]>
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 | 89 | #include <linux/slab.h> |
97d06609 | 90 | #include "slab.h" |
1da177e4 | 91 | #include <linux/mm.h> |
c9cf5528 | 92 | #include <linux/poison.h> |
1da177e4 LT |
93 | #include <linux/swap.h> |
94 | #include <linux/cache.h> | |
95 | #include <linux/interrupt.h> | |
96 | #include <linux/init.h> | |
97 | #include <linux/compiler.h> | |
101a5001 | 98 | #include <linux/cpuset.h> |
a0ec95a8 | 99 | #include <linux/proc_fs.h> |
1da177e4 LT |
100 | #include <linux/seq_file.h> |
101 | #include <linux/notifier.h> | |
102 | #include <linux/kallsyms.h> | |
103 | #include <linux/cpu.h> | |
104 | #include <linux/sysctl.h> | |
105 | #include <linux/module.h> | |
106 | #include <linux/rcupdate.h> | |
543537bd | 107 | #include <linux/string.h> |
138ae663 | 108 | #include <linux/uaccess.h> |
e498be7d | 109 | #include <linux/nodemask.h> |
d5cff635 | 110 | #include <linux/kmemleak.h> |
dc85da15 | 111 | #include <linux/mempolicy.h> |
fc0abb14 | 112 | #include <linux/mutex.h> |
8a8b6502 | 113 | #include <linux/fault-inject.h> |
e7eebaf6 | 114 | #include <linux/rtmutex.h> |
6a2d7a95 | 115 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 116 | #include <linux/debugobjects.h> |
c175eea4 | 117 | #include <linux/kmemcheck.h> |
8f9f8d9e | 118 | #include <linux/memory.h> |
268bb0ce | 119 | #include <linux/prefetch.h> |
1da177e4 | 120 | |
1da177e4 LT |
121 | #include <asm/cacheflush.h> |
122 | #include <asm/tlbflush.h> | |
123 | #include <asm/page.h> | |
124 | ||
4dee6b64 SR |
125 | #include <trace/events/kmem.h> |
126 | ||
1da177e4 | 127 | /* |
50953fe9 | 128 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
129 | * 0 for faster, smaller code (especially in the critical paths). |
130 | * | |
131 | * STATS - 1 to collect stats for /proc/slabinfo. | |
132 | * 0 for faster, smaller code (especially in the critical paths). | |
133 | * | |
134 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
135 | */ | |
136 | ||
137 | #ifdef CONFIG_DEBUG_SLAB | |
138 | #define DEBUG 1 | |
139 | #define STATS 1 | |
140 | #define FORCED_DEBUG 1 | |
141 | #else | |
142 | #define DEBUG 0 | |
143 | #define STATS 0 | |
144 | #define FORCED_DEBUG 0 | |
145 | #endif | |
146 | ||
1da177e4 LT |
147 | /* Shouldn't this be in a header file somewhere? */ |
148 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 149 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 150 | |
1da177e4 LT |
151 | #ifndef ARCH_KMALLOC_FLAGS |
152 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
153 | #endif | |
154 | ||
155 | /* Legal flag mask for kmem_cache_create(). */ | |
156 | #if DEBUG | |
50953fe9 | 157 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 158 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 159 | SLAB_CACHE_DMA | \ |
5af60839 | 160 | SLAB_STORE_USER | \ |
1da177e4 | 161 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 162 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 163 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 | 164 | #else |
ac2b898c | 165 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 166 | SLAB_CACHE_DMA | \ |
1da177e4 | 167 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 168 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 169 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 LT |
170 | #endif |
171 | ||
172 | /* | |
173 | * kmem_bufctl_t: | |
174 | * | |
175 | * Bufctl's are used for linking objs within a slab | |
176 | * linked offsets. | |
177 | * | |
178 | * This implementation relies on "struct page" for locating the cache & | |
179 | * slab an object belongs to. | |
180 | * This allows the bufctl structure to be small (one int), but limits | |
181 | * the number of objects a slab (not a cache) can contain when off-slab | |
182 | * bufctls are used. The limit is the size of the largest general cache | |
183 | * that does not use off-slab slabs. | |
184 | * For 32bit archs with 4 kB pages, is this 56. | |
185 | * This is not serious, as it is only for large objects, when it is unwise | |
186 | * to have too many per slab. | |
187 | * Note: This limit can be raised by introducing a general cache whose size | |
188 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
189 | */ | |
190 | ||
fa5b08d5 | 191 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
192 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
193 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
194 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
195 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 196 | |
1da177e4 LT |
197 | /* |
198 | * struct slab_rcu | |
199 | * | |
200 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
201 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
202 | * we need to approach a kernel structure obliquely, from its address | |
203 | * obtained without the usual locking. We can lock the structure to | |
204 | * stabilize it and check it's still at the given address, only if we | |
205 | * can be sure that the memory has not been meanwhile reused for some | |
206 | * other kind of object (which our subsystem's lock might corrupt). | |
207 | * | |
208 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
209 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
210 | */ |
211 | struct slab_rcu { | |
b28a02de | 212 | struct rcu_head head; |
343e0d7a | 213 | struct kmem_cache *cachep; |
b28a02de | 214 | void *addr; |
1da177e4 LT |
215 | }; |
216 | ||
5bfe53a7 LJ |
217 | /* |
218 | * struct slab | |
219 | * | |
220 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
221 | * for a slab, or allocated from an general cache. | |
222 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
223 | */ | |
224 | struct slab { | |
225 | union { | |
226 | struct { | |
227 | struct list_head list; | |
228 | unsigned long colouroff; | |
229 | void *s_mem; /* including colour offset */ | |
230 | unsigned int inuse; /* num of objs active in slab */ | |
231 | kmem_bufctl_t free; | |
232 | unsigned short nodeid; | |
233 | }; | |
234 | struct slab_rcu __slab_cover_slab_rcu; | |
235 | }; | |
236 | }; | |
237 | ||
1da177e4 LT |
238 | /* |
239 | * struct array_cache | |
240 | * | |
1da177e4 LT |
241 | * Purpose: |
242 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
243 | * - reduce the number of linked list operations | |
244 | * - reduce spinlock operations | |
245 | * | |
246 | * The limit is stored in the per-cpu structure to reduce the data cache | |
247 | * footprint. | |
248 | * | |
249 | */ | |
250 | struct array_cache { | |
251 | unsigned int avail; | |
252 | unsigned int limit; | |
253 | unsigned int batchcount; | |
254 | unsigned int touched; | |
e498be7d | 255 | spinlock_t lock; |
bda5b655 | 256 | void *entry[]; /* |
a737b3e2 AM |
257 | * Must have this definition in here for the proper |
258 | * alignment of array_cache. Also simplifies accessing | |
259 | * the entries. | |
a737b3e2 | 260 | */ |
1da177e4 LT |
261 | }; |
262 | ||
a737b3e2 AM |
263 | /* |
264 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
265 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
266 | */ |
267 | #define BOOT_CPUCACHE_ENTRIES 1 | |
268 | struct arraycache_init { | |
269 | struct array_cache cache; | |
b28a02de | 270 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
271 | }; |
272 | ||
273 | /* | |
e498be7d | 274 | * The slab lists for all objects. |
1da177e4 LT |
275 | */ |
276 | struct kmem_list3 { | |
b28a02de PE |
277 | struct list_head slabs_partial; /* partial list first, better asm code */ |
278 | struct list_head slabs_full; | |
279 | struct list_head slabs_free; | |
280 | unsigned long free_objects; | |
b28a02de | 281 | unsigned int free_limit; |
2e1217cf | 282 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
283 | spinlock_t list_lock; |
284 | struct array_cache *shared; /* shared per node */ | |
285 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
286 | unsigned long next_reap; /* updated without locking */ |
287 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
288 | }; |
289 | ||
e498be7d CL |
290 | /* |
291 | * Need this for bootstrapping a per node allocator. | |
292 | */ | |
556a169d | 293 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
68a1b195 | 294 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 295 | #define CACHE_CACHE 0 |
556a169d PE |
296 | #define SIZE_AC MAX_NUMNODES |
297 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 298 | |
ed11d9eb CL |
299 | static int drain_freelist(struct kmem_cache *cache, |
300 | struct kmem_list3 *l3, int tofree); | |
301 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
302 | int node); | |
83b519e8 | 303 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 304 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 305 | |
e498be7d | 306 | /* |
a737b3e2 AM |
307 | * This function must be completely optimized away if a constant is passed to |
308 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 309 | */ |
7243cc05 | 310 | static __always_inline int index_of(const size_t size) |
e498be7d | 311 | { |
5ec8a847 SR |
312 | extern void __bad_size(void); |
313 | ||
e498be7d CL |
314 | if (__builtin_constant_p(size)) { |
315 | int i = 0; | |
316 | ||
317 | #define CACHE(x) \ | |
318 | if (size <=x) \ | |
319 | return i; \ | |
320 | else \ | |
321 | i++; | |
1c61fc40 | 322 | #include <linux/kmalloc_sizes.h> |
e498be7d | 323 | #undef CACHE |
5ec8a847 | 324 | __bad_size(); |
7243cc05 | 325 | } else |
5ec8a847 | 326 | __bad_size(); |
e498be7d CL |
327 | return 0; |
328 | } | |
329 | ||
e0a42726 IM |
330 | static int slab_early_init = 1; |
331 | ||
e498be7d CL |
332 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
333 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 334 | |
5295a74c | 335 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
336 | { |
337 | INIT_LIST_HEAD(&parent->slabs_full); | |
338 | INIT_LIST_HEAD(&parent->slabs_partial); | |
339 | INIT_LIST_HEAD(&parent->slabs_free); | |
340 | parent->shared = NULL; | |
341 | parent->alien = NULL; | |
2e1217cf | 342 | parent->colour_next = 0; |
e498be7d CL |
343 | spin_lock_init(&parent->list_lock); |
344 | parent->free_objects = 0; | |
345 | parent->free_touched = 0; | |
346 | } | |
347 | ||
a737b3e2 AM |
348 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
349 | do { \ | |
350 | INIT_LIST_HEAD(listp); \ | |
351 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
352 | } while (0) |
353 | ||
a737b3e2 AM |
354 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
355 | do { \ | |
e498be7d CL |
356 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
357 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
358 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
359 | } while (0) | |
1da177e4 | 360 | |
1da177e4 LT |
361 | #define CFLGS_OFF_SLAB (0x80000000UL) |
362 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
363 | ||
364 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
365 | /* |
366 | * Optimization question: fewer reaps means less probability for unnessary | |
367 | * cpucache drain/refill cycles. | |
1da177e4 | 368 | * |
dc6f3f27 | 369 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
370 | * which could lock up otherwise freeable slabs. |
371 | */ | |
372 | #define REAPTIMEOUT_CPUC (2*HZ) | |
373 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
374 | ||
375 | #if STATS | |
376 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
377 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
378 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
379 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 380 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
381 | #define STATS_SET_HIGH(x) \ |
382 | do { \ | |
383 | if ((x)->num_active > (x)->high_mark) \ | |
384 | (x)->high_mark = (x)->num_active; \ | |
385 | } while (0) | |
1da177e4 LT |
386 | #define STATS_INC_ERR(x) ((x)->errors++) |
387 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 388 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 389 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
390 | #define STATS_SET_FREEABLE(x, i) \ |
391 | do { \ | |
392 | if ((x)->max_freeable < i) \ | |
393 | (x)->max_freeable = i; \ | |
394 | } while (0) | |
1da177e4 LT |
395 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
396 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
397 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
398 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
399 | #else | |
400 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
401 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
402 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
403 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 404 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
405 | #define STATS_SET_HIGH(x) do { } while (0) |
406 | #define STATS_INC_ERR(x) do { } while (0) | |
407 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 408 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 409 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 410 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
411 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
412 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
413 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
414 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
415 | #endif | |
416 | ||
417 | #if DEBUG | |
1da177e4 | 418 | |
a737b3e2 AM |
419 | /* |
420 | * memory layout of objects: | |
1da177e4 | 421 | * 0 : objp |
3dafccf2 | 422 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
423 | * the end of an object is aligned with the end of the real |
424 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 425 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 426 | * redzone word. |
3dafccf2 | 427 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
428 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
429 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 430 | * [BYTES_PER_WORD long] |
1da177e4 | 431 | */ |
343e0d7a | 432 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 433 | { |
3dafccf2 | 434 | return cachep->obj_offset; |
1da177e4 LT |
435 | } |
436 | ||
b46b8f19 | 437 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
438 | { |
439 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
440 | return (unsigned long long*) (objp + obj_offset(cachep) - |
441 | sizeof(unsigned long long)); | |
1da177e4 LT |
442 | } |
443 | ||
b46b8f19 | 444 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
445 | { |
446 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
447 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 448 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 449 | sizeof(unsigned long long) - |
87a927c7 | 450 | REDZONE_ALIGN); |
3b0efdfa | 451 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 452 | sizeof(unsigned long long)); |
1da177e4 LT |
453 | } |
454 | ||
343e0d7a | 455 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
456 | { |
457 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 458 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
459 | } |
460 | ||
461 | #else | |
462 | ||
3dafccf2 | 463 | #define obj_offset(x) 0 |
b46b8f19 DW |
464 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
465 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
466 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
467 | ||
468 | #endif | |
469 | ||
0f24f128 | 470 | #ifdef CONFIG_TRACING |
36555751 EGM |
471 | size_t slab_buffer_size(struct kmem_cache *cachep) |
472 | { | |
3b0efdfa | 473 | return cachep->size; |
36555751 EGM |
474 | } |
475 | EXPORT_SYMBOL(slab_buffer_size); | |
476 | #endif | |
477 | ||
1da177e4 | 478 | /* |
3df1cccd DR |
479 | * Do not go above this order unless 0 objects fit into the slab or |
480 | * overridden on the command line. | |
1da177e4 | 481 | */ |
543585cc DR |
482 | #define SLAB_MAX_ORDER_HI 1 |
483 | #define SLAB_MAX_ORDER_LO 0 | |
484 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 485 | static bool slab_max_order_set __initdata; |
1da177e4 | 486 | |
065d41cb PE |
487 | static inline struct kmem_cache *page_get_cache(struct page *page) |
488 | { | |
d85f3385 | 489 | page = compound_head(page); |
ddc2e812 | 490 | BUG_ON(!PageSlab(page)); |
e571b0ad | 491 | return page->slab_cache; |
065d41cb PE |
492 | } |
493 | ||
6ed5eb22 PE |
494 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
495 | { | |
b49af68f | 496 | struct page *page = virt_to_head_page(obj); |
35026088 | 497 | return page->slab_cache; |
6ed5eb22 PE |
498 | } |
499 | ||
500 | static inline struct slab *virt_to_slab(const void *obj) | |
501 | { | |
b49af68f | 502 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
503 | |
504 | VM_BUG_ON(!PageSlab(page)); | |
505 | return page->slab_page; | |
6ed5eb22 PE |
506 | } |
507 | ||
8fea4e96 PE |
508 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
509 | unsigned int idx) | |
510 | { | |
3b0efdfa | 511 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
512 | } |
513 | ||
6a2d7a95 | 514 | /* |
3b0efdfa CL |
515 | * We want to avoid an expensive divide : (offset / cache->size) |
516 | * Using the fact that size is a constant for a particular cache, | |
517 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
518 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
519 | */ | |
520 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
521 | const struct slab *slab, void *obj) | |
8fea4e96 | 522 | { |
6a2d7a95 ED |
523 | u32 offset = (obj - slab->s_mem); |
524 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
525 | } |
526 | ||
a737b3e2 AM |
527 | /* |
528 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
529 | */ | |
1da177e4 LT |
530 | struct cache_sizes malloc_sizes[] = { |
531 | #define CACHE(x) { .cs_size = (x) }, | |
532 | #include <linux/kmalloc_sizes.h> | |
533 | CACHE(ULONG_MAX) | |
534 | #undef CACHE | |
535 | }; | |
536 | EXPORT_SYMBOL(malloc_sizes); | |
537 | ||
538 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
539 | struct cache_names { | |
540 | char *name; | |
541 | char *name_dma; | |
542 | }; | |
543 | ||
544 | static struct cache_names __initdata cache_names[] = { | |
545 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
546 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 547 | {NULL,} |
1da177e4 LT |
548 | #undef CACHE |
549 | }; | |
550 | ||
551 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 552 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 553 | static struct arraycache_init initarray_generic = |
b28a02de | 554 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
555 | |
556 | /* internal cache of cache description objs */ | |
b56efcf0 | 557 | static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES]; |
343e0d7a | 558 | static struct kmem_cache cache_cache = { |
b56efcf0 | 559 | .nodelists = cache_cache_nodelists, |
b28a02de PE |
560 | .batchcount = 1, |
561 | .limit = BOOT_CPUCACHE_ENTRIES, | |
562 | .shared = 1, | |
3b0efdfa | 563 | .size = sizeof(struct kmem_cache), |
b28a02de | 564 | .name = "kmem_cache", |
1da177e4 LT |
565 | }; |
566 | ||
056c6241 RT |
567 | #define BAD_ALIEN_MAGIC 0x01020304ul |
568 | ||
f1aaee53 AV |
569 | #ifdef CONFIG_LOCKDEP |
570 | ||
571 | /* | |
572 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
573 | * for other slabs "off slab". | |
574 | * The locking for this is tricky in that it nests within the locks | |
575 | * of all other slabs in a few places; to deal with this special | |
576 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
577 | * |
578 | * We set lock class for alien array caches which are up during init. | |
579 | * The lock annotation will be lost if all cpus of a node goes down and | |
580 | * then comes back up during hotplug | |
f1aaee53 | 581 | */ |
056c6241 RT |
582 | static struct lock_class_key on_slab_l3_key; |
583 | static struct lock_class_key on_slab_alc_key; | |
584 | ||
83835b3d PZ |
585 | static struct lock_class_key debugobj_l3_key; |
586 | static struct lock_class_key debugobj_alc_key; | |
587 | ||
588 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
589 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
590 | int q) | |
591 | { | |
592 | struct array_cache **alc; | |
593 | struct kmem_list3 *l3; | |
594 | int r; | |
595 | ||
596 | l3 = cachep->nodelists[q]; | |
597 | if (!l3) | |
598 | return; | |
599 | ||
600 | lockdep_set_class(&l3->list_lock, l3_key); | |
601 | alc = l3->alien; | |
602 | /* | |
603 | * FIXME: This check for BAD_ALIEN_MAGIC | |
604 | * should go away when common slab code is taught to | |
605 | * work even without alien caches. | |
606 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
607 | * for alloc_alien_cache, | |
608 | */ | |
609 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
610 | return; | |
611 | for_each_node(r) { | |
612 | if (alc[r]) | |
613 | lockdep_set_class(&alc[r]->lock, alc_key); | |
614 | } | |
615 | } | |
616 | ||
617 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
618 | { | |
619 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
620 | } | |
621 | ||
622 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
623 | { | |
624 | int node; | |
625 | ||
626 | for_each_online_node(node) | |
627 | slab_set_debugobj_lock_classes_node(cachep, node); | |
628 | } | |
629 | ||
ce79ddc8 | 630 | static void init_node_lock_keys(int q) |
f1aaee53 | 631 | { |
056c6241 RT |
632 | struct cache_sizes *s = malloc_sizes; |
633 | ||
97d06609 | 634 | if (slab_state < UP) |
ce79ddc8 PE |
635 | return; |
636 | ||
637 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
ce79ddc8 | 638 | struct kmem_list3 *l3; |
ce79ddc8 PE |
639 | |
640 | l3 = s->cs_cachep->nodelists[q]; | |
641 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 642 | continue; |
83835b3d PZ |
643 | |
644 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
645 | &on_slab_alc_key, q); | |
f1aaee53 AV |
646 | } |
647 | } | |
ce79ddc8 PE |
648 | |
649 | static inline void init_lock_keys(void) | |
650 | { | |
651 | int node; | |
652 | ||
653 | for_each_node(node) | |
654 | init_node_lock_keys(node); | |
655 | } | |
f1aaee53 | 656 | #else |
ce79ddc8 PE |
657 | static void init_node_lock_keys(int q) |
658 | { | |
659 | } | |
660 | ||
056c6241 | 661 | static inline void init_lock_keys(void) |
f1aaee53 AV |
662 | { |
663 | } | |
83835b3d PZ |
664 | |
665 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
666 | { | |
667 | } | |
668 | ||
669 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
670 | { | |
671 | } | |
f1aaee53 AV |
672 | #endif |
673 | ||
8f5be20b | 674 | /* |
95402b38 | 675 | * Guard access to the cache-chain. |
8f5be20b | 676 | */ |
fc0abb14 | 677 | static DEFINE_MUTEX(cache_chain_mutex); |
1da177e4 LT |
678 | static struct list_head cache_chain; |
679 | ||
1871e52c | 680 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 681 | |
343e0d7a | 682 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
683 | { |
684 | return cachep->array[smp_processor_id()]; | |
685 | } | |
686 | ||
a737b3e2 AM |
687 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
688 | gfp_t gfpflags) | |
1da177e4 LT |
689 | { |
690 | struct cache_sizes *csizep = malloc_sizes; | |
691 | ||
692 | #if DEBUG | |
693 | /* This happens if someone tries to call | |
b28a02de PE |
694 | * kmem_cache_create(), or __kmalloc(), before |
695 | * the generic caches are initialized. | |
696 | */ | |
c7e43c78 | 697 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 698 | #endif |
6cb8f913 CL |
699 | if (!size) |
700 | return ZERO_SIZE_PTR; | |
701 | ||
1da177e4 LT |
702 | while (size > csizep->cs_size) |
703 | csizep++; | |
704 | ||
705 | /* | |
0abf40c1 | 706 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
707 | * has cs_{dma,}cachep==NULL. Thus no special case |
708 | * for large kmalloc calls required. | |
709 | */ | |
4b51d669 | 710 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
711 | if (unlikely(gfpflags & GFP_DMA)) |
712 | return csizep->cs_dmacachep; | |
4b51d669 | 713 | #endif |
1da177e4 LT |
714 | return csizep->cs_cachep; |
715 | } | |
716 | ||
b221385b | 717 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
718 | { |
719 | return __find_general_cachep(size, gfpflags); | |
720 | } | |
97e2bde4 | 721 | |
fbaccacf | 722 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 723 | { |
fbaccacf SR |
724 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
725 | } | |
1da177e4 | 726 | |
a737b3e2 AM |
727 | /* |
728 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
729 | */ | |
fbaccacf SR |
730 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
731 | size_t align, int flags, size_t *left_over, | |
732 | unsigned int *num) | |
733 | { | |
734 | int nr_objs; | |
735 | size_t mgmt_size; | |
736 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 737 | |
fbaccacf SR |
738 | /* |
739 | * The slab management structure can be either off the slab or | |
740 | * on it. For the latter case, the memory allocated for a | |
741 | * slab is used for: | |
742 | * | |
743 | * - The struct slab | |
744 | * - One kmem_bufctl_t for each object | |
745 | * - Padding to respect alignment of @align | |
746 | * - @buffer_size bytes for each object | |
747 | * | |
748 | * If the slab management structure is off the slab, then the | |
749 | * alignment will already be calculated into the size. Because | |
750 | * the slabs are all pages aligned, the objects will be at the | |
751 | * correct alignment when allocated. | |
752 | */ | |
753 | if (flags & CFLGS_OFF_SLAB) { | |
754 | mgmt_size = 0; | |
755 | nr_objs = slab_size / buffer_size; | |
756 | ||
757 | if (nr_objs > SLAB_LIMIT) | |
758 | nr_objs = SLAB_LIMIT; | |
759 | } else { | |
760 | /* | |
761 | * Ignore padding for the initial guess. The padding | |
762 | * is at most @align-1 bytes, and @buffer_size is at | |
763 | * least @align. In the worst case, this result will | |
764 | * be one greater than the number of objects that fit | |
765 | * into the memory allocation when taking the padding | |
766 | * into account. | |
767 | */ | |
768 | nr_objs = (slab_size - sizeof(struct slab)) / | |
769 | (buffer_size + sizeof(kmem_bufctl_t)); | |
770 | ||
771 | /* | |
772 | * This calculated number will be either the right | |
773 | * amount, or one greater than what we want. | |
774 | */ | |
775 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
776 | > slab_size) | |
777 | nr_objs--; | |
778 | ||
779 | if (nr_objs > SLAB_LIMIT) | |
780 | nr_objs = SLAB_LIMIT; | |
781 | ||
782 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
783 | } | |
784 | *num = nr_objs; | |
785 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
786 | } |
787 | ||
d40cee24 | 788 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 789 | |
a737b3e2 AM |
790 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
791 | char *msg) | |
1da177e4 LT |
792 | { |
793 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 794 | function, cachep->name, msg); |
1da177e4 LT |
795 | dump_stack(); |
796 | } | |
797 | ||
3395ee05 PM |
798 | /* |
799 | * By default on NUMA we use alien caches to stage the freeing of | |
800 | * objects allocated from other nodes. This causes massive memory | |
801 | * inefficiencies when using fake NUMA setup to split memory into a | |
802 | * large number of small nodes, so it can be disabled on the command | |
803 | * line | |
804 | */ | |
805 | ||
806 | static int use_alien_caches __read_mostly = 1; | |
807 | static int __init noaliencache_setup(char *s) | |
808 | { | |
809 | use_alien_caches = 0; | |
810 | return 1; | |
811 | } | |
812 | __setup("noaliencache", noaliencache_setup); | |
813 | ||
3df1cccd DR |
814 | static int __init slab_max_order_setup(char *str) |
815 | { | |
816 | get_option(&str, &slab_max_order); | |
817 | slab_max_order = slab_max_order < 0 ? 0 : | |
818 | min(slab_max_order, MAX_ORDER - 1); | |
819 | slab_max_order_set = true; | |
820 | ||
821 | return 1; | |
822 | } | |
823 | __setup("slab_max_order=", slab_max_order_setup); | |
824 | ||
8fce4d8e CL |
825 | #ifdef CONFIG_NUMA |
826 | /* | |
827 | * Special reaping functions for NUMA systems called from cache_reap(). | |
828 | * These take care of doing round robin flushing of alien caches (containing | |
829 | * objects freed on different nodes from which they were allocated) and the | |
830 | * flushing of remote pcps by calling drain_node_pages. | |
831 | */ | |
1871e52c | 832 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
833 | |
834 | static void init_reap_node(int cpu) | |
835 | { | |
836 | int node; | |
837 | ||
7d6e6d09 | 838 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 839 | if (node == MAX_NUMNODES) |
442295c9 | 840 | node = first_node(node_online_map); |
8fce4d8e | 841 | |
1871e52c | 842 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
843 | } |
844 | ||
845 | static void next_reap_node(void) | |
846 | { | |
909ea964 | 847 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 848 | |
8fce4d8e CL |
849 | node = next_node(node, node_online_map); |
850 | if (unlikely(node >= MAX_NUMNODES)) | |
851 | node = first_node(node_online_map); | |
909ea964 | 852 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
853 | } |
854 | ||
855 | #else | |
856 | #define init_reap_node(cpu) do { } while (0) | |
857 | #define next_reap_node(void) do { } while (0) | |
858 | #endif | |
859 | ||
1da177e4 LT |
860 | /* |
861 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
862 | * via the workqueue/eventd. | |
863 | * Add the CPU number into the expiration time to minimize the possibility of | |
864 | * the CPUs getting into lockstep and contending for the global cache chain | |
865 | * lock. | |
866 | */ | |
897e679b | 867 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 868 | { |
1871e52c | 869 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
870 | |
871 | /* | |
872 | * When this gets called from do_initcalls via cpucache_init(), | |
873 | * init_workqueues() has already run, so keventd will be setup | |
874 | * at that time. | |
875 | */ | |
52bad64d | 876 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 877 | init_reap_node(cpu); |
78b43536 | 878 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); |
2b284214 AV |
879 | schedule_delayed_work_on(cpu, reap_work, |
880 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
881 | } |
882 | } | |
883 | ||
e498be7d | 884 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 885 | int batchcount, gfp_t gfp) |
1da177e4 | 886 | { |
b28a02de | 887 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
888 | struct array_cache *nc = NULL; |
889 | ||
83b519e8 | 890 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
891 | /* |
892 | * The array_cache structures contain pointers to free object. | |
25985edc | 893 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
894 | * cache the pointers are not cleared and they could be counted as |
895 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
896 | * not scan such objects. | |
897 | */ | |
898 | kmemleak_no_scan(nc); | |
1da177e4 LT |
899 | if (nc) { |
900 | nc->avail = 0; | |
901 | nc->limit = entries; | |
902 | nc->batchcount = batchcount; | |
903 | nc->touched = 0; | |
e498be7d | 904 | spin_lock_init(&nc->lock); |
1da177e4 LT |
905 | } |
906 | return nc; | |
907 | } | |
908 | ||
3ded175a CL |
909 | /* |
910 | * Transfer objects in one arraycache to another. | |
911 | * Locking must be handled by the caller. | |
912 | * | |
913 | * Return the number of entries transferred. | |
914 | */ | |
915 | static int transfer_objects(struct array_cache *to, | |
916 | struct array_cache *from, unsigned int max) | |
917 | { | |
918 | /* Figure out how many entries to transfer */ | |
732eacc0 | 919 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
920 | |
921 | if (!nr) | |
922 | return 0; | |
923 | ||
924 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
925 | sizeof(void *) *nr); | |
926 | ||
927 | from->avail -= nr; | |
928 | to->avail += nr; | |
3ded175a CL |
929 | return nr; |
930 | } | |
931 | ||
765c4507 CL |
932 | #ifndef CONFIG_NUMA |
933 | ||
934 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
935 | #define reap_alien(cachep, l3) do { } while (0) | |
936 | ||
83b519e8 | 937 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
938 | { |
939 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
940 | } | |
941 | ||
942 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
943 | { | |
944 | } | |
945 | ||
946 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
947 | { | |
948 | return 0; | |
949 | } | |
950 | ||
951 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
952 | gfp_t flags) | |
953 | { | |
954 | return NULL; | |
955 | } | |
956 | ||
8b98c169 | 957 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
958 | gfp_t flags, int nodeid) |
959 | { | |
960 | return NULL; | |
961 | } | |
962 | ||
963 | #else /* CONFIG_NUMA */ | |
964 | ||
8b98c169 | 965 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 966 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 967 | |
83b519e8 | 968 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
969 | { |
970 | struct array_cache **ac_ptr; | |
8ef82866 | 971 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
972 | int i; |
973 | ||
974 | if (limit > 1) | |
975 | limit = 12; | |
f3186a9c | 976 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
977 | if (ac_ptr) { |
978 | for_each_node(i) { | |
f3186a9c | 979 | if (i == node || !node_online(i)) |
e498be7d | 980 | continue; |
83b519e8 | 981 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 982 | if (!ac_ptr[i]) { |
cc550def | 983 | for (i--; i >= 0; i--) |
e498be7d CL |
984 | kfree(ac_ptr[i]); |
985 | kfree(ac_ptr); | |
986 | return NULL; | |
987 | } | |
988 | } | |
989 | } | |
990 | return ac_ptr; | |
991 | } | |
992 | ||
5295a74c | 993 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
994 | { |
995 | int i; | |
996 | ||
997 | if (!ac_ptr) | |
998 | return; | |
e498be7d | 999 | for_each_node(i) |
b28a02de | 1000 | kfree(ac_ptr[i]); |
e498be7d CL |
1001 | kfree(ac_ptr); |
1002 | } | |
1003 | ||
343e0d7a | 1004 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1005 | struct array_cache *ac, int node) |
e498be7d CL |
1006 | { |
1007 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1008 | ||
1009 | if (ac->avail) { | |
1010 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1011 | /* |
1012 | * Stuff objects into the remote nodes shared array first. | |
1013 | * That way we could avoid the overhead of putting the objects | |
1014 | * into the free lists and getting them back later. | |
1015 | */ | |
693f7d36 JS |
1016 | if (rl3->shared) |
1017 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1018 | |
ff69416e | 1019 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1020 | ac->avail = 0; |
1021 | spin_unlock(&rl3->list_lock); | |
1022 | } | |
1023 | } | |
1024 | ||
8fce4d8e CL |
1025 | /* |
1026 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1027 | */ | |
1028 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1029 | { | |
909ea964 | 1030 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1031 | |
1032 | if (l3->alien) { | |
1033 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1034 | |
1035 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1036 | __drain_alien_cache(cachep, ac, node); |
1037 | spin_unlock_irq(&ac->lock); | |
1038 | } | |
1039 | } | |
1040 | } | |
1041 | ||
a737b3e2 AM |
1042 | static void drain_alien_cache(struct kmem_cache *cachep, |
1043 | struct array_cache **alien) | |
e498be7d | 1044 | { |
b28a02de | 1045 | int i = 0; |
e498be7d CL |
1046 | struct array_cache *ac; |
1047 | unsigned long flags; | |
1048 | ||
1049 | for_each_online_node(i) { | |
4484ebf1 | 1050 | ac = alien[i]; |
e498be7d CL |
1051 | if (ac) { |
1052 | spin_lock_irqsave(&ac->lock, flags); | |
1053 | __drain_alien_cache(cachep, ac, i); | |
1054 | spin_unlock_irqrestore(&ac->lock, flags); | |
1055 | } | |
1056 | } | |
1057 | } | |
729bd0b7 | 1058 | |
873623df | 1059 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1060 | { |
1061 | struct slab *slabp = virt_to_slab(objp); | |
1062 | int nodeid = slabp->nodeid; | |
1063 | struct kmem_list3 *l3; | |
1064 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1065 | int node; |
1066 | ||
7d6e6d09 | 1067 | node = numa_mem_id(); |
729bd0b7 PE |
1068 | |
1069 | /* | |
1070 | * Make sure we are not freeing a object from another node to the array | |
1071 | * cache on this cpu. | |
1072 | */ | |
62918a03 | 1073 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1074 | return 0; |
1075 | ||
1ca4cb24 | 1076 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1077 | STATS_INC_NODEFREES(cachep); |
1078 | if (l3->alien && l3->alien[nodeid]) { | |
1079 | alien = l3->alien[nodeid]; | |
873623df | 1080 | spin_lock(&alien->lock); |
729bd0b7 PE |
1081 | if (unlikely(alien->avail == alien->limit)) { |
1082 | STATS_INC_ACOVERFLOW(cachep); | |
1083 | __drain_alien_cache(cachep, alien, nodeid); | |
1084 | } | |
1085 | alien->entry[alien->avail++] = objp; | |
1086 | spin_unlock(&alien->lock); | |
1087 | } else { | |
1088 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1089 | free_block(cachep, &objp, 1, nodeid); | |
1090 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1091 | } | |
1092 | return 1; | |
1093 | } | |
e498be7d CL |
1094 | #endif |
1095 | ||
8f9f8d9e DR |
1096 | /* |
1097 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1098 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1099 | * will be allocated off-node since memory is not yet online for the new node. | |
1100 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1101 | * already in use. | |
1102 | * | |
1103 | * Must hold cache_chain_mutex. | |
1104 | */ | |
1105 | static int init_cache_nodelists_node(int node) | |
1106 | { | |
1107 | struct kmem_cache *cachep; | |
1108 | struct kmem_list3 *l3; | |
1109 | const int memsize = sizeof(struct kmem_list3); | |
1110 | ||
3b0efdfa | 1111 | list_for_each_entry(cachep, &cache_chain, list) { |
8f9f8d9e DR |
1112 | /* |
1113 | * Set up the size64 kmemlist for cpu before we can | |
1114 | * begin anything. Make sure some other cpu on this | |
1115 | * node has not already allocated this | |
1116 | */ | |
1117 | if (!cachep->nodelists[node]) { | |
1118 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1119 | if (!l3) | |
1120 | return -ENOMEM; | |
1121 | kmem_list3_init(l3); | |
1122 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1123 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1124 | ||
1125 | /* | |
1126 | * The l3s don't come and go as CPUs come and | |
1127 | * go. cache_chain_mutex is sufficient | |
1128 | * protection here. | |
1129 | */ | |
1130 | cachep->nodelists[node] = l3; | |
1131 | } | |
1132 | ||
1133 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1134 | cachep->nodelists[node]->free_limit = | |
1135 | (1 + nr_cpus_node(node)) * | |
1136 | cachep->batchcount + cachep->num; | |
1137 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1138 | } | |
1139 | return 0; | |
1140 | } | |
1141 | ||
fbf1e473 AM |
1142 | static void __cpuinit cpuup_canceled(long cpu) |
1143 | { | |
1144 | struct kmem_cache *cachep; | |
1145 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1146 | int node = cpu_to_mem(cpu); |
a70f7302 | 1147 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1148 | |
3b0efdfa | 1149 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1150 | struct array_cache *nc; |
1151 | struct array_cache *shared; | |
1152 | struct array_cache **alien; | |
fbf1e473 | 1153 | |
fbf1e473 AM |
1154 | /* cpu is dead; no one can alloc from it. */ |
1155 | nc = cachep->array[cpu]; | |
1156 | cachep->array[cpu] = NULL; | |
1157 | l3 = cachep->nodelists[node]; | |
1158 | ||
1159 | if (!l3) | |
1160 | goto free_array_cache; | |
1161 | ||
1162 | spin_lock_irq(&l3->list_lock); | |
1163 | ||
1164 | /* Free limit for this kmem_list3 */ | |
1165 | l3->free_limit -= cachep->batchcount; | |
1166 | if (nc) | |
1167 | free_block(cachep, nc->entry, nc->avail, node); | |
1168 | ||
58463c1f | 1169 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1170 | spin_unlock_irq(&l3->list_lock); |
1171 | goto free_array_cache; | |
1172 | } | |
1173 | ||
1174 | shared = l3->shared; | |
1175 | if (shared) { | |
1176 | free_block(cachep, shared->entry, | |
1177 | shared->avail, node); | |
1178 | l3->shared = NULL; | |
1179 | } | |
1180 | ||
1181 | alien = l3->alien; | |
1182 | l3->alien = NULL; | |
1183 | ||
1184 | spin_unlock_irq(&l3->list_lock); | |
1185 | ||
1186 | kfree(shared); | |
1187 | if (alien) { | |
1188 | drain_alien_cache(cachep, alien); | |
1189 | free_alien_cache(alien); | |
1190 | } | |
1191 | free_array_cache: | |
1192 | kfree(nc); | |
1193 | } | |
1194 | /* | |
1195 | * In the previous loop, all the objects were freed to | |
1196 | * the respective cache's slabs, now we can go ahead and | |
1197 | * shrink each nodelist to its limit. | |
1198 | */ | |
3b0efdfa | 1199 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1200 | l3 = cachep->nodelists[node]; |
1201 | if (!l3) | |
1202 | continue; | |
1203 | drain_freelist(cachep, l3, l3->free_objects); | |
1204 | } | |
1205 | } | |
1206 | ||
1207 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1208 | { |
343e0d7a | 1209 | struct kmem_cache *cachep; |
e498be7d | 1210 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1211 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1212 | int err; |
1da177e4 | 1213 | |
fbf1e473 AM |
1214 | /* |
1215 | * We need to do this right in the beginning since | |
1216 | * alloc_arraycache's are going to use this list. | |
1217 | * kmalloc_node allows us to add the slab to the right | |
1218 | * kmem_list3 and not this cpu's kmem_list3 | |
1219 | */ | |
8f9f8d9e DR |
1220 | err = init_cache_nodelists_node(node); |
1221 | if (err < 0) | |
1222 | goto bad; | |
fbf1e473 AM |
1223 | |
1224 | /* | |
1225 | * Now we can go ahead with allocating the shared arrays and | |
1226 | * array caches | |
1227 | */ | |
3b0efdfa | 1228 | list_for_each_entry(cachep, &cache_chain, list) { |
fbf1e473 AM |
1229 | struct array_cache *nc; |
1230 | struct array_cache *shared = NULL; | |
1231 | struct array_cache **alien = NULL; | |
1232 | ||
1233 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1234 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1235 | if (!nc) |
1236 | goto bad; | |
1237 | if (cachep->shared) { | |
1238 | shared = alloc_arraycache(node, | |
1239 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1240 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1241 | if (!shared) { |
1242 | kfree(nc); | |
1da177e4 | 1243 | goto bad; |
12d00f6a | 1244 | } |
fbf1e473 AM |
1245 | } |
1246 | if (use_alien_caches) { | |
83b519e8 | 1247 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1248 | if (!alien) { |
1249 | kfree(shared); | |
1250 | kfree(nc); | |
fbf1e473 | 1251 | goto bad; |
12d00f6a | 1252 | } |
fbf1e473 AM |
1253 | } |
1254 | cachep->array[cpu] = nc; | |
1255 | l3 = cachep->nodelists[node]; | |
1256 | BUG_ON(!l3); | |
1257 | ||
1258 | spin_lock_irq(&l3->list_lock); | |
1259 | if (!l3->shared) { | |
1260 | /* | |
1261 | * We are serialised from CPU_DEAD or | |
1262 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1263 | */ | |
1264 | l3->shared = shared; | |
1265 | shared = NULL; | |
1266 | } | |
4484ebf1 | 1267 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1268 | if (!l3->alien) { |
1269 | l3->alien = alien; | |
1270 | alien = NULL; | |
1da177e4 | 1271 | } |
fbf1e473 AM |
1272 | #endif |
1273 | spin_unlock_irq(&l3->list_lock); | |
1274 | kfree(shared); | |
1275 | free_alien_cache(alien); | |
83835b3d PZ |
1276 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1277 | slab_set_debugobj_lock_classes_node(cachep, node); | |
fbf1e473 | 1278 | } |
ce79ddc8 PE |
1279 | init_node_lock_keys(node); |
1280 | ||
fbf1e473 AM |
1281 | return 0; |
1282 | bad: | |
12d00f6a | 1283 | cpuup_canceled(cpu); |
fbf1e473 AM |
1284 | return -ENOMEM; |
1285 | } | |
1286 | ||
1287 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1288 | unsigned long action, void *hcpu) | |
1289 | { | |
1290 | long cpu = (long)hcpu; | |
1291 | int err = 0; | |
1292 | ||
1293 | switch (action) { | |
fbf1e473 AM |
1294 | case CPU_UP_PREPARE: |
1295 | case CPU_UP_PREPARE_FROZEN: | |
95402b38 | 1296 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1297 | err = cpuup_prepare(cpu); |
95402b38 | 1298 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
1299 | break; |
1300 | case CPU_ONLINE: | |
8bb78442 | 1301 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1302 | start_cpu_timer(cpu); |
1303 | break; | |
1304 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1305 | case CPU_DOWN_PREPARE: |
8bb78442 | 1306 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 CL |
1307 | /* |
1308 | * Shutdown cache reaper. Note that the cache_chain_mutex is | |
1309 | * held so that if cache_reap() is invoked it cannot do | |
1310 | * anything expensive but will only modify reap_work | |
1311 | * and reschedule the timer. | |
1312 | */ | |
afe2c511 | 1313 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1314 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1315 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1316 | break; |
1317 | case CPU_DOWN_FAILED: | |
8bb78442 | 1318 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1319 | start_cpu_timer(cpu); |
1320 | break; | |
1da177e4 | 1321 | case CPU_DEAD: |
8bb78442 | 1322 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1323 | /* |
1324 | * Even if all the cpus of a node are down, we don't free the | |
1325 | * kmem_list3 of any cache. This to avoid a race between | |
1326 | * cpu_down, and a kmalloc allocation from another cpu for | |
1327 | * memory from the node of the cpu going down. The list3 | |
1328 | * structure is usually allocated from kmem_cache_create() and | |
1329 | * gets destroyed at kmem_cache_destroy(). | |
1330 | */ | |
183ff22b | 1331 | /* fall through */ |
8f5be20b | 1332 | #endif |
1da177e4 | 1333 | case CPU_UP_CANCELED: |
8bb78442 | 1334 | case CPU_UP_CANCELED_FROZEN: |
95402b38 | 1335 | mutex_lock(&cache_chain_mutex); |
fbf1e473 | 1336 | cpuup_canceled(cpu); |
fc0abb14 | 1337 | mutex_unlock(&cache_chain_mutex); |
1da177e4 | 1338 | break; |
1da177e4 | 1339 | } |
eac40680 | 1340 | return notifier_from_errno(err); |
1da177e4 LT |
1341 | } |
1342 | ||
74b85f37 CS |
1343 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1344 | &cpuup_callback, NULL, 0 | |
1345 | }; | |
1da177e4 | 1346 | |
8f9f8d9e DR |
1347 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1348 | /* | |
1349 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1350 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1351 | * removed. | |
1352 | * | |
1353 | * Must hold cache_chain_mutex. | |
1354 | */ | |
1355 | static int __meminit drain_cache_nodelists_node(int node) | |
1356 | { | |
1357 | struct kmem_cache *cachep; | |
1358 | int ret = 0; | |
1359 | ||
3b0efdfa | 1360 | list_for_each_entry(cachep, &cache_chain, list) { |
8f9f8d9e DR |
1361 | struct kmem_list3 *l3; |
1362 | ||
1363 | l3 = cachep->nodelists[node]; | |
1364 | if (!l3) | |
1365 | continue; | |
1366 | ||
1367 | drain_freelist(cachep, l3, l3->free_objects); | |
1368 | ||
1369 | if (!list_empty(&l3->slabs_full) || | |
1370 | !list_empty(&l3->slabs_partial)) { | |
1371 | ret = -EBUSY; | |
1372 | break; | |
1373 | } | |
1374 | } | |
1375 | return ret; | |
1376 | } | |
1377 | ||
1378 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1379 | unsigned long action, void *arg) | |
1380 | { | |
1381 | struct memory_notify *mnb = arg; | |
1382 | int ret = 0; | |
1383 | int nid; | |
1384 | ||
1385 | nid = mnb->status_change_nid; | |
1386 | if (nid < 0) | |
1387 | goto out; | |
1388 | ||
1389 | switch (action) { | |
1390 | case MEM_GOING_ONLINE: | |
1391 | mutex_lock(&cache_chain_mutex); | |
1392 | ret = init_cache_nodelists_node(nid); | |
1393 | mutex_unlock(&cache_chain_mutex); | |
1394 | break; | |
1395 | case MEM_GOING_OFFLINE: | |
1396 | mutex_lock(&cache_chain_mutex); | |
1397 | ret = drain_cache_nodelists_node(nid); | |
1398 | mutex_unlock(&cache_chain_mutex); | |
1399 | break; | |
1400 | case MEM_ONLINE: | |
1401 | case MEM_OFFLINE: | |
1402 | case MEM_CANCEL_ONLINE: | |
1403 | case MEM_CANCEL_OFFLINE: | |
1404 | break; | |
1405 | } | |
1406 | out: | |
5fda1bd5 | 1407 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1408 | } |
1409 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1410 | ||
e498be7d CL |
1411 | /* |
1412 | * swap the static kmem_list3 with kmalloced memory | |
1413 | */ | |
8f9f8d9e DR |
1414 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1415 | int nodeid) | |
e498be7d CL |
1416 | { |
1417 | struct kmem_list3 *ptr; | |
1418 | ||
83b519e8 | 1419 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1420 | BUG_ON(!ptr); |
1421 | ||
e498be7d | 1422 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1423 | /* |
1424 | * Do not assume that spinlocks can be initialized via memcpy: | |
1425 | */ | |
1426 | spin_lock_init(&ptr->list_lock); | |
1427 | ||
e498be7d CL |
1428 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1429 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1430 | } |
1431 | ||
556a169d PE |
1432 | /* |
1433 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1434 | * size of kmem_list3. | |
1435 | */ | |
1436 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1437 | { | |
1438 | int node; | |
1439 | ||
1440 | for_each_online_node(node) { | |
1441 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1442 | cachep->nodelists[node]->next_reap = jiffies + | |
1443 | REAPTIMEOUT_LIST3 + | |
1444 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1445 | } | |
1446 | } | |
1447 | ||
a737b3e2 AM |
1448 | /* |
1449 | * Initialisation. Called after the page allocator have been initialised and | |
1450 | * before smp_init(). | |
1da177e4 LT |
1451 | */ |
1452 | void __init kmem_cache_init(void) | |
1453 | { | |
1454 | size_t left_over; | |
1455 | struct cache_sizes *sizes; | |
1456 | struct cache_names *names; | |
e498be7d | 1457 | int i; |
07ed76b2 | 1458 | int order; |
1ca4cb24 | 1459 | int node; |
e498be7d | 1460 | |
b6e68bc1 | 1461 | if (num_possible_nodes() == 1) |
62918a03 SS |
1462 | use_alien_caches = 0; |
1463 | ||
e498be7d CL |
1464 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1465 | kmem_list3_init(&initkmem_list3[i]); | |
1466 | if (i < MAX_NUMNODES) | |
1467 | cache_cache.nodelists[i] = NULL; | |
1468 | } | |
556a169d | 1469 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1470 | |
1471 | /* | |
1472 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1473 | * page orders on machines with more than 32MB of memory if |
1474 | * not overridden on the command line. | |
1da177e4 | 1475 | */ |
3df1cccd | 1476 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1477 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1478 | |
1da177e4 LT |
1479 | /* Bootstrap is tricky, because several objects are allocated |
1480 | * from caches that do not exist yet: | |
a737b3e2 AM |
1481 | * 1) initialize the cache_cache cache: it contains the struct |
1482 | * kmem_cache structures of all caches, except cache_cache itself: | |
1483 | * cache_cache is statically allocated. | |
e498be7d CL |
1484 | * Initially an __init data area is used for the head array and the |
1485 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1486 | * array at the end of the bootstrap. | |
1da177e4 | 1487 | * 2) Create the first kmalloc cache. |
343e0d7a | 1488 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1489 | * An __init data area is used for the head array. |
1490 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1491 | * head arrays. | |
1da177e4 LT |
1492 | * 4) Replace the __init data head arrays for cache_cache and the first |
1493 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1494 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1495 | * the other cache's with kmalloc allocated memory. | |
1496 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1497 | */ |
1498 | ||
7d6e6d09 | 1499 | node = numa_mem_id(); |
1ca4cb24 | 1500 | |
1da177e4 | 1501 | /* 1) create the cache_cache */ |
1da177e4 | 1502 | INIT_LIST_HEAD(&cache_chain); |
3b0efdfa | 1503 | list_add(&cache_cache.list, &cache_chain); |
1da177e4 LT |
1504 | cache_cache.colour_off = cache_line_size(); |
1505 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1506 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1507 | |
8da3430d | 1508 | /* |
b56efcf0 | 1509 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1510 | */ |
3b0efdfa | 1511 | cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
b56efcf0 | 1512 | nr_node_ids * sizeof(struct kmem_list3 *); |
3b0efdfa CL |
1513 | cache_cache.object_size = cache_cache.size; |
1514 | cache_cache.size = ALIGN(cache_cache.size, | |
a737b3e2 | 1515 | cache_line_size()); |
6a2d7a95 | 1516 | cache_cache.reciprocal_buffer_size = |
3b0efdfa | 1517 | reciprocal_value(cache_cache.size); |
1da177e4 | 1518 | |
07ed76b2 | 1519 | for (order = 0; order < MAX_ORDER; order++) { |
3b0efdfa | 1520 | cache_estimate(order, cache_cache.size, |
07ed76b2 JS |
1521 | cache_line_size(), 0, &left_over, &cache_cache.num); |
1522 | if (cache_cache.num) | |
1523 | break; | |
1524 | } | |
40094fa6 | 1525 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1526 | cache_cache.gfporder = order; |
b28a02de | 1527 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1528 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1529 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1530 | |
1531 | /* 2+3) create the kmalloc caches */ | |
1532 | sizes = malloc_sizes; | |
1533 | names = cache_names; | |
1534 | ||
a737b3e2 AM |
1535 | /* |
1536 | * Initialize the caches that provide memory for the array cache and the | |
1537 | * kmem_list3 structures first. Without this, further allocations will | |
1538 | * bug. | |
e498be7d CL |
1539 | */ |
1540 | ||
039363f3 | 1541 | sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name, |
a737b3e2 AM |
1542 | sizes[INDEX_AC].cs_size, |
1543 | ARCH_KMALLOC_MINALIGN, | |
1544 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1545 | NULL); |
e498be7d | 1546 | |
a737b3e2 | 1547 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1548 | sizes[INDEX_L3].cs_cachep = |
039363f3 | 1549 | __kmem_cache_create(names[INDEX_L3].name, |
a737b3e2 AM |
1550 | sizes[INDEX_L3].cs_size, |
1551 | ARCH_KMALLOC_MINALIGN, | |
1552 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1553 | NULL); |
a737b3e2 | 1554 | } |
e498be7d | 1555 | |
e0a42726 IM |
1556 | slab_early_init = 0; |
1557 | ||
1da177e4 | 1558 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1559 | /* |
1560 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1561 | * This should be particularly beneficial on SMP boxes, as it |
1562 | * eliminates "false sharing". | |
1563 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1564 | * allow tighter packing of the smaller caches. |
1565 | */ | |
a737b3e2 | 1566 | if (!sizes->cs_cachep) { |
039363f3 | 1567 | sizes->cs_cachep = __kmem_cache_create(names->name, |
a737b3e2 AM |
1568 | sizes->cs_size, |
1569 | ARCH_KMALLOC_MINALIGN, | |
1570 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1571 | NULL); |
a737b3e2 | 1572 | } |
4b51d669 | 1573 | #ifdef CONFIG_ZONE_DMA |
039363f3 | 1574 | sizes->cs_dmacachep = __kmem_cache_create( |
4b51d669 | 1575 | names->name_dma, |
a737b3e2 AM |
1576 | sizes->cs_size, |
1577 | ARCH_KMALLOC_MINALIGN, | |
1578 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1579 | SLAB_PANIC, | |
20c2df83 | 1580 | NULL); |
4b51d669 | 1581 | #endif |
1da177e4 LT |
1582 | sizes++; |
1583 | names++; | |
1584 | } | |
1585 | /* 4) Replace the bootstrap head arrays */ | |
1586 | { | |
2b2d5493 | 1587 | struct array_cache *ptr; |
e498be7d | 1588 | |
83b519e8 | 1589 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1590 | |
9a2dba4b PE |
1591 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1592 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1593 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1594 | /* |
1595 | * Do not assume that spinlocks can be initialized via memcpy: | |
1596 | */ | |
1597 | spin_lock_init(&ptr->lock); | |
1598 | ||
1da177e4 | 1599 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1600 | |
83b519e8 | 1601 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1602 | |
9a2dba4b | 1603 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1604 | != &initarray_generic.cache); |
9a2dba4b | 1605 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1606 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1607 | /* |
1608 | * Do not assume that spinlocks can be initialized via memcpy: | |
1609 | */ | |
1610 | spin_lock_init(&ptr->lock); | |
1611 | ||
e498be7d | 1612 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1613 | ptr; |
1da177e4 | 1614 | } |
e498be7d CL |
1615 | /* 5) Replace the bootstrap kmem_list3's */ |
1616 | { | |
1ca4cb24 PE |
1617 | int nid; |
1618 | ||
9c09a95c | 1619 | for_each_online_node(nid) { |
ec1f5eee | 1620 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1621 | |
e498be7d | 1622 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1623 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1624 | |
1625 | if (INDEX_AC != INDEX_L3) { | |
1626 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1627 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1628 | } |
1629 | } | |
1630 | } | |
1da177e4 | 1631 | |
97d06609 | 1632 | slab_state = UP; |
8429db5c PE |
1633 | } |
1634 | ||
1635 | void __init kmem_cache_init_late(void) | |
1636 | { | |
1637 | struct kmem_cache *cachep; | |
1638 | ||
97d06609 | 1639 | slab_state = UP; |
52cef189 | 1640 | |
30765b92 PZ |
1641 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1642 | init_lock_keys(); | |
1643 | ||
8429db5c PE |
1644 | /* 6) resize the head arrays to their final sizes */ |
1645 | mutex_lock(&cache_chain_mutex); | |
3b0efdfa | 1646 | list_for_each_entry(cachep, &cache_chain, list) |
8429db5c PE |
1647 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1648 | BUG(); | |
1649 | mutex_unlock(&cache_chain_mutex); | |
056c6241 | 1650 | |
97d06609 CL |
1651 | /* Done! */ |
1652 | slab_state = FULL; | |
1653 | ||
a737b3e2 AM |
1654 | /* |
1655 | * Register a cpu startup notifier callback that initializes | |
1656 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1657 | */ |
1658 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1659 | |
8f9f8d9e DR |
1660 | #ifdef CONFIG_NUMA |
1661 | /* | |
1662 | * Register a memory hotplug callback that initializes and frees | |
1663 | * nodelists. | |
1664 | */ | |
1665 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1666 | #endif | |
1667 | ||
a737b3e2 AM |
1668 | /* |
1669 | * The reap timers are started later, with a module init call: That part | |
1670 | * of the kernel is not yet operational. | |
1da177e4 LT |
1671 | */ |
1672 | } | |
1673 | ||
1674 | static int __init cpucache_init(void) | |
1675 | { | |
1676 | int cpu; | |
1677 | ||
a737b3e2 AM |
1678 | /* |
1679 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1680 | */ |
e498be7d | 1681 | for_each_online_cpu(cpu) |
a737b3e2 | 1682 | start_cpu_timer(cpu); |
a164f896 GC |
1683 | |
1684 | /* Done! */ | |
97d06609 | 1685 | slab_state = FULL; |
1da177e4 LT |
1686 | return 0; |
1687 | } | |
1da177e4 LT |
1688 | __initcall(cpucache_init); |
1689 | ||
8bdec192 RA |
1690 | static noinline void |
1691 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1692 | { | |
1693 | struct kmem_list3 *l3; | |
1694 | struct slab *slabp; | |
1695 | unsigned long flags; | |
1696 | int node; | |
1697 | ||
1698 | printk(KERN_WARNING | |
1699 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1700 | nodeid, gfpflags); | |
1701 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1702 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1703 | |
1704 | for_each_online_node(node) { | |
1705 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1706 | unsigned long active_slabs = 0, num_slabs = 0; | |
1707 | ||
1708 | l3 = cachep->nodelists[node]; | |
1709 | if (!l3) | |
1710 | continue; | |
1711 | ||
1712 | spin_lock_irqsave(&l3->list_lock, flags); | |
1713 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1714 | active_objs += cachep->num; | |
1715 | active_slabs++; | |
1716 | } | |
1717 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1718 | active_objs += slabp->inuse; | |
1719 | active_slabs++; | |
1720 | } | |
1721 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1722 | num_slabs++; | |
1723 | ||
1724 | free_objects += l3->free_objects; | |
1725 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1726 | ||
1727 | num_slabs += active_slabs; | |
1728 | num_objs = num_slabs * cachep->num; | |
1729 | printk(KERN_WARNING | |
1730 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1731 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1732 | free_objects); | |
1733 | } | |
1734 | } | |
1735 | ||
1da177e4 LT |
1736 | /* |
1737 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1738 | * | |
1739 | * If we requested dmaable memory, we will get it. Even if we | |
1740 | * did not request dmaable memory, we might get it, but that | |
1741 | * would be relatively rare and ignorable. | |
1742 | */ | |
343e0d7a | 1743 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1744 | { |
1745 | struct page *page; | |
e1b6aa6f | 1746 | int nr_pages; |
1da177e4 LT |
1747 | int i; |
1748 | ||
d6fef9da | 1749 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1750 | /* |
1751 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1752 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1753 | */ |
e1b6aa6f | 1754 | flags |= __GFP_COMP; |
d6fef9da | 1755 | #endif |
765c4507 | 1756 | |
a618e89f | 1757 | flags |= cachep->allocflags; |
e12ba74d MG |
1758 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1759 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1760 | |
517d0869 | 1761 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1762 | if (!page) { |
1763 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1764 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1765 | return NULL; |
8bdec192 | 1766 | } |
1da177e4 | 1767 | |
e1b6aa6f | 1768 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1769 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1770 | add_zone_page_state(page_zone(page), |
1771 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1772 | else | |
1773 | add_zone_page_state(page_zone(page), | |
1774 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
e1b6aa6f CH |
1775 | for (i = 0; i < nr_pages; i++) |
1776 | __SetPageSlab(page + i); | |
c175eea4 | 1777 | |
b1eeab67 VN |
1778 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1779 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1780 | ||
1781 | if (cachep->ctor) | |
1782 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1783 | else | |
1784 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1785 | } | |
c175eea4 | 1786 | |
e1b6aa6f | 1787 | return page_address(page); |
1da177e4 LT |
1788 | } |
1789 | ||
1790 | /* | |
1791 | * Interface to system's page release. | |
1792 | */ | |
343e0d7a | 1793 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1794 | { |
b28a02de | 1795 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1796 | struct page *page = virt_to_page(addr); |
1797 | const unsigned long nr_freed = i; | |
1798 | ||
b1eeab67 | 1799 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1800 | |
972d1a7b CL |
1801 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1802 | sub_zone_page_state(page_zone(page), | |
1803 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1804 | else | |
1805 | sub_zone_page_state(page_zone(page), | |
1806 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1807 | while (i--) { |
f205b2fe NP |
1808 | BUG_ON(!PageSlab(page)); |
1809 | __ClearPageSlab(page); | |
1da177e4 LT |
1810 | page++; |
1811 | } | |
1da177e4 LT |
1812 | if (current->reclaim_state) |
1813 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1814 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1815 | } |
1816 | ||
1817 | static void kmem_rcu_free(struct rcu_head *head) | |
1818 | { | |
b28a02de | 1819 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1820 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1821 | |
1822 | kmem_freepages(cachep, slab_rcu->addr); | |
1823 | if (OFF_SLAB(cachep)) | |
1824 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1825 | } | |
1826 | ||
1827 | #if DEBUG | |
1828 | ||
1829 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1830 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1831 | unsigned long caller) |
1da177e4 | 1832 | { |
8c138bc0 | 1833 | int size = cachep->object_size; |
1da177e4 | 1834 | |
3dafccf2 | 1835 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1836 | |
b28a02de | 1837 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1838 | return; |
1839 | ||
b28a02de PE |
1840 | *addr++ = 0x12345678; |
1841 | *addr++ = caller; | |
1842 | *addr++ = smp_processor_id(); | |
1843 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1844 | { |
1845 | unsigned long *sptr = &caller; | |
1846 | unsigned long svalue; | |
1847 | ||
1848 | while (!kstack_end(sptr)) { | |
1849 | svalue = *sptr++; | |
1850 | if (kernel_text_address(svalue)) { | |
b28a02de | 1851 | *addr++ = svalue; |
1da177e4 LT |
1852 | size -= sizeof(unsigned long); |
1853 | if (size <= sizeof(unsigned long)) | |
1854 | break; | |
1855 | } | |
1856 | } | |
1857 | ||
1858 | } | |
b28a02de | 1859 | *addr++ = 0x87654321; |
1da177e4 LT |
1860 | } |
1861 | #endif | |
1862 | ||
343e0d7a | 1863 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1864 | { |
8c138bc0 | 1865 | int size = cachep->object_size; |
3dafccf2 | 1866 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1867 | |
1868 | memset(addr, val, size); | |
b28a02de | 1869 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1870 | } |
1871 | ||
1872 | static void dump_line(char *data, int offset, int limit) | |
1873 | { | |
1874 | int i; | |
aa83aa40 DJ |
1875 | unsigned char error = 0; |
1876 | int bad_count = 0; | |
1877 | ||
fdde6abb | 1878 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
1879 | for (i = 0; i < limit; i++) { |
1880 | if (data[offset + i] != POISON_FREE) { | |
1881 | error = data[offset + i]; | |
1882 | bad_count++; | |
1883 | } | |
aa83aa40 | 1884 | } |
fdde6abb SAS |
1885 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1886 | &data[offset], limit, 1); | |
aa83aa40 DJ |
1887 | |
1888 | if (bad_count == 1) { | |
1889 | error ^= POISON_FREE; | |
1890 | if (!(error & (error - 1))) { | |
1891 | printk(KERN_ERR "Single bit error detected. Probably " | |
1892 | "bad RAM.\n"); | |
1893 | #ifdef CONFIG_X86 | |
1894 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1895 | "test tool.\n"); | |
1896 | #else | |
1897 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1898 | #endif | |
1899 | } | |
1900 | } | |
1da177e4 LT |
1901 | } |
1902 | #endif | |
1903 | ||
1904 | #if DEBUG | |
1905 | ||
343e0d7a | 1906 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
1907 | { |
1908 | int i, size; | |
1909 | char *realobj; | |
1910 | ||
1911 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 1912 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
1913 | *dbg_redzone1(cachep, objp), |
1914 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
1915 | } |
1916 | ||
1917 | if (cachep->flags & SLAB_STORE_USER) { | |
1918 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 1919 | *dbg_userword(cachep, objp)); |
1da177e4 | 1920 | print_symbol("(%s)", |
a737b3e2 | 1921 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
1922 | printk("\n"); |
1923 | } | |
3dafccf2 | 1924 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 1925 | size = cachep->object_size; |
b28a02de | 1926 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
1927 | int limit; |
1928 | limit = 16; | |
b28a02de PE |
1929 | if (i + limit > size) |
1930 | limit = size - i; | |
1da177e4 LT |
1931 | dump_line(realobj, i, limit); |
1932 | } | |
1933 | } | |
1934 | ||
343e0d7a | 1935 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
1936 | { |
1937 | char *realobj; | |
1938 | int size, i; | |
1939 | int lines = 0; | |
1940 | ||
3dafccf2 | 1941 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 1942 | size = cachep->object_size; |
1da177e4 | 1943 | |
b28a02de | 1944 | for (i = 0; i < size; i++) { |
1da177e4 | 1945 | char exp = POISON_FREE; |
b28a02de | 1946 | if (i == size - 1) |
1da177e4 LT |
1947 | exp = POISON_END; |
1948 | if (realobj[i] != exp) { | |
1949 | int limit; | |
1950 | /* Mismatch ! */ | |
1951 | /* Print header */ | |
1952 | if (lines == 0) { | |
b28a02de | 1953 | printk(KERN_ERR |
face37f5 DJ |
1954 | "Slab corruption (%s): %s start=%p, len=%d\n", |
1955 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
1956 | print_objinfo(cachep, objp, 0); |
1957 | } | |
1958 | /* Hexdump the affected line */ | |
b28a02de | 1959 | i = (i / 16) * 16; |
1da177e4 | 1960 | limit = 16; |
b28a02de PE |
1961 | if (i + limit > size) |
1962 | limit = size - i; | |
1da177e4 LT |
1963 | dump_line(realobj, i, limit); |
1964 | i += 16; | |
1965 | lines++; | |
1966 | /* Limit to 5 lines */ | |
1967 | if (lines > 5) | |
1968 | break; | |
1969 | } | |
1970 | } | |
1971 | if (lines != 0) { | |
1972 | /* Print some data about the neighboring objects, if they | |
1973 | * exist: | |
1974 | */ | |
6ed5eb22 | 1975 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 1976 | unsigned int objnr; |
1da177e4 | 1977 | |
8fea4e96 | 1978 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 1979 | if (objnr) { |
8fea4e96 | 1980 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 1981 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1982 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 1983 | realobj, size); |
1da177e4 LT |
1984 | print_objinfo(cachep, objp, 2); |
1985 | } | |
b28a02de | 1986 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 1987 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 1988 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 1989 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 1990 | realobj, size); |
1da177e4 LT |
1991 | print_objinfo(cachep, objp, 2); |
1992 | } | |
1993 | } | |
1994 | } | |
1995 | #endif | |
1996 | ||
12dd36fa | 1997 | #if DEBUG |
e79aec29 | 1998 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 1999 | { |
1da177e4 LT |
2000 | int i; |
2001 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2002 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2003 | |
2004 | if (cachep->flags & SLAB_POISON) { | |
2005 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2006 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2007 | OFF_SLAB(cachep)) |
b28a02de | 2008 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2009 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2010 | else |
2011 | check_poison_obj(cachep, objp); | |
2012 | #else | |
2013 | check_poison_obj(cachep, objp); | |
2014 | #endif | |
2015 | } | |
2016 | if (cachep->flags & SLAB_RED_ZONE) { | |
2017 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2018 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2019 | "was overwritten"); |
1da177e4 LT |
2020 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2021 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2022 | "was overwritten"); |
1da177e4 | 2023 | } |
1da177e4 | 2024 | } |
12dd36fa | 2025 | } |
1da177e4 | 2026 | #else |
e79aec29 | 2027 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2028 | { |
12dd36fa | 2029 | } |
1da177e4 LT |
2030 | #endif |
2031 | ||
911851e6 RD |
2032 | /** |
2033 | * slab_destroy - destroy and release all objects in a slab | |
2034 | * @cachep: cache pointer being destroyed | |
2035 | * @slabp: slab pointer being destroyed | |
2036 | * | |
12dd36fa | 2037 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2038 | * Before calling the slab must have been unlinked from the cache. The |
2039 | * cache-lock is not held/needed. | |
12dd36fa | 2040 | */ |
343e0d7a | 2041 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2042 | { |
2043 | void *addr = slabp->s_mem - slabp->colouroff; | |
2044 | ||
e79aec29 | 2045 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2046 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2047 | struct slab_rcu *slab_rcu; | |
2048 | ||
b28a02de | 2049 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2050 | slab_rcu->cachep = cachep; |
2051 | slab_rcu->addr = addr; | |
2052 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2053 | } else { | |
2054 | kmem_freepages(cachep, addr); | |
873623df IM |
2055 | if (OFF_SLAB(cachep)) |
2056 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2057 | } |
2058 | } | |
2059 | ||
117f6eb1 CL |
2060 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
2061 | { | |
2062 | int i; | |
2063 | struct kmem_list3 *l3; | |
2064 | ||
2065 | for_each_online_cpu(i) | |
2066 | kfree(cachep->array[i]); | |
2067 | ||
2068 | /* NUMA: free the list3 structures */ | |
2069 | for_each_online_node(i) { | |
2070 | l3 = cachep->nodelists[i]; | |
2071 | if (l3) { | |
2072 | kfree(l3->shared); | |
2073 | free_alien_cache(l3->alien); | |
2074 | kfree(l3); | |
2075 | } | |
2076 | } | |
2077 | kmem_cache_free(&cache_cache, cachep); | |
2078 | } | |
2079 | ||
2080 | ||
4d268eba | 2081 | /** |
a70773dd RD |
2082 | * calculate_slab_order - calculate size (page order) of slabs |
2083 | * @cachep: pointer to the cache that is being created | |
2084 | * @size: size of objects to be created in this cache. | |
2085 | * @align: required alignment for the objects. | |
2086 | * @flags: slab allocation flags | |
2087 | * | |
2088 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2089 | * |
2090 | * This could be made much more intelligent. For now, try to avoid using | |
2091 | * high order pages for slabs. When the gfp() functions are more friendly | |
2092 | * towards high-order requests, this should be changed. | |
2093 | */ | |
a737b3e2 | 2094 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2095 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2096 | { |
b1ab41c4 | 2097 | unsigned long offslab_limit; |
4d268eba | 2098 | size_t left_over = 0; |
9888e6fa | 2099 | int gfporder; |
4d268eba | 2100 | |
0aa817f0 | 2101 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2102 | unsigned int num; |
2103 | size_t remainder; | |
2104 | ||
9888e6fa | 2105 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2106 | if (!num) |
2107 | continue; | |
9888e6fa | 2108 | |
b1ab41c4 IM |
2109 | if (flags & CFLGS_OFF_SLAB) { |
2110 | /* | |
2111 | * Max number of objs-per-slab for caches which | |
2112 | * use off-slab slabs. Needed to avoid a possible | |
2113 | * looping condition in cache_grow(). | |
2114 | */ | |
2115 | offslab_limit = size - sizeof(struct slab); | |
2116 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2117 | ||
2118 | if (num > offslab_limit) | |
2119 | break; | |
2120 | } | |
4d268eba | 2121 | |
9888e6fa | 2122 | /* Found something acceptable - save it away */ |
4d268eba | 2123 | cachep->num = num; |
9888e6fa | 2124 | cachep->gfporder = gfporder; |
4d268eba PE |
2125 | left_over = remainder; |
2126 | ||
f78bb8ad LT |
2127 | /* |
2128 | * A VFS-reclaimable slab tends to have most allocations | |
2129 | * as GFP_NOFS and we really don't want to have to be allocating | |
2130 | * higher-order pages when we are unable to shrink dcache. | |
2131 | */ | |
2132 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2133 | break; | |
2134 | ||
4d268eba PE |
2135 | /* |
2136 | * Large number of objects is good, but very large slabs are | |
2137 | * currently bad for the gfp()s. | |
2138 | */ | |
543585cc | 2139 | if (gfporder >= slab_max_order) |
4d268eba PE |
2140 | break; |
2141 | ||
9888e6fa LT |
2142 | /* |
2143 | * Acceptable internal fragmentation? | |
2144 | */ | |
a737b3e2 | 2145 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2146 | break; |
2147 | } | |
2148 | return left_over; | |
2149 | } | |
2150 | ||
83b519e8 | 2151 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2152 | { |
97d06609 | 2153 | if (slab_state >= FULL) |
83b519e8 | 2154 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2155 | |
97d06609 | 2156 | if (slab_state == DOWN) { |
f30cf7d1 PE |
2157 | /* |
2158 | * Note: the first kmem_cache_create must create the cache | |
2159 | * that's used by kmalloc(24), otherwise the creation of | |
2160 | * further caches will BUG(). | |
2161 | */ | |
2162 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2163 | ||
2164 | /* | |
2165 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2166 | * the first cache, then we need to set up all its list3s, | |
2167 | * otherwise the creation of further caches will BUG(). | |
2168 | */ | |
2169 | set_up_list3s(cachep, SIZE_AC); | |
2170 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2171 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2172 | else |
97d06609 | 2173 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 PE |
2174 | } else { |
2175 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2176 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2177 | |
97d06609 | 2178 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2179 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2180 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2181 | } else { |
2182 | int node; | |
556a169d | 2183 | for_each_online_node(node) { |
f30cf7d1 PE |
2184 | cachep->nodelists[node] = |
2185 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2186 | gfp, node); |
f30cf7d1 PE |
2187 | BUG_ON(!cachep->nodelists[node]); |
2188 | kmem_list3_init(cachep->nodelists[node]); | |
2189 | } | |
2190 | } | |
2191 | } | |
7d6e6d09 | 2192 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2193 | jiffies + REAPTIMEOUT_LIST3 + |
2194 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2195 | ||
2196 | cpu_cache_get(cachep)->avail = 0; | |
2197 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2198 | cpu_cache_get(cachep)->batchcount = 1; | |
2199 | cpu_cache_get(cachep)->touched = 0; | |
2200 | cachep->batchcount = 1; | |
2201 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2202 | return 0; |
f30cf7d1 PE |
2203 | } |
2204 | ||
1da177e4 | 2205 | /** |
039363f3 | 2206 | * __kmem_cache_create - Create a cache. |
1da177e4 LT |
2207 | * @name: A string which is used in /proc/slabinfo to identify this cache. |
2208 | * @size: The size of objects to be created in this cache. | |
2209 | * @align: The required alignment for the objects. | |
2210 | * @flags: SLAB flags | |
2211 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2212 | * |
2213 | * Returns a ptr to the cache on success, NULL on failure. | |
2214 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2215 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2216 | * |
2217 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2218 | * the module calling this has to destroy the cache before getting unloaded. |
2219 | * | |
1da177e4 LT |
2220 | * The flags are |
2221 | * | |
2222 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2223 | * to catch references to uninitialised memory. | |
2224 | * | |
2225 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2226 | * for buffer overruns. | |
2227 | * | |
1da177e4 LT |
2228 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2229 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2230 | * as davem. | |
2231 | */ | |
343e0d7a | 2232 | struct kmem_cache * |
039363f3 | 2233 | __kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2234 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2235 | { |
2236 | size_t left_over, slab_size, ralign; | |
7a7c381d | 2237 | struct kmem_cache *cachep = NULL, *pc; |
83b519e8 | 2238 | gfp_t gfp; |
1da177e4 LT |
2239 | |
2240 | /* | |
2241 | * Sanity checks... these are all serious usage bugs. | |
2242 | */ | |
a737b3e2 | 2243 | if (!name || in_interrupt() || (size < BYTES_PER_WORD) || |
20c2df83 | 2244 | size > KMALLOC_MAX_SIZE) { |
d40cee24 | 2245 | printk(KERN_ERR "%s: Early error in slab %s\n", __func__, |
a737b3e2 | 2246 | name); |
b28a02de PE |
2247 | BUG(); |
2248 | } | |
1da177e4 | 2249 | |
f0188f47 | 2250 | /* |
8f5be20b | 2251 | * We use cache_chain_mutex to ensure a consistent view of |
174596a0 | 2252 | * cpu_online_mask as well. Please see cpuup_callback |
f0188f47 | 2253 | */ |
83b519e8 PE |
2254 | if (slab_is_available()) { |
2255 | get_online_cpus(); | |
2256 | mutex_lock(&cache_chain_mutex); | |
2257 | } | |
4f12bb4f | 2258 | |
3b0efdfa | 2259 | list_for_each_entry(pc, &cache_chain, list) { |
4f12bb4f AM |
2260 | char tmp; |
2261 | int res; | |
2262 | ||
2263 | /* | |
2264 | * This happens when the module gets unloaded and doesn't | |
2265 | * destroy its slab cache and no-one else reuses the vmalloc | |
2266 | * area of the module. Print a warning. | |
2267 | */ | |
138ae663 | 2268 | res = probe_kernel_address(pc->name, tmp); |
4f12bb4f | 2269 | if (res) { |
b4169525 | 2270 | printk(KERN_ERR |
2271 | "SLAB: cache with size %d has lost its name\n", | |
3b0efdfa | 2272 | pc->size); |
4f12bb4f AM |
2273 | continue; |
2274 | } | |
2275 | ||
b28a02de | 2276 | if (!strcmp(pc->name, name)) { |
b4169525 | 2277 | printk(KERN_ERR |
2278 | "kmem_cache_create: duplicate cache %s\n", name); | |
4f12bb4f AM |
2279 | dump_stack(); |
2280 | goto oops; | |
2281 | } | |
2282 | } | |
2283 | ||
1da177e4 LT |
2284 | #if DEBUG |
2285 | WARN_ON(strchr(name, ' ')); /* It confuses parsers */ | |
1da177e4 LT |
2286 | #if FORCED_DEBUG |
2287 | /* | |
2288 | * Enable redzoning and last user accounting, except for caches with | |
2289 | * large objects, if the increased size would increase the object size | |
2290 | * above the next power of two: caches with object sizes just above a | |
2291 | * power of two have a significant amount of internal fragmentation. | |
2292 | */ | |
87a927c7 DW |
2293 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2294 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2295 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2296 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2297 | flags |= SLAB_POISON; | |
2298 | #endif | |
2299 | if (flags & SLAB_DESTROY_BY_RCU) | |
2300 | BUG_ON(flags & SLAB_POISON); | |
2301 | #endif | |
1da177e4 | 2302 | /* |
a737b3e2 AM |
2303 | * Always checks flags, a caller might be expecting debug support which |
2304 | * isn't available. | |
1da177e4 | 2305 | */ |
40094fa6 | 2306 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2307 | |
a737b3e2 AM |
2308 | /* |
2309 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2310 | * unaligned accesses for some archs when redzoning is used, and makes |
2311 | * sure any on-slab bufctl's are also correctly aligned. | |
2312 | */ | |
b28a02de PE |
2313 | if (size & (BYTES_PER_WORD - 1)) { |
2314 | size += (BYTES_PER_WORD - 1); | |
2315 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2316 | } |
2317 | ||
a737b3e2 AM |
2318 | /* calculate the final buffer alignment: */ |
2319 | ||
1da177e4 LT |
2320 | /* 1) arch recommendation: can be overridden for debug */ |
2321 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2322 | /* |
2323 | * Default alignment: as specified by the arch code. Except if | |
2324 | * an object is really small, then squeeze multiple objects into | |
2325 | * one cacheline. | |
1da177e4 LT |
2326 | */ |
2327 | ralign = cache_line_size(); | |
b28a02de | 2328 | while (size <= ralign / 2) |
1da177e4 LT |
2329 | ralign /= 2; |
2330 | } else { | |
2331 | ralign = BYTES_PER_WORD; | |
2332 | } | |
ca5f9703 PE |
2333 | |
2334 | /* | |
87a927c7 DW |
2335 | * Redzoning and user store require word alignment or possibly larger. |
2336 | * Note this will be overridden by architecture or caller mandated | |
2337 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2338 | */ |
87a927c7 DW |
2339 | if (flags & SLAB_STORE_USER) |
2340 | ralign = BYTES_PER_WORD; | |
2341 | ||
2342 | if (flags & SLAB_RED_ZONE) { | |
2343 | ralign = REDZONE_ALIGN; | |
2344 | /* If redzoning, ensure that the second redzone is suitably | |
2345 | * aligned, by adjusting the object size accordingly. */ | |
2346 | size += REDZONE_ALIGN - 1; | |
2347 | size &= ~(REDZONE_ALIGN - 1); | |
2348 | } | |
ca5f9703 | 2349 | |
a44b56d3 | 2350 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2351 | if (ralign < ARCH_SLAB_MINALIGN) { |
2352 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2353 | } |
a44b56d3 | 2354 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2355 | if (ralign < align) { |
2356 | ralign = align; | |
1da177e4 | 2357 | } |
3ff84a7f PE |
2358 | /* disable debug if necessary */ |
2359 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2360 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2361 | /* |
ca5f9703 | 2362 | * 4) Store it. |
1da177e4 LT |
2363 | */ |
2364 | align = ralign; | |
2365 | ||
83b519e8 PE |
2366 | if (slab_is_available()) |
2367 | gfp = GFP_KERNEL; | |
2368 | else | |
2369 | gfp = GFP_NOWAIT; | |
2370 | ||
1da177e4 | 2371 | /* Get cache's description obj. */ |
83b519e8 | 2372 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2373 | if (!cachep) |
039363f3 | 2374 | return NULL; |
1da177e4 | 2375 | |
b56efcf0 | 2376 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
3b0efdfa CL |
2377 | cachep->object_size = size; |
2378 | cachep->align = align; | |
1da177e4 | 2379 | #if DEBUG |
1da177e4 | 2380 | |
ca5f9703 PE |
2381 | /* |
2382 | * Both debugging options require word-alignment which is calculated | |
2383 | * into align above. | |
2384 | */ | |
1da177e4 | 2385 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2386 | /* add space for red zone words */ |
3ff84a7f PE |
2387 | cachep->obj_offset += sizeof(unsigned long long); |
2388 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2389 | } |
2390 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2391 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2392 | * the real object. But if the second red zone needs to be |
2393 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2394 | */ |
87a927c7 DW |
2395 | if (flags & SLAB_RED_ZONE) |
2396 | size += REDZONE_ALIGN; | |
2397 | else | |
2398 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2399 | } |
2400 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2401 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3b0efdfa | 2402 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
1ab335d8 | 2403 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); |
1da177e4 LT |
2404 | size = PAGE_SIZE; |
2405 | } | |
2406 | #endif | |
2407 | #endif | |
2408 | ||
e0a42726 IM |
2409 | /* |
2410 | * Determine if the slab management is 'on' or 'off' slab. | |
2411 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2412 | * it too early on. Always use on-slab management when |
2413 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2414 | */ |
e7cb55b9 CM |
2415 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2416 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2417 | /* |
2418 | * Size is large, assume best to place the slab management obj | |
2419 | * off-slab (should allow better packing of objs). | |
2420 | */ | |
2421 | flags |= CFLGS_OFF_SLAB; | |
2422 | ||
2423 | size = ALIGN(size, align); | |
2424 | ||
f78bb8ad | 2425 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2426 | |
2427 | if (!cachep->num) { | |
b4169525 | 2428 | printk(KERN_ERR |
2429 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 | 2430 | kmem_cache_free(&cache_cache, cachep); |
039363f3 | 2431 | return NULL; |
1da177e4 | 2432 | } |
b28a02de PE |
2433 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2434 | + sizeof(struct slab), align); | |
1da177e4 LT |
2435 | |
2436 | /* | |
2437 | * If the slab has been placed off-slab, and we have enough space then | |
2438 | * move it on-slab. This is at the expense of any extra colouring. | |
2439 | */ | |
2440 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2441 | flags &= ~CFLGS_OFF_SLAB; | |
2442 | left_over -= slab_size; | |
2443 | } | |
2444 | ||
2445 | if (flags & CFLGS_OFF_SLAB) { | |
2446 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2447 | slab_size = |
2448 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2449 | |
2450 | #ifdef CONFIG_PAGE_POISONING | |
2451 | /* If we're going to use the generic kernel_map_pages() | |
2452 | * poisoning, then it's going to smash the contents of | |
2453 | * the redzone and userword anyhow, so switch them off. | |
2454 | */ | |
2455 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2456 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2457 | #endif | |
1da177e4 LT |
2458 | } |
2459 | ||
2460 | cachep->colour_off = cache_line_size(); | |
2461 | /* Offset must be a multiple of the alignment. */ | |
2462 | if (cachep->colour_off < align) | |
2463 | cachep->colour_off = align; | |
b28a02de | 2464 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2465 | cachep->slab_size = slab_size; |
2466 | cachep->flags = flags; | |
a618e89f | 2467 | cachep->allocflags = 0; |
4b51d669 | 2468 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2469 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2470 | cachep->size = size; |
6a2d7a95 | 2471 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2472 | |
e5ac9c5a | 2473 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2474 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2475 | /* |
2476 | * This is a possibility for one of the malloc_sizes caches. | |
2477 | * But since we go off slab only for object size greater than | |
2478 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2479 | * this should not happen at all. | |
2480 | * But leave a BUG_ON for some lucky dude. | |
2481 | */ | |
6cb8f913 | 2482 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2483 | } |
1da177e4 | 2484 | cachep->ctor = ctor; |
1da177e4 LT |
2485 | cachep->name = name; |
2486 | ||
83b519e8 | 2487 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef | 2488 | __kmem_cache_destroy(cachep); |
039363f3 | 2489 | return NULL; |
2ed3a4ef | 2490 | } |
1da177e4 | 2491 | |
83835b3d PZ |
2492 | if (flags & SLAB_DEBUG_OBJECTS) { |
2493 | /* | |
2494 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2495 | * debug_object_activate()->kmem_cache_alloc(). | |
2496 | */ | |
2497 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2498 | ||
2499 | slab_set_debugobj_lock_classes(cachep); | |
2500 | } | |
2501 | ||
1da177e4 | 2502 | /* cache setup completed, link it into the list */ |
3b0efdfa | 2503 | list_add(&cachep->list, &cache_chain); |
a737b3e2 | 2504 | oops: |
83b519e8 PE |
2505 | if (slab_is_available()) { |
2506 | mutex_unlock(&cache_chain_mutex); | |
2507 | put_online_cpus(); | |
2508 | } | |
1da177e4 LT |
2509 | return cachep; |
2510 | } | |
1da177e4 LT |
2511 | |
2512 | #if DEBUG | |
2513 | static void check_irq_off(void) | |
2514 | { | |
2515 | BUG_ON(!irqs_disabled()); | |
2516 | } | |
2517 | ||
2518 | static void check_irq_on(void) | |
2519 | { | |
2520 | BUG_ON(irqs_disabled()); | |
2521 | } | |
2522 | ||
343e0d7a | 2523 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2524 | { |
2525 | #ifdef CONFIG_SMP | |
2526 | check_irq_off(); | |
7d6e6d09 | 2527 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2528 | #endif |
2529 | } | |
e498be7d | 2530 | |
343e0d7a | 2531 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2532 | { |
2533 | #ifdef CONFIG_SMP | |
2534 | check_irq_off(); | |
2535 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2536 | #endif | |
2537 | } | |
2538 | ||
1da177e4 LT |
2539 | #else |
2540 | #define check_irq_off() do { } while(0) | |
2541 | #define check_irq_on() do { } while(0) | |
2542 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2543 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2544 | #endif |
2545 | ||
aab2207c CL |
2546 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2547 | struct array_cache *ac, | |
2548 | int force, int node); | |
2549 | ||
1da177e4 LT |
2550 | static void do_drain(void *arg) |
2551 | { | |
a737b3e2 | 2552 | struct kmem_cache *cachep = arg; |
1da177e4 | 2553 | struct array_cache *ac; |
7d6e6d09 | 2554 | int node = numa_mem_id(); |
1da177e4 LT |
2555 | |
2556 | check_irq_off(); | |
9a2dba4b | 2557 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2558 | spin_lock(&cachep->nodelists[node]->list_lock); |
2559 | free_block(cachep, ac->entry, ac->avail, node); | |
2560 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2561 | ac->avail = 0; |
2562 | } | |
2563 | ||
343e0d7a | 2564 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2565 | { |
e498be7d CL |
2566 | struct kmem_list3 *l3; |
2567 | int node; | |
2568 | ||
15c8b6c1 | 2569 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2570 | check_irq_on(); |
b28a02de | 2571 | for_each_online_node(node) { |
e498be7d | 2572 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2573 | if (l3 && l3->alien) |
2574 | drain_alien_cache(cachep, l3->alien); | |
2575 | } | |
2576 | ||
2577 | for_each_online_node(node) { | |
2578 | l3 = cachep->nodelists[node]; | |
2579 | if (l3) | |
aab2207c | 2580 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2581 | } |
1da177e4 LT |
2582 | } |
2583 | ||
ed11d9eb CL |
2584 | /* |
2585 | * Remove slabs from the list of free slabs. | |
2586 | * Specify the number of slabs to drain in tofree. | |
2587 | * | |
2588 | * Returns the actual number of slabs released. | |
2589 | */ | |
2590 | static int drain_freelist(struct kmem_cache *cache, | |
2591 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2592 | { |
ed11d9eb CL |
2593 | struct list_head *p; |
2594 | int nr_freed; | |
1da177e4 | 2595 | struct slab *slabp; |
1da177e4 | 2596 | |
ed11d9eb CL |
2597 | nr_freed = 0; |
2598 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2599 | |
ed11d9eb | 2600 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2601 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2602 | if (p == &l3->slabs_free) { |
2603 | spin_unlock_irq(&l3->list_lock); | |
2604 | goto out; | |
2605 | } | |
1da177e4 | 2606 | |
ed11d9eb | 2607 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2608 | #if DEBUG |
40094fa6 | 2609 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2610 | #endif |
2611 | list_del(&slabp->list); | |
ed11d9eb CL |
2612 | /* |
2613 | * Safe to drop the lock. The slab is no longer linked | |
2614 | * to the cache. | |
2615 | */ | |
2616 | l3->free_objects -= cache->num; | |
e498be7d | 2617 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2618 | slab_destroy(cache, slabp); |
2619 | nr_freed++; | |
1da177e4 | 2620 | } |
ed11d9eb CL |
2621 | out: |
2622 | return nr_freed; | |
1da177e4 LT |
2623 | } |
2624 | ||
8f5be20b | 2625 | /* Called with cache_chain_mutex held to protect against cpu hotplug */ |
343e0d7a | 2626 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2627 | { |
2628 | int ret = 0, i = 0; | |
2629 | struct kmem_list3 *l3; | |
2630 | ||
2631 | drain_cpu_caches(cachep); | |
2632 | ||
2633 | check_irq_on(); | |
2634 | for_each_online_node(i) { | |
2635 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2636 | if (!l3) |
2637 | continue; | |
2638 | ||
2639 | drain_freelist(cachep, l3, l3->free_objects); | |
2640 | ||
2641 | ret += !list_empty(&l3->slabs_full) || | |
2642 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2643 | } |
2644 | return (ret ? 1 : 0); | |
2645 | } | |
2646 | ||
1da177e4 LT |
2647 | /** |
2648 | * kmem_cache_shrink - Shrink a cache. | |
2649 | * @cachep: The cache to shrink. | |
2650 | * | |
2651 | * Releases as many slabs as possible for a cache. | |
2652 | * To help debugging, a zero exit status indicates all slabs were released. | |
2653 | */ | |
343e0d7a | 2654 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2655 | { |
8f5be20b | 2656 | int ret; |
40094fa6 | 2657 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2658 | |
95402b38 | 2659 | get_online_cpus(); |
8f5be20b RT |
2660 | mutex_lock(&cache_chain_mutex); |
2661 | ret = __cache_shrink(cachep); | |
2662 | mutex_unlock(&cache_chain_mutex); | |
95402b38 | 2663 | put_online_cpus(); |
8f5be20b | 2664 | return ret; |
1da177e4 LT |
2665 | } |
2666 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2667 | ||
2668 | /** | |
2669 | * kmem_cache_destroy - delete a cache | |
2670 | * @cachep: the cache to destroy | |
2671 | * | |
72fd4a35 | 2672 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2673 | * |
2674 | * It is expected this function will be called by a module when it is | |
2675 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2676 | * cache being allocated each time a module is loaded and unloaded, if the | |
2677 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2678 | * | |
2679 | * The cache must be empty before calling this function. | |
2680 | * | |
25985edc | 2681 | * The caller must guarantee that no one will allocate memory from the cache |
1da177e4 LT |
2682 | * during the kmem_cache_destroy(). |
2683 | */ | |
133d205a | 2684 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2685 | { |
40094fa6 | 2686 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2687 | |
1da177e4 | 2688 | /* Find the cache in the chain of caches. */ |
95402b38 | 2689 | get_online_cpus(); |
fc0abb14 | 2690 | mutex_lock(&cache_chain_mutex); |
1da177e4 LT |
2691 | /* |
2692 | * the chain is never empty, cache_cache is never destroyed | |
2693 | */ | |
3b0efdfa | 2694 | list_del(&cachep->list); |
1da177e4 LT |
2695 | if (__cache_shrink(cachep)) { |
2696 | slab_error(cachep, "Can't free all objects"); | |
3b0efdfa | 2697 | list_add(&cachep->list, &cache_chain); |
fc0abb14 | 2698 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2699 | put_online_cpus(); |
133d205a | 2700 | return; |
1da177e4 LT |
2701 | } |
2702 | ||
2703 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
7ed9f7e5 | 2704 | rcu_barrier(); |
1da177e4 | 2705 | |
117f6eb1 | 2706 | __kmem_cache_destroy(cachep); |
8f5be20b | 2707 | mutex_unlock(&cache_chain_mutex); |
95402b38 | 2708 | put_online_cpus(); |
1da177e4 LT |
2709 | } |
2710 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2711 | ||
e5ac9c5a RT |
2712 | /* |
2713 | * Get the memory for a slab management obj. | |
2714 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2715 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2716 | * come from the same cache which is getting created because, | |
2717 | * when we are searching for an appropriate cache for these | |
2718 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2719 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2720 | * kmem_find_general_cachep till the initialization is complete. | |
2721 | * Hence we cannot have slabp_cache same as the original cache. | |
2722 | */ | |
343e0d7a | 2723 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2724 | int colour_off, gfp_t local_flags, |
2725 | int nodeid) | |
1da177e4 LT |
2726 | { |
2727 | struct slab *slabp; | |
b28a02de | 2728 | |
1da177e4 LT |
2729 | if (OFF_SLAB(cachep)) { |
2730 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2731 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2732 | local_flags, nodeid); |
d5cff635 CM |
2733 | /* |
2734 | * If the first object in the slab is leaked (it's allocated | |
2735 | * but no one has a reference to it), we want to make sure | |
2736 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2737 | * to the object. Otherwise we will not report the leak. | |
2738 | */ | |
c017b4be CM |
2739 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2740 | local_flags); | |
1da177e4 LT |
2741 | if (!slabp) |
2742 | return NULL; | |
2743 | } else { | |
b28a02de | 2744 | slabp = objp + colour_off; |
1da177e4 LT |
2745 | colour_off += cachep->slab_size; |
2746 | } | |
2747 | slabp->inuse = 0; | |
2748 | slabp->colouroff = colour_off; | |
b28a02de | 2749 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2750 | slabp->nodeid = nodeid; |
e51bfd0a | 2751 | slabp->free = 0; |
1da177e4 LT |
2752 | return slabp; |
2753 | } | |
2754 | ||
2755 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2756 | { | |
b28a02de | 2757 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2758 | } |
2759 | ||
343e0d7a | 2760 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2761 | struct slab *slabp) |
1da177e4 LT |
2762 | { |
2763 | int i; | |
2764 | ||
2765 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2766 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2767 | #if DEBUG |
2768 | /* need to poison the objs? */ | |
2769 | if (cachep->flags & SLAB_POISON) | |
2770 | poison_obj(cachep, objp, POISON_FREE); | |
2771 | if (cachep->flags & SLAB_STORE_USER) | |
2772 | *dbg_userword(cachep, objp) = NULL; | |
2773 | ||
2774 | if (cachep->flags & SLAB_RED_ZONE) { | |
2775 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2776 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2777 | } | |
2778 | /* | |
a737b3e2 AM |
2779 | * Constructors are not allowed to allocate memory from the same |
2780 | * cache which they are a constructor for. Otherwise, deadlock. | |
2781 | * They must also be threaded. | |
1da177e4 LT |
2782 | */ |
2783 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2784 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2785 | |
2786 | if (cachep->flags & SLAB_RED_ZONE) { | |
2787 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2788 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2789 | " end of an object"); |
1da177e4 LT |
2790 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2791 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2792 | " start of an object"); |
1da177e4 | 2793 | } |
3b0efdfa | 2794 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2795 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2796 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2797 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2798 | #else |
2799 | if (cachep->ctor) | |
51cc5068 | 2800 | cachep->ctor(objp); |
1da177e4 | 2801 | #endif |
b28a02de | 2802 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2803 | } |
b28a02de | 2804 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2805 | } |
2806 | ||
343e0d7a | 2807 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2808 | { |
4b51d669 CL |
2809 | if (CONFIG_ZONE_DMA_FLAG) { |
2810 | if (flags & GFP_DMA) | |
a618e89f | 2811 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2812 | else |
a618e89f | 2813 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2814 | } |
1da177e4 LT |
2815 | } |
2816 | ||
a737b3e2 AM |
2817 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2818 | int nodeid) | |
78d382d7 | 2819 | { |
8fea4e96 | 2820 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2821 | kmem_bufctl_t next; |
2822 | ||
2823 | slabp->inuse++; | |
2824 | next = slab_bufctl(slabp)[slabp->free]; | |
2825 | #if DEBUG | |
2826 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2827 | WARN_ON(slabp->nodeid != nodeid); | |
2828 | #endif | |
2829 | slabp->free = next; | |
2830 | ||
2831 | return objp; | |
2832 | } | |
2833 | ||
a737b3e2 AM |
2834 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2835 | void *objp, int nodeid) | |
78d382d7 | 2836 | { |
8fea4e96 | 2837 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2838 | |
2839 | #if DEBUG | |
2840 | /* Verify that the slab belongs to the intended node */ | |
2841 | WARN_ON(slabp->nodeid != nodeid); | |
2842 | ||
871751e2 | 2843 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2844 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2845 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2846 | BUG(); |
2847 | } | |
2848 | #endif | |
2849 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2850 | slabp->free = objnr; | |
2851 | slabp->inuse--; | |
2852 | } | |
2853 | ||
4776874f PE |
2854 | /* |
2855 | * Map pages beginning at addr to the given cache and slab. This is required | |
2856 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2857 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2858 | */ |
2859 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2860 | void *addr) | |
1da177e4 | 2861 | { |
4776874f | 2862 | int nr_pages; |
1da177e4 LT |
2863 | struct page *page; |
2864 | ||
4776874f | 2865 | page = virt_to_page(addr); |
84097518 | 2866 | |
4776874f | 2867 | nr_pages = 1; |
84097518 | 2868 | if (likely(!PageCompound(page))) |
4776874f PE |
2869 | nr_pages <<= cache->gfporder; |
2870 | ||
1da177e4 | 2871 | do { |
35026088 CL |
2872 | page->slab_cache = cache; |
2873 | page->slab_page = slab; | |
1da177e4 | 2874 | page++; |
4776874f | 2875 | } while (--nr_pages); |
1da177e4 LT |
2876 | } |
2877 | ||
2878 | /* | |
2879 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2880 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2881 | */ | |
3c517a61 CL |
2882 | static int cache_grow(struct kmem_cache *cachep, |
2883 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2884 | { |
b28a02de | 2885 | struct slab *slabp; |
b28a02de PE |
2886 | size_t offset; |
2887 | gfp_t local_flags; | |
e498be7d | 2888 | struct kmem_list3 *l3; |
1da177e4 | 2889 | |
a737b3e2 AM |
2890 | /* |
2891 | * Be lazy and only check for valid flags here, keeping it out of the | |
2892 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2893 | */ |
6cb06229 CL |
2894 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2895 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2896 | |
2e1217cf | 2897 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2898 | check_irq_off(); |
2e1217cf RT |
2899 | l3 = cachep->nodelists[nodeid]; |
2900 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2901 | |
2902 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2903 | offset = l3->colour_next; |
2904 | l3->colour_next++; | |
2905 | if (l3->colour_next >= cachep->colour) | |
2906 | l3->colour_next = 0; | |
2907 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2908 | |
2e1217cf | 2909 | offset *= cachep->colour_off; |
1da177e4 LT |
2910 | |
2911 | if (local_flags & __GFP_WAIT) | |
2912 | local_irq_enable(); | |
2913 | ||
2914 | /* | |
2915 | * The test for missing atomic flag is performed here, rather than | |
2916 | * the more obvious place, simply to reduce the critical path length | |
2917 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2918 | * will eventually be caught here (where it matters). | |
2919 | */ | |
2920 | kmem_flagcheck(cachep, flags); | |
2921 | ||
a737b3e2 AM |
2922 | /* |
2923 | * Get mem for the objs. Attempt to allocate a physical page from | |
2924 | * 'nodeid'. | |
e498be7d | 2925 | */ |
3c517a61 | 2926 | if (!objp) |
b8c1c5da | 2927 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2928 | if (!objp) |
1da177e4 LT |
2929 | goto failed; |
2930 | ||
2931 | /* Get slab management. */ | |
3c517a61 | 2932 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2933 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2934 | if (!slabp) |
1da177e4 LT |
2935 | goto opps1; |
2936 | ||
4776874f | 2937 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2938 | |
a35afb83 | 2939 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2940 | |
2941 | if (local_flags & __GFP_WAIT) | |
2942 | local_irq_disable(); | |
2943 | check_irq_off(); | |
e498be7d | 2944 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2945 | |
2946 | /* Make slab active. */ | |
e498be7d | 2947 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2948 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2949 | l3->free_objects += cachep->num; |
2950 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2951 | return 1; |
a737b3e2 | 2952 | opps1: |
1da177e4 | 2953 | kmem_freepages(cachep, objp); |
a737b3e2 | 2954 | failed: |
1da177e4 LT |
2955 | if (local_flags & __GFP_WAIT) |
2956 | local_irq_disable(); | |
2957 | return 0; | |
2958 | } | |
2959 | ||
2960 | #if DEBUG | |
2961 | ||
2962 | /* | |
2963 | * Perform extra freeing checks: | |
2964 | * - detect bad pointers. | |
2965 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2966 | */ |
2967 | static void kfree_debugcheck(const void *objp) | |
2968 | { | |
1da177e4 LT |
2969 | if (!virt_addr_valid(objp)) { |
2970 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2971 | (unsigned long)objp); |
2972 | BUG(); | |
1da177e4 | 2973 | } |
1da177e4 LT |
2974 | } |
2975 | ||
58ce1fd5 PE |
2976 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2977 | { | |
b46b8f19 | 2978 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2979 | |
2980 | redzone1 = *dbg_redzone1(cache, obj); | |
2981 | redzone2 = *dbg_redzone2(cache, obj); | |
2982 | ||
2983 | /* | |
2984 | * Redzone is ok. | |
2985 | */ | |
2986 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2987 | return; | |
2988 | ||
2989 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2990 | slab_error(cache, "double free detected"); | |
2991 | else | |
2992 | slab_error(cache, "memory outside object was overwritten"); | |
2993 | ||
b46b8f19 | 2994 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2995 | obj, redzone1, redzone2); |
2996 | } | |
2997 | ||
343e0d7a | 2998 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 2999 | void *caller) |
1da177e4 LT |
3000 | { |
3001 | struct page *page; | |
3002 | unsigned int objnr; | |
3003 | struct slab *slabp; | |
3004 | ||
80cbd911 MW |
3005 | BUG_ON(virt_to_cache(objp) != cachep); |
3006 | ||
3dafccf2 | 3007 | objp -= obj_offset(cachep); |
1da177e4 | 3008 | kfree_debugcheck(objp); |
b49af68f | 3009 | page = virt_to_head_page(objp); |
1da177e4 | 3010 | |
35026088 | 3011 | slabp = page->slab_page; |
1da177e4 LT |
3012 | |
3013 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3014 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3015 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3016 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3017 | } | |
3018 | if (cachep->flags & SLAB_STORE_USER) | |
3019 | *dbg_userword(cachep, objp) = caller; | |
3020 | ||
8fea4e96 | 3021 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3022 | |
3023 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3024 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3025 | |
871751e2 AV |
3026 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3027 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3028 | #endif | |
1da177e4 LT |
3029 | if (cachep->flags & SLAB_POISON) { |
3030 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3031 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 3032 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 3033 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3034 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3035 | } else { |
3036 | poison_obj(cachep, objp, POISON_FREE); | |
3037 | } | |
3038 | #else | |
3039 | poison_obj(cachep, objp, POISON_FREE); | |
3040 | #endif | |
3041 | } | |
3042 | return objp; | |
3043 | } | |
3044 | ||
343e0d7a | 3045 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3046 | { |
3047 | kmem_bufctl_t i; | |
3048 | int entries = 0; | |
b28a02de | 3049 | |
1da177e4 LT |
3050 | /* Check slab's freelist to see if this obj is there. */ |
3051 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3052 | entries++; | |
3053 | if (entries > cachep->num || i >= cachep->num) | |
3054 | goto bad; | |
3055 | } | |
3056 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3057 | bad: |
3058 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3059 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3060 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3061 | print_tainted()); | |
fdde6abb SAS |
3062 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3063 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3064 | 1); | |
1da177e4 LT |
3065 | BUG(); |
3066 | } | |
3067 | } | |
3068 | #else | |
3069 | #define kfree_debugcheck(x) do { } while(0) | |
3070 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3071 | #define check_slabp(x,y) do { } while(0) | |
3072 | #endif | |
3073 | ||
343e0d7a | 3074 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 LT |
3075 | { |
3076 | int batchcount; | |
3077 | struct kmem_list3 *l3; | |
3078 | struct array_cache *ac; | |
1ca4cb24 PE |
3079 | int node; |
3080 | ||
6d2144d3 | 3081 | retry: |
1da177e4 | 3082 | check_irq_off(); |
7d6e6d09 | 3083 | node = numa_mem_id(); |
9a2dba4b | 3084 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3085 | batchcount = ac->batchcount; |
3086 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3087 | /* |
3088 | * If there was little recent activity on this cache, then | |
3089 | * perform only a partial refill. Otherwise we could generate | |
3090 | * refill bouncing. | |
1da177e4 LT |
3091 | */ |
3092 | batchcount = BATCHREFILL_LIMIT; | |
3093 | } | |
1ca4cb24 | 3094 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3095 | |
3096 | BUG_ON(ac->avail > 0 || !l3); | |
3097 | spin_lock(&l3->list_lock); | |
1da177e4 | 3098 | |
3ded175a | 3099 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3100 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3101 | l3->shared->touched = 1; | |
3ded175a | 3102 | goto alloc_done; |
44b57f1c | 3103 | } |
3ded175a | 3104 | |
1da177e4 LT |
3105 | while (batchcount > 0) { |
3106 | struct list_head *entry; | |
3107 | struct slab *slabp; | |
3108 | /* Get slab alloc is to come from. */ | |
3109 | entry = l3->slabs_partial.next; | |
3110 | if (entry == &l3->slabs_partial) { | |
3111 | l3->free_touched = 1; | |
3112 | entry = l3->slabs_free.next; | |
3113 | if (entry == &l3->slabs_free) | |
3114 | goto must_grow; | |
3115 | } | |
3116 | ||
3117 | slabp = list_entry(entry, struct slab, list); | |
3118 | check_slabp(cachep, slabp); | |
3119 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3120 | |
3121 | /* | |
3122 | * The slab was either on partial or free list so | |
3123 | * there must be at least one object available for | |
3124 | * allocation. | |
3125 | */ | |
249b9f33 | 3126 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3127 | |
1da177e4 | 3128 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3129 | STATS_INC_ALLOCED(cachep); |
3130 | STATS_INC_ACTIVE(cachep); | |
3131 | STATS_SET_HIGH(cachep); | |
3132 | ||
78d382d7 | 3133 | ac->entry[ac->avail++] = slab_get_obj(cachep, slabp, |
1ca4cb24 | 3134 | node); |
1da177e4 LT |
3135 | } |
3136 | check_slabp(cachep, slabp); | |
3137 | ||
3138 | /* move slabp to correct slabp list: */ | |
3139 | list_del(&slabp->list); | |
3140 | if (slabp->free == BUFCTL_END) | |
3141 | list_add(&slabp->list, &l3->slabs_full); | |
3142 | else | |
3143 | list_add(&slabp->list, &l3->slabs_partial); | |
3144 | } | |
3145 | ||
a737b3e2 | 3146 | must_grow: |
1da177e4 | 3147 | l3->free_objects -= ac->avail; |
a737b3e2 | 3148 | alloc_done: |
e498be7d | 3149 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3150 | |
3151 | if (unlikely(!ac->avail)) { | |
3152 | int x; | |
3c517a61 | 3153 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3154 | |
a737b3e2 | 3155 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3156 | ac = cpu_cache_get(cachep); |
a737b3e2 | 3157 | if (!x && ac->avail == 0) /* no objects in sight? abort */ |
1da177e4 LT |
3158 | return NULL; |
3159 | ||
a737b3e2 | 3160 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3161 | goto retry; |
3162 | } | |
3163 | ac->touched = 1; | |
e498be7d | 3164 | return ac->entry[--ac->avail]; |
1da177e4 LT |
3165 | } |
3166 | ||
a737b3e2 AM |
3167 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3168 | gfp_t flags) | |
1da177e4 LT |
3169 | { |
3170 | might_sleep_if(flags & __GFP_WAIT); | |
3171 | #if DEBUG | |
3172 | kmem_flagcheck(cachep, flags); | |
3173 | #endif | |
3174 | } | |
3175 | ||
3176 | #if DEBUG | |
a737b3e2 AM |
3177 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3178 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3179 | { |
b28a02de | 3180 | if (!objp) |
1da177e4 | 3181 | return objp; |
b28a02de | 3182 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3183 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3184 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3185 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3186 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3187 | else |
3188 | check_poison_obj(cachep, objp); | |
3189 | #else | |
3190 | check_poison_obj(cachep, objp); | |
3191 | #endif | |
3192 | poison_obj(cachep, objp, POISON_INUSE); | |
3193 | } | |
3194 | if (cachep->flags & SLAB_STORE_USER) | |
3195 | *dbg_userword(cachep, objp) = caller; | |
3196 | ||
3197 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3198 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3199 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3200 | slab_error(cachep, "double free, or memory outside" | |
3201 | " object was overwritten"); | |
b28a02de | 3202 | printk(KERN_ERR |
b46b8f19 | 3203 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3204 | objp, *dbg_redzone1(cachep, objp), |
3205 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3206 | } |
3207 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3208 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3209 | } | |
871751e2 AV |
3210 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3211 | { | |
3212 | struct slab *slabp; | |
3213 | unsigned objnr; | |
3214 | ||
35026088 | 3215 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3216 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3217 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3218 | } | |
3219 | #endif | |
3dafccf2 | 3220 | objp += obj_offset(cachep); |
4f104934 | 3221 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3222 | cachep->ctor(objp); |
7ea466f2 TH |
3223 | if (ARCH_SLAB_MINALIGN && |
3224 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3225 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3226 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3227 | } |
1da177e4 LT |
3228 | return objp; |
3229 | } | |
3230 | #else | |
3231 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3232 | #endif | |
3233 | ||
773ff60e | 3234 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3235 | { |
3236 | if (cachep == &cache_cache) | |
773ff60e | 3237 | return false; |
8a8b6502 | 3238 | |
8c138bc0 | 3239 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3240 | } |
3241 | ||
343e0d7a | 3242 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3243 | { |
b28a02de | 3244 | void *objp; |
1da177e4 LT |
3245 | struct array_cache *ac; |
3246 | ||
5c382300 | 3247 | check_irq_off(); |
8a8b6502 | 3248 | |
9a2dba4b | 3249 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3250 | if (likely(ac->avail)) { |
3251 | STATS_INC_ALLOCHIT(cachep); | |
3252 | ac->touched = 1; | |
e498be7d | 3253 | objp = ac->entry[--ac->avail]; |
1da177e4 LT |
3254 | } else { |
3255 | STATS_INC_ALLOCMISS(cachep); | |
3256 | objp = cache_alloc_refill(cachep, flags); | |
ddbf2e83 O |
3257 | /* |
3258 | * the 'ac' may be updated by cache_alloc_refill(), | |
3259 | * and kmemleak_erase() requires its correct value. | |
3260 | */ | |
3261 | ac = cpu_cache_get(cachep); | |
1da177e4 | 3262 | } |
d5cff635 CM |
3263 | /* |
3264 | * To avoid a false negative, if an object that is in one of the | |
3265 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3266 | * treat the array pointers as a reference to the object. | |
3267 | */ | |
f3d8b53a O |
3268 | if (objp) |
3269 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3270 | return objp; |
3271 | } | |
3272 | ||
e498be7d | 3273 | #ifdef CONFIG_NUMA |
c61afb18 | 3274 | /* |
b2455396 | 3275 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3276 | * |
3277 | * If we are in_interrupt, then process context, including cpusets and | |
3278 | * mempolicy, may not apply and should not be used for allocation policy. | |
3279 | */ | |
3280 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3281 | { | |
3282 | int nid_alloc, nid_here; | |
3283 | ||
765c4507 | 3284 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3285 | return NULL; |
7d6e6d09 | 3286 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3287 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3288 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3289 | else if (current->mempolicy) |
e7b691b0 | 3290 | nid_alloc = slab_node(); |
c61afb18 | 3291 | if (nid_alloc != nid_here) |
8b98c169 | 3292 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3293 | return NULL; |
3294 | } | |
3295 | ||
765c4507 CL |
3296 | /* |
3297 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3298 | * certain node and fall back is permitted. First we scan all the |
3299 | * available nodelists for available objects. If that fails then we | |
3300 | * perform an allocation without specifying a node. This allows the page | |
3301 | * allocator to do its reclaim / fallback magic. We then insert the | |
3302 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3303 | */ |
8c8cc2c1 | 3304 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3305 | { |
8c8cc2c1 PE |
3306 | struct zonelist *zonelist; |
3307 | gfp_t local_flags; | |
dd1a239f | 3308 | struct zoneref *z; |
54a6eb5c MG |
3309 | struct zone *zone; |
3310 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3311 | void *obj = NULL; |
3c517a61 | 3312 | int nid; |
cc9a6c87 | 3313 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3314 | |
3315 | if (flags & __GFP_THISNODE) | |
3316 | return NULL; | |
3317 | ||
6cb06229 | 3318 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3319 | |
cc9a6c87 MG |
3320 | retry_cpuset: |
3321 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3322 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3323 | |
3c517a61 CL |
3324 | retry: |
3325 | /* | |
3326 | * Look through allowed nodes for objects available | |
3327 | * from existing per node queues. | |
3328 | */ | |
54a6eb5c MG |
3329 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3330 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3331 | |
54a6eb5c | 3332 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3333 | cache->nodelists[nid] && |
481c5346 | 3334 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3335 | obj = ____cache_alloc_node(cache, |
3336 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3337 | if (obj) |
3338 | break; | |
3339 | } | |
3c517a61 CL |
3340 | } |
3341 | ||
cfce6604 | 3342 | if (!obj) { |
3c517a61 CL |
3343 | /* |
3344 | * This allocation will be performed within the constraints | |
3345 | * of the current cpuset / memory policy requirements. | |
3346 | * We may trigger various forms of reclaim on the allowed | |
3347 | * set and go into memory reserves if necessary. | |
3348 | */ | |
dd47ea75 CL |
3349 | if (local_flags & __GFP_WAIT) |
3350 | local_irq_enable(); | |
3351 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3352 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3353 | if (local_flags & __GFP_WAIT) |
3354 | local_irq_disable(); | |
3c517a61 CL |
3355 | if (obj) { |
3356 | /* | |
3357 | * Insert into the appropriate per node queues | |
3358 | */ | |
3359 | nid = page_to_nid(virt_to_page(obj)); | |
3360 | if (cache_grow(cache, flags, nid, obj)) { | |
3361 | obj = ____cache_alloc_node(cache, | |
3362 | flags | GFP_THISNODE, nid); | |
3363 | if (!obj) | |
3364 | /* | |
3365 | * Another processor may allocate the | |
3366 | * objects in the slab since we are | |
3367 | * not holding any locks. | |
3368 | */ | |
3369 | goto retry; | |
3370 | } else { | |
b6a60451 | 3371 | /* cache_grow already freed obj */ |
3c517a61 CL |
3372 | obj = NULL; |
3373 | } | |
3374 | } | |
aedb0eb1 | 3375 | } |
cc9a6c87 MG |
3376 | |
3377 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3378 | goto retry_cpuset; | |
765c4507 CL |
3379 | return obj; |
3380 | } | |
3381 | ||
e498be7d CL |
3382 | /* |
3383 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3384 | */ |
8b98c169 | 3385 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3386 | int nodeid) |
e498be7d CL |
3387 | { |
3388 | struct list_head *entry; | |
b28a02de PE |
3389 | struct slab *slabp; |
3390 | struct kmem_list3 *l3; | |
3391 | void *obj; | |
b28a02de PE |
3392 | int x; |
3393 | ||
3394 | l3 = cachep->nodelists[nodeid]; | |
3395 | BUG_ON(!l3); | |
3396 | ||
a737b3e2 | 3397 | retry: |
ca3b9b91 | 3398 | check_irq_off(); |
b28a02de PE |
3399 | spin_lock(&l3->list_lock); |
3400 | entry = l3->slabs_partial.next; | |
3401 | if (entry == &l3->slabs_partial) { | |
3402 | l3->free_touched = 1; | |
3403 | entry = l3->slabs_free.next; | |
3404 | if (entry == &l3->slabs_free) | |
3405 | goto must_grow; | |
3406 | } | |
3407 | ||
3408 | slabp = list_entry(entry, struct slab, list); | |
3409 | check_spinlock_acquired_node(cachep, nodeid); | |
3410 | check_slabp(cachep, slabp); | |
3411 | ||
3412 | STATS_INC_NODEALLOCS(cachep); | |
3413 | STATS_INC_ACTIVE(cachep); | |
3414 | STATS_SET_HIGH(cachep); | |
3415 | ||
3416 | BUG_ON(slabp->inuse == cachep->num); | |
3417 | ||
78d382d7 | 3418 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3419 | check_slabp(cachep, slabp); |
3420 | l3->free_objects--; | |
3421 | /* move slabp to correct slabp list: */ | |
3422 | list_del(&slabp->list); | |
3423 | ||
a737b3e2 | 3424 | if (slabp->free == BUFCTL_END) |
b28a02de | 3425 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3426 | else |
b28a02de | 3427 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3428 | |
b28a02de PE |
3429 | spin_unlock(&l3->list_lock); |
3430 | goto done; | |
e498be7d | 3431 | |
a737b3e2 | 3432 | must_grow: |
b28a02de | 3433 | spin_unlock(&l3->list_lock); |
3c517a61 | 3434 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3435 | if (x) |
3436 | goto retry; | |
1da177e4 | 3437 | |
8c8cc2c1 | 3438 | return fallback_alloc(cachep, flags); |
e498be7d | 3439 | |
a737b3e2 | 3440 | done: |
b28a02de | 3441 | return obj; |
e498be7d | 3442 | } |
8c8cc2c1 PE |
3443 | |
3444 | /** | |
3445 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3446 | * @cachep: The cache to allocate from. | |
3447 | * @flags: See kmalloc(). | |
3448 | * @nodeid: node number of the target node. | |
3449 | * @caller: return address of caller, used for debug information | |
3450 | * | |
3451 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3452 | * node, which can improve the performance for cpu bound structures. | |
3453 | * | |
3454 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3455 | */ | |
3456 | static __always_inline void * | |
3457 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3458 | void *caller) | |
3459 | { | |
3460 | unsigned long save_flags; | |
3461 | void *ptr; | |
7d6e6d09 | 3462 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3463 | |
dcce284a | 3464 | flags &= gfp_allowed_mask; |
7e85ee0c | 3465 | |
cf40bd16 NP |
3466 | lockdep_trace_alloc(flags); |
3467 | ||
773ff60e | 3468 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3469 | return NULL; |
3470 | ||
8c8cc2c1 PE |
3471 | cache_alloc_debugcheck_before(cachep, flags); |
3472 | local_irq_save(save_flags); | |
3473 | ||
eacbbae3 | 3474 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3475 | nodeid = slab_node; |
8c8cc2c1 PE |
3476 | |
3477 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3478 | /* Node not bootstrapped yet */ | |
3479 | ptr = fallback_alloc(cachep, flags); | |
3480 | goto out; | |
3481 | } | |
3482 | ||
7d6e6d09 | 3483 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3484 | /* |
3485 | * Use the locally cached objects if possible. | |
3486 | * However ____cache_alloc does not allow fallback | |
3487 | * to other nodes. It may fail while we still have | |
3488 | * objects on other nodes available. | |
3489 | */ | |
3490 | ptr = ____cache_alloc(cachep, flags); | |
3491 | if (ptr) | |
3492 | goto out; | |
3493 | } | |
3494 | /* ___cache_alloc_node can fall back to other nodes */ | |
3495 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3496 | out: | |
3497 | local_irq_restore(save_flags); | |
3498 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3499 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3500 | flags); |
8c8cc2c1 | 3501 | |
c175eea4 | 3502 | if (likely(ptr)) |
8c138bc0 | 3503 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3504 | |
d07dbea4 | 3505 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3506 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3507 | |
8c8cc2c1 PE |
3508 | return ptr; |
3509 | } | |
3510 | ||
3511 | static __always_inline void * | |
3512 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3513 | { | |
3514 | void *objp; | |
3515 | ||
3516 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3517 | objp = alternate_node_alloc(cache, flags); | |
3518 | if (objp) | |
3519 | goto out; | |
3520 | } | |
3521 | objp = ____cache_alloc(cache, flags); | |
3522 | ||
3523 | /* | |
3524 | * We may just have run out of memory on the local node. | |
3525 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3526 | */ | |
7d6e6d09 LS |
3527 | if (!objp) |
3528 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3529 | |
3530 | out: | |
3531 | return objp; | |
3532 | } | |
3533 | #else | |
3534 | ||
3535 | static __always_inline void * | |
3536 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3537 | { | |
3538 | return ____cache_alloc(cachep, flags); | |
3539 | } | |
3540 | ||
3541 | #endif /* CONFIG_NUMA */ | |
3542 | ||
3543 | static __always_inline void * | |
3544 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3545 | { | |
3546 | unsigned long save_flags; | |
3547 | void *objp; | |
3548 | ||
dcce284a | 3549 | flags &= gfp_allowed_mask; |
7e85ee0c | 3550 | |
cf40bd16 NP |
3551 | lockdep_trace_alloc(flags); |
3552 | ||
773ff60e | 3553 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3554 | return NULL; |
3555 | ||
8c8cc2c1 PE |
3556 | cache_alloc_debugcheck_before(cachep, flags); |
3557 | local_irq_save(save_flags); | |
3558 | objp = __do_cache_alloc(cachep, flags); | |
3559 | local_irq_restore(save_flags); | |
3560 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3561 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3562 | flags); |
8c8cc2c1 PE |
3563 | prefetchw(objp); |
3564 | ||
c175eea4 | 3565 | if (likely(objp)) |
8c138bc0 | 3566 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3567 | |
d07dbea4 | 3568 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3569 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3570 | |
8c8cc2c1 PE |
3571 | return objp; |
3572 | } | |
e498be7d CL |
3573 | |
3574 | /* | |
3575 | * Caller needs to acquire correct kmem_list's list_lock | |
3576 | */ | |
343e0d7a | 3577 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3578 | int node) |
1da177e4 LT |
3579 | { |
3580 | int i; | |
e498be7d | 3581 | struct kmem_list3 *l3; |
1da177e4 LT |
3582 | |
3583 | for (i = 0; i < nr_objects; i++) { | |
3584 | void *objp = objpp[i]; | |
3585 | struct slab *slabp; | |
1da177e4 | 3586 | |
6ed5eb22 | 3587 | slabp = virt_to_slab(objp); |
ff69416e | 3588 | l3 = cachep->nodelists[node]; |
1da177e4 | 3589 | list_del(&slabp->list); |
ff69416e | 3590 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3591 | check_slabp(cachep, slabp); |
78d382d7 | 3592 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3593 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3594 | l3->free_objects++; |
1da177e4 LT |
3595 | check_slabp(cachep, slabp); |
3596 | ||
3597 | /* fixup slab chains */ | |
3598 | if (slabp->inuse == 0) { | |
e498be7d CL |
3599 | if (l3->free_objects > l3->free_limit) { |
3600 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3601 | /* No need to drop any previously held |
3602 | * lock here, even if we have a off-slab slab | |
3603 | * descriptor it is guaranteed to come from | |
3604 | * a different cache, refer to comments before | |
3605 | * alloc_slabmgmt. | |
3606 | */ | |
1da177e4 LT |
3607 | slab_destroy(cachep, slabp); |
3608 | } else { | |
e498be7d | 3609 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3610 | } |
3611 | } else { | |
3612 | /* Unconditionally move a slab to the end of the | |
3613 | * partial list on free - maximum time for the | |
3614 | * other objects to be freed, too. | |
3615 | */ | |
e498be7d | 3616 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3617 | } |
3618 | } | |
3619 | } | |
3620 | ||
343e0d7a | 3621 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3622 | { |
3623 | int batchcount; | |
e498be7d | 3624 | struct kmem_list3 *l3; |
7d6e6d09 | 3625 | int node = numa_mem_id(); |
1da177e4 LT |
3626 | |
3627 | batchcount = ac->batchcount; | |
3628 | #if DEBUG | |
3629 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3630 | #endif | |
3631 | check_irq_off(); | |
ff69416e | 3632 | l3 = cachep->nodelists[node]; |
873623df | 3633 | spin_lock(&l3->list_lock); |
e498be7d CL |
3634 | if (l3->shared) { |
3635 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3636 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3637 | if (max) { |
3638 | if (batchcount > max) | |
3639 | batchcount = max; | |
e498be7d | 3640 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3641 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3642 | shared_array->avail += batchcount; |
3643 | goto free_done; | |
3644 | } | |
3645 | } | |
3646 | ||
ff69416e | 3647 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3648 | free_done: |
1da177e4 LT |
3649 | #if STATS |
3650 | { | |
3651 | int i = 0; | |
3652 | struct list_head *p; | |
3653 | ||
e498be7d CL |
3654 | p = l3->slabs_free.next; |
3655 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3656 | struct slab *slabp; |
3657 | ||
3658 | slabp = list_entry(p, struct slab, list); | |
3659 | BUG_ON(slabp->inuse); | |
3660 | ||
3661 | i++; | |
3662 | p = p->next; | |
3663 | } | |
3664 | STATS_SET_FREEABLE(cachep, i); | |
3665 | } | |
3666 | #endif | |
e498be7d | 3667 | spin_unlock(&l3->list_lock); |
1da177e4 | 3668 | ac->avail -= batchcount; |
a737b3e2 | 3669 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3670 | } |
3671 | ||
3672 | /* | |
a737b3e2 AM |
3673 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3674 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3675 | */ |
a947eb95 SS |
3676 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3677 | void *caller) | |
1da177e4 | 3678 | { |
9a2dba4b | 3679 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3680 | |
3681 | check_irq_off(); | |
d5cff635 | 3682 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3683 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3684 | |
8c138bc0 | 3685 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3686 | |
1807a1aa SS |
3687 | /* |
3688 | * Skip calling cache_free_alien() when the platform is not numa. | |
3689 | * This will avoid cache misses that happen while accessing slabp (which | |
3690 | * is per page memory reference) to get nodeid. Instead use a global | |
3691 | * variable to skip the call, which is mostly likely to be present in | |
3692 | * the cache. | |
3693 | */ | |
b6e68bc1 | 3694 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3695 | return; |
3696 | ||
1da177e4 LT |
3697 | if (likely(ac->avail < ac->limit)) { |
3698 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3699 | } else { |
3700 | STATS_INC_FREEMISS(cachep); | |
3701 | cache_flusharray(cachep, ac); | |
1da177e4 | 3702 | } |
42c8c99c ZJ |
3703 | |
3704 | ac->entry[ac->avail++] = objp; | |
1da177e4 LT |
3705 | } |
3706 | ||
3707 | /** | |
3708 | * kmem_cache_alloc - Allocate an object | |
3709 | * @cachep: The cache to allocate from. | |
3710 | * @flags: See kmalloc(). | |
3711 | * | |
3712 | * Allocate an object from this cache. The flags are only relevant | |
3713 | * if the cache has no available objects. | |
3714 | */ | |
343e0d7a | 3715 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3716 | { |
36555751 EGM |
3717 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3718 | ||
ca2b84cb | 3719 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3720 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3721 | |
3722 | return ret; | |
1da177e4 LT |
3723 | } |
3724 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3725 | ||
0f24f128 | 3726 | #ifdef CONFIG_TRACING |
85beb586 SR |
3727 | void * |
3728 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3729 | { |
85beb586 SR |
3730 | void *ret; |
3731 | ||
3732 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3733 | ||
3734 | trace_kmalloc(_RET_IP_, ret, | |
3735 | size, slab_buffer_size(cachep), flags); | |
3736 | return ret; | |
36555751 | 3737 | } |
85beb586 | 3738 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3739 | #endif |
3740 | ||
1da177e4 | 3741 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3742 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3743 | { | |
36555751 EGM |
3744 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3745 | __builtin_return_address(0)); | |
3746 | ||
ca2b84cb | 3747 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3748 | cachep->object_size, cachep->size, |
ca2b84cb | 3749 | flags, nodeid); |
36555751 EGM |
3750 | |
3751 | return ret; | |
8b98c169 | 3752 | } |
1da177e4 LT |
3753 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3754 | ||
0f24f128 | 3755 | #ifdef CONFIG_TRACING |
85beb586 SR |
3756 | void *kmem_cache_alloc_node_trace(size_t size, |
3757 | struct kmem_cache *cachep, | |
3758 | gfp_t flags, | |
3759 | int nodeid) | |
36555751 | 3760 | { |
85beb586 SR |
3761 | void *ret; |
3762 | ||
3763 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3764 | __builtin_return_address(0)); |
85beb586 SR |
3765 | trace_kmalloc_node(_RET_IP_, ret, |
3766 | size, slab_buffer_size(cachep), | |
3767 | flags, nodeid); | |
3768 | return ret; | |
36555751 | 3769 | } |
85beb586 | 3770 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3771 | #endif |
3772 | ||
8b98c169 CH |
3773 | static __always_inline void * |
3774 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3775 | { |
343e0d7a | 3776 | struct kmem_cache *cachep; |
97e2bde4 MS |
3777 | |
3778 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3779 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3780 | return cachep; | |
85beb586 | 3781 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3782 | } |
8b98c169 | 3783 | |
0bb38a5c | 3784 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3785 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3786 | { | |
3787 | return __do_kmalloc_node(size, flags, node, | |
3788 | __builtin_return_address(0)); | |
3789 | } | |
dbe5e69d | 3790 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3791 | |
3792 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3793 | int node, unsigned long caller) |
8b98c169 | 3794 | { |
ce71e27c | 3795 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3796 | } |
3797 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3798 | #else | |
3799 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3800 | { | |
3801 | return __do_kmalloc_node(size, flags, node, NULL); | |
3802 | } | |
3803 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3804 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3805 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3806 | |
3807 | /** | |
800590f5 | 3808 | * __do_kmalloc - allocate memory |
1da177e4 | 3809 | * @size: how many bytes of memory are required. |
800590f5 | 3810 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3811 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3812 | */ |
7fd6b141 PE |
3813 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3814 | void *caller) | |
1da177e4 | 3815 | { |
343e0d7a | 3816 | struct kmem_cache *cachep; |
36555751 | 3817 | void *ret; |
1da177e4 | 3818 | |
97e2bde4 MS |
3819 | /* If you want to save a few bytes .text space: replace |
3820 | * __ with kmem_. | |
3821 | * Then kmalloc uses the uninlined functions instead of the inline | |
3822 | * functions. | |
3823 | */ | |
3824 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3825 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3826 | return cachep; | |
36555751 EGM |
3827 | ret = __cache_alloc(cachep, flags, caller); |
3828 | ||
ca2b84cb | 3829 | trace_kmalloc((unsigned long) caller, ret, |
3b0efdfa | 3830 | size, cachep->size, flags); |
36555751 EGM |
3831 | |
3832 | return ret; | |
7fd6b141 PE |
3833 | } |
3834 | ||
7fd6b141 | 3835 | |
0bb38a5c | 3836 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3837 | void *__kmalloc(size_t size, gfp_t flags) |
3838 | { | |
871751e2 | 3839 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3840 | } |
3841 | EXPORT_SYMBOL(__kmalloc); | |
3842 | ||
ce71e27c | 3843 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3844 | { |
ce71e27c | 3845 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3846 | } |
3847 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3848 | |
3849 | #else | |
3850 | void *__kmalloc(size_t size, gfp_t flags) | |
3851 | { | |
3852 | return __do_kmalloc(size, flags, NULL); | |
3853 | } | |
3854 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3855 | #endif |
3856 | ||
1da177e4 LT |
3857 | /** |
3858 | * kmem_cache_free - Deallocate an object | |
3859 | * @cachep: The cache the allocation was from. | |
3860 | * @objp: The previously allocated object. | |
3861 | * | |
3862 | * Free an object which was previously allocated from this | |
3863 | * cache. | |
3864 | */ | |
343e0d7a | 3865 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3866 | { |
3867 | unsigned long flags; | |
3868 | ||
3869 | local_irq_save(flags); | |
d97d476b | 3870 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3871 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3872 | debug_check_no_obj_freed(objp, cachep->object_size); |
a947eb95 | 3873 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3874 | local_irq_restore(flags); |
36555751 | 3875 | |
ca2b84cb | 3876 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3877 | } |
3878 | EXPORT_SYMBOL(kmem_cache_free); | |
3879 | ||
1da177e4 LT |
3880 | /** |
3881 | * kfree - free previously allocated memory | |
3882 | * @objp: pointer returned by kmalloc. | |
3883 | * | |
80e93eff PE |
3884 | * If @objp is NULL, no operation is performed. |
3885 | * | |
1da177e4 LT |
3886 | * Don't free memory not originally allocated by kmalloc() |
3887 | * or you will run into trouble. | |
3888 | */ | |
3889 | void kfree(const void *objp) | |
3890 | { | |
343e0d7a | 3891 | struct kmem_cache *c; |
1da177e4 LT |
3892 | unsigned long flags; |
3893 | ||
2121db74 PE |
3894 | trace_kfree(_RET_IP_, objp); |
3895 | ||
6cb8f913 | 3896 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3897 | return; |
3898 | local_irq_save(flags); | |
3899 | kfree_debugcheck(objp); | |
6ed5eb22 | 3900 | c = virt_to_cache(objp); |
8c138bc0 CL |
3901 | debug_check_no_locks_freed(objp, c->object_size); |
3902 | ||
3903 | debug_check_no_obj_freed(objp, c->object_size); | |
a947eb95 | 3904 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
3905 | local_irq_restore(flags); |
3906 | } | |
3907 | EXPORT_SYMBOL(kfree); | |
3908 | ||
343e0d7a | 3909 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3910 | { |
8c138bc0 | 3911 | return cachep->object_size; |
1da177e4 LT |
3912 | } |
3913 | EXPORT_SYMBOL(kmem_cache_size); | |
3914 | ||
e498be7d | 3915 | /* |
183ff22b | 3916 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3917 | */ |
83b519e8 | 3918 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3919 | { |
3920 | int node; | |
3921 | struct kmem_list3 *l3; | |
cafeb02e | 3922 | struct array_cache *new_shared; |
3395ee05 | 3923 | struct array_cache **new_alien = NULL; |
e498be7d | 3924 | |
9c09a95c | 3925 | for_each_online_node(node) { |
cafeb02e | 3926 | |
3395ee05 | 3927 | if (use_alien_caches) { |
83b519e8 | 3928 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3929 | if (!new_alien) |
3930 | goto fail; | |
3931 | } | |
cafeb02e | 3932 | |
63109846 ED |
3933 | new_shared = NULL; |
3934 | if (cachep->shared) { | |
3935 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3936 | cachep->shared*cachep->batchcount, |
83b519e8 | 3937 | 0xbaadf00d, gfp); |
63109846 ED |
3938 | if (!new_shared) { |
3939 | free_alien_cache(new_alien); | |
3940 | goto fail; | |
3941 | } | |
0718dc2a | 3942 | } |
cafeb02e | 3943 | |
a737b3e2 AM |
3944 | l3 = cachep->nodelists[node]; |
3945 | if (l3) { | |
cafeb02e CL |
3946 | struct array_cache *shared = l3->shared; |
3947 | ||
e498be7d CL |
3948 | spin_lock_irq(&l3->list_lock); |
3949 | ||
cafeb02e | 3950 | if (shared) |
0718dc2a CL |
3951 | free_block(cachep, shared->entry, |
3952 | shared->avail, node); | |
e498be7d | 3953 | |
cafeb02e CL |
3954 | l3->shared = new_shared; |
3955 | if (!l3->alien) { | |
e498be7d CL |
3956 | l3->alien = new_alien; |
3957 | new_alien = NULL; | |
3958 | } | |
b28a02de | 3959 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3960 | cachep->batchcount + cachep->num; |
e498be7d | 3961 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3962 | kfree(shared); |
e498be7d CL |
3963 | free_alien_cache(new_alien); |
3964 | continue; | |
3965 | } | |
83b519e8 | 3966 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3967 | if (!l3) { |
3968 | free_alien_cache(new_alien); | |
3969 | kfree(new_shared); | |
e498be7d | 3970 | goto fail; |
0718dc2a | 3971 | } |
e498be7d CL |
3972 | |
3973 | kmem_list3_init(l3); | |
3974 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3975 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3976 | l3->shared = new_shared; |
e498be7d | 3977 | l3->alien = new_alien; |
b28a02de | 3978 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3979 | cachep->batchcount + cachep->num; |
e498be7d CL |
3980 | cachep->nodelists[node] = l3; |
3981 | } | |
cafeb02e | 3982 | return 0; |
0718dc2a | 3983 | |
a737b3e2 | 3984 | fail: |
3b0efdfa | 3985 | if (!cachep->list.next) { |
0718dc2a CL |
3986 | /* Cache is not active yet. Roll back what we did */ |
3987 | node--; | |
3988 | while (node >= 0) { | |
3989 | if (cachep->nodelists[node]) { | |
3990 | l3 = cachep->nodelists[node]; | |
3991 | ||
3992 | kfree(l3->shared); | |
3993 | free_alien_cache(l3->alien); | |
3994 | kfree(l3); | |
3995 | cachep->nodelists[node] = NULL; | |
3996 | } | |
3997 | node--; | |
3998 | } | |
3999 | } | |
cafeb02e | 4000 | return -ENOMEM; |
e498be7d CL |
4001 | } |
4002 | ||
1da177e4 | 4003 | struct ccupdate_struct { |
343e0d7a | 4004 | struct kmem_cache *cachep; |
acfe7d74 | 4005 | struct array_cache *new[0]; |
1da177e4 LT |
4006 | }; |
4007 | ||
4008 | static void do_ccupdate_local(void *info) | |
4009 | { | |
a737b3e2 | 4010 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4011 | struct array_cache *old; |
4012 | ||
4013 | check_irq_off(); | |
9a2dba4b | 4014 | old = cpu_cache_get(new->cachep); |
e498be7d | 4015 | |
1da177e4 LT |
4016 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4017 | new->new[smp_processor_id()] = old; | |
4018 | } | |
4019 | ||
b5d8ca7c | 4020 | /* Always called with the cache_chain_mutex held */ |
a737b3e2 | 4021 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4022 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4023 | { |
d2e7b7d0 | 4024 | struct ccupdate_struct *new; |
2ed3a4ef | 4025 | int i; |
1da177e4 | 4026 | |
acfe7d74 ED |
4027 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4028 | gfp); | |
d2e7b7d0 SS |
4029 | if (!new) |
4030 | return -ENOMEM; | |
4031 | ||
e498be7d | 4032 | for_each_online_cpu(i) { |
7d6e6d09 | 4033 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4034 | batchcount, gfp); |
d2e7b7d0 | 4035 | if (!new->new[i]) { |
b28a02de | 4036 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4037 | kfree(new->new[i]); |
4038 | kfree(new); | |
e498be7d | 4039 | return -ENOMEM; |
1da177e4 LT |
4040 | } |
4041 | } | |
d2e7b7d0 | 4042 | new->cachep = cachep; |
1da177e4 | 4043 | |
15c8b6c1 | 4044 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4045 | |
1da177e4 | 4046 | check_irq_on(); |
1da177e4 LT |
4047 | cachep->batchcount = batchcount; |
4048 | cachep->limit = limit; | |
e498be7d | 4049 | cachep->shared = shared; |
1da177e4 | 4050 | |
e498be7d | 4051 | for_each_online_cpu(i) { |
d2e7b7d0 | 4052 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4053 | if (!ccold) |
4054 | continue; | |
7d6e6d09 LS |
4055 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4056 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4057 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4058 | kfree(ccold); |
4059 | } | |
d2e7b7d0 | 4060 | kfree(new); |
83b519e8 | 4061 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4062 | } |
4063 | ||
b5d8ca7c | 4064 | /* Called with cache_chain_mutex held always */ |
83b519e8 | 4065 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4066 | { |
4067 | int err; | |
4068 | int limit, shared; | |
4069 | ||
a737b3e2 AM |
4070 | /* |
4071 | * The head array serves three purposes: | |
1da177e4 LT |
4072 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4073 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4074 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4075 | * bufctl chains: array operations are cheaper. |
4076 | * The numbers are guessed, we should auto-tune as described by | |
4077 | * Bonwick. | |
4078 | */ | |
3b0efdfa | 4079 | if (cachep->size > 131072) |
1da177e4 | 4080 | limit = 1; |
3b0efdfa | 4081 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4082 | limit = 8; |
3b0efdfa | 4083 | else if (cachep->size > 1024) |
1da177e4 | 4084 | limit = 24; |
3b0efdfa | 4085 | else if (cachep->size > 256) |
1da177e4 LT |
4086 | limit = 54; |
4087 | else | |
4088 | limit = 120; | |
4089 | ||
a737b3e2 AM |
4090 | /* |
4091 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4092 | * allocation behaviour: Most allocs on one cpu, most free operations |
4093 | * on another cpu. For these cases, an efficient object passing between | |
4094 | * cpus is necessary. This is provided by a shared array. The array | |
4095 | * replaces Bonwick's magazine layer. | |
4096 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4097 | * to a larger limit. Thus disabled by default. | |
4098 | */ | |
4099 | shared = 0; | |
3b0efdfa | 4100 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4101 | shared = 8; |
1da177e4 LT |
4102 | |
4103 | #if DEBUG | |
a737b3e2 AM |
4104 | /* |
4105 | * With debugging enabled, large batchcount lead to excessively long | |
4106 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4107 | */ |
4108 | if (limit > 32) | |
4109 | limit = 32; | |
4110 | #endif | |
83b519e8 | 4111 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4112 | if (err) |
4113 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4114 | cachep->name, -err); |
2ed3a4ef | 4115 | return err; |
1da177e4 LT |
4116 | } |
4117 | ||
1b55253a CL |
4118 | /* |
4119 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4120 | * necessary. Note that the l3 listlock also protects the array_cache |
4121 | * if drain_array() is used on the shared array. | |
1b55253a | 4122 | */ |
68a1b195 | 4123 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4124 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4125 | { |
4126 | int tofree; | |
4127 | ||
1b55253a CL |
4128 | if (!ac || !ac->avail) |
4129 | return; | |
1da177e4 LT |
4130 | if (ac->touched && !force) { |
4131 | ac->touched = 0; | |
b18e7e65 | 4132 | } else { |
1b55253a | 4133 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4134 | if (ac->avail) { |
4135 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4136 | if (tofree > ac->avail) | |
4137 | tofree = (ac->avail + 1) / 2; | |
4138 | free_block(cachep, ac->entry, tofree, node); | |
4139 | ac->avail -= tofree; | |
4140 | memmove(ac->entry, &(ac->entry[tofree]), | |
4141 | sizeof(void *) * ac->avail); | |
4142 | } | |
1b55253a | 4143 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4144 | } |
4145 | } | |
4146 | ||
4147 | /** | |
4148 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4149 | * @w: work descriptor |
1da177e4 LT |
4150 | * |
4151 | * Called from workqueue/eventd every few seconds. | |
4152 | * Purpose: | |
4153 | * - clear the per-cpu caches for this CPU. | |
4154 | * - return freeable pages to the main free memory pool. | |
4155 | * | |
a737b3e2 AM |
4156 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4157 | * again on the next iteration. | |
1da177e4 | 4158 | */ |
7c5cae36 | 4159 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4160 | { |
7a7c381d | 4161 | struct kmem_cache *searchp; |
e498be7d | 4162 | struct kmem_list3 *l3; |
7d6e6d09 | 4163 | int node = numa_mem_id(); |
bf6aede7 | 4164 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4165 | |
7c5cae36 | 4166 | if (!mutex_trylock(&cache_chain_mutex)) |
1da177e4 | 4167 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4168 | goto out; |
1da177e4 | 4169 | |
3b0efdfa | 4170 | list_for_each_entry(searchp, &cache_chain, list) { |
1da177e4 LT |
4171 | check_irq_on(); |
4172 | ||
35386e3b CL |
4173 | /* |
4174 | * We only take the l3 lock if absolutely necessary and we | |
4175 | * have established with reasonable certainty that | |
4176 | * we can do some work if the lock was obtained. | |
4177 | */ | |
aab2207c | 4178 | l3 = searchp->nodelists[node]; |
35386e3b | 4179 | |
8fce4d8e | 4180 | reap_alien(searchp, l3); |
1da177e4 | 4181 | |
aab2207c | 4182 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4183 | |
35386e3b CL |
4184 | /* |
4185 | * These are racy checks but it does not matter | |
4186 | * if we skip one check or scan twice. | |
4187 | */ | |
e498be7d | 4188 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4189 | goto next; |
1da177e4 | 4190 | |
e498be7d | 4191 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4192 | |
aab2207c | 4193 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4194 | |
ed11d9eb | 4195 | if (l3->free_touched) |
e498be7d | 4196 | l3->free_touched = 0; |
ed11d9eb CL |
4197 | else { |
4198 | int freed; | |
1da177e4 | 4199 | |
ed11d9eb CL |
4200 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4201 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4202 | STATS_ADD_REAPED(searchp, freed); | |
4203 | } | |
35386e3b | 4204 | next: |
1da177e4 LT |
4205 | cond_resched(); |
4206 | } | |
4207 | check_irq_on(); | |
fc0abb14 | 4208 | mutex_unlock(&cache_chain_mutex); |
8fce4d8e | 4209 | next_reap_node(); |
7c5cae36 | 4210 | out: |
a737b3e2 | 4211 | /* Set up the next iteration */ |
7c5cae36 | 4212 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4213 | } |
4214 | ||
158a9624 | 4215 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4216 | |
85289f98 | 4217 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4218 | { |
85289f98 PE |
4219 | /* |
4220 | * Output format version, so at least we can change it | |
4221 | * without _too_ many complaints. | |
4222 | */ | |
1da177e4 | 4223 | #if STATS |
85289f98 | 4224 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4225 | #else |
85289f98 | 4226 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4227 | #endif |
85289f98 PE |
4228 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4229 | "<objperslab> <pagesperslab>"); | |
4230 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4231 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4232 | #if STATS |
85289f98 | 4233 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4234 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4235 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4236 | #endif |
85289f98 PE |
4237 | seq_putc(m, '\n'); |
4238 | } | |
4239 | ||
4240 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4241 | { | |
4242 | loff_t n = *pos; | |
85289f98 | 4243 | |
fc0abb14 | 4244 | mutex_lock(&cache_chain_mutex); |
85289f98 PE |
4245 | if (!n) |
4246 | print_slabinfo_header(m); | |
b92151ba PE |
4247 | |
4248 | return seq_list_start(&cache_chain, *pos); | |
1da177e4 LT |
4249 | } |
4250 | ||
4251 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4252 | { | |
b92151ba | 4253 | return seq_list_next(p, &cache_chain, pos); |
1da177e4 LT |
4254 | } |
4255 | ||
4256 | static void s_stop(struct seq_file *m, void *p) | |
4257 | { | |
fc0abb14 | 4258 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4259 | } |
4260 | ||
4261 | static int s_show(struct seq_file *m, void *p) | |
4262 | { | |
3b0efdfa | 4263 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
b28a02de PE |
4264 | struct slab *slabp; |
4265 | unsigned long active_objs; | |
4266 | unsigned long num_objs; | |
4267 | unsigned long active_slabs = 0; | |
4268 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4269 | const char *name; |
1da177e4 | 4270 | char *error = NULL; |
e498be7d CL |
4271 | int node; |
4272 | struct kmem_list3 *l3; | |
1da177e4 | 4273 | |
1da177e4 LT |
4274 | active_objs = 0; |
4275 | num_slabs = 0; | |
e498be7d CL |
4276 | for_each_online_node(node) { |
4277 | l3 = cachep->nodelists[node]; | |
4278 | if (!l3) | |
4279 | continue; | |
4280 | ||
ca3b9b91 RT |
4281 | check_irq_on(); |
4282 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4283 | |
7a7c381d | 4284 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4285 | if (slabp->inuse != cachep->num && !error) |
4286 | error = "slabs_full accounting error"; | |
4287 | active_objs += cachep->num; | |
4288 | active_slabs++; | |
4289 | } | |
7a7c381d | 4290 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4291 | if (slabp->inuse == cachep->num && !error) |
4292 | error = "slabs_partial inuse accounting error"; | |
4293 | if (!slabp->inuse && !error) | |
4294 | error = "slabs_partial/inuse accounting error"; | |
4295 | active_objs += slabp->inuse; | |
4296 | active_slabs++; | |
4297 | } | |
7a7c381d | 4298 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4299 | if (slabp->inuse && !error) |
4300 | error = "slabs_free/inuse accounting error"; | |
4301 | num_slabs++; | |
4302 | } | |
4303 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4304 | if (l3->shared) |
4305 | shared_avail += l3->shared->avail; | |
e498be7d | 4306 | |
ca3b9b91 | 4307 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4308 | } |
b28a02de PE |
4309 | num_slabs += active_slabs; |
4310 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4311 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4312 | error = "free_objects accounting error"; |
4313 | ||
b28a02de | 4314 | name = cachep->name; |
1da177e4 LT |
4315 | if (error) |
4316 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4317 | ||
4318 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3b0efdfa | 4319 | name, active_objs, num_objs, cachep->size, |
b28a02de | 4320 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4321 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4322 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4323 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4324 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4325 | #if STATS |
b28a02de | 4326 | { /* list3 stats */ |
1da177e4 LT |
4327 | unsigned long high = cachep->high_mark; |
4328 | unsigned long allocs = cachep->num_allocations; | |
4329 | unsigned long grown = cachep->grown; | |
4330 | unsigned long reaped = cachep->reaped; | |
4331 | unsigned long errors = cachep->errors; | |
4332 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4333 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4334 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4335 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4336 | |
e92dd4fd JP |
4337 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4338 | "%4lu %4lu %4lu %4lu %4lu", | |
4339 | allocs, high, grown, | |
4340 | reaped, errors, max_freeable, node_allocs, | |
4341 | node_frees, overflows); | |
1da177e4 LT |
4342 | } |
4343 | /* cpu stats */ | |
4344 | { | |
4345 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4346 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4347 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4348 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4349 | ||
4350 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4351 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4352 | } |
4353 | #endif | |
4354 | seq_putc(m, '\n'); | |
1da177e4 LT |
4355 | return 0; |
4356 | } | |
4357 | ||
4358 | /* | |
4359 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4360 | * | |
4361 | * Output layout: | |
4362 | * cache-name | |
4363 | * num-active-objs | |
4364 | * total-objs | |
4365 | * object size | |
4366 | * num-active-slabs | |
4367 | * total-slabs | |
4368 | * num-pages-per-slab | |
4369 | * + further values on SMP and with statistics enabled | |
4370 | */ | |
4371 | ||
7b3c3a50 | 4372 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4373 | .start = s_start, |
4374 | .next = s_next, | |
4375 | .stop = s_stop, | |
4376 | .show = s_show, | |
1da177e4 LT |
4377 | }; |
4378 | ||
4379 | #define MAX_SLABINFO_WRITE 128 | |
4380 | /** | |
4381 | * slabinfo_write - Tuning for the slab allocator | |
4382 | * @file: unused | |
4383 | * @buffer: user buffer | |
4384 | * @count: data length | |
4385 | * @ppos: unused | |
4386 | */ | |
68a1b195 | 4387 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4388 | size_t count, loff_t *ppos) |
1da177e4 | 4389 | { |
b28a02de | 4390 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4391 | int limit, batchcount, shared, res; |
7a7c381d | 4392 | struct kmem_cache *cachep; |
b28a02de | 4393 | |
1da177e4 LT |
4394 | if (count > MAX_SLABINFO_WRITE) |
4395 | return -EINVAL; | |
4396 | if (copy_from_user(&kbuf, buffer, count)) | |
4397 | return -EFAULT; | |
b28a02de | 4398 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4399 | |
4400 | tmp = strchr(kbuf, ' '); | |
4401 | if (!tmp) | |
4402 | return -EINVAL; | |
4403 | *tmp = '\0'; | |
4404 | tmp++; | |
4405 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4406 | return -EINVAL; | |
4407 | ||
4408 | /* Find the cache in the chain of caches. */ | |
fc0abb14 | 4409 | mutex_lock(&cache_chain_mutex); |
1da177e4 | 4410 | res = -EINVAL; |
3b0efdfa | 4411 | list_for_each_entry(cachep, &cache_chain, list) { |
1da177e4 | 4412 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4413 | if (limit < 1 || batchcount < 1 || |
4414 | batchcount > limit || shared < 0) { | |
e498be7d | 4415 | res = 0; |
1da177e4 | 4416 | } else { |
e498be7d | 4417 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4418 | batchcount, shared, |
4419 | GFP_KERNEL); | |
1da177e4 LT |
4420 | } |
4421 | break; | |
4422 | } | |
4423 | } | |
fc0abb14 | 4424 | mutex_unlock(&cache_chain_mutex); |
1da177e4 LT |
4425 | if (res >= 0) |
4426 | res = count; | |
4427 | return res; | |
4428 | } | |
871751e2 | 4429 | |
7b3c3a50 AD |
4430 | static int slabinfo_open(struct inode *inode, struct file *file) |
4431 | { | |
4432 | return seq_open(file, &slabinfo_op); | |
4433 | } | |
4434 | ||
4435 | static const struct file_operations proc_slabinfo_operations = { | |
4436 | .open = slabinfo_open, | |
4437 | .read = seq_read, | |
4438 | .write = slabinfo_write, | |
4439 | .llseek = seq_lseek, | |
4440 | .release = seq_release, | |
4441 | }; | |
4442 | ||
871751e2 AV |
4443 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4444 | ||
4445 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4446 | { | |
871751e2 | 4447 | mutex_lock(&cache_chain_mutex); |
b92151ba | 4448 | return seq_list_start(&cache_chain, *pos); |
871751e2 AV |
4449 | } |
4450 | ||
4451 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4452 | { | |
4453 | unsigned long *p; | |
4454 | int l; | |
4455 | if (!v) | |
4456 | return 1; | |
4457 | l = n[1]; | |
4458 | p = n + 2; | |
4459 | while (l) { | |
4460 | int i = l/2; | |
4461 | unsigned long *q = p + 2 * i; | |
4462 | if (*q == v) { | |
4463 | q[1]++; | |
4464 | return 1; | |
4465 | } | |
4466 | if (*q > v) { | |
4467 | l = i; | |
4468 | } else { | |
4469 | p = q + 2; | |
4470 | l -= i + 1; | |
4471 | } | |
4472 | } | |
4473 | if (++n[1] == n[0]) | |
4474 | return 0; | |
4475 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4476 | p[0] = v; | |
4477 | p[1] = 1; | |
4478 | return 1; | |
4479 | } | |
4480 | ||
4481 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4482 | { | |
4483 | void *p; | |
4484 | int i; | |
4485 | if (n[0] == n[1]) | |
4486 | return; | |
3b0efdfa | 4487 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4488 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4489 | continue; | |
4490 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4491 | return; | |
4492 | } | |
4493 | } | |
4494 | ||
4495 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4496 | { | |
4497 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4498 | unsigned long offset, size; |
9281acea | 4499 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4500 | |
a5c43dae | 4501 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4502 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4503 | if (modname[0]) |
871751e2 AV |
4504 | seq_printf(m, " [%s]", modname); |
4505 | return; | |
4506 | } | |
4507 | #endif | |
4508 | seq_printf(m, "%p", (void *)address); | |
4509 | } | |
4510 | ||
4511 | static int leaks_show(struct seq_file *m, void *p) | |
4512 | { | |
0672aa7c | 4513 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 AV |
4514 | struct slab *slabp; |
4515 | struct kmem_list3 *l3; | |
4516 | const char *name; | |
4517 | unsigned long *n = m->private; | |
4518 | int node; | |
4519 | int i; | |
4520 | ||
4521 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4522 | return 0; | |
4523 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4524 | return 0; | |
4525 | ||
4526 | /* OK, we can do it */ | |
4527 | ||
4528 | n[1] = 0; | |
4529 | ||
4530 | for_each_online_node(node) { | |
4531 | l3 = cachep->nodelists[node]; | |
4532 | if (!l3) | |
4533 | continue; | |
4534 | ||
4535 | check_irq_on(); | |
4536 | spin_lock_irq(&l3->list_lock); | |
4537 | ||
7a7c381d | 4538 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4539 | handle_slab(n, cachep, slabp); |
7a7c381d | 4540 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4541 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4542 | spin_unlock_irq(&l3->list_lock); |
4543 | } | |
4544 | name = cachep->name; | |
4545 | if (n[0] == n[1]) { | |
4546 | /* Increase the buffer size */ | |
4547 | mutex_unlock(&cache_chain_mutex); | |
4548 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4549 | if (!m->private) { | |
4550 | /* Too bad, we are really out */ | |
4551 | m->private = n; | |
4552 | mutex_lock(&cache_chain_mutex); | |
4553 | return -ENOMEM; | |
4554 | } | |
4555 | *(unsigned long *)m->private = n[0] * 2; | |
4556 | kfree(n); | |
4557 | mutex_lock(&cache_chain_mutex); | |
4558 | /* Now make sure this entry will be retried */ | |
4559 | m->count = m->size; | |
4560 | return 0; | |
4561 | } | |
4562 | for (i = 0; i < n[1]; i++) { | |
4563 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4564 | show_symbol(m, n[2*i+2]); | |
4565 | seq_putc(m, '\n'); | |
4566 | } | |
d2e7b7d0 | 4567 | |
871751e2 AV |
4568 | return 0; |
4569 | } | |
4570 | ||
a0ec95a8 | 4571 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4572 | .start = leaks_start, |
4573 | .next = s_next, | |
4574 | .stop = s_stop, | |
4575 | .show = leaks_show, | |
4576 | }; | |
a0ec95a8 AD |
4577 | |
4578 | static int slabstats_open(struct inode *inode, struct file *file) | |
4579 | { | |
4580 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4581 | int ret = -ENOMEM; | |
4582 | if (n) { | |
4583 | ret = seq_open(file, &slabstats_op); | |
4584 | if (!ret) { | |
4585 | struct seq_file *m = file->private_data; | |
4586 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4587 | m->private = n; | |
4588 | n = NULL; | |
4589 | } | |
4590 | kfree(n); | |
4591 | } | |
4592 | return ret; | |
4593 | } | |
4594 | ||
4595 | static const struct file_operations proc_slabstats_operations = { | |
4596 | .open = slabstats_open, | |
4597 | .read = seq_read, | |
4598 | .llseek = seq_lseek, | |
4599 | .release = seq_release_private, | |
4600 | }; | |
4601 | #endif | |
4602 | ||
4603 | static int __init slab_proc_init(void) | |
4604 | { | |
ab067e99 | 4605 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4606 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4607 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4608 | #endif |
a0ec95a8 AD |
4609 | return 0; |
4610 | } | |
4611 | module_init(slab_proc_init); | |
1da177e4 LT |
4612 | #endif |
4613 | ||
00e145b6 MS |
4614 | /** |
4615 | * ksize - get the actual amount of memory allocated for a given object | |
4616 | * @objp: Pointer to the object | |
4617 | * | |
4618 | * kmalloc may internally round up allocations and return more memory | |
4619 | * than requested. ksize() can be used to determine the actual amount of | |
4620 | * memory allocated. The caller may use this additional memory, even though | |
4621 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4622 | * The caller must guarantee that objp points to a valid object previously | |
4623 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4624 | * must not be freed during the duration of the call. | |
4625 | */ | |
fd76bab2 | 4626 | size_t ksize(const void *objp) |
1da177e4 | 4627 | { |
ef8b4520 CL |
4628 | BUG_ON(!objp); |
4629 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4630 | return 0; |
1da177e4 | 4631 | |
8c138bc0 | 4632 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4633 | } |
b1aabecd | 4634 | EXPORT_SYMBOL(ksize); |