]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/blob - mm/slab.c
Btrfs: fix device replace of a missing RAID 5/6 device
[mirror_ubuntu-artful-kernel.git] / mm / slab.c
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_exact_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 (unlikely(page->pfmemalloc))
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->pfmemalloc)
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 if (size >= kmalloc_size(INDEX_NODE + 1)
2194 && cachep->object_size > cache_line_size()
2195 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2196 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2197 size = PAGE_SIZE;
2198 }
2199 #endif
2200 #endif
2201
2202 /*
2203 * Determine if the slab management is 'on' or 'off' slab.
2204 * (bootstrapping cannot cope with offslab caches so don't do
2205 * it too early on. Always use on-slab management when
2206 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2207 */
2208 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
2209 !(flags & SLAB_NOLEAKTRACE))
2210 /*
2211 * Size is large, assume best to place the slab management obj
2212 * off-slab (should allow better packing of objs).
2213 */
2214 flags |= CFLGS_OFF_SLAB;
2215
2216 size = ALIGN(size, cachep->align);
2217 /*
2218 * We should restrict the number of objects in a slab to implement
2219 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2220 */
2221 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2222 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2223
2224 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2225
2226 if (!cachep->num)
2227 return -E2BIG;
2228
2229 freelist_size = calculate_freelist_size(cachep->num, cachep->align);
2230
2231 /*
2232 * If the slab has been placed off-slab, and we have enough space then
2233 * move it on-slab. This is at the expense of any extra colouring.
2234 */
2235 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2236 flags &= ~CFLGS_OFF_SLAB;
2237 left_over -= freelist_size;
2238 }
2239
2240 if (flags & CFLGS_OFF_SLAB) {
2241 /* really off slab. No need for manual alignment */
2242 freelist_size = calculate_freelist_size(cachep->num, 0);
2243
2244 #ifdef CONFIG_PAGE_POISONING
2245 /* If we're going to use the generic kernel_map_pages()
2246 * poisoning, then it's going to smash the contents of
2247 * the redzone and userword anyhow, so switch them off.
2248 */
2249 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2250 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2251 #endif
2252 }
2253
2254 cachep->colour_off = cache_line_size();
2255 /* Offset must be a multiple of the alignment. */
2256 if (cachep->colour_off < cachep->align)
2257 cachep->colour_off = cachep->align;
2258 cachep->colour = left_over / cachep->colour_off;
2259 cachep->freelist_size = freelist_size;
2260 cachep->flags = flags;
2261 cachep->allocflags = __GFP_COMP;
2262 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2263 cachep->allocflags |= GFP_DMA;
2264 cachep->size = size;
2265 cachep->reciprocal_buffer_size = reciprocal_value(size);
2266
2267 if (flags & CFLGS_OFF_SLAB) {
2268 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2269 /*
2270 * This is a possibility for one of the kmalloc_{dma,}_caches.
2271 * But since we go off slab only for object size greater than
2272 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2273 * in ascending order,this should not happen at all.
2274 * But leave a BUG_ON for some lucky dude.
2275 */
2276 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2277 }
2278
2279 err = setup_cpu_cache(cachep, gfp);
2280 if (err) {
2281 __kmem_cache_shutdown(cachep);
2282 return err;
2283 }
2284
2285 return 0;
2286 }
2287
2288 #if DEBUG
2289 static void check_irq_off(void)
2290 {
2291 BUG_ON(!irqs_disabled());
2292 }
2293
2294 static void check_irq_on(void)
2295 {
2296 BUG_ON(irqs_disabled());
2297 }
2298
2299 static void check_spinlock_acquired(struct kmem_cache *cachep)
2300 {
2301 #ifdef CONFIG_SMP
2302 check_irq_off();
2303 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2304 #endif
2305 }
2306
2307 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2308 {
2309 #ifdef CONFIG_SMP
2310 check_irq_off();
2311 assert_spin_locked(&get_node(cachep, node)->list_lock);
2312 #endif
2313 }
2314
2315 #else
2316 #define check_irq_off() do { } while(0)
2317 #define check_irq_on() do { } while(0)
2318 #define check_spinlock_acquired(x) do { } while(0)
2319 #define check_spinlock_acquired_node(x, y) do { } while(0)
2320 #endif
2321
2322 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2323 struct array_cache *ac,
2324 int force, int node);
2325
2326 static void do_drain(void *arg)
2327 {
2328 struct kmem_cache *cachep = arg;
2329 struct array_cache *ac;
2330 int node = numa_mem_id();
2331 struct kmem_cache_node *n;
2332 LIST_HEAD(list);
2333
2334 check_irq_off();
2335 ac = cpu_cache_get(cachep);
2336 n = get_node(cachep, node);
2337 spin_lock(&n->list_lock);
2338 free_block(cachep, ac->entry, ac->avail, node, &list);
2339 spin_unlock(&n->list_lock);
2340 slabs_destroy(cachep, &list);
2341 ac->avail = 0;
2342 }
2343
2344 static void drain_cpu_caches(struct kmem_cache *cachep)
2345 {
2346 struct kmem_cache_node *n;
2347 int node;
2348
2349 on_each_cpu(do_drain, cachep, 1);
2350 check_irq_on();
2351 for_each_kmem_cache_node(cachep, node, n)
2352 if (n->alien)
2353 drain_alien_cache(cachep, n->alien);
2354
2355 for_each_kmem_cache_node(cachep, node, n)
2356 drain_array(cachep, n, n->shared, 1, node);
2357 }
2358
2359 /*
2360 * Remove slabs from the list of free slabs.
2361 * Specify the number of slabs to drain in tofree.
2362 *
2363 * Returns the actual number of slabs released.
2364 */
2365 static int drain_freelist(struct kmem_cache *cache,
2366 struct kmem_cache_node *n, int tofree)
2367 {
2368 struct list_head *p;
2369 int nr_freed;
2370 struct page *page;
2371
2372 nr_freed = 0;
2373 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2374
2375 spin_lock_irq(&n->list_lock);
2376 p = n->slabs_free.prev;
2377 if (p == &n->slabs_free) {
2378 spin_unlock_irq(&n->list_lock);
2379 goto out;
2380 }
2381
2382 page = list_entry(p, struct page, lru);
2383 #if DEBUG
2384 BUG_ON(page->active);
2385 #endif
2386 list_del(&page->lru);
2387 /*
2388 * Safe to drop the lock. The slab is no longer linked
2389 * to the cache.
2390 */
2391 n->free_objects -= cache->num;
2392 spin_unlock_irq(&n->list_lock);
2393 slab_destroy(cache, page);
2394 nr_freed++;
2395 }
2396 out:
2397 return nr_freed;
2398 }
2399
2400 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2401 {
2402 int ret = 0;
2403 int node;
2404 struct kmem_cache_node *n;
2405
2406 drain_cpu_caches(cachep);
2407
2408 check_irq_on();
2409 for_each_kmem_cache_node(cachep, node, n) {
2410 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2411
2412 ret += !list_empty(&n->slabs_full) ||
2413 !list_empty(&n->slabs_partial);
2414 }
2415 return (ret ? 1 : 0);
2416 }
2417
2418 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2419 {
2420 int i;
2421 struct kmem_cache_node *n;
2422 int rc = __kmem_cache_shrink(cachep, false);
2423
2424 if (rc)
2425 return rc;
2426
2427 free_percpu(cachep->cpu_cache);
2428
2429 /* NUMA: free the node structures */
2430 for_each_kmem_cache_node(cachep, i, n) {
2431 kfree(n->shared);
2432 free_alien_cache(n->alien);
2433 kfree(n);
2434 cachep->node[i] = NULL;
2435 }
2436 return 0;
2437 }
2438
2439 /*
2440 * Get the memory for a slab management obj.
2441 *
2442 * For a slab cache when the slab descriptor is off-slab, the
2443 * slab descriptor can't come from the same cache which is being created,
2444 * Because if it is the case, that means we defer the creation of
2445 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2446 * And we eventually call down to __kmem_cache_create(), which
2447 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2448 * This is a "chicken-and-egg" problem.
2449 *
2450 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2451 * which are all initialized during kmem_cache_init().
2452 */
2453 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2454 struct page *page, int colour_off,
2455 gfp_t local_flags, int nodeid)
2456 {
2457 void *freelist;
2458 void *addr = page_address(page);
2459
2460 if (OFF_SLAB(cachep)) {
2461 /* Slab management obj is off-slab. */
2462 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2463 local_flags, nodeid);
2464 if (!freelist)
2465 return NULL;
2466 } else {
2467 freelist = addr + colour_off;
2468 colour_off += cachep->freelist_size;
2469 }
2470 page->active = 0;
2471 page->s_mem = addr + colour_off;
2472 return freelist;
2473 }
2474
2475 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2476 {
2477 return ((freelist_idx_t *)page->freelist)[idx];
2478 }
2479
2480 static inline void set_free_obj(struct page *page,
2481 unsigned int idx, freelist_idx_t val)
2482 {
2483 ((freelist_idx_t *)(page->freelist))[idx] = val;
2484 }
2485
2486 static void cache_init_objs(struct kmem_cache *cachep,
2487 struct page *page)
2488 {
2489 int i;
2490
2491 for (i = 0; i < cachep->num; i++) {
2492 void *objp = index_to_obj(cachep, page, i);
2493 #if DEBUG
2494 /* need to poison the objs? */
2495 if (cachep->flags & SLAB_POISON)
2496 poison_obj(cachep, objp, POISON_FREE);
2497 if (cachep->flags & SLAB_STORE_USER)
2498 *dbg_userword(cachep, objp) = NULL;
2499
2500 if (cachep->flags & SLAB_RED_ZONE) {
2501 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2502 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2503 }
2504 /*
2505 * Constructors are not allowed to allocate memory from the same
2506 * cache which they are a constructor for. Otherwise, deadlock.
2507 * They must also be threaded.
2508 */
2509 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2510 cachep->ctor(objp + obj_offset(cachep));
2511
2512 if (cachep->flags & SLAB_RED_ZONE) {
2513 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2514 slab_error(cachep, "constructor overwrote the"
2515 " end of an object");
2516 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2517 slab_error(cachep, "constructor overwrote the"
2518 " start of an object");
2519 }
2520 if ((cachep->size % PAGE_SIZE) == 0 &&
2521 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2522 kernel_map_pages(virt_to_page(objp),
2523 cachep->size / PAGE_SIZE, 0);
2524 #else
2525 if (cachep->ctor)
2526 cachep->ctor(objp);
2527 #endif
2528 set_obj_status(page, i, OBJECT_FREE);
2529 set_free_obj(page, i, i);
2530 }
2531 }
2532
2533 static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2534 {
2535 if (CONFIG_ZONE_DMA_FLAG) {
2536 if (flags & GFP_DMA)
2537 BUG_ON(!(cachep->allocflags & GFP_DMA));
2538 else
2539 BUG_ON(cachep->allocflags & GFP_DMA);
2540 }
2541 }
2542
2543 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
2544 int nodeid)
2545 {
2546 void *objp;
2547
2548 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2549 page->active++;
2550 #if DEBUG
2551 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2552 #endif
2553
2554 return objp;
2555 }
2556
2557 static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
2558 void *objp, int nodeid)
2559 {
2560 unsigned int objnr = obj_to_index(cachep, page, objp);
2561 #if DEBUG
2562 unsigned int i;
2563
2564 /* Verify that the slab belongs to the intended node */
2565 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2566
2567 /* Verify double free bug */
2568 for (i = page->active; i < cachep->num; i++) {
2569 if (get_free_obj(page, i) == objnr) {
2570 printk(KERN_ERR "slab: double free detected in cache "
2571 "'%s', objp %p\n", cachep->name, objp);
2572 BUG();
2573 }
2574 }
2575 #endif
2576 page->active--;
2577 set_free_obj(page, page->active, objnr);
2578 }
2579
2580 /*
2581 * Map pages beginning at addr to the given cache and slab. This is required
2582 * for the slab allocator to be able to lookup the cache and slab of a
2583 * virtual address for kfree, ksize, and slab debugging.
2584 */
2585 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2586 void *freelist)
2587 {
2588 page->slab_cache = cache;
2589 page->freelist = freelist;
2590 }
2591
2592 /*
2593 * Grow (by 1) the number of slabs within a cache. This is called by
2594 * kmem_cache_alloc() when there are no active objs left in a cache.
2595 */
2596 static int cache_grow(struct kmem_cache *cachep,
2597 gfp_t flags, int nodeid, struct page *page)
2598 {
2599 void *freelist;
2600 size_t offset;
2601 gfp_t local_flags;
2602 struct kmem_cache_node *n;
2603
2604 /*
2605 * Be lazy and only check for valid flags here, keeping it out of the
2606 * critical path in kmem_cache_alloc().
2607 */
2608 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2609 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2610 BUG();
2611 }
2612 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2613
2614 /* Take the node list lock to change the colour_next on this node */
2615 check_irq_off();
2616 n = get_node(cachep, nodeid);
2617 spin_lock(&n->list_lock);
2618
2619 /* Get colour for the slab, and cal the next value. */
2620 offset = n->colour_next;
2621 n->colour_next++;
2622 if (n->colour_next >= cachep->colour)
2623 n->colour_next = 0;
2624 spin_unlock(&n->list_lock);
2625
2626 offset *= cachep->colour_off;
2627
2628 if (local_flags & __GFP_WAIT)
2629 local_irq_enable();
2630
2631 /*
2632 * The test for missing atomic flag is performed here, rather than
2633 * the more obvious place, simply to reduce the critical path length
2634 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2635 * will eventually be caught here (where it matters).
2636 */
2637 kmem_flagcheck(cachep, flags);
2638
2639 /*
2640 * Get mem for the objs. Attempt to allocate a physical page from
2641 * 'nodeid'.
2642 */
2643 if (!page)
2644 page = kmem_getpages(cachep, local_flags, nodeid);
2645 if (!page)
2646 goto failed;
2647
2648 /* Get slab management. */
2649 freelist = alloc_slabmgmt(cachep, page, offset,
2650 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2651 if (!freelist)
2652 goto opps1;
2653
2654 slab_map_pages(cachep, page, freelist);
2655
2656 cache_init_objs(cachep, page);
2657
2658 if (local_flags & __GFP_WAIT)
2659 local_irq_disable();
2660 check_irq_off();
2661 spin_lock(&n->list_lock);
2662
2663 /* Make slab active. */
2664 list_add_tail(&page->lru, &(n->slabs_free));
2665 STATS_INC_GROWN(cachep);
2666 n->free_objects += cachep->num;
2667 spin_unlock(&n->list_lock);
2668 return 1;
2669 opps1:
2670 kmem_freepages(cachep, page);
2671 failed:
2672 if (local_flags & __GFP_WAIT)
2673 local_irq_disable();
2674 return 0;
2675 }
2676
2677 #if DEBUG
2678
2679 /*
2680 * Perform extra freeing checks:
2681 * - detect bad pointers.
2682 * - POISON/RED_ZONE checking
2683 */
2684 static void kfree_debugcheck(const void *objp)
2685 {
2686 if (!virt_addr_valid(objp)) {
2687 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2688 (unsigned long)objp);
2689 BUG();
2690 }
2691 }
2692
2693 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2694 {
2695 unsigned long long redzone1, redzone2;
2696
2697 redzone1 = *dbg_redzone1(cache, obj);
2698 redzone2 = *dbg_redzone2(cache, obj);
2699
2700 /*
2701 * Redzone is ok.
2702 */
2703 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2704 return;
2705
2706 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2707 slab_error(cache, "double free detected");
2708 else
2709 slab_error(cache, "memory outside object was overwritten");
2710
2711 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2712 obj, redzone1, redzone2);
2713 }
2714
2715 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2716 unsigned long caller)
2717 {
2718 unsigned int objnr;
2719 struct page *page;
2720
2721 BUG_ON(virt_to_cache(objp) != cachep);
2722
2723 objp -= obj_offset(cachep);
2724 kfree_debugcheck(objp);
2725 page = virt_to_head_page(objp);
2726
2727 if (cachep->flags & SLAB_RED_ZONE) {
2728 verify_redzone_free(cachep, objp);
2729 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2730 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2731 }
2732 if (cachep->flags & SLAB_STORE_USER)
2733 *dbg_userword(cachep, objp) = (void *)caller;
2734
2735 objnr = obj_to_index(cachep, page, objp);
2736
2737 BUG_ON(objnr >= cachep->num);
2738 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2739
2740 set_obj_status(page, objnr, OBJECT_FREE);
2741 if (cachep->flags & SLAB_POISON) {
2742 #ifdef CONFIG_DEBUG_PAGEALLOC
2743 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2744 store_stackinfo(cachep, objp, caller);
2745 kernel_map_pages(virt_to_page(objp),
2746 cachep->size / PAGE_SIZE, 0);
2747 } else {
2748 poison_obj(cachep, objp, POISON_FREE);
2749 }
2750 #else
2751 poison_obj(cachep, objp, POISON_FREE);
2752 #endif
2753 }
2754 return objp;
2755 }
2756
2757 #else
2758 #define kfree_debugcheck(x) do { } while(0)
2759 #define cache_free_debugcheck(x,objp,z) (objp)
2760 #endif
2761
2762 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2763 bool force_refill)
2764 {
2765 int batchcount;
2766 struct kmem_cache_node *n;
2767 struct array_cache *ac;
2768 int node;
2769
2770 check_irq_off();
2771 node = numa_mem_id();
2772 if (unlikely(force_refill))
2773 goto force_grow;
2774 retry:
2775 ac = cpu_cache_get(cachep);
2776 batchcount = ac->batchcount;
2777 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2778 /*
2779 * If there was little recent activity on this cache, then
2780 * perform only a partial refill. Otherwise we could generate
2781 * refill bouncing.
2782 */
2783 batchcount = BATCHREFILL_LIMIT;
2784 }
2785 n = get_node(cachep, node);
2786
2787 BUG_ON(ac->avail > 0 || !n);
2788 spin_lock(&n->list_lock);
2789
2790 /* See if we can refill from the shared array */
2791 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2792 n->shared->touched = 1;
2793 goto alloc_done;
2794 }
2795
2796 while (batchcount > 0) {
2797 struct list_head *entry;
2798 struct page *page;
2799 /* Get slab alloc is to come from. */
2800 entry = n->slabs_partial.next;
2801 if (entry == &n->slabs_partial) {
2802 n->free_touched = 1;
2803 entry = n->slabs_free.next;
2804 if (entry == &n->slabs_free)
2805 goto must_grow;
2806 }
2807
2808 page = list_entry(entry, struct page, lru);
2809 check_spinlock_acquired(cachep);
2810
2811 /*
2812 * The slab was either on partial or free list so
2813 * there must be at least one object available for
2814 * allocation.
2815 */
2816 BUG_ON(page->active >= cachep->num);
2817
2818 while (page->active < cachep->num && batchcount--) {
2819 STATS_INC_ALLOCED(cachep);
2820 STATS_INC_ACTIVE(cachep);
2821 STATS_SET_HIGH(cachep);
2822
2823 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
2824 node));
2825 }
2826
2827 /* move slabp to correct slabp list: */
2828 list_del(&page->lru);
2829 if (page->active == cachep->num)
2830 list_add(&page->lru, &n->slabs_full);
2831 else
2832 list_add(&page->lru, &n->slabs_partial);
2833 }
2834
2835 must_grow:
2836 n->free_objects -= ac->avail;
2837 alloc_done:
2838 spin_unlock(&n->list_lock);
2839
2840 if (unlikely(!ac->avail)) {
2841 int x;
2842 force_grow:
2843 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2844
2845 /* cache_grow can reenable interrupts, then ac could change. */
2846 ac = cpu_cache_get(cachep);
2847 node = numa_mem_id();
2848
2849 /* no objects in sight? abort */
2850 if (!x && (ac->avail == 0 || force_refill))
2851 return NULL;
2852
2853 if (!ac->avail) /* objects refilled by interrupt? */
2854 goto retry;
2855 }
2856 ac->touched = 1;
2857
2858 return ac_get_obj(cachep, ac, flags, force_refill);
2859 }
2860
2861 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2862 gfp_t flags)
2863 {
2864 might_sleep_if(flags & __GFP_WAIT);
2865 #if DEBUG
2866 kmem_flagcheck(cachep, flags);
2867 #endif
2868 }
2869
2870 #if DEBUG
2871 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2872 gfp_t flags, void *objp, unsigned long caller)
2873 {
2874 struct page *page;
2875
2876 if (!objp)
2877 return objp;
2878 if (cachep->flags & SLAB_POISON) {
2879 #ifdef CONFIG_DEBUG_PAGEALLOC
2880 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2881 kernel_map_pages(virt_to_page(objp),
2882 cachep->size / PAGE_SIZE, 1);
2883 else
2884 check_poison_obj(cachep, objp);
2885 #else
2886 check_poison_obj(cachep, objp);
2887 #endif
2888 poison_obj(cachep, objp, POISON_INUSE);
2889 }
2890 if (cachep->flags & SLAB_STORE_USER)
2891 *dbg_userword(cachep, objp) = (void *)caller;
2892
2893 if (cachep->flags & SLAB_RED_ZONE) {
2894 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2895 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2896 slab_error(cachep, "double free, or memory outside"
2897 " object was overwritten");
2898 printk(KERN_ERR
2899 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2900 objp, *dbg_redzone1(cachep, objp),
2901 *dbg_redzone2(cachep, objp));
2902 }
2903 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2904 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2905 }
2906
2907 page = virt_to_head_page(objp);
2908 set_obj_status(page, obj_to_index(cachep, page, objp), OBJECT_ACTIVE);
2909 objp += obj_offset(cachep);
2910 if (cachep->ctor && cachep->flags & SLAB_POISON)
2911 cachep->ctor(objp);
2912 if (ARCH_SLAB_MINALIGN &&
2913 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2914 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2915 objp, (int)ARCH_SLAB_MINALIGN);
2916 }
2917 return objp;
2918 }
2919 #else
2920 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2921 #endif
2922
2923 static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
2924 {
2925 if (unlikely(cachep == kmem_cache))
2926 return false;
2927
2928 return should_failslab(cachep->object_size, flags, cachep->flags);
2929 }
2930
2931 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2932 {
2933 void *objp;
2934 struct array_cache *ac;
2935 bool force_refill = false;
2936
2937 check_irq_off();
2938
2939 ac = cpu_cache_get(cachep);
2940 if (likely(ac->avail)) {
2941 ac->touched = 1;
2942 objp = ac_get_obj(cachep, ac, flags, false);
2943
2944 /*
2945 * Allow for the possibility all avail objects are not allowed
2946 * by the current flags
2947 */
2948 if (objp) {
2949 STATS_INC_ALLOCHIT(cachep);
2950 goto out;
2951 }
2952 force_refill = true;
2953 }
2954
2955 STATS_INC_ALLOCMISS(cachep);
2956 objp = cache_alloc_refill(cachep, flags, force_refill);
2957 /*
2958 * the 'ac' may be updated by cache_alloc_refill(),
2959 * and kmemleak_erase() requires its correct value.
2960 */
2961 ac = cpu_cache_get(cachep);
2962
2963 out:
2964 /*
2965 * To avoid a false negative, if an object that is in one of the
2966 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2967 * treat the array pointers as a reference to the object.
2968 */
2969 if (objp)
2970 kmemleak_erase(&ac->entry[ac->avail]);
2971 return objp;
2972 }
2973
2974 #ifdef CONFIG_NUMA
2975 /*
2976 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2977 *
2978 * If we are in_interrupt, then process context, including cpusets and
2979 * mempolicy, may not apply and should not be used for allocation policy.
2980 */
2981 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2982 {
2983 int nid_alloc, nid_here;
2984
2985 if (in_interrupt() || (flags & __GFP_THISNODE))
2986 return NULL;
2987 nid_alloc = nid_here = numa_mem_id();
2988 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2989 nid_alloc = cpuset_slab_spread_node();
2990 else if (current->mempolicy)
2991 nid_alloc = mempolicy_slab_node();
2992 if (nid_alloc != nid_here)
2993 return ____cache_alloc_node(cachep, flags, nid_alloc);
2994 return NULL;
2995 }
2996
2997 /*
2998 * Fallback function if there was no memory available and no objects on a
2999 * certain node and fall back is permitted. First we scan all the
3000 * available node for available objects. If that fails then we
3001 * perform an allocation without specifying a node. This allows the page
3002 * allocator to do its reclaim / fallback magic. We then insert the
3003 * slab into the proper nodelist and then allocate from it.
3004 */
3005 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3006 {
3007 struct zonelist *zonelist;
3008 gfp_t local_flags;
3009 struct zoneref *z;
3010 struct zone *zone;
3011 enum zone_type high_zoneidx = gfp_zone(flags);
3012 void *obj = NULL;
3013 int nid;
3014 unsigned int cpuset_mems_cookie;
3015
3016 if (flags & __GFP_THISNODE)
3017 return NULL;
3018
3019 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3020
3021 retry_cpuset:
3022 cpuset_mems_cookie = read_mems_allowed_begin();
3023 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3024
3025 retry:
3026 /*
3027 * Look through allowed nodes for objects available
3028 * from existing per node queues.
3029 */
3030 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3031 nid = zone_to_nid(zone);
3032
3033 if (cpuset_zone_allowed(zone, flags) &&
3034 get_node(cache, nid) &&
3035 get_node(cache, nid)->free_objects) {
3036 obj = ____cache_alloc_node(cache,
3037 gfp_exact_node(flags), nid);
3038 if (obj)
3039 break;
3040 }
3041 }
3042
3043 if (!obj) {
3044 /*
3045 * This allocation will be performed within the constraints
3046 * of the current cpuset / memory policy requirements.
3047 * We may trigger various forms of reclaim on the allowed
3048 * set and go into memory reserves if necessary.
3049 */
3050 struct page *page;
3051
3052 if (local_flags & __GFP_WAIT)
3053 local_irq_enable();
3054 kmem_flagcheck(cache, flags);
3055 page = kmem_getpages(cache, local_flags, numa_mem_id());
3056 if (local_flags & __GFP_WAIT)
3057 local_irq_disable();
3058 if (page) {
3059 /*
3060 * Insert into the appropriate per node queues
3061 */
3062 nid = page_to_nid(page);
3063 if (cache_grow(cache, flags, nid, page)) {
3064 obj = ____cache_alloc_node(cache,
3065 gfp_exact_node(flags), nid);
3066 if (!obj)
3067 /*
3068 * Another processor may allocate the
3069 * objects in the slab since we are
3070 * not holding any locks.
3071 */
3072 goto retry;
3073 } else {
3074 /* cache_grow already freed obj */
3075 obj = NULL;
3076 }
3077 }
3078 }
3079
3080 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3081 goto retry_cpuset;
3082 return obj;
3083 }
3084
3085 /*
3086 * A interface to enable slab creation on nodeid
3087 */
3088 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3089 int nodeid)
3090 {
3091 struct list_head *entry;
3092 struct page *page;
3093 struct kmem_cache_node *n;
3094 void *obj;
3095 int x;
3096
3097 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3098 n = get_node(cachep, nodeid);
3099 BUG_ON(!n);
3100
3101 retry:
3102 check_irq_off();
3103 spin_lock(&n->list_lock);
3104 entry = n->slabs_partial.next;
3105 if (entry == &n->slabs_partial) {
3106 n->free_touched = 1;
3107 entry = n->slabs_free.next;
3108 if (entry == &n->slabs_free)
3109 goto must_grow;
3110 }
3111
3112 page = list_entry(entry, struct page, lru);
3113 check_spinlock_acquired_node(cachep, nodeid);
3114
3115 STATS_INC_NODEALLOCS(cachep);
3116 STATS_INC_ACTIVE(cachep);
3117 STATS_SET_HIGH(cachep);
3118
3119 BUG_ON(page->active == cachep->num);
3120
3121 obj = slab_get_obj(cachep, page, nodeid);
3122 n->free_objects--;
3123 /* move slabp to correct slabp list: */
3124 list_del(&page->lru);
3125
3126 if (page->active == cachep->num)
3127 list_add(&page->lru, &n->slabs_full);
3128 else
3129 list_add(&page->lru, &n->slabs_partial);
3130
3131 spin_unlock(&n->list_lock);
3132 goto done;
3133
3134 must_grow:
3135 spin_unlock(&n->list_lock);
3136 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3137 if (x)
3138 goto retry;
3139
3140 return fallback_alloc(cachep, flags);
3141
3142 done:
3143 return obj;
3144 }
3145
3146 static __always_inline void *
3147 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3148 unsigned long caller)
3149 {
3150 unsigned long save_flags;
3151 void *ptr;
3152 int slab_node = numa_mem_id();
3153
3154 flags &= gfp_allowed_mask;
3155
3156 lockdep_trace_alloc(flags);
3157
3158 if (slab_should_failslab(cachep, flags))
3159 return NULL;
3160
3161 cachep = memcg_kmem_get_cache(cachep, flags);
3162
3163 cache_alloc_debugcheck_before(cachep, flags);
3164 local_irq_save(save_flags);
3165
3166 if (nodeid == NUMA_NO_NODE)
3167 nodeid = slab_node;
3168
3169 if (unlikely(!get_node(cachep, nodeid))) {
3170 /* Node not bootstrapped yet */
3171 ptr = fallback_alloc(cachep, flags);
3172 goto out;
3173 }
3174
3175 if (nodeid == slab_node) {
3176 /*
3177 * Use the locally cached objects if possible.
3178 * However ____cache_alloc does not allow fallback
3179 * to other nodes. It may fail while we still have
3180 * objects on other nodes available.
3181 */
3182 ptr = ____cache_alloc(cachep, flags);
3183 if (ptr)
3184 goto out;
3185 }
3186 /* ___cache_alloc_node can fall back to other nodes */
3187 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3188 out:
3189 local_irq_restore(save_flags);
3190 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3191 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3192 flags);
3193
3194 if (likely(ptr)) {
3195 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3196 if (unlikely(flags & __GFP_ZERO))
3197 memset(ptr, 0, cachep->object_size);
3198 }
3199
3200 memcg_kmem_put_cache(cachep);
3201 return ptr;
3202 }
3203
3204 static __always_inline void *
3205 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3206 {
3207 void *objp;
3208
3209 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3210 objp = alternate_node_alloc(cache, flags);
3211 if (objp)
3212 goto out;
3213 }
3214 objp = ____cache_alloc(cache, flags);
3215
3216 /*
3217 * We may just have run out of memory on the local node.
3218 * ____cache_alloc_node() knows how to locate memory on other nodes
3219 */
3220 if (!objp)
3221 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3222
3223 out:
3224 return objp;
3225 }
3226 #else
3227
3228 static __always_inline void *
3229 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3230 {
3231 return ____cache_alloc(cachep, flags);
3232 }
3233
3234 #endif /* CONFIG_NUMA */
3235
3236 static __always_inline void *
3237 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3238 {
3239 unsigned long save_flags;
3240 void *objp;
3241
3242 flags &= gfp_allowed_mask;
3243
3244 lockdep_trace_alloc(flags);
3245
3246 if (slab_should_failslab(cachep, flags))
3247 return NULL;
3248
3249 cachep = memcg_kmem_get_cache(cachep, flags);
3250
3251 cache_alloc_debugcheck_before(cachep, flags);
3252 local_irq_save(save_flags);
3253 objp = __do_cache_alloc(cachep, flags);
3254 local_irq_restore(save_flags);
3255 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3256 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3257 flags);
3258 prefetchw(objp);
3259
3260 if (likely(objp)) {
3261 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3262 if (unlikely(flags & __GFP_ZERO))
3263 memset(objp, 0, cachep->object_size);
3264 }
3265
3266 memcg_kmem_put_cache(cachep);
3267 return objp;
3268 }
3269
3270 /*
3271 * Caller needs to acquire correct kmem_cache_node's list_lock
3272 * @list: List of detached free slabs should be freed by caller
3273 */
3274 static void free_block(struct kmem_cache *cachep, void **objpp,
3275 int nr_objects, int node, struct list_head *list)
3276 {
3277 int i;
3278 struct kmem_cache_node *n = get_node(cachep, node);
3279
3280 for (i = 0; i < nr_objects; i++) {
3281 void *objp;
3282 struct page *page;
3283
3284 clear_obj_pfmemalloc(&objpp[i]);
3285 objp = objpp[i];
3286
3287 page = virt_to_head_page(objp);
3288 list_del(&page->lru);
3289 check_spinlock_acquired_node(cachep, node);
3290 slab_put_obj(cachep, page, objp, node);
3291 STATS_DEC_ACTIVE(cachep);
3292 n->free_objects++;
3293
3294 /* fixup slab chains */
3295 if (page->active == 0) {
3296 if (n->free_objects > n->free_limit) {
3297 n->free_objects -= cachep->num;
3298 list_add_tail(&page->lru, list);
3299 } else {
3300 list_add(&page->lru, &n->slabs_free);
3301 }
3302 } else {
3303 /* Unconditionally move a slab to the end of the
3304 * partial list on free - maximum time for the
3305 * other objects to be freed, too.
3306 */
3307 list_add_tail(&page->lru, &n->slabs_partial);
3308 }
3309 }
3310 }
3311
3312 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3313 {
3314 int batchcount;
3315 struct kmem_cache_node *n;
3316 int node = numa_mem_id();
3317 LIST_HEAD(list);
3318
3319 batchcount = ac->batchcount;
3320 #if DEBUG
3321 BUG_ON(!batchcount || batchcount > ac->avail);
3322 #endif
3323 check_irq_off();
3324 n = get_node(cachep, node);
3325 spin_lock(&n->list_lock);
3326 if (n->shared) {
3327 struct array_cache *shared_array = n->shared;
3328 int max = shared_array->limit - shared_array->avail;
3329 if (max) {
3330 if (batchcount > max)
3331 batchcount = max;
3332 memcpy(&(shared_array->entry[shared_array->avail]),
3333 ac->entry, sizeof(void *) * batchcount);
3334 shared_array->avail += batchcount;
3335 goto free_done;
3336 }
3337 }
3338
3339 free_block(cachep, ac->entry, batchcount, node, &list);
3340 free_done:
3341 #if STATS
3342 {
3343 int i = 0;
3344 struct list_head *p;
3345
3346 p = n->slabs_free.next;
3347 while (p != &(n->slabs_free)) {
3348 struct page *page;
3349
3350 page = list_entry(p, struct page, lru);
3351 BUG_ON(page->active);
3352
3353 i++;
3354 p = p->next;
3355 }
3356 STATS_SET_FREEABLE(cachep, i);
3357 }
3358 #endif
3359 spin_unlock(&n->list_lock);
3360 slabs_destroy(cachep, &list);
3361 ac->avail -= batchcount;
3362 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3363 }
3364
3365 /*
3366 * Release an obj back to its cache. If the obj has a constructed state, it must
3367 * be in this state _before_ it is released. Called with disabled ints.
3368 */
3369 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3370 unsigned long caller)
3371 {
3372 struct array_cache *ac = cpu_cache_get(cachep);
3373
3374 check_irq_off();
3375 kmemleak_free_recursive(objp, cachep->flags);
3376 objp = cache_free_debugcheck(cachep, objp, caller);
3377
3378 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3379
3380 /*
3381 * Skip calling cache_free_alien() when the platform is not numa.
3382 * This will avoid cache misses that happen while accessing slabp (which
3383 * is per page memory reference) to get nodeid. Instead use a global
3384 * variable to skip the call, which is mostly likely to be present in
3385 * the cache.
3386 */
3387 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3388 return;
3389
3390 if (ac->avail < ac->limit) {
3391 STATS_INC_FREEHIT(cachep);
3392 } else {
3393 STATS_INC_FREEMISS(cachep);
3394 cache_flusharray(cachep, ac);
3395 }
3396
3397 ac_put_obj(cachep, ac, objp);
3398 }
3399
3400 /**
3401 * kmem_cache_alloc - Allocate an object
3402 * @cachep: The cache to allocate from.
3403 * @flags: See kmalloc().
3404 *
3405 * Allocate an object from this cache. The flags are only relevant
3406 * if the cache has no available objects.
3407 */
3408 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3409 {
3410 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3411
3412 trace_kmem_cache_alloc(_RET_IP_, ret,
3413 cachep->object_size, cachep->size, flags);
3414
3415 return ret;
3416 }
3417 EXPORT_SYMBOL(kmem_cache_alloc);
3418
3419 #ifdef CONFIG_TRACING
3420 void *
3421 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3422 {
3423 void *ret;
3424
3425 ret = slab_alloc(cachep, flags, _RET_IP_);
3426
3427 trace_kmalloc(_RET_IP_, ret,
3428 size, cachep->size, flags);
3429 return ret;
3430 }
3431 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3432 #endif
3433
3434 #ifdef CONFIG_NUMA
3435 /**
3436 * kmem_cache_alloc_node - Allocate an object on the specified node
3437 * @cachep: The cache to allocate from.
3438 * @flags: See kmalloc().
3439 * @nodeid: node number of the target node.
3440 *
3441 * Identical to kmem_cache_alloc but it will allocate memory on the given
3442 * node, which can improve the performance for cpu bound structures.
3443 *
3444 * Fallback to other node is possible if __GFP_THISNODE is not set.
3445 */
3446 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3447 {
3448 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3449
3450 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3451 cachep->object_size, cachep->size,
3452 flags, nodeid);
3453
3454 return ret;
3455 }
3456 EXPORT_SYMBOL(kmem_cache_alloc_node);
3457
3458 #ifdef CONFIG_TRACING
3459 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3460 gfp_t flags,
3461 int nodeid,
3462 size_t size)
3463 {
3464 void *ret;
3465
3466 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3467
3468 trace_kmalloc_node(_RET_IP_, ret,
3469 size, cachep->size,
3470 flags, nodeid);
3471 return ret;
3472 }
3473 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3474 #endif
3475
3476 static __always_inline void *
3477 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3478 {
3479 struct kmem_cache *cachep;
3480
3481 cachep = kmalloc_slab(size, flags);
3482 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3483 return cachep;
3484 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3485 }
3486
3487 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3488 {
3489 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3490 }
3491 EXPORT_SYMBOL(__kmalloc_node);
3492
3493 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3494 int node, unsigned long caller)
3495 {
3496 return __do_kmalloc_node(size, flags, node, caller);
3497 }
3498 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3499 #endif /* CONFIG_NUMA */
3500
3501 /**
3502 * __do_kmalloc - allocate memory
3503 * @size: how many bytes of memory are required.
3504 * @flags: the type of memory to allocate (see kmalloc).
3505 * @caller: function caller for debug tracking of the caller
3506 */
3507 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3508 unsigned long caller)
3509 {
3510 struct kmem_cache *cachep;
3511 void *ret;
3512
3513 cachep = kmalloc_slab(size, flags);
3514 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3515 return cachep;
3516 ret = slab_alloc(cachep, flags, caller);
3517
3518 trace_kmalloc(caller, ret,
3519 size, cachep->size, flags);
3520
3521 return ret;
3522 }
3523
3524 void *__kmalloc(size_t size, gfp_t flags)
3525 {
3526 return __do_kmalloc(size, flags, _RET_IP_);
3527 }
3528 EXPORT_SYMBOL(__kmalloc);
3529
3530 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3531 {
3532 return __do_kmalloc(size, flags, caller);
3533 }
3534 EXPORT_SYMBOL(__kmalloc_track_caller);
3535
3536 /**
3537 * kmem_cache_free - Deallocate an object
3538 * @cachep: The cache the allocation was from.
3539 * @objp: The previously allocated object.
3540 *
3541 * Free an object which was previously allocated from this
3542 * cache.
3543 */
3544 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3545 {
3546 unsigned long flags;
3547 cachep = cache_from_obj(cachep, objp);
3548 if (!cachep)
3549 return;
3550
3551 local_irq_save(flags);
3552 debug_check_no_locks_freed(objp, cachep->object_size);
3553 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3554 debug_check_no_obj_freed(objp, cachep->object_size);
3555 __cache_free(cachep, objp, _RET_IP_);
3556 local_irq_restore(flags);
3557
3558 trace_kmem_cache_free(_RET_IP_, objp);
3559 }
3560 EXPORT_SYMBOL(kmem_cache_free);
3561
3562 /**
3563 * kfree - free previously allocated memory
3564 * @objp: pointer returned by kmalloc.
3565 *
3566 * If @objp is NULL, no operation is performed.
3567 *
3568 * Don't free memory not originally allocated by kmalloc()
3569 * or you will run into trouble.
3570 */
3571 void kfree(const void *objp)
3572 {
3573 struct kmem_cache *c;
3574 unsigned long flags;
3575
3576 trace_kfree(_RET_IP_, objp);
3577
3578 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3579 return;
3580 local_irq_save(flags);
3581 kfree_debugcheck(objp);
3582 c = virt_to_cache(objp);
3583 debug_check_no_locks_freed(objp, c->object_size);
3584
3585 debug_check_no_obj_freed(objp, c->object_size);
3586 __cache_free(c, (void *)objp, _RET_IP_);
3587 local_irq_restore(flags);
3588 }
3589 EXPORT_SYMBOL(kfree);
3590
3591 /*
3592 * This initializes kmem_cache_node or resizes various caches for all nodes.
3593 */
3594 static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3595 {
3596 int node;
3597 struct kmem_cache_node *n;
3598 struct array_cache *new_shared;
3599 struct alien_cache **new_alien = NULL;
3600
3601 for_each_online_node(node) {
3602
3603 if (use_alien_caches) {
3604 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3605 if (!new_alien)
3606 goto fail;
3607 }
3608
3609 new_shared = NULL;
3610 if (cachep->shared) {
3611 new_shared = alloc_arraycache(node,
3612 cachep->shared*cachep->batchcount,
3613 0xbaadf00d, gfp);
3614 if (!new_shared) {
3615 free_alien_cache(new_alien);
3616 goto fail;
3617 }
3618 }
3619
3620 n = get_node(cachep, node);
3621 if (n) {
3622 struct array_cache *shared = n->shared;
3623 LIST_HEAD(list);
3624
3625 spin_lock_irq(&n->list_lock);
3626
3627 if (shared)
3628 free_block(cachep, shared->entry,
3629 shared->avail, node, &list);
3630
3631 n->shared = new_shared;
3632 if (!n->alien) {
3633 n->alien = new_alien;
3634 new_alien = NULL;
3635 }
3636 n->free_limit = (1 + nr_cpus_node(node)) *
3637 cachep->batchcount + cachep->num;
3638 spin_unlock_irq(&n->list_lock);
3639 slabs_destroy(cachep, &list);
3640 kfree(shared);
3641 free_alien_cache(new_alien);
3642 continue;
3643 }
3644 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3645 if (!n) {
3646 free_alien_cache(new_alien);
3647 kfree(new_shared);
3648 goto fail;
3649 }
3650
3651 kmem_cache_node_init(n);
3652 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3653 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3654 n->shared = new_shared;
3655 n->alien = new_alien;
3656 n->free_limit = (1 + nr_cpus_node(node)) *
3657 cachep->batchcount + cachep->num;
3658 cachep->node[node] = n;
3659 }
3660 return 0;
3661
3662 fail:
3663 if (!cachep->list.next) {
3664 /* Cache is not active yet. Roll back what we did */
3665 node--;
3666 while (node >= 0) {
3667 n = get_node(cachep, node);
3668 if (n) {
3669 kfree(n->shared);
3670 free_alien_cache(n->alien);
3671 kfree(n);
3672 cachep->node[node] = NULL;
3673 }
3674 node--;
3675 }
3676 }
3677 return -ENOMEM;
3678 }
3679
3680 /* Always called with the slab_mutex held */
3681 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3682 int batchcount, int shared, gfp_t gfp)
3683 {
3684 struct array_cache __percpu *cpu_cache, *prev;
3685 int cpu;
3686
3687 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3688 if (!cpu_cache)
3689 return -ENOMEM;
3690
3691 prev = cachep->cpu_cache;
3692 cachep->cpu_cache = cpu_cache;
3693 kick_all_cpus_sync();
3694
3695 check_irq_on();
3696 cachep->batchcount = batchcount;
3697 cachep->limit = limit;
3698 cachep->shared = shared;
3699
3700 if (!prev)
3701 goto alloc_node;
3702
3703 for_each_online_cpu(cpu) {
3704 LIST_HEAD(list);
3705 int node;
3706 struct kmem_cache_node *n;
3707 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3708
3709 node = cpu_to_mem(cpu);
3710 n = get_node(cachep, node);
3711 spin_lock_irq(&n->list_lock);
3712 free_block(cachep, ac->entry, ac->avail, node, &list);
3713 spin_unlock_irq(&n->list_lock);
3714 slabs_destroy(cachep, &list);
3715 }
3716 free_percpu(prev);
3717
3718 alloc_node:
3719 return alloc_kmem_cache_node(cachep, gfp);
3720 }
3721
3722 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3723 int batchcount, int shared, gfp_t gfp)
3724 {
3725 int ret;
3726 struct kmem_cache *c;
3727
3728 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3729
3730 if (slab_state < FULL)
3731 return ret;
3732
3733 if ((ret < 0) || !is_root_cache(cachep))
3734 return ret;
3735
3736 lockdep_assert_held(&slab_mutex);
3737 for_each_memcg_cache(c, cachep) {
3738 /* return value determined by the root cache only */
3739 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3740 }
3741
3742 return ret;
3743 }
3744
3745 /* Called with slab_mutex held always */
3746 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3747 {
3748 int err;
3749 int limit = 0;
3750 int shared = 0;
3751 int batchcount = 0;
3752
3753 if (!is_root_cache(cachep)) {
3754 struct kmem_cache *root = memcg_root_cache(cachep);
3755 limit = root->limit;
3756 shared = root->shared;
3757 batchcount = root->batchcount;
3758 }
3759
3760 if (limit && shared && batchcount)
3761 goto skip_setup;
3762 /*
3763 * The head array serves three purposes:
3764 * - create a LIFO ordering, i.e. return objects that are cache-warm
3765 * - reduce the number of spinlock operations.
3766 * - reduce the number of linked list operations on the slab and
3767 * bufctl chains: array operations are cheaper.
3768 * The numbers are guessed, we should auto-tune as described by
3769 * Bonwick.
3770 */
3771 if (cachep->size > 131072)
3772 limit = 1;
3773 else if (cachep->size > PAGE_SIZE)
3774 limit = 8;
3775 else if (cachep->size > 1024)
3776 limit = 24;
3777 else if (cachep->size > 256)
3778 limit = 54;
3779 else
3780 limit = 120;
3781
3782 /*
3783 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3784 * allocation behaviour: Most allocs on one cpu, most free operations
3785 * on another cpu. For these cases, an efficient object passing between
3786 * cpus is necessary. This is provided by a shared array. The array
3787 * replaces Bonwick's magazine layer.
3788 * On uniprocessor, it's functionally equivalent (but less efficient)
3789 * to a larger limit. Thus disabled by default.
3790 */
3791 shared = 0;
3792 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3793 shared = 8;
3794
3795 #if DEBUG
3796 /*
3797 * With debugging enabled, large batchcount lead to excessively long
3798 * periods with disabled local interrupts. Limit the batchcount
3799 */
3800 if (limit > 32)
3801 limit = 32;
3802 #endif
3803 batchcount = (limit + 1) / 2;
3804 skip_setup:
3805 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3806 if (err)
3807 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3808 cachep->name, -err);
3809 return err;
3810 }
3811
3812 /*
3813 * Drain an array if it contains any elements taking the node lock only if
3814 * necessary. Note that the node listlock also protects the array_cache
3815 * if drain_array() is used on the shared array.
3816 */
3817 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3818 struct array_cache *ac, int force, int node)
3819 {
3820 LIST_HEAD(list);
3821 int tofree;
3822
3823 if (!ac || !ac->avail)
3824 return;
3825 if (ac->touched && !force) {
3826 ac->touched = 0;
3827 } else {
3828 spin_lock_irq(&n->list_lock);
3829 if (ac->avail) {
3830 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3831 if (tofree > ac->avail)
3832 tofree = (ac->avail + 1) / 2;
3833 free_block(cachep, ac->entry, tofree, node, &list);
3834 ac->avail -= tofree;
3835 memmove(ac->entry, &(ac->entry[tofree]),
3836 sizeof(void *) * ac->avail);
3837 }
3838 spin_unlock_irq(&n->list_lock);
3839 slabs_destroy(cachep, &list);
3840 }
3841 }
3842
3843 /**
3844 * cache_reap - Reclaim memory from caches.
3845 * @w: work descriptor
3846 *
3847 * Called from workqueue/eventd every few seconds.
3848 * Purpose:
3849 * - clear the per-cpu caches for this CPU.
3850 * - return freeable pages to the main free memory pool.
3851 *
3852 * If we cannot acquire the cache chain mutex then just give up - we'll try
3853 * again on the next iteration.
3854 */
3855 static void cache_reap(struct work_struct *w)
3856 {
3857 struct kmem_cache *searchp;
3858 struct kmem_cache_node *n;
3859 int node = numa_mem_id();
3860 struct delayed_work *work = to_delayed_work(w);
3861
3862 if (!mutex_trylock(&slab_mutex))
3863 /* Give up. Setup the next iteration. */
3864 goto out;
3865
3866 list_for_each_entry(searchp, &slab_caches, list) {
3867 check_irq_on();
3868
3869 /*
3870 * We only take the node lock if absolutely necessary and we
3871 * have established with reasonable certainty that
3872 * we can do some work if the lock was obtained.
3873 */
3874 n = get_node(searchp, node);
3875
3876 reap_alien(searchp, n);
3877
3878 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3879
3880 /*
3881 * These are racy checks but it does not matter
3882 * if we skip one check or scan twice.
3883 */
3884 if (time_after(n->next_reap, jiffies))
3885 goto next;
3886
3887 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3888
3889 drain_array(searchp, n, n->shared, 0, node);
3890
3891 if (n->free_touched)
3892 n->free_touched = 0;
3893 else {
3894 int freed;
3895
3896 freed = drain_freelist(searchp, n, (n->free_limit +
3897 5 * searchp->num - 1) / (5 * searchp->num));
3898 STATS_ADD_REAPED(searchp, freed);
3899 }
3900 next:
3901 cond_resched();
3902 }
3903 check_irq_on();
3904 mutex_unlock(&slab_mutex);
3905 next_reap_node();
3906 out:
3907 /* Set up the next iteration */
3908 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3909 }
3910
3911 #ifdef CONFIG_SLABINFO
3912 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3913 {
3914 struct page *page;
3915 unsigned long active_objs;
3916 unsigned long num_objs;
3917 unsigned long active_slabs = 0;
3918 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3919 const char *name;
3920 char *error = NULL;
3921 int node;
3922 struct kmem_cache_node *n;
3923
3924 active_objs = 0;
3925 num_slabs = 0;
3926 for_each_kmem_cache_node(cachep, node, n) {
3927
3928 check_irq_on();
3929 spin_lock_irq(&n->list_lock);
3930
3931 list_for_each_entry(page, &n->slabs_full, lru) {
3932 if (page->active != cachep->num && !error)
3933 error = "slabs_full accounting error";
3934 active_objs += cachep->num;
3935 active_slabs++;
3936 }
3937 list_for_each_entry(page, &n->slabs_partial, lru) {
3938 if (page->active == cachep->num && !error)
3939 error = "slabs_partial accounting error";
3940 if (!page->active && !error)
3941 error = "slabs_partial accounting error";
3942 active_objs += page->active;
3943 active_slabs++;
3944 }
3945 list_for_each_entry(page, &n->slabs_free, lru) {
3946 if (page->active && !error)
3947 error = "slabs_free accounting error";
3948 num_slabs++;
3949 }
3950 free_objects += n->free_objects;
3951 if (n->shared)
3952 shared_avail += n->shared->avail;
3953
3954 spin_unlock_irq(&n->list_lock);
3955 }
3956 num_slabs += active_slabs;
3957 num_objs = num_slabs * cachep->num;
3958 if (num_objs - active_objs != free_objects && !error)
3959 error = "free_objects accounting error";
3960
3961 name = cachep->name;
3962 if (error)
3963 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
3964
3965 sinfo->active_objs = active_objs;
3966 sinfo->num_objs = num_objs;
3967 sinfo->active_slabs = active_slabs;
3968 sinfo->num_slabs = num_slabs;
3969 sinfo->shared_avail = shared_avail;
3970 sinfo->limit = cachep->limit;
3971 sinfo->batchcount = cachep->batchcount;
3972 sinfo->shared = cachep->shared;
3973 sinfo->objects_per_slab = cachep->num;
3974 sinfo->cache_order = cachep->gfporder;
3975 }
3976
3977 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
3978 {
3979 #if STATS
3980 { /* node stats */
3981 unsigned long high = cachep->high_mark;
3982 unsigned long allocs = cachep->num_allocations;
3983 unsigned long grown = cachep->grown;
3984 unsigned long reaped = cachep->reaped;
3985 unsigned long errors = cachep->errors;
3986 unsigned long max_freeable = cachep->max_freeable;
3987 unsigned long node_allocs = cachep->node_allocs;
3988 unsigned long node_frees = cachep->node_frees;
3989 unsigned long overflows = cachep->node_overflow;
3990
3991 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
3992 "%4lu %4lu %4lu %4lu %4lu",
3993 allocs, high, grown,
3994 reaped, errors, max_freeable, node_allocs,
3995 node_frees, overflows);
3996 }
3997 /* cpu stats */
3998 {
3999 unsigned long allochit = atomic_read(&cachep->allochit);
4000 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4001 unsigned long freehit = atomic_read(&cachep->freehit);
4002 unsigned long freemiss = atomic_read(&cachep->freemiss);
4003
4004 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4005 allochit, allocmiss, freehit, freemiss);
4006 }
4007 #endif
4008 }
4009
4010 #define MAX_SLABINFO_WRITE 128
4011 /**
4012 * slabinfo_write - Tuning for the slab allocator
4013 * @file: unused
4014 * @buffer: user buffer
4015 * @count: data length
4016 * @ppos: unused
4017 */
4018 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4019 size_t count, loff_t *ppos)
4020 {
4021 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4022 int limit, batchcount, shared, res;
4023 struct kmem_cache *cachep;
4024
4025 if (count > MAX_SLABINFO_WRITE)
4026 return -EINVAL;
4027 if (copy_from_user(&kbuf, buffer, count))
4028 return -EFAULT;
4029 kbuf[MAX_SLABINFO_WRITE] = '\0';
4030
4031 tmp = strchr(kbuf, ' ');
4032 if (!tmp)
4033 return -EINVAL;
4034 *tmp = '\0';
4035 tmp++;
4036 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4037 return -EINVAL;
4038
4039 /* Find the cache in the chain of caches. */
4040 mutex_lock(&slab_mutex);
4041 res = -EINVAL;
4042 list_for_each_entry(cachep, &slab_caches, list) {
4043 if (!strcmp(cachep->name, kbuf)) {
4044 if (limit < 1 || batchcount < 1 ||
4045 batchcount > limit || shared < 0) {
4046 res = 0;
4047 } else {
4048 res = do_tune_cpucache(cachep, limit,
4049 batchcount, shared,
4050 GFP_KERNEL);
4051 }
4052 break;
4053 }
4054 }
4055 mutex_unlock(&slab_mutex);
4056 if (res >= 0)
4057 res = count;
4058 return res;
4059 }
4060
4061 #ifdef CONFIG_DEBUG_SLAB_LEAK
4062
4063 static inline int add_caller(unsigned long *n, unsigned long v)
4064 {
4065 unsigned long *p;
4066 int l;
4067 if (!v)
4068 return 1;
4069 l = n[1];
4070 p = n + 2;
4071 while (l) {
4072 int i = l/2;
4073 unsigned long *q = p + 2 * i;
4074 if (*q == v) {
4075 q[1]++;
4076 return 1;
4077 }
4078 if (*q > v) {
4079 l = i;
4080 } else {
4081 p = q + 2;
4082 l -= i + 1;
4083 }
4084 }
4085 if (++n[1] == n[0])
4086 return 0;
4087 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4088 p[0] = v;
4089 p[1] = 1;
4090 return 1;
4091 }
4092
4093 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4094 struct page *page)
4095 {
4096 void *p;
4097 int i;
4098
4099 if (n[0] == n[1])
4100 return;
4101 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4102 if (get_obj_status(page, i) != OBJECT_ACTIVE)
4103 continue;
4104
4105 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4106 return;
4107 }
4108 }
4109
4110 static void show_symbol(struct seq_file *m, unsigned long address)
4111 {
4112 #ifdef CONFIG_KALLSYMS
4113 unsigned long offset, size;
4114 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4115
4116 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4117 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4118 if (modname[0])
4119 seq_printf(m, " [%s]", modname);
4120 return;
4121 }
4122 #endif
4123 seq_printf(m, "%p", (void *)address);
4124 }
4125
4126 static int leaks_show(struct seq_file *m, void *p)
4127 {
4128 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4129 struct page *page;
4130 struct kmem_cache_node *n;
4131 const char *name;
4132 unsigned long *x = m->private;
4133 int node;
4134 int i;
4135
4136 if (!(cachep->flags & SLAB_STORE_USER))
4137 return 0;
4138 if (!(cachep->flags & SLAB_RED_ZONE))
4139 return 0;
4140
4141 /* OK, we can do it */
4142
4143 x[1] = 0;
4144
4145 for_each_kmem_cache_node(cachep, node, n) {
4146
4147 check_irq_on();
4148 spin_lock_irq(&n->list_lock);
4149
4150 list_for_each_entry(page, &n->slabs_full, lru)
4151 handle_slab(x, cachep, page);
4152 list_for_each_entry(page, &n->slabs_partial, lru)
4153 handle_slab(x, cachep, page);
4154 spin_unlock_irq(&n->list_lock);
4155 }
4156 name = cachep->name;
4157 if (x[0] == x[1]) {
4158 /* Increase the buffer size */
4159 mutex_unlock(&slab_mutex);
4160 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4161 if (!m->private) {
4162 /* Too bad, we are really out */
4163 m->private = x;
4164 mutex_lock(&slab_mutex);
4165 return -ENOMEM;
4166 }
4167 *(unsigned long *)m->private = x[0] * 2;
4168 kfree(x);
4169 mutex_lock(&slab_mutex);
4170 /* Now make sure this entry will be retried */
4171 m->count = m->size;
4172 return 0;
4173 }
4174 for (i = 0; i < x[1]; i++) {
4175 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4176 show_symbol(m, x[2*i+2]);
4177 seq_putc(m, '\n');
4178 }
4179
4180 return 0;
4181 }
4182
4183 static const struct seq_operations slabstats_op = {
4184 .start = slab_start,
4185 .next = slab_next,
4186 .stop = slab_stop,
4187 .show = leaks_show,
4188 };
4189
4190 static int slabstats_open(struct inode *inode, struct file *file)
4191 {
4192 unsigned long *n;
4193
4194 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4195 if (!n)
4196 return -ENOMEM;
4197
4198 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4199
4200 return 0;
4201 }
4202
4203 static const struct file_operations proc_slabstats_operations = {
4204 .open = slabstats_open,
4205 .read = seq_read,
4206 .llseek = seq_lseek,
4207 .release = seq_release_private,
4208 };
4209 #endif
4210
4211 static int __init slab_proc_init(void)
4212 {
4213 #ifdef CONFIG_DEBUG_SLAB_LEAK
4214 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4215 #endif
4216 return 0;
4217 }
4218 module_init(slab_proc_init);
4219 #endif
4220
4221 /**
4222 * ksize - get the actual amount of memory allocated for a given object
4223 * @objp: Pointer to the object
4224 *
4225 * kmalloc may internally round up allocations and return more memory
4226 * than requested. ksize() can be used to determine the actual amount of
4227 * memory allocated. The caller may use this additional memory, even though
4228 * a smaller amount of memory was initially specified with the kmalloc call.
4229 * The caller must guarantee that objp points to a valid object previously
4230 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4231 * must not be freed during the duration of the call.
4232 */
4233 size_t ksize(const void *objp)
4234 {
4235 BUG_ON(!objp);
4236 if (unlikely(objp == ZERO_SIZE_PTR))
4237 return 0;
4238
4239 return virt_to_cache(objp)->object_size;
4240 }
4241 EXPORT_SYMBOL(ksize);