blob: bd498eff4d7396bacd56d43ee97f390419439a6e
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 <linux/mm.h> |
91 | #include <linux/poison.h> |
92 | #include <linux/swap.h> |
93 | #include <linux/cache.h> |
94 | #include <linux/interrupt.h> |
95 | #include <linux/init.h> |
96 | #include <linux/compiler.h> |
97 | #include <linux/cpuset.h> |
98 | #include <linux/proc_fs.h> |
99 | #include <linux/seq_file.h> |
100 | #include <linux/notifier.h> |
101 | #include <linux/kallsyms.h> |
102 | #include <linux/cpu.h> |
103 | #include <linux/sysctl.h> |
104 | #include <linux/module.h> |
105 | #include <linux/rcupdate.h> |
106 | #include <linux/string.h> |
107 | #include <linux/uaccess.h> |
108 | #include <linux/nodemask.h> |
109 | #include <linux/kmemleak.h> |
110 | #include <linux/mempolicy.h> |
111 | #include <linux/mutex.h> |
112 | #include <linux/fault-inject.h> |
113 | #include <linux/rtmutex.h> |
114 | #include <linux/reciprocal_div.h> |
115 | #include <linux/debugobjects.h> |
116 | #include <linux/kmemcheck.h> |
117 | #include <linux/memory.h> |
118 | #include <linux/prefetch.h> |
119 | |
120 | #include <net/sock.h> |
121 | |
122 | #include <asm/cacheflush.h> |
123 | #include <asm/tlbflush.h> |
124 | #include <asm/page.h> |
125 | |
126 | #include <trace/events/kmem.h> |
127 | |
128 | #include "internal.h" |
129 | |
130 | #include "slab.h" |
131 | |
132 | /* |
133 | * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON. |
134 | * 0 for faster, smaller code (especially in the critical paths). |
135 | * |
136 | * STATS - 1 to collect stats for /proc/slabinfo. |
137 | * 0 for faster, smaller code (especially in the critical paths). |
138 | * |
139 | * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible) |
140 | */ |
141 | |
142 | #ifdef CONFIG_DEBUG_SLAB |
143 | #define DEBUG 1 |
144 | #define STATS 1 |
145 | #define FORCED_DEBUG 1 |
146 | #else |
147 | #define DEBUG 0 |
148 | #define STATS 0 |
149 | #define FORCED_DEBUG 0 |
150 | #endif |
151 | |
152 | /* Shouldn't this be in a header file somewhere? */ |
153 | #define BYTES_PER_WORD sizeof(void *) |
154 | #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long)) |
155 | |
156 | #ifndef ARCH_KMALLOC_FLAGS |
157 | #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN |
158 | #endif |
159 | |
160 | #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \ |
161 | <= SLAB_OBJ_MIN_SIZE) ? 1 : 0) |
162 | |
163 | #if FREELIST_BYTE_INDEX |
164 | typedef unsigned char freelist_idx_t; |
165 | #else |
166 | typedef unsigned short freelist_idx_t; |
167 | #endif |
168 | |
169 | #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1) |
170 | |
171 | /* |
172 | * struct array_cache |
173 | * |
174 | * Purpose: |
175 | * - LIFO ordering, to hand out cache-warm objects from _alloc |
176 | * - reduce the number of linked list operations |
177 | * - reduce spinlock operations |
178 | * |
179 | * The limit is stored in the per-cpu structure to reduce the data cache |
180 | * footprint. |
181 | * |
182 | */ |
183 | struct array_cache { |
184 | unsigned int avail; |
185 | unsigned int limit; |
186 | unsigned int batchcount; |
187 | unsigned int touched; |
188 | void *entry[]; /* |
189 | * Must have this definition in here for the proper |
190 | * alignment of array_cache. Also simplifies accessing |
191 | * the entries. |
192 | */ |
193 | }; |
194 | |
195 | struct alien_cache { |
196 | spinlock_t lock; |
197 | struct array_cache ac; |
198 | }; |
199 | |
200 | /* |
201 | * Need this for bootstrapping a per node allocator. |
202 | */ |
203 | #define NUM_INIT_LISTS (2 * MAX_NUMNODES) |
204 | static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS]; |
205 | #define CACHE_CACHE 0 |
206 | #define SIZE_NODE (MAX_NUMNODES) |
207 | |
208 | static int drain_freelist(struct kmem_cache *cache, |
209 | struct kmem_cache_node *n, int tofree); |
210 | static void free_block(struct kmem_cache *cachep, void **objpp, int len, |
211 | int node, struct list_head *list); |
212 | static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list); |
213 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp); |
214 | static void cache_reap(struct work_struct *unused); |
215 | |
216 | static inline void fixup_objfreelist_debug(struct kmem_cache *cachep, |
217 | void **list); |
218 | static inline void fixup_slab_list(struct kmem_cache *cachep, |
219 | struct kmem_cache_node *n, struct page *page, |
220 | void **list); |
221 | static int slab_early_init = 1; |
222 | |
223 | #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node)) |
224 | |
225 | static void kmem_cache_node_init(struct kmem_cache_node *parent) |
226 | { |
227 | INIT_LIST_HEAD(&parent->slabs_full); |
228 | INIT_LIST_HEAD(&parent->slabs_partial); |
229 | INIT_LIST_HEAD(&parent->slabs_free); |
230 | parent->shared = NULL; |
231 | parent->alien = NULL; |
232 | parent->colour_next = 0; |
233 | spin_lock_init(&parent->list_lock); |
234 | parent->free_objects = 0; |
235 | parent->free_touched = 0; |
236 | parent->num_slabs = 0; |
237 | } |
238 | |
239 | #define MAKE_LIST(cachep, listp, slab, nodeid) \ |
240 | do { \ |
241 | INIT_LIST_HEAD(listp); \ |
242 | list_splice(&get_node(cachep, nodeid)->slab, listp); \ |
243 | } while (0) |
244 | |
245 | #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \ |
246 | do { \ |
247 | MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \ |
248 | MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \ |
249 | MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \ |
250 | } while (0) |
251 | |
252 | #define CFLGS_OBJFREELIST_SLAB (0x40000000UL) |
253 | #define CFLGS_OFF_SLAB (0x80000000UL) |
254 | #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB) |
255 | #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB) |
256 | |
257 | #define BATCHREFILL_LIMIT 16 |
258 | /* |
259 | * Optimization question: fewer reaps means less probability for unnessary |
260 | * cpucache drain/refill cycles. |
261 | * |
262 | * OTOH the cpuarrays can contain lots of objects, |
263 | * which could lock up otherwise freeable slabs. |
264 | */ |
265 | #define REAPTIMEOUT_AC (2*HZ) |
266 | #define REAPTIMEOUT_NODE (4*HZ) |
267 | |
268 | #if STATS |
269 | #define STATS_INC_ACTIVE(x) ((x)->num_active++) |
270 | #define STATS_DEC_ACTIVE(x) ((x)->num_active--) |
271 | #define STATS_INC_ALLOCED(x) ((x)->num_allocations++) |
272 | #define STATS_INC_GROWN(x) ((x)->grown++) |
273 | #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y)) |
274 | #define STATS_SET_HIGH(x) \ |
275 | do { \ |
276 | if ((x)->num_active > (x)->high_mark) \ |
277 | (x)->high_mark = (x)->num_active; \ |
278 | } while (0) |
279 | #define STATS_INC_ERR(x) ((x)->errors++) |
280 | #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++) |
281 | #define STATS_INC_NODEFREES(x) ((x)->node_frees++) |
282 | #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++) |
283 | #define STATS_SET_FREEABLE(x, i) \ |
284 | do { \ |
285 | if ((x)->max_freeable < i) \ |
286 | (x)->max_freeable = i; \ |
287 | } while (0) |
288 | #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit) |
289 | #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss) |
290 | #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit) |
291 | #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss) |
292 | #else |
293 | #define STATS_INC_ACTIVE(x) do { } while (0) |
294 | #define STATS_DEC_ACTIVE(x) do { } while (0) |
295 | #define STATS_INC_ALLOCED(x) do { } while (0) |
296 | #define STATS_INC_GROWN(x) do { } while (0) |
297 | #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0) |
298 | #define STATS_SET_HIGH(x) do { } while (0) |
299 | #define STATS_INC_ERR(x) do { } while (0) |
300 | #define STATS_INC_NODEALLOCS(x) do { } while (0) |
301 | #define STATS_INC_NODEFREES(x) do { } while (0) |
302 | #define STATS_INC_ACOVERFLOW(x) do { } while (0) |
303 | #define STATS_SET_FREEABLE(x, i) do { } while (0) |
304 | #define STATS_INC_ALLOCHIT(x) do { } while (0) |
305 | #define STATS_INC_ALLOCMISS(x) do { } while (0) |
306 | #define STATS_INC_FREEHIT(x) do { } while (0) |
307 | #define STATS_INC_FREEMISS(x) do { } while (0) |
308 | #endif |
309 | |
310 | #if DEBUG |
311 | |
312 | /* |
313 | * memory layout of objects: |
314 | * 0 : objp |
315 | * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that |
316 | * the end of an object is aligned with the end of the real |
317 | * allocation. Catches writes behind the end of the allocation. |
318 | * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1: |
319 | * redzone word. |
320 | * cachep->obj_offset: The real object. |
321 | * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long] |
322 | * cachep->size - 1* BYTES_PER_WORD: last caller address |
323 | * [BYTES_PER_WORD long] |
324 | */ |
325 | static int obj_offset(struct kmem_cache *cachep) |
326 | { |
327 | return cachep->obj_offset; |
328 | } |
329 | |
330 | static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp) |
331 | { |
332 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); |
333 | return (unsigned long long*) (objp + obj_offset(cachep) - |
334 | sizeof(unsigned long long)); |
335 | } |
336 | |
337 | static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp) |
338 | { |
339 | BUG_ON(!(cachep->flags & SLAB_RED_ZONE)); |
340 | if (cachep->flags & SLAB_STORE_USER) |
341 | return (unsigned long long *)(objp + cachep->size - |
342 | sizeof(unsigned long long) - |
343 | REDZONE_ALIGN); |
344 | return (unsigned long long *) (objp + cachep->size - |
345 | sizeof(unsigned long long)); |
346 | } |
347 | |
348 | static void **dbg_userword(struct kmem_cache *cachep, void *objp) |
349 | { |
350 | BUG_ON(!(cachep->flags & SLAB_STORE_USER)); |
351 | return (void **)(objp + cachep->size - BYTES_PER_WORD); |
352 | } |
353 | |
354 | #else |
355 | |
356 | #define obj_offset(x) 0 |
357 | #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
358 | #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;}) |
359 | #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;}) |
360 | |
361 | #endif |
362 | |
363 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
364 | |
365 | static inline bool is_store_user_clean(struct kmem_cache *cachep) |
366 | { |
367 | return atomic_read(&cachep->store_user_clean) == 1; |
368 | } |
369 | |
370 | static inline void set_store_user_clean(struct kmem_cache *cachep) |
371 | { |
372 | atomic_set(&cachep->store_user_clean, 1); |
373 | } |
374 | |
375 | static inline void set_store_user_dirty(struct kmem_cache *cachep) |
376 | { |
377 | if (is_store_user_clean(cachep)) |
378 | atomic_set(&cachep->store_user_clean, 0); |
379 | } |
380 | |
381 | #else |
382 | static inline void set_store_user_dirty(struct kmem_cache *cachep) {} |
383 | |
384 | #endif |
385 | |
386 | /* |
387 | * Do not go above this order unless 0 objects fit into the slab or |
388 | * overridden on the command line. |
389 | */ |
390 | #define SLAB_MAX_ORDER_HI 1 |
391 | #define SLAB_MAX_ORDER_LO 0 |
392 | static int slab_max_order = SLAB_MAX_ORDER_LO; |
393 | static bool slab_max_order_set __initdata; |
394 | |
395 | static inline struct kmem_cache *virt_to_cache(const void *obj) |
396 | { |
397 | struct page *page = virt_to_head_page(obj); |
398 | return page->slab_cache; |
399 | } |
400 | |
401 | static inline void *index_to_obj(struct kmem_cache *cache, struct page *page, |
402 | unsigned int idx) |
403 | { |
404 | return page->s_mem + cache->size * idx; |
405 | } |
406 | |
407 | /* |
408 | * We want to avoid an expensive divide : (offset / cache->size) |
409 | * Using the fact that size is a constant for a particular cache, |
410 | * we can replace (offset / cache->size) by |
411 | * reciprocal_divide(offset, cache->reciprocal_buffer_size) |
412 | */ |
413 | static inline unsigned int obj_to_index(const struct kmem_cache *cache, |
414 | const struct page *page, void *obj) |
415 | { |
416 | u32 offset = (obj - page->s_mem); |
417 | return reciprocal_divide(offset, cache->reciprocal_buffer_size); |
418 | } |
419 | |
420 | #define BOOT_CPUCACHE_ENTRIES 1 |
421 | /* internal cache of cache description objs */ |
422 | static struct kmem_cache kmem_cache_boot = { |
423 | .batchcount = 1, |
424 | .limit = BOOT_CPUCACHE_ENTRIES, |
425 | .shared = 1, |
426 | .size = sizeof(struct kmem_cache), |
427 | .name = "kmem_cache", |
428 | }; |
429 | |
430 | static DEFINE_PER_CPU(struct delayed_work, slab_reap_work); |
431 | |
432 | static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep) |
433 | { |
434 | return this_cpu_ptr(cachep->cpu_cache); |
435 | } |
436 | |
437 | /* |
438 | * Calculate the number of objects and left-over bytes for a given buffer size. |
439 | */ |
440 | static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size, |
441 | unsigned long flags, size_t *left_over) |
442 | { |
443 | unsigned int num; |
444 | size_t slab_size = PAGE_SIZE << gfporder; |
445 | |
446 | /* |
447 | * The slab management structure can be either off the slab or |
448 | * on it. For the latter case, the memory allocated for a |
449 | * slab is used for: |
450 | * |
451 | * - @buffer_size bytes for each object |
452 | * - One freelist_idx_t for each object |
453 | * |
454 | * We don't need to consider alignment of freelist because |
455 | * freelist will be at the end of slab page. The objects will be |
456 | * at the correct alignment. |
457 | * |
458 | * If the slab management structure is off the slab, then the |
459 | * alignment will already be calculated into the size. Because |
460 | * the slabs are all pages aligned, the objects will be at the |
461 | * correct alignment when allocated. |
462 | */ |
463 | if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) { |
464 | num = slab_size / buffer_size; |
465 | *left_over = slab_size % buffer_size; |
466 | } else { |
467 | num = slab_size / (buffer_size + sizeof(freelist_idx_t)); |
468 | *left_over = slab_size % |
469 | (buffer_size + sizeof(freelist_idx_t)); |
470 | } |
471 | |
472 | return num; |
473 | } |
474 | |
475 | #if DEBUG |
476 | #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg) |
477 | |
478 | static void __slab_error(const char *function, struct kmem_cache *cachep, |
479 | char *msg) |
480 | { |
481 | pr_err("slab error in %s(): cache `%s': %s\n", |
482 | function, cachep->name, msg); |
483 | dump_stack(); |
484 | add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE); |
485 | } |
486 | #endif |
487 | |
488 | /* |
489 | * By default on NUMA we use alien caches to stage the freeing of |
490 | * objects allocated from other nodes. This causes massive memory |
491 | * inefficiencies when using fake NUMA setup to split memory into a |
492 | * large number of small nodes, so it can be disabled on the command |
493 | * line |
494 | */ |
495 | |
496 | static int use_alien_caches __read_mostly = 1; |
497 | static int __init noaliencache_setup(char *s) |
498 | { |
499 | use_alien_caches = 0; |
500 | return 1; |
501 | } |
502 | __setup("noaliencache", noaliencache_setup); |
503 | |
504 | static int __init slab_max_order_setup(char *str) |
505 | { |
506 | get_option(&str, &slab_max_order); |
507 | slab_max_order = slab_max_order < 0 ? 0 : |
508 | min(slab_max_order, MAX_ORDER - 1); |
509 | slab_max_order_set = true; |
510 | |
511 | return 1; |
512 | } |
513 | __setup("slab_max_order=", slab_max_order_setup); |
514 | |
515 | #ifdef CONFIG_NUMA |
516 | /* |
517 | * Special reaping functions for NUMA systems called from cache_reap(). |
518 | * These take care of doing round robin flushing of alien caches (containing |
519 | * objects freed on different nodes from which they were allocated) and the |
520 | * flushing of remote pcps by calling drain_node_pages. |
521 | */ |
522 | static DEFINE_PER_CPU(unsigned long, slab_reap_node); |
523 | |
524 | static void init_reap_node(int cpu) |
525 | { |
526 | per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu), |
527 | node_online_map); |
528 | } |
529 | |
530 | static void next_reap_node(void) |
531 | { |
532 | int node = __this_cpu_read(slab_reap_node); |
533 | |
534 | node = next_node_in(node, node_online_map); |
535 | __this_cpu_write(slab_reap_node, node); |
536 | } |
537 | |
538 | #else |
539 | #define init_reap_node(cpu) do { } while (0) |
540 | #define next_reap_node(void) do { } while (0) |
541 | #endif |
542 | |
543 | /* |
544 | * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz |
545 | * via the workqueue/eventd. |
546 | * Add the CPU number into the expiration time to minimize the possibility of |
547 | * the CPUs getting into lockstep and contending for the global cache chain |
548 | * lock. |
549 | */ |
550 | static void start_cpu_timer(int cpu) |
551 | { |
552 | struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu); |
553 | |
554 | /* |
555 | * When this gets called from do_initcalls via cpucache_init(), |
556 | * init_workqueues() has already run, so keventd will be setup |
557 | * at that time. |
558 | */ |
559 | if (keventd_up() && reap_work->work.func == NULL) { |
560 | init_reap_node(cpu); |
561 | INIT_DEFERRABLE_WORK(reap_work, cache_reap); |
562 | schedule_delayed_work_on(cpu, reap_work, |
563 | __round_jiffies_relative(HZ, cpu)); |
564 | } |
565 | } |
566 | |
567 | static void init_arraycache(struct array_cache *ac, int limit, int batch) |
568 | { |
569 | if (ac) { |
570 | ac->avail = 0; |
571 | ac->limit = limit; |
572 | ac->batchcount = batch; |
573 | ac->touched = 0; |
574 | } |
575 | } |
576 | |
577 | static struct array_cache *alloc_arraycache(int node, int entries, |
578 | int batchcount, gfp_t gfp) |
579 | { |
580 | size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache); |
581 | struct array_cache *ac = NULL; |
582 | |
583 | ac = kmalloc_node(memsize, gfp, node); |
584 | /* |
585 | * The array_cache structures contain pointers to free object. |
586 | * However, when such objects are allocated or transferred to another |
587 | * cache the pointers are not cleared and they could be counted as |
588 | * valid references during a kmemleak scan. Therefore, kmemleak must |
589 | * not scan such objects. |
590 | */ |
591 | kmemleak_no_scan(ac); |
592 | init_arraycache(ac, entries, batchcount); |
593 | return ac; |
594 | } |
595 | |
596 | static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep, |
597 | struct page *page, void *objp) |
598 | { |
599 | struct kmem_cache_node *n; |
600 | int page_node; |
601 | LIST_HEAD(list); |
602 | |
603 | page_node = page_to_nid(page); |
604 | n = get_node(cachep, page_node); |
605 | |
606 | spin_lock(&n->list_lock); |
607 | free_block(cachep, &objp, 1, page_node, &list); |
608 | spin_unlock(&n->list_lock); |
609 | |
610 | slabs_destroy(cachep, &list); |
611 | } |
612 | |
613 | /* |
614 | * Transfer objects in one arraycache to another. |
615 | * Locking must be handled by the caller. |
616 | * |
617 | * Return the number of entries transferred. |
618 | */ |
619 | static int transfer_objects(struct array_cache *to, |
620 | struct array_cache *from, unsigned int max) |
621 | { |
622 | /* Figure out how many entries to transfer */ |
623 | int nr = min3(from->avail, max, to->limit - to->avail); |
624 | |
625 | if (!nr) |
626 | return 0; |
627 | |
628 | memcpy(to->entry + to->avail, from->entry + from->avail -nr, |
629 | sizeof(void *) *nr); |
630 | |
631 | from->avail -= nr; |
632 | to->avail += nr; |
633 | return nr; |
634 | } |
635 | |
636 | #ifndef CONFIG_NUMA |
637 | |
638 | #define drain_alien_cache(cachep, alien) do { } while (0) |
639 | #define reap_alien(cachep, n) do { } while (0) |
640 | |
641 | static inline struct alien_cache **alloc_alien_cache(int node, |
642 | int limit, gfp_t gfp) |
643 | { |
644 | return NULL; |
645 | } |
646 | |
647 | static inline void free_alien_cache(struct alien_cache **ac_ptr) |
648 | { |
649 | } |
650 | |
651 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
652 | { |
653 | return 0; |
654 | } |
655 | |
656 | static inline void *alternate_node_alloc(struct kmem_cache *cachep, |
657 | gfp_t flags) |
658 | { |
659 | return NULL; |
660 | } |
661 | |
662 | static inline void *____cache_alloc_node(struct kmem_cache *cachep, |
663 | gfp_t flags, int nodeid) |
664 | { |
665 | return NULL; |
666 | } |
667 | |
668 | static inline gfp_t gfp_exact_node(gfp_t flags) |
669 | { |
670 | return flags & ~__GFP_NOFAIL; |
671 | } |
672 | |
673 | #else /* CONFIG_NUMA */ |
674 | |
675 | static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int); |
676 | static void *alternate_node_alloc(struct kmem_cache *, gfp_t); |
677 | |
678 | static struct alien_cache *__alloc_alien_cache(int node, int entries, |
679 | int batch, gfp_t gfp) |
680 | { |
681 | size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache); |
682 | struct alien_cache *alc = NULL; |
683 | |
684 | alc = kmalloc_node(memsize, gfp, node); |
685 | if (alc) { |
686 | kmemleak_no_scan(alc); |
687 | init_arraycache(&alc->ac, entries, batch); |
688 | spin_lock_init(&alc->lock); |
689 | } |
690 | return alc; |
691 | } |
692 | |
693 | static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp) |
694 | { |
695 | struct alien_cache **alc_ptr; |
696 | size_t memsize = sizeof(void *) * nr_node_ids; |
697 | int i; |
698 | |
699 | if (limit > 1) |
700 | limit = 12; |
701 | alc_ptr = kzalloc_node(memsize, gfp, node); |
702 | if (!alc_ptr) |
703 | return NULL; |
704 | |
705 | for_each_node(i) { |
706 | if (i == node || !node_online(i)) |
707 | continue; |
708 | alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp); |
709 | if (!alc_ptr[i]) { |
710 | for (i--; i >= 0; i--) |
711 | kfree(alc_ptr[i]); |
712 | kfree(alc_ptr); |
713 | return NULL; |
714 | } |
715 | } |
716 | return alc_ptr; |
717 | } |
718 | |
719 | static void free_alien_cache(struct alien_cache **alc_ptr) |
720 | { |
721 | int i; |
722 | |
723 | if (!alc_ptr) |
724 | return; |
725 | for_each_node(i) |
726 | kfree(alc_ptr[i]); |
727 | kfree(alc_ptr); |
728 | } |
729 | |
730 | static void __drain_alien_cache(struct kmem_cache *cachep, |
731 | struct array_cache *ac, int node, |
732 | struct list_head *list) |
733 | { |
734 | struct kmem_cache_node *n = get_node(cachep, node); |
735 | |
736 | if (ac->avail) { |
737 | spin_lock(&n->list_lock); |
738 | /* |
739 | * Stuff objects into the remote nodes shared array first. |
740 | * That way we could avoid the overhead of putting the objects |
741 | * into the free lists and getting them back later. |
742 | */ |
743 | if (n->shared) |
744 | transfer_objects(n->shared, ac, ac->limit); |
745 | |
746 | free_block(cachep, ac->entry, ac->avail, node, list); |
747 | ac->avail = 0; |
748 | spin_unlock(&n->list_lock); |
749 | } |
750 | } |
751 | |
752 | /* |
753 | * Called from cache_reap() to regularly drain alien caches round robin. |
754 | */ |
755 | static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n) |
756 | { |
757 | int node = __this_cpu_read(slab_reap_node); |
758 | |
759 | if (n->alien) { |
760 | struct alien_cache *alc = n->alien[node]; |
761 | struct array_cache *ac; |
762 | |
763 | if (alc) { |
764 | ac = &alc->ac; |
765 | if (ac->avail && spin_trylock_irq(&alc->lock)) { |
766 | LIST_HEAD(list); |
767 | |
768 | __drain_alien_cache(cachep, ac, node, &list); |
769 | spin_unlock_irq(&alc->lock); |
770 | slabs_destroy(cachep, &list); |
771 | } |
772 | } |
773 | } |
774 | } |
775 | |
776 | static void drain_alien_cache(struct kmem_cache *cachep, |
777 | struct alien_cache **alien) |
778 | { |
779 | int i = 0; |
780 | struct alien_cache *alc; |
781 | struct array_cache *ac; |
782 | unsigned long flags; |
783 | |
784 | for_each_online_node(i) { |
785 | alc = alien[i]; |
786 | if (alc) { |
787 | LIST_HEAD(list); |
788 | |
789 | ac = &alc->ac; |
790 | spin_lock_irqsave(&alc->lock, flags); |
791 | __drain_alien_cache(cachep, ac, i, &list); |
792 | spin_unlock_irqrestore(&alc->lock, flags); |
793 | slabs_destroy(cachep, &list); |
794 | } |
795 | } |
796 | } |
797 | |
798 | static int __cache_free_alien(struct kmem_cache *cachep, void *objp, |
799 | int node, int page_node) |
800 | { |
801 | struct kmem_cache_node *n; |
802 | struct alien_cache *alien = NULL; |
803 | struct array_cache *ac; |
804 | LIST_HEAD(list); |
805 | |
806 | n = get_node(cachep, node); |
807 | STATS_INC_NODEFREES(cachep); |
808 | if (n->alien && n->alien[page_node]) { |
809 | alien = n->alien[page_node]; |
810 | ac = &alien->ac; |
811 | spin_lock(&alien->lock); |
812 | if (unlikely(ac->avail == ac->limit)) { |
813 | STATS_INC_ACOVERFLOW(cachep); |
814 | __drain_alien_cache(cachep, ac, page_node, &list); |
815 | } |
816 | ac->entry[ac->avail++] = objp; |
817 | spin_unlock(&alien->lock); |
818 | slabs_destroy(cachep, &list); |
819 | } else { |
820 | n = get_node(cachep, page_node); |
821 | spin_lock(&n->list_lock); |
822 | free_block(cachep, &objp, 1, page_node, &list); |
823 | spin_unlock(&n->list_lock); |
824 | slabs_destroy(cachep, &list); |
825 | } |
826 | return 1; |
827 | } |
828 | |
829 | static inline int cache_free_alien(struct kmem_cache *cachep, void *objp) |
830 | { |
831 | int page_node = page_to_nid(virt_to_page(objp)); |
832 | int node = numa_mem_id(); |
833 | /* |
834 | * Make sure we are not freeing a object from another node to the array |
835 | * cache on this cpu. |
836 | */ |
837 | if (likely(node == page_node)) |
838 | return 0; |
839 | |
840 | return __cache_free_alien(cachep, objp, node, page_node); |
841 | } |
842 | |
843 | /* |
844 | * Construct gfp mask to allocate from a specific node but do not reclaim or |
845 | * warn about failures. |
846 | */ |
847 | static inline gfp_t gfp_exact_node(gfp_t flags) |
848 | { |
849 | return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL); |
850 | } |
851 | #endif |
852 | |
853 | static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp) |
854 | { |
855 | struct kmem_cache_node *n; |
856 | |
857 | /* |
858 | * Set up the kmem_cache_node for cpu before we can |
859 | * begin anything. Make sure some other cpu on this |
860 | * node has not already allocated this |
861 | */ |
862 | n = get_node(cachep, node); |
863 | if (n) { |
864 | spin_lock_irq(&n->list_lock); |
865 | n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount + |
866 | cachep->num; |
867 | spin_unlock_irq(&n->list_lock); |
868 | |
869 | return 0; |
870 | } |
871 | |
872 | n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node); |
873 | if (!n) |
874 | return -ENOMEM; |
875 | |
876 | kmem_cache_node_init(n); |
877 | n->next_reap = jiffies + REAPTIMEOUT_NODE + |
878 | ((unsigned long)cachep) % REAPTIMEOUT_NODE; |
879 | |
880 | n->free_limit = |
881 | (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num; |
882 | |
883 | /* |
884 | * The kmem_cache_nodes don't come and go as CPUs |
885 | * come and go. slab_mutex is sufficient |
886 | * protection here. |
887 | */ |
888 | cachep->node[node] = n; |
889 | |
890 | return 0; |
891 | } |
892 | |
893 | #if (defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)) || defined(CONFIG_SMP) |
894 | /* |
895 | * Allocates and initializes node for a node on each slab cache, used for |
896 | * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node |
897 | * will be allocated off-node since memory is not yet online for the new node. |
898 | * When hotplugging memory or a cpu, existing node are not replaced if |
899 | * already in use. |
900 | * |
901 | * Must hold slab_mutex. |
902 | */ |
903 | static int init_cache_node_node(int node) |
904 | { |
905 | int ret; |
906 | struct kmem_cache *cachep; |
907 | |
908 | list_for_each_entry(cachep, &slab_caches, list) { |
909 | ret = init_cache_node(cachep, node, GFP_KERNEL); |
910 | if (ret) |
911 | return ret; |
912 | } |
913 | |
914 | return 0; |
915 | } |
916 | #endif |
917 | |
918 | static int setup_kmem_cache_node(struct kmem_cache *cachep, |
919 | int node, gfp_t gfp, bool force_change) |
920 | { |
921 | int ret = -ENOMEM; |
922 | struct kmem_cache_node *n; |
923 | struct array_cache *old_shared = NULL; |
924 | struct array_cache *new_shared = NULL; |
925 | struct alien_cache **new_alien = NULL; |
926 | LIST_HEAD(list); |
927 | |
928 | if (use_alien_caches) { |
929 | new_alien = alloc_alien_cache(node, cachep->limit, gfp); |
930 | if (!new_alien) |
931 | goto fail; |
932 | } |
933 | |
934 | if (cachep->shared) { |
935 | new_shared = alloc_arraycache(node, |
936 | cachep->shared * cachep->batchcount, 0xbaadf00d, gfp); |
937 | if (!new_shared) |
938 | goto fail; |
939 | } |
940 | |
941 | ret = init_cache_node(cachep, node, gfp); |
942 | if (ret) |
943 | goto fail; |
944 | |
945 | n = get_node(cachep, node); |
946 | spin_lock_irq(&n->list_lock); |
947 | if (n->shared && force_change) { |
948 | free_block(cachep, n->shared->entry, |
949 | n->shared->avail, node, &list); |
950 | n->shared->avail = 0; |
951 | } |
952 | |
953 | if (!n->shared || force_change) { |
954 | old_shared = n->shared; |
955 | n->shared = new_shared; |
956 | new_shared = NULL; |
957 | } |
958 | |
959 | if (!n->alien) { |
960 | n->alien = new_alien; |
961 | new_alien = NULL; |
962 | } |
963 | |
964 | spin_unlock_irq(&n->list_lock); |
965 | slabs_destroy(cachep, &list); |
966 | |
967 | /* |
968 | * To protect lockless access to n->shared during irq disabled context. |
969 | * If n->shared isn't NULL in irq disabled context, accessing to it is |
970 | * guaranteed to be valid until irq is re-enabled, because it will be |
971 | * freed after synchronize_sched(). |
972 | */ |
973 | if (old_shared && force_change) |
974 | synchronize_sched(); |
975 | |
976 | fail: |
977 | kfree(old_shared); |
978 | kfree(new_shared); |
979 | free_alien_cache(new_alien); |
980 | |
981 | return ret; |
982 | } |
983 | |
984 | #ifdef CONFIG_SMP |
985 | |
986 | static void cpuup_canceled(long cpu) |
987 | { |
988 | struct kmem_cache *cachep; |
989 | struct kmem_cache_node *n = NULL; |
990 | int node = cpu_to_mem(cpu); |
991 | const struct cpumask *mask = cpumask_of_node(node); |
992 | |
993 | list_for_each_entry(cachep, &slab_caches, list) { |
994 | struct array_cache *nc; |
995 | struct array_cache *shared; |
996 | struct alien_cache **alien; |
997 | LIST_HEAD(list); |
998 | |
999 | n = get_node(cachep, node); |
1000 | if (!n) |
1001 | continue; |
1002 | |
1003 | spin_lock_irq(&n->list_lock); |
1004 | |
1005 | /* Free limit for this kmem_cache_node */ |
1006 | n->free_limit -= cachep->batchcount; |
1007 | |
1008 | /* cpu is dead; no one can alloc from it. */ |
1009 | nc = per_cpu_ptr(cachep->cpu_cache, cpu); |
1010 | if (nc) { |
1011 | free_block(cachep, nc->entry, nc->avail, node, &list); |
1012 | nc->avail = 0; |
1013 | } |
1014 | |
1015 | if (!cpumask_empty(mask)) { |
1016 | spin_unlock_irq(&n->list_lock); |
1017 | goto free_slab; |
1018 | } |
1019 | |
1020 | shared = n->shared; |
1021 | if (shared) { |
1022 | free_block(cachep, shared->entry, |
1023 | shared->avail, node, &list); |
1024 | n->shared = NULL; |
1025 | } |
1026 | |
1027 | alien = n->alien; |
1028 | n->alien = NULL; |
1029 | |
1030 | spin_unlock_irq(&n->list_lock); |
1031 | |
1032 | kfree(shared); |
1033 | if (alien) { |
1034 | drain_alien_cache(cachep, alien); |
1035 | free_alien_cache(alien); |
1036 | } |
1037 | |
1038 | free_slab: |
1039 | slabs_destroy(cachep, &list); |
1040 | } |
1041 | /* |
1042 | * In the previous loop, all the objects were freed to |
1043 | * the respective cache's slabs, now we can go ahead and |
1044 | * shrink each nodelist to its limit. |
1045 | */ |
1046 | list_for_each_entry(cachep, &slab_caches, list) { |
1047 | n = get_node(cachep, node); |
1048 | if (!n) |
1049 | continue; |
1050 | drain_freelist(cachep, n, INT_MAX); |
1051 | } |
1052 | } |
1053 | |
1054 | static int cpuup_prepare(long cpu) |
1055 | { |
1056 | struct kmem_cache *cachep; |
1057 | int node = cpu_to_mem(cpu); |
1058 | int err; |
1059 | |
1060 | /* |
1061 | * We need to do this right in the beginning since |
1062 | * alloc_arraycache's are going to use this list. |
1063 | * kmalloc_node allows us to add the slab to the right |
1064 | * kmem_cache_node and not this cpu's kmem_cache_node |
1065 | */ |
1066 | err = init_cache_node_node(node); |
1067 | if (err < 0) |
1068 | goto bad; |
1069 | |
1070 | /* |
1071 | * Now we can go ahead with allocating the shared arrays and |
1072 | * array caches |
1073 | */ |
1074 | list_for_each_entry(cachep, &slab_caches, list) { |
1075 | err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false); |
1076 | if (err) |
1077 | goto bad; |
1078 | } |
1079 | |
1080 | return 0; |
1081 | bad: |
1082 | cpuup_canceled(cpu); |
1083 | return -ENOMEM; |
1084 | } |
1085 | |
1086 | int slab_prepare_cpu(unsigned int cpu) |
1087 | { |
1088 | int err; |
1089 | |
1090 | mutex_lock(&slab_mutex); |
1091 | err = cpuup_prepare(cpu); |
1092 | mutex_unlock(&slab_mutex); |
1093 | return err; |
1094 | } |
1095 | |
1096 | /* |
1097 | * This is called for a failed online attempt and for a successful |
1098 | * offline. |
1099 | * |
1100 | * Even if all the cpus of a node are down, we don't free the |
1101 | * kmem_list3 of any cache. This to avoid a race between cpu_down, and |
1102 | * a kmalloc allocation from another cpu for memory from the node of |
1103 | * the cpu going down. The list3 structure is usually allocated from |
1104 | * kmem_cache_create() and gets destroyed at kmem_cache_destroy(). |
1105 | */ |
1106 | int slab_dead_cpu(unsigned int cpu) |
1107 | { |
1108 | mutex_lock(&slab_mutex); |
1109 | cpuup_canceled(cpu); |
1110 | mutex_unlock(&slab_mutex); |
1111 | return 0; |
1112 | } |
1113 | #endif |
1114 | |
1115 | static int slab_online_cpu(unsigned int cpu) |
1116 | { |
1117 | start_cpu_timer(cpu); |
1118 | return 0; |
1119 | } |
1120 | |
1121 | static int slab_offline_cpu(unsigned int cpu) |
1122 | { |
1123 | /* |
1124 | * Shutdown cache reaper. Note that the slab_mutex is held so |
1125 | * that if cache_reap() is invoked it cannot do anything |
1126 | * expensive but will only modify reap_work and reschedule the |
1127 | * timer. |
1128 | */ |
1129 | cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu)); |
1130 | /* Now the cache_reaper is guaranteed to be not running. */ |
1131 | per_cpu(slab_reap_work, cpu).work.func = NULL; |
1132 | return 0; |
1133 | } |
1134 | |
1135 | #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG) |
1136 | /* |
1137 | * Drains freelist for a node on each slab cache, used for memory hot-remove. |
1138 | * Returns -EBUSY if all objects cannot be drained so that the node is not |
1139 | * removed. |
1140 | * |
1141 | * Must hold slab_mutex. |
1142 | */ |
1143 | static int __meminit drain_cache_node_node(int node) |
1144 | { |
1145 | struct kmem_cache *cachep; |
1146 | int ret = 0; |
1147 | |
1148 | list_for_each_entry(cachep, &slab_caches, list) { |
1149 | struct kmem_cache_node *n; |
1150 | |
1151 | n = get_node(cachep, node); |
1152 | if (!n) |
1153 | continue; |
1154 | |
1155 | drain_freelist(cachep, n, INT_MAX); |
1156 | |
1157 | if (!list_empty(&n->slabs_full) || |
1158 | !list_empty(&n->slabs_partial)) { |
1159 | ret = -EBUSY; |
1160 | break; |
1161 | } |
1162 | } |
1163 | return ret; |
1164 | } |
1165 | |
1166 | static int __meminit slab_memory_callback(struct notifier_block *self, |
1167 | unsigned long action, void *arg) |
1168 | { |
1169 | struct memory_notify *mnb = arg; |
1170 | int ret = 0; |
1171 | int nid; |
1172 | |
1173 | nid = mnb->status_change_nid; |
1174 | if (nid < 0) |
1175 | goto out; |
1176 | |
1177 | switch (action) { |
1178 | case MEM_GOING_ONLINE: |
1179 | mutex_lock(&slab_mutex); |
1180 | ret = init_cache_node_node(nid); |
1181 | mutex_unlock(&slab_mutex); |
1182 | break; |
1183 | case MEM_GOING_OFFLINE: |
1184 | mutex_lock(&slab_mutex); |
1185 | ret = drain_cache_node_node(nid); |
1186 | mutex_unlock(&slab_mutex); |
1187 | break; |
1188 | case MEM_ONLINE: |
1189 | case MEM_OFFLINE: |
1190 | case MEM_CANCEL_ONLINE: |
1191 | case MEM_CANCEL_OFFLINE: |
1192 | break; |
1193 | } |
1194 | out: |
1195 | return notifier_from_errno(ret); |
1196 | } |
1197 | #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */ |
1198 | |
1199 | /* |
1200 | * swap the static kmem_cache_node with kmalloced memory |
1201 | */ |
1202 | static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list, |
1203 | int nodeid) |
1204 | { |
1205 | struct kmem_cache_node *ptr; |
1206 | |
1207 | ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid); |
1208 | BUG_ON(!ptr); |
1209 | |
1210 | memcpy(ptr, list, sizeof(struct kmem_cache_node)); |
1211 | /* |
1212 | * Do not assume that spinlocks can be initialized via memcpy: |
1213 | */ |
1214 | spin_lock_init(&ptr->list_lock); |
1215 | |
1216 | MAKE_ALL_LISTS(cachep, ptr, nodeid); |
1217 | cachep->node[nodeid] = ptr; |
1218 | } |
1219 | |
1220 | /* |
1221 | * For setting up all the kmem_cache_node for cache whose buffer_size is same as |
1222 | * size of kmem_cache_node. |
1223 | */ |
1224 | static void __init set_up_node(struct kmem_cache *cachep, int index) |
1225 | { |
1226 | int node; |
1227 | |
1228 | for_each_online_node(node) { |
1229 | cachep->node[node] = &init_kmem_cache_node[index + node]; |
1230 | cachep->node[node]->next_reap = jiffies + |
1231 | REAPTIMEOUT_NODE + |
1232 | ((unsigned long)cachep) % REAPTIMEOUT_NODE; |
1233 | } |
1234 | } |
1235 | |
1236 | /* |
1237 | * Initialisation. Called after the page allocator have been initialised and |
1238 | * before smp_init(). |
1239 | */ |
1240 | void __init kmem_cache_init(void) |
1241 | { |
1242 | int i; |
1243 | |
1244 | BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) < |
1245 | sizeof(struct rcu_head)); |
1246 | kmem_cache = &kmem_cache_boot; |
1247 | |
1248 | if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1) |
1249 | use_alien_caches = 0; |
1250 | |
1251 | for (i = 0; i < NUM_INIT_LISTS; i++) |
1252 | kmem_cache_node_init(&init_kmem_cache_node[i]); |
1253 | |
1254 | /* |
1255 | * Fragmentation resistance on low memory - only use bigger |
1256 | * page orders on machines with more than 32MB of memory if |
1257 | * not overridden on the command line. |
1258 | */ |
1259 | if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT) |
1260 | slab_max_order = SLAB_MAX_ORDER_HI; |
1261 | |
1262 | /* Bootstrap is tricky, because several objects are allocated |
1263 | * from caches that do not exist yet: |
1264 | * 1) initialize the kmem_cache cache: it contains the struct |
1265 | * kmem_cache structures of all caches, except kmem_cache itself: |
1266 | * kmem_cache is statically allocated. |
1267 | * Initially an __init data area is used for the head array and the |
1268 | * kmem_cache_node structures, it's replaced with a kmalloc allocated |
1269 | * array at the end of the bootstrap. |
1270 | * 2) Create the first kmalloc cache. |
1271 | * The struct kmem_cache for the new cache is allocated normally. |
1272 | * An __init data area is used for the head array. |
1273 | * 3) Create the remaining kmalloc caches, with minimally sized |
1274 | * head arrays. |
1275 | * 4) Replace the __init data head arrays for kmem_cache and the first |
1276 | * kmalloc cache with kmalloc allocated arrays. |
1277 | * 5) Replace the __init data for kmem_cache_node for kmem_cache and |
1278 | * the other cache's with kmalloc allocated memory. |
1279 | * 6) Resize the head arrays of the kmalloc caches to their final sizes. |
1280 | */ |
1281 | |
1282 | /* 1) create the kmem_cache */ |
1283 | |
1284 | /* |
1285 | * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids |
1286 | */ |
1287 | create_boot_cache(kmem_cache, "kmem_cache", |
1288 | offsetof(struct kmem_cache, node) + |
1289 | nr_node_ids * sizeof(struct kmem_cache_node *), |
1290 | SLAB_HWCACHE_ALIGN); |
1291 | list_add(&kmem_cache->list, &slab_caches); |
1292 | slab_state = PARTIAL; |
1293 | |
1294 | /* |
1295 | * Initialize the caches that provide memory for the kmem_cache_node |
1296 | * structures first. Without this, further allocations will bug. |
1297 | */ |
1298 | kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node", |
1299 | kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS); |
1300 | slab_state = PARTIAL_NODE; |
1301 | setup_kmalloc_cache_index_table(); |
1302 | |
1303 | slab_early_init = 0; |
1304 | |
1305 | /* 5) Replace the bootstrap kmem_cache_node */ |
1306 | { |
1307 | int nid; |
1308 | |
1309 | for_each_online_node(nid) { |
1310 | init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid); |
1311 | |
1312 | init_list(kmalloc_caches[INDEX_NODE], |
1313 | &init_kmem_cache_node[SIZE_NODE + nid], nid); |
1314 | } |
1315 | } |
1316 | |
1317 | create_kmalloc_caches(ARCH_KMALLOC_FLAGS); |
1318 | } |
1319 | |
1320 | void __init kmem_cache_init_late(void) |
1321 | { |
1322 | struct kmem_cache *cachep; |
1323 | |
1324 | slab_state = UP; |
1325 | |
1326 | /* 6) resize the head arrays to their final sizes */ |
1327 | mutex_lock(&slab_mutex); |
1328 | list_for_each_entry(cachep, &slab_caches, list) |
1329 | if (enable_cpucache(cachep, GFP_NOWAIT)) |
1330 | BUG(); |
1331 | mutex_unlock(&slab_mutex); |
1332 | |
1333 | /* Done! */ |
1334 | slab_state = FULL; |
1335 | |
1336 | #ifdef CONFIG_NUMA |
1337 | /* |
1338 | * Register a memory hotplug callback that initializes and frees |
1339 | * node. |
1340 | */ |
1341 | hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI); |
1342 | #endif |
1343 | |
1344 | /* |
1345 | * The reap timers are started later, with a module init call: That part |
1346 | * of the kernel is not yet operational. |
1347 | */ |
1348 | } |
1349 | |
1350 | static int __init cpucache_init(void) |
1351 | { |
1352 | int ret; |
1353 | |
1354 | /* |
1355 | * Register the timers that return unneeded pages to the page allocator |
1356 | */ |
1357 | ret = cpuhp_setup_state(CPUHP_AP_ONLINE_DYN, "SLAB online", |
1358 | slab_online_cpu, slab_offline_cpu); |
1359 | WARN_ON(ret < 0); |
1360 | |
1361 | /* Done! */ |
1362 | slab_state = FULL; |
1363 | return 0; |
1364 | } |
1365 | __initcall(cpucache_init); |
1366 | |
1367 | static noinline void |
1368 | slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid) |
1369 | { |
1370 | #if DEBUG |
1371 | struct kmem_cache_node *n; |
1372 | struct page *page; |
1373 | unsigned long flags; |
1374 | int node; |
1375 | static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL, |
1376 | DEFAULT_RATELIMIT_BURST); |
1377 | |
1378 | if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs)) |
1379 | return; |
1380 | |
1381 | pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n", |
1382 | nodeid, gfpflags, &gfpflags); |
1383 | pr_warn(" cache: %s, object size: %d, order: %d\n", |
1384 | cachep->name, cachep->size, cachep->gfporder); |
1385 | |
1386 | for_each_kmem_cache_node(cachep, node, n) { |
1387 | unsigned long active_objs = 0, num_objs = 0, free_objects = 0; |
1388 | unsigned long active_slabs = 0, num_slabs = 0; |
1389 | unsigned long num_slabs_partial = 0, num_slabs_free = 0; |
1390 | unsigned long num_slabs_full; |
1391 | |
1392 | spin_lock_irqsave(&n->list_lock, flags); |
1393 | num_slabs = n->num_slabs; |
1394 | list_for_each_entry(page, &n->slabs_partial, lru) { |
1395 | active_objs += page->active; |
1396 | num_slabs_partial++; |
1397 | } |
1398 | list_for_each_entry(page, &n->slabs_free, lru) |
1399 | num_slabs_free++; |
1400 | |
1401 | free_objects += n->free_objects; |
1402 | spin_unlock_irqrestore(&n->list_lock, flags); |
1403 | |
1404 | num_objs = num_slabs * cachep->num; |
1405 | active_slabs = num_slabs - num_slabs_free; |
1406 | num_slabs_full = num_slabs - |
1407 | (num_slabs_partial + num_slabs_free); |
1408 | active_objs += (num_slabs_full * cachep->num); |
1409 | |
1410 | pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n", |
1411 | node, active_slabs, num_slabs, active_objs, num_objs, |
1412 | free_objects); |
1413 | } |
1414 | #endif |
1415 | } |
1416 | |
1417 | /* |
1418 | * Interface to system's page allocator. No need to hold the |
1419 | * kmem_cache_node ->list_lock. |
1420 | * |
1421 | * If we requested dmaable memory, we will get it. Even if we |
1422 | * did not request dmaable memory, we might get it, but that |
1423 | * would be relatively rare and ignorable. |
1424 | */ |
1425 | static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags, |
1426 | int nodeid) |
1427 | { |
1428 | struct page *page; |
1429 | int nr_pages; |
1430 | |
1431 | flags |= cachep->allocflags; |
1432 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1433 | flags |= __GFP_RECLAIMABLE; |
1434 | |
1435 | page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder); |
1436 | if (!page) { |
1437 | slab_out_of_memory(cachep, flags, nodeid); |
1438 | return NULL; |
1439 | } |
1440 | |
1441 | if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) { |
1442 | __free_pages(page, cachep->gfporder); |
1443 | return NULL; |
1444 | } |
1445 | |
1446 | nr_pages = (1 << cachep->gfporder); |
1447 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1448 | add_zone_page_state(page_zone(page), |
1449 | NR_SLAB_RECLAIMABLE, nr_pages); |
1450 | else |
1451 | add_zone_page_state(page_zone(page), |
1452 | NR_SLAB_UNRECLAIMABLE, nr_pages); |
1453 | |
1454 | __SetPageSlab(page); |
1455 | /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */ |
1456 | if (sk_memalloc_socks() && page_is_pfmemalloc(page)) |
1457 | SetPageSlabPfmemalloc(page); |
1458 | |
1459 | if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) { |
1460 | kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid); |
1461 | |
1462 | if (cachep->ctor) |
1463 | kmemcheck_mark_uninitialized_pages(page, nr_pages); |
1464 | else |
1465 | kmemcheck_mark_unallocated_pages(page, nr_pages); |
1466 | } |
1467 | |
1468 | return page; |
1469 | } |
1470 | |
1471 | /* |
1472 | * Interface to system's page release. |
1473 | */ |
1474 | static void kmem_freepages(struct kmem_cache *cachep, struct page *page) |
1475 | { |
1476 | int order = cachep->gfporder; |
1477 | unsigned long nr_freed = (1 << order); |
1478 | |
1479 | kmemcheck_free_shadow(page, order); |
1480 | |
1481 | if (cachep->flags & SLAB_RECLAIM_ACCOUNT) |
1482 | sub_zone_page_state(page_zone(page), |
1483 | NR_SLAB_RECLAIMABLE, nr_freed); |
1484 | else |
1485 | sub_zone_page_state(page_zone(page), |
1486 | NR_SLAB_UNRECLAIMABLE, nr_freed); |
1487 | |
1488 | BUG_ON(!PageSlab(page)); |
1489 | __ClearPageSlabPfmemalloc(page); |
1490 | __ClearPageSlab(page); |
1491 | page_mapcount_reset(page); |
1492 | page->mapping = NULL; |
1493 | |
1494 | if (current->reclaim_state) |
1495 | current->reclaim_state->reclaimed_slab += nr_freed; |
1496 | memcg_uncharge_slab(page, order, cachep); |
1497 | __free_pages(page, order); |
1498 | } |
1499 | |
1500 | static void kmem_rcu_free(struct rcu_head *head) |
1501 | { |
1502 | struct kmem_cache *cachep; |
1503 | struct page *page; |
1504 | |
1505 | page = container_of(head, struct page, rcu_head); |
1506 | cachep = page->slab_cache; |
1507 | |
1508 | kmem_freepages(cachep, page); |
1509 | } |
1510 | |
1511 | #if DEBUG |
1512 | static bool is_debug_pagealloc_cache(struct kmem_cache *cachep) |
1513 | { |
1514 | if (debug_pagealloc_enabled() && OFF_SLAB(cachep) && |
1515 | (cachep->size % PAGE_SIZE) == 0) |
1516 | return true; |
1517 | |
1518 | return false; |
1519 | } |
1520 | |
1521 | #ifdef CONFIG_DEBUG_PAGEALLOC |
1522 | static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr, |
1523 | unsigned long caller) |
1524 | { |
1525 | int size = cachep->object_size; |
1526 | |
1527 | addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)]; |
1528 | |
1529 | if (size < 5 * sizeof(unsigned long)) |
1530 | return; |
1531 | |
1532 | *addr++ = 0x12345678; |
1533 | *addr++ = caller; |
1534 | *addr++ = smp_processor_id(); |
1535 | size -= 3 * sizeof(unsigned long); |
1536 | { |
1537 | unsigned long *sptr = &caller; |
1538 | unsigned long svalue; |
1539 | |
1540 | while (!kstack_end(sptr)) { |
1541 | svalue = *sptr++; |
1542 | if (kernel_text_address(svalue)) { |
1543 | *addr++ = svalue; |
1544 | size -= sizeof(unsigned long); |
1545 | if (size <= sizeof(unsigned long)) |
1546 | break; |
1547 | } |
1548 | } |
1549 | |
1550 | } |
1551 | *addr++ = 0x87654321; |
1552 | } |
1553 | |
1554 | static void slab_kernel_map(struct kmem_cache *cachep, void *objp, |
1555 | int map, unsigned long caller) |
1556 | { |
1557 | if (!is_debug_pagealloc_cache(cachep)) |
1558 | return; |
1559 | |
1560 | if (caller) |
1561 | store_stackinfo(cachep, objp, caller); |
1562 | |
1563 | kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map); |
1564 | } |
1565 | |
1566 | #else |
1567 | static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp, |
1568 | int map, unsigned long caller) {} |
1569 | |
1570 | #endif |
1571 | |
1572 | static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val) |
1573 | { |
1574 | int size = cachep->object_size; |
1575 | addr = &((char *)addr)[obj_offset(cachep)]; |
1576 | |
1577 | memset(addr, val, size); |
1578 | *(unsigned char *)(addr + size - 1) = POISON_END; |
1579 | } |
1580 | |
1581 | static void dump_line(char *data, int offset, int limit) |
1582 | { |
1583 | int i; |
1584 | unsigned char error = 0; |
1585 | int bad_count = 0; |
1586 | |
1587 | pr_err("%03x: ", offset); |
1588 | for (i = 0; i < limit; i++) { |
1589 | if (data[offset + i] != POISON_FREE) { |
1590 | error = data[offset + i]; |
1591 | bad_count++; |
1592 | } |
1593 | } |
1594 | print_hex_dump(KERN_CONT, "", 0, 16, 1, |
1595 | &data[offset], limit, 1); |
1596 | |
1597 | if (bad_count == 1) { |
1598 | error ^= POISON_FREE; |
1599 | if (!(error & (error - 1))) { |
1600 | pr_err("Single bit error detected. Probably bad RAM.\n"); |
1601 | #ifdef CONFIG_X86 |
1602 | pr_err("Run memtest86+ or a similar memory test tool.\n"); |
1603 | #else |
1604 | pr_err("Run a memory test tool.\n"); |
1605 | #endif |
1606 | } |
1607 | } |
1608 | } |
1609 | #endif |
1610 | |
1611 | #if DEBUG |
1612 | |
1613 | static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines) |
1614 | { |
1615 | int i, size; |
1616 | char *realobj; |
1617 | |
1618 | if (cachep->flags & SLAB_RED_ZONE) { |
1619 | pr_err("Redzone: 0x%llx/0x%llx\n", |
1620 | *dbg_redzone1(cachep, objp), |
1621 | *dbg_redzone2(cachep, objp)); |
1622 | } |
1623 | |
1624 | if (cachep->flags & SLAB_STORE_USER) |
1625 | pr_err("Last user: (%pSR)\n", *dbg_userword(cachep, objp)); |
1626 | realobj = (char *)objp + obj_offset(cachep); |
1627 | size = cachep->object_size; |
1628 | for (i = 0; i < size && lines; i += 16, lines--) { |
1629 | int limit; |
1630 | limit = 16; |
1631 | if (i + limit > size) |
1632 | limit = size - i; |
1633 | dump_line(realobj, i, limit); |
1634 | } |
1635 | } |
1636 | |
1637 | static void check_poison_obj(struct kmem_cache *cachep, void *objp) |
1638 | { |
1639 | char *realobj; |
1640 | int size, i; |
1641 | int lines = 0; |
1642 | |
1643 | if (is_debug_pagealloc_cache(cachep)) |
1644 | return; |
1645 | |
1646 | realobj = (char *)objp + obj_offset(cachep); |
1647 | size = cachep->object_size; |
1648 | |
1649 | for (i = 0; i < size; i++) { |
1650 | char exp = POISON_FREE; |
1651 | if (i == size - 1) |
1652 | exp = POISON_END; |
1653 | if (realobj[i] != exp) { |
1654 | int limit; |
1655 | /* Mismatch ! */ |
1656 | /* Print header */ |
1657 | if (lines == 0) { |
1658 | pr_err("Slab corruption (%s): %s start=%px, len=%d\n", |
1659 | print_tainted(), cachep->name, |
1660 | realobj, size); |
1661 | print_objinfo(cachep, objp, 0); |
1662 | } |
1663 | /* Hexdump the affected line */ |
1664 | i = (i / 16) * 16; |
1665 | limit = 16; |
1666 | if (i + limit > size) |
1667 | limit = size - i; |
1668 | dump_line(realobj, i, limit); |
1669 | i += 16; |
1670 | lines++; |
1671 | /* Limit to 5 lines */ |
1672 | if (lines > 5) |
1673 | break; |
1674 | } |
1675 | } |
1676 | if (lines != 0) { |
1677 | /* Print some data about the neighboring objects, if they |
1678 | * exist: |
1679 | */ |
1680 | struct page *page = virt_to_head_page(objp); |
1681 | unsigned int objnr; |
1682 | |
1683 | objnr = obj_to_index(cachep, page, objp); |
1684 | if (objnr) { |
1685 | objp = index_to_obj(cachep, page, objnr - 1); |
1686 | realobj = (char *)objp + obj_offset(cachep); |
1687 | pr_err("Prev obj: start=%px, len=%d\n", realobj, size); |
1688 | print_objinfo(cachep, objp, 2); |
1689 | } |
1690 | if (objnr + 1 < cachep->num) { |
1691 | objp = index_to_obj(cachep, page, objnr + 1); |
1692 | realobj = (char *)objp + obj_offset(cachep); |
1693 | pr_err("Next obj: start=%px, len=%d\n", realobj, size); |
1694 | print_objinfo(cachep, objp, 2); |
1695 | } |
1696 | } |
1697 | } |
1698 | #endif |
1699 | |
1700 | #if DEBUG |
1701 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, |
1702 | struct page *page) |
1703 | { |
1704 | int i; |
1705 | |
1706 | if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) { |
1707 | poison_obj(cachep, page->freelist - obj_offset(cachep), |
1708 | POISON_FREE); |
1709 | } |
1710 | |
1711 | for (i = 0; i < cachep->num; i++) { |
1712 | void *objp = index_to_obj(cachep, page, i); |
1713 | |
1714 | if (cachep->flags & SLAB_POISON) { |
1715 | check_poison_obj(cachep, objp); |
1716 | slab_kernel_map(cachep, objp, 1, 0); |
1717 | } |
1718 | if (cachep->flags & SLAB_RED_ZONE) { |
1719 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
1720 | slab_error(cachep, "start of a freed object was overwritten"); |
1721 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
1722 | slab_error(cachep, "end of a freed object was overwritten"); |
1723 | } |
1724 | } |
1725 | } |
1726 | #else |
1727 | static void slab_destroy_debugcheck(struct kmem_cache *cachep, |
1728 | struct page *page) |
1729 | { |
1730 | } |
1731 | #endif |
1732 | |
1733 | /** |
1734 | * slab_destroy - destroy and release all objects in a slab |
1735 | * @cachep: cache pointer being destroyed |
1736 | * @page: page pointer being destroyed |
1737 | * |
1738 | * Destroy all the objs in a slab page, and release the mem back to the system. |
1739 | * Before calling the slab page must have been unlinked from the cache. The |
1740 | * kmem_cache_node ->list_lock is not held/needed. |
1741 | */ |
1742 | static void slab_destroy(struct kmem_cache *cachep, struct page *page) |
1743 | { |
1744 | void *freelist; |
1745 | |
1746 | freelist = page->freelist; |
1747 | slab_destroy_debugcheck(cachep, page); |
1748 | if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) |
1749 | call_rcu(&page->rcu_head, kmem_rcu_free); |
1750 | else |
1751 | kmem_freepages(cachep, page); |
1752 | |
1753 | /* |
1754 | * From now on, we don't use freelist |
1755 | * although actual page can be freed in rcu context |
1756 | */ |
1757 | if (OFF_SLAB(cachep)) |
1758 | kmem_cache_free(cachep->freelist_cache, freelist); |
1759 | } |
1760 | |
1761 | static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list) |
1762 | { |
1763 | struct page *page, *n; |
1764 | |
1765 | list_for_each_entry_safe(page, n, list, lru) { |
1766 | list_del(&page->lru); |
1767 | slab_destroy(cachep, page); |
1768 | } |
1769 | } |
1770 | |
1771 | /** |
1772 | * calculate_slab_order - calculate size (page order) of slabs |
1773 | * @cachep: pointer to the cache that is being created |
1774 | * @size: size of objects to be created in this cache. |
1775 | * @flags: slab allocation flags |
1776 | * |
1777 | * Also calculates the number of objects per slab. |
1778 | * |
1779 | * This could be made much more intelligent. For now, try to avoid using |
1780 | * high order pages for slabs. When the gfp() functions are more friendly |
1781 | * towards high-order requests, this should be changed. |
1782 | */ |
1783 | static size_t calculate_slab_order(struct kmem_cache *cachep, |
1784 | size_t size, unsigned long flags) |
1785 | { |
1786 | size_t left_over = 0; |
1787 | int gfporder; |
1788 | |
1789 | for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) { |
1790 | unsigned int num; |
1791 | size_t remainder; |
1792 | |
1793 | num = cache_estimate(gfporder, size, flags, &remainder); |
1794 | if (!num) |
1795 | continue; |
1796 | |
1797 | /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */ |
1798 | if (num > SLAB_OBJ_MAX_NUM) |
1799 | break; |
1800 | |
1801 | if (flags & CFLGS_OFF_SLAB) { |
1802 | struct kmem_cache *freelist_cache; |
1803 | size_t freelist_size; |
1804 | |
1805 | freelist_size = num * sizeof(freelist_idx_t); |
1806 | freelist_cache = kmalloc_slab(freelist_size, 0u); |
1807 | if (!freelist_cache) |
1808 | continue; |
1809 | |
1810 | /* |
1811 | * Needed to avoid possible looping condition |
1812 | * in cache_grow_begin() |
1813 | */ |
1814 | if (OFF_SLAB(freelist_cache)) |
1815 | continue; |
1816 | |
1817 | /* check if off slab has enough benefit */ |
1818 | if (freelist_cache->size > cachep->size / 2) |
1819 | continue; |
1820 | } |
1821 | |
1822 | /* Found something acceptable - save it away */ |
1823 | cachep->num = num; |
1824 | cachep->gfporder = gfporder; |
1825 | left_over = remainder; |
1826 | |
1827 | /* |
1828 | * A VFS-reclaimable slab tends to have most allocations |
1829 | * as GFP_NOFS and we really don't want to have to be allocating |
1830 | * higher-order pages when we are unable to shrink dcache. |
1831 | */ |
1832 | if (flags & SLAB_RECLAIM_ACCOUNT) |
1833 | break; |
1834 | |
1835 | /* |
1836 | * Large number of objects is good, but very large slabs are |
1837 | * currently bad for the gfp()s. |
1838 | */ |
1839 | if (gfporder >= slab_max_order) |
1840 | break; |
1841 | |
1842 | /* |
1843 | * Acceptable internal fragmentation? |
1844 | */ |
1845 | if (left_over * 8 <= (PAGE_SIZE << gfporder)) |
1846 | break; |
1847 | } |
1848 | return left_over; |
1849 | } |
1850 | |
1851 | static struct array_cache __percpu *alloc_kmem_cache_cpus( |
1852 | struct kmem_cache *cachep, int entries, int batchcount) |
1853 | { |
1854 | int cpu; |
1855 | size_t size; |
1856 | struct array_cache __percpu *cpu_cache; |
1857 | |
1858 | size = sizeof(void *) * entries + sizeof(struct array_cache); |
1859 | cpu_cache = __alloc_percpu(size, sizeof(void *)); |
1860 | |
1861 | if (!cpu_cache) |
1862 | return NULL; |
1863 | |
1864 | for_each_possible_cpu(cpu) { |
1865 | init_arraycache(per_cpu_ptr(cpu_cache, cpu), |
1866 | entries, batchcount); |
1867 | } |
1868 | |
1869 | return cpu_cache; |
1870 | } |
1871 | |
1872 | static int __ref setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp) |
1873 | { |
1874 | if (slab_state >= FULL) |
1875 | return enable_cpucache(cachep, gfp); |
1876 | |
1877 | cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1); |
1878 | if (!cachep->cpu_cache) |
1879 | return 1; |
1880 | |
1881 | if (slab_state == DOWN) { |
1882 | /* Creation of first cache (kmem_cache). */ |
1883 | set_up_node(kmem_cache, CACHE_CACHE); |
1884 | } else if (slab_state == PARTIAL) { |
1885 | /* For kmem_cache_node */ |
1886 | set_up_node(cachep, SIZE_NODE); |
1887 | } else { |
1888 | int node; |
1889 | |
1890 | for_each_online_node(node) { |
1891 | cachep->node[node] = kmalloc_node( |
1892 | sizeof(struct kmem_cache_node), gfp, node); |
1893 | BUG_ON(!cachep->node[node]); |
1894 | kmem_cache_node_init(cachep->node[node]); |
1895 | } |
1896 | } |
1897 | |
1898 | cachep->node[numa_mem_id()]->next_reap = |
1899 | jiffies + REAPTIMEOUT_NODE + |
1900 | ((unsigned long)cachep) % REAPTIMEOUT_NODE; |
1901 | |
1902 | cpu_cache_get(cachep)->avail = 0; |
1903 | cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES; |
1904 | cpu_cache_get(cachep)->batchcount = 1; |
1905 | cpu_cache_get(cachep)->touched = 0; |
1906 | cachep->batchcount = 1; |
1907 | cachep->limit = BOOT_CPUCACHE_ENTRIES; |
1908 | return 0; |
1909 | } |
1910 | |
1911 | unsigned long kmem_cache_flags(unsigned long object_size, |
1912 | unsigned long flags, const char *name, |
1913 | void (*ctor)(void *)) |
1914 | { |
1915 | return flags; |
1916 | } |
1917 | |
1918 | struct kmem_cache * |
1919 | __kmem_cache_alias(const char *name, size_t size, size_t align, |
1920 | unsigned long flags, void (*ctor)(void *)) |
1921 | { |
1922 | struct kmem_cache *cachep; |
1923 | |
1924 | cachep = find_mergeable(size, align, flags, name, ctor); |
1925 | if (cachep) { |
1926 | cachep->refcount++; |
1927 | |
1928 | /* |
1929 | * Adjust the object sizes so that we clear |
1930 | * the complete object on kzalloc. |
1931 | */ |
1932 | cachep->object_size = max_t(int, cachep->object_size, size); |
1933 | } |
1934 | return cachep; |
1935 | } |
1936 | |
1937 | static bool set_objfreelist_slab_cache(struct kmem_cache *cachep, |
1938 | size_t size, unsigned long flags) |
1939 | { |
1940 | size_t left; |
1941 | |
1942 | cachep->num = 0; |
1943 | |
1944 | if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU) |
1945 | return false; |
1946 | |
1947 | left = calculate_slab_order(cachep, size, |
1948 | flags | CFLGS_OBJFREELIST_SLAB); |
1949 | if (!cachep->num) |
1950 | return false; |
1951 | |
1952 | if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size) |
1953 | return false; |
1954 | |
1955 | cachep->colour = left / cachep->colour_off; |
1956 | |
1957 | return true; |
1958 | } |
1959 | |
1960 | static bool set_off_slab_cache(struct kmem_cache *cachep, |
1961 | size_t size, unsigned long flags) |
1962 | { |
1963 | size_t left; |
1964 | |
1965 | cachep->num = 0; |
1966 | |
1967 | /* |
1968 | * Always use on-slab management when SLAB_NOLEAKTRACE |
1969 | * to avoid recursive calls into kmemleak. |
1970 | */ |
1971 | if (flags & SLAB_NOLEAKTRACE) |
1972 | return false; |
1973 | |
1974 | /* |
1975 | * Size is large, assume best to place the slab management obj |
1976 | * off-slab (should allow better packing of objs). |
1977 | */ |
1978 | left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB); |
1979 | if (!cachep->num) |
1980 | return false; |
1981 | |
1982 | /* |
1983 | * If the slab has been placed off-slab, and we have enough space then |
1984 | * move it on-slab. This is at the expense of any extra colouring. |
1985 | */ |
1986 | if (left >= cachep->num * sizeof(freelist_idx_t)) |
1987 | return false; |
1988 | |
1989 | cachep->colour = left / cachep->colour_off; |
1990 | |
1991 | return true; |
1992 | } |
1993 | |
1994 | static bool set_on_slab_cache(struct kmem_cache *cachep, |
1995 | size_t size, unsigned long flags) |
1996 | { |
1997 | size_t left; |
1998 | |
1999 | cachep->num = 0; |
2000 | |
2001 | left = calculate_slab_order(cachep, size, flags); |
2002 | if (!cachep->num) |
2003 | return false; |
2004 | |
2005 | cachep->colour = left / cachep->colour_off; |
2006 | |
2007 | return true; |
2008 | } |
2009 | |
2010 | /** |
2011 | * __kmem_cache_create - Create a cache. |
2012 | * @cachep: cache management descriptor |
2013 | * @flags: SLAB flags |
2014 | * |
2015 | * Returns a ptr to the cache on success, NULL on failure. |
2016 | * Cannot be called within a int, but can be interrupted. |
2017 | * The @ctor is run when new pages are allocated by the cache. |
2018 | * |
2019 | * The flags are |
2020 | * |
2021 | * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5) |
2022 | * to catch references to uninitialised memory. |
2023 | * |
2024 | * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check |
2025 | * for buffer overruns. |
2026 | * |
2027 | * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware |
2028 | * cacheline. This can be beneficial if you're counting cycles as closely |
2029 | * as davem. |
2030 | */ |
2031 | int |
2032 | __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags) |
2033 | { |
2034 | size_t ralign = BYTES_PER_WORD; |
2035 | gfp_t gfp; |
2036 | int err; |
2037 | size_t size = cachep->size; |
2038 | |
2039 | #if DEBUG |
2040 | #if FORCED_DEBUG |
2041 | /* |
2042 | * Enable redzoning and last user accounting, except for caches with |
2043 | * large objects, if the increased size would increase the object size |
2044 | * above the next power of two: caches with object sizes just above a |
2045 | * power of two have a significant amount of internal fragmentation. |
2046 | */ |
2047 | if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN + |
2048 | 2 * sizeof(unsigned long long))) |
2049 | flags |= SLAB_RED_ZONE | SLAB_STORE_USER; |
2050 | if (!(flags & SLAB_DESTROY_BY_RCU)) |
2051 | flags |= SLAB_POISON; |
2052 | #endif |
2053 | #endif |
2054 | |
2055 | /* |
2056 | * Check that size is in terms of words. This is needed to avoid |
2057 | * unaligned accesses for some archs when redzoning is used, and makes |
2058 | * sure any on-slab bufctl's are also correctly aligned. |
2059 | */ |
2060 | if (size & (BYTES_PER_WORD - 1)) { |
2061 | size += (BYTES_PER_WORD - 1); |
2062 | size &= ~(BYTES_PER_WORD - 1); |
2063 | } |
2064 | |
2065 | if (flags & SLAB_RED_ZONE) { |
2066 | ralign = REDZONE_ALIGN; |
2067 | /* If redzoning, ensure that the second redzone is suitably |
2068 | * aligned, by adjusting the object size accordingly. */ |
2069 | size += REDZONE_ALIGN - 1; |
2070 | size &= ~(REDZONE_ALIGN - 1); |
2071 | } |
2072 | |
2073 | /* 3) caller mandated alignment */ |
2074 | if (ralign < cachep->align) { |
2075 | ralign = cachep->align; |
2076 | } |
2077 | /* disable debug if necessary */ |
2078 | if (ralign > __alignof__(unsigned long long)) |
2079 | flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
2080 | /* |
2081 | * 4) Store it. |
2082 | */ |
2083 | cachep->align = ralign; |
2084 | cachep->colour_off = cache_line_size(); |
2085 | /* Offset must be a multiple of the alignment. */ |
2086 | if (cachep->colour_off < cachep->align) |
2087 | cachep->colour_off = cachep->align; |
2088 | |
2089 | if (slab_is_available()) |
2090 | gfp = GFP_KERNEL; |
2091 | else |
2092 | gfp = GFP_NOWAIT; |
2093 | |
2094 | #if DEBUG |
2095 | |
2096 | /* |
2097 | * Both debugging options require word-alignment which is calculated |
2098 | * into align above. |
2099 | */ |
2100 | if (flags & SLAB_RED_ZONE) { |
2101 | /* add space for red zone words */ |
2102 | cachep->obj_offset += sizeof(unsigned long long); |
2103 | size += 2 * sizeof(unsigned long long); |
2104 | } |
2105 | if (flags & SLAB_STORE_USER) { |
2106 | /* user store requires one word storage behind the end of |
2107 | * the real object. But if the second red zone needs to be |
2108 | * aligned to 64 bits, we must allow that much space. |
2109 | */ |
2110 | if (flags & SLAB_RED_ZONE) |
2111 | size += REDZONE_ALIGN; |
2112 | else |
2113 | size += BYTES_PER_WORD; |
2114 | } |
2115 | #endif |
2116 | |
2117 | kasan_cache_create(cachep, &size, &flags); |
2118 | |
2119 | size = ALIGN(size, cachep->align); |
2120 | /* |
2121 | * We should restrict the number of objects in a slab to implement |
2122 | * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition. |
2123 | */ |
2124 | if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE) |
2125 | size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align); |
2126 | |
2127 | #if DEBUG |
2128 | /* |
2129 | * To activate debug pagealloc, off-slab management is necessary |
2130 | * requirement. In early phase of initialization, small sized slab |
2131 | * doesn't get initialized so it would not be possible. So, we need |
2132 | * to check size >= 256. It guarantees that all necessary small |
2133 | * sized slab is initialized in current slab initialization sequence. |
2134 | */ |
2135 | if (debug_pagealloc_enabled() && (flags & SLAB_POISON) && |
2136 | size >= 256 && cachep->object_size > cache_line_size()) { |
2137 | if (size < PAGE_SIZE || size % PAGE_SIZE == 0) { |
2138 | size_t tmp_size = ALIGN(size, PAGE_SIZE); |
2139 | |
2140 | if (set_off_slab_cache(cachep, tmp_size, flags)) { |
2141 | flags |= CFLGS_OFF_SLAB; |
2142 | cachep->obj_offset += tmp_size - size; |
2143 | size = tmp_size; |
2144 | goto done; |
2145 | } |
2146 | } |
2147 | } |
2148 | #endif |
2149 | |
2150 | if (set_objfreelist_slab_cache(cachep, size, flags)) { |
2151 | flags |= CFLGS_OBJFREELIST_SLAB; |
2152 | goto done; |
2153 | } |
2154 | |
2155 | if (set_off_slab_cache(cachep, size, flags)) { |
2156 | flags |= CFLGS_OFF_SLAB; |
2157 | goto done; |
2158 | } |
2159 | |
2160 | if (set_on_slab_cache(cachep, size, flags)) |
2161 | goto done; |
2162 | |
2163 | return -E2BIG; |
2164 | |
2165 | done: |
2166 | cachep->freelist_size = cachep->num * sizeof(freelist_idx_t); |
2167 | cachep->flags = flags; |
2168 | cachep->allocflags = __GFP_COMP; |
2169 | if (flags & SLAB_CACHE_DMA) |
2170 | cachep->allocflags |= GFP_DMA; |
2171 | cachep->size = size; |
2172 | cachep->reciprocal_buffer_size = reciprocal_value(size); |
2173 | |
2174 | #if DEBUG |
2175 | /* |
2176 | * If we're going to use the generic kernel_map_pages() |
2177 | * poisoning, then it's going to smash the contents of |
2178 | * the redzone and userword anyhow, so switch them off. |
2179 | */ |
2180 | if (IS_ENABLED(CONFIG_PAGE_POISONING) && |
2181 | (cachep->flags & SLAB_POISON) && |
2182 | is_debug_pagealloc_cache(cachep)) |
2183 | cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER); |
2184 | #endif |
2185 | |
2186 | if (OFF_SLAB(cachep)) { |
2187 | cachep->freelist_cache = |
2188 | kmalloc_slab(cachep->freelist_size, 0u); |
2189 | } |
2190 | |
2191 | err = setup_cpu_cache(cachep, gfp); |
2192 | if (err) { |
2193 | __kmem_cache_release(cachep); |
2194 | return err; |
2195 | } |
2196 | |
2197 | return 0; |
2198 | } |
2199 | |
2200 | #if DEBUG |
2201 | static void check_irq_off(void) |
2202 | { |
2203 | BUG_ON(!irqs_disabled()); |
2204 | } |
2205 | |
2206 | static void check_irq_on(void) |
2207 | { |
2208 | BUG_ON(irqs_disabled()); |
2209 | } |
2210 | |
2211 | static void check_mutex_acquired(void) |
2212 | { |
2213 | BUG_ON(!mutex_is_locked(&slab_mutex)); |
2214 | } |
2215 | |
2216 | static void check_spinlock_acquired(struct kmem_cache *cachep) |
2217 | { |
2218 | #ifdef CONFIG_SMP |
2219 | check_irq_off(); |
2220 | assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock); |
2221 | #endif |
2222 | } |
2223 | |
2224 | static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node) |
2225 | { |
2226 | #ifdef CONFIG_SMP |
2227 | check_irq_off(); |
2228 | assert_spin_locked(&get_node(cachep, node)->list_lock); |
2229 | #endif |
2230 | } |
2231 | |
2232 | #else |
2233 | #define check_irq_off() do { } while(0) |
2234 | #define check_irq_on() do { } while(0) |
2235 | #define check_mutex_acquired() do { } while(0) |
2236 | #define check_spinlock_acquired(x) do { } while(0) |
2237 | #define check_spinlock_acquired_node(x, y) do { } while(0) |
2238 | #endif |
2239 | |
2240 | static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac, |
2241 | int node, bool free_all, struct list_head *list) |
2242 | { |
2243 | int tofree; |
2244 | |
2245 | if (!ac || !ac->avail) |
2246 | return; |
2247 | |
2248 | tofree = free_all ? ac->avail : (ac->limit + 4) / 5; |
2249 | if (tofree > ac->avail) |
2250 | tofree = (ac->avail + 1) / 2; |
2251 | |
2252 | free_block(cachep, ac->entry, tofree, node, list); |
2253 | ac->avail -= tofree; |
2254 | memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail); |
2255 | } |
2256 | |
2257 | static void do_drain(void *arg) |
2258 | { |
2259 | struct kmem_cache *cachep = arg; |
2260 | struct array_cache *ac; |
2261 | int node = numa_mem_id(); |
2262 | struct kmem_cache_node *n; |
2263 | LIST_HEAD(list); |
2264 | |
2265 | check_irq_off(); |
2266 | ac = cpu_cache_get(cachep); |
2267 | n = get_node(cachep, node); |
2268 | spin_lock(&n->list_lock); |
2269 | free_block(cachep, ac->entry, ac->avail, node, &list); |
2270 | spin_unlock(&n->list_lock); |
2271 | slabs_destroy(cachep, &list); |
2272 | ac->avail = 0; |
2273 | } |
2274 | |
2275 | static void drain_cpu_caches(struct kmem_cache *cachep) |
2276 | { |
2277 | struct kmem_cache_node *n; |
2278 | int node; |
2279 | LIST_HEAD(list); |
2280 | |
2281 | on_each_cpu(do_drain, cachep, 1); |
2282 | check_irq_on(); |
2283 | for_each_kmem_cache_node(cachep, node, n) |
2284 | if (n->alien) |
2285 | drain_alien_cache(cachep, n->alien); |
2286 | |
2287 | for_each_kmem_cache_node(cachep, node, n) { |
2288 | spin_lock_irq(&n->list_lock); |
2289 | drain_array_locked(cachep, n->shared, node, true, &list); |
2290 | spin_unlock_irq(&n->list_lock); |
2291 | |
2292 | slabs_destroy(cachep, &list); |
2293 | } |
2294 | } |
2295 | |
2296 | /* |
2297 | * Remove slabs from the list of free slabs. |
2298 | * Specify the number of slabs to drain in tofree. |
2299 | * |
2300 | * Returns the actual number of slabs released. |
2301 | */ |
2302 | static int drain_freelist(struct kmem_cache *cache, |
2303 | struct kmem_cache_node *n, int tofree) |
2304 | { |
2305 | struct list_head *p; |
2306 | int nr_freed; |
2307 | struct page *page; |
2308 | |
2309 | nr_freed = 0; |
2310 | while (nr_freed < tofree && !list_empty(&n->slabs_free)) { |
2311 | |
2312 | spin_lock_irq(&n->list_lock); |
2313 | p = n->slabs_free.prev; |
2314 | if (p == &n->slabs_free) { |
2315 | spin_unlock_irq(&n->list_lock); |
2316 | goto out; |
2317 | } |
2318 | |
2319 | page = list_entry(p, struct page, lru); |
2320 | list_del(&page->lru); |
2321 | n->num_slabs--; |
2322 | /* |
2323 | * Safe to drop the lock. The slab is no longer linked |
2324 | * to the cache. |
2325 | */ |
2326 | n->free_objects -= cache->num; |
2327 | spin_unlock_irq(&n->list_lock); |
2328 | slab_destroy(cache, page); |
2329 | nr_freed++; |
2330 | } |
2331 | out: |
2332 | return nr_freed; |
2333 | } |
2334 | |
2335 | int __kmem_cache_shrink(struct kmem_cache *cachep) |
2336 | { |
2337 | int ret = 0; |
2338 | int node; |
2339 | struct kmem_cache_node *n; |
2340 | |
2341 | drain_cpu_caches(cachep); |
2342 | |
2343 | check_irq_on(); |
2344 | for_each_kmem_cache_node(cachep, node, n) { |
2345 | drain_freelist(cachep, n, INT_MAX); |
2346 | |
2347 | ret += !list_empty(&n->slabs_full) || |
2348 | !list_empty(&n->slabs_partial); |
2349 | } |
2350 | return (ret ? 1 : 0); |
2351 | } |
2352 | |
2353 | int __kmem_cache_shutdown(struct kmem_cache *cachep) |
2354 | { |
2355 | return __kmem_cache_shrink(cachep); |
2356 | } |
2357 | |
2358 | void __kmem_cache_release(struct kmem_cache *cachep) |
2359 | { |
2360 | int i; |
2361 | struct kmem_cache_node *n; |
2362 | |
2363 | cache_random_seq_destroy(cachep); |
2364 | |
2365 | free_percpu(cachep->cpu_cache); |
2366 | |
2367 | /* NUMA: free the node structures */ |
2368 | for_each_kmem_cache_node(cachep, i, n) { |
2369 | kfree(n->shared); |
2370 | free_alien_cache(n->alien); |
2371 | kfree(n); |
2372 | cachep->node[i] = NULL; |
2373 | } |
2374 | } |
2375 | |
2376 | /* |
2377 | * Get the memory for a slab management obj. |
2378 | * |
2379 | * For a slab cache when the slab descriptor is off-slab, the |
2380 | * slab descriptor can't come from the same cache which is being created, |
2381 | * Because if it is the case, that means we defer the creation of |
2382 | * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point. |
2383 | * And we eventually call down to __kmem_cache_create(), which |
2384 | * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one. |
2385 | * This is a "chicken-and-egg" problem. |
2386 | * |
2387 | * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches, |
2388 | * which are all initialized during kmem_cache_init(). |
2389 | */ |
2390 | static void *alloc_slabmgmt(struct kmem_cache *cachep, |
2391 | struct page *page, int colour_off, |
2392 | gfp_t local_flags, int nodeid) |
2393 | { |
2394 | void *freelist; |
2395 | void *addr = page_address(page); |
2396 | |
2397 | page->s_mem = addr + colour_off; |
2398 | page->active = 0; |
2399 | |
2400 | if (OBJFREELIST_SLAB(cachep)) |
2401 | freelist = NULL; |
2402 | else if (OFF_SLAB(cachep)) { |
2403 | /* Slab management obj is off-slab. */ |
2404 | freelist = kmem_cache_alloc_node(cachep->freelist_cache, |
2405 | local_flags, nodeid); |
2406 | if (!freelist) |
2407 | return NULL; |
2408 | } else { |
2409 | /* We will use last bytes at the slab for freelist */ |
2410 | freelist = addr + (PAGE_SIZE << cachep->gfporder) - |
2411 | cachep->freelist_size; |
2412 | } |
2413 | |
2414 | return freelist; |
2415 | } |
2416 | |
2417 | static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx) |
2418 | { |
2419 | return ((freelist_idx_t *)page->freelist)[idx]; |
2420 | } |
2421 | |
2422 | static inline void set_free_obj(struct page *page, |
2423 | unsigned int idx, freelist_idx_t val) |
2424 | { |
2425 | ((freelist_idx_t *)(page->freelist))[idx] = val; |
2426 | } |
2427 | |
2428 | static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page) |
2429 | { |
2430 | #if DEBUG |
2431 | int i; |
2432 | |
2433 | for (i = 0; i < cachep->num; i++) { |
2434 | void *objp = index_to_obj(cachep, page, i); |
2435 | |
2436 | if (cachep->flags & SLAB_STORE_USER) |
2437 | *dbg_userword(cachep, objp) = NULL; |
2438 | |
2439 | if (cachep->flags & SLAB_RED_ZONE) { |
2440 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2441 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; |
2442 | } |
2443 | /* |
2444 | * Constructors are not allowed to allocate memory from the same |
2445 | * cache which they are a constructor for. Otherwise, deadlock. |
2446 | * They must also be threaded. |
2447 | */ |
2448 | if (cachep->ctor && !(cachep->flags & SLAB_POISON)) { |
2449 | kasan_unpoison_object_data(cachep, |
2450 | objp + obj_offset(cachep)); |
2451 | cachep->ctor(objp + obj_offset(cachep)); |
2452 | kasan_poison_object_data( |
2453 | cachep, objp + obj_offset(cachep)); |
2454 | } |
2455 | |
2456 | if (cachep->flags & SLAB_RED_ZONE) { |
2457 | if (*dbg_redzone2(cachep, objp) != RED_INACTIVE) |
2458 | slab_error(cachep, "constructor overwrote the end of an object"); |
2459 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE) |
2460 | slab_error(cachep, "constructor overwrote the start of an object"); |
2461 | } |
2462 | /* need to poison the objs? */ |
2463 | if (cachep->flags & SLAB_POISON) { |
2464 | poison_obj(cachep, objp, POISON_FREE); |
2465 | slab_kernel_map(cachep, objp, 0, 0); |
2466 | } |
2467 | } |
2468 | #endif |
2469 | } |
2470 | |
2471 | #ifdef CONFIG_SLAB_FREELIST_RANDOM |
2472 | /* Hold information during a freelist initialization */ |
2473 | union freelist_init_state { |
2474 | struct { |
2475 | unsigned int pos; |
2476 | unsigned int *list; |
2477 | unsigned int count; |
2478 | }; |
2479 | struct rnd_state rnd_state; |
2480 | }; |
2481 | |
2482 | /* |
2483 | * Initialize the state based on the randomization methode available. |
2484 | * return true if the pre-computed list is available, false otherwize. |
2485 | */ |
2486 | static bool freelist_state_initialize(union freelist_init_state *state, |
2487 | struct kmem_cache *cachep, |
2488 | unsigned int count) |
2489 | { |
2490 | bool ret; |
2491 | unsigned int rand; |
2492 | |
2493 | /* Use best entropy available to define a random shift */ |
2494 | rand = get_random_int(); |
2495 | |
2496 | /* Use a random state if the pre-computed list is not available */ |
2497 | if (!cachep->random_seq) { |
2498 | prandom_seed_state(&state->rnd_state, rand); |
2499 | ret = false; |
2500 | } else { |
2501 | state->list = cachep->random_seq; |
2502 | state->count = count; |
2503 | state->pos = rand % count; |
2504 | ret = true; |
2505 | } |
2506 | return ret; |
2507 | } |
2508 | |
2509 | /* Get the next entry on the list and randomize it using a random shift */ |
2510 | static freelist_idx_t next_random_slot(union freelist_init_state *state) |
2511 | { |
2512 | if (state->pos >= state->count) |
2513 | state->pos = 0; |
2514 | return state->list[state->pos++]; |
2515 | } |
2516 | |
2517 | /* Swap two freelist entries */ |
2518 | static void swap_free_obj(struct page *page, unsigned int a, unsigned int b) |
2519 | { |
2520 | swap(((freelist_idx_t *)page->freelist)[a], |
2521 | ((freelist_idx_t *)page->freelist)[b]); |
2522 | } |
2523 | |
2524 | /* |
2525 | * Shuffle the freelist initialization state based on pre-computed lists. |
2526 | * return true if the list was successfully shuffled, false otherwise. |
2527 | */ |
2528 | static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page) |
2529 | { |
2530 | unsigned int objfreelist = 0, i, rand, count = cachep->num; |
2531 | union freelist_init_state state; |
2532 | bool precomputed; |
2533 | |
2534 | if (count < 2) |
2535 | return false; |
2536 | |
2537 | precomputed = freelist_state_initialize(&state, cachep, count); |
2538 | |
2539 | /* Take a random entry as the objfreelist */ |
2540 | if (OBJFREELIST_SLAB(cachep)) { |
2541 | if (!precomputed) |
2542 | objfreelist = count - 1; |
2543 | else |
2544 | objfreelist = next_random_slot(&state); |
2545 | page->freelist = index_to_obj(cachep, page, objfreelist) + |
2546 | obj_offset(cachep); |
2547 | count--; |
2548 | } |
2549 | |
2550 | /* |
2551 | * On early boot, generate the list dynamically. |
2552 | * Later use a pre-computed list for speed. |
2553 | */ |
2554 | if (!precomputed) { |
2555 | for (i = 0; i < count; i++) |
2556 | set_free_obj(page, i, i); |
2557 | |
2558 | /* Fisher-Yates shuffle */ |
2559 | for (i = count - 1; i > 0; i--) { |
2560 | rand = prandom_u32_state(&state.rnd_state); |
2561 | rand %= (i + 1); |
2562 | swap_free_obj(page, i, rand); |
2563 | } |
2564 | } else { |
2565 | for (i = 0; i < count; i++) |
2566 | set_free_obj(page, i, next_random_slot(&state)); |
2567 | } |
2568 | |
2569 | if (OBJFREELIST_SLAB(cachep)) |
2570 | set_free_obj(page, cachep->num - 1, objfreelist); |
2571 | |
2572 | return true; |
2573 | } |
2574 | #else |
2575 | static inline bool shuffle_freelist(struct kmem_cache *cachep, |
2576 | struct page *page) |
2577 | { |
2578 | return false; |
2579 | } |
2580 | #endif /* CONFIG_SLAB_FREELIST_RANDOM */ |
2581 | |
2582 | static void cache_init_objs(struct kmem_cache *cachep, |
2583 | struct page *page) |
2584 | { |
2585 | int i; |
2586 | void *objp; |
2587 | bool shuffled; |
2588 | |
2589 | cache_init_objs_debug(cachep, page); |
2590 | |
2591 | /* Try to randomize the freelist if enabled */ |
2592 | shuffled = shuffle_freelist(cachep, page); |
2593 | |
2594 | if (!shuffled && OBJFREELIST_SLAB(cachep)) { |
2595 | page->freelist = index_to_obj(cachep, page, cachep->num - 1) + |
2596 | obj_offset(cachep); |
2597 | } |
2598 | |
2599 | for (i = 0; i < cachep->num; i++) { |
2600 | objp = index_to_obj(cachep, page, i); |
2601 | kasan_init_slab_obj(cachep, objp); |
2602 | |
2603 | /* constructor could break poison info */ |
2604 | if (DEBUG == 0 && cachep->ctor) { |
2605 | kasan_unpoison_object_data(cachep, objp); |
2606 | cachep->ctor(objp); |
2607 | kasan_poison_object_data(cachep, objp); |
2608 | } |
2609 | |
2610 | if (!shuffled) |
2611 | set_free_obj(page, i, i); |
2612 | } |
2613 | } |
2614 | |
2615 | static void *slab_get_obj(struct kmem_cache *cachep, struct page *page) |
2616 | { |
2617 | void *objp; |
2618 | |
2619 | objp = index_to_obj(cachep, page, get_free_obj(page, page->active)); |
2620 | page->active++; |
2621 | |
2622 | #if DEBUG |
2623 | if (cachep->flags & SLAB_STORE_USER) |
2624 | set_store_user_dirty(cachep); |
2625 | #endif |
2626 | |
2627 | return objp; |
2628 | } |
2629 | |
2630 | static void slab_put_obj(struct kmem_cache *cachep, |
2631 | struct page *page, void *objp) |
2632 | { |
2633 | unsigned int objnr = obj_to_index(cachep, page, objp); |
2634 | #if DEBUG |
2635 | unsigned int i; |
2636 | |
2637 | /* Verify double free bug */ |
2638 | for (i = page->active; i < cachep->num; i++) { |
2639 | if (get_free_obj(page, i) == objnr) { |
2640 | pr_err("slab: double free detected in cache '%s', objp %px\n", |
2641 | cachep->name, objp); |
2642 | BUG(); |
2643 | } |
2644 | } |
2645 | #endif |
2646 | page->active--; |
2647 | if (!page->freelist) |
2648 | page->freelist = objp + obj_offset(cachep); |
2649 | |
2650 | set_free_obj(page, page->active, objnr); |
2651 | } |
2652 | |
2653 | /* |
2654 | * Map pages beginning at addr to the given cache and slab. This is required |
2655 | * for the slab allocator to be able to lookup the cache and slab of a |
2656 | * virtual address for kfree, ksize, and slab debugging. |
2657 | */ |
2658 | static void slab_map_pages(struct kmem_cache *cache, struct page *page, |
2659 | void *freelist) |
2660 | { |
2661 | page->slab_cache = cache; |
2662 | page->freelist = freelist; |
2663 | } |
2664 | |
2665 | /* |
2666 | * Grow (by 1) the number of slabs within a cache. This is called by |
2667 | * kmem_cache_alloc() when there are no active objs left in a cache. |
2668 | */ |
2669 | static struct page *cache_grow_begin(struct kmem_cache *cachep, |
2670 | gfp_t flags, int nodeid) |
2671 | { |
2672 | void *freelist; |
2673 | size_t offset; |
2674 | gfp_t local_flags; |
2675 | int page_node; |
2676 | struct kmem_cache_node *n; |
2677 | struct page *page; |
2678 | |
2679 | /* |
2680 | * Be lazy and only check for valid flags here, keeping it out of the |
2681 | * critical path in kmem_cache_alloc(). |
2682 | */ |
2683 | if (unlikely(flags & GFP_SLAB_BUG_MASK)) { |
2684 | gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK; |
2685 | flags &= ~GFP_SLAB_BUG_MASK; |
2686 | pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n", |
2687 | invalid_mask, &invalid_mask, flags, &flags); |
2688 | dump_stack(); |
2689 | } |
2690 | local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK); |
2691 | |
2692 | check_irq_off(); |
2693 | if (gfpflags_allow_blocking(local_flags)) |
2694 | local_irq_enable(); |
2695 | |
2696 | /* |
2697 | * Get mem for the objs. Attempt to allocate a physical page from |
2698 | * 'nodeid'. |
2699 | */ |
2700 | page = kmem_getpages(cachep, local_flags, nodeid); |
2701 | if (!page) |
2702 | goto failed; |
2703 | |
2704 | page_node = page_to_nid(page); |
2705 | n = get_node(cachep, page_node); |
2706 | |
2707 | /* Get colour for the slab, and cal the next value. */ |
2708 | n->colour_next++; |
2709 | if (n->colour_next >= cachep->colour) |
2710 | n->colour_next = 0; |
2711 | |
2712 | offset = n->colour_next; |
2713 | if (offset >= cachep->colour) |
2714 | offset = 0; |
2715 | |
2716 | offset *= cachep->colour_off; |
2717 | |
2718 | /* Get slab management. */ |
2719 | freelist = alloc_slabmgmt(cachep, page, offset, |
2720 | local_flags & ~GFP_CONSTRAINT_MASK, page_node); |
2721 | if (OFF_SLAB(cachep) && !freelist) |
2722 | goto opps1; |
2723 | |
2724 | slab_map_pages(cachep, page, freelist); |
2725 | |
2726 | kasan_poison_slab(page); |
2727 | cache_init_objs(cachep, page); |
2728 | |
2729 | if (gfpflags_allow_blocking(local_flags)) |
2730 | local_irq_disable(); |
2731 | |
2732 | return page; |
2733 | |
2734 | opps1: |
2735 | kmem_freepages(cachep, page); |
2736 | failed: |
2737 | if (gfpflags_allow_blocking(local_flags)) |
2738 | local_irq_disable(); |
2739 | return NULL; |
2740 | } |
2741 | |
2742 | static void cache_grow_end(struct kmem_cache *cachep, struct page *page) |
2743 | { |
2744 | struct kmem_cache_node *n; |
2745 | void *list = NULL; |
2746 | |
2747 | check_irq_off(); |
2748 | |
2749 | if (!page) |
2750 | return; |
2751 | |
2752 | INIT_LIST_HEAD(&page->lru); |
2753 | n = get_node(cachep, page_to_nid(page)); |
2754 | |
2755 | spin_lock(&n->list_lock); |
2756 | if (!page->active) |
2757 | list_add_tail(&page->lru, &(n->slabs_free)); |
2758 | else |
2759 | fixup_slab_list(cachep, n, page, &list); |
2760 | |
2761 | n->num_slabs++; |
2762 | STATS_INC_GROWN(cachep); |
2763 | n->free_objects += cachep->num - page->active; |
2764 | spin_unlock(&n->list_lock); |
2765 | |
2766 | fixup_objfreelist_debug(cachep, &list); |
2767 | } |
2768 | |
2769 | #if DEBUG |
2770 | |
2771 | /* |
2772 | * Perform extra freeing checks: |
2773 | * - detect bad pointers. |
2774 | * - POISON/RED_ZONE checking |
2775 | */ |
2776 | static void kfree_debugcheck(const void *objp) |
2777 | { |
2778 | if (!virt_addr_valid(objp)) { |
2779 | pr_err("kfree_debugcheck: out of range ptr %lxh\n", |
2780 | (unsigned long)objp); |
2781 | BUG(); |
2782 | } |
2783 | } |
2784 | |
2785 | static inline void verify_redzone_free(struct kmem_cache *cache, void *obj) |
2786 | { |
2787 | unsigned long long redzone1, redzone2; |
2788 | |
2789 | redzone1 = *dbg_redzone1(cache, obj); |
2790 | redzone2 = *dbg_redzone2(cache, obj); |
2791 | |
2792 | /* |
2793 | * Redzone is ok. |
2794 | */ |
2795 | if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE) |
2796 | return; |
2797 | |
2798 | if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE) |
2799 | slab_error(cache, "double free detected"); |
2800 | else |
2801 | slab_error(cache, "memory outside object was overwritten"); |
2802 | |
2803 | pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n", |
2804 | obj, redzone1, redzone2); |
2805 | } |
2806 | |
2807 | static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp, |
2808 | unsigned long caller) |
2809 | { |
2810 | unsigned int objnr; |
2811 | struct page *page; |
2812 | |
2813 | BUG_ON(virt_to_cache(objp) != cachep); |
2814 | |
2815 | objp -= obj_offset(cachep); |
2816 | kfree_debugcheck(objp); |
2817 | page = virt_to_head_page(objp); |
2818 | |
2819 | if (cachep->flags & SLAB_RED_ZONE) { |
2820 | verify_redzone_free(cachep, objp); |
2821 | *dbg_redzone1(cachep, objp) = RED_INACTIVE; |
2822 | *dbg_redzone2(cachep, objp) = RED_INACTIVE; |
2823 | } |
2824 | if (cachep->flags & SLAB_STORE_USER) { |
2825 | set_store_user_dirty(cachep); |
2826 | *dbg_userword(cachep, objp) = (void *)caller; |
2827 | } |
2828 | |
2829 | objnr = obj_to_index(cachep, page, objp); |
2830 | |
2831 | BUG_ON(objnr >= cachep->num); |
2832 | BUG_ON(objp != index_to_obj(cachep, page, objnr)); |
2833 | |
2834 | if (cachep->flags & SLAB_POISON) { |
2835 | poison_obj(cachep, objp, POISON_FREE); |
2836 | slab_kernel_map(cachep, objp, 0, caller); |
2837 | } |
2838 | return objp; |
2839 | } |
2840 | |
2841 | #else |
2842 | #define kfree_debugcheck(x) do { } while(0) |
2843 | #define cache_free_debugcheck(x,objp,z) (objp) |
2844 | #endif |
2845 | |
2846 | static inline void fixup_objfreelist_debug(struct kmem_cache *cachep, |
2847 | void **list) |
2848 | { |
2849 | #if DEBUG |
2850 | void *next = *list; |
2851 | void *objp; |
2852 | |
2853 | while (next) { |
2854 | objp = next - obj_offset(cachep); |
2855 | next = *(void **)next; |
2856 | poison_obj(cachep, objp, POISON_FREE); |
2857 | } |
2858 | #endif |
2859 | } |
2860 | |
2861 | static inline void fixup_slab_list(struct kmem_cache *cachep, |
2862 | struct kmem_cache_node *n, struct page *page, |
2863 | void **list) |
2864 | { |
2865 | /* move slabp to correct slabp list: */ |
2866 | list_del(&page->lru); |
2867 | if (page->active == cachep->num) { |
2868 | list_add(&page->lru, &n->slabs_full); |
2869 | if (OBJFREELIST_SLAB(cachep)) { |
2870 | #if DEBUG |
2871 | /* Poisoning will be done without holding the lock */ |
2872 | if (cachep->flags & SLAB_POISON) { |
2873 | void **objp = page->freelist; |
2874 | |
2875 | *objp = *list; |
2876 | *list = objp; |
2877 | } |
2878 | #endif |
2879 | page->freelist = NULL; |
2880 | } |
2881 | } else |
2882 | list_add(&page->lru, &n->slabs_partial); |
2883 | } |
2884 | |
2885 | /* Try to find non-pfmemalloc slab if needed */ |
2886 | static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n, |
2887 | struct page *page, bool pfmemalloc) |
2888 | { |
2889 | if (!page) |
2890 | return NULL; |
2891 | |
2892 | if (pfmemalloc) |
2893 | return page; |
2894 | |
2895 | if (!PageSlabPfmemalloc(page)) |
2896 | return page; |
2897 | |
2898 | /* No need to keep pfmemalloc slab if we have enough free objects */ |
2899 | if (n->free_objects > n->free_limit) { |
2900 | ClearPageSlabPfmemalloc(page); |
2901 | return page; |
2902 | } |
2903 | |
2904 | /* Move pfmemalloc slab to the end of list to speed up next search */ |
2905 | list_del(&page->lru); |
2906 | if (!page->active) |
2907 | list_add_tail(&page->lru, &n->slabs_free); |
2908 | else |
2909 | list_add_tail(&page->lru, &n->slabs_partial); |
2910 | |
2911 | list_for_each_entry(page, &n->slabs_partial, lru) { |
2912 | if (!PageSlabPfmemalloc(page)) |
2913 | return page; |
2914 | } |
2915 | |
2916 | list_for_each_entry(page, &n->slabs_free, lru) { |
2917 | if (!PageSlabPfmemalloc(page)) |
2918 | return page; |
2919 | } |
2920 | |
2921 | return NULL; |
2922 | } |
2923 | |
2924 | static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc) |
2925 | { |
2926 | struct page *page; |
2927 | |
2928 | page = list_first_entry_or_null(&n->slabs_partial, |
2929 | struct page, lru); |
2930 | if (!page) { |
2931 | n->free_touched = 1; |
2932 | page = list_first_entry_or_null(&n->slabs_free, |
2933 | struct page, lru); |
2934 | } |
2935 | |
2936 | if (sk_memalloc_socks()) |
2937 | return get_valid_first_slab(n, page, pfmemalloc); |
2938 | |
2939 | return page; |
2940 | } |
2941 | |
2942 | static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep, |
2943 | struct kmem_cache_node *n, gfp_t flags) |
2944 | { |
2945 | struct page *page; |
2946 | void *obj; |
2947 | void *list = NULL; |
2948 | |
2949 | if (!gfp_pfmemalloc_allowed(flags)) |
2950 | return NULL; |
2951 | |
2952 | spin_lock(&n->list_lock); |
2953 | page = get_first_slab(n, true); |
2954 | if (!page) { |
2955 | spin_unlock(&n->list_lock); |
2956 | return NULL; |
2957 | } |
2958 | |
2959 | obj = slab_get_obj(cachep, page); |
2960 | n->free_objects--; |
2961 | |
2962 | fixup_slab_list(cachep, n, page, &list); |
2963 | |
2964 | spin_unlock(&n->list_lock); |
2965 | fixup_objfreelist_debug(cachep, &list); |
2966 | |
2967 | return obj; |
2968 | } |
2969 | |
2970 | /* |
2971 | * Slab list should be fixed up by fixup_slab_list() for existing slab |
2972 | * or cache_grow_end() for new slab |
2973 | */ |
2974 | static __always_inline int alloc_block(struct kmem_cache *cachep, |
2975 | struct array_cache *ac, struct page *page, int batchcount) |
2976 | { |
2977 | /* |
2978 | * There must be at least one object available for |
2979 | * allocation. |
2980 | */ |
2981 | BUG_ON(page->active >= cachep->num); |
2982 | |
2983 | while (page->active < cachep->num && batchcount--) { |
2984 | STATS_INC_ALLOCED(cachep); |
2985 | STATS_INC_ACTIVE(cachep); |
2986 | STATS_SET_HIGH(cachep); |
2987 | |
2988 | ac->entry[ac->avail++] = slab_get_obj(cachep, page); |
2989 | } |
2990 | |
2991 | return batchcount; |
2992 | } |
2993 | |
2994 | static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags) |
2995 | { |
2996 | int batchcount; |
2997 | struct kmem_cache_node *n; |
2998 | struct array_cache *ac, *shared; |
2999 | int node; |
3000 | void *list = NULL; |
3001 | struct page *page; |
3002 | |
3003 | check_irq_off(); |
3004 | node = numa_mem_id(); |
3005 | |
3006 | ac = cpu_cache_get(cachep); |
3007 | batchcount = ac->batchcount; |
3008 | if (!ac->touched && batchcount > BATCHREFILL_LIMIT) { |
3009 | /* |
3010 | * If there was little recent activity on this cache, then |
3011 | * perform only a partial refill. Otherwise we could generate |
3012 | * refill bouncing. |
3013 | */ |
3014 | batchcount = BATCHREFILL_LIMIT; |
3015 | } |
3016 | n = get_node(cachep, node); |
3017 | |
3018 | BUG_ON(ac->avail > 0 || !n); |
3019 | shared = READ_ONCE(n->shared); |
3020 | if (!n->free_objects && (!shared || !shared->avail)) |
3021 | goto direct_grow; |
3022 | |
3023 | spin_lock(&n->list_lock); |
3024 | shared = READ_ONCE(n->shared); |
3025 | |
3026 | /* See if we can refill from the shared array */ |
3027 | if (shared && transfer_objects(ac, shared, batchcount)) { |
3028 | shared->touched = 1; |
3029 | goto alloc_done; |
3030 | } |
3031 | |
3032 | while (batchcount > 0) { |
3033 | /* Get slab alloc is to come from. */ |
3034 | page = get_first_slab(n, false); |
3035 | if (!page) |
3036 | goto must_grow; |
3037 | |
3038 | check_spinlock_acquired(cachep); |
3039 | |
3040 | batchcount = alloc_block(cachep, ac, page, batchcount); |
3041 | fixup_slab_list(cachep, n, page, &list); |
3042 | } |
3043 | |
3044 | must_grow: |
3045 | n->free_objects -= ac->avail; |
3046 | alloc_done: |
3047 | spin_unlock(&n->list_lock); |
3048 | fixup_objfreelist_debug(cachep, &list); |
3049 | |
3050 | direct_grow: |
3051 | if (unlikely(!ac->avail)) { |
3052 | /* Check if we can use obj in pfmemalloc slab */ |
3053 | if (sk_memalloc_socks()) { |
3054 | void *obj = cache_alloc_pfmemalloc(cachep, n, flags); |
3055 | |
3056 | if (obj) |
3057 | return obj; |
3058 | } |
3059 | |
3060 | page = cache_grow_begin(cachep, gfp_exact_node(flags), node); |
3061 | |
3062 | /* |
3063 | * cache_grow_begin() can reenable interrupts, |
3064 | * then ac could change. |
3065 | */ |
3066 | ac = cpu_cache_get(cachep); |
3067 | if (!ac->avail && page) |
3068 | alloc_block(cachep, ac, page, batchcount); |
3069 | cache_grow_end(cachep, page); |
3070 | |
3071 | if (!ac->avail) |
3072 | return NULL; |
3073 | } |
3074 | ac->touched = 1; |
3075 | |
3076 | return ac->entry[--ac->avail]; |
3077 | } |
3078 | |
3079 | static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep, |
3080 | gfp_t flags) |
3081 | { |
3082 | might_sleep_if(gfpflags_allow_blocking(flags)); |
3083 | } |
3084 | |
3085 | #if DEBUG |
3086 | static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep, |
3087 | gfp_t flags, void *objp, unsigned long caller) |
3088 | { |
3089 | if (!objp) |
3090 | return objp; |
3091 | if (cachep->flags & SLAB_POISON) { |
3092 | check_poison_obj(cachep, objp); |
3093 | slab_kernel_map(cachep, objp, 1, 0); |
3094 | poison_obj(cachep, objp, POISON_INUSE); |
3095 | } |
3096 | if (cachep->flags & SLAB_STORE_USER) |
3097 | *dbg_userword(cachep, objp) = (void *)caller; |
3098 | |
3099 | if (cachep->flags & SLAB_RED_ZONE) { |
3100 | if (*dbg_redzone1(cachep, objp) != RED_INACTIVE || |
3101 | *dbg_redzone2(cachep, objp) != RED_INACTIVE) { |
3102 | slab_error(cachep, "double free, or memory outside object was overwritten"); |
3103 | pr_err("%px: redzone 1:0x%llx, redzone 2:0x%llx\n", |
3104 | objp, *dbg_redzone1(cachep, objp), |
3105 | *dbg_redzone2(cachep, objp)); |
3106 | } |
3107 | *dbg_redzone1(cachep, objp) = RED_ACTIVE; |
3108 | *dbg_redzone2(cachep, objp) = RED_ACTIVE; |
3109 | } |
3110 | |
3111 | objp += obj_offset(cachep); |
3112 | if (cachep->ctor && cachep->flags & SLAB_POISON) |
3113 | cachep->ctor(objp); |
3114 | if (ARCH_SLAB_MINALIGN && |
3115 | ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) { |
3116 | pr_err("0x%px: not aligned to ARCH_SLAB_MINALIGN=%d\n", |
3117 | objp, (int)ARCH_SLAB_MINALIGN); |
3118 | } |
3119 | return objp; |
3120 | } |
3121 | #else |
3122 | #define cache_alloc_debugcheck_after(a,b,objp,d) (objp) |
3123 | #endif |
3124 | |
3125 | static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
3126 | { |
3127 | void *objp; |
3128 | struct array_cache *ac; |
3129 | |
3130 | check_irq_off(); |
3131 | |
3132 | ac = cpu_cache_get(cachep); |
3133 | if (likely(ac->avail)) { |
3134 | ac->touched = 1; |
3135 | objp = ac->entry[--ac->avail]; |
3136 | |
3137 | STATS_INC_ALLOCHIT(cachep); |
3138 | goto out; |
3139 | } |
3140 | |
3141 | STATS_INC_ALLOCMISS(cachep); |
3142 | objp = cache_alloc_refill(cachep, flags); |
3143 | /* |
3144 | * the 'ac' may be updated by cache_alloc_refill(), |
3145 | * and kmemleak_erase() requires its correct value. |
3146 | */ |
3147 | ac = cpu_cache_get(cachep); |
3148 | |
3149 | out: |
3150 | /* |
3151 | * To avoid a false negative, if an object that is in one of the |
3152 | * per-CPU caches is leaked, we need to make sure kmemleak doesn't |
3153 | * treat the array pointers as a reference to the object. |
3154 | */ |
3155 | if (objp) |
3156 | kmemleak_erase(&ac->entry[ac->avail]); |
3157 | return objp; |
3158 | } |
3159 | |
3160 | #ifdef CONFIG_NUMA |
3161 | /* |
3162 | * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set. |
3163 | * |
3164 | * If we are in_interrupt, then process context, including cpusets and |
3165 | * mempolicy, may not apply and should not be used for allocation policy. |
3166 | */ |
3167 | static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags) |
3168 | { |
3169 | int nid_alloc, nid_here; |
3170 | |
3171 | if (in_interrupt() || (flags & __GFP_THISNODE)) |
3172 | return NULL; |
3173 | nid_alloc = nid_here = numa_mem_id(); |
3174 | if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD)) |
3175 | nid_alloc = cpuset_slab_spread_node(); |
3176 | else if (current->mempolicy) |
3177 | nid_alloc = mempolicy_slab_node(); |
3178 | if (nid_alloc != nid_here) |
3179 | return ____cache_alloc_node(cachep, flags, nid_alloc); |
3180 | return NULL; |
3181 | } |
3182 | |
3183 | /* |
3184 | * Fallback function if there was no memory available and no objects on a |
3185 | * certain node and fall back is permitted. First we scan all the |
3186 | * available node for available objects. If that fails then we |
3187 | * perform an allocation without specifying a node. This allows the page |
3188 | * allocator to do its reclaim / fallback magic. We then insert the |
3189 | * slab into the proper nodelist and then allocate from it. |
3190 | */ |
3191 | static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags) |
3192 | { |
3193 | struct zonelist *zonelist; |
3194 | struct zoneref *z; |
3195 | struct zone *zone; |
3196 | enum zone_type high_zoneidx = gfp_zone(flags); |
3197 | void *obj = NULL; |
3198 | struct page *page; |
3199 | int nid; |
3200 | unsigned int cpuset_mems_cookie; |
3201 | |
3202 | if (flags & __GFP_THISNODE) |
3203 | return NULL; |
3204 | |
3205 | retry_cpuset: |
3206 | cpuset_mems_cookie = read_mems_allowed_begin(); |
3207 | zonelist = node_zonelist(mempolicy_slab_node(), flags); |
3208 | |
3209 | retry: |
3210 | /* |
3211 | * Look through allowed nodes for objects available |
3212 | * from existing per node queues. |
3213 | */ |
3214 | for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) { |
3215 | nid = zone_to_nid(zone); |
3216 | |
3217 | if (cpuset_zone_allowed(zone, flags) && |
3218 | get_node(cache, nid) && |
3219 | get_node(cache, nid)->free_objects) { |
3220 | obj = ____cache_alloc_node(cache, |
3221 | gfp_exact_node(flags), nid); |
3222 | if (obj) |
3223 | break; |
3224 | } |
3225 | } |
3226 | |
3227 | if (!obj) { |
3228 | /* |
3229 | * This allocation will be performed within the constraints |
3230 | * of the current cpuset / memory policy requirements. |
3231 | * We may trigger various forms of reclaim on the allowed |
3232 | * set and go into memory reserves if necessary. |
3233 | */ |
3234 | page = cache_grow_begin(cache, flags, numa_mem_id()); |
3235 | cache_grow_end(cache, page); |
3236 | if (page) { |
3237 | nid = page_to_nid(page); |
3238 | obj = ____cache_alloc_node(cache, |
3239 | gfp_exact_node(flags), nid); |
3240 | |
3241 | /* |
3242 | * Another processor may allocate the objects in |
3243 | * the slab since we are not holding any locks. |
3244 | */ |
3245 | if (!obj) |
3246 | goto retry; |
3247 | } |
3248 | } |
3249 | |
3250 | if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie))) |
3251 | goto retry_cpuset; |
3252 | return obj; |
3253 | } |
3254 | |
3255 | /* |
3256 | * A interface to enable slab creation on nodeid |
3257 | */ |
3258 | static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, |
3259 | int nodeid) |
3260 | { |
3261 | struct page *page; |
3262 | struct kmem_cache_node *n; |
3263 | void *obj = NULL; |
3264 | void *list = NULL; |
3265 | |
3266 | VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES); |
3267 | n = get_node(cachep, nodeid); |
3268 | BUG_ON(!n); |
3269 | |
3270 | check_irq_off(); |
3271 | spin_lock(&n->list_lock); |
3272 | page = get_first_slab(n, false); |
3273 | if (!page) |
3274 | goto must_grow; |
3275 | |
3276 | check_spinlock_acquired_node(cachep, nodeid); |
3277 | |
3278 | STATS_INC_NODEALLOCS(cachep); |
3279 | STATS_INC_ACTIVE(cachep); |
3280 | STATS_SET_HIGH(cachep); |
3281 | |
3282 | BUG_ON(page->active == cachep->num); |
3283 | |
3284 | obj = slab_get_obj(cachep, page); |
3285 | n->free_objects--; |
3286 | |
3287 | fixup_slab_list(cachep, n, page, &list); |
3288 | |
3289 | spin_unlock(&n->list_lock); |
3290 | fixup_objfreelist_debug(cachep, &list); |
3291 | return obj; |
3292 | |
3293 | must_grow: |
3294 | spin_unlock(&n->list_lock); |
3295 | page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid); |
3296 | if (page) { |
3297 | /* This slab isn't counted yet so don't update free_objects */ |
3298 | obj = slab_get_obj(cachep, page); |
3299 | } |
3300 | cache_grow_end(cachep, page); |
3301 | |
3302 | return obj ? obj : fallback_alloc(cachep, flags); |
3303 | } |
3304 | |
3305 | static __always_inline void * |
3306 | slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid, |
3307 | unsigned long caller) |
3308 | { |
3309 | unsigned long save_flags; |
3310 | void *ptr; |
3311 | int slab_node = numa_mem_id(); |
3312 | |
3313 | flags &= gfp_allowed_mask; |
3314 | cachep = slab_pre_alloc_hook(cachep, flags); |
3315 | if (unlikely(!cachep)) |
3316 | return NULL; |
3317 | |
3318 | cache_alloc_debugcheck_before(cachep, flags); |
3319 | local_irq_save(save_flags); |
3320 | |
3321 | if (nodeid == NUMA_NO_NODE) |
3322 | nodeid = slab_node; |
3323 | |
3324 | if (unlikely(!get_node(cachep, nodeid))) { |
3325 | /* Node not bootstrapped yet */ |
3326 | ptr = fallback_alloc(cachep, flags); |
3327 | goto out; |
3328 | } |
3329 | |
3330 | if (nodeid == slab_node) { |
3331 | /* |
3332 | * Use the locally cached objects if possible. |
3333 | * However ____cache_alloc does not allow fallback |
3334 | * to other nodes. It may fail while we still have |
3335 | * objects on other nodes available. |
3336 | */ |
3337 | ptr = ____cache_alloc(cachep, flags); |
3338 | if (ptr) |
3339 | goto out; |
3340 | } |
3341 | /* ___cache_alloc_node can fall back to other nodes */ |
3342 | ptr = ____cache_alloc_node(cachep, flags, nodeid); |
3343 | out: |
3344 | local_irq_restore(save_flags); |
3345 | ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller); |
3346 | |
3347 | if (unlikely(flags & __GFP_ZERO) && ptr) |
3348 | memset(ptr, 0, cachep->object_size); |
3349 | |
3350 | slab_post_alloc_hook(cachep, flags, 1, &ptr); |
3351 | return ptr; |
3352 | } |
3353 | |
3354 | static __always_inline void * |
3355 | __do_cache_alloc(struct kmem_cache *cache, gfp_t flags) |
3356 | { |
3357 | void *objp; |
3358 | |
3359 | if (current->mempolicy || cpuset_do_slab_mem_spread()) { |
3360 | objp = alternate_node_alloc(cache, flags); |
3361 | if (objp) |
3362 | goto out; |
3363 | } |
3364 | objp = ____cache_alloc(cache, flags); |
3365 | |
3366 | /* |
3367 | * We may just have run out of memory on the local node. |
3368 | * ____cache_alloc_node() knows how to locate memory on other nodes |
3369 | */ |
3370 | if (!objp) |
3371 | objp = ____cache_alloc_node(cache, flags, numa_mem_id()); |
3372 | |
3373 | out: |
3374 | return objp; |
3375 | } |
3376 | #else |
3377 | |
3378 | static __always_inline void * |
3379 | __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
3380 | { |
3381 | return ____cache_alloc(cachep, flags); |
3382 | } |
3383 | |
3384 | #endif /* CONFIG_NUMA */ |
3385 | |
3386 | static __always_inline void * |
3387 | slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller) |
3388 | { |
3389 | unsigned long save_flags; |
3390 | void *objp; |
3391 | |
3392 | flags &= gfp_allowed_mask; |
3393 | cachep = slab_pre_alloc_hook(cachep, flags); |
3394 | if (unlikely(!cachep)) |
3395 | return NULL; |
3396 | |
3397 | cache_alloc_debugcheck_before(cachep, flags); |
3398 | local_irq_save(save_flags); |
3399 | objp = __do_cache_alloc(cachep, flags); |
3400 | local_irq_restore(save_flags); |
3401 | objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller); |
3402 | prefetchw(objp); |
3403 | |
3404 | if (unlikely(flags & __GFP_ZERO) && objp) |
3405 | memset(objp, 0, cachep->object_size); |
3406 | |
3407 | slab_post_alloc_hook(cachep, flags, 1, &objp); |
3408 | return objp; |
3409 | } |
3410 | |
3411 | /* |
3412 | * Caller needs to acquire correct kmem_cache_node's list_lock |
3413 | * @list: List of detached free slabs should be freed by caller |
3414 | */ |
3415 | static void free_block(struct kmem_cache *cachep, void **objpp, |
3416 | int nr_objects, int node, struct list_head *list) |
3417 | { |
3418 | int i; |
3419 | struct kmem_cache_node *n = get_node(cachep, node); |
3420 | struct page *page; |
3421 | |
3422 | n->free_objects += nr_objects; |
3423 | |
3424 | for (i = 0; i < nr_objects; i++) { |
3425 | void *objp; |
3426 | struct page *page; |
3427 | |
3428 | objp = objpp[i]; |
3429 | |
3430 | page = virt_to_head_page(objp); |
3431 | list_del(&page->lru); |
3432 | check_spinlock_acquired_node(cachep, node); |
3433 | slab_put_obj(cachep, page, objp); |
3434 | STATS_DEC_ACTIVE(cachep); |
3435 | |
3436 | /* fixup slab chains */ |
3437 | if (page->active == 0) |
3438 | list_add(&page->lru, &n->slabs_free); |
3439 | else { |
3440 | /* Unconditionally move a slab to the end of the |
3441 | * partial list on free - maximum time for the |
3442 | * other objects to be freed, too. |
3443 | */ |
3444 | list_add_tail(&page->lru, &n->slabs_partial); |
3445 | } |
3446 | } |
3447 | |
3448 | while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) { |
3449 | n->free_objects -= cachep->num; |
3450 | |
3451 | page = list_last_entry(&n->slabs_free, struct page, lru); |
3452 | list_move(&page->lru, list); |
3453 | n->num_slabs--; |
3454 | } |
3455 | } |
3456 | |
3457 | static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac) |
3458 | { |
3459 | int batchcount; |
3460 | struct kmem_cache_node *n; |
3461 | int node = numa_mem_id(); |
3462 | LIST_HEAD(list); |
3463 | |
3464 | batchcount = ac->batchcount; |
3465 | |
3466 | check_irq_off(); |
3467 | n = get_node(cachep, node); |
3468 | spin_lock(&n->list_lock); |
3469 | if (n->shared) { |
3470 | struct array_cache *shared_array = n->shared; |
3471 | int max = shared_array->limit - shared_array->avail; |
3472 | if (max) { |
3473 | if (batchcount > max) |
3474 | batchcount = max; |
3475 | memcpy(&(shared_array->entry[shared_array->avail]), |
3476 | ac->entry, sizeof(void *) * batchcount); |
3477 | shared_array->avail += batchcount; |
3478 | goto free_done; |
3479 | } |
3480 | } |
3481 | |
3482 | free_block(cachep, ac->entry, batchcount, node, &list); |
3483 | free_done: |
3484 | #if STATS |
3485 | { |
3486 | int i = 0; |
3487 | struct page *page; |
3488 | |
3489 | list_for_each_entry(page, &n->slabs_free, lru) { |
3490 | BUG_ON(page->active); |
3491 | |
3492 | i++; |
3493 | } |
3494 | STATS_SET_FREEABLE(cachep, i); |
3495 | } |
3496 | #endif |
3497 | spin_unlock(&n->list_lock); |
3498 | slabs_destroy(cachep, &list); |
3499 | ac->avail -= batchcount; |
3500 | memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail); |
3501 | } |
3502 | |
3503 | /* |
3504 | * Release an obj back to its cache. If the obj has a constructed state, it must |
3505 | * be in this state _before_ it is released. Called with disabled ints. |
3506 | */ |
3507 | static inline void __cache_free(struct kmem_cache *cachep, void *objp, |
3508 | unsigned long caller) |
3509 | { |
3510 | /* Put the object into the quarantine, don't touch it for now. */ |
3511 | if (kasan_slab_free(cachep, objp)) |
3512 | return; |
3513 | |
3514 | ___cache_free(cachep, objp, caller); |
3515 | } |
3516 | |
3517 | void ___cache_free(struct kmem_cache *cachep, void *objp, |
3518 | unsigned long caller) |
3519 | { |
3520 | struct array_cache *ac = cpu_cache_get(cachep); |
3521 | |
3522 | check_irq_off(); |
3523 | kmemleak_free_recursive(objp, cachep->flags); |
3524 | objp = cache_free_debugcheck(cachep, objp, caller); |
3525 | |
3526 | kmemcheck_slab_free(cachep, objp, cachep->object_size); |
3527 | |
3528 | /* |
3529 | * Skip calling cache_free_alien() when the platform is not numa. |
3530 | * This will avoid cache misses that happen while accessing slabp (which |
3531 | * is per page memory reference) to get nodeid. Instead use a global |
3532 | * variable to skip the call, which is mostly likely to be present in |
3533 | * the cache. |
3534 | */ |
3535 | if (nr_online_nodes > 1 && cache_free_alien(cachep, objp)) |
3536 | return; |
3537 | |
3538 | if (ac->avail < ac->limit) { |
3539 | STATS_INC_FREEHIT(cachep); |
3540 | } else { |
3541 | STATS_INC_FREEMISS(cachep); |
3542 | cache_flusharray(cachep, ac); |
3543 | } |
3544 | |
3545 | if (sk_memalloc_socks()) { |
3546 | struct page *page = virt_to_head_page(objp); |
3547 | |
3548 | if (unlikely(PageSlabPfmemalloc(page))) { |
3549 | cache_free_pfmemalloc(cachep, page, objp); |
3550 | return; |
3551 | } |
3552 | } |
3553 | |
3554 | ac->entry[ac->avail++] = objp; |
3555 | } |
3556 | |
3557 | /** |
3558 | * kmem_cache_alloc - Allocate an object |
3559 | * @cachep: The cache to allocate from. |
3560 | * @flags: See kmalloc(). |
3561 | * |
3562 | * Allocate an object from this cache. The flags are only relevant |
3563 | * if the cache has no available objects. |
3564 | */ |
3565 | void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags) |
3566 | { |
3567 | void *ret = slab_alloc(cachep, flags, _RET_IP_); |
3568 | |
3569 | kasan_slab_alloc(cachep, ret, flags); |
3570 | trace_kmem_cache_alloc(_RET_IP_, ret, |
3571 | cachep->object_size, cachep->size, flags); |
3572 | |
3573 | return ret; |
3574 | } |
3575 | EXPORT_SYMBOL(kmem_cache_alloc); |
3576 | |
3577 | static __always_inline void |
3578 | cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags, |
3579 | size_t size, void **p, unsigned long caller) |
3580 | { |
3581 | size_t i; |
3582 | |
3583 | for (i = 0; i < size; i++) |
3584 | p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller); |
3585 | } |
3586 | |
3587 | int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size, |
3588 | void **p) |
3589 | { |
3590 | size_t i; |
3591 | |
3592 | s = slab_pre_alloc_hook(s, flags); |
3593 | if (!s) |
3594 | return 0; |
3595 | |
3596 | cache_alloc_debugcheck_before(s, flags); |
3597 | |
3598 | local_irq_disable(); |
3599 | for (i = 0; i < size; i++) { |
3600 | void *objp = __do_cache_alloc(s, flags); |
3601 | |
3602 | if (unlikely(!objp)) |
3603 | goto error; |
3604 | p[i] = objp; |
3605 | } |
3606 | local_irq_enable(); |
3607 | |
3608 | cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_); |
3609 | |
3610 | /* Clear memory outside IRQ disabled section */ |
3611 | if (unlikely(flags & __GFP_ZERO)) |
3612 | for (i = 0; i < size; i++) |
3613 | memset(p[i], 0, s->object_size); |
3614 | |
3615 | slab_post_alloc_hook(s, flags, size, p); |
3616 | /* FIXME: Trace call missing. Christoph would like a bulk variant */ |
3617 | return size; |
3618 | error: |
3619 | local_irq_enable(); |
3620 | cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_); |
3621 | slab_post_alloc_hook(s, flags, i, p); |
3622 | __kmem_cache_free_bulk(s, i, p); |
3623 | return 0; |
3624 | } |
3625 | EXPORT_SYMBOL(kmem_cache_alloc_bulk); |
3626 | |
3627 | #ifdef CONFIG_TRACING |
3628 | void * |
3629 | kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size) |
3630 | { |
3631 | void *ret; |
3632 | |
3633 | ret = slab_alloc(cachep, flags, _RET_IP_); |
3634 | |
3635 | kasan_kmalloc(cachep, ret, size, flags); |
3636 | trace_kmalloc(_RET_IP_, ret, |
3637 | size, cachep->size, flags); |
3638 | return ret; |
3639 | } |
3640 | EXPORT_SYMBOL(kmem_cache_alloc_trace); |
3641 | #endif |
3642 | |
3643 | #ifdef CONFIG_NUMA |
3644 | /** |
3645 | * kmem_cache_alloc_node - Allocate an object on the specified node |
3646 | * @cachep: The cache to allocate from. |
3647 | * @flags: See kmalloc(). |
3648 | * @nodeid: node number of the target node. |
3649 | * |
3650 | * Identical to kmem_cache_alloc but it will allocate memory on the given |
3651 | * node, which can improve the performance for cpu bound structures. |
3652 | * |
3653 | * Fallback to other node is possible if __GFP_THISNODE is not set. |
3654 | */ |
3655 | void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid) |
3656 | { |
3657 | void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
3658 | |
3659 | kasan_slab_alloc(cachep, ret, flags); |
3660 | trace_kmem_cache_alloc_node(_RET_IP_, ret, |
3661 | cachep->object_size, cachep->size, |
3662 | flags, nodeid); |
3663 | |
3664 | return ret; |
3665 | } |
3666 | EXPORT_SYMBOL(kmem_cache_alloc_node); |
3667 | |
3668 | #ifdef CONFIG_TRACING |
3669 | void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep, |
3670 | gfp_t flags, |
3671 | int nodeid, |
3672 | size_t size) |
3673 | { |
3674 | void *ret; |
3675 | |
3676 | ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_); |
3677 | |
3678 | kasan_kmalloc(cachep, ret, size, flags); |
3679 | trace_kmalloc_node(_RET_IP_, ret, |
3680 | size, cachep->size, |
3681 | flags, nodeid); |
3682 | return ret; |
3683 | } |
3684 | EXPORT_SYMBOL(kmem_cache_alloc_node_trace); |
3685 | #endif |
3686 | |
3687 | static __always_inline void * |
3688 | __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller) |
3689 | { |
3690 | struct kmem_cache *cachep; |
3691 | void *ret; |
3692 | |
3693 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) |
3694 | return NULL; |
3695 | cachep = kmalloc_slab(size, flags); |
3696 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3697 | return cachep; |
3698 | ret = kmem_cache_alloc_node_trace(cachep, flags, node, size); |
3699 | kasan_kmalloc(cachep, ret, size, flags); |
3700 | |
3701 | return ret; |
3702 | } |
3703 | |
3704 | void *__kmalloc_node(size_t size, gfp_t flags, int node) |
3705 | { |
3706 | return __do_kmalloc_node(size, flags, node, _RET_IP_); |
3707 | } |
3708 | EXPORT_SYMBOL(__kmalloc_node); |
3709 | |
3710 | void *__kmalloc_node_track_caller(size_t size, gfp_t flags, |
3711 | int node, unsigned long caller) |
3712 | { |
3713 | return __do_kmalloc_node(size, flags, node, caller); |
3714 | } |
3715 | EXPORT_SYMBOL(__kmalloc_node_track_caller); |
3716 | #endif /* CONFIG_NUMA */ |
3717 | |
3718 | /** |
3719 | * __do_kmalloc - allocate memory |
3720 | * @size: how many bytes of memory are required. |
3721 | * @flags: the type of memory to allocate (see kmalloc). |
3722 | * @caller: function caller for debug tracking of the caller |
3723 | */ |
3724 | static __always_inline void *__do_kmalloc(size_t size, gfp_t flags, |
3725 | unsigned long caller) |
3726 | { |
3727 | struct kmem_cache *cachep; |
3728 | void *ret; |
3729 | |
3730 | if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) |
3731 | return NULL; |
3732 | cachep = kmalloc_slab(size, flags); |
3733 | if (unlikely(ZERO_OR_NULL_PTR(cachep))) |
3734 | return cachep; |
3735 | ret = slab_alloc(cachep, flags, caller); |
3736 | |
3737 | kasan_kmalloc(cachep, ret, size, flags); |
3738 | trace_kmalloc(caller, ret, |
3739 | size, cachep->size, flags); |
3740 | |
3741 | return ret; |
3742 | } |
3743 | |
3744 | void *__kmalloc(size_t size, gfp_t flags) |
3745 | { |
3746 | return __do_kmalloc(size, flags, _RET_IP_); |
3747 | } |
3748 | EXPORT_SYMBOL(__kmalloc); |
3749 | |
3750 | void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller) |
3751 | { |
3752 | return __do_kmalloc(size, flags, caller); |
3753 | } |
3754 | EXPORT_SYMBOL(__kmalloc_track_caller); |
3755 | |
3756 | /** |
3757 | * kmem_cache_free - Deallocate an object |
3758 | * @cachep: The cache the allocation was from. |
3759 | * @objp: The previously allocated object. |
3760 | * |
3761 | * Free an object which was previously allocated from this |
3762 | * cache. |
3763 | */ |
3764 | void kmem_cache_free(struct kmem_cache *cachep, void *objp) |
3765 | { |
3766 | unsigned long flags; |
3767 | cachep = cache_from_obj(cachep, objp); |
3768 | if (!cachep) |
3769 | return; |
3770 | |
3771 | local_irq_save(flags); |
3772 | debug_check_no_locks_freed(objp, cachep->object_size); |
3773 | if (!(cachep->flags & SLAB_DEBUG_OBJECTS)) |
3774 | debug_check_no_obj_freed(objp, cachep->object_size); |
3775 | __cache_free(cachep, objp, _RET_IP_); |
3776 | local_irq_restore(flags); |
3777 | |
3778 | trace_kmem_cache_free(_RET_IP_, objp); |
3779 | } |
3780 | EXPORT_SYMBOL(kmem_cache_free); |
3781 | |
3782 | void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p) |
3783 | { |
3784 | struct kmem_cache *s; |
3785 | size_t i; |
3786 | |
3787 | local_irq_disable(); |
3788 | for (i = 0; i < size; i++) { |
3789 | void *objp = p[i]; |
3790 | |
3791 | if (!orig_s) /* called via kfree_bulk */ |
3792 | s = virt_to_cache(objp); |
3793 | else |
3794 | s = cache_from_obj(orig_s, objp); |
3795 | |
3796 | debug_check_no_locks_freed(objp, s->object_size); |
3797 | if (!(s->flags & SLAB_DEBUG_OBJECTS)) |
3798 | debug_check_no_obj_freed(objp, s->object_size); |
3799 | |
3800 | __cache_free(s, objp, _RET_IP_); |
3801 | } |
3802 | local_irq_enable(); |
3803 | |
3804 | /* FIXME: add tracing */ |
3805 | } |
3806 | EXPORT_SYMBOL(kmem_cache_free_bulk); |
3807 | |
3808 | /** |
3809 | * kfree - free previously allocated memory |
3810 | * @objp: pointer returned by kmalloc. |
3811 | * |
3812 | * If @objp is NULL, no operation is performed. |
3813 | * |
3814 | * Don't free memory not originally allocated by kmalloc() |
3815 | * or you will run into trouble. |
3816 | */ |
3817 | void kfree(const void *objp) |
3818 | { |
3819 | struct kmem_cache *c; |
3820 | unsigned long flags; |
3821 | |
3822 | trace_kfree(_RET_IP_, objp); |
3823 | |
3824 | if (unlikely(ZERO_OR_NULL_PTR(objp))) |
3825 | return; |
3826 | local_irq_save(flags); |
3827 | kfree_debugcheck(objp); |
3828 | c = virt_to_cache(objp); |
3829 | debug_check_no_locks_freed(objp, c->object_size); |
3830 | |
3831 | debug_check_no_obj_freed(objp, c->object_size); |
3832 | __cache_free(c, (void *)objp, _RET_IP_); |
3833 | local_irq_restore(flags); |
3834 | } |
3835 | EXPORT_SYMBOL(kfree); |
3836 | |
3837 | /* |
3838 | * This initializes kmem_cache_node or resizes various caches for all nodes. |
3839 | */ |
3840 | static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp) |
3841 | { |
3842 | int ret; |
3843 | int node; |
3844 | struct kmem_cache_node *n; |
3845 | |
3846 | for_each_online_node(node) { |
3847 | ret = setup_kmem_cache_node(cachep, node, gfp, true); |
3848 | if (ret) |
3849 | goto fail; |
3850 | |
3851 | } |
3852 | |
3853 | return 0; |
3854 | |
3855 | fail: |
3856 | if (!cachep->list.next) { |
3857 | /* Cache is not active yet. Roll back what we did */ |
3858 | node--; |
3859 | while (node >= 0) { |
3860 | n = get_node(cachep, node); |
3861 | if (n) { |
3862 | kfree(n->shared); |
3863 | free_alien_cache(n->alien); |
3864 | kfree(n); |
3865 | cachep->node[node] = NULL; |
3866 | } |
3867 | node--; |
3868 | } |
3869 | } |
3870 | return -ENOMEM; |
3871 | } |
3872 | |
3873 | /* Always called with the slab_mutex held */ |
3874 | static int __do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3875 | int batchcount, int shared, gfp_t gfp) |
3876 | { |
3877 | struct array_cache __percpu *cpu_cache, *prev; |
3878 | int cpu; |
3879 | |
3880 | cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount); |
3881 | if (!cpu_cache) |
3882 | return -ENOMEM; |
3883 | |
3884 | prev = cachep->cpu_cache; |
3885 | cachep->cpu_cache = cpu_cache; |
3886 | kick_all_cpus_sync(); |
3887 | |
3888 | check_irq_on(); |
3889 | cachep->batchcount = batchcount; |
3890 | cachep->limit = limit; |
3891 | cachep->shared = shared; |
3892 | |
3893 | if (!prev) |
3894 | goto setup_node; |
3895 | |
3896 | for_each_online_cpu(cpu) { |
3897 | LIST_HEAD(list); |
3898 | int node; |
3899 | struct kmem_cache_node *n; |
3900 | struct array_cache *ac = per_cpu_ptr(prev, cpu); |
3901 | |
3902 | node = cpu_to_mem(cpu); |
3903 | n = get_node(cachep, node); |
3904 | spin_lock_irq(&n->list_lock); |
3905 | free_block(cachep, ac->entry, ac->avail, node, &list); |
3906 | spin_unlock_irq(&n->list_lock); |
3907 | slabs_destroy(cachep, &list); |
3908 | } |
3909 | free_percpu(prev); |
3910 | |
3911 | setup_node: |
3912 | return setup_kmem_cache_nodes(cachep, gfp); |
3913 | } |
3914 | |
3915 | static int do_tune_cpucache(struct kmem_cache *cachep, int limit, |
3916 | int batchcount, int shared, gfp_t gfp) |
3917 | { |
3918 | int ret; |
3919 | struct kmem_cache *c; |
3920 | |
3921 | ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp); |
3922 | |
3923 | if (slab_state < FULL) |
3924 | return ret; |
3925 | |
3926 | if ((ret < 0) || !is_root_cache(cachep)) |
3927 | return ret; |
3928 | |
3929 | lockdep_assert_held(&slab_mutex); |
3930 | for_each_memcg_cache(c, cachep) { |
3931 | /* return value determined by the root cache only */ |
3932 | __do_tune_cpucache(c, limit, batchcount, shared, gfp); |
3933 | } |
3934 | |
3935 | return ret; |
3936 | } |
3937 | |
3938 | /* Called with slab_mutex held always */ |
3939 | static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp) |
3940 | { |
3941 | int err; |
3942 | int limit = 0; |
3943 | int shared = 0; |
3944 | int batchcount = 0; |
3945 | |
3946 | err = cache_random_seq_create(cachep, cachep->num, gfp); |
3947 | if (err) |
3948 | goto end; |
3949 | |
3950 | if (!is_root_cache(cachep)) { |
3951 | struct kmem_cache *root = memcg_root_cache(cachep); |
3952 | limit = root->limit; |
3953 | shared = root->shared; |
3954 | batchcount = root->batchcount; |
3955 | } |
3956 | |
3957 | if (limit && shared && batchcount) |
3958 | goto skip_setup; |
3959 | /* |
3960 | * The head array serves three purposes: |
3961 | * - create a LIFO ordering, i.e. return objects that are cache-warm |
3962 | * - reduce the number of spinlock operations. |
3963 | * - reduce the number of linked list operations on the slab and |
3964 | * bufctl chains: array operations are cheaper. |
3965 | * The numbers are guessed, we should auto-tune as described by |
3966 | * Bonwick. |
3967 | */ |
3968 | if (cachep->size > 131072) |
3969 | limit = 1; |
3970 | else if (cachep->size > PAGE_SIZE) |
3971 | limit = 8; |
3972 | else if (cachep->size > 1024) |
3973 | limit = 24; |
3974 | else if (cachep->size > 256) |
3975 | limit = 54; |
3976 | else |
3977 | limit = 120; |
3978 | |
3979 | /* |
3980 | * CPU bound tasks (e.g. network routing) can exhibit cpu bound |
3981 | * allocation behaviour: Most allocs on one cpu, most free operations |
3982 | * on another cpu. For these cases, an efficient object passing between |
3983 | * cpus is necessary. This is provided by a shared array. The array |
3984 | * replaces Bonwick's magazine layer. |
3985 | * On uniprocessor, it's functionally equivalent (but less efficient) |
3986 | * to a larger limit. Thus disabled by default. |
3987 | */ |
3988 | shared = 0; |
3989 | if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1) |
3990 | shared = 8; |
3991 | |
3992 | #if DEBUG |
3993 | /* |
3994 | * With debugging enabled, large batchcount lead to excessively long |
3995 | * periods with disabled local interrupts. Limit the batchcount |
3996 | */ |
3997 | if (limit > 32) |
3998 | limit = 32; |
3999 | #endif |
4000 | batchcount = (limit + 1) / 2; |
4001 | skip_setup: |
4002 | err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp); |
4003 | end: |
4004 | if (err) |
4005 | pr_err("enable_cpucache failed for %s, error %d\n", |
4006 | cachep->name, -err); |
4007 | return err; |
4008 | } |
4009 | |
4010 | /* |
4011 | * Drain an array if it contains any elements taking the node lock only if |
4012 | * necessary. Note that the node listlock also protects the array_cache |
4013 | * if drain_array() is used on the shared array. |
4014 | */ |
4015 | static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n, |
4016 | struct array_cache *ac, int node) |
4017 | { |
4018 | LIST_HEAD(list); |
4019 | |
4020 | /* ac from n->shared can be freed if we don't hold the slab_mutex. */ |
4021 | check_mutex_acquired(); |
4022 | |
4023 | if (!ac || !ac->avail) |
4024 | return; |
4025 | |
4026 | if (ac->touched) { |
4027 | ac->touched = 0; |
4028 | return; |
4029 | } |
4030 | |
4031 | spin_lock_irq(&n->list_lock); |
4032 | drain_array_locked(cachep, ac, node, false, &list); |
4033 | spin_unlock_irq(&n->list_lock); |
4034 | |
4035 | slabs_destroy(cachep, &list); |
4036 | } |
4037 | |
4038 | /** |
4039 | * cache_reap - Reclaim memory from caches. |
4040 | * @w: work descriptor |
4041 | * |
4042 | * Called from workqueue/eventd every few seconds. |
4043 | * Purpose: |
4044 | * - clear the per-cpu caches for this CPU. |
4045 | * - return freeable pages to the main free memory pool. |
4046 | * |
4047 | * If we cannot acquire the cache chain mutex then just give up - we'll try |
4048 | * again on the next iteration. |
4049 | */ |
4050 | static void cache_reap(struct work_struct *w) |
4051 | { |
4052 | struct kmem_cache *searchp; |
4053 | struct kmem_cache_node *n; |
4054 | int node = numa_mem_id(); |
4055 | struct delayed_work *work = to_delayed_work(w); |
4056 | |
4057 | if (!mutex_trylock(&slab_mutex)) |
4058 | /* Give up. Setup the next iteration. */ |
4059 | goto out; |
4060 | |
4061 | list_for_each_entry(searchp, &slab_caches, list) { |
4062 | check_irq_on(); |
4063 | |
4064 | /* |
4065 | * We only take the node lock if absolutely necessary and we |
4066 | * have established with reasonable certainty that |
4067 | * we can do some work if the lock was obtained. |
4068 | */ |
4069 | n = get_node(searchp, node); |
4070 | |
4071 | reap_alien(searchp, n); |
4072 | |
4073 | drain_array(searchp, n, cpu_cache_get(searchp), node); |
4074 | |
4075 | /* |
4076 | * These are racy checks but it does not matter |
4077 | * if we skip one check or scan twice. |
4078 | */ |
4079 | if (time_after(n->next_reap, jiffies)) |
4080 | goto next; |
4081 | |
4082 | n->next_reap = jiffies + REAPTIMEOUT_NODE; |
4083 | |
4084 | drain_array(searchp, n, n->shared, node); |
4085 | |
4086 | if (n->free_touched) |
4087 | n->free_touched = 0; |
4088 | else { |
4089 | int freed; |
4090 | |
4091 | freed = drain_freelist(searchp, n, (n->free_limit + |
4092 | 5 * searchp->num - 1) / (5 * searchp->num)); |
4093 | STATS_ADD_REAPED(searchp, freed); |
4094 | } |
4095 | next: |
4096 | cond_resched(); |
4097 | } |
4098 | check_irq_on(); |
4099 | mutex_unlock(&slab_mutex); |
4100 | next_reap_node(); |
4101 | out: |
4102 | /* Set up the next iteration */ |
4103 | schedule_delayed_work_on(smp_processor_id(), work, |
4104 | round_jiffies_relative(REAPTIMEOUT_AC)); |
4105 | } |
4106 | |
4107 | #ifdef CONFIG_SLABINFO |
4108 | void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo) |
4109 | { |
4110 | struct page *page; |
4111 | unsigned long active_objs; |
4112 | unsigned long num_objs; |
4113 | unsigned long active_slabs = 0; |
4114 | unsigned long num_slabs, free_objects = 0, shared_avail = 0; |
4115 | unsigned long num_slabs_partial = 0, num_slabs_free = 0; |
4116 | unsigned long num_slabs_full = 0; |
4117 | const char *name; |
4118 | char *error = NULL; |
4119 | int node; |
4120 | struct kmem_cache_node *n; |
4121 | |
4122 | active_objs = 0; |
4123 | num_slabs = 0; |
4124 | for_each_kmem_cache_node(cachep, node, n) { |
4125 | |
4126 | check_irq_on(); |
4127 | spin_lock_irq(&n->list_lock); |
4128 | |
4129 | num_slabs += n->num_slabs; |
4130 | |
4131 | list_for_each_entry(page, &n->slabs_partial, lru) { |
4132 | if (page->active == cachep->num && !error) |
4133 | error = "slabs_partial accounting error"; |
4134 | if (!page->active && !error) |
4135 | error = "slabs_partial accounting error"; |
4136 | active_objs += page->active; |
4137 | num_slabs_partial++; |
4138 | } |
4139 | |
4140 | list_for_each_entry(page, &n->slabs_free, lru) { |
4141 | if (page->active && !error) |
4142 | error = "slabs_free accounting error"; |
4143 | num_slabs_free++; |
4144 | } |
4145 | |
4146 | free_objects += n->free_objects; |
4147 | if (n->shared) |
4148 | shared_avail += n->shared->avail; |
4149 | |
4150 | spin_unlock_irq(&n->list_lock); |
4151 | } |
4152 | num_objs = num_slabs * cachep->num; |
4153 | active_slabs = num_slabs - num_slabs_free; |
4154 | num_slabs_full = num_slabs - (num_slabs_partial + num_slabs_free); |
4155 | active_objs += (num_slabs_full * cachep->num); |
4156 | |
4157 | if (num_objs - active_objs != free_objects && !error) |
4158 | error = "free_objects accounting error"; |
4159 | |
4160 | name = cachep->name; |
4161 | if (error) |
4162 | pr_err("slab: cache %s error: %s\n", name, error); |
4163 | |
4164 | sinfo->active_objs = active_objs; |
4165 | sinfo->num_objs = num_objs; |
4166 | sinfo->active_slabs = active_slabs; |
4167 | sinfo->num_slabs = num_slabs; |
4168 | sinfo->shared_avail = shared_avail; |
4169 | sinfo->limit = cachep->limit; |
4170 | sinfo->batchcount = cachep->batchcount; |
4171 | sinfo->shared = cachep->shared; |
4172 | sinfo->objects_per_slab = cachep->num; |
4173 | sinfo->cache_order = cachep->gfporder; |
4174 | } |
4175 | |
4176 | void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep) |
4177 | { |
4178 | #if STATS |
4179 | { /* node stats */ |
4180 | unsigned long high = cachep->high_mark; |
4181 | unsigned long allocs = cachep->num_allocations; |
4182 | unsigned long grown = cachep->grown; |
4183 | unsigned long reaped = cachep->reaped; |
4184 | unsigned long errors = cachep->errors; |
4185 | unsigned long max_freeable = cachep->max_freeable; |
4186 | unsigned long node_allocs = cachep->node_allocs; |
4187 | unsigned long node_frees = cachep->node_frees; |
4188 | unsigned long overflows = cachep->node_overflow; |
4189 | |
4190 | seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu", |
4191 | allocs, high, grown, |
4192 | reaped, errors, max_freeable, node_allocs, |
4193 | node_frees, overflows); |
4194 | } |
4195 | /* cpu stats */ |
4196 | { |
4197 | unsigned long allochit = atomic_read(&cachep->allochit); |
4198 | unsigned long allocmiss = atomic_read(&cachep->allocmiss); |
4199 | unsigned long freehit = atomic_read(&cachep->freehit); |
4200 | unsigned long freemiss = atomic_read(&cachep->freemiss); |
4201 | |
4202 | seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu", |
4203 | allochit, allocmiss, freehit, freemiss); |
4204 | } |
4205 | #endif |
4206 | } |
4207 | |
4208 | #define MAX_SLABINFO_WRITE 128 |
4209 | /** |
4210 | * slabinfo_write - Tuning for the slab allocator |
4211 | * @file: unused |
4212 | * @buffer: user buffer |
4213 | * @count: data length |
4214 | * @ppos: unused |
4215 | */ |
4216 | ssize_t slabinfo_write(struct file *file, const char __user *buffer, |
4217 | size_t count, loff_t *ppos) |
4218 | { |
4219 | char kbuf[MAX_SLABINFO_WRITE + 1], *tmp; |
4220 | int limit, batchcount, shared, res; |
4221 | struct kmem_cache *cachep; |
4222 | |
4223 | if (count > MAX_SLABINFO_WRITE) |
4224 | return -EINVAL; |
4225 | if (copy_from_user(&kbuf, buffer, count)) |
4226 | return -EFAULT; |
4227 | kbuf[MAX_SLABINFO_WRITE] = '\0'; |
4228 | |
4229 | tmp = strchr(kbuf, ' '); |
4230 | if (!tmp) |
4231 | return -EINVAL; |
4232 | *tmp = '\0'; |
4233 | tmp++; |
4234 | if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3) |
4235 | return -EINVAL; |
4236 | |
4237 | /* Find the cache in the chain of caches. */ |
4238 | mutex_lock(&slab_mutex); |
4239 | res = -EINVAL; |
4240 | list_for_each_entry(cachep, &slab_caches, list) { |
4241 | if (!strcmp(cachep->name, kbuf)) { |
4242 | if (limit < 1 || batchcount < 1 || |
4243 | batchcount > limit || shared < 0) { |
4244 | res = 0; |
4245 | } else { |
4246 | res = do_tune_cpucache(cachep, limit, |
4247 | batchcount, shared, |
4248 | GFP_KERNEL); |
4249 | } |
4250 | break; |
4251 | } |
4252 | } |
4253 | mutex_unlock(&slab_mutex); |
4254 | if (res >= 0) |
4255 | res = count; |
4256 | return res; |
4257 | } |
4258 | |
4259 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4260 | |
4261 | static inline int add_caller(unsigned long *n, unsigned long v) |
4262 | { |
4263 | unsigned long *p; |
4264 | int l; |
4265 | if (!v) |
4266 | return 1; |
4267 | l = n[1]; |
4268 | p = n + 2; |
4269 | while (l) { |
4270 | int i = l/2; |
4271 | unsigned long *q = p + 2 * i; |
4272 | if (*q == v) { |
4273 | q[1]++; |
4274 | return 1; |
4275 | } |
4276 | if (*q > v) { |
4277 | l = i; |
4278 | } else { |
4279 | p = q + 2; |
4280 | l -= i + 1; |
4281 | } |
4282 | } |
4283 | if (++n[1] == n[0]) |
4284 | return 0; |
4285 | memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n)); |
4286 | p[0] = v; |
4287 | p[1] = 1; |
4288 | return 1; |
4289 | } |
4290 | |
4291 | static void handle_slab(unsigned long *n, struct kmem_cache *c, |
4292 | struct page *page) |
4293 | { |
4294 | void *p; |
4295 | int i, j; |
4296 | unsigned long v; |
4297 | |
4298 | if (n[0] == n[1]) |
4299 | return; |
4300 | for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) { |
4301 | bool active = true; |
4302 | |
4303 | for (j = page->active; j < c->num; j++) { |
4304 | if (get_free_obj(page, j) == i) { |
4305 | active = false; |
4306 | break; |
4307 | } |
4308 | } |
4309 | |
4310 | if (!active) |
4311 | continue; |
4312 | |
4313 | /* |
4314 | * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table |
4315 | * mapping is established when actual object allocation and |
4316 | * we could mistakenly access the unmapped object in the cpu |
4317 | * cache. |
4318 | */ |
4319 | if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v))) |
4320 | continue; |
4321 | |
4322 | if (!add_caller(n, v)) |
4323 | return; |
4324 | } |
4325 | } |
4326 | |
4327 | static void show_symbol(struct seq_file *m, unsigned long address) |
4328 | { |
4329 | #ifdef CONFIG_KALLSYMS |
4330 | unsigned long offset, size; |
4331 | char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN]; |
4332 | |
4333 | if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) { |
4334 | seq_printf(m, "%s+%#lx/%#lx", name, offset, size); |
4335 | if (modname[0]) |
4336 | seq_printf(m, " [%s]", modname); |
4337 | return; |
4338 | } |
4339 | #endif |
4340 | seq_printf(m, "%px", (void *)address); |
4341 | } |
4342 | |
4343 | static int leaks_show(struct seq_file *m, void *p) |
4344 | { |
4345 | struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list); |
4346 | struct page *page; |
4347 | struct kmem_cache_node *n; |
4348 | const char *name; |
4349 | unsigned long *x = m->private; |
4350 | int node; |
4351 | int i; |
4352 | |
4353 | if (!(cachep->flags & SLAB_STORE_USER)) |
4354 | return 0; |
4355 | if (!(cachep->flags & SLAB_RED_ZONE)) |
4356 | return 0; |
4357 | |
4358 | /* |
4359 | * Set store_user_clean and start to grab stored user information |
4360 | * for all objects on this cache. If some alloc/free requests comes |
4361 | * during the processing, information would be wrong so restart |
4362 | * whole processing. |
4363 | */ |
4364 | do { |
4365 | set_store_user_clean(cachep); |
4366 | drain_cpu_caches(cachep); |
4367 | |
4368 | x[1] = 0; |
4369 | |
4370 | for_each_kmem_cache_node(cachep, node, n) { |
4371 | |
4372 | check_irq_on(); |
4373 | spin_lock_irq(&n->list_lock); |
4374 | |
4375 | list_for_each_entry(page, &n->slabs_full, lru) |
4376 | handle_slab(x, cachep, page); |
4377 | list_for_each_entry(page, &n->slabs_partial, lru) |
4378 | handle_slab(x, cachep, page); |
4379 | spin_unlock_irq(&n->list_lock); |
4380 | } |
4381 | } while (!is_store_user_clean(cachep)); |
4382 | |
4383 | name = cachep->name; |
4384 | if (x[0] == x[1]) { |
4385 | /* Increase the buffer size */ |
4386 | mutex_unlock(&slab_mutex); |
4387 | m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL); |
4388 | if (!m->private) { |
4389 | /* Too bad, we are really out */ |
4390 | m->private = x; |
4391 | mutex_lock(&slab_mutex); |
4392 | return -ENOMEM; |
4393 | } |
4394 | *(unsigned long *)m->private = x[0] * 2; |
4395 | kfree(x); |
4396 | mutex_lock(&slab_mutex); |
4397 | /* Now make sure this entry will be retried */ |
4398 | m->count = m->size; |
4399 | return 0; |
4400 | } |
4401 | for (i = 0; i < x[1]; i++) { |
4402 | seq_printf(m, "%s: %lu ", name, x[2*i+3]); |
4403 | show_symbol(m, x[2*i+2]); |
4404 | seq_putc(m, '\n'); |
4405 | } |
4406 | |
4407 | return 0; |
4408 | } |
4409 | |
4410 | static const struct seq_operations slabstats_op = { |
4411 | .start = slab_start, |
4412 | .next = slab_next, |
4413 | .stop = slab_stop, |
4414 | .show = leaks_show, |
4415 | }; |
4416 | |
4417 | static int slabstats_open(struct inode *inode, struct file *file) |
4418 | { |
4419 | unsigned long *n; |
4420 | |
4421 | n = __seq_open_private(file, &slabstats_op, PAGE_SIZE); |
4422 | if (!n) |
4423 | return -ENOMEM; |
4424 | |
4425 | *n = PAGE_SIZE / (2 * sizeof(unsigned long)); |
4426 | |
4427 | return 0; |
4428 | } |
4429 | |
4430 | static const struct file_operations proc_slabstats_operations = { |
4431 | .open = slabstats_open, |
4432 | .read = seq_read, |
4433 | .llseek = seq_lseek, |
4434 | .release = seq_release_private, |
4435 | }; |
4436 | #endif |
4437 | |
4438 | static int __init slab_proc_init(void) |
4439 | { |
4440 | #ifdef CONFIG_DEBUG_SLAB_LEAK |
4441 | proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations); |
4442 | #endif |
4443 | return 0; |
4444 | } |
4445 | module_init(slab_proc_init); |
4446 | #endif |
4447 | |
4448 | #ifdef CONFIG_HARDENED_USERCOPY |
4449 | /* |
4450 | * Rejects objects that are incorrectly sized. |
4451 | * |
4452 | * Returns NULL if check passes, otherwise const char * to name of cache |
4453 | * to indicate an error. |
4454 | */ |
4455 | const char *__check_heap_object(const void *ptr, unsigned long n, |
4456 | struct page *page) |
4457 | { |
4458 | struct kmem_cache *cachep; |
4459 | unsigned int objnr; |
4460 | unsigned long offset; |
4461 | |
4462 | /* Find and validate object. */ |
4463 | cachep = page->slab_cache; |
4464 | objnr = obj_to_index(cachep, page, (void *)ptr); |
4465 | BUG_ON(objnr >= cachep->num); |
4466 | |
4467 | /* Find offset within object. */ |
4468 | offset = ptr - index_to_obj(cachep, page, objnr) - obj_offset(cachep); |
4469 | |
4470 | /* Allow address range falling entirely within object size. */ |
4471 | if (offset <= cachep->object_size && n <= cachep->object_size - offset) |
4472 | return NULL; |
4473 | |
4474 | return cachep->name; |
4475 | } |
4476 | #endif /* CONFIG_HARDENED_USERCOPY */ |
4477 | |
4478 | /** |
4479 | * ksize - get the actual amount of memory allocated for a given object |
4480 | * @objp: Pointer to the object |
4481 | * |
4482 | * kmalloc may internally round up allocations and return more memory |
4483 | * than requested. ksize() can be used to determine the actual amount of |
4484 | * memory allocated. The caller may use this additional memory, even though |
4485 | * a smaller amount of memory was initially specified with the kmalloc call. |
4486 | * The caller must guarantee that objp points to a valid object previously |
4487 | * allocated with either kmalloc() or kmem_cache_alloc(). The object |
4488 | * must not be freed during the duration of the call. |
4489 | */ |
4490 | size_t ksize(const void *objp) |
4491 | { |
4492 | size_t size; |
4493 | |
4494 | BUG_ON(!objp); |
4495 | if (unlikely(objp == ZERO_SIZE_PTR)) |
4496 | return 0; |
4497 | |
4498 | size = virt_to_cache(objp)->object_size; |
4499 | /* We assume that ksize callers could use the whole allocated area, |
4500 | * so we need to unpoison this area. |
4501 | */ |
4502 | kasan_unpoison_shadow(objp, size); |
4503 | |
4504 | return size; |
4505 | } |
4506 | EXPORT_SYMBOL(ksize); |
4507 |