]>
Commit | Line | Data |
---|---|---|
1da177e4 LT |
1 | /* |
2 | * linux/mm/slab.c | |
3 | * Written by Mark Hemment, 1996/97. | |
4 | * (markhe@nextd.demon.co.uk) | |
5 | * | |
6 | * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli | |
7 | * | |
8 | * Major cleanup, different bufctl logic, per-cpu arrays | |
9 | * (c) 2000 Manfred Spraul | |
10 | * | |
11 | * Cleanup, make the head arrays unconditional, preparation for NUMA | |
12 | * (c) 2002 Manfred Spraul | |
13 | * | |
14 | * An implementation of the Slab Allocator as described in outline in; | |
15 | * UNIX Internals: The New Frontiers by Uresh Vahalia | |
16 | * Pub: Prentice Hall ISBN 0-13-101908-2 | |
17 | * or with a little more detail in; | |
18 | * The Slab Allocator: An Object-Caching Kernel Memory Allocator | |
19 | * Jeff Bonwick (Sun Microsystems). | |
20 | * Presented at: USENIX Summer 1994 Technical Conference | |
21 | * | |
22 | * The memory is organized in caches, one cache for each object type. | |
23 | * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct) | |
24 | * Each cache consists out of many slabs (they are small (usually one | |
25 | * page long) and always contiguous), and each slab contains multiple | |
26 | * initialized objects. | |
27 | * | |
28 | * This means, that your constructor is used only for newly allocated | |
183ff22b | 29 | * slabs and you must pass objects with the same initializations to |
1da177e4 LT |
30 | * kmem_cache_free. |
31 | * | |
32 | * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM, | |
33 | * normal). If you need a special memory type, then must create a new | |
34 | * cache for that memory type. | |
35 | * | |
36 | * In order to reduce fragmentation, the slabs are sorted in 3 groups: | |
37 | * full slabs with 0 free objects | |
38 | * partial slabs | |
39 | * empty slabs with no allocated objects | |
40 | * | |
41 | * If partial slabs exist, then new allocations come from these slabs, | |
42 | * otherwise from empty slabs or new slabs are allocated. | |
43 | * | |
44 | * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache | |
45 | * during kmem_cache_destroy(). The caller must prevent concurrent allocs. | |
46 | * | |
47 | * Each cache has a short per-cpu head array, most allocs | |
48 | * and frees go into that array, and if that array overflows, then 1/2 | |
49 | * of the entries in the array are given back into the global cache. | |
50 | * The head array is strictly LIFO and should improve the cache hit rates. | |
51 | * On SMP, it additionally reduces the spinlock operations. | |
52 | * | |
a737b3e2 | 53 | * The c_cpuarray may not be read with enabled local interrupts - |
1da177e4 LT |
54 | * it's changed with a smp_call_function(). |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
343e0d7a | 58 | * Several members in struct kmem_cache and struct slab never change, they |
1da177e4 LT |
59 | * are accessed without any locking. |
60 | * The per-cpu arrays are never accessed from the wrong cpu, no locking, | |
61 | * and local interrupts are disabled so slab code is preempt-safe. | |
62 | * The non-constant members are protected with a per-cache irq spinlock. | |
63 | * | |
64 | * Many thanks to Mark Hemment, who wrote another per-cpu slab patch | |
65 | * in 2000 - many ideas in the current implementation are derived from | |
66 | * his patch. | |
67 | * | |
68 | * Further notes from the original documentation: | |
69 | * | |
70 | * 11 April '97. Started multi-threading - markhe | |
18004c5d | 71 | * The global cache-chain is protected by the mutex 'slab_mutex'. |
1da177e4 LT |
72 | * The sem is only needed when accessing/extending the cache-chain, which |
73 | * can never happen inside an interrupt (kmem_cache_create(), | |
74 | * kmem_cache_shrink() and kmem_cache_reap()). | |
75 | * | |
76 | * At present, each engine can be growing a cache. This should be blocked. | |
77 | * | |
e498be7d CL |
78 | * 15 March 2005. NUMA slab allocator. |
79 | * Shai Fultheim <shai@scalex86.org>. | |
80 | * Shobhit Dayal <shobhit@calsoftinc.com> | |
81 | * Alok N Kataria <alokk@calsoftinc.com> | |
82 | * Christoph Lameter <christoph@lameter.com> | |
83 | * | |
84 | * Modified the slab allocator to be node aware on NUMA systems. | |
85 | * Each node has its own list of partial, free and full slabs. | |
86 | * All object allocations for a node occur from node specific slab lists. | |
1da177e4 LT |
87 | */ |
88 | ||
1da177e4 | 89 | #include <linux/slab.h> |
97d06609 | 90 | #include "slab.h" |
1da177e4 | 91 | #include <linux/mm.h> |
c9cf5528 | 92 | #include <linux/poison.h> |
1da177e4 LT |
93 | #include <linux/swap.h> |
94 | #include <linux/cache.h> | |
95 | #include <linux/interrupt.h> | |
96 | #include <linux/init.h> | |
97 | #include <linux/compiler.h> | |
101a5001 | 98 | #include <linux/cpuset.h> |
a0ec95a8 | 99 | #include <linux/proc_fs.h> |
1da177e4 LT |
100 | #include <linux/seq_file.h> |
101 | #include <linux/notifier.h> | |
102 | #include <linux/kallsyms.h> | |
103 | #include <linux/cpu.h> | |
104 | #include <linux/sysctl.h> | |
105 | #include <linux/module.h> | |
106 | #include <linux/rcupdate.h> | |
543537bd | 107 | #include <linux/string.h> |
138ae663 | 108 | #include <linux/uaccess.h> |
e498be7d | 109 | #include <linux/nodemask.h> |
d5cff635 | 110 | #include <linux/kmemleak.h> |
dc85da15 | 111 | #include <linux/mempolicy.h> |
fc0abb14 | 112 | #include <linux/mutex.h> |
8a8b6502 | 113 | #include <linux/fault-inject.h> |
e7eebaf6 | 114 | #include <linux/rtmutex.h> |
6a2d7a95 | 115 | #include <linux/reciprocal_div.h> |
3ac7fe5a | 116 | #include <linux/debugobjects.h> |
c175eea4 | 117 | #include <linux/kmemcheck.h> |
8f9f8d9e | 118 | #include <linux/memory.h> |
268bb0ce | 119 | #include <linux/prefetch.h> |
1da177e4 | 120 | |
381760ea MG |
121 | #include <net/sock.h> |
122 | ||
1da177e4 LT |
123 | #include <asm/cacheflush.h> |
124 | #include <asm/tlbflush.h> | |
125 | #include <asm/page.h> | |
126 | ||
4dee6b64 SR |
127 | #include <trace/events/kmem.h> |
128 | ||
072bb0aa MG |
129 | #include "internal.h" |
130 | ||
1da177e4 | 131 | /* |
50953fe9 | 132 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
1da177e4 LT |
133 | * 0 for faster, smaller code (especially in the critical paths). |
134 | * | |
135 | * STATS - 1 to collect stats for /proc/slabinfo. | |
136 | * 0 for faster, smaller code (especially in the critical paths). | |
137 | * | |
138 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) | |
139 | */ | |
140 | ||
141 | #ifdef CONFIG_DEBUG_SLAB | |
142 | #define DEBUG 1 | |
143 | #define STATS 1 | |
144 | #define FORCED_DEBUG 1 | |
145 | #else | |
146 | #define DEBUG 0 | |
147 | #define STATS 0 | |
148 | #define FORCED_DEBUG 0 | |
149 | #endif | |
150 | ||
1da177e4 LT |
151 | /* Shouldn't this be in a header file somewhere? */ |
152 | #define BYTES_PER_WORD sizeof(void *) | |
87a927c7 | 153 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
1da177e4 | 154 | |
1da177e4 LT |
155 | #ifndef ARCH_KMALLOC_FLAGS |
156 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
157 | #endif | |
158 | ||
072bb0aa MG |
159 | /* |
160 | * true if a page was allocated from pfmemalloc reserves for network-based | |
161 | * swap | |
162 | */ | |
163 | static bool pfmemalloc_active __read_mostly; | |
164 | ||
1da177e4 LT |
165 | /* Legal flag mask for kmem_cache_create(). */ |
166 | #if DEBUG | |
50953fe9 | 167 | # define CREATE_MASK (SLAB_RED_ZONE | \ |
1da177e4 | 168 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ |
ac2b898c | 169 | SLAB_CACHE_DMA | \ |
5af60839 | 170 | SLAB_STORE_USER | \ |
1da177e4 | 171 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 172 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 173 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 | 174 | #else |
ac2b898c | 175 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ |
5af60839 | 176 | SLAB_CACHE_DMA | \ |
1da177e4 | 177 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ |
3ac7fe5a | 178 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ |
c175eea4 | 179 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) |
1da177e4 LT |
180 | #endif |
181 | ||
182 | /* | |
183 | * kmem_bufctl_t: | |
184 | * | |
185 | * Bufctl's are used for linking objs within a slab | |
186 | * linked offsets. | |
187 | * | |
188 | * This implementation relies on "struct page" for locating the cache & | |
189 | * slab an object belongs to. | |
190 | * This allows the bufctl structure to be small (one int), but limits | |
191 | * the number of objects a slab (not a cache) can contain when off-slab | |
192 | * bufctls are used. The limit is the size of the largest general cache | |
193 | * that does not use off-slab slabs. | |
194 | * For 32bit archs with 4 kB pages, is this 56. | |
195 | * This is not serious, as it is only for large objects, when it is unwise | |
196 | * to have too many per slab. | |
197 | * Note: This limit can be raised by introducing a general cache whose size | |
198 | * is less than 512 (PAGE_SIZE<<3), but greater than 256. | |
199 | */ | |
200 | ||
fa5b08d5 | 201 | typedef unsigned int kmem_bufctl_t; |
1da177e4 LT |
202 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) |
203 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
871751e2 AV |
204 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) |
205 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
1da177e4 | 206 | |
1da177e4 LT |
207 | /* |
208 | * struct slab_rcu | |
209 | * | |
210 | * slab_destroy on a SLAB_DESTROY_BY_RCU cache uses this structure to | |
211 | * arrange for kmem_freepages to be called via RCU. This is useful if | |
212 | * we need to approach a kernel structure obliquely, from its address | |
213 | * obtained without the usual locking. We can lock the structure to | |
214 | * stabilize it and check it's still at the given address, only if we | |
215 | * can be sure that the memory has not been meanwhile reused for some | |
216 | * other kind of object (which our subsystem's lock might corrupt). | |
217 | * | |
218 | * rcu_read_lock before reading the address, then rcu_read_unlock after | |
219 | * taking the spinlock within the structure expected at that address. | |
1da177e4 LT |
220 | */ |
221 | struct slab_rcu { | |
b28a02de | 222 | struct rcu_head head; |
343e0d7a | 223 | struct kmem_cache *cachep; |
b28a02de | 224 | void *addr; |
1da177e4 LT |
225 | }; |
226 | ||
5bfe53a7 LJ |
227 | /* |
228 | * struct slab | |
229 | * | |
230 | * Manages the objs in a slab. Placed either at the beginning of mem allocated | |
231 | * for a slab, or allocated from an general cache. | |
232 | * Slabs are chained into three list: fully used, partial, fully free slabs. | |
233 | */ | |
234 | struct slab { | |
235 | union { | |
236 | struct { | |
237 | struct list_head list; | |
238 | unsigned long colouroff; | |
239 | void *s_mem; /* including colour offset */ | |
240 | unsigned int inuse; /* num of objs active in slab */ | |
241 | kmem_bufctl_t free; | |
242 | unsigned short nodeid; | |
243 | }; | |
244 | struct slab_rcu __slab_cover_slab_rcu; | |
245 | }; | |
246 | }; | |
247 | ||
1da177e4 LT |
248 | /* |
249 | * struct array_cache | |
250 | * | |
1da177e4 LT |
251 | * Purpose: |
252 | * - LIFO ordering, to hand out cache-warm objects from _alloc | |
253 | * - reduce the number of linked list operations | |
254 | * - reduce spinlock operations | |
255 | * | |
256 | * The limit is stored in the per-cpu structure to reduce the data cache | |
257 | * footprint. | |
258 | * | |
259 | */ | |
260 | struct array_cache { | |
261 | unsigned int avail; | |
262 | unsigned int limit; | |
263 | unsigned int batchcount; | |
264 | unsigned int touched; | |
e498be7d | 265 | spinlock_t lock; |
bda5b655 | 266 | void *entry[]; /* |
a737b3e2 AM |
267 | * Must have this definition in here for the proper |
268 | * alignment of array_cache. Also simplifies accessing | |
269 | * the entries. | |
072bb0aa MG |
270 | * |
271 | * Entries should not be directly dereferenced as | |
272 | * entries belonging to slabs marked pfmemalloc will | |
273 | * have the lower bits set SLAB_OBJ_PFMEMALLOC | |
a737b3e2 | 274 | */ |
1da177e4 LT |
275 | }; |
276 | ||
072bb0aa MG |
277 | #define SLAB_OBJ_PFMEMALLOC 1 |
278 | static inline bool is_obj_pfmemalloc(void *objp) | |
279 | { | |
280 | return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC; | |
281 | } | |
282 | ||
283 | static inline void set_obj_pfmemalloc(void **objp) | |
284 | { | |
285 | *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC); | |
286 | return; | |
287 | } | |
288 | ||
289 | static inline void clear_obj_pfmemalloc(void **objp) | |
290 | { | |
291 | *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC); | |
292 | } | |
293 | ||
a737b3e2 AM |
294 | /* |
295 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
296 | * cpuarrays are allocated from the generic caches... | |
1da177e4 LT |
297 | */ |
298 | #define BOOT_CPUCACHE_ENTRIES 1 | |
299 | struct arraycache_init { | |
300 | struct array_cache cache; | |
b28a02de | 301 | void *entries[BOOT_CPUCACHE_ENTRIES]; |
1da177e4 LT |
302 | }; |
303 | ||
304 | /* | |
e498be7d | 305 | * The slab lists for all objects. |
1da177e4 LT |
306 | */ |
307 | struct kmem_list3 { | |
b28a02de PE |
308 | struct list_head slabs_partial; /* partial list first, better asm code */ |
309 | struct list_head slabs_full; | |
310 | struct list_head slabs_free; | |
311 | unsigned long free_objects; | |
b28a02de | 312 | unsigned int free_limit; |
2e1217cf | 313 | unsigned int colour_next; /* Per-node cache coloring */ |
b28a02de PE |
314 | spinlock_t list_lock; |
315 | struct array_cache *shared; /* shared per node */ | |
316 | struct array_cache **alien; /* on other nodes */ | |
35386e3b CL |
317 | unsigned long next_reap; /* updated without locking */ |
318 | int free_touched; /* updated without locking */ | |
1da177e4 LT |
319 | }; |
320 | ||
e498be7d CL |
321 | /* |
322 | * Need this for bootstrapping a per node allocator. | |
323 | */ | |
556a169d | 324 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) |
68a1b195 | 325 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; |
e498be7d | 326 | #define CACHE_CACHE 0 |
556a169d PE |
327 | #define SIZE_AC MAX_NUMNODES |
328 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
e498be7d | 329 | |
ed11d9eb CL |
330 | static int drain_freelist(struct kmem_cache *cache, |
331 | struct kmem_list3 *l3, int tofree); | |
332 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, | |
333 | int node); | |
83b519e8 | 334 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
65f27f38 | 335 | static void cache_reap(struct work_struct *unused); |
ed11d9eb | 336 | |
e498be7d | 337 | /* |
a737b3e2 AM |
338 | * This function must be completely optimized away if a constant is passed to |
339 | * it. Mostly the same as what is in linux/slab.h except it returns an index. | |
e498be7d | 340 | */ |
7243cc05 | 341 | static __always_inline int index_of(const size_t size) |
e498be7d | 342 | { |
5ec8a847 SR |
343 | extern void __bad_size(void); |
344 | ||
e498be7d CL |
345 | if (__builtin_constant_p(size)) { |
346 | int i = 0; | |
347 | ||
348 | #define CACHE(x) \ | |
349 | if (size <=x) \ | |
350 | return i; \ | |
351 | else \ | |
352 | i++; | |
1c61fc40 | 353 | #include <linux/kmalloc_sizes.h> |
e498be7d | 354 | #undef CACHE |
5ec8a847 | 355 | __bad_size(); |
7243cc05 | 356 | } else |
5ec8a847 | 357 | __bad_size(); |
e498be7d CL |
358 | return 0; |
359 | } | |
360 | ||
e0a42726 IM |
361 | static int slab_early_init = 1; |
362 | ||
e498be7d CL |
363 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) |
364 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
1da177e4 | 365 | |
5295a74c | 366 | static void kmem_list3_init(struct kmem_list3 *parent) |
e498be7d CL |
367 | { |
368 | INIT_LIST_HEAD(&parent->slabs_full); | |
369 | INIT_LIST_HEAD(&parent->slabs_partial); | |
370 | INIT_LIST_HEAD(&parent->slabs_free); | |
371 | parent->shared = NULL; | |
372 | parent->alien = NULL; | |
2e1217cf | 373 | parent->colour_next = 0; |
e498be7d CL |
374 | spin_lock_init(&parent->list_lock); |
375 | parent->free_objects = 0; | |
376 | parent->free_touched = 0; | |
377 | } | |
378 | ||
a737b3e2 AM |
379 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
380 | do { \ | |
381 | INIT_LIST_HEAD(listp); \ | |
382 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
e498be7d CL |
383 | } while (0) |
384 | ||
a737b3e2 AM |
385 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
386 | do { \ | |
e498be7d CL |
387 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
388 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ | |
389 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ | |
390 | } while (0) | |
1da177e4 | 391 | |
1da177e4 LT |
392 | #define CFLGS_OFF_SLAB (0x80000000UL) |
393 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
394 | ||
395 | #define BATCHREFILL_LIMIT 16 | |
a737b3e2 AM |
396 | /* |
397 | * Optimization question: fewer reaps means less probability for unnessary | |
398 | * cpucache drain/refill cycles. | |
1da177e4 | 399 | * |
dc6f3f27 | 400 | * OTOH the cpuarrays can contain lots of objects, |
1da177e4 LT |
401 | * which could lock up otherwise freeable slabs. |
402 | */ | |
403 | #define REAPTIMEOUT_CPUC (2*HZ) | |
404 | #define REAPTIMEOUT_LIST3 (4*HZ) | |
405 | ||
406 | #if STATS | |
407 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) | |
408 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) | |
409 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) | |
410 | #define STATS_INC_GROWN(x) ((x)->grown++) | |
ed11d9eb | 411 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
a737b3e2 AM |
412 | #define STATS_SET_HIGH(x) \ |
413 | do { \ | |
414 | if ((x)->num_active > (x)->high_mark) \ | |
415 | (x)->high_mark = (x)->num_active; \ | |
416 | } while (0) | |
1da177e4 LT |
417 | #define STATS_INC_ERR(x) ((x)->errors++) |
418 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
e498be7d | 419 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
fb7faf33 | 420 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
a737b3e2 AM |
421 | #define STATS_SET_FREEABLE(x, i) \ |
422 | do { \ | |
423 | if ((x)->max_freeable < i) \ | |
424 | (x)->max_freeable = i; \ | |
425 | } while (0) | |
1da177e4 LT |
426 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
427 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) | |
428 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) | |
429 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) | |
430 | #else | |
431 | #define STATS_INC_ACTIVE(x) do { } while (0) | |
432 | #define STATS_DEC_ACTIVE(x) do { } while (0) | |
433 | #define STATS_INC_ALLOCED(x) do { } while (0) | |
434 | #define STATS_INC_GROWN(x) do { } while (0) | |
4e60c86b | 435 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
1da177e4 LT |
436 | #define STATS_SET_HIGH(x) do { } while (0) |
437 | #define STATS_INC_ERR(x) do { } while (0) | |
438 | #define STATS_INC_NODEALLOCS(x) do { } while (0) | |
e498be7d | 439 | #define STATS_INC_NODEFREES(x) do { } while (0) |
fb7faf33 | 440 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
a737b3e2 | 441 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
1da177e4 LT |
442 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
443 | #define STATS_INC_ALLOCMISS(x) do { } while (0) | |
444 | #define STATS_INC_FREEHIT(x) do { } while (0) | |
445 | #define STATS_INC_FREEMISS(x) do { } while (0) | |
446 | #endif | |
447 | ||
448 | #if DEBUG | |
1da177e4 | 449 | |
a737b3e2 AM |
450 | /* |
451 | * memory layout of objects: | |
1da177e4 | 452 | * 0 : objp |
3dafccf2 | 453 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
1da177e4 LT |
454 | * the end of an object is aligned with the end of the real |
455 | * allocation. Catches writes behind the end of the allocation. | |
3dafccf2 | 456 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
1da177e4 | 457 | * redzone word. |
3dafccf2 | 458 | * cachep->obj_offset: The real object. |
3b0efdfa CL |
459 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
460 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
a737b3e2 | 461 | * [BYTES_PER_WORD long] |
1da177e4 | 462 | */ |
343e0d7a | 463 | static int obj_offset(struct kmem_cache *cachep) |
1da177e4 | 464 | { |
3dafccf2 | 465 | return cachep->obj_offset; |
1da177e4 LT |
466 | } |
467 | ||
b46b8f19 | 468 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
469 | { |
470 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
b46b8f19 DW |
471 | return (unsigned long long*) (objp + obj_offset(cachep) - |
472 | sizeof(unsigned long long)); | |
1da177e4 LT |
473 | } |
474 | ||
b46b8f19 | 475 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
476 | { |
477 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
478 | if (cachep->flags & SLAB_STORE_USER) | |
3b0efdfa | 479 | return (unsigned long long *)(objp + cachep->size - |
b46b8f19 | 480 | sizeof(unsigned long long) - |
87a927c7 | 481 | REDZONE_ALIGN); |
3b0efdfa | 482 | return (unsigned long long *) (objp + cachep->size - |
b46b8f19 | 483 | sizeof(unsigned long long)); |
1da177e4 LT |
484 | } |
485 | ||
343e0d7a | 486 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
487 | { |
488 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
3b0efdfa | 489 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
1da177e4 LT |
490 | } |
491 | ||
492 | #else | |
493 | ||
3dafccf2 | 494 | #define obj_offset(x) 0 |
b46b8f19 DW |
495 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
496 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
1da177e4 LT |
497 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
498 | ||
499 | #endif | |
500 | ||
0f24f128 | 501 | #ifdef CONFIG_TRACING |
36555751 EGM |
502 | size_t slab_buffer_size(struct kmem_cache *cachep) |
503 | { | |
3b0efdfa | 504 | return cachep->size; |
36555751 EGM |
505 | } |
506 | EXPORT_SYMBOL(slab_buffer_size); | |
507 | #endif | |
508 | ||
1da177e4 | 509 | /* |
3df1cccd DR |
510 | * Do not go above this order unless 0 objects fit into the slab or |
511 | * overridden on the command line. | |
1da177e4 | 512 | */ |
543585cc DR |
513 | #define SLAB_MAX_ORDER_HI 1 |
514 | #define SLAB_MAX_ORDER_LO 0 | |
515 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
3df1cccd | 516 | static bool slab_max_order_set __initdata; |
1da177e4 | 517 | |
6ed5eb22 PE |
518 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
519 | { | |
b49af68f | 520 | struct page *page = virt_to_head_page(obj); |
35026088 | 521 | return page->slab_cache; |
6ed5eb22 PE |
522 | } |
523 | ||
524 | static inline struct slab *virt_to_slab(const void *obj) | |
525 | { | |
b49af68f | 526 | struct page *page = virt_to_head_page(obj); |
35026088 CL |
527 | |
528 | VM_BUG_ON(!PageSlab(page)); | |
529 | return page->slab_page; | |
6ed5eb22 PE |
530 | } |
531 | ||
8fea4e96 PE |
532 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, |
533 | unsigned int idx) | |
534 | { | |
3b0efdfa | 535 | return slab->s_mem + cache->size * idx; |
8fea4e96 PE |
536 | } |
537 | ||
6a2d7a95 | 538 | /* |
3b0efdfa CL |
539 | * We want to avoid an expensive divide : (offset / cache->size) |
540 | * Using the fact that size is a constant for a particular cache, | |
541 | * we can replace (offset / cache->size) by | |
6a2d7a95 ED |
542 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
543 | */ | |
544 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
545 | const struct slab *slab, void *obj) | |
8fea4e96 | 546 | { |
6a2d7a95 ED |
547 | u32 offset = (obj - slab->s_mem); |
548 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
8fea4e96 PE |
549 | } |
550 | ||
a737b3e2 AM |
551 | /* |
552 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
553 | */ | |
1da177e4 LT |
554 | struct cache_sizes malloc_sizes[] = { |
555 | #define CACHE(x) { .cs_size = (x) }, | |
556 | #include <linux/kmalloc_sizes.h> | |
557 | CACHE(ULONG_MAX) | |
558 | #undef CACHE | |
559 | }; | |
560 | EXPORT_SYMBOL(malloc_sizes); | |
561 | ||
562 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
563 | struct cache_names { | |
564 | char *name; | |
565 | char *name_dma; | |
566 | }; | |
567 | ||
568 | static struct cache_names __initdata cache_names[] = { | |
569 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
570 | #include <linux/kmalloc_sizes.h> | |
b28a02de | 571 | {NULL,} |
1da177e4 LT |
572 | #undef CACHE |
573 | }; | |
574 | ||
575 | static struct arraycache_init initarray_cache __initdata = | |
b28a02de | 576 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 | 577 | static struct arraycache_init initarray_generic = |
b28a02de | 578 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; |
1da177e4 LT |
579 | |
580 | /* internal cache of cache description objs */ | |
9b030cb8 CL |
581 | static struct kmem_list3 *kmem_cache_nodelists[MAX_NUMNODES]; |
582 | static struct kmem_cache kmem_cache_boot = { | |
583 | .nodelists = kmem_cache_nodelists, | |
b28a02de PE |
584 | .batchcount = 1, |
585 | .limit = BOOT_CPUCACHE_ENTRIES, | |
586 | .shared = 1, | |
3b0efdfa | 587 | .size = sizeof(struct kmem_cache), |
b28a02de | 588 | .name = "kmem_cache", |
1da177e4 LT |
589 | }; |
590 | ||
056c6241 RT |
591 | #define BAD_ALIEN_MAGIC 0x01020304ul |
592 | ||
f1aaee53 AV |
593 | #ifdef CONFIG_LOCKDEP |
594 | ||
595 | /* | |
596 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
597 | * for other slabs "off slab". | |
598 | * The locking for this is tricky in that it nests within the locks | |
599 | * of all other slabs in a few places; to deal with this special | |
600 | * locking we put on-slab caches into a separate lock-class. | |
056c6241 RT |
601 | * |
602 | * We set lock class for alien array caches which are up during init. | |
603 | * The lock annotation will be lost if all cpus of a node goes down and | |
604 | * then comes back up during hotplug | |
f1aaee53 | 605 | */ |
056c6241 RT |
606 | static struct lock_class_key on_slab_l3_key; |
607 | static struct lock_class_key on_slab_alc_key; | |
608 | ||
83835b3d PZ |
609 | static struct lock_class_key debugobj_l3_key; |
610 | static struct lock_class_key debugobj_alc_key; | |
611 | ||
612 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
613 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
614 | int q) | |
615 | { | |
616 | struct array_cache **alc; | |
617 | struct kmem_list3 *l3; | |
618 | int r; | |
619 | ||
620 | l3 = cachep->nodelists[q]; | |
621 | if (!l3) | |
622 | return; | |
623 | ||
624 | lockdep_set_class(&l3->list_lock, l3_key); | |
625 | alc = l3->alien; | |
626 | /* | |
627 | * FIXME: This check for BAD_ALIEN_MAGIC | |
628 | * should go away when common slab code is taught to | |
629 | * work even without alien caches. | |
630 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
631 | * for alloc_alien_cache, | |
632 | */ | |
633 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
634 | return; | |
635 | for_each_node(r) { | |
636 | if (alc[r]) | |
637 | lockdep_set_class(&alc[r]->lock, alc_key); | |
638 | } | |
639 | } | |
640 | ||
641 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
642 | { | |
643 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
644 | } | |
645 | ||
646 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
647 | { | |
648 | int node; | |
649 | ||
650 | for_each_online_node(node) | |
651 | slab_set_debugobj_lock_classes_node(cachep, node); | |
652 | } | |
653 | ||
ce79ddc8 | 654 | static void init_node_lock_keys(int q) |
f1aaee53 | 655 | { |
056c6241 RT |
656 | struct cache_sizes *s = malloc_sizes; |
657 | ||
97d06609 | 658 | if (slab_state < UP) |
ce79ddc8 PE |
659 | return; |
660 | ||
661 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
ce79ddc8 | 662 | struct kmem_list3 *l3; |
ce79ddc8 PE |
663 | |
664 | l3 = s->cs_cachep->nodelists[q]; | |
665 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
00afa758 | 666 | continue; |
83835b3d PZ |
667 | |
668 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
669 | &on_slab_alc_key, q); | |
f1aaee53 AV |
670 | } |
671 | } | |
ce79ddc8 PE |
672 | |
673 | static inline void init_lock_keys(void) | |
674 | { | |
675 | int node; | |
676 | ||
677 | for_each_node(node) | |
678 | init_node_lock_keys(node); | |
679 | } | |
f1aaee53 | 680 | #else |
ce79ddc8 PE |
681 | static void init_node_lock_keys(int q) |
682 | { | |
683 | } | |
684 | ||
056c6241 | 685 | static inline void init_lock_keys(void) |
f1aaee53 AV |
686 | { |
687 | } | |
83835b3d PZ |
688 | |
689 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
690 | { | |
691 | } | |
692 | ||
693 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
694 | { | |
695 | } | |
f1aaee53 AV |
696 | #endif |
697 | ||
1871e52c | 698 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
1da177e4 | 699 | |
343e0d7a | 700 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
1da177e4 LT |
701 | { |
702 | return cachep->array[smp_processor_id()]; | |
703 | } | |
704 | ||
a737b3e2 AM |
705 | static inline struct kmem_cache *__find_general_cachep(size_t size, |
706 | gfp_t gfpflags) | |
1da177e4 LT |
707 | { |
708 | struct cache_sizes *csizep = malloc_sizes; | |
709 | ||
710 | #if DEBUG | |
711 | /* This happens if someone tries to call | |
b28a02de PE |
712 | * kmem_cache_create(), or __kmalloc(), before |
713 | * the generic caches are initialized. | |
714 | */ | |
c7e43c78 | 715 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); |
1da177e4 | 716 | #endif |
6cb8f913 CL |
717 | if (!size) |
718 | return ZERO_SIZE_PTR; | |
719 | ||
1da177e4 LT |
720 | while (size > csizep->cs_size) |
721 | csizep++; | |
722 | ||
723 | /* | |
0abf40c1 | 724 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX |
1da177e4 LT |
725 | * has cs_{dma,}cachep==NULL. Thus no special case |
726 | * for large kmalloc calls required. | |
727 | */ | |
4b51d669 | 728 | #ifdef CONFIG_ZONE_DMA |
1da177e4 LT |
729 | if (unlikely(gfpflags & GFP_DMA)) |
730 | return csizep->cs_dmacachep; | |
4b51d669 | 731 | #endif |
1da177e4 LT |
732 | return csizep->cs_cachep; |
733 | } | |
734 | ||
b221385b | 735 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) |
97e2bde4 MS |
736 | { |
737 | return __find_general_cachep(size, gfpflags); | |
738 | } | |
97e2bde4 | 739 | |
fbaccacf | 740 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) |
1da177e4 | 741 | { |
fbaccacf SR |
742 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); |
743 | } | |
1da177e4 | 744 | |
a737b3e2 AM |
745 | /* |
746 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
747 | */ | |
fbaccacf SR |
748 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, |
749 | size_t align, int flags, size_t *left_over, | |
750 | unsigned int *num) | |
751 | { | |
752 | int nr_objs; | |
753 | size_t mgmt_size; | |
754 | size_t slab_size = PAGE_SIZE << gfporder; | |
1da177e4 | 755 | |
fbaccacf SR |
756 | /* |
757 | * The slab management structure can be either off the slab or | |
758 | * on it. For the latter case, the memory allocated for a | |
759 | * slab is used for: | |
760 | * | |
761 | * - The struct slab | |
762 | * - One kmem_bufctl_t for each object | |
763 | * - Padding to respect alignment of @align | |
764 | * - @buffer_size bytes for each object | |
765 | * | |
766 | * If the slab management structure is off the slab, then the | |
767 | * alignment will already be calculated into the size. Because | |
768 | * the slabs are all pages aligned, the objects will be at the | |
769 | * correct alignment when allocated. | |
770 | */ | |
771 | if (flags & CFLGS_OFF_SLAB) { | |
772 | mgmt_size = 0; | |
773 | nr_objs = slab_size / buffer_size; | |
774 | ||
775 | if (nr_objs > SLAB_LIMIT) | |
776 | nr_objs = SLAB_LIMIT; | |
777 | } else { | |
778 | /* | |
779 | * Ignore padding for the initial guess. The padding | |
780 | * is at most @align-1 bytes, and @buffer_size is at | |
781 | * least @align. In the worst case, this result will | |
782 | * be one greater than the number of objects that fit | |
783 | * into the memory allocation when taking the padding | |
784 | * into account. | |
785 | */ | |
786 | nr_objs = (slab_size - sizeof(struct slab)) / | |
787 | (buffer_size + sizeof(kmem_bufctl_t)); | |
788 | ||
789 | /* | |
790 | * This calculated number will be either the right | |
791 | * amount, or one greater than what we want. | |
792 | */ | |
793 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
794 | > slab_size) | |
795 | nr_objs--; | |
796 | ||
797 | if (nr_objs > SLAB_LIMIT) | |
798 | nr_objs = SLAB_LIMIT; | |
799 | ||
800 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
801 | } | |
802 | *num = nr_objs; | |
803 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
1da177e4 LT |
804 | } |
805 | ||
d40cee24 | 806 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
1da177e4 | 807 | |
a737b3e2 AM |
808 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
809 | char *msg) | |
1da177e4 LT |
810 | { |
811 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
b28a02de | 812 | function, cachep->name, msg); |
1da177e4 LT |
813 | dump_stack(); |
814 | } | |
815 | ||
3395ee05 PM |
816 | /* |
817 | * By default on NUMA we use alien caches to stage the freeing of | |
818 | * objects allocated from other nodes. This causes massive memory | |
819 | * inefficiencies when using fake NUMA setup to split memory into a | |
820 | * large number of small nodes, so it can be disabled on the command | |
821 | * line | |
822 | */ | |
823 | ||
824 | static int use_alien_caches __read_mostly = 1; | |
825 | static int __init noaliencache_setup(char *s) | |
826 | { | |
827 | use_alien_caches = 0; | |
828 | return 1; | |
829 | } | |
830 | __setup("noaliencache", noaliencache_setup); | |
831 | ||
3df1cccd DR |
832 | static int __init slab_max_order_setup(char *str) |
833 | { | |
834 | get_option(&str, &slab_max_order); | |
835 | slab_max_order = slab_max_order < 0 ? 0 : | |
836 | min(slab_max_order, MAX_ORDER - 1); | |
837 | slab_max_order_set = true; | |
838 | ||
839 | return 1; | |
840 | } | |
841 | __setup("slab_max_order=", slab_max_order_setup); | |
842 | ||
8fce4d8e CL |
843 | #ifdef CONFIG_NUMA |
844 | /* | |
845 | * Special reaping functions for NUMA systems called from cache_reap(). | |
846 | * These take care of doing round robin flushing of alien caches (containing | |
847 | * objects freed on different nodes from which they were allocated) and the | |
848 | * flushing of remote pcps by calling drain_node_pages. | |
849 | */ | |
1871e52c | 850 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
8fce4d8e CL |
851 | |
852 | static void init_reap_node(int cpu) | |
853 | { | |
854 | int node; | |
855 | ||
7d6e6d09 | 856 | node = next_node(cpu_to_mem(cpu), node_online_map); |
8fce4d8e | 857 | if (node == MAX_NUMNODES) |
442295c9 | 858 | node = first_node(node_online_map); |
8fce4d8e | 859 | |
1871e52c | 860 | per_cpu(slab_reap_node, cpu) = node; |
8fce4d8e CL |
861 | } |
862 | ||
863 | static void next_reap_node(void) | |
864 | { | |
909ea964 | 865 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e | 866 | |
8fce4d8e CL |
867 | node = next_node(node, node_online_map); |
868 | if (unlikely(node >= MAX_NUMNODES)) | |
869 | node = first_node(node_online_map); | |
909ea964 | 870 | __this_cpu_write(slab_reap_node, node); |
8fce4d8e CL |
871 | } |
872 | ||
873 | #else | |
874 | #define init_reap_node(cpu) do { } while (0) | |
875 | #define next_reap_node(void) do { } while (0) | |
876 | #endif | |
877 | ||
1da177e4 LT |
878 | /* |
879 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
880 | * via the workqueue/eventd. | |
881 | * Add the CPU number into the expiration time to minimize the possibility of | |
882 | * the CPUs getting into lockstep and contending for the global cache chain | |
883 | * lock. | |
884 | */ | |
897e679b | 885 | static void __cpuinit start_cpu_timer(int cpu) |
1da177e4 | 886 | { |
1871e52c | 887 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
1da177e4 LT |
888 | |
889 | /* | |
890 | * When this gets called from do_initcalls via cpucache_init(), | |
891 | * init_workqueues() has already run, so keventd will be setup | |
892 | * at that time. | |
893 | */ | |
52bad64d | 894 | if (keventd_up() && reap_work->work.func == NULL) { |
8fce4d8e | 895 | init_reap_node(cpu); |
78b43536 | 896 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); |
2b284214 AV |
897 | schedule_delayed_work_on(cpu, reap_work, |
898 | __round_jiffies_relative(HZ, cpu)); | |
1da177e4 LT |
899 | } |
900 | } | |
901 | ||
e498be7d | 902 | static struct array_cache *alloc_arraycache(int node, int entries, |
83b519e8 | 903 | int batchcount, gfp_t gfp) |
1da177e4 | 904 | { |
b28a02de | 905 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
1da177e4 LT |
906 | struct array_cache *nc = NULL; |
907 | ||
83b519e8 | 908 | nc = kmalloc_node(memsize, gfp, node); |
d5cff635 CM |
909 | /* |
910 | * The array_cache structures contain pointers to free object. | |
25985edc | 911 | * However, when such objects are allocated or transferred to another |
d5cff635 CM |
912 | * cache the pointers are not cleared and they could be counted as |
913 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
914 | * not scan such objects. | |
915 | */ | |
916 | kmemleak_no_scan(nc); | |
1da177e4 LT |
917 | if (nc) { |
918 | nc->avail = 0; | |
919 | nc->limit = entries; | |
920 | nc->batchcount = batchcount; | |
921 | nc->touched = 0; | |
e498be7d | 922 | spin_lock_init(&nc->lock); |
1da177e4 LT |
923 | } |
924 | return nc; | |
925 | } | |
926 | ||
072bb0aa MG |
927 | static inline bool is_slab_pfmemalloc(struct slab *slabp) |
928 | { | |
929 | struct page *page = virt_to_page(slabp->s_mem); | |
930 | ||
931 | return PageSlabPfmemalloc(page); | |
932 | } | |
933 | ||
934 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
935 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
936 | struct array_cache *ac) | |
937 | { | |
938 | struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()]; | |
939 | struct slab *slabp; | |
940 | unsigned long flags; | |
941 | ||
942 | if (!pfmemalloc_active) | |
943 | return; | |
944 | ||
945 | spin_lock_irqsave(&l3->list_lock, flags); | |
946 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
947 | if (is_slab_pfmemalloc(slabp)) | |
948 | goto out; | |
949 | ||
950 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
951 | if (is_slab_pfmemalloc(slabp)) | |
952 | goto out; | |
953 | ||
954 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
955 | if (is_slab_pfmemalloc(slabp)) | |
956 | goto out; | |
957 | ||
958 | pfmemalloc_active = false; | |
959 | out: | |
960 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
961 | } | |
962 | ||
381760ea | 963 | static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, |
072bb0aa MG |
964 | gfp_t flags, bool force_refill) |
965 | { | |
966 | int i; | |
967 | void *objp = ac->entry[--ac->avail]; | |
968 | ||
969 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
970 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
971 | struct kmem_list3 *l3; | |
972 | ||
973 | if (gfp_pfmemalloc_allowed(flags)) { | |
974 | clear_obj_pfmemalloc(&objp); | |
975 | return objp; | |
976 | } | |
977 | ||
978 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
979 | for (i = 1; i < ac->avail; i++) { | |
980 | /* If a !PFMEMALLOC object is found, swap them */ | |
981 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
982 | objp = ac->entry[i]; | |
983 | ac->entry[i] = ac->entry[ac->avail]; | |
984 | ac->entry[ac->avail] = objp; | |
985 | return objp; | |
986 | } | |
987 | } | |
988 | ||
989 | /* | |
990 | * If there are empty slabs on the slabs_free list and we are | |
991 | * being forced to refill the cache, mark this one !pfmemalloc. | |
992 | */ | |
993 | l3 = cachep->nodelists[numa_mem_id()]; | |
994 | if (!list_empty(&l3->slabs_free) && force_refill) { | |
995 | struct slab *slabp = virt_to_slab(objp); | |
996 | ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem)); | |
997 | clear_obj_pfmemalloc(&objp); | |
998 | recheck_pfmemalloc_active(cachep, ac); | |
999 | return objp; | |
1000 | } | |
1001 | ||
1002 | /* No !PFMEMALLOC objects available */ | |
1003 | ac->avail++; | |
1004 | objp = NULL; | |
1005 | } | |
1006 | ||
1007 | return objp; | |
1008 | } | |
1009 | ||
381760ea MG |
1010 | static inline void *ac_get_obj(struct kmem_cache *cachep, |
1011 | struct array_cache *ac, gfp_t flags, bool force_refill) | |
1012 | { | |
1013 | void *objp; | |
1014 | ||
1015 | if (unlikely(sk_memalloc_socks())) | |
1016 | objp = __ac_get_obj(cachep, ac, flags, force_refill); | |
1017 | else | |
1018 | objp = ac->entry[--ac->avail]; | |
1019 | ||
1020 | return objp; | |
1021 | } | |
1022 | ||
1023 | static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
072bb0aa MG |
1024 | void *objp) |
1025 | { | |
1026 | if (unlikely(pfmemalloc_active)) { | |
1027 | /* Some pfmemalloc slabs exist, check if this is one */ | |
1028 | struct page *page = virt_to_page(objp); | |
1029 | if (PageSlabPfmemalloc(page)) | |
1030 | set_obj_pfmemalloc(&objp); | |
1031 | } | |
1032 | ||
381760ea MG |
1033 | return objp; |
1034 | } | |
1035 | ||
1036 | static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
1037 | void *objp) | |
1038 | { | |
1039 | if (unlikely(sk_memalloc_socks())) | |
1040 | objp = __ac_put_obj(cachep, ac, objp); | |
1041 | ||
072bb0aa MG |
1042 | ac->entry[ac->avail++] = objp; |
1043 | } | |
1044 | ||
3ded175a CL |
1045 | /* |
1046 | * Transfer objects in one arraycache to another. | |
1047 | * Locking must be handled by the caller. | |
1048 | * | |
1049 | * Return the number of entries transferred. | |
1050 | */ | |
1051 | static int transfer_objects(struct array_cache *to, | |
1052 | struct array_cache *from, unsigned int max) | |
1053 | { | |
1054 | /* Figure out how many entries to transfer */ | |
732eacc0 | 1055 | int nr = min3(from->avail, max, to->limit - to->avail); |
3ded175a CL |
1056 | |
1057 | if (!nr) | |
1058 | return 0; | |
1059 | ||
1060 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1061 | sizeof(void *) *nr); | |
1062 | ||
1063 | from->avail -= nr; | |
1064 | to->avail += nr; | |
3ded175a CL |
1065 | return nr; |
1066 | } | |
1067 | ||
765c4507 CL |
1068 | #ifndef CONFIG_NUMA |
1069 | ||
1070 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1071 | #define reap_alien(cachep, l3) do { } while (0) | |
1072 | ||
83b519e8 | 1073 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
765c4507 CL |
1074 | { |
1075 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1076 | } | |
1077 | ||
1078 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1079 | { | |
1080 | } | |
1081 | ||
1082 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1083 | { | |
1084 | return 0; | |
1085 | } | |
1086 | ||
1087 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1088 | gfp_t flags) | |
1089 | { | |
1090 | return NULL; | |
1091 | } | |
1092 | ||
8b98c169 | 1093 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
765c4507 CL |
1094 | gfp_t flags, int nodeid) |
1095 | { | |
1096 | return NULL; | |
1097 | } | |
1098 | ||
1099 | #else /* CONFIG_NUMA */ | |
1100 | ||
8b98c169 | 1101 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
c61afb18 | 1102 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
dc85da15 | 1103 | |
83b519e8 | 1104 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
e498be7d CL |
1105 | { |
1106 | struct array_cache **ac_ptr; | |
8ef82866 | 1107 | int memsize = sizeof(void *) * nr_node_ids; |
e498be7d CL |
1108 | int i; |
1109 | ||
1110 | if (limit > 1) | |
1111 | limit = 12; | |
f3186a9c | 1112 | ac_ptr = kzalloc_node(memsize, gfp, node); |
e498be7d CL |
1113 | if (ac_ptr) { |
1114 | for_each_node(i) { | |
f3186a9c | 1115 | if (i == node || !node_online(i)) |
e498be7d | 1116 | continue; |
83b519e8 | 1117 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); |
e498be7d | 1118 | if (!ac_ptr[i]) { |
cc550def | 1119 | for (i--; i >= 0; i--) |
e498be7d CL |
1120 | kfree(ac_ptr[i]); |
1121 | kfree(ac_ptr); | |
1122 | return NULL; | |
1123 | } | |
1124 | } | |
1125 | } | |
1126 | return ac_ptr; | |
1127 | } | |
1128 | ||
5295a74c | 1129 | static void free_alien_cache(struct array_cache **ac_ptr) |
e498be7d CL |
1130 | { |
1131 | int i; | |
1132 | ||
1133 | if (!ac_ptr) | |
1134 | return; | |
e498be7d | 1135 | for_each_node(i) |
b28a02de | 1136 | kfree(ac_ptr[i]); |
e498be7d CL |
1137 | kfree(ac_ptr); |
1138 | } | |
1139 | ||
343e0d7a | 1140 | static void __drain_alien_cache(struct kmem_cache *cachep, |
5295a74c | 1141 | struct array_cache *ac, int node) |
e498be7d CL |
1142 | { |
1143 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1144 | ||
1145 | if (ac->avail) { | |
1146 | spin_lock(&rl3->list_lock); | |
e00946fe CL |
1147 | /* |
1148 | * Stuff objects into the remote nodes shared array first. | |
1149 | * That way we could avoid the overhead of putting the objects | |
1150 | * into the free lists and getting them back later. | |
1151 | */ | |
693f7d36 | 1152 | if (rl3->shared) |
1153 | transfer_objects(rl3->shared, ac, ac->limit); | |
e00946fe | 1154 | |
ff69416e | 1155 | free_block(cachep, ac->entry, ac->avail, node); |
e498be7d CL |
1156 | ac->avail = 0; |
1157 | spin_unlock(&rl3->list_lock); | |
1158 | } | |
1159 | } | |
1160 | ||
8fce4d8e CL |
1161 | /* |
1162 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1163 | */ | |
1164 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1165 | { | |
909ea964 | 1166 | int node = __this_cpu_read(slab_reap_node); |
8fce4d8e CL |
1167 | |
1168 | if (l3->alien) { | |
1169 | struct array_cache *ac = l3->alien[node]; | |
e00946fe CL |
1170 | |
1171 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
8fce4d8e CL |
1172 | __drain_alien_cache(cachep, ac, node); |
1173 | spin_unlock_irq(&ac->lock); | |
1174 | } | |
1175 | } | |
1176 | } | |
1177 | ||
a737b3e2 AM |
1178 | static void drain_alien_cache(struct kmem_cache *cachep, |
1179 | struct array_cache **alien) | |
e498be7d | 1180 | { |
b28a02de | 1181 | int i = 0; |
e498be7d CL |
1182 | struct array_cache *ac; |
1183 | unsigned long flags; | |
1184 | ||
1185 | for_each_online_node(i) { | |
4484ebf1 | 1186 | ac = alien[i]; |
e498be7d CL |
1187 | if (ac) { |
1188 | spin_lock_irqsave(&ac->lock, flags); | |
1189 | __drain_alien_cache(cachep, ac, i); | |
1190 | spin_unlock_irqrestore(&ac->lock, flags); | |
1191 | } | |
1192 | } | |
1193 | } | |
729bd0b7 | 1194 | |
873623df | 1195 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
729bd0b7 PE |
1196 | { |
1197 | struct slab *slabp = virt_to_slab(objp); | |
1198 | int nodeid = slabp->nodeid; | |
1199 | struct kmem_list3 *l3; | |
1200 | struct array_cache *alien = NULL; | |
1ca4cb24 PE |
1201 | int node; |
1202 | ||
7d6e6d09 | 1203 | node = numa_mem_id(); |
729bd0b7 PE |
1204 | |
1205 | /* | |
1206 | * Make sure we are not freeing a object from another node to the array | |
1207 | * cache on this cpu. | |
1208 | */ | |
62918a03 | 1209 | if (likely(slabp->nodeid == node)) |
729bd0b7 PE |
1210 | return 0; |
1211 | ||
1ca4cb24 | 1212 | l3 = cachep->nodelists[node]; |
729bd0b7 PE |
1213 | STATS_INC_NODEFREES(cachep); |
1214 | if (l3->alien && l3->alien[nodeid]) { | |
1215 | alien = l3->alien[nodeid]; | |
873623df | 1216 | spin_lock(&alien->lock); |
729bd0b7 PE |
1217 | if (unlikely(alien->avail == alien->limit)) { |
1218 | STATS_INC_ACOVERFLOW(cachep); | |
1219 | __drain_alien_cache(cachep, alien, nodeid); | |
1220 | } | |
072bb0aa | 1221 | ac_put_obj(cachep, alien, objp); |
729bd0b7 PE |
1222 | spin_unlock(&alien->lock); |
1223 | } else { | |
1224 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1225 | free_block(cachep, &objp, 1, nodeid); | |
1226 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1227 | } | |
1228 | return 1; | |
1229 | } | |
e498be7d CL |
1230 | #endif |
1231 | ||
8f9f8d9e DR |
1232 | /* |
1233 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1234 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1235 | * will be allocated off-node since memory is not yet online for the new node. | |
1236 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1237 | * already in use. | |
1238 | * | |
18004c5d | 1239 | * Must hold slab_mutex. |
8f9f8d9e DR |
1240 | */ |
1241 | static int init_cache_nodelists_node(int node) | |
1242 | { | |
1243 | struct kmem_cache *cachep; | |
1244 | struct kmem_list3 *l3; | |
1245 | const int memsize = sizeof(struct kmem_list3); | |
1246 | ||
18004c5d | 1247 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1248 | /* |
1249 | * Set up the size64 kmemlist for cpu before we can | |
1250 | * begin anything. Make sure some other cpu on this | |
1251 | * node has not already allocated this | |
1252 | */ | |
1253 | if (!cachep->nodelists[node]) { | |
1254 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1255 | if (!l3) | |
1256 | return -ENOMEM; | |
1257 | kmem_list3_init(l3); | |
1258 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1259 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1260 | ||
1261 | /* | |
1262 | * The l3s don't come and go as CPUs come and | |
18004c5d | 1263 | * go. slab_mutex is sufficient |
8f9f8d9e DR |
1264 | * protection here. |
1265 | */ | |
1266 | cachep->nodelists[node] = l3; | |
1267 | } | |
1268 | ||
1269 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1270 | cachep->nodelists[node]->free_limit = | |
1271 | (1 + nr_cpus_node(node)) * | |
1272 | cachep->batchcount + cachep->num; | |
1273 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1274 | } | |
1275 | return 0; | |
1276 | } | |
1277 | ||
fbf1e473 AM |
1278 | static void __cpuinit cpuup_canceled(long cpu) |
1279 | { | |
1280 | struct kmem_cache *cachep; | |
1281 | struct kmem_list3 *l3 = NULL; | |
7d6e6d09 | 1282 | int node = cpu_to_mem(cpu); |
a70f7302 | 1283 | const struct cpumask *mask = cpumask_of_node(node); |
fbf1e473 | 1284 | |
18004c5d | 1285 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1286 | struct array_cache *nc; |
1287 | struct array_cache *shared; | |
1288 | struct array_cache **alien; | |
fbf1e473 | 1289 | |
fbf1e473 AM |
1290 | /* cpu is dead; no one can alloc from it. */ |
1291 | nc = cachep->array[cpu]; | |
1292 | cachep->array[cpu] = NULL; | |
1293 | l3 = cachep->nodelists[node]; | |
1294 | ||
1295 | if (!l3) | |
1296 | goto free_array_cache; | |
1297 | ||
1298 | spin_lock_irq(&l3->list_lock); | |
1299 | ||
1300 | /* Free limit for this kmem_list3 */ | |
1301 | l3->free_limit -= cachep->batchcount; | |
1302 | if (nc) | |
1303 | free_block(cachep, nc->entry, nc->avail, node); | |
1304 | ||
58463c1f | 1305 | if (!cpumask_empty(mask)) { |
fbf1e473 AM |
1306 | spin_unlock_irq(&l3->list_lock); |
1307 | goto free_array_cache; | |
1308 | } | |
1309 | ||
1310 | shared = l3->shared; | |
1311 | if (shared) { | |
1312 | free_block(cachep, shared->entry, | |
1313 | shared->avail, node); | |
1314 | l3->shared = NULL; | |
1315 | } | |
1316 | ||
1317 | alien = l3->alien; | |
1318 | l3->alien = NULL; | |
1319 | ||
1320 | spin_unlock_irq(&l3->list_lock); | |
1321 | ||
1322 | kfree(shared); | |
1323 | if (alien) { | |
1324 | drain_alien_cache(cachep, alien); | |
1325 | free_alien_cache(alien); | |
1326 | } | |
1327 | free_array_cache: | |
1328 | kfree(nc); | |
1329 | } | |
1330 | /* | |
1331 | * In the previous loop, all the objects were freed to | |
1332 | * the respective cache's slabs, now we can go ahead and | |
1333 | * shrink each nodelist to its limit. | |
1334 | */ | |
18004c5d | 1335 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1336 | l3 = cachep->nodelists[node]; |
1337 | if (!l3) | |
1338 | continue; | |
1339 | drain_freelist(cachep, l3, l3->free_objects); | |
1340 | } | |
1341 | } | |
1342 | ||
1343 | static int __cpuinit cpuup_prepare(long cpu) | |
1da177e4 | 1344 | { |
343e0d7a | 1345 | struct kmem_cache *cachep; |
e498be7d | 1346 | struct kmem_list3 *l3 = NULL; |
7d6e6d09 | 1347 | int node = cpu_to_mem(cpu); |
8f9f8d9e | 1348 | int err; |
1da177e4 | 1349 | |
fbf1e473 AM |
1350 | /* |
1351 | * We need to do this right in the beginning since | |
1352 | * alloc_arraycache's are going to use this list. | |
1353 | * kmalloc_node allows us to add the slab to the right | |
1354 | * kmem_list3 and not this cpu's kmem_list3 | |
1355 | */ | |
8f9f8d9e DR |
1356 | err = init_cache_nodelists_node(node); |
1357 | if (err < 0) | |
1358 | goto bad; | |
fbf1e473 AM |
1359 | |
1360 | /* | |
1361 | * Now we can go ahead with allocating the shared arrays and | |
1362 | * array caches | |
1363 | */ | |
18004c5d | 1364 | list_for_each_entry(cachep, &slab_caches, list) { |
fbf1e473 AM |
1365 | struct array_cache *nc; |
1366 | struct array_cache *shared = NULL; | |
1367 | struct array_cache **alien = NULL; | |
1368 | ||
1369 | nc = alloc_arraycache(node, cachep->limit, | |
83b519e8 | 1370 | cachep->batchcount, GFP_KERNEL); |
fbf1e473 AM |
1371 | if (!nc) |
1372 | goto bad; | |
1373 | if (cachep->shared) { | |
1374 | shared = alloc_arraycache(node, | |
1375 | cachep->shared * cachep->batchcount, | |
83b519e8 | 1376 | 0xbaadf00d, GFP_KERNEL); |
12d00f6a AM |
1377 | if (!shared) { |
1378 | kfree(nc); | |
1da177e4 | 1379 | goto bad; |
12d00f6a | 1380 | } |
fbf1e473 AM |
1381 | } |
1382 | if (use_alien_caches) { | |
83b519e8 | 1383 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); |
12d00f6a AM |
1384 | if (!alien) { |
1385 | kfree(shared); | |
1386 | kfree(nc); | |
fbf1e473 | 1387 | goto bad; |
12d00f6a | 1388 | } |
fbf1e473 AM |
1389 | } |
1390 | cachep->array[cpu] = nc; | |
1391 | l3 = cachep->nodelists[node]; | |
1392 | BUG_ON(!l3); | |
1393 | ||
1394 | spin_lock_irq(&l3->list_lock); | |
1395 | if (!l3->shared) { | |
1396 | /* | |
1397 | * We are serialised from CPU_DEAD or | |
1398 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1399 | */ | |
1400 | l3->shared = shared; | |
1401 | shared = NULL; | |
1402 | } | |
4484ebf1 | 1403 | #ifdef CONFIG_NUMA |
fbf1e473 AM |
1404 | if (!l3->alien) { |
1405 | l3->alien = alien; | |
1406 | alien = NULL; | |
1da177e4 | 1407 | } |
fbf1e473 AM |
1408 | #endif |
1409 | spin_unlock_irq(&l3->list_lock); | |
1410 | kfree(shared); | |
1411 | free_alien_cache(alien); | |
83835b3d PZ |
1412 | if (cachep->flags & SLAB_DEBUG_OBJECTS) |
1413 | slab_set_debugobj_lock_classes_node(cachep, node); | |
fbf1e473 | 1414 | } |
ce79ddc8 PE |
1415 | init_node_lock_keys(node); |
1416 | ||
fbf1e473 AM |
1417 | return 0; |
1418 | bad: | |
12d00f6a | 1419 | cpuup_canceled(cpu); |
fbf1e473 AM |
1420 | return -ENOMEM; |
1421 | } | |
1422 | ||
1423 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1424 | unsigned long action, void *hcpu) | |
1425 | { | |
1426 | long cpu = (long)hcpu; | |
1427 | int err = 0; | |
1428 | ||
1429 | switch (action) { | |
fbf1e473 AM |
1430 | case CPU_UP_PREPARE: |
1431 | case CPU_UP_PREPARE_FROZEN: | |
18004c5d | 1432 | mutex_lock(&slab_mutex); |
fbf1e473 | 1433 | err = cpuup_prepare(cpu); |
18004c5d | 1434 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
1435 | break; |
1436 | case CPU_ONLINE: | |
8bb78442 | 1437 | case CPU_ONLINE_FROZEN: |
1da177e4 LT |
1438 | start_cpu_timer(cpu); |
1439 | break; | |
1440 | #ifdef CONFIG_HOTPLUG_CPU | |
5830c590 | 1441 | case CPU_DOWN_PREPARE: |
8bb78442 | 1442 | case CPU_DOWN_PREPARE_FROZEN: |
5830c590 | 1443 | /* |
18004c5d | 1444 | * Shutdown cache reaper. Note that the slab_mutex is |
5830c590 CL |
1445 | * held so that if cache_reap() is invoked it cannot do |
1446 | * anything expensive but will only modify reap_work | |
1447 | * and reschedule the timer. | |
1448 | */ | |
afe2c511 | 1449 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
5830c590 | 1450 | /* Now the cache_reaper is guaranteed to be not running. */ |
1871e52c | 1451 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
5830c590 CL |
1452 | break; |
1453 | case CPU_DOWN_FAILED: | |
8bb78442 | 1454 | case CPU_DOWN_FAILED_FROZEN: |
5830c590 CL |
1455 | start_cpu_timer(cpu); |
1456 | break; | |
1da177e4 | 1457 | case CPU_DEAD: |
8bb78442 | 1458 | case CPU_DEAD_FROZEN: |
4484ebf1 RT |
1459 | /* |
1460 | * Even if all the cpus of a node are down, we don't free the | |
1461 | * kmem_list3 of any cache. This to avoid a race between | |
1462 | * cpu_down, and a kmalloc allocation from another cpu for | |
1463 | * memory from the node of the cpu going down. The list3 | |
1464 | * structure is usually allocated from kmem_cache_create() and | |
1465 | * gets destroyed at kmem_cache_destroy(). | |
1466 | */ | |
183ff22b | 1467 | /* fall through */ |
8f5be20b | 1468 | #endif |
1da177e4 | 1469 | case CPU_UP_CANCELED: |
8bb78442 | 1470 | case CPU_UP_CANCELED_FROZEN: |
18004c5d | 1471 | mutex_lock(&slab_mutex); |
fbf1e473 | 1472 | cpuup_canceled(cpu); |
18004c5d | 1473 | mutex_unlock(&slab_mutex); |
1da177e4 | 1474 | break; |
1da177e4 | 1475 | } |
eac40680 | 1476 | return notifier_from_errno(err); |
1da177e4 LT |
1477 | } |
1478 | ||
74b85f37 CS |
1479 | static struct notifier_block __cpuinitdata cpucache_notifier = { |
1480 | &cpuup_callback, NULL, 0 | |
1481 | }; | |
1da177e4 | 1482 | |
8f9f8d9e DR |
1483 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1484 | /* | |
1485 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1486 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1487 | * removed. | |
1488 | * | |
18004c5d | 1489 | * Must hold slab_mutex. |
8f9f8d9e DR |
1490 | */ |
1491 | static int __meminit drain_cache_nodelists_node(int node) | |
1492 | { | |
1493 | struct kmem_cache *cachep; | |
1494 | int ret = 0; | |
1495 | ||
18004c5d | 1496 | list_for_each_entry(cachep, &slab_caches, list) { |
8f9f8d9e DR |
1497 | struct kmem_list3 *l3; |
1498 | ||
1499 | l3 = cachep->nodelists[node]; | |
1500 | if (!l3) | |
1501 | continue; | |
1502 | ||
1503 | drain_freelist(cachep, l3, l3->free_objects); | |
1504 | ||
1505 | if (!list_empty(&l3->slabs_full) || | |
1506 | !list_empty(&l3->slabs_partial)) { | |
1507 | ret = -EBUSY; | |
1508 | break; | |
1509 | } | |
1510 | } | |
1511 | return ret; | |
1512 | } | |
1513 | ||
1514 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1515 | unsigned long action, void *arg) | |
1516 | { | |
1517 | struct memory_notify *mnb = arg; | |
1518 | int ret = 0; | |
1519 | int nid; | |
1520 | ||
1521 | nid = mnb->status_change_nid; | |
1522 | if (nid < 0) | |
1523 | goto out; | |
1524 | ||
1525 | switch (action) { | |
1526 | case MEM_GOING_ONLINE: | |
18004c5d | 1527 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1528 | ret = init_cache_nodelists_node(nid); |
18004c5d | 1529 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1530 | break; |
1531 | case MEM_GOING_OFFLINE: | |
18004c5d | 1532 | mutex_lock(&slab_mutex); |
8f9f8d9e | 1533 | ret = drain_cache_nodelists_node(nid); |
18004c5d | 1534 | mutex_unlock(&slab_mutex); |
8f9f8d9e DR |
1535 | break; |
1536 | case MEM_ONLINE: | |
1537 | case MEM_OFFLINE: | |
1538 | case MEM_CANCEL_ONLINE: | |
1539 | case MEM_CANCEL_OFFLINE: | |
1540 | break; | |
1541 | } | |
1542 | out: | |
5fda1bd5 | 1543 | return notifier_from_errno(ret); |
8f9f8d9e DR |
1544 | } |
1545 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1546 | ||
e498be7d CL |
1547 | /* |
1548 | * swap the static kmem_list3 with kmalloced memory | |
1549 | */ | |
8f9f8d9e DR |
1550 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, |
1551 | int nodeid) | |
e498be7d CL |
1552 | { |
1553 | struct kmem_list3 *ptr; | |
1554 | ||
83b519e8 | 1555 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); |
e498be7d CL |
1556 | BUG_ON(!ptr); |
1557 | ||
e498be7d | 1558 | memcpy(ptr, list, sizeof(struct kmem_list3)); |
2b2d5493 IM |
1559 | /* |
1560 | * Do not assume that spinlocks can be initialized via memcpy: | |
1561 | */ | |
1562 | spin_lock_init(&ptr->list_lock); | |
1563 | ||
e498be7d CL |
1564 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1565 | cachep->nodelists[nodeid] = ptr; | |
e498be7d CL |
1566 | } |
1567 | ||
556a169d PE |
1568 | /* |
1569 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1570 | * size of kmem_list3. | |
1571 | */ | |
1572 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1573 | { | |
1574 | int node; | |
1575 | ||
1576 | for_each_online_node(node) { | |
1577 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1578 | cachep->nodelists[node]->next_reap = jiffies + | |
1579 | REAPTIMEOUT_LIST3 + | |
1580 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1581 | } | |
1582 | } | |
1583 | ||
a737b3e2 AM |
1584 | /* |
1585 | * Initialisation. Called after the page allocator have been initialised and | |
1586 | * before smp_init(). | |
1da177e4 LT |
1587 | */ |
1588 | void __init kmem_cache_init(void) | |
1589 | { | |
1590 | size_t left_over; | |
1591 | struct cache_sizes *sizes; | |
1592 | struct cache_names *names; | |
e498be7d | 1593 | int i; |
07ed76b2 | 1594 | int order; |
1ca4cb24 | 1595 | int node; |
e498be7d | 1596 | |
9b030cb8 CL |
1597 | kmem_cache = &kmem_cache_boot; |
1598 | ||
b6e68bc1 | 1599 | if (num_possible_nodes() == 1) |
62918a03 SS |
1600 | use_alien_caches = 0; |
1601 | ||
e498be7d CL |
1602 | for (i = 0; i < NUM_INIT_LISTS; i++) { |
1603 | kmem_list3_init(&initkmem_list3[i]); | |
1604 | if (i < MAX_NUMNODES) | |
9b030cb8 | 1605 | kmem_cache->nodelists[i] = NULL; |
e498be7d | 1606 | } |
9b030cb8 | 1607 | set_up_list3s(kmem_cache, CACHE_CACHE); |
1da177e4 LT |
1608 | |
1609 | /* | |
1610 | * Fragmentation resistance on low memory - only use bigger | |
3df1cccd DR |
1611 | * page orders on machines with more than 32MB of memory if |
1612 | * not overridden on the command line. | |
1da177e4 | 1613 | */ |
3df1cccd | 1614 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
543585cc | 1615 | slab_max_order = SLAB_MAX_ORDER_HI; |
1da177e4 | 1616 | |
1da177e4 LT |
1617 | /* Bootstrap is tricky, because several objects are allocated |
1618 | * from caches that do not exist yet: | |
9b030cb8 CL |
1619 | * 1) initialize the kmem_cache cache: it contains the struct |
1620 | * kmem_cache structures of all caches, except kmem_cache itself: | |
1621 | * kmem_cache is statically allocated. | |
e498be7d CL |
1622 | * Initially an __init data area is used for the head array and the |
1623 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1624 | * array at the end of the bootstrap. | |
1da177e4 | 1625 | * 2) Create the first kmalloc cache. |
343e0d7a | 1626 | * The struct kmem_cache for the new cache is allocated normally. |
e498be7d CL |
1627 | * An __init data area is used for the head array. |
1628 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1629 | * head arrays. | |
9b030cb8 | 1630 | * 4) Replace the __init data head arrays for kmem_cache and the first |
1da177e4 | 1631 | * kmalloc cache with kmalloc allocated arrays. |
9b030cb8 | 1632 | * 5) Replace the __init data for kmem_list3 for kmem_cache and |
e498be7d CL |
1633 | * the other cache's with kmalloc allocated memory. |
1634 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1da177e4 LT |
1635 | */ |
1636 | ||
7d6e6d09 | 1637 | node = numa_mem_id(); |
1ca4cb24 | 1638 | |
9b030cb8 | 1639 | /* 1) create the kmem_cache */ |
18004c5d | 1640 | INIT_LIST_HEAD(&slab_caches); |
9b030cb8 CL |
1641 | list_add(&kmem_cache->list, &slab_caches); |
1642 | kmem_cache->colour_off = cache_line_size(); | |
1643 | kmem_cache->array[smp_processor_id()] = &initarray_cache.cache; | |
1644 | kmem_cache->nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; | |
1da177e4 | 1645 | |
8da3430d | 1646 | /* |
b56efcf0 | 1647 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
8da3430d | 1648 | */ |
9b030cb8 | 1649 | kmem_cache->size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + |
b56efcf0 | 1650 | nr_node_ids * sizeof(struct kmem_list3 *); |
9b030cb8 CL |
1651 | kmem_cache->object_size = kmem_cache->size; |
1652 | kmem_cache->size = ALIGN(kmem_cache->object_size, | |
a737b3e2 | 1653 | cache_line_size()); |
9b030cb8 CL |
1654 | kmem_cache->reciprocal_buffer_size = |
1655 | reciprocal_value(kmem_cache->size); | |
1da177e4 | 1656 | |
07ed76b2 | 1657 | for (order = 0; order < MAX_ORDER; order++) { |
9b030cb8 CL |
1658 | cache_estimate(order, kmem_cache->size, |
1659 | cache_line_size(), 0, &left_over, &kmem_cache->num); | |
1660 | if (kmem_cache->num) | |
07ed76b2 JS |
1661 | break; |
1662 | } | |
9b030cb8 CL |
1663 | BUG_ON(!kmem_cache->num); |
1664 | kmem_cache->gfporder = order; | |
1665 | kmem_cache->colour = left_over / kmem_cache->colour_off; | |
1666 | kmem_cache->slab_size = ALIGN(kmem_cache->num * sizeof(kmem_bufctl_t) + | |
b28a02de | 1667 | sizeof(struct slab), cache_line_size()); |
1da177e4 LT |
1668 | |
1669 | /* 2+3) create the kmalloc caches */ | |
1670 | sizes = malloc_sizes; | |
1671 | names = cache_names; | |
1672 | ||
a737b3e2 AM |
1673 | /* |
1674 | * Initialize the caches that provide memory for the array cache and the | |
1675 | * kmem_list3 structures first. Without this, further allocations will | |
1676 | * bug. | |
e498be7d CL |
1677 | */ |
1678 | ||
039363f3 | 1679 | sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name, |
a737b3e2 AM |
1680 | sizes[INDEX_AC].cs_size, |
1681 | ARCH_KMALLOC_MINALIGN, | |
1682 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1683 | NULL); |
e498be7d | 1684 | |
7c9adf5a | 1685 | list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches); |
a737b3e2 | 1686 | if (INDEX_AC != INDEX_L3) { |
e498be7d | 1687 | sizes[INDEX_L3].cs_cachep = |
039363f3 | 1688 | __kmem_cache_create(names[INDEX_L3].name, |
a737b3e2 AM |
1689 | sizes[INDEX_L3].cs_size, |
1690 | ARCH_KMALLOC_MINALIGN, | |
1691 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1692 | NULL); |
7c9adf5a | 1693 | list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches); |
a737b3e2 | 1694 | } |
e498be7d | 1695 | |
e0a42726 IM |
1696 | slab_early_init = 0; |
1697 | ||
1da177e4 | 1698 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1699 | /* |
1700 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1701 | * This should be particularly beneficial on SMP boxes, as it |
1702 | * eliminates "false sharing". | |
1703 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1704 | * allow tighter packing of the smaller caches. |
1705 | */ | |
a737b3e2 | 1706 | if (!sizes->cs_cachep) { |
039363f3 | 1707 | sizes->cs_cachep = __kmem_cache_create(names->name, |
a737b3e2 AM |
1708 | sizes->cs_size, |
1709 | ARCH_KMALLOC_MINALIGN, | |
1710 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
20c2df83 | 1711 | NULL); |
7c9adf5a | 1712 | list_add(&sizes->cs_cachep->list, &slab_caches); |
a737b3e2 | 1713 | } |
4b51d669 | 1714 | #ifdef CONFIG_ZONE_DMA |
039363f3 | 1715 | sizes->cs_dmacachep = __kmem_cache_create( |
4b51d669 | 1716 | names->name_dma, |
a737b3e2 AM |
1717 | sizes->cs_size, |
1718 | ARCH_KMALLOC_MINALIGN, | |
1719 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1720 | SLAB_PANIC, | |
20c2df83 | 1721 | NULL); |
7c9adf5a | 1722 | list_add(&sizes->cs_dmacachep->list, &slab_caches); |
4b51d669 | 1723 | #endif |
1da177e4 LT |
1724 | sizes++; |
1725 | names++; | |
1726 | } | |
1727 | /* 4) Replace the bootstrap head arrays */ | |
1728 | { | |
2b2d5493 | 1729 | struct array_cache *ptr; |
e498be7d | 1730 | |
83b519e8 | 1731 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1732 | |
9b030cb8 CL |
1733 | BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache); |
1734 | memcpy(ptr, cpu_cache_get(kmem_cache), | |
b28a02de | 1735 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1736 | /* |
1737 | * Do not assume that spinlocks can be initialized via memcpy: | |
1738 | */ | |
1739 | spin_lock_init(&ptr->lock); | |
1740 | ||
9b030cb8 | 1741 | kmem_cache->array[smp_processor_id()] = ptr; |
e498be7d | 1742 | |
83b519e8 | 1743 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1744 | |
9a2dba4b | 1745 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1746 | != &initarray_generic.cache); |
9a2dba4b | 1747 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1748 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1749 | /* |
1750 | * Do not assume that spinlocks can be initialized via memcpy: | |
1751 | */ | |
1752 | spin_lock_init(&ptr->lock); | |
1753 | ||
e498be7d | 1754 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1755 | ptr; |
1da177e4 | 1756 | } |
e498be7d CL |
1757 | /* 5) Replace the bootstrap kmem_list3's */ |
1758 | { | |
1ca4cb24 PE |
1759 | int nid; |
1760 | ||
9c09a95c | 1761 | for_each_online_node(nid) { |
9b030cb8 | 1762 | init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1763 | |
e498be7d | 1764 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1765 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1766 | |
1767 | if (INDEX_AC != INDEX_L3) { | |
1768 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1769 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1770 | } |
1771 | } | |
1772 | } | |
1da177e4 | 1773 | |
97d06609 | 1774 | slab_state = UP; |
8429db5c PE |
1775 | } |
1776 | ||
1777 | void __init kmem_cache_init_late(void) | |
1778 | { | |
1779 | struct kmem_cache *cachep; | |
1780 | ||
97d06609 | 1781 | slab_state = UP; |
52cef189 | 1782 | |
30765b92 PZ |
1783 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1784 | init_lock_keys(); | |
1785 | ||
8429db5c | 1786 | /* 6) resize the head arrays to their final sizes */ |
18004c5d CL |
1787 | mutex_lock(&slab_mutex); |
1788 | list_for_each_entry(cachep, &slab_caches, list) | |
8429db5c PE |
1789 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1790 | BUG(); | |
18004c5d | 1791 | mutex_unlock(&slab_mutex); |
056c6241 | 1792 | |
97d06609 CL |
1793 | /* Done! */ |
1794 | slab_state = FULL; | |
1795 | ||
a737b3e2 AM |
1796 | /* |
1797 | * Register a cpu startup notifier callback that initializes | |
1798 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1799 | */ |
1800 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1801 | |
8f9f8d9e DR |
1802 | #ifdef CONFIG_NUMA |
1803 | /* | |
1804 | * Register a memory hotplug callback that initializes and frees | |
1805 | * nodelists. | |
1806 | */ | |
1807 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1808 | #endif | |
1809 | ||
a737b3e2 AM |
1810 | /* |
1811 | * The reap timers are started later, with a module init call: That part | |
1812 | * of the kernel is not yet operational. | |
1da177e4 LT |
1813 | */ |
1814 | } | |
1815 | ||
1816 | static int __init cpucache_init(void) | |
1817 | { | |
1818 | int cpu; | |
1819 | ||
a737b3e2 AM |
1820 | /* |
1821 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1822 | */ |
e498be7d | 1823 | for_each_online_cpu(cpu) |
a737b3e2 | 1824 | start_cpu_timer(cpu); |
a164f896 GC |
1825 | |
1826 | /* Done! */ | |
97d06609 | 1827 | slab_state = FULL; |
1da177e4 LT |
1828 | return 0; |
1829 | } | |
1da177e4 LT |
1830 | __initcall(cpucache_init); |
1831 | ||
8bdec192 RA |
1832 | static noinline void |
1833 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1834 | { | |
1835 | struct kmem_list3 *l3; | |
1836 | struct slab *slabp; | |
1837 | unsigned long flags; | |
1838 | int node; | |
1839 | ||
1840 | printk(KERN_WARNING | |
1841 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1842 | nodeid, gfpflags); | |
1843 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1844 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1845 | |
1846 | for_each_online_node(node) { | |
1847 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1848 | unsigned long active_slabs = 0, num_slabs = 0; | |
1849 | ||
1850 | l3 = cachep->nodelists[node]; | |
1851 | if (!l3) | |
1852 | continue; | |
1853 | ||
1854 | spin_lock_irqsave(&l3->list_lock, flags); | |
1855 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1856 | active_objs += cachep->num; | |
1857 | active_slabs++; | |
1858 | } | |
1859 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1860 | active_objs += slabp->inuse; | |
1861 | active_slabs++; | |
1862 | } | |
1863 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1864 | num_slabs++; | |
1865 | ||
1866 | free_objects += l3->free_objects; | |
1867 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1868 | ||
1869 | num_slabs += active_slabs; | |
1870 | num_objs = num_slabs * cachep->num; | |
1871 | printk(KERN_WARNING | |
1872 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1873 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1874 | free_objects); | |
1875 | } | |
1876 | } | |
1877 | ||
1da177e4 LT |
1878 | /* |
1879 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1880 | * | |
1881 | * If we requested dmaable memory, we will get it. Even if we | |
1882 | * did not request dmaable memory, we might get it, but that | |
1883 | * would be relatively rare and ignorable. | |
1884 | */ | |
343e0d7a | 1885 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1886 | { |
1887 | struct page *page; | |
e1b6aa6f | 1888 | int nr_pages; |
1da177e4 LT |
1889 | int i; |
1890 | ||
d6fef9da | 1891 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1892 | /* |
1893 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1894 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1895 | */ |
e1b6aa6f | 1896 | flags |= __GFP_COMP; |
d6fef9da | 1897 | #endif |
765c4507 | 1898 | |
a618e89f | 1899 | flags |= cachep->allocflags; |
e12ba74d MG |
1900 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1901 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1902 | |
517d0869 | 1903 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1904 | if (!page) { |
1905 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1906 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1907 | return NULL; |
8bdec192 | 1908 | } |
1da177e4 | 1909 | |
b37f1dd0 | 1910 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
072bb0aa MG |
1911 | if (unlikely(page->pfmemalloc)) |
1912 | pfmemalloc_active = true; | |
1913 | ||
e1b6aa6f | 1914 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1915 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1916 | add_zone_page_state(page_zone(page), |
1917 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1918 | else | |
1919 | add_zone_page_state(page_zone(page), | |
1920 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
072bb0aa | 1921 | for (i = 0; i < nr_pages; i++) { |
e1b6aa6f | 1922 | __SetPageSlab(page + i); |
c175eea4 | 1923 | |
072bb0aa MG |
1924 | if (page->pfmemalloc) |
1925 | SetPageSlabPfmemalloc(page + i); | |
1926 | } | |
1927 | ||
b1eeab67 VN |
1928 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1929 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1930 | ||
1931 | if (cachep->ctor) | |
1932 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1933 | else | |
1934 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1935 | } | |
c175eea4 | 1936 | |
e1b6aa6f | 1937 | return page_address(page); |
1da177e4 LT |
1938 | } |
1939 | ||
1940 | /* | |
1941 | * Interface to system's page release. | |
1942 | */ | |
343e0d7a | 1943 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1944 | { |
b28a02de | 1945 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1946 | struct page *page = virt_to_page(addr); |
1947 | const unsigned long nr_freed = i; | |
1948 | ||
b1eeab67 | 1949 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1950 | |
972d1a7b CL |
1951 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1952 | sub_zone_page_state(page_zone(page), | |
1953 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1954 | else | |
1955 | sub_zone_page_state(page_zone(page), | |
1956 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1957 | while (i--) { |
f205b2fe | 1958 | BUG_ON(!PageSlab(page)); |
072bb0aa | 1959 | __ClearPageSlabPfmemalloc(page); |
f205b2fe | 1960 | __ClearPageSlab(page); |
1da177e4 LT |
1961 | page++; |
1962 | } | |
1da177e4 LT |
1963 | if (current->reclaim_state) |
1964 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1965 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1966 | } |
1967 | ||
1968 | static void kmem_rcu_free(struct rcu_head *head) | |
1969 | { | |
b28a02de | 1970 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1971 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1972 | |
1973 | kmem_freepages(cachep, slab_rcu->addr); | |
1974 | if (OFF_SLAB(cachep)) | |
1975 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1976 | } | |
1977 | ||
1978 | #if DEBUG | |
1979 | ||
1980 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1981 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1982 | unsigned long caller) |
1da177e4 | 1983 | { |
8c138bc0 | 1984 | int size = cachep->object_size; |
1da177e4 | 1985 | |
3dafccf2 | 1986 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1987 | |
b28a02de | 1988 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1989 | return; |
1990 | ||
b28a02de PE |
1991 | *addr++ = 0x12345678; |
1992 | *addr++ = caller; | |
1993 | *addr++ = smp_processor_id(); | |
1994 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1995 | { |
1996 | unsigned long *sptr = &caller; | |
1997 | unsigned long svalue; | |
1998 | ||
1999 | while (!kstack_end(sptr)) { | |
2000 | svalue = *sptr++; | |
2001 | if (kernel_text_address(svalue)) { | |
b28a02de | 2002 | *addr++ = svalue; |
1da177e4 LT |
2003 | size -= sizeof(unsigned long); |
2004 | if (size <= sizeof(unsigned long)) | |
2005 | break; | |
2006 | } | |
2007 | } | |
2008 | ||
2009 | } | |
b28a02de | 2010 | *addr++ = 0x87654321; |
1da177e4 LT |
2011 | } |
2012 | #endif | |
2013 | ||
343e0d7a | 2014 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 2015 | { |
8c138bc0 | 2016 | int size = cachep->object_size; |
3dafccf2 | 2017 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
2018 | |
2019 | memset(addr, val, size); | |
b28a02de | 2020 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
2021 | } |
2022 | ||
2023 | static void dump_line(char *data, int offset, int limit) | |
2024 | { | |
2025 | int i; | |
aa83aa40 DJ |
2026 | unsigned char error = 0; |
2027 | int bad_count = 0; | |
2028 | ||
fdde6abb | 2029 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
2030 | for (i = 0; i < limit; i++) { |
2031 | if (data[offset + i] != POISON_FREE) { | |
2032 | error = data[offset + i]; | |
2033 | bad_count++; | |
2034 | } | |
aa83aa40 | 2035 | } |
fdde6abb SAS |
2036 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
2037 | &data[offset], limit, 1); | |
aa83aa40 DJ |
2038 | |
2039 | if (bad_count == 1) { | |
2040 | error ^= POISON_FREE; | |
2041 | if (!(error & (error - 1))) { | |
2042 | printk(KERN_ERR "Single bit error detected. Probably " | |
2043 | "bad RAM.\n"); | |
2044 | #ifdef CONFIG_X86 | |
2045 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
2046 | "test tool.\n"); | |
2047 | #else | |
2048 | printk(KERN_ERR "Run a memory test tool.\n"); | |
2049 | #endif | |
2050 | } | |
2051 | } | |
1da177e4 LT |
2052 | } |
2053 | #endif | |
2054 | ||
2055 | #if DEBUG | |
2056 | ||
343e0d7a | 2057 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
2058 | { |
2059 | int i, size; | |
2060 | char *realobj; | |
2061 | ||
2062 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 2063 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
2064 | *dbg_redzone1(cachep, objp), |
2065 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2066 | } |
2067 | ||
2068 | if (cachep->flags & SLAB_STORE_USER) { | |
2069 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 2070 | *dbg_userword(cachep, objp)); |
1da177e4 | 2071 | print_symbol("(%s)", |
a737b3e2 | 2072 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
2073 | printk("\n"); |
2074 | } | |
3dafccf2 | 2075 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2076 | size = cachep->object_size; |
b28a02de | 2077 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
2078 | int limit; |
2079 | limit = 16; | |
b28a02de PE |
2080 | if (i + limit > size) |
2081 | limit = size - i; | |
1da177e4 LT |
2082 | dump_line(realobj, i, limit); |
2083 | } | |
2084 | } | |
2085 | ||
343e0d7a | 2086 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
2087 | { |
2088 | char *realobj; | |
2089 | int size, i; | |
2090 | int lines = 0; | |
2091 | ||
3dafccf2 | 2092 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2093 | size = cachep->object_size; |
1da177e4 | 2094 | |
b28a02de | 2095 | for (i = 0; i < size; i++) { |
1da177e4 | 2096 | char exp = POISON_FREE; |
b28a02de | 2097 | if (i == size - 1) |
1da177e4 LT |
2098 | exp = POISON_END; |
2099 | if (realobj[i] != exp) { | |
2100 | int limit; | |
2101 | /* Mismatch ! */ | |
2102 | /* Print header */ | |
2103 | if (lines == 0) { | |
b28a02de | 2104 | printk(KERN_ERR |
face37f5 DJ |
2105 | "Slab corruption (%s): %s start=%p, len=%d\n", |
2106 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
2107 | print_objinfo(cachep, objp, 0); |
2108 | } | |
2109 | /* Hexdump the affected line */ | |
b28a02de | 2110 | i = (i / 16) * 16; |
1da177e4 | 2111 | limit = 16; |
b28a02de PE |
2112 | if (i + limit > size) |
2113 | limit = size - i; | |
1da177e4 LT |
2114 | dump_line(realobj, i, limit); |
2115 | i += 16; | |
2116 | lines++; | |
2117 | /* Limit to 5 lines */ | |
2118 | if (lines > 5) | |
2119 | break; | |
2120 | } | |
2121 | } | |
2122 | if (lines != 0) { | |
2123 | /* Print some data about the neighboring objects, if they | |
2124 | * exist: | |
2125 | */ | |
6ed5eb22 | 2126 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 2127 | unsigned int objnr; |
1da177e4 | 2128 | |
8fea4e96 | 2129 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2130 | if (objnr) { |
8fea4e96 | 2131 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2132 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2133 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2134 | realobj, size); |
1da177e4 LT |
2135 | print_objinfo(cachep, objp, 2); |
2136 | } | |
b28a02de | 2137 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2138 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2139 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2140 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2141 | realobj, size); |
1da177e4 LT |
2142 | print_objinfo(cachep, objp, 2); |
2143 | } | |
2144 | } | |
2145 | } | |
2146 | #endif | |
2147 | ||
12dd36fa | 2148 | #if DEBUG |
e79aec29 | 2149 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2150 | { |
1da177e4 LT |
2151 | int i; |
2152 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2153 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2154 | |
2155 | if (cachep->flags & SLAB_POISON) { | |
2156 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2157 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2158 | OFF_SLAB(cachep)) |
b28a02de | 2159 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2160 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2161 | else |
2162 | check_poison_obj(cachep, objp); | |
2163 | #else | |
2164 | check_poison_obj(cachep, objp); | |
2165 | #endif | |
2166 | } | |
2167 | if (cachep->flags & SLAB_RED_ZONE) { | |
2168 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2169 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2170 | "was overwritten"); |
1da177e4 LT |
2171 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2172 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2173 | "was overwritten"); |
1da177e4 | 2174 | } |
1da177e4 | 2175 | } |
12dd36fa | 2176 | } |
1da177e4 | 2177 | #else |
e79aec29 | 2178 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2179 | { |
12dd36fa | 2180 | } |
1da177e4 LT |
2181 | #endif |
2182 | ||
911851e6 RD |
2183 | /** |
2184 | * slab_destroy - destroy and release all objects in a slab | |
2185 | * @cachep: cache pointer being destroyed | |
2186 | * @slabp: slab pointer being destroyed | |
2187 | * | |
12dd36fa | 2188 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2189 | * Before calling the slab must have been unlinked from the cache. The |
2190 | * cache-lock is not held/needed. | |
12dd36fa | 2191 | */ |
343e0d7a | 2192 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2193 | { |
2194 | void *addr = slabp->s_mem - slabp->colouroff; | |
2195 | ||
e79aec29 | 2196 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2197 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2198 | struct slab_rcu *slab_rcu; | |
2199 | ||
b28a02de | 2200 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2201 | slab_rcu->cachep = cachep; |
2202 | slab_rcu->addr = addr; | |
2203 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2204 | } else { | |
2205 | kmem_freepages(cachep, addr); | |
873623df IM |
2206 | if (OFF_SLAB(cachep)) |
2207 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2208 | } |
2209 | } | |
2210 | ||
945cf2b6 | 2211 | void __kmem_cache_destroy(struct kmem_cache *cachep) |
117f6eb1 CL |
2212 | { |
2213 | int i; | |
2214 | struct kmem_list3 *l3; | |
2215 | ||
2216 | for_each_online_cpu(i) | |
2217 | kfree(cachep->array[i]); | |
2218 | ||
2219 | /* NUMA: free the list3 structures */ | |
2220 | for_each_online_node(i) { | |
2221 | l3 = cachep->nodelists[i]; | |
2222 | if (l3) { | |
2223 | kfree(l3->shared); | |
2224 | free_alien_cache(l3->alien); | |
2225 | kfree(l3); | |
2226 | } | |
2227 | } | |
117f6eb1 CL |
2228 | } |
2229 | ||
2230 | ||
4d268eba | 2231 | /** |
a70773dd RD |
2232 | * calculate_slab_order - calculate size (page order) of slabs |
2233 | * @cachep: pointer to the cache that is being created | |
2234 | * @size: size of objects to be created in this cache. | |
2235 | * @align: required alignment for the objects. | |
2236 | * @flags: slab allocation flags | |
2237 | * | |
2238 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2239 | * |
2240 | * This could be made much more intelligent. For now, try to avoid using | |
2241 | * high order pages for slabs. When the gfp() functions are more friendly | |
2242 | * towards high-order requests, this should be changed. | |
2243 | */ | |
a737b3e2 | 2244 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2245 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2246 | { |
b1ab41c4 | 2247 | unsigned long offslab_limit; |
4d268eba | 2248 | size_t left_over = 0; |
9888e6fa | 2249 | int gfporder; |
4d268eba | 2250 | |
0aa817f0 | 2251 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2252 | unsigned int num; |
2253 | size_t remainder; | |
2254 | ||
9888e6fa | 2255 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2256 | if (!num) |
2257 | continue; | |
9888e6fa | 2258 | |
b1ab41c4 IM |
2259 | if (flags & CFLGS_OFF_SLAB) { |
2260 | /* | |
2261 | * Max number of objs-per-slab for caches which | |
2262 | * use off-slab slabs. Needed to avoid a possible | |
2263 | * looping condition in cache_grow(). | |
2264 | */ | |
2265 | offslab_limit = size - sizeof(struct slab); | |
2266 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2267 | ||
2268 | if (num > offslab_limit) | |
2269 | break; | |
2270 | } | |
4d268eba | 2271 | |
9888e6fa | 2272 | /* Found something acceptable - save it away */ |
4d268eba | 2273 | cachep->num = num; |
9888e6fa | 2274 | cachep->gfporder = gfporder; |
4d268eba PE |
2275 | left_over = remainder; |
2276 | ||
f78bb8ad LT |
2277 | /* |
2278 | * A VFS-reclaimable slab tends to have most allocations | |
2279 | * as GFP_NOFS and we really don't want to have to be allocating | |
2280 | * higher-order pages when we are unable to shrink dcache. | |
2281 | */ | |
2282 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2283 | break; | |
2284 | ||
4d268eba PE |
2285 | /* |
2286 | * Large number of objects is good, but very large slabs are | |
2287 | * currently bad for the gfp()s. | |
2288 | */ | |
543585cc | 2289 | if (gfporder >= slab_max_order) |
4d268eba PE |
2290 | break; |
2291 | ||
9888e6fa LT |
2292 | /* |
2293 | * Acceptable internal fragmentation? | |
2294 | */ | |
a737b3e2 | 2295 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2296 | break; |
2297 | } | |
2298 | return left_over; | |
2299 | } | |
2300 | ||
83b519e8 | 2301 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2302 | { |
97d06609 | 2303 | if (slab_state >= FULL) |
83b519e8 | 2304 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2305 | |
97d06609 | 2306 | if (slab_state == DOWN) { |
f30cf7d1 PE |
2307 | /* |
2308 | * Note: the first kmem_cache_create must create the cache | |
2309 | * that's used by kmalloc(24), otherwise the creation of | |
2310 | * further caches will BUG(). | |
2311 | */ | |
2312 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2313 | ||
2314 | /* | |
2315 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2316 | * the first cache, then we need to set up all its list3s, | |
2317 | * otherwise the creation of further caches will BUG(). | |
2318 | */ | |
2319 | set_up_list3s(cachep, SIZE_AC); | |
2320 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2321 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2322 | else |
97d06609 | 2323 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 PE |
2324 | } else { |
2325 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2326 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2327 | |
97d06609 | 2328 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2329 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2330 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2331 | } else { |
2332 | int node; | |
556a169d | 2333 | for_each_online_node(node) { |
f30cf7d1 PE |
2334 | cachep->nodelists[node] = |
2335 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2336 | gfp, node); |
f30cf7d1 PE |
2337 | BUG_ON(!cachep->nodelists[node]); |
2338 | kmem_list3_init(cachep->nodelists[node]); | |
2339 | } | |
2340 | } | |
2341 | } | |
7d6e6d09 | 2342 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2343 | jiffies + REAPTIMEOUT_LIST3 + |
2344 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2345 | ||
2346 | cpu_cache_get(cachep)->avail = 0; | |
2347 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2348 | cpu_cache_get(cachep)->batchcount = 1; | |
2349 | cpu_cache_get(cachep)->touched = 0; | |
2350 | cachep->batchcount = 1; | |
2351 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2352 | return 0; |
f30cf7d1 PE |
2353 | } |
2354 | ||
1da177e4 | 2355 | /** |
039363f3 | 2356 | * __kmem_cache_create - Create a cache. |
1da177e4 LT |
2357 | * @name: A string which is used in /proc/slabinfo to identify this cache. |
2358 | * @size: The size of objects to be created in this cache. | |
2359 | * @align: The required alignment for the objects. | |
2360 | * @flags: SLAB flags | |
2361 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2362 | * |
2363 | * Returns a ptr to the cache on success, NULL on failure. | |
2364 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2365 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 LT |
2366 | * |
2367 | * @name must be valid until the cache is destroyed. This implies that | |
a737b3e2 AM |
2368 | * the module calling this has to destroy the cache before getting unloaded. |
2369 | * | |
1da177e4 LT |
2370 | * The flags are |
2371 | * | |
2372 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2373 | * to catch references to uninitialised memory. | |
2374 | * | |
2375 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2376 | * for buffer overruns. | |
2377 | * | |
1da177e4 LT |
2378 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2379 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2380 | * as davem. | |
2381 | */ | |
343e0d7a | 2382 | struct kmem_cache * |
039363f3 | 2383 | __kmem_cache_create (const char *name, size_t size, size_t align, |
51cc5068 | 2384 | unsigned long flags, void (*ctor)(void *)) |
1da177e4 LT |
2385 | { |
2386 | size_t left_over, slab_size, ralign; | |
20cea968 | 2387 | struct kmem_cache *cachep = NULL; |
83b519e8 | 2388 | gfp_t gfp; |
1da177e4 | 2389 | |
1da177e4 | 2390 | #if DEBUG |
1da177e4 LT |
2391 | #if FORCED_DEBUG |
2392 | /* | |
2393 | * Enable redzoning and last user accounting, except for caches with | |
2394 | * large objects, if the increased size would increase the object size | |
2395 | * above the next power of two: caches with object sizes just above a | |
2396 | * power of two have a significant amount of internal fragmentation. | |
2397 | */ | |
87a927c7 DW |
2398 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2399 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2400 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2401 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2402 | flags |= SLAB_POISON; | |
2403 | #endif | |
2404 | if (flags & SLAB_DESTROY_BY_RCU) | |
2405 | BUG_ON(flags & SLAB_POISON); | |
2406 | #endif | |
1da177e4 | 2407 | /* |
a737b3e2 AM |
2408 | * Always checks flags, a caller might be expecting debug support which |
2409 | * isn't available. | |
1da177e4 | 2410 | */ |
40094fa6 | 2411 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2412 | |
a737b3e2 AM |
2413 | /* |
2414 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2415 | * unaligned accesses for some archs when redzoning is used, and makes |
2416 | * sure any on-slab bufctl's are also correctly aligned. | |
2417 | */ | |
b28a02de PE |
2418 | if (size & (BYTES_PER_WORD - 1)) { |
2419 | size += (BYTES_PER_WORD - 1); | |
2420 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2421 | } |
2422 | ||
a737b3e2 AM |
2423 | /* calculate the final buffer alignment: */ |
2424 | ||
1da177e4 LT |
2425 | /* 1) arch recommendation: can be overridden for debug */ |
2426 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2427 | /* |
2428 | * Default alignment: as specified by the arch code. Except if | |
2429 | * an object is really small, then squeeze multiple objects into | |
2430 | * one cacheline. | |
1da177e4 LT |
2431 | */ |
2432 | ralign = cache_line_size(); | |
b28a02de | 2433 | while (size <= ralign / 2) |
1da177e4 LT |
2434 | ralign /= 2; |
2435 | } else { | |
2436 | ralign = BYTES_PER_WORD; | |
2437 | } | |
ca5f9703 PE |
2438 | |
2439 | /* | |
87a927c7 DW |
2440 | * Redzoning and user store require word alignment or possibly larger. |
2441 | * Note this will be overridden by architecture or caller mandated | |
2442 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2443 | */ |
87a927c7 DW |
2444 | if (flags & SLAB_STORE_USER) |
2445 | ralign = BYTES_PER_WORD; | |
2446 | ||
2447 | if (flags & SLAB_RED_ZONE) { | |
2448 | ralign = REDZONE_ALIGN; | |
2449 | /* If redzoning, ensure that the second redzone is suitably | |
2450 | * aligned, by adjusting the object size accordingly. */ | |
2451 | size += REDZONE_ALIGN - 1; | |
2452 | size &= ~(REDZONE_ALIGN - 1); | |
2453 | } | |
ca5f9703 | 2454 | |
a44b56d3 | 2455 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2456 | if (ralign < ARCH_SLAB_MINALIGN) { |
2457 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2458 | } |
a44b56d3 | 2459 | /* 3) caller mandated alignment */ |
1da177e4 LT |
2460 | if (ralign < align) { |
2461 | ralign = align; | |
1da177e4 | 2462 | } |
3ff84a7f PE |
2463 | /* disable debug if necessary */ |
2464 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2465 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2466 | /* |
ca5f9703 | 2467 | * 4) Store it. |
1da177e4 LT |
2468 | */ |
2469 | align = ralign; | |
2470 | ||
83b519e8 PE |
2471 | if (slab_is_available()) |
2472 | gfp = GFP_KERNEL; | |
2473 | else | |
2474 | gfp = GFP_NOWAIT; | |
2475 | ||
1da177e4 | 2476 | /* Get cache's description obj. */ |
9b030cb8 | 2477 | cachep = kmem_cache_zalloc(kmem_cache, gfp); |
1da177e4 | 2478 | if (!cachep) |
039363f3 | 2479 | return NULL; |
1da177e4 | 2480 | |
b56efcf0 | 2481 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
3b0efdfa CL |
2482 | cachep->object_size = size; |
2483 | cachep->align = align; | |
1da177e4 | 2484 | #if DEBUG |
1da177e4 | 2485 | |
ca5f9703 PE |
2486 | /* |
2487 | * Both debugging options require word-alignment which is calculated | |
2488 | * into align above. | |
2489 | */ | |
1da177e4 | 2490 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2491 | /* add space for red zone words */ |
3ff84a7f PE |
2492 | cachep->obj_offset += sizeof(unsigned long long); |
2493 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2494 | } |
2495 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2496 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2497 | * the real object. But if the second red zone needs to be |
2498 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2499 | */ |
87a927c7 DW |
2500 | if (flags & SLAB_RED_ZONE) |
2501 | size += REDZONE_ALIGN; | |
2502 | else | |
2503 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2504 | } |
2505 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2506 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3b0efdfa | 2507 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
1ab335d8 | 2508 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); |
1da177e4 LT |
2509 | size = PAGE_SIZE; |
2510 | } | |
2511 | #endif | |
2512 | #endif | |
2513 | ||
e0a42726 IM |
2514 | /* |
2515 | * Determine if the slab management is 'on' or 'off' slab. | |
2516 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2517 | * it too early on. Always use on-slab management when |
2518 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2519 | */ |
e7cb55b9 CM |
2520 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2521 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2522 | /* |
2523 | * Size is large, assume best to place the slab management obj | |
2524 | * off-slab (should allow better packing of objs). | |
2525 | */ | |
2526 | flags |= CFLGS_OFF_SLAB; | |
2527 | ||
2528 | size = ALIGN(size, align); | |
2529 | ||
f78bb8ad | 2530 | left_over = calculate_slab_order(cachep, size, align, flags); |
1da177e4 LT |
2531 | |
2532 | if (!cachep->num) { | |
b4169525 | 2533 | printk(KERN_ERR |
2534 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
9b030cb8 | 2535 | kmem_cache_free(kmem_cache, cachep); |
039363f3 | 2536 | return NULL; |
1da177e4 | 2537 | } |
b28a02de PE |
2538 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
2539 | + sizeof(struct slab), align); | |
1da177e4 LT |
2540 | |
2541 | /* | |
2542 | * If the slab has been placed off-slab, and we have enough space then | |
2543 | * move it on-slab. This is at the expense of any extra colouring. | |
2544 | */ | |
2545 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2546 | flags &= ~CFLGS_OFF_SLAB; | |
2547 | left_over -= slab_size; | |
2548 | } | |
2549 | ||
2550 | if (flags & CFLGS_OFF_SLAB) { | |
2551 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2552 | slab_size = |
2553 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2554 | |
2555 | #ifdef CONFIG_PAGE_POISONING | |
2556 | /* If we're going to use the generic kernel_map_pages() | |
2557 | * poisoning, then it's going to smash the contents of | |
2558 | * the redzone and userword anyhow, so switch them off. | |
2559 | */ | |
2560 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2561 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2562 | #endif | |
1da177e4 LT |
2563 | } |
2564 | ||
2565 | cachep->colour_off = cache_line_size(); | |
2566 | /* Offset must be a multiple of the alignment. */ | |
2567 | if (cachep->colour_off < align) | |
2568 | cachep->colour_off = align; | |
b28a02de | 2569 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2570 | cachep->slab_size = slab_size; |
2571 | cachep->flags = flags; | |
a618e89f | 2572 | cachep->allocflags = 0; |
4b51d669 | 2573 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2574 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2575 | cachep->size = size; |
6a2d7a95 | 2576 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2577 | |
e5ac9c5a | 2578 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2579 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2580 | /* |
2581 | * This is a possibility for one of the malloc_sizes caches. | |
2582 | * But since we go off slab only for object size greater than | |
2583 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2584 | * this should not happen at all. | |
2585 | * But leave a BUG_ON for some lucky dude. | |
2586 | */ | |
6cb8f913 | 2587 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2588 | } |
1da177e4 | 2589 | cachep->ctor = ctor; |
1da177e4 | 2590 | cachep->name = name; |
7c9adf5a | 2591 | cachep->refcount = 1; |
1da177e4 | 2592 | |
83b519e8 | 2593 | if (setup_cpu_cache(cachep, gfp)) { |
2ed3a4ef | 2594 | __kmem_cache_destroy(cachep); |
039363f3 | 2595 | return NULL; |
2ed3a4ef | 2596 | } |
1da177e4 | 2597 | |
83835b3d PZ |
2598 | if (flags & SLAB_DEBUG_OBJECTS) { |
2599 | /* | |
2600 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2601 | * debug_object_activate()->kmem_cache_alloc(). | |
2602 | */ | |
2603 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2604 | ||
2605 | slab_set_debugobj_lock_classes(cachep); | |
2606 | } | |
2607 | ||
1da177e4 LT |
2608 | return cachep; |
2609 | } | |
1da177e4 LT |
2610 | |
2611 | #if DEBUG | |
2612 | static void check_irq_off(void) | |
2613 | { | |
2614 | BUG_ON(!irqs_disabled()); | |
2615 | } | |
2616 | ||
2617 | static void check_irq_on(void) | |
2618 | { | |
2619 | BUG_ON(irqs_disabled()); | |
2620 | } | |
2621 | ||
343e0d7a | 2622 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2623 | { |
2624 | #ifdef CONFIG_SMP | |
2625 | check_irq_off(); | |
7d6e6d09 | 2626 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2627 | #endif |
2628 | } | |
e498be7d | 2629 | |
343e0d7a | 2630 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2631 | { |
2632 | #ifdef CONFIG_SMP | |
2633 | check_irq_off(); | |
2634 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2635 | #endif | |
2636 | } | |
2637 | ||
1da177e4 LT |
2638 | #else |
2639 | #define check_irq_off() do { } while(0) | |
2640 | #define check_irq_on() do { } while(0) | |
2641 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2642 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2643 | #endif |
2644 | ||
aab2207c CL |
2645 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2646 | struct array_cache *ac, | |
2647 | int force, int node); | |
2648 | ||
1da177e4 LT |
2649 | static void do_drain(void *arg) |
2650 | { | |
a737b3e2 | 2651 | struct kmem_cache *cachep = arg; |
1da177e4 | 2652 | struct array_cache *ac; |
7d6e6d09 | 2653 | int node = numa_mem_id(); |
1da177e4 LT |
2654 | |
2655 | check_irq_off(); | |
9a2dba4b | 2656 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2657 | spin_lock(&cachep->nodelists[node]->list_lock); |
2658 | free_block(cachep, ac->entry, ac->avail, node); | |
2659 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2660 | ac->avail = 0; |
2661 | } | |
2662 | ||
343e0d7a | 2663 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2664 | { |
e498be7d CL |
2665 | struct kmem_list3 *l3; |
2666 | int node; | |
2667 | ||
15c8b6c1 | 2668 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2669 | check_irq_on(); |
b28a02de | 2670 | for_each_online_node(node) { |
e498be7d | 2671 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2672 | if (l3 && l3->alien) |
2673 | drain_alien_cache(cachep, l3->alien); | |
2674 | } | |
2675 | ||
2676 | for_each_online_node(node) { | |
2677 | l3 = cachep->nodelists[node]; | |
2678 | if (l3) | |
aab2207c | 2679 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2680 | } |
1da177e4 LT |
2681 | } |
2682 | ||
ed11d9eb CL |
2683 | /* |
2684 | * Remove slabs from the list of free slabs. | |
2685 | * Specify the number of slabs to drain in tofree. | |
2686 | * | |
2687 | * Returns the actual number of slabs released. | |
2688 | */ | |
2689 | static int drain_freelist(struct kmem_cache *cache, | |
2690 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2691 | { |
ed11d9eb CL |
2692 | struct list_head *p; |
2693 | int nr_freed; | |
1da177e4 | 2694 | struct slab *slabp; |
1da177e4 | 2695 | |
ed11d9eb CL |
2696 | nr_freed = 0; |
2697 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2698 | |
ed11d9eb | 2699 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2700 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2701 | if (p == &l3->slabs_free) { |
2702 | spin_unlock_irq(&l3->list_lock); | |
2703 | goto out; | |
2704 | } | |
1da177e4 | 2705 | |
ed11d9eb | 2706 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2707 | #if DEBUG |
40094fa6 | 2708 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2709 | #endif |
2710 | list_del(&slabp->list); | |
ed11d9eb CL |
2711 | /* |
2712 | * Safe to drop the lock. The slab is no longer linked | |
2713 | * to the cache. | |
2714 | */ | |
2715 | l3->free_objects -= cache->num; | |
e498be7d | 2716 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2717 | slab_destroy(cache, slabp); |
2718 | nr_freed++; | |
1da177e4 | 2719 | } |
ed11d9eb CL |
2720 | out: |
2721 | return nr_freed; | |
1da177e4 LT |
2722 | } |
2723 | ||
18004c5d | 2724 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2725 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2726 | { |
2727 | int ret = 0, i = 0; | |
2728 | struct kmem_list3 *l3; | |
2729 | ||
2730 | drain_cpu_caches(cachep); | |
2731 | ||
2732 | check_irq_on(); | |
2733 | for_each_online_node(i) { | |
2734 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2735 | if (!l3) |
2736 | continue; | |
2737 | ||
2738 | drain_freelist(cachep, l3, l3->free_objects); | |
2739 | ||
2740 | ret += !list_empty(&l3->slabs_full) || | |
2741 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2742 | } |
2743 | return (ret ? 1 : 0); | |
2744 | } | |
2745 | ||
1da177e4 LT |
2746 | /** |
2747 | * kmem_cache_shrink - Shrink a cache. | |
2748 | * @cachep: The cache to shrink. | |
2749 | * | |
2750 | * Releases as many slabs as possible for a cache. | |
2751 | * To help debugging, a zero exit status indicates all slabs were released. | |
2752 | */ | |
343e0d7a | 2753 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2754 | { |
8f5be20b | 2755 | int ret; |
40094fa6 | 2756 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2757 | |
95402b38 | 2758 | get_online_cpus(); |
18004c5d | 2759 | mutex_lock(&slab_mutex); |
8f5be20b | 2760 | ret = __cache_shrink(cachep); |
18004c5d | 2761 | mutex_unlock(&slab_mutex); |
95402b38 | 2762 | put_online_cpus(); |
8f5be20b | 2763 | return ret; |
1da177e4 LT |
2764 | } |
2765 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2766 | ||
945cf2b6 | 2767 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
1da177e4 | 2768 | { |
945cf2b6 | 2769 | return __cache_shrink(cachep); |
1da177e4 | 2770 | } |
1da177e4 | 2771 | |
e5ac9c5a RT |
2772 | /* |
2773 | * Get the memory for a slab management obj. | |
2774 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2775 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2776 | * come from the same cache which is getting created because, | |
2777 | * when we are searching for an appropriate cache for these | |
2778 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2779 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2780 | * kmem_find_general_cachep till the initialization is complete. | |
2781 | * Hence we cannot have slabp_cache same as the original cache. | |
2782 | */ | |
343e0d7a | 2783 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2784 | int colour_off, gfp_t local_flags, |
2785 | int nodeid) | |
1da177e4 LT |
2786 | { |
2787 | struct slab *slabp; | |
b28a02de | 2788 | |
1da177e4 LT |
2789 | if (OFF_SLAB(cachep)) { |
2790 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2791 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2792 | local_flags, nodeid); |
d5cff635 CM |
2793 | /* |
2794 | * If the first object in the slab is leaked (it's allocated | |
2795 | * but no one has a reference to it), we want to make sure | |
2796 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2797 | * to the object. Otherwise we will not report the leak. | |
2798 | */ | |
c017b4be CM |
2799 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2800 | local_flags); | |
1da177e4 LT |
2801 | if (!slabp) |
2802 | return NULL; | |
2803 | } else { | |
b28a02de | 2804 | slabp = objp + colour_off; |
1da177e4 LT |
2805 | colour_off += cachep->slab_size; |
2806 | } | |
2807 | slabp->inuse = 0; | |
2808 | slabp->colouroff = colour_off; | |
b28a02de | 2809 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2810 | slabp->nodeid = nodeid; |
e51bfd0a | 2811 | slabp->free = 0; |
1da177e4 LT |
2812 | return slabp; |
2813 | } | |
2814 | ||
2815 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2816 | { | |
b28a02de | 2817 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2818 | } |
2819 | ||
343e0d7a | 2820 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2821 | struct slab *slabp) |
1da177e4 LT |
2822 | { |
2823 | int i; | |
2824 | ||
2825 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2826 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2827 | #if DEBUG |
2828 | /* need to poison the objs? */ | |
2829 | if (cachep->flags & SLAB_POISON) | |
2830 | poison_obj(cachep, objp, POISON_FREE); | |
2831 | if (cachep->flags & SLAB_STORE_USER) | |
2832 | *dbg_userword(cachep, objp) = NULL; | |
2833 | ||
2834 | if (cachep->flags & SLAB_RED_ZONE) { | |
2835 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2836 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2837 | } | |
2838 | /* | |
a737b3e2 AM |
2839 | * Constructors are not allowed to allocate memory from the same |
2840 | * cache which they are a constructor for. Otherwise, deadlock. | |
2841 | * They must also be threaded. | |
1da177e4 LT |
2842 | */ |
2843 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2844 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2845 | |
2846 | if (cachep->flags & SLAB_RED_ZONE) { | |
2847 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2848 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2849 | " end of an object"); |
1da177e4 LT |
2850 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2851 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2852 | " start of an object"); |
1da177e4 | 2853 | } |
3b0efdfa | 2854 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2855 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2856 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2857 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2858 | #else |
2859 | if (cachep->ctor) | |
51cc5068 | 2860 | cachep->ctor(objp); |
1da177e4 | 2861 | #endif |
b28a02de | 2862 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2863 | } |
b28a02de | 2864 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2865 | } |
2866 | ||
343e0d7a | 2867 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2868 | { |
4b51d669 CL |
2869 | if (CONFIG_ZONE_DMA_FLAG) { |
2870 | if (flags & GFP_DMA) | |
a618e89f | 2871 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2872 | else |
a618e89f | 2873 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2874 | } |
1da177e4 LT |
2875 | } |
2876 | ||
a737b3e2 AM |
2877 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2878 | int nodeid) | |
78d382d7 | 2879 | { |
8fea4e96 | 2880 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2881 | kmem_bufctl_t next; |
2882 | ||
2883 | slabp->inuse++; | |
2884 | next = slab_bufctl(slabp)[slabp->free]; | |
2885 | #if DEBUG | |
2886 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2887 | WARN_ON(slabp->nodeid != nodeid); | |
2888 | #endif | |
2889 | slabp->free = next; | |
2890 | ||
2891 | return objp; | |
2892 | } | |
2893 | ||
a737b3e2 AM |
2894 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2895 | void *objp, int nodeid) | |
78d382d7 | 2896 | { |
8fea4e96 | 2897 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2898 | |
2899 | #if DEBUG | |
2900 | /* Verify that the slab belongs to the intended node */ | |
2901 | WARN_ON(slabp->nodeid != nodeid); | |
2902 | ||
871751e2 | 2903 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2904 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2905 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2906 | BUG(); |
2907 | } | |
2908 | #endif | |
2909 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2910 | slabp->free = objnr; | |
2911 | slabp->inuse--; | |
2912 | } | |
2913 | ||
4776874f PE |
2914 | /* |
2915 | * Map pages beginning at addr to the given cache and slab. This is required | |
2916 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2917 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2918 | */ |
2919 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2920 | void *addr) | |
1da177e4 | 2921 | { |
4776874f | 2922 | int nr_pages; |
1da177e4 LT |
2923 | struct page *page; |
2924 | ||
4776874f | 2925 | page = virt_to_page(addr); |
84097518 | 2926 | |
4776874f | 2927 | nr_pages = 1; |
84097518 | 2928 | if (likely(!PageCompound(page))) |
4776874f PE |
2929 | nr_pages <<= cache->gfporder; |
2930 | ||
1da177e4 | 2931 | do { |
35026088 CL |
2932 | page->slab_cache = cache; |
2933 | page->slab_page = slab; | |
1da177e4 | 2934 | page++; |
4776874f | 2935 | } while (--nr_pages); |
1da177e4 LT |
2936 | } |
2937 | ||
2938 | /* | |
2939 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2940 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2941 | */ | |
3c517a61 CL |
2942 | static int cache_grow(struct kmem_cache *cachep, |
2943 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2944 | { |
b28a02de | 2945 | struct slab *slabp; |
b28a02de PE |
2946 | size_t offset; |
2947 | gfp_t local_flags; | |
e498be7d | 2948 | struct kmem_list3 *l3; |
1da177e4 | 2949 | |
a737b3e2 AM |
2950 | /* |
2951 | * Be lazy and only check for valid flags here, keeping it out of the | |
2952 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2953 | */ |
6cb06229 CL |
2954 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2955 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2956 | |
2e1217cf | 2957 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2958 | check_irq_off(); |
2e1217cf RT |
2959 | l3 = cachep->nodelists[nodeid]; |
2960 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2961 | |
2962 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2963 | offset = l3->colour_next; |
2964 | l3->colour_next++; | |
2965 | if (l3->colour_next >= cachep->colour) | |
2966 | l3->colour_next = 0; | |
2967 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2968 | |
2e1217cf | 2969 | offset *= cachep->colour_off; |
1da177e4 LT |
2970 | |
2971 | if (local_flags & __GFP_WAIT) | |
2972 | local_irq_enable(); | |
2973 | ||
2974 | /* | |
2975 | * The test for missing atomic flag is performed here, rather than | |
2976 | * the more obvious place, simply to reduce the critical path length | |
2977 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2978 | * will eventually be caught here (where it matters). | |
2979 | */ | |
2980 | kmem_flagcheck(cachep, flags); | |
2981 | ||
a737b3e2 AM |
2982 | /* |
2983 | * Get mem for the objs. Attempt to allocate a physical page from | |
2984 | * 'nodeid'. | |
e498be7d | 2985 | */ |
3c517a61 | 2986 | if (!objp) |
b8c1c5da | 2987 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2988 | if (!objp) |
1da177e4 LT |
2989 | goto failed; |
2990 | ||
2991 | /* Get slab management. */ | |
3c517a61 | 2992 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2993 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2994 | if (!slabp) |
1da177e4 LT |
2995 | goto opps1; |
2996 | ||
4776874f | 2997 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2998 | |
a35afb83 | 2999 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
3000 | |
3001 | if (local_flags & __GFP_WAIT) | |
3002 | local_irq_disable(); | |
3003 | check_irq_off(); | |
e498be7d | 3004 | spin_lock(&l3->list_lock); |
1da177e4 LT |
3005 | |
3006 | /* Make slab active. */ | |
e498be7d | 3007 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 3008 | STATS_INC_GROWN(cachep); |
e498be7d CL |
3009 | l3->free_objects += cachep->num; |
3010 | spin_unlock(&l3->list_lock); | |
1da177e4 | 3011 | return 1; |
a737b3e2 | 3012 | opps1: |
1da177e4 | 3013 | kmem_freepages(cachep, objp); |
a737b3e2 | 3014 | failed: |
1da177e4 LT |
3015 | if (local_flags & __GFP_WAIT) |
3016 | local_irq_disable(); | |
3017 | return 0; | |
3018 | } | |
3019 | ||
3020 | #if DEBUG | |
3021 | ||
3022 | /* | |
3023 | * Perform extra freeing checks: | |
3024 | * - detect bad pointers. | |
3025 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
3026 | */ |
3027 | static void kfree_debugcheck(const void *objp) | |
3028 | { | |
1da177e4 LT |
3029 | if (!virt_addr_valid(objp)) { |
3030 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
3031 | (unsigned long)objp); |
3032 | BUG(); | |
1da177e4 | 3033 | } |
1da177e4 LT |
3034 | } |
3035 | ||
58ce1fd5 PE |
3036 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
3037 | { | |
b46b8f19 | 3038 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
3039 | |
3040 | redzone1 = *dbg_redzone1(cache, obj); | |
3041 | redzone2 = *dbg_redzone2(cache, obj); | |
3042 | ||
3043 | /* | |
3044 | * Redzone is ok. | |
3045 | */ | |
3046 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3047 | return; | |
3048 | ||
3049 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3050 | slab_error(cache, "double free detected"); | |
3051 | else | |
3052 | slab_error(cache, "memory outside object was overwritten"); | |
3053 | ||
b46b8f19 | 3054 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
3055 | obj, redzone1, redzone2); |
3056 | } | |
3057 | ||
343e0d7a | 3058 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 3059 | void *caller) |
1da177e4 LT |
3060 | { |
3061 | struct page *page; | |
3062 | unsigned int objnr; | |
3063 | struct slab *slabp; | |
3064 | ||
80cbd911 MW |
3065 | BUG_ON(virt_to_cache(objp) != cachep); |
3066 | ||
3dafccf2 | 3067 | objp -= obj_offset(cachep); |
1da177e4 | 3068 | kfree_debugcheck(objp); |
b49af68f | 3069 | page = virt_to_head_page(objp); |
1da177e4 | 3070 | |
35026088 | 3071 | slabp = page->slab_page; |
1da177e4 LT |
3072 | |
3073 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3074 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3075 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3076 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3077 | } | |
3078 | if (cachep->flags & SLAB_STORE_USER) | |
3079 | *dbg_userword(cachep, objp) = caller; | |
3080 | ||
8fea4e96 | 3081 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3082 | |
3083 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3084 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3085 | |
871751e2 AV |
3086 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3087 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3088 | #endif | |
1da177e4 LT |
3089 | if (cachep->flags & SLAB_POISON) { |
3090 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3091 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 3092 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 3093 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3094 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3095 | } else { |
3096 | poison_obj(cachep, objp, POISON_FREE); | |
3097 | } | |
3098 | #else | |
3099 | poison_obj(cachep, objp, POISON_FREE); | |
3100 | #endif | |
3101 | } | |
3102 | return objp; | |
3103 | } | |
3104 | ||
343e0d7a | 3105 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3106 | { |
3107 | kmem_bufctl_t i; | |
3108 | int entries = 0; | |
b28a02de | 3109 | |
1da177e4 LT |
3110 | /* Check slab's freelist to see if this obj is there. */ |
3111 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3112 | entries++; | |
3113 | if (entries > cachep->num || i >= cachep->num) | |
3114 | goto bad; | |
3115 | } | |
3116 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3117 | bad: |
3118 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3119 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3120 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3121 | print_tainted()); | |
fdde6abb SAS |
3122 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3123 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3124 | 1); | |
1da177e4 LT |
3125 | BUG(); |
3126 | } | |
3127 | } | |
3128 | #else | |
3129 | #define kfree_debugcheck(x) do { } while(0) | |
3130 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3131 | #define check_slabp(x,y) do { } while(0) | |
3132 | #endif | |
3133 | ||
072bb0aa MG |
3134 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
3135 | bool force_refill) | |
1da177e4 LT |
3136 | { |
3137 | int batchcount; | |
3138 | struct kmem_list3 *l3; | |
3139 | struct array_cache *ac; | |
1ca4cb24 PE |
3140 | int node; |
3141 | ||
1da177e4 | 3142 | check_irq_off(); |
7d6e6d09 | 3143 | node = numa_mem_id(); |
072bb0aa MG |
3144 | if (unlikely(force_refill)) |
3145 | goto force_grow; | |
3146 | retry: | |
9a2dba4b | 3147 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3148 | batchcount = ac->batchcount; |
3149 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3150 | /* |
3151 | * If there was little recent activity on this cache, then | |
3152 | * perform only a partial refill. Otherwise we could generate | |
3153 | * refill bouncing. | |
1da177e4 LT |
3154 | */ |
3155 | batchcount = BATCHREFILL_LIMIT; | |
3156 | } | |
1ca4cb24 | 3157 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3158 | |
3159 | BUG_ON(ac->avail > 0 || !l3); | |
3160 | spin_lock(&l3->list_lock); | |
1da177e4 | 3161 | |
3ded175a | 3162 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3163 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3164 | l3->shared->touched = 1; | |
3ded175a | 3165 | goto alloc_done; |
44b57f1c | 3166 | } |
3ded175a | 3167 | |
1da177e4 LT |
3168 | while (batchcount > 0) { |
3169 | struct list_head *entry; | |
3170 | struct slab *slabp; | |
3171 | /* Get slab alloc is to come from. */ | |
3172 | entry = l3->slabs_partial.next; | |
3173 | if (entry == &l3->slabs_partial) { | |
3174 | l3->free_touched = 1; | |
3175 | entry = l3->slabs_free.next; | |
3176 | if (entry == &l3->slabs_free) | |
3177 | goto must_grow; | |
3178 | } | |
3179 | ||
3180 | slabp = list_entry(entry, struct slab, list); | |
3181 | check_slabp(cachep, slabp); | |
3182 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3183 | |
3184 | /* | |
3185 | * The slab was either on partial or free list so | |
3186 | * there must be at least one object available for | |
3187 | * allocation. | |
3188 | */ | |
249b9f33 | 3189 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3190 | |
1da177e4 | 3191 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3192 | STATS_INC_ALLOCED(cachep); |
3193 | STATS_INC_ACTIVE(cachep); | |
3194 | STATS_SET_HIGH(cachep); | |
3195 | ||
072bb0aa MG |
3196 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, |
3197 | node)); | |
1da177e4 LT |
3198 | } |
3199 | check_slabp(cachep, slabp); | |
3200 | ||
3201 | /* move slabp to correct slabp list: */ | |
3202 | list_del(&slabp->list); | |
3203 | if (slabp->free == BUFCTL_END) | |
3204 | list_add(&slabp->list, &l3->slabs_full); | |
3205 | else | |
3206 | list_add(&slabp->list, &l3->slabs_partial); | |
3207 | } | |
3208 | ||
a737b3e2 | 3209 | must_grow: |
1da177e4 | 3210 | l3->free_objects -= ac->avail; |
a737b3e2 | 3211 | alloc_done: |
e498be7d | 3212 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3213 | |
3214 | if (unlikely(!ac->avail)) { | |
3215 | int x; | |
072bb0aa | 3216 | force_grow: |
3c517a61 | 3217 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3218 | |
a737b3e2 | 3219 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3220 | ac = cpu_cache_get(cachep); |
072bb0aa MG |
3221 | |
3222 | /* no objects in sight? abort */ | |
3223 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
3224 | return NULL; |
3225 | ||
a737b3e2 | 3226 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3227 | goto retry; |
3228 | } | |
3229 | ac->touched = 1; | |
072bb0aa MG |
3230 | |
3231 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
3232 | } |
3233 | ||
a737b3e2 AM |
3234 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3235 | gfp_t flags) | |
1da177e4 LT |
3236 | { |
3237 | might_sleep_if(flags & __GFP_WAIT); | |
3238 | #if DEBUG | |
3239 | kmem_flagcheck(cachep, flags); | |
3240 | #endif | |
3241 | } | |
3242 | ||
3243 | #if DEBUG | |
a737b3e2 AM |
3244 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3245 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3246 | { |
b28a02de | 3247 | if (!objp) |
1da177e4 | 3248 | return objp; |
b28a02de | 3249 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3250 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3251 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3252 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3253 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3254 | else |
3255 | check_poison_obj(cachep, objp); | |
3256 | #else | |
3257 | check_poison_obj(cachep, objp); | |
3258 | #endif | |
3259 | poison_obj(cachep, objp, POISON_INUSE); | |
3260 | } | |
3261 | if (cachep->flags & SLAB_STORE_USER) | |
3262 | *dbg_userword(cachep, objp) = caller; | |
3263 | ||
3264 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3265 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3266 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3267 | slab_error(cachep, "double free, or memory outside" | |
3268 | " object was overwritten"); | |
b28a02de | 3269 | printk(KERN_ERR |
b46b8f19 | 3270 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3271 | objp, *dbg_redzone1(cachep, objp), |
3272 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3273 | } |
3274 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3275 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3276 | } | |
871751e2 AV |
3277 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3278 | { | |
3279 | struct slab *slabp; | |
3280 | unsigned objnr; | |
3281 | ||
35026088 | 3282 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3283 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3284 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3285 | } | |
3286 | #endif | |
3dafccf2 | 3287 | objp += obj_offset(cachep); |
4f104934 | 3288 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3289 | cachep->ctor(objp); |
7ea466f2 TH |
3290 | if (ARCH_SLAB_MINALIGN && |
3291 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3292 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3293 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3294 | } |
1da177e4 LT |
3295 | return objp; |
3296 | } | |
3297 | #else | |
3298 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3299 | #endif | |
3300 | ||
773ff60e | 3301 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 | 3302 | { |
9b030cb8 | 3303 | if (cachep == kmem_cache) |
773ff60e | 3304 | return false; |
8a8b6502 | 3305 | |
8c138bc0 | 3306 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3307 | } |
3308 | ||
343e0d7a | 3309 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3310 | { |
b28a02de | 3311 | void *objp; |
1da177e4 | 3312 | struct array_cache *ac; |
072bb0aa | 3313 | bool force_refill = false; |
1da177e4 | 3314 | |
5c382300 | 3315 | check_irq_off(); |
8a8b6502 | 3316 | |
9a2dba4b | 3317 | ac = cpu_cache_get(cachep); |
1da177e4 | 3318 | if (likely(ac->avail)) { |
1da177e4 | 3319 | ac->touched = 1; |
072bb0aa MG |
3320 | objp = ac_get_obj(cachep, ac, flags, false); |
3321 | ||
ddbf2e83 | 3322 | /* |
072bb0aa MG |
3323 | * Allow for the possibility all avail objects are not allowed |
3324 | * by the current flags | |
ddbf2e83 | 3325 | */ |
072bb0aa MG |
3326 | if (objp) { |
3327 | STATS_INC_ALLOCHIT(cachep); | |
3328 | goto out; | |
3329 | } | |
3330 | force_refill = true; | |
1da177e4 | 3331 | } |
072bb0aa MG |
3332 | |
3333 | STATS_INC_ALLOCMISS(cachep); | |
3334 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3335 | /* | |
3336 | * the 'ac' may be updated by cache_alloc_refill(), | |
3337 | * and kmemleak_erase() requires its correct value. | |
3338 | */ | |
3339 | ac = cpu_cache_get(cachep); | |
3340 | ||
3341 | out: | |
d5cff635 CM |
3342 | /* |
3343 | * To avoid a false negative, if an object that is in one of the | |
3344 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3345 | * treat the array pointers as a reference to the object. | |
3346 | */ | |
f3d8b53a O |
3347 | if (objp) |
3348 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3349 | return objp; |
3350 | } | |
3351 | ||
e498be7d | 3352 | #ifdef CONFIG_NUMA |
c61afb18 | 3353 | /* |
b2455396 | 3354 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3355 | * |
3356 | * If we are in_interrupt, then process context, including cpusets and | |
3357 | * mempolicy, may not apply and should not be used for allocation policy. | |
3358 | */ | |
3359 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3360 | { | |
3361 | int nid_alloc, nid_here; | |
3362 | ||
765c4507 | 3363 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3364 | return NULL; |
7d6e6d09 | 3365 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3366 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3367 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3368 | else if (current->mempolicy) |
e7b691b0 | 3369 | nid_alloc = slab_node(); |
c61afb18 | 3370 | if (nid_alloc != nid_here) |
8b98c169 | 3371 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3372 | return NULL; |
3373 | } | |
3374 | ||
765c4507 CL |
3375 | /* |
3376 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3377 | * certain node and fall back is permitted. First we scan all the |
3378 | * available nodelists for available objects. If that fails then we | |
3379 | * perform an allocation without specifying a node. This allows the page | |
3380 | * allocator to do its reclaim / fallback magic. We then insert the | |
3381 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3382 | */ |
8c8cc2c1 | 3383 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3384 | { |
8c8cc2c1 PE |
3385 | struct zonelist *zonelist; |
3386 | gfp_t local_flags; | |
dd1a239f | 3387 | struct zoneref *z; |
54a6eb5c MG |
3388 | struct zone *zone; |
3389 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3390 | void *obj = NULL; |
3c517a61 | 3391 | int nid; |
cc9a6c87 | 3392 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3393 | |
3394 | if (flags & __GFP_THISNODE) | |
3395 | return NULL; | |
3396 | ||
6cb06229 | 3397 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3398 | |
cc9a6c87 MG |
3399 | retry_cpuset: |
3400 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3401 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3402 | |
3c517a61 CL |
3403 | retry: |
3404 | /* | |
3405 | * Look through allowed nodes for objects available | |
3406 | * from existing per node queues. | |
3407 | */ | |
54a6eb5c MG |
3408 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3409 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3410 | |
54a6eb5c | 3411 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3412 | cache->nodelists[nid] && |
481c5346 | 3413 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3414 | obj = ____cache_alloc_node(cache, |
3415 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3416 | if (obj) |
3417 | break; | |
3418 | } | |
3c517a61 CL |
3419 | } |
3420 | ||
cfce6604 | 3421 | if (!obj) { |
3c517a61 CL |
3422 | /* |
3423 | * This allocation will be performed within the constraints | |
3424 | * of the current cpuset / memory policy requirements. | |
3425 | * We may trigger various forms of reclaim on the allowed | |
3426 | * set and go into memory reserves if necessary. | |
3427 | */ | |
dd47ea75 CL |
3428 | if (local_flags & __GFP_WAIT) |
3429 | local_irq_enable(); | |
3430 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3431 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3432 | if (local_flags & __GFP_WAIT) |
3433 | local_irq_disable(); | |
3c517a61 CL |
3434 | if (obj) { |
3435 | /* | |
3436 | * Insert into the appropriate per node queues | |
3437 | */ | |
3438 | nid = page_to_nid(virt_to_page(obj)); | |
3439 | if (cache_grow(cache, flags, nid, obj)) { | |
3440 | obj = ____cache_alloc_node(cache, | |
3441 | flags | GFP_THISNODE, nid); | |
3442 | if (!obj) | |
3443 | /* | |
3444 | * Another processor may allocate the | |
3445 | * objects in the slab since we are | |
3446 | * not holding any locks. | |
3447 | */ | |
3448 | goto retry; | |
3449 | } else { | |
b6a60451 | 3450 | /* cache_grow already freed obj */ |
3c517a61 CL |
3451 | obj = NULL; |
3452 | } | |
3453 | } | |
aedb0eb1 | 3454 | } |
cc9a6c87 MG |
3455 | |
3456 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3457 | goto retry_cpuset; | |
765c4507 CL |
3458 | return obj; |
3459 | } | |
3460 | ||
e498be7d CL |
3461 | /* |
3462 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3463 | */ |
8b98c169 | 3464 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3465 | int nodeid) |
e498be7d CL |
3466 | { |
3467 | struct list_head *entry; | |
b28a02de PE |
3468 | struct slab *slabp; |
3469 | struct kmem_list3 *l3; | |
3470 | void *obj; | |
b28a02de PE |
3471 | int x; |
3472 | ||
3473 | l3 = cachep->nodelists[nodeid]; | |
3474 | BUG_ON(!l3); | |
3475 | ||
a737b3e2 | 3476 | retry: |
ca3b9b91 | 3477 | check_irq_off(); |
b28a02de PE |
3478 | spin_lock(&l3->list_lock); |
3479 | entry = l3->slabs_partial.next; | |
3480 | if (entry == &l3->slabs_partial) { | |
3481 | l3->free_touched = 1; | |
3482 | entry = l3->slabs_free.next; | |
3483 | if (entry == &l3->slabs_free) | |
3484 | goto must_grow; | |
3485 | } | |
3486 | ||
3487 | slabp = list_entry(entry, struct slab, list); | |
3488 | check_spinlock_acquired_node(cachep, nodeid); | |
3489 | check_slabp(cachep, slabp); | |
3490 | ||
3491 | STATS_INC_NODEALLOCS(cachep); | |
3492 | STATS_INC_ACTIVE(cachep); | |
3493 | STATS_SET_HIGH(cachep); | |
3494 | ||
3495 | BUG_ON(slabp->inuse == cachep->num); | |
3496 | ||
78d382d7 | 3497 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3498 | check_slabp(cachep, slabp); |
3499 | l3->free_objects--; | |
3500 | /* move slabp to correct slabp list: */ | |
3501 | list_del(&slabp->list); | |
3502 | ||
a737b3e2 | 3503 | if (slabp->free == BUFCTL_END) |
b28a02de | 3504 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3505 | else |
b28a02de | 3506 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3507 | |
b28a02de PE |
3508 | spin_unlock(&l3->list_lock); |
3509 | goto done; | |
e498be7d | 3510 | |
a737b3e2 | 3511 | must_grow: |
b28a02de | 3512 | spin_unlock(&l3->list_lock); |
3c517a61 | 3513 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3514 | if (x) |
3515 | goto retry; | |
1da177e4 | 3516 | |
8c8cc2c1 | 3517 | return fallback_alloc(cachep, flags); |
e498be7d | 3518 | |
a737b3e2 | 3519 | done: |
b28a02de | 3520 | return obj; |
e498be7d | 3521 | } |
8c8cc2c1 PE |
3522 | |
3523 | /** | |
3524 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3525 | * @cachep: The cache to allocate from. | |
3526 | * @flags: See kmalloc(). | |
3527 | * @nodeid: node number of the target node. | |
3528 | * @caller: return address of caller, used for debug information | |
3529 | * | |
3530 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3531 | * node, which can improve the performance for cpu bound structures. | |
3532 | * | |
3533 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3534 | */ | |
3535 | static __always_inline void * | |
3536 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3537 | void *caller) | |
3538 | { | |
3539 | unsigned long save_flags; | |
3540 | void *ptr; | |
7d6e6d09 | 3541 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3542 | |
dcce284a | 3543 | flags &= gfp_allowed_mask; |
7e85ee0c | 3544 | |
cf40bd16 NP |
3545 | lockdep_trace_alloc(flags); |
3546 | ||
773ff60e | 3547 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3548 | return NULL; |
3549 | ||
8c8cc2c1 PE |
3550 | cache_alloc_debugcheck_before(cachep, flags); |
3551 | local_irq_save(save_flags); | |
3552 | ||
eacbbae3 | 3553 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3554 | nodeid = slab_node; |
8c8cc2c1 PE |
3555 | |
3556 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3557 | /* Node not bootstrapped yet */ | |
3558 | ptr = fallback_alloc(cachep, flags); | |
3559 | goto out; | |
3560 | } | |
3561 | ||
7d6e6d09 | 3562 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3563 | /* |
3564 | * Use the locally cached objects if possible. | |
3565 | * However ____cache_alloc does not allow fallback | |
3566 | * to other nodes. It may fail while we still have | |
3567 | * objects on other nodes available. | |
3568 | */ | |
3569 | ptr = ____cache_alloc(cachep, flags); | |
3570 | if (ptr) | |
3571 | goto out; | |
3572 | } | |
3573 | /* ___cache_alloc_node can fall back to other nodes */ | |
3574 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3575 | out: | |
3576 | local_irq_restore(save_flags); | |
3577 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3578 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3579 | flags); |
8c8cc2c1 | 3580 | |
c175eea4 | 3581 | if (likely(ptr)) |
8c138bc0 | 3582 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3583 | |
d07dbea4 | 3584 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3585 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3586 | |
8c8cc2c1 PE |
3587 | return ptr; |
3588 | } | |
3589 | ||
3590 | static __always_inline void * | |
3591 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3592 | { | |
3593 | void *objp; | |
3594 | ||
3595 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3596 | objp = alternate_node_alloc(cache, flags); | |
3597 | if (objp) | |
3598 | goto out; | |
3599 | } | |
3600 | objp = ____cache_alloc(cache, flags); | |
3601 | ||
3602 | /* | |
3603 | * We may just have run out of memory on the local node. | |
3604 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3605 | */ | |
7d6e6d09 LS |
3606 | if (!objp) |
3607 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3608 | |
3609 | out: | |
3610 | return objp; | |
3611 | } | |
3612 | #else | |
3613 | ||
3614 | static __always_inline void * | |
3615 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3616 | { | |
3617 | return ____cache_alloc(cachep, flags); | |
3618 | } | |
3619 | ||
3620 | #endif /* CONFIG_NUMA */ | |
3621 | ||
3622 | static __always_inline void * | |
3623 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3624 | { | |
3625 | unsigned long save_flags; | |
3626 | void *objp; | |
3627 | ||
dcce284a | 3628 | flags &= gfp_allowed_mask; |
7e85ee0c | 3629 | |
cf40bd16 NP |
3630 | lockdep_trace_alloc(flags); |
3631 | ||
773ff60e | 3632 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3633 | return NULL; |
3634 | ||
8c8cc2c1 PE |
3635 | cache_alloc_debugcheck_before(cachep, flags); |
3636 | local_irq_save(save_flags); | |
3637 | objp = __do_cache_alloc(cachep, flags); | |
3638 | local_irq_restore(save_flags); | |
3639 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3640 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3641 | flags); |
8c8cc2c1 PE |
3642 | prefetchw(objp); |
3643 | ||
c175eea4 | 3644 | if (likely(objp)) |
8c138bc0 | 3645 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3646 | |
d07dbea4 | 3647 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3648 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3649 | |
8c8cc2c1 PE |
3650 | return objp; |
3651 | } | |
e498be7d CL |
3652 | |
3653 | /* | |
3654 | * Caller needs to acquire correct kmem_list's list_lock | |
3655 | */ | |
343e0d7a | 3656 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3657 | int node) |
1da177e4 LT |
3658 | { |
3659 | int i; | |
e498be7d | 3660 | struct kmem_list3 *l3; |
1da177e4 LT |
3661 | |
3662 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3663 | void *objp; |
1da177e4 | 3664 | struct slab *slabp; |
1da177e4 | 3665 | |
072bb0aa MG |
3666 | clear_obj_pfmemalloc(&objpp[i]); |
3667 | objp = objpp[i]; | |
3668 | ||
6ed5eb22 | 3669 | slabp = virt_to_slab(objp); |
ff69416e | 3670 | l3 = cachep->nodelists[node]; |
1da177e4 | 3671 | list_del(&slabp->list); |
ff69416e | 3672 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3673 | check_slabp(cachep, slabp); |
78d382d7 | 3674 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3675 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3676 | l3->free_objects++; |
1da177e4 LT |
3677 | check_slabp(cachep, slabp); |
3678 | ||
3679 | /* fixup slab chains */ | |
3680 | if (slabp->inuse == 0) { | |
e498be7d CL |
3681 | if (l3->free_objects > l3->free_limit) { |
3682 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3683 | /* No need to drop any previously held |
3684 | * lock here, even if we have a off-slab slab | |
3685 | * descriptor it is guaranteed to come from | |
3686 | * a different cache, refer to comments before | |
3687 | * alloc_slabmgmt. | |
3688 | */ | |
1da177e4 LT |
3689 | slab_destroy(cachep, slabp); |
3690 | } else { | |
e498be7d | 3691 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3692 | } |
3693 | } else { | |
3694 | /* Unconditionally move a slab to the end of the | |
3695 | * partial list on free - maximum time for the | |
3696 | * other objects to be freed, too. | |
3697 | */ | |
e498be7d | 3698 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3699 | } |
3700 | } | |
3701 | } | |
3702 | ||
343e0d7a | 3703 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3704 | { |
3705 | int batchcount; | |
e498be7d | 3706 | struct kmem_list3 *l3; |
7d6e6d09 | 3707 | int node = numa_mem_id(); |
1da177e4 LT |
3708 | |
3709 | batchcount = ac->batchcount; | |
3710 | #if DEBUG | |
3711 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3712 | #endif | |
3713 | check_irq_off(); | |
ff69416e | 3714 | l3 = cachep->nodelists[node]; |
873623df | 3715 | spin_lock(&l3->list_lock); |
e498be7d CL |
3716 | if (l3->shared) { |
3717 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3718 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3719 | if (max) { |
3720 | if (batchcount > max) | |
3721 | batchcount = max; | |
e498be7d | 3722 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3723 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3724 | shared_array->avail += batchcount; |
3725 | goto free_done; | |
3726 | } | |
3727 | } | |
3728 | ||
ff69416e | 3729 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3730 | free_done: |
1da177e4 LT |
3731 | #if STATS |
3732 | { | |
3733 | int i = 0; | |
3734 | struct list_head *p; | |
3735 | ||
e498be7d CL |
3736 | p = l3->slabs_free.next; |
3737 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3738 | struct slab *slabp; |
3739 | ||
3740 | slabp = list_entry(p, struct slab, list); | |
3741 | BUG_ON(slabp->inuse); | |
3742 | ||
3743 | i++; | |
3744 | p = p->next; | |
3745 | } | |
3746 | STATS_SET_FREEABLE(cachep, i); | |
3747 | } | |
3748 | #endif | |
e498be7d | 3749 | spin_unlock(&l3->list_lock); |
1da177e4 | 3750 | ac->avail -= batchcount; |
a737b3e2 | 3751 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3752 | } |
3753 | ||
3754 | /* | |
a737b3e2 AM |
3755 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3756 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3757 | */ |
a947eb95 SS |
3758 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3759 | void *caller) | |
1da177e4 | 3760 | { |
9a2dba4b | 3761 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3762 | |
3763 | check_irq_off(); | |
d5cff635 | 3764 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3765 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3766 | |
8c138bc0 | 3767 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3768 | |
1807a1aa SS |
3769 | /* |
3770 | * Skip calling cache_free_alien() when the platform is not numa. | |
3771 | * This will avoid cache misses that happen while accessing slabp (which | |
3772 | * is per page memory reference) to get nodeid. Instead use a global | |
3773 | * variable to skip the call, which is mostly likely to be present in | |
3774 | * the cache. | |
3775 | */ | |
b6e68bc1 | 3776 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3777 | return; |
3778 | ||
1da177e4 LT |
3779 | if (likely(ac->avail < ac->limit)) { |
3780 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3781 | } else { |
3782 | STATS_INC_FREEMISS(cachep); | |
3783 | cache_flusharray(cachep, ac); | |
1da177e4 | 3784 | } |
42c8c99c | 3785 | |
072bb0aa | 3786 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3787 | } |
3788 | ||
3789 | /** | |
3790 | * kmem_cache_alloc - Allocate an object | |
3791 | * @cachep: The cache to allocate from. | |
3792 | * @flags: See kmalloc(). | |
3793 | * | |
3794 | * Allocate an object from this cache. The flags are only relevant | |
3795 | * if the cache has no available objects. | |
3796 | */ | |
343e0d7a | 3797 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3798 | { |
36555751 EGM |
3799 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3800 | ||
ca2b84cb | 3801 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3802 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3803 | |
3804 | return ret; | |
1da177e4 LT |
3805 | } |
3806 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3807 | ||
0f24f128 | 3808 | #ifdef CONFIG_TRACING |
85beb586 SR |
3809 | void * |
3810 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3811 | { |
85beb586 SR |
3812 | void *ret; |
3813 | ||
3814 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3815 | ||
3816 | trace_kmalloc(_RET_IP_, ret, | |
3817 | size, slab_buffer_size(cachep), flags); | |
3818 | return ret; | |
36555751 | 3819 | } |
85beb586 | 3820 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3821 | #endif |
3822 | ||
1da177e4 | 3823 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3824 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3825 | { | |
36555751 EGM |
3826 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3827 | __builtin_return_address(0)); | |
3828 | ||
ca2b84cb | 3829 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3830 | cachep->object_size, cachep->size, |
ca2b84cb | 3831 | flags, nodeid); |
36555751 EGM |
3832 | |
3833 | return ret; | |
8b98c169 | 3834 | } |
1da177e4 LT |
3835 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3836 | ||
0f24f128 | 3837 | #ifdef CONFIG_TRACING |
85beb586 SR |
3838 | void *kmem_cache_alloc_node_trace(size_t size, |
3839 | struct kmem_cache *cachep, | |
3840 | gfp_t flags, | |
3841 | int nodeid) | |
36555751 | 3842 | { |
85beb586 SR |
3843 | void *ret; |
3844 | ||
3845 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3846 | __builtin_return_address(0)); |
85beb586 SR |
3847 | trace_kmalloc_node(_RET_IP_, ret, |
3848 | size, slab_buffer_size(cachep), | |
3849 | flags, nodeid); | |
3850 | return ret; | |
36555751 | 3851 | } |
85beb586 | 3852 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3853 | #endif |
3854 | ||
8b98c169 CH |
3855 | static __always_inline void * |
3856 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3857 | { |
343e0d7a | 3858 | struct kmem_cache *cachep; |
97e2bde4 MS |
3859 | |
3860 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3861 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3862 | return cachep; | |
85beb586 | 3863 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3864 | } |
8b98c169 | 3865 | |
0bb38a5c | 3866 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3867 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3868 | { | |
3869 | return __do_kmalloc_node(size, flags, node, | |
3870 | __builtin_return_address(0)); | |
3871 | } | |
dbe5e69d | 3872 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3873 | |
3874 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3875 | int node, unsigned long caller) |
8b98c169 | 3876 | { |
ce71e27c | 3877 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3878 | } |
3879 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3880 | #else | |
3881 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3882 | { | |
3883 | return __do_kmalloc_node(size, flags, node, NULL); | |
3884 | } | |
3885 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3886 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3887 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3888 | |
3889 | /** | |
800590f5 | 3890 | * __do_kmalloc - allocate memory |
1da177e4 | 3891 | * @size: how many bytes of memory are required. |
800590f5 | 3892 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3893 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3894 | */ |
7fd6b141 PE |
3895 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3896 | void *caller) | |
1da177e4 | 3897 | { |
343e0d7a | 3898 | struct kmem_cache *cachep; |
36555751 | 3899 | void *ret; |
1da177e4 | 3900 | |
97e2bde4 MS |
3901 | /* If you want to save a few bytes .text space: replace |
3902 | * __ with kmem_. | |
3903 | * Then kmalloc uses the uninlined functions instead of the inline | |
3904 | * functions. | |
3905 | */ | |
3906 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3907 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3908 | return cachep; | |
36555751 EGM |
3909 | ret = __cache_alloc(cachep, flags, caller); |
3910 | ||
ca2b84cb | 3911 | trace_kmalloc((unsigned long) caller, ret, |
3b0efdfa | 3912 | size, cachep->size, flags); |
36555751 EGM |
3913 | |
3914 | return ret; | |
7fd6b141 PE |
3915 | } |
3916 | ||
7fd6b141 | 3917 | |
0bb38a5c | 3918 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3919 | void *__kmalloc(size_t size, gfp_t flags) |
3920 | { | |
871751e2 | 3921 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3922 | } |
3923 | EXPORT_SYMBOL(__kmalloc); | |
3924 | ||
ce71e27c | 3925 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3926 | { |
ce71e27c | 3927 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3928 | } |
3929 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3930 | |
3931 | #else | |
3932 | void *__kmalloc(size_t size, gfp_t flags) | |
3933 | { | |
3934 | return __do_kmalloc(size, flags, NULL); | |
3935 | } | |
3936 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3937 | #endif |
3938 | ||
1da177e4 LT |
3939 | /** |
3940 | * kmem_cache_free - Deallocate an object | |
3941 | * @cachep: The cache the allocation was from. | |
3942 | * @objp: The previously allocated object. | |
3943 | * | |
3944 | * Free an object which was previously allocated from this | |
3945 | * cache. | |
3946 | */ | |
343e0d7a | 3947 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3948 | { |
3949 | unsigned long flags; | |
3950 | ||
3951 | local_irq_save(flags); | |
d97d476b | 3952 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3953 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3954 | debug_check_no_obj_freed(objp, cachep->object_size); |
a947eb95 | 3955 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3956 | local_irq_restore(flags); |
36555751 | 3957 | |
ca2b84cb | 3958 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3959 | } |
3960 | EXPORT_SYMBOL(kmem_cache_free); | |
3961 | ||
1da177e4 LT |
3962 | /** |
3963 | * kfree - free previously allocated memory | |
3964 | * @objp: pointer returned by kmalloc. | |
3965 | * | |
80e93eff PE |
3966 | * If @objp is NULL, no operation is performed. |
3967 | * | |
1da177e4 LT |
3968 | * Don't free memory not originally allocated by kmalloc() |
3969 | * or you will run into trouble. | |
3970 | */ | |
3971 | void kfree(const void *objp) | |
3972 | { | |
343e0d7a | 3973 | struct kmem_cache *c; |
1da177e4 LT |
3974 | unsigned long flags; |
3975 | ||
2121db74 PE |
3976 | trace_kfree(_RET_IP_, objp); |
3977 | ||
6cb8f913 | 3978 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3979 | return; |
3980 | local_irq_save(flags); | |
3981 | kfree_debugcheck(objp); | |
6ed5eb22 | 3982 | c = virt_to_cache(objp); |
8c138bc0 CL |
3983 | debug_check_no_locks_freed(objp, c->object_size); |
3984 | ||
3985 | debug_check_no_obj_freed(objp, c->object_size); | |
a947eb95 | 3986 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
3987 | local_irq_restore(flags); |
3988 | } | |
3989 | EXPORT_SYMBOL(kfree); | |
3990 | ||
343e0d7a | 3991 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3992 | { |
8c138bc0 | 3993 | return cachep->object_size; |
1da177e4 LT |
3994 | } |
3995 | EXPORT_SYMBOL(kmem_cache_size); | |
3996 | ||
e498be7d | 3997 | /* |
183ff22b | 3998 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3999 | */ |
83b519e8 | 4000 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
4001 | { |
4002 | int node; | |
4003 | struct kmem_list3 *l3; | |
cafeb02e | 4004 | struct array_cache *new_shared; |
3395ee05 | 4005 | struct array_cache **new_alien = NULL; |
e498be7d | 4006 | |
9c09a95c | 4007 | for_each_online_node(node) { |
cafeb02e | 4008 | |
3395ee05 | 4009 | if (use_alien_caches) { |
83b519e8 | 4010 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
4011 | if (!new_alien) |
4012 | goto fail; | |
4013 | } | |
cafeb02e | 4014 | |
63109846 ED |
4015 | new_shared = NULL; |
4016 | if (cachep->shared) { | |
4017 | new_shared = alloc_arraycache(node, | |
0718dc2a | 4018 | cachep->shared*cachep->batchcount, |
83b519e8 | 4019 | 0xbaadf00d, gfp); |
63109846 ED |
4020 | if (!new_shared) { |
4021 | free_alien_cache(new_alien); | |
4022 | goto fail; | |
4023 | } | |
0718dc2a | 4024 | } |
cafeb02e | 4025 | |
a737b3e2 AM |
4026 | l3 = cachep->nodelists[node]; |
4027 | if (l3) { | |
cafeb02e CL |
4028 | struct array_cache *shared = l3->shared; |
4029 | ||
e498be7d CL |
4030 | spin_lock_irq(&l3->list_lock); |
4031 | ||
cafeb02e | 4032 | if (shared) |
0718dc2a CL |
4033 | free_block(cachep, shared->entry, |
4034 | shared->avail, node); | |
e498be7d | 4035 | |
cafeb02e CL |
4036 | l3->shared = new_shared; |
4037 | if (!l3->alien) { | |
e498be7d CL |
4038 | l3->alien = new_alien; |
4039 | new_alien = NULL; | |
4040 | } | |
b28a02de | 4041 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4042 | cachep->batchcount + cachep->num; |
e498be7d | 4043 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 4044 | kfree(shared); |
e498be7d CL |
4045 | free_alien_cache(new_alien); |
4046 | continue; | |
4047 | } | |
83b519e8 | 4048 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
4049 | if (!l3) { |
4050 | free_alien_cache(new_alien); | |
4051 | kfree(new_shared); | |
e498be7d | 4052 | goto fail; |
0718dc2a | 4053 | } |
e498be7d CL |
4054 | |
4055 | kmem_list3_init(l3); | |
4056 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 4057 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 4058 | l3->shared = new_shared; |
e498be7d | 4059 | l3->alien = new_alien; |
b28a02de | 4060 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4061 | cachep->batchcount + cachep->num; |
e498be7d CL |
4062 | cachep->nodelists[node] = l3; |
4063 | } | |
cafeb02e | 4064 | return 0; |
0718dc2a | 4065 | |
a737b3e2 | 4066 | fail: |
3b0efdfa | 4067 | if (!cachep->list.next) { |
0718dc2a CL |
4068 | /* Cache is not active yet. Roll back what we did */ |
4069 | node--; | |
4070 | while (node >= 0) { | |
4071 | if (cachep->nodelists[node]) { | |
4072 | l3 = cachep->nodelists[node]; | |
4073 | ||
4074 | kfree(l3->shared); | |
4075 | free_alien_cache(l3->alien); | |
4076 | kfree(l3); | |
4077 | cachep->nodelists[node] = NULL; | |
4078 | } | |
4079 | node--; | |
4080 | } | |
4081 | } | |
cafeb02e | 4082 | return -ENOMEM; |
e498be7d CL |
4083 | } |
4084 | ||
1da177e4 | 4085 | struct ccupdate_struct { |
343e0d7a | 4086 | struct kmem_cache *cachep; |
acfe7d74 | 4087 | struct array_cache *new[0]; |
1da177e4 LT |
4088 | }; |
4089 | ||
4090 | static void do_ccupdate_local(void *info) | |
4091 | { | |
a737b3e2 | 4092 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4093 | struct array_cache *old; |
4094 | ||
4095 | check_irq_off(); | |
9a2dba4b | 4096 | old = cpu_cache_get(new->cachep); |
e498be7d | 4097 | |
1da177e4 LT |
4098 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4099 | new->new[smp_processor_id()] = old; | |
4100 | } | |
4101 | ||
18004c5d | 4102 | /* Always called with the slab_mutex held */ |
a737b3e2 | 4103 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4104 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4105 | { |
d2e7b7d0 | 4106 | struct ccupdate_struct *new; |
2ed3a4ef | 4107 | int i; |
1da177e4 | 4108 | |
acfe7d74 ED |
4109 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4110 | gfp); | |
d2e7b7d0 SS |
4111 | if (!new) |
4112 | return -ENOMEM; | |
4113 | ||
e498be7d | 4114 | for_each_online_cpu(i) { |
7d6e6d09 | 4115 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4116 | batchcount, gfp); |
d2e7b7d0 | 4117 | if (!new->new[i]) { |
b28a02de | 4118 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4119 | kfree(new->new[i]); |
4120 | kfree(new); | |
e498be7d | 4121 | return -ENOMEM; |
1da177e4 LT |
4122 | } |
4123 | } | |
d2e7b7d0 | 4124 | new->cachep = cachep; |
1da177e4 | 4125 | |
15c8b6c1 | 4126 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4127 | |
1da177e4 | 4128 | check_irq_on(); |
1da177e4 LT |
4129 | cachep->batchcount = batchcount; |
4130 | cachep->limit = limit; | |
e498be7d | 4131 | cachep->shared = shared; |
1da177e4 | 4132 | |
e498be7d | 4133 | for_each_online_cpu(i) { |
d2e7b7d0 | 4134 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4135 | if (!ccold) |
4136 | continue; | |
7d6e6d09 LS |
4137 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4138 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4139 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4140 | kfree(ccold); |
4141 | } | |
d2e7b7d0 | 4142 | kfree(new); |
83b519e8 | 4143 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4144 | } |
4145 | ||
18004c5d | 4146 | /* Called with slab_mutex held always */ |
83b519e8 | 4147 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4148 | { |
4149 | int err; | |
4150 | int limit, shared; | |
4151 | ||
a737b3e2 AM |
4152 | /* |
4153 | * The head array serves three purposes: | |
1da177e4 LT |
4154 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4155 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4156 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4157 | * bufctl chains: array operations are cheaper. |
4158 | * The numbers are guessed, we should auto-tune as described by | |
4159 | * Bonwick. | |
4160 | */ | |
3b0efdfa | 4161 | if (cachep->size > 131072) |
1da177e4 | 4162 | limit = 1; |
3b0efdfa | 4163 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4164 | limit = 8; |
3b0efdfa | 4165 | else if (cachep->size > 1024) |
1da177e4 | 4166 | limit = 24; |
3b0efdfa | 4167 | else if (cachep->size > 256) |
1da177e4 LT |
4168 | limit = 54; |
4169 | else | |
4170 | limit = 120; | |
4171 | ||
a737b3e2 AM |
4172 | /* |
4173 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4174 | * allocation behaviour: Most allocs on one cpu, most free operations |
4175 | * on another cpu. For these cases, an efficient object passing between | |
4176 | * cpus is necessary. This is provided by a shared array. The array | |
4177 | * replaces Bonwick's magazine layer. | |
4178 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4179 | * to a larger limit. Thus disabled by default. | |
4180 | */ | |
4181 | shared = 0; | |
3b0efdfa | 4182 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4183 | shared = 8; |
1da177e4 LT |
4184 | |
4185 | #if DEBUG | |
a737b3e2 AM |
4186 | /* |
4187 | * With debugging enabled, large batchcount lead to excessively long | |
4188 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4189 | */ |
4190 | if (limit > 32) | |
4191 | limit = 32; | |
4192 | #endif | |
83b519e8 | 4193 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4194 | if (err) |
4195 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4196 | cachep->name, -err); |
2ed3a4ef | 4197 | return err; |
1da177e4 LT |
4198 | } |
4199 | ||
1b55253a CL |
4200 | /* |
4201 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4202 | * necessary. Note that the l3 listlock also protects the array_cache |
4203 | * if drain_array() is used on the shared array. | |
1b55253a | 4204 | */ |
68a1b195 | 4205 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4206 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4207 | { |
4208 | int tofree; | |
4209 | ||
1b55253a CL |
4210 | if (!ac || !ac->avail) |
4211 | return; | |
1da177e4 LT |
4212 | if (ac->touched && !force) { |
4213 | ac->touched = 0; | |
b18e7e65 | 4214 | } else { |
1b55253a | 4215 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4216 | if (ac->avail) { |
4217 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4218 | if (tofree > ac->avail) | |
4219 | tofree = (ac->avail + 1) / 2; | |
4220 | free_block(cachep, ac->entry, tofree, node); | |
4221 | ac->avail -= tofree; | |
4222 | memmove(ac->entry, &(ac->entry[tofree]), | |
4223 | sizeof(void *) * ac->avail); | |
4224 | } | |
1b55253a | 4225 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4226 | } |
4227 | } | |
4228 | ||
4229 | /** | |
4230 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4231 | * @w: work descriptor |
1da177e4 LT |
4232 | * |
4233 | * Called from workqueue/eventd every few seconds. | |
4234 | * Purpose: | |
4235 | * - clear the per-cpu caches for this CPU. | |
4236 | * - return freeable pages to the main free memory pool. | |
4237 | * | |
a737b3e2 AM |
4238 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4239 | * again on the next iteration. | |
1da177e4 | 4240 | */ |
7c5cae36 | 4241 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4242 | { |
7a7c381d | 4243 | struct kmem_cache *searchp; |
e498be7d | 4244 | struct kmem_list3 *l3; |
7d6e6d09 | 4245 | int node = numa_mem_id(); |
bf6aede7 | 4246 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4247 | |
18004c5d | 4248 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 4249 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4250 | goto out; |
1da177e4 | 4251 | |
18004c5d | 4252 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
4253 | check_irq_on(); |
4254 | ||
35386e3b CL |
4255 | /* |
4256 | * We only take the l3 lock if absolutely necessary and we | |
4257 | * have established with reasonable certainty that | |
4258 | * we can do some work if the lock was obtained. | |
4259 | */ | |
aab2207c | 4260 | l3 = searchp->nodelists[node]; |
35386e3b | 4261 | |
8fce4d8e | 4262 | reap_alien(searchp, l3); |
1da177e4 | 4263 | |
aab2207c | 4264 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4265 | |
35386e3b CL |
4266 | /* |
4267 | * These are racy checks but it does not matter | |
4268 | * if we skip one check or scan twice. | |
4269 | */ | |
e498be7d | 4270 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4271 | goto next; |
1da177e4 | 4272 | |
e498be7d | 4273 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4274 | |
aab2207c | 4275 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4276 | |
ed11d9eb | 4277 | if (l3->free_touched) |
e498be7d | 4278 | l3->free_touched = 0; |
ed11d9eb CL |
4279 | else { |
4280 | int freed; | |
1da177e4 | 4281 | |
ed11d9eb CL |
4282 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4283 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4284 | STATS_ADD_REAPED(searchp, freed); | |
4285 | } | |
35386e3b | 4286 | next: |
1da177e4 LT |
4287 | cond_resched(); |
4288 | } | |
4289 | check_irq_on(); | |
18004c5d | 4290 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4291 | next_reap_node(); |
7c5cae36 | 4292 | out: |
a737b3e2 | 4293 | /* Set up the next iteration */ |
7c5cae36 | 4294 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4295 | } |
4296 | ||
158a9624 | 4297 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4298 | |
85289f98 | 4299 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4300 | { |
85289f98 PE |
4301 | /* |
4302 | * Output format version, so at least we can change it | |
4303 | * without _too_ many complaints. | |
4304 | */ | |
1da177e4 | 4305 | #if STATS |
85289f98 | 4306 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4307 | #else |
85289f98 | 4308 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4309 | #endif |
85289f98 PE |
4310 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4311 | "<objperslab> <pagesperslab>"); | |
4312 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4313 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4314 | #if STATS |
85289f98 | 4315 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4316 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4317 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4318 | #endif |
85289f98 PE |
4319 | seq_putc(m, '\n'); |
4320 | } | |
4321 | ||
4322 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4323 | { | |
4324 | loff_t n = *pos; | |
85289f98 | 4325 | |
18004c5d | 4326 | mutex_lock(&slab_mutex); |
85289f98 PE |
4327 | if (!n) |
4328 | print_slabinfo_header(m); | |
b92151ba | 4329 | |
18004c5d | 4330 | return seq_list_start(&slab_caches, *pos); |
1da177e4 LT |
4331 | } |
4332 | ||
4333 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4334 | { | |
18004c5d | 4335 | return seq_list_next(p, &slab_caches, pos); |
1da177e4 LT |
4336 | } |
4337 | ||
4338 | static void s_stop(struct seq_file *m, void *p) | |
4339 | { | |
18004c5d | 4340 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4341 | } |
4342 | ||
4343 | static int s_show(struct seq_file *m, void *p) | |
4344 | { | |
3b0efdfa | 4345 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
b28a02de PE |
4346 | struct slab *slabp; |
4347 | unsigned long active_objs; | |
4348 | unsigned long num_objs; | |
4349 | unsigned long active_slabs = 0; | |
4350 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4351 | const char *name; |
1da177e4 | 4352 | char *error = NULL; |
e498be7d CL |
4353 | int node; |
4354 | struct kmem_list3 *l3; | |
1da177e4 | 4355 | |
1da177e4 LT |
4356 | active_objs = 0; |
4357 | num_slabs = 0; | |
e498be7d CL |
4358 | for_each_online_node(node) { |
4359 | l3 = cachep->nodelists[node]; | |
4360 | if (!l3) | |
4361 | continue; | |
4362 | ||
ca3b9b91 RT |
4363 | check_irq_on(); |
4364 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4365 | |
7a7c381d | 4366 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4367 | if (slabp->inuse != cachep->num && !error) |
4368 | error = "slabs_full accounting error"; | |
4369 | active_objs += cachep->num; | |
4370 | active_slabs++; | |
4371 | } | |
7a7c381d | 4372 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4373 | if (slabp->inuse == cachep->num && !error) |
4374 | error = "slabs_partial inuse accounting error"; | |
4375 | if (!slabp->inuse && !error) | |
4376 | error = "slabs_partial/inuse accounting error"; | |
4377 | active_objs += slabp->inuse; | |
4378 | active_slabs++; | |
4379 | } | |
7a7c381d | 4380 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4381 | if (slabp->inuse && !error) |
4382 | error = "slabs_free/inuse accounting error"; | |
4383 | num_slabs++; | |
4384 | } | |
4385 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4386 | if (l3->shared) |
4387 | shared_avail += l3->shared->avail; | |
e498be7d | 4388 | |
ca3b9b91 | 4389 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4390 | } |
b28a02de PE |
4391 | num_slabs += active_slabs; |
4392 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4393 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4394 | error = "free_objects accounting error"; |
4395 | ||
b28a02de | 4396 | name = cachep->name; |
1da177e4 LT |
4397 | if (error) |
4398 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4399 | ||
4400 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3b0efdfa | 4401 | name, active_objs, num_objs, cachep->size, |
b28a02de | 4402 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4403 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4404 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4405 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4406 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4407 | #if STATS |
b28a02de | 4408 | { /* list3 stats */ |
1da177e4 LT |
4409 | unsigned long high = cachep->high_mark; |
4410 | unsigned long allocs = cachep->num_allocations; | |
4411 | unsigned long grown = cachep->grown; | |
4412 | unsigned long reaped = cachep->reaped; | |
4413 | unsigned long errors = cachep->errors; | |
4414 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4415 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4416 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4417 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4418 | |
e92dd4fd JP |
4419 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4420 | "%4lu %4lu %4lu %4lu %4lu", | |
4421 | allocs, high, grown, | |
4422 | reaped, errors, max_freeable, node_allocs, | |
4423 | node_frees, overflows); | |
1da177e4 LT |
4424 | } |
4425 | /* cpu stats */ | |
4426 | { | |
4427 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4428 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4429 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4430 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4431 | ||
4432 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4433 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4434 | } |
4435 | #endif | |
4436 | seq_putc(m, '\n'); | |
1da177e4 LT |
4437 | return 0; |
4438 | } | |
4439 | ||
4440 | /* | |
4441 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4442 | * | |
4443 | * Output layout: | |
4444 | * cache-name | |
4445 | * num-active-objs | |
4446 | * total-objs | |
4447 | * object size | |
4448 | * num-active-slabs | |
4449 | * total-slabs | |
4450 | * num-pages-per-slab | |
4451 | * + further values on SMP and with statistics enabled | |
4452 | */ | |
4453 | ||
7b3c3a50 | 4454 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4455 | .start = s_start, |
4456 | .next = s_next, | |
4457 | .stop = s_stop, | |
4458 | .show = s_show, | |
1da177e4 LT |
4459 | }; |
4460 | ||
4461 | #define MAX_SLABINFO_WRITE 128 | |
4462 | /** | |
4463 | * slabinfo_write - Tuning for the slab allocator | |
4464 | * @file: unused | |
4465 | * @buffer: user buffer | |
4466 | * @count: data length | |
4467 | * @ppos: unused | |
4468 | */ | |
68a1b195 | 4469 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4470 | size_t count, loff_t *ppos) |
1da177e4 | 4471 | { |
b28a02de | 4472 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4473 | int limit, batchcount, shared, res; |
7a7c381d | 4474 | struct kmem_cache *cachep; |
b28a02de | 4475 | |
1da177e4 LT |
4476 | if (count > MAX_SLABINFO_WRITE) |
4477 | return -EINVAL; | |
4478 | if (copy_from_user(&kbuf, buffer, count)) | |
4479 | return -EFAULT; | |
b28a02de | 4480 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4481 | |
4482 | tmp = strchr(kbuf, ' '); | |
4483 | if (!tmp) | |
4484 | return -EINVAL; | |
4485 | *tmp = '\0'; | |
4486 | tmp++; | |
4487 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4488 | return -EINVAL; | |
4489 | ||
4490 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4491 | mutex_lock(&slab_mutex); |
1da177e4 | 4492 | res = -EINVAL; |
18004c5d | 4493 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4494 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4495 | if (limit < 1 || batchcount < 1 || |
4496 | batchcount > limit || shared < 0) { | |
e498be7d | 4497 | res = 0; |
1da177e4 | 4498 | } else { |
e498be7d | 4499 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4500 | batchcount, shared, |
4501 | GFP_KERNEL); | |
1da177e4 LT |
4502 | } |
4503 | break; | |
4504 | } | |
4505 | } | |
18004c5d | 4506 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4507 | if (res >= 0) |
4508 | res = count; | |
4509 | return res; | |
4510 | } | |
871751e2 | 4511 | |
7b3c3a50 AD |
4512 | static int slabinfo_open(struct inode *inode, struct file *file) |
4513 | { | |
4514 | return seq_open(file, &slabinfo_op); | |
4515 | } | |
4516 | ||
4517 | static const struct file_operations proc_slabinfo_operations = { | |
4518 | .open = slabinfo_open, | |
4519 | .read = seq_read, | |
4520 | .write = slabinfo_write, | |
4521 | .llseek = seq_lseek, | |
4522 | .release = seq_release, | |
4523 | }; | |
4524 | ||
871751e2 AV |
4525 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4526 | ||
4527 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4528 | { | |
18004c5d CL |
4529 | mutex_lock(&slab_mutex); |
4530 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4531 | } |
4532 | ||
4533 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4534 | { | |
4535 | unsigned long *p; | |
4536 | int l; | |
4537 | if (!v) | |
4538 | return 1; | |
4539 | l = n[1]; | |
4540 | p = n + 2; | |
4541 | while (l) { | |
4542 | int i = l/2; | |
4543 | unsigned long *q = p + 2 * i; | |
4544 | if (*q == v) { | |
4545 | q[1]++; | |
4546 | return 1; | |
4547 | } | |
4548 | if (*q > v) { | |
4549 | l = i; | |
4550 | } else { | |
4551 | p = q + 2; | |
4552 | l -= i + 1; | |
4553 | } | |
4554 | } | |
4555 | if (++n[1] == n[0]) | |
4556 | return 0; | |
4557 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4558 | p[0] = v; | |
4559 | p[1] = 1; | |
4560 | return 1; | |
4561 | } | |
4562 | ||
4563 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4564 | { | |
4565 | void *p; | |
4566 | int i; | |
4567 | if (n[0] == n[1]) | |
4568 | return; | |
3b0efdfa | 4569 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4570 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4571 | continue; | |
4572 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4573 | return; | |
4574 | } | |
4575 | } | |
4576 | ||
4577 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4578 | { | |
4579 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4580 | unsigned long offset, size; |
9281acea | 4581 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4582 | |
a5c43dae | 4583 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4584 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4585 | if (modname[0]) |
871751e2 AV |
4586 | seq_printf(m, " [%s]", modname); |
4587 | return; | |
4588 | } | |
4589 | #endif | |
4590 | seq_printf(m, "%p", (void *)address); | |
4591 | } | |
4592 | ||
4593 | static int leaks_show(struct seq_file *m, void *p) | |
4594 | { | |
0672aa7c | 4595 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 AV |
4596 | struct slab *slabp; |
4597 | struct kmem_list3 *l3; | |
4598 | const char *name; | |
4599 | unsigned long *n = m->private; | |
4600 | int node; | |
4601 | int i; | |
4602 | ||
4603 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4604 | return 0; | |
4605 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4606 | return 0; | |
4607 | ||
4608 | /* OK, we can do it */ | |
4609 | ||
4610 | n[1] = 0; | |
4611 | ||
4612 | for_each_online_node(node) { | |
4613 | l3 = cachep->nodelists[node]; | |
4614 | if (!l3) | |
4615 | continue; | |
4616 | ||
4617 | check_irq_on(); | |
4618 | spin_lock_irq(&l3->list_lock); | |
4619 | ||
7a7c381d | 4620 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4621 | handle_slab(n, cachep, slabp); |
7a7c381d | 4622 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4623 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4624 | spin_unlock_irq(&l3->list_lock); |
4625 | } | |
4626 | name = cachep->name; | |
4627 | if (n[0] == n[1]) { | |
4628 | /* Increase the buffer size */ | |
18004c5d | 4629 | mutex_unlock(&slab_mutex); |
871751e2 AV |
4630 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4631 | if (!m->private) { | |
4632 | /* Too bad, we are really out */ | |
4633 | m->private = n; | |
18004c5d | 4634 | mutex_lock(&slab_mutex); |
871751e2 AV |
4635 | return -ENOMEM; |
4636 | } | |
4637 | *(unsigned long *)m->private = n[0] * 2; | |
4638 | kfree(n); | |
18004c5d | 4639 | mutex_lock(&slab_mutex); |
871751e2 AV |
4640 | /* Now make sure this entry will be retried */ |
4641 | m->count = m->size; | |
4642 | return 0; | |
4643 | } | |
4644 | for (i = 0; i < n[1]; i++) { | |
4645 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4646 | show_symbol(m, n[2*i+2]); | |
4647 | seq_putc(m, '\n'); | |
4648 | } | |
d2e7b7d0 | 4649 | |
871751e2 AV |
4650 | return 0; |
4651 | } | |
4652 | ||
a0ec95a8 | 4653 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4654 | .start = leaks_start, |
4655 | .next = s_next, | |
4656 | .stop = s_stop, | |
4657 | .show = leaks_show, | |
4658 | }; | |
a0ec95a8 AD |
4659 | |
4660 | static int slabstats_open(struct inode *inode, struct file *file) | |
4661 | { | |
4662 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4663 | int ret = -ENOMEM; | |
4664 | if (n) { | |
4665 | ret = seq_open(file, &slabstats_op); | |
4666 | if (!ret) { | |
4667 | struct seq_file *m = file->private_data; | |
4668 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4669 | m->private = n; | |
4670 | n = NULL; | |
4671 | } | |
4672 | kfree(n); | |
4673 | } | |
4674 | return ret; | |
4675 | } | |
4676 | ||
4677 | static const struct file_operations proc_slabstats_operations = { | |
4678 | .open = slabstats_open, | |
4679 | .read = seq_read, | |
4680 | .llseek = seq_lseek, | |
4681 | .release = seq_release_private, | |
4682 | }; | |
4683 | #endif | |
4684 | ||
4685 | static int __init slab_proc_init(void) | |
4686 | { | |
ab067e99 | 4687 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4688 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4689 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4690 | #endif |
a0ec95a8 AD |
4691 | return 0; |
4692 | } | |
4693 | module_init(slab_proc_init); | |
1da177e4 LT |
4694 | #endif |
4695 | ||
00e145b6 MS |
4696 | /** |
4697 | * ksize - get the actual amount of memory allocated for a given object | |
4698 | * @objp: Pointer to the object | |
4699 | * | |
4700 | * kmalloc may internally round up allocations and return more memory | |
4701 | * than requested. ksize() can be used to determine the actual amount of | |
4702 | * memory allocated. The caller may use this additional memory, even though | |
4703 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4704 | * The caller must guarantee that objp points to a valid object previously | |
4705 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4706 | * must not be freed during the duration of the call. | |
4707 | */ | |
fd76bab2 | 4708 | size_t ksize(const void *objp) |
1da177e4 | 4709 | { |
ef8b4520 CL |
4710 | BUG_ON(!objp); |
4711 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4712 | return 0; |
1da177e4 | 4713 | |
8c138bc0 | 4714 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4715 | } |
b1aabecd | 4716 | EXPORT_SYMBOL(ksize); |