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 | ||
278b1bb1 | 1679 | sizes[INDEX_AC].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1680 | sizes[INDEX_AC].cs_cachep->name = names[INDEX_AC].name; |
1681 | sizes[INDEX_AC].cs_cachep->size = sizes[INDEX_AC].cs_size; | |
1682 | sizes[INDEX_AC].cs_cachep->object_size = sizes[INDEX_AC].cs_size; | |
1683 | sizes[INDEX_AC].cs_cachep->align = ARCH_KMALLOC_MINALIGN; | |
1684 | __kmem_cache_create(sizes[INDEX_AC].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC); | |
7c9adf5a | 1685 | list_add(&sizes[INDEX_AC].cs_cachep->list, &slab_caches); |
8a13a4cc | 1686 | |
a737b3e2 | 1687 | if (INDEX_AC != INDEX_L3) { |
278b1bb1 | 1688 | sizes[INDEX_L3].cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1689 | sizes[INDEX_L3].cs_cachep->name = names[INDEX_L3].name; |
1690 | sizes[INDEX_L3].cs_cachep->size = sizes[INDEX_L3].cs_size; | |
1691 | sizes[INDEX_L3].cs_cachep->object_size = sizes[INDEX_L3].cs_size; | |
1692 | sizes[INDEX_L3].cs_cachep->align = ARCH_KMALLOC_MINALIGN; | |
1693 | __kmem_cache_create(sizes[INDEX_L3].cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC); | |
7c9adf5a | 1694 | list_add(&sizes[INDEX_L3].cs_cachep->list, &slab_caches); |
a737b3e2 | 1695 | } |
e498be7d | 1696 | |
e0a42726 IM |
1697 | slab_early_init = 0; |
1698 | ||
1da177e4 | 1699 | while (sizes->cs_size != ULONG_MAX) { |
e498be7d CL |
1700 | /* |
1701 | * For performance, all the general caches are L1 aligned. | |
1da177e4 LT |
1702 | * This should be particularly beneficial on SMP boxes, as it |
1703 | * eliminates "false sharing". | |
1704 | * Note for systems short on memory removing the alignment will | |
e498be7d CL |
1705 | * allow tighter packing of the smaller caches. |
1706 | */ | |
a737b3e2 | 1707 | if (!sizes->cs_cachep) { |
278b1bb1 | 1708 | sizes->cs_cachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1709 | sizes->cs_cachep->name = names->name; |
1710 | sizes->cs_cachep->size = sizes->cs_size; | |
1711 | sizes->cs_cachep->object_size = sizes->cs_size; | |
1712 | sizes->cs_cachep->align = ARCH_KMALLOC_MINALIGN; | |
1713 | __kmem_cache_create(sizes->cs_cachep, ARCH_KMALLOC_FLAGS|SLAB_PANIC); | |
7c9adf5a | 1714 | list_add(&sizes->cs_cachep->list, &slab_caches); |
a737b3e2 | 1715 | } |
4b51d669 | 1716 | #ifdef CONFIG_ZONE_DMA |
278b1bb1 | 1717 | sizes->cs_dmacachep = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT); |
8a13a4cc CL |
1718 | sizes->cs_dmacachep->name = names->name_dma; |
1719 | sizes->cs_dmacachep->size = sizes->cs_size; | |
1720 | sizes->cs_dmacachep->object_size = sizes->cs_size; | |
1721 | sizes->cs_dmacachep->align = ARCH_KMALLOC_MINALIGN; | |
278b1bb1 | 1722 | __kmem_cache_create(sizes->cs_dmacachep, |
8a13a4cc | 1723 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| SLAB_PANIC); |
7c9adf5a | 1724 | list_add(&sizes->cs_dmacachep->list, &slab_caches); |
4b51d669 | 1725 | #endif |
1da177e4 LT |
1726 | sizes++; |
1727 | names++; | |
1728 | } | |
1729 | /* 4) Replace the bootstrap head arrays */ | |
1730 | { | |
2b2d5493 | 1731 | struct array_cache *ptr; |
e498be7d | 1732 | |
83b519e8 | 1733 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1734 | |
9b030cb8 CL |
1735 | BUG_ON(cpu_cache_get(kmem_cache) != &initarray_cache.cache); |
1736 | memcpy(ptr, cpu_cache_get(kmem_cache), | |
b28a02de | 1737 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1738 | /* |
1739 | * Do not assume that spinlocks can be initialized via memcpy: | |
1740 | */ | |
1741 | spin_lock_init(&ptr->lock); | |
1742 | ||
9b030cb8 | 1743 | kmem_cache->array[smp_processor_id()] = ptr; |
e498be7d | 1744 | |
83b519e8 | 1745 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); |
e498be7d | 1746 | |
9a2dba4b | 1747 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) |
b28a02de | 1748 | != &initarray_generic.cache); |
9a2dba4b | 1749 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), |
b28a02de | 1750 | sizeof(struct arraycache_init)); |
2b2d5493 IM |
1751 | /* |
1752 | * Do not assume that spinlocks can be initialized via memcpy: | |
1753 | */ | |
1754 | spin_lock_init(&ptr->lock); | |
1755 | ||
e498be7d | 1756 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = |
b28a02de | 1757 | ptr; |
1da177e4 | 1758 | } |
e498be7d CL |
1759 | /* 5) Replace the bootstrap kmem_list3's */ |
1760 | { | |
1ca4cb24 PE |
1761 | int nid; |
1762 | ||
9c09a95c | 1763 | for_each_online_node(nid) { |
9b030cb8 | 1764 | init_list(kmem_cache, &initkmem_list3[CACHE_CACHE + nid], nid); |
556a169d | 1765 | |
e498be7d | 1766 | init_list(malloc_sizes[INDEX_AC].cs_cachep, |
1ca4cb24 | 1767 | &initkmem_list3[SIZE_AC + nid], nid); |
e498be7d CL |
1768 | |
1769 | if (INDEX_AC != INDEX_L3) { | |
1770 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1ca4cb24 | 1771 | &initkmem_list3[SIZE_L3 + nid], nid); |
e498be7d CL |
1772 | } |
1773 | } | |
1774 | } | |
1da177e4 | 1775 | |
97d06609 | 1776 | slab_state = UP; |
8429db5c PE |
1777 | } |
1778 | ||
1779 | void __init kmem_cache_init_late(void) | |
1780 | { | |
1781 | struct kmem_cache *cachep; | |
1782 | ||
97d06609 | 1783 | slab_state = UP; |
52cef189 | 1784 | |
30765b92 PZ |
1785 | /* Annotate slab for lockdep -- annotate the malloc caches */ |
1786 | init_lock_keys(); | |
1787 | ||
8429db5c | 1788 | /* 6) resize the head arrays to their final sizes */ |
18004c5d CL |
1789 | mutex_lock(&slab_mutex); |
1790 | list_for_each_entry(cachep, &slab_caches, list) | |
8429db5c PE |
1791 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1792 | BUG(); | |
18004c5d | 1793 | mutex_unlock(&slab_mutex); |
056c6241 | 1794 | |
97d06609 CL |
1795 | /* Done! */ |
1796 | slab_state = FULL; | |
1797 | ||
a737b3e2 AM |
1798 | /* |
1799 | * Register a cpu startup notifier callback that initializes | |
1800 | * cpu_cache_get for all new cpus | |
1da177e4 LT |
1801 | */ |
1802 | register_cpu_notifier(&cpucache_notifier); | |
1da177e4 | 1803 | |
8f9f8d9e DR |
1804 | #ifdef CONFIG_NUMA |
1805 | /* | |
1806 | * Register a memory hotplug callback that initializes and frees | |
1807 | * nodelists. | |
1808 | */ | |
1809 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1810 | #endif | |
1811 | ||
a737b3e2 AM |
1812 | /* |
1813 | * The reap timers are started later, with a module init call: That part | |
1814 | * of the kernel is not yet operational. | |
1da177e4 LT |
1815 | */ |
1816 | } | |
1817 | ||
1818 | static int __init cpucache_init(void) | |
1819 | { | |
1820 | int cpu; | |
1821 | ||
a737b3e2 AM |
1822 | /* |
1823 | * Register the timers that return unneeded pages to the page allocator | |
1da177e4 | 1824 | */ |
e498be7d | 1825 | for_each_online_cpu(cpu) |
a737b3e2 | 1826 | start_cpu_timer(cpu); |
a164f896 GC |
1827 | |
1828 | /* Done! */ | |
97d06609 | 1829 | slab_state = FULL; |
1da177e4 LT |
1830 | return 0; |
1831 | } | |
1da177e4 LT |
1832 | __initcall(cpucache_init); |
1833 | ||
8bdec192 RA |
1834 | static noinline void |
1835 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1836 | { | |
1837 | struct kmem_list3 *l3; | |
1838 | struct slab *slabp; | |
1839 | unsigned long flags; | |
1840 | int node; | |
1841 | ||
1842 | printk(KERN_WARNING | |
1843 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1844 | nodeid, gfpflags); | |
1845 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
3b0efdfa | 1846 | cachep->name, cachep->size, cachep->gfporder); |
8bdec192 RA |
1847 | |
1848 | for_each_online_node(node) { | |
1849 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1850 | unsigned long active_slabs = 0, num_slabs = 0; | |
1851 | ||
1852 | l3 = cachep->nodelists[node]; | |
1853 | if (!l3) | |
1854 | continue; | |
1855 | ||
1856 | spin_lock_irqsave(&l3->list_lock, flags); | |
1857 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1858 | active_objs += cachep->num; | |
1859 | active_slabs++; | |
1860 | } | |
1861 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1862 | active_objs += slabp->inuse; | |
1863 | active_slabs++; | |
1864 | } | |
1865 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1866 | num_slabs++; | |
1867 | ||
1868 | free_objects += l3->free_objects; | |
1869 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1870 | ||
1871 | num_slabs += active_slabs; | |
1872 | num_objs = num_slabs * cachep->num; | |
1873 | printk(KERN_WARNING | |
1874 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1875 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1876 | free_objects); | |
1877 | } | |
1878 | } | |
1879 | ||
1da177e4 LT |
1880 | /* |
1881 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1882 | * | |
1883 | * If we requested dmaable memory, we will get it. Even if we | |
1884 | * did not request dmaable memory, we might get it, but that | |
1885 | * would be relatively rare and ignorable. | |
1886 | */ | |
343e0d7a | 1887 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
1da177e4 LT |
1888 | { |
1889 | struct page *page; | |
e1b6aa6f | 1890 | int nr_pages; |
1da177e4 LT |
1891 | int i; |
1892 | ||
d6fef9da | 1893 | #ifndef CONFIG_MMU |
e1b6aa6f CH |
1894 | /* |
1895 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1896 | * requires __GFP_COMP to properly refcount higher order allocations | |
d6fef9da | 1897 | */ |
e1b6aa6f | 1898 | flags |= __GFP_COMP; |
d6fef9da | 1899 | #endif |
765c4507 | 1900 | |
a618e89f | 1901 | flags |= cachep->allocflags; |
e12ba74d MG |
1902 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1903 | flags |= __GFP_RECLAIMABLE; | |
e1b6aa6f | 1904 | |
517d0869 | 1905 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
8bdec192 RA |
1906 | if (!page) { |
1907 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1908 | slab_out_of_memory(cachep, flags, nodeid); | |
1da177e4 | 1909 | return NULL; |
8bdec192 | 1910 | } |
1da177e4 | 1911 | |
b37f1dd0 | 1912 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
072bb0aa MG |
1913 | if (unlikely(page->pfmemalloc)) |
1914 | pfmemalloc_active = true; | |
1915 | ||
e1b6aa6f | 1916 | nr_pages = (1 << cachep->gfporder); |
1da177e4 | 1917 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
972d1a7b CL |
1918 | add_zone_page_state(page_zone(page), |
1919 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1920 | else | |
1921 | add_zone_page_state(page_zone(page), | |
1922 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
072bb0aa | 1923 | for (i = 0; i < nr_pages; i++) { |
e1b6aa6f | 1924 | __SetPageSlab(page + i); |
c175eea4 | 1925 | |
072bb0aa MG |
1926 | if (page->pfmemalloc) |
1927 | SetPageSlabPfmemalloc(page + i); | |
1928 | } | |
1929 | ||
b1eeab67 VN |
1930 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1931 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1932 | ||
1933 | if (cachep->ctor) | |
1934 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1935 | else | |
1936 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1937 | } | |
c175eea4 | 1938 | |
e1b6aa6f | 1939 | return page_address(page); |
1da177e4 LT |
1940 | } |
1941 | ||
1942 | /* | |
1943 | * Interface to system's page release. | |
1944 | */ | |
343e0d7a | 1945 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) |
1da177e4 | 1946 | { |
b28a02de | 1947 | unsigned long i = (1 << cachep->gfporder); |
1da177e4 LT |
1948 | struct page *page = virt_to_page(addr); |
1949 | const unsigned long nr_freed = i; | |
1950 | ||
b1eeab67 | 1951 | kmemcheck_free_shadow(page, cachep->gfporder); |
c175eea4 | 1952 | |
972d1a7b CL |
1953 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1954 | sub_zone_page_state(page_zone(page), | |
1955 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1956 | else | |
1957 | sub_zone_page_state(page_zone(page), | |
1958 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1da177e4 | 1959 | while (i--) { |
f205b2fe | 1960 | BUG_ON(!PageSlab(page)); |
072bb0aa | 1961 | __ClearPageSlabPfmemalloc(page); |
f205b2fe | 1962 | __ClearPageSlab(page); |
1da177e4 LT |
1963 | page++; |
1964 | } | |
1da177e4 LT |
1965 | if (current->reclaim_state) |
1966 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1967 | free_pages((unsigned long)addr, cachep->gfporder); | |
1da177e4 LT |
1968 | } |
1969 | ||
1970 | static void kmem_rcu_free(struct rcu_head *head) | |
1971 | { | |
b28a02de | 1972 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; |
343e0d7a | 1973 | struct kmem_cache *cachep = slab_rcu->cachep; |
1da177e4 LT |
1974 | |
1975 | kmem_freepages(cachep, slab_rcu->addr); | |
1976 | if (OFF_SLAB(cachep)) | |
1977 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1978 | } | |
1979 | ||
1980 | #if DEBUG | |
1981 | ||
1982 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
343e0d7a | 1983 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
b28a02de | 1984 | unsigned long caller) |
1da177e4 | 1985 | { |
8c138bc0 | 1986 | int size = cachep->object_size; |
1da177e4 | 1987 | |
3dafccf2 | 1988 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1da177e4 | 1989 | |
b28a02de | 1990 | if (size < 5 * sizeof(unsigned long)) |
1da177e4 LT |
1991 | return; |
1992 | ||
b28a02de PE |
1993 | *addr++ = 0x12345678; |
1994 | *addr++ = caller; | |
1995 | *addr++ = smp_processor_id(); | |
1996 | size -= 3 * sizeof(unsigned long); | |
1da177e4 LT |
1997 | { |
1998 | unsigned long *sptr = &caller; | |
1999 | unsigned long svalue; | |
2000 | ||
2001 | while (!kstack_end(sptr)) { | |
2002 | svalue = *sptr++; | |
2003 | if (kernel_text_address(svalue)) { | |
b28a02de | 2004 | *addr++ = svalue; |
1da177e4 LT |
2005 | size -= sizeof(unsigned long); |
2006 | if (size <= sizeof(unsigned long)) | |
2007 | break; | |
2008 | } | |
2009 | } | |
2010 | ||
2011 | } | |
b28a02de | 2012 | *addr++ = 0x87654321; |
1da177e4 LT |
2013 | } |
2014 | #endif | |
2015 | ||
343e0d7a | 2016 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1da177e4 | 2017 | { |
8c138bc0 | 2018 | int size = cachep->object_size; |
3dafccf2 | 2019 | addr = &((char *)addr)[obj_offset(cachep)]; |
1da177e4 LT |
2020 | |
2021 | memset(addr, val, size); | |
b28a02de | 2022 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1da177e4 LT |
2023 | } |
2024 | ||
2025 | static void dump_line(char *data, int offset, int limit) | |
2026 | { | |
2027 | int i; | |
aa83aa40 DJ |
2028 | unsigned char error = 0; |
2029 | int bad_count = 0; | |
2030 | ||
fdde6abb | 2031 | printk(KERN_ERR "%03x: ", offset); |
aa83aa40 DJ |
2032 | for (i = 0; i < limit; i++) { |
2033 | if (data[offset + i] != POISON_FREE) { | |
2034 | error = data[offset + i]; | |
2035 | bad_count++; | |
2036 | } | |
aa83aa40 | 2037 | } |
fdde6abb SAS |
2038 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
2039 | &data[offset], limit, 1); | |
aa83aa40 DJ |
2040 | |
2041 | if (bad_count == 1) { | |
2042 | error ^= POISON_FREE; | |
2043 | if (!(error & (error - 1))) { | |
2044 | printk(KERN_ERR "Single bit error detected. Probably " | |
2045 | "bad RAM.\n"); | |
2046 | #ifdef CONFIG_X86 | |
2047 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
2048 | "test tool.\n"); | |
2049 | #else | |
2050 | printk(KERN_ERR "Run a memory test tool.\n"); | |
2051 | #endif | |
2052 | } | |
2053 | } | |
1da177e4 LT |
2054 | } |
2055 | #endif | |
2056 | ||
2057 | #if DEBUG | |
2058 | ||
343e0d7a | 2059 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1da177e4 LT |
2060 | { |
2061 | int i, size; | |
2062 | char *realobj; | |
2063 | ||
2064 | if (cachep->flags & SLAB_RED_ZONE) { | |
b46b8f19 | 2065 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", |
a737b3e2 AM |
2066 | *dbg_redzone1(cachep, objp), |
2067 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
2068 | } |
2069 | ||
2070 | if (cachep->flags & SLAB_STORE_USER) { | |
2071 | printk(KERN_ERR "Last user: [<%p>]", | |
a737b3e2 | 2072 | *dbg_userword(cachep, objp)); |
1da177e4 | 2073 | print_symbol("(%s)", |
a737b3e2 | 2074 | (unsigned long)*dbg_userword(cachep, objp)); |
1da177e4 LT |
2075 | printk("\n"); |
2076 | } | |
3dafccf2 | 2077 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2078 | size = cachep->object_size; |
b28a02de | 2079 | for (i = 0; i < size && lines; i += 16, lines--) { |
1da177e4 LT |
2080 | int limit; |
2081 | limit = 16; | |
b28a02de PE |
2082 | if (i + limit > size) |
2083 | limit = size - i; | |
1da177e4 LT |
2084 | dump_line(realobj, i, limit); |
2085 | } | |
2086 | } | |
2087 | ||
343e0d7a | 2088 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
2089 | { |
2090 | char *realobj; | |
2091 | int size, i; | |
2092 | int lines = 0; | |
2093 | ||
3dafccf2 | 2094 | realobj = (char *)objp + obj_offset(cachep); |
8c138bc0 | 2095 | size = cachep->object_size; |
1da177e4 | 2096 | |
b28a02de | 2097 | for (i = 0; i < size; i++) { |
1da177e4 | 2098 | char exp = POISON_FREE; |
b28a02de | 2099 | if (i == size - 1) |
1da177e4 LT |
2100 | exp = POISON_END; |
2101 | if (realobj[i] != exp) { | |
2102 | int limit; | |
2103 | /* Mismatch ! */ | |
2104 | /* Print header */ | |
2105 | if (lines == 0) { | |
b28a02de | 2106 | printk(KERN_ERR |
face37f5 DJ |
2107 | "Slab corruption (%s): %s start=%p, len=%d\n", |
2108 | print_tainted(), cachep->name, realobj, size); | |
1da177e4 LT |
2109 | print_objinfo(cachep, objp, 0); |
2110 | } | |
2111 | /* Hexdump the affected line */ | |
b28a02de | 2112 | i = (i / 16) * 16; |
1da177e4 | 2113 | limit = 16; |
b28a02de PE |
2114 | if (i + limit > size) |
2115 | limit = size - i; | |
1da177e4 LT |
2116 | dump_line(realobj, i, limit); |
2117 | i += 16; | |
2118 | lines++; | |
2119 | /* Limit to 5 lines */ | |
2120 | if (lines > 5) | |
2121 | break; | |
2122 | } | |
2123 | } | |
2124 | if (lines != 0) { | |
2125 | /* Print some data about the neighboring objects, if they | |
2126 | * exist: | |
2127 | */ | |
6ed5eb22 | 2128 | struct slab *slabp = virt_to_slab(objp); |
8fea4e96 | 2129 | unsigned int objnr; |
1da177e4 | 2130 | |
8fea4e96 | 2131 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 | 2132 | if (objnr) { |
8fea4e96 | 2133 | objp = index_to_obj(cachep, slabp, objnr - 1); |
3dafccf2 | 2134 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2135 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", |
b28a02de | 2136 | realobj, size); |
1da177e4 LT |
2137 | print_objinfo(cachep, objp, 2); |
2138 | } | |
b28a02de | 2139 | if (objnr + 1 < cachep->num) { |
8fea4e96 | 2140 | objp = index_to_obj(cachep, slabp, objnr + 1); |
3dafccf2 | 2141 | realobj = (char *)objp + obj_offset(cachep); |
1da177e4 | 2142 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", |
b28a02de | 2143 | realobj, size); |
1da177e4 LT |
2144 | print_objinfo(cachep, objp, 2); |
2145 | } | |
2146 | } | |
2147 | } | |
2148 | #endif | |
2149 | ||
12dd36fa | 2150 | #if DEBUG |
e79aec29 | 2151 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 | 2152 | { |
1da177e4 LT |
2153 | int i; |
2154 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2155 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2156 | |
2157 | if (cachep->flags & SLAB_POISON) { | |
2158 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 2159 | if (cachep->size % PAGE_SIZE == 0 && |
a737b3e2 | 2160 | OFF_SLAB(cachep)) |
b28a02de | 2161 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2162 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
2163 | else |
2164 | check_poison_obj(cachep, objp); | |
2165 | #else | |
2166 | check_poison_obj(cachep, objp); | |
2167 | #endif | |
2168 | } | |
2169 | if (cachep->flags & SLAB_RED_ZONE) { | |
2170 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2171 | slab_error(cachep, "start of a freed object " | |
b28a02de | 2172 | "was overwritten"); |
1da177e4 LT |
2173 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2174 | slab_error(cachep, "end of a freed object " | |
b28a02de | 2175 | "was overwritten"); |
1da177e4 | 2176 | } |
1da177e4 | 2177 | } |
12dd36fa | 2178 | } |
1da177e4 | 2179 | #else |
e79aec29 | 2180 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa | 2181 | { |
12dd36fa | 2182 | } |
1da177e4 LT |
2183 | #endif |
2184 | ||
911851e6 RD |
2185 | /** |
2186 | * slab_destroy - destroy and release all objects in a slab | |
2187 | * @cachep: cache pointer being destroyed | |
2188 | * @slabp: slab pointer being destroyed | |
2189 | * | |
12dd36fa | 2190 | * Destroy all the objs in a slab, and release the mem back to the system. |
a737b3e2 AM |
2191 | * Before calling the slab must have been unlinked from the cache. The |
2192 | * cache-lock is not held/needed. | |
12dd36fa | 2193 | */ |
343e0d7a | 2194 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) |
12dd36fa MD |
2195 | { |
2196 | void *addr = slabp->s_mem - slabp->colouroff; | |
2197 | ||
e79aec29 | 2198 | slab_destroy_debugcheck(cachep, slabp); |
1da177e4 LT |
2199 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { |
2200 | struct slab_rcu *slab_rcu; | |
2201 | ||
b28a02de | 2202 | slab_rcu = (struct slab_rcu *)slabp; |
1da177e4 LT |
2203 | slab_rcu->cachep = cachep; |
2204 | slab_rcu->addr = addr; | |
2205 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2206 | } else { | |
2207 | kmem_freepages(cachep, addr); | |
873623df IM |
2208 | if (OFF_SLAB(cachep)) |
2209 | kmem_cache_free(cachep->slabp_cache, slabp); | |
1da177e4 LT |
2210 | } |
2211 | } | |
2212 | ||
4d268eba | 2213 | /** |
a70773dd RD |
2214 | * calculate_slab_order - calculate size (page order) of slabs |
2215 | * @cachep: pointer to the cache that is being created | |
2216 | * @size: size of objects to be created in this cache. | |
2217 | * @align: required alignment for the objects. | |
2218 | * @flags: slab allocation flags | |
2219 | * | |
2220 | * Also calculates the number of objects per slab. | |
4d268eba PE |
2221 | * |
2222 | * This could be made much more intelligent. For now, try to avoid using | |
2223 | * high order pages for slabs. When the gfp() functions are more friendly | |
2224 | * towards high-order requests, this should be changed. | |
2225 | */ | |
a737b3e2 | 2226 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
ee13d785 | 2227 | size_t size, size_t align, unsigned long flags) |
4d268eba | 2228 | { |
b1ab41c4 | 2229 | unsigned long offslab_limit; |
4d268eba | 2230 | size_t left_over = 0; |
9888e6fa | 2231 | int gfporder; |
4d268eba | 2232 | |
0aa817f0 | 2233 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
4d268eba PE |
2234 | unsigned int num; |
2235 | size_t remainder; | |
2236 | ||
9888e6fa | 2237 | cache_estimate(gfporder, size, align, flags, &remainder, &num); |
4d268eba PE |
2238 | if (!num) |
2239 | continue; | |
9888e6fa | 2240 | |
b1ab41c4 IM |
2241 | if (flags & CFLGS_OFF_SLAB) { |
2242 | /* | |
2243 | * Max number of objs-per-slab for caches which | |
2244 | * use off-slab slabs. Needed to avoid a possible | |
2245 | * looping condition in cache_grow(). | |
2246 | */ | |
2247 | offslab_limit = size - sizeof(struct slab); | |
2248 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2249 | ||
2250 | if (num > offslab_limit) | |
2251 | break; | |
2252 | } | |
4d268eba | 2253 | |
9888e6fa | 2254 | /* Found something acceptable - save it away */ |
4d268eba | 2255 | cachep->num = num; |
9888e6fa | 2256 | cachep->gfporder = gfporder; |
4d268eba PE |
2257 | left_over = remainder; |
2258 | ||
f78bb8ad LT |
2259 | /* |
2260 | * A VFS-reclaimable slab tends to have most allocations | |
2261 | * as GFP_NOFS and we really don't want to have to be allocating | |
2262 | * higher-order pages when we are unable to shrink dcache. | |
2263 | */ | |
2264 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2265 | break; | |
2266 | ||
4d268eba PE |
2267 | /* |
2268 | * Large number of objects is good, but very large slabs are | |
2269 | * currently bad for the gfp()s. | |
2270 | */ | |
543585cc | 2271 | if (gfporder >= slab_max_order) |
4d268eba PE |
2272 | break; |
2273 | ||
9888e6fa LT |
2274 | /* |
2275 | * Acceptable internal fragmentation? | |
2276 | */ | |
a737b3e2 | 2277 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
4d268eba PE |
2278 | break; |
2279 | } | |
2280 | return left_over; | |
2281 | } | |
2282 | ||
83b519e8 | 2283 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
f30cf7d1 | 2284 | { |
97d06609 | 2285 | if (slab_state >= FULL) |
83b519e8 | 2286 | return enable_cpucache(cachep, gfp); |
2ed3a4ef | 2287 | |
97d06609 | 2288 | if (slab_state == DOWN) { |
f30cf7d1 PE |
2289 | /* |
2290 | * Note: the first kmem_cache_create must create the cache | |
2291 | * that's used by kmalloc(24), otherwise the creation of | |
2292 | * further caches will BUG(). | |
2293 | */ | |
2294 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2295 | ||
2296 | /* | |
2297 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2298 | * the first cache, then we need to set up all its list3s, | |
2299 | * otherwise the creation of further caches will BUG(). | |
2300 | */ | |
2301 | set_up_list3s(cachep, SIZE_AC); | |
2302 | if (INDEX_AC == INDEX_L3) | |
97d06609 | 2303 | slab_state = PARTIAL_L3; |
f30cf7d1 | 2304 | else |
97d06609 | 2305 | slab_state = PARTIAL_ARRAYCACHE; |
f30cf7d1 PE |
2306 | } else { |
2307 | cachep->array[smp_processor_id()] = | |
83b519e8 | 2308 | kmalloc(sizeof(struct arraycache_init), gfp); |
f30cf7d1 | 2309 | |
97d06609 | 2310 | if (slab_state == PARTIAL_ARRAYCACHE) { |
f30cf7d1 | 2311 | set_up_list3s(cachep, SIZE_L3); |
97d06609 | 2312 | slab_state = PARTIAL_L3; |
f30cf7d1 PE |
2313 | } else { |
2314 | int node; | |
556a169d | 2315 | for_each_online_node(node) { |
f30cf7d1 PE |
2316 | cachep->nodelists[node] = |
2317 | kmalloc_node(sizeof(struct kmem_list3), | |
eb91f1d0 | 2318 | gfp, node); |
f30cf7d1 PE |
2319 | BUG_ON(!cachep->nodelists[node]); |
2320 | kmem_list3_init(cachep->nodelists[node]); | |
2321 | } | |
2322 | } | |
2323 | } | |
7d6e6d09 | 2324 | cachep->nodelists[numa_mem_id()]->next_reap = |
f30cf7d1 PE |
2325 | jiffies + REAPTIMEOUT_LIST3 + |
2326 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2327 | ||
2328 | cpu_cache_get(cachep)->avail = 0; | |
2329 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2330 | cpu_cache_get(cachep)->batchcount = 1; | |
2331 | cpu_cache_get(cachep)->touched = 0; | |
2332 | cachep->batchcount = 1; | |
2333 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2ed3a4ef | 2334 | return 0; |
f30cf7d1 PE |
2335 | } |
2336 | ||
1da177e4 | 2337 | /** |
039363f3 | 2338 | * __kmem_cache_create - Create a cache. |
1da177e4 LT |
2339 | * @name: A string which is used in /proc/slabinfo to identify this cache. |
2340 | * @size: The size of objects to be created in this cache. | |
2341 | * @align: The required alignment for the objects. | |
2342 | * @flags: SLAB flags | |
2343 | * @ctor: A constructor for the objects. | |
1da177e4 LT |
2344 | * |
2345 | * Returns a ptr to the cache on success, NULL on failure. | |
2346 | * Cannot be called within a int, but can be interrupted. | |
20c2df83 | 2347 | * The @ctor is run when new pages are allocated by the cache. |
1da177e4 | 2348 | * |
1da177e4 LT |
2349 | * The flags are |
2350 | * | |
2351 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2352 | * to catch references to uninitialised memory. | |
2353 | * | |
2354 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2355 | * for buffer overruns. | |
2356 | * | |
1da177e4 LT |
2357 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2358 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2359 | * as davem. | |
2360 | */ | |
278b1bb1 | 2361 | int |
8a13a4cc | 2362 | __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) |
1da177e4 LT |
2363 | { |
2364 | size_t left_over, slab_size, ralign; | |
83b519e8 | 2365 | gfp_t gfp; |
278b1bb1 | 2366 | int err; |
8a13a4cc | 2367 | size_t size = cachep->size; |
1da177e4 | 2368 | |
1da177e4 | 2369 | #if DEBUG |
1da177e4 LT |
2370 | #if FORCED_DEBUG |
2371 | /* | |
2372 | * Enable redzoning and last user accounting, except for caches with | |
2373 | * large objects, if the increased size would increase the object size | |
2374 | * above the next power of two: caches with object sizes just above a | |
2375 | * power of two have a significant amount of internal fragmentation. | |
2376 | */ | |
87a927c7 DW |
2377 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2378 | 2 * sizeof(unsigned long long))) | |
b28a02de | 2379 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
1da177e4 LT |
2380 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2381 | flags |= SLAB_POISON; | |
2382 | #endif | |
2383 | if (flags & SLAB_DESTROY_BY_RCU) | |
2384 | BUG_ON(flags & SLAB_POISON); | |
2385 | #endif | |
1da177e4 | 2386 | /* |
a737b3e2 AM |
2387 | * Always checks flags, a caller might be expecting debug support which |
2388 | * isn't available. | |
1da177e4 | 2389 | */ |
40094fa6 | 2390 | BUG_ON(flags & ~CREATE_MASK); |
1da177e4 | 2391 | |
a737b3e2 AM |
2392 | /* |
2393 | * Check that size is in terms of words. This is needed to avoid | |
1da177e4 LT |
2394 | * unaligned accesses for some archs when redzoning is used, and makes |
2395 | * sure any on-slab bufctl's are also correctly aligned. | |
2396 | */ | |
b28a02de PE |
2397 | if (size & (BYTES_PER_WORD - 1)) { |
2398 | size += (BYTES_PER_WORD - 1); | |
2399 | size &= ~(BYTES_PER_WORD - 1); | |
1da177e4 LT |
2400 | } |
2401 | ||
a737b3e2 AM |
2402 | /* calculate the final buffer alignment: */ |
2403 | ||
1da177e4 LT |
2404 | /* 1) arch recommendation: can be overridden for debug */ |
2405 | if (flags & SLAB_HWCACHE_ALIGN) { | |
a737b3e2 AM |
2406 | /* |
2407 | * Default alignment: as specified by the arch code. Except if | |
2408 | * an object is really small, then squeeze multiple objects into | |
2409 | * one cacheline. | |
1da177e4 LT |
2410 | */ |
2411 | ralign = cache_line_size(); | |
b28a02de | 2412 | while (size <= ralign / 2) |
1da177e4 LT |
2413 | ralign /= 2; |
2414 | } else { | |
2415 | ralign = BYTES_PER_WORD; | |
2416 | } | |
ca5f9703 PE |
2417 | |
2418 | /* | |
87a927c7 DW |
2419 | * Redzoning and user store require word alignment or possibly larger. |
2420 | * Note this will be overridden by architecture or caller mandated | |
2421 | * alignment if either is greater than BYTES_PER_WORD. | |
ca5f9703 | 2422 | */ |
87a927c7 DW |
2423 | if (flags & SLAB_STORE_USER) |
2424 | ralign = BYTES_PER_WORD; | |
2425 | ||
2426 | if (flags & SLAB_RED_ZONE) { | |
2427 | ralign = REDZONE_ALIGN; | |
2428 | /* If redzoning, ensure that the second redzone is suitably | |
2429 | * aligned, by adjusting the object size accordingly. */ | |
2430 | size += REDZONE_ALIGN - 1; | |
2431 | size &= ~(REDZONE_ALIGN - 1); | |
2432 | } | |
ca5f9703 | 2433 | |
a44b56d3 | 2434 | /* 2) arch mandated alignment */ |
1da177e4 LT |
2435 | if (ralign < ARCH_SLAB_MINALIGN) { |
2436 | ralign = ARCH_SLAB_MINALIGN; | |
1da177e4 | 2437 | } |
a44b56d3 | 2438 | /* 3) caller mandated alignment */ |
8a13a4cc CL |
2439 | if (ralign < cachep->align) { |
2440 | ralign = cachep->align; | |
1da177e4 | 2441 | } |
3ff84a7f PE |
2442 | /* disable debug if necessary */ |
2443 | if (ralign > __alignof__(unsigned long long)) | |
a44b56d3 | 2444 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
a737b3e2 | 2445 | /* |
ca5f9703 | 2446 | * 4) Store it. |
1da177e4 | 2447 | */ |
8a13a4cc | 2448 | cachep->align = ralign; |
1da177e4 | 2449 | |
83b519e8 PE |
2450 | if (slab_is_available()) |
2451 | gfp = GFP_KERNEL; | |
2452 | else | |
2453 | gfp = GFP_NOWAIT; | |
2454 | ||
b56efcf0 | 2455 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; |
1da177e4 | 2456 | #if DEBUG |
1da177e4 | 2457 | |
ca5f9703 PE |
2458 | /* |
2459 | * Both debugging options require word-alignment which is calculated | |
2460 | * into align above. | |
2461 | */ | |
1da177e4 | 2462 | if (flags & SLAB_RED_ZONE) { |
1da177e4 | 2463 | /* add space for red zone words */ |
3ff84a7f PE |
2464 | cachep->obj_offset += sizeof(unsigned long long); |
2465 | size += 2 * sizeof(unsigned long long); | |
1da177e4 LT |
2466 | } |
2467 | if (flags & SLAB_STORE_USER) { | |
ca5f9703 | 2468 | /* user store requires one word storage behind the end of |
87a927c7 DW |
2469 | * the real object. But if the second red zone needs to be |
2470 | * aligned to 64 bits, we must allow that much space. | |
1da177e4 | 2471 | */ |
87a927c7 DW |
2472 | if (flags & SLAB_RED_ZONE) |
2473 | size += REDZONE_ALIGN; | |
2474 | else | |
2475 | size += BYTES_PER_WORD; | |
1da177e4 LT |
2476 | } |
2477 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
b28a02de | 2478 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size |
3b0efdfa | 2479 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { |
1ab335d8 | 2480 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); |
1da177e4 LT |
2481 | size = PAGE_SIZE; |
2482 | } | |
2483 | #endif | |
2484 | #endif | |
2485 | ||
e0a42726 IM |
2486 | /* |
2487 | * Determine if the slab management is 'on' or 'off' slab. | |
2488 | * (bootstrapping cannot cope with offslab caches so don't do | |
e7cb55b9 CM |
2489 | * it too early on. Always use on-slab management when |
2490 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
e0a42726 | 2491 | */ |
e7cb55b9 CM |
2492 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && |
2493 | !(flags & SLAB_NOLEAKTRACE)) | |
1da177e4 LT |
2494 | /* |
2495 | * Size is large, assume best to place the slab management obj | |
2496 | * off-slab (should allow better packing of objs). | |
2497 | */ | |
2498 | flags |= CFLGS_OFF_SLAB; | |
2499 | ||
8a13a4cc | 2500 | size = ALIGN(size, cachep->align); |
1da177e4 | 2501 | |
8a13a4cc | 2502 | left_over = calculate_slab_order(cachep, size, cachep->align, flags); |
1da177e4 | 2503 | |
8a13a4cc | 2504 | if (!cachep->num) |
278b1bb1 | 2505 | return -E2BIG; |
8a13a4cc | 2506 | |
b28a02de | 2507 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) |
8a13a4cc | 2508 | + sizeof(struct slab), cachep->align); |
1da177e4 LT |
2509 | |
2510 | /* | |
2511 | * If the slab has been placed off-slab, and we have enough space then | |
2512 | * move it on-slab. This is at the expense of any extra colouring. | |
2513 | */ | |
2514 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2515 | flags &= ~CFLGS_OFF_SLAB; | |
2516 | left_over -= slab_size; | |
2517 | } | |
2518 | ||
2519 | if (flags & CFLGS_OFF_SLAB) { | |
2520 | /* really off slab. No need for manual alignment */ | |
b28a02de PE |
2521 | slab_size = |
2522 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
67461365 RL |
2523 | |
2524 | #ifdef CONFIG_PAGE_POISONING | |
2525 | /* If we're going to use the generic kernel_map_pages() | |
2526 | * poisoning, then it's going to smash the contents of | |
2527 | * the redzone and userword anyhow, so switch them off. | |
2528 | */ | |
2529 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2530 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2531 | #endif | |
1da177e4 LT |
2532 | } |
2533 | ||
2534 | cachep->colour_off = cache_line_size(); | |
2535 | /* Offset must be a multiple of the alignment. */ | |
8a13a4cc CL |
2536 | if (cachep->colour_off < cachep->align) |
2537 | cachep->colour_off = cachep->align; | |
b28a02de | 2538 | cachep->colour = left_over / cachep->colour_off; |
1da177e4 LT |
2539 | cachep->slab_size = slab_size; |
2540 | cachep->flags = flags; | |
a618e89f | 2541 | cachep->allocflags = 0; |
4b51d669 | 2542 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) |
a618e89f | 2543 | cachep->allocflags |= GFP_DMA; |
3b0efdfa | 2544 | cachep->size = size; |
6a2d7a95 | 2545 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
1da177e4 | 2546 | |
e5ac9c5a | 2547 | if (flags & CFLGS_OFF_SLAB) { |
b2d55073 | 2548 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); |
e5ac9c5a RT |
2549 | /* |
2550 | * This is a possibility for one of the malloc_sizes caches. | |
2551 | * But since we go off slab only for object size greater than | |
2552 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2553 | * this should not happen at all. | |
2554 | * But leave a BUG_ON for some lucky dude. | |
2555 | */ | |
6cb8f913 | 2556 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); |
e5ac9c5a | 2557 | } |
1da177e4 | 2558 | |
278b1bb1 CL |
2559 | err = setup_cpu_cache(cachep, gfp); |
2560 | if (err) { | |
12c3667f | 2561 | __kmem_cache_shutdown(cachep); |
278b1bb1 | 2562 | return err; |
2ed3a4ef | 2563 | } |
1da177e4 | 2564 | |
83835b3d PZ |
2565 | if (flags & SLAB_DEBUG_OBJECTS) { |
2566 | /* | |
2567 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2568 | * debug_object_activate()->kmem_cache_alloc(). | |
2569 | */ | |
2570 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2571 | ||
2572 | slab_set_debugobj_lock_classes(cachep); | |
2573 | } | |
2574 | ||
278b1bb1 | 2575 | return 0; |
1da177e4 | 2576 | } |
1da177e4 LT |
2577 | |
2578 | #if DEBUG | |
2579 | static void check_irq_off(void) | |
2580 | { | |
2581 | BUG_ON(!irqs_disabled()); | |
2582 | } | |
2583 | ||
2584 | static void check_irq_on(void) | |
2585 | { | |
2586 | BUG_ON(irqs_disabled()); | |
2587 | } | |
2588 | ||
343e0d7a | 2589 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
1da177e4 LT |
2590 | { |
2591 | #ifdef CONFIG_SMP | |
2592 | check_irq_off(); | |
7d6e6d09 | 2593 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); |
1da177e4 LT |
2594 | #endif |
2595 | } | |
e498be7d | 2596 | |
343e0d7a | 2597 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
e498be7d CL |
2598 | { |
2599 | #ifdef CONFIG_SMP | |
2600 | check_irq_off(); | |
2601 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2602 | #endif | |
2603 | } | |
2604 | ||
1da177e4 LT |
2605 | #else |
2606 | #define check_irq_off() do { } while(0) | |
2607 | #define check_irq_on() do { } while(0) | |
2608 | #define check_spinlock_acquired(x) do { } while(0) | |
e498be7d | 2609 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
1da177e4 LT |
2610 | #endif |
2611 | ||
aab2207c CL |
2612 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
2613 | struct array_cache *ac, | |
2614 | int force, int node); | |
2615 | ||
1da177e4 LT |
2616 | static void do_drain(void *arg) |
2617 | { | |
a737b3e2 | 2618 | struct kmem_cache *cachep = arg; |
1da177e4 | 2619 | struct array_cache *ac; |
7d6e6d09 | 2620 | int node = numa_mem_id(); |
1da177e4 LT |
2621 | |
2622 | check_irq_off(); | |
9a2dba4b | 2623 | ac = cpu_cache_get(cachep); |
ff69416e CL |
2624 | spin_lock(&cachep->nodelists[node]->list_lock); |
2625 | free_block(cachep, ac->entry, ac->avail, node); | |
2626 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
1da177e4 LT |
2627 | ac->avail = 0; |
2628 | } | |
2629 | ||
343e0d7a | 2630 | static void drain_cpu_caches(struct kmem_cache *cachep) |
1da177e4 | 2631 | { |
e498be7d CL |
2632 | struct kmem_list3 *l3; |
2633 | int node; | |
2634 | ||
15c8b6c1 | 2635 | on_each_cpu(do_drain, cachep, 1); |
1da177e4 | 2636 | check_irq_on(); |
b28a02de | 2637 | for_each_online_node(node) { |
e498be7d | 2638 | l3 = cachep->nodelists[node]; |
a4523a8b RD |
2639 | if (l3 && l3->alien) |
2640 | drain_alien_cache(cachep, l3->alien); | |
2641 | } | |
2642 | ||
2643 | for_each_online_node(node) { | |
2644 | l3 = cachep->nodelists[node]; | |
2645 | if (l3) | |
aab2207c | 2646 | drain_array(cachep, l3, l3->shared, 1, node); |
e498be7d | 2647 | } |
1da177e4 LT |
2648 | } |
2649 | ||
ed11d9eb CL |
2650 | /* |
2651 | * Remove slabs from the list of free slabs. | |
2652 | * Specify the number of slabs to drain in tofree. | |
2653 | * | |
2654 | * Returns the actual number of slabs released. | |
2655 | */ | |
2656 | static int drain_freelist(struct kmem_cache *cache, | |
2657 | struct kmem_list3 *l3, int tofree) | |
1da177e4 | 2658 | { |
ed11d9eb CL |
2659 | struct list_head *p; |
2660 | int nr_freed; | |
1da177e4 | 2661 | struct slab *slabp; |
1da177e4 | 2662 | |
ed11d9eb CL |
2663 | nr_freed = 0; |
2664 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
1da177e4 | 2665 | |
ed11d9eb | 2666 | spin_lock_irq(&l3->list_lock); |
e498be7d | 2667 | p = l3->slabs_free.prev; |
ed11d9eb CL |
2668 | if (p == &l3->slabs_free) { |
2669 | spin_unlock_irq(&l3->list_lock); | |
2670 | goto out; | |
2671 | } | |
1da177e4 | 2672 | |
ed11d9eb | 2673 | slabp = list_entry(p, struct slab, list); |
1da177e4 | 2674 | #if DEBUG |
40094fa6 | 2675 | BUG_ON(slabp->inuse); |
1da177e4 LT |
2676 | #endif |
2677 | list_del(&slabp->list); | |
ed11d9eb CL |
2678 | /* |
2679 | * Safe to drop the lock. The slab is no longer linked | |
2680 | * to the cache. | |
2681 | */ | |
2682 | l3->free_objects -= cache->num; | |
e498be7d | 2683 | spin_unlock_irq(&l3->list_lock); |
ed11d9eb CL |
2684 | slab_destroy(cache, slabp); |
2685 | nr_freed++; | |
1da177e4 | 2686 | } |
ed11d9eb CL |
2687 | out: |
2688 | return nr_freed; | |
1da177e4 LT |
2689 | } |
2690 | ||
18004c5d | 2691 | /* Called with slab_mutex held to protect against cpu hotplug */ |
343e0d7a | 2692 | static int __cache_shrink(struct kmem_cache *cachep) |
e498be7d CL |
2693 | { |
2694 | int ret = 0, i = 0; | |
2695 | struct kmem_list3 *l3; | |
2696 | ||
2697 | drain_cpu_caches(cachep); | |
2698 | ||
2699 | check_irq_on(); | |
2700 | for_each_online_node(i) { | |
2701 | l3 = cachep->nodelists[i]; | |
ed11d9eb CL |
2702 | if (!l3) |
2703 | continue; | |
2704 | ||
2705 | drain_freelist(cachep, l3, l3->free_objects); | |
2706 | ||
2707 | ret += !list_empty(&l3->slabs_full) || | |
2708 | !list_empty(&l3->slabs_partial); | |
e498be7d CL |
2709 | } |
2710 | return (ret ? 1 : 0); | |
2711 | } | |
2712 | ||
1da177e4 LT |
2713 | /** |
2714 | * kmem_cache_shrink - Shrink a cache. | |
2715 | * @cachep: The cache to shrink. | |
2716 | * | |
2717 | * Releases as many slabs as possible for a cache. | |
2718 | * To help debugging, a zero exit status indicates all slabs were released. | |
2719 | */ | |
343e0d7a | 2720 | int kmem_cache_shrink(struct kmem_cache *cachep) |
1da177e4 | 2721 | { |
8f5be20b | 2722 | int ret; |
40094fa6 | 2723 | BUG_ON(!cachep || in_interrupt()); |
1da177e4 | 2724 | |
95402b38 | 2725 | get_online_cpus(); |
18004c5d | 2726 | mutex_lock(&slab_mutex); |
8f5be20b | 2727 | ret = __cache_shrink(cachep); |
18004c5d | 2728 | mutex_unlock(&slab_mutex); |
95402b38 | 2729 | put_online_cpus(); |
8f5be20b | 2730 | return ret; |
1da177e4 LT |
2731 | } |
2732 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2733 | ||
945cf2b6 | 2734 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
1da177e4 | 2735 | { |
12c3667f CL |
2736 | int i; |
2737 | struct kmem_list3 *l3; | |
2738 | int rc = __cache_shrink(cachep); | |
2739 | ||
2740 | if (rc) | |
2741 | return rc; | |
2742 | ||
2743 | for_each_online_cpu(i) | |
2744 | kfree(cachep->array[i]); | |
2745 | ||
2746 | /* NUMA: free the list3 structures */ | |
2747 | for_each_online_node(i) { | |
2748 | l3 = cachep->nodelists[i]; | |
2749 | if (l3) { | |
2750 | kfree(l3->shared); | |
2751 | free_alien_cache(l3->alien); | |
2752 | kfree(l3); | |
2753 | } | |
2754 | } | |
2755 | return 0; | |
1da177e4 | 2756 | } |
1da177e4 | 2757 | |
e5ac9c5a RT |
2758 | /* |
2759 | * Get the memory for a slab management obj. | |
2760 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2761 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2762 | * come from the same cache which is getting created because, | |
2763 | * when we are searching for an appropriate cache for these | |
2764 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2765 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2766 | * kmem_find_general_cachep till the initialization is complete. | |
2767 | * Hence we cannot have slabp_cache same as the original cache. | |
2768 | */ | |
343e0d7a | 2769 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, |
5b74ada7 RT |
2770 | int colour_off, gfp_t local_flags, |
2771 | int nodeid) | |
1da177e4 LT |
2772 | { |
2773 | struct slab *slabp; | |
b28a02de | 2774 | |
1da177e4 LT |
2775 | if (OFF_SLAB(cachep)) { |
2776 | /* Slab management obj is off-slab. */ | |
5b74ada7 | 2777 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, |
8759ec50 | 2778 | local_flags, nodeid); |
d5cff635 CM |
2779 | /* |
2780 | * If the first object in the slab is leaked (it's allocated | |
2781 | * but no one has a reference to it), we want to make sure | |
2782 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2783 | * to the object. Otherwise we will not report the leak. | |
2784 | */ | |
c017b4be CM |
2785 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), |
2786 | local_flags); | |
1da177e4 LT |
2787 | if (!slabp) |
2788 | return NULL; | |
2789 | } else { | |
b28a02de | 2790 | slabp = objp + colour_off; |
1da177e4 LT |
2791 | colour_off += cachep->slab_size; |
2792 | } | |
2793 | slabp->inuse = 0; | |
2794 | slabp->colouroff = colour_off; | |
b28a02de | 2795 | slabp->s_mem = objp + colour_off; |
5b74ada7 | 2796 | slabp->nodeid = nodeid; |
e51bfd0a | 2797 | slabp->free = 0; |
1da177e4 LT |
2798 | return slabp; |
2799 | } | |
2800 | ||
2801 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2802 | { | |
b28a02de | 2803 | return (kmem_bufctl_t *) (slabp + 1); |
1da177e4 LT |
2804 | } |
2805 | ||
343e0d7a | 2806 | static void cache_init_objs(struct kmem_cache *cachep, |
a35afb83 | 2807 | struct slab *slabp) |
1da177e4 LT |
2808 | { |
2809 | int i; | |
2810 | ||
2811 | for (i = 0; i < cachep->num; i++) { | |
8fea4e96 | 2812 | void *objp = index_to_obj(cachep, slabp, i); |
1da177e4 LT |
2813 | #if DEBUG |
2814 | /* need to poison the objs? */ | |
2815 | if (cachep->flags & SLAB_POISON) | |
2816 | poison_obj(cachep, objp, POISON_FREE); | |
2817 | if (cachep->flags & SLAB_STORE_USER) | |
2818 | *dbg_userword(cachep, objp) = NULL; | |
2819 | ||
2820 | if (cachep->flags & SLAB_RED_ZONE) { | |
2821 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2822 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2823 | } | |
2824 | /* | |
a737b3e2 AM |
2825 | * Constructors are not allowed to allocate memory from the same |
2826 | * cache which they are a constructor for. Otherwise, deadlock. | |
2827 | * They must also be threaded. | |
1da177e4 LT |
2828 | */ |
2829 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
51cc5068 | 2830 | cachep->ctor(objp + obj_offset(cachep)); |
1da177e4 LT |
2831 | |
2832 | if (cachep->flags & SLAB_RED_ZONE) { | |
2833 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2834 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2835 | " end of an object"); |
1da177e4 LT |
2836 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2837 | slab_error(cachep, "constructor overwrote the" | |
b28a02de | 2838 | " start of an object"); |
1da177e4 | 2839 | } |
3b0efdfa | 2840 | if ((cachep->size % PAGE_SIZE) == 0 && |
a737b3e2 | 2841 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) |
b28a02de | 2842 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 2843 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
2844 | #else |
2845 | if (cachep->ctor) | |
51cc5068 | 2846 | cachep->ctor(objp); |
1da177e4 | 2847 | #endif |
b28a02de | 2848 | slab_bufctl(slabp)[i] = i + 1; |
1da177e4 | 2849 | } |
b28a02de | 2850 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; |
1da177e4 LT |
2851 | } |
2852 | ||
343e0d7a | 2853 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 2854 | { |
4b51d669 CL |
2855 | if (CONFIG_ZONE_DMA_FLAG) { |
2856 | if (flags & GFP_DMA) | |
a618e89f | 2857 | BUG_ON(!(cachep->allocflags & GFP_DMA)); |
4b51d669 | 2858 | else |
a618e89f | 2859 | BUG_ON(cachep->allocflags & GFP_DMA); |
4b51d669 | 2860 | } |
1da177e4 LT |
2861 | } |
2862 | ||
a737b3e2 AM |
2863 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, |
2864 | int nodeid) | |
78d382d7 | 2865 | { |
8fea4e96 | 2866 | void *objp = index_to_obj(cachep, slabp, slabp->free); |
78d382d7 MD |
2867 | kmem_bufctl_t next; |
2868 | ||
2869 | slabp->inuse++; | |
2870 | next = slab_bufctl(slabp)[slabp->free]; | |
2871 | #if DEBUG | |
2872 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2873 | WARN_ON(slabp->nodeid != nodeid); | |
2874 | #endif | |
2875 | slabp->free = next; | |
2876 | ||
2877 | return objp; | |
2878 | } | |
2879 | ||
a737b3e2 AM |
2880 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, |
2881 | void *objp, int nodeid) | |
78d382d7 | 2882 | { |
8fea4e96 | 2883 | unsigned int objnr = obj_to_index(cachep, slabp, objp); |
78d382d7 MD |
2884 | |
2885 | #if DEBUG | |
2886 | /* Verify that the slab belongs to the intended node */ | |
2887 | WARN_ON(slabp->nodeid != nodeid); | |
2888 | ||
871751e2 | 2889 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { |
78d382d7 | 2890 | printk(KERN_ERR "slab: double free detected in cache " |
a737b3e2 | 2891 | "'%s', objp %p\n", cachep->name, objp); |
78d382d7 MD |
2892 | BUG(); |
2893 | } | |
2894 | #endif | |
2895 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2896 | slabp->free = objnr; | |
2897 | slabp->inuse--; | |
2898 | } | |
2899 | ||
4776874f PE |
2900 | /* |
2901 | * Map pages beginning at addr to the given cache and slab. This is required | |
2902 | * for the slab allocator to be able to lookup the cache and slab of a | |
ccd35fb9 | 2903 | * virtual address for kfree, ksize, and slab debugging. |
4776874f PE |
2904 | */ |
2905 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2906 | void *addr) | |
1da177e4 | 2907 | { |
4776874f | 2908 | int nr_pages; |
1da177e4 LT |
2909 | struct page *page; |
2910 | ||
4776874f | 2911 | page = virt_to_page(addr); |
84097518 | 2912 | |
4776874f | 2913 | nr_pages = 1; |
84097518 | 2914 | if (likely(!PageCompound(page))) |
4776874f PE |
2915 | nr_pages <<= cache->gfporder; |
2916 | ||
1da177e4 | 2917 | do { |
35026088 CL |
2918 | page->slab_cache = cache; |
2919 | page->slab_page = slab; | |
1da177e4 | 2920 | page++; |
4776874f | 2921 | } while (--nr_pages); |
1da177e4 LT |
2922 | } |
2923 | ||
2924 | /* | |
2925 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2926 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2927 | */ | |
3c517a61 CL |
2928 | static int cache_grow(struct kmem_cache *cachep, |
2929 | gfp_t flags, int nodeid, void *objp) | |
1da177e4 | 2930 | { |
b28a02de | 2931 | struct slab *slabp; |
b28a02de PE |
2932 | size_t offset; |
2933 | gfp_t local_flags; | |
e498be7d | 2934 | struct kmem_list3 *l3; |
1da177e4 | 2935 | |
a737b3e2 AM |
2936 | /* |
2937 | * Be lazy and only check for valid flags here, keeping it out of the | |
2938 | * critical path in kmem_cache_alloc(). | |
1da177e4 | 2939 | */ |
6cb06229 CL |
2940 | BUG_ON(flags & GFP_SLAB_BUG_MASK); |
2941 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
1da177e4 | 2942 | |
2e1217cf | 2943 | /* Take the l3 list lock to change the colour_next on this node */ |
1da177e4 | 2944 | check_irq_off(); |
2e1217cf RT |
2945 | l3 = cachep->nodelists[nodeid]; |
2946 | spin_lock(&l3->list_lock); | |
1da177e4 LT |
2947 | |
2948 | /* Get colour for the slab, and cal the next value. */ | |
2e1217cf RT |
2949 | offset = l3->colour_next; |
2950 | l3->colour_next++; | |
2951 | if (l3->colour_next >= cachep->colour) | |
2952 | l3->colour_next = 0; | |
2953 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2954 | |
2e1217cf | 2955 | offset *= cachep->colour_off; |
1da177e4 LT |
2956 | |
2957 | if (local_flags & __GFP_WAIT) | |
2958 | local_irq_enable(); | |
2959 | ||
2960 | /* | |
2961 | * The test for missing atomic flag is performed here, rather than | |
2962 | * the more obvious place, simply to reduce the critical path length | |
2963 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
2964 | * will eventually be caught here (where it matters). | |
2965 | */ | |
2966 | kmem_flagcheck(cachep, flags); | |
2967 | ||
a737b3e2 AM |
2968 | /* |
2969 | * Get mem for the objs. Attempt to allocate a physical page from | |
2970 | * 'nodeid'. | |
e498be7d | 2971 | */ |
3c517a61 | 2972 | if (!objp) |
b8c1c5da | 2973 | objp = kmem_getpages(cachep, local_flags, nodeid); |
a737b3e2 | 2974 | if (!objp) |
1da177e4 LT |
2975 | goto failed; |
2976 | ||
2977 | /* Get slab management. */ | |
3c517a61 | 2978 | slabp = alloc_slabmgmt(cachep, objp, offset, |
6cb06229 | 2979 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); |
a737b3e2 | 2980 | if (!slabp) |
1da177e4 LT |
2981 | goto opps1; |
2982 | ||
4776874f | 2983 | slab_map_pages(cachep, slabp, objp); |
1da177e4 | 2984 | |
a35afb83 | 2985 | cache_init_objs(cachep, slabp); |
1da177e4 LT |
2986 | |
2987 | if (local_flags & __GFP_WAIT) | |
2988 | local_irq_disable(); | |
2989 | check_irq_off(); | |
e498be7d | 2990 | spin_lock(&l3->list_lock); |
1da177e4 LT |
2991 | |
2992 | /* Make slab active. */ | |
e498be7d | 2993 | list_add_tail(&slabp->list, &(l3->slabs_free)); |
1da177e4 | 2994 | STATS_INC_GROWN(cachep); |
e498be7d CL |
2995 | l3->free_objects += cachep->num; |
2996 | spin_unlock(&l3->list_lock); | |
1da177e4 | 2997 | return 1; |
a737b3e2 | 2998 | opps1: |
1da177e4 | 2999 | kmem_freepages(cachep, objp); |
a737b3e2 | 3000 | failed: |
1da177e4 LT |
3001 | if (local_flags & __GFP_WAIT) |
3002 | local_irq_disable(); | |
3003 | return 0; | |
3004 | } | |
3005 | ||
3006 | #if DEBUG | |
3007 | ||
3008 | /* | |
3009 | * Perform extra freeing checks: | |
3010 | * - detect bad pointers. | |
3011 | * - POISON/RED_ZONE checking | |
1da177e4 LT |
3012 | */ |
3013 | static void kfree_debugcheck(const void *objp) | |
3014 | { | |
1da177e4 LT |
3015 | if (!virt_addr_valid(objp)) { |
3016 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
b28a02de PE |
3017 | (unsigned long)objp); |
3018 | BUG(); | |
1da177e4 | 3019 | } |
1da177e4 LT |
3020 | } |
3021 | ||
58ce1fd5 PE |
3022 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
3023 | { | |
b46b8f19 | 3024 | unsigned long long redzone1, redzone2; |
58ce1fd5 PE |
3025 | |
3026 | redzone1 = *dbg_redzone1(cache, obj); | |
3027 | redzone2 = *dbg_redzone2(cache, obj); | |
3028 | ||
3029 | /* | |
3030 | * Redzone is ok. | |
3031 | */ | |
3032 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3033 | return; | |
3034 | ||
3035 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3036 | slab_error(cache, "double free detected"); | |
3037 | else | |
3038 | slab_error(cache, "memory outside object was overwritten"); | |
3039 | ||
b46b8f19 | 3040 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", |
58ce1fd5 PE |
3041 | obj, redzone1, redzone2); |
3042 | } | |
3043 | ||
343e0d7a | 3044 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
b28a02de | 3045 | void *caller) |
1da177e4 LT |
3046 | { |
3047 | struct page *page; | |
3048 | unsigned int objnr; | |
3049 | struct slab *slabp; | |
3050 | ||
80cbd911 MW |
3051 | BUG_ON(virt_to_cache(objp) != cachep); |
3052 | ||
3dafccf2 | 3053 | objp -= obj_offset(cachep); |
1da177e4 | 3054 | kfree_debugcheck(objp); |
b49af68f | 3055 | page = virt_to_head_page(objp); |
1da177e4 | 3056 | |
35026088 | 3057 | slabp = page->slab_page; |
1da177e4 LT |
3058 | |
3059 | if (cachep->flags & SLAB_RED_ZONE) { | |
58ce1fd5 | 3060 | verify_redzone_free(cachep, objp); |
1da177e4 LT |
3061 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
3062 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3063 | } | |
3064 | if (cachep->flags & SLAB_STORE_USER) | |
3065 | *dbg_userword(cachep, objp) = caller; | |
3066 | ||
8fea4e96 | 3067 | objnr = obj_to_index(cachep, slabp, objp); |
1da177e4 LT |
3068 | |
3069 | BUG_ON(objnr >= cachep->num); | |
8fea4e96 | 3070 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); |
1da177e4 | 3071 | |
871751e2 AV |
3072 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3073 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3074 | #endif | |
1da177e4 LT |
3075 | if (cachep->flags & SLAB_POISON) { |
3076 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3b0efdfa | 3077 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { |
1da177e4 | 3078 | store_stackinfo(cachep, objp, (unsigned long)caller); |
b28a02de | 3079 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3080 | cachep->size / PAGE_SIZE, 0); |
1da177e4 LT |
3081 | } else { |
3082 | poison_obj(cachep, objp, POISON_FREE); | |
3083 | } | |
3084 | #else | |
3085 | poison_obj(cachep, objp, POISON_FREE); | |
3086 | #endif | |
3087 | } | |
3088 | return objp; | |
3089 | } | |
3090 | ||
343e0d7a | 3091 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) |
1da177e4 LT |
3092 | { |
3093 | kmem_bufctl_t i; | |
3094 | int entries = 0; | |
b28a02de | 3095 | |
1da177e4 LT |
3096 | /* Check slab's freelist to see if this obj is there. */ |
3097 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3098 | entries++; | |
3099 | if (entries > cachep->num || i >= cachep->num) | |
3100 | goto bad; | |
3101 | } | |
3102 | if (entries != cachep->num - slabp->inuse) { | |
a737b3e2 AM |
3103 | bad: |
3104 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
face37f5 DJ |
3105 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", |
3106 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3107 | print_tainted()); | |
fdde6abb SAS |
3108 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, |
3109 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3110 | 1); | |
1da177e4 LT |
3111 | BUG(); |
3112 | } | |
3113 | } | |
3114 | #else | |
3115 | #define kfree_debugcheck(x) do { } while(0) | |
3116 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3117 | #define check_slabp(x,y) do { } while(0) | |
3118 | #endif | |
3119 | ||
072bb0aa MG |
3120 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, |
3121 | bool force_refill) | |
1da177e4 LT |
3122 | { |
3123 | int batchcount; | |
3124 | struct kmem_list3 *l3; | |
3125 | struct array_cache *ac; | |
1ca4cb24 PE |
3126 | int node; |
3127 | ||
1da177e4 | 3128 | check_irq_off(); |
7d6e6d09 | 3129 | node = numa_mem_id(); |
072bb0aa MG |
3130 | if (unlikely(force_refill)) |
3131 | goto force_grow; | |
3132 | retry: | |
9a2dba4b | 3133 | ac = cpu_cache_get(cachep); |
1da177e4 LT |
3134 | batchcount = ac->batchcount; |
3135 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
a737b3e2 AM |
3136 | /* |
3137 | * If there was little recent activity on this cache, then | |
3138 | * perform only a partial refill. Otherwise we could generate | |
3139 | * refill bouncing. | |
1da177e4 LT |
3140 | */ |
3141 | batchcount = BATCHREFILL_LIMIT; | |
3142 | } | |
1ca4cb24 | 3143 | l3 = cachep->nodelists[node]; |
e498be7d CL |
3144 | |
3145 | BUG_ON(ac->avail > 0 || !l3); | |
3146 | spin_lock(&l3->list_lock); | |
1da177e4 | 3147 | |
3ded175a | 3148 | /* See if we can refill from the shared array */ |
44b57f1c NP |
3149 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { |
3150 | l3->shared->touched = 1; | |
3ded175a | 3151 | goto alloc_done; |
44b57f1c | 3152 | } |
3ded175a | 3153 | |
1da177e4 LT |
3154 | while (batchcount > 0) { |
3155 | struct list_head *entry; | |
3156 | struct slab *slabp; | |
3157 | /* Get slab alloc is to come from. */ | |
3158 | entry = l3->slabs_partial.next; | |
3159 | if (entry == &l3->slabs_partial) { | |
3160 | l3->free_touched = 1; | |
3161 | entry = l3->slabs_free.next; | |
3162 | if (entry == &l3->slabs_free) | |
3163 | goto must_grow; | |
3164 | } | |
3165 | ||
3166 | slabp = list_entry(entry, struct slab, list); | |
3167 | check_slabp(cachep, slabp); | |
3168 | check_spinlock_acquired(cachep); | |
714b8171 PE |
3169 | |
3170 | /* | |
3171 | * The slab was either on partial or free list so | |
3172 | * there must be at least one object available for | |
3173 | * allocation. | |
3174 | */ | |
249b9f33 | 3175 | BUG_ON(slabp->inuse >= cachep->num); |
714b8171 | 3176 | |
1da177e4 | 3177 | while (slabp->inuse < cachep->num && batchcount--) { |
1da177e4 LT |
3178 | STATS_INC_ALLOCED(cachep); |
3179 | STATS_INC_ACTIVE(cachep); | |
3180 | STATS_SET_HIGH(cachep); | |
3181 | ||
072bb0aa MG |
3182 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, |
3183 | node)); | |
1da177e4 LT |
3184 | } |
3185 | check_slabp(cachep, slabp); | |
3186 | ||
3187 | /* move slabp to correct slabp list: */ | |
3188 | list_del(&slabp->list); | |
3189 | if (slabp->free == BUFCTL_END) | |
3190 | list_add(&slabp->list, &l3->slabs_full); | |
3191 | else | |
3192 | list_add(&slabp->list, &l3->slabs_partial); | |
3193 | } | |
3194 | ||
a737b3e2 | 3195 | must_grow: |
1da177e4 | 3196 | l3->free_objects -= ac->avail; |
a737b3e2 | 3197 | alloc_done: |
e498be7d | 3198 | spin_unlock(&l3->list_lock); |
1da177e4 LT |
3199 | |
3200 | if (unlikely(!ac->avail)) { | |
3201 | int x; | |
072bb0aa | 3202 | force_grow: |
3c517a61 | 3203 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); |
e498be7d | 3204 | |
a737b3e2 | 3205 | /* cache_grow can reenable interrupts, then ac could change. */ |
9a2dba4b | 3206 | ac = cpu_cache_get(cachep); |
072bb0aa MG |
3207 | |
3208 | /* no objects in sight? abort */ | |
3209 | if (!x && (ac->avail == 0 || force_refill)) | |
1da177e4 LT |
3210 | return NULL; |
3211 | ||
a737b3e2 | 3212 | if (!ac->avail) /* objects refilled by interrupt? */ |
1da177e4 LT |
3213 | goto retry; |
3214 | } | |
3215 | ac->touched = 1; | |
072bb0aa MG |
3216 | |
3217 | return ac_get_obj(cachep, ac, flags, force_refill); | |
1da177e4 LT |
3218 | } |
3219 | ||
a737b3e2 AM |
3220 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3221 | gfp_t flags) | |
1da177e4 LT |
3222 | { |
3223 | might_sleep_if(flags & __GFP_WAIT); | |
3224 | #if DEBUG | |
3225 | kmem_flagcheck(cachep, flags); | |
3226 | #endif | |
3227 | } | |
3228 | ||
3229 | #if DEBUG | |
a737b3e2 AM |
3230 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3231 | gfp_t flags, void *objp, void *caller) | |
1da177e4 | 3232 | { |
b28a02de | 3233 | if (!objp) |
1da177e4 | 3234 | return objp; |
b28a02de | 3235 | if (cachep->flags & SLAB_POISON) { |
1da177e4 | 3236 | #ifdef CONFIG_DEBUG_PAGEALLOC |
3b0efdfa | 3237 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) |
b28a02de | 3238 | kernel_map_pages(virt_to_page(objp), |
3b0efdfa | 3239 | cachep->size / PAGE_SIZE, 1); |
1da177e4 LT |
3240 | else |
3241 | check_poison_obj(cachep, objp); | |
3242 | #else | |
3243 | check_poison_obj(cachep, objp); | |
3244 | #endif | |
3245 | poison_obj(cachep, objp, POISON_INUSE); | |
3246 | } | |
3247 | if (cachep->flags & SLAB_STORE_USER) | |
3248 | *dbg_userword(cachep, objp) = caller; | |
3249 | ||
3250 | if (cachep->flags & SLAB_RED_ZONE) { | |
a737b3e2 AM |
3251 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3252 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3253 | slab_error(cachep, "double free, or memory outside" | |
3254 | " object was overwritten"); | |
b28a02de | 3255 | printk(KERN_ERR |
b46b8f19 | 3256 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", |
a737b3e2 AM |
3257 | objp, *dbg_redzone1(cachep, objp), |
3258 | *dbg_redzone2(cachep, objp)); | |
1da177e4 LT |
3259 | } |
3260 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3261 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3262 | } | |
871751e2 AV |
3263 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
3264 | { | |
3265 | struct slab *slabp; | |
3266 | unsigned objnr; | |
3267 | ||
35026088 | 3268 | slabp = virt_to_head_page(objp)->slab_page; |
3b0efdfa | 3269 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; |
871751e2 AV |
3270 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; |
3271 | } | |
3272 | #endif | |
3dafccf2 | 3273 | objp += obj_offset(cachep); |
4f104934 | 3274 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
51cc5068 | 3275 | cachep->ctor(objp); |
7ea466f2 TH |
3276 | if (ARCH_SLAB_MINALIGN && |
3277 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
a44b56d3 | 3278 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
c225150b | 3279 | objp, (int)ARCH_SLAB_MINALIGN); |
a44b56d3 | 3280 | } |
1da177e4 LT |
3281 | return objp; |
3282 | } | |
3283 | #else | |
3284 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3285 | #endif | |
3286 | ||
773ff60e | 3287 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) |
8a8b6502 | 3288 | { |
9b030cb8 | 3289 | if (cachep == kmem_cache) |
773ff60e | 3290 | return false; |
8a8b6502 | 3291 | |
8c138bc0 | 3292 | return should_failslab(cachep->object_size, flags, cachep->flags); |
8a8b6502 AM |
3293 | } |
3294 | ||
343e0d7a | 3295 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3296 | { |
b28a02de | 3297 | void *objp; |
1da177e4 | 3298 | struct array_cache *ac; |
072bb0aa | 3299 | bool force_refill = false; |
1da177e4 | 3300 | |
5c382300 | 3301 | check_irq_off(); |
8a8b6502 | 3302 | |
9a2dba4b | 3303 | ac = cpu_cache_get(cachep); |
1da177e4 | 3304 | if (likely(ac->avail)) { |
1da177e4 | 3305 | ac->touched = 1; |
072bb0aa MG |
3306 | objp = ac_get_obj(cachep, ac, flags, false); |
3307 | ||
ddbf2e83 | 3308 | /* |
072bb0aa MG |
3309 | * Allow for the possibility all avail objects are not allowed |
3310 | * by the current flags | |
ddbf2e83 | 3311 | */ |
072bb0aa MG |
3312 | if (objp) { |
3313 | STATS_INC_ALLOCHIT(cachep); | |
3314 | goto out; | |
3315 | } | |
3316 | force_refill = true; | |
1da177e4 | 3317 | } |
072bb0aa MG |
3318 | |
3319 | STATS_INC_ALLOCMISS(cachep); | |
3320 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3321 | /* | |
3322 | * the 'ac' may be updated by cache_alloc_refill(), | |
3323 | * and kmemleak_erase() requires its correct value. | |
3324 | */ | |
3325 | ac = cpu_cache_get(cachep); | |
3326 | ||
3327 | out: | |
d5cff635 CM |
3328 | /* |
3329 | * To avoid a false negative, if an object that is in one of the | |
3330 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3331 | * treat the array pointers as a reference to the object. | |
3332 | */ | |
f3d8b53a O |
3333 | if (objp) |
3334 | kmemleak_erase(&ac->entry[ac->avail]); | |
5c382300 AK |
3335 | return objp; |
3336 | } | |
3337 | ||
e498be7d | 3338 | #ifdef CONFIG_NUMA |
c61afb18 | 3339 | /* |
b2455396 | 3340 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. |
c61afb18 PJ |
3341 | * |
3342 | * If we are in_interrupt, then process context, including cpusets and | |
3343 | * mempolicy, may not apply and should not be used for allocation policy. | |
3344 | */ | |
3345 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3346 | { | |
3347 | int nid_alloc, nid_here; | |
3348 | ||
765c4507 | 3349 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
c61afb18 | 3350 | return NULL; |
7d6e6d09 | 3351 | nid_alloc = nid_here = numa_mem_id(); |
c61afb18 | 3352 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
6adef3eb | 3353 | nid_alloc = cpuset_slab_spread_node(); |
c61afb18 | 3354 | else if (current->mempolicy) |
e7b691b0 | 3355 | nid_alloc = slab_node(); |
c61afb18 | 3356 | if (nid_alloc != nid_here) |
8b98c169 | 3357 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
c61afb18 PJ |
3358 | return NULL; |
3359 | } | |
3360 | ||
765c4507 CL |
3361 | /* |
3362 | * Fallback function if there was no memory available and no objects on a | |
3c517a61 CL |
3363 | * certain node and fall back is permitted. First we scan all the |
3364 | * available nodelists for available objects. If that fails then we | |
3365 | * perform an allocation without specifying a node. This allows the page | |
3366 | * allocator to do its reclaim / fallback magic. We then insert the | |
3367 | * slab into the proper nodelist and then allocate from it. | |
765c4507 | 3368 | */ |
8c8cc2c1 | 3369 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
765c4507 | 3370 | { |
8c8cc2c1 PE |
3371 | struct zonelist *zonelist; |
3372 | gfp_t local_flags; | |
dd1a239f | 3373 | struct zoneref *z; |
54a6eb5c MG |
3374 | struct zone *zone; |
3375 | enum zone_type high_zoneidx = gfp_zone(flags); | |
765c4507 | 3376 | void *obj = NULL; |
3c517a61 | 3377 | int nid; |
cc9a6c87 | 3378 | unsigned int cpuset_mems_cookie; |
8c8cc2c1 PE |
3379 | |
3380 | if (flags & __GFP_THISNODE) | |
3381 | return NULL; | |
3382 | ||
6cb06229 | 3383 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
765c4507 | 3384 | |
cc9a6c87 MG |
3385 | retry_cpuset: |
3386 | cpuset_mems_cookie = get_mems_allowed(); | |
e7b691b0 | 3387 | zonelist = node_zonelist(slab_node(), flags); |
cc9a6c87 | 3388 | |
3c517a61 CL |
3389 | retry: |
3390 | /* | |
3391 | * Look through allowed nodes for objects available | |
3392 | * from existing per node queues. | |
3393 | */ | |
54a6eb5c MG |
3394 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3395 | nid = zone_to_nid(zone); | |
aedb0eb1 | 3396 | |
54a6eb5c | 3397 | if (cpuset_zone_allowed_hardwall(zone, flags) && |
3c517a61 | 3398 | cache->nodelists[nid] && |
481c5346 | 3399 | cache->nodelists[nid]->free_objects) { |
3c517a61 CL |
3400 | obj = ____cache_alloc_node(cache, |
3401 | flags | GFP_THISNODE, nid); | |
481c5346 CL |
3402 | if (obj) |
3403 | break; | |
3404 | } | |
3c517a61 CL |
3405 | } |
3406 | ||
cfce6604 | 3407 | if (!obj) { |
3c517a61 CL |
3408 | /* |
3409 | * This allocation will be performed within the constraints | |
3410 | * of the current cpuset / memory policy requirements. | |
3411 | * We may trigger various forms of reclaim on the allowed | |
3412 | * set and go into memory reserves if necessary. | |
3413 | */ | |
dd47ea75 CL |
3414 | if (local_flags & __GFP_WAIT) |
3415 | local_irq_enable(); | |
3416 | kmem_flagcheck(cache, flags); | |
7d6e6d09 | 3417 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); |
dd47ea75 CL |
3418 | if (local_flags & __GFP_WAIT) |
3419 | local_irq_disable(); | |
3c517a61 CL |
3420 | if (obj) { |
3421 | /* | |
3422 | * Insert into the appropriate per node queues | |
3423 | */ | |
3424 | nid = page_to_nid(virt_to_page(obj)); | |
3425 | if (cache_grow(cache, flags, nid, obj)) { | |
3426 | obj = ____cache_alloc_node(cache, | |
3427 | flags | GFP_THISNODE, nid); | |
3428 | if (!obj) | |
3429 | /* | |
3430 | * Another processor may allocate the | |
3431 | * objects in the slab since we are | |
3432 | * not holding any locks. | |
3433 | */ | |
3434 | goto retry; | |
3435 | } else { | |
b6a60451 | 3436 | /* cache_grow already freed obj */ |
3c517a61 CL |
3437 | obj = NULL; |
3438 | } | |
3439 | } | |
aedb0eb1 | 3440 | } |
cc9a6c87 MG |
3441 | |
3442 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3443 | goto retry_cpuset; | |
765c4507 CL |
3444 | return obj; |
3445 | } | |
3446 | ||
e498be7d CL |
3447 | /* |
3448 | * A interface to enable slab creation on nodeid | |
1da177e4 | 3449 | */ |
8b98c169 | 3450 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
a737b3e2 | 3451 | int nodeid) |
e498be7d CL |
3452 | { |
3453 | struct list_head *entry; | |
b28a02de PE |
3454 | struct slab *slabp; |
3455 | struct kmem_list3 *l3; | |
3456 | void *obj; | |
b28a02de PE |
3457 | int x; |
3458 | ||
3459 | l3 = cachep->nodelists[nodeid]; | |
3460 | BUG_ON(!l3); | |
3461 | ||
a737b3e2 | 3462 | retry: |
ca3b9b91 | 3463 | check_irq_off(); |
b28a02de PE |
3464 | spin_lock(&l3->list_lock); |
3465 | entry = l3->slabs_partial.next; | |
3466 | if (entry == &l3->slabs_partial) { | |
3467 | l3->free_touched = 1; | |
3468 | entry = l3->slabs_free.next; | |
3469 | if (entry == &l3->slabs_free) | |
3470 | goto must_grow; | |
3471 | } | |
3472 | ||
3473 | slabp = list_entry(entry, struct slab, list); | |
3474 | check_spinlock_acquired_node(cachep, nodeid); | |
3475 | check_slabp(cachep, slabp); | |
3476 | ||
3477 | STATS_INC_NODEALLOCS(cachep); | |
3478 | STATS_INC_ACTIVE(cachep); | |
3479 | STATS_SET_HIGH(cachep); | |
3480 | ||
3481 | BUG_ON(slabp->inuse == cachep->num); | |
3482 | ||
78d382d7 | 3483 | obj = slab_get_obj(cachep, slabp, nodeid); |
b28a02de PE |
3484 | check_slabp(cachep, slabp); |
3485 | l3->free_objects--; | |
3486 | /* move slabp to correct slabp list: */ | |
3487 | list_del(&slabp->list); | |
3488 | ||
a737b3e2 | 3489 | if (slabp->free == BUFCTL_END) |
b28a02de | 3490 | list_add(&slabp->list, &l3->slabs_full); |
a737b3e2 | 3491 | else |
b28a02de | 3492 | list_add(&slabp->list, &l3->slabs_partial); |
e498be7d | 3493 | |
b28a02de PE |
3494 | spin_unlock(&l3->list_lock); |
3495 | goto done; | |
e498be7d | 3496 | |
a737b3e2 | 3497 | must_grow: |
b28a02de | 3498 | spin_unlock(&l3->list_lock); |
3c517a61 | 3499 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); |
765c4507 CL |
3500 | if (x) |
3501 | goto retry; | |
1da177e4 | 3502 | |
8c8cc2c1 | 3503 | return fallback_alloc(cachep, flags); |
e498be7d | 3504 | |
a737b3e2 | 3505 | done: |
b28a02de | 3506 | return obj; |
e498be7d | 3507 | } |
8c8cc2c1 PE |
3508 | |
3509 | /** | |
3510 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3511 | * @cachep: The cache to allocate from. | |
3512 | * @flags: See kmalloc(). | |
3513 | * @nodeid: node number of the target node. | |
3514 | * @caller: return address of caller, used for debug information | |
3515 | * | |
3516 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3517 | * node, which can improve the performance for cpu bound structures. | |
3518 | * | |
3519 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3520 | */ | |
3521 | static __always_inline void * | |
3522 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3523 | void *caller) | |
3524 | { | |
3525 | unsigned long save_flags; | |
3526 | void *ptr; | |
7d6e6d09 | 3527 | int slab_node = numa_mem_id(); |
8c8cc2c1 | 3528 | |
dcce284a | 3529 | flags &= gfp_allowed_mask; |
7e85ee0c | 3530 | |
cf40bd16 NP |
3531 | lockdep_trace_alloc(flags); |
3532 | ||
773ff60e | 3533 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3534 | return NULL; |
3535 | ||
8c8cc2c1 PE |
3536 | cache_alloc_debugcheck_before(cachep, flags); |
3537 | local_irq_save(save_flags); | |
3538 | ||
eacbbae3 | 3539 | if (nodeid == NUMA_NO_NODE) |
7d6e6d09 | 3540 | nodeid = slab_node; |
8c8cc2c1 PE |
3541 | |
3542 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3543 | /* Node not bootstrapped yet */ | |
3544 | ptr = fallback_alloc(cachep, flags); | |
3545 | goto out; | |
3546 | } | |
3547 | ||
7d6e6d09 | 3548 | if (nodeid == slab_node) { |
8c8cc2c1 PE |
3549 | /* |
3550 | * Use the locally cached objects if possible. | |
3551 | * However ____cache_alloc does not allow fallback | |
3552 | * to other nodes. It may fail while we still have | |
3553 | * objects on other nodes available. | |
3554 | */ | |
3555 | ptr = ____cache_alloc(cachep, flags); | |
3556 | if (ptr) | |
3557 | goto out; | |
3558 | } | |
3559 | /* ___cache_alloc_node can fall back to other nodes */ | |
3560 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3561 | out: | |
3562 | local_irq_restore(save_flags); | |
3563 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
8c138bc0 | 3564 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3565 | flags); |
8c8cc2c1 | 3566 | |
c175eea4 | 3567 | if (likely(ptr)) |
8c138bc0 | 3568 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); |
c175eea4 | 3569 | |
d07dbea4 | 3570 | if (unlikely((flags & __GFP_ZERO) && ptr)) |
8c138bc0 | 3571 | memset(ptr, 0, cachep->object_size); |
d07dbea4 | 3572 | |
8c8cc2c1 PE |
3573 | return ptr; |
3574 | } | |
3575 | ||
3576 | static __always_inline void * | |
3577 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3578 | { | |
3579 | void *objp; | |
3580 | ||
3581 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3582 | objp = alternate_node_alloc(cache, flags); | |
3583 | if (objp) | |
3584 | goto out; | |
3585 | } | |
3586 | objp = ____cache_alloc(cache, flags); | |
3587 | ||
3588 | /* | |
3589 | * We may just have run out of memory on the local node. | |
3590 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3591 | */ | |
7d6e6d09 LS |
3592 | if (!objp) |
3593 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
8c8cc2c1 PE |
3594 | |
3595 | out: | |
3596 | return objp; | |
3597 | } | |
3598 | #else | |
3599 | ||
3600 | static __always_inline void * | |
3601 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3602 | { | |
3603 | return ____cache_alloc(cachep, flags); | |
3604 | } | |
3605 | ||
3606 | #endif /* CONFIG_NUMA */ | |
3607 | ||
3608 | static __always_inline void * | |
3609 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3610 | { | |
3611 | unsigned long save_flags; | |
3612 | void *objp; | |
3613 | ||
dcce284a | 3614 | flags &= gfp_allowed_mask; |
7e85ee0c | 3615 | |
cf40bd16 NP |
3616 | lockdep_trace_alloc(flags); |
3617 | ||
773ff60e | 3618 | if (slab_should_failslab(cachep, flags)) |
824ebef1 AM |
3619 | return NULL; |
3620 | ||
8c8cc2c1 PE |
3621 | cache_alloc_debugcheck_before(cachep, flags); |
3622 | local_irq_save(save_flags); | |
3623 | objp = __do_cache_alloc(cachep, flags); | |
3624 | local_irq_restore(save_flags); | |
3625 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
8c138bc0 | 3626 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, |
d5cff635 | 3627 | flags); |
8c8cc2c1 PE |
3628 | prefetchw(objp); |
3629 | ||
c175eea4 | 3630 | if (likely(objp)) |
8c138bc0 | 3631 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); |
c175eea4 | 3632 | |
d07dbea4 | 3633 | if (unlikely((flags & __GFP_ZERO) && objp)) |
8c138bc0 | 3634 | memset(objp, 0, cachep->object_size); |
d07dbea4 | 3635 | |
8c8cc2c1 PE |
3636 | return objp; |
3637 | } | |
e498be7d CL |
3638 | |
3639 | /* | |
3640 | * Caller needs to acquire correct kmem_list's list_lock | |
3641 | */ | |
343e0d7a | 3642 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, |
b28a02de | 3643 | int node) |
1da177e4 LT |
3644 | { |
3645 | int i; | |
e498be7d | 3646 | struct kmem_list3 *l3; |
1da177e4 LT |
3647 | |
3648 | for (i = 0; i < nr_objects; i++) { | |
072bb0aa | 3649 | void *objp; |
1da177e4 | 3650 | struct slab *slabp; |
1da177e4 | 3651 | |
072bb0aa MG |
3652 | clear_obj_pfmemalloc(&objpp[i]); |
3653 | objp = objpp[i]; | |
3654 | ||
6ed5eb22 | 3655 | slabp = virt_to_slab(objp); |
ff69416e | 3656 | l3 = cachep->nodelists[node]; |
1da177e4 | 3657 | list_del(&slabp->list); |
ff69416e | 3658 | check_spinlock_acquired_node(cachep, node); |
1da177e4 | 3659 | check_slabp(cachep, slabp); |
78d382d7 | 3660 | slab_put_obj(cachep, slabp, objp, node); |
1da177e4 | 3661 | STATS_DEC_ACTIVE(cachep); |
e498be7d | 3662 | l3->free_objects++; |
1da177e4 LT |
3663 | check_slabp(cachep, slabp); |
3664 | ||
3665 | /* fixup slab chains */ | |
3666 | if (slabp->inuse == 0) { | |
e498be7d CL |
3667 | if (l3->free_objects > l3->free_limit) { |
3668 | l3->free_objects -= cachep->num; | |
e5ac9c5a RT |
3669 | /* No need to drop any previously held |
3670 | * lock here, even if we have a off-slab slab | |
3671 | * descriptor it is guaranteed to come from | |
3672 | * a different cache, refer to comments before | |
3673 | * alloc_slabmgmt. | |
3674 | */ | |
1da177e4 LT |
3675 | slab_destroy(cachep, slabp); |
3676 | } else { | |
e498be7d | 3677 | list_add(&slabp->list, &l3->slabs_free); |
1da177e4 LT |
3678 | } |
3679 | } else { | |
3680 | /* Unconditionally move a slab to the end of the | |
3681 | * partial list on free - maximum time for the | |
3682 | * other objects to be freed, too. | |
3683 | */ | |
e498be7d | 3684 | list_add_tail(&slabp->list, &l3->slabs_partial); |
1da177e4 LT |
3685 | } |
3686 | } | |
3687 | } | |
3688 | ||
343e0d7a | 3689 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
1da177e4 LT |
3690 | { |
3691 | int batchcount; | |
e498be7d | 3692 | struct kmem_list3 *l3; |
7d6e6d09 | 3693 | int node = numa_mem_id(); |
1da177e4 LT |
3694 | |
3695 | batchcount = ac->batchcount; | |
3696 | #if DEBUG | |
3697 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3698 | #endif | |
3699 | check_irq_off(); | |
ff69416e | 3700 | l3 = cachep->nodelists[node]; |
873623df | 3701 | spin_lock(&l3->list_lock); |
e498be7d CL |
3702 | if (l3->shared) { |
3703 | struct array_cache *shared_array = l3->shared; | |
b28a02de | 3704 | int max = shared_array->limit - shared_array->avail; |
1da177e4 LT |
3705 | if (max) { |
3706 | if (batchcount > max) | |
3707 | batchcount = max; | |
e498be7d | 3708 | memcpy(&(shared_array->entry[shared_array->avail]), |
b28a02de | 3709 | ac->entry, sizeof(void *) * batchcount); |
1da177e4 LT |
3710 | shared_array->avail += batchcount; |
3711 | goto free_done; | |
3712 | } | |
3713 | } | |
3714 | ||
ff69416e | 3715 | free_block(cachep, ac->entry, batchcount, node); |
a737b3e2 | 3716 | free_done: |
1da177e4 LT |
3717 | #if STATS |
3718 | { | |
3719 | int i = 0; | |
3720 | struct list_head *p; | |
3721 | ||
e498be7d CL |
3722 | p = l3->slabs_free.next; |
3723 | while (p != &(l3->slabs_free)) { | |
1da177e4 LT |
3724 | struct slab *slabp; |
3725 | ||
3726 | slabp = list_entry(p, struct slab, list); | |
3727 | BUG_ON(slabp->inuse); | |
3728 | ||
3729 | i++; | |
3730 | p = p->next; | |
3731 | } | |
3732 | STATS_SET_FREEABLE(cachep, i); | |
3733 | } | |
3734 | #endif | |
e498be7d | 3735 | spin_unlock(&l3->list_lock); |
1da177e4 | 3736 | ac->avail -= batchcount; |
a737b3e2 | 3737 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
1da177e4 LT |
3738 | } |
3739 | ||
3740 | /* | |
a737b3e2 AM |
3741 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3742 | * be in this state _before_ it is released. Called with disabled ints. | |
1da177e4 | 3743 | */ |
a947eb95 SS |
3744 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3745 | void *caller) | |
1da177e4 | 3746 | { |
9a2dba4b | 3747 | struct array_cache *ac = cpu_cache_get(cachep); |
1da177e4 LT |
3748 | |
3749 | check_irq_off(); | |
d5cff635 | 3750 | kmemleak_free_recursive(objp, cachep->flags); |
a947eb95 | 3751 | objp = cache_free_debugcheck(cachep, objp, caller); |
1da177e4 | 3752 | |
8c138bc0 | 3753 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
c175eea4 | 3754 | |
1807a1aa SS |
3755 | /* |
3756 | * Skip calling cache_free_alien() when the platform is not numa. | |
3757 | * This will avoid cache misses that happen while accessing slabp (which | |
3758 | * is per page memory reference) to get nodeid. Instead use a global | |
3759 | * variable to skip the call, which is mostly likely to be present in | |
3760 | * the cache. | |
3761 | */ | |
b6e68bc1 | 3762 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
729bd0b7 PE |
3763 | return; |
3764 | ||
1da177e4 LT |
3765 | if (likely(ac->avail < ac->limit)) { |
3766 | STATS_INC_FREEHIT(cachep); | |
1da177e4 LT |
3767 | } else { |
3768 | STATS_INC_FREEMISS(cachep); | |
3769 | cache_flusharray(cachep, ac); | |
1da177e4 | 3770 | } |
42c8c99c | 3771 | |
072bb0aa | 3772 | ac_put_obj(cachep, ac, objp); |
1da177e4 LT |
3773 | } |
3774 | ||
3775 | /** | |
3776 | * kmem_cache_alloc - Allocate an object | |
3777 | * @cachep: The cache to allocate from. | |
3778 | * @flags: See kmalloc(). | |
3779 | * | |
3780 | * Allocate an object from this cache. The flags are only relevant | |
3781 | * if the cache has no available objects. | |
3782 | */ | |
343e0d7a | 3783 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
1da177e4 | 3784 | { |
36555751 EGM |
3785 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); |
3786 | ||
ca2b84cb | 3787 | trace_kmem_cache_alloc(_RET_IP_, ret, |
8c138bc0 | 3788 | cachep->object_size, cachep->size, flags); |
36555751 EGM |
3789 | |
3790 | return ret; | |
1da177e4 LT |
3791 | } |
3792 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3793 | ||
0f24f128 | 3794 | #ifdef CONFIG_TRACING |
85beb586 SR |
3795 | void * |
3796 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
36555751 | 3797 | { |
85beb586 SR |
3798 | void *ret; |
3799 | ||
3800 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3801 | ||
3802 | trace_kmalloc(_RET_IP_, ret, | |
3803 | size, slab_buffer_size(cachep), flags); | |
3804 | return ret; | |
36555751 | 3805 | } |
85beb586 | 3806 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
36555751 EGM |
3807 | #endif |
3808 | ||
1da177e4 | 3809 | #ifdef CONFIG_NUMA |
8b98c169 CH |
3810 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3811 | { | |
36555751 EGM |
3812 | void *ret = __cache_alloc_node(cachep, flags, nodeid, |
3813 | __builtin_return_address(0)); | |
3814 | ||
ca2b84cb | 3815 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
8c138bc0 | 3816 | cachep->object_size, cachep->size, |
ca2b84cb | 3817 | flags, nodeid); |
36555751 EGM |
3818 | |
3819 | return ret; | |
8b98c169 | 3820 | } |
1da177e4 LT |
3821 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3822 | ||
0f24f128 | 3823 | #ifdef CONFIG_TRACING |
85beb586 SR |
3824 | void *kmem_cache_alloc_node_trace(size_t size, |
3825 | struct kmem_cache *cachep, | |
3826 | gfp_t flags, | |
3827 | int nodeid) | |
36555751 | 3828 | { |
85beb586 SR |
3829 | void *ret; |
3830 | ||
3831 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
36555751 | 3832 | __builtin_return_address(0)); |
85beb586 SR |
3833 | trace_kmalloc_node(_RET_IP_, ret, |
3834 | size, slab_buffer_size(cachep), | |
3835 | flags, nodeid); | |
3836 | return ret; | |
36555751 | 3837 | } |
85beb586 | 3838 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
36555751 EGM |
3839 | #endif |
3840 | ||
8b98c169 CH |
3841 | static __always_inline void * |
3842 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
97e2bde4 | 3843 | { |
343e0d7a | 3844 | struct kmem_cache *cachep; |
97e2bde4 MS |
3845 | |
3846 | cachep = kmem_find_general_cachep(size, flags); | |
6cb8f913 CL |
3847 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3848 | return cachep; | |
85beb586 | 3849 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); |
97e2bde4 | 3850 | } |
8b98c169 | 3851 | |
0bb38a5c | 3852 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
8b98c169 CH |
3853 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3854 | { | |
3855 | return __do_kmalloc_node(size, flags, node, | |
3856 | __builtin_return_address(0)); | |
3857 | } | |
dbe5e69d | 3858 | EXPORT_SYMBOL(__kmalloc_node); |
8b98c169 CH |
3859 | |
3860 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
ce71e27c | 3861 | int node, unsigned long caller) |
8b98c169 | 3862 | { |
ce71e27c | 3863 | return __do_kmalloc_node(size, flags, node, (void *)caller); |
8b98c169 CH |
3864 | } |
3865 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3866 | #else | |
3867 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3868 | { | |
3869 | return __do_kmalloc_node(size, flags, node, NULL); | |
3870 | } | |
3871 | EXPORT_SYMBOL(__kmalloc_node); | |
0bb38a5c | 3872 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ |
8b98c169 | 3873 | #endif /* CONFIG_NUMA */ |
1da177e4 LT |
3874 | |
3875 | /** | |
800590f5 | 3876 | * __do_kmalloc - allocate memory |
1da177e4 | 3877 | * @size: how many bytes of memory are required. |
800590f5 | 3878 | * @flags: the type of memory to allocate (see kmalloc). |
911851e6 | 3879 | * @caller: function caller for debug tracking of the caller |
1da177e4 | 3880 | */ |
7fd6b141 PE |
3881 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3882 | void *caller) | |
1da177e4 | 3883 | { |
343e0d7a | 3884 | struct kmem_cache *cachep; |
36555751 | 3885 | void *ret; |
1da177e4 | 3886 | |
97e2bde4 MS |
3887 | /* If you want to save a few bytes .text space: replace |
3888 | * __ with kmem_. | |
3889 | * Then kmalloc uses the uninlined functions instead of the inline | |
3890 | * functions. | |
3891 | */ | |
3892 | cachep = __find_general_cachep(size, flags); | |
a5c96d8a LT |
3893 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3894 | return cachep; | |
36555751 EGM |
3895 | ret = __cache_alloc(cachep, flags, caller); |
3896 | ||
ca2b84cb | 3897 | trace_kmalloc((unsigned long) caller, ret, |
3b0efdfa | 3898 | size, cachep->size, flags); |
36555751 EGM |
3899 | |
3900 | return ret; | |
7fd6b141 PE |
3901 | } |
3902 | ||
7fd6b141 | 3903 | |
0bb38a5c | 3904 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) |
7fd6b141 PE |
3905 | void *__kmalloc(size_t size, gfp_t flags) |
3906 | { | |
871751e2 | 3907 | return __do_kmalloc(size, flags, __builtin_return_address(0)); |
1da177e4 LT |
3908 | } |
3909 | EXPORT_SYMBOL(__kmalloc); | |
3910 | ||
ce71e27c | 3911 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
7fd6b141 | 3912 | { |
ce71e27c | 3913 | return __do_kmalloc(size, flags, (void *)caller); |
7fd6b141 PE |
3914 | } |
3915 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
1d2c8eea CH |
3916 | |
3917 | #else | |
3918 | void *__kmalloc(size_t size, gfp_t flags) | |
3919 | { | |
3920 | return __do_kmalloc(size, flags, NULL); | |
3921 | } | |
3922 | EXPORT_SYMBOL(__kmalloc); | |
7fd6b141 PE |
3923 | #endif |
3924 | ||
1da177e4 LT |
3925 | /** |
3926 | * kmem_cache_free - Deallocate an object | |
3927 | * @cachep: The cache the allocation was from. | |
3928 | * @objp: The previously allocated object. | |
3929 | * | |
3930 | * Free an object which was previously allocated from this | |
3931 | * cache. | |
3932 | */ | |
343e0d7a | 3933 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
1da177e4 LT |
3934 | { |
3935 | unsigned long flags; | |
3936 | ||
3937 | local_irq_save(flags); | |
d97d476b | 3938 | debug_check_no_locks_freed(objp, cachep->object_size); |
3ac7fe5a | 3939 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
8c138bc0 | 3940 | debug_check_no_obj_freed(objp, cachep->object_size); |
a947eb95 | 3941 | __cache_free(cachep, objp, __builtin_return_address(0)); |
1da177e4 | 3942 | local_irq_restore(flags); |
36555751 | 3943 | |
ca2b84cb | 3944 | trace_kmem_cache_free(_RET_IP_, objp); |
1da177e4 LT |
3945 | } |
3946 | EXPORT_SYMBOL(kmem_cache_free); | |
3947 | ||
1da177e4 LT |
3948 | /** |
3949 | * kfree - free previously allocated memory | |
3950 | * @objp: pointer returned by kmalloc. | |
3951 | * | |
80e93eff PE |
3952 | * If @objp is NULL, no operation is performed. |
3953 | * | |
1da177e4 LT |
3954 | * Don't free memory not originally allocated by kmalloc() |
3955 | * or you will run into trouble. | |
3956 | */ | |
3957 | void kfree(const void *objp) | |
3958 | { | |
343e0d7a | 3959 | struct kmem_cache *c; |
1da177e4 LT |
3960 | unsigned long flags; |
3961 | ||
2121db74 PE |
3962 | trace_kfree(_RET_IP_, objp); |
3963 | ||
6cb8f913 | 3964 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
1da177e4 LT |
3965 | return; |
3966 | local_irq_save(flags); | |
3967 | kfree_debugcheck(objp); | |
6ed5eb22 | 3968 | c = virt_to_cache(objp); |
8c138bc0 CL |
3969 | debug_check_no_locks_freed(objp, c->object_size); |
3970 | ||
3971 | debug_check_no_obj_freed(objp, c->object_size); | |
a947eb95 | 3972 | __cache_free(c, (void *)objp, __builtin_return_address(0)); |
1da177e4 LT |
3973 | local_irq_restore(flags); |
3974 | } | |
3975 | EXPORT_SYMBOL(kfree); | |
3976 | ||
343e0d7a | 3977 | unsigned int kmem_cache_size(struct kmem_cache *cachep) |
1da177e4 | 3978 | { |
8c138bc0 | 3979 | return cachep->object_size; |
1da177e4 LT |
3980 | } |
3981 | EXPORT_SYMBOL(kmem_cache_size); | |
3982 | ||
e498be7d | 3983 | /* |
183ff22b | 3984 | * This initializes kmem_list3 or resizes various caches for all nodes. |
e498be7d | 3985 | */ |
83b519e8 | 3986 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) |
e498be7d CL |
3987 | { |
3988 | int node; | |
3989 | struct kmem_list3 *l3; | |
cafeb02e | 3990 | struct array_cache *new_shared; |
3395ee05 | 3991 | struct array_cache **new_alien = NULL; |
e498be7d | 3992 | |
9c09a95c | 3993 | for_each_online_node(node) { |
cafeb02e | 3994 | |
3395ee05 | 3995 | if (use_alien_caches) { |
83b519e8 | 3996 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
3395ee05 PM |
3997 | if (!new_alien) |
3998 | goto fail; | |
3999 | } | |
cafeb02e | 4000 | |
63109846 ED |
4001 | new_shared = NULL; |
4002 | if (cachep->shared) { | |
4003 | new_shared = alloc_arraycache(node, | |
0718dc2a | 4004 | cachep->shared*cachep->batchcount, |
83b519e8 | 4005 | 0xbaadf00d, gfp); |
63109846 ED |
4006 | if (!new_shared) { |
4007 | free_alien_cache(new_alien); | |
4008 | goto fail; | |
4009 | } | |
0718dc2a | 4010 | } |
cafeb02e | 4011 | |
a737b3e2 AM |
4012 | l3 = cachep->nodelists[node]; |
4013 | if (l3) { | |
cafeb02e CL |
4014 | struct array_cache *shared = l3->shared; |
4015 | ||
e498be7d CL |
4016 | spin_lock_irq(&l3->list_lock); |
4017 | ||
cafeb02e | 4018 | if (shared) |
0718dc2a CL |
4019 | free_block(cachep, shared->entry, |
4020 | shared->avail, node); | |
e498be7d | 4021 | |
cafeb02e CL |
4022 | l3->shared = new_shared; |
4023 | if (!l3->alien) { | |
e498be7d CL |
4024 | l3->alien = new_alien; |
4025 | new_alien = NULL; | |
4026 | } | |
b28a02de | 4027 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4028 | cachep->batchcount + cachep->num; |
e498be7d | 4029 | spin_unlock_irq(&l3->list_lock); |
cafeb02e | 4030 | kfree(shared); |
e498be7d CL |
4031 | free_alien_cache(new_alien); |
4032 | continue; | |
4033 | } | |
83b519e8 | 4034 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); |
0718dc2a CL |
4035 | if (!l3) { |
4036 | free_alien_cache(new_alien); | |
4037 | kfree(new_shared); | |
e498be7d | 4038 | goto fail; |
0718dc2a | 4039 | } |
e498be7d CL |
4040 | |
4041 | kmem_list3_init(l3); | |
4042 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
a737b3e2 | 4043 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; |
cafeb02e | 4044 | l3->shared = new_shared; |
e498be7d | 4045 | l3->alien = new_alien; |
b28a02de | 4046 | l3->free_limit = (1 + nr_cpus_node(node)) * |
a737b3e2 | 4047 | cachep->batchcount + cachep->num; |
e498be7d CL |
4048 | cachep->nodelists[node] = l3; |
4049 | } | |
cafeb02e | 4050 | return 0; |
0718dc2a | 4051 | |
a737b3e2 | 4052 | fail: |
3b0efdfa | 4053 | if (!cachep->list.next) { |
0718dc2a CL |
4054 | /* Cache is not active yet. Roll back what we did */ |
4055 | node--; | |
4056 | while (node >= 0) { | |
4057 | if (cachep->nodelists[node]) { | |
4058 | l3 = cachep->nodelists[node]; | |
4059 | ||
4060 | kfree(l3->shared); | |
4061 | free_alien_cache(l3->alien); | |
4062 | kfree(l3); | |
4063 | cachep->nodelists[node] = NULL; | |
4064 | } | |
4065 | node--; | |
4066 | } | |
4067 | } | |
cafeb02e | 4068 | return -ENOMEM; |
e498be7d CL |
4069 | } |
4070 | ||
1da177e4 | 4071 | struct ccupdate_struct { |
343e0d7a | 4072 | struct kmem_cache *cachep; |
acfe7d74 | 4073 | struct array_cache *new[0]; |
1da177e4 LT |
4074 | }; |
4075 | ||
4076 | static void do_ccupdate_local(void *info) | |
4077 | { | |
a737b3e2 | 4078 | struct ccupdate_struct *new = info; |
1da177e4 LT |
4079 | struct array_cache *old; |
4080 | ||
4081 | check_irq_off(); | |
9a2dba4b | 4082 | old = cpu_cache_get(new->cachep); |
e498be7d | 4083 | |
1da177e4 LT |
4084 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; |
4085 | new->new[smp_processor_id()] = old; | |
4086 | } | |
4087 | ||
18004c5d | 4088 | /* Always called with the slab_mutex held */ |
a737b3e2 | 4089 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
83b519e8 | 4090 | int batchcount, int shared, gfp_t gfp) |
1da177e4 | 4091 | { |
d2e7b7d0 | 4092 | struct ccupdate_struct *new; |
2ed3a4ef | 4093 | int i; |
1da177e4 | 4094 | |
acfe7d74 ED |
4095 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), |
4096 | gfp); | |
d2e7b7d0 SS |
4097 | if (!new) |
4098 | return -ENOMEM; | |
4099 | ||
e498be7d | 4100 | for_each_online_cpu(i) { |
7d6e6d09 | 4101 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, |
83b519e8 | 4102 | batchcount, gfp); |
d2e7b7d0 | 4103 | if (!new->new[i]) { |
b28a02de | 4104 | for (i--; i >= 0; i--) |
d2e7b7d0 SS |
4105 | kfree(new->new[i]); |
4106 | kfree(new); | |
e498be7d | 4107 | return -ENOMEM; |
1da177e4 LT |
4108 | } |
4109 | } | |
d2e7b7d0 | 4110 | new->cachep = cachep; |
1da177e4 | 4111 | |
15c8b6c1 | 4112 | on_each_cpu(do_ccupdate_local, (void *)new, 1); |
e498be7d | 4113 | |
1da177e4 | 4114 | check_irq_on(); |
1da177e4 LT |
4115 | cachep->batchcount = batchcount; |
4116 | cachep->limit = limit; | |
e498be7d | 4117 | cachep->shared = shared; |
1da177e4 | 4118 | |
e498be7d | 4119 | for_each_online_cpu(i) { |
d2e7b7d0 | 4120 | struct array_cache *ccold = new->new[i]; |
1da177e4 LT |
4121 | if (!ccold) |
4122 | continue; | |
7d6e6d09 LS |
4123 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); |
4124 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4125 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
1da177e4 LT |
4126 | kfree(ccold); |
4127 | } | |
d2e7b7d0 | 4128 | kfree(new); |
83b519e8 | 4129 | return alloc_kmemlist(cachep, gfp); |
1da177e4 LT |
4130 | } |
4131 | ||
18004c5d | 4132 | /* Called with slab_mutex held always */ |
83b519e8 | 4133 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
1da177e4 LT |
4134 | { |
4135 | int err; | |
4136 | int limit, shared; | |
4137 | ||
a737b3e2 AM |
4138 | /* |
4139 | * The head array serves three purposes: | |
1da177e4 LT |
4140 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
4141 | * - reduce the number of spinlock operations. | |
a737b3e2 | 4142 | * - reduce the number of linked list operations on the slab and |
1da177e4 LT |
4143 | * bufctl chains: array operations are cheaper. |
4144 | * The numbers are guessed, we should auto-tune as described by | |
4145 | * Bonwick. | |
4146 | */ | |
3b0efdfa | 4147 | if (cachep->size > 131072) |
1da177e4 | 4148 | limit = 1; |
3b0efdfa | 4149 | else if (cachep->size > PAGE_SIZE) |
1da177e4 | 4150 | limit = 8; |
3b0efdfa | 4151 | else if (cachep->size > 1024) |
1da177e4 | 4152 | limit = 24; |
3b0efdfa | 4153 | else if (cachep->size > 256) |
1da177e4 LT |
4154 | limit = 54; |
4155 | else | |
4156 | limit = 120; | |
4157 | ||
a737b3e2 AM |
4158 | /* |
4159 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
1da177e4 LT |
4160 | * allocation behaviour: Most allocs on one cpu, most free operations |
4161 | * on another cpu. For these cases, an efficient object passing between | |
4162 | * cpus is necessary. This is provided by a shared array. The array | |
4163 | * replaces Bonwick's magazine layer. | |
4164 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4165 | * to a larger limit. Thus disabled by default. | |
4166 | */ | |
4167 | shared = 0; | |
3b0efdfa | 4168 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
1da177e4 | 4169 | shared = 8; |
1da177e4 LT |
4170 | |
4171 | #if DEBUG | |
a737b3e2 AM |
4172 | /* |
4173 | * With debugging enabled, large batchcount lead to excessively long | |
4174 | * periods with disabled local interrupts. Limit the batchcount | |
1da177e4 LT |
4175 | */ |
4176 | if (limit > 32) | |
4177 | limit = 32; | |
4178 | #endif | |
83b519e8 | 4179 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); |
1da177e4 LT |
4180 | if (err) |
4181 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
b28a02de | 4182 | cachep->name, -err); |
2ed3a4ef | 4183 | return err; |
1da177e4 LT |
4184 | } |
4185 | ||
1b55253a CL |
4186 | /* |
4187 | * Drain an array if it contains any elements taking the l3 lock only if | |
b18e7e65 CL |
4188 | * necessary. Note that the l3 listlock also protects the array_cache |
4189 | * if drain_array() is used on the shared array. | |
1b55253a | 4190 | */ |
68a1b195 | 4191 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, |
1b55253a | 4192 | struct array_cache *ac, int force, int node) |
1da177e4 LT |
4193 | { |
4194 | int tofree; | |
4195 | ||
1b55253a CL |
4196 | if (!ac || !ac->avail) |
4197 | return; | |
1da177e4 LT |
4198 | if (ac->touched && !force) { |
4199 | ac->touched = 0; | |
b18e7e65 | 4200 | } else { |
1b55253a | 4201 | spin_lock_irq(&l3->list_lock); |
b18e7e65 CL |
4202 | if (ac->avail) { |
4203 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4204 | if (tofree > ac->avail) | |
4205 | tofree = (ac->avail + 1) / 2; | |
4206 | free_block(cachep, ac->entry, tofree, node); | |
4207 | ac->avail -= tofree; | |
4208 | memmove(ac->entry, &(ac->entry[tofree]), | |
4209 | sizeof(void *) * ac->avail); | |
4210 | } | |
1b55253a | 4211 | spin_unlock_irq(&l3->list_lock); |
1da177e4 LT |
4212 | } |
4213 | } | |
4214 | ||
4215 | /** | |
4216 | * cache_reap - Reclaim memory from caches. | |
05fb6bf0 | 4217 | * @w: work descriptor |
1da177e4 LT |
4218 | * |
4219 | * Called from workqueue/eventd every few seconds. | |
4220 | * Purpose: | |
4221 | * - clear the per-cpu caches for this CPU. | |
4222 | * - return freeable pages to the main free memory pool. | |
4223 | * | |
a737b3e2 AM |
4224 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4225 | * again on the next iteration. | |
1da177e4 | 4226 | */ |
7c5cae36 | 4227 | static void cache_reap(struct work_struct *w) |
1da177e4 | 4228 | { |
7a7c381d | 4229 | struct kmem_cache *searchp; |
e498be7d | 4230 | struct kmem_list3 *l3; |
7d6e6d09 | 4231 | int node = numa_mem_id(); |
bf6aede7 | 4232 | struct delayed_work *work = to_delayed_work(w); |
1da177e4 | 4233 | |
18004c5d | 4234 | if (!mutex_trylock(&slab_mutex)) |
1da177e4 | 4235 | /* Give up. Setup the next iteration. */ |
7c5cae36 | 4236 | goto out; |
1da177e4 | 4237 | |
18004c5d | 4238 | list_for_each_entry(searchp, &slab_caches, list) { |
1da177e4 LT |
4239 | check_irq_on(); |
4240 | ||
35386e3b CL |
4241 | /* |
4242 | * We only take the l3 lock if absolutely necessary and we | |
4243 | * have established with reasonable certainty that | |
4244 | * we can do some work if the lock was obtained. | |
4245 | */ | |
aab2207c | 4246 | l3 = searchp->nodelists[node]; |
35386e3b | 4247 | |
8fce4d8e | 4248 | reap_alien(searchp, l3); |
1da177e4 | 4249 | |
aab2207c | 4250 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); |
1da177e4 | 4251 | |
35386e3b CL |
4252 | /* |
4253 | * These are racy checks but it does not matter | |
4254 | * if we skip one check or scan twice. | |
4255 | */ | |
e498be7d | 4256 | if (time_after(l3->next_reap, jiffies)) |
35386e3b | 4257 | goto next; |
1da177e4 | 4258 | |
e498be7d | 4259 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; |
1da177e4 | 4260 | |
aab2207c | 4261 | drain_array(searchp, l3, l3->shared, 0, node); |
1da177e4 | 4262 | |
ed11d9eb | 4263 | if (l3->free_touched) |
e498be7d | 4264 | l3->free_touched = 0; |
ed11d9eb CL |
4265 | else { |
4266 | int freed; | |
1da177e4 | 4267 | |
ed11d9eb CL |
4268 | freed = drain_freelist(searchp, l3, (l3->free_limit + |
4269 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4270 | STATS_ADD_REAPED(searchp, freed); | |
4271 | } | |
35386e3b | 4272 | next: |
1da177e4 LT |
4273 | cond_resched(); |
4274 | } | |
4275 | check_irq_on(); | |
18004c5d | 4276 | mutex_unlock(&slab_mutex); |
8fce4d8e | 4277 | next_reap_node(); |
7c5cae36 | 4278 | out: |
a737b3e2 | 4279 | /* Set up the next iteration */ |
7c5cae36 | 4280 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); |
1da177e4 LT |
4281 | } |
4282 | ||
158a9624 | 4283 | #ifdef CONFIG_SLABINFO |
1da177e4 | 4284 | |
85289f98 | 4285 | static void print_slabinfo_header(struct seq_file *m) |
1da177e4 | 4286 | { |
85289f98 PE |
4287 | /* |
4288 | * Output format version, so at least we can change it | |
4289 | * without _too_ many complaints. | |
4290 | */ | |
1da177e4 | 4291 | #if STATS |
85289f98 | 4292 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); |
1da177e4 | 4293 | #else |
85289f98 | 4294 | seq_puts(m, "slabinfo - version: 2.1\n"); |
1da177e4 | 4295 | #endif |
85289f98 PE |
4296 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " |
4297 | "<objperslab> <pagesperslab>"); | |
4298 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4299 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
1da177e4 | 4300 | #if STATS |
85289f98 | 4301 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " |
fb7faf33 | 4302 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); |
85289f98 | 4303 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); |
1da177e4 | 4304 | #endif |
85289f98 PE |
4305 | seq_putc(m, '\n'); |
4306 | } | |
4307 | ||
4308 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4309 | { | |
4310 | loff_t n = *pos; | |
85289f98 | 4311 | |
18004c5d | 4312 | mutex_lock(&slab_mutex); |
85289f98 PE |
4313 | if (!n) |
4314 | print_slabinfo_header(m); | |
b92151ba | 4315 | |
18004c5d | 4316 | return seq_list_start(&slab_caches, *pos); |
1da177e4 LT |
4317 | } |
4318 | ||
4319 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4320 | { | |
18004c5d | 4321 | return seq_list_next(p, &slab_caches, pos); |
1da177e4 LT |
4322 | } |
4323 | ||
4324 | static void s_stop(struct seq_file *m, void *p) | |
4325 | { | |
18004c5d | 4326 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4327 | } |
4328 | ||
4329 | static int s_show(struct seq_file *m, void *p) | |
4330 | { | |
3b0efdfa | 4331 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
b28a02de PE |
4332 | struct slab *slabp; |
4333 | unsigned long active_objs; | |
4334 | unsigned long num_objs; | |
4335 | unsigned long active_slabs = 0; | |
4336 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
e498be7d | 4337 | const char *name; |
1da177e4 | 4338 | char *error = NULL; |
e498be7d CL |
4339 | int node; |
4340 | struct kmem_list3 *l3; | |
1da177e4 | 4341 | |
1da177e4 LT |
4342 | active_objs = 0; |
4343 | num_slabs = 0; | |
e498be7d CL |
4344 | for_each_online_node(node) { |
4345 | l3 = cachep->nodelists[node]; | |
4346 | if (!l3) | |
4347 | continue; | |
4348 | ||
ca3b9b91 RT |
4349 | check_irq_on(); |
4350 | spin_lock_irq(&l3->list_lock); | |
e498be7d | 4351 | |
7a7c381d | 4352 | list_for_each_entry(slabp, &l3->slabs_full, list) { |
e498be7d CL |
4353 | if (slabp->inuse != cachep->num && !error) |
4354 | error = "slabs_full accounting error"; | |
4355 | active_objs += cachep->num; | |
4356 | active_slabs++; | |
4357 | } | |
7a7c381d | 4358 | list_for_each_entry(slabp, &l3->slabs_partial, list) { |
e498be7d CL |
4359 | if (slabp->inuse == cachep->num && !error) |
4360 | error = "slabs_partial inuse accounting error"; | |
4361 | if (!slabp->inuse && !error) | |
4362 | error = "slabs_partial/inuse accounting error"; | |
4363 | active_objs += slabp->inuse; | |
4364 | active_slabs++; | |
4365 | } | |
7a7c381d | 4366 | list_for_each_entry(slabp, &l3->slabs_free, list) { |
e498be7d CL |
4367 | if (slabp->inuse && !error) |
4368 | error = "slabs_free/inuse accounting error"; | |
4369 | num_slabs++; | |
4370 | } | |
4371 | free_objects += l3->free_objects; | |
4484ebf1 RT |
4372 | if (l3->shared) |
4373 | shared_avail += l3->shared->avail; | |
e498be7d | 4374 | |
ca3b9b91 | 4375 | spin_unlock_irq(&l3->list_lock); |
1da177e4 | 4376 | } |
b28a02de PE |
4377 | num_slabs += active_slabs; |
4378 | num_objs = num_slabs * cachep->num; | |
e498be7d | 4379 | if (num_objs - active_objs != free_objects && !error) |
1da177e4 LT |
4380 | error = "free_objects accounting error"; |
4381 | ||
b28a02de | 4382 | name = cachep->name; |
1da177e4 LT |
4383 | if (error) |
4384 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4385 | ||
4386 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
3b0efdfa | 4387 | name, active_objs, num_objs, cachep->size, |
b28a02de | 4388 | cachep->num, (1 << cachep->gfporder)); |
1da177e4 | 4389 | seq_printf(m, " : tunables %4u %4u %4u", |
b28a02de | 4390 | cachep->limit, cachep->batchcount, cachep->shared); |
e498be7d | 4391 | seq_printf(m, " : slabdata %6lu %6lu %6lu", |
b28a02de | 4392 | active_slabs, num_slabs, shared_avail); |
1da177e4 | 4393 | #if STATS |
b28a02de | 4394 | { /* list3 stats */ |
1da177e4 LT |
4395 | unsigned long high = cachep->high_mark; |
4396 | unsigned long allocs = cachep->num_allocations; | |
4397 | unsigned long grown = cachep->grown; | |
4398 | unsigned long reaped = cachep->reaped; | |
4399 | unsigned long errors = cachep->errors; | |
4400 | unsigned long max_freeable = cachep->max_freeable; | |
1da177e4 | 4401 | unsigned long node_allocs = cachep->node_allocs; |
e498be7d | 4402 | unsigned long node_frees = cachep->node_frees; |
fb7faf33 | 4403 | unsigned long overflows = cachep->node_overflow; |
1da177e4 | 4404 | |
e92dd4fd JP |
4405 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " |
4406 | "%4lu %4lu %4lu %4lu %4lu", | |
4407 | allocs, high, grown, | |
4408 | reaped, errors, max_freeable, node_allocs, | |
4409 | node_frees, overflows); | |
1da177e4 LT |
4410 | } |
4411 | /* cpu stats */ | |
4412 | { | |
4413 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4414 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4415 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4416 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4417 | ||
4418 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
b28a02de | 4419 | allochit, allocmiss, freehit, freemiss); |
1da177e4 LT |
4420 | } |
4421 | #endif | |
4422 | seq_putc(m, '\n'); | |
1da177e4 LT |
4423 | return 0; |
4424 | } | |
4425 | ||
4426 | /* | |
4427 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4428 | * | |
4429 | * Output layout: | |
4430 | * cache-name | |
4431 | * num-active-objs | |
4432 | * total-objs | |
4433 | * object size | |
4434 | * num-active-slabs | |
4435 | * total-slabs | |
4436 | * num-pages-per-slab | |
4437 | * + further values on SMP and with statistics enabled | |
4438 | */ | |
4439 | ||
7b3c3a50 | 4440 | static const struct seq_operations slabinfo_op = { |
b28a02de PE |
4441 | .start = s_start, |
4442 | .next = s_next, | |
4443 | .stop = s_stop, | |
4444 | .show = s_show, | |
1da177e4 LT |
4445 | }; |
4446 | ||
4447 | #define MAX_SLABINFO_WRITE 128 | |
4448 | /** | |
4449 | * slabinfo_write - Tuning for the slab allocator | |
4450 | * @file: unused | |
4451 | * @buffer: user buffer | |
4452 | * @count: data length | |
4453 | * @ppos: unused | |
4454 | */ | |
68a1b195 | 4455 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
b28a02de | 4456 | size_t count, loff_t *ppos) |
1da177e4 | 4457 | { |
b28a02de | 4458 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
1da177e4 | 4459 | int limit, batchcount, shared, res; |
7a7c381d | 4460 | struct kmem_cache *cachep; |
b28a02de | 4461 | |
1da177e4 LT |
4462 | if (count > MAX_SLABINFO_WRITE) |
4463 | return -EINVAL; | |
4464 | if (copy_from_user(&kbuf, buffer, count)) | |
4465 | return -EFAULT; | |
b28a02de | 4466 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
1da177e4 LT |
4467 | |
4468 | tmp = strchr(kbuf, ' '); | |
4469 | if (!tmp) | |
4470 | return -EINVAL; | |
4471 | *tmp = '\0'; | |
4472 | tmp++; | |
4473 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4474 | return -EINVAL; | |
4475 | ||
4476 | /* Find the cache in the chain of caches. */ | |
18004c5d | 4477 | mutex_lock(&slab_mutex); |
1da177e4 | 4478 | res = -EINVAL; |
18004c5d | 4479 | list_for_each_entry(cachep, &slab_caches, list) { |
1da177e4 | 4480 | if (!strcmp(cachep->name, kbuf)) { |
a737b3e2 AM |
4481 | if (limit < 1 || batchcount < 1 || |
4482 | batchcount > limit || shared < 0) { | |
e498be7d | 4483 | res = 0; |
1da177e4 | 4484 | } else { |
e498be7d | 4485 | res = do_tune_cpucache(cachep, limit, |
83b519e8 PE |
4486 | batchcount, shared, |
4487 | GFP_KERNEL); | |
1da177e4 LT |
4488 | } |
4489 | break; | |
4490 | } | |
4491 | } | |
18004c5d | 4492 | mutex_unlock(&slab_mutex); |
1da177e4 LT |
4493 | if (res >= 0) |
4494 | res = count; | |
4495 | return res; | |
4496 | } | |
871751e2 | 4497 | |
7b3c3a50 AD |
4498 | static int slabinfo_open(struct inode *inode, struct file *file) |
4499 | { | |
4500 | return seq_open(file, &slabinfo_op); | |
4501 | } | |
4502 | ||
4503 | static const struct file_operations proc_slabinfo_operations = { | |
4504 | .open = slabinfo_open, | |
4505 | .read = seq_read, | |
4506 | .write = slabinfo_write, | |
4507 | .llseek = seq_lseek, | |
4508 | .release = seq_release, | |
4509 | }; | |
4510 | ||
871751e2 AV |
4511 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4512 | ||
4513 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4514 | { | |
18004c5d CL |
4515 | mutex_lock(&slab_mutex); |
4516 | return seq_list_start(&slab_caches, *pos); | |
871751e2 AV |
4517 | } |
4518 | ||
4519 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4520 | { | |
4521 | unsigned long *p; | |
4522 | int l; | |
4523 | if (!v) | |
4524 | return 1; | |
4525 | l = n[1]; | |
4526 | p = n + 2; | |
4527 | while (l) { | |
4528 | int i = l/2; | |
4529 | unsigned long *q = p + 2 * i; | |
4530 | if (*q == v) { | |
4531 | q[1]++; | |
4532 | return 1; | |
4533 | } | |
4534 | if (*q > v) { | |
4535 | l = i; | |
4536 | } else { | |
4537 | p = q + 2; | |
4538 | l -= i + 1; | |
4539 | } | |
4540 | } | |
4541 | if (++n[1] == n[0]) | |
4542 | return 0; | |
4543 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4544 | p[0] = v; | |
4545 | p[1] = 1; | |
4546 | return 1; | |
4547 | } | |
4548 | ||
4549 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4550 | { | |
4551 | void *p; | |
4552 | int i; | |
4553 | if (n[0] == n[1]) | |
4554 | return; | |
3b0efdfa | 4555 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { |
871751e2 AV |
4556 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) |
4557 | continue; | |
4558 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4559 | return; | |
4560 | } | |
4561 | } | |
4562 | ||
4563 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4564 | { | |
4565 | #ifdef CONFIG_KALLSYMS | |
871751e2 | 4566 | unsigned long offset, size; |
9281acea | 4567 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
871751e2 | 4568 | |
a5c43dae | 4569 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
871751e2 | 4570 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
a5c43dae | 4571 | if (modname[0]) |
871751e2 AV |
4572 | seq_printf(m, " [%s]", modname); |
4573 | return; | |
4574 | } | |
4575 | #endif | |
4576 | seq_printf(m, "%p", (void *)address); | |
4577 | } | |
4578 | ||
4579 | static int leaks_show(struct seq_file *m, void *p) | |
4580 | { | |
0672aa7c | 4581 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
871751e2 AV |
4582 | struct slab *slabp; |
4583 | struct kmem_list3 *l3; | |
4584 | const char *name; | |
4585 | unsigned long *n = m->private; | |
4586 | int node; | |
4587 | int i; | |
4588 | ||
4589 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4590 | return 0; | |
4591 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4592 | return 0; | |
4593 | ||
4594 | /* OK, we can do it */ | |
4595 | ||
4596 | n[1] = 0; | |
4597 | ||
4598 | for_each_online_node(node) { | |
4599 | l3 = cachep->nodelists[node]; | |
4600 | if (!l3) | |
4601 | continue; | |
4602 | ||
4603 | check_irq_on(); | |
4604 | spin_lock_irq(&l3->list_lock); | |
4605 | ||
7a7c381d | 4606 | list_for_each_entry(slabp, &l3->slabs_full, list) |
871751e2 | 4607 | handle_slab(n, cachep, slabp); |
7a7c381d | 4608 | list_for_each_entry(slabp, &l3->slabs_partial, list) |
871751e2 | 4609 | handle_slab(n, cachep, slabp); |
871751e2 AV |
4610 | spin_unlock_irq(&l3->list_lock); |
4611 | } | |
4612 | name = cachep->name; | |
4613 | if (n[0] == n[1]) { | |
4614 | /* Increase the buffer size */ | |
18004c5d | 4615 | mutex_unlock(&slab_mutex); |
871751e2 AV |
4616 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4617 | if (!m->private) { | |
4618 | /* Too bad, we are really out */ | |
4619 | m->private = n; | |
18004c5d | 4620 | mutex_lock(&slab_mutex); |
871751e2 AV |
4621 | return -ENOMEM; |
4622 | } | |
4623 | *(unsigned long *)m->private = n[0] * 2; | |
4624 | kfree(n); | |
18004c5d | 4625 | mutex_lock(&slab_mutex); |
871751e2 AV |
4626 | /* Now make sure this entry will be retried */ |
4627 | m->count = m->size; | |
4628 | return 0; | |
4629 | } | |
4630 | for (i = 0; i < n[1]; i++) { | |
4631 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4632 | show_symbol(m, n[2*i+2]); | |
4633 | seq_putc(m, '\n'); | |
4634 | } | |
d2e7b7d0 | 4635 | |
871751e2 AV |
4636 | return 0; |
4637 | } | |
4638 | ||
a0ec95a8 | 4639 | static const struct seq_operations slabstats_op = { |
871751e2 AV |
4640 | .start = leaks_start, |
4641 | .next = s_next, | |
4642 | .stop = s_stop, | |
4643 | .show = leaks_show, | |
4644 | }; | |
a0ec95a8 AD |
4645 | |
4646 | static int slabstats_open(struct inode *inode, struct file *file) | |
4647 | { | |
4648 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4649 | int ret = -ENOMEM; | |
4650 | if (n) { | |
4651 | ret = seq_open(file, &slabstats_op); | |
4652 | if (!ret) { | |
4653 | struct seq_file *m = file->private_data; | |
4654 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4655 | m->private = n; | |
4656 | n = NULL; | |
4657 | } | |
4658 | kfree(n); | |
4659 | } | |
4660 | return ret; | |
4661 | } | |
4662 | ||
4663 | static const struct file_operations proc_slabstats_operations = { | |
4664 | .open = slabstats_open, | |
4665 | .read = seq_read, | |
4666 | .llseek = seq_lseek, | |
4667 | .release = seq_release_private, | |
4668 | }; | |
4669 | #endif | |
4670 | ||
4671 | static int __init slab_proc_init(void) | |
4672 | { | |
ab067e99 | 4673 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); |
a0ec95a8 AD |
4674 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4675 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
871751e2 | 4676 | #endif |
a0ec95a8 AD |
4677 | return 0; |
4678 | } | |
4679 | module_init(slab_proc_init); | |
1da177e4 LT |
4680 | #endif |
4681 | ||
00e145b6 MS |
4682 | /** |
4683 | * ksize - get the actual amount of memory allocated for a given object | |
4684 | * @objp: Pointer to the object | |
4685 | * | |
4686 | * kmalloc may internally round up allocations and return more memory | |
4687 | * than requested. ksize() can be used to determine the actual amount of | |
4688 | * memory allocated. The caller may use this additional memory, even though | |
4689 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4690 | * The caller must guarantee that objp points to a valid object previously | |
4691 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4692 | * must not be freed during the duration of the call. | |
4693 | */ | |
fd76bab2 | 4694 | size_t ksize(const void *objp) |
1da177e4 | 4695 | { |
ef8b4520 CL |
4696 | BUG_ON(!objp); |
4697 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
00e145b6 | 4698 | return 0; |
1da177e4 | 4699 | |
8c138bc0 | 4700 | return virt_to_cache(objp)->object_size; |
1da177e4 | 4701 | } |
b1aabecd | 4702 | EXPORT_SYMBOL(ksize); |