]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
183ff22b | 29 | * slabs and you must pass objects with the same initializations to |
1da177e4 LT |
30 | * kmem_cache_free. |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
54 | * it's changed with a smp_call_function(). |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
18004c5d | 71 | * The global cache-chain is protected by the mutex 'slab_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
1da177e4 | 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 | |
381760ea MG |
121 | #include <net/sock.h> |
122 | ||
1da177e4 LT |
123 | #include <asm/cacheflush.h> |
124 | #include <asm/tlbflush.h> | |
125 | #include <asm/page.h> | |
126 | ||
4dee6b64 SR |
127 | #include <trace/events/kmem.h> |
128 | ||
072bb0aa MG |
129 | #include "internal.h" |
130 | ||
1da177e4 | 131 | /* |
50953fe9 | 132 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
133 | * 0 for faster, smaller code (especially in the critical paths). |
134 | * | |
135 | * STATS - 1 to collect stats for /proc/slabinfo. | |
136 | * 0 for faster, smaller code (especially in the critical paths). | |
137 | * | |
138 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
139 | */ | |
140 | ||
141 | #ifdef CONFIG_DEBUG_SLAB | |
142 | #define DEBUG 1 | |
143 | #define STATS 1 | |
144 | #define FORCED_DEBUG 1 | |
145 | #else | |
146 | #define DEBUG 0 | |
147 | #define STATS 0 | |
148 | #define FORCED_DEBUG 0 | |
149 | #endif | |
150 | ||
1da177e4 LT |
151 | /* Shouldn't this be in a header file somewhere? */ |
152 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 153 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 154 | |
1da177e4 LT |
155 | #ifndef ARCH_KMALLOC_FLAGS |
156 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
157 | #endif | |
158 | ||
072bb0aa MG |
159 | /* |
160 | * true if a page was allocated from pfmemalloc reserves for network-based | |
161 | * swap | |
162 | */ | |
163 | static bool pfmemalloc_active __read_mostly; | |
164 | ||
1da177e4 LT |
165 | /* |
166 | * kmem_bufctl_t: | |
167 | * | |
168 | * Bufctl's are used for linking objs within a slab | |
169 | * linked offsets. | |
170 | * | |
171 | * This implementation relies on "struct page" for locating the cache & | |
172 | * slab an object belongs to. | |
173 | * This allows the bufctl structure to be small (one int), but limits | |
174 | * the number of objects a slab (not a cache) can contain when off-slab | |
175 | * bufctls are used. The limit is the size of the largest general cache | |
176 | * that does not use off-slab slabs. | |
177 | * For 32bit archs with 4 kB pages, is this 56. | |
178 | * This is not serious, as it is only for large objects, when it is unwise | |
179 | * to have too many per slab. | |
180 | * Note: This limit can be raised by introducing a general cache whose size | |
181 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
182 | */ | |
183 | ||
fa5b08d5 | 184 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
185 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
186 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
187 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
188 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 189 | |
1da177e4 LT |
190 | /* |
191 | * struct slab_rcu | |
192 | * | |
193 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
194 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
195 | * we need to approach a kernel structure obliquely, from its address | |
196 | * obtained without the usual locking. We can lock the structure to | |
197 | * stabilize it and check it's still at the given address, only if we | |
198 | * can be sure that the memory has not been meanwhile reused for some | |
199 | * other kind of object (which our subsystem's lock might corrupt). | |
200 | * | |
201 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
202 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
203 | */ |
204 | struct slab_rcu { | |
b28a02de | 205 | struct rcu_head head; |
343e0d7a | 206 | struct kmem_cache *cachep; |
b28a02de | 207 | void *addr; |
1da177e4 LT |
208 | }; |
209 | ||
5bfe53a7 LJ |
210 | /* |
211 | * struct slab | |
212 | * | |
213 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
214 | * for a slab, or allocated from an general cache. | |
215 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
216 | */ | |
217 | struct slab { | |
218 | union { | |
219 | struct { | |
220 | struct list_head list; | |
221 | unsigned long colouroff; | |
222 | void *s_mem; /* including colour offset */ | |
223 | unsigned int inuse; /* num of objs active in slab */ | |
224 | kmem_bufctl_t free; | |
225 | unsigned short nodeid; | |
226 | }; | |
227 | struct slab_rcu __slab_cover_slab_rcu; | |
228 | }; | |
229 | }; | |
230 | ||
1da177e4 LT |
231 | /* |
232 | * struct array_cache | |
233 | * | |
1da177e4 LT |
234 | * Purpose: |
235 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
236 | * - reduce the number of linked list operations | |
237 | * - reduce spinlock operations | |
238 | * | |
239 | * The limit is stored in the per-cpu structure to reduce the data cache | |
240 | * footprint. | |
241 | * | |
242 | */ | |
243 | struct array_cache { | |
244 | unsigned int avail; | |
245 | unsigned int limit; | |
246 | unsigned int batchcount; | |
247 | unsigned int touched; | |
e498be7d | 248 | spinlock_t lock; |
bda5b655 | 249 | void *entry[]; /* |
a737b3e2 AM |
250 | * Must have this definition in here for the proper |
251 | * alignment of array_cache. Also simplifies accessing | |
252 | * the entries. | |
072bb0aa MG |
253 | * |
254 | * Entries should not be directly dereferenced as | |
255 | * entries belonging to slabs marked pfmemalloc will | |
256 | * have the lower bits set SLAB_OBJ_PFMEMALLOC | |
a737b3e2 | 257 | */ |
1da177e4 LT |
258 | }; |
259 | ||
072bb0aa MG |
260 | #define SLAB_OBJ_PFMEMALLOC 1 |
261 | static inline bool is_obj_pfmemalloc(void *objp) | |
262 | { | |
263 | return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC; | |
264 | } | |
265 | ||
266 | static inline void set_obj_pfmemalloc(void **objp) | |
267 | { | |
268 | *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC); | |
269 | return; | |
270 | } | |
271 | ||
272 | static inline void clear_obj_pfmemalloc(void **objp) | |
273 | { | |
274 | *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC); | |
275 | } | |
276 | ||
a737b3e2 AM |
277 | /* |
278 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
279 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
280 | */ |
281 | #define BOOT_CPUCACHE_ENTRIES 1 | |
282 | struct arraycache_init { | |
283 | struct array_cache cache; | |
b28a02de | 284 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
285 | }; |
286 | ||
287 | /* | |
e498be7d | 288 | * The slab lists for all objects. |
1da177e4 LT |
289 | */ |
290 | struct kmem_list3 { | |
b28a02de PE |
291 | struct list_head slabs_partial; /* partial list first, better asm code */ |
292 | struct list_head slabs_full; | |
293 | struct list_head slabs_free; | |
294 | unsigned long free_objects; | |
b28a02de | 295 | unsigned int free_limit; |
2e1217cf | 296 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
297 | spinlock_t list_lock; |
298 | struct array_cache *shared; /* shared per node */ | |
299 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
300 | unsigned long next_reap; /* updated without locking */ |
301 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
302 | }; |
303 | ||
e498be7d CL |
304 | /* |
305 | * Need this for bootstrapping a per node allocator. | |
306 | */ | |
556a169d | 307 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
68a1b195 | 308 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 309 | #define CACHE_CACHE 0 |
556a169d PE |
310 | #define SIZE_AC MAX_NUMNODES |
311 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 312 | |
ed11d9eb CL |
313 | static int drain_freelist(struct kmem_cache *cache, |
314 | struct kmem_list3 *l3, int tofree); | |
315 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
316 | int node); | |
83b519e8 | 317 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 318 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 319 | |
e498be7d | 320 | /* |
a737b3e2 AM |
321 | * This function must be completely optimized away if a constant is passed to |
322 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 323 | */ |
7243cc05 | 324 | static __always_inline int index_of(const size_t size) |
e498be7d | 325 | { |
5ec8a847 SR |
326 | extern void __bad_size(void); |
327 | ||
e498be7d CL |
328 | if (__builtin_constant_p(size)) { |
329 | int i = 0; | |
330 | ||
331 | #define CACHE(x) \ | |
332 | if (size <=x) \ | |
333 | return i; \ | |
334 | else \ | |
335 | i++; | |
1c61fc40 | 336 | #include <linux/kmalloc_sizes.h> |
e498be7d | 337 | #undef CACHE |
5ec8a847 | 338 | __bad_size(); |
7243cc05 | 339 | } else |
5ec8a847 | 340 | __bad_size(); |
e498be7d CL |
341 | return 0; |
342 | } | |
343 | ||
e0a42726 IM |
344 | static int slab_early_init = 1; |
345 | ||
e498be7d CL |
346 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
347 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 348 | |
5295a74c | 349 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
350 | { |
351 | INIT_LIST_HEAD(&parent->slabs_full); | |
352 | INIT_LIST_HEAD(&parent->slabs_partial); | |
353 | INIT_LIST_HEAD(&parent->slabs_free); | |
354 | parent->shared = NULL; | |
355 | parent->alien = NULL; | |
2e1217cf | 356 | parent->colour_next = 0; |
e498be7d CL |
357 | spin_lock_init(&parent->list_lock); |
358 | parent->free_objects = 0; | |
359 | parent->free_touched = 0; | |
360 | } | |
361 | ||
a737b3e2 AM |
362 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
363 | do { \ | |
364 | INIT_LIST_HEAD(listp); \ | |
365 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
366 | } while (0) |
367 | ||
a737b3e2 AM |
368 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
369 | do { \ | |
e498be7d CL |
370 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
371 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
372 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
373 | } while (0) | |
1da177e4 | 374 | |
1da177e4 LT |
375 | #define CFLGS_OFF_SLAB (0x80000000UL) |
376 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
377 | ||
378 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
379 | /* |
380 | * Optimization question: fewer reaps means less probability for unnessary | |
381 | * cpucache drain/refill cycles. | |
1da177e4 | 382 | * |
dc6f3f27 | 383 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
384 | * which could lock up otherwise freeable slabs. |
385 | */ | |
386 | #define REAPTIMEOUT_CPUC (2*HZ) | |
387 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
388 | ||
389 | #if STATS | |
390 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
391 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
392 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
393 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 394 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
395 | #define STATS_SET_HIGH(x) \ |
396 | do { \ | |
397 | if ((x)->num_active > (x)->high_mark) \ | |
398 | (x)->high_mark = (x)->num_active; \ | |
399 | } while (0) | |
1da177e4 LT |
400 | #define STATS_INC_ERR(x) ((x)->errors++) |
401 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 402 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 403 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
404 | #define STATS_SET_FREEABLE(x, i) \ |
405 | do { \ | |
406 | if ((x)->max_freeable < i) \ | |
407 | (x)->max_freeable = i; \ | |
408 | } while (0) | |
1da177e4 LT |
409 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
410 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
411 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
412 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
413 | #else | |
414 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
415 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
416 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
417 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 418 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
419 | #define STATS_SET_HIGH(x) do { } while (0) |
420 | #define STATS_INC_ERR(x) do { } while (0) | |
421 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 422 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 423 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 424 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
425 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
426 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
427 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
428 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
429 | #endif | |
430 | ||
431 | #if DEBUG | |
1da177e4 | 432 | |
a737b3e2 AM |
433 | /* |
434 | * memory layout of objects: | |
1da177e4 | 435 | * 0 : objp |
3dafccf2 | 436 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
437 | * the end of an object is aligned with the end of the real |
438 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 439 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 440 | * redzone word. |
3dafccf2 | 441 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
442 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
443 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 444 | * [BYTES_PER_WORD long] |
1da177e4 | 445 | */ |
343e0d7a | 446 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 447 | { |
3dafccf2 | 448 | return cachep->obj_offset; |
1da177e4 LT |
449 | } |
450 | ||
b46b8f19 | 451 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
452 | { |
453 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
454 | return (unsigned long long*) (objp + obj_offset(cachep) - |
455 | sizeof(unsigned long long)); | |
1da177e4 LT |
456 | } |
457 | ||
b46b8f19 | 458 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
459 | { |
460 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
461 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 462 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 463 | sizeof(unsigned long long) - |
87a927c7 | 464 | REDZONE_ALIGN); |
3b0efdfa | 465 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 466 | sizeof(unsigned long long)); |
1da177e4 LT |
467 | } |
468 | ||
343e0d7a | 469 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
470 | { |
471 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 472 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
473 | } |
474 | ||
475 | #else | |
476 | ||
3dafccf2 | 477 | #define obj_offset(x) 0 |
b46b8f19 DW |
478 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
479 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
480 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
481 | ||
482 | #endif | |
483 | ||
1da177e4 | 484 | /* |
3df1cccd DR |
485 | * Do not go above this order unless 0 objects fit into the slab or |
486 | * overridden on the command line. | |
1da177e4 | 487 | */ |
543585cc DR |
488 | #define SLAB_MAX_ORDER_HI 1 |
489 | #define SLAB_MAX_ORDER_LO 0 | |
490 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 491 | static bool slab_max_order_set __initdata; |
1da177e4 | 492 | |
6ed5eb22 PE |
493 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
494 | { | |
b49af68f | 495 | struct page *page = virt_to_head_page(obj); |
35026088 | 496 | return page->slab_cache; |
6ed5eb22 PE |
497 | } |
498 | ||
499 | static inline struct slab *virt_to_slab(const void *obj) | |
500 | { | |
b49af68f | 501 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
502 | |
503 | VM_BUG_ON(!PageSlab(page)); | |
504 | return page->slab_page; | |
6ed5eb22 PE |
505 | } |
506 | ||
8fea4e96 PE |
507 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
508 | unsigned int idx) | |
509 | { | |
3b0efdfa | 510 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
511 | } |
512 | ||
6a2d7a95 | 513 | /* |
3b0efdfa CL |
514 | * We want to avoid an expensive divide : (offset / cache->size) |
515 | * Using the fact that size is a constant for a particular cache, | |
516 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
517 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
518 | */ | |
519 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
520 | const struct slab *slab, void *obj) | |
8fea4e96 | 521 | { |
6a2d7a95 ED |
522 | u32 offset = (obj - slab->s_mem); |
523 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
524 | } |
525 | ||
a737b3e2 AM |
526 | /* |
527 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
528 | */ | |
1da177e4 LT |
529 | struct cache_sizes malloc_sizes[] = { |
530 | #define CACHE(x) { .cs_size = (x) }, | |
531 | #include <linux/kmalloc_sizes.h> | |
532 | CACHE(ULONG_MAX) | |
533 | #undef CACHE | |
534 | }; | |
535 | EXPORT_SYMBOL(malloc_sizes); | |
536 | ||
537 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
538 | struct cache_names { | |
539 | char *name; | |
540 | char *name_dma; | |
541 | }; | |
542 | ||
543 | static struct cache_names __initdata cache_names[] = { | |
544 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
545 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 546 | {NULL,} |
1da177e4 LT |
547 | #undef CACHE |
548 | }; | |
549 | ||
550 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 551 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 552 | static struct arraycache_init initarray_generic = |
b28a02de | 553 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
554 | |
555 | /* internal cache of cache description objs */ | |
9b030cb8 CL |
556 | static struct kmem_list3 *kmem_cache_nodelists[MAX_NUMNODES]; |
557 | static struct kmem_cache kmem_cache_boot = { | |
558 | .nodelists = kmem_cache_nodelists, | |
b28a02de PE |
559 | .batchcount = 1, |
560 | .limit = BOOT_CPUCACHE_ENTRIES, | |
561 | .shared = 1, | |
3b0efdfa | 562 | .size = sizeof(struct kmem_cache), |
b28a02de | 563 | .name = "kmem_cache", |
1da177e4 LT |
564 | }; |
565 | ||
056c6241 RT |
566 | #define BAD_ALIEN_MAGIC 0x01020304ul |
567 | ||
f1aaee53 AV |
568 | #ifdef CONFIG_LOCKDEP |
569 | ||
570 | /* | |
571 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
572 | * for other slabs "off slab". | |
573 | * The locking for this is tricky in that it nests within the locks | |
574 | * of all other slabs in a few places; to deal with this special | |
575 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
576 | * |
577 | * We set lock class for alien array caches which are up during init. | |
578 | * The lock annotation will be lost if all cpus of a node goes down and | |
579 | * then comes back up during hotplug | |
f1aaee53 | 580 | */ |
056c6241 RT |
581 | static struct lock_class_key on_slab_l3_key; |
582 | static struct lock_class_key on_slab_alc_key; | |
583 | ||
83835b3d PZ |
584 | static struct lock_class_key debugobj_l3_key; |
585 | static struct lock_class_key debugobj_alc_key; | |
586 | ||
587 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
588 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
589 | int q) | |
590 | { | |
591 | struct array_cache **alc; | |
592 | struct kmem_list3 *l3; | |
593 | int r; | |
594 | ||
595 | l3 = cachep->nodelists[q]; | |
596 | if (!l3) | |
597 | return; | |
598 | ||
599 | lockdep_set_class(&l3->list_lock, l3_key); | |
600 | alc = l3->alien; | |
601 | /* | |
602 | * FIXME: This check for BAD_ALIEN_MAGIC | |
603 | * should go away when common slab code is taught to | |
604 | * work even without alien caches. | |
605 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
606 | * for alloc_alien_cache, | |
607 | */ | |
608 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
609 | return; | |
610 | for_each_node(r) { | |
611 | if (alc[r]) | |
612 | lockdep_set_class(&alc[r]->lock, alc_key); | |
613 | } | |
614 | } | |
615 | ||
616 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
617 | { | |
618 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
619 | } | |
620 | ||
621 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
622 | { | |
623 | int node; | |
624 | ||
625 | for_each_online_node(node) | |
626 | slab_set_debugobj_lock_classes_node(cachep, node); | |
627 | } | |
628 | ||
ce79ddc8 | 629 | static void init_node_lock_keys(int q) |
f1aaee53 | 630 | { |
056c6241 RT |
631 | struct cache_sizes *s = malloc_sizes; |
632 | ||
97d06609 | 633 | if (slab_state < UP) |
ce79ddc8 PE |
634 | return; |
635 | ||
636 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
ce79ddc8 | 637 | struct kmem_list3 *l3; |
ce79ddc8 PE |
638 | |
639 | l3 = s->cs_cachep->nodelists[q]; | |
640 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 641 | continue; |
83835b3d PZ |
642 | |
643 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
644 | &on_slab_alc_key, q); | |
f1aaee53 AV |
645 | } |
646 | } | |
ce79ddc8 PE |
647 | |
648 | static inline void init_lock_keys(void) | |
649 | { | |
650 | int node; | |
651 | ||
652 | for_each_node(node) | |
653 | init_node_lock_keys(node); | |
654 | } | |
f1aaee53 | 655 | #else |
ce79ddc8 PE |
656 | static void init_node_lock_keys(int q) |
657 | { | |
658 | } | |
659 | ||
056c6241 | 660 | static inline void init_lock_keys(void) |
f1aaee53 AV |
661 | { |
662 | } | |
83835b3d PZ |
663 | |
664 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
665 | { | |
666 | } | |
667 | ||
668 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
669 | { | |
670 | } | |
f1aaee53 AV |
671 | #endif |
672 | ||
1871e52c | 673 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 674 | |
343e0d7a | 675 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
676 | { |
677 | return cachep->array[smp_processor_id()]; | |
678 | } | |
679 | ||
a737b3e2 AM |
680 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
681 | gfp_t gfpflags) | |
1da177e4 LT |
682 | { |
683 | struct cache_sizes *csizep = malloc_sizes; | |
684 | ||
685 | #if DEBUG | |
686 | /* This happens if someone tries to call | |
b28a02de PE |
687 | * kmem_cache_create(), or __kmalloc(), before |
688 | * the generic caches are initialized. | |
689 | */ | |
c7e43c78 | 690 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 691 | #endif |
6cb8f913 CL |
692 | if (!size) |
693 | return ZERO_SIZE_PTR; | |
694 | ||
1da177e4 LT |
695 | while (size > csizep->cs_size) |
696 | csizep++; | |
697 | ||
698 | /* | |
0abf40c1 | 699 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
700 | * has cs_{dma,}cachep==NULL. Thus no special case |
701 | * for large kmalloc calls required. | |
702 | */ | |
4b51d669 | 703 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
704 | if (unlikely(gfpflags & GFP_DMA)) |
705 | return csizep->cs_dmacachep; | |
4b51d669 | 706 | #endif |
1da177e4 LT |
707 | return csizep->cs_cachep; |
708 | } | |
709 | ||
b221385b | 710 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
711 | { |
712 | return __find_general_cachep(size, gfpflags); | |
713 | } | |
97e2bde4 | 714 | |
fbaccacf | 715 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 716 | { |
fbaccacf SR |
717 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
718 | } | |
1da177e4 | 719 | |
a737b3e2 AM |
720 | /* |
721 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
722 | */ | |
fbaccacf SR |
723 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
724 | size_t align, int flags, size_t *left_over, | |
725 | unsigned int *num) | |
726 | { | |
727 | int nr_objs; | |
728 | size_t mgmt_size; | |
729 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 730 | |
fbaccacf SR |
731 | /* |
732 | * The slab management structure can be either off the slab or | |
733 | * on it. For the latter case, the memory allocated for a | |
734 | * slab is used for: | |
735 | * | |
736 | * - The struct slab | |
737 | * - One kmem_bufctl_t for each object | |
738 | * - Padding to respect alignment of @align | |
739 | * - @buffer_size bytes for each object | |
740 | * | |
741 | * If the slab management structure is off the slab, then the | |
742 | * alignment will already be calculated into the size. Because | |
743 | * the slabs are all pages aligned, the objects will be at the | |
744 | * correct alignment when allocated. | |
745 | */ | |
746 | if (flags & CFLGS_OFF_SLAB) { | |
747 | mgmt_size = 0; | |
748 | nr_objs = slab_size / buffer_size; | |
749 | ||
750 | if (nr_objs > SLAB_LIMIT) | |
751 | nr_objs = SLAB_LIMIT; | |
752 | } else { | |
753 | /* | |
754 | * Ignore padding for the initial guess. The padding | |
755 | * is at most @align-1 bytes, and @buffer_size is at | |
756 | * least @align. In the worst case, this result will | |
757 | * be one greater than the number of objects that fit | |
758 | * into the memory allocation when taking the padding | |
759 | * into account. | |
760 | */ | |
761 | nr_objs = (slab_size - sizeof(struct slab)) / | |
762 | (buffer_size + sizeof(kmem_bufctl_t)); | |
763 | ||
764 | /* | |
765 | * This calculated number will be either the right | |
766 | * amount, or one greater than what we want. | |
767 | */ | |
768 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
769 | > slab_size) | |
770 | nr_objs--; | |
771 | ||
772 | if (nr_objs > SLAB_LIMIT) | |
773 | nr_objs = SLAB_LIMIT; | |
774 | ||
775 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
776 | } | |
777 | *num = nr_objs; | |
778 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
779 | } |
780 | ||
f28510d3 | 781 | #if DEBUG |
d40cee24 | 782 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 783 | |
a737b3e2 AM |
784 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
785 | char *msg) | |
1da177e4 LT |
786 | { |
787 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 788 | function, cachep->name, msg); |
1da177e4 | 789 | dump_stack(); |
645df230 | 790 | add_taint(TAINT_BAD_PAGE); |
1da177e4 | 791 | } |
f28510d3 | 792 | #endif |
1da177e4 | 793 | |
3395ee05 PM |
794 | /* |
795 | * By default on NUMA we use alien caches to stage the freeing of | |
796 | * objects allocated from other nodes. This causes massive memory | |
797 | * inefficiencies when using fake NUMA setup to split memory into a | |
798 | * large number of small nodes, so it can be disabled on the command | |
799 | * line | |
800 | */ | |
801 | ||
802 | static int use_alien_caches __read_mostly = 1; | |
803 | static int __init noaliencache_setup(char *s) | |
804 | { | |
805 | use_alien_caches = 0; | |
806 | return 1; | |
807 | } | |
808 | __setup("noaliencache", noaliencache_setup); | |
809 | ||
3df1cccd DR |
810 | static int __init slab_max_order_setup(char *str) |
811 | { | |
812 | get_option(&str, &slab_max_order); | |
813 | slab_max_order = slab_max_order < 0 ? 0 : | |
814 | min(slab_max_order, MAX_ORDER - 1); | |
815 | slab_max_order_set = true; | |
816 | ||
817 | return 1; | |
818 | } | |
819 | __setup("slab_max_order=", slab_max_order_setup); | |
820 | ||
8fce4d8e CL |
821 | #ifdef CONFIG_NUMA |
822 | /* | |
823 | * Special reaping functions for NUMA systems called from cache_reap(). | |
824 | * These take care of doing round robin flushing of alien caches (containing | |
825 | * objects freed on different nodes from which they were allocated) and the | |
826 | * flushing of remote pcps by calling drain_node_pages. | |
827 | */ | |
1871e52c | 828 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
829 | |
830 | static void init_reap_node(int cpu) | |
831 | { | |
832 | int node; | |
833 | ||
7d6e6d09 | 834 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 835 | if (node == MAX_NUMNODES) |
442295c9 | 836 | node = first_node(node_online_map); |
8fce4d8e | 837 | |
1871e52c | 838 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
839 | } |
840 | ||
841 | static void next_reap_node(void) | |
842 | { | |
909ea964 | 843 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 844 | |
8fce4d8e CL |
845 | node = next_node(node, node_online_map); |
846 | if (unlikely(node >= MAX_NUMNODES)) | |
847 | node = first_node(node_online_map); | |
909ea964 | 848 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
849 | } |
850 | ||
851 | #else | |
852 | #define init_reap_node(cpu) do { } while (0) | |
853 | #define next_reap_node(void) do { } while (0) | |
854 | #endif | |
855 | ||
1da177e4 LT |
856 | /* |
857 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
858 | * via the workqueue/eventd. | |
859 | * Add the CPU number into the expiration time to minimize the possibility of | |
860 | * the CPUs getting into lockstep and contending for the global cache chain | |
861 | * lock. | |
862 | */ | |
897e679b | 863 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 864 | { |
1871e52c | 865 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
866 | |
867 | /* | |
868 | * When this gets called from do_initcalls via cpucache_init(), | |
869 | * init_workqueues() has already run, so keventd will be setup | |
870 | * at that time. | |
871 | */ | |
52bad64d | 872 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 873 | init_reap_node(cpu); |
203b42f7 | 874 | INIT_DEFERRABLE_WORK(reap_work, cache_reap); |
2b284214 AV |
875 | schedule_delayed_work_on(cpu, reap_work, |
876 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
877 | } |
878 | } | |
879 | ||
e498be7d | 880 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 881 | int batchcount, gfp_t gfp) |
1da177e4 | 882 | { |
b28a02de | 883 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
884 | struct array_cache *nc = NULL; |
885 | ||
83b519e8 | 886 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
887 | /* |
888 | * The array_cache structures contain pointers to free object. | |
25985edc | 889 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
890 | * cache the pointers are not cleared and they could be counted as |
891 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
892 | * not scan such objects. | |
893 | */ | |
894 | kmemleak_no_scan(nc); | |
1da177e4 LT |
895 | if (nc) { |
896 | nc->avail = 0; | |
897 | nc->limit = entries; | |
898 | nc->batchcount = batchcount; | |
899 | nc->touched = 0; | |
e498be7d | 900 | spin_lock_init(&nc->lock); |
1da177e4 LT |
901 | } |
902 | return nc; | |
903 | } | |
904 | ||
072bb0aa MG |
905 | static inline bool is_slab_pfmemalloc(struct slab *slabp) |
906 | { | |
907 | struct page *page = virt_to_page(slabp->s_mem); | |
908 | ||
909 | return PageSlabPfmemalloc(page); | |
910 | } | |
911 | ||
912 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
913 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
914 | struct array_cache *ac) | |
915 | { | |
916 | struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()]; | |
917 | struct slab *slabp; | |
918 | unsigned long flags; | |
919 | ||
920 | if (!pfmemalloc_active) | |
921 | return; | |
922 | ||
923 | spin_lock_irqsave(&l3->list_lock, flags); | |
924 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
925 | if (is_slab_pfmemalloc(slabp)) | |
926 | goto out; | |
927 | ||
928 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
929 | if (is_slab_pfmemalloc(slabp)) | |
930 | goto out; | |
931 | ||
932 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
933 | if (is_slab_pfmemalloc(slabp)) | |
934 | goto out; | |
935 | ||
936 | pfmemalloc_active = false; | |
937 | out: | |
938 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
939 | } | |
940 | ||
381760ea | 941 | static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, |
072bb0aa MG |
942 | gfp_t flags, bool force_refill) |
943 | { | |
944 | int i; | |
945 | void *objp = ac->entry[--ac->avail]; | |
946 | ||
947 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
948 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
949 | struct kmem_list3 *l3; | |
950 | ||
951 | if (gfp_pfmemalloc_allowed(flags)) { | |
952 | clear_obj_pfmemalloc(&objp); | |
953 | return objp; | |
954 | } | |
955 | ||
956 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
d014dc2e | 957 | for (i = 0; i < ac->avail; i++) { |
072bb0aa MG |
958 | /* If a !PFMEMALLOC object is found, swap them */ |
959 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
960 | objp = ac->entry[i]; | |
961 | ac->entry[i] = ac->entry[ac->avail]; | |
962 | ac->entry[ac->avail] = objp; | |
963 | return objp; | |
964 | } | |
965 | } | |
966 | ||
967 | /* | |
968 | * If there are empty slabs on the slabs_free list and we are | |
969 | * being forced to refill the cache, mark this one !pfmemalloc. | |
970 | */ | |
971 | l3 = cachep->nodelists[numa_mem_id()]; | |
972 | if (!list_empty(&l3->slabs_free) && force_refill) { | |
973 | struct slab *slabp = virt_to_slab(objp); | |
30c29bea | 974 | ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem)); |
072bb0aa MG |
975 | clear_obj_pfmemalloc(&objp); |
976 | recheck_pfmemalloc_active(cachep, ac); | |
977 | return objp; | |
978 | } | |
979 | ||
980 | /* No !PFMEMALLOC objects available */ | |
981 | ac->avail++; | |
982 | objp = NULL; | |
983 | } | |
984 | ||
985 | return objp; | |
986 | } | |
987 | ||
381760ea MG |
988 | static inline void *ac_get_obj(struct kmem_cache *cachep, |
989 | struct array_cache *ac, gfp_t flags, bool force_refill) | |
990 | { | |
991 | void *objp; | |
992 | ||
993 | if (unlikely(sk_memalloc_socks())) | |
994 | objp = __ac_get_obj(cachep, ac, flags, force_refill); | |
995 | else | |
996 | objp = ac->entry[--ac->avail]; | |
997 | ||
998 | return objp; | |
999 | } | |
1000 | ||
1001 | static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
072bb0aa MG |
1002 | void *objp) |
1003 | { | |
1004 | if (unlikely(pfmemalloc_active)) { | |
1005 | /* Some pfmemalloc slabs exist, check if this is one */ | |
30c29bea | 1006 | struct page *page = virt_to_head_page(objp); |
072bb0aa MG |
1007 | if (PageSlabPfmemalloc(page)) |
1008 | set_obj_pfmemalloc(&objp); | |
1009 | } | |
1010 | ||
381760ea MG |
1011 | return objp; |
1012 | } | |
1013 | ||
1014 | static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
1015 | void *objp) | |
1016 | { | |
1017 | if (unlikely(sk_memalloc_socks())) | |
1018 | objp = __ac_put_obj(cachep, ac, objp); | |
1019 | ||
072bb0aa MG |
1020 | ac->entry[ac->avail++] = objp; |
1021 | } | |
1022 | ||
3ded175a CL |
1023 | /* |
1024 | * Transfer objects in one arraycache to another. | |
1025 | * Locking must be handled by the caller. | |
1026 | * | |
1027 | * Return the number of entries transferred. | |
1028 | */ | |
1029 | static int transfer_objects(struct array_cache *to, | |
1030 | struct array_cache *from, unsigned int max) | |
1031 | { | |
1032 | /* Figure out how many entries to transfer */ | |
732eacc0 | 1033 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
1034 | |
1035 | if (!nr) | |
1036 | return 0; | |
1037 | ||
1038 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1039 | sizeof(void *) *nr); | |
1040 | ||
1041 | from->avail -= nr; | |
1042 | to->avail += nr; | |
3ded175a CL |
1043 | return nr; |
1044 | } | |
1045 | ||
765c4507 CL |
1046 | #ifndef CONFIG_NUMA |
1047 | ||
1048 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1049 | #define reap_alien(cachep, l3) do { } while (0) | |
1050 | ||
83b519e8 | 1051 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1052 | { |
1053 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1054 | } | |
1055 | ||
1056 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1057 | { | |
1058 | } | |
1059 | ||
1060 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1061 | { | |
1062 | return 0; | |
1063 | } | |
1064 | ||
1065 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1066 | gfp_t flags) | |
1067 | { | |
1068 | return NULL; | |
1069 | } | |
1070 | ||
8b98c169 | 1071 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1072 | gfp_t flags, int nodeid) |
1073 | { | |
1074 | return NULL; | |
1075 | } | |
1076 | ||
1077 | #else /* CONFIG_NUMA */ | |
1078 | ||
8b98c169 | 1079 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1080 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1081 | |
83b519e8 | 1082 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1083 | { |
1084 | struct array_cache **ac_ptr; | |
8ef82866 | 1085 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1086 | int i; |
1087 | ||
1088 | if (limit > 1) | |
1089 | limit = 12; | |
f3186a9c | 1090 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
1091 | if (ac_ptr) { |
1092 | for_each_node(i) { | |
f3186a9c | 1093 | if (i == node || !node_online(i)) |
e498be7d | 1094 | continue; |
83b519e8 | 1095 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1096 | if (!ac_ptr[i]) { |
cc550def | 1097 | for (i--; i >= 0; i--) |
e498be7d CL |
1098 | kfree(ac_ptr[i]); |
1099 | kfree(ac_ptr); | |
1100 | return NULL; | |
1101 | } | |
1102 | } | |
1103 | } | |
1104 | return ac_ptr; | |
1105 | } | |
1106 | ||
5295a74c | 1107 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1108 | { |
1109 | int i; | |
1110 | ||
1111 | if (!ac_ptr) | |
1112 | return; | |
e498be7d | 1113 | for_each_node(i) |
b28a02de | 1114 | kfree(ac_ptr[i]); |
e498be7d CL |
1115 | kfree(ac_ptr); |
1116 | } | |
1117 | ||
343e0d7a | 1118 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1119 | struct array_cache *ac, int node) |
e498be7d CL |
1120 | { |
1121 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1122 | ||
1123 | if (ac->avail) { | |
1124 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1125 | /* |
1126 | * Stuff objects into the remote nodes shared array first. | |
1127 | * That way we could avoid the overhead of putting the objects | |
1128 | * into the free lists and getting them back later. | |
1129 | */ | |
693f7d36 JS |
1130 | if (rl3->shared) |
1131 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1132 | |
ff69416e | 1133 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1134 | ac->avail = 0; |
1135 | spin_unlock(&rl3->list_lock); | |
1136 | } | |
1137 | } | |
1138 | ||
8fce4d8e CL |
1139 | /* |
1140 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1141 | */ | |
1142 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1143 | { | |
909ea964 | 1144 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1145 | |
1146 | if (l3->alien) { | |
1147 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1148 | |
1149 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1150 | __drain_alien_cache(cachep, ac, node); |
1151 | spin_unlock_irq(&ac->lock); | |
1152 | } | |
1153 | } | |
1154 | } | |
1155 | ||
a737b3e2 AM |
1156 | static void drain_alien_cache(struct kmem_cache *cachep, |
1157 | struct array_cache **alien) | |
e498be7d | 1158 | { |
b28a02de | 1159 | int i = 0; |
e498be7d CL |
1160 | struct array_cache *ac; |
1161 | unsigned long flags; | |
1162 | ||
1163 | for_each_online_node(i) { | |
4484ebf1 | 1164 | ac = alien[i]; |
e498be7d CL |
1165 | if (ac) { |
1166 | spin_lock_irqsave(&ac->lock, flags); | |
1167 | __drain_alien_cache(cachep, ac, i); | |
1168 | spin_unlock_irqrestore(&ac->lock, flags); | |
1169 | } | |
1170 | } | |
1171 | } | |
729bd0b7 | 1172 | |
873623df | 1173 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1174 | { |
1175 | struct slab *slabp = virt_to_slab(objp); | |
1176 | int nodeid = slabp->nodeid; | |
1177 | struct kmem_list3 *l3; | |
1178 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1179 | int node; |
1180 | ||
7d6e6d09 | 1181 | node = numa_mem_id(); |
729bd0b7 PE |
1182 | |
1183 | /* | |
1184 | * Make sure we are not freeing a object from another node to the array | |
1185 | * cache on this cpu. | |
1186 | */ | |
62918a03 | 1187 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1188 | return 0; |
1189 | ||
1ca4cb24 | 1190 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1191 | STATS_INC_NODEFREES(cachep); |
1192 | if (l3->alien && l3->alien[nodeid]) { | |
1193 | alien = l3->alien[nodeid]; | |
873623df | 1194 | spin_lock(&alien->lock); |
729bd0b7 PE |
1195 | if (unlikely(alien->avail == alien->limit)) { |
1196 | STATS_INC_ACOVERFLOW(cachep); | |
1197 | __drain_alien_cache(cachep, alien, nodeid); | |
1198 | } | |
072bb0aa | 1199 | ac_put_obj(cachep, alien, objp); |
729bd0b7 PE |
1200 | spin_unlock(&alien->lock); |
1201 | } else { | |
1202 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1203 | free_block(cachep, &objp, 1, nodeid); | |
1204 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1205 | } | |
1206 | return 1; | |
1207 | } | |
e498be7d CL |
1208 | #endif |
1209 | ||
8f9f8d9e DR |
1210 | /* |
1211 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1212 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1213 | * will be allocated off-node since memory is not yet online for the new node. | |
1214 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1215 | * already in use. | |
1216 | * | |
18004c5d | 1217 | * Must hold slab_mutex. |
8f9f8d9e DR |
1218 | */ |
1219 | static int init_cache_nodelists_node(int node) | |
1220 | { | |
1221 | struct kmem_cache *cachep; | |
1222 | struct kmem_list3 *l3; | |
1223 | const int memsize = sizeof(struct kmem_list3); | |
1224 | ||
18004c5d | 1225 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1226 | /* |
1227 | * Set up the size64 kmemlist for cpu before we can | |
1228 | * begin anything. Make sure some other cpu on this | |
1229 | * node has not already allocated this | |
1230 | */ | |
1231 | if (!cachep->nodelists[node]) { | |
1232 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1233 | if (!l3) | |
1234 | return -ENOMEM; | |
1235 | kmem_list3_init(l3); | |
1236 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1237 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1238 | ||
1239 | /* | |
1240 | * The l3s don't come and go as CPUs come and | |
18004c5d | 1241 | * go. slab_mutex is sufficient |
8f9f8d9e DR |
1242 | * protection here. |
1243 | */ | |
1244 | cachep->nodelists[node] = l3; | |
1245 | } | |
1246 | ||
1247 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1248 | cachep->nodelists[node]->free_limit = | |
1249 | (1 + nr_cpus_node(node)) * | |
1250 | cachep->batchcount + cachep->num; | |
1251 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1252 | } | |
1253 | return 0; | |
1254 | } | |
1255 | ||
fbf1e473 AM |
1256 | static void __cpuinit cpuup_canceled(long cpu) |
1257 | { | |
1258 | struct kmem_cache *cachep; | |
1259 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1260 | int node = cpu_to_mem(cpu); |
a70f7302 | 1261 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1262 | |
18004c5d | 1263 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1264 | struct array_cache *nc; |
1265 | struct array_cache *shared; | |
1266 | struct array_cache **alien; | |
fbf1e473 | 1267 | |
fbf1e473 AM |
1268 | /* cpu is dead; no one can alloc from it. */ |
1269 | nc = cachep->array[cpu]; | |
1270 | cachep->array[cpu] = NULL; | |
1271 | l3 = cachep->nodelists[node]; | |
1272 | ||
1273 | if (!l3) | |
1274 | goto free_array_cache; | |
1275 | ||
1276 | spin_lock_irq(&l3->list_lock); | |
1277 | ||
1278 | /* Free limit for this kmem_list3 */ | |
1279 | l3->free_limit -= cachep->batchcount; | |
1280 | if (nc) | |
1281 | free_block(cachep, nc->entry, nc->avail, node); | |
1282 | ||
58463c1f | 1283 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1284 | spin_unlock_irq(&l3->list_lock); |
1285 | goto free_array_cache; | |
1286 | } | |
1287 | ||
1288 | shared = l3->shared; | |
1289 | if (shared) { | |
1290 | free_block(cachep, shared->entry, | |
1291 | shared->avail, node); | |
1292 | l3->shared = NULL; | |
1293 | } | |
1294 | ||
1295 | alien = l3->alien; | |
1296 | l3->alien = NULL; | |
1297 | ||
1298 | spin_unlock_irq(&l3->list_lock); | |
1299 | ||
1300 | kfree(shared); | |
1301 | if (alien) { | |
1302 | drain_alien_cache(cachep, alien); | |
1303 | free_alien_cache(alien); | |
1304 | } | |
1305 | free_array_cache: | |
1306 | kfree(nc); | |
1307 | } | |
1308 | /* | |
1309 | * In the previous loop, all the objects were freed to | |
1310 | * the respective cache's slabs, now we can go ahead and | |
1311 | * shrink each nodelist to its limit. | |
1312 | */ | |
18004c5d | 1313 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1314 | l3 = cachep->nodelists[node]; |
1315 | if (!l3) | |
1316 | continue; | |
1317 | drain_freelist(cachep, l3, l3->free_objects); | |
1318 | } | |
1319 | } | |
1320 | ||
1321 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1322 | { |
343e0d7a | 1323 | struct kmem_cache *cachep; |
e498be7d | 1324 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1325 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1326 | int err; |
1da177e4 | 1327 | |
fbf1e473 AM |
1328 | /* |
1329 | * We need to do this right in the beginning since | |
1330 | * alloc_arraycache's are going to use this list. | |
1331 | * kmalloc_node allows us to add the slab to the right | |
1332 | * kmem_list3 and not this cpu's kmem_list3 | |
1333 | */ | |
8f9f8d9e DR |
1334 | err = init_cache_nodelists_node(node); |
1335 | if (err < 0) | |
1336 | goto bad; | |
fbf1e473 AM |
1337 | |
1338 | /* | |
1339 | * Now we can go ahead with allocating the shared arrays and | |
1340 | * array caches | |
1341 | */ | |
18004c5d | 1342 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1343 | struct array_cache *nc; |
1344 | struct array_cache *shared = NULL; | |
1345 | struct array_cache **alien = NULL; | |
1346 | ||
1347 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1348 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1349 | if (!nc) |
1350 | goto bad; | |
1351 | if (cachep->shared) { | |
1352 | shared = alloc_arraycache(node, | |
1353 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1354 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1355 | if (!shared) { |
1356 | kfree(nc); | |
1da177e4 | 1357 | goto bad; |
12d00f6a | 1358 | } |
fbf1e473 AM |
1359 | } |
1360 | if (use_alien_caches) { | |
83b519e8 | 1361 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1362 | if (!alien) { |
1363 | kfree(shared); | |
1364 | kfree(nc); | |
fbf1e473 | 1365 | goto bad; |
12d00f6a | 1366 | } |
fbf1e473 AM |
1367 | } |
1368 | cachep->array[cpu] = nc; | |
1369 | l3 = cachep->nodelists[node]; | |
1370 | BUG_ON(!l3); | |
1371 | ||
1372 | spin_lock_irq(&l3->list_lock); | |
1373 | if (!l3->shared) { | |
1374 | /* | |
1375 | * We are serialised from CPU_DEAD or | |
1376 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1377 | */ | |
1378 | l3->shared = shared; | |
1379 | shared = NULL; | |
1380 | } | |
4484ebf1 | 1381 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1382 | if (!l3->alien) { |
1383 | l3->alien = alien; | |
1384 | alien = NULL; | |
1da177e4 | 1385 | } |
fbf1e473 AM |
1386 | #endif |
1387 | spin_unlock_irq(&l3->list_lock); | |
1388 | kfree(shared); | |
1389 | free_alien_cache(alien); | |
83835b3d PZ |
1390 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1391 | slab_set_debugobj_lock_classes_node(cachep, node); | |
fbf1e473 | 1392 | } |
ce79ddc8 PE |
1393 | init_node_lock_keys(node); |
1394 | ||
fbf1e473 AM |
1395 | return 0; |
1396 | bad: | |
12d00f6a | 1397 | cpuup_canceled(cpu); |
fbf1e473 AM |
1398 | return -ENOMEM; |
1399 | } | |
1400 | ||
1401 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1402 | unsigned long action, void *hcpu) | |
1403 | { | |
1404 | long cpu = (long)hcpu; | |
1405 | int err = 0; | |
1406 | ||
1407 | switch (action) { | |
fbf1e473 AM |
1408 | case CPU_UP_PREPARE: |
1409 | case CPU_UP_PREPARE_FROZEN: | |
18004c5d | 1410 | mutex_lock(&slab_mutex); |
fbf1e473 | 1411 | err = cpuup_prepare(cpu); |
18004c5d | 1412 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
1413 | break; |
1414 | case CPU_ONLINE: | |
8bb78442 | 1415 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1416 | start_cpu_timer(cpu); |
1417 | break; | |
1418 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1419 | case CPU_DOWN_PREPARE: |
8bb78442 | 1420 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 | 1421 | /* |
18004c5d | 1422 | * Shutdown cache reaper. Note that the slab_mutex is |
5830c590 CL |
1423 | * held so that if cache_reap() is invoked it cannot do |
1424 | * anything expensive but will only modify reap_work | |
1425 | * and reschedule the timer. | |
1426 | */ | |
afe2c511 | 1427 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1428 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1429 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1430 | break; |
1431 | case CPU_DOWN_FAILED: | |
8bb78442 | 1432 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1433 | start_cpu_timer(cpu); |
1434 | break; | |
1da177e4 | 1435 | case CPU_DEAD: |
8bb78442 | 1436 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1437 | /* |
1438 | * Even if all the cpus of a node are down, we don't free the | |
1439 | * kmem_list3 of any cache. This to avoid a race between | |
1440 | * cpu_down, and a kmalloc allocation from another cpu for | |
1441 | * memory from the node of the cpu going down. The list3 | |
1442 | * structure is usually allocated from kmem_cache_create() and | |
1443 | * gets destroyed at kmem_cache_destroy(). | |
1444 | */ | |
183ff22b | 1445 | /* fall through */ |
8f5be20b | 1446 | #endif |
1da177e4 | 1447 | case CPU_UP_CANCELED: |
8bb78442 | 1448 | case CPU_UP_CANCELED_FROZEN: |
18004c5d | 1449 | mutex_lock(&slab_mutex); |
fbf1e473 | 1450 | cpuup_canceled(cpu); |
18004c5d | 1451 | mutex_unlock(&slab_mutex); |
1da177e4 | 1452 | break; |
1da177e4 | 1453 | } |
eac40680 | 1454 | return notifier_from_errno(err); |
1da177e4 LT |
1455 | } |
1456 | ||
74b85f37 CS |
1457 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1458 | &cpuup_callback, NULL, 0 | |
1459 | }; | |
1da177e4 | 1460 | |
8f9f8d9e DR |
1461 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1462 | /* | |
1463 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1464 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1465 | * removed. | |
1466 | * | |
18004c5d | 1467 | * Must hold slab_mutex. |
8f9f8d9e DR |
1468 | */ |
1469 | static int __meminit drain_cache_nodelists_node(int node) | |
1470 | { | |
1471 | struct kmem_cache *cachep; | |
1472 | int ret = 0; | |
1473 | ||
18004c5d | 1474 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1475 | struct kmem_list3 *l3; |
1476 | ||
1477 | l3 = cachep->nodelists[node]; | |
1478 | if (!l3) | |
1479 | continue; | |
1480 | ||
1481 | drain_freelist(cachep, l3, l3->free_objects); | |
1482 | ||
1483 | if (!list_empty(&l3->slabs_full) || | |
1484 | !list_empty(&l3->slabs_partial)) { | |
1485 | ret = -EBUSY; | |
1486 | break; | |
1487 | } | |
1488 | } | |
1489 | return ret; | |
1490 | } | |
1491 | ||
1492 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1493 | unsigned long action, void *arg) | |
1494 | { | |
1495 | struct memory_notify *mnb = arg; | |
1496 | int ret = 0; | |
1497 | int nid; | |
1498 | ||
1499 | nid = mnb->status_change_nid; | |
1500 | if (nid < 0) | |
1501 | goto out; | |
1502 | ||
1503 | switch (action) { | |
1504 | case MEM_GOING_ONLINE: | |
18004c5d | 1505 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1506 | ret = init_cache_nodelists_node(nid); |
18004c5d | 1507 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1508 | break; |
1509 | case MEM_GOING_OFFLINE: | |
18004c5d | 1510 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1511 | ret = drain_cache_nodelists_node(nid); |
18004c5d | 1512 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1513 | break; |
1514 | case MEM_ONLINE: | |
1515 | case MEM_OFFLINE: | |
1516 | case MEM_CANCEL_ONLINE: | |
1517 | case MEM_CANCEL_OFFLINE: | |
1518 | break; | |
1519 | } | |
1520 | out: | |
5fda1bd5 | 1521 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1522 | } |
1523 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1524 | ||
e498be7d CL |
1525 | /* |
1526 | * swap the static kmem_list3 with kmalloced memory | |
1527 | */ | |
8f9f8d9e DR |
1528 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1529 | int nodeid) | |
e498be7d CL |
1530 | { |
1531 | struct kmem_list3 *ptr; | |
1532 | ||
83b519e8 | 1533 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1534 | BUG_ON(!ptr); |
1535 | ||
e498be7d | 1536 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1537 | /* |
1538 | * Do not assume that spinlocks can be initialized via memcpy: | |
1539 | */ | |
1540 | spin_lock_init(&ptr->list_lock); | |
1541 | ||
e498be7d CL |
1542 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1543 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1544 | } |
1545 | ||
556a169d PE |
1546 | /* |
1547 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1548 | * size of kmem_list3. | |
1549 | */ | |
1550 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1551 | { | |
1552 | int node; | |
1553 | ||
1554 | for_each_online_node(node) { | |
1555 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1556 | cachep->nodelists[node]->next_reap = jiffies + | |
1557 | REAPTIMEOUT_LIST3 + | |
1558 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1559 | } | |
1560 | } | |
1561 | ||
a737b3e2 AM |
1562 | /* |
1563 | * Initialisation. Called after the page allocator have been initialised and | |
1564 | * before smp_init(). | |
1da177e4 LT |
1565 | */ |
1566 | void __init kmem_cache_init(void) | |
1567 | { | |
1568 | size_t left_over; | |
1569 | struct cache_sizes *sizes; | |
1570 | struct cache_names *names; | |
e498be7d | 1571 | int i; |
07ed76b2 | 1572 | int order; |
1ca4cb24 | 1573 | int node; |
e498be7d | 1574 | |
9b030cb8 CL |
1575 | kmem_cache = &kmem_cache_boot; |
1576 | ||
b6e68bc1 | 1577 | if (num_possible_nodes() == 1) |
62918a03 SS |
1578 | use_alien_caches = 0; |
1579 | ||
e498be7d CL |
1580 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1581 | kmem_list3_init(&initkmem_list3[i]); | |
1582 | if (i < MAX_NUMNODES) | |
9b030cb8 | 1583 | kmem_cache->nodelists[i] = NULL; |
e498be7d | 1584 | } |
9b030cb8 | 1585 | set_up_list3s(kmem_cache, CACHE_CACHE); |
1da177e4 LT |
1586 | |
1587 | /* | |
1588 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1589 | * page orders on machines with more than 32MB of memory if |
1590 | * not overridden on the command line. | |
1da177e4 | 1591 | */ |
3df1cccd | 1592 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1593 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1594 | |
1da177e4 LT |
1595 | /* Bootstrap is tricky, because several objects are allocated |
1596 | * from caches that do not exist yet: | |
9b030cb8 CL |
1597 | * 1) initialize the kmem_cache cache: it contains the struct |
1598 | * kmem_cache structures of all caches, except kmem_cache itself: | |
1599 | * kmem_cache is statically allocated. | |
e498be7d CL |
1600 | * Initially an __init data area is used for the head array and the |
1601 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1602 | * array at the end of the bootstrap. | |
1da177e4 | 1603 | * 2) Create the first kmalloc cache. |
343e0d7a | 1604 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1605 | * An __init data area is used for the head array. |
1606 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1607 | * head arrays. | |
9b030cb8 | 1608 | * 4) Replace the __init data head arrays for kmem_cache and the first |
1da177e4 | 1609 | * kmalloc cache with kmalloc allocated arrays. |
9b030cb8 | 1610 | * 5) Replace the __init data for kmem_list3 for kmem_cache and |
e498be7d CL |
1611 | * the other cache's with kmalloc allocated memory. |
1612 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1613 | */ |
1614 | ||
7d6e6d09 | 1615 | node = numa_mem_id(); |
1ca4cb24 | 1616 | |
9b030cb8 | 1617 | /* 1) create the kmem_cache */ |
18004c5d | 1618 | INIT_LIST_HEAD(&slab_caches); |
9b030cb8 CL |
1619 | list_add(&kmem_cache->list, &slab_caches); |
1620 | kmem_cache->colour_off = cache_line_size(); | |
1621 | kmem_cache->array[smp_processor_id()] = &initarray_cache.cache; | |
1622 | kmem_cache->nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; | |
1da177e4 | 1623 | |
8da3430d | 1624 | /* |
b56efcf0 | 1625 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1626 | */ |
9b030cb8 | 1627 | kmem_cache->size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
b56efcf0 | 1628 | nr_node_ids * sizeof(struct kmem_list3 *); |
9b030cb8 CL |
1629 | kmem_cache->object_size = kmem_cache->size; |
1630 | kmem_cache->size = ALIGN(kmem_cache->object_size, | |
a737b3e2 | 1631 | cache_line_size()); |
9b030cb8 CL |
1632 | kmem_cache->reciprocal_buffer_size = |
1633 | reciprocal_value(kmem_cache->size); | |
1da177e4 | 1634 | |
07ed76b2 | 1635 | for (order = 0; order < MAX_ORDER; order++) { |
9b030cb8 CL |
1636 | cache_estimate(order, kmem_cache->size, |
1637 | cache_line_size(), 0, &left_over, &kmem_cache->num); | |
1638 | if (kmem_cache->num) | |
07ed76b2 JS |
1639 | break; |
1640 | } | |
9b030cb8 CL |
1641 | BUG_ON(!kmem_cache->num); |
1642 | kmem_cache->gfporder = order; | |
1643 | kmem_cache->colour = left_over / kmem_cache->colour_off; | |
1644 | kmem_cache->slab_size = ALIGN(kmem_cache->num * sizeof(kmem_bufctl_t) + | |
b28a02de | 1645 | sizeof(struct slab), cache_line_size()); |
1da177e4 LT |
1646 | |
1647 | /* 2+3) create the kmalloc caches */ | |
1648 | sizes = malloc_sizes; | |
1649 | names = cache_names; | |
1650 | ||
a737b3e2 AM |
1651 | /* |
1652 | * Initialize the caches that provide memory for the array cache and the | |
1653 | * kmem_list3 structures first. Without this, further allocations will | |
1654 | * bug. | |
e498be7d CL |
1655 | */ |
1656 | ||
278b1bb1 | 1657 | sizes[INDEX_AC].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1658 | sizes[INDEX_AC].cs_cachep->name = names[INDEX_AC].name; |
1659 | sizes[INDEX_AC].cs_cachep->size = sizes[INDEX_AC].cs_size; | |
1660 | sizes[INDEX_AC].cs_cachep->object_size = sizes[INDEX_AC].cs_size; | |
1661 | sizes[INDEX_AC].cs_cachep->align = ARCH_KMALLOC_MINALIGN; | |
1662 | __kmem_cache_create(sizes[INDEX_AC].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC); | |
7c9adf5a | 1663 | list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches); |
e498be7d | 1664 | |
a737b3e2 | 1665 | if (INDEX_AC != INDEX_L3) { |
278b1bb1 | 1666 | sizes[INDEX_L3].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1667 | sizes[INDEX_L3].cs_cachep->name = names[INDEX_L3].name; |
1668 | sizes[INDEX_L3].cs_cachep->size = sizes[INDEX_L3].cs_size; | |
1669 | sizes[INDEX_L3].cs_cachep->object_size = sizes[INDEX_L3].cs_size; | |
1670 | sizes[INDEX_L3].cs_cachep->align = ARCH_KMALLOC_MINALIGN; | |
1671 | __kmem_cache_create(sizes[INDEX_L3].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC); | |
7c9adf5a | 1672 | list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches); |
a737b3e2 | 1673 | } |
e498be7d | 1674 | |
e0a42726 IM |
1675 | slab_early_init = 0; |
1676 | ||
1da177e4 | 1677 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1678 | /* |
1679 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1680 | * This should be particularly beneficial on SMP boxes, as it |
1681 | * eliminates "false sharing". | |
1682 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1683 | * allow tighter packing of the smaller caches. |
1684 | */ | |
a737b3e2 | 1685 | if (!sizes->cs_cachep) { |
278b1bb1 | 1686 | sizes->cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1687 | sizes->cs_cachep->name = names->name; |
1688 | sizes->cs_cachep->size = sizes->cs_size; | |
1689 | sizes->cs_cachep->object_size = sizes->cs_size; | |
1690 | sizes->cs_cachep->align = ARCH_KMALLOC_MINALIGN; | |
1691 | __kmem_cache_create(sizes->cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC); | |
7c9adf5a | 1692 | list_add(&sizes->cs_cachep->list, &slab_caches); |
a737b3e2 | 1693 | } |
4b51d669 | 1694 | #ifdef CONFIG_ZONE_DMA |
278b1bb1 | 1695 | sizes->cs_dmacachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1696 | sizes->cs_dmacachep->name = names->name_dma; |
1697 | sizes->cs_dmacachep->size = sizes->cs_size; | |
1698 | sizes->cs_dmacachep->object_size = sizes->cs_size; | |
1699 | sizes->cs_dmacachep->align = ARCH_KMALLOC_MINALIGN; | |
278b1bb1 | 1700 | __kmem_cache_create(sizes->cs_dmacachep, |
8a13a4cc | 1701 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| SLAB_PANIC); |
7c9adf5a | 1702 | list_add(&sizes->cs_dmacachep->list, &slab_caches); |
4b51d669 | 1703 | #endif |
1da177e4 LT |
1704 | sizes++; |
1705 | names++; | |
1706 | } | |
1707 | /* 4) Replace the bootstrap head arrays */ | |
1708 | { | |
2b2d5493 | 1709 | struct array_cache *ptr; |
e498be7d | 1710 | |
83b519e8 | 1711 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1712 | |
9b030cb8 CL |
1713 | BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache); |
1714 | memcpy(ptr, cpu_cache_get(kmem_cache), | |
b28a02de | 1715 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1716 | /* |
1717 | * Do not assume that spinlocks can be initialized via memcpy: | |
1718 | */ | |
1719 | spin_lock_init(&ptr->lock); | |
1720 | ||
9b030cb8 | 1721 | kmem_cache->array[smp_processor_id()] = ptr; |
e498be7d | 1722 | |
83b519e8 | 1723 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1724 | |
9a2dba4b | 1725 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1726 | != &initarray_generic.cache); |
9a2dba4b | 1727 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1728 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1729 | /* |
1730 | * Do not assume that spinlocks can be initialized via memcpy: | |
1731 | */ | |
1732 | spin_lock_init(&ptr->lock); | |
1733 | ||
e498be7d | 1734 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1735 | ptr; |
1da177e4 | 1736 | } |
e498be7d CL |
1737 | /* 5) Replace the bootstrap kmem_list3's */ |
1738 | { | |
1ca4cb24 PE |
1739 | int nid; |
1740 | ||
9c09a95c | 1741 | for_each_online_node(nid) { |
9b030cb8 | 1742 | init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1743 | |
e498be7d | 1744 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1745 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1746 | |
1747 | if (INDEX_AC != INDEX_L3) { | |
1748 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1749 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1750 | } |
1751 | } | |
1752 | } | |
1da177e4 | 1753 | |
97d06609 | 1754 | slab_state = UP; |
8429db5c PE |
1755 | } |
1756 | ||
1757 | void __init kmem_cache_init_late(void) | |
1758 | { | |
1759 | struct kmem_cache *cachep; | |
1760 | ||
97d06609 | 1761 | slab_state = UP; |
52cef189 | 1762 | |
8429db5c | 1763 | /* 6) resize the head arrays to their final sizes */ |
18004c5d CL |
1764 | mutex_lock(&slab_mutex); |
1765 | list_for_each_entry(cachep, &slab_caches, list) | |
8429db5c PE |
1766 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1767 | BUG(); | |
18004c5d | 1768 | mutex_unlock(&slab_mutex); |
056c6241 | 1769 | |
947ca185 MW |
1770 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1771 | init_lock_keys(); | |
1772 | ||
97d06609 CL |
1773 | /* Done! */ |
1774 | slab_state = FULL; | |
1775 | ||
a737b3e2 AM |
1776 | /* |
1777 | * Register a cpu startup notifier callback that initializes | |
1778 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1779 | */ |
1780 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1781 | |
8f9f8d9e DR |
1782 | #ifdef CONFIG_NUMA |
1783 | /* | |
1784 | * Register a memory hotplug callback that initializes and frees | |
1785 | * nodelists. | |
1786 | */ | |
1787 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1788 | #endif | |
1789 | ||
a737b3e2 AM |
1790 | /* |
1791 | * The reap timers are started later, with a module init call: That part | |
1792 | * of the kernel is not yet operational. | |
1da177e4 LT |
1793 | */ |
1794 | } | |
1795 | ||
1796 | static int __init cpucache_init(void) | |
1797 | { | |
1798 | int cpu; | |
1799 | ||
a737b3e2 AM |
1800 | /* |
1801 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1802 | */ |
e498be7d | 1803 | for_each_online_cpu(cpu) |
a737b3e2 | 1804 | start_cpu_timer(cpu); |
a164f896 GC |
1805 | |
1806 | /* Done! */ | |
97d06609 | 1807 | slab_state = FULL; |
1da177e4 LT |
1808 | return 0; |
1809 | } | |
1da177e4 LT |
1810 | __initcall(cpucache_init); |
1811 | ||
8bdec192 RA |
1812 | static noinline void |
1813 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1814 | { | |
1815 | struct kmem_list3 *l3; | |
1816 | struct slab *slabp; | |
1817 | unsigned long flags; | |
1818 | int node; | |
1819 | ||
1820 | printk(KERN_WARNING | |
1821 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1822 | nodeid, gfpflags); | |
1823 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1824 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1825 | |
1826 | for_each_online_node(node) { | |
1827 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1828 | unsigned long active_slabs = 0, num_slabs = 0; | |
1829 | ||
1830 | l3 = cachep->nodelists[node]; | |
1831 | if (!l3) | |
1832 | continue; | |
1833 | ||
1834 | spin_lock_irqsave(&l3->list_lock, flags); | |
1835 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1836 | active_objs += cachep->num; | |
1837 | active_slabs++; | |
1838 | } | |
1839 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1840 | active_objs += slabp->inuse; | |
1841 | active_slabs++; | |
1842 | } | |
1843 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1844 | num_slabs++; | |
1845 | ||
1846 | free_objects += l3->free_objects; | |
1847 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1848 | ||
1849 | num_slabs += active_slabs; | |
1850 | num_objs = num_slabs * cachep->num; | |
1851 | printk(KERN_WARNING | |
1852 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1853 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1854 | free_objects); | |
1855 | } | |
1856 | } | |
1857 | ||
1da177e4 LT |
1858 | /* |
1859 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1860 | * | |
1861 | * If we requested dmaable memory, we will get it. Even if we | |
1862 | * did not request dmaable memory, we might get it, but that | |
1863 | * would be relatively rare and ignorable. | |
1864 | */ | |
343e0d7a | 1865 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1866 | { |
1867 | struct page *page; | |
e1b6aa6f | 1868 | int nr_pages; |
1da177e4 LT |
1869 | int i; |
1870 | ||
d6fef9da | 1871 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1872 | /* |
1873 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1874 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1875 | */ |
e1b6aa6f | 1876 | flags |= __GFP_COMP; |
d6fef9da | 1877 | #endif |
765c4507 | 1878 | |
a618e89f | 1879 | flags |= cachep->allocflags; |
e12ba74d MG |
1880 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1881 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1882 | |
517d0869 | 1883 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1884 | if (!page) { |
1885 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1886 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1887 | return NULL; |
8bdec192 | 1888 | } |
1da177e4 | 1889 | |
b37f1dd0 | 1890 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
072bb0aa MG |
1891 | if (unlikely(page->pfmemalloc)) |
1892 | pfmemalloc_active = true; | |
1893 | ||
e1b6aa6f | 1894 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1895 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1896 | add_zone_page_state(page_zone(page), |
1897 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1898 | else | |
1899 | add_zone_page_state(page_zone(page), | |
1900 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
072bb0aa | 1901 | for (i = 0; i < nr_pages; i++) { |
e1b6aa6f | 1902 | __SetPageSlab(page + i); |
c175eea4 | 1903 | |
072bb0aa MG |
1904 | if (page->pfmemalloc) |
1905 | SetPageSlabPfmemalloc(page + i); | |
1906 | } | |
1907 | ||
b1eeab67 VN |
1908 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1909 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1910 | ||
1911 | if (cachep->ctor) | |
1912 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1913 | else | |
1914 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1915 | } | |
c175eea4 | 1916 | |
e1b6aa6f | 1917 | return page_address(page); |
1da177e4 LT |
1918 | } |
1919 | ||
1920 | /* | |
1921 | * Interface to system's page release. | |
1922 | */ | |
343e0d7a | 1923 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1924 | { |
b28a02de | 1925 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1926 | struct page *page = virt_to_page(addr); |
1927 | const unsigned long nr_freed = i; | |
1928 | ||
b1eeab67 | 1929 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1930 | |
972d1a7b CL |
1931 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1932 | sub_zone_page_state(page_zone(page), | |
1933 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1934 | else | |
1935 | sub_zone_page_state(page_zone(page), | |
1936 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1937 | while (i--) { |
f205b2fe | 1938 | BUG_ON(!PageSlab(page)); |
072bb0aa | 1939 | __ClearPageSlabPfmemalloc(page); |
f205b2fe | 1940 | __ClearPageSlab(page); |
1da177e4 LT |
1941 | page++; |
1942 | } | |
1da177e4 LT |
1943 | if (current->reclaim_state) |
1944 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1945 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1946 | } |
1947 | ||
1948 | static void kmem_rcu_free(struct rcu_head *head) | |
1949 | { | |
b28a02de | 1950 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1951 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1952 | |
1953 | kmem_freepages(cachep, slab_rcu->addr); | |
1954 | if (OFF_SLAB(cachep)) | |
1955 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1956 | } | |
1957 | ||
1958 | #if DEBUG | |
1959 | ||
1960 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1961 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1962 | unsigned long caller) |
1da177e4 | 1963 | { |
8c138bc0 | 1964 | int size = cachep->object_size; |
1da177e4 | 1965 | |
3dafccf2 | 1966 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1967 | |
b28a02de | 1968 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1969 | return; |
1970 | ||
b28a02de PE |
1971 | *addr++ = 0x12345678; |
1972 | *addr++ = caller; | |
1973 | *addr++ = smp_processor_id(); | |
1974 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1975 | { |
1976 | unsigned long *sptr = &caller; | |
1977 | unsigned long svalue; | |
1978 | ||
1979 | while (!kstack_end(sptr)) { | |
1980 | svalue = *sptr++; | |
1981 | if (kernel_text_address(svalue)) { | |
b28a02de | 1982 | *addr++ = svalue; |
1da177e4 LT |
1983 | size -= sizeof(unsigned long); |
1984 | if (size <= sizeof(unsigned long)) | |
1985 | break; | |
1986 | } | |
1987 | } | |
1988 | ||
1989 | } | |
b28a02de | 1990 | *addr++ = 0x87654321; |
1da177e4 LT |
1991 | } |
1992 | #endif | |
1993 | ||
343e0d7a | 1994 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1995 | { |
8c138bc0 | 1996 | int size = cachep->object_size; |
3dafccf2 | 1997 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1998 | |
1999 | memset(addr, val, size); | |
b28a02de | 2000 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
2001 | } |
2002 | ||
2003 | static void dump_line(char *data, int offset, int limit) | |
2004 | { | |
2005 | int i; | |
aa83aa40 DJ |
2006 | unsigned char error = 0; |
2007 | int bad_count = 0; | |
2008 | ||
fdde6abb | 2009 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
2010 | for (i = 0; i < limit; i++) { |
2011 | if (data[offset + i] != POISON_FREE) { | |
2012 | error = data[offset + i]; | |
2013 | bad_count++; | |
2014 | } | |
aa83aa40 | 2015 | } |
fdde6abb SAS |
2016 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
2017 | &data[offset], limit, 1); | |
aa83aa40 DJ |
2018 | |
2019 | if (bad_count == 1) { | |
2020 | error ^= POISON_FREE; | |
2021 | if (!(error & (error - 1))) { | |
2022 | printk(KERN_ERR "Single bit error detected. Probably " | |
2023 | "bad RAM.\n"); | |
2024 | #ifdef CONFIG_X86 | |
2025 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
2026 | "test tool.\n"); | |
2027 | #else | |
2028 | printk(KERN_ERR "Run a memory test tool.\n"); | |
2029 | #endif | |
2030 | } | |
2031 | } | |
1da177e4 LT |
2032 | } |
2033 | #endif | |
2034 | ||
2035 | #if DEBUG | |
2036 | ||
343e0d7a | 2037 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
2038 | { |
2039 | int i, size; | |
2040 | char *realobj; | |
2041 | ||
2042 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 2043 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
2044 | *dbg_redzone1(cachep, objp), |
2045 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2046 | } |
2047 | ||
2048 | if (cachep->flags & SLAB_STORE_USER) { | |
2049 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 2050 | *dbg_userword(cachep, objp)); |
1da177e4 | 2051 | print_symbol("(%s)", |
a737b3e2 | 2052 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
2053 | printk("\n"); |
2054 | } | |
3dafccf2 | 2055 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2056 | size = cachep->object_size; |
b28a02de | 2057 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
2058 | int limit; |
2059 | limit = 16; | |
b28a02de PE |
2060 | if (i + limit > size) |
2061 | limit = size - i; | |
1da177e4 LT |
2062 | dump_line(realobj, i, limit); |
2063 | } | |
2064 | } | |
2065 | ||
343e0d7a | 2066 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
2067 | { |
2068 | char *realobj; | |
2069 | int size, i; | |
2070 | int lines = 0; | |
2071 | ||
3dafccf2 | 2072 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2073 | size = cachep->object_size; |
1da177e4 | 2074 | |
b28a02de | 2075 | for (i = 0; i < size; i++) { |
1da177e4 | 2076 | char exp = POISON_FREE; |
b28a02de | 2077 | if (i == size - 1) |
1da177e4 LT |
2078 | exp = POISON_END; |
2079 | if (realobj[i] != exp) { | |
2080 | int limit; | |
2081 | /* Mismatch ! */ | |
2082 | /* Print header */ | |
2083 | if (lines == 0) { | |
b28a02de | 2084 | printk(KERN_ERR |
face37f5 DJ |
2085 | "Slab corruption (%s): %s start=%p, len=%d\n", |
2086 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
2087 | print_objinfo(cachep, objp, 0); |
2088 | } | |
2089 | /* Hexdump the affected line */ | |
b28a02de | 2090 | i = (i / 16) * 16; |
1da177e4 | 2091 | limit = 16; |
b28a02de PE |
2092 | if (i + limit > size) |
2093 | limit = size - i; | |
1da177e4 LT |
2094 | dump_line(realobj, i, limit); |
2095 | i += 16; | |
2096 | lines++; | |
2097 | /* Limit to 5 lines */ | |
2098 | if (lines > 5) | |
2099 | break; | |
2100 | } | |
2101 | } | |
2102 | if (lines != 0) { | |
2103 | /* Print some data about the neighboring objects, if they | |
2104 | * exist: | |
2105 | */ | |
6ed5eb22 | 2106 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 2107 | unsigned int objnr; |
1da177e4 | 2108 | |
8fea4e96 | 2109 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2110 | if (objnr) { |
8fea4e96 | 2111 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2112 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2113 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2114 | realobj, size); |
1da177e4 LT |
2115 | print_objinfo(cachep, objp, 2); |
2116 | } | |
b28a02de | 2117 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2118 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2119 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2120 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2121 | realobj, size); |
1da177e4 LT |
2122 | print_objinfo(cachep, objp, 2); |
2123 | } | |
2124 | } | |
2125 | } | |
2126 | #endif | |
2127 | ||
12dd36fa | 2128 | #if DEBUG |
e79aec29 | 2129 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2130 | { |
1da177e4 LT |
2131 | int i; |
2132 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2133 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2134 | |
2135 | if (cachep->flags & SLAB_POISON) { | |
2136 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2137 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2138 | OFF_SLAB(cachep)) |
b28a02de | 2139 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2140 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2141 | else |
2142 | check_poison_obj(cachep, objp); | |
2143 | #else | |
2144 | check_poison_obj(cachep, objp); | |
2145 | #endif | |
2146 | } | |
2147 | if (cachep->flags & SLAB_RED_ZONE) { | |
2148 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2149 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2150 | "was overwritten"); |
1da177e4 LT |
2151 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2152 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2153 | "was overwritten"); |
1da177e4 | 2154 | } |
1da177e4 | 2155 | } |
12dd36fa | 2156 | } |
1da177e4 | 2157 | #else |
e79aec29 | 2158 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2159 | { |
12dd36fa | 2160 | } |
1da177e4 LT |
2161 | #endif |
2162 | ||
911851e6 RD |
2163 | /** |
2164 | * slab_destroy - destroy and release all objects in a slab | |
2165 | * @cachep: cache pointer being destroyed | |
2166 | * @slabp: slab pointer being destroyed | |
2167 | * | |
12dd36fa | 2168 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2169 | * Before calling the slab must have been unlinked from the cache. The |
2170 | * cache-lock is not held/needed. | |
12dd36fa | 2171 | */ |
343e0d7a | 2172 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2173 | { |
2174 | void *addr = slabp->s_mem - slabp->colouroff; | |
2175 | ||
e79aec29 | 2176 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2177 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2178 | struct slab_rcu *slab_rcu; | |
2179 | ||
b28a02de | 2180 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2181 | slab_rcu->cachep = cachep; |
2182 | slab_rcu->addr = addr; | |
2183 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2184 | } else { | |
2185 | kmem_freepages(cachep, addr); | |
873623df IM |
2186 | if (OFF_SLAB(cachep)) |
2187 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2188 | } |
2189 | } | |
2190 | ||
4d268eba | 2191 | /** |
a70773dd RD |
2192 | * calculate_slab_order - calculate size (page order) of slabs |
2193 | * @cachep: pointer to the cache that is being created | |
2194 | * @size: size of objects to be created in this cache. | |
2195 | * @align: required alignment for the objects. | |
2196 | * @flags: slab allocation flags | |
2197 | * | |
2198 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2199 | * |
2200 | * This could be made much more intelligent. For now, try to avoid using | |
2201 | * high order pages for slabs. When the gfp() functions are more friendly | |
2202 | * towards high-order requests, this should be changed. | |
2203 | */ | |
a737b3e2 | 2204 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2205 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2206 | { |
b1ab41c4 | 2207 | unsigned long offslab_limit; |
4d268eba | 2208 | size_t left_over = 0; |
9888e6fa | 2209 | int gfporder; |
4d268eba | 2210 | |
0aa817f0 | 2211 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2212 | unsigned int num; |
2213 | size_t remainder; | |
2214 | ||
9888e6fa | 2215 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2216 | if (!num) |
2217 | continue; | |
9888e6fa | 2218 | |
b1ab41c4 IM |
2219 | if (flags & CFLGS_OFF_SLAB) { |
2220 | /* | |
2221 | * Max number of objs-per-slab for caches which | |
2222 | * use off-slab slabs. Needed to avoid a possible | |
2223 | * looping condition in cache_grow(). | |
2224 | */ | |
2225 | offslab_limit = size - sizeof(struct slab); | |
2226 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2227 | ||
2228 | if (num > offslab_limit) | |
2229 | break; | |
2230 | } | |
4d268eba | 2231 | |
9888e6fa | 2232 | /* Found something acceptable - save it away */ |
4d268eba | 2233 | cachep->num = num; |
9888e6fa | 2234 | cachep->gfporder = gfporder; |
4d268eba PE |
2235 | left_over = remainder; |
2236 | ||
f78bb8ad LT |
2237 | /* |
2238 | * A VFS-reclaimable slab tends to have most allocations | |
2239 | * as GFP_NOFS and we really don't want to have to be allocating | |
2240 | * higher-order pages when we are unable to shrink dcache. | |
2241 | */ | |
2242 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2243 | break; | |
2244 | ||
4d268eba PE |
2245 | /* |
2246 | * Large number of objects is good, but very large slabs are | |
2247 | * currently bad for the gfp()s. | |
2248 | */ | |
543585cc | 2249 | if (gfporder >= slab_max_order) |
4d268eba PE |
2250 | break; |
2251 | ||
9888e6fa LT |
2252 | /* |
2253 | * Acceptable internal fragmentation? | |
2254 | */ | |
a737b3e2 | 2255 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2256 | break; |
2257 | } | |
2258 | return left_over; | |
2259 | } | |
2260 | ||
83b519e8 | 2261 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2262 | { |
97d06609 | 2263 | if (slab_state >= FULL) |
83b519e8 | 2264 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2265 | |
97d06609 | 2266 | if (slab_state == DOWN) { |
f30cf7d1 PE |
2267 | /* |
2268 | * Note: the first kmem_cache_create must create the cache | |
2269 | * that's used by kmalloc(24), otherwise the creation of | |
2270 | * further caches will BUG(). | |
2271 | */ | |
2272 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2273 | ||
2274 | /* | |
2275 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2276 | * the first cache, then we need to set up all its list3s, | |
2277 | * otherwise the creation of further caches will BUG(). | |
2278 | */ | |
2279 | set_up_list3s(cachep, SIZE_AC); | |
2280 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2281 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2282 | else |
97d06609 | 2283 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 PE |
2284 | } else { |
2285 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2286 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2287 | |
97d06609 | 2288 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2289 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2290 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2291 | } else { |
2292 | int node; | |
556a169d | 2293 | for_each_online_node(node) { |
f30cf7d1 PE |
2294 | cachep->nodelists[node] = |
2295 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2296 | gfp, node); |
f30cf7d1 PE |
2297 | BUG_ON(!cachep->nodelists[node]); |
2298 | kmem_list3_init(cachep->nodelists[node]); | |
2299 | } | |
2300 | } | |
2301 | } | |
7d6e6d09 | 2302 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2303 | jiffies + REAPTIMEOUT_LIST3 + |
2304 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2305 | ||
2306 | cpu_cache_get(cachep)->avail = 0; | |
2307 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2308 | cpu_cache_get(cachep)->batchcount = 1; | |
2309 | cpu_cache_get(cachep)->touched = 0; | |
2310 | cachep->batchcount = 1; | |
2311 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2312 | return 0; |
f30cf7d1 PE |
2313 | } |
2314 | ||
1da177e4 | 2315 | /** |
039363f3 | 2316 | * __kmem_cache_create - Create a cache. |
a755b76a | 2317 | * @cachep: cache management descriptor |
1da177e4 | 2318 | * @flags: SLAB flags |
1da177e4 LT |
2319 | * |
2320 | * Returns a ptr to the cache on success, NULL on failure. | |
2321 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2322 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 | 2323 | * |
1da177e4 LT |
2324 | * The flags are |
2325 | * | |
2326 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2327 | * to catch references to uninitialised memory. | |
2328 | * | |
2329 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2330 | * for buffer overruns. | |
2331 | * | |
1da177e4 LT |
2332 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2333 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2334 | * as davem. | |
2335 | */ | |
278b1bb1 | 2336 | int |
8a13a4cc | 2337 | __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) |
1da177e4 LT |
2338 | { |
2339 | size_t left_over, slab_size, ralign; | |
83b519e8 | 2340 | gfp_t gfp; |
278b1bb1 | 2341 | int err; |
8a13a4cc | 2342 | size_t size = cachep->size; |
1da177e4 | 2343 | |
1da177e4 | 2344 | #if DEBUG |
1da177e4 LT |
2345 | #if FORCED_DEBUG |
2346 | /* | |
2347 | * Enable redzoning and last user accounting, except for caches with | |
2348 | * large objects, if the increased size would increase the object size | |
2349 | * above the next power of two: caches with object sizes just above a | |
2350 | * power of two have a significant amount of internal fragmentation. | |
2351 | */ | |
87a927c7 DW |
2352 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2353 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2354 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2355 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2356 | flags |= SLAB_POISON; | |
2357 | #endif | |
2358 | if (flags & SLAB_DESTROY_BY_RCU) | |
2359 | BUG_ON(flags & SLAB_POISON); | |
2360 | #endif | |
1da177e4 | 2361 | |
a737b3e2 AM |
2362 | /* |
2363 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2364 | * unaligned accesses for some archs when redzoning is used, and makes |
2365 | * sure any on-slab bufctl's are also correctly aligned. | |
2366 | */ | |
b28a02de PE |
2367 | if (size & (BYTES_PER_WORD - 1)) { |
2368 | size += (BYTES_PER_WORD - 1); | |
2369 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2370 | } |
2371 | ||
a737b3e2 AM |
2372 | /* calculate the final buffer alignment: */ |
2373 | ||
1da177e4 LT |
2374 | /* 1) arch recommendation: can be overridden for debug */ |
2375 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2376 | /* |
2377 | * Default alignment: as specified by the arch code. Except if | |
2378 | * an object is really small, then squeeze multiple objects into | |
2379 | * one cacheline. | |
1da177e4 LT |
2380 | */ |
2381 | ralign = cache_line_size(); | |
b28a02de | 2382 | while (size <= ralign / 2) |
1da177e4 LT |
2383 | ralign /= 2; |
2384 | } else { | |
2385 | ralign = BYTES_PER_WORD; | |
2386 | } | |
ca5f9703 PE |
2387 | |
2388 | /* | |
87a927c7 DW |
2389 | * Redzoning and user store require word alignment or possibly larger. |
2390 | * Note this will be overridden by architecture or caller mandated | |
2391 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2392 | */ |
87a927c7 DW |
2393 | if (flags & SLAB_STORE_USER) |
2394 | ralign = BYTES_PER_WORD; | |
2395 | ||
2396 | if (flags & SLAB_RED_ZONE) { | |
2397 | ralign = REDZONE_ALIGN; | |
2398 | /* If redzoning, ensure that the second redzone is suitably | |
2399 | * aligned, by adjusting the object size accordingly. */ | |
2400 | size += REDZONE_ALIGN - 1; | |
2401 | size &= ~(REDZONE_ALIGN - 1); | |
2402 | } | |
ca5f9703 | 2403 | |
a44b56d3 | 2404 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2405 | if (ralign < ARCH_SLAB_MINALIGN) { |
2406 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2407 | } |
a44b56d3 | 2408 | /* 3) caller mandated alignment */ |
8a13a4cc CL |
2409 | if (ralign < cachep->align) { |
2410 | ralign = cachep->align; | |
1da177e4 | 2411 | } |
3ff84a7f PE |
2412 | /* disable debug if necessary */ |
2413 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2414 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2415 | /* |
ca5f9703 | 2416 | * 4) Store it. |
1da177e4 | 2417 | */ |
8a13a4cc | 2418 | cachep->align = ralign; |
1da177e4 | 2419 | |
83b519e8 PE |
2420 | if (slab_is_available()) |
2421 | gfp = GFP_KERNEL; | |
2422 | else | |
2423 | gfp = GFP_NOWAIT; | |
2424 | ||
b56efcf0 | 2425 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
1da177e4 | 2426 | #if DEBUG |
1da177e4 | 2427 | |
ca5f9703 PE |
2428 | /* |
2429 | * Both debugging options require word-alignment which is calculated | |
2430 | * into align above. | |
2431 | */ | |
1da177e4 | 2432 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2433 | /* add space for red zone words */ |
3ff84a7f PE |
2434 | cachep->obj_offset += sizeof(unsigned long long); |
2435 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2436 | } |
2437 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2438 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2439 | * the real object. But if the second red zone needs to be |
2440 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2441 | */ |
87a927c7 DW |
2442 | if (flags & SLAB_RED_ZONE) |
2443 | size += REDZONE_ALIGN; | |
2444 | else | |
2445 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2446 | } |
2447 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2448 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
608da7e3 TH |
2449 | && cachep->object_size > cache_line_size() |
2450 | && ALIGN(size, cachep->align) < PAGE_SIZE) { | |
2451 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align); | |
1da177e4 LT |
2452 | size = PAGE_SIZE; |
2453 | } | |
2454 | #endif | |
2455 | #endif | |
2456 | ||
e0a42726 IM |
2457 | /* |
2458 | * Determine if the slab management is 'on' or 'off' slab. | |
2459 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2460 | * it too early on. Always use on-slab management when |
2461 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2462 | */ |
e7cb55b9 CM |
2463 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2464 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2465 | /* |
2466 | * Size is large, assume best to place the slab management obj | |
2467 | * off-slab (should allow better packing of objs). | |
2468 | */ | |
2469 | flags |= CFLGS_OFF_SLAB; | |
2470 | ||
8a13a4cc | 2471 | size = ALIGN(size, cachep->align); |
1da177e4 | 2472 | |
8a13a4cc | 2473 | left_over = calculate_slab_order(cachep, size, cachep->align, flags); |
1da177e4 | 2474 | |
8a13a4cc | 2475 | if (!cachep->num) |
278b1bb1 | 2476 | return -E2BIG; |
1da177e4 | 2477 | |
b28a02de | 2478 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
8a13a4cc | 2479 | + sizeof(struct slab), cachep->align); |
1da177e4 LT |
2480 | |
2481 | /* | |
2482 | * If the slab has been placed off-slab, and we have enough space then | |
2483 | * move it on-slab. This is at the expense of any extra colouring. | |
2484 | */ | |
2485 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2486 | flags &= ~CFLGS_OFF_SLAB; | |
2487 | left_over -= slab_size; | |
2488 | } | |
2489 | ||
2490 | if (flags & CFLGS_OFF_SLAB) { | |
2491 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2492 | slab_size = |
2493 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2494 | |
2495 | #ifdef CONFIG_PAGE_POISONING | |
2496 | /* If we're going to use the generic kernel_map_pages() | |
2497 | * poisoning, then it's going to smash the contents of | |
2498 | * the redzone and userword anyhow, so switch them off. | |
2499 | */ | |
2500 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2501 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2502 | #endif | |
1da177e4 LT |
2503 | } |
2504 | ||
2505 | cachep->colour_off = cache_line_size(); | |
2506 | /* Offset must be a multiple of the alignment. */ | |
8a13a4cc CL |
2507 | if (cachep->colour_off < cachep->align) |
2508 | cachep->colour_off = cachep->align; | |
b28a02de | 2509 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2510 | cachep->slab_size = slab_size; |
2511 | cachep->flags = flags; | |
a618e89f | 2512 | cachep->allocflags = 0; |
4b51d669 | 2513 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2514 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2515 | cachep->size = size; |
6a2d7a95 | 2516 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2517 | |
e5ac9c5a | 2518 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2519 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2520 | /* |
2521 | * This is a possibility for one of the malloc_sizes caches. | |
2522 | * But since we go off slab only for object size greater than | |
2523 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2524 | * this should not happen at all. | |
2525 | * But leave a BUG_ON for some lucky dude. | |
2526 | */ | |
6cb8f913 | 2527 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2528 | } |
1da177e4 | 2529 | |
278b1bb1 CL |
2530 | err = setup_cpu_cache(cachep, gfp); |
2531 | if (err) { | |
12c3667f | 2532 | __kmem_cache_shutdown(cachep); |
278b1bb1 | 2533 | return err; |
2ed3a4ef | 2534 | } |
1da177e4 | 2535 | |
83835b3d PZ |
2536 | if (flags & SLAB_DEBUG_OBJECTS) { |
2537 | /* | |
2538 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2539 | * debug_object_activate()->kmem_cache_alloc(). | |
2540 | */ | |
2541 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2542 | ||
2543 | slab_set_debugobj_lock_classes(cachep); | |
2544 | } | |
2545 | ||
278b1bb1 | 2546 | return 0; |
1da177e4 | 2547 | } |
1da177e4 LT |
2548 | |
2549 | #if DEBUG | |
2550 | static void check_irq_off(void) | |
2551 | { | |
2552 | BUG_ON(!irqs_disabled()); | |
2553 | } | |
2554 | ||
2555 | static void check_irq_on(void) | |
2556 | { | |
2557 | BUG_ON(irqs_disabled()); | |
2558 | } | |
2559 | ||
343e0d7a | 2560 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2561 | { |
2562 | #ifdef CONFIG_SMP | |
2563 | check_irq_off(); | |
7d6e6d09 | 2564 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2565 | #endif |
2566 | } | |
e498be7d | 2567 | |
343e0d7a | 2568 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2569 | { |
2570 | #ifdef CONFIG_SMP | |
2571 | check_irq_off(); | |
2572 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2573 | #endif | |
2574 | } | |
2575 | ||
1da177e4 LT |
2576 | #else |
2577 | #define check_irq_off() do { } while(0) | |
2578 | #define check_irq_on() do { } while(0) | |
2579 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2580 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2581 | #endif |
2582 | ||
aab2207c CL |
2583 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2584 | struct array_cache *ac, | |
2585 | int force, int node); | |
2586 | ||
1da177e4 LT |
2587 | static void do_drain(void *arg) |
2588 | { | |
a737b3e2 | 2589 | struct kmem_cache *cachep = arg; |
1da177e4 | 2590 | struct array_cache *ac; |
7d6e6d09 | 2591 | int node = numa_mem_id(); |
1da177e4 LT |
2592 | |
2593 | check_irq_off(); | |
9a2dba4b | 2594 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2595 | spin_lock(&cachep->nodelists[node]->list_lock); |
2596 | free_block(cachep, ac->entry, ac->avail, node); | |
2597 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2598 | ac->avail = 0; |
2599 | } | |
2600 | ||
343e0d7a | 2601 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2602 | { |
e498be7d CL |
2603 | struct kmem_list3 *l3; |
2604 | int node; | |
2605 | ||
15c8b6c1 | 2606 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2607 | check_irq_on(); |
b28a02de | 2608 | for_each_online_node(node) { |
e498be7d | 2609 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2610 | if (l3 && l3->alien) |
2611 | drain_alien_cache(cachep, l3->alien); | |
2612 | } | |
2613 | ||
2614 | for_each_online_node(node) { | |
2615 | l3 = cachep->nodelists[node]; | |
2616 | if (l3) | |
aab2207c | 2617 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2618 | } |
1da177e4 LT |
2619 | } |
2620 | ||
ed11d9eb CL |
2621 | /* |
2622 | * Remove slabs from the list of free slabs. | |
2623 | * Specify the number of slabs to drain in tofree. | |
2624 | * | |
2625 | * Returns the actual number of slabs released. | |
2626 | */ | |
2627 | static int drain_freelist(struct kmem_cache *cache, | |
2628 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2629 | { |
ed11d9eb CL |
2630 | struct list_head *p; |
2631 | int nr_freed; | |
1da177e4 | 2632 | struct slab *slabp; |
1da177e4 | 2633 | |
ed11d9eb CL |
2634 | nr_freed = 0; |
2635 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2636 | |
ed11d9eb | 2637 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2638 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2639 | if (p == &l3->slabs_free) { |
2640 | spin_unlock_irq(&l3->list_lock); | |
2641 | goto out; | |
2642 | } | |
1da177e4 | 2643 | |
ed11d9eb | 2644 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2645 | #if DEBUG |
40094fa6 | 2646 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2647 | #endif |
2648 | list_del(&slabp->list); | |
ed11d9eb CL |
2649 | /* |
2650 | * Safe to drop the lock. The slab is no longer linked | |
2651 | * to the cache. | |
2652 | */ | |
2653 | l3->free_objects -= cache->num; | |
e498be7d | 2654 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2655 | slab_destroy(cache, slabp); |
2656 | nr_freed++; | |
1da177e4 | 2657 | } |
ed11d9eb CL |
2658 | out: |
2659 | return nr_freed; | |
1da177e4 LT |
2660 | } |
2661 | ||
18004c5d | 2662 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2663 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2664 | { |
2665 | int ret = 0, i = 0; | |
2666 | struct kmem_list3 *l3; | |
2667 | ||
2668 | drain_cpu_caches(cachep); | |
2669 | ||
2670 | check_irq_on(); | |
2671 | for_each_online_node(i) { | |
2672 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2673 | if (!l3) |
2674 | continue; | |
2675 | ||
2676 | drain_freelist(cachep, l3, l3->free_objects); | |
2677 | ||
2678 | ret += !list_empty(&l3->slabs_full) || | |
2679 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2680 | } |
2681 | return (ret ? 1 : 0); | |
2682 | } | |
2683 | ||
1da177e4 LT |
2684 | /** |
2685 | * kmem_cache_shrink - Shrink a cache. | |
2686 | * @cachep: The cache to shrink. | |
2687 | * | |
2688 | * Releases as many slabs as possible for a cache. | |
2689 | * To help debugging, a zero exit status indicates all slabs were released. | |
2690 | */ | |
343e0d7a | 2691 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2692 | { |
8f5be20b | 2693 | int ret; |
40094fa6 | 2694 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2695 | |
95402b38 | 2696 | get_online_cpus(); |
18004c5d | 2697 | mutex_lock(&slab_mutex); |
8f5be20b | 2698 | ret = __cache_shrink(cachep); |
18004c5d | 2699 | mutex_unlock(&slab_mutex); |
95402b38 | 2700 | put_online_cpus(); |
8f5be20b | 2701 | return ret; |
1da177e4 LT |
2702 | } |
2703 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2704 | ||
945cf2b6 | 2705 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
1da177e4 | 2706 | { |
12c3667f CL |
2707 | int i; |
2708 | struct kmem_list3 *l3; | |
2709 | int rc = __cache_shrink(cachep); | |
1da177e4 | 2710 | |
12c3667f CL |
2711 | if (rc) |
2712 | return rc; | |
1da177e4 | 2713 | |
12c3667f CL |
2714 | for_each_online_cpu(i) |
2715 | kfree(cachep->array[i]); | |
1da177e4 | 2716 | |
12c3667f CL |
2717 | /* NUMA: free the list3 structures */ |
2718 | for_each_online_node(i) { | |
2719 | l3 = cachep->nodelists[i]; | |
2720 | if (l3) { | |
2721 | kfree(l3->shared); | |
2722 | free_alien_cache(l3->alien); | |
2723 | kfree(l3); | |
2724 | } | |
2725 | } | |
2726 | return 0; | |
1da177e4 | 2727 | } |
1da177e4 | 2728 | |
e5ac9c5a RT |
2729 | /* |
2730 | * Get the memory for a slab management obj. | |
2731 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2732 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2733 | * come from the same cache which is getting created because, | |
2734 | * when we are searching for an appropriate cache for these | |
2735 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2736 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2737 | * kmem_find_general_cachep till the initialization is complete. | |
2738 | * Hence we cannot have slabp_cache same as the original cache. | |
2739 | */ | |
343e0d7a | 2740 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2741 | int colour_off, gfp_t local_flags, |
2742 | int nodeid) | |
1da177e4 LT |
2743 | { |
2744 | struct slab *slabp; | |
b28a02de | 2745 | |
1da177e4 LT |
2746 | if (OFF_SLAB(cachep)) { |
2747 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2748 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2749 | local_flags, nodeid); |
d5cff635 CM |
2750 | /* |
2751 | * If the first object in the slab is leaked (it's allocated | |
2752 | * but no one has a reference to it), we want to make sure | |
2753 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2754 | * to the object. Otherwise we will not report the leak. | |
2755 | */ | |
c017b4be CM |
2756 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2757 | local_flags); | |
1da177e4 LT |
2758 | if (!slabp) |
2759 | return NULL; | |
2760 | } else { | |
b28a02de | 2761 | slabp = objp + colour_off; |
1da177e4 LT |
2762 | colour_off += cachep->slab_size; |
2763 | } | |
2764 | slabp->inuse = 0; | |
2765 | slabp->colouroff = colour_off; | |
b28a02de | 2766 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2767 | slabp->nodeid = nodeid; |
e51bfd0a | 2768 | slabp->free = 0; |
1da177e4 LT |
2769 | return slabp; |
2770 | } | |
2771 | ||
2772 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2773 | { | |
b28a02de | 2774 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2775 | } |
2776 | ||
343e0d7a | 2777 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2778 | struct slab *slabp) |
1da177e4 LT |
2779 | { |
2780 | int i; | |
2781 | ||
2782 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2783 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2784 | #if DEBUG |
2785 | /* need to poison the objs? */ | |
2786 | if (cachep->flags & SLAB_POISON) | |
2787 | poison_obj(cachep, objp, POISON_FREE); | |
2788 | if (cachep->flags & SLAB_STORE_USER) | |
2789 | *dbg_userword(cachep, objp) = NULL; | |
2790 | ||
2791 | if (cachep->flags & SLAB_RED_ZONE) { | |
2792 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2793 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2794 | } | |
2795 | /* | |
a737b3e2 AM |
2796 | * Constructors are not allowed to allocate memory from the same |
2797 | * cache which they are a constructor for. Otherwise, deadlock. | |
2798 | * They must also be threaded. | |
1da177e4 LT |
2799 | */ |
2800 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2801 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2802 | |
2803 | if (cachep->flags & SLAB_RED_ZONE) { | |
2804 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2805 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2806 | " end of an object"); |
1da177e4 LT |
2807 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2808 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2809 | " start of an object"); |
1da177e4 | 2810 | } |
3b0efdfa | 2811 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2812 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2813 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2814 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2815 | #else |
2816 | if (cachep->ctor) | |
51cc5068 | 2817 | cachep->ctor(objp); |
1da177e4 | 2818 | #endif |
b28a02de | 2819 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2820 | } |
b28a02de | 2821 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2822 | } |
2823 | ||
343e0d7a | 2824 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2825 | { |
4b51d669 CL |
2826 | if (CONFIG_ZONE_DMA_FLAG) { |
2827 | if (flags & GFP_DMA) | |
a618e89f | 2828 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2829 | else |
a618e89f | 2830 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2831 | } |
1da177e4 LT |
2832 | } |
2833 | ||
a737b3e2 AM |
2834 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2835 | int nodeid) | |
78d382d7 | 2836 | { |
8fea4e96 | 2837 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2838 | kmem_bufctl_t next; |
2839 | ||
2840 | slabp->inuse++; | |
2841 | next = slab_bufctl(slabp)[slabp->free]; | |
2842 | #if DEBUG | |
2843 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2844 | WARN_ON(slabp->nodeid != nodeid); | |
2845 | #endif | |
2846 | slabp->free = next; | |
2847 | ||
2848 | return objp; | |
2849 | } | |
2850 | ||
a737b3e2 AM |
2851 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2852 | void *objp, int nodeid) | |
78d382d7 | 2853 | { |
8fea4e96 | 2854 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2855 | |
2856 | #if DEBUG | |
2857 | /* Verify that the slab belongs to the intended node */ | |
2858 | WARN_ON(slabp->nodeid != nodeid); | |
2859 | ||
871751e2 | 2860 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2861 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2862 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2863 | BUG(); |
2864 | } | |
2865 | #endif | |
2866 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2867 | slabp->free = objnr; | |
2868 | slabp->inuse--; | |
2869 | } | |
2870 | ||
4776874f PE |
2871 | /* |
2872 | * Map pages beginning at addr to the given cache and slab. This is required | |
2873 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2874 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2875 | */ |
2876 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2877 | void *addr) | |
1da177e4 | 2878 | { |
4776874f | 2879 | int nr_pages; |
1da177e4 LT |
2880 | struct page *page; |
2881 | ||
4776874f | 2882 | page = virt_to_page(addr); |
84097518 | 2883 | |
4776874f | 2884 | nr_pages = 1; |
84097518 | 2885 | if (likely(!PageCompound(page))) |
4776874f PE |
2886 | nr_pages <<= cache->gfporder; |
2887 | ||
1da177e4 | 2888 | do { |
35026088 CL |
2889 | page->slab_cache = cache; |
2890 | page->slab_page = slab; | |
1da177e4 | 2891 | page++; |
4776874f | 2892 | } while (--nr_pages); |
1da177e4 LT |
2893 | } |
2894 | ||
2895 | /* | |
2896 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2897 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2898 | */ | |
3c517a61 CL |
2899 | static int cache_grow(struct kmem_cache *cachep, |
2900 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2901 | { |
b28a02de | 2902 | struct slab *slabp; |
b28a02de PE |
2903 | size_t offset; |
2904 | gfp_t local_flags; | |
e498be7d | 2905 | struct kmem_list3 *l3; |
1da177e4 | 2906 | |
a737b3e2 AM |
2907 | /* |
2908 | * Be lazy and only check for valid flags here, keeping it out of the | |
2909 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2910 | */ |
6cb06229 CL |
2911 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2912 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2913 | |
2e1217cf | 2914 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2915 | check_irq_off(); |
2e1217cf RT |
2916 | l3 = cachep->nodelists[nodeid]; |
2917 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2918 | |
2919 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2920 | offset = l3->colour_next; |
2921 | l3->colour_next++; | |
2922 | if (l3->colour_next >= cachep->colour) | |
2923 | l3->colour_next = 0; | |
2924 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2925 | |
2e1217cf | 2926 | offset *= cachep->colour_off; |
1da177e4 LT |
2927 | |
2928 | if (local_flags & __GFP_WAIT) | |
2929 | local_irq_enable(); | |
2930 | ||
2931 | /* | |
2932 | * The test for missing atomic flag is performed here, rather than | |
2933 | * the more obvious place, simply to reduce the critical path length | |
2934 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2935 | * will eventually be caught here (where it matters). | |
2936 | */ | |
2937 | kmem_flagcheck(cachep, flags); | |
2938 | ||
a737b3e2 AM |
2939 | /* |
2940 | * Get mem for the objs. Attempt to allocate a physical page from | |
2941 | * 'nodeid'. | |
e498be7d | 2942 | */ |
3c517a61 | 2943 | if (!objp) |
b8c1c5da | 2944 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2945 | if (!objp) |
1da177e4 LT |
2946 | goto failed; |
2947 | ||
2948 | /* Get slab management. */ | |
3c517a61 | 2949 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2950 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2951 | if (!slabp) |
1da177e4 LT |
2952 | goto opps1; |
2953 | ||
4776874f | 2954 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2955 | |
a35afb83 | 2956 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2957 | |
2958 | if (local_flags & __GFP_WAIT) | |
2959 | local_irq_disable(); | |
2960 | check_irq_off(); | |
e498be7d | 2961 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2962 | |
2963 | /* Make slab active. */ | |
e498be7d | 2964 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2965 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2966 | l3->free_objects += cachep->num; |
2967 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2968 | return 1; |
a737b3e2 | 2969 | opps1: |
1da177e4 | 2970 | kmem_freepages(cachep, objp); |
a737b3e2 | 2971 | failed: |
1da177e4 LT |
2972 | if (local_flags & __GFP_WAIT) |
2973 | local_irq_disable(); | |
2974 | return 0; | |
2975 | } | |
2976 | ||
2977 | #if DEBUG | |
2978 | ||
2979 | /* | |
2980 | * Perform extra freeing checks: | |
2981 | * - detect bad pointers. | |
2982 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2983 | */ |
2984 | static void kfree_debugcheck(const void *objp) | |
2985 | { | |
1da177e4 LT |
2986 | if (!virt_addr_valid(objp)) { |
2987 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2988 | (unsigned long)objp); |
2989 | BUG(); | |
1da177e4 | 2990 | } |
1da177e4 LT |
2991 | } |
2992 | ||
58ce1fd5 PE |
2993 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2994 | { | |
b46b8f19 | 2995 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2996 | |
2997 | redzone1 = *dbg_redzone1(cache, obj); | |
2998 | redzone2 = *dbg_redzone2(cache, obj); | |
2999 | ||
3000 | /* | |
3001 | * Redzone is ok. | |
3002 | */ | |
3003 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3004 | return; | |
3005 | ||
3006 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3007 | slab_error(cache, "double free detected"); | |
3008 | else | |
3009 | slab_error(cache, "memory outside object was overwritten"); | |
3010 | ||
b46b8f19 | 3011 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
3012 | obj, redzone1, redzone2); |
3013 | } | |
3014 | ||
343e0d7a | 3015 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 3016 | unsigned long caller) |
1da177e4 LT |
3017 | { |
3018 | struct page *page; | |
3019 | unsigned int objnr; | |
3020 | struct slab *slabp; | |
3021 | ||
80cbd911 MW |
3022 | BUG_ON(virt_to_cache(objp) != cachep); |
3023 | ||
3dafccf2 | 3024 | objp -= obj_offset(cachep); |
1da177e4 | 3025 | kfree_debugcheck(objp); |
b49af68f | 3026 | page = virt_to_head_page(objp); |
1da177e4 | 3027 | |
35026088 | 3028 | slabp = page->slab_page; |
1da177e4 LT |
3029 | |
3030 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3031 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3032 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3033 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3034 | } | |
3035 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 3036 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 | 3037 | |
8fea4e96 | 3038 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3039 | |
3040 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3041 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3042 | |
871751e2 AV |
3043 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3044 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3045 | #endif | |
1da177e4 LT |
3046 | if (cachep->flags & SLAB_POISON) { |
3047 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3048 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
7c0cb9c6 | 3049 | store_stackinfo(cachep, objp, caller); |
b28a02de | 3050 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3051 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3052 | } else { |
3053 | poison_obj(cachep, objp, POISON_FREE); | |
3054 | } | |
3055 | #else | |
3056 | poison_obj(cachep, objp, POISON_FREE); | |
3057 | #endif | |
3058 | } | |
3059 | return objp; | |
3060 | } | |
3061 | ||
343e0d7a | 3062 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3063 | { |
3064 | kmem_bufctl_t i; | |
3065 | int entries = 0; | |
b28a02de | 3066 | |
1da177e4 LT |
3067 | /* Check slab's freelist to see if this obj is there. */ |
3068 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3069 | entries++; | |
3070 | if (entries > cachep->num || i >= cachep->num) | |
3071 | goto bad; | |
3072 | } | |
3073 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3074 | bad: |
3075 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3076 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3077 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3078 | print_tainted()); | |
fdde6abb SAS |
3079 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3080 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3081 | 1); | |
1da177e4 LT |
3082 | BUG(); |
3083 | } | |
3084 | } | |
3085 | #else | |
3086 | #define kfree_debugcheck(x) do { } while(0) | |
3087 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3088 | #define check_slabp(x,y) do { } while(0) | |
3089 | #endif | |
3090 | ||
072bb0aa MG |
3091 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
3092 | bool force_refill) | |
1da177e4 LT |
3093 | { |
3094 | int batchcount; | |
3095 | struct kmem_list3 *l3; | |
3096 | struct array_cache *ac; | |
1ca4cb24 PE |
3097 | int node; |
3098 | ||
1da177e4 | 3099 | check_irq_off(); |
7d6e6d09 | 3100 | node = numa_mem_id(); |
072bb0aa MG |
3101 | if (unlikely(force_refill)) |
3102 | goto force_grow; | |
3103 | retry: | |
9a2dba4b | 3104 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3105 | batchcount = ac->batchcount; |
3106 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3107 | /* |
3108 | * If there was little recent activity on this cache, then | |
3109 | * perform only a partial refill. Otherwise we could generate | |
3110 | * refill bouncing. | |
1da177e4 LT |
3111 | */ |
3112 | batchcount = BATCHREFILL_LIMIT; | |
3113 | } | |
1ca4cb24 | 3114 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3115 | |
3116 | BUG_ON(ac->avail > 0 || !l3); | |
3117 | spin_lock(&l3->list_lock); | |
1da177e4 | 3118 | |
3ded175a | 3119 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3120 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3121 | l3->shared->touched = 1; | |
3ded175a | 3122 | goto alloc_done; |
44b57f1c | 3123 | } |
3ded175a | 3124 | |
1da177e4 LT |
3125 | while (batchcount > 0) { |
3126 | struct list_head *entry; | |
3127 | struct slab *slabp; | |
3128 | /* Get slab alloc is to come from. */ | |
3129 | entry = l3->slabs_partial.next; | |
3130 | if (entry == &l3->slabs_partial) { | |
3131 | l3->free_touched = 1; | |
3132 | entry = l3->slabs_free.next; | |
3133 | if (entry == &l3->slabs_free) | |
3134 | goto must_grow; | |
3135 | } | |
3136 | ||
3137 | slabp = list_entry(entry, struct slab, list); | |
3138 | check_slabp(cachep, slabp); | |
3139 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3140 | |
3141 | /* | |
3142 | * The slab was either on partial or free list so | |
3143 | * there must be at least one object available for | |
3144 | * allocation. | |
3145 | */ | |
249b9f33 | 3146 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3147 | |
1da177e4 | 3148 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3149 | STATS_INC_ALLOCED(cachep); |
3150 | STATS_INC_ACTIVE(cachep); | |
3151 | STATS_SET_HIGH(cachep); | |
3152 | ||
072bb0aa MG |
3153 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, |
3154 | node)); | |
1da177e4 LT |
3155 | } |
3156 | check_slabp(cachep, slabp); | |
3157 | ||
3158 | /* move slabp to correct slabp list: */ | |
3159 | list_del(&slabp->list); | |
3160 | if (slabp->free == BUFCTL_END) | |
3161 | list_add(&slabp->list, &l3->slabs_full); | |
3162 | else | |
3163 | list_add(&slabp->list, &l3->slabs_partial); | |
3164 | } | |
3165 | ||
a737b3e2 | 3166 | must_grow: |
1da177e4 | 3167 | l3->free_objects -= ac->avail; |
a737b3e2 | 3168 | alloc_done: |
e498be7d | 3169 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3170 | |
3171 | if (unlikely(!ac->avail)) { | |
3172 | int x; | |
072bb0aa | 3173 | force_grow: |
3c517a61 | 3174 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3175 | |
a737b3e2 | 3176 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3177 | ac = cpu_cache_get(cachep); |
51cd8e6f | 3178 | node = numa_mem_id(); |
072bb0aa MG |
3179 | |
3180 | /* no objects in sight? abort */ | |
3181 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
3182 | return NULL; |
3183 | ||
a737b3e2 | 3184 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3185 | goto retry; |
3186 | } | |
3187 | ac->touched = 1; | |
072bb0aa MG |
3188 | |
3189 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
3190 | } |
3191 | ||
a737b3e2 AM |
3192 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3193 | gfp_t flags) | |
1da177e4 LT |
3194 | { |
3195 | might_sleep_if(flags & __GFP_WAIT); | |
3196 | #if DEBUG | |
3197 | kmem_flagcheck(cachep, flags); | |
3198 | #endif | |
3199 | } | |
3200 | ||
3201 | #if DEBUG | |
a737b3e2 | 3202 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
7c0cb9c6 | 3203 | gfp_t flags, void *objp, unsigned long caller) |
1da177e4 | 3204 | { |
b28a02de | 3205 | if (!objp) |
1da177e4 | 3206 | return objp; |
b28a02de | 3207 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3208 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3209 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3210 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3211 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3212 | else |
3213 | check_poison_obj(cachep, objp); | |
3214 | #else | |
3215 | check_poison_obj(cachep, objp); | |
3216 | #endif | |
3217 | poison_obj(cachep, objp, POISON_INUSE); | |
3218 | } | |
3219 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 3220 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 LT |
3221 | |
3222 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3223 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3224 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3225 | slab_error(cachep, "double free, or memory outside" | |
3226 | " object was overwritten"); | |
b28a02de | 3227 | printk(KERN_ERR |
b46b8f19 | 3228 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3229 | objp, *dbg_redzone1(cachep, objp), |
3230 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3231 | } |
3232 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3233 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3234 | } | |
871751e2 AV |
3235 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3236 | { | |
3237 | struct slab *slabp; | |
3238 | unsigned objnr; | |
3239 | ||
35026088 | 3240 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3241 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3242 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3243 | } | |
3244 | #endif | |
3dafccf2 | 3245 | objp += obj_offset(cachep); |
4f104934 | 3246 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3247 | cachep->ctor(objp); |
7ea466f2 TH |
3248 | if (ARCH_SLAB_MINALIGN && |
3249 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3250 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3251 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3252 | } |
1da177e4 LT |
3253 | return objp; |
3254 | } | |
3255 | #else | |
3256 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3257 | #endif | |
3258 | ||
773ff60e | 3259 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 | 3260 | { |
9b030cb8 | 3261 | if (cachep == kmem_cache) |
773ff60e | 3262 | return false; |
8a8b6502 | 3263 | |
8c138bc0 | 3264 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3265 | } |
3266 | ||
343e0d7a | 3267 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3268 | { |
b28a02de | 3269 | void *objp; |
1da177e4 | 3270 | struct array_cache *ac; |
072bb0aa | 3271 | bool force_refill = false; |
1da177e4 | 3272 | |
5c382300 | 3273 | check_irq_off(); |
8a8b6502 | 3274 | |
9a2dba4b | 3275 | ac = cpu_cache_get(cachep); |
1da177e4 | 3276 | if (likely(ac->avail)) { |
1da177e4 | 3277 | ac->touched = 1; |
072bb0aa MG |
3278 | objp = ac_get_obj(cachep, ac, flags, false); |
3279 | ||
ddbf2e83 | 3280 | /* |
072bb0aa MG |
3281 | * Allow for the possibility all avail objects are not allowed |
3282 | * by the current flags | |
ddbf2e83 | 3283 | */ |
072bb0aa MG |
3284 | if (objp) { |
3285 | STATS_INC_ALLOCHIT(cachep); | |
3286 | goto out; | |
3287 | } | |
3288 | force_refill = true; | |
1da177e4 | 3289 | } |
072bb0aa MG |
3290 | |
3291 | STATS_INC_ALLOCMISS(cachep); | |
3292 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3293 | /* | |
3294 | * the 'ac' may be updated by cache_alloc_refill(), | |
3295 | * and kmemleak_erase() requires its correct value. | |
3296 | */ | |
3297 | ac = cpu_cache_get(cachep); | |
3298 | ||
3299 | out: | |
d5cff635 CM |
3300 | /* |
3301 | * To avoid a false negative, if an object that is in one of the | |
3302 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3303 | * treat the array pointers as a reference to the object. | |
3304 | */ | |
f3d8b53a O |
3305 | if (objp) |
3306 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3307 | return objp; |
3308 | } | |
3309 | ||
e498be7d | 3310 | #ifdef CONFIG_NUMA |
c61afb18 | 3311 | /* |
b2455396 | 3312 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3313 | * |
3314 | * If we are in_interrupt, then process context, including cpusets and | |
3315 | * mempolicy, may not apply and should not be used for allocation policy. | |
3316 | */ | |
3317 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3318 | { | |
3319 | int nid_alloc, nid_here; | |
3320 | ||
765c4507 | 3321 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3322 | return NULL; |
7d6e6d09 | 3323 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3324 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3325 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3326 | else if (current->mempolicy) |
e7b691b0 | 3327 | nid_alloc = slab_node(); |
c61afb18 | 3328 | if (nid_alloc != nid_here) |
8b98c169 | 3329 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3330 | return NULL; |
3331 | } | |
3332 | ||
765c4507 CL |
3333 | /* |
3334 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3335 | * certain node and fall back is permitted. First we scan all the |
3336 | * available nodelists for available objects. If that fails then we | |
3337 | * perform an allocation without specifying a node. This allows the page | |
3338 | * allocator to do its reclaim / fallback magic. We then insert the | |
3339 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3340 | */ |
8c8cc2c1 | 3341 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3342 | { |
8c8cc2c1 PE |
3343 | struct zonelist *zonelist; |
3344 | gfp_t local_flags; | |
dd1a239f | 3345 | struct zoneref *z; |
54a6eb5c MG |
3346 | struct zone *zone; |
3347 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3348 | void *obj = NULL; |
3c517a61 | 3349 | int nid; |
cc9a6c87 | 3350 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3351 | |
3352 | if (flags & __GFP_THISNODE) | |
3353 | return NULL; | |
3354 | ||
6cb06229 | 3355 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3356 | |
cc9a6c87 MG |
3357 | retry_cpuset: |
3358 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3359 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3360 | |
3c517a61 CL |
3361 | retry: |
3362 | /* | |
3363 | * Look through allowed nodes for objects available | |
3364 | * from existing per node queues. | |
3365 | */ | |
54a6eb5c MG |
3366 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3367 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3368 | |
54a6eb5c | 3369 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3370 | cache->nodelists[nid] && |
481c5346 | 3371 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3372 | obj = ____cache_alloc_node(cache, |
3373 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3374 | if (obj) |
3375 | break; | |
3376 | } | |
3c517a61 CL |
3377 | } |
3378 | ||
cfce6604 | 3379 | if (!obj) { |
3c517a61 CL |
3380 | /* |
3381 | * This allocation will be performed within the constraints | |
3382 | * of the current cpuset / memory policy requirements. | |
3383 | * We may trigger various forms of reclaim on the allowed | |
3384 | * set and go into memory reserves if necessary. | |
3385 | */ | |
dd47ea75 CL |
3386 | if (local_flags & __GFP_WAIT) |
3387 | local_irq_enable(); | |
3388 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3389 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3390 | if (local_flags & __GFP_WAIT) |
3391 | local_irq_disable(); | |
3c517a61 CL |
3392 | if (obj) { |
3393 | /* | |
3394 | * Insert into the appropriate per node queues | |
3395 | */ | |
3396 | nid = page_to_nid(virt_to_page(obj)); | |
3397 | if (cache_grow(cache, flags, nid, obj)) { | |
3398 | obj = ____cache_alloc_node(cache, | |
3399 | flags | GFP_THISNODE, nid); | |
3400 | if (!obj) | |
3401 | /* | |
3402 | * Another processor may allocate the | |
3403 | * objects in the slab since we are | |
3404 | * not holding any locks. | |
3405 | */ | |
3406 | goto retry; | |
3407 | } else { | |
b6a60451 | 3408 | /* cache_grow already freed obj */ |
3c517a61 CL |
3409 | obj = NULL; |
3410 | } | |
3411 | } | |
aedb0eb1 | 3412 | } |
cc9a6c87 MG |
3413 | |
3414 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3415 | goto retry_cpuset; | |
765c4507 CL |
3416 | return obj; |
3417 | } | |
3418 | ||
e498be7d CL |
3419 | /* |
3420 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3421 | */ |
8b98c169 | 3422 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3423 | int nodeid) |
e498be7d CL |
3424 | { |
3425 | struct list_head *entry; | |
b28a02de PE |
3426 | struct slab *slabp; |
3427 | struct kmem_list3 *l3; | |
3428 | void *obj; | |
b28a02de PE |
3429 | int x; |
3430 | ||
3431 | l3 = cachep->nodelists[nodeid]; | |
3432 | BUG_ON(!l3); | |
3433 | ||
a737b3e2 | 3434 | retry: |
ca3b9b91 | 3435 | check_irq_off(); |
b28a02de PE |
3436 | spin_lock(&l3->list_lock); |
3437 | entry = l3->slabs_partial.next; | |
3438 | if (entry == &l3->slabs_partial) { | |
3439 | l3->free_touched = 1; | |
3440 | entry = l3->slabs_free.next; | |
3441 | if (entry == &l3->slabs_free) | |
3442 | goto must_grow; | |
3443 | } | |
3444 | ||
3445 | slabp = list_entry(entry, struct slab, list); | |
3446 | check_spinlock_acquired_node(cachep, nodeid); | |
3447 | check_slabp(cachep, slabp); | |
3448 | ||
3449 | STATS_INC_NODEALLOCS(cachep); | |
3450 | STATS_INC_ACTIVE(cachep); | |
3451 | STATS_SET_HIGH(cachep); | |
3452 | ||
3453 | BUG_ON(slabp->inuse == cachep->num); | |
3454 | ||
78d382d7 | 3455 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3456 | check_slabp(cachep, slabp); |
3457 | l3->free_objects--; | |
3458 | /* move slabp to correct slabp list: */ | |
3459 | list_del(&slabp->list); | |
3460 | ||
a737b3e2 | 3461 | if (slabp->free == BUFCTL_END) |
b28a02de | 3462 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3463 | else |
b28a02de | 3464 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3465 | |
b28a02de PE |
3466 | spin_unlock(&l3->list_lock); |
3467 | goto done; | |
e498be7d | 3468 | |
a737b3e2 | 3469 | must_grow: |
b28a02de | 3470 | spin_unlock(&l3->list_lock); |
3c517a61 | 3471 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3472 | if (x) |
3473 | goto retry; | |
1da177e4 | 3474 | |
8c8cc2c1 | 3475 | return fallback_alloc(cachep, flags); |
e498be7d | 3476 | |
a737b3e2 | 3477 | done: |
b28a02de | 3478 | return obj; |
e498be7d | 3479 | } |
8c8cc2c1 PE |
3480 | |
3481 | /** | |
3482 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3483 | * @cachep: The cache to allocate from. | |
3484 | * @flags: See kmalloc(). | |
3485 | * @nodeid: node number of the target node. | |
3486 | * @caller: return address of caller, used for debug information | |
3487 | * | |
3488 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3489 | * node, which can improve the performance for cpu bound structures. | |
3490 | * | |
3491 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3492 | */ | |
3493 | static __always_inline void * | |
48356303 | 3494 | slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, |
7c0cb9c6 | 3495 | unsigned long caller) |
8c8cc2c1 PE |
3496 | { |
3497 | unsigned long save_flags; | |
3498 | void *ptr; | |
7d6e6d09 | 3499 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3500 | |
dcce284a | 3501 | flags &= gfp_allowed_mask; |
7e85ee0c | 3502 | |
cf40bd16 NP |
3503 | lockdep_trace_alloc(flags); |
3504 | ||
773ff60e | 3505 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3506 | return NULL; |
3507 | ||
8c8cc2c1 PE |
3508 | cache_alloc_debugcheck_before(cachep, flags); |
3509 | local_irq_save(save_flags); | |
3510 | ||
eacbbae3 | 3511 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3512 | nodeid = slab_node; |
8c8cc2c1 PE |
3513 | |
3514 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3515 | /* Node not bootstrapped yet */ | |
3516 | ptr = fallback_alloc(cachep, flags); | |
3517 | goto out; | |
3518 | } | |
3519 | ||
7d6e6d09 | 3520 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3521 | /* |
3522 | * Use the locally cached objects if possible. | |
3523 | * However ____cache_alloc does not allow fallback | |
3524 | * to other nodes. It may fail while we still have | |
3525 | * objects on other nodes available. | |
3526 | */ | |
3527 | ptr = ____cache_alloc(cachep, flags); | |
3528 | if (ptr) | |
3529 | goto out; | |
3530 | } | |
3531 | /* ___cache_alloc_node can fall back to other nodes */ | |
3532 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3533 | out: | |
3534 | local_irq_restore(save_flags); | |
3535 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3536 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3537 | flags); |
8c8cc2c1 | 3538 | |
c175eea4 | 3539 | if (likely(ptr)) |
8c138bc0 | 3540 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3541 | |
d07dbea4 | 3542 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3543 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3544 | |
8c8cc2c1 PE |
3545 | return ptr; |
3546 | } | |
3547 | ||
3548 | static __always_inline void * | |
3549 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3550 | { | |
3551 | void *objp; | |
3552 | ||
3553 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3554 | objp = alternate_node_alloc(cache, flags); | |
3555 | if (objp) | |
3556 | goto out; | |
3557 | } | |
3558 | objp = ____cache_alloc(cache, flags); | |
3559 | ||
3560 | /* | |
3561 | * We may just have run out of memory on the local node. | |
3562 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3563 | */ | |
7d6e6d09 LS |
3564 | if (!objp) |
3565 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3566 | |
3567 | out: | |
3568 | return objp; | |
3569 | } | |
3570 | #else | |
3571 | ||
3572 | static __always_inline void * | |
3573 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3574 | { | |
3575 | return ____cache_alloc(cachep, flags); | |
3576 | } | |
3577 | ||
3578 | #endif /* CONFIG_NUMA */ | |
3579 | ||
3580 | static __always_inline void * | |
48356303 | 3581 | slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) |
8c8cc2c1 PE |
3582 | { |
3583 | unsigned long save_flags; | |
3584 | void *objp; | |
3585 | ||
dcce284a | 3586 | flags &= gfp_allowed_mask; |
7e85ee0c | 3587 | |
cf40bd16 NP |
3588 | lockdep_trace_alloc(flags); |
3589 | ||
773ff60e | 3590 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3591 | return NULL; |
3592 | ||
8c8cc2c1 PE |
3593 | cache_alloc_debugcheck_before(cachep, flags); |
3594 | local_irq_save(save_flags); | |
3595 | objp = __do_cache_alloc(cachep, flags); | |
3596 | local_irq_restore(save_flags); | |
3597 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3598 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3599 | flags); |
8c8cc2c1 PE |
3600 | prefetchw(objp); |
3601 | ||
c175eea4 | 3602 | if (likely(objp)) |
8c138bc0 | 3603 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3604 | |
d07dbea4 | 3605 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3606 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3607 | |
8c8cc2c1 PE |
3608 | return objp; |
3609 | } | |
e498be7d CL |
3610 | |
3611 | /* | |
3612 | * Caller needs to acquire correct kmem_list's list_lock | |
3613 | */ | |
343e0d7a | 3614 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3615 | int node) |
1da177e4 LT |
3616 | { |
3617 | int i; | |
e498be7d | 3618 | struct kmem_list3 *l3; |
1da177e4 LT |
3619 | |
3620 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3621 | void *objp; |
1da177e4 | 3622 | struct slab *slabp; |
1da177e4 | 3623 | |
072bb0aa MG |
3624 | clear_obj_pfmemalloc(&objpp[i]); |
3625 | objp = objpp[i]; | |
3626 | ||
6ed5eb22 | 3627 | slabp = virt_to_slab(objp); |
ff69416e | 3628 | l3 = cachep->nodelists[node]; |
1da177e4 | 3629 | list_del(&slabp->list); |
ff69416e | 3630 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3631 | check_slabp(cachep, slabp); |
78d382d7 | 3632 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3633 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3634 | l3->free_objects++; |
1da177e4 LT |
3635 | check_slabp(cachep, slabp); |
3636 | ||
3637 | /* fixup slab chains */ | |
3638 | if (slabp->inuse == 0) { | |
e498be7d CL |
3639 | if (l3->free_objects > l3->free_limit) { |
3640 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3641 | /* No need to drop any previously held |
3642 | * lock here, even if we have a off-slab slab | |
3643 | * descriptor it is guaranteed to come from | |
3644 | * a different cache, refer to comments before | |
3645 | * alloc_slabmgmt. | |
3646 | */ | |
1da177e4 LT |
3647 | slab_destroy(cachep, slabp); |
3648 | } else { | |
e498be7d | 3649 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3650 | } |
3651 | } else { | |
3652 | /* Unconditionally move a slab to the end of the | |
3653 | * partial list on free - maximum time for the | |
3654 | * other objects to be freed, too. | |
3655 | */ | |
e498be7d | 3656 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3657 | } |
3658 | } | |
3659 | } | |
3660 | ||
343e0d7a | 3661 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3662 | { |
3663 | int batchcount; | |
e498be7d | 3664 | struct kmem_list3 *l3; |
7d6e6d09 | 3665 | int node = numa_mem_id(); |
1da177e4 LT |
3666 | |
3667 | batchcount = ac->batchcount; | |
3668 | #if DEBUG | |
3669 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3670 | #endif | |
3671 | check_irq_off(); | |
ff69416e | 3672 | l3 = cachep->nodelists[node]; |
873623df | 3673 | spin_lock(&l3->list_lock); |
e498be7d CL |
3674 | if (l3->shared) { |
3675 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3676 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3677 | if (max) { |
3678 | if (batchcount > max) | |
3679 | batchcount = max; | |
e498be7d | 3680 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3681 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3682 | shared_array->avail += batchcount; |
3683 | goto free_done; | |
3684 | } | |
3685 | } | |
3686 | ||
ff69416e | 3687 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3688 | free_done: |
1da177e4 LT |
3689 | #if STATS |
3690 | { | |
3691 | int i = 0; | |
3692 | struct list_head *p; | |
3693 | ||
e498be7d CL |
3694 | p = l3->slabs_free.next; |
3695 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3696 | struct slab *slabp; |
3697 | ||
3698 | slabp = list_entry(p, struct slab, list); | |
3699 | BUG_ON(slabp->inuse); | |
3700 | ||
3701 | i++; | |
3702 | p = p->next; | |
3703 | } | |
3704 | STATS_SET_FREEABLE(cachep, i); | |
3705 | } | |
3706 | #endif | |
e498be7d | 3707 | spin_unlock(&l3->list_lock); |
1da177e4 | 3708 | ac->avail -= batchcount; |
a737b3e2 | 3709 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3710 | } |
3711 | ||
3712 | /* | |
a737b3e2 AM |
3713 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3714 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3715 | */ |
a947eb95 | 3716 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 3717 | unsigned long caller) |
1da177e4 | 3718 | { |
9a2dba4b | 3719 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3720 | |
3721 | check_irq_off(); | |
d5cff635 | 3722 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3723 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3724 | |
8c138bc0 | 3725 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3726 | |
1807a1aa SS |
3727 | /* |
3728 | * Skip calling cache_free_alien() when the platform is not numa. | |
3729 | * This will avoid cache misses that happen while accessing slabp (which | |
3730 | * is per page memory reference) to get nodeid. Instead use a global | |
3731 | * variable to skip the call, which is mostly likely to be present in | |
3732 | * the cache. | |
3733 | */ | |
b6e68bc1 | 3734 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3735 | return; |
3736 | ||
1da177e4 LT |
3737 | if (likely(ac->avail < ac->limit)) { |
3738 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3739 | } else { |
3740 | STATS_INC_FREEMISS(cachep); | |
3741 | cache_flusharray(cachep, ac); | |
1da177e4 | 3742 | } |
42c8c99c | 3743 | |
072bb0aa | 3744 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3745 | } |
3746 | ||
3747 | /** | |
3748 | * kmem_cache_alloc - Allocate an object | |
3749 | * @cachep: The cache to allocate from. | |
3750 | * @flags: See kmalloc(). | |
3751 | * | |
3752 | * Allocate an object from this cache. The flags are only relevant | |
3753 | * if the cache has no available objects. | |
3754 | */ | |
343e0d7a | 3755 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3756 | { |
48356303 | 3757 | void *ret = slab_alloc(cachep, flags, _RET_IP_); |
36555751 | 3758 | |
ca2b84cb | 3759 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3760 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3761 | |
3762 | return ret; | |
1da177e4 LT |
3763 | } |
3764 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3765 | ||
0f24f128 | 3766 | #ifdef CONFIG_TRACING |
85beb586 | 3767 | void * |
4052147c | 3768 | kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) |
36555751 | 3769 | { |
85beb586 SR |
3770 | void *ret; |
3771 | ||
48356303 | 3772 | ret = slab_alloc(cachep, flags, _RET_IP_); |
85beb586 SR |
3773 | |
3774 | trace_kmalloc(_RET_IP_, ret, | |
ff4fcd01 | 3775 | size, cachep->size, flags); |
85beb586 | 3776 | return ret; |
36555751 | 3777 | } |
85beb586 | 3778 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3779 | #endif |
3780 | ||
1da177e4 | 3781 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3782 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3783 | { | |
48356303 | 3784 | void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
36555751 | 3785 | |
ca2b84cb | 3786 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3787 | cachep->object_size, cachep->size, |
ca2b84cb | 3788 | flags, nodeid); |
36555751 EGM |
3789 | |
3790 | return ret; | |
8b98c169 | 3791 | } |
1da177e4 LT |
3792 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3793 | ||
0f24f128 | 3794 | #ifdef CONFIG_TRACING |
4052147c | 3795 | void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, |
85beb586 | 3796 | gfp_t flags, |
4052147c EG |
3797 | int nodeid, |
3798 | size_t size) | |
36555751 | 3799 | { |
85beb586 SR |
3800 | void *ret; |
3801 | ||
592f4145 | 3802 | ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
7c0cb9c6 | 3803 | |
85beb586 | 3804 | trace_kmalloc_node(_RET_IP_, ret, |
ff4fcd01 | 3805 | size, cachep->size, |
85beb586 SR |
3806 | flags, nodeid); |
3807 | return ret; | |
36555751 | 3808 | } |
85beb586 | 3809 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3810 | #endif |
3811 | ||
8b98c169 | 3812 | static __always_inline void * |
7c0cb9c6 | 3813 | __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) |
97e2bde4 | 3814 | { |
343e0d7a | 3815 | struct kmem_cache *cachep; |
97e2bde4 MS |
3816 | |
3817 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3818 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3819 | return cachep; | |
4052147c | 3820 | return kmem_cache_alloc_node_trace(cachep, flags, node, size); |
97e2bde4 | 3821 | } |
8b98c169 | 3822 | |
0bb38a5c | 3823 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3824 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3825 | { | |
7c0cb9c6 | 3826 | return __do_kmalloc_node(size, flags, node, _RET_IP_); |
8b98c169 | 3827 | } |
dbe5e69d | 3828 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3829 | |
3830 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3831 | int node, unsigned long caller) |
8b98c169 | 3832 | { |
7c0cb9c6 | 3833 | return __do_kmalloc_node(size, flags, node, caller); |
8b98c169 CH |
3834 | } |
3835 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3836 | #else | |
3837 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3838 | { | |
7c0cb9c6 | 3839 | return __do_kmalloc_node(size, flags, node, 0); |
8b98c169 CH |
3840 | } |
3841 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3842 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3843 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3844 | |
3845 | /** | |
800590f5 | 3846 | * __do_kmalloc - allocate memory |
1da177e4 | 3847 | * @size: how many bytes of memory are required. |
800590f5 | 3848 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3849 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3850 | */ |
7fd6b141 | 3851 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
7c0cb9c6 | 3852 | unsigned long caller) |
1da177e4 | 3853 | { |
343e0d7a | 3854 | struct kmem_cache *cachep; |
36555751 | 3855 | void *ret; |
1da177e4 | 3856 | |
97e2bde4 MS |
3857 | /* If you want to save a few bytes .text space: replace |
3858 | * __ with kmem_. | |
3859 | * Then kmalloc uses the uninlined functions instead of the inline | |
3860 | * functions. | |
3861 | */ | |
3862 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3863 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3864 | return cachep; | |
48356303 | 3865 | ret = slab_alloc(cachep, flags, caller); |
36555751 | 3866 | |
7c0cb9c6 | 3867 | trace_kmalloc(caller, ret, |
3b0efdfa | 3868 | size, cachep->size, flags); |
36555751 EGM |
3869 | |
3870 | return ret; | |
7fd6b141 PE |
3871 | } |
3872 | ||
7fd6b141 | 3873 | |
0bb38a5c | 3874 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3875 | void *__kmalloc(size_t size, gfp_t flags) |
3876 | { | |
7c0cb9c6 | 3877 | return __do_kmalloc(size, flags, _RET_IP_); |
1da177e4 LT |
3878 | } |
3879 | EXPORT_SYMBOL(__kmalloc); | |
3880 | ||
ce71e27c | 3881 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3882 | { |
7c0cb9c6 | 3883 | return __do_kmalloc(size, flags, caller); |
7fd6b141 PE |
3884 | } |
3885 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3886 | |
3887 | #else | |
3888 | void *__kmalloc(size_t size, gfp_t flags) | |
3889 | { | |
7c0cb9c6 | 3890 | return __do_kmalloc(size, flags, 0); |
1d2c8eea CH |
3891 | } |
3892 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3893 | #endif |
3894 | ||
1da177e4 LT |
3895 | /** |
3896 | * kmem_cache_free - Deallocate an object | |
3897 | * @cachep: The cache the allocation was from. | |
3898 | * @objp: The previously allocated object. | |
3899 | * | |
3900 | * Free an object which was previously allocated from this | |
3901 | * cache. | |
3902 | */ | |
343e0d7a | 3903 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3904 | { |
3905 | unsigned long flags; | |
3906 | ||
3907 | local_irq_save(flags); | |
d97d476b | 3908 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3909 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3910 | debug_check_no_obj_freed(objp, cachep->object_size); |
7c0cb9c6 | 3911 | __cache_free(cachep, objp, _RET_IP_); |
1da177e4 | 3912 | local_irq_restore(flags); |
36555751 | 3913 | |
ca2b84cb | 3914 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3915 | } |
3916 | EXPORT_SYMBOL(kmem_cache_free); | |
3917 | ||
1da177e4 LT |
3918 | /** |
3919 | * kfree - free previously allocated memory | |
3920 | * @objp: pointer returned by kmalloc. | |
3921 | * | |
80e93eff PE |
3922 | * If @objp is NULL, no operation is performed. |
3923 | * | |
1da177e4 LT |
3924 | * Don't free memory not originally allocated by kmalloc() |
3925 | * or you will run into trouble. | |
3926 | */ | |
3927 | void kfree(const void *objp) | |
3928 | { | |
343e0d7a | 3929 | struct kmem_cache *c; |
1da177e4 LT |
3930 | unsigned long flags; |
3931 | ||
2121db74 PE |
3932 | trace_kfree(_RET_IP_, objp); |
3933 | ||
6cb8f913 | 3934 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3935 | return; |
3936 | local_irq_save(flags); | |
3937 | kfree_debugcheck(objp); | |
6ed5eb22 | 3938 | c = virt_to_cache(objp); |
8c138bc0 CL |
3939 | debug_check_no_locks_freed(objp, c->object_size); |
3940 | ||
3941 | debug_check_no_obj_freed(objp, c->object_size); | |
7c0cb9c6 | 3942 | __cache_free(c, (void *)objp, _RET_IP_); |
1da177e4 LT |
3943 | local_irq_restore(flags); |
3944 | } | |
3945 | EXPORT_SYMBOL(kfree); | |
3946 | ||
e498be7d | 3947 | /* |
183ff22b | 3948 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3949 | */ |
83b519e8 | 3950 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3951 | { |
3952 | int node; | |
3953 | struct kmem_list3 *l3; | |
cafeb02e | 3954 | struct array_cache *new_shared; |
3395ee05 | 3955 | struct array_cache **new_alien = NULL; |
e498be7d | 3956 | |
9c09a95c | 3957 | for_each_online_node(node) { |
cafeb02e | 3958 | |
3395ee05 | 3959 | if (use_alien_caches) { |
83b519e8 | 3960 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3961 | if (!new_alien) |
3962 | goto fail; | |
3963 | } | |
cafeb02e | 3964 | |
63109846 ED |
3965 | new_shared = NULL; |
3966 | if (cachep->shared) { | |
3967 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3968 | cachep->shared*cachep->batchcount, |
83b519e8 | 3969 | 0xbaadf00d, gfp); |
63109846 ED |
3970 | if (!new_shared) { |
3971 | free_alien_cache(new_alien); | |
3972 | goto fail; | |
3973 | } | |
0718dc2a | 3974 | } |
cafeb02e | 3975 | |
a737b3e2 AM |
3976 | l3 = cachep->nodelists[node]; |
3977 | if (l3) { | |
cafeb02e CL |
3978 | struct array_cache *shared = l3->shared; |
3979 | ||
e498be7d CL |
3980 | spin_lock_irq(&l3->list_lock); |
3981 | ||
cafeb02e | 3982 | if (shared) |
0718dc2a CL |
3983 | free_block(cachep, shared->entry, |
3984 | shared->avail, node); | |
e498be7d | 3985 | |
cafeb02e CL |
3986 | l3->shared = new_shared; |
3987 | if (!l3->alien) { | |
e498be7d CL |
3988 | l3->alien = new_alien; |
3989 | new_alien = NULL; | |
3990 | } | |
b28a02de | 3991 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3992 | cachep->batchcount + cachep->num; |
e498be7d | 3993 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3994 | kfree(shared); |
e498be7d CL |
3995 | free_alien_cache(new_alien); |
3996 | continue; | |
3997 | } | |
83b519e8 | 3998 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
3999 | if (!l3) { |
4000 | free_alien_cache(new_alien); | |
4001 | kfree(new_shared); | |
e498be7d | 4002 | goto fail; |
0718dc2a | 4003 | } |
e498be7d CL |
4004 | |
4005 | kmem_list3_init(l3); | |
4006 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 4007 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 4008 | l3->shared = new_shared; |
e498be7d | 4009 | l3->alien = new_alien; |
b28a02de | 4010 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4011 | cachep->batchcount + cachep->num; |
e498be7d CL |
4012 | cachep->nodelists[node] = l3; |
4013 | } | |
cafeb02e | 4014 | return 0; |
0718dc2a | 4015 | |
a737b3e2 | 4016 | fail: |
3b0efdfa | 4017 | if (!cachep->list.next) { |
0718dc2a CL |
4018 | /* Cache is not active yet. Roll back what we did */ |
4019 | node--; | |
4020 | while (node >= 0) { | |
4021 | if (cachep->nodelists[node]) { | |
4022 | l3 = cachep->nodelists[node]; | |
4023 | ||
4024 | kfree(l3->shared); | |
4025 | free_alien_cache(l3->alien); | |
4026 | kfree(l3); | |
4027 | cachep->nodelists[node] = NULL; | |
4028 | } | |
4029 | node--; | |
4030 | } | |
4031 | } | |
cafeb02e | 4032 | return -ENOMEM; |
e498be7d CL |
4033 | } |
4034 | ||
1da177e4 | 4035 | struct ccupdate_struct { |
343e0d7a | 4036 | struct kmem_cache *cachep; |
acfe7d74 | 4037 | struct array_cache *new[0]; |
1da177e4 LT |
4038 | }; |
4039 | ||
4040 | static void do_ccupdate_local(void *info) | |
4041 | { | |
a737b3e2 | 4042 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4043 | struct array_cache *old; |
4044 | ||
4045 | check_irq_off(); | |
9a2dba4b | 4046 | old = cpu_cache_get(new->cachep); |
e498be7d | 4047 | |
1da177e4 LT |
4048 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4049 | new->new[smp_processor_id()] = old; | |
4050 | } | |
4051 | ||
18004c5d | 4052 | /* Always called with the slab_mutex held */ |
a737b3e2 | 4053 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4054 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4055 | { |
d2e7b7d0 | 4056 | struct ccupdate_struct *new; |
2ed3a4ef | 4057 | int i; |
1da177e4 | 4058 | |
acfe7d74 ED |
4059 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4060 | gfp); | |
d2e7b7d0 SS |
4061 | if (!new) |
4062 | return -ENOMEM; | |
4063 | ||
e498be7d | 4064 | for_each_online_cpu(i) { |
7d6e6d09 | 4065 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4066 | batchcount, gfp); |
d2e7b7d0 | 4067 | if (!new->new[i]) { |
b28a02de | 4068 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4069 | kfree(new->new[i]); |
4070 | kfree(new); | |
e498be7d | 4071 | return -ENOMEM; |
1da177e4 LT |
4072 | } |
4073 | } | |
d2e7b7d0 | 4074 | new->cachep = cachep; |
1da177e4 | 4075 | |
15c8b6c1 | 4076 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4077 | |
1da177e4 | 4078 | check_irq_on(); |
1da177e4 LT |
4079 | cachep->batchcount = batchcount; |
4080 | cachep->limit = limit; | |
e498be7d | 4081 | cachep->shared = shared; |
1da177e4 | 4082 | |
e498be7d | 4083 | for_each_online_cpu(i) { |
d2e7b7d0 | 4084 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4085 | if (!ccold) |
4086 | continue; | |
7d6e6d09 LS |
4087 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4088 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4089 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4090 | kfree(ccold); |
4091 | } | |
d2e7b7d0 | 4092 | kfree(new); |
83b519e8 | 4093 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4094 | } |
4095 | ||
18004c5d | 4096 | /* Called with slab_mutex held always */ |
83b519e8 | 4097 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4098 | { |
4099 | int err; | |
4100 | int limit, shared; | |
4101 | ||
a737b3e2 AM |
4102 | /* |
4103 | * The head array serves three purposes: | |
1da177e4 LT |
4104 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4105 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4106 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4107 | * bufctl chains: array operations are cheaper. |
4108 | * The numbers are guessed, we should auto-tune as described by | |
4109 | * Bonwick. | |
4110 | */ | |
3b0efdfa | 4111 | if (cachep->size > 131072) |
1da177e4 | 4112 | limit = 1; |
3b0efdfa | 4113 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4114 | limit = 8; |
3b0efdfa | 4115 | else if (cachep->size > 1024) |
1da177e4 | 4116 | limit = 24; |
3b0efdfa | 4117 | else if (cachep->size > 256) |
1da177e4 LT |
4118 | limit = 54; |
4119 | else | |
4120 | limit = 120; | |
4121 | ||
a737b3e2 AM |
4122 | /* |
4123 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4124 | * allocation behaviour: Most allocs on one cpu, most free operations |
4125 | * on another cpu. For these cases, an efficient object passing between | |
4126 | * cpus is necessary. This is provided by a shared array. The array | |
4127 | * replaces Bonwick's magazine layer. | |
4128 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4129 | * to a larger limit. Thus disabled by default. | |
4130 | */ | |
4131 | shared = 0; | |
3b0efdfa | 4132 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4133 | shared = 8; |
1da177e4 LT |
4134 | |
4135 | #if DEBUG | |
a737b3e2 AM |
4136 | /* |
4137 | * With debugging enabled, large batchcount lead to excessively long | |
4138 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4139 | */ |
4140 | if (limit > 32) | |
4141 | limit = 32; | |
4142 | #endif | |
83b519e8 | 4143 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4144 | if (err) |
4145 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4146 | cachep->name, -err); |
2ed3a4ef | 4147 | return err; |
1da177e4 LT |
4148 | } |
4149 | ||
1b55253a CL |
4150 | /* |
4151 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4152 | * necessary. Note that the l3 listlock also protects the array_cache |
4153 | * if drain_array() is used on the shared array. | |
1b55253a | 4154 | */ |
68a1b195 | 4155 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4156 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4157 | { |
4158 | int tofree; | |
4159 | ||
1b55253a CL |
4160 | if (!ac || !ac->avail) |
4161 | return; | |
1da177e4 LT |
4162 | if (ac->touched && !force) { |
4163 | ac->touched = 0; | |
b18e7e65 | 4164 | } else { |
1b55253a | 4165 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4166 | if (ac->avail) { |
4167 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4168 | if (tofree > ac->avail) | |
4169 | tofree = (ac->avail + 1) / 2; | |
4170 | free_block(cachep, ac->entry, tofree, node); | |
4171 | ac->avail -= tofree; | |
4172 | memmove(ac->entry, &(ac->entry[tofree]), | |
4173 | sizeof(void *) * ac->avail); | |
4174 | } | |
1b55253a | 4175 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4176 | } |
4177 | } | |
4178 | ||
4179 | /** | |
4180 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4181 | * @w: work descriptor |
1da177e4 LT |
4182 | * |
4183 | * Called from workqueue/eventd every few seconds. | |
4184 | * Purpose: | |
4185 | * - clear the per-cpu caches for this CPU. | |
4186 | * - return freeable pages to the main free memory pool. | |
4187 | * | |
a737b3e2 AM |
4188 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4189 | * again on the next iteration. | |
1da177e4 | 4190 | */ |
7c5cae36 | 4191 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4192 | { |
7a7c381d | 4193 | struct kmem_cache *searchp; |
e498be7d | 4194 | struct kmem_list3 *l3; |
7d6e6d09 | 4195 | int node = numa_mem_id(); |
bf6aede7 | 4196 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4197 | |
18004c5d | 4198 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 4199 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4200 | goto out; |
1da177e4 | 4201 | |
18004c5d | 4202 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
4203 | check_irq_on(); |
4204 | ||
35386e3b CL |
4205 | /* |
4206 | * We only take the l3 lock if absolutely necessary and we | |
4207 | * have established with reasonable certainty that | |
4208 | * we can do some work if the lock was obtained. | |
4209 | */ | |
aab2207c | 4210 | l3 = searchp->nodelists[node]; |
35386e3b | 4211 | |
8fce4d8e | 4212 | reap_alien(searchp, l3); |
1da177e4 | 4213 | |
aab2207c | 4214 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4215 | |
35386e3b CL |
4216 | /* |
4217 | * These are racy checks but it does not matter | |
4218 | * if we skip one check or scan twice. | |
4219 | */ | |
e498be7d | 4220 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4221 | goto next; |
1da177e4 | 4222 | |
e498be7d | 4223 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4224 | |
aab2207c | 4225 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4226 | |
ed11d9eb | 4227 | if (l3->free_touched) |
e498be7d | 4228 | l3->free_touched = 0; |
ed11d9eb CL |
4229 | else { |
4230 | int freed; | |
1da177e4 | 4231 | |
ed11d9eb CL |
4232 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4233 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4234 | STATS_ADD_REAPED(searchp, freed); | |
4235 | } | |
35386e3b | 4236 | next: |
1da177e4 LT |
4237 | cond_resched(); |
4238 | } | |
4239 | check_irq_on(); | |
18004c5d | 4240 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4241 | next_reap_node(); |
7c5cae36 | 4242 | out: |
a737b3e2 | 4243 | /* Set up the next iteration */ |
7c5cae36 | 4244 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4245 | } |
4246 | ||
158a9624 | 4247 | #ifdef CONFIG_SLABINFO |
0d7561c6 | 4248 | void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) |
1da177e4 | 4249 | { |
b28a02de PE |
4250 | struct slab *slabp; |
4251 | unsigned long active_objs; | |
4252 | unsigned long num_objs; | |
4253 | unsigned long active_slabs = 0; | |
4254 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4255 | const char *name; |
1da177e4 | 4256 | char *error = NULL; |
e498be7d CL |
4257 | int node; |
4258 | struct kmem_list3 *l3; | |
1da177e4 | 4259 | |
1da177e4 LT |
4260 | active_objs = 0; |
4261 | num_slabs = 0; | |
e498be7d CL |
4262 | for_each_online_node(node) { |
4263 | l3 = cachep->nodelists[node]; | |
4264 | if (!l3) | |
4265 | continue; | |
4266 | ||
ca3b9b91 RT |
4267 | check_irq_on(); |
4268 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4269 | |
7a7c381d | 4270 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4271 | if (slabp->inuse != cachep->num && !error) |
4272 | error = "slabs_full accounting error"; | |
4273 | active_objs += cachep->num; | |
4274 | active_slabs++; | |
4275 | } | |
7a7c381d | 4276 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4277 | if (slabp->inuse == cachep->num && !error) |
4278 | error = "slabs_partial inuse accounting error"; | |
4279 | if (!slabp->inuse && !error) | |
4280 | error = "slabs_partial/inuse accounting error"; | |
4281 | active_objs += slabp->inuse; | |
4282 | active_slabs++; | |
4283 | } | |
7a7c381d | 4284 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4285 | if (slabp->inuse && !error) |
4286 | error = "slabs_free/inuse accounting error"; | |
4287 | num_slabs++; | |
4288 | } | |
4289 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4290 | if (l3->shared) |
4291 | shared_avail += l3->shared->avail; | |
e498be7d | 4292 | |
ca3b9b91 | 4293 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4294 | } |
b28a02de PE |
4295 | num_slabs += active_slabs; |
4296 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4297 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4298 | error = "free_objects accounting error"; |
4299 | ||
b28a02de | 4300 | name = cachep->name; |
1da177e4 LT |
4301 | if (error) |
4302 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4303 | ||
0d7561c6 GC |
4304 | sinfo->active_objs = active_objs; |
4305 | sinfo->num_objs = num_objs; | |
4306 | sinfo->active_slabs = active_slabs; | |
4307 | sinfo->num_slabs = num_slabs; | |
4308 | sinfo->shared_avail = shared_avail; | |
4309 | sinfo->limit = cachep->limit; | |
4310 | sinfo->batchcount = cachep->batchcount; | |
4311 | sinfo->shared = cachep->shared; | |
4312 | sinfo->objects_per_slab = cachep->num; | |
4313 | sinfo->cache_order = cachep->gfporder; | |
4314 | } | |
4315 | ||
4316 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) | |
4317 | { | |
1da177e4 | 4318 | #if STATS |
b28a02de | 4319 | { /* list3 stats */ |
1da177e4 LT |
4320 | unsigned long high = cachep->high_mark; |
4321 | unsigned long allocs = cachep->num_allocations; | |
4322 | unsigned long grown = cachep->grown; | |
4323 | unsigned long reaped = cachep->reaped; | |
4324 | unsigned long errors = cachep->errors; | |
4325 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4326 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4327 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4328 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4329 | |
e92dd4fd JP |
4330 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4331 | "%4lu %4lu %4lu %4lu %4lu", | |
4332 | allocs, high, grown, | |
4333 | reaped, errors, max_freeable, node_allocs, | |
4334 | node_frees, overflows); | |
1da177e4 LT |
4335 | } |
4336 | /* cpu stats */ | |
4337 | { | |
4338 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4339 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4340 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4341 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4342 | ||
4343 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4344 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4345 | } |
4346 | #endif | |
1da177e4 LT |
4347 | } |
4348 | ||
1da177e4 LT |
4349 | #define MAX_SLABINFO_WRITE 128 |
4350 | /** | |
4351 | * slabinfo_write - Tuning for the slab allocator | |
4352 | * @file: unused | |
4353 | * @buffer: user buffer | |
4354 | * @count: data length | |
4355 | * @ppos: unused | |
4356 | */ | |
b7454ad3 | 4357 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4358 | size_t count, loff_t *ppos) |
1da177e4 | 4359 | { |
b28a02de | 4360 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4361 | int limit, batchcount, shared, res; |
7a7c381d | 4362 | struct kmem_cache *cachep; |
b28a02de | 4363 | |
1da177e4 LT |
4364 | if (count > MAX_SLABINFO_WRITE) |
4365 | return -EINVAL; | |
4366 | if (copy_from_user(&kbuf, buffer, count)) | |
4367 | return -EFAULT; | |
b28a02de | 4368 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4369 | |
4370 | tmp = strchr(kbuf, ' '); | |
4371 | if (!tmp) | |
4372 | return -EINVAL; | |
4373 | *tmp = '\0'; | |
4374 | tmp++; | |
4375 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4376 | return -EINVAL; | |
4377 | ||
4378 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4379 | mutex_lock(&slab_mutex); |
1da177e4 | 4380 | res = -EINVAL; |
18004c5d | 4381 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4382 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4383 | if (limit < 1 || batchcount < 1 || |
4384 | batchcount > limit || shared < 0) { | |
e498be7d | 4385 | res = 0; |
1da177e4 | 4386 | } else { |
e498be7d | 4387 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4388 | batchcount, shared, |
4389 | GFP_KERNEL); | |
1da177e4 LT |
4390 | } |
4391 | break; | |
4392 | } | |
4393 | } | |
18004c5d | 4394 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4395 | if (res >= 0) |
4396 | res = count; | |
4397 | return res; | |
4398 | } | |
871751e2 AV |
4399 | |
4400 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4401 | ||
4402 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4403 | { | |
18004c5d CL |
4404 | mutex_lock(&slab_mutex); |
4405 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4406 | } |
4407 | ||
4408 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4409 | { | |
4410 | unsigned long *p; | |
4411 | int l; | |
4412 | if (!v) | |
4413 | return 1; | |
4414 | l = n[1]; | |
4415 | p = n + 2; | |
4416 | while (l) { | |
4417 | int i = l/2; | |
4418 | unsigned long *q = p + 2 * i; | |
4419 | if (*q == v) { | |
4420 | q[1]++; | |
4421 | return 1; | |
4422 | } | |
4423 | if (*q > v) { | |
4424 | l = i; | |
4425 | } else { | |
4426 | p = q + 2; | |
4427 | l -= i + 1; | |
4428 | } | |
4429 | } | |
4430 | if (++n[1] == n[0]) | |
4431 | return 0; | |
4432 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4433 | p[0] = v; | |
4434 | p[1] = 1; | |
4435 | return 1; | |
4436 | } | |
4437 | ||
4438 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4439 | { | |
4440 | void *p; | |
4441 | int i; | |
4442 | if (n[0] == n[1]) | |
4443 | return; | |
3b0efdfa | 4444 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4445 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4446 | continue; | |
4447 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4448 | return; | |
4449 | } | |
4450 | } | |
4451 | ||
4452 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4453 | { | |
4454 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4455 | unsigned long offset, size; |
9281acea | 4456 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4457 | |
a5c43dae | 4458 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4459 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4460 | if (modname[0]) |
871751e2 AV |
4461 | seq_printf(m, " [%s]", modname); |
4462 | return; | |
4463 | } | |
4464 | #endif | |
4465 | seq_printf(m, "%p", (void *)address); | |
4466 | } | |
4467 | ||
4468 | static int leaks_show(struct seq_file *m, void *p) | |
4469 | { | |
0672aa7c | 4470 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 AV |
4471 | struct slab *slabp; |
4472 | struct kmem_list3 *l3; | |
4473 | const char *name; | |
4474 | unsigned long *n = m->private; | |
4475 | int node; | |
4476 | int i; | |
4477 | ||
4478 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4479 | return 0; | |
4480 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4481 | return 0; | |
4482 | ||
4483 | /* OK, we can do it */ | |
4484 | ||
4485 | n[1] = 0; | |
4486 | ||
4487 | for_each_online_node(node) { | |
4488 | l3 = cachep->nodelists[node]; | |
4489 | if (!l3) | |
4490 | continue; | |
4491 | ||
4492 | check_irq_on(); | |
4493 | spin_lock_irq(&l3->list_lock); | |
4494 | ||
7a7c381d | 4495 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4496 | handle_slab(n, cachep, slabp); |
7a7c381d | 4497 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4498 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4499 | spin_unlock_irq(&l3->list_lock); |
4500 | } | |
4501 | name = cachep->name; | |
4502 | if (n[0] == n[1]) { | |
4503 | /* Increase the buffer size */ | |
18004c5d | 4504 | mutex_unlock(&slab_mutex); |
871751e2 AV |
4505 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4506 | if (!m->private) { | |
4507 | /* Too bad, we are really out */ | |
4508 | m->private = n; | |
18004c5d | 4509 | mutex_lock(&slab_mutex); |
871751e2 AV |
4510 | return -ENOMEM; |
4511 | } | |
4512 | *(unsigned long *)m->private = n[0] * 2; | |
4513 | kfree(n); | |
18004c5d | 4514 | mutex_lock(&slab_mutex); |
871751e2 AV |
4515 | /* Now make sure this entry will be retried */ |
4516 | m->count = m->size; | |
4517 | return 0; | |
4518 | } | |
4519 | for (i = 0; i < n[1]; i++) { | |
4520 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4521 | show_symbol(m, n[2*i+2]); | |
4522 | seq_putc(m, '\n'); | |
4523 | } | |
d2e7b7d0 | 4524 | |
871751e2 AV |
4525 | return 0; |
4526 | } | |
4527 | ||
b7454ad3 GC |
4528 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) |
4529 | { | |
4530 | return seq_list_next(p, &slab_caches, pos); | |
4531 | } | |
4532 | ||
4533 | static void s_stop(struct seq_file *m, void *p) | |
4534 | { | |
4535 | mutex_unlock(&slab_mutex); | |
4536 | } | |
4537 | ||
a0ec95a8 | 4538 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4539 | .start = leaks_start, |
4540 | .next = s_next, | |
4541 | .stop = s_stop, | |
4542 | .show = leaks_show, | |
4543 | }; | |
a0ec95a8 AD |
4544 | |
4545 | static int slabstats_open(struct inode *inode, struct file *file) | |
4546 | { | |
4547 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4548 | int ret = -ENOMEM; | |
4549 | if (n) { | |
4550 | ret = seq_open(file, &slabstats_op); | |
4551 | if (!ret) { | |
4552 | struct seq_file *m = file->private_data; | |
4553 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4554 | m->private = n; | |
4555 | n = NULL; | |
4556 | } | |
4557 | kfree(n); | |
4558 | } | |
4559 | return ret; | |
4560 | } | |
4561 | ||
4562 | static const struct file_operations proc_slabstats_operations = { | |
4563 | .open = slabstats_open, | |
4564 | .read = seq_read, | |
4565 | .llseek = seq_lseek, | |
4566 | .release = seq_release_private, | |
4567 | }; | |
4568 | #endif | |
4569 | ||
4570 | static int __init slab_proc_init(void) | |
4571 | { | |
4572 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4573 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4574 | #endif |
a0ec95a8 AD |
4575 | return 0; |
4576 | } | |
4577 | module_init(slab_proc_init); | |
1da177e4 LT |
4578 | #endif |
4579 | ||
00e145b6 MS |
4580 | /** |
4581 | * ksize - get the actual amount of memory allocated for a given object | |
4582 | * @objp: Pointer to the object | |
4583 | * | |
4584 | * kmalloc may internally round up allocations and return more memory | |
4585 | * than requested. ksize() can be used to determine the actual amount of | |
4586 | * memory allocated. The caller may use this additional memory, even though | |
4587 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4588 | * The caller must guarantee that objp points to a valid object previously | |
4589 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4590 | * must not be freed during the duration of the call. | |
4591 | */ | |
fd76bab2 | 4592 | size_t ksize(const void *objp) |
1da177e4 | 4593 | { |
ef8b4520 CL |
4594 | BUG_ON(!objp); |
4595 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4596 | return 0; |
1da177e4 | 4597 | |
8c138bc0 | 4598 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4599 | } |
b1aabecd | 4600 | EXPORT_SYMBOL(ksize); |