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 | ||
343e0d7a | 436 | static int obj_size(struct kmem_cache *cachep) |
1da177e4 | 437 | { |
3b0efdfa | 438 | return cachep->object_size; |
1da177e4 LT |
439 | } |
440 | ||
b46b8f19 | 441 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
442 | { |
443 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
444 | return (unsigned long long*) (objp + obj_offset(cachep) - |
445 | sizeof(unsigned long long)); | |
1da177e4 LT |
446 | } |
447 | ||
b46b8f19 | 448 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
449 | { |
450 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
451 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 452 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 453 | sizeof(unsigned long long) - |
87a927c7 | 454 | REDZONE_ALIGN); |
3b0efdfa | 455 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 456 | sizeof(unsigned long long)); |
1da177e4 LT |
457 | } |
458 | ||
343e0d7a | 459 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
460 | { |
461 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 462 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
463 | } |
464 | ||
465 | #else | |
466 | ||
3dafccf2 | 467 | #define obj_offset(x) 0 |
3b0efdfa | 468 | #define obj_size(cachep) (cachep->size) |
b46b8f19 DW |
469 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
470 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
471 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
472 | ||
473 | #endif | |
474 | ||
0f24f128 | 475 | #ifdef CONFIG_TRACING |
36555751 EGM |
476 | size_t slab_buffer_size(struct kmem_cache *cachep) |
477 | { | |
3b0efdfa | 478 | return cachep->size; |
36555751 EGM |
479 | } |
480 | EXPORT_SYMBOL(slab_buffer_size); | |
481 | #endif | |
482 | ||
1da177e4 | 483 | /* |
3df1cccd DR |
484 | * Do not go above this order unless 0 objects fit into the slab or |
485 | * overridden on the command line. | |
1da177e4 | 486 | */ |
543585cc DR |
487 | #define SLAB_MAX_ORDER_HI 1 |
488 | #define SLAB_MAX_ORDER_LO 0 | |
489 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 490 | static bool slab_max_order_set __initdata; |
1da177e4 | 491 | |
065d41cb PE |
492 | static inline struct kmem_cache *page_get_cache(struct page *page) |
493 | { | |
d85f3385 | 494 | page = compound_head(page); |
ddc2e812 | 495 | BUG_ON(!PageSlab(page)); |
e571b0ad | 496 | return page->slab_cache; |
065d41cb PE |
497 | } |
498 | ||
6ed5eb22 PE |
499 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
500 | { | |
b49af68f | 501 | struct page *page = virt_to_head_page(obj); |
35026088 | 502 | return page->slab_cache; |
6ed5eb22 PE |
503 | } |
504 | ||
505 | static inline struct slab *virt_to_slab(const void *obj) | |
506 | { | |
b49af68f | 507 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
508 | |
509 | VM_BUG_ON(!PageSlab(page)); | |
510 | return page->slab_page; | |
6ed5eb22 PE |
511 | } |
512 | ||
8fea4e96 PE |
513 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
514 | unsigned int idx) | |
515 | { | |
3b0efdfa | 516 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
517 | } |
518 | ||
6a2d7a95 | 519 | /* |
3b0efdfa CL |
520 | * We want to avoid an expensive divide : (offset / cache->size) |
521 | * Using the fact that size is a constant for a particular cache, | |
522 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
523 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
524 | */ | |
525 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
526 | const struct slab *slab, void *obj) | |
8fea4e96 | 527 | { |
6a2d7a95 ED |
528 | u32 offset = (obj - slab->s_mem); |
529 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
530 | } |
531 | ||
a737b3e2 AM |
532 | /* |
533 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
534 | */ | |
1da177e4 LT |
535 | struct cache_sizes malloc_sizes[] = { |
536 | #define CACHE(x) { .cs_size = (x) }, | |
537 | #include <linux/kmalloc_sizes.h> | |
538 | CACHE(ULONG_MAX) | |
539 | #undef CACHE | |
540 | }; | |
541 | EXPORT_SYMBOL(malloc_sizes); | |
542 | ||
543 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
544 | struct cache_names { | |
545 | char *name; | |
546 | char *name_dma; | |
547 | }; | |
548 | ||
549 | static struct cache_names __initdata cache_names[] = { | |
550 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
551 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 552 | {NULL,} |
1da177e4 LT |
553 | #undef CACHE |
554 | }; | |
555 | ||
556 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 557 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 558 | static struct arraycache_init initarray_generic = |
b28a02de | 559 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
560 | |
561 | /* internal cache of cache description objs */ | |
b56efcf0 | 562 | static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES]; |
343e0d7a | 563 | static struct kmem_cache cache_cache = { |
b56efcf0 | 564 | .nodelists = cache_cache_nodelists, |
b28a02de PE |
565 | .batchcount = 1, |
566 | .limit = BOOT_CPUCACHE_ENTRIES, | |
567 | .shared = 1, | |
3b0efdfa | 568 | .size = sizeof(struct kmem_cache), |
b28a02de | 569 | .name = "kmem_cache", |
1da177e4 LT |
570 | }; |
571 | ||
056c6241 RT |
572 | #define BAD_ALIEN_MAGIC 0x01020304ul |
573 | ||
ce79ddc8 PE |
574 | /* |
575 | * chicken and egg problem: delay the per-cpu array allocation | |
576 | * until the general caches are up. | |
577 | */ | |
578 | static enum { | |
579 | NONE, | |
580 | PARTIAL_AC, | |
581 | PARTIAL_L3, | |
582 | EARLY, | |
52cef189 | 583 | LATE, |
ce79ddc8 PE |
584 | FULL |
585 | } g_cpucache_up; | |
586 | ||
587 | /* | |
588 | * used by boot code to determine if it can use slab based allocator | |
589 | */ | |
590 | int slab_is_available(void) | |
591 | { | |
592 | return g_cpucache_up >= EARLY; | |
593 | } | |
594 | ||
f1aaee53 AV |
595 | #ifdef CONFIG_LOCKDEP |
596 | ||
597 | /* | |
598 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
599 | * for other slabs "off slab". | |
600 | * The locking for this is tricky in that it nests within the locks | |
601 | * of all other slabs in a few places; to deal with this special | |
602 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
603 | * |
604 | * We set lock class for alien array caches which are up during init. | |
605 | * The lock annotation will be lost if all cpus of a node goes down and | |
606 | * then comes back up during hotplug | |
f1aaee53 | 607 | */ |
056c6241 RT |
608 | static struct lock_class_key on_slab_l3_key; |
609 | static struct lock_class_key on_slab_alc_key; | |
610 | ||
83835b3d PZ |
611 | static struct lock_class_key debugobj_l3_key; |
612 | static struct lock_class_key debugobj_alc_key; | |
613 | ||
614 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
615 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
616 | int q) | |
617 | { | |
618 | struct array_cache **alc; | |
619 | struct kmem_list3 *l3; | |
620 | int r; | |
621 | ||
622 | l3 = cachep->nodelists[q]; | |
623 | if (!l3) | |
624 | return; | |
625 | ||
626 | lockdep_set_class(&l3->list_lock, l3_key); | |
627 | alc = l3->alien; | |
628 | /* | |
629 | * FIXME: This check for BAD_ALIEN_MAGIC | |
630 | * should go away when common slab code is taught to | |
631 | * work even without alien caches. | |
632 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
633 | * for alloc_alien_cache, | |
634 | */ | |
635 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
636 | return; | |
637 | for_each_node(r) { | |
638 | if (alc[r]) | |
639 | lockdep_set_class(&alc[r]->lock, alc_key); | |
640 | } | |
641 | } | |
642 | ||
643 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
644 | { | |
645 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
646 | } | |
647 | ||
648 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
649 | { | |
650 | int node; | |
651 | ||
652 | for_each_online_node(node) | |
653 | slab_set_debugobj_lock_classes_node(cachep, node); | |
654 | } | |
655 | ||
ce79ddc8 | 656 | static void init_node_lock_keys(int q) |
f1aaee53 | 657 | { |
056c6241 RT |
658 | struct cache_sizes *s = malloc_sizes; |
659 | ||
52cef189 | 660 | if (g_cpucache_up < LATE) |
ce79ddc8 PE |
661 | return; |
662 | ||
663 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
ce79ddc8 | 664 | struct kmem_list3 *l3; |
ce79ddc8 PE |
665 | |
666 | l3 = s->cs_cachep->nodelists[q]; | |
667 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 668 | continue; |
83835b3d PZ |
669 | |
670 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
671 | &on_slab_alc_key, q); | |
f1aaee53 AV |
672 | } |
673 | } | |
ce79ddc8 PE |
674 | |
675 | static inline void init_lock_keys(void) | |
676 | { | |
677 | int node; | |
678 | ||
679 | for_each_node(node) | |
680 | init_node_lock_keys(node); | |
681 | } | |
f1aaee53 | 682 | #else |
ce79ddc8 PE |
683 | static void init_node_lock_keys(int q) |
684 | { | |
685 | } | |
686 | ||
056c6241 | 687 | static inline void init_lock_keys(void) |
f1aaee53 AV |
688 | { |
689 | } | |
83835b3d PZ |
690 | |
691 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
692 | { | |
693 | } | |
694 | ||
695 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
696 | { | |
697 | } | |
f1aaee53 AV |
698 | #endif |
699 | ||
8f5be20b | 700 | /* |
95402b38 | 701 | * Guard access to the cache-chain. |
8f5be20b | 702 | */ |
fc0abb14 | 703 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
704 | static struct list_head cache_chain; |
705 | ||
1871e52c | 706 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 707 | |
343e0d7a | 708 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
709 | { |
710 | return cachep->array[smp_processor_id()]; | |
711 | } | |
712 | ||
a737b3e2 AM |
713 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
714 | gfp_t gfpflags) | |
1da177e4 LT |
715 | { |
716 | struct cache_sizes *csizep = malloc_sizes; | |
717 | ||
718 | #if DEBUG | |
719 | /* This happens if someone tries to call | |
b28a02de PE |
720 | * kmem_cache_create(), or __kmalloc(), before |
721 | * the generic caches are initialized. | |
722 | */ | |
c7e43c78 | 723 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 724 | #endif |
6cb8f913 CL |
725 | if (!size) |
726 | return ZERO_SIZE_PTR; | |
727 | ||
1da177e4 LT |
728 | while (size > csizep->cs_size) |
729 | csizep++; | |
730 | ||
731 | /* | |
0abf40c1 | 732 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
733 | * has cs_{dma,}cachep==NULL. Thus no special case |
734 | * for large kmalloc calls required. | |
735 | */ | |
4b51d669 | 736 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
737 | if (unlikely(gfpflags & GFP_DMA)) |
738 | return csizep->cs_dmacachep; | |
4b51d669 | 739 | #endif |
1da177e4 LT |
740 | return csizep->cs_cachep; |
741 | } | |
742 | ||
b221385b | 743 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
744 | { |
745 | return __find_general_cachep(size, gfpflags); | |
746 | } | |
97e2bde4 | 747 | |
fbaccacf | 748 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 749 | { |
fbaccacf SR |
750 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
751 | } | |
1da177e4 | 752 | |
a737b3e2 AM |
753 | /* |
754 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
755 | */ | |
fbaccacf SR |
756 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
757 | size_t align, int flags, size_t *left_over, | |
758 | unsigned int *num) | |
759 | { | |
760 | int nr_objs; | |
761 | size_t mgmt_size; | |
762 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 763 | |
fbaccacf SR |
764 | /* |
765 | * The slab management structure can be either off the slab or | |
766 | * on it. For the latter case, the memory allocated for a | |
767 | * slab is used for: | |
768 | * | |
769 | * - The struct slab | |
770 | * - One kmem_bufctl_t for each object | |
771 | * - Padding to respect alignment of @align | |
772 | * - @buffer_size bytes for each object | |
773 | * | |
774 | * If the slab management structure is off the slab, then the | |
775 | * alignment will already be calculated into the size. Because | |
776 | * the slabs are all pages aligned, the objects will be at the | |
777 | * correct alignment when allocated. | |
778 | */ | |
779 | if (flags & CFLGS_OFF_SLAB) { | |
780 | mgmt_size = 0; | |
781 | nr_objs = slab_size / buffer_size; | |
782 | ||
783 | if (nr_objs > SLAB_LIMIT) | |
784 | nr_objs = SLAB_LIMIT; | |
785 | } else { | |
786 | /* | |
787 | * Ignore padding for the initial guess. The padding | |
788 | * is at most @align-1 bytes, and @buffer_size is at | |
789 | * least @align. In the worst case, this result will | |
790 | * be one greater than the number of objects that fit | |
791 | * into the memory allocation when taking the padding | |
792 | * into account. | |
793 | */ | |
794 | nr_objs = (slab_size - sizeof(struct slab)) / | |
795 | (buffer_size + sizeof(kmem_bufctl_t)); | |
796 | ||
797 | /* | |
798 | * This calculated number will be either the right | |
799 | * amount, or one greater than what we want. | |
800 | */ | |
801 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
802 | > slab_size) | |
803 | nr_objs--; | |
804 | ||
805 | if (nr_objs > SLAB_LIMIT) | |
806 | nr_objs = SLAB_LIMIT; | |
807 | ||
808 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
809 | } | |
810 | *num = nr_objs; | |
811 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
812 | } |
813 | ||
d40cee24 | 814 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 815 | |
a737b3e2 AM |
816 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
817 | char *msg) | |
1da177e4 LT |
818 | { |
819 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 820 | function, cachep->name, msg); |
1da177e4 LT |
821 | dump_stack(); |
822 | } | |
823 | ||
3395ee05 PM |
824 | /* |
825 | * By default on NUMA we use alien caches to stage the freeing of | |
826 | * objects allocated from other nodes. This causes massive memory | |
827 | * inefficiencies when using fake NUMA setup to split memory into a | |
828 | * large number of small nodes, so it can be disabled on the command | |
829 | * line | |
830 | */ | |
831 | ||
832 | static int use_alien_caches __read_mostly = 1; | |
833 | static int __init noaliencache_setup(char *s) | |
834 | { | |
835 | use_alien_caches = 0; | |
836 | return 1; | |
837 | } | |
838 | __setup("noaliencache", noaliencache_setup); | |
839 | ||
3df1cccd DR |
840 | static int __init slab_max_order_setup(char *str) |
841 | { | |
842 | get_option(&str, &slab_max_order); | |
843 | slab_max_order = slab_max_order < 0 ? 0 : | |
844 | min(slab_max_order, MAX_ORDER - 1); | |
845 | slab_max_order_set = true; | |
846 | ||
847 | return 1; | |
848 | } | |
849 | __setup("slab_max_order=", slab_max_order_setup); | |
850 | ||
8fce4d8e CL |
851 | #ifdef CONFIG_NUMA |
852 | /* | |
853 | * Special reaping functions for NUMA systems called from cache_reap(). | |
854 | * These take care of doing round robin flushing of alien caches (containing | |
855 | * objects freed on different nodes from which they were allocated) and the | |
856 | * flushing of remote pcps by calling drain_node_pages. | |
857 | */ | |
1871e52c | 858 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
859 | |
860 | static void init_reap_node(int cpu) | |
861 | { | |
862 | int node; | |
863 | ||
7d6e6d09 | 864 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 865 | if (node == MAX_NUMNODES) |
442295c9 | 866 | node = first_node(node_online_map); |
8fce4d8e | 867 | |
1871e52c | 868 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
869 | } |
870 | ||
871 | static void next_reap_node(void) | |
872 | { | |
909ea964 | 873 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 874 | |
8fce4d8e CL |
875 | node = next_node(node, node_online_map); |
876 | if (unlikely(node >= MAX_NUMNODES)) | |
877 | node = first_node(node_online_map); | |
909ea964 | 878 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
879 | } |
880 | ||
881 | #else | |
882 | #define init_reap_node(cpu) do { } while (0) | |
883 | #define next_reap_node(void) do { } while (0) | |
884 | #endif | |
885 | ||
1da177e4 LT |
886 | /* |
887 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
888 | * via the workqueue/eventd. | |
889 | * Add the CPU number into the expiration time to minimize the possibility of | |
890 | * the CPUs getting into lockstep and contending for the global cache chain | |
891 | * lock. | |
892 | */ | |
897e679b | 893 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 894 | { |
1871e52c | 895 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
896 | |
897 | /* | |
898 | * When this gets called from do_initcalls via cpucache_init(), | |
899 | * init_workqueues() has already run, so keventd will be setup | |
900 | * at that time. | |
901 | */ | |
52bad64d | 902 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 903 | init_reap_node(cpu); |
78b43536 | 904 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); |
2b284214 AV |
905 | schedule_delayed_work_on(cpu, reap_work, |
906 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
907 | } |
908 | } | |
909 | ||
e498be7d | 910 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 911 | int batchcount, gfp_t gfp) |
1da177e4 | 912 | { |
b28a02de | 913 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
914 | struct array_cache *nc = NULL; |
915 | ||
83b519e8 | 916 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
917 | /* |
918 | * The array_cache structures contain pointers to free object. | |
25985edc | 919 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
920 | * cache the pointers are not cleared and they could be counted as |
921 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
922 | * not scan such objects. | |
923 | */ | |
924 | kmemleak_no_scan(nc); | |
1da177e4 LT |
925 | if (nc) { |
926 | nc->avail = 0; | |
927 | nc->limit = entries; | |
928 | nc->batchcount = batchcount; | |
929 | nc->touched = 0; | |
e498be7d | 930 | spin_lock_init(&nc->lock); |
1da177e4 LT |
931 | } |
932 | return nc; | |
933 | } | |
934 | ||
3ded175a CL |
935 | /* |
936 | * Transfer objects in one arraycache to another. | |
937 | * Locking must be handled by the caller. | |
938 | * | |
939 | * Return the number of entries transferred. | |
940 | */ | |
941 | static int transfer_objects(struct array_cache *to, | |
942 | struct array_cache *from, unsigned int max) | |
943 | { | |
944 | /* Figure out how many entries to transfer */ | |
732eacc0 | 945 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
946 | |
947 | if (!nr) | |
948 | return 0; | |
949 | ||
950 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
951 | sizeof(void *) *nr); | |
952 | ||
953 | from->avail -= nr; | |
954 | to->avail += nr; | |
3ded175a CL |
955 | return nr; |
956 | } | |
957 | ||
765c4507 CL |
958 | #ifndef CONFIG_NUMA |
959 | ||
960 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
961 | #define reap_alien(cachep, l3) do { } while (0) | |
962 | ||
83b519e8 | 963 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
964 | { |
965 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
966 | } | |
967 | ||
968 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
969 | { | |
970 | } | |
971 | ||
972 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
973 | { | |
974 | return 0; | |
975 | } | |
976 | ||
977 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
978 | gfp_t flags) | |
979 | { | |
980 | return NULL; | |
981 | } | |
982 | ||
8b98c169 | 983 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
984 | gfp_t flags, int nodeid) |
985 | { | |
986 | return NULL; | |
987 | } | |
988 | ||
989 | #else /* CONFIG_NUMA */ | |
990 | ||
8b98c169 | 991 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 992 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 993 | |
83b519e8 | 994 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
995 | { |
996 | struct array_cache **ac_ptr; | |
8ef82866 | 997 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
998 | int i; |
999 | ||
1000 | if (limit > 1) | |
1001 | limit = 12; | |
f3186a9c | 1002 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
1003 | if (ac_ptr) { |
1004 | for_each_node(i) { | |
f3186a9c | 1005 | if (i == node || !node_online(i)) |
e498be7d | 1006 | continue; |
83b519e8 | 1007 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1008 | if (!ac_ptr[i]) { |
cc550def | 1009 | for (i--; i >= 0; i--) |
e498be7d CL |
1010 | kfree(ac_ptr[i]); |
1011 | kfree(ac_ptr); | |
1012 | return NULL; | |
1013 | } | |
1014 | } | |
1015 | } | |
1016 | return ac_ptr; | |
1017 | } | |
1018 | ||
5295a74c | 1019 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1020 | { |
1021 | int i; | |
1022 | ||
1023 | if (!ac_ptr) | |
1024 | return; | |
e498be7d | 1025 | for_each_node(i) |
b28a02de | 1026 | kfree(ac_ptr[i]); |
e498be7d CL |
1027 | kfree(ac_ptr); |
1028 | } | |
1029 | ||
343e0d7a | 1030 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1031 | struct array_cache *ac, int node) |
e498be7d CL |
1032 | { |
1033 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1034 | ||
1035 | if (ac->avail) { | |
1036 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1037 | /* |
1038 | * Stuff objects into the remote nodes shared array first. | |
1039 | * That way we could avoid the overhead of putting the objects | |
1040 | * into the free lists and getting them back later. | |
1041 | */ | |
693f7d36 | 1042 | if (rl3->shared) |
1043 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1044 | |
ff69416e | 1045 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1046 | ac->avail = 0; |
1047 | spin_unlock(&rl3->list_lock); | |
1048 | } | |
1049 | } | |
1050 | ||
8fce4d8e CL |
1051 | /* |
1052 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1053 | */ | |
1054 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1055 | { | |
909ea964 | 1056 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1057 | |
1058 | if (l3->alien) { | |
1059 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1060 | |
1061 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1062 | __drain_alien_cache(cachep, ac, node); |
1063 | spin_unlock_irq(&ac->lock); | |
1064 | } | |
1065 | } | |
1066 | } | |
1067 | ||
a737b3e2 AM |
1068 | static void drain_alien_cache(struct kmem_cache *cachep, |
1069 | struct array_cache **alien) | |
e498be7d | 1070 | { |
b28a02de | 1071 | int i = 0; |
e498be7d CL |
1072 | struct array_cache *ac; |
1073 | unsigned long flags; | |
1074 | ||
1075 | for_each_online_node(i) { | |
4484ebf1 | 1076 | ac = alien[i]; |
e498be7d CL |
1077 | if (ac) { |
1078 | spin_lock_irqsave(&ac->lock, flags); | |
1079 | __drain_alien_cache(cachep, ac, i); | |
1080 | spin_unlock_irqrestore(&ac->lock, flags); | |
1081 | } | |
1082 | } | |
1083 | } | |
729bd0b7 | 1084 | |
873623df | 1085 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1086 | { |
1087 | struct slab *slabp = virt_to_slab(objp); | |
1088 | int nodeid = slabp->nodeid; | |
1089 | struct kmem_list3 *l3; | |
1090 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1091 | int node; |
1092 | ||
7d6e6d09 | 1093 | node = numa_mem_id(); |
729bd0b7 PE |
1094 | |
1095 | /* | |
1096 | * Make sure we are not freeing a object from another node to the array | |
1097 | * cache on this cpu. | |
1098 | */ | |
62918a03 | 1099 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1100 | return 0; |
1101 | ||
1ca4cb24 | 1102 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1103 | STATS_INC_NODEFREES(cachep); |
1104 | if (l3->alien && l3->alien[nodeid]) { | |
1105 | alien = l3->alien[nodeid]; | |
873623df | 1106 | spin_lock(&alien->lock); |
729bd0b7 PE |
1107 | if (unlikely(alien->avail == alien->limit)) { |
1108 | STATS_INC_ACOVERFLOW(cachep); | |
1109 | __drain_alien_cache(cachep, alien, nodeid); | |
1110 | } | |
1111 | alien->entry[alien->avail++] = objp; | |
1112 | spin_unlock(&alien->lock); | |
1113 | } else { | |
1114 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1115 | free_block(cachep, &objp, 1, nodeid); | |
1116 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1117 | } | |
1118 | return 1; | |
1119 | } | |
e498be7d CL |
1120 | #endif |
1121 | ||
8f9f8d9e DR |
1122 | /* |
1123 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1124 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1125 | * will be allocated off-node since memory is not yet online for the new node. | |
1126 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1127 | * already in use. | |
1128 | * | |
1129 | * Must hold cache_chain_mutex. | |
1130 | */ | |
1131 | static int init_cache_nodelists_node(int node) | |
1132 | { | |
1133 | struct kmem_cache *cachep; | |
1134 | struct kmem_list3 *l3; | |
1135 | const int memsize = sizeof(struct kmem_list3); | |
1136 | ||
3b0efdfa | 1137 | list_for_each_entry(cachep, &cache_chain, list) { |
8f9f8d9e DR |
1138 | /* |
1139 | * Set up the size64 kmemlist for cpu before we can | |
1140 | * begin anything. Make sure some other cpu on this | |
1141 | * node has not already allocated this | |
1142 | */ | |
1143 | if (!cachep->nodelists[node]) { | |
1144 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1145 | if (!l3) | |
1146 | return -ENOMEM; | |
1147 | kmem_list3_init(l3); | |
1148 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1149 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1150 | ||
1151 | /* | |
1152 | * The l3s don't come and go as CPUs come and | |
1153 | * go. cache_chain_mutex is sufficient | |
1154 | * protection here. | |
1155 | */ | |
1156 | cachep->nodelists[node] = l3; | |
1157 | } | |
1158 | ||
1159 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1160 | cachep->nodelists[node]->free_limit = | |
1161 | (1 + nr_cpus_node(node)) * | |
1162 | cachep->batchcount + cachep->num; | |
1163 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1164 | } | |
1165 | return 0; | |
1166 | } | |
1167 | ||
fbf1e473 AM |
1168 | static void __cpuinit cpuup_canceled(long cpu) |
1169 | { | |
1170 | struct kmem_cache *cachep; | |
1171 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1172 | int node = cpu_to_mem(cpu); |
a70f7302 | 1173 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1174 | |
3b0efdfa | 1175 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1176 | struct array_cache *nc; |
1177 | struct array_cache *shared; | |
1178 | struct array_cache **alien; | |
fbf1e473 | 1179 | |
fbf1e473 AM |
1180 | /* cpu is dead; no one can alloc from it. */ |
1181 | nc = cachep->array[cpu]; | |
1182 | cachep->array[cpu] = NULL; | |
1183 | l3 = cachep->nodelists[node]; | |
1184 | ||
1185 | if (!l3) | |
1186 | goto free_array_cache; | |
1187 | ||
1188 | spin_lock_irq(&l3->list_lock); | |
1189 | ||
1190 | /* Free limit for this kmem_list3 */ | |
1191 | l3->free_limit -= cachep->batchcount; | |
1192 | if (nc) | |
1193 | free_block(cachep, nc->entry, nc->avail, node); | |
1194 | ||
58463c1f | 1195 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1196 | spin_unlock_irq(&l3->list_lock); |
1197 | goto free_array_cache; | |
1198 | } | |
1199 | ||
1200 | shared = l3->shared; | |
1201 | if (shared) { | |
1202 | free_block(cachep, shared->entry, | |
1203 | shared->avail, node); | |
1204 | l3->shared = NULL; | |
1205 | } | |
1206 | ||
1207 | alien = l3->alien; | |
1208 | l3->alien = NULL; | |
1209 | ||
1210 | spin_unlock_irq(&l3->list_lock); | |
1211 | ||
1212 | kfree(shared); | |
1213 | if (alien) { | |
1214 | drain_alien_cache(cachep, alien); | |
1215 | free_alien_cache(alien); | |
1216 | } | |
1217 | free_array_cache: | |
1218 | kfree(nc); | |
1219 | } | |
1220 | /* | |
1221 | * In the previous loop, all the objects were freed to | |
1222 | * the respective cache's slabs, now we can go ahead and | |
1223 | * shrink each nodelist to its limit. | |
1224 | */ | |
3b0efdfa | 1225 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1226 | l3 = cachep->nodelists[node]; |
1227 | if (!l3) | |
1228 | continue; | |
1229 | drain_freelist(cachep, l3, l3->free_objects); | |
1230 | } | |
1231 | } | |
1232 | ||
1233 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1234 | { |
343e0d7a | 1235 | struct kmem_cache *cachep; |
e498be7d | 1236 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1237 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1238 | int err; |
1da177e4 | 1239 | |
fbf1e473 AM |
1240 | /* |
1241 | * We need to do this right in the beginning since | |
1242 | * alloc_arraycache's are going to use this list. | |
1243 | * kmalloc_node allows us to add the slab to the right | |
1244 | * kmem_list3 and not this cpu's kmem_list3 | |
1245 | */ | |
8f9f8d9e DR |
1246 | err = init_cache_nodelists_node(node); |
1247 | if (err < 0) | |
1248 | goto bad; | |
fbf1e473 AM |
1249 | |
1250 | /* | |
1251 | * Now we can go ahead with allocating the shared arrays and | |
1252 | * array caches | |
1253 | */ | |
3b0efdfa | 1254 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1255 | struct array_cache *nc; |
1256 | struct array_cache *shared = NULL; | |
1257 | struct array_cache **alien = NULL; | |
1258 | ||
1259 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1260 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1261 | if (!nc) |
1262 | goto bad; | |
1263 | if (cachep->shared) { | |
1264 | shared = alloc_arraycache(node, | |
1265 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1266 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1267 | if (!shared) { |
1268 | kfree(nc); | |
1da177e4 | 1269 | goto bad; |
12d00f6a | 1270 | } |
fbf1e473 AM |
1271 | } |
1272 | if (use_alien_caches) { | |
83b519e8 | 1273 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1274 | if (!alien) { |
1275 | kfree(shared); | |
1276 | kfree(nc); | |
fbf1e473 | 1277 | goto bad; |
12d00f6a | 1278 | } |
fbf1e473 AM |
1279 | } |
1280 | cachep->array[cpu] = nc; | |
1281 | l3 = cachep->nodelists[node]; | |
1282 | BUG_ON(!l3); | |
1283 | ||
1284 | spin_lock_irq(&l3->list_lock); | |
1285 | if (!l3->shared) { | |
1286 | /* | |
1287 | * We are serialised from CPU_DEAD or | |
1288 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1289 | */ | |
1290 | l3->shared = shared; | |
1291 | shared = NULL; | |
1292 | } | |
4484ebf1 | 1293 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1294 | if (!l3->alien) { |
1295 | l3->alien = alien; | |
1296 | alien = NULL; | |
1da177e4 | 1297 | } |
fbf1e473 AM |
1298 | #endif |
1299 | spin_unlock_irq(&l3->list_lock); | |
1300 | kfree(shared); | |
1301 | free_alien_cache(alien); | |
83835b3d PZ |
1302 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1303 | slab_set_debugobj_lock_classes_node(cachep, node); | |
fbf1e473 | 1304 | } |
ce79ddc8 PE |
1305 | init_node_lock_keys(node); |
1306 | ||
fbf1e473 AM |
1307 | return 0; |
1308 | bad: | |
12d00f6a | 1309 | cpuup_canceled(cpu); |
fbf1e473 AM |
1310 | return -ENOMEM; |
1311 | } | |
1312 | ||
1313 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1314 | unsigned long action, void *hcpu) | |
1315 | { | |
1316 | long cpu = (long)hcpu; | |
1317 | int err = 0; | |
1318 | ||
1319 | switch (action) { | |
fbf1e473 AM |
1320 | case CPU_UP_PREPARE: |
1321 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1322 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1323 | err = cpuup_prepare(cpu); |
95402b38 | 1324 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1325 | break; |
1326 | case CPU_ONLINE: | |
8bb78442 | 1327 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1328 | start_cpu_timer(cpu); |
1329 | break; | |
1330 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1331 | case CPU_DOWN_PREPARE: |
8bb78442 | 1332 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1333 | /* |
1334 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1335 | * held so that if cache_reap() is invoked it cannot do | |
1336 | * anything expensive but will only modify reap_work | |
1337 | * and reschedule the timer. | |
1338 | */ | |
afe2c511 | 1339 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1340 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1341 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1342 | break; |
1343 | case CPU_DOWN_FAILED: | |
8bb78442 | 1344 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1345 | start_cpu_timer(cpu); |
1346 | break; | |
1da177e4 | 1347 | case CPU_DEAD: |
8bb78442 | 1348 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1349 | /* |
1350 | * Even if all the cpus of a node are down, we don't free the | |
1351 | * kmem_list3 of any cache. This to avoid a race between | |
1352 | * cpu_down, and a kmalloc allocation from another cpu for | |
1353 | * memory from the node of the cpu going down. The list3 | |
1354 | * structure is usually allocated from kmem_cache_create() and | |
1355 | * gets destroyed at kmem_cache_destroy(). | |
1356 | */ | |
183ff22b | 1357 | /* fall through */ |
8f5be20b | 1358 | #endif |
1da177e4 | 1359 | case CPU_UP_CANCELED: |
8bb78442 | 1360 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1361 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1362 | cpuup_canceled(cpu); |
fc0abb14 | 1363 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1364 | break; |
1da177e4 | 1365 | } |
eac40680 | 1366 | return notifier_from_errno(err); |
1da177e4 LT |
1367 | } |
1368 | ||
74b85f37 CS |
1369 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1370 | &cpuup_callback, NULL, 0 | |
1371 | }; | |
1da177e4 | 1372 | |
8f9f8d9e DR |
1373 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1374 | /* | |
1375 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1376 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1377 | * removed. | |
1378 | * | |
1379 | * Must hold cache_chain_mutex. | |
1380 | */ | |
1381 | static int __meminit drain_cache_nodelists_node(int node) | |
1382 | { | |
1383 | struct kmem_cache *cachep; | |
1384 | int ret = 0; | |
1385 | ||
3b0efdfa | 1386 | list_for_each_entry(cachep, &cache_chain, list) { |
8f9f8d9e DR |
1387 | struct kmem_list3 *l3; |
1388 | ||
1389 | l3 = cachep->nodelists[node]; | |
1390 | if (!l3) | |
1391 | continue; | |
1392 | ||
1393 | drain_freelist(cachep, l3, l3->free_objects); | |
1394 | ||
1395 | if (!list_empty(&l3->slabs_full) || | |
1396 | !list_empty(&l3->slabs_partial)) { | |
1397 | ret = -EBUSY; | |
1398 | break; | |
1399 | } | |
1400 | } | |
1401 | return ret; | |
1402 | } | |
1403 | ||
1404 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1405 | unsigned long action, void *arg) | |
1406 | { | |
1407 | struct memory_notify *mnb = arg; | |
1408 | int ret = 0; | |
1409 | int nid; | |
1410 | ||
1411 | nid = mnb->status_change_nid; | |
1412 | if (nid < 0) | |
1413 | goto out; | |
1414 | ||
1415 | switch (action) { | |
1416 | case MEM_GOING_ONLINE: | |
1417 | mutex_lock(&cache_chain_mutex); | |
1418 | ret = init_cache_nodelists_node(nid); | |
1419 | mutex_unlock(&cache_chain_mutex); | |
1420 | break; | |
1421 | case MEM_GOING_OFFLINE: | |
1422 | mutex_lock(&cache_chain_mutex); | |
1423 | ret = drain_cache_nodelists_node(nid); | |
1424 | mutex_unlock(&cache_chain_mutex); | |
1425 | break; | |
1426 | case MEM_ONLINE: | |
1427 | case MEM_OFFLINE: | |
1428 | case MEM_CANCEL_ONLINE: | |
1429 | case MEM_CANCEL_OFFLINE: | |
1430 | break; | |
1431 | } | |
1432 | out: | |
5fda1bd5 | 1433 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1434 | } |
1435 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1436 | ||
e498be7d CL |
1437 | /* |
1438 | * swap the static kmem_list3 with kmalloced memory | |
1439 | */ | |
8f9f8d9e DR |
1440 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1441 | int nodeid) | |
e498be7d CL |
1442 | { |
1443 | struct kmem_list3 *ptr; | |
1444 | ||
83b519e8 | 1445 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1446 | BUG_ON(!ptr); |
1447 | ||
e498be7d | 1448 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1449 | /* |
1450 | * Do not assume that spinlocks can be initialized via memcpy: | |
1451 | */ | |
1452 | spin_lock_init(&ptr->list_lock); | |
1453 | ||
e498be7d CL |
1454 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1455 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1456 | } |
1457 | ||
556a169d PE |
1458 | /* |
1459 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1460 | * size of kmem_list3. | |
1461 | */ | |
1462 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1463 | { | |
1464 | int node; | |
1465 | ||
1466 | for_each_online_node(node) { | |
1467 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1468 | cachep->nodelists[node]->next_reap = jiffies + | |
1469 | REAPTIMEOUT_LIST3 + | |
1470 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1471 | } | |
1472 | } | |
1473 | ||
a737b3e2 AM |
1474 | /* |
1475 | * Initialisation. Called after the page allocator have been initialised and | |
1476 | * before smp_init(). | |
1da177e4 LT |
1477 | */ |
1478 | void __init kmem_cache_init(void) | |
1479 | { | |
1480 | size_t left_over; | |
1481 | struct cache_sizes *sizes; | |
1482 | struct cache_names *names; | |
e498be7d | 1483 | int i; |
07ed76b2 | 1484 | int order; |
1ca4cb24 | 1485 | int node; |
e498be7d | 1486 | |
b6e68bc1 | 1487 | if (num_possible_nodes() == 1) |
62918a03 SS |
1488 | use_alien_caches = 0; |
1489 | ||
e498be7d CL |
1490 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1491 | kmem_list3_init(&initkmem_list3[i]); | |
1492 | if (i < MAX_NUMNODES) | |
1493 | cache_cache.nodelists[i] = NULL; | |
1494 | } | |
556a169d | 1495 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1496 | |
1497 | /* | |
1498 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1499 | * page orders on machines with more than 32MB of memory if |
1500 | * not overridden on the command line. | |
1da177e4 | 1501 | */ |
3df1cccd | 1502 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1503 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1504 | |
1da177e4 LT |
1505 | /* Bootstrap is tricky, because several objects are allocated |
1506 | * from caches that do not exist yet: | |
a737b3e2 AM |
1507 | * 1) initialize the cache_cache cache: it contains the struct |
1508 | * kmem_cache structures of all caches, except cache_cache itself: | |
1509 | * cache_cache is statically allocated. | |
e498be7d CL |
1510 | * Initially an __init data area is used for the head array and the |
1511 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1512 | * array at the end of the bootstrap. | |
1da177e4 | 1513 | * 2) Create the first kmalloc cache. |
343e0d7a | 1514 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1515 | * An __init data area is used for the head array. |
1516 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1517 | * head arrays. | |
1da177e4 LT |
1518 | * 4) Replace the __init data head arrays for cache_cache and the first |
1519 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1520 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1521 | * the other cache's with kmalloc allocated memory. | |
1522 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1523 | */ |
1524 | ||
7d6e6d09 | 1525 | node = numa_mem_id(); |
1ca4cb24 | 1526 | |
1da177e4 | 1527 | /* 1) create the cache_cache */ |
1da177e4 | 1528 | INIT_LIST_HEAD(&cache_chain); |
3b0efdfa | 1529 | list_add(&cache_cache.list, &cache_chain); |
1da177e4 LT |
1530 | cache_cache.colour_off = cache_line_size(); |
1531 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1532 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1533 | |
8da3430d | 1534 | /* |
b56efcf0 | 1535 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1536 | */ |
3b0efdfa | 1537 | cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
b56efcf0 | 1538 | nr_node_ids * sizeof(struct kmem_list3 *); |
3b0efdfa CL |
1539 | cache_cache.object_size = cache_cache.size; |
1540 | cache_cache.size = ALIGN(cache_cache.size, | |
a737b3e2 | 1541 | cache_line_size()); |
6a2d7a95 | 1542 | cache_cache.reciprocal_buffer_size = |
3b0efdfa | 1543 | reciprocal_value(cache_cache.size); |
1da177e4 | 1544 | |
07ed76b2 | 1545 | for (order = 0; order < MAX_ORDER; order++) { |
3b0efdfa | 1546 | cache_estimate(order, cache_cache.size, |
07ed76b2 JS |
1547 | cache_line_size(), 0, &left_over, &cache_cache.num); |
1548 | if (cache_cache.num) | |
1549 | break; | |
1550 | } | |
40094fa6 | 1551 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1552 | cache_cache.gfporder = order; |
b28a02de | 1553 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1554 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1555 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1556 | |
1557 | /* 2+3) create the kmalloc caches */ | |
1558 | sizes = malloc_sizes; | |
1559 | names = cache_names; | |
1560 | ||
a737b3e2 AM |
1561 | /* |
1562 | * Initialize the caches that provide memory for the array cache and the | |
1563 | * kmem_list3 structures first. Without this, further allocations will | |
1564 | * bug. | |
e498be7d CL |
1565 | */ |
1566 | ||
1567 | sizes[INDEX_AC].cs_cachep = kmem_cache_create(names[INDEX_AC].name, | |
a737b3e2 AM |
1568 | sizes[INDEX_AC].cs_size, |
1569 | ARCH_KMALLOC_MINALIGN, | |
1570 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1571 | NULL); |
e498be7d | 1572 | |
a737b3e2 | 1573 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1574 | sizes[INDEX_L3].cs_cachep = |
a737b3e2 AM |
1575 | kmem_cache_create(names[INDEX_L3].name, |
1576 | sizes[INDEX_L3].cs_size, | |
1577 | ARCH_KMALLOC_MINALIGN, | |
1578 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1579 | NULL); |
a737b3e2 | 1580 | } |
e498be7d | 1581 | |
e0a42726 IM |
1582 | slab_early_init = 0; |
1583 | ||
1da177e4 | 1584 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1585 | /* |
1586 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1587 | * This should be particularly beneficial on SMP boxes, as it |
1588 | * eliminates "false sharing". | |
1589 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1590 | * allow tighter packing of the smaller caches. |
1591 | */ | |
a737b3e2 | 1592 | if (!sizes->cs_cachep) { |
e498be7d | 1593 | sizes->cs_cachep = kmem_cache_create(names->name, |
a737b3e2 AM |
1594 | sizes->cs_size, |
1595 | ARCH_KMALLOC_MINALIGN, | |
1596 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1597 | NULL); |
a737b3e2 | 1598 | } |
4b51d669 CL |
1599 | #ifdef CONFIG_ZONE_DMA |
1600 | sizes->cs_dmacachep = kmem_cache_create( | |
1601 | names->name_dma, | |
a737b3e2 AM |
1602 | sizes->cs_size, |
1603 | ARCH_KMALLOC_MINALIGN, | |
1604 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1605 | SLAB_PANIC, | |
20c2df83 | 1606 | NULL); |
4b51d669 | 1607 | #endif |
1da177e4 LT |
1608 | sizes++; |
1609 | names++; | |
1610 | } | |
1611 | /* 4) Replace the bootstrap head arrays */ | |
1612 | { | |
2b2d5493 | 1613 | struct array_cache *ptr; |
e498be7d | 1614 | |
83b519e8 | 1615 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1616 | |
9a2dba4b PE |
1617 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1618 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1619 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1620 | /* |
1621 | * Do not assume that spinlocks can be initialized via memcpy: | |
1622 | */ | |
1623 | spin_lock_init(&ptr->lock); | |
1624 | ||
1da177e4 | 1625 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1626 | |
83b519e8 | 1627 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1628 | |
9a2dba4b | 1629 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1630 | != &initarray_generic.cache); |
9a2dba4b | 1631 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1632 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1633 | /* |
1634 | * Do not assume that spinlocks can be initialized via memcpy: | |
1635 | */ | |
1636 | spin_lock_init(&ptr->lock); | |
1637 | ||
e498be7d | 1638 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1639 | ptr; |
1da177e4 | 1640 | } |
e498be7d CL |
1641 | /* 5) Replace the bootstrap kmem_list3's */ |
1642 | { | |
1ca4cb24 PE |
1643 | int nid; |
1644 | ||
9c09a95c | 1645 | for_each_online_node(nid) { |
ec1f5eee | 1646 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1647 | |
e498be7d | 1648 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1649 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1650 | |
1651 | if (INDEX_AC != INDEX_L3) { | |
1652 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1653 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1654 | } |
1655 | } | |
1656 | } | |
1da177e4 | 1657 | |
8429db5c | 1658 | g_cpucache_up = EARLY; |
8429db5c PE |
1659 | } |
1660 | ||
1661 | void __init kmem_cache_init_late(void) | |
1662 | { | |
1663 | struct kmem_cache *cachep; | |
1664 | ||
52cef189 PZ |
1665 | g_cpucache_up = LATE; |
1666 | ||
30765b92 PZ |
1667 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1668 | init_lock_keys(); | |
1669 | ||
8429db5c PE |
1670 | /* 6) resize the head arrays to their final sizes */ |
1671 | mutex_lock(&cache_chain_mutex); | |
3b0efdfa | 1672 | list_for_each_entry(cachep, &cache_chain, list) |
8429db5c PE |
1673 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1674 | BUG(); | |
1675 | mutex_unlock(&cache_chain_mutex); | |
056c6241 | 1676 | |
1da177e4 LT |
1677 | /* Done! */ |
1678 | g_cpucache_up = FULL; | |
1679 | ||
a737b3e2 AM |
1680 | /* |
1681 | * Register a cpu startup notifier callback that initializes | |
1682 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1683 | */ |
1684 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1685 | |
8f9f8d9e DR |
1686 | #ifdef CONFIG_NUMA |
1687 | /* | |
1688 | * Register a memory hotplug callback that initializes and frees | |
1689 | * nodelists. | |
1690 | */ | |
1691 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1692 | #endif | |
1693 | ||
a737b3e2 AM |
1694 | /* |
1695 | * The reap timers are started later, with a module init call: That part | |
1696 | * of the kernel is not yet operational. | |
1da177e4 LT |
1697 | */ |
1698 | } | |
1699 | ||
1700 | static int __init cpucache_init(void) | |
1701 | { | |
1702 | int cpu; | |
1703 | ||
a737b3e2 AM |
1704 | /* |
1705 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1706 | */ |
e498be7d | 1707 | for_each_online_cpu(cpu) |
a737b3e2 | 1708 | start_cpu_timer(cpu); |
1da177e4 LT |
1709 | return 0; |
1710 | } | |
1da177e4 LT |
1711 | __initcall(cpucache_init); |
1712 | ||
8bdec192 RA |
1713 | static noinline void |
1714 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1715 | { | |
1716 | struct kmem_list3 *l3; | |
1717 | struct slab *slabp; | |
1718 | unsigned long flags; | |
1719 | int node; | |
1720 | ||
1721 | printk(KERN_WARNING | |
1722 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1723 | nodeid, gfpflags); | |
1724 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1725 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1726 | |
1727 | for_each_online_node(node) { | |
1728 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1729 | unsigned long active_slabs = 0, num_slabs = 0; | |
1730 | ||
1731 | l3 = cachep->nodelists[node]; | |
1732 | if (!l3) | |
1733 | continue; | |
1734 | ||
1735 | spin_lock_irqsave(&l3->list_lock, flags); | |
1736 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1737 | active_objs += cachep->num; | |
1738 | active_slabs++; | |
1739 | } | |
1740 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1741 | active_objs += slabp->inuse; | |
1742 | active_slabs++; | |
1743 | } | |
1744 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1745 | num_slabs++; | |
1746 | ||
1747 | free_objects += l3->free_objects; | |
1748 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1749 | ||
1750 | num_slabs += active_slabs; | |
1751 | num_objs = num_slabs * cachep->num; | |
1752 | printk(KERN_WARNING | |
1753 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1754 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1755 | free_objects); | |
1756 | } | |
1757 | } | |
1758 | ||
1da177e4 LT |
1759 | /* |
1760 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1761 | * | |
1762 | * If we requested dmaable memory, we will get it. Even if we | |
1763 | * did not request dmaable memory, we might get it, but that | |
1764 | * would be relatively rare and ignorable. | |
1765 | */ | |
343e0d7a | 1766 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1767 | { |
1768 | struct page *page; | |
e1b6aa6f | 1769 | int nr_pages; |
1da177e4 LT |
1770 | int i; |
1771 | ||
d6fef9da | 1772 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1773 | /* |
1774 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1775 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1776 | */ |
e1b6aa6f | 1777 | flags |= __GFP_COMP; |
d6fef9da | 1778 | #endif |
765c4507 | 1779 | |
3c517a61 | 1780 | flags |= cachep->gfpflags; |
e12ba74d MG |
1781 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1782 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1783 | |
517d0869 | 1784 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1785 | if (!page) { |
1786 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1787 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1788 | return NULL; |
8bdec192 | 1789 | } |
1da177e4 | 1790 | |
e1b6aa6f | 1791 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1792 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1793 | add_zone_page_state(page_zone(page), |
1794 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1795 | else | |
1796 | add_zone_page_state(page_zone(page), | |
1797 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1798 | for (i = 0; i < nr_pages; i++) |
1799 | __SetPageSlab(page + i); | |
c175eea4 | 1800 | |
b1eeab67 VN |
1801 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1802 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1803 | ||
1804 | if (cachep->ctor) | |
1805 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1806 | else | |
1807 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1808 | } | |
c175eea4 | 1809 | |
e1b6aa6f | 1810 | return page_address(page); |
1da177e4 LT |
1811 | } |
1812 | ||
1813 | /* | |
1814 | * Interface to system's page release. | |
1815 | */ | |
343e0d7a | 1816 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1817 | { |
b28a02de | 1818 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1819 | struct page *page = virt_to_page(addr); |
1820 | const unsigned long nr_freed = i; | |
1821 | ||
b1eeab67 | 1822 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1823 | |
972d1a7b CL |
1824 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1825 | sub_zone_page_state(page_zone(page), | |
1826 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1827 | else | |
1828 | sub_zone_page_state(page_zone(page), | |
1829 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1830 | while (i--) { |
f205b2fe NP |
1831 | BUG_ON(!PageSlab(page)); |
1832 | __ClearPageSlab(page); | |
1da177e4 LT |
1833 | page++; |
1834 | } | |
1da177e4 LT |
1835 | if (current->reclaim_state) |
1836 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1837 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1838 | } |
1839 | ||
1840 | static void kmem_rcu_free(struct rcu_head *head) | |
1841 | { | |
b28a02de | 1842 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1843 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1844 | |
1845 | kmem_freepages(cachep, slab_rcu->addr); | |
1846 | if (OFF_SLAB(cachep)) | |
1847 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1848 | } | |
1849 | ||
1850 | #if DEBUG | |
1851 | ||
1852 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1853 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1854 | unsigned long caller) |
1da177e4 | 1855 | { |
3dafccf2 | 1856 | int size = obj_size(cachep); |
1da177e4 | 1857 | |
3dafccf2 | 1858 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1859 | |
b28a02de | 1860 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1861 | return; |
1862 | ||
b28a02de PE |
1863 | *addr++ = 0x12345678; |
1864 | *addr++ = caller; | |
1865 | *addr++ = smp_processor_id(); | |
1866 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1867 | { |
1868 | unsigned long *sptr = &caller; | |
1869 | unsigned long svalue; | |
1870 | ||
1871 | while (!kstack_end(sptr)) { | |
1872 | svalue = *sptr++; | |
1873 | if (kernel_text_address(svalue)) { | |
b28a02de | 1874 | *addr++ = svalue; |
1da177e4 LT |
1875 | size -= sizeof(unsigned long); |
1876 | if (size <= sizeof(unsigned long)) | |
1877 | break; | |
1878 | } | |
1879 | } | |
1880 | ||
1881 | } | |
b28a02de | 1882 | *addr++ = 0x87654321; |
1da177e4 LT |
1883 | } |
1884 | #endif | |
1885 | ||
343e0d7a | 1886 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1887 | { |
3dafccf2 MS |
1888 | int size = obj_size(cachep); |
1889 | addr = &((char *)addr)[obj_offset(cachep)]; | |
1da177e4 LT |
1890 | |
1891 | memset(addr, val, size); | |
b28a02de | 1892 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1893 | } |
1894 | ||
1895 | static void dump_line(char *data, int offset, int limit) | |
1896 | { | |
1897 | int i; | |
aa83aa40 DJ |
1898 | unsigned char error = 0; |
1899 | int bad_count = 0; | |
1900 | ||
fdde6abb | 1901 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
1902 | for (i = 0; i < limit; i++) { |
1903 | if (data[offset + i] != POISON_FREE) { | |
1904 | error = data[offset + i]; | |
1905 | bad_count++; | |
1906 | } | |
aa83aa40 | 1907 | } |
fdde6abb SAS |
1908 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1909 | &data[offset], limit, 1); | |
aa83aa40 DJ |
1910 | |
1911 | if (bad_count == 1) { | |
1912 | error ^= POISON_FREE; | |
1913 | if (!(error & (error - 1))) { | |
1914 | printk(KERN_ERR "Single bit error detected. Probably " | |
1915 | "bad RAM.\n"); | |
1916 | #ifdef CONFIG_X86 | |
1917 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1918 | "test tool.\n"); | |
1919 | #else | |
1920 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1921 | #endif | |
1922 | } | |
1923 | } | |
1da177e4 LT |
1924 | } |
1925 | #endif | |
1926 | ||
1927 | #if DEBUG | |
1928 | ||
343e0d7a | 1929 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1930 | { |
1931 | int i, size; | |
1932 | char *realobj; | |
1933 | ||
1934 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1935 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1936 | *dbg_redzone1(cachep, objp), |
1937 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1938 | } |
1939 | ||
1940 | if (cachep->flags & SLAB_STORE_USER) { | |
1941 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1942 | *dbg_userword(cachep, objp)); |
1da177e4 | 1943 | print_symbol("(%s)", |
a737b3e2 | 1944 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1945 | printk("\n"); |
1946 | } | |
3dafccf2 MS |
1947 | realobj = (char *)objp + obj_offset(cachep); |
1948 | size = obj_size(cachep); | |
b28a02de | 1949 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1950 | int limit; |
1951 | limit = 16; | |
b28a02de PE |
1952 | if (i + limit > size) |
1953 | limit = size - i; | |
1da177e4 LT |
1954 | dump_line(realobj, i, limit); |
1955 | } | |
1956 | } | |
1957 | ||
343e0d7a | 1958 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1959 | { |
1960 | char *realobj; | |
1961 | int size, i; | |
1962 | int lines = 0; | |
1963 | ||
3dafccf2 MS |
1964 | realobj = (char *)objp + obj_offset(cachep); |
1965 | size = obj_size(cachep); | |
1da177e4 | 1966 | |
b28a02de | 1967 | for (i = 0; i < size; i++) { |
1da177e4 | 1968 | char exp = POISON_FREE; |
b28a02de | 1969 | if (i == size - 1) |
1da177e4 LT |
1970 | exp = POISON_END; |
1971 | if (realobj[i] != exp) { | |
1972 | int limit; | |
1973 | /* Mismatch ! */ | |
1974 | /* Print header */ | |
1975 | if (lines == 0) { | |
b28a02de | 1976 | printk(KERN_ERR |
face37f5 DJ |
1977 | "Slab corruption (%s): %s start=%p, len=%d\n", |
1978 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
1979 | print_objinfo(cachep, objp, 0); |
1980 | } | |
1981 | /* Hexdump the affected line */ | |
b28a02de | 1982 | i = (i / 16) * 16; |
1da177e4 | 1983 | limit = 16; |
b28a02de PE |
1984 | if (i + limit > size) |
1985 | limit = size - i; | |
1da177e4 LT |
1986 | dump_line(realobj, i, limit); |
1987 | i += 16; | |
1988 | lines++; | |
1989 | /* Limit to 5 lines */ | |
1990 | if (lines > 5) | |
1991 | break; | |
1992 | } | |
1993 | } | |
1994 | if (lines != 0) { | |
1995 | /* Print some data about the neighboring objects, if they | |
1996 | * exist: | |
1997 | */ | |
6ed5eb22 | 1998 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1999 | unsigned int objnr; |
1da177e4 | 2000 | |
8fea4e96 | 2001 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2002 | if (objnr) { |
8fea4e96 | 2003 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2004 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2005 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2006 | realobj, size); |
1da177e4 LT |
2007 | print_objinfo(cachep, objp, 2); |
2008 | } | |
b28a02de | 2009 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2010 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2011 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2012 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2013 | realobj, size); |
1da177e4 LT |
2014 | print_objinfo(cachep, objp, 2); |
2015 | } | |
2016 | } | |
2017 | } | |
2018 | #endif | |
2019 | ||
12dd36fa | 2020 | #if DEBUG |
e79aec29 | 2021 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2022 | { |
1da177e4 LT |
2023 | int i; |
2024 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2025 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2026 | |
2027 | if (cachep->flags & SLAB_POISON) { | |
2028 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2029 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2030 | OFF_SLAB(cachep)) |
b28a02de | 2031 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2032 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2033 | else |
2034 | check_poison_obj(cachep, objp); | |
2035 | #else | |
2036 | check_poison_obj(cachep, objp); | |
2037 | #endif | |
2038 | } | |
2039 | if (cachep->flags & SLAB_RED_ZONE) { | |
2040 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2041 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2042 | "was overwritten"); |
1da177e4 LT |
2043 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2044 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2045 | "was overwritten"); |
1da177e4 | 2046 | } |
1da177e4 | 2047 | } |
12dd36fa | 2048 | } |
1da177e4 | 2049 | #else |
e79aec29 | 2050 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2051 | { |
12dd36fa | 2052 | } |
1da177e4 LT |
2053 | #endif |
2054 | ||
911851e6 RD |
2055 | /** |
2056 | * slab_destroy - destroy and release all objects in a slab | |
2057 | * @cachep: cache pointer being destroyed | |
2058 | * @slabp: slab pointer being destroyed | |
2059 | * | |
12dd36fa | 2060 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2061 | * Before calling the slab must have been unlinked from the cache. The |
2062 | * cache-lock is not held/needed. | |
12dd36fa | 2063 | */ |
343e0d7a | 2064 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2065 | { |
2066 | void *addr = slabp->s_mem - slabp->colouroff; | |
2067 | ||
e79aec29 | 2068 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2069 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2070 | struct slab_rcu *slab_rcu; | |
2071 | ||
b28a02de | 2072 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2073 | slab_rcu->cachep = cachep; |
2074 | slab_rcu->addr = addr; | |
2075 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2076 | } else { | |
2077 | kmem_freepages(cachep, addr); | |
873623df IM |
2078 | if (OFF_SLAB(cachep)) |
2079 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2080 | } |
2081 | } | |
2082 | ||
117f6eb1 CL |
2083 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
2084 | { | |
2085 | int i; | |
2086 | struct kmem_list3 *l3; | |
2087 | ||
2088 | for_each_online_cpu(i) | |
2089 | kfree(cachep->array[i]); | |
2090 | ||
2091 | /* NUMA: free the list3 structures */ | |
2092 | for_each_online_node(i) { | |
2093 | l3 = cachep->nodelists[i]; | |
2094 | if (l3) { | |
2095 | kfree(l3->shared); | |
2096 | free_alien_cache(l3->alien); | |
2097 | kfree(l3); | |
2098 | } | |
2099 | } | |
2100 | kmem_cache_free(&cache_cache, cachep); | |
2101 | } | |
2102 | ||
2103 | ||
4d268eba | 2104 | /** |
a70773dd RD |
2105 | * calculate_slab_order - calculate size (page order) of slabs |
2106 | * @cachep: pointer to the cache that is being created | |
2107 | * @size: size of objects to be created in this cache. | |
2108 | * @align: required alignment for the objects. | |
2109 | * @flags: slab allocation flags | |
2110 | * | |
2111 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2112 | * |
2113 | * This could be made much more intelligent. For now, try to avoid using | |
2114 | * high order pages for slabs. When the gfp() functions are more friendly | |
2115 | * towards high-order requests, this should be changed. | |
2116 | */ | |
a737b3e2 | 2117 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2118 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2119 | { |
b1ab41c4 | 2120 | unsigned long offslab_limit; |
4d268eba | 2121 | size_t left_over = 0; |
9888e6fa | 2122 | int gfporder; |
4d268eba | 2123 | |
0aa817f0 | 2124 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2125 | unsigned int num; |
2126 | size_t remainder; | |
2127 | ||
9888e6fa | 2128 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2129 | if (!num) |
2130 | continue; | |
9888e6fa | 2131 | |
b1ab41c4 IM |
2132 | if (flags & CFLGS_OFF_SLAB) { |
2133 | /* | |
2134 | * Max number of objs-per-slab for caches which | |
2135 | * use off-slab slabs. Needed to avoid a possible | |
2136 | * looping condition in cache_grow(). | |
2137 | */ | |
2138 | offslab_limit = size - sizeof(struct slab); | |
2139 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2140 | ||
2141 | if (num > offslab_limit) | |
2142 | break; | |
2143 | } | |
4d268eba | 2144 | |
9888e6fa | 2145 | /* Found something acceptable - save it away */ |
4d268eba | 2146 | cachep->num = num; |
9888e6fa | 2147 | cachep->gfporder = gfporder; |
4d268eba PE |
2148 | left_over = remainder; |
2149 | ||
f78bb8ad LT |
2150 | /* |
2151 | * A VFS-reclaimable slab tends to have most allocations | |
2152 | * as GFP_NOFS and we really don't want to have to be allocating | |
2153 | * higher-order pages when we are unable to shrink dcache. | |
2154 | */ | |
2155 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2156 | break; | |
2157 | ||
4d268eba PE |
2158 | /* |
2159 | * Large number of objects is good, but very large slabs are | |
2160 | * currently bad for the gfp()s. | |
2161 | */ | |
543585cc | 2162 | if (gfporder >= slab_max_order) |
4d268eba PE |
2163 | break; |
2164 | ||
9888e6fa LT |
2165 | /* |
2166 | * Acceptable internal fragmentation? | |
2167 | */ | |
a737b3e2 | 2168 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2169 | break; |
2170 | } | |
2171 | return left_over; | |
2172 | } | |
2173 | ||
83b519e8 | 2174 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2175 | { |
2ed3a4ef | 2176 | if (g_cpucache_up == FULL) |
83b519e8 | 2177 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2178 | |
f30cf7d1 PE |
2179 | if (g_cpucache_up == NONE) { |
2180 | /* | |
2181 | * Note: the first kmem_cache_create must create the cache | |
2182 | * that's used by kmalloc(24), otherwise the creation of | |
2183 | * further caches will BUG(). | |
2184 | */ | |
2185 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2186 | ||
2187 | /* | |
2188 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2189 | * the first cache, then we need to set up all its list3s, | |
2190 | * otherwise the creation of further caches will BUG(). | |
2191 | */ | |
2192 | set_up_list3s(cachep, SIZE_AC); | |
2193 | if (INDEX_AC == INDEX_L3) | |
2194 | g_cpucache_up = PARTIAL_L3; | |
2195 | else | |
2196 | g_cpucache_up = PARTIAL_AC; | |
2197 | } else { | |
2198 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2199 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 PE |
2200 | |
2201 | if (g_cpucache_up == PARTIAL_AC) { | |
2202 | set_up_list3s(cachep, SIZE_L3); | |
2203 | g_cpucache_up = PARTIAL_L3; | |
2204 | } else { | |
2205 | int node; | |
556a169d | 2206 | for_each_online_node(node) { |
f30cf7d1 PE |
2207 | cachep->nodelists[node] = |
2208 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2209 | gfp, node); |
f30cf7d1 PE |
2210 | BUG_ON(!cachep->nodelists[node]); |
2211 | kmem_list3_init(cachep->nodelists[node]); | |
2212 | } | |
2213 | } | |
2214 | } | |
7d6e6d09 | 2215 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2216 | jiffies + REAPTIMEOUT_LIST3 + |
2217 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2218 | ||
2219 | cpu_cache_get(cachep)->avail = 0; | |
2220 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2221 | cpu_cache_get(cachep)->batchcount = 1; | |
2222 | cpu_cache_get(cachep)->touched = 0; | |
2223 | cachep->batchcount = 1; | |
2224 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2225 | return 0; |
f30cf7d1 PE |
2226 | } |
2227 | ||
1da177e4 LT |
2228 | /** |
2229 | * kmem_cache_create - Create a cache. | |
2230 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2231 | * @size: The size of objects to be created in this cache. | |
2232 | * @align: The required alignment for the objects. | |
2233 | * @flags: SLAB flags | |
2234 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2235 | * |
2236 | * Returns a ptr to the cache on success, NULL on failure. | |
2237 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2238 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2239 | * |
2240 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2241 | * the module calling this has to destroy the cache before getting unloaded. |
2242 | * | |
1da177e4 LT |
2243 | * The flags are |
2244 | * | |
2245 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2246 | * to catch references to uninitialised memory. | |
2247 | * | |
2248 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2249 | * for buffer overruns. | |
2250 | * | |
1da177e4 LT |
2251 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2252 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2253 | * as davem. | |
2254 | */ | |
343e0d7a | 2255 | struct kmem_cache * |
1da177e4 | 2256 | kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2257 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2258 | { |
2259 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2260 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2261 | gfp_t gfp; |
1da177e4 LT |
2262 | |
2263 | /* | |
2264 | * Sanity checks... these are all serious usage bugs. | |
2265 | */ | |
a737b3e2 | 2266 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2267 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2268 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2269 | name); |
b28a02de PE |
2270 | BUG(); |
2271 | } | |
1da177e4 | 2272 | |
f0188f47 | 2273 | /* |
8f5be20b | 2274 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2275 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2276 | */ |
83b519e8 PE |
2277 | if (slab_is_available()) { |
2278 | get_online_cpus(); | |
2279 | mutex_lock(&cache_chain_mutex); | |
2280 | } | |
4f12bb4f | 2281 | |
3b0efdfa | 2282 | list_for_each_entry(pc, &cache_chain, list) { |
4f12bb4f AM |
2283 | char tmp; |
2284 | int res; | |
2285 | ||
2286 | /* | |
2287 | * This happens when the module gets unloaded and doesn't | |
2288 | * destroy its slab cache and no-one else reuses the vmalloc | |
2289 | * area of the module. Print a warning. | |
2290 | */ | |
138ae663 | 2291 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2292 | if (res) { |
b4169525 | 2293 | printk(KERN_ERR |
2294 | "SLAB: cache with size %d has lost its name\n", | |
3b0efdfa | 2295 | pc->size); |
4f12bb4f AM |
2296 | continue; |
2297 | } | |
2298 | ||
b28a02de | 2299 | if (!strcmp(pc->name, name)) { |
b4169525 | 2300 | printk(KERN_ERR |
2301 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2302 | dump_stack(); |
2303 | goto oops; | |
2304 | } | |
2305 | } | |
2306 | ||
1da177e4 LT |
2307 | #if DEBUG |
2308 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2309 | #if FORCED_DEBUG |
2310 | /* | |
2311 | * Enable redzoning and last user accounting, except for caches with | |
2312 | * large objects, if the increased size would increase the object size | |
2313 | * above the next power of two: caches with object sizes just above a | |
2314 | * power of two have a significant amount of internal fragmentation. | |
2315 | */ | |
87a927c7 DW |
2316 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2317 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2318 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2319 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2320 | flags |= SLAB_POISON; | |
2321 | #endif | |
2322 | if (flags & SLAB_DESTROY_BY_RCU) | |
2323 | BUG_ON(flags & SLAB_POISON); | |
2324 | #endif | |
1da177e4 | 2325 | /* |
a737b3e2 AM |
2326 | * Always checks flags, a caller might be expecting debug support which |
2327 | * isn't available. | |
1da177e4 | 2328 | */ |
40094fa6 | 2329 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2330 | |
a737b3e2 AM |
2331 | /* |
2332 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2333 | * unaligned accesses for some archs when redzoning is used, and makes |
2334 | * sure any on-slab bufctl's are also correctly aligned. | |
2335 | */ | |
b28a02de PE |
2336 | if (size & (BYTES_PER_WORD - 1)) { |
2337 | size += (BYTES_PER_WORD - 1); | |
2338 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2339 | } |
2340 | ||
a737b3e2 AM |
2341 | /* calculate the final buffer alignment: */ |
2342 | ||
1da177e4 LT |
2343 | /* 1) arch recommendation: can be overridden for debug */ |
2344 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2345 | /* |
2346 | * Default alignment: as specified by the arch code. Except if | |
2347 | * an object is really small, then squeeze multiple objects into | |
2348 | * one cacheline. | |
1da177e4 LT |
2349 | */ |
2350 | ralign = cache_line_size(); | |
b28a02de | 2351 | while (size <= ralign / 2) |
1da177e4 LT |
2352 | ralign /= 2; |
2353 | } else { | |
2354 | ralign = BYTES_PER_WORD; | |
2355 | } | |
ca5f9703 PE |
2356 | |
2357 | /* | |
87a927c7 DW |
2358 | * Redzoning and user store require word alignment or possibly larger. |
2359 | * Note this will be overridden by architecture or caller mandated | |
2360 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2361 | */ |
87a927c7 DW |
2362 | if (flags & SLAB_STORE_USER) |
2363 | ralign = BYTES_PER_WORD; | |
2364 | ||
2365 | if (flags & SLAB_RED_ZONE) { | |
2366 | ralign = REDZONE_ALIGN; | |
2367 | /* If redzoning, ensure that the second redzone is suitably | |
2368 | * aligned, by adjusting the object size accordingly. */ | |
2369 | size += REDZONE_ALIGN - 1; | |
2370 | size &= ~(REDZONE_ALIGN - 1); | |
2371 | } | |
ca5f9703 | 2372 | |
a44b56d3 | 2373 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2374 | if (ralign < ARCH_SLAB_MINALIGN) { |
2375 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2376 | } |
a44b56d3 | 2377 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2378 | if (ralign < align) { |
2379 | ralign = align; | |
1da177e4 | 2380 | } |
3ff84a7f PE |
2381 | /* disable debug if necessary */ |
2382 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2383 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2384 | /* |
ca5f9703 | 2385 | * 4) Store it. |
1da177e4 LT |
2386 | */ |
2387 | align = ralign; | |
2388 | ||
83b519e8 PE |
2389 | if (slab_is_available()) |
2390 | gfp = GFP_KERNEL; | |
2391 | else | |
2392 | gfp = GFP_NOWAIT; | |
2393 | ||
1da177e4 | 2394 | /* Get cache's description obj. */ |
83b519e8 | 2395 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2396 | if (!cachep) |
4f12bb4f | 2397 | goto oops; |
1da177e4 | 2398 | |
b56efcf0 | 2399 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
3b0efdfa CL |
2400 | cachep->object_size = size; |
2401 | cachep->align = align; | |
1da177e4 | 2402 | #if DEBUG |
1da177e4 | 2403 | |
ca5f9703 PE |
2404 | /* |
2405 | * Both debugging options require word-alignment which is calculated | |
2406 | * into align above. | |
2407 | */ | |
1da177e4 | 2408 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2409 | /* add space for red zone words */ |
3ff84a7f PE |
2410 | cachep->obj_offset += sizeof(unsigned long long); |
2411 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2412 | } |
2413 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2414 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2415 | * the real object. But if the second red zone needs to be |
2416 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2417 | */ |
87a927c7 DW |
2418 | if (flags & SLAB_RED_ZONE) |
2419 | size += REDZONE_ALIGN; | |
2420 | else | |
2421 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2422 | } |
2423 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2424 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3b0efdfa | 2425 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
1ab335d8 | 2426 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); |
1da177e4 LT |
2427 | size = PAGE_SIZE; |
2428 | } | |
2429 | #endif | |
2430 | #endif | |
2431 | ||
e0a42726 IM |
2432 | /* |
2433 | * Determine if the slab management is 'on' or 'off' slab. | |
2434 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2435 | * it too early on. Always use on-slab management when |
2436 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2437 | */ |
e7cb55b9 CM |
2438 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2439 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2440 | /* |
2441 | * Size is large, assume best to place the slab management obj | |
2442 | * off-slab (should allow better packing of objs). | |
2443 | */ | |
2444 | flags |= CFLGS_OFF_SLAB; | |
2445 | ||
2446 | size = ALIGN(size, align); | |
2447 | ||
f78bb8ad | 2448 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2449 | |
2450 | if (!cachep->num) { | |
b4169525 | 2451 | printk(KERN_ERR |
2452 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 LT |
2453 | kmem_cache_free(&cache_cache, cachep); |
2454 | cachep = NULL; | |
4f12bb4f | 2455 | goto oops; |
1da177e4 | 2456 | } |
b28a02de PE |
2457 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2458 | + sizeof(struct slab), align); | |
1da177e4 LT |
2459 | |
2460 | /* | |
2461 | * If the slab has been placed off-slab, and we have enough space then | |
2462 | * move it on-slab. This is at the expense of any extra colouring. | |
2463 | */ | |
2464 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2465 | flags &= ~CFLGS_OFF_SLAB; | |
2466 | left_over -= slab_size; | |
2467 | } | |
2468 | ||
2469 | if (flags & CFLGS_OFF_SLAB) { | |
2470 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2471 | slab_size = |
2472 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2473 | |
2474 | #ifdef CONFIG_PAGE_POISONING | |
2475 | /* If we're going to use the generic kernel_map_pages() | |
2476 | * poisoning, then it's going to smash the contents of | |
2477 | * the redzone and userword anyhow, so switch them off. | |
2478 | */ | |
2479 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2480 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2481 | #endif | |
1da177e4 LT |
2482 | } |
2483 | ||
2484 | cachep->colour_off = cache_line_size(); | |
2485 | /* Offset must be a multiple of the alignment. */ | |
2486 | if (cachep->colour_off < align) | |
2487 | cachep->colour_off = align; | |
b28a02de | 2488 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2489 | cachep->slab_size = slab_size; |
2490 | cachep->flags = flags; | |
2491 | cachep->gfpflags = 0; | |
4b51d669 | 2492 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
1da177e4 | 2493 | cachep->gfpflags |= GFP_DMA; |
3b0efdfa | 2494 | cachep->size = size; |
6a2d7a95 | 2495 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2496 | |
e5ac9c5a | 2497 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2498 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2499 | /* |
2500 | * This is a possibility for one of the malloc_sizes caches. | |
2501 | * But since we go off slab only for object size greater than | |
2502 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2503 | * this should not happen at all. | |
2504 | * But leave a BUG_ON for some lucky dude. | |
2505 | */ | |
6cb8f913 | 2506 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2507 | } |
1da177e4 | 2508 | cachep->ctor = ctor; |
1da177e4 LT |
2509 | cachep->name = name; |
2510 | ||
83b519e8 | 2511 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef CL |
2512 | __kmem_cache_destroy(cachep); |
2513 | cachep = NULL; | |
2514 | goto oops; | |
2515 | } | |
1da177e4 | 2516 | |
83835b3d PZ |
2517 | if (flags & SLAB_DEBUG_OBJECTS) { |
2518 | /* | |
2519 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2520 | * debug_object_activate()->kmem_cache_alloc(). | |
2521 | */ | |
2522 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2523 | ||
2524 | slab_set_debugobj_lock_classes(cachep); | |
2525 | } | |
2526 | ||
1da177e4 | 2527 | /* cache setup completed, link it into the list */ |
3b0efdfa | 2528 | list_add(&cachep->list, &cache_chain); |
a737b3e2 | 2529 | oops: |
1da177e4 LT |
2530 | if (!cachep && (flags & SLAB_PANIC)) |
2531 | panic("kmem_cache_create(): failed to create slab `%s'\n", | |
b28a02de | 2532 | name); |
83b519e8 PE |
2533 | if (slab_is_available()) { |
2534 | mutex_unlock(&cache_chain_mutex); | |
2535 | put_online_cpus(); | |
2536 | } | |
1da177e4 LT |
2537 | return cachep; |
2538 | } | |
2539 | EXPORT_SYMBOL(kmem_cache_create); | |
2540 | ||
2541 | #if DEBUG | |
2542 | static void check_irq_off(void) | |
2543 | { | |
2544 | BUG_ON(!irqs_disabled()); | |
2545 | } | |
2546 | ||
2547 | static void check_irq_on(void) | |
2548 | { | |
2549 | BUG_ON(irqs_disabled()); | |
2550 | } | |
2551 | ||
343e0d7a | 2552 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2553 | { |
2554 | #ifdef CONFIG_SMP | |
2555 | check_irq_off(); | |
7d6e6d09 | 2556 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2557 | #endif |
2558 | } | |
e498be7d | 2559 | |
343e0d7a | 2560 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2561 | { |
2562 | #ifdef CONFIG_SMP | |
2563 | check_irq_off(); | |
2564 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2565 | #endif | |
2566 | } | |
2567 | ||
1da177e4 LT |
2568 | #else |
2569 | #define check_irq_off() do { } while(0) | |
2570 | #define check_irq_on() do { } while(0) | |
2571 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2572 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2573 | #endif |
2574 | ||
aab2207c CL |
2575 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2576 | struct array_cache *ac, | |
2577 | int force, int node); | |
2578 | ||
1da177e4 LT |
2579 | static void do_drain(void *arg) |
2580 | { | |
a737b3e2 | 2581 | struct kmem_cache *cachep = arg; |
1da177e4 | 2582 | struct array_cache *ac; |
7d6e6d09 | 2583 | int node = numa_mem_id(); |
1da177e4 LT |
2584 | |
2585 | check_irq_off(); | |
9a2dba4b | 2586 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2587 | spin_lock(&cachep->nodelists[node]->list_lock); |
2588 | free_block(cachep, ac->entry, ac->avail, node); | |
2589 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2590 | ac->avail = 0; |
2591 | } | |
2592 | ||
343e0d7a | 2593 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2594 | { |
e498be7d CL |
2595 | struct kmem_list3 *l3; |
2596 | int node; | |
2597 | ||
15c8b6c1 | 2598 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2599 | check_irq_on(); |
b28a02de | 2600 | for_each_online_node(node) { |
e498be7d | 2601 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2602 | if (l3 && l3->alien) |
2603 | drain_alien_cache(cachep, l3->alien); | |
2604 | } | |
2605 | ||
2606 | for_each_online_node(node) { | |
2607 | l3 = cachep->nodelists[node]; | |
2608 | if (l3) | |
aab2207c | 2609 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2610 | } |
1da177e4 LT |
2611 | } |
2612 | ||
ed11d9eb CL |
2613 | /* |
2614 | * Remove slabs from the list of free slabs. | |
2615 | * Specify the number of slabs to drain in tofree. | |
2616 | * | |
2617 | * Returns the actual number of slabs released. | |
2618 | */ | |
2619 | static int drain_freelist(struct kmem_cache *cache, | |
2620 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2621 | { |
ed11d9eb CL |
2622 | struct list_head *p; |
2623 | int nr_freed; | |
1da177e4 | 2624 | struct slab *slabp; |
1da177e4 | 2625 | |
ed11d9eb CL |
2626 | nr_freed = 0; |
2627 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2628 | |
ed11d9eb | 2629 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2630 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2631 | if (p == &l3->slabs_free) { |
2632 | spin_unlock_irq(&l3->list_lock); | |
2633 | goto out; | |
2634 | } | |
1da177e4 | 2635 | |
ed11d9eb | 2636 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2637 | #if DEBUG |
40094fa6 | 2638 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2639 | #endif |
2640 | list_del(&slabp->list); | |
ed11d9eb CL |
2641 | /* |
2642 | * Safe to drop the lock. The slab is no longer linked | |
2643 | * to the cache. | |
2644 | */ | |
2645 | l3->free_objects -= cache->num; | |
e498be7d | 2646 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2647 | slab_destroy(cache, slabp); |
2648 | nr_freed++; | |
1da177e4 | 2649 | } |
ed11d9eb CL |
2650 | out: |
2651 | return nr_freed; | |
1da177e4 LT |
2652 | } |
2653 | ||
8f5be20b | 2654 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2655 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2656 | { |
2657 | int ret = 0, i = 0; | |
2658 | struct kmem_list3 *l3; | |
2659 | ||
2660 | drain_cpu_caches(cachep); | |
2661 | ||
2662 | check_irq_on(); | |
2663 | for_each_online_node(i) { | |
2664 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2665 | if (!l3) |
2666 | continue; | |
2667 | ||
2668 | drain_freelist(cachep, l3, l3->free_objects); | |
2669 | ||
2670 | ret += !list_empty(&l3->slabs_full) || | |
2671 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2672 | } |
2673 | return (ret ? 1 : 0); | |
2674 | } | |
2675 | ||
1da177e4 LT |
2676 | /** |
2677 | * kmem_cache_shrink - Shrink a cache. | |
2678 | * @cachep: The cache to shrink. | |
2679 | * | |
2680 | * Releases as many slabs as possible for a cache. | |
2681 | * To help debugging, a zero exit status indicates all slabs were released. | |
2682 | */ | |
343e0d7a | 2683 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2684 | { |
8f5be20b | 2685 | int ret; |
40094fa6 | 2686 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2687 | |
95402b38 | 2688 | get_online_cpus(); |
8f5be20b RT |
2689 | mutex_lock(&cache_chain_mutex); |
2690 | ret = __cache_shrink(cachep); | |
2691 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2692 | put_online_cpus(); |
8f5be20b | 2693 | return ret; |
1da177e4 LT |
2694 | } |
2695 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2696 | ||
2697 | /** | |
2698 | * kmem_cache_destroy - delete a cache | |
2699 | * @cachep: the cache to destroy | |
2700 | * | |
72fd4a35 | 2701 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2702 | * |
2703 | * It is expected this function will be called by a module when it is | |
2704 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2705 | * cache being allocated each time a module is loaded and unloaded, if the | |
2706 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2707 | * | |
2708 | * The cache must be empty before calling this function. | |
2709 | * | |
25985edc | 2710 | * The caller must guarantee that no one will allocate memory from the cache |
1da177e4 LT |
2711 | * during the kmem_cache_destroy(). |
2712 | */ | |
133d205a | 2713 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2714 | { |
40094fa6 | 2715 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2716 | |
1da177e4 | 2717 | /* Find the cache in the chain of caches. */ |
95402b38 | 2718 | get_online_cpus(); |
fc0abb14 | 2719 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2720 | /* |
2721 | * the chain is never empty, cache_cache is never destroyed | |
2722 | */ | |
3b0efdfa | 2723 | list_del(&cachep->list); |
1da177e4 LT |
2724 | if (__cache_shrink(cachep)) { |
2725 | slab_error(cachep, "Can't free all objects"); | |
3b0efdfa | 2726 | list_add(&cachep->list, &cache_chain); |
fc0abb14 | 2727 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2728 | put_online_cpus(); |
133d205a | 2729 | return; |
1da177e4 LT |
2730 | } |
2731 | ||
2732 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
7ed9f7e5 | 2733 | rcu_barrier(); |
1da177e4 | 2734 | |
117f6eb1 | 2735 | __kmem_cache_destroy(cachep); |
8f5be20b | 2736 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2737 | put_online_cpus(); |
1da177e4 LT |
2738 | } |
2739 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2740 | ||
e5ac9c5a RT |
2741 | /* |
2742 | * Get the memory for a slab management obj. | |
2743 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2744 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2745 | * come from the same cache which is getting created because, | |
2746 | * when we are searching for an appropriate cache for these | |
2747 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2748 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2749 | * kmem_find_general_cachep till the initialization is complete. | |
2750 | * Hence we cannot have slabp_cache same as the original cache. | |
2751 | */ | |
343e0d7a | 2752 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2753 | int colour_off, gfp_t local_flags, |
2754 | int nodeid) | |
1da177e4 LT |
2755 | { |
2756 | struct slab *slabp; | |
b28a02de | 2757 | |
1da177e4 LT |
2758 | if (OFF_SLAB(cachep)) { |
2759 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2760 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2761 | local_flags, nodeid); |
d5cff635 CM |
2762 | /* |
2763 | * If the first object in the slab is leaked (it's allocated | |
2764 | * but no one has a reference to it), we want to make sure | |
2765 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2766 | * to the object. Otherwise we will not report the leak. | |
2767 | */ | |
c017b4be CM |
2768 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2769 | local_flags); | |
1da177e4 LT |
2770 | if (!slabp) |
2771 | return NULL; | |
2772 | } else { | |
b28a02de | 2773 | slabp = objp + colour_off; |
1da177e4 LT |
2774 | colour_off += cachep->slab_size; |
2775 | } | |
2776 | slabp->inuse = 0; | |
2777 | slabp->colouroff = colour_off; | |
b28a02de | 2778 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2779 | slabp->nodeid = nodeid; |
e51bfd0a | 2780 | slabp->free = 0; |
1da177e4 LT |
2781 | return slabp; |
2782 | } | |
2783 | ||
2784 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2785 | { | |
b28a02de | 2786 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2787 | } |
2788 | ||
343e0d7a | 2789 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2790 | struct slab *slabp) |
1da177e4 LT |
2791 | { |
2792 | int i; | |
2793 | ||
2794 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2795 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2796 | #if DEBUG |
2797 | /* need to poison the objs? */ | |
2798 | if (cachep->flags & SLAB_POISON) | |
2799 | poison_obj(cachep, objp, POISON_FREE); | |
2800 | if (cachep->flags & SLAB_STORE_USER) | |
2801 | *dbg_userword(cachep, objp) = NULL; | |
2802 | ||
2803 | if (cachep->flags & SLAB_RED_ZONE) { | |
2804 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2805 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2806 | } | |
2807 | /* | |
a737b3e2 AM |
2808 | * Constructors are not allowed to allocate memory from the same |
2809 | * cache which they are a constructor for. Otherwise, deadlock. | |
2810 | * They must also be threaded. | |
1da177e4 LT |
2811 | */ |
2812 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2813 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2814 | |
2815 | if (cachep->flags & SLAB_RED_ZONE) { | |
2816 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2817 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2818 | " end of an object"); |
1da177e4 LT |
2819 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2820 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2821 | " start of an object"); |
1da177e4 | 2822 | } |
3b0efdfa | 2823 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2824 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2825 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2826 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2827 | #else |
2828 | if (cachep->ctor) | |
51cc5068 | 2829 | cachep->ctor(objp); |
1da177e4 | 2830 | #endif |
b28a02de | 2831 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2832 | } |
b28a02de | 2833 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2834 | } |
2835 | ||
343e0d7a | 2836 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2837 | { |
4b51d669 CL |
2838 | if (CONFIG_ZONE_DMA_FLAG) { |
2839 | if (flags & GFP_DMA) | |
2840 | BUG_ON(!(cachep->gfpflags & GFP_DMA)); | |
2841 | else | |
2842 | BUG_ON(cachep->gfpflags & GFP_DMA); | |
2843 | } | |
1da177e4 LT |
2844 | } |
2845 | ||
a737b3e2 AM |
2846 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2847 | int nodeid) | |
78d382d7 | 2848 | { |
8fea4e96 | 2849 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2850 | kmem_bufctl_t next; |
2851 | ||
2852 | slabp->inuse++; | |
2853 | next = slab_bufctl(slabp)[slabp->free]; | |
2854 | #if DEBUG | |
2855 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2856 | WARN_ON(slabp->nodeid != nodeid); | |
2857 | #endif | |
2858 | slabp->free = next; | |
2859 | ||
2860 | return objp; | |
2861 | } | |
2862 | ||
a737b3e2 AM |
2863 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2864 | void *objp, int nodeid) | |
78d382d7 | 2865 | { |
8fea4e96 | 2866 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2867 | |
2868 | #if DEBUG | |
2869 | /* Verify that the slab belongs to the intended node */ | |
2870 | WARN_ON(slabp->nodeid != nodeid); | |
2871 | ||
871751e2 | 2872 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2873 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2874 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2875 | BUG(); |
2876 | } | |
2877 | #endif | |
2878 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2879 | slabp->free = objnr; | |
2880 | slabp->inuse--; | |
2881 | } | |
2882 | ||
4776874f PE |
2883 | /* |
2884 | * Map pages beginning at addr to the given cache and slab. This is required | |
2885 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2886 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2887 | */ |
2888 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2889 | void *addr) | |
1da177e4 | 2890 | { |
4776874f | 2891 | int nr_pages; |
1da177e4 LT |
2892 | struct page *page; |
2893 | ||
4776874f | 2894 | page = virt_to_page(addr); |
84097518 | 2895 | |
4776874f | 2896 | nr_pages = 1; |
84097518 | 2897 | if (likely(!PageCompound(page))) |
4776874f PE |
2898 | nr_pages <<= cache->gfporder; |
2899 | ||
1da177e4 | 2900 | do { |
35026088 CL |
2901 | page->slab_cache = cache; |
2902 | page->slab_page = slab; | |
1da177e4 | 2903 | page++; |
4776874f | 2904 | } while (--nr_pages); |
1da177e4 LT |
2905 | } |
2906 | ||
2907 | /* | |
2908 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2909 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2910 | */ | |
3c517a61 CL |
2911 | static int cache_grow(struct kmem_cache *cachep, |
2912 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2913 | { |
b28a02de | 2914 | struct slab *slabp; |
b28a02de PE |
2915 | size_t offset; |
2916 | gfp_t local_flags; | |
e498be7d | 2917 | struct kmem_list3 *l3; |
1da177e4 | 2918 | |
a737b3e2 AM |
2919 | /* |
2920 | * Be lazy and only check for valid flags here, keeping it out of the | |
2921 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2922 | */ |
6cb06229 CL |
2923 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2924 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2925 | |
2e1217cf | 2926 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2927 | check_irq_off(); |
2e1217cf RT |
2928 | l3 = cachep->nodelists[nodeid]; |
2929 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2930 | |
2931 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2932 | offset = l3->colour_next; |
2933 | l3->colour_next++; | |
2934 | if (l3->colour_next >= cachep->colour) | |
2935 | l3->colour_next = 0; | |
2936 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2937 | |
2e1217cf | 2938 | offset *= cachep->colour_off; |
1da177e4 LT |
2939 | |
2940 | if (local_flags & __GFP_WAIT) | |
2941 | local_irq_enable(); | |
2942 | ||
2943 | /* | |
2944 | * The test for missing atomic flag is performed here, rather than | |
2945 | * the more obvious place, simply to reduce the critical path length | |
2946 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2947 | * will eventually be caught here (where it matters). | |
2948 | */ | |
2949 | kmem_flagcheck(cachep, flags); | |
2950 | ||
a737b3e2 AM |
2951 | /* |
2952 | * Get mem for the objs. Attempt to allocate a physical page from | |
2953 | * 'nodeid'. | |
e498be7d | 2954 | */ |
3c517a61 | 2955 | if (!objp) |
b8c1c5da | 2956 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2957 | if (!objp) |
1da177e4 LT |
2958 | goto failed; |
2959 | ||
2960 | /* Get slab management. */ | |
3c517a61 | 2961 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2962 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2963 | if (!slabp) |
1da177e4 LT |
2964 | goto opps1; |
2965 | ||
4776874f | 2966 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2967 | |
a35afb83 | 2968 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2969 | |
2970 | if (local_flags & __GFP_WAIT) | |
2971 | local_irq_disable(); | |
2972 | check_irq_off(); | |
e498be7d | 2973 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2974 | |
2975 | /* Make slab active. */ | |
e498be7d | 2976 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2977 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2978 | l3->free_objects += cachep->num; |
2979 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2980 | return 1; |
a737b3e2 | 2981 | opps1: |
1da177e4 | 2982 | kmem_freepages(cachep, objp); |
a737b3e2 | 2983 | failed: |
1da177e4 LT |
2984 | if (local_flags & __GFP_WAIT) |
2985 | local_irq_disable(); | |
2986 | return 0; | |
2987 | } | |
2988 | ||
2989 | #if DEBUG | |
2990 | ||
2991 | /* | |
2992 | * Perform extra freeing checks: | |
2993 | * - detect bad pointers. | |
2994 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2995 | */ |
2996 | static void kfree_debugcheck(const void *objp) | |
2997 | { | |
1da177e4 LT |
2998 | if (!virt_addr_valid(objp)) { |
2999 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
3000 | (unsigned long)objp); |
3001 | BUG(); | |
1da177e4 | 3002 | } |
1da177e4 LT |
3003 | } |
3004 | ||
58ce1fd5 PE |
3005 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
3006 | { | |
b46b8f19 | 3007 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
3008 | |
3009 | redzone1 = *dbg_redzone1(cache, obj); | |
3010 | redzone2 = *dbg_redzone2(cache, obj); | |
3011 | ||
3012 | /* | |
3013 | * Redzone is ok. | |
3014 | */ | |
3015 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3016 | return; | |
3017 | ||
3018 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3019 | slab_error(cache, "double free detected"); | |
3020 | else | |
3021 | slab_error(cache, "memory outside object was overwritten"); | |
3022 | ||
b46b8f19 | 3023 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
3024 | obj, redzone1, redzone2); |
3025 | } | |
3026 | ||
343e0d7a | 3027 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 3028 | void *caller) |
1da177e4 LT |
3029 | { |
3030 | struct page *page; | |
3031 | unsigned int objnr; | |
3032 | struct slab *slabp; | |
3033 | ||
80cbd911 MW |
3034 | BUG_ON(virt_to_cache(objp) != cachep); |
3035 | ||
3dafccf2 | 3036 | objp -= obj_offset(cachep); |
1da177e4 | 3037 | kfree_debugcheck(objp); |
b49af68f | 3038 | page = virt_to_head_page(objp); |
1da177e4 | 3039 | |
35026088 | 3040 | slabp = page->slab_page; |
1da177e4 LT |
3041 | |
3042 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3043 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3044 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3045 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3046 | } | |
3047 | if (cachep->flags & SLAB_STORE_USER) | |
3048 | *dbg_userword(cachep, objp) = caller; | |
3049 | ||
8fea4e96 | 3050 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3051 | |
3052 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3053 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3054 | |
871751e2 AV |
3055 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3056 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3057 | #endif | |
1da177e4 LT |
3058 | if (cachep->flags & SLAB_POISON) { |
3059 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3060 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 3061 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 3062 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3063 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3064 | } else { |
3065 | poison_obj(cachep, objp, POISON_FREE); | |
3066 | } | |
3067 | #else | |
3068 | poison_obj(cachep, objp, POISON_FREE); | |
3069 | #endif | |
3070 | } | |
3071 | return objp; | |
3072 | } | |
3073 | ||
343e0d7a | 3074 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3075 | { |
3076 | kmem_bufctl_t i; | |
3077 | int entries = 0; | |
b28a02de | 3078 | |
1da177e4 LT |
3079 | /* Check slab's freelist to see if this obj is there. */ |
3080 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3081 | entries++; | |
3082 | if (entries > cachep->num || i >= cachep->num) | |
3083 | goto bad; | |
3084 | } | |
3085 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3086 | bad: |
3087 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3088 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3089 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3090 | print_tainted()); | |
fdde6abb SAS |
3091 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3092 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3093 | 1); | |
1da177e4 LT |
3094 | BUG(); |
3095 | } | |
3096 | } | |
3097 | #else | |
3098 | #define kfree_debugcheck(x) do { } while(0) | |
3099 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3100 | #define check_slabp(x,y) do { } while(0) | |
3101 | #endif | |
3102 | ||
343e0d7a | 3103 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
3104 | { |
3105 | int batchcount; | |
3106 | struct kmem_list3 *l3; | |
3107 | struct array_cache *ac; | |
1ca4cb24 PE |
3108 | int node; |
3109 | ||
6d2144d3 | 3110 | retry: |
1da177e4 | 3111 | check_irq_off(); |
7d6e6d09 | 3112 | node = numa_mem_id(); |
9a2dba4b | 3113 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3114 | batchcount = ac->batchcount; |
3115 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3116 | /* |
3117 | * If there was little recent activity on this cache, then | |
3118 | * perform only a partial refill. Otherwise we could generate | |
3119 | * refill bouncing. | |
1da177e4 LT |
3120 | */ |
3121 | batchcount = BATCHREFILL_LIMIT; | |
3122 | } | |
1ca4cb24 | 3123 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3124 | |
3125 | BUG_ON(ac->avail > 0 || !l3); | |
3126 | spin_lock(&l3->list_lock); | |
1da177e4 | 3127 | |
3ded175a | 3128 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3129 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3130 | l3->shared->touched = 1; | |
3ded175a | 3131 | goto alloc_done; |
44b57f1c | 3132 | } |
3ded175a | 3133 | |
1da177e4 LT |
3134 | while (batchcount > 0) { |
3135 | struct list_head *entry; | |
3136 | struct slab *slabp; | |
3137 | /* Get slab alloc is to come from. */ | |
3138 | entry = l3->slabs_partial.next; | |
3139 | if (entry == &l3->slabs_partial) { | |
3140 | l3->free_touched = 1; | |
3141 | entry = l3->slabs_free.next; | |
3142 | if (entry == &l3->slabs_free) | |
3143 | goto must_grow; | |
3144 | } | |
3145 | ||
3146 | slabp = list_entry(entry, struct slab, list); | |
3147 | check_slabp(cachep, slabp); | |
3148 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3149 | |
3150 | /* | |
3151 | * The slab was either on partial or free list so | |
3152 | * there must be at least one object available for | |
3153 | * allocation. | |
3154 | */ | |
249b9f33 | 3155 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3156 | |
1da177e4 | 3157 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3158 | STATS_INC_ALLOCED(cachep); |
3159 | STATS_INC_ACTIVE(cachep); | |
3160 | STATS_SET_HIGH(cachep); | |
3161 | ||
78d382d7 | 3162 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3163 | node); |
1da177e4 LT |
3164 | } |
3165 | check_slabp(cachep, slabp); | |
3166 | ||
3167 | /* move slabp to correct slabp list: */ | |
3168 | list_del(&slabp->list); | |
3169 | if (slabp->free == BUFCTL_END) | |
3170 | list_add(&slabp->list, &l3->slabs_full); | |
3171 | else | |
3172 | list_add(&slabp->list, &l3->slabs_partial); | |
3173 | } | |
3174 | ||
a737b3e2 | 3175 | must_grow: |
1da177e4 | 3176 | l3->free_objects -= ac->avail; |
a737b3e2 | 3177 | alloc_done: |
e498be7d | 3178 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3179 | |
3180 | if (unlikely(!ac->avail)) { | |
3181 | int x; | |
3c517a61 | 3182 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3183 | |
a737b3e2 | 3184 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3185 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3186 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3187 | return NULL; |
3188 | ||
a737b3e2 | 3189 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3190 | goto retry; |
3191 | } | |
3192 | ac->touched = 1; | |
e498be7d | 3193 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3194 | } |
3195 | ||
a737b3e2 AM |
3196 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3197 | gfp_t flags) | |
1da177e4 LT |
3198 | { |
3199 | might_sleep_if(flags & __GFP_WAIT); | |
3200 | #if DEBUG | |
3201 | kmem_flagcheck(cachep, flags); | |
3202 | #endif | |
3203 | } | |
3204 | ||
3205 | #if DEBUG | |
a737b3e2 AM |
3206 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3207 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3208 | { |
b28a02de | 3209 | if (!objp) |
1da177e4 | 3210 | return objp; |
b28a02de | 3211 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3212 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3213 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3214 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3215 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3216 | else |
3217 | check_poison_obj(cachep, objp); | |
3218 | #else | |
3219 | check_poison_obj(cachep, objp); | |
3220 | #endif | |
3221 | poison_obj(cachep, objp, POISON_INUSE); | |
3222 | } | |
3223 | if (cachep->flags & SLAB_STORE_USER) | |
3224 | *dbg_userword(cachep, objp) = caller; | |
3225 | ||
3226 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3227 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3228 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3229 | slab_error(cachep, "double free, or memory outside" | |
3230 | " object was overwritten"); | |
b28a02de | 3231 | printk(KERN_ERR |
b46b8f19 | 3232 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3233 | objp, *dbg_redzone1(cachep, objp), |
3234 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3235 | } |
3236 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3237 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3238 | } | |
871751e2 AV |
3239 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3240 | { | |
3241 | struct slab *slabp; | |
3242 | unsigned objnr; | |
3243 | ||
35026088 | 3244 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3245 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3246 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3247 | } | |
3248 | #endif | |
3dafccf2 | 3249 | objp += obj_offset(cachep); |
4f104934 | 3250 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3251 | cachep->ctor(objp); |
7ea466f2 TH |
3252 | if (ARCH_SLAB_MINALIGN && |
3253 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3254 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3255 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3256 | } |
1da177e4 LT |
3257 | return objp; |
3258 | } | |
3259 | #else | |
3260 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3261 | #endif | |
3262 | ||
773ff60e | 3263 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3264 | { |
3265 | if (cachep == &cache_cache) | |
773ff60e | 3266 | return false; |
8a8b6502 | 3267 | |
4c13dd3b | 3268 | return should_failslab(obj_size(cachep), flags, cachep->flags); |
8a8b6502 AM |
3269 | } |
3270 | ||
343e0d7a | 3271 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3272 | { |
b28a02de | 3273 | void *objp; |
1da177e4 LT |
3274 | struct array_cache *ac; |
3275 | ||
5c382300 | 3276 | check_irq_off(); |
8a8b6502 | 3277 | |
9a2dba4b | 3278 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3279 | if (likely(ac->avail)) { |
3280 | STATS_INC_ALLOCHIT(cachep); | |
3281 | ac->touched = 1; | |
e498be7d | 3282 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3283 | } else { |
3284 | STATS_INC_ALLOCMISS(cachep); | |
3285 | objp = cache_alloc_refill(cachep, flags); | |
ddbf2e83 O |
3286 | /* |
3287 | * the 'ac' may be updated by cache_alloc_refill(), | |
3288 | * and kmemleak_erase() requires its correct value. | |
3289 | */ | |
3290 | ac = cpu_cache_get(cachep); | |
1da177e4 | 3291 | } |
d5cff635 CM |
3292 | /* |
3293 | * To avoid a false negative, if an object that is in one of the | |
3294 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3295 | * treat the array pointers as a reference to the object. | |
3296 | */ | |
f3d8b53a O |
3297 | if (objp) |
3298 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3299 | return objp; |
3300 | } | |
3301 | ||
e498be7d | 3302 | #ifdef CONFIG_NUMA |
c61afb18 | 3303 | /* |
b2455396 | 3304 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3305 | * |
3306 | * If we are in_interrupt, then process context, including cpusets and | |
3307 | * mempolicy, may not apply and should not be used for allocation policy. | |
3308 | */ | |
3309 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3310 | { | |
3311 | int nid_alloc, nid_here; | |
3312 | ||
765c4507 | 3313 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3314 | return NULL; |
7d6e6d09 | 3315 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3316 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3317 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 PJ |
3318 | else if (current->mempolicy) |
3319 | nid_alloc = slab_node(current->mempolicy); | |
3320 | if (nid_alloc != nid_here) | |
8b98c169 | 3321 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3322 | return NULL; |
3323 | } | |
3324 | ||
765c4507 CL |
3325 | /* |
3326 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3327 | * certain node and fall back is permitted. First we scan all the |
3328 | * available nodelists for available objects. If that fails then we | |
3329 | * perform an allocation without specifying a node. This allows the page | |
3330 | * allocator to do its reclaim / fallback magic. We then insert the | |
3331 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3332 | */ |
8c8cc2c1 | 3333 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3334 | { |
8c8cc2c1 PE |
3335 | struct zonelist *zonelist; |
3336 | gfp_t local_flags; | |
dd1a239f | 3337 | struct zoneref *z; |
54a6eb5c MG |
3338 | struct zone *zone; |
3339 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3340 | void *obj = NULL; |
3c517a61 | 3341 | int nid; |
cc9a6c87 | 3342 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3343 | |
3344 | if (flags & __GFP_THISNODE) | |
3345 | return NULL; | |
3346 | ||
6cb06229 | 3347 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3348 | |
cc9a6c87 MG |
3349 | retry_cpuset: |
3350 | cpuset_mems_cookie = get_mems_allowed(); | |
3351 | zonelist = node_zonelist(slab_node(current->mempolicy), flags); | |
3352 | ||
3c517a61 CL |
3353 | retry: |
3354 | /* | |
3355 | * Look through allowed nodes for objects available | |
3356 | * from existing per node queues. | |
3357 | */ | |
54a6eb5c MG |
3358 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3359 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3360 | |
54a6eb5c | 3361 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3362 | cache->nodelists[nid] && |
481c5346 | 3363 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3364 | obj = ____cache_alloc_node(cache, |
3365 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3366 | if (obj) |
3367 | break; | |
3368 | } | |
3c517a61 CL |
3369 | } |
3370 | ||
cfce6604 | 3371 | if (!obj) { |
3c517a61 CL |
3372 | /* |
3373 | * This allocation will be performed within the constraints | |
3374 | * of the current cpuset / memory policy requirements. | |
3375 | * We may trigger various forms of reclaim on the allowed | |
3376 | * set and go into memory reserves if necessary. | |
3377 | */ | |
dd47ea75 CL |
3378 | if (local_flags & __GFP_WAIT) |
3379 | local_irq_enable(); | |
3380 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3381 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3382 | if (local_flags & __GFP_WAIT) |
3383 | local_irq_disable(); | |
3c517a61 CL |
3384 | if (obj) { |
3385 | /* | |
3386 | * Insert into the appropriate per node queues | |
3387 | */ | |
3388 | nid = page_to_nid(virt_to_page(obj)); | |
3389 | if (cache_grow(cache, flags, nid, obj)) { | |
3390 | obj = ____cache_alloc_node(cache, | |
3391 | flags | GFP_THISNODE, nid); | |
3392 | if (!obj) | |
3393 | /* | |
3394 | * Another processor may allocate the | |
3395 | * objects in the slab since we are | |
3396 | * not holding any locks. | |
3397 | */ | |
3398 | goto retry; | |
3399 | } else { | |
b6a60451 | 3400 | /* cache_grow already freed obj */ |
3c517a61 CL |
3401 | obj = NULL; |
3402 | } | |
3403 | } | |
aedb0eb1 | 3404 | } |
cc9a6c87 MG |
3405 | |
3406 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3407 | goto retry_cpuset; | |
765c4507 CL |
3408 | return obj; |
3409 | } | |
3410 | ||
e498be7d CL |
3411 | /* |
3412 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3413 | */ |
8b98c169 | 3414 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3415 | int nodeid) |
e498be7d CL |
3416 | { |
3417 | struct list_head *entry; | |
b28a02de PE |
3418 | struct slab *slabp; |
3419 | struct kmem_list3 *l3; | |
3420 | void *obj; | |
b28a02de PE |
3421 | int x; |
3422 | ||
3423 | l3 = cachep->nodelists[nodeid]; | |
3424 | BUG_ON(!l3); | |
3425 | ||
a737b3e2 | 3426 | retry: |
ca3b9b91 | 3427 | check_irq_off(); |
b28a02de PE |
3428 | spin_lock(&l3->list_lock); |
3429 | entry = l3->slabs_partial.next; | |
3430 | if (entry == &l3->slabs_partial) { | |
3431 | l3->free_touched = 1; | |
3432 | entry = l3->slabs_free.next; | |
3433 | if (entry == &l3->slabs_free) | |
3434 | goto must_grow; | |
3435 | } | |
3436 | ||
3437 | slabp = list_entry(entry, struct slab, list); | |
3438 | check_spinlock_acquired_node(cachep, nodeid); | |
3439 | check_slabp(cachep, slabp); | |
3440 | ||
3441 | STATS_INC_NODEALLOCS(cachep); | |
3442 | STATS_INC_ACTIVE(cachep); | |
3443 | STATS_SET_HIGH(cachep); | |
3444 | ||
3445 | BUG_ON(slabp->inuse == cachep->num); | |
3446 | ||
78d382d7 | 3447 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3448 | check_slabp(cachep, slabp); |
3449 | l3->free_objects--; | |
3450 | /* move slabp to correct slabp list: */ | |
3451 | list_del(&slabp->list); | |
3452 | ||
a737b3e2 | 3453 | if (slabp->free == BUFCTL_END) |
b28a02de | 3454 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3455 | else |
b28a02de | 3456 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3457 | |
b28a02de PE |
3458 | spin_unlock(&l3->list_lock); |
3459 | goto done; | |
e498be7d | 3460 | |
a737b3e2 | 3461 | must_grow: |
b28a02de | 3462 | spin_unlock(&l3->list_lock); |
3c517a61 | 3463 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3464 | if (x) |
3465 | goto retry; | |
1da177e4 | 3466 | |
8c8cc2c1 | 3467 | return fallback_alloc(cachep, flags); |
e498be7d | 3468 | |
a737b3e2 | 3469 | done: |
b28a02de | 3470 | return obj; |
e498be7d | 3471 | } |
8c8cc2c1 PE |
3472 | |
3473 | /** | |
3474 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3475 | * @cachep: The cache to allocate from. | |
3476 | * @flags: See kmalloc(). | |
3477 | * @nodeid: node number of the target node. | |
3478 | * @caller: return address of caller, used for debug information | |
3479 | * | |
3480 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3481 | * node, which can improve the performance for cpu bound structures. | |
3482 | * | |
3483 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3484 | */ | |
3485 | static __always_inline void * | |
3486 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3487 | void *caller) | |
3488 | { | |
3489 | unsigned long save_flags; | |
3490 | void *ptr; | |
7d6e6d09 | 3491 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3492 | |
dcce284a | 3493 | flags &= gfp_allowed_mask; |
7e85ee0c | 3494 | |
cf40bd16 NP |
3495 | lockdep_trace_alloc(flags); |
3496 | ||
773ff60e | 3497 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3498 | return NULL; |
3499 | ||
8c8cc2c1 PE |
3500 | cache_alloc_debugcheck_before(cachep, flags); |
3501 | local_irq_save(save_flags); | |
3502 | ||
eacbbae3 | 3503 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3504 | nodeid = slab_node; |
8c8cc2c1 PE |
3505 | |
3506 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3507 | /* Node not bootstrapped yet */ | |
3508 | ptr = fallback_alloc(cachep, flags); | |
3509 | goto out; | |
3510 | } | |
3511 | ||
7d6e6d09 | 3512 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3513 | /* |
3514 | * Use the locally cached objects if possible. | |
3515 | * However ____cache_alloc does not allow fallback | |
3516 | * to other nodes. It may fail while we still have | |
3517 | * objects on other nodes available. | |
3518 | */ | |
3519 | ptr = ____cache_alloc(cachep, flags); | |
3520 | if (ptr) | |
3521 | goto out; | |
3522 | } | |
3523 | /* ___cache_alloc_node can fall back to other nodes */ | |
3524 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3525 | out: | |
3526 | local_irq_restore(save_flags); | |
3527 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
d5cff635 CM |
3528 | kmemleak_alloc_recursive(ptr, obj_size(cachep), 1, cachep->flags, |
3529 | flags); | |
8c8cc2c1 | 3530 | |
c175eea4 PE |
3531 | if (likely(ptr)) |
3532 | kmemcheck_slab_alloc(cachep, flags, ptr, obj_size(cachep)); | |
3533 | ||
d07dbea4 CL |
3534 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
3535 | memset(ptr, 0, obj_size(cachep)); | |
3536 | ||
8c8cc2c1 PE |
3537 | return ptr; |
3538 | } | |
3539 | ||
3540 | static __always_inline void * | |
3541 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3542 | { | |
3543 | void *objp; | |
3544 | ||
3545 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3546 | objp = alternate_node_alloc(cache, flags); | |
3547 | if (objp) | |
3548 | goto out; | |
3549 | } | |
3550 | objp = ____cache_alloc(cache, flags); | |
3551 | ||
3552 | /* | |
3553 | * We may just have run out of memory on the local node. | |
3554 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3555 | */ | |
7d6e6d09 LS |
3556 | if (!objp) |
3557 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3558 | |
3559 | out: | |
3560 | return objp; | |
3561 | } | |
3562 | #else | |
3563 | ||
3564 | static __always_inline void * | |
3565 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3566 | { | |
3567 | return ____cache_alloc(cachep, flags); | |
3568 | } | |
3569 | ||
3570 | #endif /* CONFIG_NUMA */ | |
3571 | ||
3572 | static __always_inline void * | |
3573 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3574 | { | |
3575 | unsigned long save_flags; | |
3576 | void *objp; | |
3577 | ||
dcce284a | 3578 | flags &= gfp_allowed_mask; |
7e85ee0c | 3579 | |
cf40bd16 NP |
3580 | lockdep_trace_alloc(flags); |
3581 | ||
773ff60e | 3582 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3583 | return NULL; |
3584 | ||
8c8cc2c1 PE |
3585 | cache_alloc_debugcheck_before(cachep, flags); |
3586 | local_irq_save(save_flags); | |
3587 | objp = __do_cache_alloc(cachep, flags); | |
3588 | local_irq_restore(save_flags); | |
3589 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
d5cff635 CM |
3590 | kmemleak_alloc_recursive(objp, obj_size(cachep), 1, cachep->flags, |
3591 | flags); | |
8c8cc2c1 PE |
3592 | prefetchw(objp); |
3593 | ||
c175eea4 PE |
3594 | if (likely(objp)) |
3595 | kmemcheck_slab_alloc(cachep, flags, objp, obj_size(cachep)); | |
3596 | ||
d07dbea4 CL |
3597 | if (unlikely((flags & __GFP_ZERO) && objp)) |
3598 | memset(objp, 0, obj_size(cachep)); | |
3599 | ||
8c8cc2c1 PE |
3600 | return objp; |
3601 | } | |
e498be7d CL |
3602 | |
3603 | /* | |
3604 | * Caller needs to acquire correct kmem_list's list_lock | |
3605 | */ | |
343e0d7a | 3606 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3607 | int node) |
1da177e4 LT |
3608 | { |
3609 | int i; | |
e498be7d | 3610 | struct kmem_list3 *l3; |
1da177e4 LT |
3611 | |
3612 | for (i = 0; i < nr_objects; i++) { | |
3613 | void *objp = objpp[i]; | |
3614 | struct slab *slabp; | |
1da177e4 | 3615 | |
6ed5eb22 | 3616 | slabp = virt_to_slab(objp); |
ff69416e | 3617 | l3 = cachep->nodelists[node]; |
1da177e4 | 3618 | list_del(&slabp->list); |
ff69416e | 3619 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3620 | check_slabp(cachep, slabp); |
78d382d7 | 3621 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3622 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3623 | l3->free_objects++; |
1da177e4 LT |
3624 | check_slabp(cachep, slabp); |
3625 | ||
3626 | /* fixup slab chains */ | |
3627 | if (slabp->inuse == 0) { | |
e498be7d CL |
3628 | if (l3->free_objects > l3->free_limit) { |
3629 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3630 | /* No need to drop any previously held |
3631 | * lock here, even if we have a off-slab slab | |
3632 | * descriptor it is guaranteed to come from | |
3633 | * a different cache, refer to comments before | |
3634 | * alloc_slabmgmt. | |
3635 | */ | |
1da177e4 LT |
3636 | slab_destroy(cachep, slabp); |
3637 | } else { | |
e498be7d | 3638 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3639 | } |
3640 | } else { | |
3641 | /* Unconditionally move a slab to the end of the | |
3642 | * partial list on free - maximum time for the | |
3643 | * other objects to be freed, too. | |
3644 | */ | |
e498be7d | 3645 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3646 | } |
3647 | } | |
3648 | } | |
3649 | ||
343e0d7a | 3650 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3651 | { |
3652 | int batchcount; | |
e498be7d | 3653 | struct kmem_list3 *l3; |
7d6e6d09 | 3654 | int node = numa_mem_id(); |
1da177e4 LT |
3655 | |
3656 | batchcount = ac->batchcount; | |
3657 | #if DEBUG | |
3658 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3659 | #endif | |
3660 | check_irq_off(); | |
ff69416e | 3661 | l3 = cachep->nodelists[node]; |
873623df | 3662 | spin_lock(&l3->list_lock); |
e498be7d CL |
3663 | if (l3->shared) { |
3664 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3665 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3666 | if (max) { |
3667 | if (batchcount > max) | |
3668 | batchcount = max; | |
e498be7d | 3669 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3670 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3671 | shared_array->avail += batchcount; |
3672 | goto free_done; | |
3673 | } | |
3674 | } | |
3675 | ||
ff69416e | 3676 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3677 | free_done: |
1da177e4 LT |
3678 | #if STATS |
3679 | { | |
3680 | int i = 0; | |
3681 | struct list_head *p; | |
3682 | ||
e498be7d CL |
3683 | p = l3->slabs_free.next; |
3684 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3685 | struct slab *slabp; |
3686 | ||
3687 | slabp = list_entry(p, struct slab, list); | |
3688 | BUG_ON(slabp->inuse); | |
3689 | ||
3690 | i++; | |
3691 | p = p->next; | |
3692 | } | |
3693 | STATS_SET_FREEABLE(cachep, i); | |
3694 | } | |
3695 | #endif | |
e498be7d | 3696 | spin_unlock(&l3->list_lock); |
1da177e4 | 3697 | ac->avail -= batchcount; |
a737b3e2 | 3698 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3699 | } |
3700 | ||
3701 | /* | |
a737b3e2 AM |
3702 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3703 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3704 | */ |
a947eb95 SS |
3705 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3706 | void *caller) | |
1da177e4 | 3707 | { |
9a2dba4b | 3708 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3709 | |
3710 | check_irq_off(); | |
d5cff635 | 3711 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3712 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3713 | |
c175eea4 PE |
3714 | kmemcheck_slab_free(cachep, objp, obj_size(cachep)); |
3715 | ||
1807a1aa SS |
3716 | /* |
3717 | * Skip calling cache_free_alien() when the platform is not numa. | |
3718 | * This will avoid cache misses that happen while accessing slabp (which | |
3719 | * is per page memory reference) to get nodeid. Instead use a global | |
3720 | * variable to skip the call, which is mostly likely to be present in | |
3721 | * the cache. | |
3722 | */ | |
b6e68bc1 | 3723 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3724 | return; |
3725 | ||
1da177e4 LT |
3726 | if (likely(ac->avail < ac->limit)) { |
3727 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3728 | } else { |
3729 | STATS_INC_FREEMISS(cachep); | |
3730 | cache_flusharray(cachep, ac); | |
1da177e4 | 3731 | } |
42c8c99c ZJ |
3732 | |
3733 | ac->entry[ac->avail++] = objp; | |
1da177e4 LT |
3734 | } |
3735 | ||
3736 | /** | |
3737 | * kmem_cache_alloc - Allocate an object | |
3738 | * @cachep: The cache to allocate from. | |
3739 | * @flags: See kmalloc(). | |
3740 | * | |
3741 | * Allocate an object from this cache. The flags are only relevant | |
3742 | * if the cache has no available objects. | |
3743 | */ | |
343e0d7a | 3744 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3745 | { |
36555751 EGM |
3746 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3747 | ||
ca2b84cb | 3748 | trace_kmem_cache_alloc(_RET_IP_, ret, |
3b0efdfa | 3749 | obj_size(cachep), cachep->size, flags); |
36555751 EGM |
3750 | |
3751 | return ret; | |
1da177e4 LT |
3752 | } |
3753 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3754 | ||
0f24f128 | 3755 | #ifdef CONFIG_TRACING |
85beb586 SR |
3756 | void * |
3757 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3758 | { |
85beb586 SR |
3759 | void *ret; |
3760 | ||
3761 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3762 | ||
3763 | trace_kmalloc(_RET_IP_, ret, | |
3764 | size, slab_buffer_size(cachep), flags); | |
3765 | return ret; | |
36555751 | 3766 | } |
85beb586 | 3767 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3768 | #endif |
3769 | ||
1da177e4 | 3770 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3771 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3772 | { | |
36555751 EGM |
3773 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3774 | __builtin_return_address(0)); | |
3775 | ||
ca2b84cb | 3776 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
3b0efdfa | 3777 | obj_size(cachep), cachep->size, |
ca2b84cb | 3778 | flags, nodeid); |
36555751 EGM |
3779 | |
3780 | return ret; | |
8b98c169 | 3781 | } |
1da177e4 LT |
3782 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3783 | ||
0f24f128 | 3784 | #ifdef CONFIG_TRACING |
85beb586 SR |
3785 | void *kmem_cache_alloc_node_trace(size_t size, |
3786 | struct kmem_cache *cachep, | |
3787 | gfp_t flags, | |
3788 | int nodeid) | |
36555751 | 3789 | { |
85beb586 SR |
3790 | void *ret; |
3791 | ||
3792 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3793 | __builtin_return_address(0)); |
85beb586 SR |
3794 | trace_kmalloc_node(_RET_IP_, ret, |
3795 | size, slab_buffer_size(cachep), | |
3796 | flags, nodeid); | |
3797 | return ret; | |
36555751 | 3798 | } |
85beb586 | 3799 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3800 | #endif |
3801 | ||
8b98c169 CH |
3802 | static __always_inline void * |
3803 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3804 | { |
343e0d7a | 3805 | struct kmem_cache *cachep; |
97e2bde4 MS |
3806 | |
3807 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3808 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3809 | return cachep; | |
85beb586 | 3810 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3811 | } |
8b98c169 | 3812 | |
0bb38a5c | 3813 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3814 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3815 | { | |
3816 | return __do_kmalloc_node(size, flags, node, | |
3817 | __builtin_return_address(0)); | |
3818 | } | |
dbe5e69d | 3819 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3820 | |
3821 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3822 | int node, unsigned long caller) |
8b98c169 | 3823 | { |
ce71e27c | 3824 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3825 | } |
3826 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3827 | #else | |
3828 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3829 | { | |
3830 | return __do_kmalloc_node(size, flags, node, NULL); | |
3831 | } | |
3832 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3833 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3834 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3835 | |
3836 | /** | |
800590f5 | 3837 | * __do_kmalloc - allocate memory |
1da177e4 | 3838 | * @size: how many bytes of memory are required. |
800590f5 | 3839 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3840 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3841 | */ |
7fd6b141 PE |
3842 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3843 | void *caller) | |
1da177e4 | 3844 | { |
343e0d7a | 3845 | struct kmem_cache *cachep; |
36555751 | 3846 | void *ret; |
1da177e4 | 3847 | |
97e2bde4 MS |
3848 | /* If you want to save a few bytes .text space: replace |
3849 | * __ with kmem_. | |
3850 | * Then kmalloc uses the uninlined functions instead of the inline | |
3851 | * functions. | |
3852 | */ | |
3853 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3854 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3855 | return cachep; | |
36555751 EGM |
3856 | ret = __cache_alloc(cachep, flags, caller); |
3857 | ||
ca2b84cb | 3858 | trace_kmalloc((unsigned long) caller, ret, |
3b0efdfa | 3859 | size, cachep->size, flags); |
36555751 EGM |
3860 | |
3861 | return ret; | |
7fd6b141 PE |
3862 | } |
3863 | ||
7fd6b141 | 3864 | |
0bb38a5c | 3865 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3866 | void *__kmalloc(size_t size, gfp_t flags) |
3867 | { | |
871751e2 | 3868 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3869 | } |
3870 | EXPORT_SYMBOL(__kmalloc); | |
3871 | ||
ce71e27c | 3872 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3873 | { |
ce71e27c | 3874 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3875 | } |
3876 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3877 | |
3878 | #else | |
3879 | void *__kmalloc(size_t size, gfp_t flags) | |
3880 | { | |
3881 | return __do_kmalloc(size, flags, NULL); | |
3882 | } | |
3883 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3884 | #endif |
3885 | ||
1da177e4 LT |
3886 | /** |
3887 | * kmem_cache_free - Deallocate an object | |
3888 | * @cachep: The cache the allocation was from. | |
3889 | * @objp: The previously allocated object. | |
3890 | * | |
3891 | * Free an object which was previously allocated from this | |
3892 | * cache. | |
3893 | */ | |
343e0d7a | 3894 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3895 | { |
3896 | unsigned long flags; | |
3897 | ||
3898 | local_irq_save(flags); | |
898552c9 | 3899 | debug_check_no_locks_freed(objp, obj_size(cachep)); |
3ac7fe5a TG |
3900 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3901 | debug_check_no_obj_freed(objp, obj_size(cachep)); | |
a947eb95 | 3902 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3903 | local_irq_restore(flags); |
36555751 | 3904 | |
ca2b84cb | 3905 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3906 | } |
3907 | EXPORT_SYMBOL(kmem_cache_free); | |
3908 | ||
1da177e4 LT |
3909 | /** |
3910 | * kfree - free previously allocated memory | |
3911 | * @objp: pointer returned by kmalloc. | |
3912 | * | |
80e93eff PE |
3913 | * If @objp is NULL, no operation is performed. |
3914 | * | |
1da177e4 LT |
3915 | * Don't free memory not originally allocated by kmalloc() |
3916 | * or you will run into trouble. | |
3917 | */ | |
3918 | void kfree(const void *objp) | |
3919 | { | |
343e0d7a | 3920 | struct kmem_cache *c; |
1da177e4 LT |
3921 | unsigned long flags; |
3922 | ||
2121db74 PE |
3923 | trace_kfree(_RET_IP_, objp); |
3924 | ||
6cb8f913 | 3925 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3926 | return; |
3927 | local_irq_save(flags); | |
3928 | kfree_debugcheck(objp); | |
6ed5eb22 | 3929 | c = virt_to_cache(objp); |
f9b8404c | 3930 | debug_check_no_locks_freed(objp, obj_size(c)); |
3ac7fe5a | 3931 | debug_check_no_obj_freed(objp, obj_size(c)); |
a947eb95 | 3932 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
3933 | local_irq_restore(flags); |
3934 | } | |
3935 | EXPORT_SYMBOL(kfree); | |
3936 | ||
343e0d7a | 3937 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3938 | { |
3dafccf2 | 3939 | return obj_size(cachep); |
1da177e4 LT |
3940 | } |
3941 | EXPORT_SYMBOL(kmem_cache_size); | |
3942 | ||
e498be7d | 3943 | /* |
183ff22b | 3944 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3945 | */ |
83b519e8 | 3946 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3947 | { |
3948 | int node; | |
3949 | struct kmem_list3 *l3; | |
cafeb02e | 3950 | struct array_cache *new_shared; |
3395ee05 | 3951 | struct array_cache **new_alien = NULL; |
e498be7d | 3952 | |
9c09a95c | 3953 | for_each_online_node(node) { |
cafeb02e | 3954 | |
3395ee05 | 3955 | if (use_alien_caches) { |
83b519e8 | 3956 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3957 | if (!new_alien) |
3958 | goto fail; | |
3959 | } | |
cafeb02e | 3960 | |
63109846 ED |
3961 | new_shared = NULL; |
3962 | if (cachep->shared) { | |
3963 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3964 | cachep->shared*cachep->batchcount, |
83b519e8 | 3965 | 0xbaadf00d, gfp); |
63109846 ED |
3966 | if (!new_shared) { |
3967 | free_alien_cache(new_alien); | |
3968 | goto fail; | |
3969 | } | |
0718dc2a | 3970 | } |
cafeb02e | 3971 | |
a737b3e2 AM |
3972 | l3 = cachep->nodelists[node]; |
3973 | if (l3) { | |
cafeb02e CL |
3974 | struct array_cache *shared = l3->shared; |
3975 | ||
e498be7d CL |
3976 | spin_lock_irq(&l3->list_lock); |
3977 | ||
cafeb02e | 3978 | if (shared) |
0718dc2a CL |
3979 | free_block(cachep, shared->entry, |
3980 | shared->avail, node); | |
e498be7d | 3981 | |
cafeb02e CL |
3982 | l3->shared = new_shared; |
3983 | if (!l3->alien) { | |
e498be7d CL |
3984 | l3->alien = new_alien; |
3985 | new_alien = NULL; | |
3986 | } | |
b28a02de | 3987 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3988 | cachep->batchcount + cachep->num; |
e498be7d | 3989 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3990 | kfree(shared); |
e498be7d CL |
3991 | free_alien_cache(new_alien); |
3992 | continue; | |
3993 | } | |
83b519e8 | 3994 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3995 | if (!l3) { |
3996 | free_alien_cache(new_alien); | |
3997 | kfree(new_shared); | |
e498be7d | 3998 | goto fail; |
0718dc2a | 3999 | } |
e498be7d CL |
4000 | |
4001 | kmem_list3_init(l3); | |
4002 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 4003 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 4004 | l3->shared = new_shared; |
e498be7d | 4005 | l3->alien = new_alien; |
b28a02de | 4006 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4007 | cachep->batchcount + cachep->num; |
e498be7d CL |
4008 | cachep->nodelists[node] = l3; |
4009 | } | |
cafeb02e | 4010 | return 0; |
0718dc2a | 4011 | |
a737b3e2 | 4012 | fail: |
3b0efdfa | 4013 | if (!cachep->list.next) { |
0718dc2a CL |
4014 | /* Cache is not active yet. Roll back what we did */ |
4015 | node--; | |
4016 | while (node >= 0) { | |
4017 | if (cachep->nodelists[node]) { | |
4018 | l3 = cachep->nodelists[node]; | |
4019 | ||
4020 | kfree(l3->shared); | |
4021 | free_alien_cache(l3->alien); | |
4022 | kfree(l3); | |
4023 | cachep->nodelists[node] = NULL; | |
4024 | } | |
4025 | node--; | |
4026 | } | |
4027 | } | |
cafeb02e | 4028 | return -ENOMEM; |
e498be7d CL |
4029 | } |
4030 | ||
1da177e4 | 4031 | struct ccupdate_struct { |
343e0d7a | 4032 | struct kmem_cache *cachep; |
acfe7d74 | 4033 | struct array_cache *new[0]; |
1da177e4 LT |
4034 | }; |
4035 | ||
4036 | static void do_ccupdate_local(void *info) | |
4037 | { | |
a737b3e2 | 4038 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4039 | struct array_cache *old; |
4040 | ||
4041 | check_irq_off(); | |
9a2dba4b | 4042 | old = cpu_cache_get(new->cachep); |
e498be7d | 4043 | |
1da177e4 LT |
4044 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4045 | new->new[smp_processor_id()] = old; | |
4046 | } | |
4047 | ||
b5d8ca7c | 4048 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 4049 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4050 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4051 | { |
d2e7b7d0 | 4052 | struct ccupdate_struct *new; |
2ed3a4ef | 4053 | int i; |
1da177e4 | 4054 | |
acfe7d74 ED |
4055 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4056 | gfp); | |
d2e7b7d0 SS |
4057 | if (!new) |
4058 | return -ENOMEM; | |
4059 | ||
e498be7d | 4060 | for_each_online_cpu(i) { |
7d6e6d09 | 4061 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4062 | batchcount, gfp); |
d2e7b7d0 | 4063 | if (!new->new[i]) { |
b28a02de | 4064 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4065 | kfree(new->new[i]); |
4066 | kfree(new); | |
e498be7d | 4067 | return -ENOMEM; |
1da177e4 LT |
4068 | } |
4069 | } | |
d2e7b7d0 | 4070 | new->cachep = cachep; |
1da177e4 | 4071 | |
15c8b6c1 | 4072 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4073 | |
1da177e4 | 4074 | check_irq_on(); |
1da177e4 LT |
4075 | cachep->batchcount = batchcount; |
4076 | cachep->limit = limit; | |
e498be7d | 4077 | cachep->shared = shared; |
1da177e4 | 4078 | |
e498be7d | 4079 | for_each_online_cpu(i) { |
d2e7b7d0 | 4080 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4081 | if (!ccold) |
4082 | continue; | |
7d6e6d09 LS |
4083 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4084 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4085 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4086 | kfree(ccold); |
4087 | } | |
d2e7b7d0 | 4088 | kfree(new); |
83b519e8 | 4089 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4090 | } |
4091 | ||
b5d8ca7c | 4092 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 4093 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4094 | { |
4095 | int err; | |
4096 | int limit, shared; | |
4097 | ||
a737b3e2 AM |
4098 | /* |
4099 | * The head array serves three purposes: | |
1da177e4 LT |
4100 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4101 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4102 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4103 | * bufctl chains: array operations are cheaper. |
4104 | * The numbers are guessed, we should auto-tune as described by | |
4105 | * Bonwick. | |
4106 | */ | |
3b0efdfa | 4107 | if (cachep->size > 131072) |
1da177e4 | 4108 | limit = 1; |
3b0efdfa | 4109 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4110 | limit = 8; |
3b0efdfa | 4111 | else if (cachep->size > 1024) |
1da177e4 | 4112 | limit = 24; |
3b0efdfa | 4113 | else if (cachep->size > 256) |
1da177e4 LT |
4114 | limit = 54; |
4115 | else | |
4116 | limit = 120; | |
4117 | ||
a737b3e2 AM |
4118 | /* |
4119 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4120 | * allocation behaviour: Most allocs on one cpu, most free operations |
4121 | * on another cpu. For these cases, an efficient object passing between | |
4122 | * cpus is necessary. This is provided by a shared array. The array | |
4123 | * replaces Bonwick's magazine layer. | |
4124 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4125 | * to a larger limit. Thus disabled by default. | |
4126 | */ | |
4127 | shared = 0; | |
3b0efdfa | 4128 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4129 | shared = 8; |
1da177e4 LT |
4130 | |
4131 | #if DEBUG | |
a737b3e2 AM |
4132 | /* |
4133 | * With debugging enabled, large batchcount lead to excessively long | |
4134 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4135 | */ |
4136 | if (limit > 32) | |
4137 | limit = 32; | |
4138 | #endif | |
83b519e8 | 4139 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4140 | if (err) |
4141 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4142 | cachep->name, -err); |
2ed3a4ef | 4143 | return err; |
1da177e4 LT |
4144 | } |
4145 | ||
1b55253a CL |
4146 | /* |
4147 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4148 | * necessary. Note that the l3 listlock also protects the array_cache |
4149 | * if drain_array() is used on the shared array. | |
1b55253a | 4150 | */ |
68a1b195 | 4151 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4152 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4153 | { |
4154 | int tofree; | |
4155 | ||
1b55253a CL |
4156 | if (!ac || !ac->avail) |
4157 | return; | |
1da177e4 LT |
4158 | if (ac->touched && !force) { |
4159 | ac->touched = 0; | |
b18e7e65 | 4160 | } else { |
1b55253a | 4161 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4162 | if (ac->avail) { |
4163 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4164 | if (tofree > ac->avail) | |
4165 | tofree = (ac->avail + 1) / 2; | |
4166 | free_block(cachep, ac->entry, tofree, node); | |
4167 | ac->avail -= tofree; | |
4168 | memmove(ac->entry, &(ac->entry[tofree]), | |
4169 | sizeof(void *) * ac->avail); | |
4170 | } | |
1b55253a | 4171 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4172 | } |
4173 | } | |
4174 | ||
4175 | /** | |
4176 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4177 | * @w: work descriptor |
1da177e4 LT |
4178 | * |
4179 | * Called from workqueue/eventd every few seconds. | |
4180 | * Purpose: | |
4181 | * - clear the per-cpu caches for this CPU. | |
4182 | * - return freeable pages to the main free memory pool. | |
4183 | * | |
a737b3e2 AM |
4184 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4185 | * again on the next iteration. | |
1da177e4 | 4186 | */ |
7c5cae36 | 4187 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4188 | { |
7a7c381d | 4189 | struct kmem_cache *searchp; |
e498be7d | 4190 | struct kmem_list3 *l3; |
7d6e6d09 | 4191 | int node = numa_mem_id(); |
bf6aede7 | 4192 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4193 | |
7c5cae36 | 4194 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4195 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4196 | goto out; |
1da177e4 | 4197 | |
3b0efdfa | 4198 | list_for_each_entry(searchp, &cache_chain, list) { |
1da177e4 LT |
4199 | check_irq_on(); |
4200 | ||
35386e3b CL |
4201 | /* |
4202 | * We only take the l3 lock if absolutely necessary and we | |
4203 | * have established with reasonable certainty that | |
4204 | * we can do some work if the lock was obtained. | |
4205 | */ | |
aab2207c | 4206 | l3 = searchp->nodelists[node]; |
35386e3b | 4207 | |
8fce4d8e | 4208 | reap_alien(searchp, l3); |
1da177e4 | 4209 | |
aab2207c | 4210 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4211 | |
35386e3b CL |
4212 | /* |
4213 | * These are racy checks but it does not matter | |
4214 | * if we skip one check or scan twice. | |
4215 | */ | |
e498be7d | 4216 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4217 | goto next; |
1da177e4 | 4218 | |
e498be7d | 4219 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4220 | |
aab2207c | 4221 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4222 | |
ed11d9eb | 4223 | if (l3->free_touched) |
e498be7d | 4224 | l3->free_touched = 0; |
ed11d9eb CL |
4225 | else { |
4226 | int freed; | |
1da177e4 | 4227 | |
ed11d9eb CL |
4228 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4229 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4230 | STATS_ADD_REAPED(searchp, freed); | |
4231 | } | |
35386e3b | 4232 | next: |
1da177e4 LT |
4233 | cond_resched(); |
4234 | } | |
4235 | check_irq_on(); | |
fc0abb14 | 4236 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4237 | next_reap_node(); |
7c5cae36 | 4238 | out: |
a737b3e2 | 4239 | /* Set up the next iteration */ |
7c5cae36 | 4240 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4241 | } |
4242 | ||
158a9624 | 4243 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4244 | |
85289f98 | 4245 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4246 | { |
85289f98 PE |
4247 | /* |
4248 | * Output format version, so at least we can change it | |
4249 | * without _too_ many complaints. | |
4250 | */ | |
1da177e4 | 4251 | #if STATS |
85289f98 | 4252 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4253 | #else |
85289f98 | 4254 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4255 | #endif |
85289f98 PE |
4256 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4257 | "<objperslab> <pagesperslab>"); | |
4258 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4259 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4260 | #if STATS |
85289f98 | 4261 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4262 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4263 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4264 | #endif |
85289f98 PE |
4265 | seq_putc(m, '\n'); |
4266 | } | |
4267 | ||
4268 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4269 | { | |
4270 | loff_t n = *pos; | |
85289f98 | 4271 | |
fc0abb14 | 4272 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4273 | if (!n) |
4274 | print_slabinfo_header(m); | |
b92151ba PE |
4275 | |
4276 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4277 | } |
4278 | ||
4279 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4280 | { | |
b92151ba | 4281 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4282 | } |
4283 | ||
4284 | static void s_stop(struct seq_file *m, void *p) | |
4285 | { | |
fc0abb14 | 4286 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4287 | } |
4288 | ||
4289 | static int s_show(struct seq_file *m, void *p) | |
4290 | { | |
3b0efdfa | 4291 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
b28a02de PE |
4292 | struct slab *slabp; |
4293 | unsigned long active_objs; | |
4294 | unsigned long num_objs; | |
4295 | unsigned long active_slabs = 0; | |
4296 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4297 | const char *name; |
1da177e4 | 4298 | char *error = NULL; |
e498be7d CL |
4299 | int node; |
4300 | struct kmem_list3 *l3; | |
1da177e4 | 4301 | |
1da177e4 LT |
4302 | active_objs = 0; |
4303 | num_slabs = 0; | |
e498be7d CL |
4304 | for_each_online_node(node) { |
4305 | l3 = cachep->nodelists[node]; | |
4306 | if (!l3) | |
4307 | continue; | |
4308 | ||
ca3b9b91 RT |
4309 | check_irq_on(); |
4310 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4311 | |
7a7c381d | 4312 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4313 | if (slabp->inuse != cachep->num && !error) |
4314 | error = "slabs_full accounting error"; | |
4315 | active_objs += cachep->num; | |
4316 | active_slabs++; | |
4317 | } | |
7a7c381d | 4318 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4319 | if (slabp->inuse == cachep->num && !error) |
4320 | error = "slabs_partial inuse accounting error"; | |
4321 | if (!slabp->inuse && !error) | |
4322 | error = "slabs_partial/inuse accounting error"; | |
4323 | active_objs += slabp->inuse; | |
4324 | active_slabs++; | |
4325 | } | |
7a7c381d | 4326 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4327 | if (slabp->inuse && !error) |
4328 | error = "slabs_free/inuse accounting error"; | |
4329 | num_slabs++; | |
4330 | } | |
4331 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4332 | if (l3->shared) |
4333 | shared_avail += l3->shared->avail; | |
e498be7d | 4334 | |
ca3b9b91 | 4335 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4336 | } |
b28a02de PE |
4337 | num_slabs += active_slabs; |
4338 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4339 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4340 | error = "free_objects accounting error"; |
4341 | ||
b28a02de | 4342 | name = cachep->name; |
1da177e4 LT |
4343 | if (error) |
4344 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4345 | ||
4346 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3b0efdfa | 4347 | name, active_objs, num_objs, cachep->size, |
b28a02de | 4348 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4349 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4350 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4351 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4352 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4353 | #if STATS |
b28a02de | 4354 | { /* list3 stats */ |
1da177e4 LT |
4355 | unsigned long high = cachep->high_mark; |
4356 | unsigned long allocs = cachep->num_allocations; | |
4357 | unsigned long grown = cachep->grown; | |
4358 | unsigned long reaped = cachep->reaped; | |
4359 | unsigned long errors = cachep->errors; | |
4360 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4361 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4362 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4363 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4364 | |
e92dd4fd JP |
4365 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4366 | "%4lu %4lu %4lu %4lu %4lu", | |
4367 | allocs, high, grown, | |
4368 | reaped, errors, max_freeable, node_allocs, | |
4369 | node_frees, overflows); | |
1da177e4 LT |
4370 | } |
4371 | /* cpu stats */ | |
4372 | { | |
4373 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4374 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4375 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4376 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4377 | ||
4378 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4379 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4380 | } |
4381 | #endif | |
4382 | seq_putc(m, '\n'); | |
1da177e4 LT |
4383 | return 0; |
4384 | } | |
4385 | ||
4386 | /* | |
4387 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4388 | * | |
4389 | * Output layout: | |
4390 | * cache-name | |
4391 | * num-active-objs | |
4392 | * total-objs | |
4393 | * object size | |
4394 | * num-active-slabs | |
4395 | * total-slabs | |
4396 | * num-pages-per-slab | |
4397 | * + further values on SMP and with statistics enabled | |
4398 | */ | |
4399 | ||
7b3c3a50 | 4400 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4401 | .start = s_start, |
4402 | .next = s_next, | |
4403 | .stop = s_stop, | |
4404 | .show = s_show, | |
1da177e4 LT |
4405 | }; |
4406 | ||
4407 | #define MAX_SLABINFO_WRITE 128 | |
4408 | /** | |
4409 | * slabinfo_write - Tuning for the slab allocator | |
4410 | * @file: unused | |
4411 | * @buffer: user buffer | |
4412 | * @count: data length | |
4413 | * @ppos: unused | |
4414 | */ | |
68a1b195 | 4415 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4416 | size_t count, loff_t *ppos) |
1da177e4 | 4417 | { |
b28a02de | 4418 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4419 | int limit, batchcount, shared, res; |
7a7c381d | 4420 | struct kmem_cache *cachep; |
b28a02de | 4421 | |
1da177e4 LT |
4422 | if (count > MAX_SLABINFO_WRITE) |
4423 | return -EINVAL; | |
4424 | if (copy_from_user(&kbuf, buffer, count)) | |
4425 | return -EFAULT; | |
b28a02de | 4426 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4427 | |
4428 | tmp = strchr(kbuf, ' '); | |
4429 | if (!tmp) | |
4430 | return -EINVAL; | |
4431 | *tmp = '\0'; | |
4432 | tmp++; | |
4433 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4434 | return -EINVAL; | |
4435 | ||
4436 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4437 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4438 | res = -EINVAL; |
3b0efdfa | 4439 | list_for_each_entry(cachep, &cache_chain, list) { |
1da177e4 | 4440 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4441 | if (limit < 1 || batchcount < 1 || |
4442 | batchcount > limit || shared < 0) { | |
e498be7d | 4443 | res = 0; |
1da177e4 | 4444 | } else { |
e498be7d | 4445 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4446 | batchcount, shared, |
4447 | GFP_KERNEL); | |
1da177e4 LT |
4448 | } |
4449 | break; | |
4450 | } | |
4451 | } | |
fc0abb14 | 4452 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4453 | if (res >= 0) |
4454 | res = count; | |
4455 | return res; | |
4456 | } | |
871751e2 | 4457 | |
7b3c3a50 AD |
4458 | static int slabinfo_open(struct inode *inode, struct file *file) |
4459 | { | |
4460 | return seq_open(file, &slabinfo_op); | |
4461 | } | |
4462 | ||
4463 | static const struct file_operations proc_slabinfo_operations = { | |
4464 | .open = slabinfo_open, | |
4465 | .read = seq_read, | |
4466 | .write = slabinfo_write, | |
4467 | .llseek = seq_lseek, | |
4468 | .release = seq_release, | |
4469 | }; | |
4470 | ||
871751e2 AV |
4471 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4472 | ||
4473 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4474 | { | |
871751e2 | 4475 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4476 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4477 | } |
4478 | ||
4479 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4480 | { | |
4481 | unsigned long *p; | |
4482 | int l; | |
4483 | if (!v) | |
4484 | return 1; | |
4485 | l = n[1]; | |
4486 | p = n + 2; | |
4487 | while (l) { | |
4488 | int i = l/2; | |
4489 | unsigned long *q = p + 2 * i; | |
4490 | if (*q == v) { | |
4491 | q[1]++; | |
4492 | return 1; | |
4493 | } | |
4494 | if (*q > v) { | |
4495 | l = i; | |
4496 | } else { | |
4497 | p = q + 2; | |
4498 | l -= i + 1; | |
4499 | } | |
4500 | } | |
4501 | if (++n[1] == n[0]) | |
4502 | return 0; | |
4503 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4504 | p[0] = v; | |
4505 | p[1] = 1; | |
4506 | return 1; | |
4507 | } | |
4508 | ||
4509 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4510 | { | |
4511 | void *p; | |
4512 | int i; | |
4513 | if (n[0] == n[1]) | |
4514 | return; | |
3b0efdfa | 4515 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4516 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4517 | continue; | |
4518 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4519 | return; | |
4520 | } | |
4521 | } | |
4522 | ||
4523 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4524 | { | |
4525 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4526 | unsigned long offset, size; |
9281acea | 4527 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4528 | |
a5c43dae | 4529 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4530 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4531 | if (modname[0]) |
871751e2 AV |
4532 | seq_printf(m, " [%s]", modname); |
4533 | return; | |
4534 | } | |
4535 | #endif | |
4536 | seq_printf(m, "%p", (void *)address); | |
4537 | } | |
4538 | ||
4539 | static int leaks_show(struct seq_file *m, void *p) | |
4540 | { | |
b92151ba | 4541 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, next); |
871751e2 AV |
4542 | struct slab *slabp; |
4543 | struct kmem_list3 *l3; | |
4544 | const char *name; | |
4545 | unsigned long *n = m->private; | |
4546 | int node; | |
4547 | int i; | |
4548 | ||
4549 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4550 | return 0; | |
4551 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4552 | return 0; | |
4553 | ||
4554 | /* OK, we can do it */ | |
4555 | ||
4556 | n[1] = 0; | |
4557 | ||
4558 | for_each_online_node(node) { | |
4559 | l3 = cachep->nodelists[node]; | |
4560 | if (!l3) | |
4561 | continue; | |
4562 | ||
4563 | check_irq_on(); | |
4564 | spin_lock_irq(&l3->list_lock); | |
4565 | ||
7a7c381d | 4566 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4567 | handle_slab(n, cachep, slabp); |
7a7c381d | 4568 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4569 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4570 | spin_unlock_irq(&l3->list_lock); |
4571 | } | |
4572 | name = cachep->name; | |
4573 | if (n[0] == n[1]) { | |
4574 | /* Increase the buffer size */ | |
4575 | mutex_unlock(&cache_chain_mutex); | |
4576 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4577 | if (!m->private) { | |
4578 | /* Too bad, we are really out */ | |
4579 | m->private = n; | |
4580 | mutex_lock(&cache_chain_mutex); | |
4581 | return -ENOMEM; | |
4582 | } | |
4583 | *(unsigned long *)m->private = n[0] * 2; | |
4584 | kfree(n); | |
4585 | mutex_lock(&cache_chain_mutex); | |
4586 | /* Now make sure this entry will be retried */ | |
4587 | m->count = m->size; | |
4588 | return 0; | |
4589 | } | |
4590 | for (i = 0; i < n[1]; i++) { | |
4591 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4592 | show_symbol(m, n[2*i+2]); | |
4593 | seq_putc(m, '\n'); | |
4594 | } | |
d2e7b7d0 | 4595 | |
871751e2 AV |
4596 | return 0; |
4597 | } | |
4598 | ||
a0ec95a8 | 4599 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4600 | .start = leaks_start, |
4601 | .next = s_next, | |
4602 | .stop = s_stop, | |
4603 | .show = leaks_show, | |
4604 | }; | |
a0ec95a8 AD |
4605 | |
4606 | static int slabstats_open(struct inode *inode, struct file *file) | |
4607 | { | |
4608 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4609 | int ret = -ENOMEM; | |
4610 | if (n) { | |
4611 | ret = seq_open(file, &slabstats_op); | |
4612 | if (!ret) { | |
4613 | struct seq_file *m = file->private_data; | |
4614 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4615 | m->private = n; | |
4616 | n = NULL; | |
4617 | } | |
4618 | kfree(n); | |
4619 | } | |
4620 | return ret; | |
4621 | } | |
4622 | ||
4623 | static const struct file_operations proc_slabstats_operations = { | |
4624 | .open = slabstats_open, | |
4625 | .read = seq_read, | |
4626 | .llseek = seq_lseek, | |
4627 | .release = seq_release_private, | |
4628 | }; | |
4629 | #endif | |
4630 | ||
4631 | static int __init slab_proc_init(void) | |
4632 | { | |
ab067e99 | 4633 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4634 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4635 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4636 | #endif |
a0ec95a8 AD |
4637 | return 0; |
4638 | } | |
4639 | module_init(slab_proc_init); | |
1da177e4 LT |
4640 | #endif |
4641 | ||
00e145b6 MS |
4642 | /** |
4643 | * ksize - get the actual amount of memory allocated for a given object | |
4644 | * @objp: Pointer to the object | |
4645 | * | |
4646 | * kmalloc may internally round up allocations and return more memory | |
4647 | * than requested. ksize() can be used to determine the actual amount of | |
4648 | * memory allocated. The caller may use this additional memory, even though | |
4649 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4650 | * The caller must guarantee that objp points to a valid object previously | |
4651 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4652 | * must not be freed during the duration of the call. | |
4653 | */ | |
fd76bab2 | 4654 | size_t ksize(const void *objp) |
1da177e4 | 4655 | { |
ef8b4520 CL |
4656 | BUG_ON(!objp); |
4657 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4658 | return 0; |
1da177e4 | 4659 | |
6ed5eb22 | 4660 | return obj_size(virt_to_cache(objp)); |
1da177e4 | 4661 | } |
b1aabecd | 4662 | EXPORT_SYMBOL(ksize); |