]>
Commit | Line | Data |
---|---|---|
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 | |
29 | * slabs and you must pass objects with the same initializations to | |
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 | * | |
53 | * The c_cpuarray may not be read with enabled local interrupts - | |
54 | * it's changed with a smp_call_function(). | |
55 | * | |
56 | * SMP synchronization: | |
57 | * constructors and destructors are called without any locking. | |
58 | * Several members in struct kmem_cache and struct slab never change, they | |
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 | |
71 | * The global cache-chain is protected by the mutex 'slab_mutex'. | |
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 | * | |
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. | |
87 | */ | |
88 | ||
89 | #include <linux/slab.h> | |
90 | #include "slab.h" | |
91 | #include <linux/mm.h> | |
92 | #include <linux/poison.h> | |
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> | |
98 | #include <linux/cpuset.h> | |
99 | #include <linux/proc_fs.h> | |
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> | |
107 | #include <linux/string.h> | |
108 | #include <linux/uaccess.h> | |
109 | #include <linux/nodemask.h> | |
110 | #include <linux/kmemleak.h> | |
111 | #include <linux/mempolicy.h> | |
112 | #include <linux/mutex.h> | |
113 | #include <linux/fault-inject.h> | |
114 | #include <linux/rtmutex.h> | |
115 | #include <linux/reciprocal_div.h> | |
116 | #include <linux/debugobjects.h> | |
117 | #include <linux/kmemcheck.h> | |
118 | #include <linux/memory.h> | |
119 | #include <linux/prefetch.h> | |
120 | ||
121 | #include <net/sock.h> | |
122 | ||
123 | #include <asm/cacheflush.h> | |
124 | #include <asm/tlbflush.h> | |
125 | #include <asm/page.h> | |
126 | ||
127 | #include <trace/events/kmem.h> | |
128 | ||
129 | #include "internal.h" | |
130 | ||
131 | /* | |
132 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. | |
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 | ||
151 | /* Shouldn't this be in a header file somewhere? */ | |
152 | #define BYTES_PER_WORD sizeof(void *) | |
153 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) | |
154 | ||
155 | #ifndef ARCH_KMALLOC_FLAGS | |
156 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN | |
157 | #endif | |
158 | ||
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 | ||
165 | /* Legal flag mask for kmem_cache_create(). */ | |
166 | #if DEBUG | |
167 | # define CREATE_MASK (SLAB_RED_ZONE | \ | |
168 | SLAB_POISON | SLAB_HWCACHE_ALIGN | \ | |
169 | SLAB_CACHE_DMA | \ | |
170 | SLAB_STORE_USER | \ | |
171 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
172 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ | |
173 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) | |
174 | #else | |
175 | # define CREATE_MASK (SLAB_HWCACHE_ALIGN | \ | |
176 | SLAB_CACHE_DMA | \ | |
177 | SLAB_RECLAIM_ACCOUNT | SLAB_PANIC | \ | |
178 | SLAB_DESTROY_BY_RCU | SLAB_MEM_SPREAD | \ | |
179 | SLAB_DEBUG_OBJECTS | SLAB_NOLEAKTRACE | SLAB_NOTRACK) | |
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 | ||
201 | typedef unsigned int kmem_bufctl_t; | |
202 | #define BUFCTL_END (((kmem_bufctl_t)(~0U))-0) | |
203 | #define BUFCTL_FREE (((kmem_bufctl_t)(~0U))-1) | |
204 | #define BUFCTL_ACTIVE (((kmem_bufctl_t)(~0U))-2) | |
205 | #define SLAB_LIMIT (((kmem_bufctl_t)(~0U))-3) | |
206 | ||
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. | |
220 | */ | |
221 | struct slab_rcu { | |
222 | struct rcu_head head; | |
223 | struct kmem_cache *cachep; | |
224 | void *addr; | |
225 | }; | |
226 | ||
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 | ||
248 | /* | |
249 | * struct array_cache | |
250 | * | |
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; | |
265 | spinlock_t lock; | |
266 | void *entry[]; /* | |
267 | * Must have this definition in here for the proper | |
268 | * alignment of array_cache. Also simplifies accessing | |
269 | * the entries. | |
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 | |
274 | */ | |
275 | }; | |
276 | ||
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 | ||
294 | /* | |
295 | * bootstrap: The caches do not work without cpuarrays anymore, but the | |
296 | * cpuarrays are allocated from the generic caches... | |
297 | */ | |
298 | #define BOOT_CPUCACHE_ENTRIES 1 | |
299 | struct arraycache_init { | |
300 | struct array_cache cache; | |
301 | void *entries[BOOT_CPUCACHE_ENTRIES]; | |
302 | }; | |
303 | ||
304 | /* | |
305 | * The slab lists for all objects. | |
306 | */ | |
307 | struct kmem_list3 { | |
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; | |
312 | unsigned int free_limit; | |
313 | unsigned int colour_next; /* Per-node cache coloring */ | |
314 | spinlock_t list_lock; | |
315 | struct array_cache *shared; /* shared per node */ | |
316 | struct array_cache **alien; /* on other nodes */ | |
317 | unsigned long next_reap; /* updated without locking */ | |
318 | int free_touched; /* updated without locking */ | |
319 | }; | |
320 | ||
321 | /* | |
322 | * Need this for bootstrapping a per node allocator. | |
323 | */ | |
324 | #define NUM_INIT_LISTS (3 * MAX_NUMNODES) | |
325 | static struct kmem_list3 __initdata initkmem_list3[NUM_INIT_LISTS]; | |
326 | #define CACHE_CACHE 0 | |
327 | #define SIZE_AC MAX_NUMNODES | |
328 | #define SIZE_L3 (2 * MAX_NUMNODES) | |
329 | ||
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); | |
334 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); | |
335 | static void cache_reap(struct work_struct *unused); | |
336 | ||
337 | /* | |
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. | |
340 | */ | |
341 | static __always_inline int index_of(const size_t size) | |
342 | { | |
343 | extern void __bad_size(void); | |
344 | ||
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++; | |
353 | #include <linux/kmalloc_sizes.h> | |
354 | #undef CACHE | |
355 | __bad_size(); | |
356 | } else | |
357 | __bad_size(); | |
358 | return 0; | |
359 | } | |
360 | ||
361 | static int slab_early_init = 1; | |
362 | ||
363 | #define INDEX_AC index_of(sizeof(struct arraycache_init)) | |
364 | #define INDEX_L3 index_of(sizeof(struct kmem_list3)) | |
365 | ||
366 | static void kmem_list3_init(struct kmem_list3 *parent) | |
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; | |
373 | parent->colour_next = 0; | |
374 | spin_lock_init(&parent->list_lock); | |
375 | parent->free_objects = 0; | |
376 | parent->free_touched = 0; | |
377 | } | |
378 | ||
379 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ | |
380 | do { \ | |
381 | INIT_LIST_HEAD(listp); \ | |
382 | list_splice(&(cachep->nodelists[nodeid]->slab), listp); \ | |
383 | } while (0) | |
384 | ||
385 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ | |
386 | do { \ | |
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) | |
391 | ||
392 | #define CFLGS_OFF_SLAB (0x80000000UL) | |
393 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) | |
394 | ||
395 | #define BATCHREFILL_LIMIT 16 | |
396 | /* | |
397 | * Optimization question: fewer reaps means less probability for unnessary | |
398 | * cpucache drain/refill cycles. | |
399 | * | |
400 | * OTOH the cpuarrays can contain lots of objects, | |
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++) | |
411 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) | |
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) | |
417 | #define STATS_INC_ERR(x) ((x)->errors++) | |
418 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) | |
419 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) | |
420 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) | |
421 | #define STATS_SET_FREEABLE(x, i) \ | |
422 | do { \ | |
423 | if ((x)->max_freeable < i) \ | |
424 | (x)->max_freeable = i; \ | |
425 | } while (0) | |
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) | |
435 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) | |
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) | |
439 | #define STATS_INC_NODEFREES(x) do { } while (0) | |
440 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) | |
441 | #define STATS_SET_FREEABLE(x, i) do { } while (0) | |
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 | |
449 | ||
450 | /* | |
451 | * memory layout of objects: | |
452 | * 0 : objp | |
453 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that | |
454 | * the end of an object is aligned with the end of the real | |
455 | * allocation. Catches writes behind the end of the allocation. | |
456 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: | |
457 | * redzone word. | |
458 | * cachep->obj_offset: The real object. | |
459 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] | |
460 | * cachep->size - 1* BYTES_PER_WORD: last caller address | |
461 | * [BYTES_PER_WORD long] | |
462 | */ | |
463 | static int obj_offset(struct kmem_cache *cachep) | |
464 | { | |
465 | return cachep->obj_offset; | |
466 | } | |
467 | ||
468 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) | |
469 | { | |
470 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
471 | return (unsigned long long*) (objp + obj_offset(cachep) - | |
472 | sizeof(unsigned long long)); | |
473 | } | |
474 | ||
475 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) | |
476 | { | |
477 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); | |
478 | if (cachep->flags & SLAB_STORE_USER) | |
479 | return (unsigned long long *)(objp + cachep->size - | |
480 | sizeof(unsigned long long) - | |
481 | REDZONE_ALIGN); | |
482 | return (unsigned long long *) (objp + cachep->size - | |
483 | sizeof(unsigned long long)); | |
484 | } | |
485 | ||
486 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) | |
487 | { | |
488 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); | |
489 | return (void **)(objp + cachep->size - BYTES_PER_WORD); | |
490 | } | |
491 | ||
492 | #else | |
493 | ||
494 | #define obj_offset(x) 0 | |
495 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
496 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) | |
497 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) | |
498 | ||
499 | #endif | |
500 | ||
501 | #ifdef CONFIG_TRACING | |
502 | size_t slab_buffer_size(struct kmem_cache *cachep) | |
503 | { | |
504 | return cachep->size; | |
505 | } | |
506 | EXPORT_SYMBOL(slab_buffer_size); | |
507 | #endif | |
508 | ||
509 | /* | |
510 | * Do not go above this order unless 0 objects fit into the slab or | |
511 | * overridden on the command line. | |
512 | */ | |
513 | #define SLAB_MAX_ORDER_HI 1 | |
514 | #define SLAB_MAX_ORDER_LO 0 | |
515 | static int slab_max_order = SLAB_MAX_ORDER_LO; | |
516 | static bool slab_max_order_set __initdata; | |
517 | ||
518 | static inline struct kmem_cache *page_get_cache(struct page *page) | |
519 | { | |
520 | page = compound_head(page); | |
521 | BUG_ON(!PageSlab(page)); | |
522 | return page->slab_cache; | |
523 | } | |
524 | ||
525 | static inline struct kmem_cache *virt_to_cache(const void *obj) | |
526 | { | |
527 | struct page *page = virt_to_head_page(obj); | |
528 | return page->slab_cache; | |
529 | } | |
530 | ||
531 | static inline struct slab *virt_to_slab(const void *obj) | |
532 | { | |
533 | struct page *page = virt_to_head_page(obj); | |
534 | ||
535 | VM_BUG_ON(!PageSlab(page)); | |
536 | return page->slab_page; | |
537 | } | |
538 | ||
539 | static inline void *index_to_obj(struct kmem_cache *cache, struct slab *slab, | |
540 | unsigned int idx) | |
541 | { | |
542 | return slab->s_mem + cache->size * idx; | |
543 | } | |
544 | ||
545 | /* | |
546 | * We want to avoid an expensive divide : (offset / cache->size) | |
547 | * Using the fact that size is a constant for a particular cache, | |
548 | * we can replace (offset / cache->size) by | |
549 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) | |
550 | */ | |
551 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, | |
552 | const struct slab *slab, void *obj) | |
553 | { | |
554 | u32 offset = (obj - slab->s_mem); | |
555 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); | |
556 | } | |
557 | ||
558 | /* | |
559 | * These are the default caches for kmalloc. Custom caches can have other sizes. | |
560 | */ | |
561 | struct cache_sizes malloc_sizes[] = { | |
562 | #define CACHE(x) { .cs_size = (x) }, | |
563 | #include <linux/kmalloc_sizes.h> | |
564 | CACHE(ULONG_MAX) | |
565 | #undef CACHE | |
566 | }; | |
567 | EXPORT_SYMBOL(malloc_sizes); | |
568 | ||
569 | /* Must match cache_sizes above. Out of line to keep cache footprint low. */ | |
570 | struct cache_names { | |
571 | char *name; | |
572 | char *name_dma; | |
573 | }; | |
574 | ||
575 | static struct cache_names __initdata cache_names[] = { | |
576 | #define CACHE(x) { .name = "size-" #x, .name_dma = "size-" #x "(DMA)" }, | |
577 | #include <linux/kmalloc_sizes.h> | |
578 | {NULL,} | |
579 | #undef CACHE | |
580 | }; | |
581 | ||
582 | static struct arraycache_init initarray_cache __initdata = | |
583 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; | |
584 | static struct arraycache_init initarray_generic = | |
585 | { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} }; | |
586 | ||
587 | /* internal cache of cache description objs */ | |
588 | static struct kmem_list3 *cache_cache_nodelists[MAX_NUMNODES]; | |
589 | static struct kmem_cache cache_cache = { | |
590 | .nodelists = cache_cache_nodelists, | |
591 | .batchcount = 1, | |
592 | .limit = BOOT_CPUCACHE_ENTRIES, | |
593 | .shared = 1, | |
594 | .size = sizeof(struct kmem_cache), | |
595 | .name = "kmem_cache", | |
596 | }; | |
597 | ||
598 | #define BAD_ALIEN_MAGIC 0x01020304ul | |
599 | ||
600 | #ifdef CONFIG_LOCKDEP | |
601 | ||
602 | /* | |
603 | * Slab sometimes uses the kmalloc slabs to store the slab headers | |
604 | * for other slabs "off slab". | |
605 | * The locking for this is tricky in that it nests within the locks | |
606 | * of all other slabs in a few places; to deal with this special | |
607 | * locking we put on-slab caches into a separate lock-class. | |
608 | * | |
609 | * We set lock class for alien array caches which are up during init. | |
610 | * The lock annotation will be lost if all cpus of a node goes down and | |
611 | * then comes back up during hotplug | |
612 | */ | |
613 | static struct lock_class_key on_slab_l3_key; | |
614 | static struct lock_class_key on_slab_alc_key; | |
615 | ||
616 | static struct lock_class_key debugobj_l3_key; | |
617 | static struct lock_class_key debugobj_alc_key; | |
618 | ||
619 | static void slab_set_lock_classes(struct kmem_cache *cachep, | |
620 | struct lock_class_key *l3_key, struct lock_class_key *alc_key, | |
621 | int q) | |
622 | { | |
623 | struct array_cache **alc; | |
624 | struct kmem_list3 *l3; | |
625 | int r; | |
626 | ||
627 | l3 = cachep->nodelists[q]; | |
628 | if (!l3) | |
629 | return; | |
630 | ||
631 | lockdep_set_class(&l3->list_lock, l3_key); | |
632 | alc = l3->alien; | |
633 | /* | |
634 | * FIXME: This check for BAD_ALIEN_MAGIC | |
635 | * should go away when common slab code is taught to | |
636 | * work even without alien caches. | |
637 | * Currently, non NUMA code returns BAD_ALIEN_MAGIC | |
638 | * for alloc_alien_cache, | |
639 | */ | |
640 | if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC) | |
641 | return; | |
642 | for_each_node(r) { | |
643 | if (alc[r]) | |
644 | lockdep_set_class(&alc[r]->lock, alc_key); | |
645 | } | |
646 | } | |
647 | ||
648 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
649 | { | |
650 | slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node); | |
651 | } | |
652 | ||
653 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
654 | { | |
655 | int node; | |
656 | ||
657 | for_each_online_node(node) | |
658 | slab_set_debugobj_lock_classes_node(cachep, node); | |
659 | } | |
660 | ||
661 | static void init_node_lock_keys(int q) | |
662 | { | |
663 | struct cache_sizes *s = malloc_sizes; | |
664 | ||
665 | if (slab_state < UP) | |
666 | return; | |
667 | ||
668 | for (s = malloc_sizes; s->cs_size != ULONG_MAX; s++) { | |
669 | struct kmem_list3 *l3; | |
670 | ||
671 | l3 = s->cs_cachep->nodelists[q]; | |
672 | if (!l3 || OFF_SLAB(s->cs_cachep)) | |
673 | continue; | |
674 | ||
675 | slab_set_lock_classes(s->cs_cachep, &on_slab_l3_key, | |
676 | &on_slab_alc_key, q); | |
677 | } | |
678 | } | |
679 | ||
680 | static inline void init_lock_keys(void) | |
681 | { | |
682 | int node; | |
683 | ||
684 | for_each_node(node) | |
685 | init_node_lock_keys(node); | |
686 | } | |
687 | #else | |
688 | static void init_node_lock_keys(int q) | |
689 | { | |
690 | } | |
691 | ||
692 | static inline void init_lock_keys(void) | |
693 | { | |
694 | } | |
695 | ||
696 | static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node) | |
697 | { | |
698 | } | |
699 | ||
700 | static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep) | |
701 | { | |
702 | } | |
703 | #endif | |
704 | ||
705 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); | |
706 | ||
707 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) | |
708 | { | |
709 | return cachep->array[smp_processor_id()]; | |
710 | } | |
711 | ||
712 | static inline struct kmem_cache *__find_general_cachep(size_t size, | |
713 | gfp_t gfpflags) | |
714 | { | |
715 | struct cache_sizes *csizep = malloc_sizes; | |
716 | ||
717 | #if DEBUG | |
718 | /* This happens if someone tries to call | |
719 | * kmem_cache_create(), or __kmalloc(), before | |
720 | * the generic caches are initialized. | |
721 | */ | |
722 | BUG_ON(malloc_sizes[INDEX_AC].cs_cachep == NULL); | |
723 | #endif | |
724 | if (!size) | |
725 | return ZERO_SIZE_PTR; | |
726 | ||
727 | while (size > csizep->cs_size) | |
728 | csizep++; | |
729 | ||
730 | /* | |
731 | * Really subtle: The last entry with cs->cs_size==ULONG_MAX | |
732 | * has cs_{dma,}cachep==NULL. Thus no special case | |
733 | * for large kmalloc calls required. | |
734 | */ | |
735 | #ifdef CONFIG_ZONE_DMA | |
736 | if (unlikely(gfpflags & GFP_DMA)) | |
737 | return csizep->cs_dmacachep; | |
738 | #endif | |
739 | return csizep->cs_cachep; | |
740 | } | |
741 | ||
742 | static struct kmem_cache *kmem_find_general_cachep(size_t size, gfp_t gfpflags) | |
743 | { | |
744 | return __find_general_cachep(size, gfpflags); | |
745 | } | |
746 | ||
747 | static size_t slab_mgmt_size(size_t nr_objs, size_t align) | |
748 | { | |
749 | return ALIGN(sizeof(struct slab)+nr_objs*sizeof(kmem_bufctl_t), align); | |
750 | } | |
751 | ||
752 | /* | |
753 | * Calculate the number of objects and left-over bytes for a given buffer size. | |
754 | */ | |
755 | static void cache_estimate(unsigned long gfporder, size_t buffer_size, | |
756 | size_t align, int flags, size_t *left_over, | |
757 | unsigned int *num) | |
758 | { | |
759 | int nr_objs; | |
760 | size_t mgmt_size; | |
761 | size_t slab_size = PAGE_SIZE << gfporder; | |
762 | ||
763 | /* | |
764 | * The slab management structure can be either off the slab or | |
765 | * on it. For the latter case, the memory allocated for a | |
766 | * slab is used for: | |
767 | * | |
768 | * - The struct slab | |
769 | * - One kmem_bufctl_t for each object | |
770 | * - Padding to respect alignment of @align | |
771 | * - @buffer_size bytes for each object | |
772 | * | |
773 | * If the slab management structure is off the slab, then the | |
774 | * alignment will already be calculated into the size. Because | |
775 | * the slabs are all pages aligned, the objects will be at the | |
776 | * correct alignment when allocated. | |
777 | */ | |
778 | if (flags & CFLGS_OFF_SLAB) { | |
779 | mgmt_size = 0; | |
780 | nr_objs = slab_size / buffer_size; | |
781 | ||
782 | if (nr_objs > SLAB_LIMIT) | |
783 | nr_objs = SLAB_LIMIT; | |
784 | } else { | |
785 | /* | |
786 | * Ignore padding for the initial guess. The padding | |
787 | * is at most @align-1 bytes, and @buffer_size is at | |
788 | * least @align. In the worst case, this result will | |
789 | * be one greater than the number of objects that fit | |
790 | * into the memory allocation when taking the padding | |
791 | * into account. | |
792 | */ | |
793 | nr_objs = (slab_size - sizeof(struct slab)) / | |
794 | (buffer_size + sizeof(kmem_bufctl_t)); | |
795 | ||
796 | /* | |
797 | * This calculated number will be either the right | |
798 | * amount, or one greater than what we want. | |
799 | */ | |
800 | if (slab_mgmt_size(nr_objs, align) + nr_objs*buffer_size | |
801 | > slab_size) | |
802 | nr_objs--; | |
803 | ||
804 | if (nr_objs > SLAB_LIMIT) | |
805 | nr_objs = SLAB_LIMIT; | |
806 | ||
807 | mgmt_size = slab_mgmt_size(nr_objs, align); | |
808 | } | |
809 | *num = nr_objs; | |
810 | *left_over = slab_size - nr_objs*buffer_size - mgmt_size; | |
811 | } | |
812 | ||
813 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) | |
814 | ||
815 | static void __slab_error(const char *function, struct kmem_cache *cachep, | |
816 | char *msg) | |
817 | { | |
818 | printk(KERN_ERR "slab error in %s(): cache `%s': %s\n", | |
819 | function, cachep->name, msg); | |
820 | dump_stack(); | |
821 | } | |
822 | ||
823 | /* | |
824 | * By default on NUMA we use alien caches to stage the freeing of | |
825 | * objects allocated from other nodes. This causes massive memory | |
826 | * inefficiencies when using fake NUMA setup to split memory into a | |
827 | * large number of small nodes, so it can be disabled on the command | |
828 | * line | |
829 | */ | |
830 | ||
831 | static int use_alien_caches __read_mostly = 1; | |
832 | static int __init noaliencache_setup(char *s) | |
833 | { | |
834 | use_alien_caches = 0; | |
835 | return 1; | |
836 | } | |
837 | __setup("noaliencache", noaliencache_setup); | |
838 | ||
839 | static int __init slab_max_order_setup(char *str) | |
840 | { | |
841 | get_option(&str, &slab_max_order); | |
842 | slab_max_order = slab_max_order < 0 ? 0 : | |
843 | min(slab_max_order, MAX_ORDER - 1); | |
844 | slab_max_order_set = true; | |
845 | ||
846 | return 1; | |
847 | } | |
848 | __setup("slab_max_order=", slab_max_order_setup); | |
849 | ||
850 | #ifdef CONFIG_NUMA | |
851 | /* | |
852 | * Special reaping functions for NUMA systems called from cache_reap(). | |
853 | * These take care of doing round robin flushing of alien caches (containing | |
854 | * objects freed on different nodes from which they were allocated) and the | |
855 | * flushing of remote pcps by calling drain_node_pages. | |
856 | */ | |
857 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); | |
858 | ||
859 | static void init_reap_node(int cpu) | |
860 | { | |
861 | int node; | |
862 | ||
863 | node = next_node(cpu_to_mem(cpu), node_online_map); | |
864 | if (node == MAX_NUMNODES) | |
865 | node = first_node(node_online_map); | |
866 | ||
867 | per_cpu(slab_reap_node, cpu) = node; | |
868 | } | |
869 | ||
870 | static void next_reap_node(void) | |
871 | { | |
872 | int node = __this_cpu_read(slab_reap_node); | |
873 | ||
874 | node = next_node(node, node_online_map); | |
875 | if (unlikely(node >= MAX_NUMNODES)) | |
876 | node = first_node(node_online_map); | |
877 | __this_cpu_write(slab_reap_node, node); | |
878 | } | |
879 | ||
880 | #else | |
881 | #define init_reap_node(cpu) do { } while (0) | |
882 | #define next_reap_node(void) do { } while (0) | |
883 | #endif | |
884 | ||
885 | /* | |
886 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz | |
887 | * via the workqueue/eventd. | |
888 | * Add the CPU number into the expiration time to minimize the possibility of | |
889 | * the CPUs getting into lockstep and contending for the global cache chain | |
890 | * lock. | |
891 | */ | |
892 | static void __cpuinit start_cpu_timer(int cpu) | |
893 | { | |
894 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); | |
895 | ||
896 | /* | |
897 | * When this gets called from do_initcalls via cpucache_init(), | |
898 | * init_workqueues() has already run, so keventd will be setup | |
899 | * at that time. | |
900 | */ | |
901 | if (keventd_up() && reap_work->work.func == NULL) { | |
902 | init_reap_node(cpu); | |
903 | INIT_DELAYED_WORK_DEFERRABLE(reap_work, cache_reap); | |
904 | schedule_delayed_work_on(cpu, reap_work, | |
905 | __round_jiffies_relative(HZ, cpu)); | |
906 | } | |
907 | } | |
908 | ||
909 | static struct array_cache *alloc_arraycache(int node, int entries, | |
910 | int batchcount, gfp_t gfp) | |
911 | { | |
912 | int memsize = sizeof(void *) * entries + sizeof(struct array_cache); | |
913 | struct array_cache *nc = NULL; | |
914 | ||
915 | nc = kmalloc_node(memsize, gfp, node); | |
916 | /* | |
917 | * The array_cache structures contain pointers to free object. | |
918 | * However, when such objects are allocated or transferred to another | |
919 | * cache the pointers are not cleared and they could be counted as | |
920 | * valid references during a kmemleak scan. Therefore, kmemleak must | |
921 | * not scan such objects. | |
922 | */ | |
923 | kmemleak_no_scan(nc); | |
924 | if (nc) { | |
925 | nc->avail = 0; | |
926 | nc->limit = entries; | |
927 | nc->batchcount = batchcount; | |
928 | nc->touched = 0; | |
929 | spin_lock_init(&nc->lock); | |
930 | } | |
931 | return nc; | |
932 | } | |
933 | ||
934 | static inline bool is_slab_pfmemalloc(struct slab *slabp) | |
935 | { | |
936 | struct page *page = virt_to_page(slabp->s_mem); | |
937 | ||
938 | return PageSlabPfmemalloc(page); | |
939 | } | |
940 | ||
941 | /* Clears pfmemalloc_active if no slabs have pfmalloc set */ | |
942 | static void recheck_pfmemalloc_active(struct kmem_cache *cachep, | |
943 | struct array_cache *ac) | |
944 | { | |
945 | struct kmem_list3 *l3 = cachep->nodelists[numa_mem_id()]; | |
946 | struct slab *slabp; | |
947 | unsigned long flags; | |
948 | ||
949 | if (!pfmemalloc_active) | |
950 | return; | |
951 | ||
952 | spin_lock_irqsave(&l3->list_lock, flags); | |
953 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
954 | if (is_slab_pfmemalloc(slabp)) | |
955 | goto out; | |
956 | ||
957 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
958 | if (is_slab_pfmemalloc(slabp)) | |
959 | goto out; | |
960 | ||
961 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
962 | if (is_slab_pfmemalloc(slabp)) | |
963 | goto out; | |
964 | ||
965 | pfmemalloc_active = false; | |
966 | out: | |
967 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
968 | } | |
969 | ||
970 | static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
971 | gfp_t flags, bool force_refill) | |
972 | { | |
973 | int i; | |
974 | void *objp = ac->entry[--ac->avail]; | |
975 | ||
976 | /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */ | |
977 | if (unlikely(is_obj_pfmemalloc(objp))) { | |
978 | struct kmem_list3 *l3; | |
979 | ||
980 | if (gfp_pfmemalloc_allowed(flags)) { | |
981 | clear_obj_pfmemalloc(&objp); | |
982 | return objp; | |
983 | } | |
984 | ||
985 | /* The caller cannot use PFMEMALLOC objects, find another one */ | |
986 | for (i = 1; i < ac->avail; i++) { | |
987 | /* If a !PFMEMALLOC object is found, swap them */ | |
988 | if (!is_obj_pfmemalloc(ac->entry[i])) { | |
989 | objp = ac->entry[i]; | |
990 | ac->entry[i] = ac->entry[ac->avail]; | |
991 | ac->entry[ac->avail] = objp; | |
992 | return objp; | |
993 | } | |
994 | } | |
995 | ||
996 | /* | |
997 | * If there are empty slabs on the slabs_free list and we are | |
998 | * being forced to refill the cache, mark this one !pfmemalloc. | |
999 | */ | |
1000 | l3 = cachep->nodelists[numa_mem_id()]; | |
1001 | if (!list_empty(&l3->slabs_free) && force_refill) { | |
1002 | struct slab *slabp = virt_to_slab(objp); | |
1003 | ClearPageSlabPfmemalloc(virt_to_page(slabp->s_mem)); | |
1004 | clear_obj_pfmemalloc(&objp); | |
1005 | recheck_pfmemalloc_active(cachep, ac); | |
1006 | return objp; | |
1007 | } | |
1008 | ||
1009 | /* No !PFMEMALLOC objects available */ | |
1010 | ac->avail++; | |
1011 | objp = NULL; | |
1012 | } | |
1013 | ||
1014 | return objp; | |
1015 | } | |
1016 | ||
1017 | static inline void *ac_get_obj(struct kmem_cache *cachep, | |
1018 | struct array_cache *ac, gfp_t flags, bool force_refill) | |
1019 | { | |
1020 | void *objp; | |
1021 | ||
1022 | if (unlikely(sk_memalloc_socks())) | |
1023 | objp = __ac_get_obj(cachep, ac, flags, force_refill); | |
1024 | else | |
1025 | objp = ac->entry[--ac->avail]; | |
1026 | ||
1027 | return objp; | |
1028 | } | |
1029 | ||
1030 | static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
1031 | void *objp) | |
1032 | { | |
1033 | if (unlikely(pfmemalloc_active)) { | |
1034 | /* Some pfmemalloc slabs exist, check if this is one */ | |
1035 | struct page *page = virt_to_page(objp); | |
1036 | if (PageSlabPfmemalloc(page)) | |
1037 | set_obj_pfmemalloc(&objp); | |
1038 | } | |
1039 | ||
1040 | return objp; | |
1041 | } | |
1042 | ||
1043 | static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac, | |
1044 | void *objp) | |
1045 | { | |
1046 | if (unlikely(sk_memalloc_socks())) | |
1047 | objp = __ac_put_obj(cachep, ac, objp); | |
1048 | ||
1049 | ac->entry[ac->avail++] = objp; | |
1050 | } | |
1051 | ||
1052 | /* | |
1053 | * Transfer objects in one arraycache to another. | |
1054 | * Locking must be handled by the caller. | |
1055 | * | |
1056 | * Return the number of entries transferred. | |
1057 | */ | |
1058 | static int transfer_objects(struct array_cache *to, | |
1059 | struct array_cache *from, unsigned int max) | |
1060 | { | |
1061 | /* Figure out how many entries to transfer */ | |
1062 | int nr = min3(from->avail, max, to->limit - to->avail); | |
1063 | ||
1064 | if (!nr) | |
1065 | return 0; | |
1066 | ||
1067 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, | |
1068 | sizeof(void *) *nr); | |
1069 | ||
1070 | from->avail -= nr; | |
1071 | to->avail += nr; | |
1072 | return nr; | |
1073 | } | |
1074 | ||
1075 | #ifndef CONFIG_NUMA | |
1076 | ||
1077 | #define drain_alien_cache(cachep, alien) do { } while (0) | |
1078 | #define reap_alien(cachep, l3) do { } while (0) | |
1079 | ||
1080 | static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) | |
1081 | { | |
1082 | return (struct array_cache **)BAD_ALIEN_MAGIC; | |
1083 | } | |
1084 | ||
1085 | static inline void free_alien_cache(struct array_cache **ac_ptr) | |
1086 | { | |
1087 | } | |
1088 | ||
1089 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1090 | { | |
1091 | return 0; | |
1092 | } | |
1093 | ||
1094 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, | |
1095 | gfp_t flags) | |
1096 | { | |
1097 | return NULL; | |
1098 | } | |
1099 | ||
1100 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, | |
1101 | gfp_t flags, int nodeid) | |
1102 | { | |
1103 | return NULL; | |
1104 | } | |
1105 | ||
1106 | #else /* CONFIG_NUMA */ | |
1107 | ||
1108 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); | |
1109 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); | |
1110 | ||
1111 | static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) | |
1112 | { | |
1113 | struct array_cache **ac_ptr; | |
1114 | int memsize = sizeof(void *) * nr_node_ids; | |
1115 | int i; | |
1116 | ||
1117 | if (limit > 1) | |
1118 | limit = 12; | |
1119 | ac_ptr = kzalloc_node(memsize, gfp, node); | |
1120 | if (ac_ptr) { | |
1121 | for_each_node(i) { | |
1122 | if (i == node || !node_online(i)) | |
1123 | continue; | |
1124 | ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp); | |
1125 | if (!ac_ptr[i]) { | |
1126 | for (i--; i >= 0; i--) | |
1127 | kfree(ac_ptr[i]); | |
1128 | kfree(ac_ptr); | |
1129 | return NULL; | |
1130 | } | |
1131 | } | |
1132 | } | |
1133 | return ac_ptr; | |
1134 | } | |
1135 | ||
1136 | static void free_alien_cache(struct array_cache **ac_ptr) | |
1137 | { | |
1138 | int i; | |
1139 | ||
1140 | if (!ac_ptr) | |
1141 | return; | |
1142 | for_each_node(i) | |
1143 | kfree(ac_ptr[i]); | |
1144 | kfree(ac_ptr); | |
1145 | } | |
1146 | ||
1147 | static void __drain_alien_cache(struct kmem_cache *cachep, | |
1148 | struct array_cache *ac, int node) | |
1149 | { | |
1150 | struct kmem_list3 *rl3 = cachep->nodelists[node]; | |
1151 | ||
1152 | if (ac->avail) { | |
1153 | spin_lock(&rl3->list_lock); | |
1154 | /* | |
1155 | * Stuff objects into the remote nodes shared array first. | |
1156 | * That way we could avoid the overhead of putting the objects | |
1157 | * into the free lists and getting them back later. | |
1158 | */ | |
1159 | if (rl3->shared) | |
1160 | transfer_objects(rl3->shared, ac, ac->limit); | |
1161 | ||
1162 | free_block(cachep, ac->entry, ac->avail, node); | |
1163 | ac->avail = 0; | |
1164 | spin_unlock(&rl3->list_lock); | |
1165 | } | |
1166 | } | |
1167 | ||
1168 | /* | |
1169 | * Called from cache_reap() to regularly drain alien caches round robin. | |
1170 | */ | |
1171 | static void reap_alien(struct kmem_cache *cachep, struct kmem_list3 *l3) | |
1172 | { | |
1173 | int node = __this_cpu_read(slab_reap_node); | |
1174 | ||
1175 | if (l3->alien) { | |
1176 | struct array_cache *ac = l3->alien[node]; | |
1177 | ||
1178 | if (ac && ac->avail && spin_trylock_irq(&ac->lock)) { | |
1179 | __drain_alien_cache(cachep, ac, node); | |
1180 | spin_unlock_irq(&ac->lock); | |
1181 | } | |
1182 | } | |
1183 | } | |
1184 | ||
1185 | static void drain_alien_cache(struct kmem_cache *cachep, | |
1186 | struct array_cache **alien) | |
1187 | { | |
1188 | int i = 0; | |
1189 | struct array_cache *ac; | |
1190 | unsigned long flags; | |
1191 | ||
1192 | for_each_online_node(i) { | |
1193 | ac = alien[i]; | |
1194 | if (ac) { | |
1195 | spin_lock_irqsave(&ac->lock, flags); | |
1196 | __drain_alien_cache(cachep, ac, i); | |
1197 | spin_unlock_irqrestore(&ac->lock, flags); | |
1198 | } | |
1199 | } | |
1200 | } | |
1201 | ||
1202 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) | |
1203 | { | |
1204 | struct slab *slabp = virt_to_slab(objp); | |
1205 | int nodeid = slabp->nodeid; | |
1206 | struct kmem_list3 *l3; | |
1207 | struct array_cache *alien = NULL; | |
1208 | int node; | |
1209 | ||
1210 | node = numa_mem_id(); | |
1211 | ||
1212 | /* | |
1213 | * Make sure we are not freeing a object from another node to the array | |
1214 | * cache on this cpu. | |
1215 | */ | |
1216 | if (likely(slabp->nodeid == node)) | |
1217 | return 0; | |
1218 | ||
1219 | l3 = cachep->nodelists[node]; | |
1220 | STATS_INC_NODEFREES(cachep); | |
1221 | if (l3->alien && l3->alien[nodeid]) { | |
1222 | alien = l3->alien[nodeid]; | |
1223 | spin_lock(&alien->lock); | |
1224 | if (unlikely(alien->avail == alien->limit)) { | |
1225 | STATS_INC_ACOVERFLOW(cachep); | |
1226 | __drain_alien_cache(cachep, alien, nodeid); | |
1227 | } | |
1228 | ac_put_obj(cachep, alien, objp); | |
1229 | spin_unlock(&alien->lock); | |
1230 | } else { | |
1231 | spin_lock(&(cachep->nodelists[nodeid])->list_lock); | |
1232 | free_block(cachep, &objp, 1, nodeid); | |
1233 | spin_unlock(&(cachep->nodelists[nodeid])->list_lock); | |
1234 | } | |
1235 | return 1; | |
1236 | } | |
1237 | #endif | |
1238 | ||
1239 | /* | |
1240 | * Allocates and initializes nodelists for a node on each slab cache, used for | |
1241 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_list3 | |
1242 | * will be allocated off-node since memory is not yet online for the new node. | |
1243 | * When hotplugging memory or a cpu, existing nodelists are not replaced if | |
1244 | * already in use. | |
1245 | * | |
1246 | * Must hold slab_mutex. | |
1247 | */ | |
1248 | static int init_cache_nodelists_node(int node) | |
1249 | { | |
1250 | struct kmem_cache *cachep; | |
1251 | struct kmem_list3 *l3; | |
1252 | const int memsize = sizeof(struct kmem_list3); | |
1253 | ||
1254 | list_for_each_entry(cachep, &slab_caches, list) { | |
1255 | /* | |
1256 | * Set up the size64 kmemlist for cpu before we can | |
1257 | * begin anything. Make sure some other cpu on this | |
1258 | * node has not already allocated this | |
1259 | */ | |
1260 | if (!cachep->nodelists[node]) { | |
1261 | l3 = kmalloc_node(memsize, GFP_KERNEL, node); | |
1262 | if (!l3) | |
1263 | return -ENOMEM; | |
1264 | kmem_list3_init(l3); | |
1265 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
1266 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1267 | ||
1268 | /* | |
1269 | * The l3s don't come and go as CPUs come and | |
1270 | * go. slab_mutex is sufficient | |
1271 | * protection here. | |
1272 | */ | |
1273 | cachep->nodelists[node] = l3; | |
1274 | } | |
1275 | ||
1276 | spin_lock_irq(&cachep->nodelists[node]->list_lock); | |
1277 | cachep->nodelists[node]->free_limit = | |
1278 | (1 + nr_cpus_node(node)) * | |
1279 | cachep->batchcount + cachep->num; | |
1280 | spin_unlock_irq(&cachep->nodelists[node]->list_lock); | |
1281 | } | |
1282 | return 0; | |
1283 | } | |
1284 | ||
1285 | static void __cpuinit cpuup_canceled(long cpu) | |
1286 | { | |
1287 | struct kmem_cache *cachep; | |
1288 | struct kmem_list3 *l3 = NULL; | |
1289 | int node = cpu_to_mem(cpu); | |
1290 | const struct cpumask *mask = cpumask_of_node(node); | |
1291 | ||
1292 | list_for_each_entry(cachep, &slab_caches, list) { | |
1293 | struct array_cache *nc; | |
1294 | struct array_cache *shared; | |
1295 | struct array_cache **alien; | |
1296 | ||
1297 | /* cpu is dead; no one can alloc from it. */ | |
1298 | nc = cachep->array[cpu]; | |
1299 | cachep->array[cpu] = NULL; | |
1300 | l3 = cachep->nodelists[node]; | |
1301 | ||
1302 | if (!l3) | |
1303 | goto free_array_cache; | |
1304 | ||
1305 | spin_lock_irq(&l3->list_lock); | |
1306 | ||
1307 | /* Free limit for this kmem_list3 */ | |
1308 | l3->free_limit -= cachep->batchcount; | |
1309 | if (nc) | |
1310 | free_block(cachep, nc->entry, nc->avail, node); | |
1311 | ||
1312 | if (!cpumask_empty(mask)) { | |
1313 | spin_unlock_irq(&l3->list_lock); | |
1314 | goto free_array_cache; | |
1315 | } | |
1316 | ||
1317 | shared = l3->shared; | |
1318 | if (shared) { | |
1319 | free_block(cachep, shared->entry, | |
1320 | shared->avail, node); | |
1321 | l3->shared = NULL; | |
1322 | } | |
1323 | ||
1324 | alien = l3->alien; | |
1325 | l3->alien = NULL; | |
1326 | ||
1327 | spin_unlock_irq(&l3->list_lock); | |
1328 | ||
1329 | kfree(shared); | |
1330 | if (alien) { | |
1331 | drain_alien_cache(cachep, alien); | |
1332 | free_alien_cache(alien); | |
1333 | } | |
1334 | free_array_cache: | |
1335 | kfree(nc); | |
1336 | } | |
1337 | /* | |
1338 | * In the previous loop, all the objects were freed to | |
1339 | * the respective cache's slabs, now we can go ahead and | |
1340 | * shrink each nodelist to its limit. | |
1341 | */ | |
1342 | list_for_each_entry(cachep, &slab_caches, list) { | |
1343 | l3 = cachep->nodelists[node]; | |
1344 | if (!l3) | |
1345 | continue; | |
1346 | drain_freelist(cachep, l3, l3->free_objects); | |
1347 | } | |
1348 | } | |
1349 | ||
1350 | static int __cpuinit cpuup_prepare(long cpu) | |
1351 | { | |
1352 | struct kmem_cache *cachep; | |
1353 | struct kmem_list3 *l3 = NULL; | |
1354 | int node = cpu_to_mem(cpu); | |
1355 | int err; | |
1356 | ||
1357 | /* | |
1358 | * We need to do this right in the beginning since | |
1359 | * alloc_arraycache's are going to use this list. | |
1360 | * kmalloc_node allows us to add the slab to the right | |
1361 | * kmem_list3 and not this cpu's kmem_list3 | |
1362 | */ | |
1363 | err = init_cache_nodelists_node(node); | |
1364 | if (err < 0) | |
1365 | goto bad; | |
1366 | ||
1367 | /* | |
1368 | * Now we can go ahead with allocating the shared arrays and | |
1369 | * array caches | |
1370 | */ | |
1371 | list_for_each_entry(cachep, &slab_caches, list) { | |
1372 | struct array_cache *nc; | |
1373 | struct array_cache *shared = NULL; | |
1374 | struct array_cache **alien = NULL; | |
1375 | ||
1376 | nc = alloc_arraycache(node, cachep->limit, | |
1377 | cachep->batchcount, GFP_KERNEL); | |
1378 | if (!nc) | |
1379 | goto bad; | |
1380 | if (cachep->shared) { | |
1381 | shared = alloc_arraycache(node, | |
1382 | cachep->shared * cachep->batchcount, | |
1383 | 0xbaadf00d, GFP_KERNEL); | |
1384 | if (!shared) { | |
1385 | kfree(nc); | |
1386 | goto bad; | |
1387 | } | |
1388 | } | |
1389 | if (use_alien_caches) { | |
1390 | alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL); | |
1391 | if (!alien) { | |
1392 | kfree(shared); | |
1393 | kfree(nc); | |
1394 | goto bad; | |
1395 | } | |
1396 | } | |
1397 | cachep->array[cpu] = nc; | |
1398 | l3 = cachep->nodelists[node]; | |
1399 | BUG_ON(!l3); | |
1400 | ||
1401 | spin_lock_irq(&l3->list_lock); | |
1402 | if (!l3->shared) { | |
1403 | /* | |
1404 | * We are serialised from CPU_DEAD or | |
1405 | * CPU_UP_CANCELLED by the cpucontrol lock | |
1406 | */ | |
1407 | l3->shared = shared; | |
1408 | shared = NULL; | |
1409 | } | |
1410 | #ifdef CONFIG_NUMA | |
1411 | if (!l3->alien) { | |
1412 | l3->alien = alien; | |
1413 | alien = NULL; | |
1414 | } | |
1415 | #endif | |
1416 | spin_unlock_irq(&l3->list_lock); | |
1417 | kfree(shared); | |
1418 | free_alien_cache(alien); | |
1419 | if (cachep->flags & SLAB_DEBUG_OBJECTS) | |
1420 | slab_set_debugobj_lock_classes_node(cachep, node); | |
1421 | } | |
1422 | init_node_lock_keys(node); | |
1423 | ||
1424 | return 0; | |
1425 | bad: | |
1426 | cpuup_canceled(cpu); | |
1427 | return -ENOMEM; | |
1428 | } | |
1429 | ||
1430 | static int __cpuinit cpuup_callback(struct notifier_block *nfb, | |
1431 | unsigned long action, void *hcpu) | |
1432 | { | |
1433 | long cpu = (long)hcpu; | |
1434 | int err = 0; | |
1435 | ||
1436 | switch (action) { | |
1437 | case CPU_UP_PREPARE: | |
1438 | case CPU_UP_PREPARE_FROZEN: | |
1439 | mutex_lock(&slab_mutex); | |
1440 | err = cpuup_prepare(cpu); | |
1441 | mutex_unlock(&slab_mutex); | |
1442 | break; | |
1443 | case CPU_ONLINE: | |
1444 | case CPU_ONLINE_FROZEN: | |
1445 | start_cpu_timer(cpu); | |
1446 | break; | |
1447 | #ifdef CONFIG_HOTPLUG_CPU | |
1448 | case CPU_DOWN_PREPARE: | |
1449 | case CPU_DOWN_PREPARE_FROZEN: | |
1450 | /* | |
1451 | * Shutdown cache reaper. Note that the slab_mutex is | |
1452 | * held so that if cache_reap() is invoked it cannot do | |
1453 | * anything expensive but will only modify reap_work | |
1454 | * and reschedule the timer. | |
1455 | */ | |
1456 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); | |
1457 | /* Now the cache_reaper is guaranteed to be not running. */ | |
1458 | per_cpu(slab_reap_work, cpu).work.func = NULL; | |
1459 | break; | |
1460 | case CPU_DOWN_FAILED: | |
1461 | case CPU_DOWN_FAILED_FROZEN: | |
1462 | start_cpu_timer(cpu); | |
1463 | break; | |
1464 | case CPU_DEAD: | |
1465 | case CPU_DEAD_FROZEN: | |
1466 | /* | |
1467 | * Even if all the cpus of a node are down, we don't free the | |
1468 | * kmem_list3 of any cache. This to avoid a race between | |
1469 | * cpu_down, and a kmalloc allocation from another cpu for | |
1470 | * memory from the node of the cpu going down. The list3 | |
1471 | * structure is usually allocated from kmem_cache_create() and | |
1472 | * gets destroyed at kmem_cache_destroy(). | |
1473 | */ | |
1474 | /* fall through */ | |
1475 | #endif | |
1476 | case CPU_UP_CANCELED: | |
1477 | case CPU_UP_CANCELED_FROZEN: | |
1478 | mutex_lock(&slab_mutex); | |
1479 | cpuup_canceled(cpu); | |
1480 | mutex_unlock(&slab_mutex); | |
1481 | break; | |
1482 | } | |
1483 | return notifier_from_errno(err); | |
1484 | } | |
1485 | ||
1486 | static struct notifier_block __cpuinitdata cpucache_notifier = { | |
1487 | &cpuup_callback, NULL, 0 | |
1488 | }; | |
1489 | ||
1490 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) | |
1491 | /* | |
1492 | * Drains freelist for a node on each slab cache, used for memory hot-remove. | |
1493 | * Returns -EBUSY if all objects cannot be drained so that the node is not | |
1494 | * removed. | |
1495 | * | |
1496 | * Must hold slab_mutex. | |
1497 | */ | |
1498 | static int __meminit drain_cache_nodelists_node(int node) | |
1499 | { | |
1500 | struct kmem_cache *cachep; | |
1501 | int ret = 0; | |
1502 | ||
1503 | list_for_each_entry(cachep, &slab_caches, list) { | |
1504 | struct kmem_list3 *l3; | |
1505 | ||
1506 | l3 = cachep->nodelists[node]; | |
1507 | if (!l3) | |
1508 | continue; | |
1509 | ||
1510 | drain_freelist(cachep, l3, l3->free_objects); | |
1511 | ||
1512 | if (!list_empty(&l3->slabs_full) || | |
1513 | !list_empty(&l3->slabs_partial)) { | |
1514 | ret = -EBUSY; | |
1515 | break; | |
1516 | } | |
1517 | } | |
1518 | return ret; | |
1519 | } | |
1520 | ||
1521 | static int __meminit slab_memory_callback(struct notifier_block *self, | |
1522 | unsigned long action, void *arg) | |
1523 | { | |
1524 | struct memory_notify *mnb = arg; | |
1525 | int ret = 0; | |
1526 | int nid; | |
1527 | ||
1528 | nid = mnb->status_change_nid; | |
1529 | if (nid < 0) | |
1530 | goto out; | |
1531 | ||
1532 | switch (action) { | |
1533 | case MEM_GOING_ONLINE: | |
1534 | mutex_lock(&slab_mutex); | |
1535 | ret = init_cache_nodelists_node(nid); | |
1536 | mutex_unlock(&slab_mutex); | |
1537 | break; | |
1538 | case MEM_GOING_OFFLINE: | |
1539 | mutex_lock(&slab_mutex); | |
1540 | ret = drain_cache_nodelists_node(nid); | |
1541 | mutex_unlock(&slab_mutex); | |
1542 | break; | |
1543 | case MEM_ONLINE: | |
1544 | case MEM_OFFLINE: | |
1545 | case MEM_CANCEL_ONLINE: | |
1546 | case MEM_CANCEL_OFFLINE: | |
1547 | break; | |
1548 | } | |
1549 | out: | |
1550 | return notifier_from_errno(ret); | |
1551 | } | |
1552 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ | |
1553 | ||
1554 | /* | |
1555 | * swap the static kmem_list3 with kmalloced memory | |
1556 | */ | |
1557 | static void __init init_list(struct kmem_cache *cachep, struct kmem_list3 *list, | |
1558 | int nodeid) | |
1559 | { | |
1560 | struct kmem_list3 *ptr; | |
1561 | ||
1562 | ptr = kmalloc_node(sizeof(struct kmem_list3), GFP_NOWAIT, nodeid); | |
1563 | BUG_ON(!ptr); | |
1564 | ||
1565 | memcpy(ptr, list, sizeof(struct kmem_list3)); | |
1566 | /* | |
1567 | * Do not assume that spinlocks can be initialized via memcpy: | |
1568 | */ | |
1569 | spin_lock_init(&ptr->list_lock); | |
1570 | ||
1571 | MAKE_ALL_LISTS(cachep, ptr, nodeid); | |
1572 | cachep->nodelists[nodeid] = ptr; | |
1573 | } | |
1574 | ||
1575 | /* | |
1576 | * For setting up all the kmem_list3s for cache whose buffer_size is same as | |
1577 | * size of kmem_list3. | |
1578 | */ | |
1579 | static void __init set_up_list3s(struct kmem_cache *cachep, int index) | |
1580 | { | |
1581 | int node; | |
1582 | ||
1583 | for_each_online_node(node) { | |
1584 | cachep->nodelists[node] = &initkmem_list3[index + node]; | |
1585 | cachep->nodelists[node]->next_reap = jiffies + | |
1586 | REAPTIMEOUT_LIST3 + | |
1587 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
1588 | } | |
1589 | } | |
1590 | ||
1591 | /* | |
1592 | * Initialisation. Called after the page allocator have been initialised and | |
1593 | * before smp_init(). | |
1594 | */ | |
1595 | void __init kmem_cache_init(void) | |
1596 | { | |
1597 | size_t left_over; | |
1598 | struct cache_sizes *sizes; | |
1599 | struct cache_names *names; | |
1600 | int i; | |
1601 | int order; | |
1602 | int node; | |
1603 | ||
1604 | if (num_possible_nodes() == 1) | |
1605 | use_alien_caches = 0; | |
1606 | ||
1607 | for (i = 0; i < NUM_INIT_LISTS; i++) { | |
1608 | kmem_list3_init(&initkmem_list3[i]); | |
1609 | if (i < MAX_NUMNODES) | |
1610 | cache_cache.nodelists[i] = NULL; | |
1611 | } | |
1612 | set_up_list3s(&cache_cache, CACHE_CACHE); | |
1613 | ||
1614 | /* | |
1615 | * Fragmentation resistance on low memory - only use bigger | |
1616 | * page orders on machines with more than 32MB of memory if | |
1617 | * not overridden on the command line. | |
1618 | */ | |
1619 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) | |
1620 | slab_max_order = SLAB_MAX_ORDER_HI; | |
1621 | ||
1622 | /* Bootstrap is tricky, because several objects are allocated | |
1623 | * from caches that do not exist yet: | |
1624 | * 1) initialize the cache_cache cache: it contains the struct | |
1625 | * kmem_cache structures of all caches, except cache_cache itself: | |
1626 | * cache_cache is statically allocated. | |
1627 | * Initially an __init data area is used for the head array and the | |
1628 | * kmem_list3 structures, it's replaced with a kmalloc allocated | |
1629 | * array at the end of the bootstrap. | |
1630 | * 2) Create the first kmalloc cache. | |
1631 | * The struct kmem_cache for the new cache is allocated normally. | |
1632 | * An __init data area is used for the head array. | |
1633 | * 3) Create the remaining kmalloc caches, with minimally sized | |
1634 | * head arrays. | |
1635 | * 4) Replace the __init data head arrays for cache_cache and the first | |
1636 | * kmalloc cache with kmalloc allocated arrays. | |
1637 | * 5) Replace the __init data for kmem_list3 for cache_cache and | |
1638 | * the other cache's with kmalloc allocated memory. | |
1639 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. | |
1640 | */ | |
1641 | ||
1642 | node = numa_mem_id(); | |
1643 | ||
1644 | /* 1) create the cache_cache */ | |
1645 | INIT_LIST_HEAD(&slab_caches); | |
1646 | list_add(&cache_cache.list, &slab_caches); | |
1647 | cache_cache.colour_off = cache_line_size(); | |
1648 | cache_cache.array[smp_processor_id()] = &initarray_cache.cache; | |
1649 | cache_cache.nodelists[node] = &initkmem_list3[CACHE_CACHE + node]; | |
1650 | ||
1651 | /* | |
1652 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids | |
1653 | */ | |
1654 | cache_cache.size = offsetof(struct kmem_cache, array[nr_cpu_ids]) + | |
1655 | nr_node_ids * sizeof(struct kmem_list3 *); | |
1656 | cache_cache.object_size = cache_cache.size; | |
1657 | cache_cache.size = ALIGN(cache_cache.size, | |
1658 | cache_line_size()); | |
1659 | cache_cache.reciprocal_buffer_size = | |
1660 | reciprocal_value(cache_cache.size); | |
1661 | ||
1662 | for (order = 0; order < MAX_ORDER; order++) { | |
1663 | cache_estimate(order, cache_cache.size, | |
1664 | cache_line_size(), 0, &left_over, &cache_cache.num); | |
1665 | if (cache_cache.num) | |
1666 | break; | |
1667 | } | |
1668 | BUG_ON(!cache_cache.num); | |
1669 | cache_cache.gfporder = order; | |
1670 | cache_cache.colour = left_over / cache_cache.colour_off; | |
1671 | cache_cache.slab_size = ALIGN(cache_cache.num * sizeof(kmem_bufctl_t) + | |
1672 | sizeof(struct slab), cache_line_size()); | |
1673 | ||
1674 | /* 2+3) create the kmalloc caches */ | |
1675 | sizes = malloc_sizes; | |
1676 | names = cache_names; | |
1677 | ||
1678 | /* | |
1679 | * Initialize the caches that provide memory for the array cache and the | |
1680 | * kmem_list3 structures first. Without this, further allocations will | |
1681 | * bug. | |
1682 | */ | |
1683 | ||
1684 | sizes[INDEX_AC].cs_cachep = __kmem_cache_create(names[INDEX_AC].name, | |
1685 | sizes[INDEX_AC].cs_size, | |
1686 | ARCH_KMALLOC_MINALIGN, | |
1687 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1688 | NULL); | |
1689 | ||
1690 | if (INDEX_AC != INDEX_L3) { | |
1691 | sizes[INDEX_L3].cs_cachep = | |
1692 | __kmem_cache_create(names[INDEX_L3].name, | |
1693 | sizes[INDEX_L3].cs_size, | |
1694 | ARCH_KMALLOC_MINALIGN, | |
1695 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1696 | NULL); | |
1697 | } | |
1698 | ||
1699 | slab_early_init = 0; | |
1700 | ||
1701 | while (sizes->cs_size != ULONG_MAX) { | |
1702 | /* | |
1703 | * For performance, all the general caches are L1 aligned. | |
1704 | * This should be particularly beneficial on SMP boxes, as it | |
1705 | * eliminates "false sharing". | |
1706 | * Note for systems short on memory removing the alignment will | |
1707 | * allow tighter packing of the smaller caches. | |
1708 | */ | |
1709 | if (!sizes->cs_cachep) { | |
1710 | sizes->cs_cachep = __kmem_cache_create(names->name, | |
1711 | sizes->cs_size, | |
1712 | ARCH_KMALLOC_MINALIGN, | |
1713 | ARCH_KMALLOC_FLAGS|SLAB_PANIC, | |
1714 | NULL); | |
1715 | } | |
1716 | #ifdef CONFIG_ZONE_DMA | |
1717 | sizes->cs_dmacachep = __kmem_cache_create( | |
1718 | names->name_dma, | |
1719 | sizes->cs_size, | |
1720 | ARCH_KMALLOC_MINALIGN, | |
1721 | ARCH_KMALLOC_FLAGS|SLAB_CACHE_DMA| | |
1722 | SLAB_PANIC, | |
1723 | NULL); | |
1724 | #endif | |
1725 | sizes++; | |
1726 | names++; | |
1727 | } | |
1728 | /* 4) Replace the bootstrap head arrays */ | |
1729 | { | |
1730 | struct array_cache *ptr; | |
1731 | ||
1732 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); | |
1733 | ||
1734 | BUG_ON(cpu_cache_get(&cache_cache) != &initarray_cache.cache); | |
1735 | memcpy(ptr, cpu_cache_get(&cache_cache), | |
1736 | sizeof(struct arraycache_init)); | |
1737 | /* | |
1738 | * Do not assume that spinlocks can be initialized via memcpy: | |
1739 | */ | |
1740 | spin_lock_init(&ptr->lock); | |
1741 | ||
1742 | cache_cache.array[smp_processor_id()] = ptr; | |
1743 | ||
1744 | ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT); | |
1745 | ||
1746 | BUG_ON(cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep) | |
1747 | != &initarray_generic.cache); | |
1748 | memcpy(ptr, cpu_cache_get(malloc_sizes[INDEX_AC].cs_cachep), | |
1749 | sizeof(struct arraycache_init)); | |
1750 | /* | |
1751 | * Do not assume that spinlocks can be initialized via memcpy: | |
1752 | */ | |
1753 | spin_lock_init(&ptr->lock); | |
1754 | ||
1755 | malloc_sizes[INDEX_AC].cs_cachep->array[smp_processor_id()] = | |
1756 | ptr; | |
1757 | } | |
1758 | /* 5) Replace the bootstrap kmem_list3's */ | |
1759 | { | |
1760 | int nid; | |
1761 | ||
1762 | for_each_online_node(nid) { | |
1763 | init_list(&cache_cache, &initkmem_list3[CACHE_CACHE + nid], nid); | |
1764 | ||
1765 | init_list(malloc_sizes[INDEX_AC].cs_cachep, | |
1766 | &initkmem_list3[SIZE_AC + nid], nid); | |
1767 | ||
1768 | if (INDEX_AC != INDEX_L3) { | |
1769 | init_list(malloc_sizes[INDEX_L3].cs_cachep, | |
1770 | &initkmem_list3[SIZE_L3 + nid], nid); | |
1771 | } | |
1772 | } | |
1773 | } | |
1774 | ||
1775 | slab_state = UP; | |
1776 | } | |
1777 | ||
1778 | void __init kmem_cache_init_late(void) | |
1779 | { | |
1780 | struct kmem_cache *cachep; | |
1781 | ||
1782 | slab_state = UP; | |
1783 | ||
1784 | /* Annotate slab for lockdep -- annotate the malloc caches */ | |
1785 | init_lock_keys(); | |
1786 | ||
1787 | /* 6) resize the head arrays to their final sizes */ | |
1788 | mutex_lock(&slab_mutex); | |
1789 | list_for_each_entry(cachep, &slab_caches, list) | |
1790 | if (enable_cpucache(cachep, GFP_NOWAIT)) | |
1791 | BUG(); | |
1792 | mutex_unlock(&slab_mutex); | |
1793 | ||
1794 | /* Done! */ | |
1795 | slab_state = FULL; | |
1796 | ||
1797 | /* | |
1798 | * Register a cpu startup notifier callback that initializes | |
1799 | * cpu_cache_get for all new cpus | |
1800 | */ | |
1801 | register_cpu_notifier(&cpucache_notifier); | |
1802 | ||
1803 | #ifdef CONFIG_NUMA | |
1804 | /* | |
1805 | * Register a memory hotplug callback that initializes and frees | |
1806 | * nodelists. | |
1807 | */ | |
1808 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); | |
1809 | #endif | |
1810 | ||
1811 | /* | |
1812 | * The reap timers are started later, with a module init call: That part | |
1813 | * of the kernel is not yet operational. | |
1814 | */ | |
1815 | } | |
1816 | ||
1817 | static int __init cpucache_init(void) | |
1818 | { | |
1819 | int cpu; | |
1820 | ||
1821 | /* | |
1822 | * Register the timers that return unneeded pages to the page allocator | |
1823 | */ | |
1824 | for_each_online_cpu(cpu) | |
1825 | start_cpu_timer(cpu); | |
1826 | ||
1827 | /* Done! */ | |
1828 | slab_state = FULL; | |
1829 | return 0; | |
1830 | } | |
1831 | __initcall(cpucache_init); | |
1832 | ||
1833 | static noinline void | |
1834 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) | |
1835 | { | |
1836 | struct kmem_list3 *l3; | |
1837 | struct slab *slabp; | |
1838 | unsigned long flags; | |
1839 | int node; | |
1840 | ||
1841 | printk(KERN_WARNING | |
1842 | "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n", | |
1843 | nodeid, gfpflags); | |
1844 | printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n", | |
1845 | cachep->name, cachep->size, cachep->gfporder); | |
1846 | ||
1847 | for_each_online_node(node) { | |
1848 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; | |
1849 | unsigned long active_slabs = 0, num_slabs = 0; | |
1850 | ||
1851 | l3 = cachep->nodelists[node]; | |
1852 | if (!l3) | |
1853 | continue; | |
1854 | ||
1855 | spin_lock_irqsave(&l3->list_lock, flags); | |
1856 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
1857 | active_objs += cachep->num; | |
1858 | active_slabs++; | |
1859 | } | |
1860 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
1861 | active_objs += slabp->inuse; | |
1862 | active_slabs++; | |
1863 | } | |
1864 | list_for_each_entry(slabp, &l3->slabs_free, list) | |
1865 | num_slabs++; | |
1866 | ||
1867 | free_objects += l3->free_objects; | |
1868 | spin_unlock_irqrestore(&l3->list_lock, flags); | |
1869 | ||
1870 | num_slabs += active_slabs; | |
1871 | num_objs = num_slabs * cachep->num; | |
1872 | printk(KERN_WARNING | |
1873 | " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", | |
1874 | node, active_slabs, num_slabs, active_objs, num_objs, | |
1875 | free_objects); | |
1876 | } | |
1877 | } | |
1878 | ||
1879 | /* | |
1880 | * Interface to system's page allocator. No need to hold the cache-lock. | |
1881 | * | |
1882 | * If we requested dmaable memory, we will get it. Even if we | |
1883 | * did not request dmaable memory, we might get it, but that | |
1884 | * would be relatively rare and ignorable. | |
1885 | */ | |
1886 | static void *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, int nodeid) | |
1887 | { | |
1888 | struct page *page; | |
1889 | int nr_pages; | |
1890 | int i; | |
1891 | ||
1892 | #ifndef CONFIG_MMU | |
1893 | /* | |
1894 | * Nommu uses slab's for process anonymous memory allocations, and thus | |
1895 | * requires __GFP_COMP to properly refcount higher order allocations | |
1896 | */ | |
1897 | flags |= __GFP_COMP; | |
1898 | #endif | |
1899 | ||
1900 | flags |= cachep->allocflags; | |
1901 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1902 | flags |= __GFP_RECLAIMABLE; | |
1903 | ||
1904 | page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); | |
1905 | if (!page) { | |
1906 | if (!(flags & __GFP_NOWARN) && printk_ratelimit()) | |
1907 | slab_out_of_memory(cachep, flags, nodeid); | |
1908 | return NULL; | |
1909 | } | |
1910 | ||
1911 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ | |
1912 | if (unlikely(page->pfmemalloc)) | |
1913 | pfmemalloc_active = true; | |
1914 | ||
1915 | nr_pages = (1 << cachep->gfporder); | |
1916 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1917 | add_zone_page_state(page_zone(page), | |
1918 | NR_SLAB_RECLAIMABLE, nr_pages); | |
1919 | else | |
1920 | add_zone_page_state(page_zone(page), | |
1921 | NR_SLAB_UNRECLAIMABLE, nr_pages); | |
1922 | for (i = 0; i < nr_pages; i++) { | |
1923 | __SetPageSlab(page + i); | |
1924 | ||
1925 | if (page->pfmemalloc) | |
1926 | SetPageSlabPfmemalloc(page + i); | |
1927 | } | |
1928 | ||
1929 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { | |
1930 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); | |
1931 | ||
1932 | if (cachep->ctor) | |
1933 | kmemcheck_mark_uninitialized_pages(page, nr_pages); | |
1934 | else | |
1935 | kmemcheck_mark_unallocated_pages(page, nr_pages); | |
1936 | } | |
1937 | ||
1938 | return page_address(page); | |
1939 | } | |
1940 | ||
1941 | /* | |
1942 | * Interface to system's page release. | |
1943 | */ | |
1944 | static void kmem_freepages(struct kmem_cache *cachep, void *addr) | |
1945 | { | |
1946 | unsigned long i = (1 << cachep->gfporder); | |
1947 | struct page *page = virt_to_page(addr); | |
1948 | const unsigned long nr_freed = i; | |
1949 | ||
1950 | kmemcheck_free_shadow(page, cachep->gfporder); | |
1951 | ||
1952 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) | |
1953 | sub_zone_page_state(page_zone(page), | |
1954 | NR_SLAB_RECLAIMABLE, nr_freed); | |
1955 | else | |
1956 | sub_zone_page_state(page_zone(page), | |
1957 | NR_SLAB_UNRECLAIMABLE, nr_freed); | |
1958 | while (i--) { | |
1959 | BUG_ON(!PageSlab(page)); | |
1960 | __ClearPageSlabPfmemalloc(page); | |
1961 | __ClearPageSlab(page); | |
1962 | page++; | |
1963 | } | |
1964 | if (current->reclaim_state) | |
1965 | current->reclaim_state->reclaimed_slab += nr_freed; | |
1966 | free_pages((unsigned long)addr, cachep->gfporder); | |
1967 | } | |
1968 | ||
1969 | static void kmem_rcu_free(struct rcu_head *head) | |
1970 | { | |
1971 | struct slab_rcu *slab_rcu = (struct slab_rcu *)head; | |
1972 | struct kmem_cache *cachep = slab_rcu->cachep; | |
1973 | ||
1974 | kmem_freepages(cachep, slab_rcu->addr); | |
1975 | if (OFF_SLAB(cachep)) | |
1976 | kmem_cache_free(cachep->slabp_cache, slab_rcu); | |
1977 | } | |
1978 | ||
1979 | #if DEBUG | |
1980 | ||
1981 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
1982 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, | |
1983 | unsigned long caller) | |
1984 | { | |
1985 | int size = cachep->object_size; | |
1986 | ||
1987 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; | |
1988 | ||
1989 | if (size < 5 * sizeof(unsigned long)) | |
1990 | return; | |
1991 | ||
1992 | *addr++ = 0x12345678; | |
1993 | *addr++ = caller; | |
1994 | *addr++ = smp_processor_id(); | |
1995 | size -= 3 * sizeof(unsigned long); | |
1996 | { | |
1997 | unsigned long *sptr = &caller; | |
1998 | unsigned long svalue; | |
1999 | ||
2000 | while (!kstack_end(sptr)) { | |
2001 | svalue = *sptr++; | |
2002 | if (kernel_text_address(svalue)) { | |
2003 | *addr++ = svalue; | |
2004 | size -= sizeof(unsigned long); | |
2005 | if (size <= sizeof(unsigned long)) | |
2006 | break; | |
2007 | } | |
2008 | } | |
2009 | ||
2010 | } | |
2011 | *addr++ = 0x87654321; | |
2012 | } | |
2013 | #endif | |
2014 | ||
2015 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) | |
2016 | { | |
2017 | int size = cachep->object_size; | |
2018 | addr = &((char *)addr)[obj_offset(cachep)]; | |
2019 | ||
2020 | memset(addr, val, size); | |
2021 | *(unsigned char *)(addr + size - 1) = POISON_END; | |
2022 | } | |
2023 | ||
2024 | static void dump_line(char *data, int offset, int limit) | |
2025 | { | |
2026 | int i; | |
2027 | unsigned char error = 0; | |
2028 | int bad_count = 0; | |
2029 | ||
2030 | printk(KERN_ERR "%03x: ", offset); | |
2031 | for (i = 0; i < limit; i++) { | |
2032 | if (data[offset + i] != POISON_FREE) { | |
2033 | error = data[offset + i]; | |
2034 | bad_count++; | |
2035 | } | |
2036 | } | |
2037 | print_hex_dump(KERN_CONT, "", 0, 16, 1, | |
2038 | &data[offset], limit, 1); | |
2039 | ||
2040 | if (bad_count == 1) { | |
2041 | error ^= POISON_FREE; | |
2042 | if (!(error & (error - 1))) { | |
2043 | printk(KERN_ERR "Single bit error detected. Probably " | |
2044 | "bad RAM.\n"); | |
2045 | #ifdef CONFIG_X86 | |
2046 | printk(KERN_ERR "Run memtest86+ or a similar memory " | |
2047 | "test tool.\n"); | |
2048 | #else | |
2049 | printk(KERN_ERR "Run a memory test tool.\n"); | |
2050 | #endif | |
2051 | } | |
2052 | } | |
2053 | } | |
2054 | #endif | |
2055 | ||
2056 | #if DEBUG | |
2057 | ||
2058 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) | |
2059 | { | |
2060 | int i, size; | |
2061 | char *realobj; | |
2062 | ||
2063 | if (cachep->flags & SLAB_RED_ZONE) { | |
2064 | printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n", | |
2065 | *dbg_redzone1(cachep, objp), | |
2066 | *dbg_redzone2(cachep, objp)); | |
2067 | } | |
2068 | ||
2069 | if (cachep->flags & SLAB_STORE_USER) { | |
2070 | printk(KERN_ERR "Last user: [<%p>]", | |
2071 | *dbg_userword(cachep, objp)); | |
2072 | print_symbol("(%s)", | |
2073 | (unsigned long)*dbg_userword(cachep, objp)); | |
2074 | printk("\n"); | |
2075 | } | |
2076 | realobj = (char *)objp + obj_offset(cachep); | |
2077 | size = cachep->object_size; | |
2078 | for (i = 0; i < size && lines; i += 16, lines--) { | |
2079 | int limit; | |
2080 | limit = 16; | |
2081 | if (i + limit > size) | |
2082 | limit = size - i; | |
2083 | dump_line(realobj, i, limit); | |
2084 | } | |
2085 | } | |
2086 | ||
2087 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) | |
2088 | { | |
2089 | char *realobj; | |
2090 | int size, i; | |
2091 | int lines = 0; | |
2092 | ||
2093 | realobj = (char *)objp + obj_offset(cachep); | |
2094 | size = cachep->object_size; | |
2095 | ||
2096 | for (i = 0; i < size; i++) { | |
2097 | char exp = POISON_FREE; | |
2098 | if (i == size - 1) | |
2099 | exp = POISON_END; | |
2100 | if (realobj[i] != exp) { | |
2101 | int limit; | |
2102 | /* Mismatch ! */ | |
2103 | /* Print header */ | |
2104 | if (lines == 0) { | |
2105 | printk(KERN_ERR | |
2106 | "Slab corruption (%s): %s start=%p, len=%d\n", | |
2107 | print_tainted(), cachep->name, realobj, size); | |
2108 | print_objinfo(cachep, objp, 0); | |
2109 | } | |
2110 | /* Hexdump the affected line */ | |
2111 | i = (i / 16) * 16; | |
2112 | limit = 16; | |
2113 | if (i + limit > size) | |
2114 | limit = size - i; | |
2115 | dump_line(realobj, i, limit); | |
2116 | i += 16; | |
2117 | lines++; | |
2118 | /* Limit to 5 lines */ | |
2119 | if (lines > 5) | |
2120 | break; | |
2121 | } | |
2122 | } | |
2123 | if (lines != 0) { | |
2124 | /* Print some data about the neighboring objects, if they | |
2125 | * exist: | |
2126 | */ | |
2127 | struct slab *slabp = virt_to_slab(objp); | |
2128 | unsigned int objnr; | |
2129 | ||
2130 | objnr = obj_to_index(cachep, slabp, objp); | |
2131 | if (objnr) { | |
2132 | objp = index_to_obj(cachep, slabp, objnr - 1); | |
2133 | realobj = (char *)objp + obj_offset(cachep); | |
2134 | printk(KERN_ERR "Prev obj: start=%p, len=%d\n", | |
2135 | realobj, size); | |
2136 | print_objinfo(cachep, objp, 2); | |
2137 | } | |
2138 | if (objnr + 1 < cachep->num) { | |
2139 | objp = index_to_obj(cachep, slabp, objnr + 1); | |
2140 | realobj = (char *)objp + obj_offset(cachep); | |
2141 | printk(KERN_ERR "Next obj: start=%p, len=%d\n", | |
2142 | realobj, size); | |
2143 | print_objinfo(cachep, objp, 2); | |
2144 | } | |
2145 | } | |
2146 | } | |
2147 | #endif | |
2148 | ||
2149 | #if DEBUG | |
2150 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) | |
2151 | { | |
2152 | int i; | |
2153 | for (i = 0; i < cachep->num; i++) { | |
2154 | void *objp = index_to_obj(cachep, slabp, i); | |
2155 | ||
2156 | if (cachep->flags & SLAB_POISON) { | |
2157 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
2158 | if (cachep->size % PAGE_SIZE == 0 && | |
2159 | OFF_SLAB(cachep)) | |
2160 | kernel_map_pages(virt_to_page(objp), | |
2161 | cachep->size / PAGE_SIZE, 1); | |
2162 | else | |
2163 | check_poison_obj(cachep, objp); | |
2164 | #else | |
2165 | check_poison_obj(cachep, objp); | |
2166 | #endif | |
2167 | } | |
2168 | if (cachep->flags & SLAB_RED_ZONE) { | |
2169 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2170 | slab_error(cachep, "start of a freed object " | |
2171 | "was overwritten"); | |
2172 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2173 | slab_error(cachep, "end of a freed object " | |
2174 | "was overwritten"); | |
2175 | } | |
2176 | } | |
2177 | } | |
2178 | #else | |
2179 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, struct slab *slabp) | |
2180 | { | |
2181 | } | |
2182 | #endif | |
2183 | ||
2184 | /** | |
2185 | * slab_destroy - destroy and release all objects in a slab | |
2186 | * @cachep: cache pointer being destroyed | |
2187 | * @slabp: slab pointer being destroyed | |
2188 | * | |
2189 | * Destroy all the objs in a slab, and release the mem back to the system. | |
2190 | * Before calling the slab must have been unlinked from the cache. The | |
2191 | * cache-lock is not held/needed. | |
2192 | */ | |
2193 | static void slab_destroy(struct kmem_cache *cachep, struct slab *slabp) | |
2194 | { | |
2195 | void *addr = slabp->s_mem - slabp->colouroff; | |
2196 | ||
2197 | slab_destroy_debugcheck(cachep, slabp); | |
2198 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) { | |
2199 | struct slab_rcu *slab_rcu; | |
2200 | ||
2201 | slab_rcu = (struct slab_rcu *)slabp; | |
2202 | slab_rcu->cachep = cachep; | |
2203 | slab_rcu->addr = addr; | |
2204 | call_rcu(&slab_rcu->head, kmem_rcu_free); | |
2205 | } else { | |
2206 | kmem_freepages(cachep, addr); | |
2207 | if (OFF_SLAB(cachep)) | |
2208 | kmem_cache_free(cachep->slabp_cache, slabp); | |
2209 | } | |
2210 | } | |
2211 | ||
2212 | static void __kmem_cache_destroy(struct kmem_cache *cachep) | |
2213 | { | |
2214 | int i; | |
2215 | struct kmem_list3 *l3; | |
2216 | ||
2217 | for_each_online_cpu(i) | |
2218 | kfree(cachep->array[i]); | |
2219 | ||
2220 | /* NUMA: free the list3 structures */ | |
2221 | for_each_online_node(i) { | |
2222 | l3 = cachep->nodelists[i]; | |
2223 | if (l3) { | |
2224 | kfree(l3->shared); | |
2225 | free_alien_cache(l3->alien); | |
2226 | kfree(l3); | |
2227 | } | |
2228 | } | |
2229 | kmem_cache_free(&cache_cache, cachep); | |
2230 | } | |
2231 | ||
2232 | ||
2233 | /** | |
2234 | * calculate_slab_order - calculate size (page order) of slabs | |
2235 | * @cachep: pointer to the cache that is being created | |
2236 | * @size: size of objects to be created in this cache. | |
2237 | * @align: required alignment for the objects. | |
2238 | * @flags: slab allocation flags | |
2239 | * | |
2240 | * Also calculates the number of objects per slab. | |
2241 | * | |
2242 | * This could be made much more intelligent. For now, try to avoid using | |
2243 | * high order pages for slabs. When the gfp() functions are more friendly | |
2244 | * towards high-order requests, this should be changed. | |
2245 | */ | |
2246 | static size_t calculate_slab_order(struct kmem_cache *cachep, | |
2247 | size_t size, size_t align, unsigned long flags) | |
2248 | { | |
2249 | unsigned long offslab_limit; | |
2250 | size_t left_over = 0; | |
2251 | int gfporder; | |
2252 | ||
2253 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { | |
2254 | unsigned int num; | |
2255 | size_t remainder; | |
2256 | ||
2257 | cache_estimate(gfporder, size, align, flags, &remainder, &num); | |
2258 | if (!num) | |
2259 | continue; | |
2260 | ||
2261 | if (flags & CFLGS_OFF_SLAB) { | |
2262 | /* | |
2263 | * Max number of objs-per-slab for caches which | |
2264 | * use off-slab slabs. Needed to avoid a possible | |
2265 | * looping condition in cache_grow(). | |
2266 | */ | |
2267 | offslab_limit = size - sizeof(struct slab); | |
2268 | offslab_limit /= sizeof(kmem_bufctl_t); | |
2269 | ||
2270 | if (num > offslab_limit) | |
2271 | break; | |
2272 | } | |
2273 | ||
2274 | /* Found something acceptable - save it away */ | |
2275 | cachep->num = num; | |
2276 | cachep->gfporder = gfporder; | |
2277 | left_over = remainder; | |
2278 | ||
2279 | /* | |
2280 | * A VFS-reclaimable slab tends to have most allocations | |
2281 | * as GFP_NOFS and we really don't want to have to be allocating | |
2282 | * higher-order pages when we are unable to shrink dcache. | |
2283 | */ | |
2284 | if (flags & SLAB_RECLAIM_ACCOUNT) | |
2285 | break; | |
2286 | ||
2287 | /* | |
2288 | * Large number of objects is good, but very large slabs are | |
2289 | * currently bad for the gfp()s. | |
2290 | */ | |
2291 | if (gfporder >= slab_max_order) | |
2292 | break; | |
2293 | ||
2294 | /* | |
2295 | * Acceptable internal fragmentation? | |
2296 | */ | |
2297 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) | |
2298 | break; | |
2299 | } | |
2300 | return left_over; | |
2301 | } | |
2302 | ||
2303 | static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) | |
2304 | { | |
2305 | if (slab_state >= FULL) | |
2306 | return enable_cpucache(cachep, gfp); | |
2307 | ||
2308 | if (slab_state == DOWN) { | |
2309 | /* | |
2310 | * Note: the first kmem_cache_create must create the cache | |
2311 | * that's used by kmalloc(24), otherwise the creation of | |
2312 | * further caches will BUG(). | |
2313 | */ | |
2314 | cachep->array[smp_processor_id()] = &initarray_generic.cache; | |
2315 | ||
2316 | /* | |
2317 | * If the cache that's used by kmalloc(sizeof(kmem_list3)) is | |
2318 | * the first cache, then we need to set up all its list3s, | |
2319 | * otherwise the creation of further caches will BUG(). | |
2320 | */ | |
2321 | set_up_list3s(cachep, SIZE_AC); | |
2322 | if (INDEX_AC == INDEX_L3) | |
2323 | slab_state = PARTIAL_L3; | |
2324 | else | |
2325 | slab_state = PARTIAL_ARRAYCACHE; | |
2326 | } else { | |
2327 | cachep->array[smp_processor_id()] = | |
2328 | kmalloc(sizeof(struct arraycache_init), gfp); | |
2329 | ||
2330 | if (slab_state == PARTIAL_ARRAYCACHE) { | |
2331 | set_up_list3s(cachep, SIZE_L3); | |
2332 | slab_state = PARTIAL_L3; | |
2333 | } else { | |
2334 | int node; | |
2335 | for_each_online_node(node) { | |
2336 | cachep->nodelists[node] = | |
2337 | kmalloc_node(sizeof(struct kmem_list3), | |
2338 | gfp, node); | |
2339 | BUG_ON(!cachep->nodelists[node]); | |
2340 | kmem_list3_init(cachep->nodelists[node]); | |
2341 | } | |
2342 | } | |
2343 | } | |
2344 | cachep->nodelists[numa_mem_id()]->next_reap = | |
2345 | jiffies + REAPTIMEOUT_LIST3 + | |
2346 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
2347 | ||
2348 | cpu_cache_get(cachep)->avail = 0; | |
2349 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; | |
2350 | cpu_cache_get(cachep)->batchcount = 1; | |
2351 | cpu_cache_get(cachep)->touched = 0; | |
2352 | cachep->batchcount = 1; | |
2353 | cachep->limit = BOOT_CPUCACHE_ENTRIES; | |
2354 | return 0; | |
2355 | } | |
2356 | ||
2357 | /** | |
2358 | * __kmem_cache_create - Create a cache. | |
2359 | * @name: A string which is used in /proc/slabinfo to identify this cache. | |
2360 | * @size: The size of objects to be created in this cache. | |
2361 | * @align: The required alignment for the objects. | |
2362 | * @flags: SLAB flags | |
2363 | * @ctor: A constructor for the objects. | |
2364 | * | |
2365 | * Returns a ptr to the cache on success, NULL on failure. | |
2366 | * Cannot be called within a int, but can be interrupted. | |
2367 | * The @ctor is run when new pages are allocated by the cache. | |
2368 | * | |
2369 | * @name must be valid until the cache is destroyed. This implies that | |
2370 | * the module calling this has to destroy the cache before getting unloaded. | |
2371 | * | |
2372 | * The flags are | |
2373 | * | |
2374 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) | |
2375 | * to catch references to uninitialised memory. | |
2376 | * | |
2377 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check | |
2378 | * for buffer overruns. | |
2379 | * | |
2380 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware | |
2381 | * cacheline. This can be beneficial if you're counting cycles as closely | |
2382 | * as davem. | |
2383 | */ | |
2384 | struct kmem_cache * | |
2385 | __kmem_cache_create (const char *name, size_t size, size_t align, | |
2386 | unsigned long flags, void (*ctor)(void *)) | |
2387 | { | |
2388 | size_t left_over, slab_size, ralign; | |
2389 | struct kmem_cache *cachep = NULL; | |
2390 | gfp_t gfp; | |
2391 | ||
2392 | #if DEBUG | |
2393 | #if FORCED_DEBUG | |
2394 | /* | |
2395 | * Enable redzoning and last user accounting, except for caches with | |
2396 | * large objects, if the increased size would increase the object size | |
2397 | * above the next power of two: caches with object sizes just above a | |
2398 | * power of two have a significant amount of internal fragmentation. | |
2399 | */ | |
2400 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + | |
2401 | 2 * sizeof(unsigned long long))) | |
2402 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; | |
2403 | if (!(flags & SLAB_DESTROY_BY_RCU)) | |
2404 | flags |= SLAB_POISON; | |
2405 | #endif | |
2406 | if (flags & SLAB_DESTROY_BY_RCU) | |
2407 | BUG_ON(flags & SLAB_POISON); | |
2408 | #endif | |
2409 | /* | |
2410 | * Always checks flags, a caller might be expecting debug support which | |
2411 | * isn't available. | |
2412 | */ | |
2413 | BUG_ON(flags & ~CREATE_MASK); | |
2414 | ||
2415 | /* | |
2416 | * Check that size is in terms of words. This is needed to avoid | |
2417 | * unaligned accesses for some archs when redzoning is used, and makes | |
2418 | * sure any on-slab bufctl's are also correctly aligned. | |
2419 | */ | |
2420 | if (size & (BYTES_PER_WORD - 1)) { | |
2421 | size += (BYTES_PER_WORD - 1); | |
2422 | size &= ~(BYTES_PER_WORD - 1); | |
2423 | } | |
2424 | ||
2425 | /* calculate the final buffer alignment: */ | |
2426 | ||
2427 | /* 1) arch recommendation: can be overridden for debug */ | |
2428 | if (flags & SLAB_HWCACHE_ALIGN) { | |
2429 | /* | |
2430 | * Default alignment: as specified by the arch code. Except if | |
2431 | * an object is really small, then squeeze multiple objects into | |
2432 | * one cacheline. | |
2433 | */ | |
2434 | ralign = cache_line_size(); | |
2435 | while (size <= ralign / 2) | |
2436 | ralign /= 2; | |
2437 | } else { | |
2438 | ralign = BYTES_PER_WORD; | |
2439 | } | |
2440 | ||
2441 | /* | |
2442 | * Redzoning and user store require word alignment or possibly larger. | |
2443 | * Note this will be overridden by architecture or caller mandated | |
2444 | * alignment if either is greater than BYTES_PER_WORD. | |
2445 | */ | |
2446 | if (flags & SLAB_STORE_USER) | |
2447 | ralign = BYTES_PER_WORD; | |
2448 | ||
2449 | if (flags & SLAB_RED_ZONE) { | |
2450 | ralign = REDZONE_ALIGN; | |
2451 | /* If redzoning, ensure that the second redzone is suitably | |
2452 | * aligned, by adjusting the object size accordingly. */ | |
2453 | size += REDZONE_ALIGN - 1; | |
2454 | size &= ~(REDZONE_ALIGN - 1); | |
2455 | } | |
2456 | ||
2457 | /* 2) arch mandated alignment */ | |
2458 | if (ralign < ARCH_SLAB_MINALIGN) { | |
2459 | ralign = ARCH_SLAB_MINALIGN; | |
2460 | } | |
2461 | /* 3) caller mandated alignment */ | |
2462 | if (ralign < align) { | |
2463 | ralign = align; | |
2464 | } | |
2465 | /* disable debug if necessary */ | |
2466 | if (ralign > __alignof__(unsigned long long)) | |
2467 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2468 | /* | |
2469 | * 4) Store it. | |
2470 | */ | |
2471 | align = ralign; | |
2472 | ||
2473 | if (slab_is_available()) | |
2474 | gfp = GFP_KERNEL; | |
2475 | else | |
2476 | gfp = GFP_NOWAIT; | |
2477 | ||
2478 | /* Get cache's description obj. */ | |
2479 | cachep = kmem_cache_zalloc(&cache_cache, gfp); | |
2480 | if (!cachep) | |
2481 | return NULL; | |
2482 | ||
2483 | cachep->nodelists = (struct kmem_list3 **)&cachep->array[nr_cpu_ids]; | |
2484 | cachep->object_size = size; | |
2485 | cachep->align = align; | |
2486 | #if DEBUG | |
2487 | ||
2488 | /* | |
2489 | * Both debugging options require word-alignment which is calculated | |
2490 | * into align above. | |
2491 | */ | |
2492 | if (flags & SLAB_RED_ZONE) { | |
2493 | /* add space for red zone words */ | |
2494 | cachep->obj_offset += sizeof(unsigned long long); | |
2495 | size += 2 * sizeof(unsigned long long); | |
2496 | } | |
2497 | if (flags & SLAB_STORE_USER) { | |
2498 | /* user store requires one word storage behind the end of | |
2499 | * the real object. But if the second red zone needs to be | |
2500 | * aligned to 64 bits, we must allow that much space. | |
2501 | */ | |
2502 | if (flags & SLAB_RED_ZONE) | |
2503 | size += REDZONE_ALIGN; | |
2504 | else | |
2505 | size += BYTES_PER_WORD; | |
2506 | } | |
2507 | #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC) | |
2508 | if (size >= malloc_sizes[INDEX_L3 + 1].cs_size | |
2509 | && cachep->object_size > cache_line_size() && ALIGN(size, align) < PAGE_SIZE) { | |
2510 | cachep->obj_offset += PAGE_SIZE - ALIGN(size, align); | |
2511 | size = PAGE_SIZE; | |
2512 | } | |
2513 | #endif | |
2514 | #endif | |
2515 | ||
2516 | /* | |
2517 | * Determine if the slab management is 'on' or 'off' slab. | |
2518 | * (bootstrapping cannot cope with offslab caches so don't do | |
2519 | * it too early on. Always use on-slab management when | |
2520 | * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak) | |
2521 | */ | |
2522 | if ((size >= (PAGE_SIZE >> 3)) && !slab_early_init && | |
2523 | !(flags & SLAB_NOLEAKTRACE)) | |
2524 | /* | |
2525 | * Size is large, assume best to place the slab management obj | |
2526 | * off-slab (should allow better packing of objs). | |
2527 | */ | |
2528 | flags |= CFLGS_OFF_SLAB; | |
2529 | ||
2530 | size = ALIGN(size, align); | |
2531 | ||
2532 | left_over = calculate_slab_order(cachep, size, align, flags); | |
2533 | ||
2534 | if (!cachep->num) { | |
2535 | printk(KERN_ERR | |
2536 | "kmem_cache_create: couldn't create cache %s.\n", name); | |
2537 | kmem_cache_free(&cache_cache, cachep); | |
2538 | return NULL; | |
2539 | } | |
2540 | slab_size = ALIGN(cachep->num * sizeof(kmem_bufctl_t) | |
2541 | + sizeof(struct slab), align); | |
2542 | ||
2543 | /* | |
2544 | * If the slab has been placed off-slab, and we have enough space then | |
2545 | * move it on-slab. This is at the expense of any extra colouring. | |
2546 | */ | |
2547 | if (flags & CFLGS_OFF_SLAB && left_over >= slab_size) { | |
2548 | flags &= ~CFLGS_OFF_SLAB; | |
2549 | left_over -= slab_size; | |
2550 | } | |
2551 | ||
2552 | if (flags & CFLGS_OFF_SLAB) { | |
2553 | /* really off slab. No need for manual alignment */ | |
2554 | slab_size = | |
2555 | cachep->num * sizeof(kmem_bufctl_t) + sizeof(struct slab); | |
2556 | ||
2557 | #ifdef CONFIG_PAGE_POISONING | |
2558 | /* If we're going to use the generic kernel_map_pages() | |
2559 | * poisoning, then it's going to smash the contents of | |
2560 | * the redzone and userword anyhow, so switch them off. | |
2561 | */ | |
2562 | if (size % PAGE_SIZE == 0 && flags & SLAB_POISON) | |
2563 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); | |
2564 | #endif | |
2565 | } | |
2566 | ||
2567 | cachep->colour_off = cache_line_size(); | |
2568 | /* Offset must be a multiple of the alignment. */ | |
2569 | if (cachep->colour_off < align) | |
2570 | cachep->colour_off = align; | |
2571 | cachep->colour = left_over / cachep->colour_off; | |
2572 | cachep->slab_size = slab_size; | |
2573 | cachep->flags = flags; | |
2574 | cachep->allocflags = 0; | |
2575 | if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA)) | |
2576 | cachep->allocflags |= GFP_DMA; | |
2577 | cachep->size = size; | |
2578 | cachep->reciprocal_buffer_size = reciprocal_value(size); | |
2579 | ||
2580 | if (flags & CFLGS_OFF_SLAB) { | |
2581 | cachep->slabp_cache = kmem_find_general_cachep(slab_size, 0u); | |
2582 | /* | |
2583 | * This is a possibility for one of the malloc_sizes caches. | |
2584 | * But since we go off slab only for object size greater than | |
2585 | * PAGE_SIZE/8, and malloc_sizes gets created in ascending order, | |
2586 | * this should not happen at all. | |
2587 | * But leave a BUG_ON for some lucky dude. | |
2588 | */ | |
2589 | BUG_ON(ZERO_OR_NULL_PTR(cachep->slabp_cache)); | |
2590 | } | |
2591 | cachep->ctor = ctor; | |
2592 | cachep->name = name; | |
2593 | ||
2594 | if (setup_cpu_cache(cachep, gfp)) { | |
2595 | __kmem_cache_destroy(cachep); | |
2596 | return NULL; | |
2597 | } | |
2598 | ||
2599 | if (flags & SLAB_DEBUG_OBJECTS) { | |
2600 | /* | |
2601 | * Would deadlock through slab_destroy()->call_rcu()-> | |
2602 | * debug_object_activate()->kmem_cache_alloc(). | |
2603 | */ | |
2604 | WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU); | |
2605 | ||
2606 | slab_set_debugobj_lock_classes(cachep); | |
2607 | } | |
2608 | ||
2609 | /* cache setup completed, link it into the list */ | |
2610 | list_add(&cachep->list, &slab_caches); | |
2611 | return cachep; | |
2612 | } | |
2613 | ||
2614 | #if DEBUG | |
2615 | static void check_irq_off(void) | |
2616 | { | |
2617 | BUG_ON(!irqs_disabled()); | |
2618 | } | |
2619 | ||
2620 | static void check_irq_on(void) | |
2621 | { | |
2622 | BUG_ON(irqs_disabled()); | |
2623 | } | |
2624 | ||
2625 | static void check_spinlock_acquired(struct kmem_cache *cachep) | |
2626 | { | |
2627 | #ifdef CONFIG_SMP | |
2628 | check_irq_off(); | |
2629 | assert_spin_locked(&cachep->nodelists[numa_mem_id()]->list_lock); | |
2630 | #endif | |
2631 | } | |
2632 | ||
2633 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) | |
2634 | { | |
2635 | #ifdef CONFIG_SMP | |
2636 | check_irq_off(); | |
2637 | assert_spin_locked(&cachep->nodelists[node]->list_lock); | |
2638 | #endif | |
2639 | } | |
2640 | ||
2641 | #else | |
2642 | #define check_irq_off() do { } while(0) | |
2643 | #define check_irq_on() do { } while(0) | |
2644 | #define check_spinlock_acquired(x) do { } while(0) | |
2645 | #define check_spinlock_acquired_node(x, y) do { } while(0) | |
2646 | #endif | |
2647 | ||
2648 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
2649 | struct array_cache *ac, | |
2650 | int force, int node); | |
2651 | ||
2652 | static void do_drain(void *arg) | |
2653 | { | |
2654 | struct kmem_cache *cachep = arg; | |
2655 | struct array_cache *ac; | |
2656 | int node = numa_mem_id(); | |
2657 | ||
2658 | check_irq_off(); | |
2659 | ac = cpu_cache_get(cachep); | |
2660 | spin_lock(&cachep->nodelists[node]->list_lock); | |
2661 | free_block(cachep, ac->entry, ac->avail, node); | |
2662 | spin_unlock(&cachep->nodelists[node]->list_lock); | |
2663 | ac->avail = 0; | |
2664 | } | |
2665 | ||
2666 | static void drain_cpu_caches(struct kmem_cache *cachep) | |
2667 | { | |
2668 | struct kmem_list3 *l3; | |
2669 | int node; | |
2670 | ||
2671 | on_each_cpu(do_drain, cachep, 1); | |
2672 | check_irq_on(); | |
2673 | for_each_online_node(node) { | |
2674 | l3 = cachep->nodelists[node]; | |
2675 | if (l3 && l3->alien) | |
2676 | drain_alien_cache(cachep, l3->alien); | |
2677 | } | |
2678 | ||
2679 | for_each_online_node(node) { | |
2680 | l3 = cachep->nodelists[node]; | |
2681 | if (l3) | |
2682 | drain_array(cachep, l3, l3->shared, 1, node); | |
2683 | } | |
2684 | } | |
2685 | ||
2686 | /* | |
2687 | * Remove slabs from the list of free slabs. | |
2688 | * Specify the number of slabs to drain in tofree. | |
2689 | * | |
2690 | * Returns the actual number of slabs released. | |
2691 | */ | |
2692 | static int drain_freelist(struct kmem_cache *cache, | |
2693 | struct kmem_list3 *l3, int tofree) | |
2694 | { | |
2695 | struct list_head *p; | |
2696 | int nr_freed; | |
2697 | struct slab *slabp; | |
2698 | ||
2699 | nr_freed = 0; | |
2700 | while (nr_freed < tofree && !list_empty(&l3->slabs_free)) { | |
2701 | ||
2702 | spin_lock_irq(&l3->list_lock); | |
2703 | p = l3->slabs_free.prev; | |
2704 | if (p == &l3->slabs_free) { | |
2705 | spin_unlock_irq(&l3->list_lock); | |
2706 | goto out; | |
2707 | } | |
2708 | ||
2709 | slabp = list_entry(p, struct slab, list); | |
2710 | #if DEBUG | |
2711 | BUG_ON(slabp->inuse); | |
2712 | #endif | |
2713 | list_del(&slabp->list); | |
2714 | /* | |
2715 | * Safe to drop the lock. The slab is no longer linked | |
2716 | * to the cache. | |
2717 | */ | |
2718 | l3->free_objects -= cache->num; | |
2719 | spin_unlock_irq(&l3->list_lock); | |
2720 | slab_destroy(cache, slabp); | |
2721 | nr_freed++; | |
2722 | } | |
2723 | out: | |
2724 | return nr_freed; | |
2725 | } | |
2726 | ||
2727 | /* Called with slab_mutex held to protect against cpu hotplug */ | |
2728 | static int __cache_shrink(struct kmem_cache *cachep) | |
2729 | { | |
2730 | int ret = 0, i = 0; | |
2731 | struct kmem_list3 *l3; | |
2732 | ||
2733 | drain_cpu_caches(cachep); | |
2734 | ||
2735 | check_irq_on(); | |
2736 | for_each_online_node(i) { | |
2737 | l3 = cachep->nodelists[i]; | |
2738 | if (!l3) | |
2739 | continue; | |
2740 | ||
2741 | drain_freelist(cachep, l3, l3->free_objects); | |
2742 | ||
2743 | ret += !list_empty(&l3->slabs_full) || | |
2744 | !list_empty(&l3->slabs_partial); | |
2745 | } | |
2746 | return (ret ? 1 : 0); | |
2747 | } | |
2748 | ||
2749 | /** | |
2750 | * kmem_cache_shrink - Shrink a cache. | |
2751 | * @cachep: The cache to shrink. | |
2752 | * | |
2753 | * Releases as many slabs as possible for a cache. | |
2754 | * To help debugging, a zero exit status indicates all slabs were released. | |
2755 | */ | |
2756 | int kmem_cache_shrink(struct kmem_cache *cachep) | |
2757 | { | |
2758 | int ret; | |
2759 | BUG_ON(!cachep || in_interrupt()); | |
2760 | ||
2761 | get_online_cpus(); | |
2762 | mutex_lock(&slab_mutex); | |
2763 | ret = __cache_shrink(cachep); | |
2764 | mutex_unlock(&slab_mutex); | |
2765 | put_online_cpus(); | |
2766 | return ret; | |
2767 | } | |
2768 | EXPORT_SYMBOL(kmem_cache_shrink); | |
2769 | ||
2770 | /** | |
2771 | * kmem_cache_destroy - delete a cache | |
2772 | * @cachep: the cache to destroy | |
2773 | * | |
2774 | * Remove a &struct kmem_cache object from the slab cache. | |
2775 | * | |
2776 | * It is expected this function will be called by a module when it is | |
2777 | * unloaded. This will remove the cache completely, and avoid a duplicate | |
2778 | * cache being allocated each time a module is loaded and unloaded, if the | |
2779 | * module doesn't have persistent in-kernel storage across loads and unloads. | |
2780 | * | |
2781 | * The cache must be empty before calling this function. | |
2782 | * | |
2783 | * The caller must guarantee that no one will allocate memory from the cache | |
2784 | * during the kmem_cache_destroy(). | |
2785 | */ | |
2786 | void kmem_cache_destroy(struct kmem_cache *cachep) | |
2787 | { | |
2788 | BUG_ON(!cachep || in_interrupt()); | |
2789 | ||
2790 | /* Find the cache in the chain of caches. */ | |
2791 | get_online_cpus(); | |
2792 | mutex_lock(&slab_mutex); | |
2793 | /* | |
2794 | * the chain is never empty, cache_cache is never destroyed | |
2795 | */ | |
2796 | list_del(&cachep->list); | |
2797 | if (__cache_shrink(cachep)) { | |
2798 | slab_error(cachep, "Can't free all objects"); | |
2799 | list_add(&cachep->list, &slab_caches); | |
2800 | mutex_unlock(&slab_mutex); | |
2801 | put_online_cpus(); | |
2802 | return; | |
2803 | } | |
2804 | ||
2805 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) | |
2806 | rcu_barrier(); | |
2807 | ||
2808 | __kmem_cache_destroy(cachep); | |
2809 | mutex_unlock(&slab_mutex); | |
2810 | put_online_cpus(); | |
2811 | } | |
2812 | EXPORT_SYMBOL(kmem_cache_destroy); | |
2813 | ||
2814 | /* | |
2815 | * Get the memory for a slab management obj. | |
2816 | * For a slab cache when the slab descriptor is off-slab, slab descriptors | |
2817 | * always come from malloc_sizes caches. The slab descriptor cannot | |
2818 | * come from the same cache which is getting created because, | |
2819 | * when we are searching for an appropriate cache for these | |
2820 | * descriptors in kmem_cache_create, we search through the malloc_sizes array. | |
2821 | * If we are creating a malloc_sizes cache here it would not be visible to | |
2822 | * kmem_find_general_cachep till the initialization is complete. | |
2823 | * Hence we cannot have slabp_cache same as the original cache. | |
2824 | */ | |
2825 | static struct slab *alloc_slabmgmt(struct kmem_cache *cachep, void *objp, | |
2826 | int colour_off, gfp_t local_flags, | |
2827 | int nodeid) | |
2828 | { | |
2829 | struct slab *slabp; | |
2830 | ||
2831 | if (OFF_SLAB(cachep)) { | |
2832 | /* Slab management obj is off-slab. */ | |
2833 | slabp = kmem_cache_alloc_node(cachep->slabp_cache, | |
2834 | local_flags, nodeid); | |
2835 | /* | |
2836 | * If the first object in the slab is leaked (it's allocated | |
2837 | * but no one has a reference to it), we want to make sure | |
2838 | * kmemleak does not treat the ->s_mem pointer as a reference | |
2839 | * to the object. Otherwise we will not report the leak. | |
2840 | */ | |
2841 | kmemleak_scan_area(&slabp->list, sizeof(struct list_head), | |
2842 | local_flags); | |
2843 | if (!slabp) | |
2844 | return NULL; | |
2845 | } else { | |
2846 | slabp = objp + colour_off; | |
2847 | colour_off += cachep->slab_size; | |
2848 | } | |
2849 | slabp->inuse = 0; | |
2850 | slabp->colouroff = colour_off; | |
2851 | slabp->s_mem = objp + colour_off; | |
2852 | slabp->nodeid = nodeid; | |
2853 | slabp->free = 0; | |
2854 | return slabp; | |
2855 | } | |
2856 | ||
2857 | static inline kmem_bufctl_t *slab_bufctl(struct slab *slabp) | |
2858 | { | |
2859 | return (kmem_bufctl_t *) (slabp + 1); | |
2860 | } | |
2861 | ||
2862 | static void cache_init_objs(struct kmem_cache *cachep, | |
2863 | struct slab *slabp) | |
2864 | { | |
2865 | int i; | |
2866 | ||
2867 | for (i = 0; i < cachep->num; i++) { | |
2868 | void *objp = index_to_obj(cachep, slabp, i); | |
2869 | #if DEBUG | |
2870 | /* need to poison the objs? */ | |
2871 | if (cachep->flags & SLAB_POISON) | |
2872 | poison_obj(cachep, objp, POISON_FREE); | |
2873 | if (cachep->flags & SLAB_STORE_USER) | |
2874 | *dbg_userword(cachep, objp) = NULL; | |
2875 | ||
2876 | if (cachep->flags & SLAB_RED_ZONE) { | |
2877 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
2878 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
2879 | } | |
2880 | /* | |
2881 | * Constructors are not allowed to allocate memory from the same | |
2882 | * cache which they are a constructor for. Otherwise, deadlock. | |
2883 | * They must also be threaded. | |
2884 | */ | |
2885 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) | |
2886 | cachep->ctor(objp + obj_offset(cachep)); | |
2887 | ||
2888 | if (cachep->flags & SLAB_RED_ZONE) { | |
2889 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) | |
2890 | slab_error(cachep, "constructor overwrote the" | |
2891 | " end of an object"); | |
2892 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) | |
2893 | slab_error(cachep, "constructor overwrote the" | |
2894 | " start of an object"); | |
2895 | } | |
2896 | if ((cachep->size % PAGE_SIZE) == 0 && | |
2897 | OFF_SLAB(cachep) && cachep->flags & SLAB_POISON) | |
2898 | kernel_map_pages(virt_to_page(objp), | |
2899 | cachep->size / PAGE_SIZE, 0); | |
2900 | #else | |
2901 | if (cachep->ctor) | |
2902 | cachep->ctor(objp); | |
2903 | #endif | |
2904 | slab_bufctl(slabp)[i] = i + 1; | |
2905 | } | |
2906 | slab_bufctl(slabp)[i - 1] = BUFCTL_END; | |
2907 | } | |
2908 | ||
2909 | static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags) | |
2910 | { | |
2911 | if (CONFIG_ZONE_DMA_FLAG) { | |
2912 | if (flags & GFP_DMA) | |
2913 | BUG_ON(!(cachep->allocflags & GFP_DMA)); | |
2914 | else | |
2915 | BUG_ON(cachep->allocflags & GFP_DMA); | |
2916 | } | |
2917 | } | |
2918 | ||
2919 | static void *slab_get_obj(struct kmem_cache *cachep, struct slab *slabp, | |
2920 | int nodeid) | |
2921 | { | |
2922 | void *objp = index_to_obj(cachep, slabp, slabp->free); | |
2923 | kmem_bufctl_t next; | |
2924 | ||
2925 | slabp->inuse++; | |
2926 | next = slab_bufctl(slabp)[slabp->free]; | |
2927 | #if DEBUG | |
2928 | slab_bufctl(slabp)[slabp->free] = BUFCTL_FREE; | |
2929 | WARN_ON(slabp->nodeid != nodeid); | |
2930 | #endif | |
2931 | slabp->free = next; | |
2932 | ||
2933 | return objp; | |
2934 | } | |
2935 | ||
2936 | static void slab_put_obj(struct kmem_cache *cachep, struct slab *slabp, | |
2937 | void *objp, int nodeid) | |
2938 | { | |
2939 | unsigned int objnr = obj_to_index(cachep, slabp, objp); | |
2940 | ||
2941 | #if DEBUG | |
2942 | /* Verify that the slab belongs to the intended node */ | |
2943 | WARN_ON(slabp->nodeid != nodeid); | |
2944 | ||
2945 | if (slab_bufctl(slabp)[objnr] + 1 <= SLAB_LIMIT + 1) { | |
2946 | printk(KERN_ERR "slab: double free detected in cache " | |
2947 | "'%s', objp %p\n", cachep->name, objp); | |
2948 | BUG(); | |
2949 | } | |
2950 | #endif | |
2951 | slab_bufctl(slabp)[objnr] = slabp->free; | |
2952 | slabp->free = objnr; | |
2953 | slabp->inuse--; | |
2954 | } | |
2955 | ||
2956 | /* | |
2957 | * Map pages beginning at addr to the given cache and slab. This is required | |
2958 | * for the slab allocator to be able to lookup the cache and slab of a | |
2959 | * virtual address for kfree, ksize, and slab debugging. | |
2960 | */ | |
2961 | static void slab_map_pages(struct kmem_cache *cache, struct slab *slab, | |
2962 | void *addr) | |
2963 | { | |
2964 | int nr_pages; | |
2965 | struct page *page; | |
2966 | ||
2967 | page = virt_to_page(addr); | |
2968 | ||
2969 | nr_pages = 1; | |
2970 | if (likely(!PageCompound(page))) | |
2971 | nr_pages <<= cache->gfporder; | |
2972 | ||
2973 | do { | |
2974 | page->slab_cache = cache; | |
2975 | page->slab_page = slab; | |
2976 | page++; | |
2977 | } while (--nr_pages); | |
2978 | } | |
2979 | ||
2980 | /* | |
2981 | * Grow (by 1) the number of slabs within a cache. This is called by | |
2982 | * kmem_cache_alloc() when there are no active objs left in a cache. | |
2983 | */ | |
2984 | static int cache_grow(struct kmem_cache *cachep, | |
2985 | gfp_t flags, int nodeid, void *objp) | |
2986 | { | |
2987 | struct slab *slabp; | |
2988 | size_t offset; | |
2989 | gfp_t local_flags; | |
2990 | struct kmem_list3 *l3; | |
2991 | ||
2992 | /* | |
2993 | * Be lazy and only check for valid flags here, keeping it out of the | |
2994 | * critical path in kmem_cache_alloc(). | |
2995 | */ | |
2996 | BUG_ON(flags & GFP_SLAB_BUG_MASK); | |
2997 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
2998 | ||
2999 | /* Take the l3 list lock to change the colour_next on this node */ | |
3000 | check_irq_off(); | |
3001 | l3 = cachep->nodelists[nodeid]; | |
3002 | spin_lock(&l3->list_lock); | |
3003 | ||
3004 | /* Get colour for the slab, and cal the next value. */ | |
3005 | offset = l3->colour_next; | |
3006 | l3->colour_next++; | |
3007 | if (l3->colour_next >= cachep->colour) | |
3008 | l3->colour_next = 0; | |
3009 | spin_unlock(&l3->list_lock); | |
3010 | ||
3011 | offset *= cachep->colour_off; | |
3012 | ||
3013 | if (local_flags & __GFP_WAIT) | |
3014 | local_irq_enable(); | |
3015 | ||
3016 | /* | |
3017 | * The test for missing atomic flag is performed here, rather than | |
3018 | * the more obvious place, simply to reduce the critical path length | |
3019 | * in kmem_cache_alloc(). If a caller is seriously mis-behaving they | |
3020 | * will eventually be caught here (where it matters). | |
3021 | */ | |
3022 | kmem_flagcheck(cachep, flags); | |
3023 | ||
3024 | /* | |
3025 | * Get mem for the objs. Attempt to allocate a physical page from | |
3026 | * 'nodeid'. | |
3027 | */ | |
3028 | if (!objp) | |
3029 | objp = kmem_getpages(cachep, local_flags, nodeid); | |
3030 | if (!objp) | |
3031 | goto failed; | |
3032 | ||
3033 | /* Get slab management. */ | |
3034 | slabp = alloc_slabmgmt(cachep, objp, offset, | |
3035 | local_flags & ~GFP_CONSTRAINT_MASK, nodeid); | |
3036 | if (!slabp) | |
3037 | goto opps1; | |
3038 | ||
3039 | slab_map_pages(cachep, slabp, objp); | |
3040 | ||
3041 | cache_init_objs(cachep, slabp); | |
3042 | ||
3043 | if (local_flags & __GFP_WAIT) | |
3044 | local_irq_disable(); | |
3045 | check_irq_off(); | |
3046 | spin_lock(&l3->list_lock); | |
3047 | ||
3048 | /* Make slab active. */ | |
3049 | list_add_tail(&slabp->list, &(l3->slabs_free)); | |
3050 | STATS_INC_GROWN(cachep); | |
3051 | l3->free_objects += cachep->num; | |
3052 | spin_unlock(&l3->list_lock); | |
3053 | return 1; | |
3054 | opps1: | |
3055 | kmem_freepages(cachep, objp); | |
3056 | failed: | |
3057 | if (local_flags & __GFP_WAIT) | |
3058 | local_irq_disable(); | |
3059 | return 0; | |
3060 | } | |
3061 | ||
3062 | #if DEBUG | |
3063 | ||
3064 | /* | |
3065 | * Perform extra freeing checks: | |
3066 | * - detect bad pointers. | |
3067 | * - POISON/RED_ZONE checking | |
3068 | */ | |
3069 | static void kfree_debugcheck(const void *objp) | |
3070 | { | |
3071 | if (!virt_addr_valid(objp)) { | |
3072 | printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n", | |
3073 | (unsigned long)objp); | |
3074 | BUG(); | |
3075 | } | |
3076 | } | |
3077 | ||
3078 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) | |
3079 | { | |
3080 | unsigned long long redzone1, redzone2; | |
3081 | ||
3082 | redzone1 = *dbg_redzone1(cache, obj); | |
3083 | redzone2 = *dbg_redzone2(cache, obj); | |
3084 | ||
3085 | /* | |
3086 | * Redzone is ok. | |
3087 | */ | |
3088 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) | |
3089 | return; | |
3090 | ||
3091 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) | |
3092 | slab_error(cache, "double free detected"); | |
3093 | else | |
3094 | slab_error(cache, "memory outside object was overwritten"); | |
3095 | ||
3096 | printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n", | |
3097 | obj, redzone1, redzone2); | |
3098 | } | |
3099 | ||
3100 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, | |
3101 | void *caller) | |
3102 | { | |
3103 | struct page *page; | |
3104 | unsigned int objnr; | |
3105 | struct slab *slabp; | |
3106 | ||
3107 | BUG_ON(virt_to_cache(objp) != cachep); | |
3108 | ||
3109 | objp -= obj_offset(cachep); | |
3110 | kfree_debugcheck(objp); | |
3111 | page = virt_to_head_page(objp); | |
3112 | ||
3113 | slabp = page->slab_page; | |
3114 | ||
3115 | if (cachep->flags & SLAB_RED_ZONE) { | |
3116 | verify_redzone_free(cachep, objp); | |
3117 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; | |
3118 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; | |
3119 | } | |
3120 | if (cachep->flags & SLAB_STORE_USER) | |
3121 | *dbg_userword(cachep, objp) = caller; | |
3122 | ||
3123 | objnr = obj_to_index(cachep, slabp, objp); | |
3124 | ||
3125 | BUG_ON(objnr >= cachep->num); | |
3126 | BUG_ON(objp != index_to_obj(cachep, slabp, objnr)); | |
3127 | ||
3128 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
3129 | slab_bufctl(slabp)[objnr] = BUFCTL_FREE; | |
3130 | #endif | |
3131 | if (cachep->flags & SLAB_POISON) { | |
3132 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3133 | if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) { | |
3134 | store_stackinfo(cachep, objp, (unsigned long)caller); | |
3135 | kernel_map_pages(virt_to_page(objp), | |
3136 | cachep->size / PAGE_SIZE, 0); | |
3137 | } else { | |
3138 | poison_obj(cachep, objp, POISON_FREE); | |
3139 | } | |
3140 | #else | |
3141 | poison_obj(cachep, objp, POISON_FREE); | |
3142 | #endif | |
3143 | } | |
3144 | return objp; | |
3145 | } | |
3146 | ||
3147 | static void check_slabp(struct kmem_cache *cachep, struct slab *slabp) | |
3148 | { | |
3149 | kmem_bufctl_t i; | |
3150 | int entries = 0; | |
3151 | ||
3152 | /* Check slab's freelist to see if this obj is there. */ | |
3153 | for (i = slabp->free; i != BUFCTL_END; i = slab_bufctl(slabp)[i]) { | |
3154 | entries++; | |
3155 | if (entries > cachep->num || i >= cachep->num) | |
3156 | goto bad; | |
3157 | } | |
3158 | if (entries != cachep->num - slabp->inuse) { | |
3159 | bad: | |
3160 | printk(KERN_ERR "slab: Internal list corruption detected in " | |
3161 | "cache '%s'(%d), slabp %p(%d). Tainted(%s). Hexdump:\n", | |
3162 | cachep->name, cachep->num, slabp, slabp->inuse, | |
3163 | print_tainted()); | |
3164 | print_hex_dump(KERN_ERR, "", DUMP_PREFIX_OFFSET, 16, 1, slabp, | |
3165 | sizeof(*slabp) + cachep->num * sizeof(kmem_bufctl_t), | |
3166 | 1); | |
3167 | BUG(); | |
3168 | } | |
3169 | } | |
3170 | #else | |
3171 | #define kfree_debugcheck(x) do { } while(0) | |
3172 | #define cache_free_debugcheck(x,objp,z) (objp) | |
3173 | #define check_slabp(x,y) do { } while(0) | |
3174 | #endif | |
3175 | ||
3176 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags, | |
3177 | bool force_refill) | |
3178 | { | |
3179 | int batchcount; | |
3180 | struct kmem_list3 *l3; | |
3181 | struct array_cache *ac; | |
3182 | int node; | |
3183 | ||
3184 | check_irq_off(); | |
3185 | node = numa_mem_id(); | |
3186 | if (unlikely(force_refill)) | |
3187 | goto force_grow; | |
3188 | retry: | |
3189 | ac = cpu_cache_get(cachep); | |
3190 | batchcount = ac->batchcount; | |
3191 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { | |
3192 | /* | |
3193 | * If there was little recent activity on this cache, then | |
3194 | * perform only a partial refill. Otherwise we could generate | |
3195 | * refill bouncing. | |
3196 | */ | |
3197 | batchcount = BATCHREFILL_LIMIT; | |
3198 | } | |
3199 | l3 = cachep->nodelists[node]; | |
3200 | ||
3201 | BUG_ON(ac->avail > 0 || !l3); | |
3202 | spin_lock(&l3->list_lock); | |
3203 | ||
3204 | /* See if we can refill from the shared array */ | |
3205 | if (l3->shared && transfer_objects(ac, l3->shared, batchcount)) { | |
3206 | l3->shared->touched = 1; | |
3207 | goto alloc_done; | |
3208 | } | |
3209 | ||
3210 | while (batchcount > 0) { | |
3211 | struct list_head *entry; | |
3212 | struct slab *slabp; | |
3213 | /* Get slab alloc is to come from. */ | |
3214 | entry = l3->slabs_partial.next; | |
3215 | if (entry == &l3->slabs_partial) { | |
3216 | l3->free_touched = 1; | |
3217 | entry = l3->slabs_free.next; | |
3218 | if (entry == &l3->slabs_free) | |
3219 | goto must_grow; | |
3220 | } | |
3221 | ||
3222 | slabp = list_entry(entry, struct slab, list); | |
3223 | check_slabp(cachep, slabp); | |
3224 | check_spinlock_acquired(cachep); | |
3225 | ||
3226 | /* | |
3227 | * The slab was either on partial or free list so | |
3228 | * there must be at least one object available for | |
3229 | * allocation. | |
3230 | */ | |
3231 | BUG_ON(slabp->inuse >= cachep->num); | |
3232 | ||
3233 | while (slabp->inuse < cachep->num && batchcount--) { | |
3234 | STATS_INC_ALLOCED(cachep); | |
3235 | STATS_INC_ACTIVE(cachep); | |
3236 | STATS_SET_HIGH(cachep); | |
3237 | ||
3238 | ac_put_obj(cachep, ac, slab_get_obj(cachep, slabp, | |
3239 | node)); | |
3240 | } | |
3241 | check_slabp(cachep, slabp); | |
3242 | ||
3243 | /* move slabp to correct slabp list: */ | |
3244 | list_del(&slabp->list); | |
3245 | if (slabp->free == BUFCTL_END) | |
3246 | list_add(&slabp->list, &l3->slabs_full); | |
3247 | else | |
3248 | list_add(&slabp->list, &l3->slabs_partial); | |
3249 | } | |
3250 | ||
3251 | must_grow: | |
3252 | l3->free_objects -= ac->avail; | |
3253 | alloc_done: | |
3254 | spin_unlock(&l3->list_lock); | |
3255 | ||
3256 | if (unlikely(!ac->avail)) { | |
3257 | int x; | |
3258 | force_grow: | |
3259 | x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL); | |
3260 | ||
3261 | /* cache_grow can reenable interrupts, then ac could change. */ | |
3262 | ac = cpu_cache_get(cachep); | |
3263 | ||
3264 | /* no objects in sight? abort */ | |
3265 | if (!x && (ac->avail == 0 || force_refill)) | |
3266 | return NULL; | |
3267 | ||
3268 | if (!ac->avail) /* objects refilled by interrupt? */ | |
3269 | goto retry; | |
3270 | } | |
3271 | ac->touched = 1; | |
3272 | ||
3273 | return ac_get_obj(cachep, ac, flags, force_refill); | |
3274 | } | |
3275 | ||
3276 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, | |
3277 | gfp_t flags) | |
3278 | { | |
3279 | might_sleep_if(flags & __GFP_WAIT); | |
3280 | #if DEBUG | |
3281 | kmem_flagcheck(cachep, flags); | |
3282 | #endif | |
3283 | } | |
3284 | ||
3285 | #if DEBUG | |
3286 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, | |
3287 | gfp_t flags, void *objp, void *caller) | |
3288 | { | |
3289 | if (!objp) | |
3290 | return objp; | |
3291 | if (cachep->flags & SLAB_POISON) { | |
3292 | #ifdef CONFIG_DEBUG_PAGEALLOC | |
3293 | if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep)) | |
3294 | kernel_map_pages(virt_to_page(objp), | |
3295 | cachep->size / PAGE_SIZE, 1); | |
3296 | else | |
3297 | check_poison_obj(cachep, objp); | |
3298 | #else | |
3299 | check_poison_obj(cachep, objp); | |
3300 | #endif | |
3301 | poison_obj(cachep, objp, POISON_INUSE); | |
3302 | } | |
3303 | if (cachep->flags & SLAB_STORE_USER) | |
3304 | *dbg_userword(cachep, objp) = caller; | |
3305 | ||
3306 | if (cachep->flags & SLAB_RED_ZONE) { | |
3307 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || | |
3308 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { | |
3309 | slab_error(cachep, "double free, or memory outside" | |
3310 | " object was overwritten"); | |
3311 | printk(KERN_ERR | |
3312 | "%p: redzone 1:0x%llx, redzone 2:0x%llx\n", | |
3313 | objp, *dbg_redzone1(cachep, objp), | |
3314 | *dbg_redzone2(cachep, objp)); | |
3315 | } | |
3316 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; | |
3317 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; | |
3318 | } | |
3319 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
3320 | { | |
3321 | struct slab *slabp; | |
3322 | unsigned objnr; | |
3323 | ||
3324 | slabp = virt_to_head_page(objp)->slab_page; | |
3325 | objnr = (unsigned)(objp - slabp->s_mem) / cachep->size; | |
3326 | slab_bufctl(slabp)[objnr] = BUFCTL_ACTIVE; | |
3327 | } | |
3328 | #endif | |
3329 | objp += obj_offset(cachep); | |
3330 | if (cachep->ctor && cachep->flags & SLAB_POISON) | |
3331 | cachep->ctor(objp); | |
3332 | if (ARCH_SLAB_MINALIGN && | |
3333 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { | |
3334 | printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n", | |
3335 | objp, (int)ARCH_SLAB_MINALIGN); | |
3336 | } | |
3337 | return objp; | |
3338 | } | |
3339 | #else | |
3340 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) | |
3341 | #endif | |
3342 | ||
3343 | static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags) | |
3344 | { | |
3345 | if (cachep == &cache_cache) | |
3346 | return false; | |
3347 | ||
3348 | return should_failslab(cachep->object_size, flags, cachep->flags); | |
3349 | } | |
3350 | ||
3351 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3352 | { | |
3353 | void *objp; | |
3354 | struct array_cache *ac; | |
3355 | bool force_refill = false; | |
3356 | ||
3357 | check_irq_off(); | |
3358 | ||
3359 | ac = cpu_cache_get(cachep); | |
3360 | if (likely(ac->avail)) { | |
3361 | ac->touched = 1; | |
3362 | objp = ac_get_obj(cachep, ac, flags, false); | |
3363 | ||
3364 | /* | |
3365 | * Allow for the possibility all avail objects are not allowed | |
3366 | * by the current flags | |
3367 | */ | |
3368 | if (objp) { | |
3369 | STATS_INC_ALLOCHIT(cachep); | |
3370 | goto out; | |
3371 | } | |
3372 | force_refill = true; | |
3373 | } | |
3374 | ||
3375 | STATS_INC_ALLOCMISS(cachep); | |
3376 | objp = cache_alloc_refill(cachep, flags, force_refill); | |
3377 | /* | |
3378 | * the 'ac' may be updated by cache_alloc_refill(), | |
3379 | * and kmemleak_erase() requires its correct value. | |
3380 | */ | |
3381 | ac = cpu_cache_get(cachep); | |
3382 | ||
3383 | out: | |
3384 | /* | |
3385 | * To avoid a false negative, if an object that is in one of the | |
3386 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't | |
3387 | * treat the array pointers as a reference to the object. | |
3388 | */ | |
3389 | if (objp) | |
3390 | kmemleak_erase(&ac->entry[ac->avail]); | |
3391 | return objp; | |
3392 | } | |
3393 | ||
3394 | #ifdef CONFIG_NUMA | |
3395 | /* | |
3396 | * Try allocating on another node if PF_SPREAD_SLAB|PF_MEMPOLICY. | |
3397 | * | |
3398 | * If we are in_interrupt, then process context, including cpusets and | |
3399 | * mempolicy, may not apply and should not be used for allocation policy. | |
3400 | */ | |
3401 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3402 | { | |
3403 | int nid_alloc, nid_here; | |
3404 | ||
3405 | if (in_interrupt() || (flags & __GFP_THISNODE)) | |
3406 | return NULL; | |
3407 | nid_alloc = nid_here = numa_mem_id(); | |
3408 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) | |
3409 | nid_alloc = cpuset_slab_spread_node(); | |
3410 | else if (current->mempolicy) | |
3411 | nid_alloc = slab_node(); | |
3412 | if (nid_alloc != nid_here) | |
3413 | return ____cache_alloc_node(cachep, flags, nid_alloc); | |
3414 | return NULL; | |
3415 | } | |
3416 | ||
3417 | /* | |
3418 | * Fallback function if there was no memory available and no objects on a | |
3419 | * certain node and fall back is permitted. First we scan all the | |
3420 | * available nodelists for available objects. If that fails then we | |
3421 | * perform an allocation without specifying a node. This allows the page | |
3422 | * allocator to do its reclaim / fallback magic. We then insert the | |
3423 | * slab into the proper nodelist and then allocate from it. | |
3424 | */ | |
3425 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) | |
3426 | { | |
3427 | struct zonelist *zonelist; | |
3428 | gfp_t local_flags; | |
3429 | struct zoneref *z; | |
3430 | struct zone *zone; | |
3431 | enum zone_type high_zoneidx = gfp_zone(flags); | |
3432 | void *obj = NULL; | |
3433 | int nid; | |
3434 | unsigned int cpuset_mems_cookie; | |
3435 | ||
3436 | if (flags & __GFP_THISNODE) | |
3437 | return NULL; | |
3438 | ||
3439 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); | |
3440 | ||
3441 | retry_cpuset: | |
3442 | cpuset_mems_cookie = get_mems_allowed(); | |
3443 | zonelist = node_zonelist(slab_node(), flags); | |
3444 | ||
3445 | retry: | |
3446 | /* | |
3447 | * Look through allowed nodes for objects available | |
3448 | * from existing per node queues. | |
3449 | */ | |
3450 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { | |
3451 | nid = zone_to_nid(zone); | |
3452 | ||
3453 | if (cpuset_zone_allowed_hardwall(zone, flags) && | |
3454 | cache->nodelists[nid] && | |
3455 | cache->nodelists[nid]->free_objects) { | |
3456 | obj = ____cache_alloc_node(cache, | |
3457 | flags | GFP_THISNODE, nid); | |
3458 | if (obj) | |
3459 | break; | |
3460 | } | |
3461 | } | |
3462 | ||
3463 | if (!obj) { | |
3464 | /* | |
3465 | * This allocation will be performed within the constraints | |
3466 | * of the current cpuset / memory policy requirements. | |
3467 | * We may trigger various forms of reclaim on the allowed | |
3468 | * set and go into memory reserves if necessary. | |
3469 | */ | |
3470 | if (local_flags & __GFP_WAIT) | |
3471 | local_irq_enable(); | |
3472 | kmem_flagcheck(cache, flags); | |
3473 | obj = kmem_getpages(cache, local_flags, numa_mem_id()); | |
3474 | if (local_flags & __GFP_WAIT) | |
3475 | local_irq_disable(); | |
3476 | if (obj) { | |
3477 | /* | |
3478 | * Insert into the appropriate per node queues | |
3479 | */ | |
3480 | nid = page_to_nid(virt_to_page(obj)); | |
3481 | if (cache_grow(cache, flags, nid, obj)) { | |
3482 | obj = ____cache_alloc_node(cache, | |
3483 | flags | GFP_THISNODE, nid); | |
3484 | if (!obj) | |
3485 | /* | |
3486 | * Another processor may allocate the | |
3487 | * objects in the slab since we are | |
3488 | * not holding any locks. | |
3489 | */ | |
3490 | goto retry; | |
3491 | } else { | |
3492 | /* cache_grow already freed obj */ | |
3493 | obj = NULL; | |
3494 | } | |
3495 | } | |
3496 | } | |
3497 | ||
3498 | if (unlikely(!put_mems_allowed(cpuset_mems_cookie) && !obj)) | |
3499 | goto retry_cpuset; | |
3500 | return obj; | |
3501 | } | |
3502 | ||
3503 | /* | |
3504 | * A interface to enable slab creation on nodeid | |
3505 | */ | |
3506 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, | |
3507 | int nodeid) | |
3508 | { | |
3509 | struct list_head *entry; | |
3510 | struct slab *slabp; | |
3511 | struct kmem_list3 *l3; | |
3512 | void *obj; | |
3513 | int x; | |
3514 | ||
3515 | l3 = cachep->nodelists[nodeid]; | |
3516 | BUG_ON(!l3); | |
3517 | ||
3518 | retry: | |
3519 | check_irq_off(); | |
3520 | spin_lock(&l3->list_lock); | |
3521 | entry = l3->slabs_partial.next; | |
3522 | if (entry == &l3->slabs_partial) { | |
3523 | l3->free_touched = 1; | |
3524 | entry = l3->slabs_free.next; | |
3525 | if (entry == &l3->slabs_free) | |
3526 | goto must_grow; | |
3527 | } | |
3528 | ||
3529 | slabp = list_entry(entry, struct slab, list); | |
3530 | check_spinlock_acquired_node(cachep, nodeid); | |
3531 | check_slabp(cachep, slabp); | |
3532 | ||
3533 | STATS_INC_NODEALLOCS(cachep); | |
3534 | STATS_INC_ACTIVE(cachep); | |
3535 | STATS_SET_HIGH(cachep); | |
3536 | ||
3537 | BUG_ON(slabp->inuse == cachep->num); | |
3538 | ||
3539 | obj = slab_get_obj(cachep, slabp, nodeid); | |
3540 | check_slabp(cachep, slabp); | |
3541 | l3->free_objects--; | |
3542 | /* move slabp to correct slabp list: */ | |
3543 | list_del(&slabp->list); | |
3544 | ||
3545 | if (slabp->free == BUFCTL_END) | |
3546 | list_add(&slabp->list, &l3->slabs_full); | |
3547 | else | |
3548 | list_add(&slabp->list, &l3->slabs_partial); | |
3549 | ||
3550 | spin_unlock(&l3->list_lock); | |
3551 | goto done; | |
3552 | ||
3553 | must_grow: | |
3554 | spin_unlock(&l3->list_lock); | |
3555 | x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL); | |
3556 | if (x) | |
3557 | goto retry; | |
3558 | ||
3559 | return fallback_alloc(cachep, flags); | |
3560 | ||
3561 | done: | |
3562 | return obj; | |
3563 | } | |
3564 | ||
3565 | /** | |
3566 | * kmem_cache_alloc_node - Allocate an object on the specified node | |
3567 | * @cachep: The cache to allocate from. | |
3568 | * @flags: See kmalloc(). | |
3569 | * @nodeid: node number of the target node. | |
3570 | * @caller: return address of caller, used for debug information | |
3571 | * | |
3572 | * Identical to kmem_cache_alloc but it will allocate memory on the given | |
3573 | * node, which can improve the performance for cpu bound structures. | |
3574 | * | |
3575 | * Fallback to other node is possible if __GFP_THISNODE is not set. | |
3576 | */ | |
3577 | static __always_inline void * | |
3578 | __cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, | |
3579 | void *caller) | |
3580 | { | |
3581 | unsigned long save_flags; | |
3582 | void *ptr; | |
3583 | int slab_node = numa_mem_id(); | |
3584 | ||
3585 | flags &= gfp_allowed_mask; | |
3586 | ||
3587 | lockdep_trace_alloc(flags); | |
3588 | ||
3589 | if (slab_should_failslab(cachep, flags)) | |
3590 | return NULL; | |
3591 | ||
3592 | cache_alloc_debugcheck_before(cachep, flags); | |
3593 | local_irq_save(save_flags); | |
3594 | ||
3595 | if (nodeid == NUMA_NO_NODE) | |
3596 | nodeid = slab_node; | |
3597 | ||
3598 | if (unlikely(!cachep->nodelists[nodeid])) { | |
3599 | /* Node not bootstrapped yet */ | |
3600 | ptr = fallback_alloc(cachep, flags); | |
3601 | goto out; | |
3602 | } | |
3603 | ||
3604 | if (nodeid == slab_node) { | |
3605 | /* | |
3606 | * Use the locally cached objects if possible. | |
3607 | * However ____cache_alloc does not allow fallback | |
3608 | * to other nodes. It may fail while we still have | |
3609 | * objects on other nodes available. | |
3610 | */ | |
3611 | ptr = ____cache_alloc(cachep, flags); | |
3612 | if (ptr) | |
3613 | goto out; | |
3614 | } | |
3615 | /* ___cache_alloc_node can fall back to other nodes */ | |
3616 | ptr = ____cache_alloc_node(cachep, flags, nodeid); | |
3617 | out: | |
3618 | local_irq_restore(save_flags); | |
3619 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); | |
3620 | kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags, | |
3621 | flags); | |
3622 | ||
3623 | if (likely(ptr)) | |
3624 | kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size); | |
3625 | ||
3626 | if (unlikely((flags & __GFP_ZERO) && ptr)) | |
3627 | memset(ptr, 0, cachep->object_size); | |
3628 | ||
3629 | return ptr; | |
3630 | } | |
3631 | ||
3632 | static __always_inline void * | |
3633 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) | |
3634 | { | |
3635 | void *objp; | |
3636 | ||
3637 | if (unlikely(current->flags & (PF_SPREAD_SLAB | PF_MEMPOLICY))) { | |
3638 | objp = alternate_node_alloc(cache, flags); | |
3639 | if (objp) | |
3640 | goto out; | |
3641 | } | |
3642 | objp = ____cache_alloc(cache, flags); | |
3643 | ||
3644 | /* | |
3645 | * We may just have run out of memory on the local node. | |
3646 | * ____cache_alloc_node() knows how to locate memory on other nodes | |
3647 | */ | |
3648 | if (!objp) | |
3649 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); | |
3650 | ||
3651 | out: | |
3652 | return objp; | |
3653 | } | |
3654 | #else | |
3655 | ||
3656 | static __always_inline void * | |
3657 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3658 | { | |
3659 | return ____cache_alloc(cachep, flags); | |
3660 | } | |
3661 | ||
3662 | #endif /* CONFIG_NUMA */ | |
3663 | ||
3664 | static __always_inline void * | |
3665 | __cache_alloc(struct kmem_cache *cachep, gfp_t flags, void *caller) | |
3666 | { | |
3667 | unsigned long save_flags; | |
3668 | void *objp; | |
3669 | ||
3670 | flags &= gfp_allowed_mask; | |
3671 | ||
3672 | lockdep_trace_alloc(flags); | |
3673 | ||
3674 | if (slab_should_failslab(cachep, flags)) | |
3675 | return NULL; | |
3676 | ||
3677 | cache_alloc_debugcheck_before(cachep, flags); | |
3678 | local_irq_save(save_flags); | |
3679 | objp = __do_cache_alloc(cachep, flags); | |
3680 | local_irq_restore(save_flags); | |
3681 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); | |
3682 | kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags, | |
3683 | flags); | |
3684 | prefetchw(objp); | |
3685 | ||
3686 | if (likely(objp)) | |
3687 | kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size); | |
3688 | ||
3689 | if (unlikely((flags & __GFP_ZERO) && objp)) | |
3690 | memset(objp, 0, cachep->object_size); | |
3691 | ||
3692 | return objp; | |
3693 | } | |
3694 | ||
3695 | /* | |
3696 | * Caller needs to acquire correct kmem_list's list_lock | |
3697 | */ | |
3698 | static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects, | |
3699 | int node) | |
3700 | { | |
3701 | int i; | |
3702 | struct kmem_list3 *l3; | |
3703 | ||
3704 | for (i = 0; i < nr_objects; i++) { | |
3705 | void *objp; | |
3706 | struct slab *slabp; | |
3707 | ||
3708 | clear_obj_pfmemalloc(&objpp[i]); | |
3709 | objp = objpp[i]; | |
3710 | ||
3711 | slabp = virt_to_slab(objp); | |
3712 | l3 = cachep->nodelists[node]; | |
3713 | list_del(&slabp->list); | |
3714 | check_spinlock_acquired_node(cachep, node); | |
3715 | check_slabp(cachep, slabp); | |
3716 | slab_put_obj(cachep, slabp, objp, node); | |
3717 | STATS_DEC_ACTIVE(cachep); | |
3718 | l3->free_objects++; | |
3719 | check_slabp(cachep, slabp); | |
3720 | ||
3721 | /* fixup slab chains */ | |
3722 | if (slabp->inuse == 0) { | |
3723 | if (l3->free_objects > l3->free_limit) { | |
3724 | l3->free_objects -= cachep->num; | |
3725 | /* No need to drop any previously held | |
3726 | * lock here, even if we have a off-slab slab | |
3727 | * descriptor it is guaranteed to come from | |
3728 | * a different cache, refer to comments before | |
3729 | * alloc_slabmgmt. | |
3730 | */ | |
3731 | slab_destroy(cachep, slabp); | |
3732 | } else { | |
3733 | list_add(&slabp->list, &l3->slabs_free); | |
3734 | } | |
3735 | } else { | |
3736 | /* Unconditionally move a slab to the end of the | |
3737 | * partial list on free - maximum time for the | |
3738 | * other objects to be freed, too. | |
3739 | */ | |
3740 | list_add_tail(&slabp->list, &l3->slabs_partial); | |
3741 | } | |
3742 | } | |
3743 | } | |
3744 | ||
3745 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) | |
3746 | { | |
3747 | int batchcount; | |
3748 | struct kmem_list3 *l3; | |
3749 | int node = numa_mem_id(); | |
3750 | ||
3751 | batchcount = ac->batchcount; | |
3752 | #if DEBUG | |
3753 | BUG_ON(!batchcount || batchcount > ac->avail); | |
3754 | #endif | |
3755 | check_irq_off(); | |
3756 | l3 = cachep->nodelists[node]; | |
3757 | spin_lock(&l3->list_lock); | |
3758 | if (l3->shared) { | |
3759 | struct array_cache *shared_array = l3->shared; | |
3760 | int max = shared_array->limit - shared_array->avail; | |
3761 | if (max) { | |
3762 | if (batchcount > max) | |
3763 | batchcount = max; | |
3764 | memcpy(&(shared_array->entry[shared_array->avail]), | |
3765 | ac->entry, sizeof(void *) * batchcount); | |
3766 | shared_array->avail += batchcount; | |
3767 | goto free_done; | |
3768 | } | |
3769 | } | |
3770 | ||
3771 | free_block(cachep, ac->entry, batchcount, node); | |
3772 | free_done: | |
3773 | #if STATS | |
3774 | { | |
3775 | int i = 0; | |
3776 | struct list_head *p; | |
3777 | ||
3778 | p = l3->slabs_free.next; | |
3779 | while (p != &(l3->slabs_free)) { | |
3780 | struct slab *slabp; | |
3781 | ||
3782 | slabp = list_entry(p, struct slab, list); | |
3783 | BUG_ON(slabp->inuse); | |
3784 | ||
3785 | i++; | |
3786 | p = p->next; | |
3787 | } | |
3788 | STATS_SET_FREEABLE(cachep, i); | |
3789 | } | |
3790 | #endif | |
3791 | spin_unlock(&l3->list_lock); | |
3792 | ac->avail -= batchcount; | |
3793 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); | |
3794 | } | |
3795 | ||
3796 | /* | |
3797 | * Release an obj back to its cache. If the obj has a constructed state, it must | |
3798 | * be in this state _before_ it is released. Called with disabled ints. | |
3799 | */ | |
3800 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, | |
3801 | void *caller) | |
3802 | { | |
3803 | struct array_cache *ac = cpu_cache_get(cachep); | |
3804 | ||
3805 | check_irq_off(); | |
3806 | kmemleak_free_recursive(objp, cachep->flags); | |
3807 | objp = cache_free_debugcheck(cachep, objp, caller); | |
3808 | ||
3809 | kmemcheck_slab_free(cachep, objp, cachep->object_size); | |
3810 | ||
3811 | /* | |
3812 | * Skip calling cache_free_alien() when the platform is not numa. | |
3813 | * This will avoid cache misses that happen while accessing slabp (which | |
3814 | * is per page memory reference) to get nodeid. Instead use a global | |
3815 | * variable to skip the call, which is mostly likely to be present in | |
3816 | * the cache. | |
3817 | */ | |
3818 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) | |
3819 | return; | |
3820 | ||
3821 | if (likely(ac->avail < ac->limit)) { | |
3822 | STATS_INC_FREEHIT(cachep); | |
3823 | } else { | |
3824 | STATS_INC_FREEMISS(cachep); | |
3825 | cache_flusharray(cachep, ac); | |
3826 | } | |
3827 | ||
3828 | ac_put_obj(cachep, ac, objp); | |
3829 | } | |
3830 | ||
3831 | /** | |
3832 | * kmem_cache_alloc - Allocate an object | |
3833 | * @cachep: The cache to allocate from. | |
3834 | * @flags: See kmalloc(). | |
3835 | * | |
3836 | * Allocate an object from this cache. The flags are only relevant | |
3837 | * if the cache has no available objects. | |
3838 | */ | |
3839 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) | |
3840 | { | |
3841 | void *ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3842 | ||
3843 | trace_kmem_cache_alloc(_RET_IP_, ret, | |
3844 | cachep->object_size, cachep->size, flags); | |
3845 | ||
3846 | return ret; | |
3847 | } | |
3848 | EXPORT_SYMBOL(kmem_cache_alloc); | |
3849 | ||
3850 | #ifdef CONFIG_TRACING | |
3851 | void * | |
3852 | kmem_cache_alloc_trace(size_t size, struct kmem_cache *cachep, gfp_t flags) | |
3853 | { | |
3854 | void *ret; | |
3855 | ||
3856 | ret = __cache_alloc(cachep, flags, __builtin_return_address(0)); | |
3857 | ||
3858 | trace_kmalloc(_RET_IP_, ret, | |
3859 | size, slab_buffer_size(cachep), flags); | |
3860 | return ret; | |
3861 | } | |
3862 | EXPORT_SYMBOL(kmem_cache_alloc_trace); | |
3863 | #endif | |
3864 | ||
3865 | #ifdef CONFIG_NUMA | |
3866 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) | |
3867 | { | |
3868 | void *ret = __cache_alloc_node(cachep, flags, nodeid, | |
3869 | __builtin_return_address(0)); | |
3870 | ||
3871 | trace_kmem_cache_alloc_node(_RET_IP_, ret, | |
3872 | cachep->object_size, cachep->size, | |
3873 | flags, nodeid); | |
3874 | ||
3875 | return ret; | |
3876 | } | |
3877 | EXPORT_SYMBOL(kmem_cache_alloc_node); | |
3878 | ||
3879 | #ifdef CONFIG_TRACING | |
3880 | void *kmem_cache_alloc_node_trace(size_t size, | |
3881 | struct kmem_cache *cachep, | |
3882 | gfp_t flags, | |
3883 | int nodeid) | |
3884 | { | |
3885 | void *ret; | |
3886 | ||
3887 | ret = __cache_alloc_node(cachep, flags, nodeid, | |
3888 | __builtin_return_address(0)); | |
3889 | trace_kmalloc_node(_RET_IP_, ret, | |
3890 | size, slab_buffer_size(cachep), | |
3891 | flags, nodeid); | |
3892 | return ret; | |
3893 | } | |
3894 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); | |
3895 | #endif | |
3896 | ||
3897 | static __always_inline void * | |
3898 | __do_kmalloc_node(size_t size, gfp_t flags, int node, void *caller) | |
3899 | { | |
3900 | struct kmem_cache *cachep; | |
3901 | ||
3902 | cachep = kmem_find_general_cachep(size, flags); | |
3903 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) | |
3904 | return cachep; | |
3905 | return kmem_cache_alloc_node_trace(size, cachep, flags, node); | |
3906 | } | |
3907 | ||
3908 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) | |
3909 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3910 | { | |
3911 | return __do_kmalloc_node(size, flags, node, | |
3912 | __builtin_return_address(0)); | |
3913 | } | |
3914 | EXPORT_SYMBOL(__kmalloc_node); | |
3915 | ||
3916 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, | |
3917 | int node, unsigned long caller) | |
3918 | { | |
3919 | return __do_kmalloc_node(size, flags, node, (void *)caller); | |
3920 | } | |
3921 | EXPORT_SYMBOL(__kmalloc_node_track_caller); | |
3922 | #else | |
3923 | void *__kmalloc_node(size_t size, gfp_t flags, int node) | |
3924 | { | |
3925 | return __do_kmalloc_node(size, flags, node, NULL); | |
3926 | } | |
3927 | EXPORT_SYMBOL(__kmalloc_node); | |
3928 | #endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */ | |
3929 | #endif /* CONFIG_NUMA */ | |
3930 | ||
3931 | /** | |
3932 | * __do_kmalloc - allocate memory | |
3933 | * @size: how many bytes of memory are required. | |
3934 | * @flags: the type of memory to allocate (see kmalloc). | |
3935 | * @caller: function caller for debug tracking of the caller | |
3936 | */ | |
3937 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, | |
3938 | void *caller) | |
3939 | { | |
3940 | struct kmem_cache *cachep; | |
3941 | void *ret; | |
3942 | ||
3943 | /* If you want to save a few bytes .text space: replace | |
3944 | * __ with kmem_. | |
3945 | * Then kmalloc uses the uninlined functions instead of the inline | |
3946 | * functions. | |
3947 | */ | |
3948 | cachep = __find_general_cachep(size, flags); | |
3949 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) | |
3950 | return cachep; | |
3951 | ret = __cache_alloc(cachep, flags, caller); | |
3952 | ||
3953 | trace_kmalloc((unsigned long) caller, ret, | |
3954 | size, cachep->size, flags); | |
3955 | ||
3956 | return ret; | |
3957 | } | |
3958 | ||
3959 | ||
3960 | #if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING) | |
3961 | void *__kmalloc(size_t size, gfp_t flags) | |
3962 | { | |
3963 | return __do_kmalloc(size, flags, __builtin_return_address(0)); | |
3964 | } | |
3965 | EXPORT_SYMBOL(__kmalloc); | |
3966 | ||
3967 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) | |
3968 | { | |
3969 | return __do_kmalloc(size, flags, (void *)caller); | |
3970 | } | |
3971 | EXPORT_SYMBOL(__kmalloc_track_caller); | |
3972 | ||
3973 | #else | |
3974 | void *__kmalloc(size_t size, gfp_t flags) | |
3975 | { | |
3976 | return __do_kmalloc(size, flags, NULL); | |
3977 | } | |
3978 | EXPORT_SYMBOL(__kmalloc); | |
3979 | #endif | |
3980 | ||
3981 | /** | |
3982 | * kmem_cache_free - Deallocate an object | |
3983 | * @cachep: The cache the allocation was from. | |
3984 | * @objp: The previously allocated object. | |
3985 | * | |
3986 | * Free an object which was previously allocated from this | |
3987 | * cache. | |
3988 | */ | |
3989 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) | |
3990 | { | |
3991 | unsigned long flags; | |
3992 | ||
3993 | local_irq_save(flags); | |
3994 | debug_check_no_locks_freed(objp, cachep->object_size); | |
3995 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) | |
3996 | debug_check_no_obj_freed(objp, cachep->object_size); | |
3997 | __cache_free(cachep, objp, __builtin_return_address(0)); | |
3998 | local_irq_restore(flags); | |
3999 | ||
4000 | trace_kmem_cache_free(_RET_IP_, objp); | |
4001 | } | |
4002 | EXPORT_SYMBOL(kmem_cache_free); | |
4003 | ||
4004 | /** | |
4005 | * kfree - free previously allocated memory | |
4006 | * @objp: pointer returned by kmalloc. | |
4007 | * | |
4008 | * If @objp is NULL, no operation is performed. | |
4009 | * | |
4010 | * Don't free memory not originally allocated by kmalloc() | |
4011 | * or you will run into trouble. | |
4012 | */ | |
4013 | void kfree(const void *objp) | |
4014 | { | |
4015 | struct kmem_cache *c; | |
4016 | unsigned long flags; | |
4017 | ||
4018 | trace_kfree(_RET_IP_, objp); | |
4019 | ||
4020 | if (unlikely(ZERO_OR_NULL_PTR(objp))) | |
4021 | return; | |
4022 | local_irq_save(flags); | |
4023 | kfree_debugcheck(objp); | |
4024 | c = virt_to_cache(objp); | |
4025 | debug_check_no_locks_freed(objp, c->object_size); | |
4026 | ||
4027 | debug_check_no_obj_freed(objp, c->object_size); | |
4028 | __cache_free(c, (void *)objp, __builtin_return_address(0)); | |
4029 | local_irq_restore(flags); | |
4030 | } | |
4031 | EXPORT_SYMBOL(kfree); | |
4032 | ||
4033 | unsigned int kmem_cache_size(struct kmem_cache *cachep) | |
4034 | { | |
4035 | return cachep->object_size; | |
4036 | } | |
4037 | EXPORT_SYMBOL(kmem_cache_size); | |
4038 | ||
4039 | /* | |
4040 | * This initializes kmem_list3 or resizes various caches for all nodes. | |
4041 | */ | |
4042 | static int alloc_kmemlist(struct kmem_cache *cachep, gfp_t gfp) | |
4043 | { | |
4044 | int node; | |
4045 | struct kmem_list3 *l3; | |
4046 | struct array_cache *new_shared; | |
4047 | struct array_cache **new_alien = NULL; | |
4048 | ||
4049 | for_each_online_node(node) { | |
4050 | ||
4051 | if (use_alien_caches) { | |
4052 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); | |
4053 | if (!new_alien) | |
4054 | goto fail; | |
4055 | } | |
4056 | ||
4057 | new_shared = NULL; | |
4058 | if (cachep->shared) { | |
4059 | new_shared = alloc_arraycache(node, | |
4060 | cachep->shared*cachep->batchcount, | |
4061 | 0xbaadf00d, gfp); | |
4062 | if (!new_shared) { | |
4063 | free_alien_cache(new_alien); | |
4064 | goto fail; | |
4065 | } | |
4066 | } | |
4067 | ||
4068 | l3 = cachep->nodelists[node]; | |
4069 | if (l3) { | |
4070 | struct array_cache *shared = l3->shared; | |
4071 | ||
4072 | spin_lock_irq(&l3->list_lock); | |
4073 | ||
4074 | if (shared) | |
4075 | free_block(cachep, shared->entry, | |
4076 | shared->avail, node); | |
4077 | ||
4078 | l3->shared = new_shared; | |
4079 | if (!l3->alien) { | |
4080 | l3->alien = new_alien; | |
4081 | new_alien = NULL; | |
4082 | } | |
4083 | l3->free_limit = (1 + nr_cpus_node(node)) * | |
4084 | cachep->batchcount + cachep->num; | |
4085 | spin_unlock_irq(&l3->list_lock); | |
4086 | kfree(shared); | |
4087 | free_alien_cache(new_alien); | |
4088 | continue; | |
4089 | } | |
4090 | l3 = kmalloc_node(sizeof(struct kmem_list3), gfp, node); | |
4091 | if (!l3) { | |
4092 | free_alien_cache(new_alien); | |
4093 | kfree(new_shared); | |
4094 | goto fail; | |
4095 | } | |
4096 | ||
4097 | kmem_list3_init(l3); | |
4098 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3 + | |
4099 | ((unsigned long)cachep) % REAPTIMEOUT_LIST3; | |
4100 | l3->shared = new_shared; | |
4101 | l3->alien = new_alien; | |
4102 | l3->free_limit = (1 + nr_cpus_node(node)) * | |
4103 | cachep->batchcount + cachep->num; | |
4104 | cachep->nodelists[node] = l3; | |
4105 | } | |
4106 | return 0; | |
4107 | ||
4108 | fail: | |
4109 | if (!cachep->list.next) { | |
4110 | /* Cache is not active yet. Roll back what we did */ | |
4111 | node--; | |
4112 | while (node >= 0) { | |
4113 | if (cachep->nodelists[node]) { | |
4114 | l3 = cachep->nodelists[node]; | |
4115 | ||
4116 | kfree(l3->shared); | |
4117 | free_alien_cache(l3->alien); | |
4118 | kfree(l3); | |
4119 | cachep->nodelists[node] = NULL; | |
4120 | } | |
4121 | node--; | |
4122 | } | |
4123 | } | |
4124 | return -ENOMEM; | |
4125 | } | |
4126 | ||
4127 | struct ccupdate_struct { | |
4128 | struct kmem_cache *cachep; | |
4129 | struct array_cache *new[0]; | |
4130 | }; | |
4131 | ||
4132 | static void do_ccupdate_local(void *info) | |
4133 | { | |
4134 | struct ccupdate_struct *new = info; | |
4135 | struct array_cache *old; | |
4136 | ||
4137 | check_irq_off(); | |
4138 | old = cpu_cache_get(new->cachep); | |
4139 | ||
4140 | new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()]; | |
4141 | new->new[smp_processor_id()] = old; | |
4142 | } | |
4143 | ||
4144 | /* Always called with the slab_mutex held */ | |
4145 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, | |
4146 | int batchcount, int shared, gfp_t gfp) | |
4147 | { | |
4148 | struct ccupdate_struct *new; | |
4149 | int i; | |
4150 | ||
4151 | new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *), | |
4152 | gfp); | |
4153 | if (!new) | |
4154 | return -ENOMEM; | |
4155 | ||
4156 | for_each_online_cpu(i) { | |
4157 | new->new[i] = alloc_arraycache(cpu_to_mem(i), limit, | |
4158 | batchcount, gfp); | |
4159 | if (!new->new[i]) { | |
4160 | for (i--; i >= 0; i--) | |
4161 | kfree(new->new[i]); | |
4162 | kfree(new); | |
4163 | return -ENOMEM; | |
4164 | } | |
4165 | } | |
4166 | new->cachep = cachep; | |
4167 | ||
4168 | on_each_cpu(do_ccupdate_local, (void *)new, 1); | |
4169 | ||
4170 | check_irq_on(); | |
4171 | cachep->batchcount = batchcount; | |
4172 | cachep->limit = limit; | |
4173 | cachep->shared = shared; | |
4174 | ||
4175 | for_each_online_cpu(i) { | |
4176 | struct array_cache *ccold = new->new[i]; | |
4177 | if (!ccold) | |
4178 | continue; | |
4179 | spin_lock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
4180 | free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i)); | |
4181 | spin_unlock_irq(&cachep->nodelists[cpu_to_mem(i)]->list_lock); | |
4182 | kfree(ccold); | |
4183 | } | |
4184 | kfree(new); | |
4185 | return alloc_kmemlist(cachep, gfp); | |
4186 | } | |
4187 | ||
4188 | /* Called with slab_mutex held always */ | |
4189 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) | |
4190 | { | |
4191 | int err; | |
4192 | int limit, shared; | |
4193 | ||
4194 | /* | |
4195 | * The head array serves three purposes: | |
4196 | * - create a LIFO ordering, i.e. return objects that are cache-warm | |
4197 | * - reduce the number of spinlock operations. | |
4198 | * - reduce the number of linked list operations on the slab and | |
4199 | * bufctl chains: array operations are cheaper. | |
4200 | * The numbers are guessed, we should auto-tune as described by | |
4201 | * Bonwick. | |
4202 | */ | |
4203 | if (cachep->size > 131072) | |
4204 | limit = 1; | |
4205 | else if (cachep->size > PAGE_SIZE) | |
4206 | limit = 8; | |
4207 | else if (cachep->size > 1024) | |
4208 | limit = 24; | |
4209 | else if (cachep->size > 256) | |
4210 | limit = 54; | |
4211 | else | |
4212 | limit = 120; | |
4213 | ||
4214 | /* | |
4215 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound | |
4216 | * allocation behaviour: Most allocs on one cpu, most free operations | |
4217 | * on another cpu. For these cases, an efficient object passing between | |
4218 | * cpus is necessary. This is provided by a shared array. The array | |
4219 | * replaces Bonwick's magazine layer. | |
4220 | * On uniprocessor, it's functionally equivalent (but less efficient) | |
4221 | * to a larger limit. Thus disabled by default. | |
4222 | */ | |
4223 | shared = 0; | |
4224 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) | |
4225 | shared = 8; | |
4226 | ||
4227 | #if DEBUG | |
4228 | /* | |
4229 | * With debugging enabled, large batchcount lead to excessively long | |
4230 | * periods with disabled local interrupts. Limit the batchcount | |
4231 | */ | |
4232 | if (limit > 32) | |
4233 | limit = 32; | |
4234 | #endif | |
4235 | err = do_tune_cpucache(cachep, limit, (limit + 1) / 2, shared, gfp); | |
4236 | if (err) | |
4237 | printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n", | |
4238 | cachep->name, -err); | |
4239 | return err; | |
4240 | } | |
4241 | ||
4242 | /* | |
4243 | * Drain an array if it contains any elements taking the l3 lock only if | |
4244 | * necessary. Note that the l3 listlock also protects the array_cache | |
4245 | * if drain_array() is used on the shared array. | |
4246 | */ | |
4247 | static void drain_array(struct kmem_cache *cachep, struct kmem_list3 *l3, | |
4248 | struct array_cache *ac, int force, int node) | |
4249 | { | |
4250 | int tofree; | |
4251 | ||
4252 | if (!ac || !ac->avail) | |
4253 | return; | |
4254 | if (ac->touched && !force) { | |
4255 | ac->touched = 0; | |
4256 | } else { | |
4257 | spin_lock_irq(&l3->list_lock); | |
4258 | if (ac->avail) { | |
4259 | tofree = force ? ac->avail : (ac->limit + 4) / 5; | |
4260 | if (tofree > ac->avail) | |
4261 | tofree = (ac->avail + 1) / 2; | |
4262 | free_block(cachep, ac->entry, tofree, node); | |
4263 | ac->avail -= tofree; | |
4264 | memmove(ac->entry, &(ac->entry[tofree]), | |
4265 | sizeof(void *) * ac->avail); | |
4266 | } | |
4267 | spin_unlock_irq(&l3->list_lock); | |
4268 | } | |
4269 | } | |
4270 | ||
4271 | /** | |
4272 | * cache_reap - Reclaim memory from caches. | |
4273 | * @w: work descriptor | |
4274 | * | |
4275 | * Called from workqueue/eventd every few seconds. | |
4276 | * Purpose: | |
4277 | * - clear the per-cpu caches for this CPU. | |
4278 | * - return freeable pages to the main free memory pool. | |
4279 | * | |
4280 | * If we cannot acquire the cache chain mutex then just give up - we'll try | |
4281 | * again on the next iteration. | |
4282 | */ | |
4283 | static void cache_reap(struct work_struct *w) | |
4284 | { | |
4285 | struct kmem_cache *searchp; | |
4286 | struct kmem_list3 *l3; | |
4287 | int node = numa_mem_id(); | |
4288 | struct delayed_work *work = to_delayed_work(w); | |
4289 | ||
4290 | if (!mutex_trylock(&slab_mutex)) | |
4291 | /* Give up. Setup the next iteration. */ | |
4292 | goto out; | |
4293 | ||
4294 | list_for_each_entry(searchp, &slab_caches, list) { | |
4295 | check_irq_on(); | |
4296 | ||
4297 | /* | |
4298 | * We only take the l3 lock if absolutely necessary and we | |
4299 | * have established with reasonable certainty that | |
4300 | * we can do some work if the lock was obtained. | |
4301 | */ | |
4302 | l3 = searchp->nodelists[node]; | |
4303 | ||
4304 | reap_alien(searchp, l3); | |
4305 | ||
4306 | drain_array(searchp, l3, cpu_cache_get(searchp), 0, node); | |
4307 | ||
4308 | /* | |
4309 | * These are racy checks but it does not matter | |
4310 | * if we skip one check or scan twice. | |
4311 | */ | |
4312 | if (time_after(l3->next_reap, jiffies)) | |
4313 | goto next; | |
4314 | ||
4315 | l3->next_reap = jiffies + REAPTIMEOUT_LIST3; | |
4316 | ||
4317 | drain_array(searchp, l3, l3->shared, 0, node); | |
4318 | ||
4319 | if (l3->free_touched) | |
4320 | l3->free_touched = 0; | |
4321 | else { | |
4322 | int freed; | |
4323 | ||
4324 | freed = drain_freelist(searchp, l3, (l3->free_limit + | |
4325 | 5 * searchp->num - 1) / (5 * searchp->num)); | |
4326 | STATS_ADD_REAPED(searchp, freed); | |
4327 | } | |
4328 | next: | |
4329 | cond_resched(); | |
4330 | } | |
4331 | check_irq_on(); | |
4332 | mutex_unlock(&slab_mutex); | |
4333 | next_reap_node(); | |
4334 | out: | |
4335 | /* Set up the next iteration */ | |
4336 | schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_CPUC)); | |
4337 | } | |
4338 | ||
4339 | #ifdef CONFIG_SLABINFO | |
4340 | ||
4341 | static void print_slabinfo_header(struct seq_file *m) | |
4342 | { | |
4343 | /* | |
4344 | * Output format version, so at least we can change it | |
4345 | * without _too_ many complaints. | |
4346 | */ | |
4347 | #if STATS | |
4348 | seq_puts(m, "slabinfo - version: 2.1 (statistics)\n"); | |
4349 | #else | |
4350 | seq_puts(m, "slabinfo - version: 2.1\n"); | |
4351 | #endif | |
4352 | seq_puts(m, "# name <active_objs> <num_objs> <objsize> " | |
4353 | "<objperslab> <pagesperslab>"); | |
4354 | seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>"); | |
4355 | seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>"); | |
4356 | #if STATS | |
4357 | seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> " | |
4358 | "<error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>"); | |
4359 | seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>"); | |
4360 | #endif | |
4361 | seq_putc(m, '\n'); | |
4362 | } | |
4363 | ||
4364 | static void *s_start(struct seq_file *m, loff_t *pos) | |
4365 | { | |
4366 | loff_t n = *pos; | |
4367 | ||
4368 | mutex_lock(&slab_mutex); | |
4369 | if (!n) | |
4370 | print_slabinfo_header(m); | |
4371 | ||
4372 | return seq_list_start(&slab_caches, *pos); | |
4373 | } | |
4374 | ||
4375 | static void *s_next(struct seq_file *m, void *p, loff_t *pos) | |
4376 | { | |
4377 | return seq_list_next(p, &slab_caches, pos); | |
4378 | } | |
4379 | ||
4380 | static void s_stop(struct seq_file *m, void *p) | |
4381 | { | |
4382 | mutex_unlock(&slab_mutex); | |
4383 | } | |
4384 | ||
4385 | static int s_show(struct seq_file *m, void *p) | |
4386 | { | |
4387 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); | |
4388 | struct slab *slabp; | |
4389 | unsigned long active_objs; | |
4390 | unsigned long num_objs; | |
4391 | unsigned long active_slabs = 0; | |
4392 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; | |
4393 | const char *name; | |
4394 | char *error = NULL; | |
4395 | int node; | |
4396 | struct kmem_list3 *l3; | |
4397 | ||
4398 | active_objs = 0; | |
4399 | num_slabs = 0; | |
4400 | for_each_online_node(node) { | |
4401 | l3 = cachep->nodelists[node]; | |
4402 | if (!l3) | |
4403 | continue; | |
4404 | ||
4405 | check_irq_on(); | |
4406 | spin_lock_irq(&l3->list_lock); | |
4407 | ||
4408 | list_for_each_entry(slabp, &l3->slabs_full, list) { | |
4409 | if (slabp->inuse != cachep->num && !error) | |
4410 | error = "slabs_full accounting error"; | |
4411 | active_objs += cachep->num; | |
4412 | active_slabs++; | |
4413 | } | |
4414 | list_for_each_entry(slabp, &l3->slabs_partial, list) { | |
4415 | if (slabp->inuse == cachep->num && !error) | |
4416 | error = "slabs_partial inuse accounting error"; | |
4417 | if (!slabp->inuse && !error) | |
4418 | error = "slabs_partial/inuse accounting error"; | |
4419 | active_objs += slabp->inuse; | |
4420 | active_slabs++; | |
4421 | } | |
4422 | list_for_each_entry(slabp, &l3->slabs_free, list) { | |
4423 | if (slabp->inuse && !error) | |
4424 | error = "slabs_free/inuse accounting error"; | |
4425 | num_slabs++; | |
4426 | } | |
4427 | free_objects += l3->free_objects; | |
4428 | if (l3->shared) | |
4429 | shared_avail += l3->shared->avail; | |
4430 | ||
4431 | spin_unlock_irq(&l3->list_lock); | |
4432 | } | |
4433 | num_slabs += active_slabs; | |
4434 | num_objs = num_slabs * cachep->num; | |
4435 | if (num_objs - active_objs != free_objects && !error) | |
4436 | error = "free_objects accounting error"; | |
4437 | ||
4438 | name = cachep->name; | |
4439 | if (error) | |
4440 | printk(KERN_ERR "slab: cache %s error: %s\n", name, error); | |
4441 | ||
4442 | seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", | |
4443 | name, active_objs, num_objs, cachep->size, | |
4444 | cachep->num, (1 << cachep->gfporder)); | |
4445 | seq_printf(m, " : tunables %4u %4u %4u", | |
4446 | cachep->limit, cachep->batchcount, cachep->shared); | |
4447 | seq_printf(m, " : slabdata %6lu %6lu %6lu", | |
4448 | active_slabs, num_slabs, shared_avail); | |
4449 | #if STATS | |
4450 | { /* list3 stats */ | |
4451 | unsigned long high = cachep->high_mark; | |
4452 | unsigned long allocs = cachep->num_allocations; | |
4453 | unsigned long grown = cachep->grown; | |
4454 | unsigned long reaped = cachep->reaped; | |
4455 | unsigned long errors = cachep->errors; | |
4456 | unsigned long max_freeable = cachep->max_freeable; | |
4457 | unsigned long node_allocs = cachep->node_allocs; | |
4458 | unsigned long node_frees = cachep->node_frees; | |
4459 | unsigned long overflows = cachep->node_overflow; | |
4460 | ||
4461 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu " | |
4462 | "%4lu %4lu %4lu %4lu %4lu", | |
4463 | allocs, high, grown, | |
4464 | reaped, errors, max_freeable, node_allocs, | |
4465 | node_frees, overflows); | |
4466 | } | |
4467 | /* cpu stats */ | |
4468 | { | |
4469 | unsigned long allochit = atomic_read(&cachep->allochit); | |
4470 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); | |
4471 | unsigned long freehit = atomic_read(&cachep->freehit); | |
4472 | unsigned long freemiss = atomic_read(&cachep->freemiss); | |
4473 | ||
4474 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", | |
4475 | allochit, allocmiss, freehit, freemiss); | |
4476 | } | |
4477 | #endif | |
4478 | seq_putc(m, '\n'); | |
4479 | return 0; | |
4480 | } | |
4481 | ||
4482 | /* | |
4483 | * slabinfo_op - iterator that generates /proc/slabinfo | |
4484 | * | |
4485 | * Output layout: | |
4486 | * cache-name | |
4487 | * num-active-objs | |
4488 | * total-objs | |
4489 | * object size | |
4490 | * num-active-slabs | |
4491 | * total-slabs | |
4492 | * num-pages-per-slab | |
4493 | * + further values on SMP and with statistics enabled | |
4494 | */ | |
4495 | ||
4496 | static const struct seq_operations slabinfo_op = { | |
4497 | .start = s_start, | |
4498 | .next = s_next, | |
4499 | .stop = s_stop, | |
4500 | .show = s_show, | |
4501 | }; | |
4502 | ||
4503 | #define MAX_SLABINFO_WRITE 128 | |
4504 | /** | |
4505 | * slabinfo_write - Tuning for the slab allocator | |
4506 | * @file: unused | |
4507 | * @buffer: user buffer | |
4508 | * @count: data length | |
4509 | * @ppos: unused | |
4510 | */ | |
4511 | static ssize_t slabinfo_write(struct file *file, const char __user *buffer, | |
4512 | size_t count, loff_t *ppos) | |
4513 | { | |
4514 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; | |
4515 | int limit, batchcount, shared, res; | |
4516 | struct kmem_cache *cachep; | |
4517 | ||
4518 | if (count > MAX_SLABINFO_WRITE) | |
4519 | return -EINVAL; | |
4520 | if (copy_from_user(&kbuf, buffer, count)) | |
4521 | return -EFAULT; | |
4522 | kbuf[MAX_SLABINFO_WRITE] = '\0'; | |
4523 | ||
4524 | tmp = strchr(kbuf, ' '); | |
4525 | if (!tmp) | |
4526 | return -EINVAL; | |
4527 | *tmp = '\0'; | |
4528 | tmp++; | |
4529 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) | |
4530 | return -EINVAL; | |
4531 | ||
4532 | /* Find the cache in the chain of caches. */ | |
4533 | mutex_lock(&slab_mutex); | |
4534 | res = -EINVAL; | |
4535 | list_for_each_entry(cachep, &slab_caches, list) { | |
4536 | if (!strcmp(cachep->name, kbuf)) { | |
4537 | if (limit < 1 || batchcount < 1 || | |
4538 | batchcount > limit || shared < 0) { | |
4539 | res = 0; | |
4540 | } else { | |
4541 | res = do_tune_cpucache(cachep, limit, | |
4542 | batchcount, shared, | |
4543 | GFP_KERNEL); | |
4544 | } | |
4545 | break; | |
4546 | } | |
4547 | } | |
4548 | mutex_unlock(&slab_mutex); | |
4549 | if (res >= 0) | |
4550 | res = count; | |
4551 | return res; | |
4552 | } | |
4553 | ||
4554 | static int slabinfo_open(struct inode *inode, struct file *file) | |
4555 | { | |
4556 | return seq_open(file, &slabinfo_op); | |
4557 | } | |
4558 | ||
4559 | static const struct file_operations proc_slabinfo_operations = { | |
4560 | .open = slabinfo_open, | |
4561 | .read = seq_read, | |
4562 | .write = slabinfo_write, | |
4563 | .llseek = seq_lseek, | |
4564 | .release = seq_release, | |
4565 | }; | |
4566 | ||
4567 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4568 | ||
4569 | static void *leaks_start(struct seq_file *m, loff_t *pos) | |
4570 | { | |
4571 | mutex_lock(&slab_mutex); | |
4572 | return seq_list_start(&slab_caches, *pos); | |
4573 | } | |
4574 | ||
4575 | static inline int add_caller(unsigned long *n, unsigned long v) | |
4576 | { | |
4577 | unsigned long *p; | |
4578 | int l; | |
4579 | if (!v) | |
4580 | return 1; | |
4581 | l = n[1]; | |
4582 | p = n + 2; | |
4583 | while (l) { | |
4584 | int i = l/2; | |
4585 | unsigned long *q = p + 2 * i; | |
4586 | if (*q == v) { | |
4587 | q[1]++; | |
4588 | return 1; | |
4589 | } | |
4590 | if (*q > v) { | |
4591 | l = i; | |
4592 | } else { | |
4593 | p = q + 2; | |
4594 | l -= i + 1; | |
4595 | } | |
4596 | } | |
4597 | if (++n[1] == n[0]) | |
4598 | return 0; | |
4599 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); | |
4600 | p[0] = v; | |
4601 | p[1] = 1; | |
4602 | return 1; | |
4603 | } | |
4604 | ||
4605 | static void handle_slab(unsigned long *n, struct kmem_cache *c, struct slab *s) | |
4606 | { | |
4607 | void *p; | |
4608 | int i; | |
4609 | if (n[0] == n[1]) | |
4610 | return; | |
4611 | for (i = 0, p = s->s_mem; i < c->num; i++, p += c->size) { | |
4612 | if (slab_bufctl(s)[i] != BUFCTL_ACTIVE) | |
4613 | continue; | |
4614 | if (!add_caller(n, (unsigned long)*dbg_userword(c, p))) | |
4615 | return; | |
4616 | } | |
4617 | } | |
4618 | ||
4619 | static void show_symbol(struct seq_file *m, unsigned long address) | |
4620 | { | |
4621 | #ifdef CONFIG_KALLSYMS | |
4622 | unsigned long offset, size; | |
4623 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; | |
4624 | ||
4625 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { | |
4626 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); | |
4627 | if (modname[0]) | |
4628 | seq_printf(m, " [%s]", modname); | |
4629 | return; | |
4630 | } | |
4631 | #endif | |
4632 | seq_printf(m, "%p", (void *)address); | |
4633 | } | |
4634 | ||
4635 | static int leaks_show(struct seq_file *m, void *p) | |
4636 | { | |
4637 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); | |
4638 | struct slab *slabp; | |
4639 | struct kmem_list3 *l3; | |
4640 | const char *name; | |
4641 | unsigned long *n = m->private; | |
4642 | int node; | |
4643 | int i; | |
4644 | ||
4645 | if (!(cachep->flags & SLAB_STORE_USER)) | |
4646 | return 0; | |
4647 | if (!(cachep->flags & SLAB_RED_ZONE)) | |
4648 | return 0; | |
4649 | ||
4650 | /* OK, we can do it */ | |
4651 | ||
4652 | n[1] = 0; | |
4653 | ||
4654 | for_each_online_node(node) { | |
4655 | l3 = cachep->nodelists[node]; | |
4656 | if (!l3) | |
4657 | continue; | |
4658 | ||
4659 | check_irq_on(); | |
4660 | spin_lock_irq(&l3->list_lock); | |
4661 | ||
4662 | list_for_each_entry(slabp, &l3->slabs_full, list) | |
4663 | handle_slab(n, cachep, slabp); | |
4664 | list_for_each_entry(slabp, &l3->slabs_partial, list) | |
4665 | handle_slab(n, cachep, slabp); | |
4666 | spin_unlock_irq(&l3->list_lock); | |
4667 | } | |
4668 | name = cachep->name; | |
4669 | if (n[0] == n[1]) { | |
4670 | /* Increase the buffer size */ | |
4671 | mutex_unlock(&slab_mutex); | |
4672 | m->private = kzalloc(n[0] * 4 * sizeof(unsigned long), GFP_KERNEL); | |
4673 | if (!m->private) { | |
4674 | /* Too bad, we are really out */ | |
4675 | m->private = n; | |
4676 | mutex_lock(&slab_mutex); | |
4677 | return -ENOMEM; | |
4678 | } | |
4679 | *(unsigned long *)m->private = n[0] * 2; | |
4680 | kfree(n); | |
4681 | mutex_lock(&slab_mutex); | |
4682 | /* Now make sure this entry will be retried */ | |
4683 | m->count = m->size; | |
4684 | return 0; | |
4685 | } | |
4686 | for (i = 0; i < n[1]; i++) { | |
4687 | seq_printf(m, "%s: %lu ", name, n[2*i+3]); | |
4688 | show_symbol(m, n[2*i+2]); | |
4689 | seq_putc(m, '\n'); | |
4690 | } | |
4691 | ||
4692 | return 0; | |
4693 | } | |
4694 | ||
4695 | static const struct seq_operations slabstats_op = { | |
4696 | .start = leaks_start, | |
4697 | .next = s_next, | |
4698 | .stop = s_stop, | |
4699 | .show = leaks_show, | |
4700 | }; | |
4701 | ||
4702 | static int slabstats_open(struct inode *inode, struct file *file) | |
4703 | { | |
4704 | unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL); | |
4705 | int ret = -ENOMEM; | |
4706 | if (n) { | |
4707 | ret = seq_open(file, &slabstats_op); | |
4708 | if (!ret) { | |
4709 | struct seq_file *m = file->private_data; | |
4710 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); | |
4711 | m->private = n; | |
4712 | n = NULL; | |
4713 | } | |
4714 | kfree(n); | |
4715 | } | |
4716 | return ret; | |
4717 | } | |
4718 | ||
4719 | static const struct file_operations proc_slabstats_operations = { | |
4720 | .open = slabstats_open, | |
4721 | .read = seq_read, | |
4722 | .llseek = seq_lseek, | |
4723 | .release = seq_release_private, | |
4724 | }; | |
4725 | #endif | |
4726 | ||
4727 | static int __init slab_proc_init(void) | |
4728 | { | |
4729 | proc_create("slabinfo",S_IWUSR|S_IRUSR,NULL,&proc_slabinfo_operations); | |
4730 | #ifdef CONFIG_DEBUG_SLAB_LEAK | |
4731 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); | |
4732 | #endif | |
4733 | return 0; | |
4734 | } | |
4735 | module_init(slab_proc_init); | |
4736 | #endif | |
4737 | ||
4738 | /** | |
4739 | * ksize - get the actual amount of memory allocated for a given object | |
4740 | * @objp: Pointer to the object | |
4741 | * | |
4742 | * kmalloc may internally round up allocations and return more memory | |
4743 | * than requested. ksize() can be used to determine the actual amount of | |
4744 | * memory allocated. The caller may use this additional memory, even though | |
4745 | * a smaller amount of memory was initially specified with the kmalloc call. | |
4746 | * The caller must guarantee that objp points to a valid object previously | |
4747 | * allocated with either kmalloc() or kmem_cache_alloc(). The object | |
4748 | * must not be freed during the duration of the call. | |
4749 | */ | |
4750 | size_t ksize(const void *objp) | |
4751 | { | |
4752 | BUG_ON(!objp); | |
4753 | if (unlikely(objp == ZERO_SIZE_PTR)) | |
4754 | return 0; | |
4755 | ||
4756 | return virt_to_cache(objp)->object_size; | |
4757 | } | |
4758 | EXPORT_SYMBOL(ksize); |