |
7a7c381d | 4343 | list_for_each_entry(cachep, &cache_chain, next) { |
1da177e4 | 4344 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4345 | if (limit < 1 || batchcount < 1 || |
4346 | batchcount > limit || shared < 0) { | |
e498be7d | 4347 | res = 0; |
1da177e4 | 4348 | } else { |
e498be7d | 4349 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4350 | batchcount, shared, |
4351 | GFP_KERNEL); | |
1da177e4 LT |
4352 | } |
4353 | break; | |
4354 | } | |
4355 | } | |
fc0abb14 | 4356 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4357 | if (res >= 0) |
4358 | res = count; | |
4359 | return res; | |
4360 | } | |
871751e2 | 4361 | |
7b3c3a50 AD |
4362 | static int slabinfo_open(struct inode *inode, struct file *file) |
4363 | { | |
4364 | return seq_open(file, &slabinfo_op); | |
4365 | } | |
4366 | ||
4367 | static const struct file_operations proc_slabinfo_operations = { | |
4368 | .open = slabinfo_open, | |
4369 | .read = seq_read, | |
4370 | .write = slabinfo_write, | |
4371 | .llseek = seq_lseek, | |
4372 | .release = seq_release, | |
4373 | }; | |
4374 | ||
871751e2 AV |
4375 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4376 | ||
4377 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4378 | { | |
871751e2 | 4379 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4380 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4381 | } |
4382 | ||
4383 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4384 | { | |
4385 | unsigned long *p; | |
4386 | int l; | |
4387 | if (!v) | |
4388 | return 1; | |
4389 | l = n[1]; | |
4390 | p = n + 2; | |
4391 | while (l) { | |
4392 | int i = l/2; | |
4393 | unsigned long *q = p + 2 * i; | |
4394 | if (*q == v) { | |
4395 | q[1]++; | |
4396 | return 1; | |
4397 | } | |
4398 | if (*q > v) { | |
4399 | l = i; | |
4400 | } else { | |
4401 | p = q + 2; | |
4402 | l -= i + 1; | |
4403 | } | |
4404 | } | |
4405 | if (++n[1] == n[0]) | |
4406 | return 0; | |
4407 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4408 | p[0] = v; | |
4409 | p[1] = 1; | |
4410 | return 1; | |
4411 | } | |
4412 | ||
4413 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4414 | { | |
4415 | void *p; | |
4416 | int i; | |
4417 | if (n[0] == n[1]) | |
4418 | return; | |
4419 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->buffer_size) { | |
4420 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4421 | continue; | |
4422 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4423 | return; | |
4424 | } | |
4425 | } | |
4426 | ||
4427 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4428 | { | |
4429 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4430 | unsigned long offset, size; |
9281acea | 4431 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4432 | |
a5c43dae | 4433 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4434 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4435 | if (modname[0]) |
871751e2 AV |
4436 | seq_printf(m, " [%s]", modname); |
4437 | return; | |
4438 | } | |
4439 | #endif | |
4440 | seq_printf(m, "%p", (void *)address); | |
4441 | } | |
4442 | ||
4443 | static int leaks_show(struct seq_file *m, void *p) | |
4444 | { | |
b92151ba | 4445 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4446 | struct slab *slabp; |
4447 | struct kmem_list3 *l3; | |
4448 | const char *name; | |
4449 | unsigned long *n = m->private; | |
4450 | int node; | |
4451 | int i; | |
4452 | ||
4453 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4454 | return 0; | |
4455 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4456 | return 0; | |
4457 | ||
4458 | /* OK, we can do it */ | |
4459 | ||
4460 | n[1] = 0; | |
4461 | ||
4462 | for_each_online_node(node) { | |
4463 | l3 = cachep->nodelists[node]; | |
4464 | if (!l3) | |
4465 | continue; | |
4466 | ||
4467 | check_irq_on(); | |
4468 | spin_lock_irq(&l3->list_lock); | |
4469 | ||
7a7c381d | 4470 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4471 | handle_slab(n, cachep, slabp); |
7a7c381d | 4472 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4473 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4474 | spin_unlock_irq(&l3->list_lock); |
4475 | } | |
4476 | name = cachep->name; | |
4477 | if (n[0] == n[1]) { | |
4478 | /* Increase the buffer size */ | |
4479 | mutex_unlock(&cache_chain_mutex); | |
4480 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4481 | if (!m->private) { | |
4482 | /* Too bad, we are really out */ | |
4483 | m->private = n; | |
4484 | mutex_lock(&cache_chain_mutex); | |
4485 | return -ENOMEM; | |
4486 | } | |
4487 | *(unsigned long *)m->private = n[0] * 2; | |
4488 | kfree(n); | |
4489 | mutex_lock(&cache_chain_mutex); | |
4490 | /* Now make sure this entry will be retried */ | |
4491 | m->count = m->size; | |
4492 | return 0; | |
4493 | } | |
4494 | for (i = 0; i < n[1]; i++) { | |
4495 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4496 | show_symbol(m, n[2*i+2]); | |
4497 | seq_putc(m, '\n'); | |
4498 | } | |
d2e7b7d0 | 4499 | |
871751e2 AV |
4500 | return 0; |
4501 | } | |
4502 | ||
a0ec95a8 | 4503 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4504 | .start = leaks_start, |
4505 | .next = s_next, | |
4506 | .stop = s_stop, | |
4507 | .show = leaks_show, | |
4508 | }; | |
a0ec95a8 AD |
4509 | |
4510 | static int slabstats_open(struct inode *inode, struct file *file) | |
4511 | { | |
4512 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4513 | int ret = -ENOMEM; | |
4514 | if (n) { | |
4515 | ret = seq_open(file, &slabstats_op); | |
4516 | if (!ret) { | |
4517 | struct seq_file *m = file->private_data; | |
4518 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4519 | m->private = n; | |
4520 | n = NULL; | |
4521 | } | |
4522 | kfree(n); | |
4523 | } | |
4524 | return ret; | |
4525 | } | |
4526 | ||
4527 | static const struct file_operations proc_slabstats_operations = { | |
4528 | .open = slabstats_open, | |
4529 | .read = seq_read, | |
4530 | .llseek = seq_lseek, | |
4531 | .release = seq_release_private, | |
4532 | }; | |
4533 | #endif | |
4534 | ||
4535 | static int __init slab_proc_init(void) | |
4536 | { | |
7b3c3a50 | 4537 | proc_create("slabinfo",S_IWUSR|S_IRUGO,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4538 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4539 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4540 | #endif |
a0ec95a8 AD |
4541 | return 0; |
4542 | } | |
4543 | module_init(slab_proc_init); | |
1da177e4 LT |
4544 | #endif |
4545 | ||
00e145b6 MS |
4546 | /** |
4547 | * ksize - get the actual amount of memory allocated for a given object | |
4548 | * @objp: Pointer to the object | |
4549 | * | |
4550 | * kmalloc may internally round up allocations and return more memory | |
4551 | * than requested. ksize() can be used to determine the actual amount of | |
4552 | * memory allocated. The caller may use this additional memory, even though | |
4553 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4554 | * The caller must guarantee that objp points to a valid object previously | |
4555 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4556 | * must not be freed during the duration of the call. | |
4557 | */ | |
fd76bab2 | 4558 | size_t ksize(const void *objp) |
1da177e4 | 4559 | { |
ef8b4520 CL |
4560 | BUG_ON(!objp); |
4561 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4562 | return 0; |
1da177e4 | 4563 | |
6ed5eb22 | 4564 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4565 | } |
b1aabecd | 4566 | EXPORT_SYMBOL(ksize); |