]>
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 | |
1da177e4 LT |
121 | #include <asm/cacheflush.h> |
122 | #include <asm/tlbflush.h> | |
123 | #include <asm/page.h> | |
124 | ||
4dee6b64 SR |
125 | #include <trace/events/kmem.h> |
126 | ||
072bb0aa MG |
127 | #include "internal.h" |
128 | ||
1da177e4 | 129 | /* |
50953fe9 | 130 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
131 | * 0 for faster, smaller code (especially in the critical paths). |
132 | * | |
133 | * STATS - 1 to collect stats for /proc/slabinfo. | |
134 | * 0 for faster, smaller code (especially in the critical paths). | |
135 | * | |
136 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
137 | */ | |
138 | ||
139 | #ifdef CONFIG_DEBUG_SLAB | |
140 | #define DEBUG 1 | |
141 | #define STATS 1 | |
142 | #define FORCED_DEBUG 1 | |
143 | #else | |
144 | #define DEBUG 0 | |
145 | #define STATS 0 | |
146 | #define FORCED_DEBUG 0 | |
147 | #endif | |
148 | ||
1da177e4 LT |
149 | /* Shouldn't this be in a header file somewhere? */ |
150 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 151 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 152 | |
1da177e4 LT |
153 | #ifndef ARCH_KMALLOC_FLAGS |
154 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
155 | #endif | |
156 | ||
072bb0aa MG |
157 | /* |
158 | * true if a page was allocated from pfmemalloc reserves for network-based | |
159 | * swap | |
160 | */ | |
161 | static bool pfmemalloc_active __read_mostly; | |
162 | ||
1da177e4 LT |
163 | /* Legal flag mask for kmem_cache_create(). */ |
164 | #if DEBUG | |
50953fe9 | 165 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 166 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 167 | SLAB_CACHE_DMA | \ |
5af60839 | 168 | SLAB_STORE_USER | \ |
1da177e4 | 169 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 170 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 171 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 | 172 | #else |
ac2b898c | 173 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 174 | SLAB_CACHE_DMA | \ |
1da177e4 | 175 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 176 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 177 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 LT |
178 | #endif |
179 | ||
180 | /* | |
181 | * kmem_bufctl_t: | |
182 | * | |
183 | * Bufctl's are used for linking objs within a slab | |
184 | * linked offsets. | |
185 | * | |
186 | * This implementation relies on "struct page" for locating the cache & | |
187 | * slab an object belongs to. | |
188 | * This allows the bufctl structure to be small (one int), but limits | |
189 | * the number of objects a slab (not a cache) can contain when off-slab | |
190 | * bufctls are used. The limit is the size of the largest general cache | |
191 | * that does not use off-slab slabs. | |
192 | * For 32bit archs with 4 kB pages, is this 56. | |
193 | * This is not serious, as it is only for large objects, when it is unwise | |
194 | * to have too many per slab. | |
195 | * Note: This limit can be raised by introducing a general cache whose size | |
196 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
197 | */ | |
198 | ||
fa5b08d5 | 199 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
200 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
201 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
202 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
203 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 204 | |
1da177e4 LT |
205 | /* |
206 | * struct slab_rcu | |
207 | * | |
208 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
209 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
210 | * we need to approach a kernel structure obliquely, from its address | |
211 | * obtained without the usual locking. We can lock the structure to | |
212 | * stabilize it and check it's still at the given address, only if we | |
213 | * can be sure that the memory has not been meanwhile reused for some | |
214 | * other kind of object (which our subsystem's lock might corrupt). | |
215 | * | |
216 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
217 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
218 | */ |
219 | struct slab_rcu { | |
b28a02de | 220 | struct rcu_head head; |
343e0d7a | 221 | struct kmem_cache *cachep; |
b28a02de | 222 | void *addr; |
1da177e4 LT |
223 | }; |
224 | ||
5bfe53a7 LJ |
225 | /* |
226 | * struct slab | |
227 | * | |
228 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
229 | * for a slab, or allocated from an general cache. | |
230 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
231 | */ | |
232 | struct slab { | |
233 | union { | |
234 | struct { | |
235 | struct list_head list; | |
236 | unsigned long colouroff; | |
237 | void *s_mem; /* including colour offset */ | |
238 | unsigned int inuse; /* num of objs active in slab */ | |
239 | kmem_bufctl_t free; | |
240 | unsigned short nodeid; | |
241 | }; | |
242 | struct slab_rcu __slab_cover_slab_rcu; | |
243 | }; | |
244 | }; | |
245 | ||
1da177e4 LT |
246 | /* |
247 | * struct array_cache | |
248 | * | |
1da177e4 LT |
249 | * Purpose: |
250 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
251 | * - reduce the number of linked list operations | |
252 | * - reduce spinlock operations | |
253 | * | |
254 | * The limit is stored in the per-cpu structure to reduce the data cache | |
255 | * footprint. | |
256 | * | |
257 | */ | |
258 | struct array_cache { | |
259 | unsigned int avail; | |
260 | unsigned int limit; | |
261 | unsigned int batchcount; | |
262 | unsigned int touched; | |
e498be7d | 263 | spinlock_t lock; |
bda5b655 | 264 | void *entry[]; /* |
a737b3e2 AM |
265 | * Must have this definition in here for the proper |
266 | * alignment of array_cache. Also simplifies accessing | |
267 | * the entries. | |
072bb0aa MG |
268 | * |
269 | * Entries should not be directly dereferenced as | |
270 | * entries belonging to slabs marked pfmemalloc will | |
271 | * have the lower bits set SLAB_OBJ_PFMEMALLOC | |
a737b3e2 | 272 | */ |
1da177e4 LT |
273 | }; |
274 | ||
072bb0aa MG |
275 | #define SLAB_OBJ_PFMEMALLOC 1 |
276 | static inline bool is_obj_pfmemalloc(void *objp) | |
277 | { | |
278 | return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC; | |
279 | } | |
280 | ||
281 | static inline void set_obj_pfmemalloc(void **objp) | |
282 | { | |
283 | *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC); | |
284 | return; | |
285 | } | |
286 | ||
287 | static inline void clear_obj_pfmemalloc(void **objp) | |
288 | { | |
289 | *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC); | |
290 | } | |
291 | ||
a737b3e2 AM |
292 | /* |
293 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
294 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
295 | */ |
296 | #define BOOT_CPUCACHE_ENTRIES 1 | |
297 | struct arraycache_init { | |
298 | struct array_cache cache; | |
b28a02de | 299 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
300 | }; |
301 | ||
302 | /* | |
e498be7d | 303 | * The slab lists for all objects. |
1da177e4 LT |
304 | */ |
305 | struct kmem_list3 { | |
b28a02de PE |
306 | struct list_head slabs_partial; /* partial list first, better asm code */ |
307 | struct list_head slabs_full; | |
308 | struct list_head slabs_free; | |
309 | unsigned long free_objects; | |
b28a02de | 310 | unsigned int free_limit; |
2e1217cf | 311 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
312 | spinlock_t list_lock; |
313 | struct array_cache *shared; /* shared per node */ | |
314 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
315 | unsigned long next_reap; /* updated without locking */ |
316 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
317 | }; |
318 | ||
e498be7d CL |
319 | /* |
320 | * Need this for bootstrapping a per node allocator. | |
321 | */ | |
556a169d | 322 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
68a1b195 | 323 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 324 | #define CACHE_CACHE 0 |
556a169d PE |
325 | #define SIZE_AC MAX_NUMNODES |
326 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 327 | |
ed11d9eb CL |
328 | static int drain_freelist(struct kmem_cache *cache, |
329 | struct kmem_list3 *l3, int tofree); | |
330 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
331 | int node); | |
83b519e8 | 332 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 333 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 334 | |
e498be7d | 335 | /* |
a737b3e2 AM |
336 | * This function must be completely optimized away if a constant is passed to |
337 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 338 | */ |
7243cc05 | 339 | static __always_inline int index_of(const size_t size) |
e498be7d | 340 | { |
5ec8a847 SR |
341 | extern void __bad_size(void); |
342 | ||
e498be7d CL |
343 | if (__builtin_constant_p(size)) { |
344 | int i = 0; | |
345 | ||
346 | #define CACHE(x) \ | |
347 | if (size <=x) \ | |
348 | return i; \ | |
349 | else \ | |
350 | i++; | |
1c61fc40 | 351 | #include <linux/kmalloc_sizes.h> |
e498be7d | 352 | #undef CACHE |
5ec8a847 | 353 | __bad_size(); |
7243cc05 | 354 | } else |
5ec8a847 | 355 | __bad_size(); |
e498be7d CL |
356 | return 0; |
357 | } | |
358 | ||
e0a42726 IM |
359 | static int slab_early_init = 1; |
360 | ||
e498be7d CL |
361 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
362 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 363 | |
5295a74c | 364 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
365 | { |
366 | INIT_LIST_HEAD(&parent->slabs_full); | |
367 | INIT_LIST_HEAD(&parent->slabs_partial); | |
368 | INIT_LIST_HEAD(&parent->slabs_free); | |
369 | parent->shared = NULL; | |
370 | parent->alien = NULL; | |
2e1217cf | 371 | parent->colour_next = 0; |
e498be7d CL |
372 | spin_lock_init(&parent->list_lock); |
373 | parent->free_objects = 0; | |
374 | parent->free_touched = 0; | |
375 | } | |
376 | ||
a737b3e2 AM |
377 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
378 | do { \ | |
379 | INIT_LIST_HEAD(listp); \ | |
380 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
381 | } while (0) |
382 | ||
a737b3e2 AM |
383 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
384 | do { \ | |
e498be7d CL |
385 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
386 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
387 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
388 | } while (0) | |
1da177e4 | 389 | |
1da177e4 LT |
390 | #define CFLGS_OFF_SLAB (0x80000000UL) |
391 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
392 | ||
393 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
394 | /* |
395 | * Optimization question: fewer reaps means less probability for unnessary | |
396 | * cpucache drain/refill cycles. | |
1da177e4 | 397 | * |
dc6f3f27 | 398 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
399 | * which could lock up otherwise freeable slabs. |
400 | */ | |
401 | #define REAPTIMEOUT_CPUC (2*HZ) | |
402 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
403 | ||
404 | #if STATS | |
405 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
406 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
407 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
408 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 409 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
410 | #define STATS_SET_HIGH(x) \ |
411 | do { \ | |
412 | if ((x)->num_active > (x)->high_mark) \ | |
413 | (x)->high_mark = (x)->num_active; \ | |
414 | } while (0) | |
1da177e4 LT |
415 | #define STATS_INC_ERR(x) ((x)->errors++) |
416 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 417 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 418 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
419 | #define STATS_SET_FREEABLE(x, i) \ |
420 | do { \ | |
421 | if ((x)->max_freeable < i) \ | |
422 | (x)->max_freeable = i; \ | |
423 | } while (0) | |
1da177e4 LT |
424 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
425 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
426 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
427 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
428 | #else | |
429 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
430 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
431 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
432 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 433 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
434 | #define STATS_SET_HIGH(x) do { } while (0) |
435 | #define STATS_INC_ERR(x) do { } while (0) | |
436 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 437 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 438 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 439 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
440 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
441 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
442 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
443 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
444 | #endif | |
445 | ||
446 | #if DEBUG | |
1da177e4 | 447 | |
a737b3e2 AM |
448 | /* |
449 | * memory layout of objects: | |
1da177e4 | 450 | * 0 : objp |
3dafccf2 | 451 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
452 | * the end of an object is aligned with the end of the real |
453 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 454 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 455 | * redzone word. |
3dafccf2 | 456 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
457 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
458 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 459 | * [BYTES_PER_WORD long] |
1da177e4 | 460 | */ |
343e0d7a | 461 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 462 | { |
3dafccf2 | 463 | return cachep->obj_offset; |
1da177e4 LT |
464 | } |
465 | ||
b46b8f19 | 466 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
467 | { |
468 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
469 | return (unsigned long long*) (objp + obj_offset(cachep) - |
470 | sizeof(unsigned long long)); | |
1da177e4 LT |
471 | } |
472 | ||
b46b8f19 | 473 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
474 | { |
475 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
476 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 477 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 478 | sizeof(unsigned long long) - |
87a927c7 | 479 | REDZONE_ALIGN); |
3b0efdfa | 480 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 481 | sizeof(unsigned long long)); |
1da177e4 LT |
482 | } |
483 | ||
343e0d7a | 484 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
485 | { |
486 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 487 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
488 | } |
489 | ||
490 | #else | |
491 | ||
3dafccf2 | 492 | #define obj_offset(x) 0 |
b46b8f19 DW |
493 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
494 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
495 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
496 | ||
497 | #endif | |
498 | ||
0f24f128 | 499 | #ifdef CONFIG_TRACING |
36555751 EGM |
500 | size_t slab_buffer_size(struct kmem_cache *cachep) |
501 | { | |
3b0efdfa | 502 | return cachep->size; |
36555751 EGM |
503 | } |
504 | EXPORT_SYMBOL(slab_buffer_size); | |
505 | #endif | |
506 | ||
1da177e4 | 507 | /* |
3df1cccd DR |
508 | * Do not go above this order unless 0 objects fit into the slab or |
509 | * overridden on the command line. | |
1da177e4 | 510 | */ |
543585cc DR |
511 | #define SLAB_MAX_ORDER_HI 1 |
512 | #define SLAB_MAX_ORDER_LO 0 | |
513 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 514 | static bool slab_max_order_set __initdata; |
1da177e4 | 515 | |
065d41cb PE |
516 | static inline struct kmem_cache *page_get_cache(struct page *page) |
517 | { | |
d85f3385 | 518 | page = compound_head(page); |
ddc2e812 | 519 | BUG_ON(!PageSlab(page)); |
e571b0ad | 520 | return page->slab_cache; |
065d41cb PE |
521 | } |
522 | ||
6ed5eb22 PE |
523 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
524 | { | |
b49af68f | 525 | struct page *page = virt_to_head_page(obj); |
35026088 | 526 | return page->slab_cache; |
6ed5eb22 PE |
527 | } |
528 | ||
529 | static inline struct slab *virt_to_slab(const void *obj) | |
530 | { | |
b49af68f | 531 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
532 | |
533 | VM_BUG_ON(!PageSlab(page)); | |
534 | return page->slab_page; | |
6ed5eb22 PE |
535 | } |
536 | ||
8fea4e96 PE |
537 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
538 | unsigned int idx) | |
539 | { | |
3b0efdfa | 540 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
541 | } |
542 | ||
6a2d7a95 | 543 | /* |
3b0efdfa CL |
544 | * We want to avoid an expensive divide : (offset / cache->size) |
545 | * Using the fact that size is a constant for a particular cache, | |
546 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
547 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
548 | */ | |
549 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
550 | const struct slab *slab, void *obj) | |
8fea4e96 | 551 | { |
6a2d7a95 ED |
552 | u32 offset = (obj - slab->s_mem); |
553 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
554 | } |
555 | ||
a737b3e2 AM |
556 | /* |
557 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
558 | */ | |
1da177e4 LT |
559 | struct cache_sizes malloc_sizes[] = { |
560 | #define CACHE(x) { .cs_size = (x) }, | |
561 | #include <linux/kmalloc_sizes.h> | |
562 | CACHE(ULONG_MAX) | |
563 | #undef CACHE | |
564 | }; | |
565 | EXPORT_SYMBOL(malloc_sizes); | |
566 | ||
567 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
568 | struct cache_names { | |
569 | char *name; | |
570 | char *name_dma; | |
571 | }; | |
572 | ||
573 | static struct cache_names __initdata cache_names[] = { | |
574 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
575 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 576 | {NULL,} |
1da177e4 LT |
577 | #undef CACHE |
578 | }; | |
579 | ||
580 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 581 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 582 | static struct arraycache_init initarray_generic = |
b28a02de | 583 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
584 | |
585 | /* internal cache of cache description objs */ | |
b56efcf0 | 586 | static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES]; |
343e0d7a | 587 | static struct kmem_cache cache_cache = { |
b56efcf0 | 588 | .nodelists = cache_cache_nodelists, |
b28a02de PE |
589 | .batchcount = 1, |
590 | .limit = BOOT_CPUCACHE_ENTRIES, | |
591 | .shared = 1, | |
3b0efdfa | 592 | .size = sizeof(struct kmem_cache), |
b28a02de | 593 | .name = "kmem_cache", |
1da177e4 LT |
594 | }; |
595 | ||
056c6241 RT |
596 | #define BAD_ALIEN_MAGIC 0x01020304ul |
597 | ||
f1aaee53 AV |
598 | #ifdef CONFIG_LOCKDEP |
599 | ||
600 | /* | |
601 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
602 | * for other slabs "off slab". | |
603 | * The locking for this is tricky in that it nests within the locks | |
604 | * of all other slabs in a few places; to deal with this special | |
605 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
606 | * |
607 | * We set lock class for alien array caches which are up during init. | |
608 | * The lock annotation will be lost if all cpus of a node goes down and | |
609 | * then comes back up during hotplug | |
f1aaee53 | 610 | */ |
056c6241 RT |
611 | static struct lock_class_key on_slab_l3_key; |
612 | static struct lock_class_key on_slab_alc_key; | |
613 | ||
83835b3d PZ |
614 | static struct lock_class_key debugobj_l3_key; |
615 | static struct lock_class_key debugobj_alc_key; | |
616 | ||
617 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
618 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
619 | int q) | |
620 | { | |
621 | struct array_cache **alc; | |
622 | struct kmem_list3 *l3; | |
623 | int r; | |
624 | ||
625 | l3 = cachep->nodelists[q]; | |
626 | if (!l3) | |
627 | return; | |
628 | ||
629 | lockdep_set_class(&l3->list_lock, l3_key); | |
630 | alc = l3->alien; | |
631 | /* | |
632 | * FIXME: This check for BAD_ALIEN_MAGIC | |
633 | * should go away when common slab code is taught to | |
634 | * work even without alien caches. | |
635 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
636 | * for alloc_alien_cache, | |
637 | */ | |
638 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
639 | return; | |
640 | for_each_node(r) { | |
641 | if (alc[r]) | |
642 | lockdep_set_class(&alc[r]->lock, alc_key); | |
643 | } | |
644 | } | |
645 | ||
646 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
647 | { | |
648 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
649 | } | |
650 | ||
651 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
652 | { | |
653 | int node; | |
654 | ||
655 | for_each_online_node(node) | |
656 | slab_set_debugobj_lock_classes_node(cachep, node); | |
657 | } | |
658 | ||
ce79ddc8 | 659 | static void init_node_lock_keys(int q) |
f1aaee53 | 660 | { |
056c6241 RT |
661 | struct cache_sizes *s = malloc_sizes; |
662 | ||
97d06609 | 663 | if (slab_state < UP) |
ce79ddc8 PE |
664 | return; |
665 | ||
666 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
ce79ddc8 | 667 | struct kmem_list3 *l3; |
ce79ddc8 PE |
668 | |
669 | l3 = s->cs_cachep->nodelists[q]; | |
670 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 671 | continue; |
83835b3d PZ |
672 | |
673 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
674 | &on_slab_alc_key, q); | |
f1aaee53 AV |
675 | } |
676 | } | |
ce79ddc8 PE |
677 | |
678 | static inline void init_lock_keys(void) | |
679 | { | |
680 | int node; | |
681 | ||
682 | for_each_node(node) | |
683 | init_node_lock_keys(node); | |
684 | } | |
f1aaee53 | 685 | #else |
ce79ddc8 PE |
686 | static void init_node_lock_keys(int q) |
687 | { | |
688 | } | |
689 | ||
056c6241 | 690 | static inline void init_lock_keys(void) |
f1aaee53 AV |
691 | { |
692 | } | |
83835b3d PZ |
693 | |
694 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
695 | { | |
696 | } | |
697 | ||
698 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
699 | { | |
700 | } | |
f1aaee53 AV |
701 | #endif |
702 | ||
1871e52c | 703 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 704 | |
343e0d7a | 705 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
706 | { |
707 | return cachep->array[smp_processor_id()]; | |
708 | } | |
709 | ||
a737b3e2 AM |
710 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
711 | gfp_t gfpflags) | |
1da177e4 LT |
712 | { |
713 | struct cache_sizes *csizep = malloc_sizes; | |
714 | ||
715 | #if DEBUG | |
716 | /* This happens if someone tries to call | |
b28a02de PE |
717 | * kmem_cache_create(), or __kmalloc(), before |
718 | * the generic caches are initialized. | |
719 | */ | |
c7e43c78 | 720 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 721 | #endif |
6cb8f913 CL |
722 | if (!size) |
723 | return ZERO_SIZE_PTR; | |
724 | ||
1da177e4 LT |
725 | while (size > csizep->cs_size) |
726 | csizep++; | |
727 | ||
728 | /* | |
0abf40c1 | 729 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
730 | * has cs_{dma,}cachep==NULL. Thus no special case |
731 | * for large kmalloc calls required. | |
732 | */ | |
4b51d669 | 733 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
734 | if (unlikely(gfpflags & GFP_DMA)) |
735 | return csizep->cs_dmacachep; | |
4b51d669 | 736 | #endif |
1da177e4 LT |
737 | return csizep->cs_cachep; |
738 | } | |
739 | ||
b221385b | 740 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
741 | { |
742 | return __find_general_cachep(size, gfpflags); | |
743 | } | |
97e2bde4 | 744 | |
fbaccacf | 745 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 746 | { |
fbaccacf SR |
747 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
748 | } | |
1da177e4 | 749 | |
a737b3e2 AM |
750 | /* |
751 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
752 | */ | |
fbaccacf SR |
753 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
754 | size_t align, int flags, size_t *left_over, | |
755 | unsigned int *num) | |
756 | { | |
757 | int nr_objs; | |
758 | size_t mgmt_size; | |
759 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 760 | |
fbaccacf SR |
761 | /* |
762 | * The slab management structure can be either off the slab or | |
763 | * on it. For the latter case, the memory allocated for a | |
764 | * slab is used for: | |
765 | * | |
766 | * - The struct slab | |
767 | * - One kmem_bufctl_t for each object | |
768 | * - Padding to respect alignment of @align | |
769 | * - @buffer_size bytes for each object | |
770 | * | |
771 | * If the slab management structure is off the slab, then the | |
772 | * alignment will already be calculated into the size. Because | |
773 | * the slabs are all pages aligned, the objects will be at the | |
774 | * correct alignment when allocated. | |
775 | */ | |
776 | if (flags & CFLGS_OFF_SLAB) { | |
777 | mgmt_size = 0; | |
778 | nr_objs = slab_size / buffer_size; | |
779 | ||
780 | if (nr_objs > SLAB_LIMIT) | |
781 | nr_objs = SLAB_LIMIT; | |
782 | } else { | |
783 | /* | |
784 | * Ignore padding for the initial guess. The padding | |
785 | * is at most @align-1 bytes, and @buffer_size is at | |
786 | * least @align. In the worst case, this result will | |
787 | * be one greater than the number of objects that fit | |
788 | * into the memory allocation when taking the padding | |
789 | * into account. | |
790 | */ | |
791 | nr_objs = (slab_size - sizeof(struct slab)) / | |
792 | (buffer_size + sizeof(kmem_bufctl_t)); | |
793 | ||
794 | /* | |
795 | * This calculated number will be either the right | |
796 | * amount, or one greater than what we want. | |
797 | */ | |
798 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
799 | > slab_size) | |
800 | nr_objs--; | |
801 | ||
802 | if (nr_objs > SLAB_LIMIT) | |
803 | nr_objs = SLAB_LIMIT; | |
804 | ||
805 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
806 | } | |
807 | *num = nr_objs; | |
808 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
809 | } |
810 | ||
d40cee24 | 811 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 812 | |
a737b3e2 AM |
813 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
814 | char *msg) | |
1da177e4 LT |
815 | { |
816 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 817 | function, cachep->name, msg); |
1da177e4 LT |
818 | dump_stack(); |
819 | } | |
820 | ||
3395ee05 PM |
821 | /* |
822 | * By default on NUMA we use alien caches to stage the freeing of | |
823 | * objects allocated from other nodes. This causes massive memory | |
824 | * inefficiencies when using fake NUMA setup to split memory into a | |
825 | * large number of small nodes, so it can be disabled on the command | |
826 | * line | |
827 | */ | |
828 | ||
829 | static int use_alien_caches __read_mostly = 1; | |
830 | static int __init noaliencache_setup(char *s) | |
831 | { | |
832 | use_alien_caches = 0; | |
833 | return 1; | |
834 | } | |
835 | __setup("noaliencache", noaliencache_setup); | |
836 | ||
3df1cccd DR |
837 | static int __init slab_max_order_setup(char *str) |
838 | { | |
839 | get_option(&str, &slab_max_order); | |
840 | slab_max_order = slab_max_order < 0 ? 0 : | |
841 | min(slab_max_order, MAX_ORDER - 1); | |
842 | slab_max_order_set = true; | |
843 | ||
844 | return 1; | |
845 | } | |
846 | __setup("slab_max_order=", slab_max_order_setup); | |
847 | ||
8fce4d8e CL |
848 | #ifdef CONFIG_NUMA |
849 | /* | |
850 | * Special reaping functions for NUMA systems called from cache_reap(). | |
851 | * These take care of doing round robin flushing of alien caches (containing | |
852 | * objects freed on different nodes from which they were allocated) and the | |
853 | * flushing of remote pcps by calling drain_node_pages. | |
854 | */ | |
1871e52c | 855 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
856 | |
857 | static void init_reap_node(int cpu) | |
858 | { | |
859 | int node; | |
860 | ||
7d6e6d09 | 861 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 862 | if (node == MAX_NUMNODES) |
442295c9 | 863 | node = first_node(node_online_map); |
8fce4d8e | 864 | |
1871e52c | 865 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
866 | } |
867 | ||
868 | static void next_reap_node(void) | |
869 | { | |
909ea964 | 870 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 871 | |
8fce4d8e CL |
872 | node = next_node(node, node_online_map); |
873 | if (unlikely(node >= MAX_NUMNODES)) | |
874 | node = first_node(node_online_map); | |
909ea964 | 875 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
876 | } |
877 | ||
878 | #else | |
879 | #define init_reap_node(cpu) do { } while (0) | |
880 | #define next_reap_node(void) do { } while (0) | |
881 | #endif | |
882 | ||
1da177e4 LT |
883 | /* |
884 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
885 | * via the workqueue/eventd. | |
886 | * Add the CPU number into the expiration time to minimize the possibility of | |
887 | * the CPUs getting into lockstep and contending for the global cache chain | |
888 | * lock. | |
889 | */ | |
897e679b | 890 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 891 | { |
1871e52c | 892 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
893 | |
894 | /* | |
895 | * When this gets called from do_initcalls via cpucache_init(), | |
896 | * init_workqueues() has already run, so keventd will be setup | |
897 | * at that time. | |
898 | */ | |
52bad64d | 899 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 900 | init_reap_node(cpu); |
78b43536 | 901 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); |
2b284214 AV |
902 | schedule_delayed_work_on(cpu, reap_work, |
903 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
904 | } |
905 | } | |
906 | ||
e498be7d | 907 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 908 | int batchcount, gfp_t gfp) |
1da177e4 | 909 | { |
b28a02de | 910 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
911 | struct array_cache *nc = NULL; |
912 | ||
83b519e8 | 913 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
914 | /* |
915 | * The array_cache structures contain pointers to free object. | |
25985edc | 916 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
917 | * cache the pointers are not cleared and they could be counted as |
918 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
919 | * not scan such objects. | |
920 | */ | |
921 | kmemleak_no_scan(nc); | |
1da177e4 LT |
922 | if (nc) { |
923 | nc->avail = 0; | |
924 | nc->limit = entries; | |
925 | nc->batchcount = batchcount; | |
926 | nc->touched = 0; | |
e498be7d | 927 | spin_lock_init(&nc->lock); |
1da177e4 LT |
928 | } |
929 | return nc; | |
930 | } | |
931 | ||
072bb0aa MG |
932 | static inline bool is_slab_pfmemalloc(struct slab *slabp) |
933 | { | |
934 | struct page *page = virt_to_page(slabp->s_mem); | |
935 | ||
936 | return PageSlabPfmemalloc(page); | |
937 | } | |
938 | ||
939 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
940 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
941 | struct array_cache *ac) | |
942 | { | |
943 | struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()]; | |
944 | struct slab *slabp; | |
945 | unsigned long flags; | |
946 | ||
947 | if (!pfmemalloc_active) | |
948 | return; | |
949 | ||
950 | spin_lock_irqsave(&l3->list_lock, flags); | |
951 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
952 | if (is_slab_pfmemalloc(slabp)) | |
953 | goto out; | |
954 | ||
955 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
956 | if (is_slab_pfmemalloc(slabp)) | |
957 | goto out; | |
958 | ||
959 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
960 | if (is_slab_pfmemalloc(slabp)) | |
961 | goto out; | |
962 | ||
963 | pfmemalloc_active = false; | |
964 | out: | |
965 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
966 | } | |
967 | ||
968 | static void *ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
969 | gfp_t flags, bool force_refill) | |
970 | { | |
971 | int i; | |
972 | void *objp = ac->entry[--ac->avail]; | |
973 | ||
974 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
975 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
976 | struct kmem_list3 *l3; | |
977 | ||
978 | if (gfp_pfmemalloc_allowed(flags)) { | |
979 | clear_obj_pfmemalloc(&objp); | |
980 | return objp; | |
981 | } | |
982 | ||
983 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
984 | for (i = 1; i < ac->avail; i++) { | |
985 | /* If a !PFMEMALLOC object is found, swap them */ | |
986 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
987 | objp = ac->entry[i]; | |
988 | ac->entry[i] = ac->entry[ac->avail]; | |
989 | ac->entry[ac->avail] = objp; | |
990 | return objp; | |
991 | } | |
992 | } | |
993 | ||
994 | /* | |
995 | * If there are empty slabs on the slabs_free list and we are | |
996 | * being forced to refill the cache, mark this one !pfmemalloc. | |
997 | */ | |
998 | l3 = cachep->nodelists[numa_mem_id()]; | |
999 | if (!list_empty(&l3->slabs_free) && force_refill) { | |
1000 | struct slab *slabp = virt_to_slab(objp); | |
1001 | ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem)); | |
1002 | clear_obj_pfmemalloc(&objp); | |
1003 | recheck_pfmemalloc_active(cachep, ac); | |
1004 | return objp; | |
1005 | } | |
1006 | ||
1007 | /* No !PFMEMALLOC objects available */ | |
1008 | ac->avail++; | |
1009 | objp = NULL; | |
1010 | } | |
1011 | ||
1012 | return objp; | |
1013 | } | |
1014 | ||
1015 | static void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
1016 | void *objp) | |
1017 | { | |
1018 | if (unlikely(pfmemalloc_active)) { | |
1019 | /* Some pfmemalloc slabs exist, check if this is one */ | |
1020 | struct page *page = virt_to_page(objp); | |
1021 | if (PageSlabPfmemalloc(page)) | |
1022 | set_obj_pfmemalloc(&objp); | |
1023 | } | |
1024 | ||
1025 | ac->entry[ac->avail++] = objp; | |
1026 | } | |
1027 | ||
3ded175a CL |
1028 | /* |
1029 | * Transfer objects in one arraycache to another. | |
1030 | * Locking must be handled by the caller. | |
1031 | * | |
1032 | * Return the number of entries transferred. | |
1033 | */ | |
1034 | static int transfer_objects(struct array_cache *to, | |
1035 | struct array_cache *from, unsigned int max) | |
1036 | { | |
1037 | /* Figure out how many entries to transfer */ | |
732eacc0 | 1038 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
1039 | |
1040 | if (!nr) | |
1041 | return 0; | |
1042 | ||
1043 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1044 | sizeof(void *) *nr); | |
1045 | ||
1046 | from->avail -= nr; | |
1047 | to->avail += nr; | |
3ded175a CL |
1048 | return nr; |
1049 | } | |
1050 | ||
765c4507 CL |
1051 | #ifndef CONFIG_NUMA |
1052 | ||
1053 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1054 | #define reap_alien(cachep, l3) do { } while (0) | |
1055 | ||
83b519e8 | 1056 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1057 | { |
1058 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1059 | } | |
1060 | ||
1061 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1062 | { | |
1063 | } | |
1064 | ||
1065 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1066 | { | |
1067 | return 0; | |
1068 | } | |
1069 | ||
1070 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1071 | gfp_t flags) | |
1072 | { | |
1073 | return NULL; | |
1074 | } | |
1075 | ||
8b98c169 | 1076 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1077 | gfp_t flags, int nodeid) |
1078 | { | |
1079 | return NULL; | |
1080 | } | |
1081 | ||
1082 | #else /* CONFIG_NUMA */ | |
1083 | ||
8b98c169 | 1084 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1085 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1086 | |
83b519e8 | 1087 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1088 | { |
1089 | struct array_cache **ac_ptr; | |
8ef82866 | 1090 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1091 | int i; |
1092 | ||
1093 | if (limit > 1) | |
1094 | limit = 12; | |
f3186a9c | 1095 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
1096 | if (ac_ptr) { |
1097 | for_each_node(i) { | |
f3186a9c | 1098 | if (i == node || !node_online(i)) |
e498be7d | 1099 | continue; |
83b519e8 | 1100 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1101 | if (!ac_ptr[i]) { |
cc550def | 1102 | for (i--; i >= 0; i--) |
e498be7d CL |
1103 | kfree(ac_ptr[i]); |
1104 | kfree(ac_ptr); | |
1105 | return NULL; | |
1106 | } | |
1107 | } | |
1108 | } | |
1109 | return ac_ptr; | |
1110 | } | |
1111 | ||
5295a74c | 1112 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1113 | { |
1114 | int i; | |
1115 | ||
1116 | if (!ac_ptr) | |
1117 | return; | |
e498be7d | 1118 | for_each_node(i) |
b28a02de | 1119 | kfree(ac_ptr[i]); |
e498be7d CL |
1120 | kfree(ac_ptr); |
1121 | } | |
1122 | ||
343e0d7a | 1123 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1124 | struct array_cache *ac, int node) |
e498be7d CL |
1125 | { |
1126 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1127 | ||
1128 | if (ac->avail) { | |
1129 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1130 | /* |
1131 | * Stuff objects into the remote nodes shared array first. | |
1132 | * That way we could avoid the overhead of putting the objects | |
1133 | * into the free lists and getting them back later. | |
1134 | */ | |
693f7d36 JS |
1135 | if (rl3->shared) |
1136 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1137 | |
ff69416e | 1138 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1139 | ac->avail = 0; |
1140 | spin_unlock(&rl3->list_lock); | |
1141 | } | |
1142 | } | |
1143 | ||
8fce4d8e CL |
1144 | /* |
1145 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1146 | */ | |
1147 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1148 | { | |
909ea964 | 1149 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1150 | |
1151 | if (l3->alien) { | |
1152 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1153 | |
1154 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1155 | __drain_alien_cache(cachep, ac, node); |
1156 | spin_unlock_irq(&ac->lock); | |
1157 | } | |
1158 | } | |
1159 | } | |
1160 | ||
a737b3e2 AM |
1161 | static void drain_alien_cache(struct kmem_cache *cachep, |
1162 | struct array_cache **alien) | |
e498be7d | 1163 | { |
b28a02de | 1164 | int i = 0; |
e498be7d CL |
1165 | struct array_cache *ac; |
1166 | unsigned long flags; | |
1167 | ||
1168 | for_each_online_node(i) { | |
4484ebf1 | 1169 | ac = alien[i]; |
e498be7d CL |
1170 | if (ac) { |
1171 | spin_lock_irqsave(&ac->lock, flags); | |
1172 | __drain_alien_cache(cachep, ac, i); | |
1173 | spin_unlock_irqrestore(&ac->lock, flags); | |
1174 | } | |
1175 | } | |
1176 | } | |
729bd0b7 | 1177 | |
873623df | 1178 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1179 | { |
1180 | struct slab *slabp = virt_to_slab(objp); | |
1181 | int nodeid = slabp->nodeid; | |
1182 | struct kmem_list3 *l3; | |
1183 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1184 | int node; |
1185 | ||
7d6e6d09 | 1186 | node = numa_mem_id(); |
729bd0b7 PE |
1187 | |
1188 | /* | |
1189 | * Make sure we are not freeing a object from another node to the array | |
1190 | * cache on this cpu. | |
1191 | */ | |
62918a03 | 1192 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1193 | return 0; |
1194 | ||
1ca4cb24 | 1195 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1196 | STATS_INC_NODEFREES(cachep); |
1197 | if (l3->alien && l3->alien[nodeid]) { | |
1198 | alien = l3->alien[nodeid]; | |
873623df | 1199 | spin_lock(&alien->lock); |
729bd0b7 PE |
1200 | if (unlikely(alien->avail == alien->limit)) { |
1201 | STATS_INC_ACOVERFLOW(cachep); | |
1202 | __drain_alien_cache(cachep, alien, nodeid); | |
1203 | } | |
072bb0aa | 1204 | ac_put_obj(cachep, alien, objp); |
729bd0b7 PE |
1205 | spin_unlock(&alien->lock); |
1206 | } else { | |
1207 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1208 | free_block(cachep, &objp, 1, nodeid); | |
1209 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1210 | } | |
1211 | return 1; | |
1212 | } | |
e498be7d CL |
1213 | #endif |
1214 | ||
8f9f8d9e DR |
1215 | /* |
1216 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1217 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1218 | * will be allocated off-node since memory is not yet online for the new node. | |
1219 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1220 | * already in use. | |
1221 | * | |
18004c5d | 1222 | * Must hold slab_mutex. |
8f9f8d9e DR |
1223 | */ |
1224 | static int init_cache_nodelists_node(int node) | |
1225 | { | |
1226 | struct kmem_cache *cachep; | |
1227 | struct kmem_list3 *l3; | |
1228 | const int memsize = sizeof(struct kmem_list3); | |
1229 | ||
18004c5d | 1230 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1231 | /* |
1232 | * Set up the size64 kmemlist for cpu before we can | |
1233 | * begin anything. Make sure some other cpu on this | |
1234 | * node has not already allocated this | |
1235 | */ | |
1236 | if (!cachep->nodelists[node]) { | |
1237 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1238 | if (!l3) | |
1239 | return -ENOMEM; | |
1240 | kmem_list3_init(l3); | |
1241 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1242 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1243 | ||
1244 | /* | |
1245 | * The l3s don't come and go as CPUs come and | |
18004c5d | 1246 | * go. slab_mutex is sufficient |
8f9f8d9e DR |
1247 | * protection here. |
1248 | */ | |
1249 | cachep->nodelists[node] = l3; | |
1250 | } | |
1251 | ||
1252 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1253 | cachep->nodelists[node]->free_limit = | |
1254 | (1 + nr_cpus_node(node)) * | |
1255 | cachep->batchcount + cachep->num; | |
1256 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1257 | } | |
1258 | return 0; | |
1259 | } | |
1260 | ||
fbf1e473 AM |
1261 | static void __cpuinit cpuup_canceled(long cpu) |
1262 | { | |
1263 | struct kmem_cache *cachep; | |
1264 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1265 | int node = cpu_to_mem(cpu); |
a70f7302 | 1266 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1267 | |
18004c5d | 1268 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1269 | struct array_cache *nc; |
1270 | struct array_cache *shared; | |
1271 | struct array_cache **alien; | |
fbf1e473 | 1272 | |
fbf1e473 AM |
1273 | /* cpu is dead; no one can alloc from it. */ |
1274 | nc = cachep->array[cpu]; | |
1275 | cachep->array[cpu] = NULL; | |
1276 | l3 = cachep->nodelists[node]; | |
1277 | ||
1278 | if (!l3) | |
1279 | goto free_array_cache; | |
1280 | ||
1281 | spin_lock_irq(&l3->list_lock); | |
1282 | ||
1283 | /* Free limit for this kmem_list3 */ | |
1284 | l3->free_limit -= cachep->batchcount; | |
1285 | if (nc) | |
1286 | free_block(cachep, nc->entry, nc->avail, node); | |
1287 | ||
58463c1f | 1288 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1289 | spin_unlock_irq(&l3->list_lock); |
1290 | goto free_array_cache; | |
1291 | } | |
1292 | ||
1293 | shared = l3->shared; | |
1294 | if (shared) { | |
1295 | free_block(cachep, shared->entry, | |
1296 | shared->avail, node); | |
1297 | l3->shared = NULL; | |
1298 | } | |
1299 | ||
1300 | alien = l3->alien; | |
1301 | l3->alien = NULL; | |
1302 | ||
1303 | spin_unlock_irq(&l3->list_lock); | |
1304 | ||
1305 | kfree(shared); | |
1306 | if (alien) { | |
1307 | drain_alien_cache(cachep, alien); | |
1308 | free_alien_cache(alien); | |
1309 | } | |
1310 | free_array_cache: | |
1311 | kfree(nc); | |
1312 | } | |
1313 | /* | |
1314 | * In the previous loop, all the objects were freed to | |
1315 | * the respective cache's slabs, now we can go ahead and | |
1316 | * shrink each nodelist to its limit. | |
1317 | */ | |
18004c5d | 1318 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1319 | l3 = cachep->nodelists[node]; |
1320 | if (!l3) | |
1321 | continue; | |
1322 | drain_freelist(cachep, l3, l3->free_objects); | |
1323 | } | |
1324 | } | |
1325 | ||
1326 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1327 | { |
343e0d7a | 1328 | struct kmem_cache *cachep; |
e498be7d | 1329 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1330 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1331 | int err; |
1da177e4 | 1332 | |
fbf1e473 AM |
1333 | /* |
1334 | * We need to do this right in the beginning since | |
1335 | * alloc_arraycache's are going to use this list. | |
1336 | * kmalloc_node allows us to add the slab to the right | |
1337 | * kmem_list3 and not this cpu's kmem_list3 | |
1338 | */ | |
8f9f8d9e DR |
1339 | err = init_cache_nodelists_node(node); |
1340 | if (err < 0) | |
1341 | goto bad; | |
fbf1e473 AM |
1342 | |
1343 | /* | |
1344 | * Now we can go ahead with allocating the shared arrays and | |
1345 | * array caches | |
1346 | */ | |
18004c5d | 1347 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1348 | struct array_cache *nc; |
1349 | struct array_cache *shared = NULL; | |
1350 | struct array_cache **alien = NULL; | |
1351 | ||
1352 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1353 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1354 | if (!nc) |
1355 | goto bad; | |
1356 | if (cachep->shared) { | |
1357 | shared = alloc_arraycache(node, | |
1358 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1359 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1360 | if (!shared) { |
1361 | kfree(nc); | |
1da177e4 | 1362 | goto bad; |
12d00f6a | 1363 | } |
fbf1e473 AM |
1364 | } |
1365 | if (use_alien_caches) { | |
83b519e8 | 1366 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1367 | if (!alien) { |
1368 | kfree(shared); | |
1369 | kfree(nc); | |
fbf1e473 | 1370 | goto bad; |
12d00f6a | 1371 | } |
fbf1e473 AM |
1372 | } |
1373 | cachep->array[cpu] = nc; | |
1374 | l3 = cachep->nodelists[node]; | |
1375 | BUG_ON(!l3); | |
1376 | ||
1377 | spin_lock_irq(&l3->list_lock); | |
1378 | if (!l3->shared) { | |
1379 | /* | |
1380 | * We are serialised from CPU_DEAD or | |
1381 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1382 | */ | |
1383 | l3->shared = shared; | |
1384 | shared = NULL; | |
1385 | } | |
4484ebf1 | 1386 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1387 | if (!l3->alien) { |
1388 | l3->alien = alien; | |
1389 | alien = NULL; | |
1da177e4 | 1390 | } |
fbf1e473 AM |
1391 | #endif |
1392 | spin_unlock_irq(&l3->list_lock); | |
1393 | kfree(shared); | |
1394 | free_alien_cache(alien); | |
83835b3d PZ |
1395 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1396 | slab_set_debugobj_lock_classes_node(cachep, node); | |
fbf1e473 | 1397 | } |
ce79ddc8 PE |
1398 | init_node_lock_keys(node); |
1399 | ||
fbf1e473 AM |
1400 | return 0; |
1401 | bad: | |
12d00f6a | 1402 | cpuup_canceled(cpu); |
fbf1e473 AM |
1403 | return -ENOMEM; |
1404 | } | |
1405 | ||
1406 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1407 | unsigned long action, void *hcpu) | |
1408 | { | |
1409 | long cpu = (long)hcpu; | |
1410 | int err = 0; | |
1411 | ||
1412 | switch (action) { | |
fbf1e473 AM |
1413 | case CPU_UP_PREPARE: |
1414 | case CPU_UP_PREPARE_FROZEN: | |
18004c5d | 1415 | mutex_lock(&slab_mutex); |
fbf1e473 | 1416 | err = cpuup_prepare(cpu); |
18004c5d | 1417 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
1418 | break; |
1419 | case CPU_ONLINE: | |
8bb78442 | 1420 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1421 | start_cpu_timer(cpu); |
1422 | break; | |
1423 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1424 | case CPU_DOWN_PREPARE: |
8bb78442 | 1425 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 | 1426 | /* |
18004c5d | 1427 | * Shutdown cache reaper. Note that the slab_mutex is |
5830c590 CL |
1428 | * held so that if cache_reap() is invoked it cannot do |
1429 | * anything expensive but will only modify reap_work | |
1430 | * and reschedule the timer. | |
1431 | */ | |
afe2c511 | 1432 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1433 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1434 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1435 | break; |
1436 | case CPU_DOWN_FAILED: | |
8bb78442 | 1437 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1438 | start_cpu_timer(cpu); |
1439 | break; | |
1da177e4 | 1440 | case CPU_DEAD: |
8bb78442 | 1441 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1442 | /* |
1443 | * Even if all the cpus of a node are down, we don't free the | |
1444 | * kmem_list3 of any cache. This to avoid a race between | |
1445 | * cpu_down, and a kmalloc allocation from another cpu for | |
1446 | * memory from the node of the cpu going down. The list3 | |
1447 | * structure is usually allocated from kmem_cache_create() and | |
1448 | * gets destroyed at kmem_cache_destroy(). | |
1449 | */ | |
183ff22b | 1450 | /* fall through */ |
8f5be20b | 1451 | #endif |
1da177e4 | 1452 | case CPU_UP_CANCELED: |
8bb78442 | 1453 | case CPU_UP_CANCELED_FROZEN: |
18004c5d | 1454 | mutex_lock(&slab_mutex); |
fbf1e473 | 1455 | cpuup_canceled(cpu); |
18004c5d | 1456 | mutex_unlock(&slab_mutex); |
1da177e4 | 1457 | break; |
1da177e4 | 1458 | } |
eac40680 | 1459 | return notifier_from_errno(err); |
1da177e4 LT |
1460 | } |
1461 | ||
74b85f37 CS |
1462 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1463 | &cpuup_callback, NULL, 0 | |
1464 | }; | |
1da177e4 | 1465 | |
8f9f8d9e DR |
1466 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1467 | /* | |
1468 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1469 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1470 | * removed. | |
1471 | * | |
18004c5d | 1472 | * Must hold slab_mutex. |
8f9f8d9e DR |
1473 | */ |
1474 | static int __meminit drain_cache_nodelists_node(int node) | |
1475 | { | |
1476 | struct kmem_cache *cachep; | |
1477 | int ret = 0; | |
1478 | ||
18004c5d | 1479 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1480 | struct kmem_list3 *l3; |
1481 | ||
1482 | l3 = cachep->nodelists[node]; | |
1483 | if (!l3) | |
1484 | continue; | |
1485 | ||
1486 | drain_freelist(cachep, l3, l3->free_objects); | |
1487 | ||
1488 | if (!list_empty(&l3->slabs_full) || | |
1489 | !list_empty(&l3->slabs_partial)) { | |
1490 | ret = -EBUSY; | |
1491 | break; | |
1492 | } | |
1493 | } | |
1494 | return ret; | |
1495 | } | |
1496 | ||
1497 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1498 | unsigned long action, void *arg) | |
1499 | { | |
1500 | struct memory_notify *mnb = arg; | |
1501 | int ret = 0; | |
1502 | int nid; | |
1503 | ||
1504 | nid = mnb->status_change_nid; | |
1505 | if (nid < 0) | |
1506 | goto out; | |
1507 | ||
1508 | switch (action) { | |
1509 | case MEM_GOING_ONLINE: | |
18004c5d | 1510 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1511 | ret = init_cache_nodelists_node(nid); |
18004c5d | 1512 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1513 | break; |
1514 | case MEM_GOING_OFFLINE: | |
18004c5d | 1515 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1516 | ret = drain_cache_nodelists_node(nid); |
18004c5d | 1517 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1518 | break; |
1519 | case MEM_ONLINE: | |
1520 | case MEM_OFFLINE: | |
1521 | case MEM_CANCEL_ONLINE: | |
1522 | case MEM_CANCEL_OFFLINE: | |
1523 | break; | |
1524 | } | |
1525 | out: | |
5fda1bd5 | 1526 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1527 | } |
1528 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1529 | ||
e498be7d CL |
1530 | /* |
1531 | * swap the static kmem_list3 with kmalloced memory | |
1532 | */ | |
8f9f8d9e DR |
1533 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1534 | int nodeid) | |
e498be7d CL |
1535 | { |
1536 | struct kmem_list3 *ptr; | |
1537 | ||
83b519e8 | 1538 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1539 | BUG_ON(!ptr); |
1540 | ||
e498be7d | 1541 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1542 | /* |
1543 | * Do not assume that spinlocks can be initialized via memcpy: | |
1544 | */ | |
1545 | spin_lock_init(&ptr->list_lock); | |
1546 | ||
e498be7d CL |
1547 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1548 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1549 | } |
1550 | ||
556a169d PE |
1551 | /* |
1552 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1553 | * size of kmem_list3. | |
1554 | */ | |
1555 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1556 | { | |
1557 | int node; | |
1558 | ||
1559 | for_each_online_node(node) { | |
1560 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1561 | cachep->nodelists[node]->next_reap = jiffies + | |
1562 | REAPTIMEOUT_LIST3 + | |
1563 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1564 | } | |
1565 | } | |
1566 | ||
a737b3e2 AM |
1567 | /* |
1568 | * Initialisation. Called after the page allocator have been initialised and | |
1569 | * before smp_init(). | |
1da177e4 LT |
1570 | */ |
1571 | void __init kmem_cache_init(void) | |
1572 | { | |
1573 | size_t left_over; | |
1574 | struct cache_sizes *sizes; | |
1575 | struct cache_names *names; | |
e498be7d | 1576 | int i; |
07ed76b2 | 1577 | int order; |
1ca4cb24 | 1578 | int node; |
e498be7d | 1579 | |
b6e68bc1 | 1580 | if (num_possible_nodes() == 1) |
62918a03 SS |
1581 | use_alien_caches = 0; |
1582 | ||
e498be7d CL |
1583 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1584 | kmem_list3_init(&initkmem_list3[i]); | |
1585 | if (i < MAX_NUMNODES) | |
1586 | cache_cache.nodelists[i] = NULL; | |
1587 | } | |
556a169d | 1588 | set_up_list3s(&cache_cache, CACHE_CACHE); |
1da177e4 LT |
1589 | |
1590 | /* | |
1591 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1592 | * page orders on machines with more than 32MB of memory if |
1593 | * not overridden on the command line. | |
1da177e4 | 1594 | */ |
3df1cccd | 1595 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1596 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1597 | |
1da177e4 LT |
1598 | /* Bootstrap is tricky, because several objects are allocated |
1599 | * from caches that do not exist yet: | |
a737b3e2 AM |
1600 | * 1) initialize the cache_cache cache: it contains the struct |
1601 | * kmem_cache structures of all caches, except cache_cache itself: | |
1602 | * cache_cache is statically allocated. | |
e498be7d CL |
1603 | * Initially an __init data area is used for the head array and the |
1604 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1605 | * array at the end of the bootstrap. | |
1da177e4 | 1606 | * 2) Create the first kmalloc cache. |
343e0d7a | 1607 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1608 | * An __init data area is used for the head array. |
1609 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1610 | * head arrays. | |
1da177e4 LT |
1611 | * 4) Replace the __init data head arrays for cache_cache and the first |
1612 | * kmalloc cache with kmalloc allocated arrays. | |
e498be7d CL |
1613 | * 5) Replace the __init data for kmem_list3 for cache_cache and |
1614 | * the other cache's with kmalloc allocated memory. | |
1615 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1616 | */ |
1617 | ||
7d6e6d09 | 1618 | node = numa_mem_id(); |
1ca4cb24 | 1619 | |
1da177e4 | 1620 | /* 1) create the cache_cache */ |
18004c5d CL |
1621 | INIT_LIST_HEAD(&slab_caches); |
1622 | list_add(&cache_cache.list, &slab_caches); | |
1da177e4 LT |
1623 | cache_cache.colour_off = cache_line_size(); |
1624 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
ec1f5eee | 1625 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; |
1da177e4 | 1626 | |
8da3430d | 1627 | /* |
b56efcf0 | 1628 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1629 | */ |
3b0efdfa | 1630 | cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
b56efcf0 | 1631 | nr_node_ids * sizeof(struct kmem_list3 *); |
3b0efdfa CL |
1632 | cache_cache.object_size = cache_cache.size; |
1633 | cache_cache.size = ALIGN(cache_cache.size, | |
a737b3e2 | 1634 | cache_line_size()); |
6a2d7a95 | 1635 | cache_cache.reciprocal_buffer_size = |
3b0efdfa | 1636 | reciprocal_value(cache_cache.size); |
1da177e4 | 1637 | |
07ed76b2 | 1638 | for (order = 0; order < MAX_ORDER; order++) { |
3b0efdfa | 1639 | cache_estimate(order, cache_cache.size, |
07ed76b2 JS |
1640 | cache_line_size(), 0, &left_over, &cache_cache.num); |
1641 | if (cache_cache.num) | |
1642 | break; | |
1643 | } | |
40094fa6 | 1644 | BUG_ON(!cache_cache.num); |
07ed76b2 | 1645 | cache_cache.gfporder = order; |
b28a02de | 1646 | cache_cache.colour = left_over / cache_cache.colour_off; |
b28a02de PE |
1647 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + |
1648 | sizeof(struct slab), cache_line_size()); | |
1da177e4 LT |
1649 | |
1650 | /* 2+3) create the kmalloc caches */ | |
1651 | sizes = malloc_sizes; | |
1652 | names = cache_names; | |
1653 | ||
a737b3e2 AM |
1654 | /* |
1655 | * Initialize the caches that provide memory for the array cache and the | |
1656 | * kmem_list3 structures first. Without this, further allocations will | |
1657 | * bug. | |
e498be7d CL |
1658 | */ |
1659 | ||
039363f3 | 1660 | sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name, |
a737b3e2 AM |
1661 | sizes[INDEX_AC].cs_size, |
1662 | ARCH_KMALLOC_MINALIGN, | |
1663 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1664 | NULL); |
e498be7d | 1665 | |
a737b3e2 | 1666 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1667 | sizes[INDEX_L3].cs_cachep = |
039363f3 | 1668 | __kmem_cache_create(names[INDEX_L3].name, |
a737b3e2 AM |
1669 | sizes[INDEX_L3].cs_size, |
1670 | ARCH_KMALLOC_MINALIGN, | |
1671 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1672 | NULL); |
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) { |
039363f3 | 1686 | sizes->cs_cachep = __kmem_cache_create(names->name, |
a737b3e2 AM |
1687 | sizes->cs_size, |
1688 | ARCH_KMALLOC_MINALIGN, | |
1689 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1690 | NULL); |
a737b3e2 | 1691 | } |
4b51d669 | 1692 | #ifdef CONFIG_ZONE_DMA |
039363f3 | 1693 | sizes->cs_dmacachep = __kmem_cache_create( |
4b51d669 | 1694 | names->name_dma, |
a737b3e2 AM |
1695 | sizes->cs_size, |
1696 | ARCH_KMALLOC_MINALIGN, | |
1697 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1698 | SLAB_PANIC, | |
20c2df83 | 1699 | NULL); |
4b51d669 | 1700 | #endif |
1da177e4 LT |
1701 | sizes++; |
1702 | names++; | |
1703 | } | |
1704 | /* 4) Replace the bootstrap head arrays */ | |
1705 | { | |
2b2d5493 | 1706 | struct array_cache *ptr; |
e498be7d | 1707 | |
83b519e8 | 1708 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1709 | |
9a2dba4b PE |
1710 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); |
1711 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
b28a02de | 1712 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1713 | /* |
1714 | * Do not assume that spinlocks can be initialized via memcpy: | |
1715 | */ | |
1716 | spin_lock_init(&ptr->lock); | |
1717 | ||
1da177e4 | 1718 | cache_cache.array[smp_processor_id()] = ptr; |
e498be7d | 1719 | |
83b519e8 | 1720 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1721 | |
9a2dba4b | 1722 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1723 | != &initarray_generic.cache); |
9a2dba4b | 1724 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1725 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1726 | /* |
1727 | * Do not assume that spinlocks can be initialized via memcpy: | |
1728 | */ | |
1729 | spin_lock_init(&ptr->lock); | |
1730 | ||
e498be7d | 1731 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1732 | ptr; |
1da177e4 | 1733 | } |
e498be7d CL |
1734 | /* 5) Replace the bootstrap kmem_list3's */ |
1735 | { | |
1ca4cb24 PE |
1736 | int nid; |
1737 | ||
9c09a95c | 1738 | for_each_online_node(nid) { |
ec1f5eee | 1739 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1740 | |
e498be7d | 1741 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1742 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1743 | |
1744 | if (INDEX_AC != INDEX_L3) { | |
1745 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1746 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1747 | } |
1748 | } | |
1749 | } | |
1da177e4 | 1750 | |
97d06609 | 1751 | slab_state = UP; |
8429db5c PE |
1752 | } |
1753 | ||
1754 | void __init kmem_cache_init_late(void) | |
1755 | { | |
1756 | struct kmem_cache *cachep; | |
1757 | ||
97d06609 | 1758 | slab_state = UP; |
52cef189 | 1759 | |
30765b92 PZ |
1760 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1761 | init_lock_keys(); | |
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 | |
97d06609 CL |
1770 | /* Done! */ |
1771 | slab_state = FULL; | |
1772 | ||
a737b3e2 AM |
1773 | /* |
1774 | * Register a cpu startup notifier callback that initializes | |
1775 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1776 | */ |
1777 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1778 | |
8f9f8d9e DR |
1779 | #ifdef CONFIG_NUMA |
1780 | /* | |
1781 | * Register a memory hotplug callback that initializes and frees | |
1782 | * nodelists. | |
1783 | */ | |
1784 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1785 | #endif | |
1786 | ||
a737b3e2 AM |
1787 | /* |
1788 | * The reap timers are started later, with a module init call: That part | |
1789 | * of the kernel is not yet operational. | |
1da177e4 LT |
1790 | */ |
1791 | } | |
1792 | ||
1793 | static int __init cpucache_init(void) | |
1794 | { | |
1795 | int cpu; | |
1796 | ||
a737b3e2 AM |
1797 | /* |
1798 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1799 | */ |
e498be7d | 1800 | for_each_online_cpu(cpu) |
a737b3e2 | 1801 | start_cpu_timer(cpu); |
a164f896 GC |
1802 | |
1803 | /* Done! */ | |
97d06609 | 1804 | slab_state = FULL; |
1da177e4 LT |
1805 | return 0; |
1806 | } | |
1da177e4 LT |
1807 | __initcall(cpucache_init); |
1808 | ||
8bdec192 RA |
1809 | static noinline void |
1810 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1811 | { | |
1812 | struct kmem_list3 *l3; | |
1813 | struct slab *slabp; | |
1814 | unsigned long flags; | |
1815 | int node; | |
1816 | ||
1817 | printk(KERN_WARNING | |
1818 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1819 | nodeid, gfpflags); | |
1820 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1821 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1822 | |
1823 | for_each_online_node(node) { | |
1824 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1825 | unsigned long active_slabs = 0, num_slabs = 0; | |
1826 | ||
1827 | l3 = cachep->nodelists[node]; | |
1828 | if (!l3) | |
1829 | continue; | |
1830 | ||
1831 | spin_lock_irqsave(&l3->list_lock, flags); | |
1832 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1833 | active_objs += cachep->num; | |
1834 | active_slabs++; | |
1835 | } | |
1836 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1837 | active_objs += slabp->inuse; | |
1838 | active_slabs++; | |
1839 | } | |
1840 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1841 | num_slabs++; | |
1842 | ||
1843 | free_objects += l3->free_objects; | |
1844 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1845 | ||
1846 | num_slabs += active_slabs; | |
1847 | num_objs = num_slabs * cachep->num; | |
1848 | printk(KERN_WARNING | |
1849 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1850 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1851 | free_objects); | |
1852 | } | |
1853 | } | |
1854 | ||
1da177e4 LT |
1855 | /* |
1856 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1857 | * | |
1858 | * If we requested dmaable memory, we will get it. Even if we | |
1859 | * did not request dmaable memory, we might get it, but that | |
1860 | * would be relatively rare and ignorable. | |
1861 | */ | |
343e0d7a | 1862 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1863 | { |
1864 | struct page *page; | |
e1b6aa6f | 1865 | int nr_pages; |
1da177e4 LT |
1866 | int i; |
1867 | ||
d6fef9da | 1868 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1869 | /* |
1870 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1871 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1872 | */ |
e1b6aa6f | 1873 | flags |= __GFP_COMP; |
d6fef9da | 1874 | #endif |
765c4507 | 1875 | |
a618e89f | 1876 | flags |= cachep->allocflags; |
e12ba74d MG |
1877 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1878 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1879 | |
517d0869 | 1880 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1881 | if (!page) { |
1882 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1883 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1884 | return NULL; |
8bdec192 | 1885 | } |
1da177e4 | 1886 | |
072bb0aa MG |
1887 | /* Record if ALLOC_PFMEMALLOC was set when allocating the slab */ |
1888 | if (unlikely(page->pfmemalloc)) | |
1889 | pfmemalloc_active = true; | |
1890 | ||
e1b6aa6f | 1891 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1892 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1893 | add_zone_page_state(page_zone(page), |
1894 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1895 | else | |
1896 | add_zone_page_state(page_zone(page), | |
1897 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
072bb0aa | 1898 | for (i = 0; i < nr_pages; i++) { |
e1b6aa6f | 1899 | __SetPageSlab(page + i); |
c175eea4 | 1900 | |
072bb0aa MG |
1901 | if (page->pfmemalloc) |
1902 | SetPageSlabPfmemalloc(page + i); | |
1903 | } | |
1904 | ||
b1eeab67 VN |
1905 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1906 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1907 | ||
1908 | if (cachep->ctor) | |
1909 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1910 | else | |
1911 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1912 | } | |
c175eea4 | 1913 | |
e1b6aa6f | 1914 | return page_address(page); |
1da177e4 LT |
1915 | } |
1916 | ||
1917 | /* | |
1918 | * Interface to system's page release. | |
1919 | */ | |
343e0d7a | 1920 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1921 | { |
b28a02de | 1922 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1923 | struct page *page = virt_to_page(addr); |
1924 | const unsigned long nr_freed = i; | |
1925 | ||
b1eeab67 | 1926 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1927 | |
972d1a7b CL |
1928 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1929 | sub_zone_page_state(page_zone(page), | |
1930 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1931 | else | |
1932 | sub_zone_page_state(page_zone(page), | |
1933 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1934 | while (i--) { |
f205b2fe | 1935 | BUG_ON(!PageSlab(page)); |
072bb0aa | 1936 | __ClearPageSlabPfmemalloc(page); |
f205b2fe | 1937 | __ClearPageSlab(page); |
1da177e4 LT |
1938 | page++; |
1939 | } | |
1da177e4 LT |
1940 | if (current->reclaim_state) |
1941 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1942 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1943 | } |
1944 | ||
1945 | static void kmem_rcu_free(struct rcu_head *head) | |
1946 | { | |
b28a02de | 1947 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1948 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1949 | |
1950 | kmem_freepages(cachep, slab_rcu->addr); | |
1951 | if (OFF_SLAB(cachep)) | |
1952 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1953 | } | |
1954 | ||
1955 | #if DEBUG | |
1956 | ||
1957 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1958 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1959 | unsigned long caller) |
1da177e4 | 1960 | { |
8c138bc0 | 1961 | int size = cachep->object_size; |
1da177e4 | 1962 | |
3dafccf2 | 1963 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1964 | |
b28a02de | 1965 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1966 | return; |
1967 | ||
b28a02de PE |
1968 | *addr++ = 0x12345678; |
1969 | *addr++ = caller; | |
1970 | *addr++ = smp_processor_id(); | |
1971 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1972 | { |
1973 | unsigned long *sptr = &caller; | |
1974 | unsigned long svalue; | |
1975 | ||
1976 | while (!kstack_end(sptr)) { | |
1977 | svalue = *sptr++; | |
1978 | if (kernel_text_address(svalue)) { | |
b28a02de | 1979 | *addr++ = svalue; |
1da177e4 LT |
1980 | size -= sizeof(unsigned long); |
1981 | if (size <= sizeof(unsigned long)) | |
1982 | break; | |
1983 | } | |
1984 | } | |
1985 | ||
1986 | } | |
b28a02de | 1987 | *addr++ = 0x87654321; |
1da177e4 LT |
1988 | } |
1989 | #endif | |
1990 | ||
343e0d7a | 1991 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1992 | { |
8c138bc0 | 1993 | int size = cachep->object_size; |
3dafccf2 | 1994 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1995 | |
1996 | memset(addr, val, size); | |
b28a02de | 1997 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1998 | } |
1999 | ||
2000 | static void dump_line(char *data, int offset, int limit) | |
2001 | { | |
2002 | int i; | |
aa83aa40 DJ |
2003 | unsigned char error = 0; |
2004 | int bad_count = 0; | |
2005 | ||
fdde6abb | 2006 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
2007 | for (i = 0; i < limit; i++) { |
2008 | if (data[offset + i] != POISON_FREE) { | |
2009 | error = data[offset + i]; | |
2010 | bad_count++; | |
2011 | } | |
aa83aa40 | 2012 | } |
fdde6abb SAS |
2013 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
2014 | &data[offset], limit, 1); | |
aa83aa40 DJ |
2015 | |
2016 | if (bad_count == 1) { | |
2017 | error ^= POISON_FREE; | |
2018 | if (!(error & (error - 1))) { | |
2019 | printk(KERN_ERR "Single bit error detected. Probably " | |
2020 | "bad RAM.\n"); | |
2021 | #ifdef CONFIG_X86 | |
2022 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
2023 | "test tool.\n"); | |
2024 | #else | |
2025 | printk(KERN_ERR "Run a memory test tool.\n"); | |
2026 | #endif | |
2027 | } | |
2028 | } | |
1da177e4 LT |
2029 | } |
2030 | #endif | |
2031 | ||
2032 | #if DEBUG | |
2033 | ||
343e0d7a | 2034 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
2035 | { |
2036 | int i, size; | |
2037 | char *realobj; | |
2038 | ||
2039 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 2040 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
2041 | *dbg_redzone1(cachep, objp), |
2042 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2043 | } |
2044 | ||
2045 | if (cachep->flags & SLAB_STORE_USER) { | |
2046 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 2047 | *dbg_userword(cachep, objp)); |
1da177e4 | 2048 | print_symbol("(%s)", |
a737b3e2 | 2049 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
2050 | printk("\n"); |
2051 | } | |
3dafccf2 | 2052 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2053 | size = cachep->object_size; |
b28a02de | 2054 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
2055 | int limit; |
2056 | limit = 16; | |
b28a02de PE |
2057 | if (i + limit > size) |
2058 | limit = size - i; | |
1da177e4 LT |
2059 | dump_line(realobj, i, limit); |
2060 | } | |
2061 | } | |
2062 | ||
343e0d7a | 2063 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
2064 | { |
2065 | char *realobj; | |
2066 | int size, i; | |
2067 | int lines = 0; | |
2068 | ||
3dafccf2 | 2069 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2070 | size = cachep->object_size; |
1da177e4 | 2071 | |
b28a02de | 2072 | for (i = 0; i < size; i++) { |
1da177e4 | 2073 | char exp = POISON_FREE; |
b28a02de | 2074 | if (i == size - 1) |
1da177e4 LT |
2075 | exp = POISON_END; |
2076 | if (realobj[i] != exp) { | |
2077 | int limit; | |
2078 | /* Mismatch ! */ | |
2079 | /* Print header */ | |
2080 | if (lines == 0) { | |
b28a02de | 2081 | printk(KERN_ERR |
face37f5 DJ |
2082 | "Slab corruption (%s): %s start=%p, len=%d\n", |
2083 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
2084 | print_objinfo(cachep, objp, 0); |
2085 | } | |
2086 | /* Hexdump the affected line */ | |
b28a02de | 2087 | i = (i / 16) * 16; |
1da177e4 | 2088 | limit = 16; |
b28a02de PE |
2089 | if (i + limit > size) |
2090 | limit = size - i; | |
1da177e4 LT |
2091 | dump_line(realobj, i, limit); |
2092 | i += 16; | |
2093 | lines++; | |
2094 | /* Limit to 5 lines */ | |
2095 | if (lines > 5) | |
2096 | break; | |
2097 | } | |
2098 | } | |
2099 | if (lines != 0) { | |
2100 | /* Print some data about the neighboring objects, if they | |
2101 | * exist: | |
2102 | */ | |
6ed5eb22 | 2103 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 2104 | unsigned int objnr; |
1da177e4 | 2105 | |
8fea4e96 | 2106 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2107 | if (objnr) { |
8fea4e96 | 2108 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2109 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2110 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2111 | realobj, size); |
1da177e4 LT |
2112 | print_objinfo(cachep, objp, 2); |
2113 | } | |
b28a02de | 2114 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2115 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2116 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2117 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2118 | realobj, size); |
1da177e4 LT |
2119 | print_objinfo(cachep, objp, 2); |
2120 | } | |
2121 | } | |
2122 | } | |
2123 | #endif | |
2124 | ||
12dd36fa | 2125 | #if DEBUG |
e79aec29 | 2126 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2127 | { |
1da177e4 LT |
2128 | int i; |
2129 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2130 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2131 | |
2132 | if (cachep->flags & SLAB_POISON) { | |
2133 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2134 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2135 | OFF_SLAB(cachep)) |
b28a02de | 2136 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2137 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2138 | else |
2139 | check_poison_obj(cachep, objp); | |
2140 | #else | |
2141 | check_poison_obj(cachep, objp); | |
2142 | #endif | |
2143 | } | |
2144 | if (cachep->flags & SLAB_RED_ZONE) { | |
2145 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2146 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2147 | "was overwritten"); |
1da177e4 LT |
2148 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2149 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2150 | "was overwritten"); |
1da177e4 | 2151 | } |
1da177e4 | 2152 | } |
12dd36fa | 2153 | } |
1da177e4 | 2154 | #else |
e79aec29 | 2155 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2156 | { |
12dd36fa | 2157 | } |
1da177e4 LT |
2158 | #endif |
2159 | ||
911851e6 RD |
2160 | /** |
2161 | * slab_destroy - destroy and release all objects in a slab | |
2162 | * @cachep: cache pointer being destroyed | |
2163 | * @slabp: slab pointer being destroyed | |
2164 | * | |
12dd36fa | 2165 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2166 | * Before calling the slab must have been unlinked from the cache. The |
2167 | * cache-lock is not held/needed. | |
12dd36fa | 2168 | */ |
343e0d7a | 2169 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2170 | { |
2171 | void *addr = slabp->s_mem - slabp->colouroff; | |
2172 | ||
e79aec29 | 2173 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2174 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2175 | struct slab_rcu *slab_rcu; | |
2176 | ||
b28a02de | 2177 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2178 | slab_rcu->cachep = cachep; |
2179 | slab_rcu->addr = addr; | |
2180 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2181 | } else { | |
2182 | kmem_freepages(cachep, addr); | |
873623df IM |
2183 | if (OFF_SLAB(cachep)) |
2184 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2185 | } |
2186 | } | |
2187 | ||
117f6eb1 CL |
2188 | static void __kmem_cache_destroy(struct kmem_cache *cachep) |
2189 | { | |
2190 | int i; | |
2191 | struct kmem_list3 *l3; | |
2192 | ||
2193 | for_each_online_cpu(i) | |
2194 | kfree(cachep->array[i]); | |
2195 | ||
2196 | /* NUMA: free the list3 structures */ | |
2197 | for_each_online_node(i) { | |
2198 | l3 = cachep->nodelists[i]; | |
2199 | if (l3) { | |
2200 | kfree(l3->shared); | |
2201 | free_alien_cache(l3->alien); | |
2202 | kfree(l3); | |
2203 | } | |
2204 | } | |
2205 | kmem_cache_free(&cache_cache, cachep); | |
2206 | } | |
2207 | ||
2208 | ||
4d268eba | 2209 | /** |
a70773dd RD |
2210 | * calculate_slab_order - calculate size (page order) of slabs |
2211 | * @cachep: pointer to the cache that is being created | |
2212 | * @size: size of objects to be created in this cache. | |
2213 | * @align: required alignment for the objects. | |
2214 | * @flags: slab allocation flags | |
2215 | * | |
2216 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2217 | * |
2218 | * This could be made much more intelligent. For now, try to avoid using | |
2219 | * high order pages for slabs. When the gfp() functions are more friendly | |
2220 | * towards high-order requests, this should be changed. | |
2221 | */ | |
a737b3e2 | 2222 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2223 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2224 | { |
b1ab41c4 | 2225 | unsigned long offslab_limit; |
4d268eba | 2226 | size_t left_over = 0; |
9888e6fa | 2227 | int gfporder; |
4d268eba | 2228 | |
0aa817f0 | 2229 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2230 | unsigned int num; |
2231 | size_t remainder; | |
2232 | ||
9888e6fa | 2233 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2234 | if (!num) |
2235 | continue; | |
9888e6fa | 2236 | |
b1ab41c4 IM |
2237 | if (flags & CFLGS_OFF_SLAB) { |
2238 | /* | |
2239 | * Max number of objs-per-slab for caches which | |
2240 | * use off-slab slabs. Needed to avoid a possible | |
2241 | * looping condition in cache_grow(). | |
2242 | */ | |
2243 | offslab_limit = size - sizeof(struct slab); | |
2244 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2245 | ||
2246 | if (num > offslab_limit) | |
2247 | break; | |
2248 | } | |
4d268eba | 2249 | |
9888e6fa | 2250 | /* Found something acceptable - save it away */ |
4d268eba | 2251 | cachep->num = num; |
9888e6fa | 2252 | cachep->gfporder = gfporder; |
4d268eba PE |
2253 | left_over = remainder; |
2254 | ||
f78bb8ad LT |
2255 | /* |
2256 | * A VFS-reclaimable slab tends to have most allocations | |
2257 | * as GFP_NOFS and we really don't want to have to be allocating | |
2258 | * higher-order pages when we are unable to shrink dcache. | |
2259 | */ | |
2260 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2261 | break; | |
2262 | ||
4d268eba PE |
2263 | /* |
2264 | * Large number of objects is good, but very large slabs are | |
2265 | * currently bad for the gfp()s. | |
2266 | */ | |
543585cc | 2267 | if (gfporder >= slab_max_order) |
4d268eba PE |
2268 | break; |
2269 | ||
9888e6fa LT |
2270 | /* |
2271 | * Acceptable internal fragmentation? | |
2272 | */ | |
a737b3e2 | 2273 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2274 | break; |
2275 | } | |
2276 | return left_over; | |
2277 | } | |
2278 | ||
83b519e8 | 2279 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2280 | { |
97d06609 | 2281 | if (slab_state >= FULL) |
83b519e8 | 2282 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2283 | |
97d06609 | 2284 | if (slab_state == DOWN) { |
f30cf7d1 PE |
2285 | /* |
2286 | * Note: the first kmem_cache_create must create the cache | |
2287 | * that's used by kmalloc(24), otherwise the creation of | |
2288 | * further caches will BUG(). | |
2289 | */ | |
2290 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2291 | ||
2292 | /* | |
2293 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2294 | * the first cache, then we need to set up all its list3s, | |
2295 | * otherwise the creation of further caches will BUG(). | |
2296 | */ | |
2297 | set_up_list3s(cachep, SIZE_AC); | |
2298 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2299 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2300 | else |
97d06609 | 2301 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 PE |
2302 | } else { |
2303 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2304 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2305 | |
97d06609 | 2306 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2307 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2308 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2309 | } else { |
2310 | int node; | |
556a169d | 2311 | for_each_online_node(node) { |
f30cf7d1 PE |
2312 | cachep->nodelists[node] = |
2313 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2314 | gfp, node); |
f30cf7d1 PE |
2315 | BUG_ON(!cachep->nodelists[node]); |
2316 | kmem_list3_init(cachep->nodelists[node]); | |
2317 | } | |
2318 | } | |
2319 | } | |
7d6e6d09 | 2320 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2321 | jiffies + REAPTIMEOUT_LIST3 + |
2322 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2323 | ||
2324 | cpu_cache_get(cachep)->avail = 0; | |
2325 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2326 | cpu_cache_get(cachep)->batchcount = 1; | |
2327 | cpu_cache_get(cachep)->touched = 0; | |
2328 | cachep->batchcount = 1; | |
2329 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2330 | return 0; |
f30cf7d1 PE |
2331 | } |
2332 | ||
1da177e4 | 2333 | /** |
039363f3 | 2334 | * __kmem_cache_create - Create a cache. |
1da177e4 LT |
2335 | * @name: A string which is used in /proc/slabinfo to identify this cache. |
2336 | * @size: The size of objects to be created in this cache. | |
2337 | * @align: The required alignment for the objects. | |
2338 | * @flags: SLAB flags | |
2339 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2340 | * |
2341 | * Returns a ptr to the cache on success, NULL on failure. | |
2342 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2343 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2344 | * |
2345 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2346 | * the module calling this has to destroy the cache before getting unloaded. |
2347 | * | |
1da177e4 LT |
2348 | * The flags are |
2349 | * | |
2350 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2351 | * to catch references to uninitialised memory. | |
2352 | * | |
2353 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2354 | * for buffer overruns. | |
2355 | * | |
1da177e4 LT |
2356 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2357 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2358 | * as davem. | |
2359 | */ | |
343e0d7a | 2360 | struct kmem_cache * |
039363f3 | 2361 | __kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2362 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2363 | { |
2364 | size_t left_over, slab_size, ralign; | |
20cea968 | 2365 | struct kmem_cache *cachep = NULL; |
83b519e8 | 2366 | gfp_t gfp; |
1da177e4 | 2367 | |
1da177e4 | 2368 | #if DEBUG |
1da177e4 LT |
2369 | #if FORCED_DEBUG |
2370 | /* | |
2371 | * Enable redzoning and last user accounting, except for caches with | |
2372 | * large objects, if the increased size would increase the object size | |
2373 | * above the next power of two: caches with object sizes just above a | |
2374 | * power of two have a significant amount of internal fragmentation. | |
2375 | */ | |
87a927c7 DW |
2376 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2377 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2378 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2379 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2380 | flags |= SLAB_POISON; | |
2381 | #endif | |
2382 | if (flags & SLAB_DESTROY_BY_RCU) | |
2383 | BUG_ON(flags & SLAB_POISON); | |
2384 | #endif | |
1da177e4 | 2385 | /* |
a737b3e2 AM |
2386 | * Always checks flags, a caller might be expecting debug support which |
2387 | * isn't available. | |
1da177e4 | 2388 | */ |
40094fa6 | 2389 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2390 | |
a737b3e2 AM |
2391 | /* |
2392 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2393 | * unaligned accesses for some archs when redzoning is used, and makes |
2394 | * sure any on-slab bufctl's are also correctly aligned. | |
2395 | */ | |
b28a02de PE |
2396 | if (size & (BYTES_PER_WORD - 1)) { |
2397 | size += (BYTES_PER_WORD - 1); | |
2398 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2399 | } |
2400 | ||
a737b3e2 AM |
2401 | /* calculate the final buffer alignment: */ |
2402 | ||
1da177e4 LT |
2403 | /* 1) arch recommendation: can be overridden for debug */ |
2404 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2405 | /* |
2406 | * Default alignment: as specified by the arch code. Except if | |
2407 | * an object is really small, then squeeze multiple objects into | |
2408 | * one cacheline. | |
1da177e4 LT |
2409 | */ |
2410 | ralign = cache_line_size(); | |
b28a02de | 2411 | while (size <= ralign / 2) |
1da177e4 LT |
2412 | ralign /= 2; |
2413 | } else { | |
2414 | ralign = BYTES_PER_WORD; | |
2415 | } | |
ca5f9703 PE |
2416 | |
2417 | /* | |
87a927c7 DW |
2418 | * Redzoning and user store require word alignment or possibly larger. |
2419 | * Note this will be overridden by architecture or caller mandated | |
2420 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2421 | */ |
87a927c7 DW |
2422 | if (flags & SLAB_STORE_USER) |
2423 | ralign = BYTES_PER_WORD; | |
2424 | ||
2425 | if (flags & SLAB_RED_ZONE) { | |
2426 | ralign = REDZONE_ALIGN; | |
2427 | /* If redzoning, ensure that the second redzone is suitably | |
2428 | * aligned, by adjusting the object size accordingly. */ | |
2429 | size += REDZONE_ALIGN - 1; | |
2430 | size &= ~(REDZONE_ALIGN - 1); | |
2431 | } | |
ca5f9703 | 2432 | |
a44b56d3 | 2433 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2434 | if (ralign < ARCH_SLAB_MINALIGN) { |
2435 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2436 | } |
a44b56d3 | 2437 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2438 | if (ralign < align) { |
2439 | ralign = align; | |
1da177e4 | 2440 | } |
3ff84a7f PE |
2441 | /* disable debug if necessary */ |
2442 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2443 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2444 | /* |
ca5f9703 | 2445 | * 4) Store it. |
1da177e4 LT |
2446 | */ |
2447 | align = ralign; | |
2448 | ||
83b519e8 PE |
2449 | if (slab_is_available()) |
2450 | gfp = GFP_KERNEL; | |
2451 | else | |
2452 | gfp = GFP_NOWAIT; | |
2453 | ||
1da177e4 | 2454 | /* Get cache's description obj. */ |
83b519e8 | 2455 | cachep = kmem_cache_zalloc(&cache_cache, gfp); |
1da177e4 | 2456 | if (!cachep) |
039363f3 | 2457 | return NULL; |
1da177e4 | 2458 | |
b56efcf0 | 2459 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
3b0efdfa CL |
2460 | cachep->object_size = size; |
2461 | cachep->align = align; | |
1da177e4 | 2462 | #if DEBUG |
1da177e4 | 2463 | |
ca5f9703 PE |
2464 | /* |
2465 | * Both debugging options require word-alignment which is calculated | |
2466 | * into align above. | |
2467 | */ | |
1da177e4 | 2468 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2469 | /* add space for red zone words */ |
3ff84a7f PE |
2470 | cachep->obj_offset += sizeof(unsigned long long); |
2471 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2472 | } |
2473 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2474 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2475 | * the real object. But if the second red zone needs to be |
2476 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2477 | */ |
87a927c7 DW |
2478 | if (flags & SLAB_RED_ZONE) |
2479 | size += REDZONE_ALIGN; | |
2480 | else | |
2481 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2482 | } |
2483 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2484 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3b0efdfa | 2485 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
1ab335d8 | 2486 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); |
1da177e4 LT |
2487 | size = PAGE_SIZE; |
2488 | } | |
2489 | #endif | |
2490 | #endif | |
2491 | ||
e0a42726 IM |
2492 | /* |
2493 | * Determine if the slab management is 'on' or 'off' slab. | |
2494 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2495 | * it too early on. Always use on-slab management when |
2496 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2497 | */ |
e7cb55b9 CM |
2498 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2499 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2500 | /* |
2501 | * Size is large, assume best to place the slab management obj | |
2502 | * off-slab (should allow better packing of objs). | |
2503 | */ | |
2504 | flags |= CFLGS_OFF_SLAB; | |
2505 | ||
2506 | size = ALIGN(size, align); | |
2507 | ||
f78bb8ad | 2508 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2509 | |
2510 | if (!cachep->num) { | |
b4169525 | 2511 | printk(KERN_ERR |
2512 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
1da177e4 | 2513 | kmem_cache_free(&cache_cache, cachep); |
039363f3 | 2514 | return NULL; |
1da177e4 | 2515 | } |
b28a02de PE |
2516 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2517 | + sizeof(struct slab), align); | |
1da177e4 LT |
2518 | |
2519 | /* | |
2520 | * If the slab has been placed off-slab, and we have enough space then | |
2521 | * move it on-slab. This is at the expense of any extra colouring. | |
2522 | */ | |
2523 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2524 | flags &= ~CFLGS_OFF_SLAB; | |
2525 | left_over -= slab_size; | |
2526 | } | |
2527 | ||
2528 | if (flags & CFLGS_OFF_SLAB) { | |
2529 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2530 | slab_size = |
2531 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2532 | |
2533 | #ifdef CONFIG_PAGE_POISONING | |
2534 | /* If we're going to use the generic kernel_map_pages() | |
2535 | * poisoning, then it's going to smash the contents of | |
2536 | * the redzone and userword anyhow, so switch them off. | |
2537 | */ | |
2538 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2539 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2540 | #endif | |
1da177e4 LT |
2541 | } |
2542 | ||
2543 | cachep->colour_off = cache_line_size(); | |
2544 | /* Offset must be a multiple of the alignment. */ | |
2545 | if (cachep->colour_off < align) | |
2546 | cachep->colour_off = align; | |
b28a02de | 2547 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2548 | cachep->slab_size = slab_size; |
2549 | cachep->flags = flags; | |
a618e89f | 2550 | cachep->allocflags = 0; |
4b51d669 | 2551 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2552 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2553 | cachep->size = size; |
6a2d7a95 | 2554 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2555 | |
e5ac9c5a | 2556 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2557 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2558 | /* |
2559 | * This is a possibility for one of the malloc_sizes caches. | |
2560 | * But since we go off slab only for object size greater than | |
2561 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2562 | * this should not happen at all. | |
2563 | * But leave a BUG_ON for some lucky dude. | |
2564 | */ | |
6cb8f913 | 2565 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2566 | } |
1da177e4 | 2567 | cachep->ctor = ctor; |
1da177e4 LT |
2568 | cachep->name = name; |
2569 | ||
83b519e8 | 2570 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef | 2571 | __kmem_cache_destroy(cachep); |
039363f3 | 2572 | return NULL; |
2ed3a4ef | 2573 | } |
1da177e4 | 2574 | |
83835b3d PZ |
2575 | if (flags & SLAB_DEBUG_OBJECTS) { |
2576 | /* | |
2577 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2578 | * debug_object_activate()->kmem_cache_alloc(). | |
2579 | */ | |
2580 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2581 | ||
2582 | slab_set_debugobj_lock_classes(cachep); | |
2583 | } | |
2584 | ||
1da177e4 | 2585 | /* cache setup completed, link it into the list */ |
18004c5d | 2586 | list_add(&cachep->list, &slab_caches); |
1da177e4 LT |
2587 | return cachep; |
2588 | } | |
1da177e4 LT |
2589 | |
2590 | #if DEBUG | |
2591 | static void check_irq_off(void) | |
2592 | { | |
2593 | BUG_ON(!irqs_disabled()); | |
2594 | } | |
2595 | ||
2596 | static void check_irq_on(void) | |
2597 | { | |
2598 | BUG_ON(irqs_disabled()); | |
2599 | } | |
2600 | ||
343e0d7a | 2601 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2602 | { |
2603 | #ifdef CONFIG_SMP | |
2604 | check_irq_off(); | |
7d6e6d09 | 2605 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2606 | #endif |
2607 | } | |
e498be7d | 2608 | |
343e0d7a | 2609 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2610 | { |
2611 | #ifdef CONFIG_SMP | |
2612 | check_irq_off(); | |
2613 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2614 | #endif | |
2615 | } | |
2616 | ||
1da177e4 LT |
2617 | #else |
2618 | #define check_irq_off() do { } while(0) | |
2619 | #define check_irq_on() do { } while(0) | |
2620 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2621 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2622 | #endif |
2623 | ||
aab2207c CL |
2624 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2625 | struct array_cache *ac, | |
2626 | int force, int node); | |
2627 | ||
1da177e4 LT |
2628 | static void do_drain(void *arg) |
2629 | { | |
a737b3e2 | 2630 | struct kmem_cache *cachep = arg; |
1da177e4 | 2631 | struct array_cache *ac; |
7d6e6d09 | 2632 | int node = numa_mem_id(); |
1da177e4 LT |
2633 | |
2634 | check_irq_off(); | |
9a2dba4b | 2635 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2636 | spin_lock(&cachep->nodelists[node]->list_lock); |
2637 | free_block(cachep, ac->entry, ac->avail, node); | |
2638 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2639 | ac->avail = 0; |
2640 | } | |
2641 | ||
343e0d7a | 2642 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2643 | { |
e498be7d CL |
2644 | struct kmem_list3 *l3; |
2645 | int node; | |
2646 | ||
15c8b6c1 | 2647 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2648 | check_irq_on(); |
b28a02de | 2649 | for_each_online_node(node) { |
e498be7d | 2650 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2651 | if (l3 && l3->alien) |
2652 | drain_alien_cache(cachep, l3->alien); | |
2653 | } | |
2654 | ||
2655 | for_each_online_node(node) { | |
2656 | l3 = cachep->nodelists[node]; | |
2657 | if (l3) | |
aab2207c | 2658 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2659 | } |
1da177e4 LT |
2660 | } |
2661 | ||
ed11d9eb CL |
2662 | /* |
2663 | * Remove slabs from the list of free slabs. | |
2664 | * Specify the number of slabs to drain in tofree. | |
2665 | * | |
2666 | * Returns the actual number of slabs released. | |
2667 | */ | |
2668 | static int drain_freelist(struct kmem_cache *cache, | |
2669 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2670 | { |
ed11d9eb CL |
2671 | struct list_head *p; |
2672 | int nr_freed; | |
1da177e4 | 2673 | struct slab *slabp; |
1da177e4 | 2674 | |
ed11d9eb CL |
2675 | nr_freed = 0; |
2676 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2677 | |
ed11d9eb | 2678 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2679 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2680 | if (p == &l3->slabs_free) { |
2681 | spin_unlock_irq(&l3->list_lock); | |
2682 | goto out; | |
2683 | } | |
1da177e4 | 2684 | |
ed11d9eb | 2685 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2686 | #if DEBUG |
40094fa6 | 2687 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2688 | #endif |
2689 | list_del(&slabp->list); | |
ed11d9eb CL |
2690 | /* |
2691 | * Safe to drop the lock. The slab is no longer linked | |
2692 | * to the cache. | |
2693 | */ | |
2694 | l3->free_objects -= cache->num; | |
e498be7d | 2695 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2696 | slab_destroy(cache, slabp); |
2697 | nr_freed++; | |
1da177e4 | 2698 | } |
ed11d9eb CL |
2699 | out: |
2700 | return nr_freed; | |
1da177e4 LT |
2701 | } |
2702 | ||
18004c5d | 2703 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2704 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2705 | { |
2706 | int ret = 0, i = 0; | |
2707 | struct kmem_list3 *l3; | |
2708 | ||
2709 | drain_cpu_caches(cachep); | |
2710 | ||
2711 | check_irq_on(); | |
2712 | for_each_online_node(i) { | |
2713 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2714 | if (!l3) |
2715 | continue; | |
2716 | ||
2717 | drain_freelist(cachep, l3, l3->free_objects); | |
2718 | ||
2719 | ret += !list_empty(&l3->slabs_full) || | |
2720 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2721 | } |
2722 | return (ret ? 1 : 0); | |
2723 | } | |
2724 | ||
1da177e4 LT |
2725 | /** |
2726 | * kmem_cache_shrink - Shrink a cache. | |
2727 | * @cachep: The cache to shrink. | |
2728 | * | |
2729 | * Releases as many slabs as possible for a cache. | |
2730 | * To help debugging, a zero exit status indicates all slabs were released. | |
2731 | */ | |
343e0d7a | 2732 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2733 | { |
8f5be20b | 2734 | int ret; |
40094fa6 | 2735 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2736 | |
95402b38 | 2737 | get_online_cpus(); |
18004c5d | 2738 | mutex_lock(&slab_mutex); |
8f5be20b | 2739 | ret = __cache_shrink(cachep); |
18004c5d | 2740 | mutex_unlock(&slab_mutex); |
95402b38 | 2741 | put_online_cpus(); |
8f5be20b | 2742 | return ret; |
1da177e4 LT |
2743 | } |
2744 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2745 | ||
2746 | /** | |
2747 | * kmem_cache_destroy - delete a cache | |
2748 | * @cachep: the cache to destroy | |
2749 | * | |
72fd4a35 | 2750 | * Remove a &struct kmem_cache object from the slab cache. |
1da177e4 LT |
2751 | * |
2752 | * It is expected this function will be called by a module when it is | |
2753 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2754 | * cache being allocated each time a module is loaded and unloaded, if the | |
2755 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2756 | * | |
2757 | * The cache must be empty before calling this function. | |
2758 | * | |
25985edc | 2759 | * The caller must guarantee that no one will allocate memory from the cache |
1da177e4 LT |
2760 | * during the kmem_cache_destroy(). |
2761 | */ | |
133d205a | 2762 | void kmem_cache_destroy(struct kmem_cache *cachep) |
1da177e4 | 2763 | { |
40094fa6 | 2764 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2765 | |
1da177e4 | 2766 | /* Find the cache in the chain of caches. */ |
95402b38 | 2767 | get_online_cpus(); |
18004c5d | 2768 | mutex_lock(&slab_mutex); |
1da177e4 LT |
2769 | /* |
2770 | * the chain is never empty, cache_cache is never destroyed | |
2771 | */ | |
3b0efdfa | 2772 | list_del(&cachep->list); |
1da177e4 LT |
2773 | if (__cache_shrink(cachep)) { |
2774 | slab_error(cachep, "Can't free all objects"); | |
18004c5d CL |
2775 | list_add(&cachep->list, &slab_caches); |
2776 | mutex_unlock(&slab_mutex); | |
95402b38 | 2777 | put_online_cpus(); |
133d205a | 2778 | return; |
1da177e4 LT |
2779 | } |
2780 | ||
2781 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
7ed9f7e5 | 2782 | rcu_barrier(); |
1da177e4 | 2783 | |
117f6eb1 | 2784 | __kmem_cache_destroy(cachep); |
18004c5d | 2785 | mutex_unlock(&slab_mutex); |
95402b38 | 2786 | put_online_cpus(); |
1da177e4 LT |
2787 | } |
2788 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2789 | ||
e5ac9c5a RT |
2790 | /* |
2791 | * Get the memory for a slab management obj. | |
2792 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2793 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2794 | * come from the same cache which is getting created because, | |
2795 | * when we are searching for an appropriate cache for these | |
2796 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2797 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2798 | * kmem_find_general_cachep till the initialization is complete. | |
2799 | * Hence we cannot have slabp_cache same as the original cache. | |
2800 | */ | |
343e0d7a | 2801 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2802 | int colour_off, gfp_t local_flags, |
2803 | int nodeid) | |
1da177e4 LT |
2804 | { |
2805 | struct slab *slabp; | |
b28a02de | 2806 | |
1da177e4 LT |
2807 | if (OFF_SLAB(cachep)) { |
2808 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2809 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2810 | local_flags, nodeid); |
d5cff635 CM |
2811 | /* |
2812 | * If the first object in the slab is leaked (it's allocated | |
2813 | * but no one has a reference to it), we want to make sure | |
2814 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2815 | * to the object. Otherwise we will not report the leak. | |
2816 | */ | |
c017b4be CM |
2817 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2818 | local_flags); | |
1da177e4 LT |
2819 | if (!slabp) |
2820 | return NULL; | |
2821 | } else { | |
b28a02de | 2822 | slabp = objp + colour_off; |
1da177e4 LT |
2823 | colour_off += cachep->slab_size; |
2824 | } | |
2825 | slabp->inuse = 0; | |
2826 | slabp->colouroff = colour_off; | |
b28a02de | 2827 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2828 | slabp->nodeid = nodeid; |
e51bfd0a | 2829 | slabp->free = 0; |
1da177e4 LT |
2830 | return slabp; |
2831 | } | |
2832 | ||
2833 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2834 | { | |
b28a02de | 2835 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2836 | } |
2837 | ||
343e0d7a | 2838 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2839 | struct slab *slabp) |
1da177e4 LT |
2840 | { |
2841 | int i; | |
2842 | ||
2843 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2844 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2845 | #if DEBUG |
2846 | /* need to poison the objs? */ | |
2847 | if (cachep->flags & SLAB_POISON) | |
2848 | poison_obj(cachep, objp, POISON_FREE); | |
2849 | if (cachep->flags & SLAB_STORE_USER) | |
2850 | *dbg_userword(cachep, objp) = NULL; | |
2851 | ||
2852 | if (cachep->flags & SLAB_RED_ZONE) { | |
2853 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2854 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2855 | } | |
2856 | /* | |
a737b3e2 AM |
2857 | * Constructors are not allowed to allocate memory from the same |
2858 | * cache which they are a constructor for. Otherwise, deadlock. | |
2859 | * They must also be threaded. | |
1da177e4 LT |
2860 | */ |
2861 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2862 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2863 | |
2864 | if (cachep->flags & SLAB_RED_ZONE) { | |
2865 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2866 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2867 | " end of an object"); |
1da177e4 LT |
2868 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2869 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2870 | " start of an object"); |
1da177e4 | 2871 | } |
3b0efdfa | 2872 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2873 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2874 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2875 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2876 | #else |
2877 | if (cachep->ctor) | |
51cc5068 | 2878 | cachep->ctor(objp); |
1da177e4 | 2879 | #endif |
b28a02de | 2880 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2881 | } |
b28a02de | 2882 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2883 | } |
2884 | ||
343e0d7a | 2885 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2886 | { |
4b51d669 CL |
2887 | if (CONFIG_ZONE_DMA_FLAG) { |
2888 | if (flags & GFP_DMA) | |
a618e89f | 2889 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2890 | else |
a618e89f | 2891 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2892 | } |
1da177e4 LT |
2893 | } |
2894 | ||
a737b3e2 AM |
2895 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2896 | int nodeid) | |
78d382d7 | 2897 | { |
8fea4e96 | 2898 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2899 | kmem_bufctl_t next; |
2900 | ||
2901 | slabp->inuse++; | |
2902 | next = slab_bufctl(slabp)[slabp->free]; | |
2903 | #if DEBUG | |
2904 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2905 | WARN_ON(slabp->nodeid != nodeid); | |
2906 | #endif | |
2907 | slabp->free = next; | |
2908 | ||
2909 | return objp; | |
2910 | } | |
2911 | ||
a737b3e2 AM |
2912 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2913 | void *objp, int nodeid) | |
78d382d7 | 2914 | { |
8fea4e96 | 2915 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2916 | |
2917 | #if DEBUG | |
2918 | /* Verify that the slab belongs to the intended node */ | |
2919 | WARN_ON(slabp->nodeid != nodeid); | |
2920 | ||
871751e2 | 2921 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2922 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2923 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2924 | BUG(); |
2925 | } | |
2926 | #endif | |
2927 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2928 | slabp->free = objnr; | |
2929 | slabp->inuse--; | |
2930 | } | |
2931 | ||
4776874f PE |
2932 | /* |
2933 | * Map pages beginning at addr to the given cache and slab. This is required | |
2934 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2935 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2936 | */ |
2937 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2938 | void *addr) | |
1da177e4 | 2939 | { |
4776874f | 2940 | int nr_pages; |
1da177e4 LT |
2941 | struct page *page; |
2942 | ||
4776874f | 2943 | page = virt_to_page(addr); |
84097518 | 2944 | |
4776874f | 2945 | nr_pages = 1; |
84097518 | 2946 | if (likely(!PageCompound(page))) |
4776874f PE |
2947 | nr_pages <<= cache->gfporder; |
2948 | ||
1da177e4 | 2949 | do { |
35026088 CL |
2950 | page->slab_cache = cache; |
2951 | page->slab_page = slab; | |
1da177e4 | 2952 | page++; |
4776874f | 2953 | } while (--nr_pages); |
1da177e4 LT |
2954 | } |
2955 | ||
2956 | /* | |
2957 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2958 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2959 | */ | |
3c517a61 CL |
2960 | static int cache_grow(struct kmem_cache *cachep, |
2961 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2962 | { |
b28a02de | 2963 | struct slab *slabp; |
b28a02de PE |
2964 | size_t offset; |
2965 | gfp_t local_flags; | |
e498be7d | 2966 | struct kmem_list3 *l3; |
1da177e4 | 2967 | |
a737b3e2 AM |
2968 | /* |
2969 | * Be lazy and only check for valid flags here, keeping it out of the | |
2970 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2971 | */ |
6cb06229 CL |
2972 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2973 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2974 | |
2e1217cf | 2975 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2976 | check_irq_off(); |
2e1217cf RT |
2977 | l3 = cachep->nodelists[nodeid]; |
2978 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2979 | |
2980 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2981 | offset = l3->colour_next; |
2982 | l3->colour_next++; | |
2983 | if (l3->colour_next >= cachep->colour) | |
2984 | l3->colour_next = 0; | |
2985 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2986 | |
2e1217cf | 2987 | offset *= cachep->colour_off; |
1da177e4 LT |
2988 | |
2989 | if (local_flags & __GFP_WAIT) | |
2990 | local_irq_enable(); | |
2991 | ||
2992 | /* | |
2993 | * The test for missing atomic flag is performed here, rather than | |
2994 | * the more obvious place, simply to reduce the critical path length | |
2995 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2996 | * will eventually be caught here (where it matters). | |
2997 | */ | |
2998 | kmem_flagcheck(cachep, flags); | |
2999 | ||
a737b3e2 AM |
3000 | /* |
3001 | * Get mem for the objs. Attempt to allocate a physical page from | |
3002 | * 'nodeid'. | |
e498be7d | 3003 | */ |
3c517a61 | 3004 | if (!objp) |
b8c1c5da | 3005 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 3006 | if (!objp) |
1da177e4 LT |
3007 | goto failed; |
3008 | ||
3009 | /* Get slab management. */ | |
3c517a61 | 3010 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 3011 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 3012 | if (!slabp) |
1da177e4 LT |
3013 | goto opps1; |
3014 | ||
4776874f | 3015 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 3016 | |
a35afb83 | 3017 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
3018 | |
3019 | if (local_flags & __GFP_WAIT) | |
3020 | local_irq_disable(); | |
3021 | check_irq_off(); | |
e498be7d | 3022 | spin_lock(&l3->list_lock); |
1da177e4 LT |
3023 | |
3024 | /* Make slab active. */ | |
e498be7d | 3025 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 3026 | STATS_INC_GROWN(cachep); |
e498be7d CL |
3027 | l3->free_objects += cachep->num; |
3028 | spin_unlock(&l3->list_lock); | |
1da177e4 | 3029 | return 1; |
a737b3e2 | 3030 | opps1: |
1da177e4 | 3031 | kmem_freepages(cachep, objp); |
a737b3e2 | 3032 | failed: |
1da177e4 LT |
3033 | if (local_flags & __GFP_WAIT) |
3034 | local_irq_disable(); | |
3035 | return 0; | |
3036 | } | |
3037 | ||
3038 | #if DEBUG | |
3039 | ||
3040 | /* | |
3041 | * Perform extra freeing checks: | |
3042 | * - detect bad pointers. | |
3043 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
3044 | */ |
3045 | static void kfree_debugcheck(const void *objp) | |
3046 | { | |
1da177e4 LT |
3047 | if (!virt_addr_valid(objp)) { |
3048 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
3049 | (unsigned long)objp); |
3050 | BUG(); | |
1da177e4 | 3051 | } |
1da177e4 LT |
3052 | } |
3053 | ||
58ce1fd5 PE |
3054 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
3055 | { | |
b46b8f19 | 3056 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
3057 | |
3058 | redzone1 = *dbg_redzone1(cache, obj); | |
3059 | redzone2 = *dbg_redzone2(cache, obj); | |
3060 | ||
3061 | /* | |
3062 | * Redzone is ok. | |
3063 | */ | |
3064 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3065 | return; | |
3066 | ||
3067 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3068 | slab_error(cache, "double free detected"); | |
3069 | else | |
3070 | slab_error(cache, "memory outside object was overwritten"); | |
3071 | ||
b46b8f19 | 3072 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
3073 | obj, redzone1, redzone2); |
3074 | } | |
3075 | ||
343e0d7a | 3076 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 3077 | void *caller) |
1da177e4 LT |
3078 | { |
3079 | struct page *page; | |
3080 | unsigned int objnr; | |
3081 | struct slab *slabp; | |
3082 | ||
80cbd911 MW |
3083 | BUG_ON(virt_to_cache(objp) != cachep); |
3084 | ||
3dafccf2 | 3085 | objp -= obj_offset(cachep); |
1da177e4 | 3086 | kfree_debugcheck(objp); |
b49af68f | 3087 | page = virt_to_head_page(objp); |
1da177e4 | 3088 | |
35026088 | 3089 | slabp = page->slab_page; |
1da177e4 LT |
3090 | |
3091 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3092 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3093 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3094 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3095 | } | |
3096 | if (cachep->flags & SLAB_STORE_USER) | |
3097 | *dbg_userword(cachep, objp) = caller; | |
3098 | ||
8fea4e96 | 3099 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3100 | |
3101 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3102 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3103 | |
871751e2 AV |
3104 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3105 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3106 | #endif | |
1da177e4 LT |
3107 | if (cachep->flags & SLAB_POISON) { |
3108 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3109 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 3110 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 3111 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3112 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3113 | } else { |
3114 | poison_obj(cachep, objp, POISON_FREE); | |
3115 | } | |
3116 | #else | |
3117 | poison_obj(cachep, objp, POISON_FREE); | |
3118 | #endif | |
3119 | } | |
3120 | return objp; | |
3121 | } | |
3122 | ||
343e0d7a | 3123 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3124 | { |
3125 | kmem_bufctl_t i; | |
3126 | int entries = 0; | |
b28a02de | 3127 | |
1da177e4 LT |
3128 | /* Check slab's freelist to see if this obj is there. */ |
3129 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3130 | entries++; | |
3131 | if (entries > cachep->num || i >= cachep->num) | |
3132 | goto bad; | |
3133 | } | |
3134 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3135 | bad: |
3136 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3137 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3138 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3139 | print_tainted()); | |
fdde6abb SAS |
3140 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3141 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3142 | 1); | |
1da177e4 LT |
3143 | BUG(); |
3144 | } | |
3145 | } | |
3146 | #else | |
3147 | #define kfree_debugcheck(x) do { } while(0) | |
3148 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3149 | #define check_slabp(x,y) do { } while(0) | |
3150 | #endif | |
3151 | ||
072bb0aa MG |
3152 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
3153 | bool force_refill) | |
1da177e4 LT |
3154 | { |
3155 | int batchcount; | |
3156 | struct kmem_list3 *l3; | |
3157 | struct array_cache *ac; | |
1ca4cb24 PE |
3158 | int node; |
3159 | ||
1da177e4 | 3160 | check_irq_off(); |
7d6e6d09 | 3161 | node = numa_mem_id(); |
072bb0aa MG |
3162 | if (unlikely(force_refill)) |
3163 | goto force_grow; | |
3164 | retry: | |
9a2dba4b | 3165 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3166 | batchcount = ac->batchcount; |
3167 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3168 | /* |
3169 | * If there was little recent activity on this cache, then | |
3170 | * perform only a partial refill. Otherwise we could generate | |
3171 | * refill bouncing. | |
1da177e4 LT |
3172 | */ |
3173 | batchcount = BATCHREFILL_LIMIT; | |
3174 | } | |
1ca4cb24 | 3175 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3176 | |
3177 | BUG_ON(ac->avail > 0 || !l3); | |
3178 | spin_lock(&l3->list_lock); | |
1da177e4 | 3179 | |
3ded175a | 3180 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3181 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3182 | l3->shared->touched = 1; | |
3ded175a | 3183 | goto alloc_done; |
44b57f1c | 3184 | } |
3ded175a | 3185 | |
1da177e4 LT |
3186 | while (batchcount > 0) { |
3187 | struct list_head *entry; | |
3188 | struct slab *slabp; | |
3189 | /* Get slab alloc is to come from. */ | |
3190 | entry = l3->slabs_partial.next; | |
3191 | if (entry == &l3->slabs_partial) { | |
3192 | l3->free_touched = 1; | |
3193 | entry = l3->slabs_free.next; | |
3194 | if (entry == &l3->slabs_free) | |
3195 | goto must_grow; | |
3196 | } | |
3197 | ||
3198 | slabp = list_entry(entry, struct slab, list); | |
3199 | check_slabp(cachep, slabp); | |
3200 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3201 | |
3202 | /* | |
3203 | * The slab was either on partial or free list so | |
3204 | * there must be at least one object available for | |
3205 | * allocation. | |
3206 | */ | |
249b9f33 | 3207 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3208 | |
1da177e4 | 3209 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3210 | STATS_INC_ALLOCED(cachep); |
3211 | STATS_INC_ACTIVE(cachep); | |
3212 | STATS_SET_HIGH(cachep); | |
3213 | ||
072bb0aa MG |
3214 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, |
3215 | node)); | |
1da177e4 LT |
3216 | } |
3217 | check_slabp(cachep, slabp); | |
3218 | ||
3219 | /* move slabp to correct slabp list: */ | |
3220 | list_del(&slabp->list); | |
3221 | if (slabp->free == BUFCTL_END) | |
3222 | list_add(&slabp->list, &l3->slabs_full); | |
3223 | else | |
3224 | list_add(&slabp->list, &l3->slabs_partial); | |
3225 | } | |
3226 | ||
a737b3e2 | 3227 | must_grow: |
1da177e4 | 3228 | l3->free_objects -= ac->avail; |
a737b3e2 | 3229 | alloc_done: |
e498be7d | 3230 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3231 | |
3232 | if (unlikely(!ac->avail)) { | |
3233 | int x; | |
072bb0aa | 3234 | force_grow: |
3c517a61 | 3235 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3236 | |
a737b3e2 | 3237 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3238 | ac = cpu_cache_get(cachep); |
072bb0aa MG |
3239 | |
3240 | /* no objects in sight? abort */ | |
3241 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
3242 | return NULL; |
3243 | ||
a737b3e2 | 3244 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3245 | goto retry; |
3246 | } | |
3247 | ac->touched = 1; | |
072bb0aa MG |
3248 | |
3249 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
3250 | } |
3251 | ||
a737b3e2 AM |
3252 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3253 | gfp_t flags) | |
1da177e4 LT |
3254 | { |
3255 | might_sleep_if(flags & __GFP_WAIT); | |
3256 | #if DEBUG | |
3257 | kmem_flagcheck(cachep, flags); | |
3258 | #endif | |
3259 | } | |
3260 | ||
3261 | #if DEBUG | |
a737b3e2 AM |
3262 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3263 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3264 | { |
b28a02de | 3265 | if (!objp) |
1da177e4 | 3266 | return objp; |
b28a02de | 3267 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3268 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3269 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3270 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3271 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3272 | else |
3273 | check_poison_obj(cachep, objp); | |
3274 | #else | |
3275 | check_poison_obj(cachep, objp); | |
3276 | #endif | |
3277 | poison_obj(cachep, objp, POISON_INUSE); | |
3278 | } | |
3279 | if (cachep->flags & SLAB_STORE_USER) | |
3280 | *dbg_userword(cachep, objp) = caller; | |
3281 | ||
3282 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3283 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3284 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3285 | slab_error(cachep, "double free, or memory outside" | |
3286 | " object was overwritten"); | |
b28a02de | 3287 | printk(KERN_ERR |
b46b8f19 | 3288 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3289 | objp, *dbg_redzone1(cachep, objp), |
3290 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3291 | } |
3292 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3293 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3294 | } | |
871751e2 AV |
3295 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3296 | { | |
3297 | struct slab *slabp; | |
3298 | unsigned objnr; | |
3299 | ||
35026088 | 3300 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3301 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3302 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3303 | } | |
3304 | #endif | |
3dafccf2 | 3305 | objp += obj_offset(cachep); |
4f104934 | 3306 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3307 | cachep->ctor(objp); |
7ea466f2 TH |
3308 | if (ARCH_SLAB_MINALIGN && |
3309 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3310 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3311 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3312 | } |
1da177e4 LT |
3313 | return objp; |
3314 | } | |
3315 | #else | |
3316 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3317 | #endif | |
3318 | ||
773ff60e | 3319 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 AM |
3320 | { |
3321 | if (cachep == &cache_cache) | |
773ff60e | 3322 | return false; |
8a8b6502 | 3323 | |
8c138bc0 | 3324 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3325 | } |
3326 | ||
343e0d7a | 3327 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3328 | { |
b28a02de | 3329 | void *objp; |
1da177e4 | 3330 | struct array_cache *ac; |
072bb0aa | 3331 | bool force_refill = false; |
1da177e4 | 3332 | |
5c382300 | 3333 | check_irq_off(); |
8a8b6502 | 3334 | |
9a2dba4b | 3335 | ac = cpu_cache_get(cachep); |
1da177e4 | 3336 | if (likely(ac->avail)) { |
1da177e4 | 3337 | ac->touched = 1; |
072bb0aa MG |
3338 | objp = ac_get_obj(cachep, ac, flags, false); |
3339 | ||
ddbf2e83 | 3340 | /* |
072bb0aa MG |
3341 | * Allow for the possibility all avail objects are not allowed |
3342 | * by the current flags | |
ddbf2e83 | 3343 | */ |
072bb0aa MG |
3344 | if (objp) { |
3345 | STATS_INC_ALLOCHIT(cachep); | |
3346 | goto out; | |
3347 | } | |
3348 | force_refill = true; | |
1da177e4 | 3349 | } |
072bb0aa MG |
3350 | |
3351 | STATS_INC_ALLOCMISS(cachep); | |
3352 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3353 | /* | |
3354 | * the 'ac' may be updated by cache_alloc_refill(), | |
3355 | * and kmemleak_erase() requires its correct value. | |
3356 | */ | |
3357 | ac = cpu_cache_get(cachep); | |
3358 | ||
3359 | out: | |
d5cff635 CM |
3360 | /* |
3361 | * To avoid a false negative, if an object that is in one of the | |
3362 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3363 | * treat the array pointers as a reference to the object. | |
3364 | */ | |
f3d8b53a O |
3365 | if (objp) |
3366 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3367 | return objp; |
3368 | } | |
3369 | ||
e498be7d | 3370 | #ifdef CONFIG_NUMA |
c61afb18 | 3371 | /* |
b2455396 | 3372 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3373 | * |
3374 | * If we are in_interrupt, then process context, including cpusets and | |
3375 | * mempolicy, may not apply and should not be used for allocation policy. | |
3376 | */ | |
3377 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3378 | { | |
3379 | int nid_alloc, nid_here; | |
3380 | ||
765c4507 | 3381 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3382 | return NULL; |
7d6e6d09 | 3383 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3384 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3385 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3386 | else if (current->mempolicy) |
e7b691b0 | 3387 | nid_alloc = slab_node(); |
c61afb18 | 3388 | if (nid_alloc != nid_here) |
8b98c169 | 3389 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3390 | return NULL; |
3391 | } | |
3392 | ||
765c4507 CL |
3393 | /* |
3394 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3395 | * certain node and fall back is permitted. First we scan all the |
3396 | * available nodelists for available objects. If that fails then we | |
3397 | * perform an allocation without specifying a node. This allows the page | |
3398 | * allocator to do its reclaim / fallback magic. We then insert the | |
3399 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3400 | */ |
8c8cc2c1 | 3401 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3402 | { |
8c8cc2c1 PE |
3403 | struct zonelist *zonelist; |
3404 | gfp_t local_flags; | |
dd1a239f | 3405 | struct zoneref *z; |
54a6eb5c MG |
3406 | struct zone *zone; |
3407 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3408 | void *obj = NULL; |
3c517a61 | 3409 | int nid; |
cc9a6c87 | 3410 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3411 | |
3412 | if (flags & __GFP_THISNODE) | |
3413 | return NULL; | |
3414 | ||
6cb06229 | 3415 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3416 | |
cc9a6c87 MG |
3417 | retry_cpuset: |
3418 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3419 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3420 | |
3c517a61 CL |
3421 | retry: |
3422 | /* | |
3423 | * Look through allowed nodes for objects available | |
3424 | * from existing per node queues. | |
3425 | */ | |
54a6eb5c MG |
3426 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3427 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3428 | |
54a6eb5c | 3429 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3430 | cache->nodelists[nid] && |
481c5346 | 3431 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3432 | obj = ____cache_alloc_node(cache, |
3433 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3434 | if (obj) |
3435 | break; | |
3436 | } | |
3c517a61 CL |
3437 | } |
3438 | ||
cfce6604 | 3439 | if (!obj) { |
3c517a61 CL |
3440 | /* |
3441 | * This allocation will be performed within the constraints | |
3442 | * of the current cpuset / memory policy requirements. | |
3443 | * We may trigger various forms of reclaim on the allowed | |
3444 | * set and go into memory reserves if necessary. | |
3445 | */ | |
dd47ea75 CL |
3446 | if (local_flags & __GFP_WAIT) |
3447 | local_irq_enable(); | |
3448 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3449 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3450 | if (local_flags & __GFP_WAIT) |
3451 | local_irq_disable(); | |
3c517a61 CL |
3452 | if (obj) { |
3453 | /* | |
3454 | * Insert into the appropriate per node queues | |
3455 | */ | |
3456 | nid = page_to_nid(virt_to_page(obj)); | |
3457 | if (cache_grow(cache, flags, nid, obj)) { | |
3458 | obj = ____cache_alloc_node(cache, | |
3459 | flags | GFP_THISNODE, nid); | |
3460 | if (!obj) | |
3461 | /* | |
3462 | * Another processor may allocate the | |
3463 | * objects in the slab since we are | |
3464 | * not holding any locks. | |
3465 | */ | |
3466 | goto retry; | |
3467 | } else { | |
b6a60451 | 3468 | /* cache_grow already freed obj */ |
3c517a61 CL |
3469 | obj = NULL; |
3470 | } | |
3471 | } | |
aedb0eb1 | 3472 | } |
cc9a6c87 MG |
3473 | |
3474 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3475 | goto retry_cpuset; | |
765c4507 CL |
3476 | return obj; |
3477 | } | |
3478 | ||
e498be7d CL |
3479 | /* |
3480 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3481 | */ |
8b98c169 | 3482 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3483 | int nodeid) |
e498be7d CL |
3484 | { |
3485 | struct list_head *entry; | |
b28a02de PE |
3486 | struct slab *slabp; |
3487 | struct kmem_list3 *l3; | |
3488 | void *obj; | |
b28a02de PE |
3489 | int x; |
3490 | ||
3491 | l3 = cachep->nodelists[nodeid]; | |
3492 | BUG_ON(!l3); | |
3493 | ||
a737b3e2 | 3494 | retry: |
ca3b9b91 | 3495 | check_irq_off(); |
b28a02de PE |
3496 | spin_lock(&l3->list_lock); |
3497 | entry = l3->slabs_partial.next; | |
3498 | if (entry == &l3->slabs_partial) { | |
3499 | l3->free_touched = 1; | |
3500 | entry = l3->slabs_free.next; | |
3501 | if (entry == &l3->slabs_free) | |
3502 | goto must_grow; | |
3503 | } | |
3504 | ||
3505 | slabp = list_entry(entry, struct slab, list); | |
3506 | check_spinlock_acquired_node(cachep, nodeid); | |
3507 | check_slabp(cachep, slabp); | |
3508 | ||
3509 | STATS_INC_NODEALLOCS(cachep); | |
3510 | STATS_INC_ACTIVE(cachep); | |
3511 | STATS_SET_HIGH(cachep); | |
3512 | ||
3513 | BUG_ON(slabp->inuse == cachep->num); | |
3514 | ||
78d382d7 | 3515 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3516 | check_slabp(cachep, slabp); |
3517 | l3->free_objects--; | |
3518 | /* move slabp to correct slabp list: */ | |
3519 | list_del(&slabp->list); | |
3520 | ||
a737b3e2 | 3521 | if (slabp->free == BUFCTL_END) |
b28a02de | 3522 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3523 | else |
b28a02de | 3524 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3525 | |
b28a02de PE |
3526 | spin_unlock(&l3->list_lock); |
3527 | goto done; | |
e498be7d | 3528 | |
a737b3e2 | 3529 | must_grow: |
b28a02de | 3530 | spin_unlock(&l3->list_lock); |
3c517a61 | 3531 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3532 | if (x) |
3533 | goto retry; | |
1da177e4 | 3534 | |
8c8cc2c1 | 3535 | return fallback_alloc(cachep, flags); |
e498be7d | 3536 | |
a737b3e2 | 3537 | done: |
b28a02de | 3538 | return obj; |
e498be7d | 3539 | } |
8c8cc2c1 PE |
3540 | |
3541 | /** | |
3542 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3543 | * @cachep: The cache to allocate from. | |
3544 | * @flags: See kmalloc(). | |
3545 | * @nodeid: node number of the target node. | |
3546 | * @caller: return address of caller, used for debug information | |
3547 | * | |
3548 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3549 | * node, which can improve the performance for cpu bound structures. | |
3550 | * | |
3551 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3552 | */ | |
3553 | static __always_inline void * | |
3554 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3555 | void *caller) | |
3556 | { | |
3557 | unsigned long save_flags; | |
3558 | void *ptr; | |
7d6e6d09 | 3559 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3560 | |
dcce284a | 3561 | flags &= gfp_allowed_mask; |
7e85ee0c | 3562 | |
cf40bd16 NP |
3563 | lockdep_trace_alloc(flags); |
3564 | ||
773ff60e | 3565 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3566 | return NULL; |
3567 | ||
8c8cc2c1 PE |
3568 | cache_alloc_debugcheck_before(cachep, flags); |
3569 | local_irq_save(save_flags); | |
3570 | ||
eacbbae3 | 3571 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3572 | nodeid = slab_node; |
8c8cc2c1 PE |
3573 | |
3574 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3575 | /* Node not bootstrapped yet */ | |
3576 | ptr = fallback_alloc(cachep, flags); | |
3577 | goto out; | |
3578 | } | |
3579 | ||
7d6e6d09 | 3580 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3581 | /* |
3582 | * Use the locally cached objects if possible. | |
3583 | * However ____cache_alloc does not allow fallback | |
3584 | * to other nodes. It may fail while we still have | |
3585 | * objects on other nodes available. | |
3586 | */ | |
3587 | ptr = ____cache_alloc(cachep, flags); | |
3588 | if (ptr) | |
3589 | goto out; | |
3590 | } | |
3591 | /* ___cache_alloc_node can fall back to other nodes */ | |
3592 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3593 | out: | |
3594 | local_irq_restore(save_flags); | |
3595 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3596 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3597 | flags); |
8c8cc2c1 | 3598 | |
c175eea4 | 3599 | if (likely(ptr)) |
8c138bc0 | 3600 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3601 | |
d07dbea4 | 3602 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3603 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3604 | |
8c8cc2c1 PE |
3605 | return ptr; |
3606 | } | |
3607 | ||
3608 | static __always_inline void * | |
3609 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3610 | { | |
3611 | void *objp; | |
3612 | ||
3613 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3614 | objp = alternate_node_alloc(cache, flags); | |
3615 | if (objp) | |
3616 | goto out; | |
3617 | } | |
3618 | objp = ____cache_alloc(cache, flags); | |
3619 | ||
3620 | /* | |
3621 | * We may just have run out of memory on the local node. | |
3622 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3623 | */ | |
7d6e6d09 LS |
3624 | if (!objp) |
3625 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3626 | |
3627 | out: | |
3628 | return objp; | |
3629 | } | |
3630 | #else | |
3631 | ||
3632 | static __always_inline void * | |
3633 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3634 | { | |
3635 | return ____cache_alloc(cachep, flags); | |
3636 | } | |
3637 | ||
3638 | #endif /* CONFIG_NUMA */ | |
3639 | ||
3640 | static __always_inline void * | |
3641 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3642 | { | |
3643 | unsigned long save_flags; | |
3644 | void *objp; | |
3645 | ||
dcce284a | 3646 | flags &= gfp_allowed_mask; |
7e85ee0c | 3647 | |
cf40bd16 NP |
3648 | lockdep_trace_alloc(flags); |
3649 | ||
773ff60e | 3650 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3651 | return NULL; |
3652 | ||
8c8cc2c1 PE |
3653 | cache_alloc_debugcheck_before(cachep, flags); |
3654 | local_irq_save(save_flags); | |
3655 | objp = __do_cache_alloc(cachep, flags); | |
3656 | local_irq_restore(save_flags); | |
3657 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3658 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3659 | flags); |
8c8cc2c1 PE |
3660 | prefetchw(objp); |
3661 | ||
c175eea4 | 3662 | if (likely(objp)) |
8c138bc0 | 3663 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3664 | |
d07dbea4 | 3665 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3666 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3667 | |
8c8cc2c1 PE |
3668 | return objp; |
3669 | } | |
e498be7d CL |
3670 | |
3671 | /* | |
3672 | * Caller needs to acquire correct kmem_list's list_lock | |
3673 | */ | |
343e0d7a | 3674 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3675 | int node) |
1da177e4 LT |
3676 | { |
3677 | int i; | |
e498be7d | 3678 | struct kmem_list3 *l3; |
1da177e4 LT |
3679 | |
3680 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3681 | void *objp; |
1da177e4 | 3682 | struct slab *slabp; |
1da177e4 | 3683 | |
072bb0aa MG |
3684 | clear_obj_pfmemalloc(&objpp[i]); |
3685 | objp = objpp[i]; | |
3686 | ||
6ed5eb22 | 3687 | slabp = virt_to_slab(objp); |
ff69416e | 3688 | l3 = cachep->nodelists[node]; |
1da177e4 | 3689 | list_del(&slabp->list); |
ff69416e | 3690 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3691 | check_slabp(cachep, slabp); |
78d382d7 | 3692 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3693 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3694 | l3->free_objects++; |
1da177e4 LT |
3695 | check_slabp(cachep, slabp); |
3696 | ||
3697 | /* fixup slab chains */ | |
3698 | if (slabp->inuse == 0) { | |
e498be7d CL |
3699 | if (l3->free_objects > l3->free_limit) { |
3700 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3701 | /* No need to drop any previously held |
3702 | * lock here, even if we have a off-slab slab | |
3703 | * descriptor it is guaranteed to come from | |
3704 | * a different cache, refer to comments before | |
3705 | * alloc_slabmgmt. | |
3706 | */ | |
1da177e4 LT |
3707 | slab_destroy(cachep, slabp); |
3708 | } else { | |
e498be7d | 3709 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3710 | } |
3711 | } else { | |
3712 | /* Unconditionally move a slab to the end of the | |
3713 | * partial list on free - maximum time for the | |
3714 | * other objects to be freed, too. | |
3715 | */ | |
e498be7d | 3716 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3717 | } |
3718 | } | |
3719 | } | |
3720 | ||
343e0d7a | 3721 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3722 | { |
3723 | int batchcount; | |
e498be7d | 3724 | struct kmem_list3 *l3; |
7d6e6d09 | 3725 | int node = numa_mem_id(); |
1da177e4 LT |
3726 | |
3727 | batchcount = ac->batchcount; | |
3728 | #if DEBUG | |
3729 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3730 | #endif | |
3731 | check_irq_off(); | |
ff69416e | 3732 | l3 = cachep->nodelists[node]; |
873623df | 3733 | spin_lock(&l3->list_lock); |
e498be7d CL |
3734 | if (l3->shared) { |
3735 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3736 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3737 | if (max) { |
3738 | if (batchcount > max) | |
3739 | batchcount = max; | |
e498be7d | 3740 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3741 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3742 | shared_array->avail += batchcount; |
3743 | goto free_done; | |
3744 | } | |
3745 | } | |
3746 | ||
ff69416e | 3747 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3748 | free_done: |
1da177e4 LT |
3749 | #if STATS |
3750 | { | |
3751 | int i = 0; | |
3752 | struct list_head *p; | |
3753 | ||
e498be7d CL |
3754 | p = l3->slabs_free.next; |
3755 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3756 | struct slab *slabp; |
3757 | ||
3758 | slabp = list_entry(p, struct slab, list); | |
3759 | BUG_ON(slabp->inuse); | |
3760 | ||
3761 | i++; | |
3762 | p = p->next; | |
3763 | } | |
3764 | STATS_SET_FREEABLE(cachep, i); | |
3765 | } | |
3766 | #endif | |
e498be7d | 3767 | spin_unlock(&l3->list_lock); |
1da177e4 | 3768 | ac->avail -= batchcount; |
a737b3e2 | 3769 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3770 | } |
3771 | ||
3772 | /* | |
a737b3e2 AM |
3773 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3774 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3775 | */ |
a947eb95 SS |
3776 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3777 | void *caller) | |
1da177e4 | 3778 | { |
9a2dba4b | 3779 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3780 | |
3781 | check_irq_off(); | |
d5cff635 | 3782 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3783 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3784 | |
8c138bc0 | 3785 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3786 | |
1807a1aa SS |
3787 | /* |
3788 | * Skip calling cache_free_alien() when the platform is not numa. | |
3789 | * This will avoid cache misses that happen while accessing slabp (which | |
3790 | * is per page memory reference) to get nodeid. Instead use a global | |
3791 | * variable to skip the call, which is mostly likely to be present in | |
3792 | * the cache. | |
3793 | */ | |
b6e68bc1 | 3794 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3795 | return; |
3796 | ||
1da177e4 LT |
3797 | if (likely(ac->avail < ac->limit)) { |
3798 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3799 | } else { |
3800 | STATS_INC_FREEMISS(cachep); | |
3801 | cache_flusharray(cachep, ac); | |
1da177e4 | 3802 | } |
42c8c99c | 3803 | |
072bb0aa | 3804 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3805 | } |
3806 | ||
3807 | /** | |
3808 | * kmem_cache_alloc - Allocate an object | |
3809 | * @cachep: The cache to allocate from. | |
3810 | * @flags: See kmalloc(). | |
3811 | * | |
3812 | * Allocate an object from this cache. The flags are only relevant | |
3813 | * if the cache has no available objects. | |
3814 | */ | |
343e0d7a | 3815 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3816 | { |
36555751 EGM |
3817 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3818 | ||
ca2b84cb | 3819 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3820 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3821 | |
3822 | return ret; | |
1da177e4 LT |
3823 | } |
3824 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3825 | ||
0f24f128 | 3826 | #ifdef CONFIG_TRACING |
85beb586 SR |
3827 | void * |
3828 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3829 | { |
85beb586 SR |
3830 | void *ret; |
3831 | ||
3832 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3833 | ||
3834 | trace_kmalloc(_RET_IP_, ret, | |
3835 | size, slab_buffer_size(cachep), flags); | |
3836 | return ret; | |
36555751 | 3837 | } |
85beb586 | 3838 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3839 | #endif |
3840 | ||
1da177e4 | 3841 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3842 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3843 | { | |
36555751 EGM |
3844 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3845 | __builtin_return_address(0)); | |
3846 | ||
ca2b84cb | 3847 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3848 | cachep->object_size, cachep->size, |
ca2b84cb | 3849 | flags, nodeid); |
36555751 EGM |
3850 | |
3851 | return ret; | |
8b98c169 | 3852 | } |
1da177e4 LT |
3853 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3854 | ||
0f24f128 | 3855 | #ifdef CONFIG_TRACING |
85beb586 SR |
3856 | void *kmem_cache_alloc_node_trace(size_t size, |
3857 | struct kmem_cache *cachep, | |
3858 | gfp_t flags, | |
3859 | int nodeid) | |
36555751 | 3860 | { |
85beb586 SR |
3861 | void *ret; |
3862 | ||
3863 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3864 | __builtin_return_address(0)); |
85beb586 SR |
3865 | trace_kmalloc_node(_RET_IP_, ret, |
3866 | size, slab_buffer_size(cachep), | |
3867 | flags, nodeid); | |
3868 | return ret; | |
36555751 | 3869 | } |
85beb586 | 3870 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3871 | #endif |
3872 | ||
8b98c169 CH |
3873 | static __always_inline void * |
3874 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3875 | { |
343e0d7a | 3876 | struct kmem_cache *cachep; |
97e2bde4 MS |
3877 | |
3878 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3879 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3880 | return cachep; | |
85beb586 | 3881 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3882 | } |
8b98c169 | 3883 | |
0bb38a5c | 3884 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3885 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3886 | { | |
3887 | return __do_kmalloc_node(size, flags, node, | |
3888 | __builtin_return_address(0)); | |
3889 | } | |
dbe5e69d | 3890 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3891 | |
3892 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3893 | int node, unsigned long caller) |
8b98c169 | 3894 | { |
ce71e27c | 3895 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3896 | } |
3897 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3898 | #else | |
3899 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3900 | { | |
3901 | return __do_kmalloc_node(size, flags, node, NULL); | |
3902 | } | |
3903 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3904 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3905 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3906 | |
3907 | /** | |
800590f5 | 3908 | * __do_kmalloc - allocate memory |
1da177e4 | 3909 | * @size: how many bytes of memory are required. |
800590f5 | 3910 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3911 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3912 | */ |
7fd6b141 PE |
3913 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3914 | void *caller) | |
1da177e4 | 3915 | { |
343e0d7a | 3916 | struct kmem_cache *cachep; |
36555751 | 3917 | void *ret; |
1da177e4 | 3918 | |
97e2bde4 MS |
3919 | /* If you want to save a few bytes .text space: replace |
3920 | * __ with kmem_. | |
3921 | * Then kmalloc uses the uninlined functions instead of the inline | |
3922 | * functions. | |
3923 | */ | |
3924 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3925 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3926 | return cachep; | |
36555751 EGM |
3927 | ret = __cache_alloc(cachep, flags, caller); |
3928 | ||
ca2b84cb | 3929 | trace_kmalloc((unsigned long) caller, ret, |
3b0efdfa | 3930 | size, cachep->size, flags); |
36555751 EGM |
3931 | |
3932 | return ret; | |
7fd6b141 PE |
3933 | } |
3934 | ||
7fd6b141 | 3935 | |
0bb38a5c | 3936 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3937 | void *__kmalloc(size_t size, gfp_t flags) |
3938 | { | |
871751e2 | 3939 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3940 | } |
3941 | EXPORT_SYMBOL(__kmalloc); | |
3942 | ||
ce71e27c | 3943 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3944 | { |
ce71e27c | 3945 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3946 | } |
3947 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3948 | |
3949 | #else | |
3950 | void *__kmalloc(size_t size, gfp_t flags) | |
3951 | { | |
3952 | return __do_kmalloc(size, flags, NULL); | |
3953 | } | |
3954 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3955 | #endif |
3956 | ||
1da177e4 LT |
3957 | /** |
3958 | * kmem_cache_free - Deallocate an object | |
3959 | * @cachep: The cache the allocation was from. | |
3960 | * @objp: The previously allocated object. | |
3961 | * | |
3962 | * Free an object which was previously allocated from this | |
3963 | * cache. | |
3964 | */ | |
343e0d7a | 3965 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3966 | { |
3967 | unsigned long flags; | |
3968 | ||
3969 | local_irq_save(flags); | |
d97d476b | 3970 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3971 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3972 | debug_check_no_obj_freed(objp, cachep->object_size); |
a947eb95 | 3973 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3974 | local_irq_restore(flags); |
36555751 | 3975 | |
ca2b84cb | 3976 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3977 | } |
3978 | EXPORT_SYMBOL(kmem_cache_free); | |
3979 | ||
1da177e4 LT |
3980 | /** |
3981 | * kfree - free previously allocated memory | |
3982 | * @objp: pointer returned by kmalloc. | |
3983 | * | |
80e93eff PE |
3984 | * If @objp is NULL, no operation is performed. |
3985 | * | |
1da177e4 LT |
3986 | * Don't free memory not originally allocated by kmalloc() |
3987 | * or you will run into trouble. | |
3988 | */ | |
3989 | void kfree(const void *objp) | |
3990 | { | |
343e0d7a | 3991 | struct kmem_cache *c; |
1da177e4 LT |
3992 | unsigned long flags; |
3993 | ||
2121db74 PE |
3994 | trace_kfree(_RET_IP_, objp); |
3995 | ||
6cb8f913 | 3996 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3997 | return; |
3998 | local_irq_save(flags); | |
3999 | kfree_debugcheck(objp); | |
6ed5eb22 | 4000 | c = virt_to_cache(objp); |
8c138bc0 CL |
4001 | debug_check_no_locks_freed(objp, c->object_size); |
4002 | ||
4003 | debug_check_no_obj_freed(objp, c->object_size); | |
a947eb95 | 4004 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
4005 | local_irq_restore(flags); |
4006 | } | |
4007 | EXPORT_SYMBOL(kfree); | |
4008 | ||
343e0d7a | 4009 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 4010 | { |
8c138bc0 | 4011 | return cachep->object_size; |
1da177e4 LT |
4012 | } |
4013 | EXPORT_SYMBOL(kmem_cache_size); | |
4014 | ||
e498be7d | 4015 | /* |
183ff22b | 4016 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 4017 | */ |
83b519e8 | 4018 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
4019 | { |
4020 | int node; | |
4021 | struct kmem_list3 *l3; | |
cafeb02e | 4022 | struct array_cache *new_shared; |
3395ee05 | 4023 | struct array_cache **new_alien = NULL; |
e498be7d | 4024 | |
9c09a95c | 4025 | for_each_online_node(node) { |
cafeb02e | 4026 | |
3395ee05 | 4027 | if (use_alien_caches) { |
83b519e8 | 4028 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
4029 | if (!new_alien) |
4030 | goto fail; | |
4031 | } | |
cafeb02e | 4032 | |
63109846 ED |
4033 | new_shared = NULL; |
4034 | if (cachep->shared) { | |
4035 | new_shared = alloc_arraycache(node, | |
0718dc2a | 4036 | cachep->shared*cachep->batchcount, |
83b519e8 | 4037 | 0xbaadf00d, gfp); |
63109846 ED |
4038 | if (!new_shared) { |
4039 | free_alien_cache(new_alien); | |
4040 | goto fail; | |
4041 | } | |
0718dc2a | 4042 | } |
cafeb02e | 4043 | |
a737b3e2 AM |
4044 | l3 = cachep->nodelists[node]; |
4045 | if (l3) { | |
cafeb02e CL |
4046 | struct array_cache *shared = l3->shared; |
4047 | ||
e498be7d CL |
4048 | spin_lock_irq(&l3->list_lock); |
4049 | ||
cafeb02e | 4050 | if (shared) |
0718dc2a CL |
4051 | free_block(cachep, shared->entry, |
4052 | shared->avail, node); | |
e498be7d | 4053 | |
cafeb02e CL |
4054 | l3->shared = new_shared; |
4055 | if (!l3->alien) { | |
e498be7d CL |
4056 | l3->alien = new_alien; |
4057 | new_alien = NULL; | |
4058 | } | |
b28a02de | 4059 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4060 | cachep->batchcount + cachep->num; |
e498be7d | 4061 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 4062 | kfree(shared); |
e498be7d CL |
4063 | free_alien_cache(new_alien); |
4064 | continue; | |
4065 | } | |
83b519e8 | 4066 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
4067 | if (!l3) { |
4068 | free_alien_cache(new_alien); | |
4069 | kfree(new_shared); | |
e498be7d | 4070 | goto fail; |
0718dc2a | 4071 | } |
e498be7d CL |
4072 | |
4073 | kmem_list3_init(l3); | |
4074 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 4075 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 4076 | l3->shared = new_shared; |
e498be7d | 4077 | l3->alien = new_alien; |
b28a02de | 4078 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4079 | cachep->batchcount + cachep->num; |
e498be7d CL |
4080 | cachep->nodelists[node] = l3; |
4081 | } | |
cafeb02e | 4082 | return 0; |
0718dc2a | 4083 | |
a737b3e2 | 4084 | fail: |
3b0efdfa | 4085 | if (!cachep->list.next) { |
0718dc2a CL |
4086 | /* Cache is not active yet. Roll back what we did */ |
4087 | node--; | |
4088 | while (node >= 0) { | |
4089 | if (cachep->nodelists[node]) { | |
4090 | l3 = cachep->nodelists[node]; | |
4091 | ||
4092 | kfree(l3->shared); | |
4093 | free_alien_cache(l3->alien); | |
4094 | kfree(l3); | |
4095 | cachep->nodelists[node] = NULL; | |
4096 | } | |
4097 | node--; | |
4098 | } | |
4099 | } | |
cafeb02e | 4100 | return -ENOMEM; |
e498be7d CL |
4101 | } |
4102 | ||
1da177e4 | 4103 | struct ccupdate_struct { |
343e0d7a | 4104 | struct kmem_cache *cachep; |
acfe7d74 | 4105 | struct array_cache *new[0]; |
1da177e4 LT |
4106 | }; |
4107 | ||
4108 | static void do_ccupdate_local(void *info) | |
4109 | { | |
a737b3e2 | 4110 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4111 | struct array_cache *old; |
4112 | ||
4113 | check_irq_off(); | |
9a2dba4b | 4114 | old = cpu_cache_get(new->cachep); |
e498be7d | 4115 | |
1da177e4 LT |
4116 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4117 | new->new[smp_processor_id()] = old; | |
4118 | } | |
4119 | ||
18004c5d | 4120 | /* Always called with the slab_mutex held */ |
a737b3e2 | 4121 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4122 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4123 | { |
d2e7b7d0 | 4124 | struct ccupdate_struct *new; |
2ed3a4ef | 4125 | int i; |
1da177e4 | 4126 | |
acfe7d74 ED |
4127 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4128 | gfp); | |
d2e7b7d0 SS |
4129 | if (!new) |
4130 | return -ENOMEM; | |
4131 | ||
e498be7d | 4132 | for_each_online_cpu(i) { |
7d6e6d09 | 4133 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4134 | batchcount, gfp); |
d2e7b7d0 | 4135 | if (!new->new[i]) { |
b28a02de | 4136 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4137 | kfree(new->new[i]); |
4138 | kfree(new); | |
e498be7d | 4139 | return -ENOMEM; |
1da177e4 LT |
4140 | } |
4141 | } | |
d2e7b7d0 | 4142 | new->cachep = cachep; |
1da177e4 | 4143 | |
15c8b6c1 | 4144 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4145 | |
1da177e4 | 4146 | check_irq_on(); |
1da177e4 LT |
4147 | cachep->batchcount = batchcount; |
4148 | cachep->limit = limit; | |
e498be7d | 4149 | cachep->shared = shared; |
1da177e4 | 4150 | |
e498be7d | 4151 | for_each_online_cpu(i) { |
d2e7b7d0 | 4152 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4153 | if (!ccold) |
4154 | continue; | |
7d6e6d09 LS |
4155 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4156 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4157 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4158 | kfree(ccold); |
4159 | } | |
d2e7b7d0 | 4160 | kfree(new); |
83b519e8 | 4161 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4162 | } |
4163 | ||
18004c5d | 4164 | /* Called with slab_mutex held always */ |
83b519e8 | 4165 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4166 | { |
4167 | int err; | |
4168 | int limit, shared; | |
4169 | ||
a737b3e2 AM |
4170 | /* |
4171 | * The head array serves three purposes: | |
1da177e4 LT |
4172 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4173 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4174 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4175 | * bufctl chains: array operations are cheaper. |
4176 | * The numbers are guessed, we should auto-tune as described by | |
4177 | * Bonwick. | |
4178 | */ | |
3b0efdfa | 4179 | if (cachep->size > 131072) |
1da177e4 | 4180 | limit = 1; |
3b0efdfa | 4181 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4182 | limit = 8; |
3b0efdfa | 4183 | else if (cachep->size > 1024) |
1da177e4 | 4184 | limit = 24; |
3b0efdfa | 4185 | else if (cachep->size > 256) |
1da177e4 LT |
4186 | limit = 54; |
4187 | else | |
4188 | limit = 120; | |
4189 | ||
a737b3e2 AM |
4190 | /* |
4191 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4192 | * allocation behaviour: Most allocs on one cpu, most free operations |
4193 | * on another cpu. For these cases, an efficient object passing between | |
4194 | * cpus is necessary. This is provided by a shared array. The array | |
4195 | * replaces Bonwick's magazine layer. | |
4196 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4197 | * to a larger limit. Thus disabled by default. | |
4198 | */ | |
4199 | shared = 0; | |
3b0efdfa | 4200 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4201 | shared = 8; |
1da177e4 LT |
4202 | |
4203 | #if DEBUG | |
a737b3e2 AM |
4204 | /* |
4205 | * With debugging enabled, large batchcount lead to excessively long | |
4206 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4207 | */ |
4208 | if (limit > 32) | |
4209 | limit = 32; | |
4210 | #endif | |
83b519e8 | 4211 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4212 | if (err) |
4213 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4214 | cachep->name, -err); |
2ed3a4ef | 4215 | return err; |
1da177e4 LT |
4216 | } |
4217 | ||
1b55253a CL |
4218 | /* |
4219 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4220 | * necessary. Note that the l3 listlock also protects the array_cache |
4221 | * if drain_array() is used on the shared array. | |
1b55253a | 4222 | */ |
68a1b195 | 4223 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4224 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4225 | { |
4226 | int tofree; | |
4227 | ||
1b55253a CL |
4228 | if (!ac || !ac->avail) |
4229 | return; | |
1da177e4 LT |
4230 | if (ac->touched && !force) { |
4231 | ac->touched = 0; | |
b18e7e65 | 4232 | } else { |
1b55253a | 4233 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4234 | if (ac->avail) { |
4235 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4236 | if (tofree > ac->avail) | |
4237 | tofree = (ac->avail + 1) / 2; | |
4238 | free_block(cachep, ac->entry, tofree, node); | |
4239 | ac->avail -= tofree; | |
4240 | memmove(ac->entry, &(ac->entry[tofree]), | |
4241 | sizeof(void *) * ac->avail); | |
4242 | } | |
1b55253a | 4243 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4244 | } |
4245 | } | |
4246 | ||
4247 | /** | |
4248 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4249 | * @w: work descriptor |
1da177e4 LT |
4250 | * |
4251 | * Called from workqueue/eventd every few seconds. | |
4252 | * Purpose: | |
4253 | * - clear the per-cpu caches for this CPU. | |
4254 | * - return freeable pages to the main free memory pool. | |
4255 | * | |
a737b3e2 AM |
4256 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4257 | * again on the next iteration. | |
1da177e4 | 4258 | */ |
7c5cae36 | 4259 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4260 | { |
7a7c381d | 4261 | struct kmem_cache *searchp; |
e498be7d | 4262 | struct kmem_list3 *l3; |
7d6e6d09 | 4263 | int node = numa_mem_id(); |
bf6aede7 | 4264 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4265 | |
18004c5d | 4266 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 4267 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4268 | goto out; |
1da177e4 | 4269 | |
18004c5d | 4270 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
4271 | check_irq_on(); |
4272 | ||
35386e3b CL |
4273 | /* |
4274 | * We only take the l3 lock if absolutely necessary and we | |
4275 | * have established with reasonable certainty that | |
4276 | * we can do some work if the lock was obtained. | |
4277 | */ | |
aab2207c | 4278 | l3 = searchp->nodelists[node]; |
35386e3b | 4279 | |
8fce4d8e | 4280 | reap_alien(searchp, l3); |
1da177e4 | 4281 | |
aab2207c | 4282 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4283 | |
35386e3b CL |
4284 | /* |
4285 | * These are racy checks but it does not matter | |
4286 | * if we skip one check or scan twice. | |
4287 | */ | |
e498be7d | 4288 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4289 | goto next; |
1da177e4 | 4290 | |
e498be7d | 4291 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4292 | |
aab2207c | 4293 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4294 | |
ed11d9eb | 4295 | if (l3->free_touched) |
e498be7d | 4296 | l3->free_touched = 0; |
ed11d9eb CL |
4297 | else { |
4298 | int freed; | |
1da177e4 | 4299 | |
ed11d9eb CL |
4300 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4301 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4302 | STATS_ADD_REAPED(searchp, freed); | |
4303 | } | |
35386e3b | 4304 | next: |
1da177e4 LT |
4305 | cond_resched(); |
4306 | } | |
4307 | check_irq_on(); | |
18004c5d | 4308 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4309 | next_reap_node(); |
7c5cae36 | 4310 | out: |
a737b3e2 | 4311 | /* Set up the next iteration */ |
7c5cae36 | 4312 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4313 | } |
4314 | ||
158a9624 | 4315 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4316 | |
85289f98 | 4317 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4318 | { |
85289f98 PE |
4319 | /* |
4320 | * Output format version, so at least we can change it | |
4321 | * without _too_ many complaints. | |
4322 | */ | |
1da177e4 | 4323 | #if STATS |
85289f98 | 4324 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4325 | #else |
85289f98 | 4326 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4327 | #endif |
85289f98 PE |
4328 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4329 | "<objperslab> <pagesperslab>"); | |
4330 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4331 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4332 | #if STATS |
85289f98 | 4333 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4334 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4335 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4336 | #endif |
85289f98 PE |
4337 | seq_putc(m, '\n'); |
4338 | } | |
4339 | ||
4340 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4341 | { | |
4342 | loff_t n = *pos; | |
85289f98 | 4343 | |
18004c5d | 4344 | mutex_lock(&slab_mutex); |
85289f98 PE |
4345 | if (!n) |
4346 | print_slabinfo_header(m); | |
b92151ba | 4347 | |
18004c5d | 4348 | return seq_list_start(&slab_caches, *pos); |
1da177e4 LT |
4349 | } |
4350 | ||
4351 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4352 | { | |
18004c5d | 4353 | return seq_list_next(p, &slab_caches, pos); |
1da177e4 LT |
4354 | } |
4355 | ||
4356 | static void s_stop(struct seq_file *m, void *p) | |
4357 | { | |
18004c5d | 4358 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4359 | } |
4360 | ||
4361 | static int s_show(struct seq_file *m, void *p) | |
4362 | { | |
3b0efdfa | 4363 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
b28a02de PE |
4364 | struct slab *slabp; |
4365 | unsigned long active_objs; | |
4366 | unsigned long num_objs; | |
4367 | unsigned long active_slabs = 0; | |
4368 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4369 | const char *name; |
1da177e4 | 4370 | char *error = NULL; |
e498be7d CL |
4371 | int node; |
4372 | struct kmem_list3 *l3; | |
1da177e4 | 4373 | |
1da177e4 LT |
4374 | active_objs = 0; |
4375 | num_slabs = 0; | |
e498be7d CL |
4376 | for_each_online_node(node) { |
4377 | l3 = cachep->nodelists[node]; | |
4378 | if (!l3) | |
4379 | continue; | |
4380 | ||
ca3b9b91 RT |
4381 | check_irq_on(); |
4382 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4383 | |
7a7c381d | 4384 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4385 | if (slabp->inuse != cachep->num && !error) |
4386 | error = "slabs_full accounting error"; | |
4387 | active_objs += cachep->num; | |
4388 | active_slabs++; | |
4389 | } | |
7a7c381d | 4390 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4391 | if (slabp->inuse == cachep->num && !error) |
4392 | error = "slabs_partial inuse accounting error"; | |
4393 | if (!slabp->inuse && !error) | |
4394 | error = "slabs_partial/inuse accounting error"; | |
4395 | active_objs += slabp->inuse; | |
4396 | active_slabs++; | |
4397 | } | |
7a7c381d | 4398 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4399 | if (slabp->inuse && !error) |
4400 | error = "slabs_free/inuse accounting error"; | |
4401 | num_slabs++; | |
4402 | } | |
4403 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4404 | if (l3->shared) |
4405 | shared_avail += l3->shared->avail; | |
e498be7d | 4406 | |
ca3b9b91 | 4407 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4408 | } |
b28a02de PE |
4409 | num_slabs += active_slabs; |
4410 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4411 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4412 | error = "free_objects accounting error"; |
4413 | ||
b28a02de | 4414 | name = cachep->name; |
1da177e4 LT |
4415 | if (error) |
4416 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4417 | ||
4418 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3b0efdfa | 4419 | name, active_objs, num_objs, cachep->size, |
b28a02de | 4420 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4421 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4422 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4423 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4424 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4425 | #if STATS |
b28a02de | 4426 | { /* list3 stats */ |
1da177e4 LT |
4427 | unsigned long high = cachep->high_mark; |
4428 | unsigned long allocs = cachep->num_allocations; | |
4429 | unsigned long grown = cachep->grown; | |
4430 | unsigned long reaped = cachep->reaped; | |
4431 | unsigned long errors = cachep->errors; | |
4432 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4433 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4434 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4435 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4436 | |
e92dd4fd JP |
4437 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4438 | "%4lu %4lu %4lu %4lu %4lu", | |
4439 | allocs, high, grown, | |
4440 | reaped, errors, max_freeable, node_allocs, | |
4441 | node_frees, overflows); | |
1da177e4 LT |
4442 | } |
4443 | /* cpu stats */ | |
4444 | { | |
4445 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4446 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4447 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4448 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4449 | ||
4450 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4451 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4452 | } |
4453 | #endif | |
4454 | seq_putc(m, '\n'); | |
1da177e4 LT |
4455 | return 0; |
4456 | } | |
4457 | ||
4458 | /* | |
4459 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4460 | * | |
4461 | * Output layout: | |
4462 | * cache-name | |
4463 | * num-active-objs | |
4464 | * total-objs | |
4465 | * object size | |
4466 | * num-active-slabs | |
4467 | * total-slabs | |
4468 | * num-pages-per-slab | |
4469 | * + further values on SMP and with statistics enabled | |
4470 | */ | |
4471 | ||
7b3c3a50 | 4472 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4473 | .start = s_start, |
4474 | .next = s_next, | |
4475 | .stop = s_stop, | |
4476 | .show = s_show, | |
1da177e4 LT |
4477 | }; |
4478 | ||
4479 | #define MAX_SLABINFO_WRITE 128 | |
4480 | /** | |
4481 | * slabinfo_write - Tuning for the slab allocator | |
4482 | * @file: unused | |
4483 | * @buffer: user buffer | |
4484 | * @count: data length | |
4485 | * @ppos: unused | |
4486 | */ | |
68a1b195 | 4487 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4488 | size_t count, loff_t *ppos) |
1da177e4 | 4489 | { |
b28a02de | 4490 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4491 | int limit, batchcount, shared, res; |
7a7c381d | 4492 | struct kmem_cache *cachep; |
b28a02de | 4493 | |
1da177e4 LT |
4494 | if (count > MAX_SLABINFO_WRITE) |
4495 | return -EINVAL; | |
4496 | if (copy_from_user(&kbuf, buffer, count)) | |
4497 | return -EFAULT; | |
b28a02de | 4498 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4499 | |
4500 | tmp = strchr(kbuf, ' '); | |
4501 | if (!tmp) | |
4502 | return -EINVAL; | |
4503 | *tmp = '\0'; | |
4504 | tmp++; | |
4505 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4506 | return -EINVAL; | |
4507 | ||
4508 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4509 | mutex_lock(&slab_mutex); |
1da177e4 | 4510 | res = -EINVAL; |
18004c5d | 4511 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4512 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4513 | if (limit < 1 || batchcount < 1 || |
4514 | batchcount > limit || shared < 0) { | |
e498be7d | 4515 | res = 0; |
1da177e4 | 4516 | } else { |
e498be7d | 4517 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4518 | batchcount, shared, |
4519 | GFP_KERNEL); | |
1da177e4 LT |
4520 | } |
4521 | break; | |
4522 | } | |
4523 | } | |
18004c5d | 4524 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4525 | if (res >= 0) |
4526 | res = count; | |
4527 | return res; | |
4528 | } | |
871751e2 | 4529 | |
7b3c3a50 AD |
4530 | static int slabinfo_open(struct inode *inode, struct file *file) |
4531 | { | |
4532 | return seq_open(file, &slabinfo_op); | |
4533 | } | |
4534 | ||
4535 | static const struct file_operations proc_slabinfo_operations = { | |
4536 | .open = slabinfo_open, | |
4537 | .read = seq_read, | |
4538 | .write = slabinfo_write, | |
4539 | .llseek = seq_lseek, | |
4540 | .release = seq_release, | |
4541 | }; | |
4542 | ||
871751e2 AV |
4543 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4544 | ||
4545 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4546 | { | |
18004c5d CL |
4547 | mutex_lock(&slab_mutex); |
4548 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4549 | } |
4550 | ||
4551 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4552 | { | |
4553 | unsigned long *p; | |
4554 | int l; | |
4555 | if (!v) | |
4556 | return 1; | |
4557 | l = n[1]; | |
4558 | p = n + 2; | |
4559 | while (l) { | |
4560 | int i = l/2; | |
4561 | unsigned long *q = p + 2 * i; | |
4562 | if (*q == v) { | |
4563 | q[1]++; | |
4564 | return 1; | |
4565 | } | |
4566 | if (*q > v) { | |
4567 | l = i; | |
4568 | } else { | |
4569 | p = q + 2; | |
4570 | l -= i + 1; | |
4571 | } | |
4572 | } | |
4573 | if (++n[1] == n[0]) | |
4574 | return 0; | |
4575 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4576 | p[0] = v; | |
4577 | p[1] = 1; | |
4578 | return 1; | |
4579 | } | |
4580 | ||
4581 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4582 | { | |
4583 | void *p; | |
4584 | int i; | |
4585 | if (n[0] == n[1]) | |
4586 | return; | |
3b0efdfa | 4587 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4588 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4589 | continue; | |
4590 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4591 | return; | |
4592 | } | |
4593 | } | |
4594 | ||
4595 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4596 | { | |
4597 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4598 | unsigned long offset, size; |
9281acea | 4599 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4600 | |
a5c43dae | 4601 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4602 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4603 | if (modname[0]) |
871751e2 AV |
4604 | seq_printf(m, " [%s]", modname); |
4605 | return; | |
4606 | } | |
4607 | #endif | |
4608 | seq_printf(m, "%p", (void *)address); | |
4609 | } | |
4610 | ||
4611 | static int leaks_show(struct seq_file *m, void *p) | |
4612 | { | |
0672aa7c | 4613 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 AV |
4614 | struct slab *slabp; |
4615 | struct kmem_list3 *l3; | |
4616 | const char *name; | |
4617 | unsigned long *n = m->private; | |
4618 | int node; | |
4619 | int i; | |
4620 | ||
4621 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4622 | return 0; | |
4623 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4624 | return 0; | |
4625 | ||
4626 | /* OK, we can do it */ | |
4627 | ||
4628 | n[1] = 0; | |
4629 | ||
4630 | for_each_online_node(node) { | |
4631 | l3 = cachep->nodelists[node]; | |
4632 | if (!l3) | |
4633 | continue; | |
4634 | ||
4635 | check_irq_on(); | |
4636 | spin_lock_irq(&l3->list_lock); | |
4637 | ||
7a7c381d | 4638 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4639 | handle_slab(n, cachep, slabp); |
7a7c381d | 4640 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4641 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4642 | spin_unlock_irq(&l3->list_lock); |
4643 | } | |
4644 | name = cachep->name; | |
4645 | if (n[0] == n[1]) { | |
4646 | /* Increase the buffer size */ | |
18004c5d | 4647 | mutex_unlock(&slab_mutex); |
871751e2 AV |
4648 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4649 | if (!m->private) { | |
4650 | /* Too bad, we are really out */ | |
4651 | m->private = n; | |
18004c5d | 4652 | mutex_lock(&slab_mutex); |
871751e2 AV |
4653 | return -ENOMEM; |
4654 | } | |
4655 | *(unsigned long *)m->private = n[0] * 2; | |
4656 | kfree(n); | |
18004c5d | 4657 | mutex_lock(&slab_mutex); |
871751e2 AV |
4658 | /* Now make sure this entry will be retried */ |
4659 | m->count = m->size; | |
4660 | return 0; | |
4661 | } | |
4662 | for (i = 0; i < n[1]; i++) { | |
4663 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4664 | show_symbol(m, n[2*i+2]); | |
4665 | seq_putc(m, '\n'); | |
4666 | } | |
d2e7b7d0 | 4667 | |
871751e2 AV |
4668 | return 0; |
4669 | } | |
4670 | ||
a0ec95a8 | 4671 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4672 | .start = leaks_start, |
4673 | .next = s_next, | |
4674 | .stop = s_stop, | |
4675 | .show = leaks_show, | |
4676 | }; | |
a0ec95a8 AD |
4677 | |
4678 | static int slabstats_open(struct inode *inode, struct file *file) | |
4679 | { | |
4680 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4681 | int ret = -ENOMEM; | |
4682 | if (n) { | |
4683 | ret = seq_open(file, &slabstats_op); | |
4684 | if (!ret) { | |
4685 | struct seq_file *m = file->private_data; | |
4686 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4687 | m->private = n; | |
4688 | n = NULL; | |
4689 | } | |
4690 | kfree(n); | |
4691 | } | |
4692 | return ret; | |
4693 | } | |
4694 | ||
4695 | static const struct file_operations proc_slabstats_operations = { | |
4696 | .open = slabstats_open, | |
4697 | .read = seq_read, | |
4698 | .llseek = seq_lseek, | |
4699 | .release = seq_release_private, | |
4700 | }; | |
4701 | #endif | |
4702 | ||
4703 | static int __init slab_proc_init(void) | |
4704 | { | |
ab067e99 | 4705 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4706 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4707 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4708 | #endif |
a0ec95a8 AD |
4709 | return 0; |
4710 | } | |
4711 | module_init(slab_proc_init); | |
1da177e4 LT |
4712 | #endif |
4713 | ||
00e145b6 MS |
4714 | /** |
4715 | * ksize - get the actual amount of memory allocated for a given object | |
4716 | * @objp: Pointer to the object | |
4717 | * | |
4718 | * kmalloc may internally round up allocations and return more memory | |
4719 | * than requested. ksize() can be used to determine the actual amount of | |
4720 | * memory allocated. The caller may use this additional memory, even though | |
4721 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4722 | * The caller must guarantee that objp points to a valid object previously | |
4723 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4724 | * must not be freed during the duration of the call. | |
4725 | */ | |
fd76bab2 | 4726 | size_t ksize(const void *objp) |
1da177e4 | 4727 | { |
ef8b4520 CL |
4728 | BUG_ON(!objp); |
4729 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4730 | return 0; |
1da177e4 | 4731 | |
8c138bc0 | 4732 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4733 | } |
b1aabecd | 4734 | EXPORT_SYMBOL(ksize); |