]>
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 | ||
4dee6b64 SR |
124 | #include <trace/events/kmem.h> |
125 | ||
1da177e4 | 126 | /* |
50953fe9 | 127 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
128 | * 0 for faster, smaller code (especially in the critical paths). |
129 | * | |
130 | * STATS - 1 to collect stats for /proc/slabinfo. | |
131 | * 0 for faster, smaller code (especially in the critical paths). | |
132 | * | |
133 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
134 | */ | |
135 | ||
136 | #ifdef CONFIG_DEBUG_SLAB | |
137 | #define DEBUG 1 | |
138 | #define STATS 1 | |
139 | #define FORCED_DEBUG 1 | |
140 | #else | |
141 | #define DEBUG 0 | |
142 | #define STATS 0 | |
143 | #define FORCED_DEBUG 0 | |
144 | #endif | |
145 | ||
1da177e4 LT |
146 | /* Shouldn't this be in a header file somewhere? */ |
147 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 148 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 149 | |
1da177e4 LT |
150 | #ifndef ARCH_KMALLOC_FLAGS |
151 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
152 | #endif | |
153 | ||
154 | /* Legal flag mask for kmem_cache_create(). */ | |
155 | #if DEBUG | |
50953fe9 | 156 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 157 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 158 | SLAB_CACHE_DMA | \ |
5af60839 | 159 | SLAB_STORE_USER | \ |
1da177e4 | 160 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 161 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 162 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 | 163 | #else |
ac2b898c | 164 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 165 | SLAB_CACHE_DMA | \ |
1da177e4 | 166 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 167 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 168 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 LT |
169 | #endif |
170 | ||
171 | /* | |
172 | * kmem_bufctl_t: | |
173 | * | |
174 | * Bufctl's are used for linking objs within a slab | |
175 | * linked offsets. | |
176 | * | |
177 | * This implementation relies on "struct page" for locating the cache & | |
178 | * slab an object belongs to. | |
179 | * This allows the bufctl structure to be small (one int), but limits | |
180 | * the number of objects a slab (not a cache) can contain when off-slab | |
181 | * bufctls are used. The limit is the size of the largest general cache | |
182 | * that does not use off-slab slabs. | |
183 | * For 32bit archs with 4 kB pages, is this 56. | |
184 | * This is not serious, as it is only for large objects, when it is unwise | |
185 | * to have too many per slab. | |
186 | * Note: This limit can be raised by introducing a general cache whose size | |
187 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
188 | */ | |
189 | ||
fa5b08d5 | 190 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
191 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
192 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
193 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
194 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 195 | |
1da177e4 LT |
196 | /* |
197 | * struct slab_rcu | |
198 | * | |
199 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
200 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
201 | * we need to approach a kernel structure obliquely, from its address | |
202 | * obtained without the usual locking. We can lock the structure to | |
203 | * stabilize it and check it's still at the given address, only if we | |
204 | * can be sure that the memory has not been meanwhile reused for some | |
205 | * other kind of object (which our subsystem's lock might corrupt). | |
206 | * | |
207 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
208 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
209 | */ |
210 | struct slab_rcu { | |
b28a02de | 211 | struct rcu_head head; |
343e0d7a | 212 | struct kmem_cache *cachep; |
b28a02de | 213 | void *addr; |
1da177e4 LT |
214 | }; |
215 | ||
5bfe53a7 LJ |
216 | /* |
217 | * struct slab | |
218 | * | |
219 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
220 | * for a slab, or allocated from an general cache. | |
221 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
222 | */ | |
223 | struct slab { | |
224 | union { | |
225 | struct { | |
226 | struct list_head list; | |
227 | unsigned long colouroff; | |
228 | void *s_mem; /* including colour offset */ | |
229 | unsigned int inuse; /* num of objs active in slab */ | |
230 | kmem_bufctl_t free; | |
231 | unsigned short nodeid; | |
232 | }; | |
233 | struct slab_rcu __slab_cover_slab_rcu; | |
234 | }; | |
235 | }; | |
236 | ||
1da177e4 LT |
237 | /* |
238 | * struct array_cache | |
239 | * | |
1da177e4 LT |
240 | * Purpose: |
241 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
242 | * - reduce the number of linked list operations | |
243 | * - reduce spinlock operations | |
244 | * | |
245 | * The limit is stored in the per-cpu structure to reduce the data cache | |
246 | * footprint. | |
247 | * | |
248 | */ | |
249 | struct array_cache { | |
250 | unsigned int avail; | |
251 | unsigned int limit; | |
252 | unsigned int batchcount; | |
253 | unsigned int touched; | |
e498be7d | 254 | spinlock_t lock; |
bda5b655 | 255 | void *entry[]; /* |
a737b3e2 AM |
256 | * Must have this definition in here for the proper |
257 | * alignment of array_cache. Also simplifies accessing | |
258 | * the entries. | |
a737b3e2 | 259 | */ |
1da177e4 LT |
260 | }; |
261 | ||
a737b3e2 AM |
262 | /* |
263 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
264 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
265 | */ |
266 | #define BOOT_CPUCACHE_ENTRIES 1 | |
267 | struct arraycache_init { | |
268 | struct array_cache cache; | |
b28a02de | 269 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
270 | }; |
271 | ||
272 | /* | |
e498be7d | 273 | * The slab lists for all objects. |
1da177e4 LT |
274 | */ |
275 | struct kmem_list3 { | |
b28a02de PE |
276 | struct list_head slabs_partial; /* partial list first, better asm code */ |
277 | struct list_head slabs_full; | |
278 | struct list_head slabs_free; | |
279 | unsigned long free_objects; | |
b28a02de | 280 | unsigned int free_limit; |
2e1217cf | 281 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
282 | spinlock_t list_lock; |
283 | struct array_cache *shared; /* shared per node */ | |
284 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
285 | unsigned long next_reap; /* updated without locking */ |
286 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
287 | }; |
288 | ||
e498be7d CL |
289 | /* |
290 | * Need this for bootstrapping a per node allocator. | |
291 | */ | |
556a169d | 292 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
68a1b195 | 293 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 294 | #define CACHE_CACHE 0 |
556a169d PE |
295 | #define SIZE_AC MAX_NUMNODES |
296 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 297 | |
ed11d9eb CL |
298 | static int drain_freelist(struct kmem_cache *cache, |
299 | struct kmem_list3 *l3, int tofree); | |
300 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
301 | int node); | |
83b519e8 | 302 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 303 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 304 | |
e498be7d | 305 | /* |
a737b3e2 AM |
306 | * This function must be completely optimized away if a constant is passed to |
307 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 308 | */ |
7243cc05 | 309 | static __always_inline int index_of(const size_t size) |
e498be7d | 310 | { |
5ec8a847 SR |
311 | extern void __bad_size(void); |
312 | ||
e498be7d CL |
313 | if (__builtin_constant_p(size)) { |
314 | int i = 0; | |
315 | ||
316 | #define CACHE(x) \ | |
317 | if (size <=x) \ | |
318 | return i; \ | |
319 | else \ | |
320 | i++; | |
1c61fc40 | 321 | #include <linux/kmalloc_sizes.h> |
e498be7d | 322 | #undef CACHE |
5ec8a847 | 323 | __bad_size(); |
7243cc05 | 324 | } else |
5ec8a847 | 325 | __bad_size(); |
e498be7d CL |
326 | return 0; |
327 | } | |
328 | ||
e0a42726 IM |
329 | static int slab_early_init = 1; |
330 | ||
e498be7d CL |
331 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
332 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 333 | |
5295a74c | 334 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
335 | { |
336 | INIT_LIST_HEAD(&parent->slabs_full); | |
337 | INIT_LIST_HEAD(&parent->slabs_partial); | |
338 | INIT_LIST_HEAD(&parent->slabs_free); | |
339 | parent->shared = NULL; | |
340 | parent->alien = NULL; | |
2e1217cf | 341 | parent->colour_next = 0; |
e498be7d CL |
342 | spin_lock_init(&parent->list_lock); |
343 | parent->free_objects = 0; | |
344 | parent->free_touched = 0; | |
345 | } | |
346 | ||
a737b3e2 AM |
347 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
348 | do { \ | |
349 | INIT_LIST_HEAD(listp); \ | |
350 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
351 | } while (0) |
352 | ||
a737b3e2 AM |
353 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
354 | do { \ | |
e498be7d CL |
355 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
356 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
357 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
358 | } while (0) | |
1da177e4 | 359 | |
1da177e4 LT |
360 | #define CFLGS_OFF_SLAB (0x80000000UL) |
361 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
362 | ||
363 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
364 | /* |
365 | * Optimization question: fewer reaps means less probability for unnessary | |
366 | * cpucache drain/refill cycles. | |
1da177e4 | 367 | * |
dc6f3f27 | 368 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
369 | * which could lock up otherwise freeable slabs. |
370 | */ | |
371 | #define REAPTIMEOUT_CPUC (2*HZ) | |
372 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
373 | ||
374 | #if STATS | |
375 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
376 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
377 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
378 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 379 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
380 | #define STATS_SET_HIGH(x) \ |
381 | do { \ | |
382 | if ((x)->num_active > (x)->high_mark) \ | |
383 | (x)->high_mark = (x)->num_active; \ | |
384 | } while (0) | |
1da177e4 LT |
385 | #define STATS_INC_ERR(x) ((x)->errors++) |
386 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 387 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 388 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
389 | #define STATS_SET_FREEABLE(x, i) \ |
390 | do { \ | |
391 | if ((x)->max_freeable < i) \ | |
392 | (x)->max_freeable = i; \ | |
393 | } while (0) | |
1da177e4 LT |
394 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
395 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
396 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
397 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
398 | #else | |
399 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
400 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
401 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
402 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 403 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
404 | #define STATS_SET_HIGH(x) do { } while (0) |
405 | #define STATS_INC_ERR(x) do { } while (0) | |
406 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 407 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 408 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 409 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
410 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
411 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
412 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
413 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
414 | #endif | |
415 | ||
416 | #if DEBUG | |
1da177e4 | 417 | |
a737b3e2 AM |
418 | /* |
419 | * memory layout of objects: | |
1da177e4 | 420 | * 0 : objp |
3dafccf2 | 421 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
422 | * the end of an object is aligned with the end of the real |
423 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 424 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 425 | * redzone word. |
3dafccf2 | 426 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
427 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
428 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 429 | * [BYTES_PER_WORD long] |
1da177e4 | 430 | */ |
343e0d7a | 431 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 432 | { |
3dafccf2 | 433 | return cachep->obj_offset; |
1da177e4 LT |
434 | } |
435 | ||
b46b8f19 | 436 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
437 | { |
438 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
439 | return (unsigned long long*) (objp + obj_offset(cachep) - |
440 | sizeof(unsigned long long)); | |
1da177e4 LT |
441 | } |
442 | ||
b46b8f19 | 443 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
444 | { |
445 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
446 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 447 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 448 | sizeof(unsigned long long) - |
87a927c7 | 449 | REDZONE_ALIGN); |
3b0efdfa | 450 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 451 | sizeof(unsigned long long)); |
1da177e4 LT |
452 | } |
453 | ||
343e0d7a | 454 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
455 | { |
456 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 457 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
458 | } |
459 | ||
460 | #else | |
461 | ||
3dafccf2 | 462 | #define obj_offset(x) 0 |
b46b8f19 DW |
463 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
464 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
465 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
466 | ||
467 | #endif | |
468 | ||
0f24f128 | 469 | #ifdef CONFIG_TRACING |
36555751 EGM |
470 | size_t slab_buffer_size(struct kmem_cache *cachep) |
471 | { | |
3b0efdfa | 472 | return cachep->size; |
36555751 EGM |
473 | } |
474 | EXPORT_SYMBOL(slab_buffer_size); | |
475 | #endif | |
476 | ||
1da177e4 | 477 | /* |
3df1cccd DR |
478 | * Do not go above this order unless 0 objects fit into the slab or |
479 | * overridden on the command line. | |
1da177e4 | 480 | */ |
543585cc DR |
481 | #define SLAB_MAX_ORDER_HI 1 |
482 | #define SLAB_MAX_ORDER_LO 0 | |
483 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 484 | static bool slab_max_order_set __initdata; |
1da177e4 | 485 | |
065d41cb PE |
486 | static inline struct kmem_cache *page_get_cache(struct page *page) |
487 | { | |
d85f3385 | 488 | page = compound_head(page); |
ddc2e812 | 489 | BUG_ON(!PageSlab(page)); |
e571b0ad | 490 | return page->slab_cache; |
065d41cb PE |
491 | } |
492 | ||
6ed5eb22 PE |
493 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
494 | { | |
b49af68f | 495 | struct page *page = virt_to_head_page(obj); |
35026088 | 496 | return page->slab_cache; |
6ed5eb22 PE |
497 | } |
498 | ||
499 | static inline struct slab *virt_to_slab(const void *obj) | |
500 | { | |
b49af68f | 501 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
502 | |
503 | VM_BUG_ON(!PageSlab(page)); | |
504 | return page->slab_page; | |
6ed5eb22 PE |
505 | } |
506 | ||
8fea4e96 PE |
507 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
508 | unsigned int idx) | |
509 | { | |
3b0efdfa | 510 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
511 | } |
512 | ||
6a2d7a95 | 513 | /* |
3b0efdfa CL |
514 | * We want to avoid an expensive divide : (offset / cache->size) |
515 | * Using the fact that size is a constant for a particular cache, | |
516 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
517 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
518 | */ | |
519 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
520 | const struct slab *slab, void *obj) | |
8fea4e96 | 521 | { |
6a2d7a95 ED |
522 | u32 offset = (obj - slab->s_mem); |
523 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
524 | } |
525 | ||
a737b3e2 AM |
526 | /* |
527 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
528 | */ | |
1da177e4 LT |
529 | struct cache_sizes malloc_sizes[] = { |
530 | #define CACHE(x) { .cs_size = (x) }, | |
531 | #include <linux/kmalloc_sizes.h> | |
532 | CACHE(ULONG_MAX) | |
533 | #undef CACHE | |
534 | }; | |
535 | EXPORT_SYMBOL(malloc_sizes); | |
536 | ||
537 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
538 | struct cache_names { | |
539 | char *name; | |
540 | char *name_dma; | |
541 | }; | |
542 | ||
543 | static struct cache_names __initdata cache_names[] = { | |
544 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
545 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 546 | {NULL,} |
1da177e4 LT |
547 | #undef CACHE |
548 | }; | |
549 | ||
550 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 551 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 552 | static struct arraycache_init initarray_generic = |
b28a02de | 553 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
554 | |
555 | /* internal cache of cache description objs */ | |
b56efcf0 | 556 | static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES]; |
343e0d7a | 557 | static struct kmem_cache cache_cache = { |
b56efcf0 | 558 | .nodelists = cache_cache_nodelists, |
b28a02de PE |
559 | .batchcount = 1, |
560 | .limit = BOOT_CPUCACHE_ENTRIES, | |
561 | .shared = 1, | |
3b0efdfa | 562 | .size = sizeof(struct kmem_cache), |
b28a02de | 563 | .name = "kmem_cache", |
1da177e4 LT |
564 | }; |
565 | ||
056c6241 RT |
566 | #define BAD_ALIEN_MAGIC 0x01020304ul |
567 | ||
ce79ddc8 PE |
568 | /* |
569 | * chicken and egg problem: delay the per-cpu array allocation | |
570 | * until the general caches are up. | |
571 | */ | |
572 | static enum { | |
573 | NONE, | |
574 | PARTIAL_AC, | |
575 | PARTIAL_L3, | |
576 | EARLY, | |
52cef189 | 577 | LATE, |
ce79ddc8 PE |
578 | FULL |
579 | } g_cpucache_up; | |
580 | ||
581 | /* | |
582 | * used by boot code to determine if it can use slab based allocator | |
583 | */ | |
584 | int slab_is_available(void) | |
585 | { | |
586 | return g_cpucache_up >= EARLY; | |
587 | } | |
588 | ||
f1aaee53 AV |
589 | #ifdef CONFIG_LOCKDEP |
590 | ||
591 | /* | |
592 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
593 | * for other slabs "off slab". | |
594 | * The locking for this is tricky in that it nests within the locks | |
595 | * of all other slabs in a few places; to deal with this special | |
596 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
597 | * |
598 | * We set lock class for alien array caches which are up during init. | |
599 | * The lock annotation will be lost if all cpus of a node goes down and | |
600 | * then comes back up during hotplug | |
f1aaee53 | 601 | */ |
056c6241 RT |
602 | static struct lock_class_key on_slab_l3_key; |
603 | static struct lock_class_key on_slab_alc_key; | |
604 | ||
83835b3d PZ |
605 | static struct lock_class_key debugobj_l3_key; |
606 | static struct lock_class_key debugobj_alc_key; | |
607 | ||
608 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
609 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
610 | int q) | |
611 | { | |
612 | struct array_cache **alc; | |
613 | struct kmem_list3 *l3; | |
614 | int r; | |
615 | ||
616 | l3 = cachep->nodelists[q]; | |
617 | if (!l3) | |
618 | return; | |
619 | ||
620 | lockdep_set_class(&l3->list_lock, l3_key); | |
621 | alc = l3->alien; | |
622 | /* | |
623 | * FIXME: This check for BAD_ALIEN_MAGIC | |
624 | * should go away when common slab code is taught to | |
625 | * work even without alien caches. | |
626 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
627 | * for alloc_alien_cache, | |
628 | */ | |
629 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
630 | return; | |
631 | for_each_node(r) { | |
632 | if (alc[r]) | |
633 | lockdep_set_class(&alc[r]->lock, alc_key); | |
634 | } | |
635 | } | |
636 | ||
637 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
638 | { | |
639 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
640 | } | |
641 | ||
642 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
643 | { | |
644 | int node; | |
645 | ||
646 | for_each_online_node(node) | |
647 | slab_set_debugobj_lock_classes_node(cachep, node); | |
648 | } | |
649 | ||
ce79ddc8 | 650 | static void init_node_lock_keys(int q) |
f1aaee53 | 651 | { |
056c6241 RT |
652 | struct cache_sizes *s = malloc_sizes; |
653 | ||
52cef189 | 654 | if (g_cpucache_up < LATE) |
ce79ddc8 PE |
655 | return; |
656 | ||
657 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
ce79ddc8 | 658 | struct kmem_list3 *l3; |
ce79ddc8 PE |
659 | |
660 | l3 = s->cs_cachep->nodelists[q]; | |
661 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 662 | continue; |
83835b3d PZ |
663 | |
664 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
665 | &on_slab_alc_key, q); | |
f1aaee53 AV |
666 | } |
667 | } | |
ce79ddc8 PE |
668 | |
669 | static inline void init_lock_keys(void) | |
670 | { | |
671 | int node; | |
672 | ||
673 | for_each_node(node) | |
674 | init_node_lock_keys(node); | |
675 | } | |
f1aaee53 | 676 | #else |
ce79ddc8 PE |
677 | static void init_node_lock_keys(int q) |
678 | { | |
679 | } | |
680 | ||
056c6241 | 681 | static inline void init_lock_keys(void) |
f1aaee53 AV |
682 | { |
683 | } | |
83835b3d PZ |
684 | |
685 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
686 | { | |
687 | } | |
688 | ||
689 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
690 | { | |
691 | } | |
f1aaee53 AV |
692 | #endif |
693 | ||
8f5be20b | 694 | /* |
95402b38 | 695 | * Guard access to the cache-chain. |
8f5be20b | 696 | */ |
fc0abb14 | 697 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
698 | static struct list_head cache_chain; |
699 | ||
1871e52c | 700 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 701 | |
343e0d7a | 702 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
703 | { |
704 | return cachep->array[smp_processor_id()]; | |
705 | } | |
706 | ||
a737b3e2 AM |
707 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
708 | gfp_t gfpflags) | |
1da177e4 LT |
709 | { |
710 | struct cache_sizes *csizep = malloc_sizes; | |
711 | ||
712 | #if DEBUG | |
713 | /* This happens if someone tries to call | |
b28a02de PE |
714 | * kmem_cache_create(), or __kmalloc(), before |
715 | * the generic caches are initialized. | |
716 | */ | |
c7e43c78 | 717 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 718 | #endif |
6cb8f913 CL |
719 | if (!size) |
720 | return ZERO_SIZE_PTR; | |
721 | ||
1da177e4 LT |
722 | while (size > csizep->cs_size) |
723 | csizep++; | |
724 | ||
725 | /* | |
0abf40c1 | 726 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
727 | * has cs_{dma,}cachep==NULL. Thus no special case |
728 | * for large kmalloc calls required. | |
729 | */ | |
4b51d669 | 730 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
731 | if (unlikely(gfpflags & GFP_DMA)) |
732 | return csizep->cs_dmacachep; | |
4b51d669 | 733 | #endif |
1da177e4 LT |
734 | return csizep->cs_cachep; |
735 | } | |
736 | ||
b221385b | 737 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
738 | { |
739 | return __find_general_cachep(size, gfpflags); | |
740 | } | |
97e2bde4 | 741 | |
fbaccacf | 742 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 743 | { |
fbaccacf SR |
744 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
745 | } | |
1da177e4 | 746 | |
a737b3e2 AM |
747 | /* |
748 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
749 | */ | |
fbaccacf SR |
750 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
751 | size_t align, int flags, size_t *left_over, | |
752 | unsigned int *num) | |
753 | { | |
754 | int nr_objs; | |
755 | size_t mgmt_size; | |
756 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 757 | |
fbaccacf SR |
758 | /* |
759 | * The slab management structure can be either off the slab or | |
760 | * on it. For the latter case, the memory allocated for a | |
761 | * slab is used for: | |
762 | * | |
763 | * - The struct slab | |
764 | * - One kmem_bufctl_t for each object | |
765 | * - Padding to respect alignment of @align | |
766 | * - @buffer_size bytes for each object | |
767 | * | |
768 | * If the slab management structure is off the slab, then the | |
769 | * alignment will already be calculated into the size. Because | |
770 | * the slabs are all pages aligned, the objects will be at the | |
771 | * correct alignment when allocated. | |
772 | */ | |
773 | if (flags & CFLGS_OFF_SLAB) { | |
774 | mgmt_size = 0; | |
775 | nr_objs = slab_size / buffer_size; | |
776 | ||
777 | if (nr_objs > SLAB_LIMIT) | |
778 | nr_objs = SLAB_LIMIT; | |
779 | } else { | |
780 | /* | |
781 | * Ignore padding for the initial guess. The padding | |
782 | * is at most @align-1 bytes, and @buffer_size is at | |
783 | * least @align. In the worst case, this result will | |
784 | * be one greater than the number of objects that fit | |
785 | * into the memory allocation when taking the padding | |
786 | * into account. | |
787 | */ | |
788 | nr_objs = (slab_size - sizeof(struct slab)) / | |
789 | (buffer_size + sizeof(kmem_bufctl_t)); | |
790 | ||
791 | /* | |
792 | * This calculated number will be either the right | |
793 | * amount, or one greater than what we want. | |
794 | */ | |
795 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
796 | > slab_size) | |
797 | nr_objs--; | |
798 | ||
799 | if (nr_objs > SLAB_LIMIT) | |
800 | nr_objs = SLAB_LIMIT; | |
801 | ||
802 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
803 | } | |
804 | *num = nr_objs; | |
805 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
806 | } |
807 | ||
d40cee24 | 808 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 809 | |
a737b3e2 AM |
810 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
811 | char *msg) | |
1da177e4 LT |
812 | { |
813 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 814 | function, cachep->name, msg); |
1da177e4 LT |
815 | dump_stack(); |
816 | } | |
817 | ||
3395ee05 PM |
818 | /* |
819 | * By default on NUMA we use alien caches to stage the freeing of | |
820 | * objects allocated from other nodes. This causes massive memory | |
821 | * inefficiencies when using fake NUMA setup to split memory into a | |
822 | * large number of small nodes, so it can be disabled on the command | |
823 | * line | |
824 | */ | |
825 | ||
826 | static int use_alien_caches __read_mostly = 1; | |
827 | static int __init noaliencache_setup(char *s) | |
828 | { | |
829 | use_alien_caches = 0; | |
830 | return 1; | |
831 | } | |
832 | __setup("noaliencache", noaliencache_setup); | |
833 | ||
3df1cccd DR |
834 | static int __init slab_max_order_setup(char *str) |
835 | { | |
836 | get_option(&str, &slab_max_order); | |
837 | slab_max_order = slab_max_order < 0 ? 0 : | |
838 | min(slab_max_order, MAX_ORDER - 1); | |
839 | slab_max_order_set = true; | |
840 | ||
841 | return 1; | |
842 | } | |
843 | __setup("slab_max_order=", slab_max_order_setup); | |
844 | ||
8fce4d8e CL |
845 | #ifdef CONFIG_NUMA |
846 | /* | |
847 | * Special reaping functions for NUMA systems called from cache_reap(). | |
848 | * These take care of doing round robin flushing of alien caches (containing | |
849 | * objects freed on different nodes from which they were allocated) and the | |
850 | * flushing of remote pcps by calling drain_node_pages. | |
851 | */ | |
1871e52c | 852 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
853 | |
854 | static void init_reap_node(int cpu) | |
855 | { | |
856 | int node; | |
857 | ||
7d6e6d09 | 858 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 859 | if (node == MAX_NUMNODES) |
442295c9 | 860 | node = first_node(node_online_map); |
8fce4d8e | 861 | |
1871e52c | 862 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
863 | } |
864 | ||
865 | static void next_reap_node(void) | |
866 | { | |
909ea964 | 867 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 868 | |
8fce4d8e CL |
869 | node = next_node(node, node_online_map); |
870 | if (unlikely(node >= MAX_NUMNODES)) | |
871 | node = first_node(node_online_map); | |
909ea964 | 872 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
873 | } |
874 | ||
875 | #else | |
876 | #define init_reap_node(cpu) do { } while (0) | |
877 | #define next_reap_node(void) do { } while (0) | |
878 | #endif | |
879 | ||
1da177e4 LT |
880 | /* |
881 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
882 | * via the workqueue/eventd. | |
883 | * Add the CPU number into the expiration time to minimize the possibility of | |
884 | * the CPUs getting into lockstep and contending for the global cache chain | |
885 | * lock. | |
886 | */ | |
897e679b | 887 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 888 | { |
1871e52c | 889 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
890 | |
891 | /* | |
892 | * When this gets called from do_initcalls via cpucache_init(), | |
893 | * init_workqueues() has already run, so keventd will be setup | |
894 | * at that time. | |
895 | */ | |
52bad64d | 896 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 897 | init_reap_node(cpu); |
78b43536 | 898 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); |
2b284214 AV |
899 | schedule_delayed_work_on(cpu, reap_work, |
900 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
901 | } |
902 | } | |
903 | ||
e498be7d | 904 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 905 | int batchcount, gfp_t gfp) |
1da177e4 | 906 | { |
b28a02de | 907 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
908 | struct array_cache *nc = NULL; |
909 | ||
83b519e8 | 910 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
911 | /* |
912 | * The array_cache structures contain pointers to free object. | |
25985edc | 913 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
914 | * cache the pointers are not cleared and they could be counted as |
915 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
916 | * not scan such objects. | |
917 | */ | |
918 | kmemleak_no_scan(nc); | |
1da177e4 LT |
919 | if (nc) { |
920 | nc->avail = 0; | |
921 | nc->limit = entries; | |
922 | nc->batchcount = batchcount; | |
923 | nc->touched = 0; | |
e498be7d | 924 | spin_lock_init(&nc->lock); |
1da177e4 LT |
925 | } |
926 | return nc; | |
927 | } | |
928 | ||
3ded175a CL |
929 | /* |
930 | * Transfer objects in one arraycache to another. | |
931 | * Locking must be handled by the caller. | |
932 | * | |
933 | * Return the number of entries transferred. | |
934 | */ | |
935 | static int transfer_objects(struct array_cache *to, | |
936 | struct array_cache *from, unsigned int max) | |
937 | { | |
938 | /* Figure out how many entries to transfer */ | |
732eacc0 | 939 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
940 | |
941 | if (!nr) | |
942 | return 0; | |
943 | ||
944 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
945 | sizeof(void *) *nr); | |
946 | ||
947 | from->avail -= nr; | |
948 | to->avail += nr; | |
3ded175a CL |
949 | return nr; |
950 | } | |
951 | ||
765c4507 CL |
952 | #ifndef CONFIG_NUMA |
953 | ||
954 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
955 | #define reap_alien(cachep, l3) do { } while (0) | |
956 | ||
83b519e8 | 957 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
958 | { |
959 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
960 | } | |
961 | ||
962 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
963 | { | |
964 | } | |
965 | ||
966 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
967 | { | |
968 | return 0; | |
969 | } | |
970 | ||
971 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
972 | gfp_t flags) | |
973 | { | |
974 | return NULL; | |
975 | } | |
976 | ||
8b98c169 | 977 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
978 | gfp_t flags, int nodeid) |
979 | { | |
980 | return NULL; | |
981 | } | |
982 | ||
983 | #else /* CONFIG_NUMA */ | |
984 | ||
8b98c169 | 985 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 986 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 987 | |
83b519e8 | 988 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
989 | { |
990 | struct array_cache **ac_ptr; | |
8ef82866 | 991 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
992 | int i; |
993 | ||
994 | if (limit > 1) | |
995 | limit = 12; | |
f3186a9c | 996 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
997 | if (ac_ptr) { |
998 | for_each_node(i) { | |
f3186a9c | 999 | if (i == node || !node_online(i)) |
e498be7d | 1000 | continue; |
83b519e8 | 1001 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1002 | if (!ac_ptr[i]) { |
cc550def | 1003 | for (i--; i >= 0; i--) |
e498be7d CL |
1004 | kfree(ac_ptr[i]); |
1005 | kfree(ac_ptr); | |
1006 | return NULL; | |
1007 | } | |
1008 | } | |
1009 | } | |
1010 | return ac_ptr; | |
1011 | } | |
1012 | ||
5295a74c | 1013 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1014 | { |
1015 | int i; | |
1016 | ||
1017 | if (!ac_ptr) | |
1018 | return; | |
e498be7d | 1019 | for_each_node(i) |
b28a02de | 1020 | kfree(ac_ptr[i]); |
e498be7d CL |
1021 | kfree(ac_ptr); |
1022 | } | |
1023 | ||
343e0d7a | 1024 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1025 | struct array_cache *ac, int node) |
e498be7d CL |
1026 | { |
1027 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1028 | ||
1029 | if (ac->avail) { | |
1030 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1031 | /* |
1032 | * Stuff objects into the remote nodes shared array first. | |
1033 | * That way we could avoid the overhead of putting the objects | |
1034 | * into the free lists and getting them back later. | |
1035 | */ | |
693f7d36 | 1036 | if (rl3->shared) |
1037 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1038 | |
ff69416e | 1039 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1040 | ac->avail = 0; |
1041 | spin_unlock(&rl3->list_lock); | |
1042 | } | |
1043 | } | |
1044 | ||
8fce4d8e CL |
1045 | /* |
1046 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1047 | */ | |
1048 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1049 | { | |
909ea964 | 1050 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1051 | |
1052 | if (l3->alien) { | |
1053 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1054 | |
1055 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1056 | __drain_alien_cache(cachep, ac, node); |
1057 | spin_unlock_irq(&ac->lock); | |
1058 | } | |
1059 | } | |
1060 | } | |
1061 | ||
a737b3e2 AM |
1062 | static void drain_alien_cache(struct kmem_cache *cachep, |
1063 | struct array_cache **alien) | |
e498be7d | 1064 | { |
b28a02de | 1065 | int i = 0; |
e498be7d CL |
1066 | struct array_cache *ac; |
1067 | unsigned long flags; | |
1068 | ||
1069 | for_each_online_node(i) { | |
4484ebf1 | 1070 | ac = alien[i]; |
e498be7d CL |
1071 | if (ac) { |
1072 | spin_lock_irqsave(&ac->lock, flags); | |
1073 | __drain_alien_cache(cachep, ac, i); | |
1074 | spin_unlock_irqrestore(&ac->lock, flags); | |
1075 | } | |
1076 | } | |
1077 | } | |
729bd0b7 | 1078 | |
873623df | 1079 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1080 | { |
1081 | struct slab *slabp = virt_to_slab(objp); | |
1082 | int nodeid = slabp->nodeid; | |
1083 | struct kmem_list3 *l3; | |
1084 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1085 | int node; |
1086 | ||
7d6e6d09 | 1087 | node = numa_mem_id(); |
729bd0b7 PE |
1088 | |
1089 | /* | |
1090 | * Make sure we are not freeing a object from another node to the array | |
1091 | * cache on this cpu. | |
1092 | */ | |
62918a03 | 1093 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1094 | return 0; |
1095 | ||
1ca4cb24 | 1096 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1097 | STATS_INC_NODEFREES(cachep); |
1098 | if (l3->alien && l3->alien[nodeid]) { | |
1099 | alien = l3->alien[nodeid]; | |
873623df | 1100 | spin_lock(&alien->lock); |
729bd0b7 PE |
1101 | if (unlikely(alien->avail == alien->limit)) { |
1102 | STATS_INC_ACOVERFLOW(cachep); | |
1103 | __drain_alien_cache(cachep, alien, nodeid); | |
1104 | } | |
1105 | alien->entry[alien->avail++] = objp; | |
1106 | spin_unlock(&alien->lock); | |
1107 | } else { | |
1108 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1109 | free_block(cachep, &objp, 1, nodeid); | |
1110 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1111 | } | |
1112 | return 1; | |
1113 | } | |
e498be7d CL |
1114 | #endif |
1115 | ||
8f9f8d9e DR |
1116 | /* |
1117 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1118 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1119 | * will be allocated off-node since memory is not yet online for the new node. | |
1120 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1121 | * already in use. | |
1122 | * | |
1123 | * Must hold cache_chain_mutex. | |
1124 | */ | |
1125 | static int init_cache_nodelists_node(int node) | |
1126 | { | |
1127 | struct kmem_cache *cachep; | |
1128 | struct kmem_list3 *l3; | |
1129 | const int memsize = sizeof(struct kmem_list3); | |
1130 | ||
3b0efdfa | 1131 | list_for_each_entry(cachep, &cache_chain, list) { |
8f9f8d9e DR |
1132 | /* |
1133 | * Set up the size64 kmemlist for cpu before we can | |
1134 | * begin anything. Make sure some other cpu on this | |
1135 | * node has not already allocated this | |
1136 | */ | |
1137 | if (!cachep->nodelists[node]) { | |
1138 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1139 | if (!l3) | |
1140 | return -ENOMEM; | |
1141 | kmem_list3_init(l3); | |
1142 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1143 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1144 | ||
1145 | /* | |
1146 | * The l3s don't come and go as CPUs come and | |
1147 | * go. cache_chain_mutex is sufficient | |
1148 | * protection here. | |
1149 | */ | |
1150 | cachep->nodelists[node] = l3; | |
1151 | } | |
1152 | ||
1153 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1154 | cachep->nodelists[node]->free_limit = | |
1155 | (1 + nr_cpus_node(node)) * | |
1156 | cachep->batchcount + cachep->num; | |
1157 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1158 | } | |
1159 | return 0; | |
1160 | } | |
1161 | ||
fbf1e473 AM |
1162 | static void __cpuinit cpuup_canceled(long cpu) |
1163 | { | |
1164 | struct kmem_cache *cachep; | |
1165 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1166 | int node = cpu_to_mem(cpu); |
a70f7302 | 1167 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1168 | |
3b0efdfa | 1169 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1170 | struct array_cache *nc; |
1171 | struct array_cache *shared; | |
1172 | struct array_cache **alien; | |
fbf1e473 | 1173 | |
fbf1e473 AM |
1174 | /* cpu is dead; no one can alloc from it. */ |
1175 | nc = cachep->array[cpu]; | |
1176 | cachep->array[cpu] = NULL; | |
1177 | l3 = cachep->nodelists[node]; | |
1178 | ||
1179 | if (!l3) | |
1180 | goto free_array_cache; | |
1181 | ||
1182 | spin_lock_irq(&l3->list_lock); | |
1183 | ||
1184 | /* Free limit for this kmem_list3 */ | |
1185 | l3->free_limit -= cachep->batchcount; | |
1186 | if (nc) | |
1187 | free_block(cachep, nc->entry, nc->avail, node); | |
1188 | ||
58463c1f | 1189 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1190 | spin_unlock_irq(&l3->list_lock); |
1191 | goto free_array_cache; | |
1192 | } | |
1193 | ||
1194 | shared = l3->shared; | |
1195 | if (shared) { | |
1196 | free_block(cachep, shared->entry, | |
1197 | shared->avail, node); | |
1198 | l3->shared = NULL; | |
1199 | } | |
1200 | ||
1201 | alien = l3->alien; | |
1202 | l3->alien = NULL; | |
1203 | ||
1204 | spin_unlock_irq(&l3->list_lock); | |
1205 | ||
1206 | kfree(shared); | |
1207 | if (alien) { | |
1208 | drain_alien_cache(cachep, alien); | |
1209 | free_alien_cache(alien); | |
1210 | } | |
1211 | free_array_cache: | |
1212 | kfree(nc); | |
1213 | } | |
1214 | /* | |
1215 | * In the previous loop, all the objects were freed to | |
1216 | * the respective cache's slabs, now we can go ahead and | |
1217 | * shrink each nodelist to its limit. | |
1218 | */ | |
3b0efdfa | 1219 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1220 | l3 = cachep->nodelists[node]; |
1221 | if (!l3) | |
1222 | continue; | |
1223 | drain_freelist(cachep, l3, l3->free_objects); | |
1224 | } | |
1225 | } | |
1226 | ||
1227 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1228 | { |
343e0d7a | 1229 | struct kmem_cache *cachep; |
e498be7d | 1230 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1231 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1232 | int err; |
1da177e4 | 1233 | |
fbf1e473 AM |
1234 | /* |
1235 | * We need to do this right in the beginning since | |
1236 | * alloc_arraycache's are going to use this list. | |
1237 | * kmalloc_node allows us to add the slab to the right | |
1238 | * kmem_list3 and not this cpu's kmem_list3 | |
1239 | */ | |
8f9f8d9e DR |
1240 | err = init_cache_nodelists_node(node); |
1241 | if (err < 0) | |
1242 | goto bad; | |
fbf1e473 AM |
1243 | |
1244 | /* | |
1245 | * Now we can go ahead with allocating the shared arrays and | |
1246 | * array caches | |
1247 | */ | |
3b0efdfa | 1248 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1249 | struct array_cache *nc; |
1250 | struct array_cache *shared = NULL; | |
1251 | struct array_cache **alien = NULL; | |
1252 | ||
1253 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1254 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1255 | if (!nc) |
1256 | goto bad; | |
1257 | if (cachep->shared) { | |
1258 | shared = alloc_arraycache(node, | |
1259 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1260 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1261 | if (!shared) { |
1262 | kfree(nc); | |
1da177e4 | 1263 | goto bad; |
12d00f6a | 1264 | } |
fbf1e473 AM |
1265 | } |
1266 | if (use_alien_caches) { | |
83b519e8 | 1267 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1268 | if (!alien) { |
1269 | kfree(shared); | |
1270 | kfree(nc); | |
fbf1e473 | 1271 | goto bad; |
12d00f6a | 1272 | } |
fbf1e473 AM |
1273 | } |
1274 | cachep->array[cpu] = nc; | |
1275 | l3 = cachep->nodelists[node]; | |
1276 | BUG_ON(!l3); | |
1277 | ||
1278 | spin_lock_irq(&l3->list_lock); | |
1279 | if (!l3->shared) { | |
1280 | /* | |
1281 | * We are serialised from CPU_DEAD or | |
1282 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1283 | */ | |
1284 | l3->shared = shared; | |
1285 | shared = NULL; | |
1286 | } | |
4484ebf1 | 1287 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1288 | if (!l3->alien) { |
1289 | l3->alien = alien; | |
1290 | alien = NULL; | |
1da177e4 | 1291 | } |
fbf1e473 AM |
1292 | #endif |
1293 | spin_unlock_irq(&l3->list_lock); | |
1294 | kfree(shared); | |
1295 | free_alien_cache(alien); | |
83835b3d PZ |
1296 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1297 | slab_set_debugobj_lock_classes_node(cachep, node); | |
fbf1e473 | 1298 | } |
ce79ddc8 PE |
1299 | init_node_lock_keys(node); |
1300 | ||
fbf1e473 AM |
1301 | return 0; |
1302 | bad: | |
12d00f6a | 1303 | cpuup_canceled(cpu); |
fbf1e473 AM |
1304 | return -ENOMEM; |
1305 | } | |
1306 | ||
1307 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1308 | unsigned long action, void *hcpu) | |
1309 | { | |
1310 | long cpu = (long)hcpu; | |
1311 | int err = 0; | |
1312 | ||
1313 | switch (action) { | |
fbf1e473 AM |
1314 | case CPU_UP_PREPARE: |
1315 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1316 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1317 | err = cpuup_prepare(cpu); |
95402b38 | 1318 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1319 | break; |
1320 | case CPU_ONLINE: | |
8bb78442 | 1321 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1322 | start_cpu_timer(cpu); |
1323 | break; | |
1324 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1325 | case CPU_DOWN_PREPARE: |
8bb78442 | 1326 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1327 | /* |
1328 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1329 | * held so that if cache_reap() is invoked it cannot do | |
1330 | * anything expensive but will only modify reap_work | |
1331 | * and reschedule the timer. | |
1332 | */ | |
afe2c511 | 1333 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1334 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1335 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1336 | break; |
1337 | case CPU_DOWN_FAILED: | |
8bb78442 | 1338 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1339 | start_cpu_timer(cpu); |
1340 | break; | |
1da177e4 | 1341 | case CPU_DEAD: |
8bb78442 | 1342 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1343 | /* |
1344 | * Even if all the cpus of a node are down, we don't free the | |
1345 | * kmem_list3 of any cache. This to avoid a race between | |
1346 | * cpu_down, and a kmalloc allocation from another cpu for | |
1347 | * memory from the node of the cpu going down. The list3 | |
1348 | * structure is usually allocated from kmem_cache_create() and | |
1349 | * gets destroyed at kmem_cache_destroy(). | |
1350 | */ | |
183ff22b | 1351 | /* fall through */ |
8f5be20b | 1352 | #endif |
1da177e4 | 1353 | case CPU_UP_CANCELED: |
8bb78442 | 1354 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1355 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1356 | cpuup_canceled(cpu); |
fc0abb14 | 1357 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1358 | break; |
1da177e4 | 1359 | } |
eac40680 | 1360 | return notifier_from_errno(err); |
1da177e4 LT |
1361 | } |
1362 | ||
74b85f37 CS |
1363 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1364 | &cpuup_callback, NULL, 0 | |
1365 | }; | |
1da177e4 | 1366 | |
8f9f8d9e DR |
1367 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1368 | /* | |
1369 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1370 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1371 | * removed. | |
1372 | * | |
1373 | * Must hold cache_chain_mutex. | |
1374 | */ | |
1375 | static int __meminit drain_cache_nodelists_node(int node) | |
1376 | { | |
1377 | struct kmem_cache *cachep; | |
1378 | int ret = 0; | |
1379 | ||
3b0efdfa | 1380 | list_for_each_entry(cachep, &cache_chain, list) { |
8f9f8d9e DR |
1381 | struct kmem_list3 *l3; |
1382 | ||
1383 | l3 = cachep->nodelists[node]; | |
1384 | if (!l3) | |
1385 | continue; | |
1386 | ||
1387 | drain_freelist(cachep, l3, l3->free_objects); | |
1388 | ||
1389 | if (!list_empty(&l3->slabs_full) || | |
1390 | !list_empty(&l3->slabs_partial)) { | |
1391 | ret = -EBUSY; | |
1392 | break; | |
1393 | } | |
1394 | } | |
1395 | return ret; | |
1396 | } | |
1397 | ||
1398 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1399 | unsigned long action, void *arg) | |
1400 | { | |
1401 | struct memory_notify *mnb = arg; | |
1402 | int ret = 0; | |
1403 | int nid; | |
1404 | ||
1405 | nid = mnb->status_change_nid; | |
1406 | if (nid < 0) | |
1407 | goto out; | |
1408 | ||
1409 | switch (action) { | |
1410 | case MEM_GOING_ONLINE: | |
1411 | mutex_lock(&cache_chain_mutex); | |
1412 | ret = init_cache_nodelists_node(nid); | |
1413 | mutex_unlock(&cache_chain_mutex); | |
1414 | break; | |
1415 | case MEM_GOING_OFFLINE: | |
1416 | mutex_lock(&cache_chain_mutex); | |
1417 | ret = drain_cache_nodelists_node(nid); | |
1418 | mutex_unlock(&cache_chain_mutex); | |
1419 | break; | |
1420 | case MEM_ONLINE: | |
1421 | case MEM_OFFLINE: | |
1422 | case MEM_CANCEL_ONLINE: | |
1423 | case MEM_CANCEL_OFFLINE: | |
1424 | break; | |
1425 | } | |
1426 | out: | |
5fda1bd5 | 1427 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1428 | } |
1429 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1430 | ||
e498be7d CL |
1431 | /* |
1432 | * swap the static kmem_list3 with kmalloced memory | |
1433 | */ | |
8f9f8d9e DR |
1434 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1435 | int nodeid) | |
e498be7d CL |
1436 | { |
1437 | struct kmem_list3 *ptr; | |
1438 | ||
83b519e8 | 1439 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1440 | BUG_ON(!ptr); |
1441 | ||
e498be7d | 1442 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1443 | /* |
1444 | * Do not assume that spinlocks can be initialized via memcpy: | |
1445 | */ | |
1446 | spin_lock_init(&ptr->list_lock); | |
1447 | ||
e498be7d CL |
1448 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1449 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1450 | } |
1451 | ||
556a169d PE |
1452 | /* |
1453 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1454 | * size of kmem_list3. | |
1455 | */ | |
1456 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1457 | { | |
1458 | int node; | |
1459 | ||
1460 | for_each_online_node(node) { | |
1461 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1462 | cachep->nodelists[node]->next_reap = jiffies + | |
1463 | REAPTIMEOUT_LIST3 + | |
1464 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1465 | } | |
1466 | } | |
1467 | ||
a737b3e2 AM |
1468 | /* |
1469 | * Initialisation. Called after the page allocator have been initialised and | |
1470 | * before smp_init(). | |
1da177e4 LT |
1471 | */ |
1472 | void __init kmem_cache_init(void) | |
1473 | { | |
1474 | size_t left_over; | |
1475 | struct cache_sizes *sizes; | |
1476 | struct cache_names *names; | |
e498be7d | 1477 | int i; |
07ed76b2 | 1478 | int order; |
1ca4cb24 | 1479 | int node; |
e498be7d | 1480 | |
b6e68bc1 | 1481 | if (num_possible_nodes() == 1) |
62918a03 SS |
1482 | use_alien_caches = 0; |
1483 | ||
e498be7d CL |
1484 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1485 | kmem_list3_init(&initkmem_list3[i]); | |
1486 | if (i < MAX_NUMNODES) | |
1487 | cache_cache.nodelists[i] = NULL; | |
1488 | } | |
556a169d | 1489 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1490 | |
1491 | /* | |
1492 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1493 | * page orders on machines with more than 32MB of memory if |
1494 | * not overridden on the command line. | |
1da177e4 | 1495 | */ |
3df1cccd | 1496 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1497 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1498 | |
1da177e4 LT |
1499 | /* Bootstrap is tricky, because several objects are allocated |
1500 | * from caches that do not exist yet: | |
a737b3e2 AM |
1501 | * 1) initialize the cache_cache cache: it contains the struct |
1502 | * kmem_cache structures of all caches, except cache_cache itself: | |
1503 | * cache_cache is statically allocated. | |
e498be7d CL |
1504 | * Initially an __init data area is used for the head array and the |
1505 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1506 | * array at the end of the bootstrap. | |
1da177e4 | 1507 | * 2) Create the first kmalloc cache. |
343e0d7a | 1508 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1509 | * An __init data area is used for the head array. |
1510 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1511 | * head arrays. | |
1da177e4 LT |
1512 | * 4) Replace the __init data head arrays for cache_cache and the first |
1513 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1514 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1515 | * the other cache's with kmalloc allocated memory. | |
1516 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1517 | */ |
1518 | ||
7d6e6d09 | 1519 | node = numa_mem_id(); |
1ca4cb24 | 1520 | |
1da177e4 | 1521 | /* 1) create the cache_cache */ |
1da177e4 | 1522 | INIT_LIST_HEAD(&cache_chain); |
3b0efdfa | 1523 | list_add(&cache_cache.list, &cache_chain); |
1da177e4 LT |
1524 | cache_cache.colour_off = cache_line_size(); |
1525 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1526 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1527 | |
8da3430d | 1528 | /* |
b56efcf0 | 1529 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1530 | */ |
3b0efdfa | 1531 | cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
b56efcf0 | 1532 | nr_node_ids * sizeof(struct kmem_list3 *); |
3b0efdfa CL |
1533 | cache_cache.object_size = cache_cache.size; |
1534 | cache_cache.size = ALIGN(cache_cache.size, | |
a737b3e2 | 1535 | cache_line_size()); |
6a2d7a95 | 1536 | cache_cache.reciprocal_buffer_size = |
3b0efdfa | 1537 | reciprocal_value(cache_cache.size); |
1da177e4 | 1538 | |
07ed76b2 | 1539 | for (order = 0; order < MAX_ORDER; order++) { |
3b0efdfa | 1540 | cache_estimate(order, cache_cache.size, |
07ed76b2 JS |
1541 | cache_line_size(), 0, &left_over, &cache_cache.num); |
1542 | if (cache_cache.num) | |
1543 | break; | |
1544 | } | |
40094fa6 | 1545 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1546 | cache_cache.gfporder = order; |
b28a02de | 1547 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1548 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1549 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1550 | |
1551 | /* 2+3) create the kmalloc caches */ | |
1552 | sizes = malloc_sizes; | |
1553 | names = cache_names; | |
1554 | ||
a737b3e2 AM |
1555 | /* |
1556 | * Initialize the caches that provide memory for the array cache and the | |
1557 | * kmem_list3 structures first. Without this, further allocations will | |
1558 | * bug. | |
e498be7d CL |
1559 | */ |
1560 | ||
1561 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1562 | sizes[INDEX_AC].cs_size, |
1563 | ARCH_KMALLOC_MINALIGN, | |
1564 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1565 | NULL); |
e498be7d | 1566 | |
a737b3e2 | 1567 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1568 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1569 | kmem_cache_create(names[INDEX_L3].name, |
1570 | sizes[INDEX_L3].cs_size, | |
1571 | ARCH_KMALLOC_MINALIGN, | |
1572 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1573 | NULL); |
a737b3e2 | 1574 | } |
e498be7d | 1575 | |
e0a42726 IM |
1576 | slab_early_init = 0; |
1577 | ||
1da177e4 | 1578 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1579 | /* |
1580 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1581 | * This should be particularly beneficial on SMP boxes, as it |
1582 | * eliminates "false sharing". | |
1583 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1584 | * allow tighter packing of the smaller caches. |
1585 | */ | |
a737b3e2 | 1586 | if (!sizes->cs_cachep) { |
e498be7d | 1587 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1588 | sizes->cs_size, |
1589 | ARCH_KMALLOC_MINALIGN, | |
1590 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1591 | NULL); |
a737b3e2 | 1592 | } |
4b51d669 CL |
1593 | #ifdef CONFIG_ZONE_DMA |
1594 | sizes->cs_dmacachep = kmem_cache_create( | |
1595 | names->name_dma, | |
a737b3e2 AM |
1596 | sizes->cs_size, |
1597 | ARCH_KMALLOC_MINALIGN, | |
1598 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1599 | SLAB_PANIC, | |
20c2df83 | 1600 | NULL); |
4b51d669 | 1601 | #endif |
1da177e4 LT |
1602 | sizes++; |
1603 | names++; | |
1604 | } | |
1605 | /* 4) Replace the bootstrap head arrays */ | |
1606 | { | |
2b2d5493 | 1607 | struct array_cache *ptr; |
e498be7d | 1608 | |
83b519e8 | 1609 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1610 | |
9a2dba4b PE |
1611 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1612 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1613 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1614 | /* |
1615 | * Do not assume that spinlocks can be initialized via memcpy: | |
1616 | */ | |
1617 | spin_lock_init(&ptr->lock); | |
1618 | ||
1da177e4 | 1619 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1620 | |
83b519e8 | 1621 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1622 | |
9a2dba4b | 1623 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1624 | != &initarray_generic.cache); |
9a2dba4b | 1625 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1626 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1627 | /* |
1628 | * Do not assume that spinlocks can be initialized via memcpy: | |
1629 | */ | |
1630 | spin_lock_init(&ptr->lock); | |
1631 | ||
e498be7d | 1632 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1633 | ptr; |
1da177e4 | 1634 | } |
e498be7d CL |
1635 | /* 5) Replace the bootstrap kmem_list3's */ |
1636 | { | |
1ca4cb24 PE |
1637 | int nid; |
1638 | ||
9c09a95c | 1639 | for_each_online_node(nid) { |
ec1f5eee | 1640 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1641 | |
e498be7d | 1642 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1643 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1644 | |
1645 | if (INDEX_AC != INDEX_L3) { | |
1646 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1647 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1648 | } |
1649 | } | |
1650 | } | |
1da177e4 | 1651 | |
8429db5c | 1652 | g_cpucache_up = EARLY; |
8429db5c PE |
1653 | } |
1654 | ||
1655 | void __init kmem_cache_init_late(void) | |
1656 | { | |
1657 | struct kmem_cache *cachep; | |
1658 | ||
52cef189 PZ |
1659 | g_cpucache_up = LATE; |
1660 | ||
30765b92 PZ |
1661 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1662 | init_lock_keys(); | |
1663 | ||
8429db5c PE |
1664 | /* 6) resize the head arrays to their final sizes */ |
1665 | mutex_lock(&cache_chain_mutex); | |
3b0efdfa | 1666 | list_for_each_entry(cachep, &cache_chain, list) |
8429db5c PE |
1667 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1668 | BUG(); | |
1669 | mutex_unlock(&cache_chain_mutex); | |
056c6241 | 1670 | |
1da177e4 LT |
1671 | /* Done! */ |
1672 | g_cpucache_up = FULL; | |
1673 | ||
a737b3e2 AM |
1674 | /* |
1675 | * Register a cpu startup notifier callback that initializes | |
1676 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1677 | */ |
1678 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1679 | |
8f9f8d9e DR |
1680 | #ifdef CONFIG_NUMA |
1681 | /* | |
1682 | * Register a memory hotplug callback that initializes and frees | |
1683 | * nodelists. | |
1684 | */ | |
1685 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1686 | #endif | |
1687 | ||
a737b3e2 AM |
1688 | /* |
1689 | * The reap timers are started later, with a module init call: That part | |
1690 | * of the kernel is not yet operational. | |
1da177e4 LT |
1691 | */ |
1692 | } | |
1693 | ||
1694 | static int __init cpucache_init(void) | |
1695 | { | |
1696 | int cpu; | |
1697 | ||
a737b3e2 AM |
1698 | /* |
1699 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1700 | */ |
e498be7d | 1701 | for_each_online_cpu(cpu) |
a737b3e2 | 1702 | start_cpu_timer(cpu); |
1da177e4 LT |
1703 | return 0; |
1704 | } | |
1da177e4 LT |
1705 | __initcall(cpucache_init); |
1706 | ||
8bdec192 RA |
1707 | static noinline void |
1708 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1709 | { | |
1710 | struct kmem_list3 *l3; | |
1711 | struct slab *slabp; | |
1712 | unsigned long flags; | |
1713 | int node; | |
1714 | ||
1715 | printk(KERN_WARNING | |
1716 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1717 | nodeid, gfpflags); | |
1718 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1719 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1720 | |
1721 | for_each_online_node(node) { | |
1722 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1723 | unsigned long active_slabs = 0, num_slabs = 0; | |
1724 | ||
1725 | l3 = cachep->nodelists[node]; | |
1726 | if (!l3) | |
1727 | continue; | |
1728 | ||
1729 | spin_lock_irqsave(&l3->list_lock, flags); | |
1730 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1731 | active_objs += cachep->num; | |
1732 | active_slabs++; | |
1733 | } | |
1734 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1735 | active_objs += slabp->inuse; | |
1736 | active_slabs++; | |
1737 | } | |
1738 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1739 | num_slabs++; | |
1740 | ||
1741 | free_objects += l3->free_objects; | |
1742 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1743 | ||
1744 | num_slabs += active_slabs; | |
1745 | num_objs = num_slabs * cachep->num; | |
1746 | printk(KERN_WARNING | |
1747 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1748 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1749 | free_objects); | |
1750 | } | |
1751 | } | |
1752 | ||
1da177e4 LT |
1753 | /* |
1754 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1755 | * | |
1756 | * If we requested dmaable memory, we will get it. Even if we | |
1757 | * did not request dmaable memory, we might get it, but that | |
1758 | * would be relatively rare and ignorable. | |
1759 | */ | |
343e0d7a | 1760 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1761 | { |
1762 | struct page *page; | |
e1b6aa6f | 1763 | int nr_pages; |
1da177e4 LT |
1764 | int i; |
1765 | ||
d6fef9da | 1766 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1767 | /* |
1768 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1769 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1770 | */ |
e1b6aa6f | 1771 | flags |= __GFP_COMP; |
d6fef9da | 1772 | #endif |
765c4507 | 1773 | |
3c517a61 | 1774 | flags |= cachep->gfpflags; |
e12ba74d MG |
1775 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1776 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1777 | |
517d0869 | 1778 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1779 | if (!page) { |
1780 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1781 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1782 | return NULL; |
8bdec192 | 1783 | } |
1da177e4 | 1784 | |
e1b6aa6f | 1785 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1786 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1787 | add_zone_page_state(page_zone(page), |
1788 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1789 | else | |
1790 | add_zone_page_state(page_zone(page), | |
1791 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1792 | for (i = 0; i < nr_pages; i++) |
1793 | __SetPageSlab(page + i); | |
c175eea4 | 1794 | |
b1eeab67 VN |
1795 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1796 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1797 | ||
1798 | if (cachep->ctor) | |
1799 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1800 | else | |
1801 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1802 | } | |
c175eea4 | 1803 | |
e1b6aa6f | 1804 | return page_address(page); |
1da177e4 LT |
1805 | } |
1806 | ||
1807 | /* | |
1808 | * Interface to system's page release. | |
1809 | */ | |
343e0d7a | 1810 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1811 | { |
b28a02de | 1812 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1813 | struct page *page = virt_to_page(addr); |
1814 | const unsigned long nr_freed = i; | |
1815 | ||
b1eeab67 | 1816 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1817 | |
972d1a7b CL |
1818 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1819 | sub_zone_page_state(page_zone(page), | |
1820 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1821 | else | |
1822 | sub_zone_page_state(page_zone(page), | |
1823 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1824 | while (i--) { |
f205b2fe NP |
1825 | BUG_ON(!PageSlab(page)); |
1826 | __ClearPageSlab(page); | |
1da177e4 LT |
1827 | page++; |
1828 | } | |
1da177e4 LT |
1829 | if (current->reclaim_state) |
1830 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1831 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1832 | } |
1833 | ||
1834 | static void kmem_rcu_free(struct rcu_head *head) | |
1835 | { | |
b28a02de | 1836 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1837 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1838 | |
1839 | kmem_freepages(cachep, slab_rcu->addr); | |
1840 | if (OFF_SLAB(cachep)) | |
1841 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1842 | } | |
1843 | ||
1844 | #if DEBUG | |
1845 | ||
1846 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1847 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1848 | unsigned long caller) |
1da177e4 | 1849 | { |
8c138bc0 | 1850 | int size = cachep->object_size; |
1da177e4 | 1851 | |
3dafccf2 | 1852 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1853 | |
b28a02de | 1854 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1855 | return; |
1856 | ||
b28a02de PE |
1857 | *addr++ = 0x12345678; |
1858 | *addr++ = caller; | |
1859 | *addr++ = smp_processor_id(); | |
1860 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1861 | { |
1862 | unsigned long *sptr = &caller; | |
1863 | unsigned long svalue; | |
1864 | ||
1865 | while (!kstack_end(sptr)) { | |
1866 | svalue = *sptr++; | |
1867 | if (kernel_text_address(svalue)) { | |
b28a02de | 1868 | *addr++ = svalue; |
1da177e4 LT |
1869 | size -= sizeof(unsigned long); |
1870 | if (size <= sizeof(unsigned long)) | |
1871 | break; | |
1872 | } | |
1873 | } | |
1874 | ||
1875 | } | |
b28a02de | 1876 | *addr++ = 0x87654321; |
1da177e4 LT |
1877 | } |
1878 | #endif | |
1879 | ||
343e0d7a | 1880 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1881 | { |
8c138bc0 | 1882 | int size = cachep->object_size; |
3dafccf2 | 1883 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1884 | |
1885 | memset(addr, val, size); | |
b28a02de | 1886 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1887 | } |
1888 | ||
1889 | static void dump_line(char *data, int offset, int limit) | |
1890 | { | |
1891 | int i; | |
aa83aa40 DJ |
1892 | unsigned char error = 0; |
1893 | int bad_count = 0; | |
1894 | ||
fdde6abb | 1895 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
1896 | for (i = 0; i < limit; i++) { |
1897 | if (data[offset + i] != POISON_FREE) { | |
1898 | error = data[offset + i]; | |
1899 | bad_count++; | |
1900 | } | |
aa83aa40 | 1901 | } |
fdde6abb SAS |
1902 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1903 | &data[offset], limit, 1); | |
aa83aa40 DJ |
1904 | |
1905 | if (bad_count == 1) { | |
1906 | error ^= POISON_FREE; | |
1907 | if (!(error & (error - 1))) { | |
1908 | printk(KERN_ERR "Single bit error detected. Probably " | |
1909 | "bad RAM.\n"); | |
1910 | #ifdef CONFIG_X86 | |
1911 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1912 | "test tool.\n"); | |
1913 | #else | |
1914 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1915 | #endif | |
1916 | } | |
1917 | } | |
1da177e4 LT |
1918 | } |
1919 | #endif | |
1920 | ||
1921 | #if DEBUG | |
1922 | ||
343e0d7a | 1923 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1924 | { |
1925 | int i, size; | |
1926 | char *realobj; | |
1927 | ||
1928 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1929 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1930 | *dbg_redzone1(cachep, objp), |
1931 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1932 | } |
1933 | ||
1934 | if (cachep->flags & SLAB_STORE_USER) { | |
1935 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1936 | *dbg_userword(cachep, objp)); |
1da177e4 | 1937 | print_symbol("(%s)", |
a737b3e2 | 1938 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1939 | printk("\n"); |
1940 | } | |
3dafccf2 | 1941 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 1942 | size = cachep->object_size; |
b28a02de | 1943 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1944 | int limit; |
1945 | limit = 16; | |
b28a02de PE |
1946 | if (i + limit > size) |
1947 | limit = size - i; | |
1da177e4 LT |
1948 | dump_line(realobj, i, limit); |
1949 | } | |
1950 | } | |
1951 | ||
343e0d7a | 1952 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1953 | { |
1954 | char *realobj; | |
1955 | int size, i; | |
1956 | int lines = 0; | |
1957 | ||
3dafccf2 | 1958 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 1959 | size = cachep->object_size; |
1da177e4 | 1960 | |
b28a02de | 1961 | for (i = 0; i < size; i++) { |
1da177e4 | 1962 | char exp = POISON_FREE; |
b28a02de | 1963 | if (i == size - 1) |
1da177e4 LT |
1964 | exp = POISON_END; |
1965 | if (realobj[i] != exp) { | |
1966 | int limit; | |
1967 | /* Mismatch ! */ | |
1968 | /* Print header */ | |
1969 | if (lines == 0) { | |
b28a02de | 1970 | printk(KERN_ERR |
face37f5 DJ |
1971 | "Slab corruption (%s): %s start=%p, len=%d\n", |
1972 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
1973 | print_objinfo(cachep, objp, 0); |
1974 | } | |
1975 | /* Hexdump the affected line */ | |
b28a02de | 1976 | i = (i / 16) * 16; |
1da177e4 | 1977 | limit = 16; |
b28a02de PE |
1978 | if (i + limit > size) |
1979 | limit = size - i; | |
1da177e4 LT |
1980 | dump_line(realobj, i, limit); |
1981 | i += 16; | |
1982 | lines++; | |
1983 | /* Limit to 5 lines */ | |
1984 | if (lines > 5) | |
1985 | break; | |
1986 | } | |
1987 | } | |
1988 | if (lines != 0) { | |
1989 | /* Print some data about the neighboring objects, if they | |
1990 | * exist: | |
1991 | */ | |
6ed5eb22 | 1992 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1993 | unsigned int objnr; |
1da177e4 | 1994 | |
8fea4e96 | 1995 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1996 | if (objnr) { |
8fea4e96 | 1997 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1998 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1999 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2000 | realobj, size); |
1da177e4 LT |
2001 | print_objinfo(cachep, objp, 2); |
2002 | } | |
b28a02de | 2003 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2004 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2005 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2006 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2007 | realobj, size); |
1da177e4 LT |
2008 | print_objinfo(cachep, objp, 2); |
2009 | } | |
2010 | } | |
2011 | } | |
2012 | #endif | |
2013 | ||
12dd36fa | 2014 | #if DEBUG |
e79aec29 | 2015 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2016 | { |
1da177e4 LT |
2017 | int i; |
2018 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2019 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2020 | |
2021 | if (cachep->flags & SLAB_POISON) { | |
2022 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2023 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2024 | OFF_SLAB(cachep)) |
b28a02de | 2025 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2026 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2027 | else |
2028 | check_poison_obj(cachep, objp); | |
2029 | #else | |
2030 | check_poison_obj(cachep, objp); | |
2031 | #endif | |
2032 | } | |
2033 | if (cachep->flags & SLAB_RED_ZONE) { | |
2034 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2035 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2036 | "was overwritten"); |
1da177e4 LT |
2037 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2038 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2039 | "was overwritten"); |
1da177e4 | 2040 | } |
1da177e4 | 2041 | } |
12dd36fa | 2042 | } |
1da177e4 | 2043 | #else |
e79aec29 | 2044 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2045 | { |
12dd36fa | 2046 | } |
1da177e4 LT |
2047 | #endif |
2048 | ||
911851e6 RD |
2049 | /** |
2050 | * slab_destroy - destroy and release all objects in a slab | |
2051 | * @cachep: cache pointer being destroyed | |
2052 | * @slabp: slab pointer being destroyed | |
2053 | * | |
12dd36fa | 2054 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2055 | * Before calling the slab must have been unlinked from the cache. The |
2056 | * cache-lock is not held/needed. | |
12dd36fa | 2057 | */ |
343e0d7a | 2058 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2059 | { |
2060 | void *addr = slabp->s_mem - slabp->colouroff; | |
2061 | ||
e79aec29 | 2062 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2063 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2064 | struct slab_rcu *slab_rcu; | |
2065 | ||
b28a02de | 2066 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2067 | slab_rcu->cachep = cachep; |
2068 | slab_rcu->addr = addr; | |
2069 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2070 | } else { | |
2071 | kmem_freepages(cachep, addr); | |
873623df IM |
2072 | if (OFF_SLAB(cachep)) |
2073 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2074 | } |
2075 | } | |
2076 | ||
117f6eb1 CL |
2077 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
2078 | { | |
2079 | int i; | |
2080 | struct kmem_list3 *l3; | |
2081 | ||
2082 | for_each_online_cpu(i) | |
2083 | kfree(cachep->array[i]); | |
2084 | ||
2085 | /* NUMA: free the list3 structures */ | |
2086 | for_each_online_node(i) { | |
2087 | l3 = cachep->nodelists[i]; | |
2088 | if (l3) { | |
2089 | kfree(l3->shared); | |
2090 | free_alien_cache(l3->alien); | |
2091 | kfree(l3); | |
2092 | } | |
2093 | } | |
2094 | kmem_cache_free(&cache_cache, cachep); | |
2095 | } | |
2096 | ||
2097 | ||
4d268eba | 2098 | /** |
a70773dd RD |
2099 | * calculate_slab_order - calculate size (page order) of slabs |
2100 | * @cachep: pointer to the cache that is being created | |
2101 | * @size: size of objects to be created in this cache. | |
2102 | * @align: required alignment for the objects. | |
2103 | * @flags: slab allocation flags | |
2104 | * | |
2105 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2106 | * |
2107 | * This could be made much more intelligent. For now, try to avoid using | |
2108 | * high order pages for slabs. When the gfp() functions are more friendly | |
2109 | * towards high-order requests, this should be changed. | |
2110 | */ | |
a737b3e2 | 2111 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2112 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2113 | { |
b1ab41c4 | 2114 | unsigned long offslab_limit; |
4d268eba | 2115 | size_t left_over = 0; |
9888e6fa | 2116 | int gfporder; |
4d268eba | 2117 | |
0aa817f0 | 2118 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2119 | unsigned int num; |
2120 | size_t remainder; | |
2121 | ||
9888e6fa | 2122 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2123 | if (!num) |
2124 | continue; | |
9888e6fa | 2125 | |
b1ab41c4 IM |
2126 | if (flags & CFLGS_OFF_SLAB) { |
2127 | /* | |
2128 | * Max number of objs-per-slab for caches which | |
2129 | * use off-slab slabs. Needed to avoid a possible | |
2130 | * looping condition in cache_grow(). | |
2131 | */ | |
2132 | offslab_limit = size - sizeof(struct slab); | |
2133 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2134 | ||
2135 | if (num > offslab_limit) | |
2136 | break; | |
2137 | } | |
4d268eba | 2138 | |
9888e6fa | 2139 | /* Found something acceptable - save it away */ |
4d268eba | 2140 | cachep->num = num; |
9888e6fa | 2141 | cachep->gfporder = gfporder; |
4d268eba PE |
2142 | left_over = remainder; |
2143 | ||
f78bb8ad LT |
2144 | /* |
2145 | * A VFS-reclaimable slab tends to have most allocations | |
2146 | * as GFP_NOFS and we really don't want to have to be allocating | |
2147 | * higher-order pages when we are unable to shrink dcache. | |
2148 | */ | |
2149 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2150 | break; | |
2151 | ||
4d268eba PE |
2152 | /* |
2153 | * Large number of objects is good, but very large slabs are | |
2154 | * currently bad for the gfp()s. | |
2155 | */ | |
543585cc | 2156 | if (gfporder >= slab_max_order) |
4d268eba PE |
2157 | break; |
2158 | ||
9888e6fa LT |
2159 | /* |
2160 | * Acceptable internal fragmentation? | |
2161 | */ | |
a737b3e2 | 2162 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2163 | break; |
2164 | } | |
2165 | return left_over; | |
2166 | } | |
2167 | ||
83b519e8 | 2168 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2169 | { |
2ed3a4ef | 2170 | if (g_cpucache_up == FULL) |
83b519e8 | 2171 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2172 | |
f30cf7d1 PE |
2173 | if (g_cpucache_up == NONE) { |
2174 | /* | |
2175 | * Note: the first kmem_cache_create must create the cache | |
2176 | * that's used by kmalloc(24), otherwise the creation of | |
2177 | * further caches will BUG(). | |
2178 | */ | |
2179 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2180 | ||
2181 | /* | |
2182 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2183 | * the first cache, then we need to set up all its list3s, | |
2184 | * otherwise the creation of further caches will BUG(). | |
2185 | */ | |
2186 | set_up_list3s(cachep, SIZE_AC); | |
2187 | if (INDEX_AC == INDEX_L3) | |
2188 | g_cpucache_up = PARTIAL_L3; | |
2189 | else | |
2190 | g_cpucache_up = PARTIAL_AC; | |
2191 | } else { | |
2192 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2193 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 PE |
2194 | |
2195 | if (g_cpucache_up == PARTIAL_AC) { | |
2196 | set_up_list3s(cachep, SIZE_L3); | |
2197 | g_cpucache_up = PARTIAL_L3; | |
2198 | } else { | |
2199 | int node; | |
556a169d | 2200 | for_each_online_node(node) { |
f30cf7d1 PE |
2201 | cachep->nodelists[node] = |
2202 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2203 | gfp, node); |
f30cf7d1 PE |
2204 | BUG_ON(!cachep->nodelists[node]); |
2205 | kmem_list3_init(cachep->nodelists[node]); | |
2206 | } | |
2207 | } | |
2208 | } | |
7d6e6d09 | 2209 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2210 | jiffies + REAPTIMEOUT_LIST3 + |
2211 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2212 | ||
2213 | cpu_cache_get(cachep)->avail = 0; | |
2214 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2215 | cpu_cache_get(cachep)->batchcount = 1; | |
2216 | cpu_cache_get(cachep)->touched = 0; | |
2217 | cachep->batchcount = 1; | |
2218 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2219 | return 0; |
f30cf7d1 PE |
2220 | } |
2221 | ||
1da177e4 LT |
2222 | /** |
2223 | * kmem_cache_create - Create a cache. | |
2224 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2225 | * @size: The size of objects to be created in this cache. | |
2226 | * @align: The required alignment for the objects. | |
2227 | * @flags: SLAB flags | |
2228 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2229 | * |
2230 | * Returns a ptr to the cache on success, NULL on failure. | |
2231 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2232 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2233 | * |
2234 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2235 | * the module calling this has to destroy the cache before getting unloaded. |
2236 | * | |
1da177e4 LT |
2237 | * The flags are |
2238 | * | |
2239 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2240 | * to catch references to uninitialised memory. | |
2241 | * | |
2242 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2243 | * for buffer overruns. | |
2244 | * | |
1da177e4 LT |
2245 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2246 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2247 | * as davem. | |
2248 | */ | |
343e0d7a | 2249 | struct kmem_cache * |
1da177e4 | 2250 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2251 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2252 | { |
2253 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2254 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2255 | gfp_t gfp; |
1da177e4 LT |
2256 | |
2257 | /* | |
2258 | * Sanity checks... these are all serious usage bugs. | |
2259 | */ | |
a737b3e2 | 2260 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2261 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2262 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2263 | name); |
b28a02de PE |
2264 | BUG(); |
2265 | } | |
1da177e4 | 2266 | |
f0188f47 | 2267 | /* |
8f5be20b | 2268 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2269 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2270 | */ |
83b519e8 PE |
2271 | if (slab_is_available()) { |
2272 | get_online_cpus(); | |
2273 | mutex_lock(&cache_chain_mutex); | |
2274 | } | |
4f12bb4f | 2275 | |
3b0efdfa | 2276 | list_for_each_entry(pc, &cache_chain, list) { |
4f12bb4f AM |
2277 | char tmp; |
2278 | int res; | |
2279 | ||
2280 | /* | |
2281 | * This happens when the module gets unloaded and doesn't | |
2282 | * destroy its slab cache and no-one else reuses the vmalloc | |
2283 | * area of the module. Print a warning. | |
2284 | */ | |
138ae663 | 2285 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2286 | if (res) { |
b4169525 | 2287 | printk(KERN_ERR |
2288 | "SLAB: cache with size %d has lost its name\n", | |
3b0efdfa | 2289 | pc->size); |
4f12bb4f AM |
2290 | continue; |
2291 | } | |
2292 | ||
b28a02de | 2293 | if (!strcmp(pc->name, name)) { |
b4169525 | 2294 | printk(KERN_ERR |
2295 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2296 | dump_stack(); |
2297 | goto oops; | |
2298 | } | |
2299 | } | |
2300 | ||
1da177e4 LT |
2301 | #if DEBUG |
2302 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2303 | #if FORCED_DEBUG |
2304 | /* | |
2305 | * Enable redzoning and last user accounting, except for caches with | |
2306 | * large objects, if the increased size would increase the object size | |
2307 | * above the next power of two: caches with object sizes just above a | |
2308 | * power of two have a significant amount of internal fragmentation. | |
2309 | */ | |
87a927c7 DW |
2310 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2311 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2312 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2313 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2314 | flags |= SLAB_POISON; | |
2315 | #endif | |
2316 | if (flags & SLAB_DESTROY_BY_RCU) | |
2317 | BUG_ON(flags & SLAB_POISON); | |
2318 | #endif | |
1da177e4 | 2319 | /* |
a737b3e2 AM |
2320 | * Always checks flags, a caller might be expecting debug support which |
2321 | * isn't available. | |
1da177e4 | 2322 | */ |
40094fa6 | 2323 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2324 | |
a737b3e2 AM |
2325 | /* |
2326 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2327 | * unaligned accesses for some archs when redzoning is used, and makes |
2328 | * sure any on-slab bufctl's are also correctly aligned. | |
2329 | */ | |
b28a02de PE |
2330 | if (size & (BYTES_PER_WORD - 1)) { |
2331 | size += (BYTES_PER_WORD - 1); | |
2332 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2333 | } |
2334 | ||
a737b3e2 AM |
2335 | /* calculate the final buffer alignment: */ |
2336 | ||
1da177e4 LT |
2337 | /* 1) arch recommendation: can be overridden for debug */ |
2338 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2339 | /* |
2340 | * Default alignment: as specified by the arch code. Except if | |
2341 | * an object is really small, then squeeze multiple objects into | |
2342 | * one cacheline. | |
1da177e4 LT |
2343 | */ |
2344 | ralign = cache_line_size(); | |
b28a02de | 2345 | while (size <= ralign / 2) |
1da177e4 LT |
2346 | ralign /= 2; |
2347 | } else { | |
2348 | ralign = BYTES_PER_WORD; | |
2349 | } | |
ca5f9703 PE |
2350 | |
2351 | /* | |
87a927c7 DW |
2352 | * Redzoning and user store require word alignment or possibly larger. |
2353 | * Note this will be overridden by architecture or caller mandated | |
2354 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2355 | */ |
87a927c7 DW |
2356 | if (flags & SLAB_STORE_USER) |
2357 | ralign = BYTES_PER_WORD; | |
2358 | ||
2359 | if (flags & SLAB_RED_ZONE) { | |
2360 | ralign = REDZONE_ALIGN; | |
2361 | /* If redzoning, ensure that the second redzone is suitably | |
2362 | * aligned, by adjusting the object size accordingly. */ | |
2363 | size += REDZONE_ALIGN - 1; | |
2364 | size &= ~(REDZONE_ALIGN - 1); | |
2365 | } | |
ca5f9703 | 2366 | |
a44b56d3 | 2367 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2368 | if (ralign < ARCH_SLAB_MINALIGN) { |
2369 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2370 | } |
a44b56d3 | 2371 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2372 | if (ralign < align) { |
2373 | ralign = align; | |
1da177e4 | 2374 | } |
3ff84a7f PE |
2375 | /* disable debug if necessary */ |
2376 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2377 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2378 | /* |
ca5f9703 | 2379 | * 4) Store it. |
1da177e4 LT |
2380 | */ |
2381 | align = ralign; | |
2382 | ||
83b519e8 PE |
2383 | if (slab_is_available()) |
2384 | gfp = GFP_KERNEL; | |
2385 | else | |
2386 | gfp = GFP_NOWAIT; | |
2387 | ||
1da177e4 | 2388 | /* Get cache's description obj. */ |
83b519e8 | 2389 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2390 | if (!cachep) |
4f12bb4f | 2391 | goto oops; |
1da177e4 | 2392 | |
b56efcf0 | 2393 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
3b0efdfa CL |
2394 | cachep->object_size = size; |
2395 | cachep->align = align; | |
1da177e4 | 2396 | #if DEBUG |
1da177e4 | 2397 | |
ca5f9703 PE |
2398 | /* |
2399 | * Both debugging options require word-alignment which is calculated | |
2400 | * into align above. | |
2401 | */ | |
1da177e4 | 2402 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2403 | /* add space for red zone words */ |
3ff84a7f PE |
2404 | cachep->obj_offset += sizeof(unsigned long long); |
2405 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2406 | } |
2407 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2408 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2409 | * the real object. But if the second red zone needs to be |
2410 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2411 | */ |
87a927c7 DW |
2412 | if (flags & SLAB_RED_ZONE) |
2413 | size += REDZONE_ALIGN; | |
2414 | else | |
2415 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2416 | } |
2417 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2418 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3b0efdfa | 2419 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
1ab335d8 | 2420 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); |
1da177e4 LT |
2421 | size = PAGE_SIZE; |
2422 | } | |
2423 | #endif | |
2424 | #endif | |
2425 | ||
e0a42726 IM |
2426 | /* |
2427 | * Determine if the slab management is 'on' or 'off' slab. | |
2428 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2429 | * it too early on. Always use on-slab management when |
2430 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2431 | */ |
e7cb55b9 CM |
2432 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2433 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2434 | /* |
2435 | * Size is large, assume best to place the slab management obj | |
2436 | * off-slab (should allow better packing of objs). | |
2437 | */ | |
2438 | flags |= CFLGS_OFF_SLAB; | |
2439 | ||
2440 | size = ALIGN(size, align); | |
2441 | ||
f78bb8ad | 2442 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2443 | |
2444 | if (!cachep->num) { | |
b4169525 | 2445 | printk(KERN_ERR |
2446 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2447 | kmem_cache_free(&cache_cache, cachep); |
2448 | cachep = NULL; | |
4f12bb4f | 2449 | goto oops; |
1da177e4 | 2450 | } |
b28a02de PE |
2451 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2452 | + sizeof(struct slab), align); | |
1da177e4 LT |
2453 | |
2454 | /* | |
2455 | * If the slab has been placed off-slab, and we have enough space then | |
2456 | * move it on-slab. This is at the expense of any extra colouring. | |
2457 | */ | |
2458 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2459 | flags &= ~CFLGS_OFF_SLAB; | |
2460 | left_over -= slab_size; | |
2461 | } | |
2462 | ||
2463 | if (flags & CFLGS_OFF_SLAB) { | |
2464 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2465 | slab_size = |
2466 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2467 | |
2468 | #ifdef CONFIG_PAGE_POISONING | |
2469 | /* If we're going to use the generic kernel_map_pages() | |
2470 | * poisoning, then it's going to smash the contents of | |
2471 | * the redzone and userword anyhow, so switch them off. | |
2472 | */ | |
2473 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2474 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2475 | #endif | |
1da177e4 LT |
2476 | } |
2477 | ||
2478 | cachep->colour_off = cache_line_size(); | |
2479 | /* Offset must be a multiple of the alignment. */ | |
2480 | if (cachep->colour_off < align) | |
2481 | cachep->colour_off = align; | |
b28a02de | 2482 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2483 | cachep->slab_size = slab_size; |
2484 | cachep->flags = flags; | |
2485 | cachep->gfpflags = 0; | |
4b51d669 | 2486 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2487 | cachep->gfpflags |= GFP_DMA; |
3b0efdfa | 2488 | cachep->size = size; |
6a2d7a95 | 2489 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2490 | |
e5ac9c5a | 2491 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2492 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2493 | /* |
2494 | * This is a possibility for one of the malloc_sizes caches. | |
2495 | * But since we go off slab only for object size greater than | |
2496 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2497 | * this should not happen at all. | |
2498 | * But leave a BUG_ON for some lucky dude. | |
2499 | */ | |
6cb8f913 | 2500 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2501 | } |
1da177e4 | 2502 | cachep->ctor = ctor; |
1da177e4 LT |
2503 | cachep->name = name; |
2504 | ||
83b519e8 | 2505 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef CL |
2506 | __kmem_cache_destroy(cachep); |
2507 | cachep = NULL; | |
2508 | goto oops; | |
2509 | } | |
1da177e4 | 2510 | |
83835b3d PZ |
2511 | if (flags & SLAB_DEBUG_OBJECTS) { |
2512 | /* | |
2513 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2514 | * debug_object_activate()->kmem_cache_alloc(). | |
2515 | */ | |
2516 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2517 | ||
2518 | slab_set_debugobj_lock_classes(cachep); | |
2519 | } | |
2520 | ||
1da177e4 | 2521 | /* cache setup completed, link it into the list */ |
3b0efdfa | 2522 | list_add(&cachep->list, &cache_chain); |
a737b3e2 | 2523 | oops: |
1da177e4 LT |
2524 | if (!cachep && (flags & SLAB_PANIC)) |
2525 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2526 | name); |
83b519e8 PE |
2527 | if (slab_is_available()) { |
2528 | mutex_unlock(&cache_chain_mutex); | |
2529 | put_online_cpus(); | |
2530 | } | |
1da177e4 LT |
2531 | return cachep; |
2532 | } | |
2533 | EXPORT_SYMBOL(kmem_cache_create); | |
2534 | ||
2535 | #if DEBUG | |
2536 | static void check_irq_off(void) | |
2537 | { | |
2538 | BUG_ON(!irqs_disabled()); | |
2539 | } | |
2540 | ||
2541 | static void check_irq_on(void) | |
2542 | { | |
2543 | BUG_ON(irqs_disabled()); | |
2544 | } | |
2545 | ||
343e0d7a | 2546 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2547 | { |
2548 | #ifdef CONFIG_SMP | |
2549 | check_irq_off(); | |
7d6e6d09 | 2550 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2551 | #endif |
2552 | } | |
e498be7d | 2553 | |
343e0d7a | 2554 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2555 | { |
2556 | #ifdef CONFIG_SMP | |
2557 | check_irq_off(); | |
2558 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2559 | #endif | |
2560 | } | |
2561 | ||
1da177e4 LT |
2562 | #else |
2563 | #define check_irq_off() do { } while(0) | |
2564 | #define check_irq_on() do { } while(0) | |
2565 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2566 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2567 | #endif |
2568 | ||
aab2207c CL |
2569 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2570 | struct array_cache *ac, | |
2571 | int force, int node); | |
2572 | ||
1da177e4 LT |
2573 | static void do_drain(void *arg) |
2574 | { | |
a737b3e2 | 2575 | struct kmem_cache *cachep = arg; |
1da177e4 | 2576 | struct array_cache *ac; |
7d6e6d09 | 2577 | int node = numa_mem_id(); |
1da177e4 LT |
2578 | |
2579 | check_irq_off(); | |
9a2dba4b | 2580 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2581 | spin_lock(&cachep->nodelists[node]->list_lock); |
2582 | free_block(cachep, ac->entry, ac->avail, node); | |
2583 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2584 | ac->avail = 0; |
2585 | } | |
2586 | ||
343e0d7a | 2587 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2588 | { |
e498be7d CL |
2589 | struct kmem_list3 *l3; |
2590 | int node; | |
2591 | ||
15c8b6c1 | 2592 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2593 | check_irq_on(); |
b28a02de | 2594 | for_each_online_node(node) { |
e498be7d | 2595 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2596 | if (l3 && l3->alien) |
2597 | drain_alien_cache(cachep, l3->alien); | |
2598 | } | |
2599 | ||
2600 | for_each_online_node(node) { | |
2601 | l3 = cachep->nodelists[node]; | |
2602 | if (l3) | |
aab2207c | 2603 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2604 | } |
1da177e4 LT |
2605 | } |
2606 | ||
ed11d9eb CL |
2607 | /* |
2608 | * Remove slabs from the list of free slabs. | |
2609 | * Specify the number of slabs to drain in tofree. | |
2610 | * | |
2611 | * Returns the actual number of slabs released. | |
2612 | */ | |
2613 | static int drain_freelist(struct kmem_cache *cache, | |
2614 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2615 | { |
ed11d9eb CL |
2616 | struct list_head *p; |
2617 | int nr_freed; | |
1da177e4 | 2618 | struct slab *slabp; |
1da177e4 | 2619 | |
ed11d9eb CL |
2620 | nr_freed = 0; |
2621 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2622 | |
ed11d9eb | 2623 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2624 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2625 | if (p == &l3->slabs_free) { |
2626 | spin_unlock_irq(&l3->list_lock); | |
2627 | goto out; | |
2628 | } | |
1da177e4 | 2629 | |
ed11d9eb | 2630 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2631 | #if DEBUG |
40094fa6 | 2632 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2633 | #endif |
2634 | list_del(&slabp->list); | |
ed11d9eb CL |
2635 | /* |
2636 | * Safe to drop the lock. The slab is no longer linked | |
2637 | * to the cache. | |
2638 | */ | |
2639 | l3->free_objects -= cache->num; | |
e498be7d | 2640 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2641 | slab_destroy(cache, slabp); |
2642 | nr_freed++; | |
1da177e4 | 2643 | } |
ed11d9eb CL |
2644 | out: |
2645 | return nr_freed; | |
1da177e4 LT |
2646 | } |
2647 | ||
8f5be20b | 2648 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2649 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2650 | { |
2651 | int ret = 0, i = 0; | |
2652 | struct kmem_list3 *l3; | |
2653 | ||
2654 | drain_cpu_caches(cachep); | |
2655 | ||
2656 | check_irq_on(); | |
2657 | for_each_online_node(i) { | |
2658 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2659 | if (!l3) |
2660 | continue; | |
2661 | ||
2662 | drain_freelist(cachep, l3, l3->free_objects); | |
2663 | ||
2664 | ret += !list_empty(&l3->slabs_full) || | |
2665 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2666 | } |
2667 | return (ret ? 1 : 0); | |
2668 | } | |
2669 | ||
1da177e4 LT |
2670 | /** |
2671 | * kmem_cache_shrink - Shrink a cache. | |
2672 | * @cachep: The cache to shrink. | |
2673 | * | |
2674 | * Releases as many slabs as possible for a cache. | |
2675 | * To help debugging, a zero exit status indicates all slabs were released. | |
2676 | */ | |
343e0d7a | 2677 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2678 | { |
8f5be20b | 2679 | int ret; |
40094fa6 | 2680 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2681 | |
95402b38 | 2682 | get_online_cpus(); |
8f5be20b RT |
2683 | mutex_lock(&cache_chain_mutex); |
2684 | ret = __cache_shrink(cachep); | |
2685 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2686 | put_online_cpus(); |
8f5be20b | 2687 | return ret; |
1da177e4 LT |
2688 | } |
2689 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2690 | ||
2691 | /** | |
2692 | * kmem_cache_destroy - delete a cache | |
2693 | * @cachep: the cache to destroy | |
2694 | * | |
72fd4a35 | 2695 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2696 | * |
2697 | * It is expected this function will be called by a module when it is | |
2698 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2699 | * cache being allocated each time a module is loaded and unloaded, if the | |
2700 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2701 | * | |
2702 | * The cache must be empty before calling this function. | |
2703 | * | |
25985edc | 2704 | * The caller must guarantee that no one will allocate memory from the cache |
1da177e4 LT |
2705 | * during the kmem_cache_destroy(). |
2706 | */ | |
133d205a | 2707 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2708 | { |
40094fa6 | 2709 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2710 | |
1da177e4 | 2711 | /* Find the cache in the chain of caches. */ |
95402b38 | 2712 | get_online_cpus(); |
fc0abb14 | 2713 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2714 | /* |
2715 | * the chain is never empty, cache_cache is never destroyed | |
2716 | */ | |
3b0efdfa | 2717 | list_del(&cachep->list); |
1da177e4 LT |
2718 | if (__cache_shrink(cachep)) { |
2719 | slab_error(cachep, "Can't free all objects"); | |
3b0efdfa | 2720 | list_add(&cachep->list, &cache_chain); |
fc0abb14 | 2721 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2722 | put_online_cpus(); |
133d205a | 2723 | return; |
1da177e4 LT |
2724 | } |
2725 | ||
2726 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
7ed9f7e5 | 2727 | rcu_barrier(); |
1da177e4 | 2728 | |
117f6eb1 | 2729 | __kmem_cache_destroy(cachep); |
8f5be20b | 2730 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2731 | put_online_cpus(); |
1da177e4 LT |
2732 | } |
2733 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2734 | ||
e5ac9c5a RT |
2735 | /* |
2736 | * Get the memory for a slab management obj. | |
2737 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2738 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2739 | * come from the same cache which is getting created because, | |
2740 | * when we are searching for an appropriate cache for these | |
2741 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2742 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2743 | * kmem_find_general_cachep till the initialization is complete. | |
2744 | * Hence we cannot have slabp_cache same as the original cache. | |
2745 | */ | |
343e0d7a | 2746 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2747 | int colour_off, gfp_t local_flags, |
2748 | int nodeid) | |
1da177e4 LT |
2749 | { |
2750 | struct slab *slabp; | |
b28a02de | 2751 | |
1da177e4 LT |
2752 | if (OFF_SLAB(cachep)) { |
2753 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2754 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2755 | local_flags, nodeid); |
d5cff635 CM |
2756 | /* |
2757 | * If the first object in the slab is leaked (it's allocated | |
2758 | * but no one has a reference to it), we want to make sure | |
2759 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2760 | * to the object. Otherwise we will not report the leak. | |
2761 | */ | |
c017b4be CM |
2762 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2763 | local_flags); | |
1da177e4 LT |
2764 | if (!slabp) |
2765 | return NULL; | |
2766 | } else { | |
b28a02de | 2767 | slabp = objp + colour_off; |
1da177e4 LT |
2768 | colour_off += cachep->slab_size; |
2769 | } | |
2770 | slabp->inuse = 0; | |
2771 | slabp->colouroff = colour_off; | |
b28a02de | 2772 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2773 | slabp->nodeid = nodeid; |
e51bfd0a | 2774 | slabp->free = 0; |
1da177e4 LT |
2775 | return slabp; |
2776 | } | |
2777 | ||
2778 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2779 | { | |
b28a02de | 2780 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2781 | } |
2782 | ||
343e0d7a | 2783 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2784 | struct slab *slabp) |
1da177e4 LT |
2785 | { |
2786 | int i; | |
2787 | ||
2788 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2789 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2790 | #if DEBUG |
2791 | /* need to poison the objs? */ | |
2792 | if (cachep->flags & SLAB_POISON) | |
2793 | poison_obj(cachep, objp, POISON_FREE); | |
2794 | if (cachep->flags & SLAB_STORE_USER) | |
2795 | *dbg_userword(cachep, objp) = NULL; | |
2796 | ||
2797 | if (cachep->flags & SLAB_RED_ZONE) { | |
2798 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2799 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2800 | } | |
2801 | /* | |
a737b3e2 AM |
2802 | * Constructors are not allowed to allocate memory from the same |
2803 | * cache which they are a constructor for. Otherwise, deadlock. | |
2804 | * They must also be threaded. | |
1da177e4 LT |
2805 | */ |
2806 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2807 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2808 | |
2809 | if (cachep->flags & SLAB_RED_ZONE) { | |
2810 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2811 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2812 | " end of an object"); |
1da177e4 LT |
2813 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2814 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2815 | " start of an object"); |
1da177e4 | 2816 | } |
3b0efdfa | 2817 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2818 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2819 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2820 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2821 | #else |
2822 | if (cachep->ctor) | |
51cc5068 | 2823 | cachep->ctor(objp); |
1da177e4 | 2824 | #endif |
b28a02de | 2825 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2826 | } |
b28a02de | 2827 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2828 | } |
2829 | ||
343e0d7a | 2830 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2831 | { |
4b51d669 CL |
2832 | if (CONFIG_ZONE_DMA_FLAG) { |
2833 | if (flags & GFP_DMA) | |
2834 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2835 | else | |
2836 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2837 | } | |
1da177e4 LT |
2838 | } |
2839 | ||
a737b3e2 AM |
2840 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2841 | int nodeid) | |
78d382d7 | 2842 | { |
8fea4e96 | 2843 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2844 | kmem_bufctl_t next; |
2845 | ||
2846 | slabp->inuse++; | |
2847 | next = slab_bufctl(slabp)[slabp->free]; | |
2848 | #if DEBUG | |
2849 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2850 | WARN_ON(slabp->nodeid != nodeid); | |
2851 | #endif | |
2852 | slabp->free = next; | |
2853 | ||
2854 | return objp; | |
2855 | } | |
2856 | ||
a737b3e2 AM |
2857 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2858 | void *objp, int nodeid) | |
78d382d7 | 2859 | { |
8fea4e96 | 2860 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2861 | |
2862 | #if DEBUG | |
2863 | /* Verify that the slab belongs to the intended node */ | |
2864 | WARN_ON(slabp->nodeid != nodeid); | |
2865 | ||
871751e2 | 2866 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2867 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2868 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2869 | BUG(); |
2870 | } | |
2871 | #endif | |
2872 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2873 | slabp->free = objnr; | |
2874 | slabp->inuse--; | |
2875 | } | |
2876 | ||
4776874f PE |
2877 | /* |
2878 | * Map pages beginning at addr to the given cache and slab. This is required | |
2879 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2880 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2881 | */ |
2882 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2883 | void *addr) | |
1da177e4 | 2884 | { |
4776874f | 2885 | int nr_pages; |
1da177e4 LT |
2886 | struct page *page; |
2887 | ||
4776874f | 2888 | page = virt_to_page(addr); |
84097518 | 2889 | |
4776874f | 2890 | nr_pages = 1; |
84097518 | 2891 | if (likely(!PageCompound(page))) |
4776874f PE |
2892 | nr_pages <<= cache->gfporder; |
2893 | ||
1da177e4 | 2894 | do { |
35026088 CL |
2895 | page->slab_cache = cache; |
2896 | page->slab_page = slab; | |
1da177e4 | 2897 | page++; |
4776874f | 2898 | } while (--nr_pages); |
1da177e4 LT |
2899 | } |
2900 | ||
2901 | /* | |
2902 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2903 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2904 | */ | |
3c517a61 CL |
2905 | static int cache_grow(struct kmem_cache *cachep, |
2906 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2907 | { |
b28a02de | 2908 | struct slab *slabp; |
b28a02de PE |
2909 | size_t offset; |
2910 | gfp_t local_flags; | |
e498be7d | 2911 | struct kmem_list3 *l3; |
1da177e4 | 2912 | |
a737b3e2 AM |
2913 | /* |
2914 | * Be lazy and only check for valid flags here, keeping it out of the | |
2915 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2916 | */ |
6cb06229 CL |
2917 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2918 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2919 | |
2e1217cf | 2920 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2921 | check_irq_off(); |
2e1217cf RT |
2922 | l3 = cachep->nodelists[nodeid]; |
2923 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2924 | |
2925 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2926 | offset = l3->colour_next; |
2927 | l3->colour_next++; | |
2928 | if (l3->colour_next >= cachep->colour) | |
2929 | l3->colour_next = 0; | |
2930 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2931 | |
2e1217cf | 2932 | offset *= cachep->colour_off; |
1da177e4 LT |
2933 | |
2934 | if (local_flags & __GFP_WAIT) | |
2935 | local_irq_enable(); | |
2936 | ||
2937 | /* | |
2938 | * The test for missing atomic flag is performed here, rather than | |
2939 | * the more obvious place, simply to reduce the critical path length | |
2940 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2941 | * will eventually be caught here (where it matters). | |
2942 | */ | |
2943 | kmem_flagcheck(cachep, flags); | |
2944 | ||
a737b3e2 AM |
2945 | /* |
2946 | * Get mem for the objs. Attempt to allocate a physical page from | |
2947 | * 'nodeid'. | |
e498be7d | 2948 | */ |
3c517a61 | 2949 | if (!objp) |
b8c1c5da | 2950 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2951 | if (!objp) |
1da177e4 LT |
2952 | goto failed; |
2953 | ||
2954 | /* Get slab management. */ | |
3c517a61 | 2955 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2956 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2957 | if (!slabp) |
1da177e4 LT |
2958 | goto opps1; |
2959 | ||
4776874f | 2960 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2961 | |
a35afb83 | 2962 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2963 | |
2964 | if (local_flags & __GFP_WAIT) | |
2965 | local_irq_disable(); | |
2966 | check_irq_off(); | |
e498be7d | 2967 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2968 | |
2969 | /* Make slab active. */ | |
e498be7d | 2970 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2971 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2972 | l3->free_objects += cachep->num; |
2973 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2974 | return 1; |
a737b3e2 | 2975 | opps1: |
1da177e4 | 2976 | kmem_freepages(cachep, objp); |
a737b3e2 | 2977 | failed: |
1da177e4 LT |
2978 | if (local_flags & __GFP_WAIT) |
2979 | local_irq_disable(); | |
2980 | return 0; | |
2981 | } | |
2982 | ||
2983 | #if DEBUG | |
2984 | ||
2985 | /* | |
2986 | * Perform extra freeing checks: | |
2987 | * - detect bad pointers. | |
2988 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2989 | */ |
2990 | static void kfree_debugcheck(const void *objp) | |
2991 | { | |
1da177e4 LT |
2992 | if (!virt_addr_valid(objp)) { |
2993 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2994 | (unsigned long)objp); |
2995 | BUG(); | |
1da177e4 | 2996 | } |
1da177e4 LT |
2997 | } |
2998 | ||
58ce1fd5 PE |
2999 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
3000 | { | |
b46b8f19 | 3001 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
3002 | |
3003 | redzone1 = *dbg_redzone1(cache, obj); | |
3004 | redzone2 = *dbg_redzone2(cache, obj); | |
3005 | ||
3006 | /* | |
3007 | * Redzone is ok. | |
3008 | */ | |
3009 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3010 | return; | |
3011 | ||
3012 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3013 | slab_error(cache, "double free detected"); | |
3014 | else | |
3015 | slab_error(cache, "memory outside object was overwritten"); | |
3016 | ||
b46b8f19 | 3017 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
3018 | obj, redzone1, redzone2); |
3019 | } | |
3020 | ||
343e0d7a | 3021 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 3022 | void *caller) |
1da177e4 LT |
3023 | { |
3024 | struct page *page; | |
3025 | unsigned int objnr; | |
3026 | struct slab *slabp; | |
3027 | ||
80cbd911 MW |
3028 | BUG_ON(virt_to_cache(objp) != cachep); |
3029 | ||
3dafccf2 | 3030 | objp -= obj_offset(cachep); |
1da177e4 | 3031 | kfree_debugcheck(objp); |
b49af68f | 3032 | page = virt_to_head_page(objp); |
1da177e4 | 3033 | |
35026088 | 3034 | slabp = page->slab_page; |
1da177e4 LT |
3035 | |
3036 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3037 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3038 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3039 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3040 | } | |
3041 | if (cachep->flags & SLAB_STORE_USER) | |
3042 | *dbg_userword(cachep, objp) = caller; | |
3043 | ||
8fea4e96 | 3044 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3045 | |
3046 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3047 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3048 | |
871751e2 AV |
3049 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3050 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3051 | #endif | |
1da177e4 LT |
3052 | if (cachep->flags & SLAB_POISON) { |
3053 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3054 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 3055 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 3056 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3057 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3058 | } else { |
3059 | poison_obj(cachep, objp, POISON_FREE); | |
3060 | } | |
3061 | #else | |
3062 | poison_obj(cachep, objp, POISON_FREE); | |
3063 | #endif | |
3064 | } | |
3065 | return objp; | |
3066 | } | |
3067 | ||
343e0d7a | 3068 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3069 | { |
3070 | kmem_bufctl_t i; | |
3071 | int entries = 0; | |
b28a02de | 3072 | |
1da177e4 LT |
3073 | /* Check slab's freelist to see if this obj is there. */ |
3074 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3075 | entries++; | |
3076 | if (entries > cachep->num || i >= cachep->num) | |
3077 | goto bad; | |
3078 | } | |
3079 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3080 | bad: |
3081 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3082 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3083 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3084 | print_tainted()); | |
fdde6abb SAS |
3085 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3086 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3087 | 1); | |
1da177e4 LT |
3088 | BUG(); |
3089 | } | |
3090 | } | |
3091 | #else | |
3092 | #define kfree_debugcheck(x) do { } while(0) | |
3093 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3094 | #define check_slabp(x,y) do { } while(0) | |
3095 | #endif | |
3096 | ||
343e0d7a | 3097 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
3098 | { |
3099 | int batchcount; | |
3100 | struct kmem_list3 *l3; | |
3101 | struct array_cache *ac; | |
1ca4cb24 PE |
3102 | int node; |
3103 | ||
6d2144d3 | 3104 | retry: |
1da177e4 | 3105 | check_irq_off(); |
7d6e6d09 | 3106 | node = numa_mem_id(); |
9a2dba4b | 3107 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3108 | batchcount = ac->batchcount; |
3109 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3110 | /* |
3111 | * If there was little recent activity on this cache, then | |
3112 | * perform only a partial refill. Otherwise we could generate | |
3113 | * refill bouncing. | |
1da177e4 LT |
3114 | */ |
3115 | batchcount = BATCHREFILL_LIMIT; | |
3116 | } | |
1ca4cb24 | 3117 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3118 | |
3119 | BUG_ON(ac->avail > 0 || !l3); | |
3120 | spin_lock(&l3->list_lock); | |
1da177e4 | 3121 | |
3ded175a | 3122 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3123 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3124 | l3->shared->touched = 1; | |
3ded175a | 3125 | goto alloc_done; |
44b57f1c | 3126 | } |
3ded175a | 3127 | |
1da177e4 LT |
3128 | while (batchcount > 0) { |
3129 | struct list_head *entry; | |
3130 | struct slab *slabp; | |
3131 | /* Get slab alloc is to come from. */ | |
3132 | entry = l3->slabs_partial.next; | |
3133 | if (entry == &l3->slabs_partial) { | |
3134 | l3->free_touched = 1; | |
3135 | entry = l3->slabs_free.next; | |
3136 | if (entry == &l3->slabs_free) | |
3137 | goto must_grow; | |
3138 | } | |
3139 | ||
3140 | slabp = list_entry(entry, struct slab, list); | |
3141 | check_slabp(cachep, slabp); | |
3142 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3143 | |
3144 | /* | |
3145 | * The slab was either on partial or free list so | |
3146 | * there must be at least one object available for | |
3147 | * allocation. | |
3148 | */ | |
249b9f33 | 3149 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3150 | |
1da177e4 | 3151 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3152 | STATS_INC_ALLOCED(cachep); |
3153 | STATS_INC_ACTIVE(cachep); | |
3154 | STATS_SET_HIGH(cachep); | |
3155 | ||
78d382d7 | 3156 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3157 | node); |
1da177e4 LT |
3158 | } |
3159 | check_slabp(cachep, slabp); | |
3160 | ||
3161 | /* move slabp to correct slabp list: */ | |
3162 | list_del(&slabp->list); | |
3163 | if (slabp->free == BUFCTL_END) | |
3164 | list_add(&slabp->list, &l3->slabs_full); | |
3165 | else | |
3166 | list_add(&slabp->list, &l3->slabs_partial); | |
3167 | } | |
3168 | ||
a737b3e2 | 3169 | must_grow: |
1da177e4 | 3170 | l3->free_objects -= ac->avail; |
a737b3e2 | 3171 | alloc_done: |
e498be7d | 3172 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3173 | |
3174 | if (unlikely(!ac->avail)) { | |
3175 | int x; | |
3c517a61 | 3176 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3177 | |
a737b3e2 | 3178 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3179 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3180 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3181 | return NULL; |
3182 | ||
a737b3e2 | 3183 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3184 | goto retry; |
3185 | } | |
3186 | ac->touched = 1; | |
e498be7d | 3187 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3188 | } |
3189 | ||
a737b3e2 AM |
3190 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3191 | gfp_t flags) | |
1da177e4 LT |
3192 | { |
3193 | might_sleep_if(flags & __GFP_WAIT); | |
3194 | #if DEBUG | |
3195 | kmem_flagcheck(cachep, flags); | |
3196 | #endif | |
3197 | } | |
3198 | ||
3199 | #if DEBUG | |
a737b3e2 AM |
3200 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3201 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3202 | { |
b28a02de | 3203 | if (!objp) |
1da177e4 | 3204 | return objp; |
b28a02de | 3205 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3206 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3207 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3208 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3209 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3210 | else |
3211 | check_poison_obj(cachep, objp); | |
3212 | #else | |
3213 | check_poison_obj(cachep, objp); | |
3214 | #endif | |
3215 | poison_obj(cachep, objp, POISON_INUSE); | |
3216 | } | |
3217 | if (cachep->flags & SLAB_STORE_USER) | |
3218 | *dbg_userword(cachep, objp) = caller; | |
3219 | ||
3220 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3221 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3222 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3223 | slab_error(cachep, "double free, or memory outside" | |
3224 | " object was overwritten"); | |
b28a02de | 3225 | printk(KERN_ERR |
b46b8f19 | 3226 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3227 | objp, *dbg_redzone1(cachep, objp), |
3228 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3229 | } |
3230 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3231 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3232 | } | |
871751e2 AV |
3233 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3234 | { | |
3235 | struct slab *slabp; | |
3236 | unsigned objnr; | |
3237 | ||
35026088 | 3238 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3239 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3240 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3241 | } | |
3242 | #endif | |
3dafccf2 | 3243 | objp += obj_offset(cachep); |
4f104934 | 3244 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3245 | cachep->ctor(objp); |
7ea466f2 TH |
3246 | if (ARCH_SLAB_MINALIGN && |
3247 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3248 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3249 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3250 | } |
1da177e4 LT |
3251 | return objp; |
3252 | } | |
3253 | #else | |
3254 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3255 | #endif | |
3256 | ||
773ff60e | 3257 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3258 | { |
3259 | if (cachep == &cache_cache) | |
773ff60e | 3260 | return false; |
8a8b6502 | 3261 | |
8c138bc0 | 3262 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3263 | } |
3264 | ||
343e0d7a | 3265 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3266 | { |
b28a02de | 3267 | void *objp; |
1da177e4 LT |
3268 | struct array_cache *ac; |
3269 | ||
5c382300 | 3270 | check_irq_off(); |
8a8b6502 | 3271 | |
9a2dba4b | 3272 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3273 | if (likely(ac->avail)) { |
3274 | STATS_INC_ALLOCHIT(cachep); | |
3275 | ac->touched = 1; | |
e498be7d | 3276 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3277 | } else { |
3278 | STATS_INC_ALLOCMISS(cachep); | |
3279 | objp = cache_alloc_refill(cachep, flags); | |
ddbf2e83 O |
3280 | /* |
3281 | * the 'ac' may be updated by cache_alloc_refill(), | |
3282 | * and kmemleak_erase() requires its correct value. | |
3283 | */ | |
3284 | ac = cpu_cache_get(cachep); | |
1da177e4 | 3285 | } |
d5cff635 CM |
3286 | /* |
3287 | * To avoid a false negative, if an object that is in one of the | |
3288 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3289 | * treat the array pointers as a reference to the object. | |
3290 | */ | |
f3d8b53a O |
3291 | if (objp) |
3292 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3293 | return objp; |
3294 | } | |
3295 | ||
e498be7d | 3296 | #ifdef CONFIG_NUMA |
c61afb18 | 3297 | /* |
b2455396 | 3298 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3299 | * |
3300 | * If we are in_interrupt, then process context, including cpusets and | |
3301 | * mempolicy, may not apply and should not be used for allocation policy. | |
3302 | */ | |
3303 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3304 | { | |
3305 | int nid_alloc, nid_here; | |
3306 | ||
765c4507 | 3307 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3308 | return NULL; |
7d6e6d09 | 3309 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3310 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3311 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3312 | else if (current->mempolicy) |
e7b691b0 | 3313 | nid_alloc = slab_node(); |
c61afb18 | 3314 | if (nid_alloc != nid_here) |
8b98c169 | 3315 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3316 | return NULL; |
3317 | } | |
3318 | ||
765c4507 CL |
3319 | /* |
3320 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3321 | * certain node and fall back is permitted. First we scan all the |
3322 | * available nodelists for available objects. If that fails then we | |
3323 | * perform an allocation without specifying a node. This allows the page | |
3324 | * allocator to do its reclaim / fallback magic. We then insert the | |
3325 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3326 | */ |
8c8cc2c1 | 3327 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3328 | { |
8c8cc2c1 PE |
3329 | struct zonelist *zonelist; |
3330 | gfp_t local_flags; | |
dd1a239f | 3331 | struct zoneref *z; |
54a6eb5c MG |
3332 | struct zone *zone; |
3333 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3334 | void *obj = NULL; |
3c517a61 | 3335 | int nid; |
cc9a6c87 | 3336 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3337 | |
3338 | if (flags & __GFP_THISNODE) | |
3339 | return NULL; | |
3340 | ||
6cb06229 | 3341 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3342 | |
cc9a6c87 MG |
3343 | retry_cpuset: |
3344 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3345 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3346 | |
3c517a61 CL |
3347 | retry: |
3348 | /* | |
3349 | * Look through allowed nodes for objects available | |
3350 | * from existing per node queues. | |
3351 | */ | |
54a6eb5c MG |
3352 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3353 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3354 | |
54a6eb5c | 3355 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3356 | cache->nodelists[nid] && |
481c5346 | 3357 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3358 | obj = ____cache_alloc_node(cache, |
3359 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3360 | if (obj) |
3361 | break; | |
3362 | } | |
3c517a61 CL |
3363 | } |
3364 | ||
cfce6604 | 3365 | if (!obj) { |
3c517a61 CL |
3366 | /* |
3367 | * This allocation will be performed within the constraints | |
3368 | * of the current cpuset / memory policy requirements. | |
3369 | * We may trigger various forms of reclaim on the allowed | |
3370 | * set and go into memory reserves if necessary. | |
3371 | */ | |
dd47ea75 CL |
3372 | if (local_flags & __GFP_WAIT) |
3373 | local_irq_enable(); | |
3374 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3375 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3376 | if (local_flags & __GFP_WAIT) |
3377 | local_irq_disable(); | |
3c517a61 CL |
3378 | if (obj) { |
3379 | /* | |
3380 | * Insert into the appropriate per node queues | |
3381 | */ | |
3382 | nid = page_to_nid(virt_to_page(obj)); | |
3383 | if (cache_grow(cache, flags, nid, obj)) { | |
3384 | obj = ____cache_alloc_node(cache, | |
3385 | flags | GFP_THISNODE, nid); | |
3386 | if (!obj) | |
3387 | /* | |
3388 | * Another processor may allocate the | |
3389 | * objects in the slab since we are | |
3390 | * not holding any locks. | |
3391 | */ | |
3392 | goto retry; | |
3393 | } else { | |
b6a60451 | 3394 | /* cache_grow already freed obj */ |
3c517a61 CL |
3395 | obj = NULL; |
3396 | } | |
3397 | } | |
aedb0eb1 | 3398 | } |
cc9a6c87 MG |
3399 | |
3400 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3401 | goto retry_cpuset; | |
765c4507 CL |
3402 | return obj; |
3403 | } | |
3404 | ||
e498be7d CL |
3405 | /* |
3406 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3407 | */ |
8b98c169 | 3408 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3409 | int nodeid) |
e498be7d CL |
3410 | { |
3411 | struct list_head *entry; | |
b28a02de PE |
3412 | struct slab *slabp; |
3413 | struct kmem_list3 *l3; | |
3414 | void *obj; | |
b28a02de PE |
3415 | int x; |
3416 | ||
3417 | l3 = cachep->nodelists[nodeid]; | |
3418 | BUG_ON(!l3); | |
3419 | ||
a737b3e2 | 3420 | retry: |
ca3b9b91 | 3421 | check_irq_off(); |
b28a02de PE |
3422 | spin_lock(&l3->list_lock); |
3423 | entry = l3->slabs_partial.next; | |
3424 | if (entry == &l3->slabs_partial) { | |
3425 | l3->free_touched = 1; | |
3426 | entry = l3->slabs_free.next; | |
3427 | if (entry == &l3->slabs_free) | |
3428 | goto must_grow; | |
3429 | } | |
3430 | ||
3431 | slabp = list_entry(entry, struct slab, list); | |
3432 | check_spinlock_acquired_node(cachep, nodeid); | |
3433 | check_slabp(cachep, slabp); | |
3434 | ||
3435 | STATS_INC_NODEALLOCS(cachep); | |
3436 | STATS_INC_ACTIVE(cachep); | |
3437 | STATS_SET_HIGH(cachep); | |
3438 | ||
3439 | BUG_ON(slabp->inuse == cachep->num); | |
3440 | ||
78d382d7 | 3441 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3442 | check_slabp(cachep, slabp); |
3443 | l3->free_objects--; | |
3444 | /* move slabp to correct slabp list: */ | |
3445 | list_del(&slabp->list); | |
3446 | ||
a737b3e2 | 3447 | if (slabp->free == BUFCTL_END) |
b28a02de | 3448 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3449 | else |
b28a02de | 3450 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3451 | |
b28a02de PE |
3452 | spin_unlock(&l3->list_lock); |
3453 | goto done; | |
e498be7d | 3454 | |
a737b3e2 | 3455 | must_grow: |
b28a02de | 3456 | spin_unlock(&l3->list_lock); |
3c517a61 | 3457 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3458 | if (x) |
3459 | goto retry; | |
1da177e4 | 3460 | |
8c8cc2c1 | 3461 | return fallback_alloc(cachep, flags); |
e498be7d | 3462 | |
a737b3e2 | 3463 | done: |
b28a02de | 3464 | return obj; |
e498be7d | 3465 | } |
8c8cc2c1 PE |
3466 | |
3467 | /** | |
3468 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3469 | * @cachep: The cache to allocate from. | |
3470 | * @flags: See kmalloc(). | |
3471 | * @nodeid: node number of the target node. | |
3472 | * @caller: return address of caller, used for debug information | |
3473 | * | |
3474 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3475 | * node, which can improve the performance for cpu bound structures. | |
3476 | * | |
3477 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3478 | */ | |
3479 | static __always_inline void * | |
3480 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3481 | void *caller) | |
3482 | { | |
3483 | unsigned long save_flags; | |
3484 | void *ptr; | |
7d6e6d09 | 3485 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3486 | |
dcce284a | 3487 | flags &= gfp_allowed_mask; |
7e85ee0c | 3488 | |
cf40bd16 NP |
3489 | lockdep_trace_alloc(flags); |
3490 | ||
773ff60e | 3491 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3492 | return NULL; |
3493 | ||
8c8cc2c1 PE |
3494 | cache_alloc_debugcheck_before(cachep, flags); |
3495 | local_irq_save(save_flags); | |
3496 | ||
eacbbae3 | 3497 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3498 | nodeid = slab_node; |
8c8cc2c1 PE |
3499 | |
3500 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3501 | /* Node not bootstrapped yet */ | |
3502 | ptr = fallback_alloc(cachep, flags); | |
3503 | goto out; | |
3504 | } | |
3505 | ||
7d6e6d09 | 3506 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3507 | /* |
3508 | * Use the locally cached objects if possible. | |
3509 | * However ____cache_alloc does not allow fallback | |
3510 | * to other nodes. It may fail while we still have | |
3511 | * objects on other nodes available. | |
3512 | */ | |
3513 | ptr = ____cache_alloc(cachep, flags); | |
3514 | if (ptr) | |
3515 | goto out; | |
3516 | } | |
3517 | /* ___cache_alloc_node can fall back to other nodes */ | |
3518 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3519 | out: | |
3520 | local_irq_restore(save_flags); | |
3521 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3522 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3523 | flags); |
8c8cc2c1 | 3524 | |
c175eea4 | 3525 | if (likely(ptr)) |
8c138bc0 | 3526 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3527 | |
d07dbea4 | 3528 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3529 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3530 | |
8c8cc2c1 PE |
3531 | return ptr; |
3532 | } | |
3533 | ||
3534 | static __always_inline void * | |
3535 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3536 | { | |
3537 | void *objp; | |
3538 | ||
3539 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3540 | objp = alternate_node_alloc(cache, flags); | |
3541 | if (objp) | |
3542 | goto out; | |
3543 | } | |
3544 | objp = ____cache_alloc(cache, flags); | |
3545 | ||
3546 | /* | |
3547 | * We may just have run out of memory on the local node. | |
3548 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3549 | */ | |
7d6e6d09 LS |
3550 | if (!objp) |
3551 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3552 | |
3553 | out: | |
3554 | return objp; | |
3555 | } | |
3556 | #else | |
3557 | ||
3558 | static __always_inline void * | |
3559 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3560 | { | |
3561 | return ____cache_alloc(cachep, flags); | |
3562 | } | |
3563 | ||
3564 | #endif /* CONFIG_NUMA */ | |
3565 | ||
3566 | static __always_inline void * | |
3567 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3568 | { | |
3569 | unsigned long save_flags; | |
3570 | void *objp; | |
3571 | ||
dcce284a | 3572 | flags &= gfp_allowed_mask; |
7e85ee0c | 3573 | |
cf40bd16 NP |
3574 | lockdep_trace_alloc(flags); |
3575 | ||
773ff60e | 3576 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3577 | return NULL; |
3578 | ||
8c8cc2c1 PE |
3579 | cache_alloc_debugcheck_before(cachep, flags); |
3580 | local_irq_save(save_flags); | |
3581 | objp = __do_cache_alloc(cachep, flags); | |
3582 | local_irq_restore(save_flags); | |
3583 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3584 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3585 | flags); |
8c8cc2c1 PE |
3586 | prefetchw(objp); |
3587 | ||
c175eea4 | 3588 | if (likely(objp)) |
8c138bc0 | 3589 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3590 | |
d07dbea4 | 3591 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3592 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3593 | |
8c8cc2c1 PE |
3594 | return objp; |
3595 | } | |
e498be7d CL |
3596 | |
3597 | /* | |
3598 | * Caller needs to acquire correct kmem_list's list_lock | |
3599 | */ | |
343e0d7a | 3600 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3601 | int node) |
1da177e4 LT |
3602 | { |
3603 | int i; | |
e498be7d | 3604 | struct kmem_list3 *l3; |
1da177e4 LT |
3605 | |
3606 | for (i = 0; i < nr_objects; i++) { | |
3607 | void *objp = objpp[i]; | |
3608 | struct slab *slabp; | |
1da177e4 | 3609 | |
6ed5eb22 | 3610 | slabp = virt_to_slab(objp); |
ff69416e | 3611 | l3 = cachep->nodelists[node]; |
1da177e4 | 3612 | list_del(&slabp->list); |
ff69416e | 3613 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3614 | check_slabp(cachep, slabp); |
78d382d7 | 3615 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3616 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3617 | l3->free_objects++; |
1da177e4 LT |
3618 | check_slabp(cachep, slabp); |
3619 | ||
3620 | /* fixup slab chains */ | |
3621 | if (slabp->inuse == 0) { | |
e498be7d CL |
3622 | if (l3->free_objects > l3->free_limit) { |
3623 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3624 | /* No need to drop any previously held |
3625 | * lock here, even if we have a off-slab slab | |
3626 | * descriptor it is guaranteed to come from | |
3627 | * a different cache, refer to comments before | |
3628 | * alloc_slabmgmt. | |
3629 | */ | |
1da177e4 LT |
3630 | slab_destroy(cachep, slabp); |
3631 | } else { | |
e498be7d | 3632 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3633 | } |
3634 | } else { | |
3635 | /* Unconditionally move a slab to the end of the | |
3636 | * partial list on free - maximum time for the | |
3637 | * other objects to be freed, too. | |
3638 | */ | |
e498be7d | 3639 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3640 | } |
3641 | } | |
3642 | } | |
3643 | ||
343e0d7a | 3644 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3645 | { |
3646 | int batchcount; | |
e498be7d | 3647 | struct kmem_list3 *l3; |
7d6e6d09 | 3648 | int node = numa_mem_id(); |
1da177e4 LT |
3649 | |
3650 | batchcount = ac->batchcount; | |
3651 | #if DEBUG | |
3652 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3653 | #endif | |
3654 | check_irq_off(); | |
ff69416e | 3655 | l3 = cachep->nodelists[node]; |
873623df | 3656 | spin_lock(&l3->list_lock); |
e498be7d CL |
3657 | if (l3->shared) { |
3658 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3659 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3660 | if (max) { |
3661 | if (batchcount > max) | |
3662 | batchcount = max; | |
e498be7d | 3663 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3664 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3665 | shared_array->avail += batchcount; |
3666 | goto free_done; | |
3667 | } | |
3668 | } | |
3669 | ||
ff69416e | 3670 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3671 | free_done: |
1da177e4 LT |
3672 | #if STATS |
3673 | { | |
3674 | int i = 0; | |
3675 | struct list_head *p; | |
3676 | ||
e498be7d CL |
3677 | p = l3->slabs_free.next; |
3678 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3679 | struct slab *slabp; |
3680 | ||
3681 | slabp = list_entry(p, struct slab, list); | |
3682 | BUG_ON(slabp->inuse); | |
3683 | ||
3684 | i++; | |
3685 | p = p->next; | |
3686 | } | |
3687 | STATS_SET_FREEABLE(cachep, i); | |
3688 | } | |
3689 | #endif | |
e498be7d | 3690 | spin_unlock(&l3->list_lock); |
1da177e4 | 3691 | ac->avail -= batchcount; |
a737b3e2 | 3692 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3693 | } |
3694 | ||
3695 | /* | |
a737b3e2 AM |
3696 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3697 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3698 | */ |
a947eb95 SS |
3699 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3700 | void *caller) | |
1da177e4 | 3701 | { |
9a2dba4b | 3702 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3703 | |
3704 | check_irq_off(); | |
d5cff635 | 3705 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3706 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3707 | |
8c138bc0 | 3708 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3709 | |
1807a1aa SS |
3710 | /* |
3711 | * Skip calling cache_free_alien() when the platform is not numa. | |
3712 | * This will avoid cache misses that happen while accessing slabp (which | |
3713 | * is per page memory reference) to get nodeid. Instead use a global | |
3714 | * variable to skip the call, which is mostly likely to be present in | |
3715 | * the cache. | |
3716 | */ | |
b6e68bc1 | 3717 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3718 | return; |
3719 | ||
1da177e4 LT |
3720 | if (likely(ac->avail < ac->limit)) { |
3721 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3722 | } else { |
3723 | STATS_INC_FREEMISS(cachep); | |
3724 | cache_flusharray(cachep, ac); | |
1da177e4 | 3725 | } |
42c8c99c ZJ |
3726 | |
3727 | ac->entry[ac->avail++] = objp; | |
1da177e4 LT |
3728 | } |
3729 | ||
3730 | /** | |
3731 | * kmem_cache_alloc - Allocate an object | |
3732 | * @cachep: The cache to allocate from. | |
3733 | * @flags: See kmalloc(). | |
3734 | * | |
3735 | * Allocate an object from this cache. The flags are only relevant | |
3736 | * if the cache has no available objects. | |
3737 | */ | |
343e0d7a | 3738 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3739 | { |
36555751 EGM |
3740 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3741 | ||
ca2b84cb | 3742 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3743 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3744 | |
3745 | return ret; | |
1da177e4 LT |
3746 | } |
3747 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3748 | ||
0f24f128 | 3749 | #ifdef CONFIG_TRACING |
85beb586 SR |
3750 | void * |
3751 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3752 | { |
85beb586 SR |
3753 | void *ret; |
3754 | ||
3755 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3756 | ||
3757 | trace_kmalloc(_RET_IP_, ret, | |
3758 | size, slab_buffer_size(cachep), flags); | |
3759 | return ret; | |
36555751 | 3760 | } |
85beb586 | 3761 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3762 | #endif |
3763 | ||
1da177e4 | 3764 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3765 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3766 | { | |
36555751 EGM |
3767 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3768 | __builtin_return_address(0)); | |
3769 | ||
ca2b84cb | 3770 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3771 | cachep->object_size, cachep->size, |
ca2b84cb | 3772 | flags, nodeid); |
36555751 EGM |
3773 | |
3774 | return ret; | |
8b98c169 | 3775 | } |
1da177e4 LT |
3776 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3777 | ||
0f24f128 | 3778 | #ifdef CONFIG_TRACING |
85beb586 SR |
3779 | void *kmem_cache_alloc_node_trace(size_t size, |
3780 | struct kmem_cache *cachep, | |
3781 | gfp_t flags, | |
3782 | int nodeid) | |
36555751 | 3783 | { |
85beb586 SR |
3784 | void *ret; |
3785 | ||
3786 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3787 | __builtin_return_address(0)); |
85beb586 SR |
3788 | trace_kmalloc_node(_RET_IP_, ret, |
3789 | size, slab_buffer_size(cachep), | |
3790 | flags, nodeid); | |
3791 | return ret; | |
36555751 | 3792 | } |
85beb586 | 3793 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3794 | #endif |
3795 | ||
8b98c169 CH |
3796 | static __always_inline void * |
3797 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3798 | { |
343e0d7a | 3799 | struct kmem_cache *cachep; |
97e2bde4 MS |
3800 | |
3801 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3802 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3803 | return cachep; | |
85beb586 | 3804 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3805 | } |
8b98c169 | 3806 | |
0bb38a5c | 3807 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3808 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3809 | { | |
3810 | return __do_kmalloc_node(size, flags, node, | |
3811 | __builtin_return_address(0)); | |
3812 | } | |
dbe5e69d | 3813 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3814 | |
3815 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3816 | int node, unsigned long caller) |
8b98c169 | 3817 | { |
ce71e27c | 3818 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3819 | } |
3820 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3821 | #else | |
3822 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3823 | { | |
3824 | return __do_kmalloc_node(size, flags, node, NULL); | |
3825 | } | |
3826 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3827 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3828 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3829 | |
3830 | /** | |
800590f5 | 3831 | * __do_kmalloc - allocate memory |
1da177e4 | 3832 | * @size: how many bytes of memory are required. |
800590f5 | 3833 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3834 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3835 | */ |
7fd6b141 PE |
3836 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3837 | void *caller) | |
1da177e4 | 3838 | { |
343e0d7a | 3839 | struct kmem_cache *cachep; |
36555751 | 3840 | void *ret; |
1da177e4 | 3841 | |
97e2bde4 MS |
3842 | /* If you want to save a few bytes .text space: replace |
3843 | * __ with kmem_. | |
3844 | * Then kmalloc uses the uninlined functions instead of the inline | |
3845 | * functions. | |
3846 | */ | |
3847 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3848 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3849 | return cachep; | |
36555751 EGM |
3850 | ret = __cache_alloc(cachep, flags, caller); |
3851 | ||
ca2b84cb | 3852 | trace_kmalloc((unsigned long) caller, ret, |
3b0efdfa | 3853 | size, cachep->size, flags); |
36555751 EGM |
3854 | |
3855 | return ret; | |
7fd6b141 PE |
3856 | } |
3857 | ||
7fd6b141 | 3858 | |
0bb38a5c | 3859 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3860 | void *__kmalloc(size_t size, gfp_t flags) |
3861 | { | |
871751e2 | 3862 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3863 | } |
3864 | EXPORT_SYMBOL(__kmalloc); | |
3865 | ||
ce71e27c | 3866 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3867 | { |
ce71e27c | 3868 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3869 | } |
3870 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3871 | |
3872 | #else | |
3873 | void *__kmalloc(size_t size, gfp_t flags) | |
3874 | { | |
3875 | return __do_kmalloc(size, flags, NULL); | |
3876 | } | |
3877 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3878 | #endif |
3879 | ||
1da177e4 LT |
3880 | /** |
3881 | * kmem_cache_free - Deallocate an object | |
3882 | * @cachep: The cache the allocation was from. | |
3883 | * @objp: The previously allocated object. | |
3884 | * | |
3885 | * Free an object which was previously allocated from this | |
3886 | * cache. | |
3887 | */ | |
343e0d7a | 3888 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3889 | { |
3890 | unsigned long flags; | |
3891 | ||
3892 | local_irq_save(flags); | |
8c138bc0 | 3893 | debug_check_no_locks_freed(objp, cachep->size); |
3ac7fe5a | 3894 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3895 | debug_check_no_obj_freed(objp, cachep->object_size); |
a947eb95 | 3896 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3897 | local_irq_restore(flags); |
36555751 | 3898 | |
ca2b84cb | 3899 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3900 | } |
3901 | EXPORT_SYMBOL(kmem_cache_free); | |
3902 | ||
1da177e4 LT |
3903 | /** |
3904 | * kfree - free previously allocated memory | |
3905 | * @objp: pointer returned by kmalloc. | |
3906 | * | |
80e93eff PE |
3907 | * If @objp is NULL, no operation is performed. |
3908 | * | |
1da177e4 LT |
3909 | * Don't free memory not originally allocated by kmalloc() |
3910 | * or you will run into trouble. | |
3911 | */ | |
3912 | void kfree(const void *objp) | |
3913 | { | |
343e0d7a | 3914 | struct kmem_cache *c; |
1da177e4 LT |
3915 | unsigned long flags; |
3916 | ||
2121db74 PE |
3917 | trace_kfree(_RET_IP_, objp); |
3918 | ||
6cb8f913 | 3919 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3920 | return; |
3921 | local_irq_save(flags); | |
3922 | kfree_debugcheck(objp); | |
6ed5eb22 | 3923 | c = virt_to_cache(objp); |
8c138bc0 CL |
3924 | debug_check_no_locks_freed(objp, c->object_size); |
3925 | ||
3926 | debug_check_no_obj_freed(objp, c->object_size); | |
a947eb95 | 3927 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
3928 | local_irq_restore(flags); |
3929 | } | |
3930 | EXPORT_SYMBOL(kfree); | |
3931 | ||
343e0d7a | 3932 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3933 | { |
8c138bc0 | 3934 | return cachep->object_size; |
1da177e4 LT |
3935 | } |
3936 | EXPORT_SYMBOL(kmem_cache_size); | |
3937 | ||
e498be7d | 3938 | /* |
183ff22b | 3939 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3940 | */ |
83b519e8 | 3941 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3942 | { |
3943 | int node; | |
3944 | struct kmem_list3 *l3; | |
cafeb02e | 3945 | struct array_cache *new_shared; |
3395ee05 | 3946 | struct array_cache **new_alien = NULL; |
e498be7d | 3947 | |
9c09a95c | 3948 | for_each_online_node(node) { |
cafeb02e | 3949 | |
3395ee05 | 3950 | if (use_alien_caches) { |
83b519e8 | 3951 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3952 | if (!new_alien) |
3953 | goto fail; | |
3954 | } | |
cafeb02e | 3955 | |
63109846 ED |
3956 | new_shared = NULL; |
3957 | if (cachep->shared) { | |
3958 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3959 | cachep->shared*cachep->batchcount, |
83b519e8 | 3960 | 0xbaadf00d, gfp); |
63109846 ED |
3961 | if (!new_shared) { |
3962 | free_alien_cache(new_alien); | |
3963 | goto fail; | |
3964 | } | |
0718dc2a | 3965 | } |
cafeb02e | 3966 | |
a737b3e2 AM |
3967 | l3 = cachep->nodelists[node]; |
3968 | if (l3) { | |
cafeb02e CL |
3969 | struct array_cache *shared = l3->shared; |
3970 | ||
e498be7d CL |
3971 | spin_lock_irq(&l3->list_lock); |
3972 | ||
cafeb02e | 3973 | if (shared) |
0718dc2a CL |
3974 | free_block(cachep, shared->entry, |
3975 | shared->avail, node); | |
e498be7d | 3976 | |
cafeb02e CL |
3977 | l3->shared = new_shared; |
3978 | if (!l3->alien) { | |
e498be7d CL |
3979 | l3->alien = new_alien; |
3980 | new_alien = NULL; | |
3981 | } | |
b28a02de | 3982 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3983 | cachep->batchcount + cachep->num; |
e498be7d | 3984 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3985 | kfree(shared); |
e498be7d CL |
3986 | free_alien_cache(new_alien); |
3987 | continue; | |
3988 | } | |
83b519e8 | 3989 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3990 | if (!l3) { |
3991 | free_alien_cache(new_alien); | |
3992 | kfree(new_shared); | |
e498be7d | 3993 | goto fail; |
0718dc2a | 3994 | } |
e498be7d CL |
3995 | |
3996 | kmem_list3_init(l3); | |
3997 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3998 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3999 | l3->shared = new_shared; |
e498be7d | 4000 | l3->alien = new_alien; |
b28a02de | 4001 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4002 | cachep->batchcount + cachep->num; |
e498be7d CL |
4003 | cachep->nodelists[node] = l3; |
4004 | } | |
cafeb02e | 4005 | return 0; |
0718dc2a | 4006 | |
a737b3e2 | 4007 | fail: |
3b0efdfa | 4008 | if (!cachep->list.next) { |
0718dc2a CL |
4009 | /* Cache is not active yet. Roll back what we did */ |
4010 | node--; | |
4011 | while (node >= 0) { | |
4012 | if (cachep->nodelists[node]) { | |
4013 | l3 = cachep->nodelists[node]; | |
4014 | ||
4015 | kfree(l3->shared); | |
4016 | free_alien_cache(l3->alien); | |
4017 | kfree(l3); | |
4018 | cachep->nodelists[node] = NULL; | |
4019 | } | |
4020 | node--; | |
4021 | } | |
4022 | } | |
cafeb02e | 4023 | return -ENOMEM; |
e498be7d CL |
4024 | } |
4025 | ||
1da177e4 | 4026 | struct ccupdate_struct { |
343e0d7a | 4027 | struct kmem_cache *cachep; |
acfe7d74 | 4028 | struct array_cache *new[0]; |
1da177e4 LT |
4029 | }; |
4030 | ||
4031 | static void do_ccupdate_local(void *info) | |
4032 | { | |
a737b3e2 | 4033 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4034 | struct array_cache *old; |
4035 | ||
4036 | check_irq_off(); | |
9a2dba4b | 4037 | old = cpu_cache_get(new->cachep); |
e498be7d | 4038 | |
1da177e4 LT |
4039 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4040 | new->new[smp_processor_id()] = old; | |
4041 | } | |
4042 | ||
b5d8ca7c | 4043 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 4044 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4045 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4046 | { |
d2e7b7d0 | 4047 | struct ccupdate_struct *new; |
2ed3a4ef | 4048 | int i; |
1da177e4 | 4049 | |
acfe7d74 ED |
4050 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4051 | gfp); | |
d2e7b7d0 SS |
4052 | if (!new) |
4053 | return -ENOMEM; | |
4054 | ||
e498be7d | 4055 | for_each_online_cpu(i) { |
7d6e6d09 | 4056 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4057 | batchcount, gfp); |
d2e7b7d0 | 4058 | if (!new->new[i]) { |
b28a02de | 4059 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4060 | kfree(new->new[i]); |
4061 | kfree(new); | |
e498be7d | 4062 | return -ENOMEM; |
1da177e4 LT |
4063 | } |
4064 | } | |
d2e7b7d0 | 4065 | new->cachep = cachep; |
1da177e4 | 4066 | |
15c8b6c1 | 4067 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4068 | |
1da177e4 | 4069 | check_irq_on(); |
1da177e4 LT |
4070 | cachep->batchcount = batchcount; |
4071 | cachep->limit = limit; | |
e498be7d | 4072 | cachep->shared = shared; |
1da177e4 | 4073 | |
e498be7d | 4074 | for_each_online_cpu(i) { |
d2e7b7d0 | 4075 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4076 | if (!ccold) |
4077 | continue; | |
7d6e6d09 LS |
4078 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4079 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4080 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4081 | kfree(ccold); |
4082 | } | |
d2e7b7d0 | 4083 | kfree(new); |
83b519e8 | 4084 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4085 | } |
4086 | ||
b5d8ca7c | 4087 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 4088 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4089 | { |
4090 | int err; | |
4091 | int limit, shared; | |
4092 | ||
a737b3e2 AM |
4093 | /* |
4094 | * The head array serves three purposes: | |
1da177e4 LT |
4095 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4096 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4097 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4098 | * bufctl chains: array operations are cheaper. |
4099 | * The numbers are guessed, we should auto-tune as described by | |
4100 | * Bonwick. | |
4101 | */ | |
3b0efdfa | 4102 | if (cachep->size > 131072) |
1da177e4 | 4103 | limit = 1; |
3b0efdfa | 4104 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4105 | limit = 8; |
3b0efdfa | 4106 | else if (cachep->size > 1024) |
1da177e4 | 4107 | limit = 24; |
3b0efdfa | 4108 | else if (cachep->size > 256) |
1da177e4 LT |
4109 | limit = 54; |
4110 | else | |
4111 | limit = 120; | |
4112 | ||
a737b3e2 AM |
4113 | /* |
4114 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4115 | * allocation behaviour: Most allocs on one cpu, most free operations |
4116 | * on another cpu. For these cases, an efficient object passing between | |
4117 | * cpus is necessary. This is provided by a shared array. The array | |
4118 | * replaces Bonwick's magazine layer. | |
4119 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4120 | * to a larger limit. Thus disabled by default. | |
4121 | */ | |
4122 | shared = 0; | |
3b0efdfa | 4123 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4124 | shared = 8; |
1da177e4 LT |
4125 | |
4126 | #if DEBUG | |
a737b3e2 AM |
4127 | /* |
4128 | * With debugging enabled, large batchcount lead to excessively long | |
4129 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4130 | */ |
4131 | if (limit > 32) | |
4132 | limit = 32; | |
4133 | #endif | |
83b519e8 | 4134 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4135 | if (err) |
4136 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4137 | cachep->name, -err); |
2ed3a4ef | 4138 | return err; |
1da177e4 LT |
4139 | } |
4140 | ||
1b55253a CL |
4141 | /* |
4142 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4143 | * necessary. Note that the l3 listlock also protects the array_cache |
4144 | * if drain_array() is used on the shared array. | |
1b55253a | 4145 | */ |
68a1b195 | 4146 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4147 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4148 | { |
4149 | int tofree; | |
4150 | ||
1b55253a CL |
4151 | if (!ac || !ac->avail) |
4152 | return; | |
1da177e4 LT |
4153 | if (ac->touched && !force) { |
4154 | ac->touched = 0; | |
b18e7e65 | 4155 | } else { |
1b55253a | 4156 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4157 | if (ac->avail) { |
4158 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4159 | if (tofree > ac->avail) | |
4160 | tofree = (ac->avail + 1) / 2; | |
4161 | free_block(cachep, ac->entry, tofree, node); | |
4162 | ac->avail -= tofree; | |
4163 | memmove(ac->entry, &(ac->entry[tofree]), | |
4164 | sizeof(void *) * ac->avail); | |
4165 | } | |
1b55253a | 4166 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4167 | } |
4168 | } | |
4169 | ||
4170 | /** | |
4171 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4172 | * @w: work descriptor |
1da177e4 LT |
4173 | * |
4174 | * Called from workqueue/eventd every few seconds. | |
4175 | * Purpose: | |
4176 | * - clear the per-cpu caches for this CPU. | |
4177 | * - return freeable pages to the main free memory pool. | |
4178 | * | |
a737b3e2 AM |
4179 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4180 | * again on the next iteration. | |
1da177e4 | 4181 | */ |
7c5cae36 | 4182 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4183 | { |
7a7c381d | 4184 | struct kmem_cache *searchp; |
e498be7d | 4185 | struct kmem_list3 *l3; |
7d6e6d09 | 4186 | int node = numa_mem_id(); |
bf6aede7 | 4187 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4188 | |
7c5cae36 | 4189 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4190 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4191 | goto out; |
1da177e4 | 4192 | |
3b0efdfa | 4193 | list_for_each_entry(searchp, &cache_chain, list) { |
1da177e4 LT |
4194 | check_irq_on(); |
4195 | ||
35386e3b CL |
4196 | /* |
4197 | * We only take the l3 lock if absolutely necessary and we | |
4198 | * have established with reasonable certainty that | |
4199 | * we can do some work if the lock was obtained. | |
4200 | */ | |
aab2207c | 4201 | l3 = searchp->nodelists[node]; |
35386e3b | 4202 | |
8fce4d8e | 4203 | reap_alien(searchp, l3); |
1da177e4 | 4204 | |
aab2207c | 4205 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4206 | |
35386e3b CL |
4207 | /* |
4208 | * These are racy checks but it does not matter | |
4209 | * if we skip one check or scan twice. | |
4210 | */ | |
e498be7d | 4211 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4212 | goto next; |
1da177e4 | 4213 | |
e498be7d | 4214 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4215 | |
aab2207c | 4216 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4217 | |
ed11d9eb | 4218 | if (l3->free_touched) |
e498be7d | 4219 | l3->free_touched = 0; |
ed11d9eb CL |
4220 | else { |
4221 | int freed; | |
1da177e4 | 4222 | |
ed11d9eb CL |
4223 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4224 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4225 | STATS_ADD_REAPED(searchp, freed); | |
4226 | } | |
35386e3b | 4227 | next: |
1da177e4 LT |
4228 | cond_resched(); |
4229 | } | |
4230 | check_irq_on(); | |
fc0abb14 | 4231 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4232 | next_reap_node(); |
7c5cae36 | 4233 | out: |
a737b3e2 | 4234 | /* Set up the next iteration */ |
7c5cae36 | 4235 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4236 | } |
4237 | ||
158a9624 | 4238 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4239 | |
85289f98 | 4240 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4241 | { |
85289f98 PE |
4242 | /* |
4243 | * Output format version, so at least we can change it | |
4244 | * without _too_ many complaints. | |
4245 | */ | |
1da177e4 | 4246 | #if STATS |
85289f98 | 4247 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4248 | #else |
85289f98 | 4249 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4250 | #endif |
85289f98 PE |
4251 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4252 | "<objperslab> <pagesperslab>"); | |
4253 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4254 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4255 | #if STATS |
85289f98 | 4256 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4257 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4258 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4259 | #endif |
85289f98 PE |
4260 | seq_putc(m, '\n'); |
4261 | } | |
4262 | ||
4263 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4264 | { | |
4265 | loff_t n = *pos; | |
85289f98 | 4266 | |
fc0abb14 | 4267 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4268 | if (!n) |
4269 | print_slabinfo_header(m); | |
b92151ba PE |
4270 | |
4271 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4272 | } |
4273 | ||
4274 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4275 | { | |
b92151ba | 4276 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4277 | } |
4278 | ||
4279 | static void s_stop(struct seq_file *m, void *p) | |
4280 | { | |
fc0abb14 | 4281 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4282 | } |
4283 | ||
4284 | static int s_show(struct seq_file *m, void *p) | |
4285 | { | |
3b0efdfa | 4286 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
b28a02de PE |
4287 | struct slab *slabp; |
4288 | unsigned long active_objs; | |
4289 | unsigned long num_objs; | |
4290 | unsigned long active_slabs = 0; | |
4291 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4292 | const char *name; |
1da177e4 | 4293 | char *error = NULL; |
e498be7d CL |
4294 | int node; |
4295 | struct kmem_list3 *l3; | |
1da177e4 | 4296 | |
1da177e4 LT |
4297 | active_objs = 0; |
4298 | num_slabs = 0; | |
e498be7d CL |
4299 | for_each_online_node(node) { |
4300 | l3 = cachep->nodelists[node]; | |
4301 | if (!l3) | |
4302 | continue; | |
4303 | ||
ca3b9b91 RT |
4304 | check_irq_on(); |
4305 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4306 | |
7a7c381d | 4307 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4308 | if (slabp->inuse != cachep->num && !error) |
4309 | error = "slabs_full accounting error"; | |
4310 | active_objs += cachep->num; | |
4311 | active_slabs++; | |
4312 | } | |
7a7c381d | 4313 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4314 | if (slabp->inuse == cachep->num && !error) |
4315 | error = "slabs_partial inuse accounting error"; | |
4316 | if (!slabp->inuse && !error) | |
4317 | error = "slabs_partial/inuse accounting error"; | |
4318 | active_objs += slabp->inuse; | |
4319 | active_slabs++; | |
4320 | } | |
7a7c381d | 4321 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4322 | if (slabp->inuse && !error) |
4323 | error = "slabs_free/inuse accounting error"; | |
4324 | num_slabs++; | |
4325 | } | |
4326 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4327 | if (l3->shared) |
4328 | shared_avail += l3->shared->avail; | |
e498be7d | 4329 | |
ca3b9b91 | 4330 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4331 | } |
b28a02de PE |
4332 | num_slabs += active_slabs; |
4333 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4334 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4335 | error = "free_objects accounting error"; |
4336 | ||
b28a02de | 4337 | name = cachep->name; |
1da177e4 LT |
4338 | if (error) |
4339 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4340 | ||
4341 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3b0efdfa | 4342 | name, active_objs, num_objs, cachep->size, |
b28a02de | 4343 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4344 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4345 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4346 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4347 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4348 | #if STATS |
b28a02de | 4349 | { /* list3 stats */ |
1da177e4 LT |
4350 | unsigned long high = cachep->high_mark; |
4351 | unsigned long allocs = cachep->num_allocations; | |
4352 | unsigned long grown = cachep->grown; | |
4353 | unsigned long reaped = cachep->reaped; | |
4354 | unsigned long errors = cachep->errors; | |
4355 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4356 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4357 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4358 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4359 | |
e92dd4fd JP |
4360 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4361 | "%4lu %4lu %4lu %4lu %4lu", | |
4362 | allocs, high, grown, | |
4363 | reaped, errors, max_freeable, node_allocs, | |
4364 | node_frees, overflows); | |
1da177e4 LT |
4365 | } |
4366 | /* cpu stats */ | |
4367 | { | |
4368 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4369 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4370 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4371 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4372 | ||
4373 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4374 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4375 | } |
4376 | #endif | |
4377 | seq_putc(m, '\n'); | |
1da177e4 LT |
4378 | return 0; |
4379 | } | |
4380 | ||
4381 | /* | |
4382 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4383 | * | |
4384 | * Output layout: | |
4385 | * cache-name | |
4386 | * num-active-objs | |
4387 | * total-objs | |
4388 | * object size | |
4389 | * num-active-slabs | |
4390 | * total-slabs | |
4391 | * num-pages-per-slab | |
4392 | * + further values on SMP and with statistics enabled | |
4393 | */ | |
4394 | ||
7b3c3a50 | 4395 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4396 | .start = s_start, |
4397 | .next = s_next, | |
4398 | .stop = s_stop, | |
4399 | .show = s_show, | |
1da177e4 LT |
4400 | }; |
4401 | ||
4402 | #define MAX_SLABINFO_WRITE 128 | |
4403 | /** | |
4404 | * slabinfo_write - Tuning for the slab allocator | |
4405 | * @file: unused | |
4406 | * @buffer: user buffer | |
4407 | * @count: data length | |
4408 | * @ppos: unused | |
4409 | */ | |
68a1b195 | 4410 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4411 | size_t count, loff_t *ppos) |
1da177e4 | 4412 | { |
b28a02de | 4413 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4414 | int limit, batchcount, shared, res; |
7a7c381d | 4415 | struct kmem_cache *cachep; |
b28a02de | 4416 | |
1da177e4 LT |
4417 | if (count > MAX_SLABINFO_WRITE) |
4418 | return -EINVAL; | |
4419 | if (copy_from_user(&kbuf, buffer, count)) | |
4420 | return -EFAULT; | |
b28a02de | 4421 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4422 | |
4423 | tmp = strchr(kbuf, ' '); | |
4424 | if (!tmp) | |
4425 | return -EINVAL; | |
4426 | *tmp = '\0'; | |
4427 | tmp++; | |
4428 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4429 | return -EINVAL; | |
4430 | ||
4431 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4432 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4433 | res = -EINVAL; |
3b0efdfa | 4434 | list_for_each_entry(cachep, &cache_chain, list) { |
1da177e4 | 4435 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4436 | if (limit < 1 || batchcount < 1 || |
4437 | batchcount > limit || shared < 0) { | |
e498be7d | 4438 | res = 0; |
1da177e4 | 4439 | } else { |
e498be7d | 4440 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4441 | batchcount, shared, |
4442 | GFP_KERNEL); | |
1da177e4 LT |
4443 | } |
4444 | break; | |
4445 | } | |
4446 | } | |
fc0abb14 | 4447 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4448 | if (res >= 0) |
4449 | res = count; | |
4450 | return res; | |
4451 | } | |
871751e2 | 4452 | |
7b3c3a50 AD |
4453 | static int slabinfo_open(struct inode *inode, struct file *file) |
4454 | { | |
4455 | return seq_open(file, &slabinfo_op); | |
4456 | } | |
4457 | ||
4458 | static const struct file_operations proc_slabinfo_operations = { | |
4459 | .open = slabinfo_open, | |
4460 | .read = seq_read, | |
4461 | .write = slabinfo_write, | |
4462 | .llseek = seq_lseek, | |
4463 | .release = seq_release, | |
4464 | }; | |
4465 | ||
871751e2 AV |
4466 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4467 | ||
4468 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4469 | { | |
871751e2 | 4470 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4471 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4472 | } |
4473 | ||
4474 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4475 | { | |
4476 | unsigned long *p; | |
4477 | int l; | |
4478 | if (!v) | |
4479 | return 1; | |
4480 | l = n[1]; | |
4481 | p = n + 2; | |
4482 | while (l) { | |
4483 | int i = l/2; | |
4484 | unsigned long *q = p + 2 * i; | |
4485 | if (*q == v) { | |
4486 | q[1]++; | |
4487 | return 1; | |
4488 | } | |
4489 | if (*q > v) { | |
4490 | l = i; | |
4491 | } else { | |
4492 | p = q + 2; | |
4493 | l -= i + 1; | |
4494 | } | |
4495 | } | |
4496 | if (++n[1] == n[0]) | |
4497 | return 0; | |
4498 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4499 | p[0] = v; | |
4500 | p[1] = 1; | |
4501 | return 1; | |
4502 | } | |
4503 | ||
4504 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4505 | { | |
4506 | void *p; | |
4507 | int i; | |
4508 | if (n[0] == n[1]) | |
4509 | return; | |
3b0efdfa | 4510 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4511 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4512 | continue; | |
4513 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4514 | return; | |
4515 | } | |
4516 | } | |
4517 | ||
4518 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4519 | { | |
4520 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4521 | unsigned long offset, size; |
9281acea | 4522 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4523 | |
a5c43dae | 4524 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4525 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4526 | if (modname[0]) |
871751e2 AV |
4527 | seq_printf(m, " [%s]", modname); |
4528 | return; | |
4529 | } | |
4530 | #endif | |
4531 | seq_printf(m, "%p", (void *)address); | |
4532 | } | |
4533 | ||
4534 | static int leaks_show(struct seq_file *m, void *p) | |
4535 | { | |
b92151ba | 4536 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4537 | struct slab *slabp; |
4538 | struct kmem_list3 *l3; | |
4539 | const char *name; | |
4540 | unsigned long *n = m->private; | |
4541 | int node; | |
4542 | int i; | |
4543 | ||
4544 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4545 | return 0; | |
4546 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4547 | return 0; | |
4548 | ||
4549 | /* OK, we can do it */ | |
4550 | ||
4551 | n[1] = 0; | |
4552 | ||
4553 | for_each_online_node(node) { | |
4554 | l3 = cachep->nodelists[node]; | |
4555 | if (!l3) | |
4556 | continue; | |
4557 | ||
4558 | check_irq_on(); | |
4559 | spin_lock_irq(&l3->list_lock); | |
4560 | ||
7a7c381d | 4561 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4562 | handle_slab(n, cachep, slabp); |
7a7c381d | 4563 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4564 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4565 | spin_unlock_irq(&l3->list_lock); |
4566 | } | |
4567 | name = cachep->name; | |
4568 | if (n[0] == n[1]) { | |
4569 | /* Increase the buffer size */ | |
4570 | mutex_unlock(&cache_chain_mutex); | |
4571 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4572 | if (!m->private) { | |
4573 | /* Too bad, we are really out */ | |
4574 | m->private = n; | |
4575 | mutex_lock(&cache_chain_mutex); | |
4576 | return -ENOMEM; | |
4577 | } | |
4578 | *(unsigned long *)m->private = n[0] * 2; | |
4579 | kfree(n); | |
4580 | mutex_lock(&cache_chain_mutex); | |
4581 | /* Now make sure this entry will be retried */ | |
4582 | m->count = m->size; | |
4583 | return 0; | |
4584 | } | |
4585 | for (i = 0; i < n[1]; i++) { | |
4586 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4587 | show_symbol(m, n[2*i+2]); | |
4588 | seq_putc(m, '\n'); | |
4589 | } | |
d2e7b7d0 | 4590 | |
871751e2 AV |
4591 | return 0; |
4592 | } | |
4593 | ||
a0ec95a8 | 4594 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4595 | .start = leaks_start, |
4596 | .next = s_next, | |
4597 | .stop = s_stop, | |
4598 | .show = leaks_show, | |
4599 | }; | |
a0ec95a8 AD |
4600 | |
4601 | static int slabstats_open(struct inode *inode, struct file *file) | |
4602 | { | |
4603 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4604 | int ret = -ENOMEM; | |
4605 | if (n) { | |
4606 | ret = seq_open(file, &slabstats_op); | |
4607 | if (!ret) { | |
4608 | struct seq_file *m = file->private_data; | |
4609 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4610 | m->private = n; | |
4611 | n = NULL; | |
4612 | } | |
4613 | kfree(n); | |
4614 | } | |
4615 | return ret; | |
4616 | } | |
4617 | ||
4618 | static const struct file_operations proc_slabstats_operations = { | |
4619 | .open = slabstats_open, | |
4620 | .read = seq_read, | |
4621 | .llseek = seq_lseek, | |
4622 | .release = seq_release_private, | |
4623 | }; | |
4624 | #endif | |
4625 | ||
4626 | static int __init slab_proc_init(void) | |
4627 | { | |
ab067e99 | 4628 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4629 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4630 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4631 | #endif |
a0ec95a8 AD |
4632 | return 0; |
4633 | } | |
4634 | module_init(slab_proc_init); | |
1da177e4 LT |
4635 | #endif |
4636 | ||
00e145b6 MS |
4637 | /** |
4638 | * ksize - get the actual amount of memory allocated for a given object | |
4639 | * @objp: Pointer to the object | |
4640 | * | |
4641 | * kmalloc may internally round up allocations and return more memory | |
4642 | * than requested. ksize() can be used to determine the actual amount of | |
4643 | * memory allocated. The caller may use this additional memory, even though | |
4644 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4645 | * The caller must guarantee that objp points to a valid object previously | |
4646 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4647 | * must not be freed during the duration of the call. | |
4648 | */ | |
fd76bab2 | 4649 | size_t ksize(const void *objp) |
1da177e4 | 4650 | { |
ef8b4520 CL |
4651 | BUG_ON(!objp); |
4652 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4653 | return 0; |
1da177e4 | 4654 | |
8c138bc0 | 4655 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4656 | } |
b1aabecd | 4657 | EXPORT_SYMBOL(ksize); |