xref: /openbmc/linux/mm/slab.c (revision a2cce7a9)
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  * true if a page was allocated from pfmemalloc reserves for network-based
173  * swap
174  */
175 static bool pfmemalloc_active __read_mostly;
176 
177 /*
178  * struct array_cache
179  *
180  * Purpose:
181  * - LIFO ordering, to hand out cache-warm objects from _alloc
182  * - reduce the number of linked list operations
183  * - reduce spinlock operations
184  *
185  * The limit is stored in the per-cpu structure to reduce the data cache
186  * footprint.
187  *
188  */
189 struct array_cache {
190 	unsigned int avail;
191 	unsigned int limit;
192 	unsigned int batchcount;
193 	unsigned int touched;
194 	void *entry[];	/*
195 			 * Must have this definition in here for the proper
196 			 * alignment of array_cache. Also simplifies accessing
197 			 * the entries.
198 			 *
199 			 * Entries should not be directly dereferenced as
200 			 * entries belonging to slabs marked pfmemalloc will
201 			 * have the lower bits set SLAB_OBJ_PFMEMALLOC
202 			 */
203 };
204 
205 struct alien_cache {
206 	spinlock_t lock;
207 	struct array_cache ac;
208 };
209 
210 #define SLAB_OBJ_PFMEMALLOC	1
211 static inline bool is_obj_pfmemalloc(void *objp)
212 {
213 	return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
214 }
215 
216 static inline void set_obj_pfmemalloc(void **objp)
217 {
218 	*objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
219 	return;
220 }
221 
222 static inline void clear_obj_pfmemalloc(void **objp)
223 {
224 	*objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
225 }
226 
227 /*
228  * bootstrap: The caches do not work without cpuarrays anymore, but the
229  * cpuarrays are allocated from the generic caches...
230  */
231 #define BOOT_CPUCACHE_ENTRIES	1
232 struct arraycache_init {
233 	struct array_cache cache;
234 	void *entries[BOOT_CPUCACHE_ENTRIES];
235 };
236 
237 /*
238  * Need this for bootstrapping a per node allocator.
239  */
240 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
241 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
242 #define	CACHE_CACHE 0
243 #define	SIZE_NODE (MAX_NUMNODES)
244 
245 static int drain_freelist(struct kmem_cache *cache,
246 			struct kmem_cache_node *n, int tofree);
247 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
248 			int node, struct list_head *list);
249 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
250 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
251 static void cache_reap(struct work_struct *unused);
252 
253 static int slab_early_init = 1;
254 
255 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
256 
257 static void kmem_cache_node_init(struct kmem_cache_node *parent)
258 {
259 	INIT_LIST_HEAD(&parent->slabs_full);
260 	INIT_LIST_HEAD(&parent->slabs_partial);
261 	INIT_LIST_HEAD(&parent->slabs_free);
262 	parent->shared = NULL;
263 	parent->alien = NULL;
264 	parent->colour_next = 0;
265 	spin_lock_init(&parent->list_lock);
266 	parent->free_objects = 0;
267 	parent->free_touched = 0;
268 }
269 
270 #define MAKE_LIST(cachep, listp, slab, nodeid)				\
271 	do {								\
272 		INIT_LIST_HEAD(listp);					\
273 		list_splice(&get_node(cachep, nodeid)->slab, listp);	\
274 	} while (0)
275 
276 #define	MAKE_ALL_LISTS(cachep, ptr, nodeid)				\
277 	do {								\
278 	MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid);	\
279 	MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
280 	MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid);	\
281 	} while (0)
282 
283 #define CFLGS_OFF_SLAB		(0x80000000UL)
284 #define	OFF_SLAB(x)	((x)->flags & CFLGS_OFF_SLAB)
285 
286 #define BATCHREFILL_LIMIT	16
287 /*
288  * Optimization question: fewer reaps means less probability for unnessary
289  * cpucache drain/refill cycles.
290  *
291  * OTOH the cpuarrays can contain lots of objects,
292  * which could lock up otherwise freeable slabs.
293  */
294 #define REAPTIMEOUT_AC		(2*HZ)
295 #define REAPTIMEOUT_NODE	(4*HZ)
296 
297 #if STATS
298 #define	STATS_INC_ACTIVE(x)	((x)->num_active++)
299 #define	STATS_DEC_ACTIVE(x)	((x)->num_active--)
300 #define	STATS_INC_ALLOCED(x)	((x)->num_allocations++)
301 #define	STATS_INC_GROWN(x)	((x)->grown++)
302 #define	STATS_ADD_REAPED(x,y)	((x)->reaped += (y))
303 #define	STATS_SET_HIGH(x)						\
304 	do {								\
305 		if ((x)->num_active > (x)->high_mark)			\
306 			(x)->high_mark = (x)->num_active;		\
307 	} while (0)
308 #define	STATS_INC_ERR(x)	((x)->errors++)
309 #define	STATS_INC_NODEALLOCS(x)	((x)->node_allocs++)
310 #define	STATS_INC_NODEFREES(x)	((x)->node_frees++)
311 #define STATS_INC_ACOVERFLOW(x)   ((x)->node_overflow++)
312 #define	STATS_SET_FREEABLE(x, i)					\
313 	do {								\
314 		if ((x)->max_freeable < i)				\
315 			(x)->max_freeable = i;				\
316 	} while (0)
317 #define STATS_INC_ALLOCHIT(x)	atomic_inc(&(x)->allochit)
318 #define STATS_INC_ALLOCMISS(x)	atomic_inc(&(x)->allocmiss)
319 #define STATS_INC_FREEHIT(x)	atomic_inc(&(x)->freehit)
320 #define STATS_INC_FREEMISS(x)	atomic_inc(&(x)->freemiss)
321 #else
322 #define	STATS_INC_ACTIVE(x)	do { } while (0)
323 #define	STATS_DEC_ACTIVE(x)	do { } while (0)
324 #define	STATS_INC_ALLOCED(x)	do { } while (0)
325 #define	STATS_INC_GROWN(x)	do { } while (0)
326 #define	STATS_ADD_REAPED(x,y)	do { (void)(y); } while (0)
327 #define	STATS_SET_HIGH(x)	do { } while (0)
328 #define	STATS_INC_ERR(x)	do { } while (0)
329 #define	STATS_INC_NODEALLOCS(x)	do { } while (0)
330 #define	STATS_INC_NODEFREES(x)	do { } while (0)
331 #define STATS_INC_ACOVERFLOW(x)   do { } while (0)
332 #define	STATS_SET_FREEABLE(x, i) do { } while (0)
333 #define STATS_INC_ALLOCHIT(x)	do { } while (0)
334 #define STATS_INC_ALLOCMISS(x)	do { } while (0)
335 #define STATS_INC_FREEHIT(x)	do { } while (0)
336 #define STATS_INC_FREEMISS(x)	do { } while (0)
337 #endif
338 
339 #if DEBUG
340 
341 /*
342  * memory layout of objects:
343  * 0		: objp
344  * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
345  * 		the end of an object is aligned with the end of the real
346  * 		allocation. Catches writes behind the end of the allocation.
347  * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
348  * 		redzone word.
349  * cachep->obj_offset: The real object.
350  * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
351  * cachep->size - 1* BYTES_PER_WORD: last caller address
352  *					[BYTES_PER_WORD long]
353  */
354 static int obj_offset(struct kmem_cache *cachep)
355 {
356 	return cachep->obj_offset;
357 }
358 
359 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
360 {
361 	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
362 	return (unsigned long long*) (objp + obj_offset(cachep) -
363 				      sizeof(unsigned long long));
364 }
365 
366 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
367 {
368 	BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
369 	if (cachep->flags & SLAB_STORE_USER)
370 		return (unsigned long long *)(objp + cachep->size -
371 					      sizeof(unsigned long long) -
372 					      REDZONE_ALIGN);
373 	return (unsigned long long *) (objp + cachep->size -
374 				       sizeof(unsigned long long));
375 }
376 
377 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
378 {
379 	BUG_ON(!(cachep->flags & SLAB_STORE_USER));
380 	return (void **)(objp + cachep->size - BYTES_PER_WORD);
381 }
382 
383 #else
384 
385 #define obj_offset(x)			0
386 #define dbg_redzone1(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
387 #define dbg_redzone2(cachep, objp)	({BUG(); (unsigned long long *)NULL;})
388 #define dbg_userword(cachep, objp)	({BUG(); (void **)NULL;})
389 
390 #endif
391 
392 #define OBJECT_FREE (0)
393 #define OBJECT_ACTIVE (1)
394 
395 #ifdef CONFIG_DEBUG_SLAB_LEAK
396 
397 static void set_obj_status(struct page *page, int idx, int val)
398 {
399 	int freelist_size;
400 	char *status;
401 	struct kmem_cache *cachep = page->slab_cache;
402 
403 	freelist_size = cachep->num * sizeof(freelist_idx_t);
404 	status = (char *)page->freelist + freelist_size;
405 	status[idx] = val;
406 }
407 
408 static inline unsigned int get_obj_status(struct page *page, int idx)
409 {
410 	int freelist_size;
411 	char *status;
412 	struct kmem_cache *cachep = page->slab_cache;
413 
414 	freelist_size = cachep->num * sizeof(freelist_idx_t);
415 	status = (char *)page->freelist + freelist_size;
416 
417 	return status[idx];
418 }
419 
420 #else
421 static inline void set_obj_status(struct page *page, int idx, int val) {}
422 
423 #endif
424 
425 /*
426  * Do not go above this order unless 0 objects fit into the slab or
427  * overridden on the command line.
428  */
429 #define	SLAB_MAX_ORDER_HI	1
430 #define	SLAB_MAX_ORDER_LO	0
431 static int slab_max_order = SLAB_MAX_ORDER_LO;
432 static bool slab_max_order_set __initdata;
433 
434 static inline struct kmem_cache *virt_to_cache(const void *obj)
435 {
436 	struct page *page = virt_to_head_page(obj);
437 	return page->slab_cache;
438 }
439 
440 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
441 				 unsigned int idx)
442 {
443 	return page->s_mem + cache->size * idx;
444 }
445 
446 /*
447  * We want to avoid an expensive divide : (offset / cache->size)
448  *   Using the fact that size is a constant for a particular cache,
449  *   we can replace (offset / cache->size) by
450  *   reciprocal_divide(offset, cache->reciprocal_buffer_size)
451  */
452 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
453 					const struct page *page, void *obj)
454 {
455 	u32 offset = (obj - page->s_mem);
456 	return reciprocal_divide(offset, cache->reciprocal_buffer_size);
457 }
458 
459 /* internal cache of cache description objs */
460 static struct kmem_cache kmem_cache_boot = {
461 	.batchcount = 1,
462 	.limit = BOOT_CPUCACHE_ENTRIES,
463 	.shared = 1,
464 	.size = sizeof(struct kmem_cache),
465 	.name = "kmem_cache",
466 };
467 
468 #define BAD_ALIEN_MAGIC 0x01020304ul
469 
470 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
471 
472 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
473 {
474 	return this_cpu_ptr(cachep->cpu_cache);
475 }
476 
477 static size_t calculate_freelist_size(int nr_objs, size_t align)
478 {
479 	size_t freelist_size;
480 
481 	freelist_size = nr_objs * sizeof(freelist_idx_t);
482 	if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
483 		freelist_size += nr_objs * sizeof(char);
484 
485 	if (align)
486 		freelist_size = ALIGN(freelist_size, align);
487 
488 	return freelist_size;
489 }
490 
491 static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
492 				size_t idx_size, size_t align)
493 {
494 	int nr_objs;
495 	size_t remained_size;
496 	size_t freelist_size;
497 	int extra_space = 0;
498 
499 	if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
500 		extra_space = sizeof(char);
501 	/*
502 	 * Ignore padding for the initial guess. The padding
503 	 * is at most @align-1 bytes, and @buffer_size is at
504 	 * least @align. In the worst case, this result will
505 	 * be one greater than the number of objects that fit
506 	 * into the memory allocation when taking the padding
507 	 * into account.
508 	 */
509 	nr_objs = slab_size / (buffer_size + idx_size + extra_space);
510 
511 	/*
512 	 * This calculated number will be either the right
513 	 * amount, or one greater than what we want.
514 	 */
515 	remained_size = slab_size - nr_objs * buffer_size;
516 	freelist_size = calculate_freelist_size(nr_objs, align);
517 	if (remained_size < freelist_size)
518 		nr_objs--;
519 
520 	return nr_objs;
521 }
522 
523 /*
524  * Calculate the number of objects and left-over bytes for a given buffer size.
525  */
526 static void cache_estimate(unsigned long gfporder, size_t buffer_size,
527 			   size_t align, int flags, size_t *left_over,
528 			   unsigned int *num)
529 {
530 	int nr_objs;
531 	size_t mgmt_size;
532 	size_t slab_size = PAGE_SIZE << gfporder;
533 
534 	/*
535 	 * The slab management structure can be either off the slab or
536 	 * on it. For the latter case, the memory allocated for a
537 	 * slab is used for:
538 	 *
539 	 * - One unsigned int for each object
540 	 * - Padding to respect alignment of @align
541 	 * - @buffer_size bytes for each object
542 	 *
543 	 * If the slab management structure is off the slab, then the
544 	 * alignment will already be calculated into the size. Because
545 	 * the slabs are all pages aligned, the objects will be at the
546 	 * correct alignment when allocated.
547 	 */
548 	if (flags & CFLGS_OFF_SLAB) {
549 		mgmt_size = 0;
550 		nr_objs = slab_size / buffer_size;
551 
552 	} else {
553 		nr_objs = calculate_nr_objs(slab_size, buffer_size,
554 					sizeof(freelist_idx_t), align);
555 		mgmt_size = calculate_freelist_size(nr_objs, align);
556 	}
557 	*num = nr_objs;
558 	*left_over = slab_size - nr_objs*buffer_size - mgmt_size;
559 }
560 
561 #if DEBUG
562 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
563 
564 static void __slab_error(const char *function, struct kmem_cache *cachep,
565 			char *msg)
566 {
567 	printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
568 	       function, cachep->name, msg);
569 	dump_stack();
570 	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
571 }
572 #endif
573 
574 /*
575  * By default on NUMA we use alien caches to stage the freeing of
576  * objects allocated from other nodes. This causes massive memory
577  * inefficiencies when using fake NUMA setup to split memory into a
578  * large number of small nodes, so it can be disabled on the command
579  * line
580   */
581 
582 static int use_alien_caches __read_mostly = 1;
583 static int __init noaliencache_setup(char *s)
584 {
585 	use_alien_caches = 0;
586 	return 1;
587 }
588 __setup("noaliencache", noaliencache_setup);
589 
590 static int __init slab_max_order_setup(char *str)
591 {
592 	get_option(&str, &slab_max_order);
593 	slab_max_order = slab_max_order < 0 ? 0 :
594 				min(slab_max_order, MAX_ORDER - 1);
595 	slab_max_order_set = true;
596 
597 	return 1;
598 }
599 __setup("slab_max_order=", slab_max_order_setup);
600 
601 #ifdef CONFIG_NUMA
602 /*
603  * Special reaping functions for NUMA systems called from cache_reap().
604  * These take care of doing round robin flushing of alien caches (containing
605  * objects freed on different nodes from which they were allocated) and the
606  * flushing of remote pcps by calling drain_node_pages.
607  */
608 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
609 
610 static void init_reap_node(int cpu)
611 {
612 	int node;
613 
614 	node = next_node(cpu_to_mem(cpu), node_online_map);
615 	if (node == MAX_NUMNODES)
616 		node = first_node(node_online_map);
617 
618 	per_cpu(slab_reap_node, cpu) = node;
619 }
620 
621 static void next_reap_node(void)
622 {
623 	int node = __this_cpu_read(slab_reap_node);
624 
625 	node = next_node(node, node_online_map);
626 	if (unlikely(node >= MAX_NUMNODES))
627 		node = first_node(node_online_map);
628 	__this_cpu_write(slab_reap_node, node);
629 }
630 
631 #else
632 #define init_reap_node(cpu) do { } while (0)
633 #define next_reap_node(void) do { } while (0)
634 #endif
635 
636 /*
637  * Initiate the reap timer running on the target CPU.  We run at around 1 to 2Hz
638  * via the workqueue/eventd.
639  * Add the CPU number into the expiration time to minimize the possibility of
640  * the CPUs getting into lockstep and contending for the global cache chain
641  * lock.
642  */
643 static void start_cpu_timer(int cpu)
644 {
645 	struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
646 
647 	/*
648 	 * When this gets called from do_initcalls via cpucache_init(),
649 	 * init_workqueues() has already run, so keventd will be setup
650 	 * at that time.
651 	 */
652 	if (keventd_up() && reap_work->work.func == NULL) {
653 		init_reap_node(cpu);
654 		INIT_DEFERRABLE_WORK(reap_work, cache_reap);
655 		schedule_delayed_work_on(cpu, reap_work,
656 					__round_jiffies_relative(HZ, cpu));
657 	}
658 }
659 
660 static void init_arraycache(struct array_cache *ac, int limit, int batch)
661 {
662 	/*
663 	 * The array_cache structures contain pointers to free object.
664 	 * However, when such objects are allocated or transferred to another
665 	 * cache the pointers are not cleared and they could be counted as
666 	 * valid references during a kmemleak scan. Therefore, kmemleak must
667 	 * not scan such objects.
668 	 */
669 	kmemleak_no_scan(ac);
670 	if (ac) {
671 		ac->avail = 0;
672 		ac->limit = limit;
673 		ac->batchcount = batch;
674 		ac->touched = 0;
675 	}
676 }
677 
678 static struct array_cache *alloc_arraycache(int node, int entries,
679 					    int batchcount, gfp_t gfp)
680 {
681 	size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
682 	struct array_cache *ac = NULL;
683 
684 	ac = kmalloc_node(memsize, gfp, node);
685 	init_arraycache(ac, entries, batchcount);
686 	return ac;
687 }
688 
689 static inline bool is_slab_pfmemalloc(struct page *page)
690 {
691 	return PageSlabPfmemalloc(page);
692 }
693 
694 /* Clears pfmemalloc_active if no slabs have pfmalloc set */
695 static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
696 						struct array_cache *ac)
697 {
698 	struct kmem_cache_node *n = get_node(cachep, numa_mem_id());
699 	struct page *page;
700 	unsigned long flags;
701 
702 	if (!pfmemalloc_active)
703 		return;
704 
705 	spin_lock_irqsave(&n->list_lock, flags);
706 	list_for_each_entry(page, &n->slabs_full, lru)
707 		if (is_slab_pfmemalloc(page))
708 			goto out;
709 
710 	list_for_each_entry(page, &n->slabs_partial, lru)
711 		if (is_slab_pfmemalloc(page))
712 			goto out;
713 
714 	list_for_each_entry(page, &n->slabs_free, lru)
715 		if (is_slab_pfmemalloc(page))
716 			goto out;
717 
718 	pfmemalloc_active = false;
719 out:
720 	spin_unlock_irqrestore(&n->list_lock, flags);
721 }
722 
723 static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
724 						gfp_t flags, bool force_refill)
725 {
726 	int i;
727 	void *objp = ac->entry[--ac->avail];
728 
729 	/* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
730 	if (unlikely(is_obj_pfmemalloc(objp))) {
731 		struct kmem_cache_node *n;
732 
733 		if (gfp_pfmemalloc_allowed(flags)) {
734 			clear_obj_pfmemalloc(&objp);
735 			return objp;
736 		}
737 
738 		/* The caller cannot use PFMEMALLOC objects, find another one */
739 		for (i = 0; i < ac->avail; i++) {
740 			/* If a !PFMEMALLOC object is found, swap them */
741 			if (!is_obj_pfmemalloc(ac->entry[i])) {
742 				objp = ac->entry[i];
743 				ac->entry[i] = ac->entry[ac->avail];
744 				ac->entry[ac->avail] = objp;
745 				return objp;
746 			}
747 		}
748 
749 		/*
750 		 * If there are empty slabs on the slabs_free list and we are
751 		 * being forced to refill the cache, mark this one !pfmemalloc.
752 		 */
753 		n = get_node(cachep, numa_mem_id());
754 		if (!list_empty(&n->slabs_free) && force_refill) {
755 			struct page *page = virt_to_head_page(objp);
756 			ClearPageSlabPfmemalloc(page);
757 			clear_obj_pfmemalloc(&objp);
758 			recheck_pfmemalloc_active(cachep, ac);
759 			return objp;
760 		}
761 
762 		/* No !PFMEMALLOC objects available */
763 		ac->avail++;
764 		objp = NULL;
765 	}
766 
767 	return objp;
768 }
769 
770 static inline void *ac_get_obj(struct kmem_cache *cachep,
771 			struct array_cache *ac, gfp_t flags, bool force_refill)
772 {
773 	void *objp;
774 
775 	if (unlikely(sk_memalloc_socks()))
776 		objp = __ac_get_obj(cachep, ac, flags, force_refill);
777 	else
778 		objp = ac->entry[--ac->avail];
779 
780 	return objp;
781 }
782 
783 static noinline void *__ac_put_obj(struct kmem_cache *cachep,
784 			struct array_cache *ac, void *objp)
785 {
786 	if (unlikely(pfmemalloc_active)) {
787 		/* Some pfmemalloc slabs exist, check if this is one */
788 		struct page *page = virt_to_head_page(objp);
789 		if (PageSlabPfmemalloc(page))
790 			set_obj_pfmemalloc(&objp);
791 	}
792 
793 	return objp;
794 }
795 
796 static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
797 								void *objp)
798 {
799 	if (unlikely(sk_memalloc_socks()))
800 		objp = __ac_put_obj(cachep, ac, objp);
801 
802 	ac->entry[ac->avail++] = objp;
803 }
804 
805 /*
806  * Transfer objects in one arraycache to another.
807  * Locking must be handled by the caller.
808  *
809  * Return the number of entries transferred.
810  */
811 static int transfer_objects(struct array_cache *to,
812 		struct array_cache *from, unsigned int max)
813 {
814 	/* Figure out how many entries to transfer */
815 	int nr = min3(from->avail, max, to->limit - to->avail);
816 
817 	if (!nr)
818 		return 0;
819 
820 	memcpy(to->entry + to->avail, from->entry + from->avail -nr,
821 			sizeof(void *) *nr);
822 
823 	from->avail -= nr;
824 	to->avail += nr;
825 	return nr;
826 }
827 
828 #ifndef CONFIG_NUMA
829 
830 #define drain_alien_cache(cachep, alien) do { } while (0)
831 #define reap_alien(cachep, n) do { } while (0)
832 
833 static inline struct alien_cache **alloc_alien_cache(int node,
834 						int limit, gfp_t gfp)
835 {
836 	return (struct alien_cache **)BAD_ALIEN_MAGIC;
837 }
838 
839 static inline void free_alien_cache(struct alien_cache **ac_ptr)
840 {
841 }
842 
843 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
844 {
845 	return 0;
846 }
847 
848 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
849 		gfp_t flags)
850 {
851 	return NULL;
852 }
853 
854 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
855 		 gfp_t flags, int nodeid)
856 {
857 	return NULL;
858 }
859 
860 static inline gfp_t gfp_exact_node(gfp_t flags)
861 {
862 	return flags;
863 }
864 
865 #else	/* CONFIG_NUMA */
866 
867 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
868 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
869 
870 static struct alien_cache *__alloc_alien_cache(int node, int entries,
871 						int batch, gfp_t gfp)
872 {
873 	size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
874 	struct alien_cache *alc = NULL;
875 
876 	alc = kmalloc_node(memsize, gfp, node);
877 	init_arraycache(&alc->ac, entries, batch);
878 	spin_lock_init(&alc->lock);
879 	return alc;
880 }
881 
882 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
883 {
884 	struct alien_cache **alc_ptr;
885 	size_t memsize = sizeof(void *) * nr_node_ids;
886 	int i;
887 
888 	if (limit > 1)
889 		limit = 12;
890 	alc_ptr = kzalloc_node(memsize, gfp, node);
891 	if (!alc_ptr)
892 		return NULL;
893 
894 	for_each_node(i) {
895 		if (i == node || !node_online(i))
896 			continue;
897 		alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
898 		if (!alc_ptr[i]) {
899 			for (i--; i >= 0; i--)
900 				kfree(alc_ptr[i]);
901 			kfree(alc_ptr);
902 			return NULL;
903 		}
904 	}
905 	return alc_ptr;
906 }
907 
908 static void free_alien_cache(struct alien_cache **alc_ptr)
909 {
910 	int i;
911 
912 	if (!alc_ptr)
913 		return;
914 	for_each_node(i)
915 	    kfree(alc_ptr[i]);
916 	kfree(alc_ptr);
917 }
918 
919 static void __drain_alien_cache(struct kmem_cache *cachep,
920 				struct array_cache *ac, int node,
921 				struct list_head *list)
922 {
923 	struct kmem_cache_node *n = get_node(cachep, node);
924 
925 	if (ac->avail) {
926 		spin_lock(&n->list_lock);
927 		/*
928 		 * Stuff objects into the remote nodes shared array first.
929 		 * That way we could avoid the overhead of putting the objects
930 		 * into the free lists and getting them back later.
931 		 */
932 		if (n->shared)
933 			transfer_objects(n->shared, ac, ac->limit);
934 
935 		free_block(cachep, ac->entry, ac->avail, node, list);
936 		ac->avail = 0;
937 		spin_unlock(&n->list_lock);
938 	}
939 }
940 
941 /*
942  * Called from cache_reap() to regularly drain alien caches round robin.
943  */
944 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
945 {
946 	int node = __this_cpu_read(slab_reap_node);
947 
948 	if (n->alien) {
949 		struct alien_cache *alc = n->alien[node];
950 		struct array_cache *ac;
951 
952 		if (alc) {
953 			ac = &alc->ac;
954 			if (ac->avail && spin_trylock_irq(&alc->lock)) {
955 				LIST_HEAD(list);
956 
957 				__drain_alien_cache(cachep, ac, node, &list);
958 				spin_unlock_irq(&alc->lock);
959 				slabs_destroy(cachep, &list);
960 			}
961 		}
962 	}
963 }
964 
965 static void drain_alien_cache(struct kmem_cache *cachep,
966 				struct alien_cache **alien)
967 {
968 	int i = 0;
969 	struct alien_cache *alc;
970 	struct array_cache *ac;
971 	unsigned long flags;
972 
973 	for_each_online_node(i) {
974 		alc = alien[i];
975 		if (alc) {
976 			LIST_HEAD(list);
977 
978 			ac = &alc->ac;
979 			spin_lock_irqsave(&alc->lock, flags);
980 			__drain_alien_cache(cachep, ac, i, &list);
981 			spin_unlock_irqrestore(&alc->lock, flags);
982 			slabs_destroy(cachep, &list);
983 		}
984 	}
985 }
986 
987 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
988 				int node, int page_node)
989 {
990 	struct kmem_cache_node *n;
991 	struct alien_cache *alien = NULL;
992 	struct array_cache *ac;
993 	LIST_HEAD(list);
994 
995 	n = get_node(cachep, node);
996 	STATS_INC_NODEFREES(cachep);
997 	if (n->alien && n->alien[page_node]) {
998 		alien = n->alien[page_node];
999 		ac = &alien->ac;
1000 		spin_lock(&alien->lock);
1001 		if (unlikely(ac->avail == ac->limit)) {
1002 			STATS_INC_ACOVERFLOW(cachep);
1003 			__drain_alien_cache(cachep, ac, page_node, &list);
1004 		}
1005 		ac_put_obj(cachep, ac, objp);
1006 		spin_unlock(&alien->lock);
1007 		slabs_destroy(cachep, &list);
1008 	} else {
1009 		n = get_node(cachep, page_node);
1010 		spin_lock(&n->list_lock);
1011 		free_block(cachep, &objp, 1, page_node, &list);
1012 		spin_unlock(&n->list_lock);
1013 		slabs_destroy(cachep, &list);
1014 	}
1015 	return 1;
1016 }
1017 
1018 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1019 {
1020 	int page_node = page_to_nid(virt_to_page(objp));
1021 	int node = numa_mem_id();
1022 	/*
1023 	 * Make sure we are not freeing a object from another node to the array
1024 	 * cache on this cpu.
1025 	 */
1026 	if (likely(node == page_node))
1027 		return 0;
1028 
1029 	return __cache_free_alien(cachep, objp, node, page_node);
1030 }
1031 
1032 /*
1033  * Construct gfp mask to allocate from a specific node but do not invoke reclaim
1034  * or warn about failures.
1035  */
1036 static inline gfp_t gfp_exact_node(gfp_t flags)
1037 {
1038 	return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~__GFP_WAIT;
1039 }
1040 #endif
1041 
1042 /*
1043  * Allocates and initializes node for a node on each slab cache, used for
1044  * either memory or cpu hotplug.  If memory is being hot-added, the kmem_cache_node
1045  * will be allocated off-node since memory is not yet online for the new node.
1046  * When hotplugging memory or a cpu, existing node are not replaced if
1047  * already in use.
1048  *
1049  * Must hold slab_mutex.
1050  */
1051 static int init_cache_node_node(int node)
1052 {
1053 	struct kmem_cache *cachep;
1054 	struct kmem_cache_node *n;
1055 	const size_t memsize = sizeof(struct kmem_cache_node);
1056 
1057 	list_for_each_entry(cachep, &slab_caches, list) {
1058 		/*
1059 		 * Set up the kmem_cache_node for cpu before we can
1060 		 * begin anything. Make sure some other cpu on this
1061 		 * node has not already allocated this
1062 		 */
1063 		n = get_node(cachep, node);
1064 		if (!n) {
1065 			n = kmalloc_node(memsize, GFP_KERNEL, node);
1066 			if (!n)
1067 				return -ENOMEM;
1068 			kmem_cache_node_init(n);
1069 			n->next_reap = jiffies + REAPTIMEOUT_NODE +
1070 			    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1071 
1072 			/*
1073 			 * The kmem_cache_nodes don't come and go as CPUs
1074 			 * come and go.  slab_mutex is sufficient
1075 			 * protection here.
1076 			 */
1077 			cachep->node[node] = n;
1078 		}
1079 
1080 		spin_lock_irq(&n->list_lock);
1081 		n->free_limit =
1082 			(1 + nr_cpus_node(node)) *
1083 			cachep->batchcount + cachep->num;
1084 		spin_unlock_irq(&n->list_lock);
1085 	}
1086 	return 0;
1087 }
1088 
1089 static inline int slabs_tofree(struct kmem_cache *cachep,
1090 						struct kmem_cache_node *n)
1091 {
1092 	return (n->free_objects + cachep->num - 1) / cachep->num;
1093 }
1094 
1095 static void cpuup_canceled(long cpu)
1096 {
1097 	struct kmem_cache *cachep;
1098 	struct kmem_cache_node *n = NULL;
1099 	int node = cpu_to_mem(cpu);
1100 	const struct cpumask *mask = cpumask_of_node(node);
1101 
1102 	list_for_each_entry(cachep, &slab_caches, list) {
1103 		struct array_cache *nc;
1104 		struct array_cache *shared;
1105 		struct alien_cache **alien;
1106 		LIST_HEAD(list);
1107 
1108 		n = get_node(cachep, node);
1109 		if (!n)
1110 			continue;
1111 
1112 		spin_lock_irq(&n->list_lock);
1113 
1114 		/* Free limit for this kmem_cache_node */
1115 		n->free_limit -= cachep->batchcount;
1116 
1117 		/* cpu is dead; no one can alloc from it. */
1118 		nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1119 		if (nc) {
1120 			free_block(cachep, nc->entry, nc->avail, node, &list);
1121 			nc->avail = 0;
1122 		}
1123 
1124 		if (!cpumask_empty(mask)) {
1125 			spin_unlock_irq(&n->list_lock);
1126 			goto free_slab;
1127 		}
1128 
1129 		shared = n->shared;
1130 		if (shared) {
1131 			free_block(cachep, shared->entry,
1132 				   shared->avail, node, &list);
1133 			n->shared = NULL;
1134 		}
1135 
1136 		alien = n->alien;
1137 		n->alien = NULL;
1138 
1139 		spin_unlock_irq(&n->list_lock);
1140 
1141 		kfree(shared);
1142 		if (alien) {
1143 			drain_alien_cache(cachep, alien);
1144 			free_alien_cache(alien);
1145 		}
1146 
1147 free_slab:
1148 		slabs_destroy(cachep, &list);
1149 	}
1150 	/*
1151 	 * In the previous loop, all the objects were freed to
1152 	 * the respective cache's slabs,  now we can go ahead and
1153 	 * shrink each nodelist to its limit.
1154 	 */
1155 	list_for_each_entry(cachep, &slab_caches, list) {
1156 		n = get_node(cachep, node);
1157 		if (!n)
1158 			continue;
1159 		drain_freelist(cachep, n, slabs_tofree(cachep, n));
1160 	}
1161 }
1162 
1163 static int cpuup_prepare(long cpu)
1164 {
1165 	struct kmem_cache *cachep;
1166 	struct kmem_cache_node *n = NULL;
1167 	int node = cpu_to_mem(cpu);
1168 	int err;
1169 
1170 	/*
1171 	 * We need to do this right in the beginning since
1172 	 * alloc_arraycache's are going to use this list.
1173 	 * kmalloc_node allows us to add the slab to the right
1174 	 * kmem_cache_node and not this cpu's kmem_cache_node
1175 	 */
1176 	err = init_cache_node_node(node);
1177 	if (err < 0)
1178 		goto bad;
1179 
1180 	/*
1181 	 * Now we can go ahead with allocating the shared arrays and
1182 	 * array caches
1183 	 */
1184 	list_for_each_entry(cachep, &slab_caches, list) {
1185 		struct array_cache *shared = NULL;
1186 		struct alien_cache **alien = NULL;
1187 
1188 		if (cachep->shared) {
1189 			shared = alloc_arraycache(node,
1190 				cachep->shared * cachep->batchcount,
1191 				0xbaadf00d, GFP_KERNEL);
1192 			if (!shared)
1193 				goto bad;
1194 		}
1195 		if (use_alien_caches) {
1196 			alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1197 			if (!alien) {
1198 				kfree(shared);
1199 				goto bad;
1200 			}
1201 		}
1202 		n = get_node(cachep, node);
1203 		BUG_ON(!n);
1204 
1205 		spin_lock_irq(&n->list_lock);
1206 		if (!n->shared) {
1207 			/*
1208 			 * We are serialised from CPU_DEAD or
1209 			 * CPU_UP_CANCELLED by the cpucontrol lock
1210 			 */
1211 			n->shared = shared;
1212 			shared = NULL;
1213 		}
1214 #ifdef CONFIG_NUMA
1215 		if (!n->alien) {
1216 			n->alien = alien;
1217 			alien = NULL;
1218 		}
1219 #endif
1220 		spin_unlock_irq(&n->list_lock);
1221 		kfree(shared);
1222 		free_alien_cache(alien);
1223 	}
1224 
1225 	return 0;
1226 bad:
1227 	cpuup_canceled(cpu);
1228 	return -ENOMEM;
1229 }
1230 
1231 static int cpuup_callback(struct notifier_block *nfb,
1232 				    unsigned long action, void *hcpu)
1233 {
1234 	long cpu = (long)hcpu;
1235 	int err = 0;
1236 
1237 	switch (action) {
1238 	case CPU_UP_PREPARE:
1239 	case CPU_UP_PREPARE_FROZEN:
1240 		mutex_lock(&slab_mutex);
1241 		err = cpuup_prepare(cpu);
1242 		mutex_unlock(&slab_mutex);
1243 		break;
1244 	case CPU_ONLINE:
1245 	case CPU_ONLINE_FROZEN:
1246 		start_cpu_timer(cpu);
1247 		break;
1248 #ifdef CONFIG_HOTPLUG_CPU
1249   	case CPU_DOWN_PREPARE:
1250   	case CPU_DOWN_PREPARE_FROZEN:
1251 		/*
1252 		 * Shutdown cache reaper. Note that the slab_mutex is
1253 		 * held so that if cache_reap() is invoked it cannot do
1254 		 * anything expensive but will only modify reap_work
1255 		 * and reschedule the timer.
1256 		*/
1257 		cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1258 		/* Now the cache_reaper is guaranteed to be not running. */
1259 		per_cpu(slab_reap_work, cpu).work.func = NULL;
1260   		break;
1261   	case CPU_DOWN_FAILED:
1262   	case CPU_DOWN_FAILED_FROZEN:
1263 		start_cpu_timer(cpu);
1264   		break;
1265 	case CPU_DEAD:
1266 	case CPU_DEAD_FROZEN:
1267 		/*
1268 		 * Even if all the cpus of a node are down, we don't free the
1269 		 * kmem_cache_node of any cache. This to avoid a race between
1270 		 * cpu_down, and a kmalloc allocation from another cpu for
1271 		 * memory from the node of the cpu going down.  The node
1272 		 * structure is usually allocated from kmem_cache_create() and
1273 		 * gets destroyed at kmem_cache_destroy().
1274 		 */
1275 		/* fall through */
1276 #endif
1277 	case CPU_UP_CANCELED:
1278 	case CPU_UP_CANCELED_FROZEN:
1279 		mutex_lock(&slab_mutex);
1280 		cpuup_canceled(cpu);
1281 		mutex_unlock(&slab_mutex);
1282 		break;
1283 	}
1284 	return notifier_from_errno(err);
1285 }
1286 
1287 static struct notifier_block cpucache_notifier = {
1288 	&cpuup_callback, NULL, 0
1289 };
1290 
1291 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1292 /*
1293  * Drains freelist for a node on each slab cache, used for memory hot-remove.
1294  * Returns -EBUSY if all objects cannot be drained so that the node is not
1295  * removed.
1296  *
1297  * Must hold slab_mutex.
1298  */
1299 static int __meminit drain_cache_node_node(int node)
1300 {
1301 	struct kmem_cache *cachep;
1302 	int ret = 0;
1303 
1304 	list_for_each_entry(cachep, &slab_caches, list) {
1305 		struct kmem_cache_node *n;
1306 
1307 		n = get_node(cachep, node);
1308 		if (!n)
1309 			continue;
1310 
1311 		drain_freelist(cachep, n, slabs_tofree(cachep, n));
1312 
1313 		if (!list_empty(&n->slabs_full) ||
1314 		    !list_empty(&n->slabs_partial)) {
1315 			ret = -EBUSY;
1316 			break;
1317 		}
1318 	}
1319 	return ret;
1320 }
1321 
1322 static int __meminit slab_memory_callback(struct notifier_block *self,
1323 					unsigned long action, void *arg)
1324 {
1325 	struct memory_notify *mnb = arg;
1326 	int ret = 0;
1327 	int nid;
1328 
1329 	nid = mnb->status_change_nid;
1330 	if (nid < 0)
1331 		goto out;
1332 
1333 	switch (action) {
1334 	case MEM_GOING_ONLINE:
1335 		mutex_lock(&slab_mutex);
1336 		ret = init_cache_node_node(nid);
1337 		mutex_unlock(&slab_mutex);
1338 		break;
1339 	case MEM_GOING_OFFLINE:
1340 		mutex_lock(&slab_mutex);
1341 		ret = drain_cache_node_node(nid);
1342 		mutex_unlock(&slab_mutex);
1343 		break;
1344 	case MEM_ONLINE:
1345 	case MEM_OFFLINE:
1346 	case MEM_CANCEL_ONLINE:
1347 	case MEM_CANCEL_OFFLINE:
1348 		break;
1349 	}
1350 out:
1351 	return notifier_from_errno(ret);
1352 }
1353 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1354 
1355 /*
1356  * swap the static kmem_cache_node with kmalloced memory
1357  */
1358 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1359 				int nodeid)
1360 {
1361 	struct kmem_cache_node *ptr;
1362 
1363 	ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1364 	BUG_ON(!ptr);
1365 
1366 	memcpy(ptr, list, sizeof(struct kmem_cache_node));
1367 	/*
1368 	 * Do not assume that spinlocks can be initialized via memcpy:
1369 	 */
1370 	spin_lock_init(&ptr->list_lock);
1371 
1372 	MAKE_ALL_LISTS(cachep, ptr, nodeid);
1373 	cachep->node[nodeid] = ptr;
1374 }
1375 
1376 /*
1377  * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1378  * size of kmem_cache_node.
1379  */
1380 static void __init set_up_node(struct kmem_cache *cachep, int index)
1381 {
1382 	int node;
1383 
1384 	for_each_online_node(node) {
1385 		cachep->node[node] = &init_kmem_cache_node[index + node];
1386 		cachep->node[node]->next_reap = jiffies +
1387 		    REAPTIMEOUT_NODE +
1388 		    ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1389 	}
1390 }
1391 
1392 /*
1393  * Initialisation.  Called after the page allocator have been initialised and
1394  * before smp_init().
1395  */
1396 void __init kmem_cache_init(void)
1397 {
1398 	int i;
1399 
1400 	BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1401 					sizeof(struct rcu_head));
1402 	kmem_cache = &kmem_cache_boot;
1403 
1404 	if (num_possible_nodes() == 1)
1405 		use_alien_caches = 0;
1406 
1407 	for (i = 0; i < NUM_INIT_LISTS; i++)
1408 		kmem_cache_node_init(&init_kmem_cache_node[i]);
1409 
1410 	/*
1411 	 * Fragmentation resistance on low memory - only use bigger
1412 	 * page orders on machines with more than 32MB of memory if
1413 	 * not overridden on the command line.
1414 	 */
1415 	if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1416 		slab_max_order = SLAB_MAX_ORDER_HI;
1417 
1418 	/* Bootstrap is tricky, because several objects are allocated
1419 	 * from caches that do not exist yet:
1420 	 * 1) initialize the kmem_cache cache: it contains the struct
1421 	 *    kmem_cache structures of all caches, except kmem_cache itself:
1422 	 *    kmem_cache is statically allocated.
1423 	 *    Initially an __init data area is used for the head array and the
1424 	 *    kmem_cache_node structures, it's replaced with a kmalloc allocated
1425 	 *    array at the end of the bootstrap.
1426 	 * 2) Create the first kmalloc cache.
1427 	 *    The struct kmem_cache for the new cache is allocated normally.
1428 	 *    An __init data area is used for the head array.
1429 	 * 3) Create the remaining kmalloc caches, with minimally sized
1430 	 *    head arrays.
1431 	 * 4) Replace the __init data head arrays for kmem_cache and the first
1432 	 *    kmalloc cache with kmalloc allocated arrays.
1433 	 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1434 	 *    the other cache's with kmalloc allocated memory.
1435 	 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1436 	 */
1437 
1438 	/* 1) create the kmem_cache */
1439 
1440 	/*
1441 	 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1442 	 */
1443 	create_boot_cache(kmem_cache, "kmem_cache",
1444 		offsetof(struct kmem_cache, node) +
1445 				  nr_node_ids * sizeof(struct kmem_cache_node *),
1446 				  SLAB_HWCACHE_ALIGN);
1447 	list_add(&kmem_cache->list, &slab_caches);
1448 	slab_state = PARTIAL;
1449 
1450 	/*
1451 	 * Initialize the caches that provide memory for the  kmem_cache_node
1452 	 * structures first.  Without this, further allocations will bug.
1453 	 */
1454 	kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1455 				kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1456 	slab_state = PARTIAL_NODE;
1457 	setup_kmalloc_cache_index_table();
1458 
1459 	slab_early_init = 0;
1460 
1461 	/* 5) Replace the bootstrap kmem_cache_node */
1462 	{
1463 		int nid;
1464 
1465 		for_each_online_node(nid) {
1466 			init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1467 
1468 			init_list(kmalloc_caches[INDEX_NODE],
1469 					  &init_kmem_cache_node[SIZE_NODE + nid], nid);
1470 		}
1471 	}
1472 
1473 	create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1474 }
1475 
1476 void __init kmem_cache_init_late(void)
1477 {
1478 	struct kmem_cache *cachep;
1479 
1480 	slab_state = UP;
1481 
1482 	/* 6) resize the head arrays to their final sizes */
1483 	mutex_lock(&slab_mutex);
1484 	list_for_each_entry(cachep, &slab_caches, list)
1485 		if (enable_cpucache(cachep, GFP_NOWAIT))
1486 			BUG();
1487 	mutex_unlock(&slab_mutex);
1488 
1489 	/* Done! */
1490 	slab_state = FULL;
1491 
1492 	/*
1493 	 * Register a cpu startup notifier callback that initializes
1494 	 * cpu_cache_get for all new cpus
1495 	 */
1496 	register_cpu_notifier(&cpucache_notifier);
1497 
1498 #ifdef CONFIG_NUMA
1499 	/*
1500 	 * Register a memory hotplug callback that initializes and frees
1501 	 * node.
1502 	 */
1503 	hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1504 #endif
1505 
1506 	/*
1507 	 * The reap timers are started later, with a module init call: That part
1508 	 * of the kernel is not yet operational.
1509 	 */
1510 }
1511 
1512 static int __init cpucache_init(void)
1513 {
1514 	int cpu;
1515 
1516 	/*
1517 	 * Register the timers that return unneeded pages to the page allocator
1518 	 */
1519 	for_each_online_cpu(cpu)
1520 		start_cpu_timer(cpu);
1521 
1522 	/* Done! */
1523 	slab_state = FULL;
1524 	return 0;
1525 }
1526 __initcall(cpucache_init);
1527 
1528 static noinline void
1529 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1530 {
1531 #if DEBUG
1532 	struct kmem_cache_node *n;
1533 	struct page *page;
1534 	unsigned long flags;
1535 	int node;
1536 	static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1537 				      DEFAULT_RATELIMIT_BURST);
1538 
1539 	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1540 		return;
1541 
1542 	printk(KERN_WARNING
1543 		"SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1544 		nodeid, gfpflags);
1545 	printk(KERN_WARNING "  cache: %s, object size: %d, order: %d\n",
1546 		cachep->name, cachep->size, cachep->gfporder);
1547 
1548 	for_each_kmem_cache_node(cachep, node, n) {
1549 		unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1550 		unsigned long active_slabs = 0, num_slabs = 0;
1551 
1552 		spin_lock_irqsave(&n->list_lock, flags);
1553 		list_for_each_entry(page, &n->slabs_full, lru) {
1554 			active_objs += cachep->num;
1555 			active_slabs++;
1556 		}
1557 		list_for_each_entry(page, &n->slabs_partial, lru) {
1558 			active_objs += page->active;
1559 			active_slabs++;
1560 		}
1561 		list_for_each_entry(page, &n->slabs_free, lru)
1562 			num_slabs++;
1563 
1564 		free_objects += n->free_objects;
1565 		spin_unlock_irqrestore(&n->list_lock, flags);
1566 
1567 		num_slabs += active_slabs;
1568 		num_objs = num_slabs * cachep->num;
1569 		printk(KERN_WARNING
1570 			"  node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1571 			node, active_slabs, num_slabs, active_objs, num_objs,
1572 			free_objects);
1573 	}
1574 #endif
1575 }
1576 
1577 /*
1578  * Interface to system's page allocator. No need to hold the
1579  * kmem_cache_node ->list_lock.
1580  *
1581  * If we requested dmaable memory, we will get it. Even if we
1582  * did not request dmaable memory, we might get it, but that
1583  * would be relatively rare and ignorable.
1584  */
1585 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1586 								int nodeid)
1587 {
1588 	struct page *page;
1589 	int nr_pages;
1590 
1591 	flags |= cachep->allocflags;
1592 	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1593 		flags |= __GFP_RECLAIMABLE;
1594 
1595 	if (memcg_charge_slab(cachep, flags, cachep->gfporder))
1596 		return NULL;
1597 
1598 	page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1599 	if (!page) {
1600 		memcg_uncharge_slab(cachep, cachep->gfporder);
1601 		slab_out_of_memory(cachep, flags, nodeid);
1602 		return NULL;
1603 	}
1604 
1605 	/* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1606 	if (page_is_pfmemalloc(page))
1607 		pfmemalloc_active = true;
1608 
1609 	nr_pages = (1 << cachep->gfporder);
1610 	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1611 		add_zone_page_state(page_zone(page),
1612 			NR_SLAB_RECLAIMABLE, nr_pages);
1613 	else
1614 		add_zone_page_state(page_zone(page),
1615 			NR_SLAB_UNRECLAIMABLE, nr_pages);
1616 	__SetPageSlab(page);
1617 	if (page_is_pfmemalloc(page))
1618 		SetPageSlabPfmemalloc(page);
1619 
1620 	if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1621 		kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1622 
1623 		if (cachep->ctor)
1624 			kmemcheck_mark_uninitialized_pages(page, nr_pages);
1625 		else
1626 			kmemcheck_mark_unallocated_pages(page, nr_pages);
1627 	}
1628 
1629 	return page;
1630 }
1631 
1632 /*
1633  * Interface to system's page release.
1634  */
1635 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1636 {
1637 	const unsigned long nr_freed = (1 << cachep->gfporder);
1638 
1639 	kmemcheck_free_shadow(page, cachep->gfporder);
1640 
1641 	if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1642 		sub_zone_page_state(page_zone(page),
1643 				NR_SLAB_RECLAIMABLE, nr_freed);
1644 	else
1645 		sub_zone_page_state(page_zone(page),
1646 				NR_SLAB_UNRECLAIMABLE, nr_freed);
1647 
1648 	BUG_ON(!PageSlab(page));
1649 	__ClearPageSlabPfmemalloc(page);
1650 	__ClearPageSlab(page);
1651 	page_mapcount_reset(page);
1652 	page->mapping = NULL;
1653 
1654 	if (current->reclaim_state)
1655 		current->reclaim_state->reclaimed_slab += nr_freed;
1656 	__free_pages(page, cachep->gfporder);
1657 	memcg_uncharge_slab(cachep, cachep->gfporder);
1658 }
1659 
1660 static void kmem_rcu_free(struct rcu_head *head)
1661 {
1662 	struct kmem_cache *cachep;
1663 	struct page *page;
1664 
1665 	page = container_of(head, struct page, rcu_head);
1666 	cachep = page->slab_cache;
1667 
1668 	kmem_freepages(cachep, page);
1669 }
1670 
1671 #if DEBUG
1672 
1673 #ifdef CONFIG_DEBUG_PAGEALLOC
1674 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1675 			    unsigned long caller)
1676 {
1677 	int size = cachep->object_size;
1678 
1679 	addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1680 
1681 	if (size < 5 * sizeof(unsigned long))
1682 		return;
1683 
1684 	*addr++ = 0x12345678;
1685 	*addr++ = caller;
1686 	*addr++ = smp_processor_id();
1687 	size -= 3 * sizeof(unsigned long);
1688 	{
1689 		unsigned long *sptr = &caller;
1690 		unsigned long svalue;
1691 
1692 		while (!kstack_end(sptr)) {
1693 			svalue = *sptr++;
1694 			if (kernel_text_address(svalue)) {
1695 				*addr++ = svalue;
1696 				size -= sizeof(unsigned long);
1697 				if (size <= sizeof(unsigned long))
1698 					break;
1699 			}
1700 		}
1701 
1702 	}
1703 	*addr++ = 0x87654321;
1704 }
1705 #endif
1706 
1707 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1708 {
1709 	int size = cachep->object_size;
1710 	addr = &((char *)addr)[obj_offset(cachep)];
1711 
1712 	memset(addr, val, size);
1713 	*(unsigned char *)(addr + size - 1) = POISON_END;
1714 }
1715 
1716 static void dump_line(char *data, int offset, int limit)
1717 {
1718 	int i;
1719 	unsigned char error = 0;
1720 	int bad_count = 0;
1721 
1722 	printk(KERN_ERR "%03x: ", offset);
1723 	for (i = 0; i < limit; i++) {
1724 		if (data[offset + i] != POISON_FREE) {
1725 			error = data[offset + i];
1726 			bad_count++;
1727 		}
1728 	}
1729 	print_hex_dump(KERN_CONT, "", 0, 16, 1,
1730 			&data[offset], limit, 1);
1731 
1732 	if (bad_count == 1) {
1733 		error ^= POISON_FREE;
1734 		if (!(error & (error - 1))) {
1735 			printk(KERN_ERR "Single bit error detected. Probably "
1736 					"bad RAM.\n");
1737 #ifdef CONFIG_X86
1738 			printk(KERN_ERR "Run memtest86+ or a similar memory "
1739 					"test tool.\n");
1740 #else
1741 			printk(KERN_ERR "Run a memory test tool.\n");
1742 #endif
1743 		}
1744 	}
1745 }
1746 #endif
1747 
1748 #if DEBUG
1749 
1750 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1751 {
1752 	int i, size;
1753 	char *realobj;
1754 
1755 	if (cachep->flags & SLAB_RED_ZONE) {
1756 		printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1757 			*dbg_redzone1(cachep, objp),
1758 			*dbg_redzone2(cachep, objp));
1759 	}
1760 
1761 	if (cachep->flags & SLAB_STORE_USER) {
1762 		printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1763 		       *dbg_userword(cachep, objp),
1764 		       *dbg_userword(cachep, objp));
1765 	}
1766 	realobj = (char *)objp + obj_offset(cachep);
1767 	size = cachep->object_size;
1768 	for (i = 0; i < size && lines; i += 16, lines--) {
1769 		int limit;
1770 		limit = 16;
1771 		if (i + limit > size)
1772 			limit = size - i;
1773 		dump_line(realobj, i, limit);
1774 	}
1775 }
1776 
1777 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1778 {
1779 	char *realobj;
1780 	int size, i;
1781 	int lines = 0;
1782 
1783 	realobj = (char *)objp + obj_offset(cachep);
1784 	size = cachep->object_size;
1785 
1786 	for (i = 0; i < size; i++) {
1787 		char exp = POISON_FREE;
1788 		if (i == size - 1)
1789 			exp = POISON_END;
1790 		if (realobj[i] != exp) {
1791 			int limit;
1792 			/* Mismatch ! */
1793 			/* Print header */
1794 			if (lines == 0) {
1795 				printk(KERN_ERR
1796 					"Slab corruption (%s): %s start=%p, len=%d\n",
1797 					print_tainted(), cachep->name, realobj, size);
1798 				print_objinfo(cachep, objp, 0);
1799 			}
1800 			/* Hexdump the affected line */
1801 			i = (i / 16) * 16;
1802 			limit = 16;
1803 			if (i + limit > size)
1804 				limit = size - i;
1805 			dump_line(realobj, i, limit);
1806 			i += 16;
1807 			lines++;
1808 			/* Limit to 5 lines */
1809 			if (lines > 5)
1810 				break;
1811 		}
1812 	}
1813 	if (lines != 0) {
1814 		/* Print some data about the neighboring objects, if they
1815 		 * exist:
1816 		 */
1817 		struct page *page = virt_to_head_page(objp);
1818 		unsigned int objnr;
1819 
1820 		objnr = obj_to_index(cachep, page, objp);
1821 		if (objnr) {
1822 			objp = index_to_obj(cachep, page, objnr - 1);
1823 			realobj = (char *)objp + obj_offset(cachep);
1824 			printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1825 			       realobj, size);
1826 			print_objinfo(cachep, objp, 2);
1827 		}
1828 		if (objnr + 1 < cachep->num) {
1829 			objp = index_to_obj(cachep, page, objnr + 1);
1830 			realobj = (char *)objp + obj_offset(cachep);
1831 			printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1832 			       realobj, size);
1833 			print_objinfo(cachep, objp, 2);
1834 		}
1835 	}
1836 }
1837 #endif
1838 
1839 #if DEBUG
1840 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1841 						struct page *page)
1842 {
1843 	int i;
1844 	for (i = 0; i < cachep->num; i++) {
1845 		void *objp = index_to_obj(cachep, page, i);
1846 
1847 		if (cachep->flags & SLAB_POISON) {
1848 #ifdef CONFIG_DEBUG_PAGEALLOC
1849 			if (cachep->size % PAGE_SIZE == 0 &&
1850 					OFF_SLAB(cachep))
1851 				kernel_map_pages(virt_to_page(objp),
1852 					cachep->size / PAGE_SIZE, 1);
1853 			else
1854 				check_poison_obj(cachep, objp);
1855 #else
1856 			check_poison_obj(cachep, objp);
1857 #endif
1858 		}
1859 		if (cachep->flags & SLAB_RED_ZONE) {
1860 			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1861 				slab_error(cachep, "start of a freed object "
1862 					   "was overwritten");
1863 			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1864 				slab_error(cachep, "end of a freed object "
1865 					   "was overwritten");
1866 		}
1867 	}
1868 }
1869 #else
1870 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1871 						struct page *page)
1872 {
1873 }
1874 #endif
1875 
1876 /**
1877  * slab_destroy - destroy and release all objects in a slab
1878  * @cachep: cache pointer being destroyed
1879  * @page: page pointer being destroyed
1880  *
1881  * Destroy all the objs in a slab page, and release the mem back to the system.
1882  * Before calling the slab page must have been unlinked from the cache. The
1883  * kmem_cache_node ->list_lock is not held/needed.
1884  */
1885 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1886 {
1887 	void *freelist;
1888 
1889 	freelist = page->freelist;
1890 	slab_destroy_debugcheck(cachep, page);
1891 	if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1892 		struct rcu_head *head;
1893 
1894 		/*
1895 		 * RCU free overloads the RCU head over the LRU.
1896 		 * slab_page has been overloeaded over the LRU,
1897 		 * however it is not used from now on so that
1898 		 * we can use it safely.
1899 		 */
1900 		head = (void *)&page->rcu_head;
1901 		call_rcu(head, kmem_rcu_free);
1902 
1903 	} else {
1904 		kmem_freepages(cachep, page);
1905 	}
1906 
1907 	/*
1908 	 * From now on, we don't use freelist
1909 	 * although actual page can be freed in rcu context
1910 	 */
1911 	if (OFF_SLAB(cachep))
1912 		kmem_cache_free(cachep->freelist_cache, freelist);
1913 }
1914 
1915 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1916 {
1917 	struct page *page, *n;
1918 
1919 	list_for_each_entry_safe(page, n, list, lru) {
1920 		list_del(&page->lru);
1921 		slab_destroy(cachep, page);
1922 	}
1923 }
1924 
1925 /**
1926  * calculate_slab_order - calculate size (page order) of slabs
1927  * @cachep: pointer to the cache that is being created
1928  * @size: size of objects to be created in this cache.
1929  * @align: required alignment for the objects.
1930  * @flags: slab allocation flags
1931  *
1932  * Also calculates the number of objects per slab.
1933  *
1934  * This could be made much more intelligent.  For now, try to avoid using
1935  * high order pages for slabs.  When the gfp() functions are more friendly
1936  * towards high-order requests, this should be changed.
1937  */
1938 static size_t calculate_slab_order(struct kmem_cache *cachep,
1939 			size_t size, size_t align, unsigned long flags)
1940 {
1941 	unsigned long offslab_limit;
1942 	size_t left_over = 0;
1943 	int gfporder;
1944 
1945 	for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1946 		unsigned int num;
1947 		size_t remainder;
1948 
1949 		cache_estimate(gfporder, size, align, flags, &remainder, &num);
1950 		if (!num)
1951 			continue;
1952 
1953 		/* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1954 		if (num > SLAB_OBJ_MAX_NUM)
1955 			break;
1956 
1957 		if (flags & CFLGS_OFF_SLAB) {
1958 			size_t freelist_size_per_obj = sizeof(freelist_idx_t);
1959 			/*
1960 			 * Max number of objs-per-slab for caches which
1961 			 * use off-slab slabs. Needed to avoid a possible
1962 			 * looping condition in cache_grow().
1963 			 */
1964 			if (IS_ENABLED(CONFIG_DEBUG_SLAB_LEAK))
1965 				freelist_size_per_obj += sizeof(char);
1966 			offslab_limit = size;
1967 			offslab_limit /= freelist_size_per_obj;
1968 
1969  			if (num > offslab_limit)
1970 				break;
1971 		}
1972 
1973 		/* Found something acceptable - save it away */
1974 		cachep->num = num;
1975 		cachep->gfporder = gfporder;
1976 		left_over = remainder;
1977 
1978 		/*
1979 		 * A VFS-reclaimable slab tends to have most allocations
1980 		 * as GFP_NOFS and we really don't want to have to be allocating
1981 		 * higher-order pages when we are unable to shrink dcache.
1982 		 */
1983 		if (flags & SLAB_RECLAIM_ACCOUNT)
1984 			break;
1985 
1986 		/*
1987 		 * Large number of objects is good, but very large slabs are
1988 		 * currently bad for the gfp()s.
1989 		 */
1990 		if (gfporder >= slab_max_order)
1991 			break;
1992 
1993 		/*
1994 		 * Acceptable internal fragmentation?
1995 		 */
1996 		if (left_over * 8 <= (PAGE_SIZE << gfporder))
1997 			break;
1998 	}
1999 	return left_over;
2000 }
2001 
2002 static struct array_cache __percpu *alloc_kmem_cache_cpus(
2003 		struct kmem_cache *cachep, int entries, int batchcount)
2004 {
2005 	int cpu;
2006 	size_t size;
2007 	struct array_cache __percpu *cpu_cache;
2008 
2009 	size = sizeof(void *) * entries + sizeof(struct array_cache);
2010 	cpu_cache = __alloc_percpu(size, sizeof(void *));
2011 
2012 	if (!cpu_cache)
2013 		return NULL;
2014 
2015 	for_each_possible_cpu(cpu) {
2016 		init_arraycache(per_cpu_ptr(cpu_cache, cpu),
2017 				entries, batchcount);
2018 	}
2019 
2020 	return cpu_cache;
2021 }
2022 
2023 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2024 {
2025 	if (slab_state >= FULL)
2026 		return enable_cpucache(cachep, gfp);
2027 
2028 	cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
2029 	if (!cachep->cpu_cache)
2030 		return 1;
2031 
2032 	if (slab_state == DOWN) {
2033 		/* Creation of first cache (kmem_cache). */
2034 		set_up_node(kmem_cache, CACHE_CACHE);
2035 	} else if (slab_state == PARTIAL) {
2036 		/* For kmem_cache_node */
2037 		set_up_node(cachep, SIZE_NODE);
2038 	} else {
2039 		int node;
2040 
2041 		for_each_online_node(node) {
2042 			cachep->node[node] = kmalloc_node(
2043 				sizeof(struct kmem_cache_node), gfp, node);
2044 			BUG_ON(!cachep->node[node]);
2045 			kmem_cache_node_init(cachep->node[node]);
2046 		}
2047 	}
2048 
2049 	cachep->node[numa_mem_id()]->next_reap =
2050 			jiffies + REAPTIMEOUT_NODE +
2051 			((unsigned long)cachep) % REAPTIMEOUT_NODE;
2052 
2053 	cpu_cache_get(cachep)->avail = 0;
2054 	cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2055 	cpu_cache_get(cachep)->batchcount = 1;
2056 	cpu_cache_get(cachep)->touched = 0;
2057 	cachep->batchcount = 1;
2058 	cachep->limit = BOOT_CPUCACHE_ENTRIES;
2059 	return 0;
2060 }
2061 
2062 unsigned long kmem_cache_flags(unsigned long object_size,
2063 	unsigned long flags, const char *name,
2064 	void (*ctor)(void *))
2065 {
2066 	return flags;
2067 }
2068 
2069 struct kmem_cache *
2070 __kmem_cache_alias(const char *name, size_t size, size_t align,
2071 		   unsigned long flags, void (*ctor)(void *))
2072 {
2073 	struct kmem_cache *cachep;
2074 
2075 	cachep = find_mergeable(size, align, flags, name, ctor);
2076 	if (cachep) {
2077 		cachep->refcount++;
2078 
2079 		/*
2080 		 * Adjust the object sizes so that we clear
2081 		 * the complete object on kzalloc.
2082 		 */
2083 		cachep->object_size = max_t(int, cachep->object_size, size);
2084 	}
2085 	return cachep;
2086 }
2087 
2088 /**
2089  * __kmem_cache_create - Create a cache.
2090  * @cachep: cache management descriptor
2091  * @flags: SLAB flags
2092  *
2093  * Returns a ptr to the cache on success, NULL on failure.
2094  * Cannot be called within a int, but can be interrupted.
2095  * The @ctor is run when new pages are allocated by the cache.
2096  *
2097  * The flags are
2098  *
2099  * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2100  * to catch references to uninitialised memory.
2101  *
2102  * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2103  * for buffer overruns.
2104  *
2105  * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2106  * cacheline.  This can be beneficial if you're counting cycles as closely
2107  * as davem.
2108  */
2109 int
2110 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2111 {
2112 	size_t left_over, freelist_size;
2113 	size_t ralign = BYTES_PER_WORD;
2114 	gfp_t gfp;
2115 	int err;
2116 	size_t size = cachep->size;
2117 
2118 #if DEBUG
2119 #if FORCED_DEBUG
2120 	/*
2121 	 * Enable redzoning and last user accounting, except for caches with
2122 	 * large objects, if the increased size would increase the object size
2123 	 * above the next power of two: caches with object sizes just above a
2124 	 * power of two have a significant amount of internal fragmentation.
2125 	 */
2126 	if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2127 						2 * sizeof(unsigned long long)))
2128 		flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2129 	if (!(flags & SLAB_DESTROY_BY_RCU))
2130 		flags |= SLAB_POISON;
2131 #endif
2132 	if (flags & SLAB_DESTROY_BY_RCU)
2133 		BUG_ON(flags & SLAB_POISON);
2134 #endif
2135 
2136 	/*
2137 	 * Check that size is in terms of words.  This is needed to avoid
2138 	 * unaligned accesses for some archs when redzoning is used, and makes
2139 	 * sure any on-slab bufctl's are also correctly aligned.
2140 	 */
2141 	if (size & (BYTES_PER_WORD - 1)) {
2142 		size += (BYTES_PER_WORD - 1);
2143 		size &= ~(BYTES_PER_WORD - 1);
2144 	}
2145 
2146 	if (flags & SLAB_RED_ZONE) {
2147 		ralign = REDZONE_ALIGN;
2148 		/* If redzoning, ensure that the second redzone is suitably
2149 		 * aligned, by adjusting the object size accordingly. */
2150 		size += REDZONE_ALIGN - 1;
2151 		size &= ~(REDZONE_ALIGN - 1);
2152 	}
2153 
2154 	/* 3) caller mandated alignment */
2155 	if (ralign < cachep->align) {
2156 		ralign = cachep->align;
2157 	}
2158 	/* disable debug if necessary */
2159 	if (ralign > __alignof__(unsigned long long))
2160 		flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2161 	/*
2162 	 * 4) Store it.
2163 	 */
2164 	cachep->align = ralign;
2165 
2166 	if (slab_is_available())
2167 		gfp = GFP_KERNEL;
2168 	else
2169 		gfp = GFP_NOWAIT;
2170 
2171 #if DEBUG
2172 
2173 	/*
2174 	 * Both debugging options require word-alignment which is calculated
2175 	 * into align above.
2176 	 */
2177 	if (flags & SLAB_RED_ZONE) {
2178 		/* add space for red zone words */
2179 		cachep->obj_offset += sizeof(unsigned long long);
2180 		size += 2 * sizeof(unsigned long long);
2181 	}
2182 	if (flags & SLAB_STORE_USER) {
2183 		/* user store requires one word storage behind the end of
2184 		 * the real object. But if the second red zone needs to be
2185 		 * aligned to 64 bits, we must allow that much space.
2186 		 */
2187 		if (flags & SLAB_RED_ZONE)
2188 			size += REDZONE_ALIGN;
2189 		else
2190 			size += BYTES_PER_WORD;
2191 	}
2192 #if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2193 	/*
2194 	 * To activate debug pagealloc, off-slab management is necessary
2195 	 * requirement. In early phase of initialization, small sized slab
2196 	 * doesn't get initialized so it would not be possible. So, we need
2197 	 * to check size >= 256. It guarantees that all necessary small
2198 	 * sized slab is initialized in current slab initialization sequence.
2199 	 */
2200 	if (!slab_early_init && size >= kmalloc_size(INDEX_NODE) &&
2201 		size >= 256 && cachep->object_size > cache_line_size() &&
2202 		ALIGN(size, cachep->align) < PAGE_SIZE) {
2203 		cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2204 		size = PAGE_SIZE;
2205 	}
2206 #endif
2207 #endif
2208 
2209 	/*
2210 	 * Determine if the slab management is 'on' or 'off' slab.
2211 	 * (bootstrapping cannot cope with offslab caches so don't do
2212 	 * it too early on. Always use on-slab management when
2213 	 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2214 	 */
2215 	if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
2216 	    !(flags & SLAB_NOLEAKTRACE))
2217 		/*
2218 		 * Size is large, assume best to place the slab management obj
2219 		 * off-slab (should allow better packing of objs).
2220 		 */
2221 		flags |= CFLGS_OFF_SLAB;
2222 
2223 	size = ALIGN(size, cachep->align);
2224 	/*
2225 	 * We should restrict the number of objects in a slab to implement
2226 	 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2227 	 */
2228 	if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2229 		size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2230 
2231 	left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2232 
2233 	if (!cachep->num)
2234 		return -E2BIG;
2235 
2236 	freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2237 
2238 	/*
2239 	 * If the slab has been placed off-slab, and we have enough space then
2240 	 * move it on-slab. This is at the expense of any extra colouring.
2241 	 */
2242 	if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2243 		flags &= ~CFLGS_OFF_SLAB;
2244 		left_over -= freelist_size;
2245 	}
2246 
2247 	if (flags & CFLGS_OFF_SLAB) {
2248 		/* really off slab. No need for manual alignment */
2249 		freelist_size = calculate_freelist_size(cachep->num, 0);
2250 
2251 #ifdef CONFIG_PAGE_POISONING
2252 		/* If we're going to use the generic kernel_map_pages()
2253 		 * poisoning, then it's going to smash the contents of
2254 		 * the redzone and userword anyhow, so switch them off.
2255 		 */
2256 		if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2257 			flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2258 #endif
2259 	}
2260 
2261 	cachep->colour_off = cache_line_size();
2262 	/* Offset must be a multiple of the alignment. */
2263 	if (cachep->colour_off < cachep->align)
2264 		cachep->colour_off = cachep->align;
2265 	cachep->colour = left_over / cachep->colour_off;
2266 	cachep->freelist_size = freelist_size;
2267 	cachep->flags = flags;
2268 	cachep->allocflags = __GFP_COMP;
2269 	if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2270 		cachep->allocflags |= GFP_DMA;
2271 	cachep->size = size;
2272 	cachep->reciprocal_buffer_size = reciprocal_value(size);
2273 
2274 	if (flags & CFLGS_OFF_SLAB) {
2275 		cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2276 		/*
2277 		 * This is a possibility for one of the kmalloc_{dma,}_caches.
2278 		 * But since we go off slab only for object size greater than
2279 		 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2280 		 * in ascending order,this should not happen at all.
2281 		 * But leave a BUG_ON for some lucky dude.
2282 		 */
2283 		BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2284 	}
2285 
2286 	err = setup_cpu_cache(cachep, gfp);
2287 	if (err) {
2288 		__kmem_cache_shutdown(cachep);
2289 		return err;
2290 	}
2291 
2292 	return 0;
2293 }
2294 
2295 #if DEBUG
2296 static void check_irq_off(void)
2297 {
2298 	BUG_ON(!irqs_disabled());
2299 }
2300 
2301 static void check_irq_on(void)
2302 {
2303 	BUG_ON(irqs_disabled());
2304 }
2305 
2306 static void check_spinlock_acquired(struct kmem_cache *cachep)
2307 {
2308 #ifdef CONFIG_SMP
2309 	check_irq_off();
2310 	assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2311 #endif
2312 }
2313 
2314 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2315 {
2316 #ifdef CONFIG_SMP
2317 	check_irq_off();
2318 	assert_spin_locked(&get_node(cachep, node)->list_lock);
2319 #endif
2320 }
2321 
2322 #else
2323 #define check_irq_off()	do { } while(0)
2324 #define check_irq_on()	do { } while(0)
2325 #define check_spinlock_acquired(x) do { } while(0)
2326 #define check_spinlock_acquired_node(x, y) do { } while(0)
2327 #endif
2328 
2329 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2330 			struct array_cache *ac,
2331 			int force, int node);
2332 
2333 static void do_drain(void *arg)
2334 {
2335 	struct kmem_cache *cachep = arg;
2336 	struct array_cache *ac;
2337 	int node = numa_mem_id();
2338 	struct kmem_cache_node *n;
2339 	LIST_HEAD(list);
2340 
2341 	check_irq_off();
2342 	ac = cpu_cache_get(cachep);
2343 	n = get_node(cachep, node);
2344 	spin_lock(&n->list_lock);
2345 	free_block(cachep, ac->entry, ac->avail, node, &list);
2346 	spin_unlock(&n->list_lock);
2347 	slabs_destroy(cachep, &list);
2348 	ac->avail = 0;
2349 }
2350 
2351 static void drain_cpu_caches(struct kmem_cache *cachep)
2352 {
2353 	struct kmem_cache_node *n;
2354 	int node;
2355 
2356 	on_each_cpu(do_drain, cachep, 1);
2357 	check_irq_on();
2358 	for_each_kmem_cache_node(cachep, node, n)
2359 		if (n->alien)
2360 			drain_alien_cache(cachep, n->alien);
2361 
2362 	for_each_kmem_cache_node(cachep, node, n)
2363 		drain_array(cachep, n, n->shared, 1, node);
2364 }
2365 
2366 /*
2367  * Remove slabs from the list of free slabs.
2368  * Specify the number of slabs to drain in tofree.
2369  *
2370  * Returns the actual number of slabs released.
2371  */
2372 static int drain_freelist(struct kmem_cache *cache,
2373 			struct kmem_cache_node *n, int tofree)
2374 {
2375 	struct list_head *p;
2376 	int nr_freed;
2377 	struct page *page;
2378 
2379 	nr_freed = 0;
2380 	while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2381 
2382 		spin_lock_irq(&n->list_lock);
2383 		p = n->slabs_free.prev;
2384 		if (p == &n->slabs_free) {
2385 			spin_unlock_irq(&n->list_lock);
2386 			goto out;
2387 		}
2388 
2389 		page = list_entry(p, struct page, lru);
2390 #if DEBUG
2391 		BUG_ON(page->active);
2392 #endif
2393 		list_del(&page->lru);
2394 		/*
2395 		 * Safe to drop the lock. The slab is no longer linked
2396 		 * to the cache.
2397 		 */
2398 		n->free_objects -= cache->num;
2399 		spin_unlock_irq(&n->list_lock);
2400 		slab_destroy(cache, page);
2401 		nr_freed++;
2402 	}
2403 out:
2404 	return nr_freed;
2405 }
2406 
2407 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2408 {
2409 	int ret = 0;
2410 	int node;
2411 	struct kmem_cache_node *n;
2412 
2413 	drain_cpu_caches(cachep);
2414 
2415 	check_irq_on();
2416 	for_each_kmem_cache_node(cachep, node, n) {
2417 		drain_freelist(cachep, n, slabs_tofree(cachep, n));
2418 
2419 		ret += !list_empty(&n->slabs_full) ||
2420 			!list_empty(&n->slabs_partial);
2421 	}
2422 	return (ret ? 1 : 0);
2423 }
2424 
2425 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2426 {
2427 	int i;
2428 	struct kmem_cache_node *n;
2429 	int rc = __kmem_cache_shrink(cachep, false);
2430 
2431 	if (rc)
2432 		return rc;
2433 
2434 	free_percpu(cachep->cpu_cache);
2435 
2436 	/* NUMA: free the node structures */
2437 	for_each_kmem_cache_node(cachep, i, n) {
2438 		kfree(n->shared);
2439 		free_alien_cache(n->alien);
2440 		kfree(n);
2441 		cachep->node[i] = NULL;
2442 	}
2443 	return 0;
2444 }
2445 
2446 /*
2447  * Get the memory for a slab management obj.
2448  *
2449  * For a slab cache when the slab descriptor is off-slab, the
2450  * slab descriptor can't come from the same cache which is being created,
2451  * Because if it is the case, that means we defer the creation of
2452  * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2453  * And we eventually call down to __kmem_cache_create(), which
2454  * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2455  * This is a "chicken-and-egg" problem.
2456  *
2457  * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2458  * which are all initialized during kmem_cache_init().
2459  */
2460 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2461 				   struct page *page, int colour_off,
2462 				   gfp_t local_flags, int nodeid)
2463 {
2464 	void *freelist;
2465 	void *addr = page_address(page);
2466 
2467 	if (OFF_SLAB(cachep)) {
2468 		/* Slab management obj is off-slab. */
2469 		freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2470 					      local_flags, nodeid);
2471 		if (!freelist)
2472 			return NULL;
2473 	} else {
2474 		freelist = addr + colour_off;
2475 		colour_off += cachep->freelist_size;
2476 	}
2477 	page->active = 0;
2478 	page->s_mem = addr + colour_off;
2479 	return freelist;
2480 }
2481 
2482 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2483 {
2484 	return ((freelist_idx_t *)page->freelist)[idx];
2485 }
2486 
2487 static inline void set_free_obj(struct page *page,
2488 					unsigned int idx, freelist_idx_t val)
2489 {
2490 	((freelist_idx_t *)(page->freelist))[idx] = val;
2491 }
2492 
2493 static void cache_init_objs(struct kmem_cache *cachep,
2494 			    struct page *page)
2495 {
2496 	int i;
2497 
2498 	for (i = 0; i < cachep->num; i++) {
2499 		void *objp = index_to_obj(cachep, page, i);
2500 #if DEBUG
2501 		/* need to poison the objs? */
2502 		if (cachep->flags & SLAB_POISON)
2503 			poison_obj(cachep, objp, POISON_FREE);
2504 		if (cachep->flags & SLAB_STORE_USER)
2505 			*dbg_userword(cachep, objp) = NULL;
2506 
2507 		if (cachep->flags & SLAB_RED_ZONE) {
2508 			*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2509 			*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2510 		}
2511 		/*
2512 		 * Constructors are not allowed to allocate memory from the same
2513 		 * cache which they are a constructor for.  Otherwise, deadlock.
2514 		 * They must also be threaded.
2515 		 */
2516 		if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2517 			cachep->ctor(objp + obj_offset(cachep));
2518 
2519 		if (cachep->flags & SLAB_RED_ZONE) {
2520 			if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2521 				slab_error(cachep, "constructor overwrote the"
2522 					   " end of an object");
2523 			if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2524 				slab_error(cachep, "constructor overwrote the"
2525 					   " start of an object");
2526 		}
2527 		if ((cachep->size % PAGE_SIZE) == 0 &&
2528 			    OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2529 			kernel_map_pages(virt_to_page(objp),
2530 					 cachep->size / PAGE_SIZE, 0);
2531 #else
2532 		if (cachep->ctor)
2533 			cachep->ctor(objp);
2534 #endif
2535 		set_obj_status(page, i, OBJECT_FREE);
2536 		set_free_obj(page, i, i);
2537 	}
2538 }
2539 
2540 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2541 {
2542 	if (CONFIG_ZONE_DMA_FLAG) {
2543 		if (flags & GFP_DMA)
2544 			BUG_ON(!(cachep->allocflags & GFP_DMA));
2545 		else
2546 			BUG_ON(cachep->allocflags & GFP_DMA);
2547 	}
2548 }
2549 
2550 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
2551 				int nodeid)
2552 {
2553 	void *objp;
2554 
2555 	objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2556 	page->active++;
2557 #if DEBUG
2558 	WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2559 #endif
2560 
2561 	return objp;
2562 }
2563 
2564 static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
2565 				void *objp, int nodeid)
2566 {
2567 	unsigned int objnr = obj_to_index(cachep, page, objp);
2568 #if DEBUG
2569 	unsigned int i;
2570 
2571 	/* Verify that the slab belongs to the intended node */
2572 	WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2573 
2574 	/* Verify double free bug */
2575 	for (i = page->active; i < cachep->num; i++) {
2576 		if (get_free_obj(page, i) == objnr) {
2577 			printk(KERN_ERR "slab: double free detected in cache "
2578 					"'%s', objp %p\n", cachep->name, objp);
2579 			BUG();
2580 		}
2581 	}
2582 #endif
2583 	page->active--;
2584 	set_free_obj(page, page->active, objnr);
2585 }
2586 
2587 /*
2588  * Map pages beginning at addr to the given cache and slab. This is required
2589  * for the slab allocator to be able to lookup the cache and slab of a
2590  * virtual address for kfree, ksize, and slab debugging.
2591  */
2592 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2593 			   void *freelist)
2594 {
2595 	page->slab_cache = cache;
2596 	page->freelist = freelist;
2597 }
2598 
2599 /*
2600  * Grow (by 1) the number of slabs within a cache.  This is called by
2601  * kmem_cache_alloc() when there are no active objs left in a cache.
2602  */
2603 static int cache_grow(struct kmem_cache *cachep,
2604 		gfp_t flags, int nodeid, struct page *page)
2605 {
2606 	void *freelist;
2607 	size_t offset;
2608 	gfp_t local_flags;
2609 	struct kmem_cache_node *n;
2610 
2611 	/*
2612 	 * Be lazy and only check for valid flags here,  keeping it out of the
2613 	 * critical path in kmem_cache_alloc().
2614 	 */
2615 	if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2616 		pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2617 		BUG();
2618 	}
2619 	local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2620 
2621 	/* Take the node list lock to change the colour_next on this node */
2622 	check_irq_off();
2623 	n = get_node(cachep, nodeid);
2624 	spin_lock(&n->list_lock);
2625 
2626 	/* Get colour for the slab, and cal the next value. */
2627 	offset = n->colour_next;
2628 	n->colour_next++;
2629 	if (n->colour_next >= cachep->colour)
2630 		n->colour_next = 0;
2631 	spin_unlock(&n->list_lock);
2632 
2633 	offset *= cachep->colour_off;
2634 
2635 	if (local_flags & __GFP_WAIT)
2636 		local_irq_enable();
2637 
2638 	/*
2639 	 * The test for missing atomic flag is performed here, rather than
2640 	 * the more obvious place, simply to reduce the critical path length
2641 	 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2642 	 * will eventually be caught here (where it matters).
2643 	 */
2644 	kmem_flagcheck(cachep, flags);
2645 
2646 	/*
2647 	 * Get mem for the objs.  Attempt to allocate a physical page from
2648 	 * 'nodeid'.
2649 	 */
2650 	if (!page)
2651 		page = kmem_getpages(cachep, local_flags, nodeid);
2652 	if (!page)
2653 		goto failed;
2654 
2655 	/* Get slab management. */
2656 	freelist = alloc_slabmgmt(cachep, page, offset,
2657 			local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2658 	if (!freelist)
2659 		goto opps1;
2660 
2661 	slab_map_pages(cachep, page, freelist);
2662 
2663 	cache_init_objs(cachep, page);
2664 
2665 	if (local_flags & __GFP_WAIT)
2666 		local_irq_disable();
2667 	check_irq_off();
2668 	spin_lock(&n->list_lock);
2669 
2670 	/* Make slab active. */
2671 	list_add_tail(&page->lru, &(n->slabs_free));
2672 	STATS_INC_GROWN(cachep);
2673 	n->free_objects += cachep->num;
2674 	spin_unlock(&n->list_lock);
2675 	return 1;
2676 opps1:
2677 	kmem_freepages(cachep, page);
2678 failed:
2679 	if (local_flags & __GFP_WAIT)
2680 		local_irq_disable();
2681 	return 0;
2682 }
2683 
2684 #if DEBUG
2685 
2686 /*
2687  * Perform extra freeing checks:
2688  * - detect bad pointers.
2689  * - POISON/RED_ZONE checking
2690  */
2691 static void kfree_debugcheck(const void *objp)
2692 {
2693 	if (!virt_addr_valid(objp)) {
2694 		printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2695 		       (unsigned long)objp);
2696 		BUG();
2697 	}
2698 }
2699 
2700 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2701 {
2702 	unsigned long long redzone1, redzone2;
2703 
2704 	redzone1 = *dbg_redzone1(cache, obj);
2705 	redzone2 = *dbg_redzone2(cache, obj);
2706 
2707 	/*
2708 	 * Redzone is ok.
2709 	 */
2710 	if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2711 		return;
2712 
2713 	if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2714 		slab_error(cache, "double free detected");
2715 	else
2716 		slab_error(cache, "memory outside object was overwritten");
2717 
2718 	printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2719 			obj, redzone1, redzone2);
2720 }
2721 
2722 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2723 				   unsigned long caller)
2724 {
2725 	unsigned int objnr;
2726 	struct page *page;
2727 
2728 	BUG_ON(virt_to_cache(objp) != cachep);
2729 
2730 	objp -= obj_offset(cachep);
2731 	kfree_debugcheck(objp);
2732 	page = virt_to_head_page(objp);
2733 
2734 	if (cachep->flags & SLAB_RED_ZONE) {
2735 		verify_redzone_free(cachep, objp);
2736 		*dbg_redzone1(cachep, objp) = RED_INACTIVE;
2737 		*dbg_redzone2(cachep, objp) = RED_INACTIVE;
2738 	}
2739 	if (cachep->flags & SLAB_STORE_USER)
2740 		*dbg_userword(cachep, objp) = (void *)caller;
2741 
2742 	objnr = obj_to_index(cachep, page, objp);
2743 
2744 	BUG_ON(objnr >= cachep->num);
2745 	BUG_ON(objp != index_to_obj(cachep, page, objnr));
2746 
2747 	set_obj_status(page, objnr, OBJECT_FREE);
2748 	if (cachep->flags & SLAB_POISON) {
2749 #ifdef CONFIG_DEBUG_PAGEALLOC
2750 		if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2751 			store_stackinfo(cachep, objp, caller);
2752 			kernel_map_pages(virt_to_page(objp),
2753 					 cachep->size / PAGE_SIZE, 0);
2754 		} else {
2755 			poison_obj(cachep, objp, POISON_FREE);
2756 		}
2757 #else
2758 		poison_obj(cachep, objp, POISON_FREE);
2759 #endif
2760 	}
2761 	return objp;
2762 }
2763 
2764 #else
2765 #define kfree_debugcheck(x) do { } while(0)
2766 #define cache_free_debugcheck(x,objp,z) (objp)
2767 #endif
2768 
2769 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2770 							bool force_refill)
2771 {
2772 	int batchcount;
2773 	struct kmem_cache_node *n;
2774 	struct array_cache *ac;
2775 	int node;
2776 
2777 	check_irq_off();
2778 	node = numa_mem_id();
2779 	if (unlikely(force_refill))
2780 		goto force_grow;
2781 retry:
2782 	ac = cpu_cache_get(cachep);
2783 	batchcount = ac->batchcount;
2784 	if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2785 		/*
2786 		 * If there was little recent activity on this cache, then
2787 		 * perform only a partial refill.  Otherwise we could generate
2788 		 * refill bouncing.
2789 		 */
2790 		batchcount = BATCHREFILL_LIMIT;
2791 	}
2792 	n = get_node(cachep, node);
2793 
2794 	BUG_ON(ac->avail > 0 || !n);
2795 	spin_lock(&n->list_lock);
2796 
2797 	/* See if we can refill from the shared array */
2798 	if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2799 		n->shared->touched = 1;
2800 		goto alloc_done;
2801 	}
2802 
2803 	while (batchcount > 0) {
2804 		struct list_head *entry;
2805 		struct page *page;
2806 		/* Get slab alloc is to come from. */
2807 		entry = n->slabs_partial.next;
2808 		if (entry == &n->slabs_partial) {
2809 			n->free_touched = 1;
2810 			entry = n->slabs_free.next;
2811 			if (entry == &n->slabs_free)
2812 				goto must_grow;
2813 		}
2814 
2815 		page = list_entry(entry, struct page, lru);
2816 		check_spinlock_acquired(cachep);
2817 
2818 		/*
2819 		 * The slab was either on partial or free list so
2820 		 * there must be at least one object available for
2821 		 * allocation.
2822 		 */
2823 		BUG_ON(page->active >= cachep->num);
2824 
2825 		while (page->active < cachep->num && batchcount--) {
2826 			STATS_INC_ALLOCED(cachep);
2827 			STATS_INC_ACTIVE(cachep);
2828 			STATS_SET_HIGH(cachep);
2829 
2830 			ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
2831 									node));
2832 		}
2833 
2834 		/* move slabp to correct slabp list: */
2835 		list_del(&page->lru);
2836 		if (page->active == cachep->num)
2837 			list_add(&page->lru, &n->slabs_full);
2838 		else
2839 			list_add(&page->lru, &n->slabs_partial);
2840 	}
2841 
2842 must_grow:
2843 	n->free_objects -= ac->avail;
2844 alloc_done:
2845 	spin_unlock(&n->list_lock);
2846 
2847 	if (unlikely(!ac->avail)) {
2848 		int x;
2849 force_grow:
2850 		x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2851 
2852 		/* cache_grow can reenable interrupts, then ac could change. */
2853 		ac = cpu_cache_get(cachep);
2854 		node = numa_mem_id();
2855 
2856 		/* no objects in sight? abort */
2857 		if (!x && (ac->avail == 0 || force_refill))
2858 			return NULL;
2859 
2860 		if (!ac->avail)		/* objects refilled by interrupt? */
2861 			goto retry;
2862 	}
2863 	ac->touched = 1;
2864 
2865 	return ac_get_obj(cachep, ac, flags, force_refill);
2866 }
2867 
2868 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2869 						gfp_t flags)
2870 {
2871 	might_sleep_if(flags & __GFP_WAIT);
2872 #if DEBUG
2873 	kmem_flagcheck(cachep, flags);
2874 #endif
2875 }
2876 
2877 #if DEBUG
2878 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2879 				gfp_t flags, void *objp, unsigned long caller)
2880 {
2881 	struct page *page;
2882 
2883 	if (!objp)
2884 		return objp;
2885 	if (cachep->flags & SLAB_POISON) {
2886 #ifdef CONFIG_DEBUG_PAGEALLOC
2887 		if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2888 			kernel_map_pages(virt_to_page(objp),
2889 					 cachep->size / PAGE_SIZE, 1);
2890 		else
2891 			check_poison_obj(cachep, objp);
2892 #else
2893 		check_poison_obj(cachep, objp);
2894 #endif
2895 		poison_obj(cachep, objp, POISON_INUSE);
2896 	}
2897 	if (cachep->flags & SLAB_STORE_USER)
2898 		*dbg_userword(cachep, objp) = (void *)caller;
2899 
2900 	if (cachep->flags & SLAB_RED_ZONE) {
2901 		if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2902 				*dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2903 			slab_error(cachep, "double free, or memory outside"
2904 						" object was overwritten");
2905 			printk(KERN_ERR
2906 				"%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2907 				objp, *dbg_redzone1(cachep, objp),
2908 				*dbg_redzone2(cachep, objp));
2909 		}
2910 		*dbg_redzone1(cachep, objp) = RED_ACTIVE;
2911 		*dbg_redzone2(cachep, objp) = RED_ACTIVE;
2912 	}
2913 
2914 	page = virt_to_head_page(objp);
2915 	set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
2916 	objp += obj_offset(cachep);
2917 	if (cachep->ctor && cachep->flags & SLAB_POISON)
2918 		cachep->ctor(objp);
2919 	if (ARCH_SLAB_MINALIGN &&
2920 	    ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2921 		printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2922 		       objp, (int)ARCH_SLAB_MINALIGN);
2923 	}
2924 	return objp;
2925 }
2926 #else
2927 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2928 #endif
2929 
2930 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
2931 {
2932 	if (unlikely(cachep == kmem_cache))
2933 		return false;
2934 
2935 	return should_failslab(cachep->object_size, flags, cachep->flags);
2936 }
2937 
2938 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2939 {
2940 	void *objp;
2941 	struct array_cache *ac;
2942 	bool force_refill = false;
2943 
2944 	check_irq_off();
2945 
2946 	ac = cpu_cache_get(cachep);
2947 	if (likely(ac->avail)) {
2948 		ac->touched = 1;
2949 		objp = ac_get_obj(cachep, ac, flags, false);
2950 
2951 		/*
2952 		 * Allow for the possibility all avail objects are not allowed
2953 		 * by the current flags
2954 		 */
2955 		if (objp) {
2956 			STATS_INC_ALLOCHIT(cachep);
2957 			goto out;
2958 		}
2959 		force_refill = true;
2960 	}
2961 
2962 	STATS_INC_ALLOCMISS(cachep);
2963 	objp = cache_alloc_refill(cachep, flags, force_refill);
2964 	/*
2965 	 * the 'ac' may be updated by cache_alloc_refill(),
2966 	 * and kmemleak_erase() requires its correct value.
2967 	 */
2968 	ac = cpu_cache_get(cachep);
2969 
2970 out:
2971 	/*
2972 	 * To avoid a false negative, if an object that is in one of the
2973 	 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2974 	 * treat the array pointers as a reference to the object.
2975 	 */
2976 	if (objp)
2977 		kmemleak_erase(&ac->entry[ac->avail]);
2978 	return objp;
2979 }
2980 
2981 #ifdef CONFIG_NUMA
2982 /*
2983  * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2984  *
2985  * If we are in_interrupt, then process context, including cpusets and
2986  * mempolicy, may not apply and should not be used for allocation policy.
2987  */
2988 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2989 {
2990 	int nid_alloc, nid_here;
2991 
2992 	if (in_interrupt() || (flags & __GFP_THISNODE))
2993 		return NULL;
2994 	nid_alloc = nid_here = numa_mem_id();
2995 	if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2996 		nid_alloc = cpuset_slab_spread_node();
2997 	else if (current->mempolicy)
2998 		nid_alloc = mempolicy_slab_node();
2999 	if (nid_alloc != nid_here)
3000 		return ____cache_alloc_node(cachep, flags, nid_alloc);
3001 	return NULL;
3002 }
3003 
3004 /*
3005  * Fallback function if there was no memory available and no objects on a
3006  * certain node and fall back is permitted. First we scan all the
3007  * available node for available objects. If that fails then we
3008  * perform an allocation without specifying a node. This allows the page
3009  * allocator to do its reclaim / fallback magic. We then insert the
3010  * slab into the proper nodelist and then allocate from it.
3011  */
3012 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3013 {
3014 	struct zonelist *zonelist;
3015 	gfp_t local_flags;
3016 	struct zoneref *z;
3017 	struct zone *zone;
3018 	enum zone_type high_zoneidx = gfp_zone(flags);
3019 	void *obj = NULL;
3020 	int nid;
3021 	unsigned int cpuset_mems_cookie;
3022 
3023 	if (flags & __GFP_THISNODE)
3024 		return NULL;
3025 
3026 	local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3027 
3028 retry_cpuset:
3029 	cpuset_mems_cookie = read_mems_allowed_begin();
3030 	zonelist = node_zonelist(mempolicy_slab_node(), flags);
3031 
3032 retry:
3033 	/*
3034 	 * Look through allowed nodes for objects available
3035 	 * from existing per node queues.
3036 	 */
3037 	for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3038 		nid = zone_to_nid(zone);
3039 
3040 		if (cpuset_zone_allowed(zone, flags) &&
3041 			get_node(cache, nid) &&
3042 			get_node(cache, nid)->free_objects) {
3043 				obj = ____cache_alloc_node(cache,
3044 					gfp_exact_node(flags), nid);
3045 				if (obj)
3046 					break;
3047 		}
3048 	}
3049 
3050 	if (!obj) {
3051 		/*
3052 		 * This allocation will be performed within the constraints
3053 		 * of the current cpuset / memory policy requirements.
3054 		 * We may trigger various forms of reclaim on the allowed
3055 		 * set and go into memory reserves if necessary.
3056 		 */
3057 		struct page *page;
3058 
3059 		if (local_flags & __GFP_WAIT)
3060 			local_irq_enable();
3061 		kmem_flagcheck(cache, flags);
3062 		page = kmem_getpages(cache, local_flags, numa_mem_id());
3063 		if (local_flags & __GFP_WAIT)
3064 			local_irq_disable();
3065 		if (page) {
3066 			/*
3067 			 * Insert into the appropriate per node queues
3068 			 */
3069 			nid = page_to_nid(page);
3070 			if (cache_grow(cache, flags, nid, page)) {
3071 				obj = ____cache_alloc_node(cache,
3072 					gfp_exact_node(flags), nid);
3073 				if (!obj)
3074 					/*
3075 					 * Another processor may allocate the
3076 					 * objects in the slab since we are
3077 					 * not holding any locks.
3078 					 */
3079 					goto retry;
3080 			} else {
3081 				/* cache_grow already freed obj */
3082 				obj = NULL;
3083 			}
3084 		}
3085 	}
3086 
3087 	if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3088 		goto retry_cpuset;
3089 	return obj;
3090 }
3091 
3092 /*
3093  * A interface to enable slab creation on nodeid
3094  */
3095 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3096 				int nodeid)
3097 {
3098 	struct list_head *entry;
3099 	struct page *page;
3100 	struct kmem_cache_node *n;
3101 	void *obj;
3102 	int x;
3103 
3104 	VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3105 	n = get_node(cachep, nodeid);
3106 	BUG_ON(!n);
3107 
3108 retry:
3109 	check_irq_off();
3110 	spin_lock(&n->list_lock);
3111 	entry = n->slabs_partial.next;
3112 	if (entry == &n->slabs_partial) {
3113 		n->free_touched = 1;
3114 		entry = n->slabs_free.next;
3115 		if (entry == &n->slabs_free)
3116 			goto must_grow;
3117 	}
3118 
3119 	page = list_entry(entry, struct page, lru);
3120 	check_spinlock_acquired_node(cachep, nodeid);
3121 
3122 	STATS_INC_NODEALLOCS(cachep);
3123 	STATS_INC_ACTIVE(cachep);
3124 	STATS_SET_HIGH(cachep);
3125 
3126 	BUG_ON(page->active == cachep->num);
3127 
3128 	obj = slab_get_obj(cachep, page, nodeid);
3129 	n->free_objects--;
3130 	/* move slabp to correct slabp list: */
3131 	list_del(&page->lru);
3132 
3133 	if (page->active == cachep->num)
3134 		list_add(&page->lru, &n->slabs_full);
3135 	else
3136 		list_add(&page->lru, &n->slabs_partial);
3137 
3138 	spin_unlock(&n->list_lock);
3139 	goto done;
3140 
3141 must_grow:
3142 	spin_unlock(&n->list_lock);
3143 	x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3144 	if (x)
3145 		goto retry;
3146 
3147 	return fallback_alloc(cachep, flags);
3148 
3149 done:
3150 	return obj;
3151 }
3152 
3153 static __always_inline void *
3154 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3155 		   unsigned long caller)
3156 {
3157 	unsigned long save_flags;
3158 	void *ptr;
3159 	int slab_node = numa_mem_id();
3160 
3161 	flags &= gfp_allowed_mask;
3162 
3163 	lockdep_trace_alloc(flags);
3164 
3165 	if (slab_should_failslab(cachep, flags))
3166 		return NULL;
3167 
3168 	cachep = memcg_kmem_get_cache(cachep, flags);
3169 
3170 	cache_alloc_debugcheck_before(cachep, flags);
3171 	local_irq_save(save_flags);
3172 
3173 	if (nodeid == NUMA_NO_NODE)
3174 		nodeid = slab_node;
3175 
3176 	if (unlikely(!get_node(cachep, nodeid))) {
3177 		/* Node not bootstrapped yet */
3178 		ptr = fallback_alloc(cachep, flags);
3179 		goto out;
3180 	}
3181 
3182 	if (nodeid == slab_node) {
3183 		/*
3184 		 * Use the locally cached objects if possible.
3185 		 * However ____cache_alloc does not allow fallback
3186 		 * to other nodes. It may fail while we still have
3187 		 * objects on other nodes available.
3188 		 */
3189 		ptr = ____cache_alloc(cachep, flags);
3190 		if (ptr)
3191 			goto out;
3192 	}
3193 	/* ___cache_alloc_node can fall back to other nodes */
3194 	ptr = ____cache_alloc_node(cachep, flags, nodeid);
3195   out:
3196 	local_irq_restore(save_flags);
3197 	ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3198 	kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3199 				 flags);
3200 
3201 	if (likely(ptr)) {
3202 		kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3203 		if (unlikely(flags & __GFP_ZERO))
3204 			memset(ptr, 0, cachep->object_size);
3205 	}
3206 
3207 	memcg_kmem_put_cache(cachep);
3208 	return ptr;
3209 }
3210 
3211 static __always_inline void *
3212 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3213 {
3214 	void *objp;
3215 
3216 	if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3217 		objp = alternate_node_alloc(cache, flags);
3218 		if (objp)
3219 			goto out;
3220 	}
3221 	objp = ____cache_alloc(cache, flags);
3222 
3223 	/*
3224 	 * We may just have run out of memory on the local node.
3225 	 * ____cache_alloc_node() knows how to locate memory on other nodes
3226 	 */
3227 	if (!objp)
3228 		objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3229 
3230   out:
3231 	return objp;
3232 }
3233 #else
3234 
3235 static __always_inline void *
3236 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3237 {
3238 	return ____cache_alloc(cachep, flags);
3239 }
3240 
3241 #endif /* CONFIG_NUMA */
3242 
3243 static __always_inline void *
3244 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3245 {
3246 	unsigned long save_flags;
3247 	void *objp;
3248 
3249 	flags &= gfp_allowed_mask;
3250 
3251 	lockdep_trace_alloc(flags);
3252 
3253 	if (slab_should_failslab(cachep, flags))
3254 		return NULL;
3255 
3256 	cachep = memcg_kmem_get_cache(cachep, flags);
3257 
3258 	cache_alloc_debugcheck_before(cachep, flags);
3259 	local_irq_save(save_flags);
3260 	objp = __do_cache_alloc(cachep, flags);
3261 	local_irq_restore(save_flags);
3262 	objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3263 	kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3264 				 flags);
3265 	prefetchw(objp);
3266 
3267 	if (likely(objp)) {
3268 		kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3269 		if (unlikely(flags & __GFP_ZERO))
3270 			memset(objp, 0, cachep->object_size);
3271 	}
3272 
3273 	memcg_kmem_put_cache(cachep);
3274 	return objp;
3275 }
3276 
3277 /*
3278  * Caller needs to acquire correct kmem_cache_node's list_lock
3279  * @list: List of detached free slabs should be freed by caller
3280  */
3281 static void free_block(struct kmem_cache *cachep, void **objpp,
3282 			int nr_objects, int node, struct list_head *list)
3283 {
3284 	int i;
3285 	struct kmem_cache_node *n = get_node(cachep, node);
3286 
3287 	for (i = 0; i < nr_objects; i++) {
3288 		void *objp;
3289 		struct page *page;
3290 
3291 		clear_obj_pfmemalloc(&objpp[i]);
3292 		objp = objpp[i];
3293 
3294 		page = virt_to_head_page(objp);
3295 		list_del(&page->lru);
3296 		check_spinlock_acquired_node(cachep, node);
3297 		slab_put_obj(cachep, page, objp, node);
3298 		STATS_DEC_ACTIVE(cachep);
3299 		n->free_objects++;
3300 
3301 		/* fixup slab chains */
3302 		if (page->active == 0) {
3303 			if (n->free_objects > n->free_limit) {
3304 				n->free_objects -= cachep->num;
3305 				list_add_tail(&page->lru, list);
3306 			} else {
3307 				list_add(&page->lru, &n->slabs_free);
3308 			}
3309 		} else {
3310 			/* Unconditionally move a slab to the end of the
3311 			 * partial list on free - maximum time for the
3312 			 * other objects to be freed, too.
3313 			 */
3314 			list_add_tail(&page->lru, &n->slabs_partial);
3315 		}
3316 	}
3317 }
3318 
3319 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3320 {
3321 	int batchcount;
3322 	struct kmem_cache_node *n;
3323 	int node = numa_mem_id();
3324 	LIST_HEAD(list);
3325 
3326 	batchcount = ac->batchcount;
3327 #if DEBUG
3328 	BUG_ON(!batchcount || batchcount > ac->avail);
3329 #endif
3330 	check_irq_off();
3331 	n = get_node(cachep, node);
3332 	spin_lock(&n->list_lock);
3333 	if (n->shared) {
3334 		struct array_cache *shared_array = n->shared;
3335 		int max = shared_array->limit - shared_array->avail;
3336 		if (max) {
3337 			if (batchcount > max)
3338 				batchcount = max;
3339 			memcpy(&(shared_array->entry[shared_array->avail]),
3340 			       ac->entry, sizeof(void *) * batchcount);
3341 			shared_array->avail += batchcount;
3342 			goto free_done;
3343 		}
3344 	}
3345 
3346 	free_block(cachep, ac->entry, batchcount, node, &list);
3347 free_done:
3348 #if STATS
3349 	{
3350 		int i = 0;
3351 		struct list_head *p;
3352 
3353 		p = n->slabs_free.next;
3354 		while (p != &(n->slabs_free)) {
3355 			struct page *page;
3356 
3357 			page = list_entry(p, struct page, lru);
3358 			BUG_ON(page->active);
3359 
3360 			i++;
3361 			p = p->next;
3362 		}
3363 		STATS_SET_FREEABLE(cachep, i);
3364 	}
3365 #endif
3366 	spin_unlock(&n->list_lock);
3367 	slabs_destroy(cachep, &list);
3368 	ac->avail -= batchcount;
3369 	memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3370 }
3371 
3372 /*
3373  * Release an obj back to its cache. If the obj has a constructed state, it must
3374  * be in this state _before_ it is released.  Called with disabled ints.
3375  */
3376 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3377 				unsigned long caller)
3378 {
3379 	struct array_cache *ac = cpu_cache_get(cachep);
3380 
3381 	check_irq_off();
3382 	kmemleak_free_recursive(objp, cachep->flags);
3383 	objp = cache_free_debugcheck(cachep, objp, caller);
3384 
3385 	kmemcheck_slab_free(cachep, objp, cachep->object_size);
3386 
3387 	/*
3388 	 * Skip calling cache_free_alien() when the platform is not numa.
3389 	 * This will avoid cache misses that happen while accessing slabp (which
3390 	 * is per page memory  reference) to get nodeid. Instead use a global
3391 	 * variable to skip the call, which is mostly likely to be present in
3392 	 * the cache.
3393 	 */
3394 	if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3395 		return;
3396 
3397 	if (ac->avail < ac->limit) {
3398 		STATS_INC_FREEHIT(cachep);
3399 	} else {
3400 		STATS_INC_FREEMISS(cachep);
3401 		cache_flusharray(cachep, ac);
3402 	}
3403 
3404 	ac_put_obj(cachep, ac, objp);
3405 }
3406 
3407 /**
3408  * kmem_cache_alloc - Allocate an object
3409  * @cachep: The cache to allocate from.
3410  * @flags: See kmalloc().
3411  *
3412  * Allocate an object from this cache.  The flags are only relevant
3413  * if the cache has no available objects.
3414  */
3415 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3416 {
3417 	void *ret = slab_alloc(cachep, flags, _RET_IP_);
3418 
3419 	trace_kmem_cache_alloc(_RET_IP_, ret,
3420 			       cachep->object_size, cachep->size, flags);
3421 
3422 	return ret;
3423 }
3424 EXPORT_SYMBOL(kmem_cache_alloc);
3425 
3426 void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3427 {
3428 	__kmem_cache_free_bulk(s, size, p);
3429 }
3430 EXPORT_SYMBOL(kmem_cache_free_bulk);
3431 
3432 bool kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3433 								void **p)
3434 {
3435 	return __kmem_cache_alloc_bulk(s, flags, size, p);
3436 }
3437 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3438 
3439 #ifdef CONFIG_TRACING
3440 void *
3441 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3442 {
3443 	void *ret;
3444 
3445 	ret = slab_alloc(cachep, flags, _RET_IP_);
3446 
3447 	trace_kmalloc(_RET_IP_, ret,
3448 		      size, cachep->size, flags);
3449 	return ret;
3450 }
3451 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3452 #endif
3453 
3454 #ifdef CONFIG_NUMA
3455 /**
3456  * kmem_cache_alloc_node - Allocate an object on the specified node
3457  * @cachep: The cache to allocate from.
3458  * @flags: See kmalloc().
3459  * @nodeid: node number of the target node.
3460  *
3461  * Identical to kmem_cache_alloc but it will allocate memory on the given
3462  * node, which can improve the performance for cpu bound structures.
3463  *
3464  * Fallback to other node is possible if __GFP_THISNODE is not set.
3465  */
3466 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3467 {
3468 	void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3469 
3470 	trace_kmem_cache_alloc_node(_RET_IP_, ret,
3471 				    cachep->object_size, cachep->size,
3472 				    flags, nodeid);
3473 
3474 	return ret;
3475 }
3476 EXPORT_SYMBOL(kmem_cache_alloc_node);
3477 
3478 #ifdef CONFIG_TRACING
3479 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3480 				  gfp_t flags,
3481 				  int nodeid,
3482 				  size_t size)
3483 {
3484 	void *ret;
3485 
3486 	ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3487 
3488 	trace_kmalloc_node(_RET_IP_, ret,
3489 			   size, cachep->size,
3490 			   flags, nodeid);
3491 	return ret;
3492 }
3493 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3494 #endif
3495 
3496 static __always_inline void *
3497 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3498 {
3499 	struct kmem_cache *cachep;
3500 
3501 	cachep = kmalloc_slab(size, flags);
3502 	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3503 		return cachep;
3504 	return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3505 }
3506 
3507 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3508 {
3509 	return __do_kmalloc_node(size, flags, node, _RET_IP_);
3510 }
3511 EXPORT_SYMBOL(__kmalloc_node);
3512 
3513 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3514 		int node, unsigned long caller)
3515 {
3516 	return __do_kmalloc_node(size, flags, node, caller);
3517 }
3518 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3519 #endif /* CONFIG_NUMA */
3520 
3521 /**
3522  * __do_kmalloc - allocate memory
3523  * @size: how many bytes of memory are required.
3524  * @flags: the type of memory to allocate (see kmalloc).
3525  * @caller: function caller for debug tracking of the caller
3526  */
3527 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3528 					  unsigned long caller)
3529 {
3530 	struct kmem_cache *cachep;
3531 	void *ret;
3532 
3533 	cachep = kmalloc_slab(size, flags);
3534 	if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3535 		return cachep;
3536 	ret = slab_alloc(cachep, flags, caller);
3537 
3538 	trace_kmalloc(caller, ret,
3539 		      size, cachep->size, flags);
3540 
3541 	return ret;
3542 }
3543 
3544 void *__kmalloc(size_t size, gfp_t flags)
3545 {
3546 	return __do_kmalloc(size, flags, _RET_IP_);
3547 }
3548 EXPORT_SYMBOL(__kmalloc);
3549 
3550 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3551 {
3552 	return __do_kmalloc(size, flags, caller);
3553 }
3554 EXPORT_SYMBOL(__kmalloc_track_caller);
3555 
3556 /**
3557  * kmem_cache_free - Deallocate an object
3558  * @cachep: The cache the allocation was from.
3559  * @objp: The previously allocated object.
3560  *
3561  * Free an object which was previously allocated from this
3562  * cache.
3563  */
3564 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3565 {
3566 	unsigned long flags;
3567 	cachep = cache_from_obj(cachep, objp);
3568 	if (!cachep)
3569 		return;
3570 
3571 	local_irq_save(flags);
3572 	debug_check_no_locks_freed(objp, cachep->object_size);
3573 	if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3574 		debug_check_no_obj_freed(objp, cachep->object_size);
3575 	__cache_free(cachep, objp, _RET_IP_);
3576 	local_irq_restore(flags);
3577 
3578 	trace_kmem_cache_free(_RET_IP_, objp);
3579 }
3580 EXPORT_SYMBOL(kmem_cache_free);
3581 
3582 /**
3583  * kfree - free previously allocated memory
3584  * @objp: pointer returned by kmalloc.
3585  *
3586  * If @objp is NULL, no operation is performed.
3587  *
3588  * Don't free memory not originally allocated by kmalloc()
3589  * or you will run into trouble.
3590  */
3591 void kfree(const void *objp)
3592 {
3593 	struct kmem_cache *c;
3594 	unsigned long flags;
3595 
3596 	trace_kfree(_RET_IP_, objp);
3597 
3598 	if (unlikely(ZERO_OR_NULL_PTR(objp)))
3599 		return;
3600 	local_irq_save(flags);
3601 	kfree_debugcheck(objp);
3602 	c = virt_to_cache(objp);
3603 	debug_check_no_locks_freed(objp, c->object_size);
3604 
3605 	debug_check_no_obj_freed(objp, c->object_size);
3606 	__cache_free(c, (void *)objp, _RET_IP_);
3607 	local_irq_restore(flags);
3608 }
3609 EXPORT_SYMBOL(kfree);
3610 
3611 /*
3612  * This initializes kmem_cache_node or resizes various caches for all nodes.
3613  */
3614 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3615 {
3616 	int node;
3617 	struct kmem_cache_node *n;
3618 	struct array_cache *new_shared;
3619 	struct alien_cache **new_alien = NULL;
3620 
3621 	for_each_online_node(node) {
3622 
3623 		if (use_alien_caches) {
3624 			new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3625 			if (!new_alien)
3626 				goto fail;
3627 		}
3628 
3629 		new_shared = NULL;
3630 		if (cachep->shared) {
3631 			new_shared = alloc_arraycache(node,
3632 				cachep->shared*cachep->batchcount,
3633 					0xbaadf00d, gfp);
3634 			if (!new_shared) {
3635 				free_alien_cache(new_alien);
3636 				goto fail;
3637 			}
3638 		}
3639 
3640 		n = get_node(cachep, node);
3641 		if (n) {
3642 			struct array_cache *shared = n->shared;
3643 			LIST_HEAD(list);
3644 
3645 			spin_lock_irq(&n->list_lock);
3646 
3647 			if (shared)
3648 				free_block(cachep, shared->entry,
3649 						shared->avail, node, &list);
3650 
3651 			n->shared = new_shared;
3652 			if (!n->alien) {
3653 				n->alien = new_alien;
3654 				new_alien = NULL;
3655 			}
3656 			n->free_limit = (1 + nr_cpus_node(node)) *
3657 					cachep->batchcount + cachep->num;
3658 			spin_unlock_irq(&n->list_lock);
3659 			slabs_destroy(cachep, &list);
3660 			kfree(shared);
3661 			free_alien_cache(new_alien);
3662 			continue;
3663 		}
3664 		n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3665 		if (!n) {
3666 			free_alien_cache(new_alien);
3667 			kfree(new_shared);
3668 			goto fail;
3669 		}
3670 
3671 		kmem_cache_node_init(n);
3672 		n->next_reap = jiffies + REAPTIMEOUT_NODE +
3673 				((unsigned long)cachep) % REAPTIMEOUT_NODE;
3674 		n->shared = new_shared;
3675 		n->alien = new_alien;
3676 		n->free_limit = (1 + nr_cpus_node(node)) *
3677 					cachep->batchcount + cachep->num;
3678 		cachep->node[node] = n;
3679 	}
3680 	return 0;
3681 
3682 fail:
3683 	if (!cachep->list.next) {
3684 		/* Cache is not active yet. Roll back what we did */
3685 		node--;
3686 		while (node >= 0) {
3687 			n = get_node(cachep, node);
3688 			if (n) {
3689 				kfree(n->shared);
3690 				free_alien_cache(n->alien);
3691 				kfree(n);
3692 				cachep->node[node] = NULL;
3693 			}
3694 			node--;
3695 		}
3696 	}
3697 	return -ENOMEM;
3698 }
3699 
3700 /* Always called with the slab_mutex held */
3701 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3702 				int batchcount, int shared, gfp_t gfp)
3703 {
3704 	struct array_cache __percpu *cpu_cache, *prev;
3705 	int cpu;
3706 
3707 	cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3708 	if (!cpu_cache)
3709 		return -ENOMEM;
3710 
3711 	prev = cachep->cpu_cache;
3712 	cachep->cpu_cache = cpu_cache;
3713 	kick_all_cpus_sync();
3714 
3715 	check_irq_on();
3716 	cachep->batchcount = batchcount;
3717 	cachep->limit = limit;
3718 	cachep->shared = shared;
3719 
3720 	if (!prev)
3721 		goto alloc_node;
3722 
3723 	for_each_online_cpu(cpu) {
3724 		LIST_HEAD(list);
3725 		int node;
3726 		struct kmem_cache_node *n;
3727 		struct array_cache *ac = per_cpu_ptr(prev, cpu);
3728 
3729 		node = cpu_to_mem(cpu);
3730 		n = get_node(cachep, node);
3731 		spin_lock_irq(&n->list_lock);
3732 		free_block(cachep, ac->entry, ac->avail, node, &list);
3733 		spin_unlock_irq(&n->list_lock);
3734 		slabs_destroy(cachep, &list);
3735 	}
3736 	free_percpu(prev);
3737 
3738 alloc_node:
3739 	return alloc_kmem_cache_node(cachep, gfp);
3740 }
3741 
3742 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3743 				int batchcount, int shared, gfp_t gfp)
3744 {
3745 	int ret;
3746 	struct kmem_cache *c;
3747 
3748 	ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3749 
3750 	if (slab_state < FULL)
3751 		return ret;
3752 
3753 	if ((ret < 0) || !is_root_cache(cachep))
3754 		return ret;
3755 
3756 	lockdep_assert_held(&slab_mutex);
3757 	for_each_memcg_cache(c, cachep) {
3758 		/* return value determined by the root cache only */
3759 		__do_tune_cpucache(c, limit, batchcount, shared, gfp);
3760 	}
3761 
3762 	return ret;
3763 }
3764 
3765 /* Called with slab_mutex held always */
3766 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3767 {
3768 	int err;
3769 	int limit = 0;
3770 	int shared = 0;
3771 	int batchcount = 0;
3772 
3773 	if (!is_root_cache(cachep)) {
3774 		struct kmem_cache *root = memcg_root_cache(cachep);
3775 		limit = root->limit;
3776 		shared = root->shared;
3777 		batchcount = root->batchcount;
3778 	}
3779 
3780 	if (limit && shared && batchcount)
3781 		goto skip_setup;
3782 	/*
3783 	 * The head array serves three purposes:
3784 	 * - create a LIFO ordering, i.e. return objects that are cache-warm
3785 	 * - reduce the number of spinlock operations.
3786 	 * - reduce the number of linked list operations on the slab and
3787 	 *   bufctl chains: array operations are cheaper.
3788 	 * The numbers are guessed, we should auto-tune as described by
3789 	 * Bonwick.
3790 	 */
3791 	if (cachep->size > 131072)
3792 		limit = 1;
3793 	else if (cachep->size > PAGE_SIZE)
3794 		limit = 8;
3795 	else if (cachep->size > 1024)
3796 		limit = 24;
3797 	else if (cachep->size > 256)
3798 		limit = 54;
3799 	else
3800 		limit = 120;
3801 
3802 	/*
3803 	 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3804 	 * allocation behaviour: Most allocs on one cpu, most free operations
3805 	 * on another cpu. For these cases, an efficient object passing between
3806 	 * cpus is necessary. This is provided by a shared array. The array
3807 	 * replaces Bonwick's magazine layer.
3808 	 * On uniprocessor, it's functionally equivalent (but less efficient)
3809 	 * to a larger limit. Thus disabled by default.
3810 	 */
3811 	shared = 0;
3812 	if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3813 		shared = 8;
3814 
3815 #if DEBUG
3816 	/*
3817 	 * With debugging enabled, large batchcount lead to excessively long
3818 	 * periods with disabled local interrupts. Limit the batchcount
3819 	 */
3820 	if (limit > 32)
3821 		limit = 32;
3822 #endif
3823 	batchcount = (limit + 1) / 2;
3824 skip_setup:
3825 	err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3826 	if (err)
3827 		printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3828 		       cachep->name, -err);
3829 	return err;
3830 }
3831 
3832 /*
3833  * Drain an array if it contains any elements taking the node lock only if
3834  * necessary. Note that the node listlock also protects the array_cache
3835  * if drain_array() is used on the shared array.
3836  */
3837 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3838 			 struct array_cache *ac, int force, int node)
3839 {
3840 	LIST_HEAD(list);
3841 	int tofree;
3842 
3843 	if (!ac || !ac->avail)
3844 		return;
3845 	if (ac->touched && !force) {
3846 		ac->touched = 0;
3847 	} else {
3848 		spin_lock_irq(&n->list_lock);
3849 		if (ac->avail) {
3850 			tofree = force ? ac->avail : (ac->limit + 4) / 5;
3851 			if (tofree > ac->avail)
3852 				tofree = (ac->avail + 1) / 2;
3853 			free_block(cachep, ac->entry, tofree, node, &list);
3854 			ac->avail -= tofree;
3855 			memmove(ac->entry, &(ac->entry[tofree]),
3856 				sizeof(void *) * ac->avail);
3857 		}
3858 		spin_unlock_irq(&n->list_lock);
3859 		slabs_destroy(cachep, &list);
3860 	}
3861 }
3862 
3863 /**
3864  * cache_reap - Reclaim memory from caches.
3865  * @w: work descriptor
3866  *
3867  * Called from workqueue/eventd every few seconds.
3868  * Purpose:
3869  * - clear the per-cpu caches for this CPU.
3870  * - return freeable pages to the main free memory pool.
3871  *
3872  * If we cannot acquire the cache chain mutex then just give up - we'll try
3873  * again on the next iteration.
3874  */
3875 static void cache_reap(struct work_struct *w)
3876 {
3877 	struct kmem_cache *searchp;
3878 	struct kmem_cache_node *n;
3879 	int node = numa_mem_id();
3880 	struct delayed_work *work = to_delayed_work(w);
3881 
3882 	if (!mutex_trylock(&slab_mutex))
3883 		/* Give up. Setup the next iteration. */
3884 		goto out;
3885 
3886 	list_for_each_entry(searchp, &slab_caches, list) {
3887 		check_irq_on();
3888 
3889 		/*
3890 		 * We only take the node lock if absolutely necessary and we
3891 		 * have established with reasonable certainty that
3892 		 * we can do some work if the lock was obtained.
3893 		 */
3894 		n = get_node(searchp, node);
3895 
3896 		reap_alien(searchp, n);
3897 
3898 		drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3899 
3900 		/*
3901 		 * These are racy checks but it does not matter
3902 		 * if we skip one check or scan twice.
3903 		 */
3904 		if (time_after(n->next_reap, jiffies))
3905 			goto next;
3906 
3907 		n->next_reap = jiffies + REAPTIMEOUT_NODE;
3908 
3909 		drain_array(searchp, n, n->shared, 0, node);
3910 
3911 		if (n->free_touched)
3912 			n->free_touched = 0;
3913 		else {
3914 			int freed;
3915 
3916 			freed = drain_freelist(searchp, n, (n->free_limit +
3917 				5 * searchp->num - 1) / (5 * searchp->num));
3918 			STATS_ADD_REAPED(searchp, freed);
3919 		}
3920 next:
3921 		cond_resched();
3922 	}
3923 	check_irq_on();
3924 	mutex_unlock(&slab_mutex);
3925 	next_reap_node();
3926 out:
3927 	/* Set up the next iteration */
3928 	schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3929 }
3930 
3931 #ifdef CONFIG_SLABINFO
3932 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3933 {
3934 	struct page *page;
3935 	unsigned long active_objs;
3936 	unsigned long num_objs;
3937 	unsigned long active_slabs = 0;
3938 	unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3939 	const char *name;
3940 	char *error = NULL;
3941 	int node;
3942 	struct kmem_cache_node *n;
3943 
3944 	active_objs = 0;
3945 	num_slabs = 0;
3946 	for_each_kmem_cache_node(cachep, node, n) {
3947 
3948 		check_irq_on();
3949 		spin_lock_irq(&n->list_lock);
3950 
3951 		list_for_each_entry(page, &n->slabs_full, lru) {
3952 			if (page->active != cachep->num && !error)
3953 				error = "slabs_full accounting error";
3954 			active_objs += cachep->num;
3955 			active_slabs++;
3956 		}
3957 		list_for_each_entry(page, &n->slabs_partial, lru) {
3958 			if (page->active == cachep->num && !error)
3959 				error = "slabs_partial accounting error";
3960 			if (!page->active && !error)
3961 				error = "slabs_partial accounting error";
3962 			active_objs += page->active;
3963 			active_slabs++;
3964 		}
3965 		list_for_each_entry(page, &n->slabs_free, lru) {
3966 			if (page->active && !error)
3967 				error = "slabs_free accounting error";
3968 			num_slabs++;
3969 		}
3970 		free_objects += n->free_objects;
3971 		if (n->shared)
3972 			shared_avail += n->shared->avail;
3973 
3974 		spin_unlock_irq(&n->list_lock);
3975 	}
3976 	num_slabs += active_slabs;
3977 	num_objs = num_slabs * cachep->num;
3978 	if (num_objs - active_objs != free_objects && !error)
3979 		error = "free_objects accounting error";
3980 
3981 	name = cachep->name;
3982 	if (error)
3983 		printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3984 
3985 	sinfo->active_objs = active_objs;
3986 	sinfo->num_objs = num_objs;
3987 	sinfo->active_slabs = active_slabs;
3988 	sinfo->num_slabs = num_slabs;
3989 	sinfo->shared_avail = shared_avail;
3990 	sinfo->limit = cachep->limit;
3991 	sinfo->batchcount = cachep->batchcount;
3992 	sinfo->shared = cachep->shared;
3993 	sinfo->objects_per_slab = cachep->num;
3994 	sinfo->cache_order = cachep->gfporder;
3995 }
3996 
3997 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3998 {
3999 #if STATS
4000 	{			/* node stats */
4001 		unsigned long high = cachep->high_mark;
4002 		unsigned long allocs = cachep->num_allocations;
4003 		unsigned long grown = cachep->grown;
4004 		unsigned long reaped = cachep->reaped;
4005 		unsigned long errors = cachep->errors;
4006 		unsigned long max_freeable = cachep->max_freeable;
4007 		unsigned long node_allocs = cachep->node_allocs;
4008 		unsigned long node_frees = cachep->node_frees;
4009 		unsigned long overflows = cachep->node_overflow;
4010 
4011 		seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4012 			   "%4lu %4lu %4lu %4lu %4lu",
4013 			   allocs, high, grown,
4014 			   reaped, errors, max_freeable, node_allocs,
4015 			   node_frees, overflows);
4016 	}
4017 	/* cpu stats */
4018 	{
4019 		unsigned long allochit = atomic_read(&cachep->allochit);
4020 		unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4021 		unsigned long freehit = atomic_read(&cachep->freehit);
4022 		unsigned long freemiss = atomic_read(&cachep->freemiss);
4023 
4024 		seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4025 			   allochit, allocmiss, freehit, freemiss);
4026 	}
4027 #endif
4028 }
4029 
4030 #define MAX_SLABINFO_WRITE 128
4031 /**
4032  * slabinfo_write - Tuning for the slab allocator
4033  * @file: unused
4034  * @buffer: user buffer
4035  * @count: data length
4036  * @ppos: unused
4037  */
4038 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4039 		       size_t count, loff_t *ppos)
4040 {
4041 	char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4042 	int limit, batchcount, shared, res;
4043 	struct kmem_cache *cachep;
4044 
4045 	if (count > MAX_SLABINFO_WRITE)
4046 		return -EINVAL;
4047 	if (copy_from_user(&kbuf, buffer, count))
4048 		return -EFAULT;
4049 	kbuf[MAX_SLABINFO_WRITE] = '\0';
4050 
4051 	tmp = strchr(kbuf, ' ');
4052 	if (!tmp)
4053 		return -EINVAL;
4054 	*tmp = '\0';
4055 	tmp++;
4056 	if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4057 		return -EINVAL;
4058 
4059 	/* Find the cache in the chain of caches. */
4060 	mutex_lock(&slab_mutex);
4061 	res = -EINVAL;
4062 	list_for_each_entry(cachep, &slab_caches, list) {
4063 		if (!strcmp(cachep->name, kbuf)) {
4064 			if (limit < 1 || batchcount < 1 ||
4065 					batchcount > limit || shared < 0) {
4066 				res = 0;
4067 			} else {
4068 				res = do_tune_cpucache(cachep, limit,
4069 						       batchcount, shared,
4070 						       GFP_KERNEL);
4071 			}
4072 			break;
4073 		}
4074 	}
4075 	mutex_unlock(&slab_mutex);
4076 	if (res >= 0)
4077 		res = count;
4078 	return res;
4079 }
4080 
4081 #ifdef CONFIG_DEBUG_SLAB_LEAK
4082 
4083 static inline int add_caller(unsigned long *n, unsigned long v)
4084 {
4085 	unsigned long *p;
4086 	int l;
4087 	if (!v)
4088 		return 1;
4089 	l = n[1];
4090 	p = n + 2;
4091 	while (l) {
4092 		int i = l/2;
4093 		unsigned long *q = p + 2 * i;
4094 		if (*q == v) {
4095 			q[1]++;
4096 			return 1;
4097 		}
4098 		if (*q > v) {
4099 			l = i;
4100 		} else {
4101 			p = q + 2;
4102 			l -= i + 1;
4103 		}
4104 	}
4105 	if (++n[1] == n[0])
4106 		return 0;
4107 	memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4108 	p[0] = v;
4109 	p[1] = 1;
4110 	return 1;
4111 }
4112 
4113 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4114 						struct page *page)
4115 {
4116 	void *p;
4117 	int i;
4118 
4119 	if (n[0] == n[1])
4120 		return;
4121 	for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4122 		if (get_obj_status(page, i) != OBJECT_ACTIVE)
4123 			continue;
4124 
4125 		if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4126 			return;
4127 	}
4128 }
4129 
4130 static void show_symbol(struct seq_file *m, unsigned long address)
4131 {
4132 #ifdef CONFIG_KALLSYMS
4133 	unsigned long offset, size;
4134 	char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4135 
4136 	if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4137 		seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4138 		if (modname[0])
4139 			seq_printf(m, " [%s]", modname);
4140 		return;
4141 	}
4142 #endif
4143 	seq_printf(m, "%p", (void *)address);
4144 }
4145 
4146 static int leaks_show(struct seq_file *m, void *p)
4147 {
4148 	struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4149 	struct page *page;
4150 	struct kmem_cache_node *n;
4151 	const char *name;
4152 	unsigned long *x = m->private;
4153 	int node;
4154 	int i;
4155 
4156 	if (!(cachep->flags & SLAB_STORE_USER))
4157 		return 0;
4158 	if (!(cachep->flags & SLAB_RED_ZONE))
4159 		return 0;
4160 
4161 	/* OK, we can do it */
4162 
4163 	x[1] = 0;
4164 
4165 	for_each_kmem_cache_node(cachep, node, n) {
4166 
4167 		check_irq_on();
4168 		spin_lock_irq(&n->list_lock);
4169 
4170 		list_for_each_entry(page, &n->slabs_full, lru)
4171 			handle_slab(x, cachep, page);
4172 		list_for_each_entry(page, &n->slabs_partial, lru)
4173 			handle_slab(x, cachep, page);
4174 		spin_unlock_irq(&n->list_lock);
4175 	}
4176 	name = cachep->name;
4177 	if (x[0] == x[1]) {
4178 		/* Increase the buffer size */
4179 		mutex_unlock(&slab_mutex);
4180 		m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4181 		if (!m->private) {
4182 			/* Too bad, we are really out */
4183 			m->private = x;
4184 			mutex_lock(&slab_mutex);
4185 			return -ENOMEM;
4186 		}
4187 		*(unsigned long *)m->private = x[0] * 2;
4188 		kfree(x);
4189 		mutex_lock(&slab_mutex);
4190 		/* Now make sure this entry will be retried */
4191 		m->count = m->size;
4192 		return 0;
4193 	}
4194 	for (i = 0; i < x[1]; i++) {
4195 		seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4196 		show_symbol(m, x[2*i+2]);
4197 		seq_putc(m, '\n');
4198 	}
4199 
4200 	return 0;
4201 }
4202 
4203 static const struct seq_operations slabstats_op = {
4204 	.start = slab_start,
4205 	.next = slab_next,
4206 	.stop = slab_stop,
4207 	.show = leaks_show,
4208 };
4209 
4210 static int slabstats_open(struct inode *inode, struct file *file)
4211 {
4212 	unsigned long *n;
4213 
4214 	n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4215 	if (!n)
4216 		return -ENOMEM;
4217 
4218 	*n = PAGE_SIZE / (2 * sizeof(unsigned long));
4219 
4220 	return 0;
4221 }
4222 
4223 static const struct file_operations proc_slabstats_operations = {
4224 	.open		= slabstats_open,
4225 	.read		= seq_read,
4226 	.llseek		= seq_lseek,
4227 	.release	= seq_release_private,
4228 };
4229 #endif
4230 
4231 static int __init slab_proc_init(void)
4232 {
4233 #ifdef CONFIG_DEBUG_SLAB_LEAK
4234 	proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4235 #endif
4236 	return 0;
4237 }
4238 module_init(slab_proc_init);
4239 #endif
4240 
4241 /**
4242  * ksize - get the actual amount of memory allocated for a given object
4243  * @objp: Pointer to the object
4244  *
4245  * kmalloc may internally round up allocations and return more memory
4246  * than requested. ksize() can be used to determine the actual amount of
4247  * memory allocated. The caller may use this additional memory, even though
4248  * a smaller amount of memory was initially specified with the kmalloc call.
4249  * The caller must guarantee that objp points to a valid object previously
4250  * allocated with either kmalloc() or kmem_cache_alloc(). The object
4251  * must not be freed during the duration of the call.
4252  */
4253 size_t ksize(const void *objp)
4254 {
4255 	BUG_ON(!objp);
4256 	if (unlikely(objp == ZERO_SIZE_PTR))
4257 		return 0;
4258 
4259 	return virt_to_cache(objp)->object_size;
4260 }
4261 EXPORT_SYMBOL(ksize);
4262