]>
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 | |
e0a42726 IM |
321 | static int slab_early_init = 1; |
322 | ||
e3366016 | 323 | #define INDEX_AC kmalloc_index(sizeof(struct arraycache_init)) |
6744f087 | 324 | #define INDEX_L3 kmalloc_index(sizeof(struct kmem_cache_node)) |
1da177e4 | 325 | |
6744f087 | 326 | static void kmem_list3_init(struct kmem_cache_node *parent) |
e498be7d CL |
327 | { |
328 | INIT_LIST_HEAD(&parent->slabs_full); | |
329 | INIT_LIST_HEAD(&parent->slabs_partial); | |
330 | INIT_LIST_HEAD(&parent->slabs_free); | |
331 | parent->shared = NULL; | |
332 | parent->alien = NULL; | |
2e1217cf | 333 | parent->colour_next = 0; |
e498be7d CL |
334 | spin_lock_init(&parent->list_lock); |
335 | parent->free_objects = 0; | |
336 | parent->free_touched = 0; | |
337 | } | |
338 | ||
a737b3e2 AM |
339 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
340 | do { \ | |
341 | INIT_LIST_HEAD(listp); \ | |
6a67368c | 342 | list_splice(&(cachep->node[nodeid]->slab), listp); \ |
e498be7d CL |
343 | } while (0) |
344 | ||
a737b3e2 AM |
345 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
346 | do { \ | |
e498be7d CL |
347 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
348 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
349 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
350 | } while (0) | |
1da177e4 | 351 | |
1da177e4 LT |
352 | #define CFLGS_OFF_SLAB (0x80000000UL) |
353 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
354 | ||
355 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
356 | /* |
357 | * Optimization question: fewer reaps means less probability for unnessary | |
358 | * cpucache drain/refill cycles. | |
1da177e4 | 359 | * |
dc6f3f27 | 360 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
361 | * which could lock up otherwise freeable slabs. |
362 | */ | |
363 | #define REAPTIMEOUT_CPUC (2*HZ) | |
364 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
365 | ||
366 | #if STATS | |
367 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
368 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
369 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
370 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 371 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
372 | #define STATS_SET_HIGH(x) \ |
373 | do { \ | |
374 | if ((x)->num_active > (x)->high_mark) \ | |
375 | (x)->high_mark = (x)->num_active; \ | |
376 | } while (0) | |
1da177e4 LT |
377 | #define STATS_INC_ERR(x) ((x)->errors++) |
378 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 379 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 380 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
381 | #define STATS_SET_FREEABLE(x, i) \ |
382 | do { \ | |
383 | if ((x)->max_freeable < i) \ | |
384 | (x)->max_freeable = i; \ | |
385 | } while (0) | |
1da177e4 LT |
386 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
387 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
388 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
389 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
390 | #else | |
391 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
392 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
393 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
394 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 395 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
396 | #define STATS_SET_HIGH(x) do { } while (0) |
397 | #define STATS_INC_ERR(x) do { } while (0) | |
398 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 399 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 400 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 401 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
402 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
403 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
404 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
405 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
406 | #endif | |
407 | ||
408 | #if DEBUG | |
1da177e4 | 409 | |
a737b3e2 AM |
410 | /* |
411 | * memory layout of objects: | |
1da177e4 | 412 | * 0 : objp |
3dafccf2 | 413 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
414 | * the end of an object is aligned with the end of the real |
415 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 416 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 417 | * redzone word. |
3dafccf2 | 418 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
419 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
420 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 421 | * [BYTES_PER_WORD long] |
1da177e4 | 422 | */ |
343e0d7a | 423 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 424 | { |
3dafccf2 | 425 | return cachep->obj_offset; |
1da177e4 LT |
426 | } |
427 | ||
b46b8f19 | 428 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
429 | { |
430 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
431 | return (unsigned long long*) (objp + obj_offset(cachep) - |
432 | sizeof(unsigned long long)); | |
1da177e4 LT |
433 | } |
434 | ||
b46b8f19 | 435 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
436 | { |
437 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
438 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 439 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 440 | sizeof(unsigned long long) - |
87a927c7 | 441 | REDZONE_ALIGN); |
3b0efdfa | 442 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 443 | sizeof(unsigned long long)); |
1da177e4 LT |
444 | } |
445 | ||
343e0d7a | 446 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
447 | { |
448 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 449 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
450 | } |
451 | ||
452 | #else | |
453 | ||
3dafccf2 | 454 | #define obj_offset(x) 0 |
b46b8f19 DW |
455 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
456 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
457 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
458 | ||
459 | #endif | |
460 | ||
1da177e4 | 461 | /* |
3df1cccd DR |
462 | * Do not go above this order unless 0 objects fit into the slab or |
463 | * overridden on the command line. | |
1da177e4 | 464 | */ |
543585cc DR |
465 | #define SLAB_MAX_ORDER_HI 1 |
466 | #define SLAB_MAX_ORDER_LO 0 | |
467 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 468 | static bool slab_max_order_set __initdata; |
1da177e4 | 469 | |
6ed5eb22 PE |
470 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
471 | { | |
b49af68f | 472 | struct page *page = virt_to_head_page(obj); |
35026088 | 473 | return page->slab_cache; |
6ed5eb22 PE |
474 | } |
475 | ||
476 | static inline struct slab *virt_to_slab(const void *obj) | |
477 | { | |
b49af68f | 478 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
479 | |
480 | VM_BUG_ON(!PageSlab(page)); | |
481 | return page->slab_page; | |
6ed5eb22 PE |
482 | } |
483 | ||
8fea4e96 PE |
484 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
485 | unsigned int idx) | |
486 | { | |
3b0efdfa | 487 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
488 | } |
489 | ||
6a2d7a95 | 490 | /* |
3b0efdfa CL |
491 | * We want to avoid an expensive divide : (offset / cache->size) |
492 | * Using the fact that size is a constant for a particular cache, | |
493 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
494 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
495 | */ | |
496 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
497 | const struct slab *slab, void *obj) | |
8fea4e96 | 498 | { |
6a2d7a95 ED |
499 | u32 offset = (obj - slab->s_mem); |
500 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
501 | } |
502 | ||
1da177e4 | 503 | static struct arraycache_init initarray_generic = |
b28a02de | 504 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
505 | |
506 | /* internal cache of cache description objs */ | |
9b030cb8 | 507 | static struct kmem_cache kmem_cache_boot = { |
b28a02de PE |
508 | .batchcount = 1, |
509 | .limit = BOOT_CPUCACHE_ENTRIES, | |
510 | .shared = 1, | |
3b0efdfa | 511 | .size = sizeof(struct kmem_cache), |
b28a02de | 512 | .name = "kmem_cache", |
1da177e4 LT |
513 | }; |
514 | ||
056c6241 RT |
515 | #define BAD_ALIEN_MAGIC 0x01020304ul |
516 | ||
f1aaee53 AV |
517 | #ifdef CONFIG_LOCKDEP |
518 | ||
519 | /* | |
520 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
521 | * for other slabs "off slab". | |
522 | * The locking for this is tricky in that it nests within the locks | |
523 | * of all other slabs in a few places; to deal with this special | |
524 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
525 | * |
526 | * We set lock class for alien array caches which are up during init. | |
527 | * The lock annotation will be lost if all cpus of a node goes down and | |
528 | * then comes back up during hotplug | |
f1aaee53 | 529 | */ |
056c6241 RT |
530 | static struct lock_class_key on_slab_l3_key; |
531 | static struct lock_class_key on_slab_alc_key; | |
532 | ||
83835b3d PZ |
533 | static struct lock_class_key debugobj_l3_key; |
534 | static struct lock_class_key debugobj_alc_key; | |
535 | ||
536 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
537 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
538 | int q) | |
539 | { | |
540 | struct array_cache **alc; | |
6744f087 | 541 | struct kmem_cache_node *l3; |
83835b3d PZ |
542 | int r; |
543 | ||
6a67368c | 544 | l3 = cachep->node[q]; |
83835b3d PZ |
545 | if (!l3) |
546 | return; | |
547 | ||
548 | lockdep_set_class(&l3->list_lock, l3_key); | |
549 | alc = l3->alien; | |
550 | /* | |
551 | * FIXME: This check for BAD_ALIEN_MAGIC | |
552 | * should go away when common slab code is taught to | |
553 | * work even without alien caches. | |
554 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
555 | * for alloc_alien_cache, | |
556 | */ | |
557 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
558 | return; | |
559 | for_each_node(r) { | |
560 | if (alc[r]) | |
561 | lockdep_set_class(&alc[r]->lock, alc_key); | |
562 | } | |
563 | } | |
564 | ||
565 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
566 | { | |
567 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
568 | } | |
569 | ||
570 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
571 | { | |
572 | int node; | |
573 | ||
574 | for_each_online_node(node) | |
575 | slab_set_debugobj_lock_classes_node(cachep, node); | |
576 | } | |
577 | ||
ce79ddc8 | 578 | static void init_node_lock_keys(int q) |
f1aaee53 | 579 | { |
e3366016 | 580 | int i; |
056c6241 | 581 | |
97d06609 | 582 | if (slab_state < UP) |
ce79ddc8 PE |
583 | return; |
584 | ||
e3366016 | 585 | for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) { |
6744f087 | 586 | struct kmem_cache_node *l3; |
e3366016 CL |
587 | struct kmem_cache *cache = kmalloc_caches[i]; |
588 | ||
589 | if (!cache) | |
590 | continue; | |
ce79ddc8 | 591 | |
6a67368c | 592 | l3 = cache->node[q]; |
e3366016 | 593 | if (!l3 || OFF_SLAB(cache)) |
00afa758 | 594 | continue; |
83835b3d | 595 | |
e3366016 | 596 | slab_set_lock_classes(cache, &on_slab_l3_key, |
83835b3d | 597 | &on_slab_alc_key, q); |
f1aaee53 AV |
598 | } |
599 | } | |
ce79ddc8 | 600 | |
6ccfb5bc GC |
601 | static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q) |
602 | { | |
6a67368c | 603 | if (!cachep->node[q]) |
6ccfb5bc GC |
604 | return; |
605 | ||
606 | slab_set_lock_classes(cachep, &on_slab_l3_key, | |
607 | &on_slab_alc_key, q); | |
608 | } | |
609 | ||
610 | static inline void on_slab_lock_classes(struct kmem_cache *cachep) | |
611 | { | |
612 | int node; | |
613 | ||
614 | VM_BUG_ON(OFF_SLAB(cachep)); | |
615 | for_each_node(node) | |
616 | on_slab_lock_classes_node(cachep, node); | |
617 | } | |
618 | ||
ce79ddc8 PE |
619 | static inline void init_lock_keys(void) |
620 | { | |
621 | int node; | |
622 | ||
623 | for_each_node(node) | |
624 | init_node_lock_keys(node); | |
625 | } | |
f1aaee53 | 626 | #else |
ce79ddc8 PE |
627 | static void init_node_lock_keys(int q) |
628 | { | |
629 | } | |
630 | ||
056c6241 | 631 | static inline void init_lock_keys(void) |
f1aaee53 AV |
632 | { |
633 | } | |
83835b3d | 634 | |
6ccfb5bc GC |
635 | static inline void on_slab_lock_classes(struct kmem_cache *cachep) |
636 | { | |
637 | } | |
638 | ||
639 | static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node) | |
640 | { | |
641 | } | |
642 | ||
83835b3d PZ |
643 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) |
644 | { | |
645 | } | |
646 | ||
647 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
648 | { | |
649 | } | |
f1aaee53 AV |
650 | #endif |
651 | ||
1871e52c | 652 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 653 | |
343e0d7a | 654 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
655 | { |
656 | return cachep->array[smp_processor_id()]; | |
657 | } | |
658 | ||
a737b3e2 AM |
659 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
660 | gfp_t gfpflags) | |
1da177e4 | 661 | { |
e3366016 | 662 | int i; |
1da177e4 LT |
663 | |
664 | #if DEBUG | |
665 | /* This happens if someone tries to call | |
b28a02de PE |
666 | * kmem_cache_create(), or __kmalloc(), before |
667 | * the generic caches are initialized. | |
668 | */ | |
e3366016 | 669 | BUG_ON(kmalloc_caches[INDEX_AC] == NULL); |
1da177e4 | 670 | #endif |
6cb8f913 CL |
671 | if (!size) |
672 | return ZERO_SIZE_PTR; | |
673 | ||
e3366016 | 674 | i = kmalloc_index(size); |
1da177e4 LT |
675 | |
676 | /* | |
0abf40c1 | 677 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
678 | * has cs_{dma,}cachep==NULL. Thus no special case |
679 | * for large kmalloc calls required. | |
680 | */ | |
4b51d669 | 681 | #ifdef CONFIG_ZONE_DMA |
1da177e4 | 682 | if (unlikely(gfpflags & GFP_DMA)) |
e3366016 | 683 | return kmalloc_dma_caches[i]; |
4b51d669 | 684 | #endif |
e3366016 | 685 | return kmalloc_caches[i]; |
1da177e4 LT |
686 | } |
687 | ||
b221385b | 688 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
689 | { |
690 | return __find_general_cachep(size, gfpflags); | |
691 | } | |
97e2bde4 | 692 | |
fbaccacf | 693 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 694 | { |
fbaccacf SR |
695 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
696 | } | |
1da177e4 | 697 | |
a737b3e2 AM |
698 | /* |
699 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
700 | */ | |
fbaccacf SR |
701 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
702 | size_t align, int flags, size_t *left_over, | |
703 | unsigned int *num) | |
704 | { | |
705 | int nr_objs; | |
706 | size_t mgmt_size; | |
707 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 708 | |
fbaccacf SR |
709 | /* |
710 | * The slab management structure can be either off the slab or | |
711 | * on it. For the latter case, the memory allocated for a | |
712 | * slab is used for: | |
713 | * | |
714 | * - The struct slab | |
715 | * - One kmem_bufctl_t for each object | |
716 | * - Padding to respect alignment of @align | |
717 | * - @buffer_size bytes for each object | |
718 | * | |
719 | * If the slab management structure is off the slab, then the | |
720 | * alignment will already be calculated into the size. Because | |
721 | * the slabs are all pages aligned, the objects will be at the | |
722 | * correct alignment when allocated. | |
723 | */ | |
724 | if (flags & CFLGS_OFF_SLAB) { | |
725 | mgmt_size = 0; | |
726 | nr_objs = slab_size / buffer_size; | |
727 | ||
728 | if (nr_objs > SLAB_LIMIT) | |
729 | nr_objs = SLAB_LIMIT; | |
730 | } else { | |
731 | /* | |
732 | * Ignore padding for the initial guess. The padding | |
733 | * is at most @align-1 bytes, and @buffer_size is at | |
734 | * least @align. In the worst case, this result will | |
735 | * be one greater than the number of objects that fit | |
736 | * into the memory allocation when taking the padding | |
737 | * into account. | |
738 | */ | |
739 | nr_objs = (slab_size - sizeof(struct slab)) / | |
740 | (buffer_size + sizeof(kmem_bufctl_t)); | |
741 | ||
742 | /* | |
743 | * This calculated number will be either the right | |
744 | * amount, or one greater than what we want. | |
745 | */ | |
746 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
747 | > slab_size) | |
748 | nr_objs--; | |
749 | ||
750 | if (nr_objs > SLAB_LIMIT) | |
751 | nr_objs = SLAB_LIMIT; | |
752 | ||
753 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
754 | } | |
755 | *num = nr_objs; | |
756 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
757 | } |
758 | ||
f28510d3 | 759 | #if DEBUG |
d40cee24 | 760 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 761 | |
a737b3e2 AM |
762 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
763 | char *msg) | |
1da177e4 LT |
764 | { |
765 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 766 | function, cachep->name, msg); |
1da177e4 | 767 | dump_stack(); |
645df230 | 768 | add_taint(TAINT_BAD_PAGE); |
1da177e4 | 769 | } |
f28510d3 | 770 | #endif |
1da177e4 | 771 | |
3395ee05 PM |
772 | /* |
773 | * By default on NUMA we use alien caches to stage the freeing of | |
774 | * objects allocated from other nodes. This causes massive memory | |
775 | * inefficiencies when using fake NUMA setup to split memory into a | |
776 | * large number of small nodes, so it can be disabled on the command | |
777 | * line | |
778 | */ | |
779 | ||
780 | static int use_alien_caches __read_mostly = 1; | |
781 | static int __init noaliencache_setup(char *s) | |
782 | { | |
783 | use_alien_caches = 0; | |
784 | return 1; | |
785 | } | |
786 | __setup("noaliencache", noaliencache_setup); | |
787 | ||
3df1cccd DR |
788 | static int __init slab_max_order_setup(char *str) |
789 | { | |
790 | get_option(&str, &slab_max_order); | |
791 | slab_max_order = slab_max_order < 0 ? 0 : | |
792 | min(slab_max_order, MAX_ORDER - 1); | |
793 | slab_max_order_set = true; | |
794 | ||
795 | return 1; | |
796 | } | |
797 | __setup("slab_max_order=", slab_max_order_setup); | |
798 | ||
8fce4d8e CL |
799 | #ifdef CONFIG_NUMA |
800 | /* | |
801 | * Special reaping functions for NUMA systems called from cache_reap(). | |
802 | * These take care of doing round robin flushing of alien caches (containing | |
803 | * objects freed on different nodes from which they were allocated) and the | |
804 | * flushing of remote pcps by calling drain_node_pages. | |
805 | */ | |
1871e52c | 806 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
807 | |
808 | static void init_reap_node(int cpu) | |
809 | { | |
810 | int node; | |
811 | ||
7d6e6d09 | 812 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 813 | if (node == MAX_NUMNODES) |
442295c9 | 814 | node = first_node(node_online_map); |
8fce4d8e | 815 | |
1871e52c | 816 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
817 | } |
818 | ||
819 | static void next_reap_node(void) | |
820 | { | |
909ea964 | 821 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 822 | |
8fce4d8e CL |
823 | node = next_node(node, node_online_map); |
824 | if (unlikely(node >= MAX_NUMNODES)) | |
825 | node = first_node(node_online_map); | |
909ea964 | 826 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
827 | } |
828 | ||
829 | #else | |
830 | #define init_reap_node(cpu) do { } while (0) | |
831 | #define next_reap_node(void) do { } while (0) | |
832 | #endif | |
833 | ||
1da177e4 LT |
834 | /* |
835 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
836 | * via the workqueue/eventd. | |
837 | * Add the CPU number into the expiration time to minimize the possibility of | |
838 | * the CPUs getting into lockstep and contending for the global cache chain | |
839 | * lock. | |
840 | */ | |
897e679b | 841 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 842 | { |
1871e52c | 843 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
844 | |
845 | /* | |
846 | * When this gets called from do_initcalls via cpucache_init(), | |
847 | * init_workqueues() has already run, so keventd will be setup | |
848 | * at that time. | |
849 | */ | |
52bad64d | 850 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 851 | init_reap_node(cpu); |
203b42f7 | 852 | INIT_DEFERRABLE_WORK(reap_work, cache_reap); |
2b284214 AV |
853 | schedule_delayed_work_on(cpu, reap_work, |
854 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
855 | } |
856 | } | |
857 | ||
e498be7d | 858 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 859 | int batchcount, gfp_t gfp) |
1da177e4 | 860 | { |
b28a02de | 861 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
862 | struct array_cache *nc = NULL; |
863 | ||
83b519e8 | 864 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
865 | /* |
866 | * The array_cache structures contain pointers to free object. | |
25985edc | 867 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
868 | * cache the pointers are not cleared and they could be counted as |
869 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
870 | * not scan such objects. | |
871 | */ | |
872 | kmemleak_no_scan(nc); | |
1da177e4 LT |
873 | if (nc) { |
874 | nc->avail = 0; | |
875 | nc->limit = entries; | |
876 | nc->batchcount = batchcount; | |
877 | nc->touched = 0; | |
e498be7d | 878 | spin_lock_init(&nc->lock); |
1da177e4 LT |
879 | } |
880 | return nc; | |
881 | } | |
882 | ||
072bb0aa MG |
883 | static inline bool is_slab_pfmemalloc(struct slab *slabp) |
884 | { | |
885 | struct page *page = virt_to_page(slabp->s_mem); | |
886 | ||
887 | return PageSlabPfmemalloc(page); | |
888 | } | |
889 | ||
890 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
891 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
892 | struct array_cache *ac) | |
893 | { | |
6a67368c | 894 | struct kmem_cache_node *l3 = cachep->node[numa_mem_id()]; |
072bb0aa MG |
895 | struct slab *slabp; |
896 | unsigned long flags; | |
897 | ||
898 | if (!pfmemalloc_active) | |
899 | return; | |
900 | ||
901 | spin_lock_irqsave(&l3->list_lock, flags); | |
902 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
903 | if (is_slab_pfmemalloc(slabp)) | |
904 | goto out; | |
905 | ||
906 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
907 | if (is_slab_pfmemalloc(slabp)) | |
908 | goto out; | |
909 | ||
910 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
911 | if (is_slab_pfmemalloc(slabp)) | |
912 | goto out; | |
913 | ||
914 | pfmemalloc_active = false; | |
915 | out: | |
916 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
917 | } | |
918 | ||
381760ea | 919 | static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, |
072bb0aa MG |
920 | gfp_t flags, bool force_refill) |
921 | { | |
922 | int i; | |
923 | void *objp = ac->entry[--ac->avail]; | |
924 | ||
925 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
926 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
6744f087 | 927 | struct kmem_cache_node *l3; |
072bb0aa MG |
928 | |
929 | if (gfp_pfmemalloc_allowed(flags)) { | |
930 | clear_obj_pfmemalloc(&objp); | |
931 | return objp; | |
932 | } | |
933 | ||
934 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
d014dc2e | 935 | for (i = 0; i < ac->avail; i++) { |
072bb0aa MG |
936 | /* If a !PFMEMALLOC object is found, swap them */ |
937 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
938 | objp = ac->entry[i]; | |
939 | ac->entry[i] = ac->entry[ac->avail]; | |
940 | ac->entry[ac->avail] = objp; | |
941 | return objp; | |
942 | } | |
943 | } | |
944 | ||
945 | /* | |
946 | * If there are empty slabs on the slabs_free list and we are | |
947 | * being forced to refill the cache, mark this one !pfmemalloc. | |
948 | */ | |
6a67368c | 949 | l3 = cachep->node[numa_mem_id()]; |
072bb0aa MG |
950 | if (!list_empty(&l3->slabs_free) && force_refill) { |
951 | struct slab *slabp = virt_to_slab(objp); | |
30c29bea | 952 | ClearPageSlabPfmemalloc(virt_to_head_page(slabp->s_mem)); |
072bb0aa MG |
953 | clear_obj_pfmemalloc(&objp); |
954 | recheck_pfmemalloc_active(cachep, ac); | |
955 | return objp; | |
956 | } | |
957 | ||
958 | /* No !PFMEMALLOC objects available */ | |
959 | ac->avail++; | |
960 | objp = NULL; | |
961 | } | |
962 | ||
963 | return objp; | |
964 | } | |
965 | ||
381760ea MG |
966 | static inline void *ac_get_obj(struct kmem_cache *cachep, |
967 | struct array_cache *ac, gfp_t flags, bool force_refill) | |
968 | { | |
969 | void *objp; | |
970 | ||
971 | if (unlikely(sk_memalloc_socks())) | |
972 | objp = __ac_get_obj(cachep, ac, flags, force_refill); | |
973 | else | |
974 | objp = ac->entry[--ac->avail]; | |
975 | ||
976 | return objp; | |
977 | } | |
978 | ||
979 | static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
072bb0aa MG |
980 | void *objp) |
981 | { | |
982 | if (unlikely(pfmemalloc_active)) { | |
983 | /* Some pfmemalloc slabs exist, check if this is one */ | |
30c29bea | 984 | struct page *page = virt_to_head_page(objp); |
072bb0aa MG |
985 | if (PageSlabPfmemalloc(page)) |
986 | set_obj_pfmemalloc(&objp); | |
987 | } | |
988 | ||
381760ea MG |
989 | return objp; |
990 | } | |
991 | ||
992 | static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
993 | void *objp) | |
994 | { | |
995 | if (unlikely(sk_memalloc_socks())) | |
996 | objp = __ac_put_obj(cachep, ac, objp); | |
997 | ||
072bb0aa MG |
998 | ac->entry[ac->avail++] = objp; |
999 | } | |
1000 | ||
3ded175a CL |
1001 | /* |
1002 | * Transfer objects in one arraycache to another. | |
1003 | * Locking must be handled by the caller. | |
1004 | * | |
1005 | * Return the number of entries transferred. | |
1006 | */ | |
1007 | static int transfer_objects(struct array_cache *to, | |
1008 | struct array_cache *from, unsigned int max) | |
1009 | { | |
1010 | /* Figure out how many entries to transfer */ | |
732eacc0 | 1011 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
1012 | |
1013 | if (!nr) | |
1014 | return 0; | |
1015 | ||
1016 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1017 | sizeof(void *) *nr); | |
1018 | ||
1019 | from->avail -= nr; | |
1020 | to->avail += nr; | |
3ded175a CL |
1021 | return nr; |
1022 | } | |
1023 | ||
765c4507 CL |
1024 | #ifndef CONFIG_NUMA |
1025 | ||
1026 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1027 | #define reap_alien(cachep, l3) do { } while (0) | |
1028 | ||
83b519e8 | 1029 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1030 | { |
1031 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1032 | } | |
1033 | ||
1034 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1035 | { | |
1036 | } | |
1037 | ||
1038 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1039 | { | |
1040 | return 0; | |
1041 | } | |
1042 | ||
1043 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1044 | gfp_t flags) | |
1045 | { | |
1046 | return NULL; | |
1047 | } | |
1048 | ||
8b98c169 | 1049 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1050 | gfp_t flags, int nodeid) |
1051 | { | |
1052 | return NULL; | |
1053 | } | |
1054 | ||
1055 | #else /* CONFIG_NUMA */ | |
1056 | ||
8b98c169 | 1057 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1058 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1059 | |
83b519e8 | 1060 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1061 | { |
1062 | struct array_cache **ac_ptr; | |
8ef82866 | 1063 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1064 | int i; |
1065 | ||
1066 | if (limit > 1) | |
1067 | limit = 12; | |
f3186a9c | 1068 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
1069 | if (ac_ptr) { |
1070 | for_each_node(i) { | |
f3186a9c | 1071 | if (i == node || !node_online(i)) |
e498be7d | 1072 | continue; |
83b519e8 | 1073 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1074 | if (!ac_ptr[i]) { |
cc550def | 1075 | for (i--; i >= 0; i--) |
e498be7d CL |
1076 | kfree(ac_ptr[i]); |
1077 | kfree(ac_ptr); | |
1078 | return NULL; | |
1079 | } | |
1080 | } | |
1081 | } | |
1082 | return ac_ptr; | |
1083 | } | |
1084 | ||
5295a74c | 1085 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1086 | { |
1087 | int i; | |
1088 | ||
1089 | if (!ac_ptr) | |
1090 | return; | |
e498be7d | 1091 | for_each_node(i) |
b28a02de | 1092 | kfree(ac_ptr[i]); |
e498be7d CL |
1093 | kfree(ac_ptr); |
1094 | } | |
1095 | ||
343e0d7a | 1096 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1097 | struct array_cache *ac, int node) |
e498be7d | 1098 | { |
6a67368c | 1099 | struct kmem_cache_node *rl3 = cachep->node[node]; |
e498be7d CL |
1100 | |
1101 | if (ac->avail) { | |
1102 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1103 | /* |
1104 | * Stuff objects into the remote nodes shared array first. | |
1105 | * That way we could avoid the overhead of putting the objects | |
1106 | * into the free lists and getting them back later. | |
1107 | */ | |
693f7d36 JS |
1108 | if (rl3->shared) |
1109 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1110 | |
ff69416e | 1111 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1112 | ac->avail = 0; |
1113 | spin_unlock(&rl3->list_lock); | |
1114 | } | |
1115 | } | |
1116 | ||
8fce4d8e CL |
1117 | /* |
1118 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1119 | */ | |
6744f087 | 1120 | static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *l3) |
8fce4d8e | 1121 | { |
909ea964 | 1122 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1123 | |
1124 | if (l3->alien) { | |
1125 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1126 | |
1127 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1128 | __drain_alien_cache(cachep, ac, node); |
1129 | spin_unlock_irq(&ac->lock); | |
1130 | } | |
1131 | } | |
1132 | } | |
1133 | ||
a737b3e2 AM |
1134 | static void drain_alien_cache(struct kmem_cache *cachep, |
1135 | struct array_cache **alien) | |
e498be7d | 1136 | { |
b28a02de | 1137 | int i = 0; |
e498be7d CL |
1138 | struct array_cache *ac; |
1139 | unsigned long flags; | |
1140 | ||
1141 | for_each_online_node(i) { | |
4484ebf1 | 1142 | ac = alien[i]; |
e498be7d CL |
1143 | if (ac) { |
1144 | spin_lock_irqsave(&ac->lock, flags); | |
1145 | __drain_alien_cache(cachep, ac, i); | |
1146 | spin_unlock_irqrestore(&ac->lock, flags); | |
1147 | } | |
1148 | } | |
1149 | } | |
729bd0b7 | 1150 | |
873623df | 1151 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1152 | { |
1153 | struct slab *slabp = virt_to_slab(objp); | |
1154 | int nodeid = slabp->nodeid; | |
6744f087 | 1155 | struct kmem_cache_node *l3; |
729bd0b7 | 1156 | struct array_cache *alien = NULL; |
1ca4cb24 PE |
1157 | int node; |
1158 | ||
7d6e6d09 | 1159 | node = numa_mem_id(); |
729bd0b7 PE |
1160 | |
1161 | /* | |
1162 | * Make sure we are not freeing a object from another node to the array | |
1163 | * cache on this cpu. | |
1164 | */ | |
62918a03 | 1165 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1166 | return 0; |
1167 | ||
6a67368c | 1168 | l3 = cachep->node[node]; |
729bd0b7 PE |
1169 | STATS_INC_NODEFREES(cachep); |
1170 | if (l3->alien && l3->alien[nodeid]) { | |
1171 | alien = l3->alien[nodeid]; | |
873623df | 1172 | spin_lock(&alien->lock); |
729bd0b7 PE |
1173 | if (unlikely(alien->avail == alien->limit)) { |
1174 | STATS_INC_ACOVERFLOW(cachep); | |
1175 | __drain_alien_cache(cachep, alien, nodeid); | |
1176 | } | |
072bb0aa | 1177 | ac_put_obj(cachep, alien, objp); |
729bd0b7 PE |
1178 | spin_unlock(&alien->lock); |
1179 | } else { | |
6a67368c | 1180 | spin_lock(&(cachep->node[nodeid])->list_lock); |
729bd0b7 | 1181 | free_block(cachep, &objp, 1, nodeid); |
6a67368c | 1182 | spin_unlock(&(cachep->node[nodeid])->list_lock); |
729bd0b7 PE |
1183 | } |
1184 | return 1; | |
1185 | } | |
e498be7d CL |
1186 | #endif |
1187 | ||
8f9f8d9e | 1188 | /* |
6a67368c | 1189 | * Allocates and initializes node for a node on each slab cache, used for |
8f9f8d9e DR |
1190 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 |
1191 | * will be allocated off-node since memory is not yet online for the new node. | |
6a67368c | 1192 | * When hotplugging memory or a cpu, existing node are not replaced if |
8f9f8d9e DR |
1193 | * already in use. |
1194 | * | |
18004c5d | 1195 | * Must hold slab_mutex. |
8f9f8d9e | 1196 | */ |
6a67368c | 1197 | static int init_cache_node_node(int node) |
8f9f8d9e DR |
1198 | { |
1199 | struct kmem_cache *cachep; | |
6744f087 CL |
1200 | struct kmem_cache_node *l3; |
1201 | const int memsize = sizeof(struct kmem_cache_node); | |
8f9f8d9e | 1202 | |
18004c5d | 1203 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1204 | /* |
1205 | * Set up the size64 kmemlist for cpu before we can | |
1206 | * begin anything. Make sure some other cpu on this | |
1207 | * node has not already allocated this | |
1208 | */ | |
6a67368c | 1209 | if (!cachep->node[node]) { |
8f9f8d9e DR |
1210 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); |
1211 | if (!l3) | |
1212 | return -ENOMEM; | |
1213 | kmem_list3_init(l3); | |
1214 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1215 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1216 | ||
1217 | /* | |
1218 | * The l3s don't come and go as CPUs come and | |
18004c5d | 1219 | * go. slab_mutex is sufficient |
8f9f8d9e DR |
1220 | * protection here. |
1221 | */ | |
6a67368c | 1222 | cachep->node[node] = l3; |
8f9f8d9e DR |
1223 | } |
1224 | ||
6a67368c CL |
1225 | spin_lock_irq(&cachep->node[node]->list_lock); |
1226 | cachep->node[node]->free_limit = | |
8f9f8d9e DR |
1227 | (1 + nr_cpus_node(node)) * |
1228 | cachep->batchcount + cachep->num; | |
6a67368c | 1229 | spin_unlock_irq(&cachep->node[node]->list_lock); |
8f9f8d9e DR |
1230 | } |
1231 | return 0; | |
1232 | } | |
1233 | ||
fbf1e473 AM |
1234 | static void __cpuinit cpuup_canceled(long cpu) |
1235 | { | |
1236 | struct kmem_cache *cachep; | |
6744f087 | 1237 | struct kmem_cache_node *l3 = NULL; |
7d6e6d09 | 1238 | int node = cpu_to_mem(cpu); |
a70f7302 | 1239 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1240 | |
18004c5d | 1241 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1242 | struct array_cache *nc; |
1243 | struct array_cache *shared; | |
1244 | struct array_cache **alien; | |
fbf1e473 | 1245 | |
fbf1e473 AM |
1246 | /* cpu is dead; no one can alloc from it. */ |
1247 | nc = cachep->array[cpu]; | |
1248 | cachep->array[cpu] = NULL; | |
6a67368c | 1249 | l3 = cachep->node[node]; |
fbf1e473 AM |
1250 | |
1251 | if (!l3) | |
1252 | goto free_array_cache; | |
1253 | ||
1254 | spin_lock_irq(&l3->list_lock); | |
1255 | ||
1256 | /* Free limit for this kmem_list3 */ | |
1257 | l3->free_limit -= cachep->batchcount; | |
1258 | if (nc) | |
1259 | free_block(cachep, nc->entry, nc->avail, node); | |
1260 | ||
58463c1f | 1261 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1262 | spin_unlock_irq(&l3->list_lock); |
1263 | goto free_array_cache; | |
1264 | } | |
1265 | ||
1266 | shared = l3->shared; | |
1267 | if (shared) { | |
1268 | free_block(cachep, shared->entry, | |
1269 | shared->avail, node); | |
1270 | l3->shared = NULL; | |
1271 | } | |
1272 | ||
1273 | alien = l3->alien; | |
1274 | l3->alien = NULL; | |
1275 | ||
1276 | spin_unlock_irq(&l3->list_lock); | |
1277 | ||
1278 | kfree(shared); | |
1279 | if (alien) { | |
1280 | drain_alien_cache(cachep, alien); | |
1281 | free_alien_cache(alien); | |
1282 | } | |
1283 | free_array_cache: | |
1284 | kfree(nc); | |
1285 | } | |
1286 | /* | |
1287 | * In the previous loop, all the objects were freed to | |
1288 | * the respective cache's slabs, now we can go ahead and | |
1289 | * shrink each nodelist to its limit. | |
1290 | */ | |
18004c5d | 1291 | list_for_each_entry(cachep, &slab_caches, list) { |
6a67368c | 1292 | l3 = cachep->node[node]; |
fbf1e473 AM |
1293 | if (!l3) |
1294 | continue; | |
1295 | drain_freelist(cachep, l3, l3->free_objects); | |
1296 | } | |
1297 | } | |
1298 | ||
1299 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1300 | { |
343e0d7a | 1301 | struct kmem_cache *cachep; |
6744f087 | 1302 | struct kmem_cache_node *l3 = NULL; |
7d6e6d09 | 1303 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1304 | int err; |
1da177e4 | 1305 | |
fbf1e473 AM |
1306 | /* |
1307 | * We need to do this right in the beginning since | |
1308 | * alloc_arraycache's are going to use this list. | |
1309 | * kmalloc_node allows us to add the slab to the right | |
1310 | * kmem_list3 and not this cpu's kmem_list3 | |
1311 | */ | |
6a67368c | 1312 | err = init_cache_node_node(node); |
8f9f8d9e DR |
1313 | if (err < 0) |
1314 | goto bad; | |
fbf1e473 AM |
1315 | |
1316 | /* | |
1317 | * Now we can go ahead with allocating the shared arrays and | |
1318 | * array caches | |
1319 | */ | |
18004c5d | 1320 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1321 | struct array_cache *nc; |
1322 | struct array_cache *shared = NULL; | |
1323 | struct array_cache **alien = NULL; | |
1324 | ||
1325 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1326 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1327 | if (!nc) |
1328 | goto bad; | |
1329 | if (cachep->shared) { | |
1330 | shared = alloc_arraycache(node, | |
1331 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1332 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1333 | if (!shared) { |
1334 | kfree(nc); | |
1da177e4 | 1335 | goto bad; |
12d00f6a | 1336 | } |
fbf1e473 AM |
1337 | } |
1338 | if (use_alien_caches) { | |
83b519e8 | 1339 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1340 | if (!alien) { |
1341 | kfree(shared); | |
1342 | kfree(nc); | |
fbf1e473 | 1343 | goto bad; |
12d00f6a | 1344 | } |
fbf1e473 AM |
1345 | } |
1346 | cachep->array[cpu] = nc; | |
6a67368c | 1347 | l3 = cachep->node[node]; |
fbf1e473 AM |
1348 | BUG_ON(!l3); |
1349 | ||
1350 | spin_lock_irq(&l3->list_lock); | |
1351 | if (!l3->shared) { | |
1352 | /* | |
1353 | * We are serialised from CPU_DEAD or | |
1354 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1355 | */ | |
1356 | l3->shared = shared; | |
1357 | shared = NULL; | |
1358 | } | |
4484ebf1 | 1359 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1360 | if (!l3->alien) { |
1361 | l3->alien = alien; | |
1362 | alien = NULL; | |
1da177e4 | 1363 | } |
fbf1e473 AM |
1364 | #endif |
1365 | spin_unlock_irq(&l3->list_lock); | |
1366 | kfree(shared); | |
1367 | free_alien_cache(alien); | |
83835b3d PZ |
1368 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1369 | slab_set_debugobj_lock_classes_node(cachep, node); | |
6ccfb5bc GC |
1370 | else if (!OFF_SLAB(cachep) && |
1371 | !(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
1372 | on_slab_lock_classes_node(cachep, node); | |
fbf1e473 | 1373 | } |
ce79ddc8 PE |
1374 | init_node_lock_keys(node); |
1375 | ||
fbf1e473 AM |
1376 | return 0; |
1377 | bad: | |
12d00f6a | 1378 | cpuup_canceled(cpu); |
fbf1e473 AM |
1379 | return -ENOMEM; |
1380 | } | |
1381 | ||
1382 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1383 | unsigned long action, void *hcpu) | |
1384 | { | |
1385 | long cpu = (long)hcpu; | |
1386 | int err = 0; | |
1387 | ||
1388 | switch (action) { | |
fbf1e473 AM |
1389 | case CPU_UP_PREPARE: |
1390 | case CPU_UP_PREPARE_FROZEN: | |
18004c5d | 1391 | mutex_lock(&slab_mutex); |
fbf1e473 | 1392 | err = cpuup_prepare(cpu); |
18004c5d | 1393 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
1394 | break; |
1395 | case CPU_ONLINE: | |
8bb78442 | 1396 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1397 | start_cpu_timer(cpu); |
1398 | break; | |
1399 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1400 | case CPU_DOWN_PREPARE: |
8bb78442 | 1401 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 | 1402 | /* |
18004c5d | 1403 | * Shutdown cache reaper. Note that the slab_mutex is |
5830c590 CL |
1404 | * held so that if cache_reap() is invoked it cannot do |
1405 | * anything expensive but will only modify reap_work | |
1406 | * and reschedule the timer. | |
1407 | */ | |
afe2c511 | 1408 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1409 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1410 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1411 | break; |
1412 | case CPU_DOWN_FAILED: | |
8bb78442 | 1413 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1414 | start_cpu_timer(cpu); |
1415 | break; | |
1da177e4 | 1416 | case CPU_DEAD: |
8bb78442 | 1417 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1418 | /* |
1419 | * Even if all the cpus of a node are down, we don't free the | |
1420 | * kmem_list3 of any cache. This to avoid a race between | |
1421 | * cpu_down, and a kmalloc allocation from another cpu for | |
1422 | * memory from the node of the cpu going down. The list3 | |
1423 | * structure is usually allocated from kmem_cache_create() and | |
1424 | * gets destroyed at kmem_cache_destroy(). | |
1425 | */ | |
183ff22b | 1426 | /* fall through */ |
8f5be20b | 1427 | #endif |
1da177e4 | 1428 | case CPU_UP_CANCELED: |
8bb78442 | 1429 | case CPU_UP_CANCELED_FROZEN: |
18004c5d | 1430 | mutex_lock(&slab_mutex); |
fbf1e473 | 1431 | cpuup_canceled(cpu); |
18004c5d | 1432 | mutex_unlock(&slab_mutex); |
1da177e4 | 1433 | break; |
1da177e4 | 1434 | } |
eac40680 | 1435 | return notifier_from_errno(err); |
1da177e4 LT |
1436 | } |
1437 | ||
74b85f37 CS |
1438 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1439 | &cpuup_callback, NULL, 0 | |
1440 | }; | |
1da177e4 | 1441 | |
8f9f8d9e DR |
1442 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1443 | /* | |
1444 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1445 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1446 | * removed. | |
1447 | * | |
18004c5d | 1448 | * Must hold slab_mutex. |
8f9f8d9e | 1449 | */ |
6a67368c | 1450 | static int __meminit drain_cache_node_node(int node) |
8f9f8d9e DR |
1451 | { |
1452 | struct kmem_cache *cachep; | |
1453 | int ret = 0; | |
1454 | ||
18004c5d | 1455 | list_for_each_entry(cachep, &slab_caches, list) { |
6744f087 | 1456 | struct kmem_cache_node *l3; |
8f9f8d9e | 1457 | |
6a67368c | 1458 | l3 = cachep->node[node]; |
8f9f8d9e DR |
1459 | if (!l3) |
1460 | continue; | |
1461 | ||
1462 | drain_freelist(cachep, l3, l3->free_objects); | |
1463 | ||
1464 | if (!list_empty(&l3->slabs_full) || | |
1465 | !list_empty(&l3->slabs_partial)) { | |
1466 | ret = -EBUSY; | |
1467 | break; | |
1468 | } | |
1469 | } | |
1470 | return ret; | |
1471 | } | |
1472 | ||
1473 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1474 | unsigned long action, void *arg) | |
1475 | { | |
1476 | struct memory_notify *mnb = arg; | |
1477 | int ret = 0; | |
1478 | int nid; | |
1479 | ||
1480 | nid = mnb->status_change_nid; | |
1481 | if (nid < 0) | |
1482 | goto out; | |
1483 | ||
1484 | switch (action) { | |
1485 | case MEM_GOING_ONLINE: | |
18004c5d | 1486 | mutex_lock(&slab_mutex); |
6a67368c | 1487 | ret = init_cache_node_node(nid); |
18004c5d | 1488 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1489 | break; |
1490 | case MEM_GOING_OFFLINE: | |
18004c5d | 1491 | mutex_lock(&slab_mutex); |
6a67368c | 1492 | ret = drain_cache_node_node(nid); |
18004c5d | 1493 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1494 | break; |
1495 | case MEM_ONLINE: | |
1496 | case MEM_OFFLINE: | |
1497 | case MEM_CANCEL_ONLINE: | |
1498 | case MEM_CANCEL_OFFLINE: | |
1499 | break; | |
1500 | } | |
1501 | out: | |
5fda1bd5 | 1502 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1503 | } |
1504 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1505 | ||
e498be7d CL |
1506 | /* |
1507 | * swap the static kmem_list3 with kmalloced memory | |
1508 | */ | |
6744f087 | 1509 | static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list, |
8f9f8d9e | 1510 | int nodeid) |
e498be7d | 1511 | { |
6744f087 | 1512 | struct kmem_cache_node *ptr; |
e498be7d | 1513 | |
6744f087 | 1514 | ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid); |
e498be7d CL |
1515 | BUG_ON(!ptr); |
1516 | ||
6744f087 | 1517 | memcpy(ptr, list, sizeof(struct kmem_cache_node)); |
2b2d5493 IM |
1518 | /* |
1519 | * Do not assume that spinlocks can be initialized via memcpy: | |
1520 | */ | |
1521 | spin_lock_init(&ptr->list_lock); | |
1522 | ||
e498be7d | 1523 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
6a67368c | 1524 | cachep->node[nodeid] = ptr; |
e498be7d CL |
1525 | } |
1526 | ||
556a169d PE |
1527 | /* |
1528 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1529 | * size of kmem_list3. | |
1530 | */ | |
1531 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1532 | { | |
1533 | int node; | |
1534 | ||
1535 | for_each_online_node(node) { | |
6a67368c CL |
1536 | cachep->node[node] = &initkmem_list3[index + node]; |
1537 | cachep->node[node]->next_reap = jiffies + | |
556a169d PE |
1538 | REAPTIMEOUT_LIST3 + |
1539 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1540 | } | |
1541 | } | |
1542 | ||
3c583465 CL |
1543 | /* |
1544 | * The memory after the last cpu cache pointer is used for the | |
6a67368c | 1545 | * the node pointer. |
3c583465 | 1546 | */ |
6a67368c | 1547 | static void setup_node_pointer(struct kmem_cache *cachep) |
3c583465 | 1548 | { |
6a67368c | 1549 | cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids]; |
3c583465 CL |
1550 | } |
1551 | ||
a737b3e2 AM |
1552 | /* |
1553 | * Initialisation. Called after the page allocator have been initialised and | |
1554 | * before smp_init(). | |
1da177e4 LT |
1555 | */ |
1556 | void __init kmem_cache_init(void) | |
1557 | { | |
e498be7d CL |
1558 | int i; |
1559 | ||
9b030cb8 | 1560 | kmem_cache = &kmem_cache_boot; |
6a67368c | 1561 | setup_node_pointer(kmem_cache); |
9b030cb8 | 1562 | |
b6e68bc1 | 1563 | if (num_possible_nodes() == 1) |
62918a03 SS |
1564 | use_alien_caches = 0; |
1565 | ||
3c583465 | 1566 | for (i = 0; i < NUM_INIT_LISTS; i++) |
e498be7d | 1567 | kmem_list3_init(&initkmem_list3[i]); |
3c583465 | 1568 | |
9b030cb8 | 1569 | set_up_list3s(kmem_cache, CACHE_CACHE); |
1da177e4 LT |
1570 | |
1571 | /* | |
1572 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1573 | * page orders on machines with more than 32MB of memory if |
1574 | * not overridden on the command line. | |
1da177e4 | 1575 | */ |
3df1cccd | 1576 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1577 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1578 | |
1da177e4 LT |
1579 | /* Bootstrap is tricky, because several objects are allocated |
1580 | * from caches that do not exist yet: | |
9b030cb8 CL |
1581 | * 1) initialize the kmem_cache cache: it contains the struct |
1582 | * kmem_cache structures of all caches, except kmem_cache itself: | |
1583 | * kmem_cache is statically allocated. | |
e498be7d CL |
1584 | * Initially an __init data area is used for the head array and the |
1585 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1586 | * array at the end of the bootstrap. | |
1da177e4 | 1587 | * 2) Create the first kmalloc cache. |
343e0d7a | 1588 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1589 | * An __init data area is used for the head array. |
1590 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1591 | * head arrays. | |
9b030cb8 | 1592 | * 4) Replace the __init data head arrays for kmem_cache and the first |
1da177e4 | 1593 | * kmalloc cache with kmalloc allocated arrays. |
9b030cb8 | 1594 | * 5) Replace the __init data for kmem_list3 for kmem_cache and |
e498be7d CL |
1595 | * the other cache's with kmalloc allocated memory. |
1596 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1597 | */ |
1598 | ||
9b030cb8 | 1599 | /* 1) create the kmem_cache */ |
1da177e4 | 1600 | |
8da3430d | 1601 | /* |
b56efcf0 | 1602 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1603 | */ |
2f9baa9f CL |
1604 | create_boot_cache(kmem_cache, "kmem_cache", |
1605 | offsetof(struct kmem_cache, array[nr_cpu_ids]) + | |
6744f087 | 1606 | nr_node_ids * sizeof(struct kmem_cache_node *), |
2f9baa9f CL |
1607 | SLAB_HWCACHE_ALIGN); |
1608 | list_add(&kmem_cache->list, &slab_caches); | |
1da177e4 LT |
1609 | |
1610 | /* 2+3) create the kmalloc caches */ | |
1da177e4 | 1611 | |
a737b3e2 AM |
1612 | /* |
1613 | * Initialize the caches that provide memory for the array cache and the | |
1614 | * kmem_list3 structures first. Without this, further allocations will | |
1615 | * bug. | |
e498be7d CL |
1616 | */ |
1617 | ||
e3366016 CL |
1618 | kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac", |
1619 | kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS); | |
45530c44 CL |
1620 | |
1621 | if (INDEX_AC != INDEX_L3) | |
e3366016 CL |
1622 | kmalloc_caches[INDEX_L3] = |
1623 | create_kmalloc_cache("kmalloc-l3", | |
1624 | kmalloc_size(INDEX_L3), ARCH_KMALLOC_FLAGS); | |
e498be7d | 1625 | |
e0a42726 IM |
1626 | slab_early_init = 0; |
1627 | ||
e3366016 CL |
1628 | for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) { |
1629 | size_t cs_size = kmalloc_size(i); | |
1630 | ||
1631 | if (cs_size < KMALLOC_MIN_SIZE) | |
1632 | continue; | |
1633 | ||
1634 | if (!kmalloc_caches[i]) { | |
1635 | /* | |
1636 | * For performance, all the general caches are L1 aligned. | |
1637 | * This should be particularly beneficial on SMP boxes, as it | |
1638 | * eliminates "false sharing". | |
1639 | * Note for systems short on memory removing the alignment will | |
1640 | * allow tighter packing of the smaller caches. | |
1641 | */ | |
1642 | kmalloc_caches[i] = create_kmalloc_cache("kmalloc", | |
1643 | cs_size, ARCH_KMALLOC_FLAGS); | |
1644 | } | |
45530c44 | 1645 | |
4b51d669 | 1646 | #ifdef CONFIG_ZONE_DMA |
e3366016 CL |
1647 | kmalloc_dma_caches[i] = create_kmalloc_cache( |
1648 | "kmalloc-dma", cs_size, | |
45530c44 | 1649 | SLAB_CACHE_DMA|ARCH_KMALLOC_FLAGS); |
4b51d669 | 1650 | #endif |
1da177e4 LT |
1651 | } |
1652 | /* 4) Replace the bootstrap head arrays */ | |
1653 | { | |
2b2d5493 | 1654 | struct array_cache *ptr; |
e498be7d | 1655 | |
83b519e8 | 1656 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1657 | |
9b030cb8 | 1658 | memcpy(ptr, cpu_cache_get(kmem_cache), |
b28a02de | 1659 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1660 | /* |
1661 | * Do not assume that spinlocks can be initialized via memcpy: | |
1662 | */ | |
1663 | spin_lock_init(&ptr->lock); | |
1664 | ||
9b030cb8 | 1665 | kmem_cache->array[smp_processor_id()] = ptr; |
e498be7d | 1666 | |
83b519e8 | 1667 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1668 | |
e3366016 | 1669 | BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC]) |
b28a02de | 1670 | != &initarray_generic.cache); |
e3366016 | 1671 | memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]), |
b28a02de | 1672 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1673 | /* |
1674 | * Do not assume that spinlocks can be initialized via memcpy: | |
1675 | */ | |
1676 | spin_lock_init(&ptr->lock); | |
1677 | ||
e3366016 | 1678 | kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr; |
1da177e4 | 1679 | } |
e498be7d CL |
1680 | /* 5) Replace the bootstrap kmem_list3's */ |
1681 | { | |
1ca4cb24 PE |
1682 | int nid; |
1683 | ||
9c09a95c | 1684 | for_each_online_node(nid) { |
9b030cb8 | 1685 | init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1686 | |
e3366016 | 1687 | init_list(kmalloc_caches[INDEX_AC], |
1ca4cb24 | 1688 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1689 | |
1690 | if (INDEX_AC != INDEX_L3) { | |
e3366016 | 1691 | init_list(kmalloc_caches[INDEX_L3], |
1ca4cb24 | 1692 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1693 | } |
1694 | } | |
1695 | } | |
1da177e4 | 1696 | |
97d06609 | 1697 | slab_state = UP; |
e3366016 CL |
1698 | |
1699 | /* Create the proper names */ | |
1700 | for (i = 1; i < PAGE_SHIFT + MAX_ORDER; i++) { | |
1701 | char *s; | |
1702 | struct kmem_cache *c = kmalloc_caches[i]; | |
1703 | ||
1704 | if (!c) | |
1705 | continue; | |
1706 | ||
1707 | s = kasprintf(GFP_NOWAIT, "kmalloc-%d", kmalloc_size(i)); | |
1708 | ||
1709 | BUG_ON(!s); | |
1710 | c->name = s; | |
1711 | ||
1712 | #ifdef CONFIG_ZONE_DMA | |
1713 | c = kmalloc_dma_caches[i]; | |
1714 | BUG_ON(!c); | |
1715 | s = kasprintf(GFP_NOWAIT, "dma-kmalloc-%d", kmalloc_size(i)); | |
1716 | BUG_ON(!s); | |
1717 | c->name = s; | |
1718 | #endif | |
1719 | } | |
8429db5c PE |
1720 | } |
1721 | ||
1722 | void __init kmem_cache_init_late(void) | |
1723 | { | |
1724 | struct kmem_cache *cachep; | |
1725 | ||
97d06609 | 1726 | slab_state = UP; |
52cef189 | 1727 | |
8429db5c | 1728 | /* 6) resize the head arrays to their final sizes */ |
18004c5d CL |
1729 | mutex_lock(&slab_mutex); |
1730 | list_for_each_entry(cachep, &slab_caches, list) | |
8429db5c PE |
1731 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1732 | BUG(); | |
18004c5d | 1733 | mutex_unlock(&slab_mutex); |
056c6241 | 1734 | |
947ca185 MW |
1735 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1736 | init_lock_keys(); | |
1737 | ||
97d06609 CL |
1738 | /* Done! */ |
1739 | slab_state = FULL; | |
1740 | ||
a737b3e2 AM |
1741 | /* |
1742 | * Register a cpu startup notifier callback that initializes | |
1743 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1744 | */ |
1745 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1746 | |
8f9f8d9e DR |
1747 | #ifdef CONFIG_NUMA |
1748 | /* | |
1749 | * Register a memory hotplug callback that initializes and frees | |
6a67368c | 1750 | * node. |
8f9f8d9e DR |
1751 | */ |
1752 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1753 | #endif | |
1754 | ||
a737b3e2 AM |
1755 | /* |
1756 | * The reap timers are started later, with a module init call: That part | |
1757 | * of the kernel is not yet operational. | |
1da177e4 LT |
1758 | */ |
1759 | } | |
1760 | ||
1761 | static int __init cpucache_init(void) | |
1762 | { | |
1763 | int cpu; | |
1764 | ||
a737b3e2 AM |
1765 | /* |
1766 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1767 | */ |
e498be7d | 1768 | for_each_online_cpu(cpu) |
a737b3e2 | 1769 | start_cpu_timer(cpu); |
a164f896 GC |
1770 | |
1771 | /* Done! */ | |
97d06609 | 1772 | slab_state = FULL; |
1da177e4 LT |
1773 | return 0; |
1774 | } | |
1da177e4 LT |
1775 | __initcall(cpucache_init); |
1776 | ||
8bdec192 RA |
1777 | static noinline void |
1778 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1779 | { | |
6744f087 | 1780 | struct kmem_cache_node *l3; |
8bdec192 RA |
1781 | struct slab *slabp; |
1782 | unsigned long flags; | |
1783 | int node; | |
1784 | ||
1785 | printk(KERN_WARNING | |
1786 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1787 | nodeid, gfpflags); | |
1788 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1789 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1790 | |
1791 | for_each_online_node(node) { | |
1792 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1793 | unsigned long active_slabs = 0, num_slabs = 0; | |
1794 | ||
6a67368c | 1795 | l3 = cachep->node[node]; |
8bdec192 RA |
1796 | if (!l3) |
1797 | continue; | |
1798 | ||
1799 | spin_lock_irqsave(&l3->list_lock, flags); | |
1800 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1801 | active_objs += cachep->num; | |
1802 | active_slabs++; | |
1803 | } | |
1804 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1805 | active_objs += slabp->inuse; | |
1806 | active_slabs++; | |
1807 | } | |
1808 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1809 | num_slabs++; | |
1810 | ||
1811 | free_objects += l3->free_objects; | |
1812 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1813 | ||
1814 | num_slabs += active_slabs; | |
1815 | num_objs = num_slabs * cachep->num; | |
1816 | printk(KERN_WARNING | |
1817 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1818 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1819 | free_objects); | |
1820 | } | |
1821 | } | |
1822 | ||
1da177e4 LT |
1823 | /* |
1824 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1825 | * | |
1826 | * If we requested dmaable memory, we will get it. Even if we | |
1827 | * did not request dmaable memory, we might get it, but that | |
1828 | * would be relatively rare and ignorable. | |
1829 | */ | |
343e0d7a | 1830 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1831 | { |
1832 | struct page *page; | |
e1b6aa6f | 1833 | int nr_pages; |
1da177e4 LT |
1834 | int i; |
1835 | ||
d6fef9da | 1836 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1837 | /* |
1838 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1839 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1840 | */ |
e1b6aa6f | 1841 | flags |= __GFP_COMP; |
d6fef9da | 1842 | #endif |
765c4507 | 1843 | |
a618e89f | 1844 | flags |= cachep->allocflags; |
e12ba74d MG |
1845 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1846 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1847 | |
517d0869 | 1848 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1849 | if (!page) { |
1850 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1851 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1852 | return NULL; |
8bdec192 | 1853 | } |
1da177e4 | 1854 | |
b37f1dd0 | 1855 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
072bb0aa MG |
1856 | if (unlikely(page->pfmemalloc)) |
1857 | pfmemalloc_active = true; | |
1858 | ||
e1b6aa6f | 1859 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1860 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1861 | add_zone_page_state(page_zone(page), |
1862 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1863 | else | |
1864 | add_zone_page_state(page_zone(page), | |
1865 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
072bb0aa | 1866 | for (i = 0; i < nr_pages; i++) { |
e1b6aa6f | 1867 | __SetPageSlab(page + i); |
c175eea4 | 1868 | |
072bb0aa MG |
1869 | if (page->pfmemalloc) |
1870 | SetPageSlabPfmemalloc(page + i); | |
1871 | } | |
1f458cbf | 1872 | memcg_bind_pages(cachep, cachep->gfporder); |
072bb0aa | 1873 | |
b1eeab67 VN |
1874 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1875 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1876 | ||
1877 | if (cachep->ctor) | |
1878 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1879 | else | |
1880 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1881 | } | |
c175eea4 | 1882 | |
e1b6aa6f | 1883 | return page_address(page); |
1da177e4 LT |
1884 | } |
1885 | ||
1886 | /* | |
1887 | * Interface to system's page release. | |
1888 | */ | |
343e0d7a | 1889 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1890 | { |
b28a02de | 1891 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1892 | struct page *page = virt_to_page(addr); |
1893 | const unsigned long nr_freed = i; | |
1894 | ||
b1eeab67 | 1895 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1896 | |
972d1a7b CL |
1897 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1898 | sub_zone_page_state(page_zone(page), | |
1899 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1900 | else | |
1901 | sub_zone_page_state(page_zone(page), | |
1902 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1903 | while (i--) { |
f205b2fe | 1904 | BUG_ON(!PageSlab(page)); |
072bb0aa | 1905 | __ClearPageSlabPfmemalloc(page); |
f205b2fe | 1906 | __ClearPageSlab(page); |
1da177e4 LT |
1907 | page++; |
1908 | } | |
1f458cbf GC |
1909 | |
1910 | memcg_release_pages(cachep, cachep->gfporder); | |
1da177e4 LT |
1911 | if (current->reclaim_state) |
1912 | current->reclaim_state->reclaimed_slab += nr_freed; | |
d79923fa | 1913 | free_memcg_kmem_pages((unsigned long)addr, cachep->gfporder); |
1da177e4 LT |
1914 | } |
1915 | ||
1916 | static void kmem_rcu_free(struct rcu_head *head) | |
1917 | { | |
b28a02de | 1918 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1919 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1920 | |
1921 | kmem_freepages(cachep, slab_rcu->addr); | |
1922 | if (OFF_SLAB(cachep)) | |
1923 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1924 | } | |
1925 | ||
1926 | #if DEBUG | |
1927 | ||
1928 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1929 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1930 | unsigned long caller) |
1da177e4 | 1931 | { |
8c138bc0 | 1932 | int size = cachep->object_size; |
1da177e4 | 1933 | |
3dafccf2 | 1934 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1935 | |
b28a02de | 1936 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1937 | return; |
1938 | ||
b28a02de PE |
1939 | *addr++ = 0x12345678; |
1940 | *addr++ = caller; | |
1941 | *addr++ = smp_processor_id(); | |
1942 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1943 | { |
1944 | unsigned long *sptr = &caller; | |
1945 | unsigned long svalue; | |
1946 | ||
1947 | while (!kstack_end(sptr)) { | |
1948 | svalue = *sptr++; | |
1949 | if (kernel_text_address(svalue)) { | |
b28a02de | 1950 | *addr++ = svalue; |
1da177e4 LT |
1951 | size -= sizeof(unsigned long); |
1952 | if (size <= sizeof(unsigned long)) | |
1953 | break; | |
1954 | } | |
1955 | } | |
1956 | ||
1957 | } | |
b28a02de | 1958 | *addr++ = 0x87654321; |
1da177e4 LT |
1959 | } |
1960 | #endif | |
1961 | ||
343e0d7a | 1962 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 1963 | { |
8c138bc0 | 1964 | int size = cachep->object_size; |
3dafccf2 | 1965 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
1966 | |
1967 | memset(addr, val, size); | |
b28a02de | 1968 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
1969 | } |
1970 | ||
1971 | static void dump_line(char *data, int offset, int limit) | |
1972 | { | |
1973 | int i; | |
aa83aa40 DJ |
1974 | unsigned char error = 0; |
1975 | int bad_count = 0; | |
1976 | ||
fdde6abb | 1977 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
1978 | for (i = 0; i < limit; i++) { |
1979 | if (data[offset + i] != POISON_FREE) { | |
1980 | error = data[offset + i]; | |
1981 | bad_count++; | |
1982 | } | |
aa83aa40 | 1983 | } |
fdde6abb SAS |
1984 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1985 | &data[offset], limit, 1); | |
aa83aa40 DJ |
1986 | |
1987 | if (bad_count == 1) { | |
1988 | error ^= POISON_FREE; | |
1989 | if (!(error & (error - 1))) { | |
1990 | printk(KERN_ERR "Single bit error detected. Probably " | |
1991 | "bad RAM.\n"); | |
1992 | #ifdef CONFIG_X86 | |
1993 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
1994 | "test tool.\n"); | |
1995 | #else | |
1996 | printk(KERN_ERR "Run a memory test tool.\n"); | |
1997 | #endif | |
1998 | } | |
1999 | } | |
1da177e4 LT |
2000 | } |
2001 | #endif | |
2002 | ||
2003 | #if DEBUG | |
2004 | ||
343e0d7a | 2005 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
2006 | { |
2007 | int i, size; | |
2008 | char *realobj; | |
2009 | ||
2010 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 2011 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
2012 | *dbg_redzone1(cachep, objp), |
2013 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2014 | } |
2015 | ||
2016 | if (cachep->flags & SLAB_STORE_USER) { | |
2017 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 2018 | *dbg_userword(cachep, objp)); |
1da177e4 | 2019 | print_symbol("(%s)", |
a737b3e2 | 2020 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
2021 | printk("\n"); |
2022 | } | |
3dafccf2 | 2023 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2024 | size = cachep->object_size; |
b28a02de | 2025 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
2026 | int limit; |
2027 | limit = 16; | |
b28a02de PE |
2028 | if (i + limit > size) |
2029 | limit = size - i; | |
1da177e4 LT |
2030 | dump_line(realobj, i, limit); |
2031 | } | |
2032 | } | |
2033 | ||
343e0d7a | 2034 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
2035 | { |
2036 | char *realobj; | |
2037 | int size, i; | |
2038 | int lines = 0; | |
2039 | ||
3dafccf2 | 2040 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2041 | size = cachep->object_size; |
1da177e4 | 2042 | |
b28a02de | 2043 | for (i = 0; i < size; i++) { |
1da177e4 | 2044 | char exp = POISON_FREE; |
b28a02de | 2045 | if (i == size - 1) |
1da177e4 LT |
2046 | exp = POISON_END; |
2047 | if (realobj[i] != exp) { | |
2048 | int limit; | |
2049 | /* Mismatch ! */ | |
2050 | /* Print header */ | |
2051 | if (lines == 0) { | |
b28a02de | 2052 | printk(KERN_ERR |
face37f5 DJ |
2053 | "Slab corruption (%s): %s start=%p, len=%d\n", |
2054 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
2055 | print_objinfo(cachep, objp, 0); |
2056 | } | |
2057 | /* Hexdump the affected line */ | |
b28a02de | 2058 | i = (i / 16) * 16; |
1da177e4 | 2059 | limit = 16; |
b28a02de PE |
2060 | if (i + limit > size) |
2061 | limit = size - i; | |
1da177e4 LT |
2062 | dump_line(realobj, i, limit); |
2063 | i += 16; | |
2064 | lines++; | |
2065 | /* Limit to 5 lines */ | |
2066 | if (lines > 5) | |
2067 | break; | |
2068 | } | |
2069 | } | |
2070 | if (lines != 0) { | |
2071 | /* Print some data about the neighboring objects, if they | |
2072 | * exist: | |
2073 | */ | |
6ed5eb22 | 2074 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 2075 | unsigned int objnr; |
1da177e4 | 2076 | |
8fea4e96 | 2077 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2078 | if (objnr) { |
8fea4e96 | 2079 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2080 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2081 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2082 | realobj, size); |
1da177e4 LT |
2083 | print_objinfo(cachep, objp, 2); |
2084 | } | |
b28a02de | 2085 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2086 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2087 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2088 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2089 | realobj, size); |
1da177e4 LT |
2090 | print_objinfo(cachep, objp, 2); |
2091 | } | |
2092 | } | |
2093 | } | |
2094 | #endif | |
2095 | ||
12dd36fa | 2096 | #if DEBUG |
e79aec29 | 2097 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2098 | { |
1da177e4 LT |
2099 | int i; |
2100 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2101 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2102 | |
2103 | if (cachep->flags & SLAB_POISON) { | |
2104 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2105 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2106 | OFF_SLAB(cachep)) |
b28a02de | 2107 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2108 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2109 | else |
2110 | check_poison_obj(cachep, objp); | |
2111 | #else | |
2112 | check_poison_obj(cachep, objp); | |
2113 | #endif | |
2114 | } | |
2115 | if (cachep->flags & SLAB_RED_ZONE) { | |
2116 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2117 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2118 | "was overwritten"); |
1da177e4 LT |
2119 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2120 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2121 | "was overwritten"); |
1da177e4 | 2122 | } |
1da177e4 | 2123 | } |
12dd36fa | 2124 | } |
1da177e4 | 2125 | #else |
e79aec29 | 2126 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2127 | { |
12dd36fa | 2128 | } |
1da177e4 LT |
2129 | #endif |
2130 | ||
911851e6 RD |
2131 | /** |
2132 | * slab_destroy - destroy and release all objects in a slab | |
2133 | * @cachep: cache pointer being destroyed | |
2134 | * @slabp: slab pointer being destroyed | |
2135 | * | |
12dd36fa | 2136 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2137 | * Before calling the slab must have been unlinked from the cache. The |
2138 | * cache-lock is not held/needed. | |
12dd36fa | 2139 | */ |
343e0d7a | 2140 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2141 | { |
2142 | void *addr = slabp->s_mem - slabp->colouroff; | |
2143 | ||
e79aec29 | 2144 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2145 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2146 | struct slab_rcu *slab_rcu; | |
2147 | ||
b28a02de | 2148 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2149 | slab_rcu->cachep = cachep; |
2150 | slab_rcu->addr = addr; | |
2151 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2152 | } else { | |
2153 | kmem_freepages(cachep, addr); | |
873623df IM |
2154 | if (OFF_SLAB(cachep)) |
2155 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2156 | } |
2157 | } | |
2158 | ||
4d268eba | 2159 | /** |
a70773dd RD |
2160 | * calculate_slab_order - calculate size (page order) of slabs |
2161 | * @cachep: pointer to the cache that is being created | |
2162 | * @size: size of objects to be created in this cache. | |
2163 | * @align: required alignment for the objects. | |
2164 | * @flags: slab allocation flags | |
2165 | * | |
2166 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2167 | * |
2168 | * This could be made much more intelligent. For now, try to avoid using | |
2169 | * high order pages for slabs. When the gfp() functions are more friendly | |
2170 | * towards high-order requests, this should be changed. | |
2171 | */ | |
a737b3e2 | 2172 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2173 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2174 | { |
b1ab41c4 | 2175 | unsigned long offslab_limit; |
4d268eba | 2176 | size_t left_over = 0; |
9888e6fa | 2177 | int gfporder; |
4d268eba | 2178 | |
0aa817f0 | 2179 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2180 | unsigned int num; |
2181 | size_t remainder; | |
2182 | ||
9888e6fa | 2183 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2184 | if (!num) |
2185 | continue; | |
9888e6fa | 2186 | |
b1ab41c4 IM |
2187 | if (flags & CFLGS_OFF_SLAB) { |
2188 | /* | |
2189 | * Max number of objs-per-slab for caches which | |
2190 | * use off-slab slabs. Needed to avoid a possible | |
2191 | * looping condition in cache_grow(). | |
2192 | */ | |
2193 | offslab_limit = size - sizeof(struct slab); | |
2194 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2195 | ||
2196 | if (num > offslab_limit) | |
2197 | break; | |
2198 | } | |
4d268eba | 2199 | |
9888e6fa | 2200 | /* Found something acceptable - save it away */ |
4d268eba | 2201 | cachep->num = num; |
9888e6fa | 2202 | cachep->gfporder = gfporder; |
4d268eba PE |
2203 | left_over = remainder; |
2204 | ||
f78bb8ad LT |
2205 | /* |
2206 | * A VFS-reclaimable slab tends to have most allocations | |
2207 | * as GFP_NOFS and we really don't want to have to be allocating | |
2208 | * higher-order pages when we are unable to shrink dcache. | |
2209 | */ | |
2210 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2211 | break; | |
2212 | ||
4d268eba PE |
2213 | /* |
2214 | * Large number of objects is good, but very large slabs are | |
2215 | * currently bad for the gfp()s. | |
2216 | */ | |
543585cc | 2217 | if (gfporder >= slab_max_order) |
4d268eba PE |
2218 | break; |
2219 | ||
9888e6fa LT |
2220 | /* |
2221 | * Acceptable internal fragmentation? | |
2222 | */ | |
a737b3e2 | 2223 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2224 | break; |
2225 | } | |
2226 | return left_over; | |
2227 | } | |
2228 | ||
83b519e8 | 2229 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2230 | { |
97d06609 | 2231 | if (slab_state >= FULL) |
83b519e8 | 2232 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2233 | |
97d06609 | 2234 | if (slab_state == DOWN) { |
f30cf7d1 | 2235 | /* |
2f9baa9f CL |
2236 | * Note: Creation of first cache (kmem_cache). |
2237 | * The setup_list3s is taken care | |
2238 | * of by the caller of __kmem_cache_create | |
2239 | */ | |
2240 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2241 | slab_state = PARTIAL; | |
2242 | } else if (slab_state == PARTIAL) { | |
2243 | /* | |
2244 | * Note: the second kmem_cache_create must create the cache | |
f30cf7d1 PE |
2245 | * that's used by kmalloc(24), otherwise the creation of |
2246 | * further caches will BUG(). | |
2247 | */ | |
2248 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2249 | ||
2250 | /* | |
2251 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2f9baa9f | 2252 | * the second cache, then we need to set up all its list3s, |
f30cf7d1 PE |
2253 | * otherwise the creation of further caches will BUG(). |
2254 | */ | |
2255 | set_up_list3s(cachep, SIZE_AC); | |
2256 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2257 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2258 | else |
97d06609 | 2259 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 | 2260 | } else { |
2f9baa9f | 2261 | /* Remaining boot caches */ |
f30cf7d1 | 2262 | cachep->array[smp_processor_id()] = |
83b519e8 | 2263 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2264 | |
97d06609 | 2265 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2266 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2267 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2268 | } else { |
2269 | int node; | |
556a169d | 2270 | for_each_online_node(node) { |
6a67368c | 2271 | cachep->node[node] = |
6744f087 | 2272 | kmalloc_node(sizeof(struct kmem_cache_node), |
eb91f1d0 | 2273 | gfp, node); |
6a67368c CL |
2274 | BUG_ON(!cachep->node[node]); |
2275 | kmem_list3_init(cachep->node[node]); | |
f30cf7d1 PE |
2276 | } |
2277 | } | |
2278 | } | |
6a67368c | 2279 | cachep->node[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2280 | jiffies + REAPTIMEOUT_LIST3 + |
2281 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2282 | ||
2283 | cpu_cache_get(cachep)->avail = 0; | |
2284 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2285 | cpu_cache_get(cachep)->batchcount = 1; | |
2286 | cpu_cache_get(cachep)->touched = 0; | |
2287 | cachep->batchcount = 1; | |
2288 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2289 | return 0; |
f30cf7d1 PE |
2290 | } |
2291 | ||
1da177e4 | 2292 | /** |
039363f3 | 2293 | * __kmem_cache_create - Create a cache. |
a755b76a | 2294 | * @cachep: cache management descriptor |
1da177e4 | 2295 | * @flags: SLAB flags |
1da177e4 LT |
2296 | * |
2297 | * Returns a ptr to the cache on success, NULL on failure. | |
2298 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2299 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 | 2300 | * |
1da177e4 LT |
2301 | * The flags are |
2302 | * | |
2303 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2304 | * to catch references to uninitialised memory. | |
2305 | * | |
2306 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2307 | * for buffer overruns. | |
2308 | * | |
1da177e4 LT |
2309 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2310 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2311 | * as davem. | |
2312 | */ | |
278b1bb1 | 2313 | int |
8a13a4cc | 2314 | __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) |
1da177e4 LT |
2315 | { |
2316 | size_t left_over, slab_size, ralign; | |
83b519e8 | 2317 | gfp_t gfp; |
278b1bb1 | 2318 | int err; |
8a13a4cc | 2319 | size_t size = cachep->size; |
1da177e4 | 2320 | |
1da177e4 | 2321 | #if DEBUG |
1da177e4 LT |
2322 | #if FORCED_DEBUG |
2323 | /* | |
2324 | * Enable redzoning and last user accounting, except for caches with | |
2325 | * large objects, if the increased size would increase the object size | |
2326 | * above the next power of two: caches with object sizes just above a | |
2327 | * power of two have a significant amount of internal fragmentation. | |
2328 | */ | |
87a927c7 DW |
2329 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2330 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2331 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2332 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2333 | flags |= SLAB_POISON; | |
2334 | #endif | |
2335 | if (flags & SLAB_DESTROY_BY_RCU) | |
2336 | BUG_ON(flags & SLAB_POISON); | |
2337 | #endif | |
1da177e4 | 2338 | |
a737b3e2 AM |
2339 | /* |
2340 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2341 | * unaligned accesses for some archs when redzoning is used, and makes |
2342 | * sure any on-slab bufctl's are also correctly aligned. | |
2343 | */ | |
b28a02de PE |
2344 | if (size & (BYTES_PER_WORD - 1)) { |
2345 | size += (BYTES_PER_WORD - 1); | |
2346 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2347 | } |
2348 | ||
ca5f9703 | 2349 | /* |
87a927c7 DW |
2350 | * Redzoning and user store require word alignment or possibly larger. |
2351 | * Note this will be overridden by architecture or caller mandated | |
2352 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2353 | */ |
87a927c7 DW |
2354 | if (flags & SLAB_STORE_USER) |
2355 | ralign = BYTES_PER_WORD; | |
2356 | ||
2357 | if (flags & SLAB_RED_ZONE) { | |
2358 | ralign = REDZONE_ALIGN; | |
2359 | /* If redzoning, ensure that the second redzone is suitably | |
2360 | * aligned, by adjusting the object size accordingly. */ | |
2361 | size += REDZONE_ALIGN - 1; | |
2362 | size &= ~(REDZONE_ALIGN - 1); | |
2363 | } | |
ca5f9703 | 2364 | |
a44b56d3 | 2365 | /* 3) caller mandated alignment */ |
8a13a4cc CL |
2366 | if (ralign < cachep->align) { |
2367 | ralign = cachep->align; | |
1da177e4 | 2368 | } |
3ff84a7f PE |
2369 | /* disable debug if necessary */ |
2370 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2371 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2372 | /* |
ca5f9703 | 2373 | * 4) Store it. |
1da177e4 | 2374 | */ |
8a13a4cc | 2375 | cachep->align = ralign; |
1da177e4 | 2376 | |
83b519e8 PE |
2377 | if (slab_is_available()) |
2378 | gfp = GFP_KERNEL; | |
2379 | else | |
2380 | gfp = GFP_NOWAIT; | |
2381 | ||
6a67368c | 2382 | setup_node_pointer(cachep); |
1da177e4 | 2383 | #if DEBUG |
1da177e4 | 2384 | |
ca5f9703 PE |
2385 | /* |
2386 | * Both debugging options require word-alignment which is calculated | |
2387 | * into align above. | |
2388 | */ | |
1da177e4 | 2389 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2390 | /* add space for red zone words */ |
3ff84a7f PE |
2391 | cachep->obj_offset += sizeof(unsigned long long); |
2392 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2393 | } |
2394 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2395 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2396 | * the real object. But if the second red zone needs to be |
2397 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2398 | */ |
87a927c7 DW |
2399 | if (flags & SLAB_RED_ZONE) |
2400 | size += REDZONE_ALIGN; | |
2401 | else | |
2402 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2403 | } |
2404 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
e3366016 CL |
2405 | if (size >= kmalloc_size(INDEX_L3 + 1) |
2406 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { | |
2407 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); | |
1da177e4 LT |
2408 | size = PAGE_SIZE; |
2409 | } | |
2410 | #endif | |
2411 | #endif | |
2412 | ||
e0a42726 IM |
2413 | /* |
2414 | * Determine if the slab management is 'on' or 'off' slab. | |
2415 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2416 | * it too early on. Always use on-slab management when |
2417 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2418 | */ |
e7cb55b9 CM |
2419 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2420 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2421 | /* |
2422 | * Size is large, assume best to place the slab management obj | |
2423 | * off-slab (should allow better packing of objs). | |
2424 | */ | |
2425 | flags |= CFLGS_OFF_SLAB; | |
2426 | ||
8a13a4cc | 2427 | size = ALIGN(size, cachep->align); |
1da177e4 | 2428 | |
8a13a4cc | 2429 | left_over = calculate_slab_order(cachep, size, cachep->align, flags); |
1da177e4 | 2430 | |
8a13a4cc | 2431 | if (!cachep->num) |
278b1bb1 | 2432 | return -E2BIG; |
1da177e4 | 2433 | |
b28a02de | 2434 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
8a13a4cc | 2435 | + sizeof(struct slab), cachep->align); |
1da177e4 LT |
2436 | |
2437 | /* | |
2438 | * If the slab has been placed off-slab, and we have enough space then | |
2439 | * move it on-slab. This is at the expense of any extra colouring. | |
2440 | */ | |
2441 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2442 | flags &= ~CFLGS_OFF_SLAB; | |
2443 | left_over -= slab_size; | |
2444 | } | |
2445 | ||
2446 | if (flags & CFLGS_OFF_SLAB) { | |
2447 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2448 | slab_size = |
2449 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2450 | |
2451 | #ifdef CONFIG_PAGE_POISONING | |
2452 | /* If we're going to use the generic kernel_map_pages() | |
2453 | * poisoning, then it's going to smash the contents of | |
2454 | * the redzone and userword anyhow, so switch them off. | |
2455 | */ | |
2456 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2457 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2458 | #endif | |
1da177e4 LT |
2459 | } |
2460 | ||
2461 | cachep->colour_off = cache_line_size(); | |
2462 | /* Offset must be a multiple of the alignment. */ | |
8a13a4cc CL |
2463 | if (cachep->colour_off < cachep->align) |
2464 | cachep->colour_off = cachep->align; | |
b28a02de | 2465 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2466 | cachep->slab_size = slab_size; |
2467 | cachep->flags = flags; | |
a618e89f | 2468 | cachep->allocflags = 0; |
4b51d669 | 2469 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2470 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2471 | cachep->size = size; |
6a2d7a95 | 2472 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2473 | |
e5ac9c5a | 2474 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2475 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2476 | /* |
2477 | * This is a possibility for one of the malloc_sizes caches. | |
2478 | * But since we go off slab only for object size greater than | |
2479 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2480 | * this should not happen at all. | |
2481 | * But leave a BUG_ON for some lucky dude. | |
2482 | */ | |
6cb8f913 | 2483 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2484 | } |
1da177e4 | 2485 | |
278b1bb1 CL |
2486 | err = setup_cpu_cache(cachep, gfp); |
2487 | if (err) { | |
12c3667f | 2488 | __kmem_cache_shutdown(cachep); |
278b1bb1 | 2489 | return err; |
2ed3a4ef | 2490 | } |
1da177e4 | 2491 | |
83835b3d PZ |
2492 | if (flags & SLAB_DEBUG_OBJECTS) { |
2493 | /* | |
2494 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2495 | * debug_object_activate()->kmem_cache_alloc(). | |
2496 | */ | |
2497 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2498 | ||
2499 | slab_set_debugobj_lock_classes(cachep); | |
6ccfb5bc GC |
2500 | } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU)) |
2501 | on_slab_lock_classes(cachep); | |
83835b3d | 2502 | |
278b1bb1 | 2503 | return 0; |
1da177e4 | 2504 | } |
1da177e4 LT |
2505 | |
2506 | #if DEBUG | |
2507 | static void check_irq_off(void) | |
2508 | { | |
2509 | BUG_ON(!irqs_disabled()); | |
2510 | } | |
2511 | ||
2512 | static void check_irq_on(void) | |
2513 | { | |
2514 | BUG_ON(irqs_disabled()); | |
2515 | } | |
2516 | ||
343e0d7a | 2517 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2518 | { |
2519 | #ifdef CONFIG_SMP | |
2520 | check_irq_off(); | |
6a67368c | 2521 | assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock); |
1da177e4 LT |
2522 | #endif |
2523 | } | |
e498be7d | 2524 | |
343e0d7a | 2525 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2526 | { |
2527 | #ifdef CONFIG_SMP | |
2528 | check_irq_off(); | |
6a67368c | 2529 | assert_spin_locked(&cachep->node[node]->list_lock); |
e498be7d CL |
2530 | #endif |
2531 | } | |
2532 | ||
1da177e4 LT |
2533 | #else |
2534 | #define check_irq_off() do { } while(0) | |
2535 | #define check_irq_on() do { } while(0) | |
2536 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2537 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2538 | #endif |
2539 | ||
6744f087 | 2540 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3, |
aab2207c CL |
2541 | struct array_cache *ac, |
2542 | int force, int node); | |
2543 | ||
1da177e4 LT |
2544 | static void do_drain(void *arg) |
2545 | { | |
a737b3e2 | 2546 | struct kmem_cache *cachep = arg; |
1da177e4 | 2547 | struct array_cache *ac; |
7d6e6d09 | 2548 | int node = numa_mem_id(); |
1da177e4 LT |
2549 | |
2550 | check_irq_off(); | |
9a2dba4b | 2551 | ac = cpu_cache_get(cachep); |
6a67368c | 2552 | spin_lock(&cachep->node[node]->list_lock); |
ff69416e | 2553 | free_block(cachep, ac->entry, ac->avail, node); |
6a67368c | 2554 | spin_unlock(&cachep->node[node]->list_lock); |
1da177e4 LT |
2555 | ac->avail = 0; |
2556 | } | |
2557 | ||
343e0d7a | 2558 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2559 | { |
6744f087 | 2560 | struct kmem_cache_node *l3; |
e498be7d CL |
2561 | int node; |
2562 | ||
15c8b6c1 | 2563 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2564 | check_irq_on(); |
b28a02de | 2565 | for_each_online_node(node) { |
6a67368c | 2566 | l3 = cachep->node[node]; |
a4523a8b RD |
2567 | if (l3 && l3->alien) |
2568 | drain_alien_cache(cachep, l3->alien); | |
2569 | } | |
2570 | ||
2571 | for_each_online_node(node) { | |
6a67368c | 2572 | l3 = cachep->node[node]; |
a4523a8b | 2573 | if (l3) |
aab2207c | 2574 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2575 | } |
1da177e4 LT |
2576 | } |
2577 | ||
ed11d9eb CL |
2578 | /* |
2579 | * Remove slabs from the list of free slabs. | |
2580 | * Specify the number of slabs to drain in tofree. | |
2581 | * | |
2582 | * Returns the actual number of slabs released. | |
2583 | */ | |
2584 | static int drain_freelist(struct kmem_cache *cache, | |
6744f087 | 2585 | struct kmem_cache_node *l3, int tofree) |
1da177e4 | 2586 | { |
ed11d9eb CL |
2587 | struct list_head *p; |
2588 | int nr_freed; | |
1da177e4 | 2589 | struct slab *slabp; |
1da177e4 | 2590 | |
ed11d9eb CL |
2591 | nr_freed = 0; |
2592 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2593 | |
ed11d9eb | 2594 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2595 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2596 | if (p == &l3->slabs_free) { |
2597 | spin_unlock_irq(&l3->list_lock); | |
2598 | goto out; | |
2599 | } | |
1da177e4 | 2600 | |
ed11d9eb | 2601 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2602 | #if DEBUG |
40094fa6 | 2603 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2604 | #endif |
2605 | list_del(&slabp->list); | |
ed11d9eb CL |
2606 | /* |
2607 | * Safe to drop the lock. The slab is no longer linked | |
2608 | * to the cache. | |
2609 | */ | |
2610 | l3->free_objects -= cache->num; | |
e498be7d | 2611 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2612 | slab_destroy(cache, slabp); |
2613 | nr_freed++; | |
1da177e4 | 2614 | } |
ed11d9eb CL |
2615 | out: |
2616 | return nr_freed; | |
1da177e4 LT |
2617 | } |
2618 | ||
18004c5d | 2619 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2620 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2621 | { |
2622 | int ret = 0, i = 0; | |
6744f087 | 2623 | struct kmem_cache_node *l3; |
e498be7d CL |
2624 | |
2625 | drain_cpu_caches(cachep); | |
2626 | ||
2627 | check_irq_on(); | |
2628 | for_each_online_node(i) { | |
6a67368c | 2629 | l3 = cachep->node[i]; |
ed11d9eb CL |
2630 | if (!l3) |
2631 | continue; | |
2632 | ||
2633 | drain_freelist(cachep, l3, l3->free_objects); | |
2634 | ||
2635 | ret += !list_empty(&l3->slabs_full) || | |
2636 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2637 | } |
2638 | return (ret ? 1 : 0); | |
2639 | } | |
2640 | ||
1da177e4 LT |
2641 | /** |
2642 | * kmem_cache_shrink - Shrink a cache. | |
2643 | * @cachep: The cache to shrink. | |
2644 | * | |
2645 | * Releases as many slabs as possible for a cache. | |
2646 | * To help debugging, a zero exit status indicates all slabs were released. | |
2647 | */ | |
343e0d7a | 2648 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2649 | { |
8f5be20b | 2650 | int ret; |
40094fa6 | 2651 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2652 | |
95402b38 | 2653 | get_online_cpus(); |
18004c5d | 2654 | mutex_lock(&slab_mutex); |
8f5be20b | 2655 | ret = __cache_shrink(cachep); |
18004c5d | 2656 | mutex_unlock(&slab_mutex); |
95402b38 | 2657 | put_online_cpus(); |
8f5be20b | 2658 | return ret; |
1da177e4 LT |
2659 | } |
2660 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2661 | ||
945cf2b6 | 2662 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
1da177e4 | 2663 | { |
12c3667f | 2664 | int i; |
6744f087 | 2665 | struct kmem_cache_node *l3; |
12c3667f | 2666 | int rc = __cache_shrink(cachep); |
1da177e4 | 2667 | |
12c3667f CL |
2668 | if (rc) |
2669 | return rc; | |
1da177e4 | 2670 | |
12c3667f CL |
2671 | for_each_online_cpu(i) |
2672 | kfree(cachep->array[i]); | |
1da177e4 | 2673 | |
12c3667f CL |
2674 | /* NUMA: free the list3 structures */ |
2675 | for_each_online_node(i) { | |
6a67368c | 2676 | l3 = cachep->node[i]; |
12c3667f CL |
2677 | if (l3) { |
2678 | kfree(l3->shared); | |
2679 | free_alien_cache(l3->alien); | |
2680 | kfree(l3); | |
2681 | } | |
2682 | } | |
2683 | return 0; | |
1da177e4 | 2684 | } |
1da177e4 | 2685 | |
e5ac9c5a RT |
2686 | /* |
2687 | * Get the memory for a slab management obj. | |
2688 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2689 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2690 | * come from the same cache which is getting created because, | |
2691 | * when we are searching for an appropriate cache for these | |
2692 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2693 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2694 | * kmem_find_general_cachep till the initialization is complete. | |
2695 | * Hence we cannot have slabp_cache same as the original cache. | |
2696 | */ | |
343e0d7a | 2697 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2698 | int colour_off, gfp_t local_flags, |
2699 | int nodeid) | |
1da177e4 LT |
2700 | { |
2701 | struct slab *slabp; | |
b28a02de | 2702 | |
1da177e4 LT |
2703 | if (OFF_SLAB(cachep)) { |
2704 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2705 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2706 | local_flags, nodeid); |
d5cff635 CM |
2707 | /* |
2708 | * If the first object in the slab is leaked (it's allocated | |
2709 | * but no one has a reference to it), we want to make sure | |
2710 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2711 | * to the object. Otherwise we will not report the leak. | |
2712 | */ | |
c017b4be CM |
2713 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2714 | local_flags); | |
1da177e4 LT |
2715 | if (!slabp) |
2716 | return NULL; | |
2717 | } else { | |
b28a02de | 2718 | slabp = objp + colour_off; |
1da177e4 LT |
2719 | colour_off += cachep->slab_size; |
2720 | } | |
2721 | slabp->inuse = 0; | |
2722 | slabp->colouroff = colour_off; | |
b28a02de | 2723 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2724 | slabp->nodeid = nodeid; |
e51bfd0a | 2725 | slabp->free = 0; |
1da177e4 LT |
2726 | return slabp; |
2727 | } | |
2728 | ||
2729 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2730 | { | |
b28a02de | 2731 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2732 | } |
2733 | ||
343e0d7a | 2734 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2735 | struct slab *slabp) |
1da177e4 LT |
2736 | { |
2737 | int i; | |
2738 | ||
2739 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2740 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2741 | #if DEBUG |
2742 | /* need to poison the objs? */ | |
2743 | if (cachep->flags & SLAB_POISON) | |
2744 | poison_obj(cachep, objp, POISON_FREE); | |
2745 | if (cachep->flags & SLAB_STORE_USER) | |
2746 | *dbg_userword(cachep, objp) = NULL; | |
2747 | ||
2748 | if (cachep->flags & SLAB_RED_ZONE) { | |
2749 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2750 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2751 | } | |
2752 | /* | |
a737b3e2 AM |
2753 | * Constructors are not allowed to allocate memory from the same |
2754 | * cache which they are a constructor for. Otherwise, deadlock. | |
2755 | * They must also be threaded. | |
1da177e4 LT |
2756 | */ |
2757 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2758 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2759 | |
2760 | if (cachep->flags & SLAB_RED_ZONE) { | |
2761 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2762 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2763 | " end of an object"); |
1da177e4 LT |
2764 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2765 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2766 | " start of an object"); |
1da177e4 | 2767 | } |
3b0efdfa | 2768 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2769 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2770 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2771 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2772 | #else |
2773 | if (cachep->ctor) | |
51cc5068 | 2774 | cachep->ctor(objp); |
1da177e4 | 2775 | #endif |
b28a02de | 2776 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2777 | } |
b28a02de | 2778 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2779 | } |
2780 | ||
343e0d7a | 2781 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2782 | { |
4b51d669 CL |
2783 | if (CONFIG_ZONE_DMA_FLAG) { |
2784 | if (flags & GFP_DMA) | |
a618e89f | 2785 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2786 | else |
a618e89f | 2787 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2788 | } |
1da177e4 LT |
2789 | } |
2790 | ||
a737b3e2 AM |
2791 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2792 | int nodeid) | |
78d382d7 | 2793 | { |
8fea4e96 | 2794 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2795 | kmem_bufctl_t next; |
2796 | ||
2797 | slabp->inuse++; | |
2798 | next = slab_bufctl(slabp)[slabp->free]; | |
2799 | #if DEBUG | |
2800 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2801 | WARN_ON(slabp->nodeid != nodeid); | |
2802 | #endif | |
2803 | slabp->free = next; | |
2804 | ||
2805 | return objp; | |
2806 | } | |
2807 | ||
a737b3e2 AM |
2808 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2809 | void *objp, int nodeid) | |
78d382d7 | 2810 | { |
8fea4e96 | 2811 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2812 | |
2813 | #if DEBUG | |
2814 | /* Verify that the slab belongs to the intended node */ | |
2815 | WARN_ON(slabp->nodeid != nodeid); | |
2816 | ||
871751e2 | 2817 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2818 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2819 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2820 | BUG(); |
2821 | } | |
2822 | #endif | |
2823 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2824 | slabp->free = objnr; | |
2825 | slabp->inuse--; | |
2826 | } | |
2827 | ||
4776874f PE |
2828 | /* |
2829 | * Map pages beginning at addr to the given cache and slab. This is required | |
2830 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2831 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2832 | */ |
2833 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2834 | void *addr) | |
1da177e4 | 2835 | { |
4776874f | 2836 | int nr_pages; |
1da177e4 LT |
2837 | struct page *page; |
2838 | ||
4776874f | 2839 | page = virt_to_page(addr); |
84097518 | 2840 | |
4776874f | 2841 | nr_pages = 1; |
84097518 | 2842 | if (likely(!PageCompound(page))) |
4776874f PE |
2843 | nr_pages <<= cache->gfporder; |
2844 | ||
1da177e4 | 2845 | do { |
35026088 CL |
2846 | page->slab_cache = cache; |
2847 | page->slab_page = slab; | |
1da177e4 | 2848 | page++; |
4776874f | 2849 | } while (--nr_pages); |
1da177e4 LT |
2850 | } |
2851 | ||
2852 | /* | |
2853 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2854 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2855 | */ | |
3c517a61 CL |
2856 | static int cache_grow(struct kmem_cache *cachep, |
2857 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2858 | { |
b28a02de | 2859 | struct slab *slabp; |
b28a02de PE |
2860 | size_t offset; |
2861 | gfp_t local_flags; | |
6744f087 | 2862 | struct kmem_cache_node *l3; |
1da177e4 | 2863 | |
a737b3e2 AM |
2864 | /* |
2865 | * Be lazy and only check for valid flags here, keeping it out of the | |
2866 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2867 | */ |
6cb06229 CL |
2868 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2869 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2870 | |
2e1217cf | 2871 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2872 | check_irq_off(); |
6a67368c | 2873 | l3 = cachep->node[nodeid]; |
2e1217cf | 2874 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2875 | |
2876 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2877 | offset = l3->colour_next; |
2878 | l3->colour_next++; | |
2879 | if (l3->colour_next >= cachep->colour) | |
2880 | l3->colour_next = 0; | |
2881 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2882 | |
2e1217cf | 2883 | offset *= cachep->colour_off; |
1da177e4 LT |
2884 | |
2885 | if (local_flags & __GFP_WAIT) | |
2886 | local_irq_enable(); | |
2887 | ||
2888 | /* | |
2889 | * The test for missing atomic flag is performed here, rather than | |
2890 | * the more obvious place, simply to reduce the critical path length | |
2891 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2892 | * will eventually be caught here (where it matters). | |
2893 | */ | |
2894 | kmem_flagcheck(cachep, flags); | |
2895 | ||
a737b3e2 AM |
2896 | /* |
2897 | * Get mem for the objs. Attempt to allocate a physical page from | |
2898 | * 'nodeid'. | |
e498be7d | 2899 | */ |
3c517a61 | 2900 | if (!objp) |
b8c1c5da | 2901 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2902 | if (!objp) |
1da177e4 LT |
2903 | goto failed; |
2904 | ||
2905 | /* Get slab management. */ | |
3c517a61 | 2906 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2907 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2908 | if (!slabp) |
1da177e4 LT |
2909 | goto opps1; |
2910 | ||
4776874f | 2911 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2912 | |
a35afb83 | 2913 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2914 | |
2915 | if (local_flags & __GFP_WAIT) | |
2916 | local_irq_disable(); | |
2917 | check_irq_off(); | |
e498be7d | 2918 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2919 | |
2920 | /* Make slab active. */ | |
e498be7d | 2921 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2922 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2923 | l3->free_objects += cachep->num; |
2924 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2925 | return 1; |
a737b3e2 | 2926 | opps1: |
1da177e4 | 2927 | kmem_freepages(cachep, objp); |
a737b3e2 | 2928 | failed: |
1da177e4 LT |
2929 | if (local_flags & __GFP_WAIT) |
2930 | local_irq_disable(); | |
2931 | return 0; | |
2932 | } | |
2933 | ||
2934 | #if DEBUG | |
2935 | ||
2936 | /* | |
2937 | * Perform extra freeing checks: | |
2938 | * - detect bad pointers. | |
2939 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
2940 | */ |
2941 | static void kfree_debugcheck(const void *objp) | |
2942 | { | |
1da177e4 LT |
2943 | if (!virt_addr_valid(objp)) { |
2944 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
2945 | (unsigned long)objp); |
2946 | BUG(); | |
1da177e4 | 2947 | } |
1da177e4 LT |
2948 | } |
2949 | ||
58ce1fd5 PE |
2950 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2951 | { | |
b46b8f19 | 2952 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
2953 | |
2954 | redzone1 = *dbg_redzone1(cache, obj); | |
2955 | redzone2 = *dbg_redzone2(cache, obj); | |
2956 | ||
2957 | /* | |
2958 | * Redzone is ok. | |
2959 | */ | |
2960 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
2961 | return; | |
2962 | ||
2963 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
2964 | slab_error(cache, "double free detected"); | |
2965 | else | |
2966 | slab_error(cache, "memory outside object was overwritten"); | |
2967 | ||
b46b8f19 | 2968 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
2969 | obj, redzone1, redzone2); |
2970 | } | |
2971 | ||
343e0d7a | 2972 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 2973 | unsigned long caller) |
1da177e4 LT |
2974 | { |
2975 | struct page *page; | |
2976 | unsigned int objnr; | |
2977 | struct slab *slabp; | |
2978 | ||
80cbd911 MW |
2979 | BUG_ON(virt_to_cache(objp) != cachep); |
2980 | ||
3dafccf2 | 2981 | objp -= obj_offset(cachep); |
1da177e4 | 2982 | kfree_debugcheck(objp); |
b49af68f | 2983 | page = virt_to_head_page(objp); |
1da177e4 | 2984 | |
35026088 | 2985 | slabp = page->slab_page; |
1da177e4 LT |
2986 | |
2987 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 2988 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
2989 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2990 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2991 | } | |
2992 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 2993 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 | 2994 | |
8fea4e96 | 2995 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
2996 | |
2997 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 2998 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 2999 | |
871751e2 AV |
3000 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3001 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3002 | #endif | |
1da177e4 LT |
3003 | if (cachep->flags & SLAB_POISON) { |
3004 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3005 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
7c0cb9c6 | 3006 | store_stackinfo(cachep, objp, caller); |
b28a02de | 3007 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3008 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3009 | } else { |
3010 | poison_obj(cachep, objp, POISON_FREE); | |
3011 | } | |
3012 | #else | |
3013 | poison_obj(cachep, objp, POISON_FREE); | |
3014 | #endif | |
3015 | } | |
3016 | return objp; | |
3017 | } | |
3018 | ||
343e0d7a | 3019 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3020 | { |
3021 | kmem_bufctl_t i; | |
3022 | int entries = 0; | |
b28a02de | 3023 | |
1da177e4 LT |
3024 | /* Check slab's freelist to see if this obj is there. */ |
3025 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3026 | entries++; | |
3027 | if (entries > cachep->num || i >= cachep->num) | |
3028 | goto bad; | |
3029 | } | |
3030 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3031 | bad: |
3032 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3033 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3034 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3035 | print_tainted()); | |
fdde6abb SAS |
3036 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3037 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3038 | 1); | |
1da177e4 LT |
3039 | BUG(); |
3040 | } | |
3041 | } | |
3042 | #else | |
3043 | #define kfree_debugcheck(x) do { } while(0) | |
3044 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3045 | #define check_slabp(x,y) do { } while(0) | |
3046 | #endif | |
3047 | ||
072bb0aa MG |
3048 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
3049 | bool force_refill) | |
1da177e4 LT |
3050 | { |
3051 | int batchcount; | |
6744f087 | 3052 | struct kmem_cache_node *l3; |
1da177e4 | 3053 | struct array_cache *ac; |
1ca4cb24 PE |
3054 | int node; |
3055 | ||
1da177e4 | 3056 | check_irq_off(); |
7d6e6d09 | 3057 | node = numa_mem_id(); |
072bb0aa MG |
3058 | if (unlikely(force_refill)) |
3059 | goto force_grow; | |
3060 | retry: | |
9a2dba4b | 3061 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3062 | batchcount = ac->batchcount; |
3063 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3064 | /* |
3065 | * If there was little recent activity on this cache, then | |
3066 | * perform only a partial refill. Otherwise we could generate | |
3067 | * refill bouncing. | |
1da177e4 LT |
3068 | */ |
3069 | batchcount = BATCHREFILL_LIMIT; | |
3070 | } | |
6a67368c | 3071 | l3 = cachep->node[node]; |
e498be7d CL |
3072 | |
3073 | BUG_ON(ac->avail > 0 || !l3); | |
3074 | spin_lock(&l3->list_lock); | |
1da177e4 | 3075 | |
3ded175a | 3076 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3077 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3078 | l3->shared->touched = 1; | |
3ded175a | 3079 | goto alloc_done; |
44b57f1c | 3080 | } |
3ded175a | 3081 | |
1da177e4 LT |
3082 | while (batchcount > 0) { |
3083 | struct list_head *entry; | |
3084 | struct slab *slabp; | |
3085 | /* Get slab alloc is to come from. */ | |
3086 | entry = l3->slabs_partial.next; | |
3087 | if (entry == &l3->slabs_partial) { | |
3088 | l3->free_touched = 1; | |
3089 | entry = l3->slabs_free.next; | |
3090 | if (entry == &l3->slabs_free) | |
3091 | goto must_grow; | |
3092 | } | |
3093 | ||
3094 | slabp = list_entry(entry, struct slab, list); | |
3095 | check_slabp(cachep, slabp); | |
3096 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3097 | |
3098 | /* | |
3099 | * The slab was either on partial or free list so | |
3100 | * there must be at least one object available for | |
3101 | * allocation. | |
3102 | */ | |
249b9f33 | 3103 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3104 | |
1da177e4 | 3105 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3106 | STATS_INC_ALLOCED(cachep); |
3107 | STATS_INC_ACTIVE(cachep); | |
3108 | STATS_SET_HIGH(cachep); | |
3109 | ||
072bb0aa MG |
3110 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, |
3111 | node)); | |
1da177e4 LT |
3112 | } |
3113 | check_slabp(cachep, slabp); | |
3114 | ||
3115 | /* move slabp to correct slabp list: */ | |
3116 | list_del(&slabp->list); | |
3117 | if (slabp->free == BUFCTL_END) | |
3118 | list_add(&slabp->list, &l3->slabs_full); | |
3119 | else | |
3120 | list_add(&slabp->list, &l3->slabs_partial); | |
3121 | } | |
3122 | ||
a737b3e2 | 3123 | must_grow: |
1da177e4 | 3124 | l3->free_objects -= ac->avail; |
a737b3e2 | 3125 | alloc_done: |
e498be7d | 3126 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3127 | |
3128 | if (unlikely(!ac->avail)) { | |
3129 | int x; | |
072bb0aa | 3130 | force_grow: |
3c517a61 | 3131 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3132 | |
a737b3e2 | 3133 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3134 | ac = cpu_cache_get(cachep); |
51cd8e6f | 3135 | node = numa_mem_id(); |
072bb0aa MG |
3136 | |
3137 | /* no objects in sight? abort */ | |
3138 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
3139 | return NULL; |
3140 | ||
a737b3e2 | 3141 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3142 | goto retry; |
3143 | } | |
3144 | ac->touched = 1; | |
072bb0aa MG |
3145 | |
3146 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
3147 | } |
3148 | ||
a737b3e2 AM |
3149 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3150 | gfp_t flags) | |
1da177e4 LT |
3151 | { |
3152 | might_sleep_if(flags & __GFP_WAIT); | |
3153 | #if DEBUG | |
3154 | kmem_flagcheck(cachep, flags); | |
3155 | #endif | |
3156 | } | |
3157 | ||
3158 | #if DEBUG | |
a737b3e2 | 3159 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
7c0cb9c6 | 3160 | gfp_t flags, void *objp, unsigned long caller) |
1da177e4 | 3161 | { |
b28a02de | 3162 | if (!objp) |
1da177e4 | 3163 | return objp; |
b28a02de | 3164 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3165 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3166 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3167 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3168 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3169 | else |
3170 | check_poison_obj(cachep, objp); | |
3171 | #else | |
3172 | check_poison_obj(cachep, objp); | |
3173 | #endif | |
3174 | poison_obj(cachep, objp, POISON_INUSE); | |
3175 | } | |
3176 | if (cachep->flags & SLAB_STORE_USER) | |
7c0cb9c6 | 3177 | *dbg_userword(cachep, objp) = (void *)caller; |
1da177e4 LT |
3178 | |
3179 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3180 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3181 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3182 | slab_error(cachep, "double free, or memory outside" | |
3183 | " object was overwritten"); | |
b28a02de | 3184 | printk(KERN_ERR |
b46b8f19 | 3185 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3186 | objp, *dbg_redzone1(cachep, objp), |
3187 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3188 | } |
3189 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3190 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3191 | } | |
871751e2 AV |
3192 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3193 | { | |
3194 | struct slab *slabp; | |
3195 | unsigned objnr; | |
3196 | ||
35026088 | 3197 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3198 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3199 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3200 | } | |
3201 | #endif | |
3dafccf2 | 3202 | objp += obj_offset(cachep); |
4f104934 | 3203 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3204 | cachep->ctor(objp); |
7ea466f2 TH |
3205 | if (ARCH_SLAB_MINALIGN && |
3206 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3207 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3208 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3209 | } |
1da177e4 LT |
3210 | return objp; |
3211 | } | |
3212 | #else | |
3213 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3214 | #endif | |
3215 | ||
773ff60e | 3216 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 | 3217 | { |
9b030cb8 | 3218 | if (cachep == kmem_cache) |
773ff60e | 3219 | return false; |
8a8b6502 | 3220 | |
8c138bc0 | 3221 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3222 | } |
3223 | ||
343e0d7a | 3224 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3225 | { |
b28a02de | 3226 | void *objp; |
1da177e4 | 3227 | struct array_cache *ac; |
072bb0aa | 3228 | bool force_refill = false; |
1da177e4 | 3229 | |
5c382300 | 3230 | check_irq_off(); |
8a8b6502 | 3231 | |
9a2dba4b | 3232 | ac = cpu_cache_get(cachep); |
1da177e4 | 3233 | if (likely(ac->avail)) { |
1da177e4 | 3234 | ac->touched = 1; |
072bb0aa MG |
3235 | objp = ac_get_obj(cachep, ac, flags, false); |
3236 | ||
ddbf2e83 | 3237 | /* |
072bb0aa MG |
3238 | * Allow for the possibility all avail objects are not allowed |
3239 | * by the current flags | |
ddbf2e83 | 3240 | */ |
072bb0aa MG |
3241 | if (objp) { |
3242 | STATS_INC_ALLOCHIT(cachep); | |
3243 | goto out; | |
3244 | } | |
3245 | force_refill = true; | |
1da177e4 | 3246 | } |
072bb0aa MG |
3247 | |
3248 | STATS_INC_ALLOCMISS(cachep); | |
3249 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3250 | /* | |
3251 | * the 'ac' may be updated by cache_alloc_refill(), | |
3252 | * and kmemleak_erase() requires its correct value. | |
3253 | */ | |
3254 | ac = cpu_cache_get(cachep); | |
3255 | ||
3256 | out: | |
d5cff635 CM |
3257 | /* |
3258 | * To avoid a false negative, if an object that is in one of the | |
3259 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3260 | * treat the array pointers as a reference to the object. | |
3261 | */ | |
f3d8b53a O |
3262 | if (objp) |
3263 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3264 | return objp; |
3265 | } | |
3266 | ||
e498be7d | 3267 | #ifdef CONFIG_NUMA |
c61afb18 | 3268 | /* |
b2455396 | 3269 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3270 | * |
3271 | * If we are in_interrupt, then process context, including cpusets and | |
3272 | * mempolicy, may not apply and should not be used for allocation policy. | |
3273 | */ | |
3274 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3275 | { | |
3276 | int nid_alloc, nid_here; | |
3277 | ||
765c4507 | 3278 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3279 | return NULL; |
7d6e6d09 | 3280 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3281 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3282 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3283 | else if (current->mempolicy) |
e7b691b0 | 3284 | nid_alloc = slab_node(); |
c61afb18 | 3285 | if (nid_alloc != nid_here) |
8b98c169 | 3286 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3287 | return NULL; |
3288 | } | |
3289 | ||
765c4507 CL |
3290 | /* |
3291 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 | 3292 | * certain node and fall back is permitted. First we scan all the |
6a67368c | 3293 | * available node for available objects. If that fails then we |
3c517a61 CL |
3294 | * perform an allocation without specifying a node. This allows the page |
3295 | * allocator to do its reclaim / fallback magic. We then insert the | |
3296 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3297 | */ |
8c8cc2c1 | 3298 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3299 | { |
8c8cc2c1 PE |
3300 | struct zonelist *zonelist; |
3301 | gfp_t local_flags; | |
dd1a239f | 3302 | struct zoneref *z; |
54a6eb5c MG |
3303 | struct zone *zone; |
3304 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3305 | void *obj = NULL; |
3c517a61 | 3306 | int nid; |
cc9a6c87 | 3307 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3308 | |
3309 | if (flags & __GFP_THISNODE) | |
3310 | return NULL; | |
3311 | ||
6cb06229 | 3312 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3313 | |
cc9a6c87 MG |
3314 | retry_cpuset: |
3315 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3316 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3317 | |
3c517a61 CL |
3318 | retry: |
3319 | /* | |
3320 | * Look through allowed nodes for objects available | |
3321 | * from existing per node queues. | |
3322 | */ | |
54a6eb5c MG |
3323 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3324 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3325 | |
54a6eb5c | 3326 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
6a67368c CL |
3327 | cache->node[nid] && |
3328 | cache->node[nid]->free_objects) { | |
3c517a61 CL |
3329 | obj = ____cache_alloc_node(cache, |
3330 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3331 | if (obj) |
3332 | break; | |
3333 | } | |
3c517a61 CL |
3334 | } |
3335 | ||
cfce6604 | 3336 | if (!obj) { |
3c517a61 CL |
3337 | /* |
3338 | * This allocation will be performed within the constraints | |
3339 | * of the current cpuset / memory policy requirements. | |
3340 | * We may trigger various forms of reclaim on the allowed | |
3341 | * set and go into memory reserves if necessary. | |
3342 | */ | |
dd47ea75 CL |
3343 | if (local_flags & __GFP_WAIT) |
3344 | local_irq_enable(); | |
3345 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3346 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3347 | if (local_flags & __GFP_WAIT) |
3348 | local_irq_disable(); | |
3c517a61 CL |
3349 | if (obj) { |
3350 | /* | |
3351 | * Insert into the appropriate per node queues | |
3352 | */ | |
3353 | nid = page_to_nid(virt_to_page(obj)); | |
3354 | if (cache_grow(cache, flags, nid, obj)) { | |
3355 | obj = ____cache_alloc_node(cache, | |
3356 | flags | GFP_THISNODE, nid); | |
3357 | if (!obj) | |
3358 | /* | |
3359 | * Another processor may allocate the | |
3360 | * objects in the slab since we are | |
3361 | * not holding any locks. | |
3362 | */ | |
3363 | goto retry; | |
3364 | } else { | |
b6a60451 | 3365 | /* cache_grow already freed obj */ |
3c517a61 CL |
3366 | obj = NULL; |
3367 | } | |
3368 | } | |
aedb0eb1 | 3369 | } |
cc9a6c87 MG |
3370 | |
3371 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3372 | goto retry_cpuset; | |
765c4507 CL |
3373 | return obj; |
3374 | } | |
3375 | ||
e498be7d CL |
3376 | /* |
3377 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3378 | */ |
8b98c169 | 3379 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3380 | int nodeid) |
e498be7d CL |
3381 | { |
3382 | struct list_head *entry; | |
b28a02de | 3383 | struct slab *slabp; |
6744f087 | 3384 | struct kmem_cache_node *l3; |
b28a02de | 3385 | void *obj; |
b28a02de PE |
3386 | int x; |
3387 | ||
6a67368c | 3388 | l3 = cachep->node[nodeid]; |
b28a02de PE |
3389 | BUG_ON(!l3); |
3390 | ||
a737b3e2 | 3391 | retry: |
ca3b9b91 | 3392 | check_irq_off(); |
b28a02de PE |
3393 | spin_lock(&l3->list_lock); |
3394 | entry = l3->slabs_partial.next; | |
3395 | if (entry == &l3->slabs_partial) { | |
3396 | l3->free_touched = 1; | |
3397 | entry = l3->slabs_free.next; | |
3398 | if (entry == &l3->slabs_free) | |
3399 | goto must_grow; | |
3400 | } | |
3401 | ||
3402 | slabp = list_entry(entry, struct slab, list); | |
3403 | check_spinlock_acquired_node(cachep, nodeid); | |
3404 | check_slabp(cachep, slabp); | |
3405 | ||
3406 | STATS_INC_NODEALLOCS(cachep); | |
3407 | STATS_INC_ACTIVE(cachep); | |
3408 | STATS_SET_HIGH(cachep); | |
3409 | ||
3410 | BUG_ON(slabp->inuse == cachep->num); | |
3411 | ||
78d382d7 | 3412 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3413 | check_slabp(cachep, slabp); |
3414 | l3->free_objects--; | |
3415 | /* move slabp to correct slabp list: */ | |
3416 | list_del(&slabp->list); | |
3417 | ||
a737b3e2 | 3418 | if (slabp->free == BUFCTL_END) |
b28a02de | 3419 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3420 | else |
b28a02de | 3421 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3422 | |
b28a02de PE |
3423 | spin_unlock(&l3->list_lock); |
3424 | goto done; | |
e498be7d | 3425 | |
a737b3e2 | 3426 | must_grow: |
b28a02de | 3427 | spin_unlock(&l3->list_lock); |
3c517a61 | 3428 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3429 | if (x) |
3430 | goto retry; | |
1da177e4 | 3431 | |
8c8cc2c1 | 3432 | return fallback_alloc(cachep, flags); |
e498be7d | 3433 | |
a737b3e2 | 3434 | done: |
b28a02de | 3435 | return obj; |
e498be7d | 3436 | } |
8c8cc2c1 PE |
3437 | |
3438 | /** | |
3439 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3440 | * @cachep: The cache to allocate from. | |
3441 | * @flags: See kmalloc(). | |
3442 | * @nodeid: node number of the target node. | |
3443 | * @caller: return address of caller, used for debug information | |
3444 | * | |
3445 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3446 | * node, which can improve the performance for cpu bound structures. | |
3447 | * | |
3448 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3449 | */ | |
3450 | static __always_inline void * | |
48356303 | 3451 | slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, |
7c0cb9c6 | 3452 | unsigned long caller) |
8c8cc2c1 PE |
3453 | { |
3454 | unsigned long save_flags; | |
3455 | void *ptr; | |
7d6e6d09 | 3456 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3457 | |
dcce284a | 3458 | flags &= gfp_allowed_mask; |
7e85ee0c | 3459 | |
cf40bd16 NP |
3460 | lockdep_trace_alloc(flags); |
3461 | ||
773ff60e | 3462 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3463 | return NULL; |
3464 | ||
d79923fa GC |
3465 | cachep = memcg_kmem_get_cache(cachep, flags); |
3466 | ||
8c8cc2c1 PE |
3467 | cache_alloc_debugcheck_before(cachep, flags); |
3468 | local_irq_save(save_flags); | |
3469 | ||
eacbbae3 | 3470 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3471 | nodeid = slab_node; |
8c8cc2c1 | 3472 | |
6a67368c | 3473 | if (unlikely(!cachep->node[nodeid])) { |
8c8cc2c1 PE |
3474 | /* Node not bootstrapped yet */ |
3475 | ptr = fallback_alloc(cachep, flags); | |
3476 | goto out; | |
3477 | } | |
3478 | ||
7d6e6d09 | 3479 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3480 | /* |
3481 | * Use the locally cached objects if possible. | |
3482 | * However ____cache_alloc does not allow fallback | |
3483 | * to other nodes. It may fail while we still have | |
3484 | * objects on other nodes available. | |
3485 | */ | |
3486 | ptr = ____cache_alloc(cachep, flags); | |
3487 | if (ptr) | |
3488 | goto out; | |
3489 | } | |
3490 | /* ___cache_alloc_node can fall back to other nodes */ | |
3491 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3492 | out: | |
3493 | local_irq_restore(save_flags); | |
3494 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3495 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3496 | flags); |
8c8cc2c1 | 3497 | |
c175eea4 | 3498 | if (likely(ptr)) |
8c138bc0 | 3499 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3500 | |
d07dbea4 | 3501 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3502 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3503 | |
8c8cc2c1 PE |
3504 | return ptr; |
3505 | } | |
3506 | ||
3507 | static __always_inline void * | |
3508 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3509 | { | |
3510 | void *objp; | |
3511 | ||
3512 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3513 | objp = alternate_node_alloc(cache, flags); | |
3514 | if (objp) | |
3515 | goto out; | |
3516 | } | |
3517 | objp = ____cache_alloc(cache, flags); | |
3518 | ||
3519 | /* | |
3520 | * We may just have run out of memory on the local node. | |
3521 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3522 | */ | |
7d6e6d09 LS |
3523 | if (!objp) |
3524 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3525 | |
3526 | out: | |
3527 | return objp; | |
3528 | } | |
3529 | #else | |
3530 | ||
3531 | static __always_inline void * | |
3532 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3533 | { | |
3534 | return ____cache_alloc(cachep, flags); | |
3535 | } | |
3536 | ||
3537 | #endif /* CONFIG_NUMA */ | |
3538 | ||
3539 | static __always_inline void * | |
48356303 | 3540 | slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) |
8c8cc2c1 PE |
3541 | { |
3542 | unsigned long save_flags; | |
3543 | void *objp; | |
3544 | ||
dcce284a | 3545 | flags &= gfp_allowed_mask; |
7e85ee0c | 3546 | |
cf40bd16 NP |
3547 | lockdep_trace_alloc(flags); |
3548 | ||
773ff60e | 3549 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3550 | return NULL; |
3551 | ||
d79923fa GC |
3552 | cachep = memcg_kmem_get_cache(cachep, flags); |
3553 | ||
8c8cc2c1 PE |
3554 | cache_alloc_debugcheck_before(cachep, flags); |
3555 | local_irq_save(save_flags); | |
3556 | objp = __do_cache_alloc(cachep, flags); | |
3557 | local_irq_restore(save_flags); | |
3558 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3559 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3560 | flags); |
8c8cc2c1 PE |
3561 | prefetchw(objp); |
3562 | ||
c175eea4 | 3563 | if (likely(objp)) |
8c138bc0 | 3564 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3565 | |
d07dbea4 | 3566 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3567 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3568 | |
8c8cc2c1 PE |
3569 | return objp; |
3570 | } | |
e498be7d CL |
3571 | |
3572 | /* | |
3573 | * Caller needs to acquire correct kmem_list's list_lock | |
3574 | */ | |
343e0d7a | 3575 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3576 | int node) |
1da177e4 LT |
3577 | { |
3578 | int i; | |
6744f087 | 3579 | struct kmem_cache_node *l3; |
1da177e4 LT |
3580 | |
3581 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3582 | void *objp; |
1da177e4 | 3583 | struct slab *slabp; |
1da177e4 | 3584 | |
072bb0aa MG |
3585 | clear_obj_pfmemalloc(&objpp[i]); |
3586 | objp = objpp[i]; | |
3587 | ||
6ed5eb22 | 3588 | slabp = virt_to_slab(objp); |
6a67368c | 3589 | l3 = cachep->node[node]; |
1da177e4 | 3590 | list_del(&slabp->list); |
ff69416e | 3591 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3592 | check_slabp(cachep, slabp); |
78d382d7 | 3593 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3594 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3595 | l3->free_objects++; |
1da177e4 LT |
3596 | check_slabp(cachep, slabp); |
3597 | ||
3598 | /* fixup slab chains */ | |
3599 | if (slabp->inuse == 0) { | |
e498be7d CL |
3600 | if (l3->free_objects > l3->free_limit) { |
3601 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3602 | /* No need to drop any previously held |
3603 | * lock here, even if we have a off-slab slab | |
3604 | * descriptor it is guaranteed to come from | |
3605 | * a different cache, refer to comments before | |
3606 | * alloc_slabmgmt. | |
3607 | */ | |
1da177e4 LT |
3608 | slab_destroy(cachep, slabp); |
3609 | } else { | |
e498be7d | 3610 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3611 | } |
3612 | } else { | |
3613 | /* Unconditionally move a slab to the end of the | |
3614 | * partial list on free - maximum time for the | |
3615 | * other objects to be freed, too. | |
3616 | */ | |
e498be7d | 3617 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3618 | } |
3619 | } | |
3620 | } | |
3621 | ||
343e0d7a | 3622 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3623 | { |
3624 | int batchcount; | |
6744f087 | 3625 | struct kmem_cache_node *l3; |
7d6e6d09 | 3626 | int node = numa_mem_id(); |
1da177e4 LT |
3627 | |
3628 | batchcount = ac->batchcount; | |
3629 | #if DEBUG | |
3630 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3631 | #endif | |
3632 | check_irq_off(); | |
6a67368c | 3633 | l3 = cachep->node[node]; |
873623df | 3634 | spin_lock(&l3->list_lock); |
e498be7d CL |
3635 | if (l3->shared) { |
3636 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3637 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3638 | if (max) { |
3639 | if (batchcount > max) | |
3640 | batchcount = max; | |
e498be7d | 3641 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3642 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3643 | shared_array->avail += batchcount; |
3644 | goto free_done; | |
3645 | } | |
3646 | } | |
3647 | ||
ff69416e | 3648 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3649 | free_done: |
1da177e4 LT |
3650 | #if STATS |
3651 | { | |
3652 | int i = 0; | |
3653 | struct list_head *p; | |
3654 | ||
e498be7d CL |
3655 | p = l3->slabs_free.next; |
3656 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3657 | struct slab *slabp; |
3658 | ||
3659 | slabp = list_entry(p, struct slab, list); | |
3660 | BUG_ON(slabp->inuse); | |
3661 | ||
3662 | i++; | |
3663 | p = p->next; | |
3664 | } | |
3665 | STATS_SET_FREEABLE(cachep, i); | |
3666 | } | |
3667 | #endif | |
e498be7d | 3668 | spin_unlock(&l3->list_lock); |
1da177e4 | 3669 | ac->avail -= batchcount; |
a737b3e2 | 3670 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3671 | } |
3672 | ||
3673 | /* | |
a737b3e2 AM |
3674 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3675 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3676 | */ |
a947eb95 | 3677 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
7c0cb9c6 | 3678 | unsigned long caller) |
1da177e4 | 3679 | { |
9a2dba4b | 3680 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3681 | |
3682 | check_irq_off(); | |
d5cff635 | 3683 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3684 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3685 | |
8c138bc0 | 3686 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3687 | |
1807a1aa SS |
3688 | /* |
3689 | * Skip calling cache_free_alien() when the platform is not numa. | |
3690 | * This will avoid cache misses that happen while accessing slabp (which | |
3691 | * is per page memory reference) to get nodeid. Instead use a global | |
3692 | * variable to skip the call, which is mostly likely to be present in | |
3693 | * the cache. | |
3694 | */ | |
b6e68bc1 | 3695 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3696 | return; |
3697 | ||
1da177e4 LT |
3698 | if (likely(ac->avail < ac->limit)) { |
3699 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3700 | } else { |
3701 | STATS_INC_FREEMISS(cachep); | |
3702 | cache_flusharray(cachep, ac); | |
1da177e4 | 3703 | } |
42c8c99c | 3704 | |
072bb0aa | 3705 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3706 | } |
3707 | ||
3708 | /** | |
3709 | * kmem_cache_alloc - Allocate an object | |
3710 | * @cachep: The cache to allocate from. | |
3711 | * @flags: See kmalloc(). | |
3712 | * | |
3713 | * Allocate an object from this cache. The flags are only relevant | |
3714 | * if the cache has no available objects. | |
3715 | */ | |
343e0d7a | 3716 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3717 | { |
48356303 | 3718 | void *ret = slab_alloc(cachep, flags, _RET_IP_); |
36555751 | 3719 | |
ca2b84cb | 3720 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3721 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3722 | |
3723 | return ret; | |
1da177e4 LT |
3724 | } |
3725 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3726 | ||
0f24f128 | 3727 | #ifdef CONFIG_TRACING |
85beb586 | 3728 | void * |
4052147c | 3729 | kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) |
36555751 | 3730 | { |
85beb586 SR |
3731 | void *ret; |
3732 | ||
48356303 | 3733 | ret = slab_alloc(cachep, flags, _RET_IP_); |
85beb586 SR |
3734 | |
3735 | trace_kmalloc(_RET_IP_, ret, | |
ff4fcd01 | 3736 | size, cachep->size, flags); |
85beb586 | 3737 | return ret; |
36555751 | 3738 | } |
85beb586 | 3739 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3740 | #endif |
3741 | ||
1da177e4 | 3742 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3743 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3744 | { | |
48356303 | 3745 | void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
36555751 | 3746 | |
ca2b84cb | 3747 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3748 | cachep->object_size, cachep->size, |
ca2b84cb | 3749 | flags, nodeid); |
36555751 EGM |
3750 | |
3751 | return ret; | |
8b98c169 | 3752 | } |
1da177e4 LT |
3753 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3754 | ||
0f24f128 | 3755 | #ifdef CONFIG_TRACING |
4052147c | 3756 | void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, |
85beb586 | 3757 | gfp_t flags, |
4052147c EG |
3758 | int nodeid, |
3759 | size_t size) | |
36555751 | 3760 | { |
85beb586 SR |
3761 | void *ret; |
3762 | ||
592f4145 | 3763 | ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
7c0cb9c6 | 3764 | |
85beb586 | 3765 | trace_kmalloc_node(_RET_IP_, ret, |
ff4fcd01 | 3766 | size, cachep->size, |
85beb586 SR |
3767 | flags, nodeid); |
3768 | return ret; | |
36555751 | 3769 | } |
85beb586 | 3770 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3771 | #endif |
3772 | ||
8b98c169 | 3773 | static __always_inline void * |
7c0cb9c6 | 3774 | __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) |
97e2bde4 | 3775 | { |
343e0d7a | 3776 | struct kmem_cache *cachep; |
97e2bde4 MS |
3777 | |
3778 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3779 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3780 | return cachep; | |
4052147c | 3781 | return kmem_cache_alloc_node_trace(cachep, flags, node, size); |
97e2bde4 | 3782 | } |
8b98c169 | 3783 | |
0bb38a5c | 3784 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3785 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3786 | { | |
7c0cb9c6 | 3787 | return __do_kmalloc_node(size, flags, node, _RET_IP_); |
8b98c169 | 3788 | } |
dbe5e69d | 3789 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3790 | |
3791 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3792 | int node, unsigned long caller) |
8b98c169 | 3793 | { |
7c0cb9c6 | 3794 | return __do_kmalloc_node(size, flags, node, caller); |
8b98c169 CH |
3795 | } |
3796 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3797 | #else | |
3798 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3799 | { | |
7c0cb9c6 | 3800 | return __do_kmalloc_node(size, flags, node, 0); |
8b98c169 CH |
3801 | } |
3802 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3803 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3804 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3805 | |
3806 | /** | |
800590f5 | 3807 | * __do_kmalloc - allocate memory |
1da177e4 | 3808 | * @size: how many bytes of memory are required. |
800590f5 | 3809 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3810 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3811 | */ |
7fd6b141 | 3812 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
7c0cb9c6 | 3813 | unsigned long caller) |
1da177e4 | 3814 | { |
343e0d7a | 3815 | struct kmem_cache *cachep; |
36555751 | 3816 | void *ret; |
1da177e4 | 3817 | |
97e2bde4 MS |
3818 | /* If you want to save a few bytes .text space: replace |
3819 | * __ with kmem_. | |
3820 | * Then kmalloc uses the uninlined functions instead of the inline | |
3821 | * functions. | |
3822 | */ | |
3823 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3824 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3825 | return cachep; | |
48356303 | 3826 | ret = slab_alloc(cachep, flags, caller); |
36555751 | 3827 | |
7c0cb9c6 | 3828 | trace_kmalloc(caller, ret, |
3b0efdfa | 3829 | size, cachep->size, flags); |
36555751 EGM |
3830 | |
3831 | return ret; | |
7fd6b141 PE |
3832 | } |
3833 | ||
7fd6b141 | 3834 | |
0bb38a5c | 3835 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3836 | void *__kmalloc(size_t size, gfp_t flags) |
3837 | { | |
7c0cb9c6 | 3838 | return __do_kmalloc(size, flags, _RET_IP_); |
1da177e4 LT |
3839 | } |
3840 | EXPORT_SYMBOL(__kmalloc); | |
3841 | ||
ce71e27c | 3842 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3843 | { |
7c0cb9c6 | 3844 | return __do_kmalloc(size, flags, caller); |
7fd6b141 PE |
3845 | } |
3846 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3847 | |
3848 | #else | |
3849 | void *__kmalloc(size_t size, gfp_t flags) | |
3850 | { | |
7c0cb9c6 | 3851 | return __do_kmalloc(size, flags, 0); |
1d2c8eea CH |
3852 | } |
3853 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3854 | #endif |
3855 | ||
1da177e4 LT |
3856 | /** |
3857 | * kmem_cache_free - Deallocate an object | |
3858 | * @cachep: The cache the allocation was from. | |
3859 | * @objp: The previously allocated object. | |
3860 | * | |
3861 | * Free an object which was previously allocated from this | |
3862 | * cache. | |
3863 | */ | |
343e0d7a | 3864 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3865 | { |
3866 | unsigned long flags; | |
b9ce5ef4 GC |
3867 | cachep = cache_from_obj(cachep, objp); |
3868 | if (!cachep) | |
3869 | return; | |
1da177e4 LT |
3870 | |
3871 | local_irq_save(flags); | |
d97d476b | 3872 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3873 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3874 | debug_check_no_obj_freed(objp, cachep->object_size); |
7c0cb9c6 | 3875 | __cache_free(cachep, objp, _RET_IP_); |
1da177e4 | 3876 | local_irq_restore(flags); |
36555751 | 3877 | |
ca2b84cb | 3878 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3879 | } |
3880 | EXPORT_SYMBOL(kmem_cache_free); | |
3881 | ||
1da177e4 LT |
3882 | /** |
3883 | * kfree - free previously allocated memory | |
3884 | * @objp: pointer returned by kmalloc. | |
3885 | * | |
80e93eff PE |
3886 | * If @objp is NULL, no operation is performed. |
3887 | * | |
1da177e4 LT |
3888 | * Don't free memory not originally allocated by kmalloc() |
3889 | * or you will run into trouble. | |
3890 | */ | |
3891 | void kfree(const void *objp) | |
3892 | { | |
343e0d7a | 3893 | struct kmem_cache *c; |
1da177e4 LT |
3894 | unsigned long flags; |
3895 | ||
2121db74 PE |
3896 | trace_kfree(_RET_IP_, objp); |
3897 | ||
6cb8f913 | 3898 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3899 | return; |
3900 | local_irq_save(flags); | |
3901 | kfree_debugcheck(objp); | |
6ed5eb22 | 3902 | c = virt_to_cache(objp); |
8c138bc0 CL |
3903 | debug_check_no_locks_freed(objp, c->object_size); |
3904 | ||
3905 | debug_check_no_obj_freed(objp, c->object_size); | |
7c0cb9c6 | 3906 | __cache_free(c, (void *)objp, _RET_IP_); |
1da177e4 LT |
3907 | local_irq_restore(flags); |
3908 | } | |
3909 | EXPORT_SYMBOL(kfree); | |
3910 | ||
e498be7d | 3911 | /* |
183ff22b | 3912 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3913 | */ |
83b519e8 | 3914 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3915 | { |
3916 | int node; | |
6744f087 | 3917 | struct kmem_cache_node *l3; |
cafeb02e | 3918 | struct array_cache *new_shared; |
3395ee05 | 3919 | struct array_cache **new_alien = NULL; |
e498be7d | 3920 | |
9c09a95c | 3921 | for_each_online_node(node) { |
cafeb02e | 3922 | |
3395ee05 | 3923 | if (use_alien_caches) { |
83b519e8 | 3924 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3925 | if (!new_alien) |
3926 | goto fail; | |
3927 | } | |
cafeb02e | 3928 | |
63109846 ED |
3929 | new_shared = NULL; |
3930 | if (cachep->shared) { | |
3931 | new_shared = alloc_arraycache(node, | |
0718dc2a | 3932 | cachep->shared*cachep->batchcount, |
83b519e8 | 3933 | 0xbaadf00d, gfp); |
63109846 ED |
3934 | if (!new_shared) { |
3935 | free_alien_cache(new_alien); | |
3936 | goto fail; | |
3937 | } | |
0718dc2a | 3938 | } |
cafeb02e | 3939 | |
6a67368c | 3940 | l3 = cachep->node[node]; |
a737b3e2 | 3941 | if (l3) { |
cafeb02e CL |
3942 | struct array_cache *shared = l3->shared; |
3943 | ||
e498be7d CL |
3944 | spin_lock_irq(&l3->list_lock); |
3945 | ||
cafeb02e | 3946 | if (shared) |
0718dc2a CL |
3947 | free_block(cachep, shared->entry, |
3948 | shared->avail, node); | |
e498be7d | 3949 | |
cafeb02e CL |
3950 | l3->shared = new_shared; |
3951 | if (!l3->alien) { | |
e498be7d CL |
3952 | l3->alien = new_alien; |
3953 | new_alien = NULL; | |
3954 | } | |
b28a02de | 3955 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3956 | cachep->batchcount + cachep->num; |
e498be7d | 3957 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 3958 | kfree(shared); |
e498be7d CL |
3959 | free_alien_cache(new_alien); |
3960 | continue; | |
3961 | } | |
6744f087 | 3962 | l3 = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node); |
0718dc2a CL |
3963 | if (!l3) { |
3964 | free_alien_cache(new_alien); | |
3965 | kfree(new_shared); | |
e498be7d | 3966 | goto fail; |
0718dc2a | 3967 | } |
e498be7d CL |
3968 | |
3969 | kmem_list3_init(l3); | |
3970 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 3971 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 3972 | l3->shared = new_shared; |
e498be7d | 3973 | l3->alien = new_alien; |
b28a02de | 3974 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 3975 | cachep->batchcount + cachep->num; |
6a67368c | 3976 | cachep->node[node] = l3; |
e498be7d | 3977 | } |
cafeb02e | 3978 | return 0; |
0718dc2a | 3979 | |
a737b3e2 | 3980 | fail: |
3b0efdfa | 3981 | if (!cachep->list.next) { |
0718dc2a CL |
3982 | /* Cache is not active yet. Roll back what we did */ |
3983 | node--; | |
3984 | while (node >= 0) { | |
6a67368c CL |
3985 | if (cachep->node[node]) { |
3986 | l3 = cachep->node[node]; | |
0718dc2a CL |
3987 | |
3988 | kfree(l3->shared); | |
3989 | free_alien_cache(l3->alien); | |
3990 | kfree(l3); | |
6a67368c | 3991 | cachep->node[node] = NULL; |
0718dc2a CL |
3992 | } |
3993 | node--; | |
3994 | } | |
3995 | } | |
cafeb02e | 3996 | return -ENOMEM; |
e498be7d CL |
3997 | } |
3998 | ||
1da177e4 | 3999 | struct ccupdate_struct { |
343e0d7a | 4000 | struct kmem_cache *cachep; |
acfe7d74 | 4001 | struct array_cache *new[0]; |
1da177e4 LT |
4002 | }; |
4003 | ||
4004 | static void do_ccupdate_local(void *info) | |
4005 | { | |
a737b3e2 | 4006 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4007 | struct array_cache *old; |
4008 | ||
4009 | check_irq_off(); | |
9a2dba4b | 4010 | old = cpu_cache_get(new->cachep); |
e498be7d | 4011 | |
1da177e4 LT |
4012 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4013 | new->new[smp_processor_id()] = old; | |
4014 | } | |
4015 | ||
18004c5d | 4016 | /* Always called with the slab_mutex held */ |
943a451a | 4017 | static int __do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4018 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4019 | { |
d2e7b7d0 | 4020 | struct ccupdate_struct *new; |
2ed3a4ef | 4021 | int i; |
1da177e4 | 4022 | |
acfe7d74 ED |
4023 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4024 | gfp); | |
d2e7b7d0 SS |
4025 | if (!new) |
4026 | return -ENOMEM; | |
4027 | ||
e498be7d | 4028 | for_each_online_cpu(i) { |
7d6e6d09 | 4029 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4030 | batchcount, gfp); |
d2e7b7d0 | 4031 | if (!new->new[i]) { |
b28a02de | 4032 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4033 | kfree(new->new[i]); |
4034 | kfree(new); | |
e498be7d | 4035 | return -ENOMEM; |
1da177e4 LT |
4036 | } |
4037 | } | |
d2e7b7d0 | 4038 | new->cachep = cachep; |
1da177e4 | 4039 | |
15c8b6c1 | 4040 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4041 | |
1da177e4 | 4042 | check_irq_on(); |
1da177e4 LT |
4043 | cachep->batchcount = batchcount; |
4044 | cachep->limit = limit; | |
e498be7d | 4045 | cachep->shared = shared; |
1da177e4 | 4046 | |
e498be7d | 4047 | for_each_online_cpu(i) { |
d2e7b7d0 | 4048 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4049 | if (!ccold) |
4050 | continue; | |
6a67368c | 4051 | spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock); |
7d6e6d09 | 4052 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); |
6a67368c | 4053 | spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock); |
1da177e4 LT |
4054 | kfree(ccold); |
4055 | } | |
d2e7b7d0 | 4056 | kfree(new); |
83b519e8 | 4057 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4058 | } |
4059 | ||
943a451a GC |
4060 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
4061 | int batchcount, int shared, gfp_t gfp) | |
4062 | { | |
4063 | int ret; | |
4064 | struct kmem_cache *c = NULL; | |
4065 | int i = 0; | |
4066 | ||
4067 | ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp); | |
4068 | ||
4069 | if (slab_state < FULL) | |
4070 | return ret; | |
4071 | ||
4072 | if ((ret < 0) || !is_root_cache(cachep)) | |
4073 | return ret; | |
4074 | ||
ebe945c2 | 4075 | VM_BUG_ON(!mutex_is_locked(&slab_mutex)); |
943a451a GC |
4076 | for_each_memcg_cache_index(i) { |
4077 | c = cache_from_memcg(cachep, i); | |
4078 | if (c) | |
4079 | /* return value determined by the parent cache only */ | |
4080 | __do_tune_cpucache(c, limit, batchcount, shared, gfp); | |
4081 | } | |
4082 | ||
4083 | return ret; | |
4084 | } | |
4085 | ||
18004c5d | 4086 | /* Called with slab_mutex held always */ |
83b519e8 | 4087 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4088 | { |
4089 | int err; | |
943a451a GC |
4090 | int limit = 0; |
4091 | int shared = 0; | |
4092 | int batchcount = 0; | |
4093 | ||
4094 | if (!is_root_cache(cachep)) { | |
4095 | struct kmem_cache *root = memcg_root_cache(cachep); | |
4096 | limit = root->limit; | |
4097 | shared = root->shared; | |
4098 | batchcount = root->batchcount; | |
4099 | } | |
1da177e4 | 4100 | |
943a451a GC |
4101 | if (limit && shared && batchcount) |
4102 | goto skip_setup; | |
a737b3e2 AM |
4103 | /* |
4104 | * The head array serves three purposes: | |
1da177e4 LT |
4105 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4106 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4107 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4108 | * bufctl chains: array operations are cheaper. |
4109 | * The numbers are guessed, we should auto-tune as described by | |
4110 | * Bonwick. | |
4111 | */ | |
3b0efdfa | 4112 | if (cachep->size > 131072) |
1da177e4 | 4113 | limit = 1; |
3b0efdfa | 4114 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4115 | limit = 8; |
3b0efdfa | 4116 | else if (cachep->size > 1024) |
1da177e4 | 4117 | limit = 24; |
3b0efdfa | 4118 | else if (cachep->size > 256) |
1da177e4 LT |
4119 | limit = 54; |
4120 | else | |
4121 | limit = 120; | |
4122 | ||
a737b3e2 AM |
4123 | /* |
4124 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4125 | * allocation behaviour: Most allocs on one cpu, most free operations |
4126 | * on another cpu. For these cases, an efficient object passing between | |
4127 | * cpus is necessary. This is provided by a shared array. The array | |
4128 | * replaces Bonwick's magazine layer. | |
4129 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4130 | * to a larger limit. Thus disabled by default. | |
4131 | */ | |
4132 | shared = 0; | |
3b0efdfa | 4133 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4134 | shared = 8; |
1da177e4 LT |
4135 | |
4136 | #if DEBUG | |
a737b3e2 AM |
4137 | /* |
4138 | * With debugging enabled, large batchcount lead to excessively long | |
4139 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4140 | */ |
4141 | if (limit > 32) | |
4142 | limit = 32; | |
4143 | #endif | |
943a451a GC |
4144 | batchcount = (limit + 1) / 2; |
4145 | skip_setup: | |
4146 | err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp); | |
1da177e4 LT |
4147 | if (err) |
4148 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4149 | cachep->name, -err); |
2ed3a4ef | 4150 | return err; |
1da177e4 LT |
4151 | } |
4152 | ||
1b55253a CL |
4153 | /* |
4154 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4155 | * necessary. Note that the l3 listlock also protects the array_cache |
4156 | * if drain_array() is used on the shared array. | |
1b55253a | 4157 | */ |
6744f087 | 4158 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *l3, |
1b55253a | 4159 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4160 | { |
4161 | int tofree; | |
4162 | ||
1b55253a CL |
4163 | if (!ac || !ac->avail) |
4164 | return; | |
1da177e4 LT |
4165 | if (ac->touched && !force) { |
4166 | ac->touched = 0; | |
b18e7e65 | 4167 | } else { |
1b55253a | 4168 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4169 | if (ac->avail) { |
4170 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4171 | if (tofree > ac->avail) | |
4172 | tofree = (ac->avail + 1) / 2; | |
4173 | free_block(cachep, ac->entry, tofree, node); | |
4174 | ac->avail -= tofree; | |
4175 | memmove(ac->entry, &(ac->entry[tofree]), | |
4176 | sizeof(void *) * ac->avail); | |
4177 | } | |
1b55253a | 4178 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4179 | } |
4180 | } | |
4181 | ||
4182 | /** | |
4183 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4184 | * @w: work descriptor |
1da177e4 LT |
4185 | * |
4186 | * Called from workqueue/eventd every few seconds. | |
4187 | * Purpose: | |
4188 | * - clear the per-cpu caches for this CPU. | |
4189 | * - return freeable pages to the main free memory pool. | |
4190 | * | |
a737b3e2 AM |
4191 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4192 | * again on the next iteration. | |
1da177e4 | 4193 | */ |
7c5cae36 | 4194 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4195 | { |
7a7c381d | 4196 | struct kmem_cache *searchp; |
6744f087 | 4197 | struct kmem_cache_node *l3; |
7d6e6d09 | 4198 | int node = numa_mem_id(); |
bf6aede7 | 4199 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4200 | |
18004c5d | 4201 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 4202 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4203 | goto out; |
1da177e4 | 4204 | |
18004c5d | 4205 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
4206 | check_irq_on(); |
4207 | ||
35386e3b CL |
4208 | /* |
4209 | * We only take the l3 lock if absolutely necessary and we | |
4210 | * have established with reasonable certainty that | |
4211 | * we can do some work if the lock was obtained. | |
4212 | */ | |
6a67368c | 4213 | l3 = searchp->node[node]; |
35386e3b | 4214 | |
8fce4d8e | 4215 | reap_alien(searchp, l3); |
1da177e4 | 4216 | |
aab2207c | 4217 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4218 | |
35386e3b CL |
4219 | /* |
4220 | * These are racy checks but it does not matter | |
4221 | * if we skip one check or scan twice. | |
4222 | */ | |
e498be7d | 4223 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4224 | goto next; |
1da177e4 | 4225 | |
e498be7d | 4226 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4227 | |
aab2207c | 4228 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4229 | |
ed11d9eb | 4230 | if (l3->free_touched) |
e498be7d | 4231 | l3->free_touched = 0; |
ed11d9eb CL |
4232 | else { |
4233 | int freed; | |
1da177e4 | 4234 | |
ed11d9eb CL |
4235 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4236 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4237 | STATS_ADD_REAPED(searchp, freed); | |
4238 | } | |
35386e3b | 4239 | next: |
1da177e4 LT |
4240 | cond_resched(); |
4241 | } | |
4242 | check_irq_on(); | |
18004c5d | 4243 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4244 | next_reap_node(); |
7c5cae36 | 4245 | out: |
a737b3e2 | 4246 | /* Set up the next iteration */ |
7c5cae36 | 4247 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4248 | } |
4249 | ||
158a9624 | 4250 | #ifdef CONFIG_SLABINFO |
0d7561c6 | 4251 | void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) |
1da177e4 | 4252 | { |
b28a02de PE |
4253 | struct slab *slabp; |
4254 | unsigned long active_objs; | |
4255 | unsigned long num_objs; | |
4256 | unsigned long active_slabs = 0; | |
4257 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4258 | const char *name; |
1da177e4 | 4259 | char *error = NULL; |
e498be7d | 4260 | int node; |
6744f087 | 4261 | struct kmem_cache_node *l3; |
1da177e4 | 4262 | |
1da177e4 LT |
4263 | active_objs = 0; |
4264 | num_slabs = 0; | |
e498be7d | 4265 | for_each_online_node(node) { |
6a67368c | 4266 | l3 = cachep->node[node]; |
e498be7d CL |
4267 | if (!l3) |
4268 | continue; | |
4269 | ||
ca3b9b91 RT |
4270 | check_irq_on(); |
4271 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4272 | |
7a7c381d | 4273 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4274 | if (slabp->inuse != cachep->num && !error) |
4275 | error = "slabs_full accounting error"; | |
4276 | active_objs += cachep->num; | |
4277 | active_slabs++; | |
4278 | } | |
7a7c381d | 4279 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4280 | if (slabp->inuse == cachep->num && !error) |
4281 | error = "slabs_partial inuse accounting error"; | |
4282 | if (!slabp->inuse && !error) | |
4283 | error = "slabs_partial/inuse accounting error"; | |
4284 | active_objs += slabp->inuse; | |
4285 | active_slabs++; | |
4286 | } | |
7a7c381d | 4287 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4288 | if (slabp->inuse && !error) |
4289 | error = "slabs_free/inuse accounting error"; | |
4290 | num_slabs++; | |
4291 | } | |
4292 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4293 | if (l3->shared) |
4294 | shared_avail += l3->shared->avail; | |
e498be7d | 4295 | |
ca3b9b91 | 4296 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4297 | } |
b28a02de PE |
4298 | num_slabs += active_slabs; |
4299 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4300 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4301 | error = "free_objects accounting error"; |
4302 | ||
b28a02de | 4303 | name = cachep->name; |
1da177e4 LT |
4304 | if (error) |
4305 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4306 | ||
0d7561c6 GC |
4307 | sinfo->active_objs = active_objs; |
4308 | sinfo->num_objs = num_objs; | |
4309 | sinfo->active_slabs = active_slabs; | |
4310 | sinfo->num_slabs = num_slabs; | |
4311 | sinfo->shared_avail = shared_avail; | |
4312 | sinfo->limit = cachep->limit; | |
4313 | sinfo->batchcount = cachep->batchcount; | |
4314 | sinfo->shared = cachep->shared; | |
4315 | sinfo->objects_per_slab = cachep->num; | |
4316 | sinfo->cache_order = cachep->gfporder; | |
4317 | } | |
4318 | ||
4319 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) | |
4320 | { | |
1da177e4 | 4321 | #if STATS |
b28a02de | 4322 | { /* list3 stats */ |
1da177e4 LT |
4323 | unsigned long high = cachep->high_mark; |
4324 | unsigned long allocs = cachep->num_allocations; | |
4325 | unsigned long grown = cachep->grown; | |
4326 | unsigned long reaped = cachep->reaped; | |
4327 | unsigned long errors = cachep->errors; | |
4328 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4329 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4330 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4331 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4332 | |
e92dd4fd JP |
4333 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4334 | "%4lu %4lu %4lu %4lu %4lu", | |
4335 | allocs, high, grown, | |
4336 | reaped, errors, max_freeable, node_allocs, | |
4337 | node_frees, overflows); | |
1da177e4 LT |
4338 | } |
4339 | /* cpu stats */ | |
4340 | { | |
4341 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4342 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4343 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4344 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4345 | ||
4346 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4347 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4348 | } |
4349 | #endif | |
1da177e4 LT |
4350 | } |
4351 | ||
1da177e4 LT |
4352 | #define MAX_SLABINFO_WRITE 128 |
4353 | /** | |
4354 | * slabinfo_write - Tuning for the slab allocator | |
4355 | * @file: unused | |
4356 | * @buffer: user buffer | |
4357 | * @count: data length | |
4358 | * @ppos: unused | |
4359 | */ | |
b7454ad3 | 4360 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4361 | size_t count, loff_t *ppos) |
1da177e4 | 4362 | { |
b28a02de | 4363 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4364 | int limit, batchcount, shared, res; |
7a7c381d | 4365 | struct kmem_cache *cachep; |
b28a02de | 4366 | |
1da177e4 LT |
4367 | if (count > MAX_SLABINFO_WRITE) |
4368 | return -EINVAL; | |
4369 | if (copy_from_user(&kbuf, buffer, count)) | |
4370 | return -EFAULT; | |
b28a02de | 4371 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4372 | |
4373 | tmp = strchr(kbuf, ' '); | |
4374 | if (!tmp) | |
4375 | return -EINVAL; | |
4376 | *tmp = '\0'; | |
4377 | tmp++; | |
4378 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4379 | return -EINVAL; | |
4380 | ||
4381 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4382 | mutex_lock(&slab_mutex); |
1da177e4 | 4383 | res = -EINVAL; |
18004c5d | 4384 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4385 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4386 | if (limit < 1 || batchcount < 1 || |
4387 | batchcount > limit || shared < 0) { | |
e498be7d | 4388 | res = 0; |
1da177e4 | 4389 | } else { |
e498be7d | 4390 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4391 | batchcount, shared, |
4392 | GFP_KERNEL); | |
1da177e4 LT |
4393 | } |
4394 | break; | |
4395 | } | |
4396 | } | |
18004c5d | 4397 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4398 | if (res >= 0) |
4399 | res = count; | |
4400 | return res; | |
4401 | } | |
871751e2 AV |
4402 | |
4403 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4404 | ||
4405 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4406 | { | |
18004c5d CL |
4407 | mutex_lock(&slab_mutex); |
4408 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4409 | } |
4410 | ||
4411 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4412 | { | |
4413 | unsigned long *p; | |
4414 | int l; | |
4415 | if (!v) | |
4416 | return 1; | |
4417 | l = n[1]; | |
4418 | p = n + 2; | |
4419 | while (l) { | |
4420 | int i = l/2; | |
4421 | unsigned long *q = p + 2 * i; | |
4422 | if (*q == v) { | |
4423 | q[1]++; | |
4424 | return 1; | |
4425 | } | |
4426 | if (*q > v) { | |
4427 | l = i; | |
4428 | } else { | |
4429 | p = q + 2; | |
4430 | l -= i + 1; | |
4431 | } | |
4432 | } | |
4433 | if (++n[1] == n[0]) | |
4434 | return 0; | |
4435 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4436 | p[0] = v; | |
4437 | p[1] = 1; | |
4438 | return 1; | |
4439 | } | |
4440 | ||
4441 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4442 | { | |
4443 | void *p; | |
4444 | int i; | |
4445 | if (n[0] == n[1]) | |
4446 | return; | |
3b0efdfa | 4447 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4448 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4449 | continue; | |
4450 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4451 | return; | |
4452 | } | |
4453 | } | |
4454 | ||
4455 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4456 | { | |
4457 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4458 | unsigned long offset, size; |
9281acea | 4459 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4460 | |
a5c43dae | 4461 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4462 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4463 | if (modname[0]) |
871751e2 AV |
4464 | seq_printf(m, " [%s]", modname); |
4465 | return; | |
4466 | } | |
4467 | #endif | |
4468 | seq_printf(m, "%p", (void *)address); | |
4469 | } | |
4470 | ||
4471 | static int leaks_show(struct seq_file *m, void *p) | |
4472 | { | |
0672aa7c | 4473 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 | 4474 | struct slab *slabp; |
6744f087 | 4475 | struct kmem_cache_node *l3; |
871751e2 AV |
4476 | const char *name; |
4477 | unsigned long *n = m->private; | |
4478 | int node; | |
4479 | int i; | |
4480 | ||
4481 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4482 | return 0; | |
4483 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4484 | return 0; | |
4485 | ||
4486 | /* OK, we can do it */ | |
4487 | ||
4488 | n[1] = 0; | |
4489 | ||
4490 | for_each_online_node(node) { | |
6a67368c | 4491 | l3 = cachep->node[node]; |
871751e2 AV |
4492 | if (!l3) |
4493 | continue; | |
4494 | ||
4495 | check_irq_on(); | |
4496 | spin_lock_irq(&l3->list_lock); | |
4497 | ||
7a7c381d | 4498 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4499 | handle_slab(n, cachep, slabp); |
7a7c381d | 4500 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4501 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4502 | spin_unlock_irq(&l3->list_lock); |
4503 | } | |
4504 | name = cachep->name; | |
4505 | if (n[0] == n[1]) { | |
4506 | /* Increase the buffer size */ | |
18004c5d | 4507 | mutex_unlock(&slab_mutex); |
871751e2 AV |
4508 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4509 | if (!m->private) { | |
4510 | /* Too bad, we are really out */ | |
4511 | m->private = n; | |
18004c5d | 4512 | mutex_lock(&slab_mutex); |
871751e2 AV |
4513 | return -ENOMEM; |
4514 | } | |
4515 | *(unsigned long *)m->private = n[0] * 2; | |
4516 | kfree(n); | |
18004c5d | 4517 | mutex_lock(&slab_mutex); |
871751e2 AV |
4518 | /* Now make sure this entry will be retried */ |
4519 | m->count = m->size; | |
4520 | return 0; | |
4521 | } | |
4522 | for (i = 0; i < n[1]; i++) { | |
4523 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4524 | show_symbol(m, n[2*i+2]); | |
4525 | seq_putc(m, '\n'); | |
4526 | } | |
d2e7b7d0 | 4527 | |
871751e2 AV |
4528 | return 0; |
4529 | } | |
4530 | ||
b7454ad3 GC |
4531 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) |
4532 | { | |
4533 | return seq_list_next(p, &slab_caches, pos); | |
4534 | } | |
4535 | ||
4536 | static void s_stop(struct seq_file *m, void *p) | |
4537 | { | |
4538 | mutex_unlock(&slab_mutex); | |
4539 | } | |
4540 | ||
a0ec95a8 | 4541 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4542 | .start = leaks_start, |
4543 | .next = s_next, | |
4544 | .stop = s_stop, | |
4545 | .show = leaks_show, | |
4546 | }; | |
a0ec95a8 AD |
4547 | |
4548 | static int slabstats_open(struct inode *inode, struct file *file) | |
4549 | { | |
4550 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4551 | int ret = -ENOMEM; | |
4552 | if (n) { | |
4553 | ret = seq_open(file, &slabstats_op); | |
4554 | if (!ret) { | |
4555 | struct seq_file *m = file->private_data; | |
4556 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4557 | m->private = n; | |
4558 | n = NULL; | |
4559 | } | |
4560 | kfree(n); | |
4561 | } | |
4562 | return ret; | |
4563 | } | |
4564 | ||
4565 | static const struct file_operations proc_slabstats_operations = { | |
4566 | .open = slabstats_open, | |
4567 | .read = seq_read, | |
4568 | .llseek = seq_lseek, | |
4569 | .release = seq_release_private, | |
4570 | }; | |
4571 | #endif | |
4572 | ||
4573 | static int __init slab_proc_init(void) | |
4574 | { | |
4575 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4576 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4577 | #endif |
a0ec95a8 AD |
4578 | return 0; |
4579 | } | |
4580 | module_init(slab_proc_init); | |
1da177e4 LT |
4581 | #endif |
4582 | ||
00e145b6 MS |
4583 | /** |
4584 | * ksize - get the actual amount of memory allocated for a given object | |
4585 | * @objp: Pointer to the object | |
4586 | * | |
4587 | * kmalloc may internally round up allocations and return more memory | |
4588 | * than requested. ksize() can be used to determine the actual amount of | |
4589 | * memory allocated. The caller may use this additional memory, even though | |
4590 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4591 | * The caller must guarantee that objp points to a valid object previously | |
4592 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4593 | * must not be freed during the duration of the call. | |
4594 | */ | |
fd76bab2 | 4595 | size_t ksize(const void *objp) |
1da177e4 | 4596 | { |
ef8b4520 CL |
4597 | BUG_ON(!objp); |
4598 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4599 | return 0; |
1da177e4 | 4600 | |
8c138bc0 | 4601 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4602 | } |
b1aabecd | 4603 | EXPORT_SYMBOL(ksize); |