]>
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 LT |
89 | #include <linux/slab.h> |
90 | #include <linux/mm.h> | |
c9cf5528 | 91 | #include <linux/poison.h> |
1da177e4 LT |
92 | #include <linux/swap.h> |
93 | #include <linux/cache.h> | |
94 | #include <linux/interrupt.h> | |
95 | #include <linux/init.h> | |
96 | #include <linux/compiler.h> | |
101a5001 | 97 | #include <linux/cpuset.h> |
a0ec95a8 | 98 | #include <linux/proc_fs.h> |
1da177e4 LT |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> | |
101 | #include <linux/kallsyms.h> | |
102 | #include <linux/cpu.h> | |
103 | #include <linux/sysctl.h> | |
104 | #include <linux/module.h> | |
105 | #include <linux/rcupdate.h> | |
543537bd | 106 | #include <linux/string.h> |
138ae663 | 107 | #include <linux/uaccess.h> |
e498be7d | 108 | #include <linux/nodemask.h> |
d5cff635 | 109 | #include <linux/kmemleak.h> |
dc85da15 | 110 | #include <linux/mempolicy.h> |
fc0abb14 | 111 | #include <linux/mutex.h> |
8a8b6502 | 112 | #include <linux/fault-inject.h> |
e7eebaf6 | 113 | #include <linux/rtmutex.h> |
6a2d7a95 | 114 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 115 | #include <linux/debugobjects.h> |
c175eea4 | 116 | #include <linux/kmemcheck.h> |
8f9f8d9e | 117 | #include <linux/memory.h> |
268bb0ce | 118 | #include <linux/prefetch.h> |
1da177e4 | 119 | |
381760ea MG |
120 | #include <net/sock.h> |
121 | ||
1da177e4 LT |
122 | #include <asm/cacheflush.h> |
123 | #include <asm/tlbflush.h> | |
124 | #include <asm/page.h> | |
125 | ||
4dee6b64 SR |
126 | #include <trace/events/kmem.h> |
127 | ||
072bb0aa MG |
128 | #include "internal.h" |
129 | ||
b9ce5ef4 GC |
130 | #include "slab.h" |
131 | ||
1da177e4 | 132 | /* |
50953fe9 | 133 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
134 | * 0 for faster, smaller code (especially in the critical paths). |
135 | * | |
136 | * STATS - 1 to collect stats for /proc/slabinfo. | |
137 | * 0 for faster, smaller code (especially in the critical paths). | |
138 | * | |
139 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
140 | */ | |
141 | ||
142 | #ifdef CONFIG_DEBUG_SLAB | |
143 | #define DEBUG 1 | |
144 | #define STATS 1 | |
145 | #define FORCED_DEBUG 1 | |
146 | #else | |
147 | #define DEBUG 0 | |
148 | #define STATS 0 | |
149 | #define FORCED_DEBUG 0 | |
150 | #endif | |
151 | ||
1da177e4 LT |
152 | /* Shouldn't this be in a header file somewhere? */ |
153 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 154 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 155 | |
1da177e4 LT |
156 | #ifndef ARCH_KMALLOC_FLAGS |
157 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
158 | #endif | |
159 | ||
072bb0aa MG |
160 | /* |
161 | * true if a page was allocated from pfmemalloc reserves for network-based | |
162 | * swap | |
163 | */ | |
164 | static bool pfmemalloc_active __read_mostly; | |
165 | ||
1da177e4 LT |
166 | /* |
167 | * kmem_bufctl_t: | |
168 | * | |
169 | * Bufctl's are used for linking objs within a slab | |
170 | * linked offsets. | |
171 | * | |
172 | * This implementation relies on "struct page" for locating the cache & | |
173 | * slab an object belongs to. | |
174 | * This allows the bufctl structure to be small (one int), but limits | |
175 | * the number of objects a slab (not a cache) can contain when off-slab | |
176 | * bufctls are used. The limit is the size of the largest general cache | |
177 | * that does not use off-slab slabs. | |
178 | * For 32bit archs with 4 kB pages, is this 56. | |
179 | * This is not serious, as it is only for large objects, when it is unwise | |
180 | * to have too many per slab. | |
181 | * Note: This limit can be raised by introducing a general cache whose size | |
182 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
183 | */ | |
184 | ||
fa5b08d5 | 185 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
186 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
187 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
188 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
189 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 190 | |
1da177e4 LT |
191 | /* |
192 | * struct slab_rcu | |
193 | * | |
194 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
195 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
196 | * we need to approach a kernel structure obliquely, from its address | |
197 | * obtained without the usual locking. We can lock the structure to | |
198 | * stabilize it and check it's still at the given address, only if we | |
199 | * can be sure that the memory has not been meanwhile reused for some | |
200 | * other kind of object (which our subsystem's lock might corrupt). | |
201 | * | |
202 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
203 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
204 | */ |
205 | struct slab_rcu { | |
b28a02de | 206 | struct rcu_head head; |
343e0d7a | 207 | struct kmem_cache *cachep; |
b28a02de | 208 | void *addr; |
1da177e4 LT |
209 | }; |
210 | ||
5bfe53a7 LJ |
211 | /* |
212 | * struct slab | |
213 | * | |
214 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
215 | * for a slab, or allocated from an general cache. | |
216 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
217 | */ | |
218 | struct slab { | |
219 | union { | |
220 | struct { | |
221 | struct list_head list; | |
222 | unsigned long colouroff; | |
223 | void *s_mem; /* including colour offset */ | |
224 | unsigned int inuse; /* num of objs active in slab */ | |
225 | kmem_bufctl_t free; | |
226 | unsigned short nodeid; | |
227 | }; | |
228 | struct slab_rcu __slab_cover_slab_rcu; | |
229 | }; | |
230 | }; | |
231 | ||
1da177e4 LT |
232 | /* |
233 | * struct array_cache | |
234 | * | |
1da177e4 LT |
235 | * Purpose: |
236 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
237 | * - reduce the number of linked list operations | |
238 | * - reduce spinlock operations | |
239 | * | |
240 | * The limit is stored in the per-cpu structure to reduce the data cache | |
241 | * footprint. | |
242 | * | |
243 | */ | |
244 | struct array_cache { | |
245 | unsigned int avail; | |
246 | unsigned int limit; | |
247 | unsigned int batchcount; | |
248 | unsigned int touched; | |
e498be7d | 249 | spinlock_t lock; |
bda5b655 | 250 | void *entry[]; /* |
a737b3e2 AM |
251 | * Must have this definition in here for the proper |
252 | * alignment of array_cache. Also simplifies accessing | |
253 | * the entries. | |
072bb0aa MG |
254 | * |
255 | * Entries should not be directly dereferenced as | |
256 | * entries belonging to slabs marked pfmemalloc will | |
257 | * have the lower bits set SLAB_OBJ_PFMEMALLOC | |
a737b3e2 | 258 | */ |
1da177e4 LT |
259 | }; |
260 | ||
072bb0aa MG |
261 | #define SLAB_OBJ_PFMEMALLOC 1 |
262 | static inline bool is_obj_pfmemalloc(void *objp) | |
263 | { | |
264 | return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC; | |
265 | } | |
266 | ||
267 | static inline void set_obj_pfmemalloc(void **objp) | |
268 | { | |
269 | *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC); | |
270 | return; | |
271 | } | |
272 | ||
273 | static inline void clear_obj_pfmemalloc(void **objp) | |
274 | { | |
275 | *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC); | |
276 | } | |
277 | ||
a737b3e2 AM |
278 | /* |
279 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
280 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
281 | */ |
282 | #define BOOT_CPUCACHE_ENTRIES 1 | |
283 | struct arraycache_init { | |
284 | struct array_cache cache; | |
b28a02de | 285 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
286 | }; |
287 | ||
288 | /* | |
e498be7d | 289 | * The slab lists for all objects. |
1da177e4 | 290 | */ |
6744f087 | 291 | struct kmem_cache_node { |
b28a02de PE |
292 | struct list_head slabs_partial; /* partial list first, better asm code */ |
293 | struct list_head slabs_full; | |
294 | struct list_head slabs_free; | |
295 | unsigned long free_objects; | |
b28a02de | 296 | unsigned int free_limit; |
2e1217cf | 297 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
298 | spinlock_t list_lock; |
299 | struct array_cache *shared; /* shared per node */ | |
300 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
301 | unsigned long next_reap; /* updated without locking */ |
302 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
303 | }; |
304 | ||
e498be7d CL |
305 | /* |
306 | * Need this for bootstrapping a per node allocator. | |
307 | */ | |
556a169d | 308 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
6744f087 | 309 | static struct kmem_cache_node __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 310 | #define CACHE_CACHE 0 |
556a169d PE |
311 | #define SIZE_AC MAX_NUMNODES |
312 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 313 | |
ed11d9eb | 314 | static int drain_freelist(struct kmem_cache *cache, |
6744f087 | 315 | struct kmem_cache_node *l3, int tofree); |
ed11d9eb CL |
316 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, |
317 | int node); | |
83b519e8 | 318 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 319 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 320 | |
e3366016 CL |
321 | struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1]; |
322 | EXPORT_SYMBOL(kmalloc_caches); | |
323 | ||
324 | #ifdef CONFIG_ZONE_DMA | |
325 | struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1]; | |
326 | EXPORT_SYMBOL(kmalloc_dma_caches); | |
327 | #endif | |
e498be7d | 328 | |
e0a42726 IM |
329 | static int slab_early_init = 1; |
330 | ||
e3366016 | 331 | #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init)) |
6744f087 | 332 | #define INDEX_L3 kmalloc_index(sizeof(struct kmem_cache_node)) |
1da177e4 | 333 | |
6744f087 | 334 | static void kmem_list3_init(struct kmem_cache_node *parent) |
e498be7d CL |
335 | { |
336 | INIT_LIST_HEAD(&parent->slabs_full); | |
337 | INIT_LIST_HEAD(&parent->slabs_partial); | |
338 | INIT_LIST_HEAD(&parent->slabs_free); | |
339 | parent->shared = NULL; | |
340 | parent->alien = NULL; | |
2e1217cf | 341 | parent->colour_next = 0; |
e498be7d CL |
342 | spin_lock_init(&parent->list_lock); |
343 | parent->free_objects = 0; | |
344 | parent->free_touched = 0; | |
345 | } | |
346 | ||
a737b3e2 AM |
347 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
348 | do { \ | |
349 | INIT_LIST_HEAD(listp); \ | |
350 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
351 | } while (0) |
352 | ||
a737b3e2 AM |
353 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
354 | do { \ | |
e498be7d CL |
355 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
356 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
357 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
358 | } while (0) | |
1da177e4 | 359 | |
1da177e4 LT |
360 | #define CFLGS_OFF_SLAB (0x80000000UL) |
361 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
362 | ||
363 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
364 | /* |
365 | * Optimization question: fewer reaps means less probability for unnessary | |
366 | * cpucache drain/refill cycles. | |
1da177e4 | 367 | * |
dc6f3f27 | 368 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
369 | * which could lock up otherwise freeable slabs. |
370 | */ | |
371 | #define REAPTIMEOUT_CPUC (2*HZ) | |
372 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
373 | ||
374 | #if STATS | |
375 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
376 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
377 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
378 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 379 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
380 | #define STATS_SET_HIGH(x) \ |
381 | do { \ | |
382 | if ((x)->num_active > (x)->high_mark) \ | |
383 | (x)->high_mark = (x)->num_active; \ | |
384 | } while (0) | |
1da177e4 LT |
385 | #define STATS_INC_ERR(x) ((x)->errors++) |
386 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 387 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 388 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
389 | #define STATS_SET_FREEABLE(x, i) \ |
390 | do { \ | |
391 | if ((x)->max_freeable < i) \ | |
392 | (x)->max_freeable = i; \ | |
393 | } while (0) | |
1da177e4 LT |
394 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
395 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
396 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
397 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
398 | #else | |
399 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
400 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
401 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
402 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 403 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
404 | #define STATS_SET_HIGH(x) do { } while (0) |
405 | #define STATS_INC_ERR(x) do { } while (0) | |
406 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 407 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 408 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 409 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
410 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
411 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
412 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
413 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
414 | #endif | |
415 | ||
416 | #if DEBUG | |
1da177e4 | 417 | |
a737b3e2 AM |
418 | /* |
419 | * memory layout of objects: | |
1da177e4 | 420 | * 0 : objp |
3dafccf2 | 421 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
422 | * the end of an object is aligned with the end of the real |
423 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 424 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 425 | * redzone word. |
3dafccf2 | 426 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
427 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
428 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 429 | * [BYTES_PER_WORD long] |
1da177e4 | 430 | */ |
343e0d7a | 431 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 432 | { |
3dafccf2 | 433 | return cachep->obj_offset; |
1da177e4 LT |
434 | } |
435 | ||
b46b8f19 | 436 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
437 | { |
438 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
439 | return (unsigned long long*) (objp + obj_offset(cachep) - |
440 | sizeof(unsigned long long)); | |
1da177e4 LT |
441 | } |
442 | ||
b46b8f19 | 443 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
444 | { |
445 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
446 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 447 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 448 | sizeof(unsigned long long) - |
87a927c7 | 449 | REDZONE_ALIGN); |
3b0efdfa | 450 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 451 | sizeof(unsigned long long)); |
1da177e4 LT |
452 | } |
453 | ||
343e0d7a | 454 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
455 | { |
456 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 457 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
458 | } |
459 | ||
460 | #else | |
461 | ||
3dafccf2 | 462 | #define obj_offset(x) 0 |
b46b8f19 DW |
463 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
464 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
465 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
466 | ||
467 | #endif | |
468 | ||
1da177e4 | 469 | /* |
3df1cccd DR |
470 | * Do not go above this order unless 0 objects fit into the slab or |
471 | * overridden on the command line. | |
1da177e4 | 472 | */ |
543585cc DR |
473 | #define SLAB_MAX_ORDER_HI 1 |
474 | #define SLAB_MAX_ORDER_LO 0 | |
475 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 476 | static bool slab_max_order_set __initdata; |
1da177e4 | 477 | |
6ed5eb22 PE |
478 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
479 | { | |
b49af68f | 480 | struct page *page = virt_to_head_page(obj); |
35026088 | 481 | return page->slab_cache; |
6ed5eb22 PE |
482 | } |
483 | ||
484 | static inline struct slab *virt_to_slab(const void *obj) | |
485 | { | |
b49af68f | 486 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
487 | |
488 | VM_BUG_ON(!PageSlab(page)); | |
489 | return page->slab_page; | |
6ed5eb22 PE |
490 | } |
491 | ||
8fea4e96 PE |
492 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
493 | unsigned int idx) | |
494 | { | |
3b0efdfa | 495 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
496 | } |
497 | ||
6a2d7a95 | 498 | /* |
3b0efdfa CL |
499 | * We want to avoid an expensive divide : (offset / cache->size) |
500 | * Using the fact that size is a constant for a particular cache, | |
501 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
502 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
503 | */ | |
504 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
505 | const struct slab *slab, void *obj) | |
8fea4e96 | 506 | { |
6a2d7a95 ED |
507 | u32 offset = (obj - slab->s_mem); |
508 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
509 | } |
510 | ||
1da177e4 | 511 | static struct arraycache_init initarray_generic = |
b28a02de | 512 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
513 | |
514 | /* internal cache of cache description objs */ | |
9b030cb8 | 515 | static struct kmem_cache kmem_cache_boot = { |
b28a02de PE |
516 | .batchcount = 1, |
517 | .limit = BOOT_CPUCACHE_ENTRIES, | |
518 | .shared = 1, | |
3b0efdfa | 519 | .size = sizeof(struct kmem_cache), |
b28a02de | 520 | .name = "kmem_cache", |
1da177e4 LT |
521 | }; |
522 | ||
056c6241 RT |
523 | #define BAD_ALIEN_MAGIC 0x01020304ul |
524 | ||
f1aaee53 AV |
525 | #ifdef CONFIG_LOCKDEP |
526 | ||
527 | /* | |
528 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
529 | * for other slabs "off slab". | |
530 | * The locking for this is tricky in that it nests within the locks | |
531 | * of all other slabs in a few places; to deal with this special | |
532 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
533 | * |
534 | * We set lock class for alien array caches which are up during init. | |
535 | * The lock annotation will be lost if all cpus of a node goes down and | |
536 | * then comes back up during hotplug | |
f1aaee53 | 537 | */ |
056c6241 RT |
538 | static struct lock_class_key on_slab_l3_key; |
539 | static struct lock_class_key on_slab_alc_key; | |
540 | ||
83835b3d PZ |
541 | static struct lock_class_key debugobj_l3_key; |
542 | static struct lock_class_key debugobj_alc_key; | |
543 | ||
544 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
545 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
546 | int q) | |
547 | { | |
548 | struct array_cache **alc; | |
6744f087 | 549 | struct kmem_cache_node *l3; |
83835b3d PZ |
550 | int r; |
551 | ||
552 | l3 = cachep->nodelists[q]; | |
553 | if (!l3) | |
554 | return; | |
555 | ||
556 | lockdep_set_class(&l3->list_lock, l3_key); | |
557 | alc = l3->alien; | |
558 | /* | |
559 | * FIXME: This check for BAD_ALIEN_MAGIC | |
560 | * should go away when common slab code is taught to | |
561 | * work even without alien caches. | |
562 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
563 | * for alloc_alien_cache, | |
564 | */ | |
565 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
566 | return; | |
567 | for_each_node(r) { | |
568 | if (alc[r]) | |
569 | lockdep_set_class(&alc[r]->lock, alc_key); | |
570 | } | |
571 | } | |
572 | ||
573 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
574 | { | |
575 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
576 | } | |
577 | ||
578 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
579 | { | |
580 | int node; | |
581 | ||
582 | for_each_online_node(node) | |
583 | slab_set_debugobj_lock_classes_node(cachep, node); | |
584 | } | |
585 | ||
ce79ddc8 | 586 | static void init_node_lock_keys(int q) |
f1aaee53 | 587 | { |
e3366016 | 588 | int i; |
056c6241 | 589 | |
97d06609 | 590 | if (slab_state < UP) |
ce79ddc8 PE |
591 | return; |
592 | ||
e3366016 | 593 | for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) { |
6744f087 | 594 | struct kmem_cache_node *l3; |
e3366016 CL |
595 | struct kmem_cache *cache = kmalloc_caches[i]; |
596 | ||
597 | if (!cache) | |
598 | continue; | |
ce79ddc8 | 599 | |
e3366016 CL |
600 | l3 = cache->nodelists[q]; |
601 | if (!l3 || OFF_SLAB(cache)) | |
00afa758 | 602 | continue; |
83835b3d | 603 | |
e3366016 | 604 | slab_set_lock_classes(cache, &on_slab_l3_key, |
83835b3d | 605 | &on_slab_alc_key, q); |
f1aaee53 AV |
606 | } |
607 | } | |
ce79ddc8 | 608 | |
6ccfb5bc GC |
609 | static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q) |
610 | { | |
6744f087 CL |
611 | |
612 | if (!cachep->nodelists[q]) | |
6ccfb5bc GC |
613 | return; |
614 | ||
615 | slab_set_lock_classes(cachep, &on_slab_l3_key, | |
616 | &on_slab_alc_key, q); | |
617 | } | |
618 | ||
619 | static inline void on_slab_lock_classes(struct kmem_cache *cachep) | |
620 | { | |
621 | int node; | |
622 | ||
623 | VM_BUG_ON(OFF_SLAB(cachep)); | |
624 | for_each_node(node) | |
625 | on_slab_lock_classes_node(cachep, node); | |
626 | } | |
627 | ||
ce79ddc8 PE |
628 | static inline void init_lock_keys(void) |
629 | { | |
630 | int node; | |
631 | ||
632 | for_each_node(node) | |
633 | init_node_lock_keys(node); | |
634 | } | |
f1aaee53 | 635 | #else |
ce79ddc8 PE |
636 | static void init_node_lock_keys(int q) |
637 | { | |
638 | } | |
639 | ||
056c6241 | 640 | static inline void init_lock_keys(void) |
f1aaee53 AV |
641 | { |
642 | } | |
83835b3d | 643 | |
6ccfb5bc GC |
644 | static inline void on_slab_lock_classes(struct kmem_cache *cachep) |
645 | { | |
646 | } | |
647 | ||
648 | static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node) | |
649 | { | |
650 | } | |
651 | ||
83835b3d PZ |
652 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) |
653 | { | |
654 | } | |
655 | ||
656 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
657 | { | |
658 | } | |
f1aaee53 AV |
659 | #endif |
660 | ||
1871e52c | 661 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 662 | |
343e0d7a | 663 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
664 | { |
665 | return cachep->array[smp_processor_id()]; | |
666 | } | |
667 | ||
a737b3e2 AM |
668 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
669 | gfp_t gfpflags) | |
1da177e4 | 670 | { |
e3366016 | 671 | int i; |
1da177e4 LT |
672 | |
673 | #if DEBUG | |
674 | /* This happens if someone tries to call | |
b28a02de PE |
675 | * kmem_cache_create(), or __kmalloc(), before |
676 | * the generic caches are initialized. | |
677 | */ | |
e3366016 | 678 | BUG_ON(kmalloc_caches[INDEX_AC] == NULL); |
1da177e4 | 679 | #endif |
6cb8f913 CL |
680 | if (!size) |
681 | return ZERO_SIZE_PTR; | |
682 | ||
e3366016 | 683 | i = kmalloc_index(size); |
1da177e4 LT |
684 | |
685 | /* | |
0abf40c1 | 686 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
687 | * has cs_{dma,}cachep==NULL. Thus no special case |
688 | * for large kmalloc calls required. | |
689 | */ | |
4b51d669 | 690 | #ifdef CONFIG_ZONE_DMA |
1da177e4 | 691 | if (unlikely(gfpflags & GFP_DMA)) |
e3366016 | 692 | return kmalloc_dma_caches[i]; |
4b51d669 | 693 | #endif |
e3366016 | 694 | return kmalloc_caches[i]; |
1da177e4 LT |
695 | } |
696 | ||
b221385b | 697 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
698 | { |
699 | return __find_general_cachep(size, gfpflags); | |
700 | } | |
97e2bde4 | 701 | |
fbaccacf | 702 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 703 | { |
fbaccacf SR |
704 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
705 | } | |
1da177e4 | 706 | |
a737b3e2 AM |
707 | /* |
708 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
709 | */ | |
fbaccacf SR |
710 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
711 | size_t align, int flags, size_t *left_over, | |
712 | unsigned int *num) | |
713 | { | |
714 | int nr_objs; | |
715 | size_t mgmt_size; | |
716 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 717 | |
fbaccacf SR |
718 | /* |
719 | * The slab management structure can be either off the slab or | |
720 | * on it. For the latter case, the memory allocated for a | |
721 | * slab is used for: | |
722 | * | |
723 | * - The struct slab | |
724 | * - One kmem_bufctl_t for each object | |
725 | * - Padding to respect alignment of @align | |
726 | * - @buffer_size bytes for each object | |
727 | * | |
728 | * If the slab management structure is off the slab, then the | |
729 | * alignment will already be calculated into the size. Because | |
730 | * the slabs are all pages aligned, the objects will be at the | |
731 | * correct alignment when allocated. | |
732 | */ | |
733 | if (flags & CFLGS_OFF_SLAB) { | |
734 | mgmt_size = 0; | |
735 | nr_objs = slab_size / buffer_size; | |
736 | ||
737 | if (nr_objs > SLAB_LIMIT) | |
738 | nr_objs = SLAB_LIMIT; | |
739 | } else { | |
740 | /* | |
741 | * Ignore padding for the initial guess. The padding | |
742 | * is at most @align-1 bytes, and @buffer_size is at | |
743 | * least @align. In the worst case, this result will | |
744 | * be one greater than the number of objects that fit | |
745 | * into the memory allocation when taking the padding | |
746 | * into account. | |
747 | */ | |
748 | nr_objs = (slab_size - sizeof(struct slab)) / | |
749 | (buffer_size + sizeof(kmem_bufctl_t)); | |
750 | ||
751 | /* | |
752 | * This calculated number will be either the right | |
753 | * amount, or one greater than what we want. | |
754 | */ | |
755 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
756 | > slab_size) | |
757 | nr_objs--; | |
758 | ||
759 | if (nr_objs > SLAB_LIMIT) | |
760 | nr_objs = SLAB_LIMIT; | |
761 | ||
762 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
763 | } | |
764 | *num = nr_objs; | |
765 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
766 | } |
767 | ||
f28510d3 | 768 | #if DEBUG |
d40cee24 | 769 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 770 | |
a737b3e2 AM |
771 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
772 | char *msg) | |
1da177e4 LT |
773 | { |
774 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 775 | function, cachep->name, msg); |
1da177e4 | 776 | dump_stack(); |
645df230 | 777 | add_taint(TAINT_BAD_PAGE); |
1da177e4 | 778 | } |
f28510d3 | 779 | #endif |
1da177e4 | 780 | |
3395ee05 PM |
781 | /* |
782 | * By default on NUMA we use alien caches to stage the freeing of | |
783 | * objects allocated from other nodes. This causes massive memory | |
784 | * inefficiencies when using fake NUMA setup to split memory into a | |
785 | * large number of small nodes, so it can be disabled on the command | |
786 | * line | |
787 | */ | |
788 | ||
789 | static int use_alien_caches __read_mostly = 1; | |
790 | static int __init noaliencache_setup(char *s) | |
791 | { | |
792 | use_alien_caches = 0; | |
793 | return 1; | |
794 | } | |
795 | __setup("noaliencache", noaliencache_setup); | |
796 | ||
3df1cccd DR |
797 | static int __init slab_max_order_setup(char *str) |
798 | { | |
799 | get_option(&str, &slab_max_order); | |
800 | slab_max_order = slab_max_order < 0 ? 0 : | |
801 | min(slab_max_order, MAX_ORDER - 1); | |
802 | slab_max_order_set = true; | |
803 | ||
804 | return 1; | |
805 | } | |
806 | __setup("slab_max_order=", slab_max_order_setup); | |
807 | ||
8fce4d8e CL |
808 | #ifdef CONFIG_NUMA |
809 | /* | |
810 | * Special reaping functions for NUMA systems called from cache_reap(). | |
811 | * These take care of doing round robin flushing of alien caches (containing | |
812 | * objects freed on different nodes from which they were allocated) and the | |
813 | * flushing of remote pcps by calling drain_node_pages. | |
814 | */ | |
1871e52c | 815 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
816 | |
817 | static void init_reap_node(int cpu) | |
818 | { | |
819 | int node; | |
820 | ||
7d6e6d09 | 821 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 822 | if (node == MAX_NUMNODES) |
442295c9 | 823 | node = first_node(node_online_map); |
8fce4d8e | 824 | |
1871e52c | 825 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
826 | } |
827 | ||
828 | static void next_reap_node(void) | |
829 | { | |
909ea964 | 830 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 831 | |
8fce4d8e CL |
832 | node = next_node(node, node_online_map); |
833 | if (unlikely(node >= MAX_NUMNODES)) | |
834 | node = first_node(node_online_map); | |
909ea964 | 835 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
836 | } |
837 | ||
838 | #else | |
839 | #define init_reap_node(cpu) do { } while (0) | |
840 | #define next_reap_node(void) do { } while (0) | |
841 | #endif | |
842 | ||
1da177e4 LT |
843 | /* |
844 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
845 | * via the workqueue/eventd. | |
846 | * Add the CPU number into the expiration time to minimize the possibility of | |
847 | * the CPUs getting into lockstep and contending for the global cache chain | |
848 | * lock. | |
849 | */ | |
897e679b | 850 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 851 | { |
1871e52c | 852 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
853 | |
854 | /* | |
855 | * When this gets called from do_initcalls via cpucache_init(), | |
856 | * init_workqueues() has already run, so keventd will be setup | |
857 | * at that time. | |
858 | */ | |
52bad64d | 859 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 860 | init_reap_node(cpu); |
203b42f7 | 861 | INIT_DEFERRABLE_WORK(reap_work, cache_reap); |
2b284214 AV |
862 | schedule_delayed_work_on(cpu, reap_work, |
863 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
864 | } |
865 | } | |
866 | ||
e498be7d | 867 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 868 | int batchcount, gfp_t gfp) |
1da177e4 | 869 | { |
b28a02de | 870 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
871 | struct array_cache *nc = NULL; |
872 | ||
83b519e8 | 873 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
874 | /* |
875 | * The array_cache structures contain pointers to free object. | |
25985edc | 876 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
877 | * cache the pointers are not cleared and they could be counted as |
878 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
879 | * not scan such objects. | |
880 | */ | |
881 | kmemleak_no_scan(nc); | |
1da177e4 LT |
882 | if (nc) { |
883 | nc->avail = 0; | |
884 | nc->limit = entries; | |
885 | nc->batchcount = batchcount; | |
886 | nc->touched = 0; | |
e498be7d | 887 | spin_lock_init(&nc->lock); |
1da177e4 LT |
888 | } |
889 | return nc; | |
890 | } | |
891 | ||
072bb0aa MG |
892 | static inline bool is_slab_pfmemalloc(struct slab *slabp) |
893 | { | |
894 | struct page *page = virt_to_page(slabp->s_mem); | |
895 | ||
896 | return PageSlabPfmemalloc(page); | |
897 | } | |
898 | ||
899 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
900 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
901 | struct array_cache *ac) | |
902 | { | |
6744f087 | 903 | struct kmem_cache_node *l3 = cachep->nodelists[numa_mem_id()]; |
072bb0aa MG |
904 | struct slab *slabp; |
905 | unsigned long flags; | |
906 | ||
907 | if (!pfmemalloc_active) | |
908 | return; | |
909 | ||
910 | spin_lock_irqsave(&l3->list_lock, flags); | |
911 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
912 | if (is_slab_pfmemalloc(slabp)) | |
913 | goto out; | |
914 | ||
915 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
916 | if (is_slab_pfmemalloc(slabp)) | |
917 | goto out; | |
918 | ||
919 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
920 | if (is_slab_pfmemalloc(slabp)) | |
921 | goto out; | |
922 | ||
923 | pfmemalloc_active = false; | |
924 | out: | |
925 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
926 | } | |
927 | ||
381760ea | 928 | static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, |
072bb0aa MG |
929 | gfp_t flags, bool force_refill) |
930 | { | |
931 | int i; | |
932 | void *objp = ac->entry[--ac->avail]; | |
933 | ||
934 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
935 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
6744f087 | 936 | struct kmem_cache_node *l3; |
072bb0aa MG |
937 | |
938 | if (gfp_pfmemalloc_allowed(flags)) { | |
939 | clear_obj_pfmemalloc(&objp); | |
940 | return objp; | |
941 | } | |
942 | ||
943 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
d014dc2e | 944 | for (i = 0; i < ac->avail; i++) { |
072bb0aa MG |
945 | /* If a !PFMEMALLOC object is found, swap them */ |
946 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
947 | objp = ac->entry[i]; | |
948 | ac->entry[i] = ac->entry[ac->avail]; | |
949 | ac->entry[ac->avail] = objp; | |
950 | return objp; | |
951 | } | |
952 | } | |
953 | ||
954 | /* | |
955 | * If there are empty slabs on the slabs_free list and we are | |
956 | * being forced to refill the cache, mark this one !pfmemalloc. | |
957 | */ | |
958 | l3 = cachep->nodelists[numa_mem_id()]; | |
959 | if (!list_empty(&l3->slabs_free) && force_refill) { | |
960 | struct slab *slabp = virt_to_slab(objp); | |
30c29bea | 961 | ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem)); |
072bb0aa MG |
962 | clear_obj_pfmemalloc(&objp); |
963 | recheck_pfmemalloc_active(cachep, ac); | |
964 | return objp; | |
965 | } | |
966 | ||
967 | /* No !PFMEMALLOC objects available */ | |
968 | ac->avail++; | |
969 | objp = NULL; | |
970 | } | |
971 | ||
972 | return objp; | |
973 | } | |
974 | ||
381760ea MG |
975 | static inline void *ac_get_obj(struct kmem_cache *cachep, |
976 | struct array_cache *ac, gfp_t flags, bool force_refill) | |
977 | { | |
978 | void *objp; | |
979 | ||
980 | if (unlikely(sk_memalloc_socks())) | |
981 | objp = __ac_get_obj(cachep, ac, flags, force_refill); | |
982 | else | |
983 | objp = ac->entry[--ac->avail]; | |
984 | ||
985 | return objp; | |
986 | } | |
987 | ||
988 | static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
072bb0aa MG |
989 | void *objp) |
990 | { | |
991 | if (unlikely(pfmemalloc_active)) { | |
992 | /* Some pfmemalloc slabs exist, check if this is one */ | |
30c29bea | 993 | struct page *page = virt_to_head_page(objp); |
072bb0aa MG |
994 | if (PageSlabPfmemalloc(page)) |
995 | set_obj_pfmemalloc(&objp); | |
996 | } | |
997 | ||
381760ea MG |
998 | return objp; |
999 | } | |
1000 | ||
1001 | static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
1002 | void *objp) | |
1003 | { | |
1004 | if (unlikely(sk_memalloc_socks())) | |
1005 | objp = __ac_put_obj(cachep, ac, objp); | |
1006 | ||
072bb0aa MG |
1007 | ac->entry[ac->avail++] = objp; |
1008 | } | |
1009 | ||
3ded175a CL |
1010 | /* |
1011 | * Transfer objects in one arraycache to another. | |
1012 | * Locking must be handled by the caller. | |
1013 | * | |
1014 | * Return the number of entries transferred. | |
1015 | */ | |
1016 | static int transfer_objects(struct array_cache *to, | |
1017 | struct array_cache *from, unsigned int max) | |
1018 | { | |
1019 | /* Figure out how many entries to transfer */ | |
732eacc0 | 1020 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
1021 | |
1022 | if (!nr) | |
1023 | return 0; | |
1024 | ||
1025 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1026 | sizeof(void *) *nr); | |
1027 | ||
1028 | from->avail -= nr; | |
1029 | to->avail += nr; | |
3ded175a CL |
1030 | return nr; |
1031 | } | |
1032 | ||
765c4507 CL |
1033 | #ifndef CONFIG_NUMA |
1034 | ||
1035 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1036 | #define reap_alien(cachep, l3) do { } while (0) | |
1037 | ||
83b519e8 | 1038 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1039 | { |
1040 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1041 | } | |
1042 | ||
1043 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1044 | { | |
1045 | } | |
1046 | ||
1047 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1048 | { | |
1049 | return 0; | |
1050 | } | |
1051 | ||
1052 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1053 | gfp_t flags) | |
1054 | { | |
1055 | return NULL; | |
1056 | } | |
1057 | ||
8b98c169 | 1058 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1059 | gfp_t flags, int nodeid) |
1060 | { | |
1061 | return NULL; | |
1062 | } | |
1063 | ||
1064 | #else /* CONFIG_NUMA */ | |
1065 | ||
8b98c169 | 1066 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1067 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1068 | |
83b519e8 | 1069 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1070 | { |
1071 | struct array_cache **ac_ptr; | |
8ef82866 | 1072 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1073 | int i; |
1074 | ||
1075 | if (limit > 1) | |
1076 | limit = 12; | |
f3186a9c | 1077 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
1078 | if (ac_ptr) { |
1079 | for_each_node(i) { | |
f3186a9c | 1080 | if (i == node || !node_online(i)) |
e498be7d | 1081 | continue; |
83b519e8 | 1082 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1083 | if (!ac_ptr[i]) { |
cc550def | 1084 | for (i--; i >= 0; i--) |
e498be7d CL |
1085 | kfree(ac_ptr[i]); |
1086 | kfree(ac_ptr); | |
1087 | return NULL; | |
1088 | } | |
1089 | } | |
1090 | } | |
1091 | return ac_ptr; | |
1092 | } | |
1093 | ||
5295a74c | 1094 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1095 | { |
1096 | int i; | |
1097 | ||
1098 | if (!ac_ptr) | |
1099 | return; | |
e498be7d | 1100 | for_each_node(i) |
b28a02de | 1101 | kfree(ac_ptr[i]); |
e498be7d CL |
1102 | kfree(ac_ptr); |
1103 | } | |
1104 | ||
343e0d7a | 1105 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1106 | struct array_cache *ac, int node) |
e498be7d | 1107 | { |
6744f087 | 1108 | struct kmem_cache_node *rl3 = cachep->nodelists[node]; |
e498be7d CL |
1109 | |
1110 | if (ac->avail) { | |
1111 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1112 | /* |
1113 | * Stuff objects into the remote nodes shared array first. | |
1114 | * That way we could avoid the overhead of putting the objects | |
1115 | * into the free lists and getting them back later. | |
1116 | */ | |
693f7d36 JS |
1117 | if (rl3->shared) |
1118 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1119 | |
ff69416e | 1120 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1121 | ac->avail = 0; |
1122 | spin_unlock(&rl3->list_lock); | |
1123 | } | |
1124 | } | |
1125 | ||
8fce4d8e CL |
1126 | /* |
1127 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1128 | */ | |
6744f087 | 1129 | static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *l3) |
8fce4d8e | 1130 | { |
909ea964 | 1131 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1132 | |
1133 | if (l3->alien) { | |
1134 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1135 | |
1136 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1137 | __drain_alien_cache(cachep, ac, node); |
1138 | spin_unlock_irq(&ac->lock); | |
1139 | } | |
1140 | } | |
1141 | } | |
1142 | ||
a737b3e2 AM |
1143 | static void drain_alien_cache(struct kmem_cache *cachep, |
1144 | struct array_cache **alien) | |
e498be7d | 1145 | { |
b28a02de | 1146 | int i = 0; |
e498be7d CL |
1147 | struct array_cache *ac; |
1148 | unsigned long flags; | |
1149 | ||
1150 | for_each_online_node(i) { | |
4484ebf1 | 1151 | ac = alien[i]; |
e498be7d CL |
1152 | if (ac) { |
1153 | spin_lock_irqsave(&ac->lock, flags); | |
1154 | __drain_alien_cache(cachep, ac, i); | |
1155 | spin_unlock_irqrestore(&ac->lock, flags); | |
1156 | } | |
1157 | } | |
1158 | } | |
729bd0b7 | 1159 | |
873623df | 1160 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1161 | { |
1162 | struct slab *slabp = virt_to_slab(objp); | |
1163 | int nodeid = slabp->nodeid; | |
6744f087 | 1164 | struct kmem_cache_node *l3; |
729bd0b7 | 1165 | struct array_cache *alien = NULL; |
1ca4cb24 PE |
1166 | int node; |
1167 | ||
7d6e6d09 | 1168 | node = numa_mem_id(); |
729bd0b7 PE |
1169 | |
1170 | /* | |
1171 | * Make sure we are not freeing a object from another node to the array | |
1172 | * cache on this cpu. | |
1173 | */ | |
62918a03 | 1174 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1175 | return 0; |
1176 | ||
1ca4cb24 | 1177 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1178 | STATS_INC_NODEFREES(cachep); |
1179 | if (l3->alien && l3->alien[nodeid]) { | |
1180 | alien = l3->alien[nodeid]; | |
873623df | 1181 | spin_lock(&alien->lock); |
729bd0b7 PE |
1182 | if (unlikely(alien->avail == alien->limit)) { |
1183 | STATS_INC_ACOVERFLOW(cachep); | |
1184 | __drain_alien_cache(cachep, alien, nodeid); | |
1185 | } | |
072bb0aa | 1186 | ac_put_obj(cachep, alien, objp); |
729bd0b7 PE |
1187 | spin_unlock(&alien->lock); |
1188 | } else { | |
1189 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1190 | free_block(cachep, &objp, 1, nodeid); | |
1191 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1192 | } | |
1193 | return 1; | |
1194 | } | |
e498be7d CL |
1195 | #endif |
1196 | ||
8f9f8d9e DR |
1197 | /* |
1198 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1199 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1200 | * will be allocated off-node since memory is not yet online for the new node. | |
1201 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1202 | * already in use. | |
1203 | * | |
18004c5d | 1204 | * Must hold slab_mutex. |
8f9f8d9e DR |
1205 | */ |
1206 | static int init_cache_nodelists_node(int node) | |
1207 | { | |
1208 | struct kmem_cache *cachep; | |
6744f087 CL |
1209 | struct kmem_cache_node *l3; |
1210 | const int memsize = sizeof(struct kmem_cache_node); | |
8f9f8d9e | 1211 | |
18004c5d | 1212 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1213 | /* |
1214 | * Set up the size64 kmemlist for cpu before we can | |
1215 | * begin anything. Make sure some other cpu on this | |
1216 | * node has not already allocated this | |
1217 | */ | |
1218 | if (!cachep->nodelists[node]) { | |
1219 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1220 | if (!l3) | |
1221 | return -ENOMEM; | |
1222 | kmem_list3_init(l3); | |
1223 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1224 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1225 | ||
1226 | /* | |
1227 | * The l3s don't come and go as CPUs come and | |
18004c5d | 1228 | * go. slab_mutex is sufficient |
8f9f8d9e DR |
1229 | * protection here. |
1230 | */ | |
1231 | cachep->nodelists[node] = l3; | |
1232 | } | |
1233 | ||
1234 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1235 | cachep->nodelists[node]->free_limit = | |
1236 | (1 + nr_cpus_node(node)) * | |
1237 | cachep->batchcount + cachep->num; | |
1238 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1239 | } | |
1240 | return 0; | |
1241 | } | |
1242 | ||
fbf1e473 AM |
1243 | static void __cpuinit cpuup_canceled(long cpu) |
1244 | { | |
1245 | struct kmem_cache *cachep; | |
6744f087 | 1246 | struct kmem_cache_node *l3 = NULL; |
7d6e6d09 | 1247 | int node = cpu_to_mem(cpu); |
a70f7302 | 1248 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1249 | |
18004c5d | 1250 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1251 | struct array_cache *nc; |
1252 | struct array_cache *shared; | |
1253 | struct array_cache **alien; | |
fbf1e473 | 1254 | |
fbf1e473 AM |
1255 | /* cpu is dead; no one can alloc from it. */ |
1256 | nc = cachep->array[cpu]; | |
1257 | cachep->array[cpu] = NULL; | |
1258 | l3 = cachep->nodelists[node]; | |
1259 | ||
1260 | if (!l3) | |
1261 | goto free_array_cache; | |
1262 | ||
1263 | spin_lock_irq(&l3->list_lock); | |
1264 | ||
1265 | /* Free limit for this kmem_list3 */ | |
1266 | l3->free_limit -= cachep->batchcount; | |
1267 | if (nc) | |
1268 | free_block(cachep, nc->entry, nc->avail, node); | |
1269 | ||
58463c1f | 1270 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1271 | spin_unlock_irq(&l3->list_lock); |
1272 | goto free_array_cache; | |
1273 | } | |
1274 | ||
1275 | shared = l3->shared; | |
1276 | if (shared) { | |
1277 | free_block(cachep, shared->entry, | |
1278 | shared->avail, node); | |
1279 | l3->shared = NULL; | |
1280 | } | |
1281 | ||
1282 | alien = l3->alien; | |
1283 | l3->alien = NULL; | |
1284 | ||
1285 | spin_unlock_irq(&l3->list_lock); | |
1286 | ||
1287 | kfree(shared); | |
1288 | if (alien) { | |
1289 | drain_alien_cache(cachep, alien); | |
1290 | free_alien_cache(alien); | |
1291 | } | |
1292 | free_array_cache: | |
1293 | kfree(nc); | |
1294 | } | |
1295 | /* | |
1296 | * In the previous loop, all the objects were freed to | |
1297 | * the respective cache's slabs, now we can go ahead and | |
1298 | * shrink each nodelist to its limit. | |
1299 | */ | |
18004c5d | 1300 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1301 | l3 = cachep->nodelists[node]; |
1302 | if (!l3) | |
1303 | continue; | |
1304 | drain_freelist(cachep, l3, l3->free_objects); | |
1305 | } | |
1306 | } | |
1307 | ||
1308 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1309 | { |
343e0d7a | 1310 | struct kmem_cache *cachep; |
6744f087 | 1311 | struct kmem_cache_node *l3 = NULL; |
7d6e6d09 | 1312 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1313 | int err; |
1da177e4 | 1314 | |
fbf1e473 AM |
1315 | /* |
1316 | * We need to do this right in the beginning since | |
1317 | * alloc_arraycache's are going to use this list. | |
1318 | * kmalloc_node allows us to add the slab to the right | |
1319 | * kmem_list3 and not this cpu's kmem_list3 | |
1320 | */ | |
8f9f8d9e DR |
1321 | err = init_cache_nodelists_node(node); |
1322 | if (err < 0) | |
1323 | goto bad; | |
fbf1e473 AM |
1324 | |
1325 | /* | |
1326 | * Now we can go ahead with allocating the shared arrays and | |
1327 | * array caches | |
1328 | */ | |
18004c5d | 1329 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1330 | struct array_cache *nc; |
1331 | struct array_cache *shared = NULL; | |
1332 | struct array_cache **alien = NULL; | |
1333 | ||
1334 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1335 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1336 | if (!nc) |
1337 | goto bad; | |
1338 | if (cachep->shared) { | |
1339 | shared = alloc_arraycache(node, | |
1340 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1341 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1342 | if (!shared) { |
1343 | kfree(nc); | |
1da177e4 | 1344 | goto bad; |
12d00f6a | 1345 | } |
fbf1e473 AM |
1346 | } |
1347 | if (use_alien_caches) { | |
83b519e8 | 1348 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1349 | if (!alien) { |
1350 | kfree(shared); | |
1351 | kfree(nc); | |
fbf1e473 | 1352 | goto bad; |
12d00f6a | 1353 | } |
fbf1e473 AM |
1354 | } |
1355 | cachep->array[cpu] = nc; | |
1356 | l3 = cachep->nodelists[node]; | |
1357 | BUG_ON(!l3); | |
1358 | ||
1359 | spin_lock_irq(&l3->list_lock); | |
1360 | if (!l3->shared) { | |
1361 | /* | |
1362 | * We are serialised from CPU_DEAD or | |
1363 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1364 | */ | |
1365 | l3->shared = shared; | |
1366 | shared = NULL; | |
1367 | } | |
4484ebf1 | 1368 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1369 | if (!l3->alien) { |
1370 | l3->alien = alien; | |
1371 | alien = NULL; | |
1da177e4 | 1372 | } |
fbf1e473 AM |
1373 | #endif |
1374 | spin_unlock_irq(&l3->list_lock); | |
1375 | kfree(shared); | |
1376 | free_alien_cache(alien); | |
83835b3d PZ |
1377 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1378 | slab_set_debugobj_lock_classes_node(cachep, node); | |
6ccfb5bc GC |
1379 | else if (!OFF_SLAB(cachep) && |
1380 | !(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
1381 | on_slab_lock_classes_node(cachep, node); | |
fbf1e473 | 1382 | } |
ce79ddc8 PE |
1383 | init_node_lock_keys(node); |
1384 | ||
fbf1e473 AM |
1385 | return 0; |
1386 | bad: | |
12d00f6a | 1387 | cpuup_canceled(cpu); |
fbf1e473 AM |
1388 | return -ENOMEM; |
1389 | } | |
1390 | ||
1391 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1392 | unsigned long action, void *hcpu) | |
1393 | { | |
1394 | long cpu = (long)hcpu; | |
1395 | int err = 0; | |
1396 | ||
1397 | switch (action) { | |
fbf1e473 AM |
1398 | case CPU_UP_PREPARE: |
1399 | case CPU_UP_PREPARE_FROZEN: | |
18004c5d | 1400 | mutex_lock(&slab_mutex); |
fbf1e473 | 1401 | err = cpuup_prepare(cpu); |
18004c5d | 1402 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
1403 | break; |
1404 | case CPU_ONLINE: | |
8bb78442 | 1405 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1406 | start_cpu_timer(cpu); |
1407 | break; | |
1408 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1409 | case CPU_DOWN_PREPARE: |
8bb78442 | 1410 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 | 1411 | /* |
18004c5d | 1412 | * Shutdown cache reaper. Note that the slab_mutex is |
5830c590 CL |
1413 | * held so that if cache_reap() is invoked it cannot do |
1414 | * anything expensive but will only modify reap_work | |
1415 | * and reschedule the timer. | |
1416 | */ | |
afe2c511 | 1417 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1418 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1419 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1420 | break; |
1421 | case CPU_DOWN_FAILED: | |
8bb78442 | 1422 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1423 | start_cpu_timer(cpu); |
1424 | break; | |
1da177e4 | 1425 | case CPU_DEAD: |
8bb78442 | 1426 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1427 | /* |
1428 | * Even if all the cpus of a node are down, we don't free the | |
1429 | * kmem_list3 of any cache. This to avoid a race between | |
1430 | * cpu_down, and a kmalloc allocation from another cpu for | |
1431 | * memory from the node of the cpu going down. The list3 | |
1432 | * structure is usually allocated from kmem_cache_create() and | |
1433 | * gets destroyed at kmem_cache_destroy(). | |
1434 | */ | |
183ff22b | 1435 | /* fall through */ |
8f5be20b | 1436 | #endif |
1da177e4 | 1437 | case CPU_UP_CANCELED: |
8bb78442 | 1438 | case CPU_UP_CANCELED_FROZEN: |
18004c5d | 1439 | mutex_lock(&slab_mutex); |
fbf1e473 | 1440 | cpuup_canceled(cpu); |
18004c5d | 1441 | mutex_unlock(&slab_mutex); |
1da177e4 | 1442 | break; |
1da177e4 | 1443 | } |
eac40680 | 1444 | return notifier_from_errno(err); |
1da177e4 LT |
1445 | } |
1446 | ||
74b85f37 CS |
1447 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1448 | &cpuup_callback, NULL, 0 | |
1449 | }; | |
1da177e4 | 1450 | |
8f9f8d9e DR |
1451 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1452 | /* | |
1453 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1454 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1455 | * removed. | |
1456 | * | |
18004c5d | 1457 | * Must hold slab_mutex. |
8f9f8d9e DR |
1458 | */ |
1459 | static int __meminit drain_cache_nodelists_node(int node) | |
1460 | { | |
1461 | struct kmem_cache *cachep; | |
1462 | int ret = 0; | |
1463 | ||
18004c5d | 1464 | list_for_each_entry(cachep, &slab_caches, list) { |
6744f087 | 1465 | struct kmem_cache_node *l3; |
8f9f8d9e DR |
1466 | |
1467 | l3 = cachep->nodelists[node]; | |
1468 | if (!l3) | |
1469 | continue; | |
1470 | ||
1471 | drain_freelist(cachep, l3, l3->free_objects); | |
1472 | ||
1473 | if (!list_empty(&l3->slabs_full) || | |
1474 | !list_empty(&l3->slabs_partial)) { | |
1475 | ret = -EBUSY; | |
1476 | break; | |
1477 | } | |
1478 | } | |
1479 | return ret; | |
1480 | } | |
1481 | ||
1482 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1483 | unsigned long action, void *arg) | |
1484 | { | |
1485 | struct memory_notify *mnb = arg; | |
1486 | int ret = 0; | |
1487 | int nid; | |
1488 | ||
1489 | nid = mnb->status_change_nid; | |
1490 | if (nid < 0) | |
1491 | goto out; | |
1492 | ||
1493 | switch (action) { | |
1494 | case MEM_GOING_ONLINE: | |
18004c5d | 1495 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1496 | ret = init_cache_nodelists_node(nid); |
18004c5d | 1497 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1498 | break; |
1499 | case MEM_GOING_OFFLINE: | |
18004c5d | 1500 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1501 | ret = drain_cache_nodelists_node(nid); |
18004c5d | 1502 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1503 | break; |
1504 | case MEM_ONLINE: | |
1505 | case MEM_OFFLINE: | |
1506 | case MEM_CANCEL_ONLINE: | |
1507 | case MEM_CANCEL_OFFLINE: | |
1508 | break; | |
1509 | } | |
1510 | out: | |
5fda1bd5 | 1511 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1512 | } |
1513 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1514 | ||
e498be7d CL |
1515 | /* |
1516 | * swap the static kmem_list3 with kmalloced memory | |
1517 | */ | |
6744f087 | 1518 | static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list, |
8f9f8d9e | 1519 | int nodeid) |
e498be7d | 1520 | { |
6744f087 | 1521 | struct kmem_cache_node *ptr; |
e498be7d | 1522 | |
6744f087 | 1523 | ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid); |
e498be7d CL |
1524 | BUG_ON(!ptr); |
1525 | ||
6744f087 | 1526 | memcpy(ptr, list, sizeof(struct kmem_cache_node)); |
2b2d5493 IM |
1527 | /* |
1528 | * Do not assume that spinlocks can be initialized via memcpy: | |
1529 | */ | |
1530 | spin_lock_init(&ptr->list_lock); | |
1531 | ||
e498be7d CL |
1532 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1533 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1534 | } |
1535 | ||
556a169d PE |
1536 | /* |
1537 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1538 | * size of kmem_list3. | |
1539 | */ | |
1540 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1541 | { | |
1542 | int node; | |
1543 | ||
1544 | for_each_online_node(node) { | |
1545 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1546 | cachep->nodelists[node]->next_reap = jiffies + | |
1547 | REAPTIMEOUT_LIST3 + | |
1548 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1549 | } | |
1550 | } | |
1551 | ||
3c583465 CL |
1552 | /* |
1553 | * The memory after the last cpu cache pointer is used for the | |
1554 | * the nodelists pointer. | |
1555 | */ | |
1556 | static void setup_nodelists_pointer(struct kmem_cache *cachep) | |
1557 | { | |
6744f087 | 1558 | cachep->nodelists = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids]; |
3c583465 CL |
1559 | } |
1560 | ||
a737b3e2 AM |
1561 | /* |
1562 | * Initialisation. Called after the page allocator have been initialised and | |
1563 | * before smp_init(). | |
1da177e4 LT |
1564 | */ |
1565 | void __init kmem_cache_init(void) | |
1566 | { | |
e498be7d CL |
1567 | int i; |
1568 | ||
9b030cb8 | 1569 | kmem_cache = &kmem_cache_boot; |
3c583465 | 1570 | setup_nodelists_pointer(kmem_cache); |
9b030cb8 | 1571 | |
b6e68bc1 | 1572 | if (num_possible_nodes() == 1) |
62918a03 SS |
1573 | use_alien_caches = 0; |
1574 | ||
3c583465 | 1575 | for (i = 0; i < NUM_INIT_LISTS; i++) |
e498be7d | 1576 | kmem_list3_init(&initkmem_list3[i]); |
3c583465 | 1577 | |
9b030cb8 | 1578 | set_up_list3s(kmem_cache, CACHE_CACHE); |
1da177e4 LT |
1579 | |
1580 | /* | |
1581 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1582 | * page orders on machines with more than 32MB of memory if |
1583 | * not overridden on the command line. | |
1da177e4 | 1584 | */ |
3df1cccd | 1585 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1586 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1587 | |
1da177e4 LT |
1588 | /* Bootstrap is tricky, because several objects are allocated |
1589 | * from caches that do not exist yet: | |
9b030cb8 CL |
1590 | * 1) initialize the kmem_cache cache: it contains the struct |
1591 | * kmem_cache structures of all caches, except kmem_cache itself: | |
1592 | * kmem_cache is statically allocated. | |
e498be7d CL |
1593 | * Initially an __init data area is used for the head array and the |
1594 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1595 | * array at the end of the bootstrap. | |
1da177e4 | 1596 | * 2) Create the first kmalloc cache. |
343e0d7a | 1597 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1598 | * An __init data area is used for the head array. |
1599 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1600 | * head arrays. | |
9b030cb8 | 1601 | * 4) Replace the __init data head arrays for kmem_cache and the first |
1da177e4 | 1602 | * kmalloc cache with kmalloc allocated arrays. |
9b030cb8 | 1603 | * 5) Replace the __init data for kmem_list3 for kmem_cache and |
e498be7d CL |
1604 | * the other cache's with kmalloc allocated memory. |
1605 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1606 | */ |
1607 | ||
9b030cb8 | 1608 | /* 1) create the kmem_cache */ |
1da177e4 | 1609 | |
8da3430d | 1610 | /* |
b56efcf0 | 1611 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1612 | */ |
2f9baa9f CL |
1613 | create_boot_cache(kmem_cache, "kmem_cache", |
1614 | offsetof(struct kmem_cache, array[nr_cpu_ids]) + | |
6744f087 | 1615 | nr_node_ids * sizeof(struct kmem_cache_node *), |
2f9baa9f CL |
1616 | SLAB_HWCACHE_ALIGN); |
1617 | list_add(&kmem_cache->list, &slab_caches); | |
1da177e4 LT |
1618 | |
1619 | /* 2+3) create the kmalloc caches */ | |
1da177e4 | 1620 | |
a737b3e2 AM |
1621 | /* |
1622 | * Initialize the caches that provide memory for the array cache and the | |
1623 | * kmem_list3 structures first. Without this, further allocations will | |
1624 | * bug. | |
e498be7d CL |
1625 | */ |
1626 | ||
e3366016 CL |
1627 | kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac", |
1628 | kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS); | |
45530c44 CL |
1629 | |
1630 | if (INDEX_AC != INDEX_L3) | |
e3366016 CL |
1631 | kmalloc_caches[INDEX_L3] = |
1632 | create_kmalloc_cache("kmalloc-l3", | |
1633 | kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS); | |
e498be7d | 1634 | |
e0a42726 IM |
1635 | slab_early_init = 0; |
1636 | ||
e3366016 CL |
1637 | for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) { |
1638 | size_t cs_size = kmalloc_size(i); | |
1639 | ||
1640 | if (cs_size < KMALLOC_MIN_SIZE) | |
1641 | continue; | |
1642 | ||
1643 | if (!kmalloc_caches[i]) { | |
1644 | /* | |
1645 | * For performance, all the general caches are L1 aligned. | |
1646 | * This should be particularly beneficial on SMP boxes, as it | |
1647 | * eliminates "false sharing". | |
1648 | * Note for systems short on memory removing the alignment will | |
1649 | * allow tighter packing of the smaller caches. | |
1650 | */ | |
1651 | kmalloc_caches[i] = create_kmalloc_cache("kmalloc", | |
1652 | cs_size, ARCH_KMALLOC_FLAGS); | |
1653 | } | |
45530c44 | 1654 | |
4b51d669 | 1655 | #ifdef CONFIG_ZONE_DMA |
e3366016 CL |
1656 | kmalloc_dma_caches[i] = create_kmalloc_cache( |
1657 | "kmalloc-dma", cs_size, | |
45530c44 | 1658 | SLAB_CACHE_DMA|ARCH_KMALLOC_FLAGS); |
4b51d669 | 1659 | #endif |
1da177e4 LT |
1660 | } |
1661 | /* 4) Replace the bootstrap head arrays */ | |
1662 | { | |
2b2d5493 | 1663 | struct array_cache *ptr; |
e498be7d | 1664 | |
83b519e8 | 1665 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1666 | |
9b030cb8 | 1667 | memcpy(ptr, cpu_cache_get(kmem_cache), |
b28a02de | 1668 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1669 | /* |
1670 | * Do not assume that spinlocks can be initialized via memcpy: | |
1671 | */ | |
1672 | spin_lock_init(&ptr->lock); | |
1673 | ||
9b030cb8 | 1674 | kmem_cache->array[smp_processor_id()] = ptr; |
e498be7d | 1675 | |
83b519e8 | 1676 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1677 | |
e3366016 | 1678 | BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC]) |
b28a02de | 1679 | != &initarray_generic.cache); |
e3366016 | 1680 | memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]), |
b28a02de | 1681 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1682 | /* |
1683 | * Do not assume that spinlocks can be initialized via memcpy: | |
1684 | */ | |
1685 | spin_lock_init(&ptr->lock); | |
1686 | ||
e3366016 | 1687 | kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr; |
1da177e4 | 1688 | } |
e498be7d CL |
1689 | /* 5) Replace the bootstrap kmem_list3's */ |
1690 | { | |
1ca4cb24 PE |
1691 | int nid; |
1692 | ||
9c09a95c | 1693 | for_each_online_node(nid) { |
9b030cb8 | 1694 | init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1695 | |
e3366016 | 1696 | init_list(kmalloc_caches[INDEX_AC], |
1ca4cb24 | 1697 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1698 | |
1699 | if (INDEX_AC != INDEX_L3) { | |
e3366016 | 1700 | init_list(kmalloc_caches[INDEX_L3], |
1ca4cb24 | 1701 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1702 | } |
1703 | } | |
1704 | } | |
1da177e4 | 1705 | |
97d06609 | 1706 | slab_state = UP; |
e3366016 CL |
1707 | |
1708 | /* Create the proper names */ | |
1709 | for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) { | |
1710 | char *s; | |
1711 | struct kmem_cache *c = kmalloc_caches[i]; | |
1712 | ||
1713 | if (!c) | |
1714 | continue; | |
1715 | ||
1716 | s = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i)); | |
1717 | ||
1718 | BUG_ON(!s); | |
1719 | c->name = s; | |
1720 | ||
1721 | #ifdef CONFIG_ZONE_DMA | |
1722 | c = kmalloc_dma_caches[i]; | |
1723 | BUG_ON(!c); | |
1724 | s = kasprintf(GFP_NOWAIT, "dma-kmalloc-%d", kmalloc_size(i)); | |
1725 | BUG_ON(!s); | |
1726 | c->name = s; | |
1727 | #endif | |
1728 | } | |
8429db5c PE |
1729 | } |
1730 | ||
1731 | void __init kmem_cache_init_late(void) | |
1732 | { | |
1733 | struct kmem_cache *cachep; | |
1734 | ||
97d06609 | 1735 | slab_state = UP; |
52cef189 | 1736 | |
8429db5c | 1737 | /* 6) resize the head arrays to their final sizes */ |
18004c5d CL |
1738 | mutex_lock(&slab_mutex); |
1739 | list_for_each_entry(cachep, &slab_caches, list) | |
8429db5c PE |
1740 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1741 | BUG(); | |
18004c5d | 1742 | mutex_unlock(&slab_mutex); |
056c6241 | 1743 | |
947ca185 MW |
1744 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1745 | init_lock_keys(); | |
1746 | ||
97d06609 CL |
1747 | /* Done! */ |
1748 | slab_state = FULL; | |
1749 | ||
a737b3e2 AM |
1750 | /* |
1751 | * Register a cpu startup notifier callback that initializes | |
1752 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1753 | */ |
1754 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1755 | |
8f9f8d9e DR |
1756 | #ifdef CONFIG_NUMA |
1757 | /* | |
1758 | * Register a memory hotplug callback that initializes and frees | |
1759 | * nodelists. | |
1760 | */ | |
1761 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1762 | #endif | |
1763 | ||
a737b3e2 AM |
1764 | /* |
1765 | * The reap timers are started later, with a module init call: That part | |
1766 | * of the kernel is not yet operational. | |
1da177e4 LT |
1767 | */ |
1768 | } | |
1769 | ||
1770 | static int __init cpucache_init(void) | |
1771 | { | |
1772 | int cpu; | |
1773 | ||
a737b3e2 AM |
1774 | /* |
1775 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1776 | */ |
e498be7d | 1777 | for_each_online_cpu(cpu) |
a737b3e2 | 1778 | start_cpu_timer(cpu); |
a164f896 GC |
1779 | |
1780 | /* Done! */ | |
97d06609 | 1781 | slab_state = FULL; |
1da177e4 LT |
1782 | return 0; |
1783 | } | |
1da177e4 LT |
1784 | __initcall(cpucache_init); |
1785 | ||
8bdec192 RA |
1786 | static noinline void |
1787 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1788 | { | |
6744f087 | 1789 | struct kmem_cache_node *l3; |
8bdec192 RA |
1790 | struct slab *slabp; |
1791 | unsigned long flags; | |
1792 | int node; | |
1793 | ||
1794 | printk(KERN_WARNING | |
1795 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1796 | nodeid, gfpflags); | |
1797 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1798 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1799 | |
1800 | for_each_online_node(node) { | |
1801 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1802 | unsigned long active_slabs = 0, num_slabs = 0; | |
1803 | ||
1804 | l3 = cachep->nodelists[node]; | |
1805 | if (!l3) | |
1806 | continue; | |
1807 | ||
1808 | spin_lock_irqsave(&l3->list_lock, flags); | |
1809 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1810 | active_objs += cachep->num; | |
1811 | active_slabs++; | |
1812 | } | |
1813 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1814 | active_objs += slabp->inuse; | |
1815 | active_slabs++; | |
1816 | } | |
1817 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1818 | num_slabs++; | |
1819 | ||
1820 | free_objects += l3->free_objects; | |
1821 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1822 | ||
1823 | num_slabs += active_slabs; | |
1824 | num_objs = num_slabs * cachep->num; | |
1825 | printk(KERN_WARNING | |
1826 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1827 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1828 | free_objects); | |
1829 | } | |
1830 | } | |
1831 | ||
1da177e4 LT |
1832 | /* |
1833 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1834 | * | |
1835 | * If we requested dmaable memory, we will get it. Even if we | |
1836 | * did not request dmaable memory, we might get it, but that | |
1837 | * would be relatively rare and ignorable. | |
1838 | */ | |
343e0d7a | 1839 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1840 | { |
1841 | struct page *page; | |
e1b6aa6f | 1842 | int nr_pages; |
1da177e4 LT |
1843 | int i; |
1844 | ||
d6fef9da | 1845 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1846 | /* |
1847 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1848 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1849 | */ |
e1b6aa6f | 1850 | flags |= __GFP_COMP; |
d6fef9da | 1851 | #endif |
765c4507 | 1852 | |
a618e89f | 1853 | flags |= cachep->allocflags; |
e12ba74d MG |
1854 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1855 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1856 | |
517d0869 | 1857 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1858 | if (!page) { |
1859 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1860 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1861 | return NULL; |
8bdec192 | 1862 | } |
1da177e4 | 1863 | |
b37f1dd0 | 1864 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
072bb0aa MG |
1865 | if (unlikely(page->pfmemalloc)) |
1866 | pfmemalloc_active = true; | |
1867 | ||
e1b6aa6f | 1868 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1869 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1870 | add_zone_page_state(page_zone(page), |
1871 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1872 | else | |
1873 | add_zone_page_state(page_zone(page), | |
1874 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
072bb0aa | 1875 | for (i = 0; i < nr_pages; i++) { |
e1b6aa6f | 1876 | __SetPageSlab(page + i); |
c175eea4 | 1877 | |
072bb0aa MG |
1878 | if (page->pfmemalloc) |
1879 | SetPageSlabPfmemalloc(page + i); | |
1880 | } | |
1f458cbf | 1881 | memcg_bind_pages(cachep, cachep->gfporder); |
072bb0aa | 1882 | |
b1eeab67 VN |
1883 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1884 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1885 | ||
1886 | if (cachep->ctor) | |
1887 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1888 | else | |
1889 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1890 | } | |
c175eea4 | 1891 | |
e1b6aa6f | 1892 | return page_address(page); |
1da177e4 LT |
1893 | } |
1894 | ||
1895 | /* | |
1896 | * Interface to system's page release. | |
1897 | */ | |
343e0d7a | 1898 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1899 | { |
b28a02de | 1900 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1901 | struct page *page = virt_to_page(addr); |
1902 | const unsigned long nr_freed = i; | |
1903 | ||
b1eeab67 | 1904 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1905 | |
972d1a7b CL |
1906 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1907 | sub_zone_page_state(page_zone(page), | |
1908 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1909 | else | |
1910 | sub_zone_page_state(page_zone(page), | |
1911 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1912 | while (i--) { |
f205b2fe | 1913 | BUG_ON(!PageSlab(page)); |
072bb0aa | 1914 | __ClearPageSlabPfmemalloc(page); |
f205b2fe | 1915 | __ClearPageSlab(page); |
1da177e4 LT |
1916 | page++; |
1917 | } | |
1f458cbf GC |
1918 | |
1919 | memcg_release_pages(cachep, cachep->gfporder); | |
1da177e4 LT |
1920 | if (current->reclaim_state) |
1921 | current->reclaim_state->reclaimed_slab += nr_freed; | |
d79923fa | 1922 | free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder); |
1da177e4 LT |
1923 | } |
1924 | ||
1925 | static void kmem_rcu_free(struct rcu_head *head) | |
1926 | { | |
b28a02de | 1927 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1928 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1929 | |
1930 | kmem_freepages(cachep, slab_rcu->addr); | |
1931 | if (OFF_SLAB(cachep)) | |
1932 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1933 | } | |
1934 | ||
1935 | #if DEBUG | |
1936 | ||
1937 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1938 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1939 | unsigned long caller) |
1da177e4 | 1940 | { |
8c138bc0 | 1941 | int size = cachep->object_size; |
1da177e4 | 1942 | |
3dafccf2 | 1943 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1944 | |
b28a02de | 1945 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1946 | return; |
1947 | ||
b28a02de PE |
1948 | *addr++ = 0x12345678; |
1949 | *addr++ = caller; | |
1950 | *addr++ = smp_processor_id(); | |
1951 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1952 | { |
1953 | unsigned long *sptr = &caller; | |
1954 | unsigned long svalue; | |
1955 | ||
1956 | while (!kstack_end(sptr)) { | |
1957 | svalue = *sptr++; | |
1958 | if (kernel_text_address(svalue)) { | |
b28a02de | 1959 | *addr++ = svalue; |
1da177e4 LT |
1960 | size -= sizeof(unsigned long); |
1961 | if (size <= sizeof(unsigned long)) | |
1962 | break; | |
1963 | } | |
1964 | } | |
1965 | ||
1966 | } | |
b28a02de | 1967 | *addr++ = 0x87654321; |
1da177e4 LT |
1968 | } |
1969 | #endif | |
1970 | ||
343e0d7a | 1971 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1972 | { |
8c138bc0 | 1973 | int size = cachep->object_size; |
3dafccf2 | 1974 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1975 | |
1976 | memset(addr, val, size); | |
b28a02de | 1977 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1978 | } |
1979 | ||
1980 | static void dump_line(char *data, int offset, int limit) | |
1981 | { | |
1982 | int i; | |
aa83aa40 DJ |
1983 | unsigned char error = 0; |
1984 | int bad_count = 0; | |
1985 | ||
fdde6abb | 1986 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
1987 | for (i = 0; i < limit; i++) { |
1988 | if (data[offset + i] != POISON_FREE) { | |
1989 | error = data[offset + i]; | |
1990 | bad_count++; | |
1991 | } | |
aa83aa40 | 1992 | } |
fdde6abb SAS |
1993 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1994 | &data[offset], limit, 1); | |
aa83aa40 DJ |
1995 | |
1996 | if (bad_count == 1) { | |
1997 | error ^= POISON_FREE; | |
1998 | if (!(error & (error - 1))) { | |
1999 | printk(KERN_ERR "Single bit error detected. Probably " | |
2000 | "bad RAM.\n"); | |
2001 | #ifdef CONFIG_X86 | |
2002 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
2003 | "test tool.\n"); | |
2004 | #else | |
2005 | printk(KERN_ERR "Run a memory test tool.\n"); | |
2006 | #endif | |
2007 | } | |
2008 | } | |
1da177e4 LT |
2009 | } |
2010 | #endif | |
2011 | ||
2012 | #if DEBUG | |
2013 | ||
343e0d7a | 2014 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
2015 | { |
2016 | int i, size; | |
2017 | char *realobj; | |
2018 | ||
2019 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 2020 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
2021 | *dbg_redzone1(cachep, objp), |
2022 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2023 | } |
2024 | ||
2025 | if (cachep->flags & SLAB_STORE_USER) { | |
2026 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 2027 | *dbg_userword(cachep, objp)); |
1da177e4 | 2028 | print_symbol("(%s)", |
a737b3e2 | 2029 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
2030 | printk("\n"); |
2031 | } | |
3dafccf2 | 2032 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2033 | size = cachep->object_size; |
b28a02de | 2034 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
2035 | int limit; |
2036 | limit = 16; | |
b28a02de PE |
2037 | if (i + limit > size) |
2038 | limit = size - i; | |
1da177e4 LT |
2039 | dump_line(realobj, i, limit); |
2040 | } | |
2041 | } | |
2042 | ||
343e0d7a | 2043 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
2044 | { |
2045 | char *realobj; | |
2046 | int size, i; | |
2047 | int lines = 0; | |
2048 | ||
3dafccf2 | 2049 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2050 | size = cachep->object_size; |
1da177e4 | 2051 | |
b28a02de | 2052 | for (i = 0; i < size; i++) { |
1da177e4 | 2053 | char exp = POISON_FREE; |
b28a02de | 2054 | if (i == size - 1) |
1da177e4 LT |
2055 | exp = POISON_END; |
2056 | if (realobj[i] != exp) { | |
2057 | int limit; | |
2058 | /* Mismatch ! */ | |
2059 | /* Print header */ | |
2060 | if (lines == 0) { | |
b28a02de | 2061 | printk(KERN_ERR |
face37f5 DJ |
2062 | "Slab corruption (%s): %s start=%p, len=%d\n", |
2063 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
2064 | print_objinfo(cachep, objp, 0); |
2065 | } | |
2066 | /* Hexdump the affected line */ | |
b28a02de | 2067 | i = (i / 16) * 16; |
1da177e4 | 2068 | limit = 16; |
b28a02de PE |
2069 | if (i + limit > size) |
2070 | limit = size - i; | |
1da177e4 LT |
2071 | dump_line(realobj, i, limit); |
2072 | i += 16; | |
2073 | lines++; | |
2074 | /* Limit to 5 lines */ | |
2075 | if (lines > 5) | |
2076 | break; | |
2077 | } | |
2078 | } | |
2079 | if (lines != 0) { | |
2080 | /* Print some data about the neighboring objects, if they | |
2081 | * exist: | |
2082 | */ | |
6ed5eb22 | 2083 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 2084 | unsigned int objnr; |
1da177e4 | 2085 | |
8fea4e96 | 2086 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2087 | if (objnr) { |
8fea4e96 | 2088 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2089 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2090 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2091 | realobj, size); |
1da177e4 LT |
2092 | print_objinfo(cachep, objp, 2); |
2093 | } | |
b28a02de | 2094 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2095 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2096 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2097 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2098 | realobj, size); |
1da177e4 LT |
2099 | print_objinfo(cachep, objp, 2); |
2100 | } | |
2101 | } | |
2102 | } | |
2103 | #endif | |
2104 | ||
12dd36fa | 2105 | #if DEBUG |
e79aec29 | 2106 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2107 | { |
1da177e4 LT |
2108 | int i; |
2109 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2110 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2111 | |
2112 | if (cachep->flags & SLAB_POISON) { | |
2113 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2114 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2115 | OFF_SLAB(cachep)) |
b28a02de | 2116 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2117 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2118 | else |
2119 | check_poison_obj(cachep, objp); | |
2120 | #else | |
2121 | check_poison_obj(cachep, objp); | |
2122 | #endif | |
2123 | } | |
2124 | if (cachep->flags & SLAB_RED_ZONE) { | |
2125 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2126 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2127 | "was overwritten"); |
1da177e4 LT |
2128 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2129 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2130 | "was overwritten"); |
1da177e4 | 2131 | } |
1da177e4 | 2132 | } |
12dd36fa | 2133 | } |
1da177e4 | 2134 | #else |
e79aec29 | 2135 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2136 | { |
12dd36fa | 2137 | } |
1da177e4 LT |
2138 | #endif |
2139 | ||
911851e6 RD |
2140 | /** |
2141 | * slab_destroy - destroy and release all objects in a slab | |
2142 | * @cachep: cache pointer being destroyed | |
2143 | * @slabp: slab pointer being destroyed | |
2144 | * | |
12dd36fa | 2145 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2146 | * Before calling the slab must have been unlinked from the cache. The |
2147 | * cache-lock is not held/needed. | |
12dd36fa | 2148 | */ |
343e0d7a | 2149 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2150 | { |
2151 | void *addr = slabp->s_mem - slabp->colouroff; | |
2152 | ||
e79aec29 | 2153 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2154 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2155 | struct slab_rcu *slab_rcu; | |
2156 | ||
b28a02de | 2157 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2158 | slab_rcu->cachep = cachep; |
2159 | slab_rcu->addr = addr; | |
2160 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2161 | } else { | |
2162 | kmem_freepages(cachep, addr); | |
873623df IM |
2163 | if (OFF_SLAB(cachep)) |
2164 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2165 | } |
2166 | } | |
2167 | ||
4d268eba | 2168 | /** |
a70773dd RD |
2169 | * calculate_slab_order - calculate size (page order) of slabs |
2170 | * @cachep: pointer to the cache that is being created | |
2171 | * @size: size of objects to be created in this cache. | |
2172 | * @align: required alignment for the objects. | |
2173 | * @flags: slab allocation flags | |
2174 | * | |
2175 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2176 | * |
2177 | * This could be made much more intelligent. For now, try to avoid using | |
2178 | * high order pages for slabs. When the gfp() functions are more friendly | |
2179 | * towards high-order requests, this should be changed. | |
2180 | */ | |
a737b3e2 | 2181 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2182 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2183 | { |
b1ab41c4 | 2184 | unsigned long offslab_limit; |
4d268eba | 2185 | size_t left_over = 0; |
9888e6fa | 2186 | int gfporder; |
4d268eba | 2187 | |
0aa817f0 | 2188 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2189 | unsigned int num; |
2190 | size_t remainder; | |
2191 | ||
9888e6fa | 2192 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2193 | if (!num) |
2194 | continue; | |
9888e6fa | 2195 | |
b1ab41c4 IM |
2196 | if (flags & CFLGS_OFF_SLAB) { |
2197 | /* | |
2198 | * Max number of objs-per-slab for caches which | |
2199 | * use off-slab slabs. Needed to avoid a possible | |
2200 | * looping condition in cache_grow(). | |
2201 | */ | |
2202 | offslab_limit = size - sizeof(struct slab); | |
2203 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2204 | ||
2205 | if (num > offslab_limit) | |
2206 | break; | |
2207 | } | |
4d268eba | 2208 | |
9888e6fa | 2209 | /* Found something acceptable - save it away */ |
4d268eba | 2210 | cachep->num = num; |
9888e6fa | 2211 | cachep->gfporder = gfporder; |
4d268eba PE |
2212 | left_over = remainder; |
2213 | ||
f78bb8ad LT |
2214 | /* |
2215 | * A VFS-reclaimable slab tends to have most allocations | |
2216 | * as GFP_NOFS and we really don't want to have to be allocating | |
2217 | * higher-order pages when we are unable to shrink dcache. | |
2218 | */ | |
2219 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2220 | break; | |
2221 | ||
4d268eba PE |
2222 | /* |
2223 | * Large number of objects is good, but very large slabs are | |
2224 | * currently bad for the gfp()s. | |
2225 | */ | |
543585cc | 2226 | if (gfporder >= slab_max_order) |
4d268eba PE |
2227 | break; |
2228 | ||
9888e6fa LT |
2229 | /* |
2230 | * Acceptable internal fragmentation? | |
2231 | */ | |
a737b3e2 | 2232 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2233 | break; |
2234 | } | |
2235 | return left_over; | |
2236 | } | |
2237 | ||
83b519e8 | 2238 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2239 | { |
97d06609 | 2240 | if (slab_state >= FULL) |
83b519e8 | 2241 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2242 | |
97d06609 | 2243 | if (slab_state == DOWN) { |
f30cf7d1 | 2244 | /* |
2f9baa9f CL |
2245 | * Note: Creation of first cache (kmem_cache). |
2246 | * The setup_list3s is taken care | |
2247 | * of by the caller of __kmem_cache_create | |
2248 | */ | |
2249 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2250 | slab_state = PARTIAL; | |
2251 | } else if (slab_state == PARTIAL) { | |
2252 | /* | |
2253 | * Note: the second kmem_cache_create must create the cache | |
f30cf7d1 PE |
2254 | * that's used by kmalloc(24), otherwise the creation of |
2255 | * further caches will BUG(). | |
2256 | */ | |
2257 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2258 | ||
2259 | /* | |
2260 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2f9baa9f | 2261 | * the second cache, then we need to set up all its list3s, |
f30cf7d1 PE |
2262 | * otherwise the creation of further caches will BUG(). |
2263 | */ | |
2264 | set_up_list3s(cachep, SIZE_AC); | |
2265 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2266 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2267 | else |
97d06609 | 2268 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 | 2269 | } else { |
2f9baa9f | 2270 | /* Remaining boot caches */ |
f30cf7d1 | 2271 | cachep->array[smp_processor_id()] = |
83b519e8 | 2272 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2273 | |
97d06609 | 2274 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2275 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2276 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2277 | } else { |
2278 | int node; | |
556a169d | 2279 | for_each_online_node(node) { |
f30cf7d1 | 2280 | cachep->nodelists[node] = |
6744f087 | 2281 | kmalloc_node(sizeof(struct kmem_cache_node), |
eb91f1d0 | 2282 | gfp, node); |
f30cf7d1 PE |
2283 | BUG_ON(!cachep->nodelists[node]); |
2284 | kmem_list3_init(cachep->nodelists[node]); | |
2285 | } | |
2286 | } | |
2287 | } | |
7d6e6d09 | 2288 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2289 | jiffies + REAPTIMEOUT_LIST3 + |
2290 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2291 | ||
2292 | cpu_cache_get(cachep)->avail = 0; | |
2293 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2294 | cpu_cache_get(cachep)->batchcount = 1; | |
2295 | cpu_cache_get(cachep)->touched = 0; | |
2296 | cachep->batchcount = 1; | |
2297 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2298 | return 0; |
f30cf7d1 PE |
2299 | } |
2300 | ||
1da177e4 | 2301 | /** |
039363f3 | 2302 | * __kmem_cache_create - Create a cache. |
a755b76a | 2303 | * @cachep: cache management descriptor |
1da177e4 | 2304 | * @flags: SLAB flags |
1da177e4 LT |
2305 | * |
2306 | * Returns a ptr to the cache on success, NULL on failure. | |
2307 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2308 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 | 2309 | * |
1da177e4 LT |
2310 | * The flags are |
2311 | * | |
2312 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2313 | * to catch references to uninitialised memory. | |
2314 | * | |
2315 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2316 | * for buffer overruns. | |
2317 | * | |
1da177e4 LT |
2318 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2319 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2320 | * as davem. | |
2321 | */ | |
278b1bb1 | 2322 | int |
8a13a4cc | 2323 | __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) |
1da177e4 LT |
2324 | { |
2325 | size_t left_over, slab_size, ralign; | |
83b519e8 | 2326 | gfp_t gfp; |
278b1bb1 | 2327 | int err; |
8a13a4cc | 2328 | size_t size = cachep->size; |
1da177e4 | 2329 | |
1da177e4 | 2330 | #if DEBUG |
1da177e4 LT |
2331 | #if FORCED_DEBUG |
2332 | /* | |
2333 | * Enable redzoning and last user accounting, except for caches with | |
2334 | * large objects, if the increased size would increase the object size | |
2335 | * above the next power of two: caches with object sizes just above a | |
2336 | * power of two have a significant amount of internal fragmentation. | |
2337 | */ | |
87a927c7 DW |
2338 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2339 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2340 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2341 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2342 | flags |= SLAB_POISON; | |
2343 | #endif | |
2344 | if (flags & SLAB_DESTROY_BY_RCU) | |
2345 | BUG_ON(flags & SLAB_POISON); | |
2346 | #endif | |
1da177e4 | 2347 | |
a737b3e2 AM |
2348 | /* |
2349 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2350 | * unaligned accesses for some archs when redzoning is used, and makes |
2351 | * sure any on-slab bufctl's are also correctly aligned. | |
2352 | */ | |
b28a02de PE |
2353 | if (size & (BYTES_PER_WORD - 1)) { |
2354 | size += (BYTES_PER_WORD - 1); | |
2355 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2356 | } |
2357 | ||
ca5f9703 | 2358 | /* |
87a927c7 DW |
2359 | * Redzoning and user store require word alignment or possibly larger. |
2360 | * Note this will be overridden by architecture or caller mandated | |
2361 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2362 | */ |
87a927c7 DW |
2363 | if (flags & SLAB_STORE_USER) |
2364 | ralign = BYTES_PER_WORD; | |
2365 | ||
2366 | if (flags & SLAB_RED_ZONE) { | |
2367 | ralign = REDZONE_ALIGN; | |
2368 | /* If redzoning, ensure that the second redzone is suitably | |
2369 | * aligned, by adjusting the object size accordingly. */ | |
2370 | size += REDZONE_ALIGN - 1; | |
2371 | size &= ~(REDZONE_ALIGN - 1); | |
2372 | } | |
ca5f9703 | 2373 | |
a44b56d3 | 2374 | /* 3) caller mandated alignment */ |
8a13a4cc CL |
2375 | if (ralign < cachep->align) { |
2376 | ralign = cachep->align; | |
1da177e4 | 2377 | } |
3ff84a7f PE |
2378 | /* disable debug if necessary */ |
2379 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2380 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2381 | /* |
ca5f9703 | 2382 | * 4) Store it. |
1da177e4 | 2383 | */ |
8a13a4cc | 2384 | cachep->align = ralign; |
1da177e4 | 2385 | |
83b519e8 PE |
2386 | if (slab_is_available()) |
2387 | gfp = GFP_KERNEL; | |
2388 | else | |
2389 | gfp = GFP_NOWAIT; | |
2390 | ||
3c583465 | 2391 | setup_nodelists_pointer(cachep); |
1da177e4 | 2392 | #if DEBUG |
1da177e4 | 2393 | |
ca5f9703 PE |
2394 | /* |
2395 | * Both debugging options require word-alignment which is calculated | |
2396 | * into align above. | |
2397 | */ | |
1da177e4 | 2398 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2399 | /* add space for red zone words */ |
3ff84a7f PE |
2400 | cachep->obj_offset += sizeof(unsigned long long); |
2401 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2402 | } |
2403 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2404 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2405 | * the real object. But if the second red zone needs to be |
2406 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2407 | */ |
87a927c7 DW |
2408 | if (flags & SLAB_RED_ZONE) |
2409 | size += REDZONE_ALIGN; | |
2410 | else | |
2411 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2412 | } |
2413 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
e3366016 CL |
2414 | if (size >= kmalloc_size(INDEX_L3 + 1) |
2415 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { | |
2416 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); | |
1da177e4 LT |
2417 | size = PAGE_SIZE; |
2418 | } | |
2419 | #endif | |
2420 | #endif | |
2421 | ||
e0a42726 IM |
2422 | /* |
2423 | * Determine if the slab management is 'on' or 'off' slab. | |
2424 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2425 | * it too early on. Always use on-slab management when |
2426 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2427 | */ |
e7cb55b9 CM |
2428 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2429 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2430 | /* |
2431 | * Size is large, assume best to place the slab management obj | |
2432 | * off-slab (should allow better packing of objs). | |
2433 | */ | |
2434 | flags |= CFLGS_OFF_SLAB; | |
2435 | ||
8a13a4cc | 2436 | size = ALIGN(size, cachep->align); |
1da177e4 | 2437 | |
8a13a4cc | 2438 | left_over = calculate_slab_order(cachep, size, cachep->align, flags); |
1da177e4 | 2439 | |
8a13a4cc | 2440 | if (!cachep->num) |
278b1bb1 | 2441 | return -E2BIG; |
1da177e4 | 2442 | |
b28a02de | 2443 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
8a13a4cc | 2444 | + sizeof(struct slab), cachep->align); |
1da177e4 LT |
2445 | |
2446 | /* | |
2447 | * If the slab has been placed off-slab, and we have enough space then | |
2448 | * move it on-slab. This is at the expense of any extra colouring. | |
2449 | */ | |
2450 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2451 | flags &= ~CFLGS_OFF_SLAB; | |
2452 | left_over -= slab_size; | |
2453 | } | |
2454 | ||
2455 | if (flags & CFLGS_OFF_SLAB) { | |
2456 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2457 | slab_size = |
2458 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2459 | |
2460 | #ifdef CONFIG_PAGE_POISONING | |
2461 | /* If we're going to use the generic kernel_map_pages() | |
2462 | * poisoning, then it's going to smash the contents of | |
2463 | * the redzone and userword anyhow, so switch them off. | |
2464 | */ | |
2465 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2466 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2467 | #endif | |
1da177e4 LT |
2468 | } |
2469 | ||
2470 | cachep->colour_off = cache_line_size(); | |
2471 | /* Offset must be a multiple of the alignment. */ | |
8a13a4cc CL |
2472 | if (cachep->colour_off < cachep->align) |
2473 | cachep->colour_off = cachep->align; | |
b28a02de | 2474 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2475 | cachep->slab_size = slab_size; |
2476 | cachep->flags = flags; | |
a618e89f | 2477 | cachep->allocflags = 0; |
4b51d669 | 2478 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2479 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2480 | cachep->size = size; |
6a2d7a95 | 2481 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2482 | |
e5ac9c5a | 2483 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2484 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2485 | /* |
2486 | * This is a possibility for one of the malloc_sizes caches. | |
2487 | * But since we go off slab only for object size greater than | |
2488 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2489 | * this should not happen at all. | |
2490 | * But leave a BUG_ON for some lucky dude. | |
2491 | */ | |
6cb8f913 | 2492 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2493 | } |
1da177e4 | 2494 | |
278b1bb1 CL |
2495 | err = setup_cpu_cache(cachep, gfp); |
2496 | if (err) { | |
12c3667f | 2497 | __kmem_cache_shutdown(cachep); |
278b1bb1 | 2498 | return err; |
2ed3a4ef | 2499 | } |
1da177e4 | 2500 | |
83835b3d PZ |
2501 | if (flags & SLAB_DEBUG_OBJECTS) { |
2502 | /* | |
2503 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2504 | * debug_object_activate()->kmem_cache_alloc(). | |
2505 | */ | |
2506 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2507 | ||
2508 | slab_set_debugobj_lock_classes(cachep); | |
6ccfb5bc GC |
2509 | } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU)) |
2510 | on_slab_lock_classes(cachep); | |
83835b3d | 2511 | |
278b1bb1 | 2512 | return 0; |
1da177e4 | 2513 | } |
1da177e4 LT |
2514 | |
2515 | #if DEBUG | |
2516 | static void check_irq_off(void) | |
2517 | { | |
2518 | BUG_ON(!irqs_disabled()); | |
2519 | } | |
2520 | ||
2521 | static void check_irq_on(void) | |
2522 | { | |
2523 | BUG_ON(irqs_disabled()); | |
2524 | } | |
2525 | ||
343e0d7a | 2526 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2527 | { |
2528 | #ifdef CONFIG_SMP | |
2529 | check_irq_off(); | |
7d6e6d09 | 2530 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2531 | #endif |
2532 | } | |
e498be7d | 2533 | |
343e0d7a | 2534 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2535 | { |
2536 | #ifdef CONFIG_SMP | |
2537 | check_irq_off(); | |
2538 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2539 | #endif | |
2540 | } | |
2541 | ||
1da177e4 LT |
2542 | #else |
2543 | #define check_irq_off() do { } while(0) | |
2544 | #define check_irq_on() do { } while(0) | |
2545 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2546 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2547 | #endif |
2548 | ||
6744f087 | 2549 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3, |
aab2207c CL |
2550 | struct array_cache *ac, |
2551 | int force, int node); | |
2552 | ||
1da177e4 LT |
2553 | static void do_drain(void *arg) |
2554 | { | |
a737b3e2 | 2555 | struct kmem_cache *cachep = arg; |
1da177e4 | 2556 | struct array_cache *ac; |
7d6e6d09 | 2557 | int node = numa_mem_id(); |
1da177e4 LT |
2558 | |
2559 | check_irq_off(); | |
9a2dba4b | 2560 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2561 | spin_lock(&cachep->nodelists[node]->list_lock); |
2562 | free_block(cachep, ac->entry, ac->avail, node); | |
2563 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2564 | ac->avail = 0; |
2565 | } | |
2566 | ||
343e0d7a | 2567 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2568 | { |
6744f087 | 2569 | struct kmem_cache_node *l3; |
e498be7d CL |
2570 | int node; |
2571 | ||
15c8b6c1 | 2572 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2573 | check_irq_on(); |
b28a02de | 2574 | for_each_online_node(node) { |
e498be7d | 2575 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2576 | if (l3 && l3->alien) |
2577 | drain_alien_cache(cachep, l3->alien); | |
2578 | } | |
2579 | ||
2580 | for_each_online_node(node) { | |
2581 | l3 = cachep->nodelists[node]; | |
2582 | if (l3) | |
aab2207c | 2583 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2584 | } |
1da177e4 LT |
2585 | } |
2586 | ||
ed11d9eb CL |
2587 | /* |
2588 | * Remove slabs from the list of free slabs. | |
2589 | * Specify the number of slabs to drain in tofree. | |
2590 | * | |
2591 | * Returns the actual number of slabs released. | |
2592 | */ | |
2593 | static int drain_freelist(struct kmem_cache *cache, | |
6744f087 | 2594 | struct kmem_cache_node *l3, int tofree) |
1da177e4 | 2595 | { |
ed11d9eb CL |
2596 | struct list_head *p; |
2597 | int nr_freed; | |
1da177e4 | 2598 | struct slab *slabp; |
1da177e4 | 2599 | |
ed11d9eb CL |
2600 | nr_freed = 0; |
2601 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2602 | |
ed11d9eb | 2603 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2604 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2605 | if (p == &l3->slabs_free) { |
2606 | spin_unlock_irq(&l3->list_lock); | |
2607 | goto out; | |
2608 | } | |
1da177e4 | 2609 | |
ed11d9eb | 2610 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2611 | #if DEBUG |
40094fa6 | 2612 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2613 | #endif |
2614 | list_del(&slabp->list); | |
ed11d9eb CL |
2615 | /* |
2616 | * Safe to drop the lock. The slab is no longer linked | |
2617 | * to the cache. | |
2618 | */ | |
2619 | l3->free_objects -= cache->num; | |
e498be7d | 2620 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2621 | slab_destroy(cache, slabp); |
2622 | nr_freed++; | |
1da177e4 | 2623 | } |
ed11d9eb CL |
2624 | out: |
2625 | return nr_freed; | |
1da177e4 LT |
2626 | } |
2627 | ||
18004c5d | 2628 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2629 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2630 | { |
2631 | int ret = 0, i = 0; | |
6744f087 | 2632 | struct kmem_cache_node *l3; |
e498be7d CL |
2633 | |
2634 | drain_cpu_caches(cachep); | |
2635 | ||
2636 | check_irq_on(); | |
2637 | for_each_online_node(i) { | |
2638 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2639 | if (!l3) |
2640 | continue; | |
2641 | ||
2642 | drain_freelist(cachep, l3, l3->free_objects); | |
2643 | ||
2644 | ret += !list_empty(&l3->slabs_full) || | |
2645 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2646 | } |
2647 | return (ret ? 1 : 0); | |
2648 | } | |
2649 | ||
1da177e4 LT |
2650 | /** |
2651 | * kmem_cache_shrink - Shrink a cache. | |
2652 | * @cachep: The cache to shrink. | |
2653 | * | |
2654 | * Releases as many slabs as possible for a cache. | |
2655 | * To help debugging, a zero exit status indicates all slabs were released. | |
2656 | */ | |
343e0d7a | 2657 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2658 | { |
8f5be20b | 2659 | int ret; |
40094fa6 | 2660 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2661 | |
95402b38 | 2662 | get_online_cpus(); |
18004c5d | 2663 | mutex_lock(&slab_mutex); |
8f5be20b | 2664 | ret = __cache_shrink(cachep); |
18004c5d | 2665 | mutex_unlock(&slab_mutex); |
95402b38 | 2666 | put_online_cpus(); |
8f5be20b | 2667 | return ret; |
1da177e4 LT |
2668 | } |
2669 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2670 | ||
945cf2b6 | 2671 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
1da177e4 | 2672 | { |
12c3667f | 2673 | int i; |
6744f087 | 2674 | struct kmem_cache_node *l3; |
12c3667f | 2675 | int rc = __cache_shrink(cachep); |
1da177e4 | 2676 | |
12c3667f CL |
2677 | if (rc) |
2678 | return rc; | |
1da177e4 | 2679 | |
12c3667f CL |
2680 | for_each_online_cpu(i) |
2681 | kfree(cachep->array[i]); | |
1da177e4 | 2682 | |
12c3667f CL |
2683 | /* NUMA: free the list3 structures */ |
2684 | for_each_online_node(i) { | |
2685 | l3 = cachep->nodelists[i]; | |
2686 | if (l3) { | |
2687 | kfree(l3->shared); | |
2688 | free_alien_cache(l3->alien); | |
2689 | kfree(l3); | |
2690 | } | |
2691 | } | |
2692 | return 0; | |
1da177e4 | 2693 | } |
1da177e4 | 2694 | |
e5ac9c5a RT |
2695 | /* |
2696 | * Get the memory for a slab management obj. | |
2697 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2698 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2699 | * come from the same cache which is getting created because, | |
2700 | * when we are searching for an appropriate cache for these | |
2701 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2702 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2703 | * kmem_find_general_cachep till the initialization is complete. | |
2704 | * Hence we cannot have slabp_cache same as the original cache. | |
2705 | */ | |
343e0d7a | 2706 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2707 | int colour_off, gfp_t local_flags, |
2708 | int nodeid) | |
1da177e4 LT |
2709 | { |
2710 | struct slab *slabp; | |
b28a02de | 2711 | |
1da177e4 LT |
2712 | if (OFF_SLAB(cachep)) { |
2713 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2714 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2715 | local_flags, nodeid); |
d5cff635 CM |
2716 | /* |
2717 | * If the first object in the slab is leaked (it's allocated | |
2718 | * but no one has a reference to it), we want to make sure | |
2719 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2720 | * to the object. Otherwise we will not report the leak. | |
2721 | */ | |
c017b4be CM |
2722 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2723 | local_flags); | |
1da177e4 LT |
2724 | if (!slabp) |
2725 | return NULL; | |
2726 | } else { | |
b28a02de | 2727 | slabp = objp + colour_off; |
1da177e4 LT |
2728 | colour_off += cachep->slab_size; |
2729 | } | |
2730 | slabp->inuse = 0; | |
2731 | slabp->colouroff = colour_off; | |
b28a02de | 2732 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2733 | slabp->nodeid = nodeid; |
e51bfd0a | 2734 | slabp->free = 0; |
1da177e4 LT |
2735 | return slabp; |
2736 | } | |
2737 | ||
2738 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2739 | { | |
b28a02de | 2740 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2741 | } |
2742 | ||
343e0d7a | 2743 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2744 | struct slab *slabp) |
1da177e4 LT |
2745 | { |
2746 | int i; | |
2747 | ||
2748 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2749 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2750 | #if DEBUG |
2751 | /* need to poison the objs? */ | |
2752 | if (cachep->flags & SLAB_POISON) | |
2753 | poison_obj(cachep, objp, POISON_FREE); | |
2754 | if (cachep->flags & SLAB_STORE_USER) | |
2755 | *dbg_userword(cachep, objp) = NULL; | |
2756 | ||
2757 | if (cachep->flags & SLAB_RED_ZONE) { | |
2758 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2759 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2760 | } | |
2761 | /* | |
a737b3e2 AM |
2762 | * Constructors are not allowed to allocate memory from the same |
2763 | * cache which they are a constructor for. Otherwise, deadlock. | |
2764 | * They must also be threaded. | |
1da177e4 LT |
2765 | */ |
2766 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2767 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2768 | |
2769 | if (cachep->flags & SLAB_RED_ZONE) { | |
2770 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2771 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2772 | " end of an object"); |
1da177e4 LT |
2773 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2774 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2775 | " start of an object"); |
1da177e4 | 2776 | } |
3b0efdfa | 2777 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2778 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2779 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2780 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2781 | #else |
2782 | if (cachep->ctor) | |
51cc5068 | 2783 | cachep->ctor(objp); |
1da177e4 | 2784 | #endif |
b28a02de | 2785 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2786 | } |
b28a02de | 2787 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2788 | } |
2789 | ||
343e0d7a | 2790 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2791 | { |
4b51d669 CL |
2792 | if (CONFIG_ZONE_DMA_FLAG) { |
2793 | if (flags & GFP_DMA) | |
a618e89f | 2794 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2795 | else |
a618e89f | 2796 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2797 | } |
1da177e4 LT |
2798 | } |
2799 | ||
a737b3e2 AM |
2800 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2801 | int nodeid) | |
78d382d7 | 2802 | { |
8fea4e96 | 2803 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2804 | kmem_bufctl_t next; |
2805 | ||
2806 | slabp->inuse++; | |
2807 | next = slab_bufctl(slabp)[slabp->free]; | |
2808 | #if DEBUG | |
2809 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2810 | WARN_ON(slabp->nodeid != nodeid); | |
2811 | #endif | |
2812 | slabp->free = next; | |
2813 | ||
2814 | return objp; | |
2815 | } | |
2816 | ||
a737b3e2 AM |
2817 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2818 | void *objp, int nodeid) | |
78d382d7 | 2819 | { |
8fea4e96 | 2820 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2821 | |
2822 | #if DEBUG | |
2823 | /* Verify that the slab belongs to the intended node */ | |
2824 | WARN_ON(slabp->nodeid != nodeid); | |
2825 | ||
871751e2 | 2826 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2827 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2828 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2829 | BUG(); |
2830 | } | |
2831 | #endif | |
2832 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2833 | slabp->free = objnr; | |
2834 | slabp->inuse--; | |
2835 | } | |
2836 | ||
4776874f PE |
2837 | /* |
2838 | * Map pages beginning at addr to the given cache and slab. This is required | |
2839 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2840 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2841 | */ |
2842 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2843 | void *addr) | |
1da177e4 | 2844 | { |
4776874f | 2845 | int nr_pages; |
1da177e4 LT |
2846 | struct page *page; |
2847 | ||
4776874f | 2848 | page = virt_to_page(addr); |
84097518 | 2849 | |
4776874f | 2850 | nr_pages = 1; |
84097518 | 2851 | if (likely(!PageCompound(page))) |
4776874f PE |
2852 | nr_pages <<= cache->gfporder; |
2853 | ||
1da177e4 | 2854 | do { |
35026088 CL |
2855 | page->slab_cache = cache; |
2856 | page->slab_page = slab; | |
1da177e4 | 2857 | page++; |
4776874f | 2858 | } while (--nr_pages); |
1da177e4 LT |
2859 | } |
2860 | ||
2861 | /* | |
2862 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2863 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2864 | */ | |
3c517a61 CL |
2865 | static int cache_grow(struct kmem_cache *cachep, |
2866 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2867 | { |
b28a02de | 2868 | struct slab *slabp; |
b28a02de PE |
2869 | size_t offset; |
2870 | gfp_t local_flags; | |
6744f087 | 2871 | struct kmem_cache_node *l3; |
1da177e4 | 2872 | |
a737b3e2 AM |
2873 | /* |
2874 | * Be lazy and only check for valid flags here, keeping it out of the | |
2875 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2876 | */ |
6cb06229 CL |
2877 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2878 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2879 | |
2e1217cf | 2880 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2881 | check_irq_off(); |
2e1217cf RT |
2882 | l3 = cachep->nodelists[nodeid]; |
2883 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2884 | |
2885 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2886 | offset = l3->colour_next; |
2887 | l3->colour_next++; | |
2888 | if (l3->colour_next >= cachep->colour) | |
2889 | l3->colour_next = 0; | |
2890 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2891 | |
2e1217cf | 2892 | offset *= cachep->colour_off; |
1da177e4 LT |
2893 | |
2894 | if (local_flags & __GFP_WAIT) | |
2895 | local_irq_enable(); | |
2896 | ||
2897 | /* | |
2898 | * The test for missing atomic flag is performed here, rather than | |
2899 | * the more obvious place, simply to reduce the critical path length | |
2900 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2901 | * will eventually be caught here (where it matters). | |
2902 | */ | |
2903 | kmem_flagcheck(cachep, flags); | |
2904 | ||
a737b3e2 AM |
2905 | /* |
2906 | * Get mem for the objs. Attempt to allocate a physical page from | |
2907 | * 'nodeid'. | |
e498be7d | 2908 | */ |
3c517a61 | 2909 | if (!objp) |
b8c1c5da | 2910 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2911 | if (!objp) |
1da177e4 LT |
2912 | goto failed; |
2913 | ||
2914 | /* Get slab management. */ | |
3c517a61 | 2915 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2916 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2917 | if (!slabp) |
1da177e4 LT |
2918 | goto opps1; |
2919 | ||
4776874f | 2920 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2921 | |
a35afb83 | 2922 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2923 | |
2924 | if (local_flags & __GFP_WAIT) | |
2925 | local_irq_disable(); | |
2926 | check_irq_off(); | |
e498be7d | 2927 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2928 | |
2929 | /* Make slab active. */ | |
e498be7d | 2930 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2931 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2932 | l3->free_objects += cachep->num; |
2933 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2934 | return 1; |
a737b3e2 | 2935 | opps1: |
1da177e4 | 2936 | kmem_freepages(cachep, objp); |
a737b3e2 | 2937 | failed: |
1da177e4 LT |
2938 | if (local_flags & __GFP_WAIT) |
2939 | local_irq_disable(); | |
2940 | return 0; | |
2941 | } | |
2942 | ||
2943 | #if DEBUG | |
2944 | ||
2945 | /* | |
2946 | * Perform extra freeing checks: | |
2947 | * - detect bad pointers. | |
2948 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2949 | */ |
2950 | static void kfree_debugcheck(const void *objp) | |
2951 | { | |
1da177e4 LT |
2952 | if (!virt_addr_valid(objp)) { |
2953 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2954 | (unsigned long)objp); |
2955 | BUG(); | |
1da177e4 | 2956 | } |
1da177e4 LT |
2957 | } |
2958 | ||
58ce1fd5 PE |
2959 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2960 | { | |
b46b8f19 | 2961 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2962 | |
2963 | redzone1 = *dbg_redzone1(cache, obj); | |
2964 | redzone2 = *dbg_redzone2(cache, obj); | |
2965 | ||
2966 | /* | |
2967 | * Redzone is ok. | |
2968 | */ | |
2969 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2970 | return; | |
2971 | ||
2972 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2973 | slab_error(cache, "double free detected"); | |
2974 | else | |
2975 | slab_error(cache, "memory outside object was overwritten"); | |
2976 | ||
b46b8f19 | 2977 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2978 | obj, redzone1, redzone2); |
2979 | } | |
2980 | ||
343e0d7a | 2981 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 2982 | unsigned long caller) |
1da177e4 LT |
2983 | { |
2984 | struct page *page; | |
2985 | unsigned int objnr; | |
2986 | struct slab *slabp; | |
2987 | ||
80cbd911 MW |
2988 | BUG_ON(virt_to_cache(objp) != cachep); |
2989 | ||
3dafccf2 | 2990 | objp -= obj_offset(cachep); |
1da177e4 | 2991 | kfree_debugcheck(objp); |
b49af68f | 2992 | page = virt_to_head_page(objp); |
1da177e4 | 2993 | |
35026088 | 2994 | slabp = page->slab_page; |
1da177e4 LT |
2995 | |
2996 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2997 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2998 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2999 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3000 | } | |
3001 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 3002 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 | 3003 | |
8fea4e96 | 3004 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3005 | |
3006 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3007 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3008 | |
871751e2 AV |
3009 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3010 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3011 | #endif | |
1da177e4 LT |
3012 | if (cachep->flags & SLAB_POISON) { |
3013 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3014 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
7c0cb9c6 | 3015 | store_stackinfo(cachep, objp, caller); |
b28a02de | 3016 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3017 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3018 | } else { |
3019 | poison_obj(cachep, objp, POISON_FREE); | |
3020 | } | |
3021 | #else | |
3022 | poison_obj(cachep, objp, POISON_FREE); | |
3023 | #endif | |
3024 | } | |
3025 | return objp; | |
3026 | } | |
3027 | ||
343e0d7a | 3028 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3029 | { |
3030 | kmem_bufctl_t i; | |
3031 | int entries = 0; | |
b28a02de | 3032 | |
1da177e4 LT |
3033 | /* Check slab's freelist to see if this obj is there. */ |
3034 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3035 | entries++; | |
3036 | if (entries > cachep->num || i >= cachep->num) | |
3037 | goto bad; | |
3038 | } | |
3039 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3040 | bad: |
3041 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3042 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3043 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3044 | print_tainted()); | |
fdde6abb SAS |
3045 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3046 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3047 | 1); | |
1da177e4 LT |
3048 | BUG(); |
3049 | } | |
3050 | } | |
3051 | #else | |
3052 | #define kfree_debugcheck(x) do { } while(0) | |
3053 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3054 | #define check_slabp(x,y) do { } while(0) | |
3055 | #endif | |
3056 | ||
072bb0aa MG |
3057 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
3058 | bool force_refill) | |
1da177e4 LT |
3059 | { |
3060 | int batchcount; | |
6744f087 | 3061 | struct kmem_cache_node *l3; |
1da177e4 | 3062 | struct array_cache *ac; |
1ca4cb24 PE |
3063 | int node; |
3064 | ||
1da177e4 | 3065 | check_irq_off(); |
7d6e6d09 | 3066 | node = numa_mem_id(); |
072bb0aa MG |
3067 | if (unlikely(force_refill)) |
3068 | goto force_grow; | |
3069 | retry: | |
9a2dba4b | 3070 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3071 | batchcount = ac->batchcount; |
3072 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3073 | /* |
3074 | * If there was little recent activity on this cache, then | |
3075 | * perform only a partial refill. Otherwise we could generate | |
3076 | * refill bouncing. | |
1da177e4 LT |
3077 | */ |
3078 | batchcount = BATCHREFILL_LIMIT; | |
3079 | } | |
1ca4cb24 | 3080 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3081 | |
3082 | BUG_ON(ac->avail > 0 || !l3); | |
3083 | spin_lock(&l3->list_lock); | |
1da177e4 | 3084 | |
3ded175a | 3085 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3086 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3087 | l3->shared->touched = 1; | |
3ded175a | 3088 | goto alloc_done; |
44b57f1c | 3089 | } |
3ded175a | 3090 | |
1da177e4 LT |
3091 | while (batchcount > 0) { |
3092 | struct list_head *entry; | |
3093 | struct slab *slabp; | |
3094 | /* Get slab alloc is to come from. */ | |
3095 | entry = l3->slabs_partial.next; | |
3096 | if (entry == &l3->slabs_partial) { | |
3097 | l3->free_touched = 1; | |
3098 | entry = l3->slabs_free.next; | |
3099 | if (entry == &l3->slabs_free) | |
3100 | goto must_grow; | |
3101 | } | |
3102 | ||
3103 | slabp = list_entry(entry, struct slab, list); | |
3104 | check_slabp(cachep, slabp); | |
3105 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3106 | |
3107 | /* | |
3108 | * The slab was either on partial or free list so | |
3109 | * there must be at least one object available for | |
3110 | * allocation. | |
3111 | */ | |
249b9f33 | 3112 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3113 | |
1da177e4 | 3114 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3115 | STATS_INC_ALLOCED(cachep); |
3116 | STATS_INC_ACTIVE(cachep); | |
3117 | STATS_SET_HIGH(cachep); | |
3118 | ||
072bb0aa MG |
3119 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, |
3120 | node)); | |
1da177e4 LT |
3121 | } |
3122 | check_slabp(cachep, slabp); | |
3123 | ||
3124 | /* move slabp to correct slabp list: */ | |
3125 | list_del(&slabp->list); | |
3126 | if (slabp->free == BUFCTL_END) | |
3127 | list_add(&slabp->list, &l3->slabs_full); | |
3128 | else | |
3129 | list_add(&slabp->list, &l3->slabs_partial); | |
3130 | } | |
3131 | ||
a737b3e2 | 3132 | must_grow: |
1da177e4 | 3133 | l3->free_objects -= ac->avail; |
a737b3e2 | 3134 | alloc_done: |
e498be7d | 3135 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3136 | |
3137 | if (unlikely(!ac->avail)) { | |
3138 | int x; | |
072bb0aa | 3139 | force_grow: |
3c517a61 | 3140 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3141 | |
a737b3e2 | 3142 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3143 | ac = cpu_cache_get(cachep); |
51cd8e6f | 3144 | node = numa_mem_id(); |
072bb0aa MG |
3145 | |
3146 | /* no objects in sight? abort */ | |
3147 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
3148 | return NULL; |
3149 | ||
a737b3e2 | 3150 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3151 | goto retry; |
3152 | } | |
3153 | ac->touched = 1; | |
072bb0aa MG |
3154 | |
3155 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
3156 | } |
3157 | ||
a737b3e2 AM |
3158 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3159 | gfp_t flags) | |
1da177e4 LT |
3160 | { |
3161 | might_sleep_if(flags & __GFP_WAIT); | |
3162 | #if DEBUG | |
3163 | kmem_flagcheck(cachep, flags); | |
3164 | #endif | |
3165 | } | |
3166 | ||
3167 | #if DEBUG | |
a737b3e2 | 3168 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
7c0cb9c6 | 3169 | gfp_t flags, void *objp, unsigned long caller) |
1da177e4 | 3170 | { |
b28a02de | 3171 | if (!objp) |
1da177e4 | 3172 | return objp; |
b28a02de | 3173 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3174 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3175 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3176 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3177 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3178 | else |
3179 | check_poison_obj(cachep, objp); | |
3180 | #else | |
3181 | check_poison_obj(cachep, objp); | |
3182 | #endif | |
3183 | poison_obj(cachep, objp, POISON_INUSE); | |
3184 | } | |
3185 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 3186 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 LT |
3187 | |
3188 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3189 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3190 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3191 | slab_error(cachep, "double free, or memory outside" | |
3192 | " object was overwritten"); | |
b28a02de | 3193 | printk(KERN_ERR |
b46b8f19 | 3194 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3195 | objp, *dbg_redzone1(cachep, objp), |
3196 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3197 | } |
3198 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3199 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3200 | } | |
871751e2 AV |
3201 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3202 | { | |
3203 | struct slab *slabp; | |
3204 | unsigned objnr; | |
3205 | ||
35026088 | 3206 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3207 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3208 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3209 | } | |
3210 | #endif | |
3dafccf2 | 3211 | objp += obj_offset(cachep); |
4f104934 | 3212 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3213 | cachep->ctor(objp); |
7ea466f2 TH |
3214 | if (ARCH_SLAB_MINALIGN && |
3215 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3216 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3217 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3218 | } |
1da177e4 LT |
3219 | return objp; |
3220 | } | |
3221 | #else | |
3222 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3223 | #endif | |
3224 | ||
773ff60e | 3225 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 | 3226 | { |
9b030cb8 | 3227 | if (cachep == kmem_cache) |
773ff60e | 3228 | return false; |
8a8b6502 | 3229 | |
8c138bc0 | 3230 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3231 | } |
3232 | ||
343e0d7a | 3233 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3234 | { |
b28a02de | 3235 | void *objp; |
1da177e4 | 3236 | struct array_cache *ac; |
072bb0aa | 3237 | bool force_refill = false; |
1da177e4 | 3238 | |
5c382300 | 3239 | check_irq_off(); |
8a8b6502 | 3240 | |
9a2dba4b | 3241 | ac = cpu_cache_get(cachep); |
1da177e4 | 3242 | if (likely(ac->avail)) { |
1da177e4 | 3243 | ac->touched = 1; |
072bb0aa MG |
3244 | objp = ac_get_obj(cachep, ac, flags, false); |
3245 | ||
ddbf2e83 | 3246 | /* |
072bb0aa MG |
3247 | * Allow for the possibility all avail objects are not allowed |
3248 | * by the current flags | |
ddbf2e83 | 3249 | */ |
072bb0aa MG |
3250 | if (objp) { |
3251 | STATS_INC_ALLOCHIT(cachep); | |
3252 | goto out; | |
3253 | } | |
3254 | force_refill = true; | |
1da177e4 | 3255 | } |
072bb0aa MG |
3256 | |
3257 | STATS_INC_ALLOCMISS(cachep); | |
3258 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3259 | /* | |
3260 | * the 'ac' may be updated by cache_alloc_refill(), | |
3261 | * and kmemleak_erase() requires its correct value. | |
3262 | */ | |
3263 | ac = cpu_cache_get(cachep); | |
3264 | ||
3265 | out: | |
d5cff635 CM |
3266 | /* |
3267 | * To avoid a false negative, if an object that is in one of the | |
3268 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3269 | * treat the array pointers as a reference to the object. | |
3270 | */ | |
f3d8b53a O |
3271 | if (objp) |
3272 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3273 | return objp; |
3274 | } | |
3275 | ||
e498be7d | 3276 | #ifdef CONFIG_NUMA |
c61afb18 | 3277 | /* |
b2455396 | 3278 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3279 | * |
3280 | * If we are in_interrupt, then process context, including cpusets and | |
3281 | * mempolicy, may not apply and should not be used for allocation policy. | |
3282 | */ | |
3283 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3284 | { | |
3285 | int nid_alloc, nid_here; | |
3286 | ||
765c4507 | 3287 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3288 | return NULL; |
7d6e6d09 | 3289 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3290 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3291 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3292 | else if (current->mempolicy) |
e7b691b0 | 3293 | nid_alloc = slab_node(); |
c61afb18 | 3294 | if (nid_alloc != nid_here) |
8b98c169 | 3295 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3296 | return NULL; |
3297 | } | |
3298 | ||
765c4507 CL |
3299 | /* |
3300 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3301 | * certain node and fall back is permitted. First we scan all the |
3302 | * available nodelists for available objects. If that fails then we | |
3303 | * perform an allocation without specifying a node. This allows the page | |
3304 | * allocator to do its reclaim / fallback magic. We then insert the | |
3305 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3306 | */ |
8c8cc2c1 | 3307 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3308 | { |
8c8cc2c1 PE |
3309 | struct zonelist *zonelist; |
3310 | gfp_t local_flags; | |
dd1a239f | 3311 | struct zoneref *z; |
54a6eb5c MG |
3312 | struct zone *zone; |
3313 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3314 | void *obj = NULL; |
3c517a61 | 3315 | int nid; |
cc9a6c87 | 3316 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3317 | |
3318 | if (flags & __GFP_THISNODE) | |
3319 | return NULL; | |
3320 | ||
6cb06229 | 3321 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3322 | |
cc9a6c87 MG |
3323 | retry_cpuset: |
3324 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3325 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3326 | |
3c517a61 CL |
3327 | retry: |
3328 | /* | |
3329 | * Look through allowed nodes for objects available | |
3330 | * from existing per node queues. | |
3331 | */ | |
54a6eb5c MG |
3332 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3333 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3334 | |
54a6eb5c | 3335 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3336 | cache->nodelists[nid] && |
481c5346 | 3337 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3338 | obj = ____cache_alloc_node(cache, |
3339 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3340 | if (obj) |
3341 | break; | |
3342 | } | |
3c517a61 CL |
3343 | } |
3344 | ||
cfce6604 | 3345 | if (!obj) { |
3c517a61 CL |
3346 | /* |
3347 | * This allocation will be performed within the constraints | |
3348 | * of the current cpuset / memory policy requirements. | |
3349 | * We may trigger various forms of reclaim on the allowed | |
3350 | * set and go into memory reserves if necessary. | |
3351 | */ | |
dd47ea75 CL |
3352 | if (local_flags & __GFP_WAIT) |
3353 | local_irq_enable(); | |
3354 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3355 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3356 | if (local_flags & __GFP_WAIT) |
3357 | local_irq_disable(); | |
3c517a61 CL |
3358 | if (obj) { |
3359 | /* | |
3360 | * Insert into the appropriate per node queues | |
3361 | */ | |
3362 | nid = page_to_nid(virt_to_page(obj)); | |
3363 | if (cache_grow(cache, flags, nid, obj)) { | |
3364 | obj = ____cache_alloc_node(cache, | |
3365 | flags | GFP_THISNODE, nid); | |
3366 | if (!obj) | |
3367 | /* | |
3368 | * Another processor may allocate the | |
3369 | * objects in the slab since we are | |
3370 | * not holding any locks. | |
3371 | */ | |
3372 | goto retry; | |
3373 | } else { | |
b6a60451 | 3374 | /* cache_grow already freed obj */ |
3c517a61 CL |
3375 | obj = NULL; |
3376 | } | |
3377 | } | |
aedb0eb1 | 3378 | } |
cc9a6c87 MG |
3379 | |
3380 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3381 | goto retry_cpuset; | |
765c4507 CL |
3382 | return obj; |
3383 | } | |
3384 | ||
e498be7d CL |
3385 | /* |
3386 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3387 | */ |
8b98c169 | 3388 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3389 | int nodeid) |
e498be7d CL |
3390 | { |
3391 | struct list_head *entry; | |
b28a02de | 3392 | struct slab *slabp; |
6744f087 | 3393 | struct kmem_cache_node *l3; |
b28a02de | 3394 | void *obj; |
b28a02de PE |
3395 | int x; |
3396 | ||
3397 | l3 = cachep->nodelists[nodeid]; | |
3398 | BUG_ON(!l3); | |
3399 | ||
a737b3e2 | 3400 | retry: |
ca3b9b91 | 3401 | check_irq_off(); |
b28a02de PE |
3402 | spin_lock(&l3->list_lock); |
3403 | entry = l3->slabs_partial.next; | |
3404 | if (entry == &l3->slabs_partial) { | |
3405 | l3->free_touched = 1; | |
3406 | entry = l3->slabs_free.next; | |
3407 | if (entry == &l3->slabs_free) | |
3408 | goto must_grow; | |
3409 | } | |
3410 | ||
3411 | slabp = list_entry(entry, struct slab, list); | |
3412 | check_spinlock_acquired_node(cachep, nodeid); | |
3413 | check_slabp(cachep, slabp); | |
3414 | ||
3415 | STATS_INC_NODEALLOCS(cachep); | |
3416 | STATS_INC_ACTIVE(cachep); | |
3417 | STATS_SET_HIGH(cachep); | |
3418 | ||
3419 | BUG_ON(slabp->inuse == cachep->num); | |
3420 | ||
78d382d7 | 3421 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3422 | check_slabp(cachep, slabp); |
3423 | l3->free_objects--; | |
3424 | /* move slabp to correct slabp list: */ | |
3425 | list_del(&slabp->list); | |
3426 | ||
a737b3e2 | 3427 | if (slabp->free == BUFCTL_END) |
b28a02de | 3428 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3429 | else |
b28a02de | 3430 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3431 | |
b28a02de PE |
3432 | spin_unlock(&l3->list_lock); |
3433 | goto done; | |
e498be7d | 3434 | |
a737b3e2 | 3435 | must_grow: |
b28a02de | 3436 | spin_unlock(&l3->list_lock); |
3c517a61 | 3437 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3438 | if (x) |
3439 | goto retry; | |
1da177e4 | 3440 | |
8c8cc2c1 | 3441 | return fallback_alloc(cachep, flags); |
e498be7d | 3442 | |
a737b3e2 | 3443 | done: |
b28a02de | 3444 | return obj; |
e498be7d | 3445 | } |
8c8cc2c1 PE |
3446 | |
3447 | /** | |
3448 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3449 | * @cachep: The cache to allocate from. | |
3450 | * @flags: See kmalloc(). | |
3451 | * @nodeid: node number of the target node. | |
3452 | * @caller: return address of caller, used for debug information | |
3453 | * | |
3454 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3455 | * node, which can improve the performance for cpu bound structures. | |
3456 | * | |
3457 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3458 | */ | |
3459 | static __always_inline void * | |
48356303 | 3460 | slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, |
7c0cb9c6 | 3461 | unsigned long caller) |
8c8cc2c1 PE |
3462 | { |
3463 | unsigned long save_flags; | |
3464 | void *ptr; | |
7d6e6d09 | 3465 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3466 | |
dcce284a | 3467 | flags &= gfp_allowed_mask; |
7e85ee0c | 3468 | |
cf40bd16 NP |
3469 | lockdep_trace_alloc(flags); |
3470 | ||
773ff60e | 3471 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3472 | return NULL; |
3473 | ||
d79923fa GC |
3474 | cachep = memcg_kmem_get_cache(cachep, flags); |
3475 | ||
8c8cc2c1 PE |
3476 | cache_alloc_debugcheck_before(cachep, flags); |
3477 | local_irq_save(save_flags); | |
3478 | ||
eacbbae3 | 3479 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3480 | nodeid = slab_node; |
8c8cc2c1 PE |
3481 | |
3482 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3483 | /* Node not bootstrapped yet */ | |
3484 | ptr = fallback_alloc(cachep, flags); | |
3485 | goto out; | |
3486 | } | |
3487 | ||
7d6e6d09 | 3488 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3489 | /* |
3490 | * Use the locally cached objects if possible. | |
3491 | * However ____cache_alloc does not allow fallback | |
3492 | * to other nodes. It may fail while we still have | |
3493 | * objects on other nodes available. | |
3494 | */ | |
3495 | ptr = ____cache_alloc(cachep, flags); | |
3496 | if (ptr) | |
3497 | goto out; | |
3498 | } | |
3499 | /* ___cache_alloc_node can fall back to other nodes */ | |
3500 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3501 | out: | |
3502 | local_irq_restore(save_flags); | |
3503 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3504 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3505 | flags); |
8c8cc2c1 | 3506 | |
c175eea4 | 3507 | if (likely(ptr)) |
8c138bc0 | 3508 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3509 | |
d07dbea4 | 3510 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3511 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3512 | |
8c8cc2c1 PE |
3513 | return ptr; |
3514 | } | |
3515 | ||
3516 | static __always_inline void * | |
3517 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3518 | { | |
3519 | void *objp; | |
3520 | ||
3521 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3522 | objp = alternate_node_alloc(cache, flags); | |
3523 | if (objp) | |
3524 | goto out; | |
3525 | } | |
3526 | objp = ____cache_alloc(cache, flags); | |
3527 | ||
3528 | /* | |
3529 | * We may just have run out of memory on the local node. | |
3530 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3531 | */ | |
7d6e6d09 LS |
3532 | if (!objp) |
3533 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3534 | |
3535 | out: | |
3536 | return objp; | |
3537 | } | |
3538 | #else | |
3539 | ||
3540 | static __always_inline void * | |
3541 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3542 | { | |
3543 | return ____cache_alloc(cachep, flags); | |
3544 | } | |
3545 | ||
3546 | #endif /* CONFIG_NUMA */ | |
3547 | ||
3548 | static __always_inline void * | |
48356303 | 3549 | slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) |
8c8cc2c1 PE |
3550 | { |
3551 | unsigned long save_flags; | |
3552 | void *objp; | |
3553 | ||
dcce284a | 3554 | flags &= gfp_allowed_mask; |
7e85ee0c | 3555 | |
cf40bd16 NP |
3556 | lockdep_trace_alloc(flags); |
3557 | ||
773ff60e | 3558 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3559 | return NULL; |
3560 | ||
d79923fa GC |
3561 | cachep = memcg_kmem_get_cache(cachep, flags); |
3562 | ||
8c8cc2c1 PE |
3563 | cache_alloc_debugcheck_before(cachep, flags); |
3564 | local_irq_save(save_flags); | |
3565 | objp = __do_cache_alloc(cachep, flags); | |
3566 | local_irq_restore(save_flags); | |
3567 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3568 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3569 | flags); |
8c8cc2c1 PE |
3570 | prefetchw(objp); |
3571 | ||
c175eea4 | 3572 | if (likely(objp)) |
8c138bc0 | 3573 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3574 | |
d07dbea4 | 3575 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3576 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3577 | |
8c8cc2c1 PE |
3578 | return objp; |
3579 | } | |
e498be7d CL |
3580 | |
3581 | /* | |
3582 | * Caller needs to acquire correct kmem_list's list_lock | |
3583 | */ | |
343e0d7a | 3584 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3585 | int node) |
1da177e4 LT |
3586 | { |
3587 | int i; | |
6744f087 | 3588 | struct kmem_cache_node *l3; |
1da177e4 LT |
3589 | |
3590 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3591 | void *objp; |
1da177e4 | 3592 | struct slab *slabp; |
1da177e4 | 3593 | |
072bb0aa MG |
3594 | clear_obj_pfmemalloc(&objpp[i]); |
3595 | objp = objpp[i]; | |
3596 | ||
6ed5eb22 | 3597 | slabp = virt_to_slab(objp); |
ff69416e | 3598 | l3 = cachep->nodelists[node]; |
1da177e4 | 3599 | list_del(&slabp->list); |
ff69416e | 3600 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3601 | check_slabp(cachep, slabp); |
78d382d7 | 3602 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3603 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3604 | l3->free_objects++; |
1da177e4 LT |
3605 | check_slabp(cachep, slabp); |
3606 | ||
3607 | /* fixup slab chains */ | |
3608 | if (slabp->inuse == 0) { | |
e498be7d CL |
3609 | if (l3->free_objects > l3->free_limit) { |
3610 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3611 | /* No need to drop any previously held |
3612 | * lock here, even if we have a off-slab slab | |
3613 | * descriptor it is guaranteed to come from | |
3614 | * a different cache, refer to comments before | |
3615 | * alloc_slabmgmt. | |
3616 | */ | |
1da177e4 LT |
3617 | slab_destroy(cachep, slabp); |
3618 | } else { | |
e498be7d | 3619 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3620 | } |
3621 | } else { | |
3622 | /* Unconditionally move a slab to the end of the | |
3623 | * partial list on free - maximum time for the | |
3624 | * other objects to be freed, too. | |
3625 | */ | |
e498be7d | 3626 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3627 | } |
3628 | } | |
3629 | } | |
3630 | ||
343e0d7a | 3631 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3632 | { |
3633 | int batchcount; | |
6744f087 | 3634 | struct kmem_cache_node *l3; |
7d6e6d09 | 3635 | int node = numa_mem_id(); |
1da177e4 LT |
3636 | |
3637 | batchcount = ac->batchcount; | |
3638 | #if DEBUG | |
3639 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3640 | #endif | |
3641 | check_irq_off(); | |
ff69416e | 3642 | l3 = cachep->nodelists[node]; |
873623df | 3643 | spin_lock(&l3->list_lock); |
e498be7d CL |
3644 | if (l3->shared) { |
3645 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3646 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3647 | if (max) { |
3648 | if (batchcount > max) | |
3649 | batchcount = max; | |
e498be7d | 3650 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3651 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3652 | shared_array->avail += batchcount; |
3653 | goto free_done; | |
3654 | } | |
3655 | } | |
3656 | ||
ff69416e | 3657 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3658 | free_done: |
1da177e4 LT |
3659 | #if STATS |
3660 | { | |
3661 | int i = 0; | |
3662 | struct list_head *p; | |
3663 | ||
e498be7d CL |
3664 | p = l3->slabs_free.next; |
3665 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3666 | struct slab *slabp; |
3667 | ||
3668 | slabp = list_entry(p, struct slab, list); | |
3669 | BUG_ON(slabp->inuse); | |
3670 | ||
3671 | i++; | |
3672 | p = p->next; | |
3673 | } | |
3674 | STATS_SET_FREEABLE(cachep, i); | |
3675 | } | |
3676 | #endif | |
e498be7d | 3677 | spin_unlock(&l3->list_lock); |
1da177e4 | 3678 | ac->avail -= batchcount; |
a737b3e2 | 3679 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3680 | } |
3681 | ||
3682 | /* | |
a737b3e2 AM |
3683 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3684 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3685 | */ |
a947eb95 | 3686 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 3687 | unsigned long caller) |
1da177e4 | 3688 | { |
9a2dba4b | 3689 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3690 | |
3691 | check_irq_off(); | |
d5cff635 | 3692 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3693 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3694 | |
8c138bc0 | 3695 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3696 | |
1807a1aa SS |
3697 | /* |
3698 | * Skip calling cache_free_alien() when the platform is not numa. | |
3699 | * This will avoid cache misses that happen while accessing slabp (which | |
3700 | * is per page memory reference) to get nodeid. Instead use a global | |
3701 | * variable to skip the call, which is mostly likely to be present in | |
3702 | * the cache. | |
3703 | */ | |
b6e68bc1 | 3704 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3705 | return; |
3706 | ||
1da177e4 LT |
3707 | if (likely(ac->avail < ac->limit)) { |
3708 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3709 | } else { |
3710 | STATS_INC_FREEMISS(cachep); | |
3711 | cache_flusharray(cachep, ac); | |
1da177e4 | 3712 | } |
42c8c99c | 3713 | |
072bb0aa | 3714 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3715 | } |
3716 | ||
3717 | /** | |
3718 | * kmem_cache_alloc - Allocate an object | |
3719 | * @cachep: The cache to allocate from. | |
3720 | * @flags: See kmalloc(). | |
3721 | * | |
3722 | * Allocate an object from this cache. The flags are only relevant | |
3723 | * if the cache has no available objects. | |
3724 | */ | |
343e0d7a | 3725 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3726 | { |
48356303 | 3727 | void *ret = slab_alloc(cachep, flags, _RET_IP_); |
36555751 | 3728 | |
ca2b84cb | 3729 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3730 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3731 | |
3732 | return ret; | |
1da177e4 LT |
3733 | } |
3734 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3735 | ||
0f24f128 | 3736 | #ifdef CONFIG_TRACING |
85beb586 | 3737 | void * |
4052147c | 3738 | kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) |
36555751 | 3739 | { |
85beb586 SR |
3740 | void *ret; |
3741 | ||
48356303 | 3742 | ret = slab_alloc(cachep, flags, _RET_IP_); |
85beb586 SR |
3743 | |
3744 | trace_kmalloc(_RET_IP_, ret, | |
ff4fcd01 | 3745 | size, cachep->size, flags); |
85beb586 | 3746 | return ret; |
36555751 | 3747 | } |
85beb586 | 3748 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3749 | #endif |
3750 | ||
1da177e4 | 3751 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3752 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3753 | { | |
48356303 | 3754 | void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
36555751 | 3755 | |
ca2b84cb | 3756 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3757 | cachep->object_size, cachep->size, |
ca2b84cb | 3758 | flags, nodeid); |
36555751 EGM |
3759 | |
3760 | return ret; | |
8b98c169 | 3761 | } |
1da177e4 LT |
3762 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3763 | ||
0f24f128 | 3764 | #ifdef CONFIG_TRACING |
4052147c | 3765 | void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, |
85beb586 | 3766 | gfp_t flags, |
4052147c EG |
3767 | int nodeid, |
3768 | size_t size) | |
36555751 | 3769 | { |
85beb586 SR |
3770 | void *ret; |
3771 | ||
592f4145 | 3772 | ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
7c0cb9c6 | 3773 | |
85beb586 | 3774 | trace_kmalloc_node(_RET_IP_, ret, |
ff4fcd01 | 3775 | size, cachep->size, |
85beb586 SR |
3776 | flags, nodeid); |
3777 | return ret; | |
36555751 | 3778 | } |
85beb586 | 3779 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3780 | #endif |
3781 | ||
8b98c169 | 3782 | static __always_inline void * |
7c0cb9c6 | 3783 | __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) |
97e2bde4 | 3784 | { |
343e0d7a | 3785 | struct kmem_cache *cachep; |
97e2bde4 MS |
3786 | |
3787 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3788 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3789 | return cachep; | |
4052147c | 3790 | return kmem_cache_alloc_node_trace(cachep, flags, node, size); |
97e2bde4 | 3791 | } |
8b98c169 | 3792 | |
0bb38a5c | 3793 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3794 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3795 | { | |
7c0cb9c6 | 3796 | return __do_kmalloc_node(size, flags, node, _RET_IP_); |
8b98c169 | 3797 | } |
dbe5e69d | 3798 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3799 | |
3800 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3801 | int node, unsigned long caller) |
8b98c169 | 3802 | { |
7c0cb9c6 | 3803 | return __do_kmalloc_node(size, flags, node, caller); |
8b98c169 CH |
3804 | } |
3805 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3806 | #else | |
3807 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3808 | { | |
7c0cb9c6 | 3809 | return __do_kmalloc_node(size, flags, node, 0); |
8b98c169 CH |
3810 | } |
3811 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3812 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3813 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3814 | |
3815 | /** | |
800590f5 | 3816 | * __do_kmalloc - allocate memory |
1da177e4 | 3817 | * @size: how many bytes of memory are required. |
800590f5 | 3818 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3819 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3820 | */ |
7fd6b141 | 3821 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
7c0cb9c6 | 3822 | unsigned long caller) |
1da177e4 | 3823 | { |
343e0d7a | 3824 | struct kmem_cache *cachep; |
36555751 | 3825 | void *ret; |
1da177e4 | 3826 | |
97e2bde4 MS |
3827 | /* If you want to save a few bytes .text space: replace |
3828 | * __ with kmem_. | |
3829 | * Then kmalloc uses the uninlined functions instead of the inline | |
3830 | * functions. | |
3831 | */ | |
3832 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3833 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3834 | return cachep; | |
48356303 | 3835 | ret = slab_alloc(cachep, flags, caller); |
36555751 | 3836 | |
7c0cb9c6 | 3837 | trace_kmalloc(caller, ret, |
3b0efdfa | 3838 | size, cachep->size, flags); |
36555751 EGM |
3839 | |
3840 | return ret; | |
7fd6b141 PE |
3841 | } |
3842 | ||
7fd6b141 | 3843 | |
0bb38a5c | 3844 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3845 | void *__kmalloc(size_t size, gfp_t flags) |
3846 | { | |
7c0cb9c6 | 3847 | return __do_kmalloc(size, flags, _RET_IP_); |
1da177e4 LT |
3848 | } |
3849 | EXPORT_SYMBOL(__kmalloc); | |
3850 | ||
ce71e27c | 3851 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3852 | { |
7c0cb9c6 | 3853 | return __do_kmalloc(size, flags, caller); |
7fd6b141 PE |
3854 | } |
3855 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3856 | |
3857 | #else | |
3858 | void *__kmalloc(size_t size, gfp_t flags) | |
3859 | { | |
7c0cb9c6 | 3860 | return __do_kmalloc(size, flags, 0); |
1d2c8eea CH |
3861 | } |
3862 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3863 | #endif |
3864 | ||
1da177e4 LT |
3865 | /** |
3866 | * kmem_cache_free - Deallocate an object | |
3867 | * @cachep: The cache the allocation was from. | |
3868 | * @objp: The previously allocated object. | |
3869 | * | |
3870 | * Free an object which was previously allocated from this | |
3871 | * cache. | |
3872 | */ | |
343e0d7a | 3873 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3874 | { |
3875 | unsigned long flags; | |
b9ce5ef4 GC |
3876 | cachep = cache_from_obj(cachep, objp); |
3877 | if (!cachep) | |
3878 | return; | |
1da177e4 LT |
3879 | |
3880 | local_irq_save(flags); | |
d97d476b | 3881 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3882 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3883 | debug_check_no_obj_freed(objp, cachep->object_size); |
7c0cb9c6 | 3884 | __cache_free(cachep, objp, _RET_IP_); |
1da177e4 | 3885 | local_irq_restore(flags); |
36555751 | 3886 | |
ca2b84cb | 3887 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3888 | } |
3889 | EXPORT_SYMBOL(kmem_cache_free); | |
3890 | ||
1da177e4 LT |
3891 | /** |
3892 | * kfree - free previously allocated memory | |
3893 | * @objp: pointer returned by kmalloc. | |
3894 | * | |
80e93eff PE |
3895 | * If @objp is NULL, no operation is performed. |
3896 | * | |
1da177e4 LT |
3897 | * Don't free memory not originally allocated by kmalloc() |
3898 | * or you will run into trouble. | |
3899 | */ | |
3900 | void kfree(const void *objp) | |
3901 | { | |
343e0d7a | 3902 | struct kmem_cache *c; |
1da177e4 LT |
3903 | unsigned long flags; |
3904 | ||
2121db74 PE |
3905 | trace_kfree(_RET_IP_, objp); |
3906 | ||
6cb8f913 | 3907 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3908 | return; |
3909 | local_irq_save(flags); | |
3910 | kfree_debugcheck(objp); | |
6ed5eb22 | 3911 | c = virt_to_cache(objp); |
8c138bc0 CL |
3912 | debug_check_no_locks_freed(objp, c->object_size); |
3913 | ||
3914 | debug_check_no_obj_freed(objp, c->object_size); | |
7c0cb9c6 | 3915 | __cache_free(c, (void *)objp, _RET_IP_); |
1da177e4 LT |
3916 | local_irq_restore(flags); |
3917 | } | |
3918 | EXPORT_SYMBOL(kfree); | |
3919 | ||
e498be7d | 3920 | /* |
183ff22b | 3921 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3922 | */ |
83b519e8 | 3923 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3924 | { |
3925 | int node; | |
6744f087 | 3926 | struct kmem_cache_node *l3; |
cafeb02e | 3927 | struct array_cache *new_shared; |
3395ee05 | 3928 | struct array_cache **new_alien = NULL; |
e498be7d | 3929 | |
9c09a95c | 3930 | for_each_online_node(node) { |
cafeb02e | 3931 | |
3395ee05 | 3932 | if (use_alien_caches) { |
83b519e8 | 3933 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3934 | if (!new_alien) |
3935 | goto fail; | |
3936 | } | |
cafeb02e | 3937 | |
63109846 ED |
3938 | new_shared = NULL; |
3939 | if (cachep->shared) { | |
3940 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3941 | cachep->shared*cachep->batchcount, |
83b519e8 | 3942 | 0xbaadf00d, gfp); |
63109846 ED |
3943 | if (!new_shared) { |
3944 | free_alien_cache(new_alien); | |
3945 | goto fail; | |
3946 | } | |
0718dc2a | 3947 | } |
cafeb02e | 3948 | |
a737b3e2 AM |
3949 | l3 = cachep->nodelists[node]; |
3950 | if (l3) { | |
cafeb02e CL |
3951 | struct array_cache *shared = l3->shared; |
3952 | ||
e498be7d CL |
3953 | spin_lock_irq(&l3->list_lock); |
3954 | ||
cafeb02e | 3955 | if (shared) |
0718dc2a CL |
3956 | free_block(cachep, shared->entry, |
3957 | shared->avail, node); | |
e498be7d | 3958 | |
cafeb02e CL |
3959 | l3->shared = new_shared; |
3960 | if (!l3->alien) { | |
e498be7d CL |
3961 | l3->alien = new_alien; |
3962 | new_alien = NULL; | |
3963 | } | |
b28a02de | 3964 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3965 | cachep->batchcount + cachep->num; |
e498be7d | 3966 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3967 | kfree(shared); |
e498be7d CL |
3968 | free_alien_cache(new_alien); |
3969 | continue; | |
3970 | } | |
6744f087 | 3971 | l3 = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node); |
0718dc2a CL |
3972 | if (!l3) { |
3973 | free_alien_cache(new_alien); | |
3974 | kfree(new_shared); | |
e498be7d | 3975 | goto fail; |
0718dc2a | 3976 | } |
e498be7d CL |
3977 | |
3978 | kmem_list3_init(l3); | |
3979 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3980 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3981 | l3->shared = new_shared; |
e498be7d | 3982 | l3->alien = new_alien; |
b28a02de | 3983 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3984 | cachep->batchcount + cachep->num; |
e498be7d CL |
3985 | cachep->nodelists[node] = l3; |
3986 | } | |
cafeb02e | 3987 | return 0; |
0718dc2a | 3988 | |
a737b3e2 | 3989 | fail: |
3b0efdfa | 3990 | if (!cachep->list.next) { |
0718dc2a CL |
3991 | /* Cache is not active yet. Roll back what we did */ |
3992 | node--; | |
3993 | while (node >= 0) { | |
3994 | if (cachep->nodelists[node]) { | |
3995 | l3 = cachep->nodelists[node]; | |
3996 | ||
3997 | kfree(l3->shared); | |
3998 | free_alien_cache(l3->alien); | |
3999 | kfree(l3); | |
4000 | cachep->nodelists[node] = NULL; | |
4001 | } | |
4002 | node--; | |
4003 | } | |
4004 | } | |
cafeb02e | 4005 | return -ENOMEM; |
e498be7d CL |
4006 | } |
4007 | ||
1da177e4 | 4008 | struct ccupdate_struct { |
343e0d7a | 4009 | struct kmem_cache *cachep; |
acfe7d74 | 4010 | struct array_cache *new[0]; |
1da177e4 LT |
4011 | }; |
4012 | ||
4013 | static void do_ccupdate_local(void *info) | |
4014 | { | |
a737b3e2 | 4015 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4016 | struct array_cache *old; |
4017 | ||
4018 | check_irq_off(); | |
9a2dba4b | 4019 | old = cpu_cache_get(new->cachep); |
e498be7d | 4020 | |
1da177e4 LT |
4021 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4022 | new->new[smp_processor_id()] = old; | |
4023 | } | |
4024 | ||
18004c5d | 4025 | /* Always called with the slab_mutex held */ |
943a451a | 4026 | static int __do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4027 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4028 | { |
d2e7b7d0 | 4029 | struct ccupdate_struct *new; |
2ed3a4ef | 4030 | int i; |
1da177e4 | 4031 | |
acfe7d74 ED |
4032 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4033 | gfp); | |
d2e7b7d0 SS |
4034 | if (!new) |
4035 | return -ENOMEM; | |
4036 | ||
e498be7d | 4037 | for_each_online_cpu(i) { |
7d6e6d09 | 4038 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4039 | batchcount, gfp); |
d2e7b7d0 | 4040 | if (!new->new[i]) { |
b28a02de | 4041 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4042 | kfree(new->new[i]); |
4043 | kfree(new); | |
e498be7d | 4044 | return -ENOMEM; |
1da177e4 LT |
4045 | } |
4046 | } | |
d2e7b7d0 | 4047 | new->cachep = cachep; |
1da177e4 | 4048 | |
15c8b6c1 | 4049 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4050 | |
1da177e4 | 4051 | check_irq_on(); |
1da177e4 LT |
4052 | cachep->batchcount = batchcount; |
4053 | cachep->limit = limit; | |
e498be7d | 4054 | cachep->shared = shared; |
1da177e4 | 4055 | |
e498be7d | 4056 | for_each_online_cpu(i) { |
d2e7b7d0 | 4057 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4058 | if (!ccold) |
4059 | continue; | |
7d6e6d09 LS |
4060 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4061 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4062 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4063 | kfree(ccold); |
4064 | } | |
d2e7b7d0 | 4065 | kfree(new); |
83b519e8 | 4066 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4067 | } |
4068 | ||
943a451a GC |
4069 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
4070 | int batchcount, int shared, gfp_t gfp) | |
4071 | { | |
4072 | int ret; | |
4073 | struct kmem_cache *c = NULL; | |
4074 | int i = 0; | |
4075 | ||
4076 | ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp); | |
4077 | ||
4078 | if (slab_state < FULL) | |
4079 | return ret; | |
4080 | ||
4081 | if ((ret < 0) || !is_root_cache(cachep)) | |
4082 | return ret; | |
4083 | ||
ebe945c2 | 4084 | VM_BUG_ON(!mutex_is_locked(&slab_mutex)); |
943a451a GC |
4085 | for_each_memcg_cache_index(i) { |
4086 | c = cache_from_memcg(cachep, i); | |
4087 | if (c) | |
4088 | /* return value determined by the parent cache only */ | |
4089 | __do_tune_cpucache(c, limit, batchcount, shared, gfp); | |
4090 | } | |
4091 | ||
4092 | return ret; | |
4093 | } | |
4094 | ||
18004c5d | 4095 | /* Called with slab_mutex held always */ |
83b519e8 | 4096 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4097 | { |
4098 | int err; | |
943a451a GC |
4099 | int limit = 0; |
4100 | int shared = 0; | |
4101 | int batchcount = 0; | |
4102 | ||
4103 | if (!is_root_cache(cachep)) { | |
4104 | struct kmem_cache *root = memcg_root_cache(cachep); | |
4105 | limit = root->limit; | |
4106 | shared = root->shared; | |
4107 | batchcount = root->batchcount; | |
4108 | } | |
1da177e4 | 4109 | |
943a451a GC |
4110 | if (limit && shared && batchcount) |
4111 | goto skip_setup; | |
a737b3e2 AM |
4112 | /* |
4113 | * The head array serves three purposes: | |
1da177e4 LT |
4114 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4115 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4116 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4117 | * bufctl chains: array operations are cheaper. |
4118 | * The numbers are guessed, we should auto-tune as described by | |
4119 | * Bonwick. | |
4120 | */ | |
3b0efdfa | 4121 | if (cachep->size > 131072) |
1da177e4 | 4122 | limit = 1; |
3b0efdfa | 4123 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4124 | limit = 8; |
3b0efdfa | 4125 | else if (cachep->size > 1024) |
1da177e4 | 4126 | limit = 24; |
3b0efdfa | 4127 | else if (cachep->size > 256) |
1da177e4 LT |
4128 | limit = 54; |
4129 | else | |
4130 | limit = 120; | |
4131 | ||
a737b3e2 AM |
4132 | /* |
4133 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4134 | * allocation behaviour: Most allocs on one cpu, most free operations |
4135 | * on another cpu. For these cases, an efficient object passing between | |
4136 | * cpus is necessary. This is provided by a shared array. The array | |
4137 | * replaces Bonwick's magazine layer. | |
4138 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4139 | * to a larger limit. Thus disabled by default. | |
4140 | */ | |
4141 | shared = 0; | |
3b0efdfa | 4142 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4143 | shared = 8; |
1da177e4 LT |
4144 | |
4145 | #if DEBUG | |
a737b3e2 AM |
4146 | /* |
4147 | * With debugging enabled, large batchcount lead to excessively long | |
4148 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4149 | */ |
4150 | if (limit > 32) | |
4151 | limit = 32; | |
4152 | #endif | |
943a451a GC |
4153 | batchcount = (limit + 1) / 2; |
4154 | skip_setup: | |
4155 | err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp); | |
1da177e4 LT |
4156 | if (err) |
4157 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4158 | cachep->name, -err); |
2ed3a4ef | 4159 | return err; |
1da177e4 LT |
4160 | } |
4161 | ||
1b55253a CL |
4162 | /* |
4163 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4164 | * necessary. Note that the l3 listlock also protects the array_cache |
4165 | * if drain_array() is used on the shared array. | |
1b55253a | 4166 | */ |
6744f087 | 4167 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3, |
1b55253a | 4168 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4169 | { |
4170 | int tofree; | |
4171 | ||
1b55253a CL |
4172 | if (!ac || !ac->avail) |
4173 | return; | |
1da177e4 LT |
4174 | if (ac->touched && !force) { |
4175 | ac->touched = 0; | |
b18e7e65 | 4176 | } else { |
1b55253a | 4177 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4178 | if (ac->avail) { |
4179 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4180 | if (tofree > ac->avail) | |
4181 | tofree = (ac->avail + 1) / 2; | |
4182 | free_block(cachep, ac->entry, tofree, node); | |
4183 | ac->avail -= tofree; | |
4184 | memmove(ac->entry, &(ac->entry[tofree]), | |
4185 | sizeof(void *) * ac->avail); | |
4186 | } | |
1b55253a | 4187 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4188 | } |
4189 | } | |
4190 | ||
4191 | /** | |
4192 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4193 | * @w: work descriptor |
1da177e4 LT |
4194 | * |
4195 | * Called from workqueue/eventd every few seconds. | |
4196 | * Purpose: | |
4197 | * - clear the per-cpu caches for this CPU. | |
4198 | * - return freeable pages to the main free memory pool. | |
4199 | * | |
a737b3e2 AM |
4200 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4201 | * again on the next iteration. | |
1da177e4 | 4202 | */ |
7c5cae36 | 4203 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4204 | { |
7a7c381d | 4205 | struct kmem_cache *searchp; |
6744f087 | 4206 | struct kmem_cache_node *l3; |
7d6e6d09 | 4207 | int node = numa_mem_id(); |
bf6aede7 | 4208 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4209 | |
18004c5d | 4210 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 4211 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4212 | goto out; |
1da177e4 | 4213 | |
18004c5d | 4214 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
4215 | check_irq_on(); |
4216 | ||
35386e3b CL |
4217 | /* |
4218 | * We only take the l3 lock if absolutely necessary and we | |
4219 | * have established with reasonable certainty that | |
4220 | * we can do some work if the lock was obtained. | |
4221 | */ | |
aab2207c | 4222 | l3 = searchp->nodelists[node]; |
35386e3b | 4223 | |
8fce4d8e | 4224 | reap_alien(searchp, l3); |
1da177e4 | 4225 | |
aab2207c | 4226 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4227 | |
35386e3b CL |
4228 | /* |
4229 | * These are racy checks but it does not matter | |
4230 | * if we skip one check or scan twice. | |
4231 | */ | |
e498be7d | 4232 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4233 | goto next; |
1da177e4 | 4234 | |
e498be7d | 4235 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4236 | |
aab2207c | 4237 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4238 | |
ed11d9eb | 4239 | if (l3->free_touched) |
e498be7d | 4240 | l3->free_touched = 0; |
ed11d9eb CL |
4241 | else { |
4242 | int freed; | |
1da177e4 | 4243 | |
ed11d9eb CL |
4244 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4245 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4246 | STATS_ADD_REAPED(searchp, freed); | |
4247 | } | |
35386e3b | 4248 | next: |
1da177e4 LT |
4249 | cond_resched(); |
4250 | } | |
4251 | check_irq_on(); | |
18004c5d | 4252 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4253 | next_reap_node(); |
7c5cae36 | 4254 | out: |
a737b3e2 | 4255 | /* Set up the next iteration */ |
7c5cae36 | 4256 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4257 | } |
4258 | ||
158a9624 | 4259 | #ifdef CONFIG_SLABINFO |
0d7561c6 | 4260 | void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) |
1da177e4 | 4261 | { |
b28a02de PE |
4262 | struct slab *slabp; |
4263 | unsigned long active_objs; | |
4264 | unsigned long num_objs; | |
4265 | unsigned long active_slabs = 0; | |
4266 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4267 | const char *name; |
1da177e4 | 4268 | char *error = NULL; |
e498be7d | 4269 | int node; |
6744f087 | 4270 | struct kmem_cache_node *l3; |
1da177e4 | 4271 | |
1da177e4 LT |
4272 | active_objs = 0; |
4273 | num_slabs = 0; | |
e498be7d CL |
4274 | for_each_online_node(node) { |
4275 | l3 = cachep->nodelists[node]; | |
4276 | if (!l3) | |
4277 | continue; | |
4278 | ||
ca3b9b91 RT |
4279 | check_irq_on(); |
4280 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4281 | |
7a7c381d | 4282 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4283 | if (slabp->inuse != cachep->num && !error) |
4284 | error = "slabs_full accounting error"; | |
4285 | active_objs += cachep->num; | |
4286 | active_slabs++; | |
4287 | } | |
7a7c381d | 4288 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4289 | if (slabp->inuse == cachep->num && !error) |
4290 | error = "slabs_partial inuse accounting error"; | |
4291 | if (!slabp->inuse && !error) | |
4292 | error = "slabs_partial/inuse accounting error"; | |
4293 | active_objs += slabp->inuse; | |
4294 | active_slabs++; | |
4295 | } | |
7a7c381d | 4296 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4297 | if (slabp->inuse && !error) |
4298 | error = "slabs_free/inuse accounting error"; | |
4299 | num_slabs++; | |
4300 | } | |
4301 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4302 | if (l3->shared) |
4303 | shared_avail += l3->shared->avail; | |
e498be7d | 4304 | |
ca3b9b91 | 4305 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4306 | } |
b28a02de PE |
4307 | num_slabs += active_slabs; |
4308 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4309 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4310 | error = "free_objects accounting error"; |
4311 | ||
b28a02de | 4312 | name = cachep->name; |
1da177e4 LT |
4313 | if (error) |
4314 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4315 | ||
0d7561c6 GC |
4316 | sinfo->active_objs = active_objs; |
4317 | sinfo->num_objs = num_objs; | |
4318 | sinfo->active_slabs = active_slabs; | |
4319 | sinfo->num_slabs = num_slabs; | |
4320 | sinfo->shared_avail = shared_avail; | |
4321 | sinfo->limit = cachep->limit; | |
4322 | sinfo->batchcount = cachep->batchcount; | |
4323 | sinfo->shared = cachep->shared; | |
4324 | sinfo->objects_per_slab = cachep->num; | |
4325 | sinfo->cache_order = cachep->gfporder; | |
4326 | } | |
4327 | ||
4328 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) | |
4329 | { | |
1da177e4 | 4330 | #if STATS |
b28a02de | 4331 | { /* list3 stats */ |
1da177e4 LT |
4332 | unsigned long high = cachep->high_mark; |
4333 | unsigned long allocs = cachep->num_allocations; | |
4334 | unsigned long grown = cachep->grown; | |
4335 | unsigned long reaped = cachep->reaped; | |
4336 | unsigned long errors = cachep->errors; | |
4337 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4338 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4339 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4340 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4341 | |
e92dd4fd JP |
4342 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4343 | "%4lu %4lu %4lu %4lu %4lu", | |
4344 | allocs, high, grown, | |
4345 | reaped, errors, max_freeable, node_allocs, | |
4346 | node_frees, overflows); | |
1da177e4 LT |
4347 | } |
4348 | /* cpu stats */ | |
4349 | { | |
4350 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4351 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4352 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4353 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4354 | ||
4355 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4356 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4357 | } |
4358 | #endif | |
1da177e4 LT |
4359 | } |
4360 | ||
1da177e4 LT |
4361 | #define MAX_SLABINFO_WRITE 128 |
4362 | /** | |
4363 | * slabinfo_write - Tuning for the slab allocator | |
4364 | * @file: unused | |
4365 | * @buffer: user buffer | |
4366 | * @count: data length | |
4367 | * @ppos: unused | |
4368 | */ | |
b7454ad3 | 4369 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4370 | size_t count, loff_t *ppos) |
1da177e4 | 4371 | { |
b28a02de | 4372 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4373 | int limit, batchcount, shared, res; |
7a7c381d | 4374 | struct kmem_cache *cachep; |
b28a02de | 4375 | |
1da177e4 LT |
4376 | if (count > MAX_SLABINFO_WRITE) |
4377 | return -EINVAL; | |
4378 | if (copy_from_user(&kbuf, buffer, count)) | |
4379 | return -EFAULT; | |
b28a02de | 4380 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4381 | |
4382 | tmp = strchr(kbuf, ' '); | |
4383 | if (!tmp) | |
4384 | return -EINVAL; | |
4385 | *tmp = '\0'; | |
4386 | tmp++; | |
4387 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4388 | return -EINVAL; | |
4389 | ||
4390 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4391 | mutex_lock(&slab_mutex); |
1da177e4 | 4392 | res = -EINVAL; |
18004c5d | 4393 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4394 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4395 | if (limit < 1 || batchcount < 1 || |
4396 | batchcount > limit || shared < 0) { | |
e498be7d | 4397 | res = 0; |
1da177e4 | 4398 | } else { |
e498be7d | 4399 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4400 | batchcount, shared, |
4401 | GFP_KERNEL); | |
1da177e4 LT |
4402 | } |
4403 | break; | |
4404 | } | |
4405 | } | |
18004c5d | 4406 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4407 | if (res >= 0) |
4408 | res = count; | |
4409 | return res; | |
4410 | } | |
871751e2 AV |
4411 | |
4412 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4413 | ||
4414 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4415 | { | |
18004c5d CL |
4416 | mutex_lock(&slab_mutex); |
4417 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4418 | } |
4419 | ||
4420 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4421 | { | |
4422 | unsigned long *p; | |
4423 | int l; | |
4424 | if (!v) | |
4425 | return 1; | |
4426 | l = n[1]; | |
4427 | p = n + 2; | |
4428 | while (l) { | |
4429 | int i = l/2; | |
4430 | unsigned long *q = p + 2 * i; | |
4431 | if (*q == v) { | |
4432 | q[1]++; | |
4433 | return 1; | |
4434 | } | |
4435 | if (*q > v) { | |
4436 | l = i; | |
4437 | } else { | |
4438 | p = q + 2; | |
4439 | l -= i + 1; | |
4440 | } | |
4441 | } | |
4442 | if (++n[1] == n[0]) | |
4443 | return 0; | |
4444 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4445 | p[0] = v; | |
4446 | p[1] = 1; | |
4447 | return 1; | |
4448 | } | |
4449 | ||
4450 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4451 | { | |
4452 | void *p; | |
4453 | int i; | |
4454 | if (n[0] == n[1]) | |
4455 | return; | |
3b0efdfa | 4456 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4457 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4458 | continue; | |
4459 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4460 | return; | |
4461 | } | |
4462 | } | |
4463 | ||
4464 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4465 | { | |
4466 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4467 | unsigned long offset, size; |
9281acea | 4468 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4469 | |
a5c43dae | 4470 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4471 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4472 | if (modname[0]) |
871751e2 AV |
4473 | seq_printf(m, " [%s]", modname); |
4474 | return; | |
4475 | } | |
4476 | #endif | |
4477 | seq_printf(m, "%p", (void *)address); | |
4478 | } | |
4479 | ||
4480 | static int leaks_show(struct seq_file *m, void *p) | |
4481 | { | |
0672aa7c | 4482 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 | 4483 | struct slab *slabp; |
6744f087 | 4484 | struct kmem_cache_node *l3; |
871751e2 AV |
4485 | const char *name; |
4486 | unsigned long *n = m->private; | |
4487 | int node; | |
4488 | int i; | |
4489 | ||
4490 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4491 | return 0; | |
4492 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4493 | return 0; | |
4494 | ||
4495 | /* OK, we can do it */ | |
4496 | ||
4497 | n[1] = 0; | |
4498 | ||
4499 | for_each_online_node(node) { | |
4500 | l3 = cachep->nodelists[node]; | |
4501 | if (!l3) | |
4502 | continue; | |
4503 | ||
4504 | check_irq_on(); | |
4505 | spin_lock_irq(&l3->list_lock); | |
4506 | ||
7a7c381d | 4507 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4508 | handle_slab(n, cachep, slabp); |
7a7c381d | 4509 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4510 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4511 | spin_unlock_irq(&l3->list_lock); |
4512 | } | |
4513 | name = cachep->name; | |
4514 | if (n[0] == n[1]) { | |
4515 | /* Increase the buffer size */ | |
18004c5d | 4516 | mutex_unlock(&slab_mutex); |
871751e2 AV |
4517 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4518 | if (!m->private) { | |
4519 | /* Too bad, we are really out */ | |
4520 | m->private = n; | |
18004c5d | 4521 | mutex_lock(&slab_mutex); |
871751e2 AV |
4522 | return -ENOMEM; |
4523 | } | |
4524 | *(unsigned long *)m->private = n[0] * 2; | |
4525 | kfree(n); | |
18004c5d | 4526 | mutex_lock(&slab_mutex); |
871751e2 AV |
4527 | /* Now make sure this entry will be retried */ |
4528 | m->count = m->size; | |
4529 | return 0; | |
4530 | } | |
4531 | for (i = 0; i < n[1]; i++) { | |
4532 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4533 | show_symbol(m, n[2*i+2]); | |
4534 | seq_putc(m, '\n'); | |
4535 | } | |
d2e7b7d0 | 4536 | |
871751e2 AV |
4537 | return 0; |
4538 | } | |
4539 | ||
b7454ad3 GC |
4540 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) |
4541 | { | |
4542 | return seq_list_next(p, &slab_caches, pos); | |
4543 | } | |
4544 | ||
4545 | static void s_stop(struct seq_file *m, void *p) | |
4546 | { | |
4547 | mutex_unlock(&slab_mutex); | |
4548 | } | |
4549 | ||
a0ec95a8 | 4550 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4551 | .start = leaks_start, |
4552 | .next = s_next, | |
4553 | .stop = s_stop, | |
4554 | .show = leaks_show, | |
4555 | }; | |
a0ec95a8 AD |
4556 | |
4557 | static int slabstats_open(struct inode *inode, struct file *file) | |
4558 | { | |
4559 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4560 | int ret = -ENOMEM; | |
4561 | if (n) { | |
4562 | ret = seq_open(file, &slabstats_op); | |
4563 | if (!ret) { | |
4564 | struct seq_file *m = file->private_data; | |
4565 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4566 | m->private = n; | |
4567 | n = NULL; | |
4568 | } | |
4569 | kfree(n); | |
4570 | } | |
4571 | return ret; | |
4572 | } | |
4573 | ||
4574 | static const struct file_operations proc_slabstats_operations = { | |
4575 | .open = slabstats_open, | |
4576 | .read = seq_read, | |
4577 | .llseek = seq_lseek, | |
4578 | .release = seq_release_private, | |
4579 | }; | |
4580 | #endif | |
4581 | ||
4582 | static int __init slab_proc_init(void) | |
4583 | { | |
4584 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4585 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4586 | #endif |
a0ec95a8 AD |
4587 | return 0; |
4588 | } | |
4589 | module_init(slab_proc_init); | |
1da177e4 LT |
4590 | #endif |
4591 | ||
00e145b6 MS |
4592 | /** |
4593 | * ksize - get the actual amount of memory allocated for a given object | |
4594 | * @objp: Pointer to the object | |
4595 | * | |
4596 | * kmalloc may internally round up allocations and return more memory | |
4597 | * than requested. ksize() can be used to determine the actual amount of | |
4598 | * memory allocated. The caller may use this additional memory, even though | |
4599 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4600 | * The caller must guarantee that objp points to a valid object previously | |
4601 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4602 | * must not be freed during the duration of the call. | |
4603 | */ | |
fd76bab2 | 4604 | size_t ksize(const void *objp) |
1da177e4 | 4605 | { |
ef8b4520 CL |
4606 | BUG_ON(!objp); |
4607 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4608 | return 0; |
1da177e4 | 4609 | |
8c138bc0 | 4610 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4611 | } |
b1aabecd | 4612 | EXPORT_SYMBOL(ksize); |