Loading...
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/__KEEPIDENTS__B.h>
90#include <linux/__KEEPIDENTS__C.h>
91#include <linux/__KEEPIDENTS__D.h>
92#include <linux/__KEEPIDENTS__E.h>
93#include <linux/__KEEPIDENTS__F.h>
94#include <linux/__KEEPIDENTS__G.h>
95#include <linux/__KEEPIDENTS__H.h>
96#include <linux/__KEEPIDENTS__I.h>
97#include <linux/__KEEPIDENTS__J.h>
98#include <linux/proc_fs.h>
99#include <linux/__KEEPIDENTS__BA.h>
100#include <linux/__KEEPIDENTS__BB.h>
101#include <linux/__KEEPIDENTS__BC.h>
102#include <linux/cpu.h>
103#include <linux/__KEEPIDENTS__BD.h>
104#include <linux/__KEEPIDENTS__BE.h>
105#include <linux/rcupdate.h>
106#include <linux/__KEEPIDENTS__BF.h>
107#include <linux/__KEEPIDENTS__BG.h>
108#include <linux/__KEEPIDENTS__BH.h>
109#include <linux/kmemleak.h>
110#include <linux/__KEEPIDENTS__BI.h>
111#include <linux/__KEEPIDENTS__BJ.h>
112#include <linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
113#include <linux/__KEEPIDENTS__CC.h>
114#include <linux/reciprocal_div.h>
115#include <linux/debugobjects.h>
116#include <linux/kmemcheck.h>
117#include <linux/__KEEPIDENTS__CD.h>
118#include <linux/__KEEPIDENTS__CE.h>
119
120#include <net/__KEEPIDENTS__CF.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
164typedef unsigned char freelist_idx_t;
165#else
166typedef 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 */
175static 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 */
189struct array_cache {
190 unsigned int avail;
191 unsigned int limit;
192 unsigned int batchcount;
193 unsigned int touched;
194 spinlock_t lock;
195 void *entry[]; /*
196 * Must have this definition in here for the proper
197 * alignment of array_cache. Also simplifies accessing
198 * the entries.
199 *
200 * Entries should not be directly dereferenced as
201 * entries belonging to slabs marked pfmemalloc will
202 * have the lower bits set SLAB_OBJ_PFMEMALLOC
203 */
204};
205
206#define SLAB_OBJ_PFMEMALLOC 1
207static inline bool is_obj_pfmemalloc(void *objp)
208{
209 return (unsigned long)objp & SLAB_OBJ_PFMEMALLOC;
210}
211
212static inline void set_obj_pfmemalloc(void **objp)
213{
214 *objp = (void *)((unsigned long)*objp | SLAB_OBJ_PFMEMALLOC);
215 return;
216}
217
218static inline void clear_obj_pfmemalloc(void **objp)
219{
220 *objp = (void *)((unsigned long)*objp & ~SLAB_OBJ_PFMEMALLOC);
221}
222
223/*
224 * bootstrap: The caches do not work without cpuarrays anymore, but the
225 * cpuarrays are allocated from the generic caches...
226 */
227#define BOOT_CPUCACHE_ENTRIES 1
228struct arraycache_init {
229 struct array_cache cache;
230 void *entries[BOOT_CPUCACHE_ENTRIES];
231};
232
233/*
234 * Need this for bootstrapping a per node allocator.
235 */
236#define NUM_INIT_LISTS (3 * MAX_NUMNODES)
237static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
238#define CACHE_CACHE 0
239#define SIZE_AC MAX_NUMNODES
240#define SIZE_NODE (2 * MAX_NUMNODES)
241
242static int drain_freelist(struct kmem_cache *cache,
243 struct kmem_cache_node *n, int tofree);
244static void free_block(struct kmem_cache *cachep, void **objpp, int len,
245 int node);
246static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
247static void cache_reap(struct work_struct *unused);
248
249static int slab_early_init = 1;
250
251#define INDEX_AC kmalloc_index(sizeof(struct arraycache_init))
252#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
253
254static void kmem_cache_node_init(struct kmem_cache_node *parent)
255{
256 INIT_LIST_HEAD(&parent->slabs_full);
257 INIT_LIST_HEAD(&parent->slabs_partial);
258 INIT_LIST_HEAD(&parent->slabs_free);
259 parent->shared = NULL;
260 parent->alien = NULL;
261 parent->colour_next = 0;
262 spin_lock_init(&parent->list_lock);
263 parent->free_objects = 0;
264 parent->free_touched = 0;
265}
266
267#define MAKE_LIST(cachep, listp, slab, nodeid) \
268 do { \
269 INIT_LIST_HEAD(listp); \
270 list_splice(&(cachep->node[nodeid]->slab), listp); \
271 } while (0)
272
273#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
274 do { \
275 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
276 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
277 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
278 } while (0)
279
280#define CFLGS_OFF_SLAB (0x80000000UL)
281#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
282
283#define BATCHREFILL_LIMIT 16
284/*
285 * Optimization question: fewer reaps means less probability for unnessary
286 * cpucache drain/refill cycles.
287 *
288 * OTOH the cpuarrays can contain lots of objects,
289 * which could lock up otherwise freeable slabs.
290 */
291#define REAPTIMEOUT_AC (2*HZ)
292#define REAPTIMEOUT_NODE (4*HZ)
293
294#if STATS
295#define STATS_INC_ACTIVE(x) ((x)->num_active++)
296#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
297#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
298#define STATS_INC_GROWN(x) ((x)->grown++)
299#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
300#define STATS_SET_HIGH(x) \
301 do { \
302 if ((x)->num_active > (x)->high_mark) \
303 (x)->high_mark = (x)->num_active; \
304 } while (0)
305#define STATS_INC_ERR(x) ((x)->errors++)
306#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
307#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
308#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
309#define STATS_SET_FREEABLE(x, i) \
310 do { \
311 if ((x)->max_freeable < i) \
312 (x)->max_freeable = i; \
313 } while (0)
314#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
315#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
316#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
317#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
318#else
319#define STATS_INC_ACTIVE(x) do { } while (0)
320#define STATS_DEC_ACTIVE(x) do { } while (0)
321#define STATS_INC_ALLOCED(x) do { } while (0)
322#define STATS_INC_GROWN(x) do { } while (0)
323#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
324#define STATS_SET_HIGH(x) do { } while (0)
325#define STATS_INC_ERR(x) do { } while (0)
326#define STATS_INC_NODEALLOCS(x) do { } while (0)
327#define STATS_INC_NODEFREES(x) do { } while (0)
328#define STATS_INC_ACOVERFLOW(x) do { } while (0)
329#define STATS_SET_FREEABLE(x, i) do { } while (0)
330#define STATS_INC_ALLOCHIT(x) do { } while (0)
331#define STATS_INC_ALLOCMISS(x) do { } while (0)
332#define STATS_INC_FREEHIT(x) do { } while (0)
333#define STATS_INC_FREEMISS(x) do { } while (0)
334#endif
335
336#if DEBUG
337
338/*
339 * memory layout of objects:
340 * 0 : objp
341 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
342 * the end of an object is aligned with the end of the real
343 * allocation. Catches writes behind the end of the allocation.
344 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
345 * redzone word.
346 * cachep->obj_offset: The real object.
347 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
348 * cachep->size - 1* BYTES_PER_WORD: last caller address
349 * [BYTES_PER_WORD long]
350 */
351static int obj_offset(struct kmem_cache *cachep)
352{
353 return cachep->obj_offset;
354}
355
356static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
357{
358 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
359 return (unsigned long long*) (objp + obj_offset(cachep) -
360 sizeof(unsigned long long));
361}
362
363static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
364{
365 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
366 if (cachep->flags & SLAB_STORE_USER)
367 return (unsigned long long *)(objp + cachep->size -
368 sizeof(unsigned long long) -
369 REDZONE_ALIGN);
370 return (unsigned long long *) (objp + cachep->size -
371 sizeof(unsigned long long));
372}
373
374static void **dbg_userword(struct kmem_cache *cachep, void *objp)
375{
376 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
377 return (void **)(objp + cachep->size - BYTES_PER_WORD);
378}
379
380#else
381
382#define obj_offset(x) 0
383#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
384#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
385#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
386
387#endif
388
389/*
390 * Do not go above this order unless 0 objects fit into the slab or
391 * overridden on the command line.
392 */
393#define SLAB_MAX_ORDER_HI 1
394#define SLAB_MAX_ORDER_LO 0
395static int slab_max_order = SLAB_MAX_ORDER_LO;
396static bool slab_max_order_set __initdata;
397
398static inline struct kmem_cache *virt_to_cache(const void *obj)
399{
400 struct page *page = virt_to_head_page(obj);
401 return page->slab_cache;
402}
403
404static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
405 unsigned int idx)
406{
407 return page->s_mem + cache->size * idx;
408}
409
410/*
411 * We want to avoid an expensive divide : (offset / cache->size)
412 * Using the fact that size is a constant for a particular cache,
413 * we can replace (offset / cache->size) by
414 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
415 */
416static inline unsigned int obj_to_index(const struct kmem_cache *cache,
417 const struct page *page, void *obj)
418{
419 u32 offset = (obj - page->s_mem);
420 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
421}
422
423static struct arraycache_init initarray_generic =
424 { {0, BOOT_CPUCACHE_ENTRIES, 1, 0} };
425
426/* internal cache of cache description objs */
427static struct kmem_cache kmem_cache_boot = {
428 .batchcount = 1,
429 .limit = BOOT_CPUCACHE_ENTRIES,
430 .shared = 1,
431 .size = sizeof(struct kmem_cache),
432 .name = "kmem_cache",
433};
434
435#define BAD_ALIEN_MAGIC 0x01020304ul
436
437#ifdef CONFIG_LOCKDEP
438
439/*
440 * Slab sometimes uses the kmalloc slabs to store the slab headers
441 * for other slabs "off slab".
442 * The locking for this is tricky in that it nests within the locks
443 * of all other slabs in a few places; to deal with this special
444 * locking we put on-slab caches into a separate lock-class.
445 *
446 * We set lock class for alien array caches which are up during init.
447 * The lock annotation will be lost if all cpus of a node goes down and
448 * then comes back up during hotplug
449 */
450static struct lock_class_key on_slab_l3_key;
451static struct lock_class_key on_slab_alc_key;
452
453static struct lock_class_key debugobj_l3_key;
454static struct lock_class_key debugobj_alc_key;
455
456static void slab_set_lock_classes(struct kmem_cache *cachep,
457 struct lock_class_key *l3_key, struct lock_class_key *alc_key,
458 int q)
459{
460 struct array_cache **alc;
461 struct kmem_cache_node *n;
462 int r;
463
464 n = cachep->node[q];
465 if (!n)
466 return;
467
468 lockdep_set_class(&n->list_lock, l3_key);
469 alc = n->alien;
470 /*
471 * FIXME: This check for BAD_ALIEN_MAGIC
472 * should go away when common slab code is taught to
473 * work even without alien caches.
474 * Currently, non NUMA code returns BAD_ALIEN_MAGIC
475 * for alloc_alien_cache,
476 */
477 if (!alc || (unsigned long)alc == BAD_ALIEN_MAGIC)
478 return;
479 for_each_node(r) {
480 if (alc[r])
481 lockdep_set_class(&alc[r]->lock, alc_key);
482 }
483}
484
485static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
486{
487 slab_set_lock_classes(cachep, &debugobj_l3_key, &debugobj_alc_key, node);
488}
489
490static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
491{
492 int node;
493
494 for_each_online_node(node)
495 slab_set_debugobj_lock_classes_node(cachep, node);
496}
497
498static void init_node_lock_keys(int q)
499{
500 int i;
501
502 if (slab_state < UP)
503 return;
504
505 for (i = 1; i <= KMALLOC_SHIFT_HIGH; i++) {
506 struct kmem_cache_node *n;
507 struct kmem_cache *cache = kmalloc_caches[i];
508
509 if (!cache)
510 continue;
511
512 n = cache->node[q];
513 if (!n || OFF_SLAB(cache))
514 continue;
515
516 slab_set_lock_classes(cache, &on_slab_l3_key,
517 &on_slab_alc_key, q);
518 }
519}
520
521static void on_slab_lock_classes_node(struct kmem_cache *cachep, int q)
522{
523 if (!cachep->node[q])
524 return;
525
526 slab_set_lock_classes(cachep, &on_slab_l3_key,
527 &on_slab_alc_key, q);
528}
529
530static inline void on_slab_lock_classes(struct kmem_cache *cachep)
531{
532 int node;
533
534 VM_BUG_ON(OFF_SLAB(cachep));
535 for_each_node(node)
536 on_slab_lock_classes_node(cachep, node);
537}
538
539static inline void init_lock_keys(void)
540{
541 int node;
542
543 for_each_node(node)
544 init_node_lock_keys(node);
545}
546#else
547static void init_node_lock_keys(int q)
548{
549}
550
551static inline void init_lock_keys(void)
552{
553}
554
555static inline void on_slab_lock_classes(struct kmem_cache *cachep)
556{
557}
558
559static inline void on_slab_lock_classes_node(struct kmem_cache *cachep, int node)
560{
561}
562
563static void slab_set_debugobj_lock_classes_node(struct kmem_cache *cachep, int node)
564{
565}
566
567static void slab_set_debugobj_lock_classes(struct kmem_cache *cachep)
568{
569}
570#endif
571
572static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
573
574static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
575{
576 return cachep->array[smp_processor_id()];
577}
578
579static int calculate_nr_objs(size_t slab_size, size_t buffer_size,
580 size_t idx_size, size_t align)
581{
582 int nr_objs;
583 size_t freelist_size;
584
585 /*
586 * Ignore padding for the initial guess. The padding
587 * is at most @align-1 bytes, and @buffer_size is at
588 * least @align. In the worst case, this result will
589 * be one greater than the number of objects that fit
590 * into the memory allocation when taking the padding
591 * into account.
592 */
593 nr_objs = slab_size / (buffer_size + idx_size);
594
595 /*
596 * This calculated number will be either the right
597 * amount, or one greater than what we want.
598 */
599 freelist_size = slab_size - nr_objs * buffer_size;
600 if (freelist_size < ALIGN(nr_objs * idx_size, align))
601 nr_objs--;
602
603 return nr_objs;
604}
605
606/*
607 * Calculate the number of objects and left-over bytes for a given buffer size.
608 */
609static void cache_estimate(unsigned long gfporder, size_t buffer_size,
610 size_t align, int flags, size_t *left_over,
611 unsigned int *num)
612{
613 int nr_objs;
614 size_t mgmt_size;
615 size_t slab_size = PAGE_SIZE << gfporder;
616
617 /*
618 * The slab management structure can be either off the slab or
619 * on it. For the latter case, the memory allocated for a
620 * slab is used for:
621 *
622 * - One unsigned int for each object
623 * - Padding to respect alignment of @align
624 * - @buffer_size bytes for each object
625 *
626 * If the slab management structure is off the slab, then the
627 * alignment will already be calculated into the size. Because
628 * the slabs are all pages aligned, the objects will be at the
629 * correct alignment when allocated.
630 */
631 if (flags & CFLGS_OFF_SLAB) {
632 mgmt_size = 0;
633 nr_objs = slab_size / buffer_size;
634
635 } else {
636 nr_objs = calculate_nr_objs(slab_size, buffer_size,
637 sizeof(freelist_idx_t), align);
638 mgmt_size = ALIGN(nr_objs * sizeof(freelist_idx_t), align);
639 }
640 *num = nr_objs;
641 *left_over = slab_size - nr_objs*buffer_size - mgmt_size;
642}
643
644#if DEBUG
645#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
646
647static void __slab_error(const char *function, struct kmem_cache *cachep,
648 char *msg)
649{
650 printk(KERN_ERR "slab error in %s(): cache `%s': %s\n",
651 function, cachep->name, msg);
652 dump_stack();
653 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
654}
655#endif
656
657/*
658 * By default on NUMA we use alien caches to stage the freeing of
659 * objects allocated from other nodes. This causes massive memory
660 * inefficiencies when using fake NUMA setup to split memory into a
661 * large number of small nodes, so it can be disabled on the command
662 * line
663 */
664
665static int use_alien_caches __read_mostly = 1;
666static int __init noaliencache_setup(char *s)
667{
668 use_alien_caches = 0;
669 return 1;
670}
671__setup("noaliencache", noaliencache_setup);
672
673static int __init slab_max_order_setup(char *str)
674{
675 get_option(&str, &slab_max_order);
676 slab_max_order = slab_max_order < 0 ? 0 :
677 min(slab_max_order, MAX_ORDER - 1);
678 slab_max_order_set = true;
679
680 return 1;
681}
682__setup("slab_max_order=", slab_max_order_setup);
683
684#ifdef CONFIG_NUMA
685/*
686 * Special reaping functions for NUMA systems called from cache_reap().
687 * These take care of doing round robin flushing of alien caches (containing
688 * objects freed on different nodes from which they were allocated) and the
689 * flushing of remote pcps by calling drain_node_pages.
690 */
691static DEFINE_PER_CPU(unsigned long, slab_reap_node);
692
693static void init_reap_node(int cpu)
694{
695 int node;
696
697 node = next_node(cpu_to_mem(cpu), node_online_map);
698 if (node == MAX_NUMNODES)
699 node = first_node(node_online_map);
700
701 per_cpu(slab_reap_node, cpu) = node;
702}
703
704static void next_reap_node(void)
705{
706 int node = __this_cpu_read(slab_reap_node);
707
708 node = next_node(node, node_online_map);
709 if (unlikely(node >= MAX_NUMNODES))
710 node = first_node(node_online_map);
711 __this_cpu_write(slab_reap_node, node);
712}
713
714#else
715#define init_reap_node(cpu) do { } while (0)
716#define next_reap_node(void) do { } while (0)
717#endif
718
719/*
720 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
721 * via the workqueue/eventd.
722 * Add the CPU number into the expiration time to minimize the possibility of
723 * the CPUs getting into lockstep and contending for the global cache chain
724 * lock.
725 */
726static void start_cpu_timer(int cpu)
727{
728 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
729
730 /*
731 * When this gets called from do_initcalls via cpucache_init(),
732 * init_workqueues() has already run, so keventd will be setup
733 * at that time.
734 */
735 if (keventd_up() && reap_work->work.func == NULL) {
736 init_reap_node(cpu);
737 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
738 schedule_delayed_work_on(cpu, reap_work,
739 __round_jiffies_relative(HZ, cpu));
740 }
741}
742
743static struct array_cache *alloc_arraycache(int node, int entries,
744 int batchcount, gfp_t gfp)
745{
746 int memsize = sizeof(void *) * entries + sizeof(struct array_cache);
747 struct array_cache *nc = NULL;
748
749 nc = kmalloc_node(memsize, gfp, node);
750 /*
751 * The array_cache structures contain pointers to free object.
752 * However, when such objects are allocated or transferred to another
753 * cache the pointers are not cleared and they could be counted as
754 * valid references during a kmemleak scan. Therefore, kmemleak must
755 * not scan such objects.
756 */
757 kmemleak_no_scan(nc);
758 if (nc) {
759 nc->avail = 0;
760 nc->limit = entries;
761 nc->batchcount = batchcount;
762 nc->touched = 0;
763 spin_lock_init(&nc->lock);
764 }
765 return nc;
766}
767
768static inline bool is_slab_pfmemalloc(struct page *page)
769{
770 return PageSlabPfmemalloc(page);
771}
772
773/* Clears pfmemalloc_active if no slabs have pfmalloc set */
774static void recheck_pfmemalloc_active(struct kmem_cache *cachep,
775 struct array_cache *ac)
776{
777 struct kmem_cache_node *n = cachep->node[numa_mem_id()];
778 struct page *page;
779 unsigned long flags;
780
781 if (!pfmemalloc_active)
782 return;
783
784 spin_lock_irqsave(&n->list_lock, flags);
785 list_for_each_entry(page, &n->slabs_full, lru)
786 if (is_slab_pfmemalloc(page))
787 goto out;
788
789 list_for_each_entry(page, &n->slabs_partial, lru)
790 if (is_slab_pfmemalloc(page))
791 goto out;
792
793 list_for_each_entry(page, &n->slabs_free, lru)
794 if (is_slab_pfmemalloc(page))
795 goto out;
796
797 pfmemalloc_active = false;
798out:
799 spin_unlock_irqrestore(&n->list_lock, flags);
800}
801
802static void *__ac_get_obj(struct kmem_cache *cachep, struct array_cache *ac,
803 gfp_t flags, bool force_refill)
804{
805 int i;
806 void *objp = ac->entry[--ac->avail];
807
808 /* Ensure the caller is allowed to use objects from PFMEMALLOC slab */
809 if (unlikely(is_obj_pfmemalloc(objp))) {
810 struct kmem_cache_node *n;
811
812 if (gfp_pfmemalloc_allowed(flags)) {
813 clear_obj_pfmemalloc(&objp);
814 return objp;
815 }
816
817 /* The caller cannot use PFMEMALLOC objects, find another one */
818 for (i = 0; i < ac->avail; i++) {
819 /* If a !PFMEMALLOC object is found, swap them */
820 if (!is_obj_pfmemalloc(ac->entry[i])) {
821 objp = ac->entry[i];
822 ac->entry[i] = ac->entry[ac->avail];
823 ac->entry[ac->avail] = objp;
824 return objp;
825 }
826 }
827
828 /*
829 * If there are empty slabs on the slabs_free list and we are
830 * being forced to refill the cache, mark this one !pfmemalloc.
831 */
832 n = cachep->node[numa_mem_id()];
833 if (!list_empty(&n->slabs_free) && force_refill) {
834 struct page *page = virt_to_head_page(objp);
835 ClearPageSlabPfmemalloc(page);
836 clear_obj_pfmemalloc(&objp);
837 recheck_pfmemalloc_active(cachep, ac);
838 return objp;
839 }
840
841 /* No !PFMEMALLOC objects available */
842 ac->avail++;
843 objp = NULL;
844 }
845
846 return objp;
847}
848
849static inline void *ac_get_obj(struct kmem_cache *cachep,
850 struct array_cache *ac, gfp_t flags, bool force_refill)
851{
852 void *objp;
853
854 if (unlikely(sk_memalloc_socks()))
855 objp = __ac_get_obj(cachep, ac, flags, force_refill);
856 else
857 objp = ac->entry[--ac->avail];
858
859 return objp;
860}
861
862static void *__ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
863 void *objp)
864{
865 if (unlikely(pfmemalloc_active)) {
866 /* Some pfmemalloc slabs exist, check if this is one */
867 struct page *page = virt_to_head_page(objp);
868 if (PageSlabPfmemalloc(page))
869 set_obj_pfmemalloc(&objp);
870 }
871
872 return objp;
873}
874
875static inline void ac_put_obj(struct kmem_cache *cachep, struct array_cache *ac,
876 void *objp)
877{
878 if (unlikely(sk_memalloc_socks()))
879 objp = __ac_put_obj(cachep, ac, objp);
880
881 ac->entry[ac->avail++] = objp;
882}
883
884/*
885 * Transfer objects in one arraycache to another.
886 * Locking must be handled by the caller.
887 *
888 * Return the number of entries transferred.
889 */
890static int transfer_objects(struct array_cache *to,
891 struct array_cache *from, unsigned int max)
892{
893 /* Figure out how many entries to transfer */
894 int nr = min3(from->avail, max, to->limit - to->avail);
895
896 if (!nr)
897 return 0;
898
899 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
900 sizeof(void *) *nr);
901
902 from->avail -= nr;
903 to->avail += nr;
904 return nr;
905}
906
907#ifndef CONFIG_NUMA
908
909#define drain_alien_cache(cachep, alien) do { } while (0)
910#define reap_alien(cachep, n) do { } while (0)
911
912static inline struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
913{
914 return (struct array_cache **)BAD_ALIEN_MAGIC;
915}
916
917static inline void free_alien_cache(struct array_cache **ac_ptr)
918{
919}
920
921static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
922{
923 return 0;
924}
925
926static inline void *alternate_node_alloc(struct kmem_cache *cachep,
927 gfp_t flags)
928{
929 return NULL;
930}
931
932static inline void *____cache_alloc_node(struct kmem_cache *cachep,
933 gfp_t flags, int nodeid)
934{
935 return NULL;
936}
937
938#else /* CONFIG_NUMA */
939
940static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
941static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
942
943static struct array_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
944{
945 struct array_cache **ac_ptr;
946 int memsize = sizeof(void *) * nr_node_ids;
947 int i;
948
949 if (limit > 1)
950 limit = 12;
951 ac_ptr = kzalloc_node(memsize, gfp, node);
952 if (ac_ptr) {
953 for_each_node(i) {
954 if (i == node || !node_online(i))
955 continue;
956 ac_ptr[i] = alloc_arraycache(node, limit, 0xbaadf00d, gfp);
957 if (!ac_ptr[i]) {
958 for (i--; i >= 0; i--)
959 kfree(ac_ptr[i]);
960 kfree(ac_ptr);
961 return NULL;
962 }
963 }
964 }
965 return ac_ptr;
966}
967
968static void free_alien_cache(struct array_cache **ac_ptr)
969{
970 int i;
971
972 if (!ac_ptr)
973 return;
974 for_each_node(i)
975 kfree(ac_ptr[i]);
976 kfree(ac_ptr);
977}
978
979static void __drain_alien_cache(struct kmem_cache *cachep,
980 struct array_cache *ac, int node)
981{
982 struct kmem_cache_node *n = cachep->node[node];
983
984 if (ac->avail) {
985 spin_lock(&n->list_lock);
986 /*
987 * Stuff objects into the remote nodes shared array first.
988 * That way we could avoid the overhead of putting the objects
989 * into the free lists and getting them back later.
990 */
991 if (n->shared)
992 transfer_objects(n->shared, ac, ac->limit);
993
994 free_block(cachep, ac->entry, ac->avail, node);
995 ac->avail = 0;
996 spin_unlock(&n->list_lock);
997 }
998}
999
1000/*
1001 * Called from cache_reap() to regularly drain alien caches round robin.
1002 */
1003static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
1004{
1005 int node = __this_cpu_read(slab_reap_node);
1006
1007 if (n->alien) {
1008 struct array_cache *ac = n->alien[node];
1009
1010 if (ac && ac->avail && spin_trylock_irq(&ac->lock)) {
1011 __drain_alien_cache(cachep, ac, node);
1012 spin_unlock_irq(&ac->lock);
1013 }
1014 }
1015}
1016
1017static void drain_alien_cache(struct kmem_cache *cachep,
1018 struct array_cache **alien)
1019{
1020 int i = 0;
1021 struct array_cache *ac;
1022 unsigned long flags;
1023
1024 for_each_online_node(i) {
1025 ac = alien[i];
1026 if (ac) {
1027 spin_lock_irqsave(&ac->lock, flags);
1028 __drain_alien_cache(cachep, ac, i);
1029 spin_unlock_irqrestore(&ac->lock, flags);
1030 }
1031 }
1032}
1033
1034static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
1035{
1036 int nodeid = page_to_nid(virt_to_page(objp));
1037 struct kmem_cache_node *n;
1038 struct array_cache *alien = NULL;
1039 int node;
1040
1041 node = numa_mem_id();
1042
1043 /*
1044 * Make sure we are not freeing a object from another node to the array
1045 * cache on this cpu.
1046 */
1047 if (likely(nodeid == node))
1048 return 0;
1049
1050 n = cachep->node[node];
1051 STATS_INC_NODEFREES(cachep);
1052 if (n->alien && n->alien[nodeid]) {
1053 alien = n->alien[nodeid];
1054 spin_lock(&alien->lock);
1055 if (unlikely(alien->avail == alien->limit)) {
1056 STATS_INC_ACOVERFLOW(cachep);
1057 __drain_alien_cache(cachep, alien, nodeid);
1058 }
1059 ac_put_obj(cachep, alien, objp);
1060 spin_unlock(&alien->lock);
1061 } else {
1062 spin_lock(&(cachep->node[nodeid])->list_lock);
1063 free_block(cachep, &objp, 1, nodeid);
1064 spin_unlock(&(cachep->node[nodeid])->list_lock);
1065 }
1066 return 1;
1067}
1068#endif
1069
1070/*
1071 * Allocates and initializes node for a node on each slab cache, used for
1072 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
1073 * will be allocated off-node since memory is not yet online for the new node.
1074 * When hotplugging memory or a cpu, existing node are not replaced if
1075 * already in use.
1076 *
1077 * Must hold slab_mutex.
1078 */
1079static int init_cache_node_node(int node)
1080{
1081 struct kmem_cache *cachep;
1082 struct kmem_cache_node *n;
1083 const int memsize = sizeof(struct kmem_cache_node);
1084
1085 list_for_each_entry(cachep, &slab_caches, list) {
1086 /*
1087 * Set up the kmem_cache_node for cpu before we can
1088 * begin anything. Make sure some other cpu on this
1089 * node has not already allocated this
1090 */
1091 if (!cachep->node[node]) {
1092 n = kmalloc_node(memsize, GFP_KERNEL, node);
1093 if (!n)
1094 return -ENOMEM;
1095 kmem_cache_node_init(n);
1096 n->next_reap = jiffies + REAPTIMEOUT_NODE +
1097 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1098
1099 /*
1100 * The kmem_cache_nodes don't come and go as CPUs
1101 * come and go. slab_mutex is sufficient
1102 * protection here.
1103 */
1104 cachep->node[node] = n;
1105 }
1106
1107 spin_lock_irq(&cachep->node[node]->list_lock);
1108 cachep->node[node]->free_limit =
1109 (1 + nr_cpus_node(node)) *
1110 cachep->batchcount + cachep->num;
1111 spin_unlock_irq(&cachep->node[node]->list_lock);
1112 }
1113 return 0;
1114}
1115
1116static inline int slabs_tofree(struct kmem_cache *cachep,
1117 struct kmem_cache_node *n)
1118{
1119 return (n->free_objects + cachep->num - 1) / cachep->num;
1120}
1121
1122static void cpuup_canceled(long cpu)
1123{
1124 struct kmem_cache *cachep;
1125 struct kmem_cache_node *n = NULL;
1126 int node = cpu_to_mem(cpu);
1127 const struct cpumask *mask = cpumask_of_node(node);
1128
1129 list_for_each_entry(cachep, &slab_caches, list) {
1130 struct array_cache *nc;
1131 struct array_cache *shared;
1132 struct array_cache **alien;
1133
1134 /* cpu is dead; no one can alloc from it. */
1135 nc = cachep->array[cpu];
1136 cachep->array[cpu] = NULL;
1137 n = cachep->node[node];
1138
1139 if (!n)
1140 goto free_array_cache;
1141
1142 spin_lock_irq(&n->list_lock);
1143
1144 /* Free limit for this kmem_cache_node */
1145 n->free_limit -= cachep->batchcount;
1146 if (nc)
1147 free_block(cachep, nc->entry, nc->avail, node);
1148
1149 if (!cpumask_empty(mask)) {
1150 spin_unlock_irq(&n->list_lock);
1151 goto free_array_cache;
1152 }
1153
1154 shared = n->shared;
1155 if (shared) {
1156 free_block(cachep, shared->entry,
1157 shared->avail, node);
1158 n->shared = NULL;
1159 }
1160
1161 alien = n->alien;
1162 n->alien = NULL;
1163
1164 spin_unlock_irq(&n->list_lock);
1165
1166 kfree(shared);
1167 if (alien) {
1168 drain_alien_cache(cachep, alien);
1169 free_alien_cache(alien);
1170 }
1171free_array_cache:
1172 kfree(nc);
1173 }
1174 /*
1175 * In the previous loop, all the objects were freed to
1176 * the respective cache's slabs, now we can go ahead and
1177 * shrink each nodelist to its limit.
1178 */
1179 list_for_each_entry(cachep, &slab_caches, list) {
1180 n = cachep->node[node];
1181 if (!n)
1182 continue;
1183 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1184 }
1185}
1186
1187static int cpuup_prepare(long cpu)
1188{
1189 struct kmem_cache *cachep;
1190 struct kmem_cache_node *n = NULL;
1191 int node = cpu_to_mem(cpu);
1192 int err;
1193
1194 /*
1195 * We need to do this right in the beginning since
1196 * alloc_arraycache's are going to use this list.
1197 * kmalloc_node allows us to add the slab to the right
1198 * kmem_cache_node and not this cpu's kmem_cache_node
1199 */
1200 err = init_cache_node_node(node);
1201 if (err < 0)
1202 goto bad;
1203
1204 /*
1205 * Now we can go ahead with allocating the shared arrays and
1206 * array caches
1207 */
1208 list_for_each_entry(cachep, &slab_caches, list) {
1209 struct array_cache *nc;
1210 struct array_cache *shared = NULL;
1211 struct array_cache **alien = NULL;
1212
1213 nc = alloc_arraycache(node, cachep->limit,
1214 cachep->batchcount, GFP_KERNEL);
1215 if (!nc)
1216 goto bad;
1217 if (cachep->shared) {
1218 shared = alloc_arraycache(node,
1219 cachep->shared * cachep->batchcount,
1220 0xbaadf00d, GFP_KERNEL);
1221 if (!shared) {
1222 kfree(nc);
1223 goto bad;
1224 }
1225 }
1226 if (use_alien_caches) {
1227 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1228 if (!alien) {
1229 kfree(shared);
1230 kfree(nc);
1231 goto bad;
1232 }
1233 }
1234 cachep->array[cpu] = nc;
1235 n = cachep->node[node];
1236 BUG_ON(!n);
1237
1238 spin_lock_irq(&n->list_lock);
1239 if (!n->shared) {
1240 /*
1241 * We are serialised from CPU_DEAD or
1242 * CPU_UP_CANCELLED by the cpucontrol lock
1243 */
1244 n->shared = shared;
1245 shared = NULL;
1246 }
1247#ifdef CONFIG_NUMA
1248 if (!n->alien) {
1249 n->alien = alien;
1250 alien = NULL;
1251 }
1252#endif
1253 spin_unlock_irq(&n->list_lock);
1254 kfree(shared);
1255 free_alien_cache(alien);
1256 if (cachep->flags & SLAB_DEBUG_OBJECTS)
1257 slab_set_debugobj_lock_classes_node(cachep, node);
1258 else if (!OFF_SLAB(cachep) &&
1259 !(cachep->flags & SLAB_DESTROY_BY_RCU))
1260 on_slab_lock_classes_node(cachep, node);
1261 }
1262 init_node_lock_keys(node);
1263
1264 return 0;
1265bad:
1266 cpuup_canceled(cpu);
1267 return -ENOMEM;
1268}
1269
1270static int cpuup_callback(struct notifier_block *nfb,
1271 unsigned long action, void *hcpu)
1272{
1273 long cpu = (long)hcpu;
1274 int err = 0;
1275
1276 switch (action) {
1277 case CPU_UP_PREPARE:
1278 case CPU_UP_PREPARE_FROZEN:
1279 mutex_lock(&slab_mutex);
1280 err = cpuup_prepare(cpu);
1281 mutex_unlock(&slab_mutex);
1282 break;
1283 case CPU_ONLINE:
1284 case CPU_ONLINE_FROZEN:
1285 start_cpu_timer(cpu);
1286 break;
1287#ifdef CONFIG_HOTPLUG_CPU
1288 case CPU_DOWN_PREPARE:
1289 case CPU_DOWN_PREPARE_FROZEN:
1290 /*
1291 * Shutdown cache reaper. Note that the slab_mutex is
1292 * held so that if cache_reap() is invoked it cannot do
1293 * anything expensive but will only modify reap_work
1294 * and reschedule the timer.
1295 */
1296 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1297 /* Now the cache_reaper is guaranteed to be not running. */
1298 per_cpu(slab_reap_work, cpu).work.func = NULL;
1299 break;
1300 case CPU_DOWN_FAILED:
1301 case CPU_DOWN_FAILED_FROZEN:
1302 start_cpu_timer(cpu);
1303 break;
1304 case CPU_DEAD:
1305 case CPU_DEAD_FROZEN:
1306 /*
1307 * Even if all the cpus of a node are down, we don't free the
1308 * kmem_cache_node of any cache. This to avoid a race between
1309 * cpu_down, and a kmalloc allocation from another cpu for
1310 * memory from the node of the cpu going down. The node
1311 * structure is usually allocated from kmem_cache_create() and
1312 * gets destroyed at kmem_cache_destroy().
1313 */
1314 /* fall through */
1315#endif
1316 case CPU_UP_CANCELED:
1317 case CPU_UP_CANCELED_FROZEN:
1318 mutex_lock(&slab_mutex);
1319 cpuup_canceled(cpu);
1320 mutex_unlock(&slab_mutex);
1321 break;
1322 }
1323 return notifier_from_errno(err);
1324}
1325
1326static struct notifier_block cpucache_notifier = {
1327 &cpuup_callback, NULL, 0
1328};
1329
1330#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1331/*
1332 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1333 * Returns -EBUSY if all objects cannot be drained so that the node is not
1334 * removed.
1335 *
1336 * Must hold slab_mutex.
1337 */
1338static int __meminit drain_cache_node_node(int node)
1339{
1340 struct kmem_cache *cachep;
1341 int ret = 0;
1342
1343 list_for_each_entry(cachep, &slab_caches, list) {
1344 struct kmem_cache_node *n;
1345
1346 n = cachep->node[node];
1347 if (!n)
1348 continue;
1349
1350 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1351
1352 if (!list_empty(&n->slabs_full) ||
1353 !list_empty(&n->slabs_partial)) {
1354 ret = -EBUSY;
1355 break;
1356 }
1357 }
1358 return ret;
1359}
1360
1361static int __meminit slab_memory_callback(struct notifier_block *self,
1362 unsigned long action, void *arg)
1363{
1364 struct memory_notify *mnb = arg;
1365 int ret = 0;
1366 int nid;
1367
1368 nid = mnb->status_change_nid;
1369 if (nid < 0)
1370 goto out;
1371
1372 switch (action) {
1373 case MEM_GOING_ONLINE:
1374 mutex_lock(&slab_mutex);
1375 ret = init_cache_node_node(nid);
1376 mutex_unlock(&slab_mutex);
1377 break;
1378 case MEM_GOING_OFFLINE:
1379 mutex_lock(&slab_mutex);
1380 ret = drain_cache_node_node(nid);
1381 mutex_unlock(&slab_mutex);
1382 break;
1383 case MEM_ONLINE:
1384 case MEM_OFFLINE:
1385 case MEM_CANCEL_ONLINE:
1386 case MEM_CANCEL_OFFLINE:
1387 break;
1388 }
1389out:
1390 return notifier_from_errno(ret);
1391}
1392#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1393
1394/*
1395 * swap the static kmem_cache_node with kmalloced memory
1396 */
1397static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1398 int nodeid)
1399{
1400 struct kmem_cache_node *ptr;
1401
1402 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1403 BUG_ON(!ptr);
1404
1405 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1406 /*
1407 * Do not assume that spinlocks can be initialized via memcpy:
1408 */
1409 spin_lock_init(&ptr->list_lock);
1410
1411 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1412 cachep->node[nodeid] = ptr;
1413}
1414
1415/*
1416 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1417 * size of kmem_cache_node.
1418 */
1419static void __init set_up_node(struct kmem_cache *cachep, int index)
1420{
1421 int node;
1422
1423 for_each_online_node(node) {
1424 cachep->node[node] = &init_kmem_cache_node[index + node];
1425 cachep->node[node]->next_reap = jiffies +
1426 REAPTIMEOUT_NODE +
1427 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1428 }
1429}
1430
1431/*
1432 * The memory after the last cpu cache pointer is used for the
1433 * the node pointer.
1434 */
1435static void setup_node_pointer(struct kmem_cache *cachep)
1436{
1437 cachep->node = (struct kmem_cache_node **)&cachep->array[nr_cpu_ids];
1438}
1439
1440/*
1441 * Initialisation. Called after the page allocator have been initialised and
1442 * before smp_init().
1443 */
1444void __init kmem_cache_init(void)
1445{
1446 int i;
1447
1448 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1449 sizeof(struct rcu_head));
1450 kmem_cache = &kmem_cache_boot;
1451 setup_node_pointer(kmem_cache);
1452
1453 if (num_possible_nodes() == 1)
1454 use_alien_caches = 0;
1455
1456 for (i = 0; i < NUM_INIT_LISTS; i++)
1457 kmem_cache_node_init(&init_kmem_cache_node[i]);
1458
1459 set_up_node(kmem_cache, CACHE_CACHE);
1460
1461 /*
1462 * Fragmentation resistance on low memory - only use bigger
1463 * page orders on machines with more than 32MB of memory if
1464 * not overridden on the command line.
1465 */
1466 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1467 slab_max_order = SLAB_MAX_ORDER_HI;
1468
1469 /* Bootstrap is tricky, because several objects are allocated
1470 * from caches that do not exist yet:
1471 * 1) initialize the kmem_cache cache: it contains the struct
1472 * kmem_cache structures of all caches, except kmem_cache itself:
1473 * kmem_cache is statically allocated.
1474 * Initially an __init data area is used for the head array and the
1475 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1476 * array at the end of the bootstrap.
1477 * 2) Create the first kmalloc cache.
1478 * The struct kmem_cache for the new cache is allocated normally.
1479 * An __init data area is used for the head array.
1480 * 3) Create the remaining kmalloc caches, with minimally sized
1481 * head arrays.
1482 * 4) Replace the __init data head arrays for kmem_cache and the first
1483 * kmalloc cache with kmalloc allocated arrays.
1484 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1485 * the other cache's with kmalloc allocated memory.
1486 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1487 */
1488
1489 /* 1) create the kmem_cache */
1490
1491 /*
1492 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1493 */
1494 create_boot_cache(kmem_cache, "kmem_cache",
1495 offsetof(struct kmem_cache, array[nr_cpu_ids]) +
1496 nr_node_ids * sizeof(struct kmem_cache_node *),
1497 SLAB_HWCACHE_ALIGN);
1498 list_add(&kmem_cache->list, &slab_caches);
1499
1500 /* 2+3) create the kmalloc caches */
1501
1502 /*
1503 * Initialize the caches that provide memory for the array cache and the
1504 * kmem_cache_node structures first. Without this, further allocations will
1505 * bug.
1506 */
1507
1508 kmalloc_caches[INDEX_AC] = create_kmalloc_cache("kmalloc-ac",
1509 kmalloc_size(INDEX_AC), ARCH_KMALLOC_FLAGS);
1510
1511 if (INDEX_AC != INDEX_NODE)
1512 kmalloc_caches[INDEX_NODE] =
1513 create_kmalloc_cache("kmalloc-node",
1514 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1515
1516 slab_early_init = 0;
1517
1518 /* 4) Replace the bootstrap head arrays */
1519 {
1520 struct array_cache *ptr;
1521
1522 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1523
1524 memcpy(ptr, cpu_cache_get(kmem_cache),
1525 sizeof(struct arraycache_init));
1526 /*
1527 * Do not assume that spinlocks can be initialized via memcpy:
1528 */
1529 spin_lock_init(&ptr->lock);
1530
1531 kmem_cache->array[smp_processor_id()] = ptr;
1532
1533 ptr = kmalloc(sizeof(struct arraycache_init), GFP_NOWAIT);
1534
1535 BUG_ON(cpu_cache_get(kmalloc_caches[INDEX_AC])
1536 != &initarray_generic.cache);
1537 memcpy(ptr, cpu_cache_get(kmalloc_caches[INDEX_AC]),
1538 sizeof(struct arraycache_init));
1539 /*
1540 * Do not assume that spinlocks can be initialized via memcpy:
1541 */
1542 spin_lock_init(&ptr->lock);
1543
1544 kmalloc_caches[INDEX_AC]->array[smp_processor_id()] = ptr;
1545 }
1546 /* 5) Replace the bootstrap kmem_cache_node */
1547 {
1548 int nid;
1549
1550 for_each_online_node(nid) {
1551 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1552
1553 init_list(kmalloc_caches[INDEX_AC],
1554 &init_kmem_cache_node[SIZE_AC + nid], nid);
1555
1556 if (INDEX_AC != INDEX_NODE) {
1557 init_list(kmalloc_caches[INDEX_NODE],
1558 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1559 }
1560 }
1561 }
1562
1563 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1564}
1565
1566void __init kmem_cache_init_late(void)
1567{
1568 struct kmem_cache *cachep;
1569
1570 slab_state = UP;
1571
1572 /* 6) resize the head arrays to their final sizes */
1573 mutex_lock(&slab_mutex);
1574 list_for_each_entry(cachep, &slab_caches, list)
1575 if (enable_cpucache(cachep, GFP_NOWAIT))
1576 BUG();
1577 mutex_unlock(&slab_mutex);
1578
1579 /* Annotate slab for lockdep -- annotate the malloc caches */
1580 init_lock_keys();
1581
1582 /* Done! */
1583 slab_state = FULL;
1584
1585 /*
1586 * Register a cpu startup notifier callback that initializes
1587 * cpu_cache_get for all new cpus
1588 */
1589 register_cpu_notifier(&cpucache_notifier);
1590
1591#ifdef CONFIG_NUMA
1592 /*
1593 * Register a memory hotplug callback that initializes and frees
1594 * node.
1595 */
1596 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1597#endif
1598
1599 /*
1600 * The reap timers are started later, with a module init call: That part
1601 * of the kernel is not yet operational.
1602 */
1603}
1604
1605static int __init cpucache_init(void)
1606{
1607 int cpu;
1608
1609 /*
1610 * Register the timers that return unneeded pages to the page allocator
1611 */
1612 for_each_online_cpu(cpu)
1613 start_cpu_timer(cpu);
1614
1615 /* Done! */
1616 slab_state = FULL;
1617 return 0;
1618}
1619__initcall(cpucache_init);
1620
1621static noinline void
1622slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1623{
1624 struct kmem_cache_node *n;
1625 struct page *page;
1626 unsigned long flags;
1627 int node;
1628
1629 printk(KERN_WARNING
1630 "SLAB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1631 nodeid, gfpflags);
1632 printk(KERN_WARNING " cache: %s, object size: %d, order: %d\n",
1633 cachep->name, cachep->size, cachep->gfporder);
1634
1635 for_each_online_node(node) {
1636 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1637 unsigned long active_slabs = 0, num_slabs = 0;
1638
1639 n = cachep->node[node];
1640 if (!n)
1641 continue;
1642
1643 spin_lock_irqsave(&n->list_lock, flags);
1644 list_for_each_entry(page, &n->slabs_full, lru) {
1645 active_objs += cachep->num;
1646 active_slabs++;
1647 }
1648 list_for_each_entry(page, &n->slabs_partial, lru) {
1649 active_objs += page->active;
1650 active_slabs++;
1651 }
1652 list_for_each_entry(page, &n->slabs_free, lru)
1653 num_slabs++;
1654
1655 free_objects += n->free_objects;
1656 spin_unlock_irqrestore(&n->list_lock, flags);
1657
1658 num_slabs += active_slabs;
1659 num_objs = num_slabs * cachep->num;
1660 printk(KERN_WARNING
1661 " node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1662 node, active_slabs, num_slabs, active_objs, num_objs,
1663 free_objects);
1664 }
1665}
1666
1667/*
1668 * Interface to system's page allocator. No need to hold the cache-lock.
1669 *
1670 * If we requested dmaable memory, we will get it. Even if we
1671 * did not request dmaable memory, we might get it, but that
1672 * would be relatively rare and ignorable.
1673 */
1674static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1675 int nodeid)
1676{
1677 struct page *page;
1678 int nr_pages;
1679
1680 flags |= cachep->allocflags;
1681 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1682 flags |= __GFP_RECLAIMABLE;
1683
1684 page = alloc_pages_exact_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1685 if (!page) {
1686 if (!(flags & __GFP_NOWARN) && printk_ratelimit())
1687 slab_out_of_memory(cachep, flags, nodeid);
1688 return NULL;
1689 }
1690
1691 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1692 if (unlikely(page->pfmemalloc))
1693 pfmemalloc_active = true;
1694
1695 nr_pages = (1 << cachep->gfporder);
1696 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1697 add_zone_page_state(page_zone(page),
1698 NR_SLAB_RECLAIMABLE, nr_pages);
1699 else
1700 add_zone_page_state(page_zone(page),
1701 NR_SLAB_UNRECLAIMABLE, nr_pages);
1702 __SetPageSlab(page);
1703 if (page->pfmemalloc)
1704 SetPageSlabPfmemalloc(page);
1705 memcg_bind_pages(cachep, cachep->gfporder);
1706
1707 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1708 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1709
1710 if (cachep->ctor)
1711 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1712 else
1713 kmemcheck_mark_unallocated_pages(page, nr_pages);
1714 }
1715
1716 return page;
1717}
1718
1719/*
1720 * Interface to system's page release.
1721 */
1722static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1723{
1724 const unsigned long nr_freed = (1 << cachep->gfporder);
1725
1726 kmemcheck_free_shadow(page, cachep->gfporder);
1727
1728 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1729 sub_zone_page_state(page_zone(page),
1730 NR_SLAB_RECLAIMABLE, nr_freed);
1731 else
1732 sub_zone_page_state(page_zone(page),
1733 NR_SLAB_UNRECLAIMABLE, nr_freed);
1734
1735 BUG_ON(!PageSlab(page));
1736 __ClearPageSlabPfmemalloc(page);
1737 __ClearPageSlab(page);
1738 page_mapcount_reset(page);
1739 page->mapping = NULL;
1740
1741 memcg_release_pages(cachep, cachep->gfporder);
1742 if (current->reclaim_state)
1743 current->reclaim_state->reclaimed_slab += nr_freed;
1744 __free_memcg_kmem_pages(page, cachep->gfporder);
1745}
1746
1747static void kmem_rcu_free(struct rcu_head *head)
1748{
1749 struct kmem_cache *cachep;
1750 struct page *page;
1751
1752 page = container_of(head, struct page, rcu_head);
1753 cachep = page->slab_cache;
1754
1755 kmem_freepages(cachep, page);
1756}
1757
1758#if DEBUG
1759
1760#ifdef CONFIG_DEBUG_PAGEALLOC
1761static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1762 unsigned long caller)
1763{
1764 int size = cachep->object_size;
1765
1766 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1767
1768 if (size < 5 * sizeof(unsigned long))
1769 return;
1770
1771 *addr++ = 0x12345678;
1772 *addr++ = caller;
1773 *addr++ = smp_processor_id();
1774 size -= 3 * sizeof(unsigned long);
1775 {
1776 unsigned long *sptr = &caller;
1777 unsigned long svalue;
1778
1779 while (!kstack_end(sptr)) {
1780 svalue = *sptr++;
1781 if (kernel_text_address(svalue)) {
1782 *addr++ = svalue;
1783 size -= sizeof(unsigned long);
1784 if (size <= sizeof(unsigned long))
1785 break;
1786 }
1787 }
1788
1789 }
1790 *addr++ = 0x87654321;
1791}
1792#endif
1793
1794static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1795{
1796 int size = cachep->object_size;
1797 addr = &((char *)addr)[obj_offset(cachep)];
1798
1799 memset(addr, val, size);
1800 *(unsigned char *)(addr + size - 1) = POISON_END;
1801}
1802
1803static void dump_line(char *data, int offset, int limit)
1804{
1805 int i;
1806 unsigned char error = 0;
1807 int bad_count = 0;
1808
1809 printk(KERN_ERR "%03x: ", offset);
1810 for (i = 0; i < limit; i++) {
1811 if (data[offset + i] != POISON_FREE) {
1812 error = data[offset + i];
1813 bad_count++;
1814 }
1815 }
1816 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1817 &data[offset], limit, 1);
1818
1819 if (bad_count == 1) {
1820 error ^= POISON_FREE;
1821 if (!(error & (error - 1))) {
1822 printk(KERN_ERR "Single bit error detected. Probably "
1823 "bad RAM.\n");
1824#ifdef CONFIG_X86
1825 printk(KERN_ERR "Run memtest86+ or a similar memory "
1826 "test tool.\n");
1827#else
1828 printk(KERN_ERR "Run a memory test tool.\n");
1829#endif
1830 }
1831 }
1832}
1833#endif
1834
1835#if DEBUG
1836
1837static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1838{
1839 int i, size;
1840 char *realobj;
1841
1842 if (cachep->flags & SLAB_RED_ZONE) {
1843 printk(KERN_ERR "Redzone: 0x%llx/0x%llx.\n",
1844 *dbg_redzone1(cachep, objp),
1845 *dbg_redzone2(cachep, objp));
1846 }
1847
1848 if (cachep->flags & SLAB_STORE_USER) {
1849 printk(KERN_ERR "Last user: [<%p>](%pSR)\n",
1850 *dbg_userword(cachep, objp),
1851 *dbg_userword(cachep, objp));
1852 }
1853 realobj = (char *)objp + obj_offset(cachep);
1854 size = cachep->object_size;
1855 for (i = 0; i < size && lines; i += 16, lines--) {
1856 int limit;
1857 limit = 16;
1858 if (i + limit > size)
1859 limit = size - i;
1860 dump_line(realobj, i, limit);
1861 }
1862}
1863
1864static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1865{
1866 char *realobj;
1867 int size, i;
1868 int lines = 0;
1869
1870 realobj = (char *)objp + obj_offset(cachep);
1871 size = cachep->object_size;
1872
1873 for (i = 0; i < size; i++) {
1874 char exp = POISON_FREE;
1875 if (i == size - 1)
1876 exp = POISON_END;
1877 if (realobj[i] != exp) {
1878 int limit;
1879 /* Mismatch ! */
1880 /* Print header */
1881 if (lines == 0) {
1882 printk(KERN_ERR
1883 "Slab corruption (%s): %s start=%p, len=%d\n",
1884 print_tainted(), cachep->name, realobj, size);
1885 print_objinfo(cachep, objp, 0);
1886 }
1887 /* Hexdump the affected line */
1888 i = (i / 16) * 16;
1889 limit = 16;
1890 if (i + limit > size)
1891 limit = size - i;
1892 dump_line(realobj, i, limit);
1893 i += 16;
1894 lines++;
1895 /* Limit to 5 lines */
1896 if (lines > 5)
1897 break;
1898 }
1899 }
1900 if (lines != 0) {
1901 /* Print some data about the neighboring objects, if they
1902 * exist:
1903 */
1904 struct page *page = virt_to_head_page(objp);
1905 unsigned int objnr;
1906
1907 objnr = obj_to_index(cachep, page, objp);
1908 if (objnr) {
1909 objp = index_to_obj(cachep, page, objnr - 1);
1910 realobj = (char *)objp + obj_offset(cachep);
1911 printk(KERN_ERR "Prev obj: start=%p, len=%d\n",
1912 realobj, size);
1913 print_objinfo(cachep, objp, 2);
1914 }
1915 if (objnr + 1 < cachep->num) {
1916 objp = index_to_obj(cachep, page, objnr + 1);
1917 realobj = (char *)objp + obj_offset(cachep);
1918 printk(KERN_ERR "Next obj: start=%p, len=%d\n",
1919 realobj, size);
1920 print_objinfo(cachep, objp, 2);
1921 }
1922 }
1923}
1924#endif
1925
1926#if DEBUG
1927static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1928 struct page *page)
1929{
1930 int i;
1931 for (i = 0; i < cachep->num; i++) {
1932 void *objp = index_to_obj(cachep, page, i);
1933
1934 if (cachep->flags & SLAB_POISON) {
1935#ifdef CONFIG_DEBUG_PAGEALLOC
1936 if (cachep->size % PAGE_SIZE == 0 &&
1937 OFF_SLAB(cachep))
1938 kernel_map_pages(virt_to_page(objp),
1939 cachep->size / PAGE_SIZE, 1);
1940 else
1941 check_poison_obj(cachep, objp);
1942#else
1943 check_poison_obj(cachep, objp);
1944#endif
1945 }
1946 if (cachep->flags & SLAB_RED_ZONE) {
1947 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1948 slab_error(cachep, "start of a freed object "
1949 "was overwritten");
1950 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1951 slab_error(cachep, "end of a freed object "
1952 "was overwritten");
1953 }
1954 }
1955}
1956#else
1957static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1958 struct page *page)
1959{
1960}
1961#endif
1962
1963/**
1964 * slab_destroy - destroy and release all objects in a slab
1965 * @cachep: cache pointer being destroyed
1966 * @page: page pointer being destroyed
1967 *
1968 * Destroy all the objs in a slab, and release the mem back to the system.
1969 * Before calling the slab must have been unlinked from the cache. The
1970 * cache-lock is not held/needed.
1971 */
1972static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1973{
1974 void *freelist;
1975
1976 freelist = page->freelist;
1977 slab_destroy_debugcheck(cachep, page);
1978 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU)) {
1979 struct rcu_head *head;
1980
1981 /*
1982 * RCU free overloads the RCU head over the LRU.
1983 * slab_page has been overloeaded over the LRU,
1984 * however it is not used from now on so that
1985 * we can use it safely.
1986 */
1987 head = (void *)&page->rcu_head;
1988 call_rcu(head, kmem_rcu_free);
1989
1990 } else {
1991 kmem_freepages(cachep, page);
1992 }
1993
1994 /*
1995 * From now on, we don't use freelist
1996 * although actual page can be freed in rcu context
1997 */
1998 if (OFF_SLAB(cachep))
1999 kmem_cache_free(cachep->freelist_cache, freelist);
2000}
2001
2002/**
2003 * calculate_slab_order - calculate size (page order) of slabs
2004 * @cachep: pointer to the cache that is being created
2005 * @size: size of objects to be created in this cache.
2006 * @align: required alignment for the objects.
2007 * @flags: slab allocation flags
2008 *
2009 * Also calculates the number of objects per slab.
2010 *
2011 * This could be made much more intelligent. For now, try to avoid using
2012 * high order pages for slabs. When the gfp() functions are more friendly
2013 * towards high-order requests, this should be changed.
2014 */
2015static size_t calculate_slab_order(struct kmem_cache *cachep,
2016 size_t size, size_t align, unsigned long flags)
2017{
2018 unsigned long offslab_limit;
2019 size_t left_over = 0;
2020 int gfporder;
2021
2022 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
2023 unsigned int num;
2024 size_t remainder;
2025
2026 cache_estimate(gfporder, size, align, flags, &remainder, &num);
2027 if (!num)
2028 continue;
2029
2030 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
2031 if (num > SLAB_OBJ_MAX_NUM)
2032 break;
2033
2034 if (flags & CFLGS_OFF_SLAB) {
2035 /*
2036 * Max number of objs-per-slab for caches which
2037 * use off-slab slabs. Needed to avoid a possible
2038 * looping condition in cache_grow().
2039 */
2040 offslab_limit = size;
2041 offslab_limit /= sizeof(freelist_idx_t);
2042
2043 if (num > offslab_limit)
2044 break;
2045 }
2046
2047 /* Found something acceptable - save it away */
2048 cachep->num = num;
2049 cachep->gfporder = gfporder;
2050 left_over = remainder;
2051
2052 /*
2053 * A VFS-reclaimable slab tends to have most allocations
2054 * as GFP_NOFS and we really don't want to have to be allocating
2055 * higher-order pages when we are unable to shrink dcache.
2056 */
2057 if (flags & SLAB_RECLAIM_ACCOUNT)
2058 break;
2059
2060 /*
2061 * Large number of objects is good, but very large slabs are
2062 * currently bad for the gfp()s.
2063 */
2064 if (gfporder >= slab_max_order)
2065 break;
2066
2067 /*
2068 * Acceptable internal fragmentation?
2069 */
2070 if (left_over * 8 <= (PAGE_SIZE << gfporder))
2071 break;
2072 }
2073 return left_over;
2074}
2075
2076static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
2077{
2078 if (slab_state >= FULL)
2079 return enable_cpucache(cachep, gfp);
2080
2081 if (slab_state == DOWN) {
2082 /*
2083 * Note: Creation of first cache (kmem_cache).
2084 * The setup_node is taken care
2085 * of by the caller of __kmem_cache_create
2086 */
2087 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2088 slab_state = PARTIAL;
2089 } else if (slab_state == PARTIAL) {
2090 /*
2091 * Note: the second kmem_cache_create must create the cache
2092 * that's used by kmalloc(24), otherwise the creation of
2093 * further caches will BUG().
2094 */
2095 cachep->array[smp_processor_id()] = &initarray_generic.cache;
2096
2097 /*
2098 * If the cache that's used by kmalloc(sizeof(kmem_cache_node)) is
2099 * the second cache, then we need to set up all its node/,
2100 * otherwise the creation of further caches will BUG().
2101 */
2102 set_up_node(cachep, SIZE_AC);
2103 if (INDEX_AC == INDEX_NODE)
2104 slab_state = PARTIAL_NODE;
2105 else
2106 slab_state = PARTIAL_ARRAYCACHE;
2107 } else {
2108 /* Remaining boot caches */
2109 cachep->array[smp_processor_id()] =
2110 kmalloc(sizeof(struct arraycache_init), gfp);
2111
2112 if (slab_state == PARTIAL_ARRAYCACHE) {
2113 set_up_node(cachep, SIZE_NODE);
2114 slab_state = PARTIAL_NODE;
2115 } else {
2116 int node;
2117 for_each_online_node(node) {
2118 cachep->node[node] =
2119 kmalloc_node(sizeof(struct kmem_cache_node),
2120 gfp, node);
2121 BUG_ON(!cachep->node[node]);
2122 kmem_cache_node_init(cachep->node[node]);
2123 }
2124 }
2125 }
2126 cachep->node[numa_mem_id()]->next_reap =
2127 jiffies + REAPTIMEOUT_NODE +
2128 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
2129
2130 cpu_cache_get(cachep)->avail = 0;
2131 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
2132 cpu_cache_get(cachep)->batchcount = 1;
2133 cpu_cache_get(cachep)->touched = 0;
2134 cachep->batchcount = 1;
2135 cachep->limit = BOOT_CPUCACHE_ENTRIES;
2136 return 0;
2137}
2138
2139/**
2140 * __kmem_cache_create - Create a cache.
2141 * @cachep: cache management descriptor
2142 * @flags: SLAB flags
2143 *
2144 * Returns a ptr to the cache on success, NULL on failure.
2145 * Cannot be called within a int, but can be interrupted.
2146 * The @ctor is run when new pages are allocated by the cache.
2147 *
2148 * The flags are
2149 *
2150 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2151 * to catch references to uninitialised memory.
2152 *
2153 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2154 * for buffer overruns.
2155 *
2156 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2157 * cacheline. This can be beneficial if you're counting cycles as closely
2158 * as davem.
2159 */
2160int
2161__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2162{
2163 size_t left_over, freelist_size, ralign;
2164 gfp_t gfp;
2165 int err;
2166 size_t size = cachep->size;
2167
2168#if DEBUG
2169#if FORCED_DEBUG
2170 /*
2171 * Enable redzoning and last user accounting, except for caches with
2172 * large objects, if the increased size would increase the object size
2173 * above the next power of two: caches with object sizes just above a
2174 * power of two have a significant amount of internal fragmentation.
2175 */
2176 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2177 2 * sizeof(unsigned long long)))
2178 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2179 if (!(flags & SLAB_DESTROY_BY_RCU))
2180 flags |= SLAB_POISON;
2181#endif
2182 if (flags & SLAB_DESTROY_BY_RCU)
2183 BUG_ON(flags & SLAB_POISON);
2184#endif
2185
2186 /*
2187 * Check that size is in terms of words. This is needed to avoid
2188 * unaligned accesses for some archs when redzoning is used, and makes
2189 * sure any on-slab bufctl's are also correctly aligned.
2190 */
2191 if (size & (BYTES_PER_WORD - 1)) {
2192 size += (BYTES_PER_WORD - 1);
2193 size &= ~(BYTES_PER_WORD - 1);
2194 }
2195
2196 /*
2197 * Redzoning and user store require word alignment or possibly larger.
2198 * Note this will be overridden by architecture or caller mandated
2199 * alignment if either is greater than BYTES_PER_WORD.
2200 */
2201 if (flags & SLAB_STORE_USER)
2202 ralign = BYTES_PER_WORD;
2203
2204 if (flags & SLAB_RED_ZONE) {
2205 ralign = REDZONE_ALIGN;
2206 /* If redzoning, ensure that the second redzone is suitably
2207 * aligned, by adjusting the object size accordingly. */
2208 size += REDZONE_ALIGN - 1;
2209 size &= ~(REDZONE_ALIGN - 1);
2210 }
2211
2212 /* 3) caller mandated alignment */
2213 if (ralign < cachep->align) {
2214 ralign = cachep->align;
2215 }
2216 /* disable debug if necessary */
2217 if (ralign > __alignof__(unsigned long long))
2218 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2219 /*
2220 * 4) Store it.
2221 */
2222 cachep->align = ralign;
2223
2224 if (slab_is_available())
2225 gfp = GFP_KERNEL;
2226 else
2227 gfp = GFP_NOWAIT;
2228
2229 setup_node_pointer(cachep);
2230#if DEBUG
2231
2232 /*
2233 * Both debugging options require word-alignment which is calculated
2234 * into align above.
2235 */
2236 if (flags & SLAB_RED_ZONE) {
2237 /* add space for red zone words */
2238 cachep->obj_offset += sizeof(unsigned long long);
2239 size += 2 * sizeof(unsigned long long);
2240 }
2241 if (flags & SLAB_STORE_USER) {
2242 /* user store requires one word storage behind the end of
2243 * the real object. But if the second red zone needs to be
2244 * aligned to 64 bits, we must allow that much space.
2245 */
2246 if (flags & SLAB_RED_ZONE)
2247 size += REDZONE_ALIGN;
2248 else
2249 size += BYTES_PER_WORD;
2250 }
2251#if FORCED_DEBUG && defined(CONFIG_DEBUG_PAGEALLOC)
2252 if (size >= kmalloc_size(INDEX_NODE + 1)
2253 && cachep->object_size > cache_line_size()
2254 && ALIGN(size, cachep->align) < PAGE_SIZE) {
2255 cachep->obj_offset += PAGE_SIZE - ALIGN(size, cachep->align);
2256 size = PAGE_SIZE;
2257 }
2258#endif
2259#endif
2260
2261 /*
2262 * Determine if the slab management is 'on' or 'off' slab.
2263 * (bootstrapping cannot cope with offslab caches so don't do
2264 * it too early on. Always use on-slab management when
2265 * SLAB_NOLEAKTRACE to avoid recursive calls into kmemleak)
2266 */
2267 if ((size >= (PAGE_SIZE >> 5)) && !slab_early_init &&
2268 !(flags & SLAB_NOLEAKTRACE))
2269 /*
2270 * Size is large, assume best to place the slab management obj
2271 * off-slab (should allow better packing of objs).
2272 */
2273 flags |= CFLGS_OFF_SLAB;
2274
2275 size = ALIGN(size, cachep->align);
2276 /*
2277 * We should restrict the number of objects in a slab to implement
2278 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2279 */
2280 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2281 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2282
2283 left_over = calculate_slab_order(cachep, size, cachep->align, flags);
2284
2285 if (!cachep->num)
2286 return -E2BIG;
2287
2288 freelist_size =
2289 ALIGN(cachep->num * sizeof(freelist_idx_t), cachep->align);
2290
2291 /*
2292 * If the slab has been placed off-slab, and we have enough space then
2293 * move it on-slab. This is at the expense of any extra colouring.
2294 */
2295 if (flags & CFLGS_OFF_SLAB && left_over >= freelist_size) {
2296 flags &= ~CFLGS_OFF_SLAB;
2297 left_over -= freelist_size;
2298 }
2299
2300 if (flags & CFLGS_OFF_SLAB) {
2301 /* really off slab. No need for manual alignment */
2302 freelist_size = cachep->num * sizeof(freelist_idx_t);
2303
2304#ifdef CONFIG_PAGE_POISONING
2305 /* If we're going to use the generic kernel_map_pages()
2306 * poisoning, then it's going to smash the contents of
2307 * the redzone and userword anyhow, so switch them off.
2308 */
2309 if (size % PAGE_SIZE == 0 && flags & SLAB_POISON)
2310 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2311#endif
2312 }
2313
2314 cachep->colour_off = cache_line_size();
2315 /* Offset must be a multiple of the alignment. */
2316 if (cachep->colour_off < cachep->align)
2317 cachep->colour_off = cachep->align;
2318 cachep->colour = left_over / cachep->colour_off;
2319 cachep->freelist_size = freelist_size;
2320 cachep->flags = flags;
2321 cachep->allocflags = __GFP_COMP;
2322 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2323 cachep->allocflags |= GFP_DMA;
2324 cachep->size = size;
2325 cachep->reciprocal_buffer_size = reciprocal_value(size);
2326
2327 if (flags & CFLGS_OFF_SLAB) {
2328 cachep->freelist_cache = kmalloc_slab(freelist_size, 0u);
2329 /*
2330 * This is a possibility for one of the kmalloc_{dma,}_caches.
2331 * But since we go off slab only for object size greater than
2332 * PAGE_SIZE/8, and kmalloc_{dma,}_caches get created
2333 * in ascending order,this should not happen at all.
2334 * But leave a BUG_ON for some lucky dude.
2335 */
2336 BUG_ON(ZERO_OR_NULL_PTR(cachep->freelist_cache));
2337 }
2338
2339 err = setup_cpu_cache(cachep, gfp);
2340 if (err) {
2341 __kmem_cache_shutdown(cachep);
2342 return err;
2343 }
2344
2345 if (flags & SLAB_DEBUG_OBJECTS) {
2346 /*
2347 * Would deadlock through slab_destroy()->call_rcu()->
2348 * debug_object_activate()->kmem_cache_alloc().
2349 */
2350 WARN_ON_ONCE(flags & SLAB_DESTROY_BY_RCU);
2351
2352 slab_set_debugobj_lock_classes(cachep);
2353 } else if (!OFF_SLAB(cachep) && !(flags & SLAB_DESTROY_BY_RCU))
2354 on_slab_lock_classes(cachep);
2355
2356 return 0;
2357}
2358
2359#if DEBUG
2360static void check_irq_off(void)
2361{
2362 BUG_ON(!irqs_disabled());
2363}
2364
2365static void check_irq_on(void)
2366{
2367 BUG_ON(irqs_disabled());
2368}
2369
2370static void check_spinlock_acquired(struct kmem_cache *cachep)
2371{
2372#ifdef CONFIG_SMP
2373 check_irq_off();
2374 assert_spin_locked(&cachep->node[numa_mem_id()]->list_lock);
2375#endif
2376}
2377
2378static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2379{
2380#ifdef CONFIG_SMP
2381 check_irq_off();
2382 assert_spin_locked(&cachep->node[node]->list_lock);
2383#endif
2384}
2385
2386#else
2387#define check_irq_off() do { } while(0)
2388#define check_irq_on() do { } while(0)
2389#define check_spinlock_acquired(x) do { } while(0)
2390#define check_spinlock_acquired_node(x, y) do { } while(0)
2391#endif
2392
2393static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2394 struct array_cache *ac,
2395 int force, int node);
2396
2397static void do_drain(void *arg)
2398{
2399 struct kmem_cache *cachep = arg;
2400 struct array_cache *ac;
2401 int node = numa_mem_id();
2402
2403 check_irq_off();
2404 ac = cpu_cache_get(cachep);
2405 spin_lock(&cachep->node[node]->list_lock);
2406 free_block(cachep, ac->entry, ac->avail, node);
2407 spin_unlock(&cachep->node[node]->list_lock);
2408 ac->avail = 0;
2409}
2410
2411static void drain_cpu_caches(struct kmem_cache *cachep)
2412{
2413 struct kmem_cache_node *n;
2414 int node;
2415
2416 on_each_cpu(do_drain, cachep, 1);
2417 check_irq_on();
2418 for_each_online_node(node) {
2419 n = cachep->node[node];
2420 if (n && n->alien)
2421 drain_alien_cache(cachep, n->alien);
2422 }
2423
2424 for_each_online_node(node) {
2425 n = cachep->node[node];
2426 if (n)
2427 drain_array(cachep, n, n->shared, 1, node);
2428 }
2429}
2430
2431/*
2432 * Remove slabs from the list of free slabs.
2433 * Specify the number of slabs to drain in tofree.
2434 *
2435 * Returns the actual number of slabs released.
2436 */
2437static int drain_freelist(struct kmem_cache *cache,
2438 struct kmem_cache_node *n, int tofree)
2439{
2440 struct list_head *p;
2441 int nr_freed;
2442 struct page *page;
2443
2444 nr_freed = 0;
2445 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2446
2447 spin_lock_irq(&n->list_lock);
2448 p = n->slabs_free.prev;
2449 if (p == &n->slabs_free) {
2450 spin_unlock_irq(&n->list_lock);
2451 goto out;
2452 }
2453
2454 page = list_entry(p, struct page, lru);
2455#if DEBUG
2456 BUG_ON(page->active);
2457#endif
2458 list_del(&page->lru);
2459 /*
2460 * Safe to drop the lock. The slab is no longer linked
2461 * to the cache.
2462 */
2463 n->free_objects -= cache->num;
2464 spin_unlock_irq(&n->list_lock);
2465 slab_destroy(cache, page);
2466 nr_freed++;
2467 }
2468out:
2469 return nr_freed;
2470}
2471
2472/* Called with slab_mutex held to protect against cpu hotplug */
2473static int __cache_shrink(struct kmem_cache *cachep)
2474{
2475 int ret = 0, i = 0;
2476 struct kmem_cache_node *n;
2477
2478 drain_cpu_caches(cachep);
2479
2480 check_irq_on();
2481 for_each_online_node(i) {
2482 n = cachep->node[i];
2483 if (!n)
2484 continue;
2485
2486 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2487
2488 ret += !list_empty(&n->slabs_full) ||
2489 !list_empty(&n->slabs_partial);
2490 }
2491 return (ret ? 1 : 0);
2492}
2493
2494/**
2495 * kmem_cache_shrink - Shrink a cache.
2496 * @cachep: The cache to shrink.
2497 *
2498 * Releases as many slabs as possible for a cache.
2499 * To help debugging, a zero exit status indicates all slabs were released.
2500 */
2501int kmem_cache_shrink(struct kmem_cache *cachep)
2502{
2503 int ret;
2504 BUG_ON(!cachep || in_interrupt());
2505
2506 get_online_cpus();
2507 mutex_lock(&slab_mutex);
2508 ret = __cache_shrink(cachep);
2509 mutex_unlock(&slab_mutex);
2510 put_online_cpus();
2511 return ret;
2512}
2513EXPORT_SYMBOL(kmem_cache_shrink);
2514
2515int __kmem_cache_shutdown(struct kmem_cache *cachep)
2516{
2517 int i;
2518 struct kmem_cache_node *n;
2519 int rc = __cache_shrink(cachep);
2520
2521 if (rc)
2522 return rc;
2523
2524 for_each_online_cpu(i)
2525 kfree(cachep->array[i]);
2526
2527 /* NUMA: free the node structures */
2528 for_each_online_node(i) {
2529 n = cachep->node[i];
2530 if (n) {
2531 kfree(n->shared);
2532 free_alien_cache(n->alien);
2533 kfree(n);
2534 }
2535 }
2536 return 0;
2537}
2538
2539/*
2540 * Get the memory for a slab management obj.
2541 *
2542 * For a slab cache when the slab descriptor is off-slab, the
2543 * slab descriptor can't come from the same cache which is being created,
2544 * Because if it is the case, that means we defer the creation of
2545 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2546 * And we eventually call down to __kmem_cache_create(), which
2547 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2548 * This is a "chicken-and-egg" problem.
2549 *
2550 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2551 * which are all initialized during kmem_cache_init().
2552 */
2553static void *alloc_slabmgmt(struct kmem_cache *cachep,
2554 struct page *page, int colour_off,
2555 gfp_t local_flags, int nodeid)
2556{
2557 void *freelist;
2558 void *addr = page_address(page);
2559
2560 if (OFF_SLAB(cachep)) {
2561 /* Slab management obj is off-slab. */
2562 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2563 local_flags, nodeid);
2564 if (!freelist)
2565 return NULL;
2566 } else {
2567 freelist = addr + colour_off;
2568 colour_off += cachep->freelist_size;
2569 }
2570 page->active = 0;
2571 page->s_mem = addr + colour_off;
2572 return freelist;
2573}
2574
2575static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2576{
2577 return ((freelist_idx_t *)page->freelist)[idx];
2578}
2579
2580static inline void set_free_obj(struct page *page,
2581 unsigned int idx, freelist_idx_t val)
2582{
2583 ((freelist_idx_t *)(page->freelist))[idx] = val;
2584}
2585
2586static void cache_init_objs(struct kmem_cache *cachep,
2587 struct page *page)
2588{
2589 int i;
2590
2591 for (i = 0; i < cachep->num; i++) {
2592 void *objp = index_to_obj(cachep, page, i);
2593#if DEBUG
2594 /* need to poison the objs? */
2595 if (cachep->flags & SLAB_POISON)
2596 poison_obj(cachep, objp, POISON_FREE);
2597 if (cachep->flags & SLAB_STORE_USER)
2598 *dbg_userword(cachep, objp) = NULL;
2599
2600 if (cachep->flags & SLAB_RED_ZONE) {
2601 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2602 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2603 }
2604 /*
2605 * Constructors are not allowed to allocate memory from the same
2606 * cache which they are a constructor for. Otherwise, deadlock.
2607 * They must also be threaded.
2608 */
2609 if (cachep->ctor && !(cachep->flags & SLAB_POISON))
2610 cachep->ctor(objp + obj_offset(cachep));
2611
2612 if (cachep->flags & SLAB_RED_ZONE) {
2613 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2614 slab_error(cachep, "constructor overwrote the"
2615 " end of an object");
2616 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2617 slab_error(cachep, "constructor overwrote the"
2618 " start of an object");
2619 }
2620 if ((cachep->size % PAGE_SIZE) == 0 &&
2621 OFF_SLAB(cachep) && cachep->flags & SLAB_POISON)
2622 kernel_map_pages(virt_to_page(objp),
2623 cachep->size / PAGE_SIZE, 0);
2624#else
2625 if (cachep->ctor)
2626 cachep->ctor(objp);
2627#endif
2628 set_free_obj(page, i, i);
2629 }
2630}
2631
2632static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2633{
2634 if (CONFIG_ZONE_DMA_FLAG) {
2635 if (flags & GFP_DMA)
2636 BUG_ON(!(cachep->allocflags & GFP_DMA));
2637 else
2638 BUG_ON(cachep->allocflags & GFP_DMA);
2639 }
2640}
2641
2642static void *slab_get_obj(struct kmem_cache *cachep, struct page *page,
2643 int nodeid)
2644{
2645 void *objp;
2646
2647 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2648 page->active++;
2649#if DEBUG
2650 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2651#endif
2652
2653 return objp;
2654}
2655
2656static void slab_put_obj(struct kmem_cache *cachep, struct page *page,
2657 void *objp, int nodeid)
2658{
2659 unsigned int objnr = obj_to_index(cachep, page, objp);
2660#if DEBUG
2661 unsigned int i;
2662
2663 /* Verify that the slab belongs to the intended node */
2664 WARN_ON(page_to_nid(virt_to_page(objp)) != nodeid);
2665
2666 /* Verify double free bug */
2667 for (i = page->active; i < cachep->num; i++) {
2668 if (get_free_obj(page, i) == objnr) {
2669 printk(KERN_ERR "slab: double free detected in cache "
2670 "'%s', objp %p\n", cachep->name, objp);
2671 BUG();
2672 }
2673 }
2674#endif
2675 page->active--;
2676 set_free_obj(page, page->active, objnr);
2677}
2678
2679/*
2680 * Map pages beginning at addr to the given cache and slab. This is required
2681 * for the slab allocator to be able to lookup the cache and slab of a
2682 * virtual address for kfree, ksize, and slab debugging.
2683 */
2684static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2685 void *freelist)
2686{
2687 page->slab_cache = cache;
2688 page->freelist = freelist;
2689}
2690
2691/*
2692 * Grow (by 1) the number of slabs within a cache. This is called by
2693 * kmem_cache_alloc() when there are no active objs left in a cache.
2694 */
2695static int cache_grow(struct kmem_cache *cachep,
2696 gfp_t flags, int nodeid, struct page *page)
2697{
2698 void *freelist;
2699 size_t offset;
2700 gfp_t local_flags;
2701 struct kmem_cache_node *n;
2702
2703 /*
2704 * Be lazy and only check for valid flags here, keeping it out of the
2705 * critical path in kmem_cache_alloc().
2706 */
2707 BUG_ON(flags & GFP_SLAB_BUG_MASK);
2708 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2709
2710 /* Take the node list lock to change the colour_next on this node */
2711 check_irq_off();
2712 n = cachep->node[nodeid];
2713 spin_lock(&n->list_lock);
2714
2715 /* Get colour for the slab, and cal the next value. */
2716 offset = n->colour_next;
2717 n->colour_next++;
2718 if (n->colour_next >= cachep->colour)
2719 n->colour_next = 0;
2720 spin_unlock(&n->list_lock);
2721
2722 offset *= cachep->colour_off;
2723
2724 if (local_flags & __GFP_WAIT)
2725 local_irq_enable();
2726
2727 /*
2728 * The test for missing atomic flag is performed here, rather than
2729 * the more obvious place, simply to reduce the critical path length
2730 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2731 * will eventually be caught here (where it matters).
2732 */
2733 kmem_flagcheck(cachep, flags);
2734
2735 /*
2736 * Get mem for the objs. Attempt to allocate a physical page from
2737 * 'nodeid'.
2738 */
2739 if (!page)
2740 page = kmem_getpages(cachep, local_flags, nodeid);
2741 if (!page)
2742 goto failed;
2743
2744 /* Get slab management. */
2745 freelist = alloc_slabmgmt(cachep, page, offset,
2746 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2747 if (!freelist)
2748 goto opps1;
2749
2750 slab_map_pages(cachep, page, freelist);
2751
2752 cache_init_objs(cachep, page);
2753
2754 if (local_flags & __GFP_WAIT)
2755 local_irq_disable();
2756 check_irq_off();
2757 spin_lock(&n->list_lock);
2758
2759 /* Make slab active. */
2760 list_add_tail(&page->lru, &(n->slabs_free));
2761 STATS_INC_GROWN(cachep);
2762 n->free_objects += cachep->num;
2763 spin_unlock(&n->list_lock);
2764 return 1;
2765opps1:
2766 kmem_freepages(cachep, page);
2767failed:
2768 if (local_flags & __GFP_WAIT)
2769 local_irq_disable();
2770 return 0;
2771}
2772
2773#if DEBUG
2774
2775/*
2776 * Perform extra freeing checks:
2777 * - detect bad pointers.
2778 * - POISON/RED_ZONE checking
2779 */
2780static void kfree_debugcheck(const void *objp)
2781{
2782 if (!virt_addr_valid(objp)) {
2783 printk(KERN_ERR "kfree_debugcheck: out of range ptr %lxh.\n",
2784 (unsigned long)objp);
2785 BUG();
2786 }
2787}
2788
2789static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2790{
2791 unsigned long long redzone1, redzone2;
2792
2793 redzone1 = *dbg_redzone1(cache, obj);
2794 redzone2 = *dbg_redzone2(cache, obj);
2795
2796 /*
2797 * Redzone is ok.
2798 */
2799 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2800 return;
2801
2802 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2803 slab_error(cache, "double free detected");
2804 else
2805 slab_error(cache, "memory outside object was overwritten");
2806
2807 printk(KERN_ERR "%p: redzone 1:0x%llx, redzone 2:0x%llx.\n",
2808 obj, redzone1, redzone2);
2809}
2810
2811static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2812 unsigned long caller)
2813{
2814 unsigned int objnr;
2815 struct page *page;
2816
2817 BUG_ON(virt_to_cache(objp) != cachep);
2818
2819 objp -= obj_offset(cachep);
2820 kfree_debugcheck(objp);
2821 page = virt_to_head_page(objp);
2822
2823 if (cachep->flags & SLAB_RED_ZONE) {
2824 verify_redzone_free(cachep, objp);
2825 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2826 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2827 }
2828 if (cachep->flags & SLAB_STORE_USER)
2829 *dbg_userword(cachep, objp) = (void *)caller;
2830
2831 objnr = obj_to_index(cachep, page, objp);
2832
2833 BUG_ON(objnr >= cachep->num);
2834 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2835
2836 if (cachep->flags & SLAB_POISON) {
2837#ifdef CONFIG_DEBUG_PAGEALLOC
2838 if ((cachep->size % PAGE_SIZE)==0 && OFF_SLAB(cachep)) {
2839 store_stackinfo(cachep, objp, caller);
2840 kernel_map_pages(virt_to_page(objp),
2841 cachep->size / PAGE_SIZE, 0);
2842 } else {
2843 poison_obj(cachep, objp, POISON_FREE);
2844 }
2845#else
2846 poison_obj(cachep, objp, POISON_FREE);
2847#endif
2848 }
2849 return objp;
2850}
2851
2852#else
2853#define kfree_debugcheck(x) do { } while(0)
2854#define cache_free_debugcheck(x,objp,z) (objp)
2855#endif
2856
2857static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags,
2858 bool force_refill)
2859{
2860 int batchcount;
2861 struct kmem_cache_node *n;
2862 struct array_cache *ac;
2863 int node;
2864
2865 check_irq_off();
2866 node = numa_mem_id();
2867 if (unlikely(force_refill))
2868 goto force_grow;
2869retry:
2870 ac = cpu_cache_get(cachep);
2871 batchcount = ac->batchcount;
2872 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2873 /*
2874 * If there was little recent activity on this cache, then
2875 * perform only a partial refill. Otherwise we could generate
2876 * refill bouncing.
2877 */
2878 batchcount = BATCHREFILL_LIMIT;
2879 }
2880 n = cachep->node[node];
2881
2882 BUG_ON(ac->avail > 0 || !n);
2883 spin_lock(&n->list_lock);
2884
2885 /* See if we can refill from the shared array */
2886 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2887 n->shared->touched = 1;
2888 goto alloc_done;
2889 }
2890
2891 while (batchcount > 0) {
2892 struct list_head *entry;
2893 struct page *page;
2894 /* Get slab alloc is to come from. */
2895 entry = n->slabs_partial.next;
2896 if (entry == &n->slabs_partial) {
2897 n->free_touched = 1;
2898 entry = n->slabs_free.next;
2899 if (entry == &n->slabs_free)
2900 goto must_grow;
2901 }
2902
2903 page = list_entry(entry, struct page, lru);
2904 check_spinlock_acquired(cachep);
2905
2906 /*
2907 * The slab was either on partial or free list so
2908 * there must be at least one object available for
2909 * allocation.
2910 */
2911 BUG_ON(page->active >= cachep->num);
2912
2913 while (page->active < cachep->num && batchcount--) {
2914 STATS_INC_ALLOCED(cachep);
2915 STATS_INC_ACTIVE(cachep);
2916 STATS_SET_HIGH(cachep);
2917
2918 ac_put_obj(cachep, ac, slab_get_obj(cachep, page,
2919 node));
2920 }
2921
2922 /* move slabp to correct slabp list: */
2923 list_del(&page->lru);
2924 if (page->active == cachep->num)
2925 list_add(&page->lru, &n->slabs_full);
2926 else
2927 list_add(&page->lru, &n->slabs_partial);
2928 }
2929
2930must_grow:
2931 n->free_objects -= ac->avail;
2932alloc_done:
2933 spin_unlock(&n->list_lock);
2934
2935 if (unlikely(!ac->avail)) {
2936 int x;
2937force_grow:
2938 x = cache_grow(cachep, flags | GFP_THISNODE, node, NULL);
2939
2940 /* cache_grow can reenable interrupts, then ac could change. */
2941 ac = cpu_cache_get(cachep);
2942 node = numa_mem_id();
2943
2944 /* no objects in sight? abort */
2945 if (!x && (ac->avail == 0 || force_refill))
2946 return NULL;
2947
2948 if (!ac->avail) /* objects refilled by interrupt? */
2949 goto retry;
2950 }
2951 ac->touched = 1;
2952
2953 return ac_get_obj(cachep, ac, flags, force_refill);
2954}
2955
2956static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2957 gfp_t flags)
2958{
2959 might_sleep_if(flags & __GFP_WAIT);
2960#if DEBUG
2961 kmem_flagcheck(cachep, flags);
2962#endif
2963}
2964
2965#if DEBUG
2966static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2967 gfp_t flags, void *objp, unsigned long caller)
2968{
2969 if (!objp)
2970 return objp;
2971 if (cachep->flags & SLAB_POISON) {
2972#ifdef CONFIG_DEBUG_PAGEALLOC
2973 if ((cachep->size % PAGE_SIZE) == 0 && OFF_SLAB(cachep))
2974 kernel_map_pages(virt_to_page(objp),
2975 cachep->size / PAGE_SIZE, 1);
2976 else
2977 check_poison_obj(cachep, objp);
2978#else
2979 check_poison_obj(cachep, objp);
2980#endif
2981 poison_obj(cachep, objp, POISON_INUSE);
2982 }
2983 if (cachep->flags & SLAB_STORE_USER)
2984 *dbg_userword(cachep, objp) = (void *)caller;
2985
2986 if (cachep->flags & SLAB_RED_ZONE) {
2987 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2988 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2989 slab_error(cachep, "double free, or memory outside"
2990 " object was overwritten");
2991 printk(KERN_ERR
2992 "%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2993 objp, *dbg_redzone1(cachep, objp),
2994 *dbg_redzone2(cachep, objp));
2995 }
2996 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2997 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2998 }
2999 objp += obj_offset(cachep);
3000 if (cachep->ctor && cachep->flags & SLAB_POISON)
3001 cachep->ctor(objp);
3002 if (ARCH_SLAB_MINALIGN &&
3003 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3004 printk(KERN_ERR "0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3005 objp, (int)ARCH_SLAB_MINALIGN);
3006 }
3007 return objp;
3008}
3009#else
3010#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3011#endif
3012
3013static bool slab_should_failslab(struct kmem_cache *cachep, gfp_t flags)
3014{
3015 if (cachep == kmem_cache)
3016 return false;
3017
3018 return should_failslab(cachep->object_size, flags, cachep->flags);
3019}
3020
3021static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3022{
3023 void *objp;
3024 struct array_cache *ac;
3025 bool force_refill = false;
3026
3027 check_irq_off();
3028
3029 ac = cpu_cache_get(cachep);
3030 if (likely(ac->avail)) {
3031 ac->touched = 1;
3032 objp = ac_get_obj(cachep, ac, flags, false);
3033
3034 /*
3035 * Allow for the possibility all avail objects are not allowed
3036 * by the current flags
3037 */
3038 if (objp) {
3039 STATS_INC_ALLOCHIT(cachep);
3040 goto out;
3041 }
3042 force_refill = true;
3043 }
3044
3045 STATS_INC_ALLOCMISS(cachep);
3046 objp = cache_alloc_refill(cachep, flags, force_refill);
3047 /*
3048 * the 'ac' may be updated by cache_alloc_refill(),
3049 * and kmemleak_erase() requires its correct value.
3050 */
3051 ac = cpu_cache_get(cachep);
3052
3053out:
3054 /*
3055 * To avoid a false negative, if an object that is in one of the
3056 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3057 * treat the array pointers as a reference to the object.
3058 */
3059 if (objp)
3060 kmemleak_erase(&ac->entry[ac->avail]);
3061 return objp;
3062}
3063
3064#ifdef CONFIG_NUMA
3065/*
3066 * Try allocating on another node if PF_SPREAD_SLAB is a mempolicy is set.
3067 *
3068 * If we are in_interrupt, then process context, including cpusets and
3069 * mempolicy, may not apply and should not be used for allocation policy.
3070 */
3071static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3072{
3073 int nid_alloc, nid_here;
3074
3075 if (in_interrupt() || (flags & __GFP_THISNODE))
3076 return NULL;
3077 nid_alloc = nid_here = numa_mem_id();
3078 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3079 nid_alloc = cpuset_slab_spread_node();
3080 else if (current->mempolicy)
3081 nid_alloc = mempolicy_slab_node();
3082 if (nid_alloc != nid_here)
3083 return ____cache_alloc_node(cachep, flags, nid_alloc);
3084 return NULL;
3085}
3086
3087/*
3088 * Fallback function if there was no memory available and no objects on a
3089 * certain node and fall back is permitted. First we scan all the
3090 * available node for available objects. If that fails then we
3091 * perform an allocation without specifying a node. This allows the page
3092 * allocator to do its reclaim / fallback magic. We then insert the
3093 * slab into the proper nodelist and then allocate from it.
3094 */
3095static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3096{
3097 struct zonelist *zonelist;
3098 gfp_t local_flags;
3099 struct zoneref *z;
3100 struct zone *zone;
3101 enum zone_type high_zoneidx = gfp_zone(flags);
3102 void *obj = NULL;
3103 int nid;
3104 unsigned int cpuset_mems_cookie;
3105
3106 if (flags & __GFP_THISNODE)
3107 return NULL;
3108
3109 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3110
3111retry_cpuset:
3112 cpuset_mems_cookie = read_mems_allowed_begin();
3113 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3114
3115retry:
3116 /*
3117 * Look through allowed nodes for objects available
3118 * from existing per node queues.
3119 */
3120 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3121 nid = zone_to_nid(zone);
3122
3123 if (cpuset_zone_allowed_hardwall(zone, flags) &&
3124 cache->node[nid] &&
3125 cache->node[nid]->free_objects) {
3126 obj = ____cache_alloc_node(cache,
3127 flags | GFP_THISNODE, nid);
3128 if (obj)
3129 break;
3130 }
3131 }
3132
3133 if (!obj) {
3134 /*
3135 * This allocation will be performed within the constraints
3136 * of the current cpuset / memory policy requirements.
3137 * We may trigger various forms of reclaim on the allowed
3138 * set and go into memory reserves if necessary.
3139 */
3140 struct page *page;
3141
3142 if (local_flags & __GFP_WAIT)
3143 local_irq_enable();
3144 kmem_flagcheck(cache, flags);
3145 page = kmem_getpages(cache, local_flags, numa_mem_id());
3146 if (local_flags & __GFP_WAIT)
3147 local_irq_disable();
3148 if (page) {
3149 /*
3150 * Insert into the appropriate per node queues
3151 */
3152 nid = page_to_nid(page);
3153 if (cache_grow(cache, flags, nid, page)) {
3154 obj = ____cache_alloc_node(cache,
3155 flags | GFP_THISNODE, nid);
3156 if (!obj)
3157 /*
3158 * Another processor may allocate the
3159 * objects in the slab since we are
3160 * not holding any locks.
3161 */
3162 goto retry;
3163 } else {
3164 /* cache_grow already freed obj */
3165 obj = NULL;
3166 }
3167 }
3168 }
3169
3170 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3171 goto retry_cpuset;
3172 return obj;
3173}
3174
3175/*
3176 * A interface to enable slab creation on nodeid
3177 */
3178static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3179 int nodeid)
3180{
3181 struct list_head *entry;
3182 struct page *page;
3183 struct kmem_cache_node *n;
3184 void *obj;
3185 int x;
3186
3187 VM_BUG_ON(nodeid > num_online_nodes());
3188 n = cachep->node[nodeid];
3189 BUG_ON(!n);
3190
3191retry:
3192 check_irq_off();
3193 spin_lock(&n->list_lock);
3194 entry = n->slabs_partial.next;
3195 if (entry == &n->slabs_partial) {
3196 n->free_touched = 1;
3197 entry = n->slabs_free.next;
3198 if (entry == &n->slabs_free)
3199 goto must_grow;
3200 }
3201
3202 page = list_entry(entry, struct page, lru);
3203 check_spinlock_acquired_node(cachep, nodeid);
3204
3205 STATS_INC_NODEALLOCS(cachep);
3206 STATS_INC_ACTIVE(cachep);
3207 STATS_SET_HIGH(cachep);
3208
3209 BUG_ON(page->active == cachep->num);
3210
3211 obj = slab_get_obj(cachep, page, nodeid);
3212 n->free_objects--;
3213 /* move slabp to correct slabp list: */
3214 list_del(&page->lru);
3215
3216 if (page->active == cachep->num)
3217 list_add(&page->lru, &n->slabs_full);
3218 else
3219 list_add(&page->lru, &n->slabs_partial);
3220
3221 spin_unlock(&n->list_lock);
3222 goto done;
3223
3224must_grow:
3225 spin_unlock(&n->list_lock);
3226 x = cache_grow(cachep, flags | GFP_THISNODE, nodeid, NULL);
3227 if (x)
3228 goto retry;
3229
3230 return fallback_alloc(cachep, flags);
3231
3232done:
3233 return obj;
3234}
3235
3236static __always_inline void *
3237slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3238 unsigned long caller)
3239{
3240 unsigned long save_flags;
3241 void *ptr;
3242 int slab_node = numa_mem_id();
3243
3244 flags &= gfp_allowed_mask;
3245
3246 lockdep_trace_alloc(flags);
3247
3248 if (slab_should_failslab(cachep, flags))
3249 return NULL;
3250
3251 cachep = memcg_kmem_get_cache(cachep, flags);
3252
3253 cache_alloc_debugcheck_before(cachep, flags);
3254 local_irq_save(save_flags);
3255
3256 if (nodeid == NUMA_NO_NODE)
3257 nodeid = slab_node;
3258
3259 if (unlikely(!cachep->node[nodeid])) {
3260 /* Node not bootstrapped yet */
3261 ptr = fallback_alloc(cachep, flags);
3262 goto out;
3263 }
3264
3265 if (nodeid == slab_node) {
3266 /*
3267 * Use the locally cached objects if possible.
3268 * However ____cache_alloc does not allow fallback
3269 * to other nodes. It may fail while we still have
3270 * objects on other nodes available.
3271 */
3272 ptr = ____cache_alloc(cachep, flags);
3273 if (ptr)
3274 goto out;
3275 }
3276 /* ___cache_alloc_node can fall back to other nodes */
3277 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3278 out:
3279 local_irq_restore(save_flags);
3280 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3281 kmemleak_alloc_recursive(ptr, cachep->object_size, 1, cachep->flags,
3282 flags);
3283
3284 if (likely(ptr)) {
3285 kmemcheck_slab_alloc(cachep, flags, ptr, cachep->object_size);
3286 if (unlikely(flags & __GFP_ZERO))
3287 memset(ptr, 0, cachep->object_size);
3288 }
3289
3290 return ptr;
3291}
3292
3293static __always_inline void *
3294__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3295{
3296 void *objp;
3297
3298 if (current->mempolicy || unlikely(current->flags & PF_SPREAD_SLAB)) {
3299 objp = alternate_node_alloc(cache, flags);
3300 if (objp)
3301 goto out;
3302 }
3303 objp = ____cache_alloc(cache, flags);
3304
3305 /*
3306 * We may just have run out of memory on the local node.
3307 * ____cache_alloc_node() knows how to locate memory on other nodes
3308 */
3309 if (!objp)
3310 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3311
3312 out:
3313 return objp;
3314}
3315#else
3316
3317static __always_inline void *
3318__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3319{
3320 return ____cache_alloc(cachep, flags);
3321}
3322
3323#endif /* CONFIG_NUMA */
3324
3325static __always_inline void *
3326slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3327{
3328 unsigned long save_flags;
3329 void *objp;
3330
3331 flags &= gfp_allowed_mask;
3332
3333 lockdep_trace_alloc(flags);
3334
3335 if (slab_should_failslab(cachep, flags))
3336 return NULL;
3337
3338 cachep = memcg_kmem_get_cache(cachep, flags);
3339
3340 cache_alloc_debugcheck_before(cachep, flags);
3341 local_irq_save(save_flags);
3342 objp = __do_cache_alloc(cachep, flags);
3343 local_irq_restore(save_flags);
3344 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3345 kmemleak_alloc_recursive(objp, cachep->object_size, 1, cachep->flags,
3346 flags);
3347 prefetchw(objp);
3348
3349 if (likely(objp)) {
3350 kmemcheck_slab_alloc(cachep, flags, objp, cachep->object_size);
3351 if (unlikely(flags & __GFP_ZERO))
3352 memset(objp, 0, cachep->object_size);
3353 }
3354
3355 return objp;
3356}
3357
3358/*
3359 * Caller needs to acquire correct kmem_cache_node's list_lock
3360 */
3361static void free_block(struct kmem_cache *cachep, void **objpp, int nr_objects,
3362 int node)
3363{
3364 int i;
3365 struct kmem_cache_node *n;
3366
3367 for (i = 0; i < nr_objects; i++) {
3368 void *objp;
3369 struct page *page;
3370
3371 clear_obj_pfmemalloc(&objpp[i]);
3372 objp = objpp[i];
3373
3374 page = virt_to_head_page(objp);
3375 n = cachep->node[node];
3376 list_del(&page->lru);
3377 check_spinlock_acquired_node(cachep, node);
3378 slab_put_obj(cachep, page, objp, node);
3379 STATS_DEC_ACTIVE(cachep);
3380 n->free_objects++;
3381
3382 /* fixup slab chains */
3383 if (page->active == 0) {
3384 if (n->free_objects > n->free_limit) {
3385 n->free_objects -= cachep->num;
3386 /* No need to drop any previously held
3387 * lock here, even if we have a off-slab slab
3388 * descriptor it is guaranteed to come from
3389 * a different cache, refer to comments before
3390 * alloc_slabmgmt.
3391 */
3392 slab_destroy(cachep, page);
3393 } else {
3394 list_add(&page->lru, &n->slabs_free);
3395 }
3396 } else {
3397 /* Unconditionally move a slab to the end of the
3398 * partial list on free - maximum time for the
3399 * other objects to be freed, too.
3400 */
3401 list_add_tail(&page->lru, &n->slabs_partial);
3402 }
3403 }
3404}
3405
3406static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3407{
3408 int batchcount;
3409 struct kmem_cache_node *n;
3410 int node = numa_mem_id();
3411
3412 batchcount = ac->batchcount;
3413#if DEBUG
3414 BUG_ON(!batchcount || batchcount > ac->avail);
3415#endif
3416 check_irq_off();
3417 n = cachep->node[node];
3418 spin_lock(&n->list_lock);
3419 if (n->shared) {
3420 struct array_cache *shared_array = n->shared;
3421 int max = shared_array->limit - shared_array->avail;
3422 if (max) {
3423 if (batchcount > max)
3424 batchcount = max;
3425 memcpy(&(shared_array->entry[shared_array->avail]),
3426 ac->entry, sizeof(void *) * batchcount);
3427 shared_array->avail += batchcount;
3428 goto free_done;
3429 }
3430 }
3431
3432 free_block(cachep, ac->entry, batchcount, node);
3433free_done:
3434#if STATS
3435 {
3436 int i = 0;
3437 struct list_head *p;
3438
3439 p = n->slabs_free.next;
3440 while (p != &(n->slabs_free)) {
3441 struct page *page;
3442
3443 page = list_entry(p, struct page, lru);
3444 BUG_ON(page->active);
3445
3446 i++;
3447 p = p->next;
3448 }
3449 STATS_SET_FREEABLE(cachep, i);
3450 }
3451#endif
3452 spin_unlock(&n->list_lock);
3453 ac->avail -= batchcount;
3454 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3455}
3456
3457/*
3458 * Release an obj back to its cache. If the obj has a constructed state, it must
3459 * be in this state _before_ it is released. Called with disabled ints.
3460 */
3461static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3462 unsigned long caller)
3463{
3464 struct array_cache *ac = cpu_cache_get(cachep);
3465
3466 check_irq_off();
3467 kmemleak_free_recursive(objp, cachep->flags);
3468 objp = cache_free_debugcheck(cachep, objp, caller);
3469
3470 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3471
3472 /*
3473 * Skip calling cache_free_alien() when the platform is not numa.
3474 * This will avoid cache misses that happen while accessing slabp (which
3475 * is per page memory reference) to get nodeid. Instead use a global
3476 * variable to skip the call, which is mostly likely to be present in
3477 * the cache.
3478 */
3479 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3480 return;
3481
3482 if (likely(ac->avail < ac->limit)) {
3483 STATS_INC_FREEHIT(cachep);
3484 } else {
3485 STATS_INC_FREEMISS(cachep);
3486 cache_flusharray(cachep, ac);
3487 }
3488
3489 ac_put_obj(cachep, ac, objp);
3490}
3491
3492/**
3493 * kmem_cache_alloc - Allocate an object
3494 * @cachep: The cache to allocate from.
3495 * @flags: See kmalloc().
3496 *
3497 * Allocate an object from this cache. The flags are only relevant
3498 * if the cache has no available objects.
3499 */
3500void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3501{
3502 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3503
3504 trace_kmem_cache_alloc(_RET_IP_, ret,
3505 cachep->object_size, cachep->size, flags);
3506
3507 return ret;
3508}
3509EXPORT_SYMBOL(kmem_cache_alloc);
3510
3511#ifdef CONFIG_TRACING
3512void *
3513kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3514{
3515 void *ret;
3516
3517 ret = slab_alloc(cachep, flags, _RET_IP_);
3518
3519 trace_kmalloc(_RET_IP_, ret,
3520 size, cachep->size, flags);
3521 return ret;
3522}
3523EXPORT_SYMBOL(kmem_cache_alloc_trace);
3524#endif
3525
3526#ifdef CONFIG_NUMA
3527/**
3528 * kmem_cache_alloc_node - Allocate an object on the specified node
3529 * @cachep: The cache to allocate from.
3530 * @flags: See kmalloc().
3531 * @nodeid: node number of the target node.
3532 *
3533 * Identical to kmem_cache_alloc but it will allocate memory on the given
3534 * node, which can improve the performance for cpu bound structures.
3535 *
3536 * Fallback to other node is possible if __GFP_THISNODE is not set.
3537 */
3538void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3539{
3540 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3541
3542 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3543 cachep->object_size, cachep->size,
3544 flags, nodeid);
3545
3546 return ret;
3547}
3548EXPORT_SYMBOL(kmem_cache_alloc_node);
3549
3550#ifdef CONFIG_TRACING
3551void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3552 gfp_t flags,
3553 int nodeid,
3554 size_t size)
3555{
3556 void *ret;
3557
3558 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3559
3560 trace_kmalloc_node(_RET_IP_, ret,
3561 size, cachep->size,
3562 flags, nodeid);
3563 return ret;
3564}
3565EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3566#endif
3567
3568static __always_inline void *
3569__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3570{
3571 struct kmem_cache *cachep;
3572
3573 cachep = kmalloc_slab(size, flags);
3574 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3575 return cachep;
3576 return kmem_cache_alloc_node_trace(cachep, flags, node, size);
3577}
3578
3579#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3580void *__kmalloc_node(size_t size, gfp_t flags, int node)
3581{
3582 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3583}
3584EXPORT_SYMBOL(__kmalloc_node);
3585
3586void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3587 int node, unsigned long caller)
3588{
3589 return __do_kmalloc_node(size, flags, node, caller);
3590}
3591EXPORT_SYMBOL(__kmalloc_node_track_caller);
3592#else
3593void *__kmalloc_node(size_t size, gfp_t flags, int node)
3594{
3595 return __do_kmalloc_node(size, flags, node, 0);
3596}
3597EXPORT_SYMBOL(__kmalloc_node);
3598#endif /* CONFIG_DEBUG_SLAB || CONFIG_TRACING */
3599#endif /* CONFIG_NUMA */
3600
3601/**
3602 * __do_kmalloc - allocate memory
3603 * @size: how many bytes of memory are required.
3604 * @flags: the type of memory to allocate (see kmalloc).
3605 * @caller: function caller for debug tracking of the caller
3606 */
3607static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3608 unsigned long caller)
3609{
3610 struct kmem_cache *cachep;
3611 void *ret;
3612
3613 cachep = kmalloc_slab(size, flags);
3614 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3615 return cachep;
3616 ret = slab_alloc(cachep, flags, caller);
3617
3618 trace_kmalloc(caller, ret,
3619 size, cachep->size, flags);
3620
3621 return ret;
3622}
3623
3624
3625#if defined(CONFIG_DEBUG_SLAB) || defined(CONFIG_TRACING)
3626void *__kmalloc(size_t size, gfp_t flags)
3627{
3628 return __do_kmalloc(size, flags, _RET_IP_);
3629}
3630EXPORT_SYMBOL(__kmalloc);
3631
3632void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3633{
3634 return __do_kmalloc(size, flags, caller);
3635}
3636EXPORT_SYMBOL(__kmalloc_track_caller);
3637
3638#else
3639void *__kmalloc(size_t size, gfp_t flags)
3640{
3641 return __do_kmalloc(size, flags, 0);
3642}
3643EXPORT_SYMBOL(__kmalloc);
3644#endif
3645
3646/**
3647 * kmem_cache_free - Deallocate an object
3648 * @cachep: The cache the allocation was from.
3649 * @objp: The previously allocated object.
3650 *
3651 * Free an object which was previously allocated from this
3652 * cache.
3653 */
3654void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3655{
3656 unsigned long flags;
3657 cachep = cache_from_obj(cachep, objp);
3658 if (!cachep)
3659 return;
3660
3661 local_irq_save(flags);
3662 debug_check_no_locks_freed(objp, cachep->object_size);
3663 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3664 debug_check_no_obj_freed(objp, cachep->object_size);
3665 __cache_free(cachep, objp, _RET_IP_);
3666 local_irq_restore(flags);
3667
3668 trace_kmem_cache_free(_RET_IP_, objp);
3669}
3670EXPORT_SYMBOL(kmem_cache_free);
3671
3672/**
3673 * kfree - free previously allocated memory
3674 * @objp: pointer returned by kmalloc.
3675 *
3676 * If @objp is NULL, no operation is performed.
3677 *
3678 * Don't free memory not originally allocated by kmalloc()
3679 * or you will run into trouble.
3680 */
3681void kfree(const void *objp)
3682{
3683 struct kmem_cache *c;
3684 unsigned long flags;
3685
3686 trace_kfree(_RET_IP_, objp);
3687
3688 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3689 return;
3690 local_irq_save(flags);
3691 kfree_debugcheck(objp);
3692 c = virt_to_cache(objp);
3693 debug_check_no_locks_freed(objp, c->object_size);
3694
3695 debug_check_no_obj_freed(objp, c->object_size);
3696 __cache_free(c, (void *)objp, _RET_IP_);
3697 local_irq_restore(flags);
3698}
3699EXPORT_SYMBOL(kfree);
3700
3701/*
3702 * This initializes kmem_cache_node or resizes various caches for all nodes.
3703 */
3704static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3705{
3706 int node;
3707 struct kmem_cache_node *n;
3708 struct array_cache *new_shared;
3709 struct array_cache **new_alien = NULL;
3710
3711 for_each_online_node(node) {
3712
3713 if (use_alien_caches) {
3714 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3715 if (!new_alien)
3716 goto fail;
3717 }
3718
3719 new_shared = NULL;
3720 if (cachep->shared) {
3721 new_shared = alloc_arraycache(node,
3722 cachep->shared*cachep->batchcount,
3723 0xbaadf00d, gfp);
3724 if (!new_shared) {
3725 free_alien_cache(new_alien);
3726 goto fail;
3727 }
3728 }
3729
3730 n = cachep->node[node];
3731 if (n) {
3732 struct array_cache *shared = n->shared;
3733
3734 spin_lock_irq(&n->list_lock);
3735
3736 if (shared)
3737 free_block(cachep, shared->entry,
3738 shared->avail, node);
3739
3740 n->shared = new_shared;
3741 if (!n->alien) {
3742 n->alien = new_alien;
3743 new_alien = NULL;
3744 }
3745 n->free_limit = (1 + nr_cpus_node(node)) *
3746 cachep->batchcount + cachep->num;
3747 spin_unlock_irq(&n->list_lock);
3748 kfree(shared);
3749 free_alien_cache(new_alien);
3750 continue;
3751 }
3752 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3753 if (!n) {
3754 free_alien_cache(new_alien);
3755 kfree(new_shared);
3756 goto fail;
3757 }
3758
3759 kmem_cache_node_init(n);
3760 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3761 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3762 n->shared = new_shared;
3763 n->alien = new_alien;
3764 n->free_limit = (1 + nr_cpus_node(node)) *
3765 cachep->batchcount + cachep->num;
3766 cachep->node[node] = n;
3767 }
3768 return 0;
3769
3770fail:
3771 if (!cachep->list.next) {
3772 /* Cache is not active yet. Roll back what we did */
3773 node--;
3774 while (node >= 0) {
3775 if (cachep->node[node]) {
3776 n = cachep->node[node];
3777
3778 kfree(n->shared);
3779 free_alien_cache(n->alien);
3780 kfree(n);
3781 cachep->node[node] = NULL;
3782 }
3783 node--;
3784 }
3785 }
3786 return -ENOMEM;
3787}
3788
3789struct ccupdate_struct {
3790 struct kmem_cache *cachep;
3791 struct array_cache *new[0];
3792};
3793
3794static void do_ccupdate_local(void *info)
3795{
3796 struct ccupdate_struct *new = info;
3797 struct array_cache *old;
3798
3799 check_irq_off();
3800 old = cpu_cache_get(new->cachep);
3801
3802 new->cachep->array[smp_processor_id()] = new->new[smp_processor_id()];
3803 new->new[smp_processor_id()] = old;
3804}
3805
3806/* Always called with the slab_mutex held */
3807static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3808 int batchcount, int shared, gfp_t gfp)
3809{
3810 struct ccupdate_struct *new;
3811 int i;
3812
3813 new = kzalloc(sizeof(*new) + nr_cpu_ids * sizeof(struct array_cache *),
3814 gfp);
3815 if (!new)
3816 return -ENOMEM;
3817
3818 for_each_online_cpu(i) {
3819 new->new[i] = alloc_arraycache(cpu_to_mem(i), limit,
3820 batchcount, gfp);
3821 if (!new->new[i]) {
3822 for (i--; i >= 0; i--)
3823 kfree(new->new[i]);
3824 kfree(new);
3825 return -ENOMEM;
3826 }
3827 }
3828 new->cachep = cachep;
3829
3830 on_each_cpu(do_ccupdate_local, (void *)new, 1);
3831
3832 check_irq_on();
3833 cachep->batchcount = batchcount;
3834 cachep->limit = limit;
3835 cachep->shared = shared;
3836
3837 for_each_online_cpu(i) {
3838 struct array_cache *ccold = new->new[i];
3839 if (!ccold)
3840 continue;
3841 spin_lock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3842 free_block(cachep, ccold->entry, ccold->avail, cpu_to_mem(i));
3843 spin_unlock_irq(&cachep->node[cpu_to_mem(i)]->list_lock);
3844 kfree(ccold);
3845 }
3846 kfree(new);
3847 return alloc_kmem_cache_node(cachep, gfp);
3848}
3849
3850static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3851 int batchcount, int shared, gfp_t gfp)
3852{
3853 int ret;
3854 struct kmem_cache *c = NULL;
3855 int i = 0;
3856
3857 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3858
3859 if (slab_state < FULL)
3860 return ret;
3861
3862 if ((ret < 0) || !is_root_cache(cachep))
3863 return ret;
3864
3865 VM_BUG_ON(!mutex_is_locked(&slab_mutex));
3866 for_each_memcg_cache_index(i) {
3867 c = cache_from_memcg_idx(cachep, i);
3868 if (c)
3869 /* return value determined by the parent cache only */
3870 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3871 }
3872
3873 return ret;
3874}
3875
3876/* Called with slab_mutex held always */
3877static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3878{
3879 int err;
3880 int limit = 0;
3881 int shared = 0;
3882 int batchcount = 0;
3883
3884 if (!is_root_cache(cachep)) {
3885 struct kmem_cache *root = memcg_root_cache(cachep);
3886 limit = root->limit;
3887 shared = root->shared;
3888 batchcount = root->batchcount;
3889 }
3890
3891 if (limit && shared && batchcount)
3892 goto skip_setup;
3893 /*
3894 * The head array serves three purposes:
3895 * - create a LIFO ordering, i.e. return objects that are cache-warm
3896 * - reduce the number of spinlock operations.
3897 * - reduce the number of linked list operations on the slab and
3898 * bufctl chains: array operations are cheaper.
3899 * The numbers are guessed, we should auto-tune as described by
3900 * Bonwick.
3901 */
3902 if (cachep->size > 131072)
3903 limit = 1;
3904 else if (cachep->size > PAGE_SIZE)
3905 limit = 8;
3906 else if (cachep->size > 1024)
3907 limit = 24;
3908 else if (cachep->size > 256)
3909 limit = 54;
3910 else
3911 limit = 120;
3912
3913 /*
3914 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3915 * allocation behaviour: Most allocs on one cpu, most free operations
3916 * on another cpu. For these cases, an efficient object passing between
3917 * cpus is necessary. This is provided by a shared array. The array
3918 * replaces Bonwick's magazine layer.
3919 * On uniprocessor, it's functionally equivalent (but less efficient)
3920 * to a larger limit. Thus disabled by default.
3921 */
3922 shared = 0;
3923 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3924 shared = 8;
3925
3926#if DEBUG
3927 /*
3928 * With debugging enabled, large batchcount lead to excessively long
3929 * periods with disabled local interrupts. Limit the batchcount
3930 */
3931 if (limit > 32)
3932 limit = 32;
3933#endif
3934 batchcount = (limit + 1) / 2;
3935skip_setup:
3936 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3937 if (err)
3938 printk(KERN_ERR "enable_cpucache failed for %s, error %d.\n",
3939 cachep->name, -err);
3940 return err;
3941}
3942
3943/*
3944 * Drain an array if it contains any elements taking the node lock only if
3945 * necessary. Note that the node listlock also protects the array_cache
3946 * if drain_array() is used on the shared array.
3947 */
3948static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3949 struct array_cache *ac, int force, int node)
3950{
3951 int tofree;
3952
3953 if (!ac || !ac->avail)
3954 return;
3955 if (ac->touched && !force) {
3956 ac->touched = 0;
3957 } else {
3958 spin_lock_irq(&n->list_lock);
3959 if (ac->avail) {
3960 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3961 if (tofree > ac->avail)
3962 tofree = (ac->avail + 1) / 2;
3963 free_block(cachep, ac->entry, tofree, node);
3964 ac->avail -= tofree;
3965 memmove(ac->entry, &(ac->entry[tofree]),
3966 sizeof(void *) * ac->avail);
3967 }
3968 spin_unlock_irq(&n->list_lock);
3969 }
3970}
3971
3972/**
3973 * cache_reap - Reclaim memory from caches.
3974 * @w: work descriptor
3975 *
3976 * Called from workqueue/eventd every few seconds.
3977 * Purpose:
3978 * - clear the per-cpu caches for this CPU.
3979 * - return freeable pages to the main free memory pool.
3980 *
3981 * If we cannot acquire the cache chain mutex then just give up - we'll try
3982 * again on the next iteration.
3983 */
3984static void cache_reap(struct work_struct *w)
3985{
3986 struct kmem_cache *searchp;
3987 struct kmem_cache_node *n;
3988 int node = numa_mem_id();
3989 struct delayed_work *work = to_delayed_work(w);
3990
3991 if (!mutex_trylock(&slab_mutex))
3992 /* Give up. Setup the next iteration. */
3993 goto out;
3994
3995 list_for_each_entry(searchp, &slab_caches, list) {
3996 check_irq_on();
3997
3998 /*
3999 * We only take the node lock if absolutely necessary and we
4000 * have established with reasonable certainty that
4001 * we can do some work if the lock was obtained.
4002 */
4003 n = searchp->node[node];
4004
4005 reap_alien(searchp, n);
4006
4007 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
4008
4009 /*
4010 * These are racy checks but it does not matter
4011 * if we skip one check or scan twice.
4012 */
4013 if (time_after(n->next_reap, jiffies))
4014 goto next;
4015
4016 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4017
4018 drain_array(searchp, n, n->shared, 0, node);
4019
4020 if (n->free_touched)
4021 n->free_touched = 0;
4022 else {
4023 int freed;
4024
4025 freed = drain_freelist(searchp, n, (n->free_limit +
4026 5 * searchp->num - 1) / (5 * searchp->num));
4027 STATS_ADD_REAPED(searchp, freed);
4028 }
4029next:
4030 cond_resched();
4031 }
4032 check_irq_on();
4033 mutex_unlock(&slab_mutex);
4034 next_reap_node();
4035out:
4036 /* Set up the next iteration */
4037 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4038}
4039
4040#ifdef CONFIG_SLABINFO
4041void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4042{
4043 struct page *page;
4044 unsigned long active_objs;
4045 unsigned long num_objs;
4046 unsigned long active_slabs = 0;
4047 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4048 const char *name;
4049 char *error = NULL;
4050 int node;
4051 struct kmem_cache_node *n;
4052
4053 active_objs = 0;
4054 num_slabs = 0;
4055 for_each_online_node(node) {
4056 n = cachep->node[node];
4057 if (!n)
4058 continue;
4059
4060 check_irq_on();
4061 spin_lock_irq(&n->list_lock);
4062
4063 list_for_each_entry(page, &n->slabs_full, lru) {
4064 if (page->active != cachep->num && !error)
4065 error = "slabs_full accounting error";
4066 active_objs += cachep->num;
4067 active_slabs++;
4068 }
4069 list_for_each_entry(page, &n->slabs_partial, lru) {
4070 if (page->active == cachep->num && !error)
4071 error = "slabs_partial accounting error";
4072 if (!page->active && !error)
4073 error = "slabs_partial accounting error";
4074 active_objs += page->active;
4075 active_slabs++;
4076 }
4077 list_for_each_entry(page, &n->slabs_free, lru) {
4078 if (page->active && !error)
4079 error = "slabs_free accounting error";
4080 num_slabs++;
4081 }
4082 free_objects += n->free_objects;
4083 if (n->shared)
4084 shared_avail += n->shared->avail;
4085
4086 spin_unlock_irq(&n->list_lock);
4087 }
4088 num_slabs += active_slabs;
4089 num_objs = num_slabs * cachep->num;
4090 if (num_objs - active_objs != free_objects && !error)
4091 error = "free_objects accounting error";
4092
4093 name = cachep->name;
4094 if (error)
4095 printk(KERN_ERR "slab: cache %s error: %s\n", name, error);
4096
4097 sinfo->active_objs = active_objs;
4098 sinfo->num_objs = num_objs;
4099 sinfo->active_slabs = active_slabs;
4100 sinfo->num_slabs = num_slabs;
4101 sinfo->shared_avail = shared_avail;
4102 sinfo->limit = cachep->limit;
4103 sinfo->batchcount = cachep->batchcount;
4104 sinfo->shared = cachep->shared;
4105 sinfo->objects_per_slab = cachep->num;
4106 sinfo->cache_order = cachep->gfporder;
4107}
4108
4109void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4110{
4111#if STATS
4112 { /* node stats */
4113 unsigned long high = cachep->high_mark;
4114 unsigned long allocs = cachep->num_allocations;
4115 unsigned long grown = cachep->grown;
4116 unsigned long reaped = cachep->reaped;
4117 unsigned long errors = cachep->errors;
4118 unsigned long max_freeable = cachep->max_freeable;
4119 unsigned long node_allocs = cachep->node_allocs;
4120 unsigned long node_frees = cachep->node_frees;
4121 unsigned long overflows = cachep->node_overflow;
4122
4123 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu "
4124 "%4lu %4lu %4lu %4lu %4lu",
4125 allocs, high, grown,
4126 reaped, errors, max_freeable, node_allocs,
4127 node_frees, overflows);
4128 }
4129 /* cpu stats */
4130 {
4131 unsigned long allochit = atomic_read(&cachep->allochit);
4132 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4133 unsigned long freehit = atomic_read(&cachep->freehit);
4134 unsigned long freemiss = atomic_read(&cachep->freemiss);
4135
4136 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4137 allochit, allocmiss, freehit, freemiss);
4138 }
4139#endif
4140}
4141
4142#define MAX_SLABINFO_WRITE 128
4143/**
4144 * slabinfo_write - Tuning for the slab allocator
4145 * @file: unused
4146 * @buffer: user buffer
4147 * @count: data length
4148 * @ppos: unused
4149 */
4150ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4151 size_t count, loff_t *ppos)
4152{
4153 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4154 int limit, batchcount, shared, res;
4155 struct kmem_cache *cachep;
4156
4157 if (count > MAX_SLABINFO_WRITE)
4158 return -EINVAL;
4159 if (copy_from_user(&kbuf, buffer, count))
4160 return -EFAULT;
4161 kbuf[MAX_SLABINFO_WRITE] = '\0';
4162
4163 tmp = strchr(kbuf, ' ');
4164 if (!tmp)
4165 return -EINVAL;
4166 *tmp = '\0';
4167 tmp++;
4168 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4169 return -EINVAL;
4170
4171 /* Find the cache in the chain of caches. */
4172 mutex_lock(&slab_mutex);
4173 res = -EINVAL;
4174 list_for_each_entry(cachep, &slab_caches, list) {
4175 if (!strcmp(cachep->name, kbuf)) {
4176 if (limit < 1 || batchcount < 1 ||
4177 batchcount > limit || shared < 0) {
4178 res = 0;
4179 } else {
4180 res = do_tune_cpucache(cachep, limit,
4181 batchcount, shared,
4182 GFP_KERNEL);
4183 }
4184 break;
4185 }
4186 }
4187 mutex_unlock(&slab_mutex);
4188 if (res >= 0)
4189 res = count;
4190 return res;
4191}
4192
4193#ifdef CONFIG_DEBUG_SLAB_LEAK
4194
4195static void *leaks_start(struct seq_file *m, loff_t *pos)
4196{
4197 mutex_lock(&slab_mutex);
4198 return seq_list_start(&slab_caches, *pos);
4199}
4200
4201static inline int add_caller(unsigned long *n, unsigned long v)
4202{
4203 unsigned long *p;
4204 int l;
4205 if (!v)
4206 return 1;
4207 l = n[1];
4208 p = n + 2;
4209 while (l) {
4210 int i = l/2;
4211 unsigned long *q = p + 2 * i;
4212 if (*q == v) {
4213 q[1]++;
4214 return 1;
4215 }
4216 if (*q > v) {
4217 l = i;
4218 } else {
4219 p = q + 2;
4220 l -= i + 1;
4221 }
4222 }
4223 if (++n[1] == n[0])
4224 return 0;
4225 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4226 p[0] = v;
4227 p[1] = 1;
4228 return 1;
4229}
4230
4231static void handle_slab(unsigned long *n, struct kmem_cache *c,
4232 struct page *page)
4233{
4234 void *p;
4235 int i, j;
4236
4237 if (n[0] == n[1])
4238 return;
4239 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4240 bool active = true;
4241
4242 for (j = page->active; j < c->num; j++) {
4243 /* Skip freed item */
4244 if (get_free_obj(page, j) == i) {
4245 active = false;
4246 break;
4247 }
4248 }
4249 if (!active)
4250 continue;
4251
4252 if (!add_caller(n, (unsigned long)*dbg_userword(c, p)))
4253 return;
4254 }
4255}
4256
4257static void show_symbol(struct seq_file *m, unsigned long address)
4258{
4259#ifdef CONFIG_KALLSYMS
4260 unsigned long offset, size;
4261 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4262
4263 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4264 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4265 if (modname[0])
4266 seq_printf(m, " [%s]", modname);
4267 return;
4268 }
4269#endif
4270 seq_printf(m, "%p", (void *)address);
4271}
4272
4273static int leaks_show(struct seq_file *m, void *p)
4274{
4275 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4276 struct page *page;
4277 struct kmem_cache_node *n;
4278 const char *name;
4279 unsigned long *x = m->private;
4280 int node;
4281 int i;
4282
4283 if (!(cachep->flags & SLAB_STORE_USER))
4284 return 0;
4285 if (!(cachep->flags & SLAB_RED_ZONE))
4286 return 0;
4287
4288 /* OK, we can do it */
4289
4290 x[1] = 0;
4291
4292 for_each_online_node(node) {
4293 n = cachep->node[node];
4294 if (!n)
4295 continue;
4296
4297 check_irq_on();
4298 spin_lock_irq(&n->list_lock);
4299
4300 list_for_each_entry(page, &n->slabs_full, lru)
4301 handle_slab(x, cachep, page);
4302 list_for_each_entry(page, &n->slabs_partial, lru)
4303 handle_slab(x, cachep, page);
4304 spin_unlock_irq(&n->list_lock);
4305 }
4306 name = cachep->name;
4307 if (x[0] == x[1]) {
4308 /* Increase the buffer size */
4309 mutex_unlock(&slab_mutex);
4310 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4311 if (!m->private) {
4312 /* Too bad, we are really out */
4313 m->private = x;
4314 mutex_lock(&slab_mutex);
4315 return -ENOMEM;
4316 }
4317 *(unsigned long *)m->private = x[0] * 2;
4318 kfree(x);
4319 mutex_lock(&slab_mutex);
4320 /* Now make sure this entry will be retried */
4321 m->count = m->size;
4322 return 0;
4323 }
4324 for (i = 0; i < x[1]; i++) {
4325 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4326 show_symbol(m, x[2*i+2]);
4327 seq_putc(m, '\n');
4328 }
4329
4330 return 0;
4331}
4332
4333static const struct seq_operations slabstats_op = {
4334 .start = leaks_start,
4335 .next = slab_next,
4336 .stop = slab_stop,
4337 .show = leaks_show,
4338};
4339
4340static int slabstats_open(struct inode *inode, struct file *file)
4341{
4342 unsigned long *n = kzalloc(PAGE_SIZE, GFP_KERNEL);
4343 int ret = -ENOMEM;
4344 if (n) {
4345 ret = seq_open(file, &slabstats_op);
4346 if (!ret) {
4347 struct seq_file *m = file->private_data;
4348 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4349 m->private = n;
4350 n = NULL;
4351 }
4352 kfree(n);
4353 }
4354 return ret;
4355}
4356
4357static const struct file_operations proc_slabstats_operations = {
4358 .open = slabstats_open,
4359 .read = seq_read,
4360 .llseek = seq_lseek,
4361 .release = seq_release_private,
4362};
4363#endif
4364
4365static int __init slab_proc_init(void)
4366{
4367#ifdef CONFIG_DEBUG_SLAB_LEAK
4368 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4369#endif
4370 return 0;
4371}
4372module_init(slab_proc_init);
4373#endif
4374
4375/**
4376 * ksize - get the actual amount of memory allocated for a given object
4377 * @objp: Pointer to the object
4378 *
4379 * kmalloc may internally round up allocations and return more memory
4380 * than requested. ksize() can be used to determine the actual amount of
4381 * memory allocated. The caller may use this additional memory, even though
4382 * a smaller amount of memory was initially specified with the kmalloc call.
4383 * The caller must guarantee that objp points to a valid object previously
4384 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4385 * must not be freed during the duration of the call.
4386 */
4387size_t ksize(const void *objp)
4388{
4389 BUG_ON(!objp);
4390 if (unlikely(objp == ZERO_SIZE_PTR))
4391 return 0;
4392
4393 return virt_to_cache(objp)->object_size;
4394}
4395EXPORT_SYMBOL(ksize);
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/__KEEPIDENTS__B.h>
90#include <linux/__KEEPIDENTS__C.h>
91#include <linux/__KEEPIDENTS__D.h>
92#include <linux/__KEEPIDENTS__E.h>
93#include <linux/__KEEPIDENTS__F.h>
94#include <linux/__KEEPIDENTS__G.h>
95#include <linux/__KEEPIDENTS__H.h>
96#include <linux/__KEEPIDENTS__I.h>
97#include <linux/__KEEPIDENTS__J.h>
98#include <linux/proc_fs.h>
99#include <linux/__KEEPIDENTS__BA.h>
100#include <linux/__KEEPIDENTS__BB.h>
101#include <linux/__KEEPIDENTS__BC.h>
102#include <linux/cpu.h>
103#include <linux/__KEEPIDENTS__BD.h>
104#include <linux/__KEEPIDENTS__BE.h>
105#include <linux/rcupdate.h>
106#include <linux/__KEEPIDENTS__BF.h>
107#include <linux/__KEEPIDENTS__BG.h>
108#include <linux/__KEEPIDENTS__BH.h>
109#include <linux/kmemleak.h>
110#include <linux/__KEEPIDENTS__BI.h>
111#include <linux/__KEEPIDENTS__BJ.h>
112#include <linux/__KEEPIDENTS__CA-__KEEPIDENTS__CB.h>
113#include <linux/__KEEPIDENTS__CC.h>
114#include <linux/reciprocal_div.h>
115#include <linux/debugobjects.h>
116#include <linux/kmemcheck.h>
117#include <linux/__KEEPIDENTS__CD.h>
118#include <linux/__KEEPIDENTS__CE.h>
119
120#include <net/__KEEPIDENTS__CF.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
164typedef unsigned char freelist_idx_t;
165#else
166typedef unsigned short freelist_idx_t;
167#endif
168
169#define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
170
171/*
172 * struct array_cache
173 *
174 * Purpose:
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
178 *
179 * The limit is stored in the per-cpu structure to reduce the data cache
180 * footprint.
181 *
182 */
183struct array_cache {
184 unsigned int avail;
185 unsigned int limit;
186 unsigned int batchcount;
187 unsigned int touched;
188 void *entry[]; /*
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
191 * the entries.
192 */
193};
194
195struct alien_cache {
196 spinlock_t lock;
197 struct array_cache ac;
198};
199
200/*
201 * Need this for bootstrapping a per node allocator.
202 */
203#define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
205#define CACHE_CACHE 0
206#define SIZE_NODE (MAX_NUMNODES)
207
208static int drain_freelist(struct kmem_cache *cache,
209 struct kmem_cache_node *n, int tofree);
210static void free_block(struct kmem_cache *cachep, void **objpp, int len,
211 int node, struct list_head *list);
212static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
213static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
214static void cache_reap(struct work_struct *unused);
215
216static int slab_early_init = 1;
217
218#define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
219
220static void kmem_cache_node_init(struct kmem_cache_node *parent)
221{
222 INIT_LIST_HEAD(&parent->slabs_full);
223 INIT_LIST_HEAD(&parent->slabs_partial);
224 INIT_LIST_HEAD(&parent->slabs_free);
225 parent->shared = NULL;
226 parent->alien = NULL;
227 parent->colour_next = 0;
228 spin_lock_init(&parent->list_lock);
229 parent->free_objects = 0;
230 parent->free_touched = 0;
231}
232
233#define MAKE_LIST(cachep, listp, slab, nodeid) \
234 do { \
235 INIT_LIST_HEAD(listp); \
236 list_splice(&get_node(cachep, nodeid)->slab, listp); \
237 } while (0)
238
239#define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
240 do { \
241 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
242 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
243 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
244 } while (0)
245
246#define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
247#define CFLGS_OFF_SLAB (0x80000000UL)
248#define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
249#define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
250
251#define BATCHREFILL_LIMIT 16
252/*
253 * Optimization question: fewer reaps means less probability for unnessary
254 * cpucache drain/refill cycles.
255 *
256 * OTOH the cpuarrays can contain lots of objects,
257 * which could lock up otherwise freeable slabs.
258 */
259#define REAPTIMEOUT_AC (2*HZ)
260#define REAPTIMEOUT_NODE (4*HZ)
261
262#if STATS
263#define STATS_INC_ACTIVE(x) ((x)->num_active++)
264#define STATS_DEC_ACTIVE(x) ((x)->num_active--)
265#define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
266#define STATS_INC_GROWN(x) ((x)->grown++)
267#define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
268#define STATS_SET_HIGH(x) \
269 do { \
270 if ((x)->num_active > (x)->high_mark) \
271 (x)->high_mark = (x)->num_active; \
272 } while (0)
273#define STATS_INC_ERR(x) ((x)->errors++)
274#define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
275#define STATS_INC_NODEFREES(x) ((x)->node_frees++)
276#define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
277#define STATS_SET_FREEABLE(x, i) \
278 do { \
279 if ((x)->max_freeable < i) \
280 (x)->max_freeable = i; \
281 } while (0)
282#define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
283#define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
284#define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
285#define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
286#else
287#define STATS_INC_ACTIVE(x) do { } while (0)
288#define STATS_DEC_ACTIVE(x) do { } while (0)
289#define STATS_INC_ALLOCED(x) do { } while (0)
290#define STATS_INC_GROWN(x) do { } while (0)
291#define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
292#define STATS_SET_HIGH(x) do { } while (0)
293#define STATS_INC_ERR(x) do { } while (0)
294#define STATS_INC_NODEALLOCS(x) do { } while (0)
295#define STATS_INC_NODEFREES(x) do { } while (0)
296#define STATS_INC_ACOVERFLOW(x) do { } while (0)
297#define STATS_SET_FREEABLE(x, i) do { } while (0)
298#define STATS_INC_ALLOCHIT(x) do { } while (0)
299#define STATS_INC_ALLOCMISS(x) do { } while (0)
300#define STATS_INC_FREEHIT(x) do { } while (0)
301#define STATS_INC_FREEMISS(x) do { } while (0)
302#endif
303
304#if DEBUG
305
306/*
307 * memory layout of objects:
308 * 0 : objp
309 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
310 * the end of an object is aligned with the end of the real
311 * allocation. Catches writes behind the end of the allocation.
312 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
313 * redzone word.
314 * cachep->obj_offset: The real object.
315 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
316 * cachep->size - 1* BYTES_PER_WORD: last caller address
317 * [BYTES_PER_WORD long]
318 */
319static int obj_offset(struct kmem_cache *cachep)
320{
321 return cachep->obj_offset;
322}
323
324static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
325{
326 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
327 return (unsigned long long*) (objp + obj_offset(cachep) -
328 sizeof(unsigned long long));
329}
330
331static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
332{
333 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
334 if (cachep->flags & SLAB_STORE_USER)
335 return (unsigned long long *)(objp + cachep->size -
336 sizeof(unsigned long long) -
337 REDZONE_ALIGN);
338 return (unsigned long long *) (objp + cachep->size -
339 sizeof(unsigned long long));
340}
341
342static void **dbg_userword(struct kmem_cache *cachep, void *objp)
343{
344 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
345 return (void **)(objp + cachep->size - BYTES_PER_WORD);
346}
347
348#else
349
350#define obj_offset(x) 0
351#define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
352#define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
353#define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
354
355#endif
356
357#ifdef CONFIG_DEBUG_SLAB_LEAK
358
359static inline bool is_store_user_clean(struct kmem_cache *cachep)
360{
361 return atomic_read(&cachep->store_user_clean) == 1;
362}
363
364static inline void set_store_user_clean(struct kmem_cache *cachep)
365{
366 atomic_set(&cachep->store_user_clean, 1);
367}
368
369static inline void set_store_user_dirty(struct kmem_cache *cachep)
370{
371 if (is_store_user_clean(cachep))
372 atomic_set(&cachep->store_user_clean, 0);
373}
374
375#else
376static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
377
378#endif
379
380/*
381 * Do not go above this order unless 0 objects fit into the slab or
382 * overridden on the command line.
383 */
384#define SLAB_MAX_ORDER_HI 1
385#define SLAB_MAX_ORDER_LO 0
386static int slab_max_order = SLAB_MAX_ORDER_LO;
387static bool slab_max_order_set __initdata;
388
389static inline struct kmem_cache *virt_to_cache(const void *obj)
390{
391 struct page *page = virt_to_head_page(obj);
392 return page->slab_cache;
393}
394
395static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
396 unsigned int idx)
397{
398 return page->s_mem + cache->size * idx;
399}
400
401/*
402 * We want to avoid an expensive divide : (offset / cache->size)
403 * Using the fact that size is a constant for a particular cache,
404 * we can replace (offset / cache->size) by
405 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
406 */
407static inline unsigned int obj_to_index(const struct kmem_cache *cache,
408 const struct page *page, void *obj)
409{
410 u32 offset = (obj - page->s_mem);
411 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
412}
413
414#define BOOT_CPUCACHE_ENTRIES 1
415/* internal cache of cache description objs */
416static struct kmem_cache kmem_cache_boot = {
417 .batchcount = 1,
418 .limit = BOOT_CPUCACHE_ENTRIES,
419 .shared = 1,
420 .size = sizeof(struct kmem_cache),
421 .name = "kmem_cache",
422};
423
424#define BAD_ALIEN_MAGIC 0x01020304ul
425
426static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
427
428static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
429{
430 return this_cpu_ptr(cachep->cpu_cache);
431}
432
433/*
434 * Calculate the number of objects and left-over bytes for a given buffer size.
435 */
436static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
437 unsigned long flags, size_t *left_over)
438{
439 unsigned int num;
440 size_t slab_size = PAGE_SIZE << gfporder;
441
442 /*
443 * The slab management structure can be either off the slab or
444 * on it. For the latter case, the memory allocated for a
445 * slab is used for:
446 *
447 * - @buffer_size bytes for each object
448 * - One freelist_idx_t for each object
449 *
450 * We don't need to consider alignment of freelist because
451 * freelist will be at the end of slab page. The objects will be
452 * at the correct alignment.
453 *
454 * If the slab management structure is off the slab, then the
455 * alignment will already be calculated into the size. Because
456 * the slabs are all pages aligned, the objects will be at the
457 * correct alignment when allocated.
458 */
459 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
460 num = slab_size / buffer_size;
461 *left_over = slab_size % buffer_size;
462 } else {
463 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
464 *left_over = slab_size %
465 (buffer_size + sizeof(freelist_idx_t));
466 }
467
468 return num;
469}
470
471#if DEBUG
472#define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
473
474static void __slab_error(const char *function, struct kmem_cache *cachep,
475 char *msg)
476{
477 pr_err("slab error in %s(): cache `%s': %s\n",
478 function, cachep->name, msg);
479 dump_stack();
480 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
481}
482#endif
483
484/*
485 * By default on NUMA we use alien caches to stage the freeing of
486 * objects allocated from other nodes. This causes massive memory
487 * inefficiencies when using fake NUMA setup to split memory into a
488 * large number of small nodes, so it can be disabled on the command
489 * line
490 */
491
492static int use_alien_caches __read_mostly = 1;
493static int __init noaliencache_setup(char *s)
494{
495 use_alien_caches = 0;
496 return 1;
497}
498__setup("noaliencache", noaliencache_setup);
499
500static int __init slab_max_order_setup(char *str)
501{
502 get_option(&str, &slab_max_order);
503 slab_max_order = slab_max_order < 0 ? 0 :
504 min(slab_max_order, MAX_ORDER - 1);
505 slab_max_order_set = true;
506
507 return 1;
508}
509__setup("slab_max_order=", slab_max_order_setup);
510
511#ifdef CONFIG_NUMA
512/*
513 * Special reaping functions for NUMA systems called from cache_reap().
514 * These take care of doing round robin flushing of alien caches (containing
515 * objects freed on different nodes from which they were allocated) and the
516 * flushing of remote pcps by calling drain_node_pages.
517 */
518static DEFINE_PER_CPU(unsigned long, slab_reap_node);
519
520static void init_reap_node(int cpu)
521{
522 int node;
523
524 node = next_node(cpu_to_mem(cpu), node_online_map);
525 if (node == MAX_NUMNODES)
526 node = first_node(node_online_map);
527
528 per_cpu(slab_reap_node, cpu) = node;
529}
530
531static void next_reap_node(void)
532{
533 int node = __this_cpu_read(slab_reap_node);
534
535 node = next_node(node, node_online_map);
536 if (unlikely(node >= MAX_NUMNODES))
537 node = first_node(node_online_map);
538 __this_cpu_write(slab_reap_node, node);
539}
540
541#else
542#define init_reap_node(cpu) do { } while (0)
543#define next_reap_node(void) do { } while (0)
544#endif
545
546/*
547 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
548 * via the workqueue/eventd.
549 * Add the CPU number into the expiration time to minimize the possibility of
550 * the CPUs getting into lockstep and contending for the global cache chain
551 * lock.
552 */
553static void start_cpu_timer(int cpu)
554{
555 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
556
557 /*
558 * When this gets called from do_initcalls via cpucache_init(),
559 * init_workqueues() has already run, so keventd will be setup
560 * at that time.
561 */
562 if (keventd_up() && reap_work->work.func == NULL) {
563 init_reap_node(cpu);
564 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
565 schedule_delayed_work_on(cpu, reap_work,
566 __round_jiffies_relative(HZ, cpu));
567 }
568}
569
570static void init_arraycache(struct array_cache *ac, int limit, int batch)
571{
572 /*
573 * The array_cache structures contain pointers to free object.
574 * However, when such objects are allocated or transferred to another
575 * cache the pointers are not cleared and they could be counted as
576 * valid references during a kmemleak scan. Therefore, kmemleak must
577 * not scan such objects.
578 */
579 kmemleak_no_scan(ac);
580 if (ac) {
581 ac->avail = 0;
582 ac->limit = limit;
583 ac->batchcount = batch;
584 ac->touched = 0;
585 }
586}
587
588static struct array_cache *alloc_arraycache(int node, int entries,
589 int batchcount, gfp_t gfp)
590{
591 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
592 struct array_cache *ac = NULL;
593
594 ac = kmalloc_node(memsize, gfp, node);
595 init_arraycache(ac, entries, batchcount);
596 return ac;
597}
598
599static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
600 struct page *page, void *objp)
601{
602 struct kmem_cache_node *n;
603 int page_node;
604 LIST_HEAD(list);
605
606 page_node = page_to_nid(page);
607 n = get_node(cachep, page_node);
608
609 spin_lock(&n->list_lock);
610 free_block(cachep, &objp, 1, page_node, &list);
611 spin_unlock(&n->list_lock);
612
613 slabs_destroy(cachep, &list);
614}
615
616/*
617 * Transfer objects in one arraycache to another.
618 * Locking must be handled by the caller.
619 *
620 * Return the number of entries transferred.
621 */
622static int transfer_objects(struct array_cache *to,
623 struct array_cache *from, unsigned int max)
624{
625 /* Figure out how many entries to transfer */
626 int nr = min3(from->avail, max, to->limit - to->avail);
627
628 if (!nr)
629 return 0;
630
631 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
632 sizeof(void *) *nr);
633
634 from->avail -= nr;
635 to->avail += nr;
636 return nr;
637}
638
639#ifndef CONFIG_NUMA
640
641#define drain_alien_cache(cachep, alien) do { } while (0)
642#define reap_alien(cachep, n) do { } while (0)
643
644static inline struct alien_cache **alloc_alien_cache(int node,
645 int limit, gfp_t gfp)
646{
647 return (struct alien_cache **)BAD_ALIEN_MAGIC;
648}
649
650static inline void free_alien_cache(struct alien_cache **ac_ptr)
651{
652}
653
654static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
655{
656 return 0;
657}
658
659static inline void *alternate_node_alloc(struct kmem_cache *cachep,
660 gfp_t flags)
661{
662 return NULL;
663}
664
665static inline void *____cache_alloc_node(struct kmem_cache *cachep,
666 gfp_t flags, int nodeid)
667{
668 return NULL;
669}
670
671static inline gfp_t gfp_exact_node(gfp_t flags)
672{
673 return flags & ~__GFP_NOFAIL;
674}
675
676#else /* CONFIG_NUMA */
677
678static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
679static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
680
681static struct alien_cache *__alloc_alien_cache(int node, int entries,
682 int batch, gfp_t gfp)
683{
684 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
685 struct alien_cache *alc = NULL;
686
687 alc = kmalloc_node(memsize, gfp, node);
688 init_arraycache(&alc->ac, entries, batch);
689 spin_lock_init(&alc->lock);
690 return alc;
691}
692
693static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
694{
695 struct alien_cache **alc_ptr;
696 size_t memsize = sizeof(void *) * nr_node_ids;
697 int i;
698
699 if (limit > 1)
700 limit = 12;
701 alc_ptr = kzalloc_node(memsize, gfp, node);
702 if (!alc_ptr)
703 return NULL;
704
705 for_each_node(i) {
706 if (i == node || !node_online(i))
707 continue;
708 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
709 if (!alc_ptr[i]) {
710 for (i--; i >= 0; i--)
711 kfree(alc_ptr[i]);
712 kfree(alc_ptr);
713 return NULL;
714 }
715 }
716 return alc_ptr;
717}
718
719static void free_alien_cache(struct alien_cache **alc_ptr)
720{
721 int i;
722
723 if (!alc_ptr)
724 return;
725 for_each_node(i)
726 kfree(alc_ptr[i]);
727 kfree(alc_ptr);
728}
729
730static void __drain_alien_cache(struct kmem_cache *cachep,
731 struct array_cache *ac, int node,
732 struct list_head *list)
733{
734 struct kmem_cache_node *n = get_node(cachep, node);
735
736 if (ac->avail) {
737 spin_lock(&n->list_lock);
738 /*
739 * Stuff objects into the remote nodes shared array first.
740 * That way we could avoid the overhead of putting the objects
741 * into the free lists and getting them back later.
742 */
743 if (n->shared)
744 transfer_objects(n->shared, ac, ac->limit);
745
746 free_block(cachep, ac->entry, ac->avail, node, list);
747 ac->avail = 0;
748 spin_unlock(&n->list_lock);
749 }
750}
751
752/*
753 * Called from cache_reap() to regularly drain alien caches round robin.
754 */
755static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
756{
757 int node = __this_cpu_read(slab_reap_node);
758
759 if (n->alien) {
760 struct alien_cache *alc = n->alien[node];
761 struct array_cache *ac;
762
763 if (alc) {
764 ac = &alc->ac;
765 if (ac->avail && spin_trylock_irq(&alc->lock)) {
766 LIST_HEAD(list);
767
768 __drain_alien_cache(cachep, ac, node, &list);
769 spin_unlock_irq(&alc->lock);
770 slabs_destroy(cachep, &list);
771 }
772 }
773 }
774}
775
776static void drain_alien_cache(struct kmem_cache *cachep,
777 struct alien_cache **alien)
778{
779 int i = 0;
780 struct alien_cache *alc;
781 struct array_cache *ac;
782 unsigned long flags;
783
784 for_each_online_node(i) {
785 alc = alien[i];
786 if (alc) {
787 LIST_HEAD(list);
788
789 ac = &alc->ac;
790 spin_lock_irqsave(&alc->lock, flags);
791 __drain_alien_cache(cachep, ac, i, &list);
792 spin_unlock_irqrestore(&alc->lock, flags);
793 slabs_destroy(cachep, &list);
794 }
795 }
796}
797
798static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
799 int node, int page_node)
800{
801 struct kmem_cache_node *n;
802 struct alien_cache *alien = NULL;
803 struct array_cache *ac;
804 LIST_HEAD(list);
805
806 n = get_node(cachep, node);
807 STATS_INC_NODEFREES(cachep);
808 if (n->alien && n->alien[page_node]) {
809 alien = n->alien[page_node];
810 ac = &alien->ac;
811 spin_lock(&alien->lock);
812 if (unlikely(ac->avail == ac->limit)) {
813 STATS_INC_ACOVERFLOW(cachep);
814 __drain_alien_cache(cachep, ac, page_node, &list);
815 }
816 ac->entry[ac->avail++] = objp;
817 spin_unlock(&alien->lock);
818 slabs_destroy(cachep, &list);
819 } else {
820 n = get_node(cachep, page_node);
821 spin_lock(&n->list_lock);
822 free_block(cachep, &objp, 1, page_node, &list);
823 spin_unlock(&n->list_lock);
824 slabs_destroy(cachep, &list);
825 }
826 return 1;
827}
828
829static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
830{
831 int page_node = page_to_nid(virt_to_page(objp));
832 int node = numa_mem_id();
833 /*
834 * Make sure we are not freeing a object from another node to the array
835 * cache on this cpu.
836 */
837 if (likely(node == page_node))
838 return 0;
839
840 return __cache_free_alien(cachep, objp, node, page_node);
841}
842
843/*
844 * Construct gfp mask to allocate from a specific node but do not reclaim or
845 * warn about failures.
846 */
847static inline gfp_t gfp_exact_node(gfp_t flags)
848{
849 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
850}
851#endif
852
853/*
854 * Allocates and initializes node for a node on each slab cache, used for
855 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
856 * will be allocated off-node since memory is not yet online for the new node.
857 * When hotplugging memory or a cpu, existing node are not replaced if
858 * already in use.
859 *
860 * Must hold slab_mutex.
861 */
862static int init_cache_node_node(int node)
863{
864 struct kmem_cache *cachep;
865 struct kmem_cache_node *n;
866 const size_t memsize = sizeof(struct kmem_cache_node);
867
868 list_for_each_entry(cachep, &slab_caches, list) {
869 /*
870 * Set up the kmem_cache_node for cpu before we can
871 * begin anything. Make sure some other cpu on this
872 * node has not already allocated this
873 */
874 n = get_node(cachep, node);
875 if (!n) {
876 n = kmalloc_node(memsize, GFP_KERNEL, node);
877 if (!n)
878 return -ENOMEM;
879 kmem_cache_node_init(n);
880 n->next_reap = jiffies + REAPTIMEOUT_NODE +
881 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
882
883 /*
884 * The kmem_cache_nodes don't come and go as CPUs
885 * come and go. slab_mutex is sufficient
886 * protection here.
887 */
888 cachep->node[node] = n;
889 }
890
891 spin_lock_irq(&n->list_lock);
892 n->free_limit =
893 (1 + nr_cpus_node(node)) *
894 cachep->batchcount + cachep->num;
895 spin_unlock_irq(&n->list_lock);
896 }
897 return 0;
898}
899
900static inline int slabs_tofree(struct kmem_cache *cachep,
901 struct kmem_cache_node *n)
902{
903 return (n->free_objects + cachep->num - 1) / cachep->num;
904}
905
906static void cpuup_canceled(long cpu)
907{
908 struct kmem_cache *cachep;
909 struct kmem_cache_node *n = NULL;
910 int node = cpu_to_mem(cpu);
911 const struct cpumask *mask = cpumask_of_node(node);
912
913 list_for_each_entry(cachep, &slab_caches, list) {
914 struct array_cache *nc;
915 struct array_cache *shared;
916 struct alien_cache **alien;
917 LIST_HEAD(list);
918
919 n = get_node(cachep, node);
920 if (!n)
921 continue;
922
923 spin_lock_irq(&n->list_lock);
924
925 /* Free limit for this kmem_cache_node */
926 n->free_limit -= cachep->batchcount;
927
928 /* cpu is dead; no one can alloc from it. */
929 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
930 if (nc) {
931 free_block(cachep, nc->entry, nc->avail, node, &list);
932 nc->avail = 0;
933 }
934
935 if (!cpumask_empty(mask)) {
936 spin_unlock_irq(&n->list_lock);
937 goto free_slab;
938 }
939
940 shared = n->shared;
941 if (shared) {
942 free_block(cachep, shared->entry,
943 shared->avail, node, &list);
944 n->shared = NULL;
945 }
946
947 alien = n->alien;
948 n->alien = NULL;
949
950 spin_unlock_irq(&n->list_lock);
951
952 kfree(shared);
953 if (alien) {
954 drain_alien_cache(cachep, alien);
955 free_alien_cache(alien);
956 }
957
958free_slab:
959 slabs_destroy(cachep, &list);
960 }
961 /*
962 * In the previous loop, all the objects were freed to
963 * the respective cache's slabs, now we can go ahead and
964 * shrink each nodelist to its limit.
965 */
966 list_for_each_entry(cachep, &slab_caches, list) {
967 n = get_node(cachep, node);
968 if (!n)
969 continue;
970 drain_freelist(cachep, n, slabs_tofree(cachep, n));
971 }
972}
973
974static int cpuup_prepare(long cpu)
975{
976 struct kmem_cache *cachep;
977 struct kmem_cache_node *n = NULL;
978 int node = cpu_to_mem(cpu);
979 int err;
980
981 /*
982 * We need to do this right in the beginning since
983 * alloc_arraycache's are going to use this list.
984 * kmalloc_node allows us to add the slab to the right
985 * kmem_cache_node and not this cpu's kmem_cache_node
986 */
987 err = init_cache_node_node(node);
988 if (err < 0)
989 goto bad;
990
991 /*
992 * Now we can go ahead with allocating the shared arrays and
993 * array caches
994 */
995 list_for_each_entry(cachep, &slab_caches, list) {
996 struct array_cache *shared = NULL;
997 struct alien_cache **alien = NULL;
998
999 if (cachep->shared) {
1000 shared = alloc_arraycache(node,
1001 cachep->shared * cachep->batchcount,
1002 0xbaadf00d, GFP_KERNEL);
1003 if (!shared)
1004 goto bad;
1005 }
1006 if (use_alien_caches) {
1007 alien = alloc_alien_cache(node, cachep->limit, GFP_KERNEL);
1008 if (!alien) {
1009 kfree(shared);
1010 goto bad;
1011 }
1012 }
1013 n = get_node(cachep, node);
1014 BUG_ON(!n);
1015
1016 spin_lock_irq(&n->list_lock);
1017 if (!n->shared) {
1018 /*
1019 * We are serialised from CPU_DEAD or
1020 * CPU_UP_CANCELLED by the cpucontrol lock
1021 */
1022 n->shared = shared;
1023 shared = NULL;
1024 }
1025#ifdef CONFIG_NUMA
1026 if (!n->alien) {
1027 n->alien = alien;
1028 alien = NULL;
1029 }
1030#endif
1031 spin_unlock_irq(&n->list_lock);
1032 kfree(shared);
1033 free_alien_cache(alien);
1034 }
1035
1036 return 0;
1037bad:
1038 cpuup_canceled(cpu);
1039 return -ENOMEM;
1040}
1041
1042static int cpuup_callback(struct notifier_block *nfb,
1043 unsigned long action, void *hcpu)
1044{
1045 long cpu = (long)hcpu;
1046 int err = 0;
1047
1048 switch (action) {
1049 case CPU_UP_PREPARE:
1050 case CPU_UP_PREPARE_FROZEN:
1051 mutex_lock(&slab_mutex);
1052 err = cpuup_prepare(cpu);
1053 mutex_unlock(&slab_mutex);
1054 break;
1055 case CPU_ONLINE:
1056 case CPU_ONLINE_FROZEN:
1057 start_cpu_timer(cpu);
1058 break;
1059#ifdef CONFIG_HOTPLUG_CPU
1060 case CPU_DOWN_PREPARE:
1061 case CPU_DOWN_PREPARE_FROZEN:
1062 /*
1063 * Shutdown cache reaper. Note that the slab_mutex is
1064 * held so that if cache_reap() is invoked it cannot do
1065 * anything expensive but will only modify reap_work
1066 * and reschedule the timer.
1067 */
1068 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1069 /* Now the cache_reaper is guaranteed to be not running. */
1070 per_cpu(slab_reap_work, cpu).work.func = NULL;
1071 break;
1072 case CPU_DOWN_FAILED:
1073 case CPU_DOWN_FAILED_FROZEN:
1074 start_cpu_timer(cpu);
1075 break;
1076 case CPU_DEAD:
1077 case CPU_DEAD_FROZEN:
1078 /*
1079 * Even if all the cpus of a node are down, we don't free the
1080 * kmem_cache_node of any cache. This to avoid a race between
1081 * cpu_down, and a kmalloc allocation from another cpu for
1082 * memory from the node of the cpu going down. The node
1083 * structure is usually allocated from kmem_cache_create() and
1084 * gets destroyed at kmem_cache_destroy().
1085 */
1086 /* fall through */
1087#endif
1088 case CPU_UP_CANCELED:
1089 case CPU_UP_CANCELED_FROZEN:
1090 mutex_lock(&slab_mutex);
1091 cpuup_canceled(cpu);
1092 mutex_unlock(&slab_mutex);
1093 break;
1094 }
1095 return notifier_from_errno(err);
1096}
1097
1098static struct notifier_block cpucache_notifier = {
1099 &cpuup_callback, NULL, 0
1100};
1101
1102#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1103/*
1104 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1105 * Returns -EBUSY if all objects cannot be drained so that the node is not
1106 * removed.
1107 *
1108 * Must hold slab_mutex.
1109 */
1110static int __meminit drain_cache_node_node(int node)
1111{
1112 struct kmem_cache *cachep;
1113 int ret = 0;
1114
1115 list_for_each_entry(cachep, &slab_caches, list) {
1116 struct kmem_cache_node *n;
1117
1118 n = get_node(cachep, node);
1119 if (!n)
1120 continue;
1121
1122 drain_freelist(cachep, n, slabs_tofree(cachep, n));
1123
1124 if (!list_empty(&n->slabs_full) ||
1125 !list_empty(&n->slabs_partial)) {
1126 ret = -EBUSY;
1127 break;
1128 }
1129 }
1130 return ret;
1131}
1132
1133static int __meminit slab_memory_callback(struct notifier_block *self,
1134 unsigned long action, void *arg)
1135{
1136 struct memory_notify *mnb = arg;
1137 int ret = 0;
1138 int nid;
1139
1140 nid = mnb->status_change_nid;
1141 if (nid < 0)
1142 goto out;
1143
1144 switch (action) {
1145 case MEM_GOING_ONLINE:
1146 mutex_lock(&slab_mutex);
1147 ret = init_cache_node_node(nid);
1148 mutex_unlock(&slab_mutex);
1149 break;
1150 case MEM_GOING_OFFLINE:
1151 mutex_lock(&slab_mutex);
1152 ret = drain_cache_node_node(nid);
1153 mutex_unlock(&slab_mutex);
1154 break;
1155 case MEM_ONLINE:
1156 case MEM_OFFLINE:
1157 case MEM_CANCEL_ONLINE:
1158 case MEM_CANCEL_OFFLINE:
1159 break;
1160 }
1161out:
1162 return notifier_from_errno(ret);
1163}
1164#endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1165
1166/*
1167 * swap the static kmem_cache_node with kmalloced memory
1168 */
1169static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1170 int nodeid)
1171{
1172 struct kmem_cache_node *ptr;
1173
1174 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1175 BUG_ON(!ptr);
1176
1177 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1178 /*
1179 * Do not assume that spinlocks can be initialized via memcpy:
1180 */
1181 spin_lock_init(&ptr->list_lock);
1182
1183 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1184 cachep->node[nodeid] = ptr;
1185}
1186
1187/*
1188 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1189 * size of kmem_cache_node.
1190 */
1191static void __init set_up_node(struct kmem_cache *cachep, int index)
1192{
1193 int node;
1194
1195 for_each_online_node(node) {
1196 cachep->node[node] = &init_kmem_cache_node[index + node];
1197 cachep->node[node]->next_reap = jiffies +
1198 REAPTIMEOUT_NODE +
1199 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1200 }
1201}
1202
1203/*
1204 * Initialisation. Called after the page allocator have been initialised and
1205 * before smp_init().
1206 */
1207void __init kmem_cache_init(void)
1208{
1209 int i;
1210
1211 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1212 sizeof(struct rcu_head));
1213 kmem_cache = &kmem_cache_boot;
1214
1215 if (num_possible_nodes() == 1)
1216 use_alien_caches = 0;
1217
1218 for (i = 0; i < NUM_INIT_LISTS; i++)
1219 kmem_cache_node_init(&init_kmem_cache_node[i]);
1220
1221 /*
1222 * Fragmentation resistance on low memory - only use bigger
1223 * page orders on machines with more than 32MB of memory if
1224 * not overridden on the command line.
1225 */
1226 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1227 slab_max_order = SLAB_MAX_ORDER_HI;
1228
1229 /* Bootstrap is tricky, because several objects are allocated
1230 * from caches that do not exist yet:
1231 * 1) initialize the kmem_cache cache: it contains the struct
1232 * kmem_cache structures of all caches, except kmem_cache itself:
1233 * kmem_cache is statically allocated.
1234 * Initially an __init data area is used for the head array and the
1235 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1236 * array at the end of the bootstrap.
1237 * 2) Create the first kmalloc cache.
1238 * The struct kmem_cache for the new cache is allocated normally.
1239 * An __init data area is used for the head array.
1240 * 3) Create the remaining kmalloc caches, with minimally sized
1241 * head arrays.
1242 * 4) Replace the __init data head arrays for kmem_cache and the first
1243 * kmalloc cache with kmalloc allocated arrays.
1244 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1245 * the other cache's with kmalloc allocated memory.
1246 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1247 */
1248
1249 /* 1) create the kmem_cache */
1250
1251 /*
1252 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1253 */
1254 create_boot_cache(kmem_cache, "kmem_cache",
1255 offsetof(struct kmem_cache, node) +
1256 nr_node_ids * sizeof(struct kmem_cache_node *),
1257 SLAB_HWCACHE_ALIGN);
1258 list_add(&kmem_cache->list, &slab_caches);
1259 slab_state = PARTIAL;
1260
1261 /*
1262 * Initialize the caches that provide memory for the kmem_cache_node
1263 * structures first. Without this, further allocations will bug.
1264 */
1265 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1266 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1267 slab_state = PARTIAL_NODE;
1268 setup_kmalloc_cache_index_table();
1269
1270 slab_early_init = 0;
1271
1272 /* 5) Replace the bootstrap kmem_cache_node */
1273 {
1274 int nid;
1275
1276 for_each_online_node(nid) {
1277 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1278
1279 init_list(kmalloc_caches[INDEX_NODE],
1280 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1281 }
1282 }
1283
1284 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1285}
1286
1287void __init kmem_cache_init_late(void)
1288{
1289 struct kmem_cache *cachep;
1290
1291 slab_state = UP;
1292
1293 /* 6) resize the head arrays to their final sizes */
1294 mutex_lock(&slab_mutex);
1295 list_for_each_entry(cachep, &slab_caches, list)
1296 if (enable_cpucache(cachep, GFP_NOWAIT))
1297 BUG();
1298 mutex_unlock(&slab_mutex);
1299
1300 /* Done! */
1301 slab_state = FULL;
1302
1303 /*
1304 * Register a cpu startup notifier callback that initializes
1305 * cpu_cache_get for all new cpus
1306 */
1307 register_cpu_notifier(&cpucache_notifier);
1308
1309#ifdef CONFIG_NUMA
1310 /*
1311 * Register a memory hotplug callback that initializes and frees
1312 * node.
1313 */
1314 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1315#endif
1316
1317 /*
1318 * The reap timers are started later, with a module init call: That part
1319 * of the kernel is not yet operational.
1320 */
1321}
1322
1323static int __init cpucache_init(void)
1324{
1325 int cpu;
1326
1327 /*
1328 * Register the timers that return unneeded pages to the page allocator
1329 */
1330 for_each_online_cpu(cpu)
1331 start_cpu_timer(cpu);
1332
1333 /* Done! */
1334 slab_state = FULL;
1335 return 0;
1336}
1337__initcall(cpucache_init);
1338
1339static noinline void
1340slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1341{
1342#if DEBUG
1343 struct kmem_cache_node *n;
1344 struct page *page;
1345 unsigned long flags;
1346 int node;
1347 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1348 DEFAULT_RATELIMIT_BURST);
1349
1350 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1351 return;
1352
1353 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1354 nodeid, gfpflags, &gfpflags);
1355 pr_warn(" cache: %s, object size: %d, order: %d\n",
1356 cachep->name, cachep->size, cachep->gfporder);
1357
1358 for_each_kmem_cache_node(cachep, node, n) {
1359 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1360 unsigned long active_slabs = 0, num_slabs = 0;
1361
1362 spin_lock_irqsave(&n->list_lock, flags);
1363 list_for_each_entry(page, &n->slabs_full, lru) {
1364 active_objs += cachep->num;
1365 active_slabs++;
1366 }
1367 list_for_each_entry(page, &n->slabs_partial, lru) {
1368 active_objs += page->active;
1369 active_slabs++;
1370 }
1371 list_for_each_entry(page, &n->slabs_free, lru)
1372 num_slabs++;
1373
1374 free_objects += n->free_objects;
1375 spin_unlock_irqrestore(&n->list_lock, flags);
1376
1377 num_slabs += active_slabs;
1378 num_objs = num_slabs * cachep->num;
1379 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1380 node, active_slabs, num_slabs, active_objs, num_objs,
1381 free_objects);
1382 }
1383#endif
1384}
1385
1386/*
1387 * Interface to system's page allocator. No need to hold the
1388 * kmem_cache_node ->list_lock.
1389 *
1390 * If we requested dmaable memory, we will get it. Even if we
1391 * did not request dmaable memory, we might get it, but that
1392 * would be relatively rare and ignorable.
1393 */
1394static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1395 int nodeid)
1396{
1397 struct page *page;
1398 int nr_pages;
1399
1400 flags |= cachep->allocflags;
1401 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1402 flags |= __GFP_RECLAIMABLE;
1403
1404 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1405 if (!page) {
1406 slab_out_of_memory(cachep, flags, nodeid);
1407 return NULL;
1408 }
1409
1410 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1411 __free_pages(page, cachep->gfporder);
1412 return NULL;
1413 }
1414
1415 nr_pages = (1 << cachep->gfporder);
1416 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1417 add_zone_page_state(page_zone(page),
1418 NR_SLAB_RECLAIMABLE, nr_pages);
1419 else
1420 add_zone_page_state(page_zone(page),
1421 NR_SLAB_UNRECLAIMABLE, nr_pages);
1422
1423 __SetPageSlab(page);
1424 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1425 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1426 SetPageSlabPfmemalloc(page);
1427
1428 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1429 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1430
1431 if (cachep->ctor)
1432 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1433 else
1434 kmemcheck_mark_unallocated_pages(page, nr_pages);
1435 }
1436
1437 return page;
1438}
1439
1440/*
1441 * Interface to system's page release.
1442 */
1443static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1444{
1445 int order = cachep->gfporder;
1446 unsigned long nr_freed = (1 << order);
1447
1448 kmemcheck_free_shadow(page, order);
1449
1450 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1451 sub_zone_page_state(page_zone(page),
1452 NR_SLAB_RECLAIMABLE, nr_freed);
1453 else
1454 sub_zone_page_state(page_zone(page),
1455 NR_SLAB_UNRECLAIMABLE, nr_freed);
1456
1457 BUG_ON(!PageSlab(page));
1458 __ClearPageSlabPfmemalloc(page);
1459 __ClearPageSlab(page);
1460 page_mapcount_reset(page);
1461 page->mapping = NULL;
1462
1463 if (current->reclaim_state)
1464 current->reclaim_state->reclaimed_slab += nr_freed;
1465 memcg_uncharge_slab(page, order, cachep);
1466 __free_pages(page, order);
1467}
1468
1469static void kmem_rcu_free(struct rcu_head *head)
1470{
1471 struct kmem_cache *cachep;
1472 struct page *page;
1473
1474 page = container_of(head, struct page, rcu_head);
1475 cachep = page->slab_cache;
1476
1477 kmem_freepages(cachep, page);
1478}
1479
1480#if DEBUG
1481static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1482{
1483 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1484 (cachep->size % PAGE_SIZE) == 0)
1485 return true;
1486
1487 return false;
1488}
1489
1490#ifdef CONFIG_DEBUG_PAGEALLOC
1491static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1492 unsigned long caller)
1493{
1494 int size = cachep->object_size;
1495
1496 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1497
1498 if (size < 5 * sizeof(unsigned long))
1499 return;
1500
1501 *addr++ = 0x12345678;
1502 *addr++ = caller;
1503 *addr++ = smp_processor_id();
1504 size -= 3 * sizeof(unsigned long);
1505 {
1506 unsigned long *sptr = &caller;
1507 unsigned long svalue;
1508
1509 while (!kstack_end(sptr)) {
1510 svalue = *sptr++;
1511 if (kernel_text_address(svalue)) {
1512 *addr++ = svalue;
1513 size -= sizeof(unsigned long);
1514 if (size <= sizeof(unsigned long))
1515 break;
1516 }
1517 }
1518
1519 }
1520 *addr++ = 0x87654321;
1521}
1522
1523static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1524 int map, unsigned long caller)
1525{
1526 if (!is_debug_pagealloc_cache(cachep))
1527 return;
1528
1529 if (caller)
1530 store_stackinfo(cachep, objp, caller);
1531
1532 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1533}
1534
1535#else
1536static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1537 int map, unsigned long caller) {}
1538
1539#endif
1540
1541static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1542{
1543 int size = cachep->object_size;
1544 addr = &((char *)addr)[obj_offset(cachep)];
1545
1546 memset(addr, val, size);
1547 *(unsigned char *)(addr + size - 1) = POISON_END;
1548}
1549
1550static void dump_line(char *data, int offset, int limit)
1551{
1552 int i;
1553 unsigned char error = 0;
1554 int bad_count = 0;
1555
1556 pr_err("%03x: ", offset);
1557 for (i = 0; i < limit; i++) {
1558 if (data[offset + i] != POISON_FREE) {
1559 error = data[offset + i];
1560 bad_count++;
1561 }
1562 }
1563 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1564 &data[offset], limit, 1);
1565
1566 if (bad_count == 1) {
1567 error ^= POISON_FREE;
1568 if (!(error & (error - 1))) {
1569 pr_err("Single bit error detected. Probably bad RAM.\n");
1570#ifdef CONFIG_X86
1571 pr_err("Run memtest86+ or a similar memory test tool.\n");
1572#else
1573 pr_err("Run a memory test tool.\n");
1574#endif
1575 }
1576 }
1577}
1578#endif
1579
1580#if DEBUG
1581
1582static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1583{
1584 int i, size;
1585 char *realobj;
1586
1587 if (cachep->flags & SLAB_RED_ZONE) {
1588 pr_err("Redzone: 0x%llx/0x%llx\n",
1589 *dbg_redzone1(cachep, objp),
1590 *dbg_redzone2(cachep, objp));
1591 }
1592
1593 if (cachep->flags & SLAB_STORE_USER) {
1594 pr_err("Last user: [<%p>](%pSR)\n",
1595 *dbg_userword(cachep, objp),
1596 *dbg_userword(cachep, objp));
1597 }
1598 realobj = (char *)objp + obj_offset(cachep);
1599 size = cachep->object_size;
1600 for (i = 0; i < size && lines; i += 16, lines--) {
1601 int limit;
1602 limit = 16;
1603 if (i + limit > size)
1604 limit = size - i;
1605 dump_line(realobj, i, limit);
1606 }
1607}
1608
1609static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1610{
1611 char *realobj;
1612 int size, i;
1613 int lines = 0;
1614
1615 if (is_debug_pagealloc_cache(cachep))
1616 return;
1617
1618 realobj = (char *)objp + obj_offset(cachep);
1619 size = cachep->object_size;
1620
1621 for (i = 0; i < size; i++) {
1622 char exp = POISON_FREE;
1623 if (i == size - 1)
1624 exp = POISON_END;
1625 if (realobj[i] != exp) {
1626 int limit;
1627 /* Mismatch ! */
1628 /* Print header */
1629 if (lines == 0) {
1630 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1631 print_tainted(), cachep->name,
1632 realobj, size);
1633 print_objinfo(cachep, objp, 0);
1634 }
1635 /* Hexdump the affected line */
1636 i = (i / 16) * 16;
1637 limit = 16;
1638 if (i + limit > size)
1639 limit = size - i;
1640 dump_line(realobj, i, limit);
1641 i += 16;
1642 lines++;
1643 /* Limit to 5 lines */
1644 if (lines > 5)
1645 break;
1646 }
1647 }
1648 if (lines != 0) {
1649 /* Print some data about the neighboring objects, if they
1650 * exist:
1651 */
1652 struct page *page = virt_to_head_page(objp);
1653 unsigned int objnr;
1654
1655 objnr = obj_to_index(cachep, page, objp);
1656 if (objnr) {
1657 objp = index_to_obj(cachep, page, objnr - 1);
1658 realobj = (char *)objp + obj_offset(cachep);
1659 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1660 print_objinfo(cachep, objp, 2);
1661 }
1662 if (objnr + 1 < cachep->num) {
1663 objp = index_to_obj(cachep, page, objnr + 1);
1664 realobj = (char *)objp + obj_offset(cachep);
1665 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1666 print_objinfo(cachep, objp, 2);
1667 }
1668 }
1669}
1670#endif
1671
1672#if DEBUG
1673static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1674 struct page *page)
1675{
1676 int i;
1677
1678 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1679 poison_obj(cachep, page->freelist - obj_offset(cachep),
1680 POISON_FREE);
1681 }
1682
1683 for (i = 0; i < cachep->num; i++) {
1684 void *objp = index_to_obj(cachep, page, i);
1685
1686 if (cachep->flags & SLAB_POISON) {
1687 check_poison_obj(cachep, objp);
1688 slab_kernel_map(cachep, objp, 1, 0);
1689 }
1690 if (cachep->flags & SLAB_RED_ZONE) {
1691 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1692 slab_error(cachep, "start of a freed object was overwritten");
1693 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1694 slab_error(cachep, "end of a freed object was overwritten");
1695 }
1696 }
1697}
1698#else
1699static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1700 struct page *page)
1701{
1702}
1703#endif
1704
1705/**
1706 * slab_destroy - destroy and release all objects in a slab
1707 * @cachep: cache pointer being destroyed
1708 * @page: page pointer being destroyed
1709 *
1710 * Destroy all the objs in a slab page, and release the mem back to the system.
1711 * Before calling the slab page must have been unlinked from the cache. The
1712 * kmem_cache_node ->list_lock is not held/needed.
1713 */
1714static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1715{
1716 void *freelist;
1717
1718 freelist = page->freelist;
1719 slab_destroy_debugcheck(cachep, page);
1720 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1721 call_rcu(&page->rcu_head, kmem_rcu_free);
1722 else
1723 kmem_freepages(cachep, page);
1724
1725 /*
1726 * From now on, we don't use freelist
1727 * although actual page can be freed in rcu context
1728 */
1729 if (OFF_SLAB(cachep))
1730 kmem_cache_free(cachep->freelist_cache, freelist);
1731}
1732
1733static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1734{
1735 struct page *page, *n;
1736
1737 list_for_each_entry_safe(page, n, list, lru) {
1738 list_del(&page->lru);
1739 slab_destroy(cachep, page);
1740 }
1741}
1742
1743/**
1744 * calculate_slab_order - calculate size (page order) of slabs
1745 * @cachep: pointer to the cache that is being created
1746 * @size: size of objects to be created in this cache.
1747 * @flags: slab allocation flags
1748 *
1749 * Also calculates the number of objects per slab.
1750 *
1751 * This could be made much more intelligent. For now, try to avoid using
1752 * high order pages for slabs. When the gfp() functions are more friendly
1753 * towards high-order requests, this should be changed.
1754 */
1755static size_t calculate_slab_order(struct kmem_cache *cachep,
1756 size_t size, unsigned long flags)
1757{
1758 size_t left_over = 0;
1759 int gfporder;
1760
1761 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1762 unsigned int num;
1763 size_t remainder;
1764
1765 num = cache_estimate(gfporder, size, flags, &remainder);
1766 if (!num)
1767 continue;
1768
1769 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1770 if (num > SLAB_OBJ_MAX_NUM)
1771 break;
1772
1773 if (flags & CFLGS_OFF_SLAB) {
1774 struct kmem_cache *freelist_cache;
1775 size_t freelist_size;
1776
1777 freelist_size = num * sizeof(freelist_idx_t);
1778 freelist_cache = kmalloc_slab(freelist_size, 0u);
1779 if (!freelist_cache)
1780 continue;
1781
1782 /*
1783 * Needed to avoid possible looping condition
1784 * in cache_grow()
1785 */
1786 if (OFF_SLAB(freelist_cache))
1787 continue;
1788
1789 /* check if off slab has enough benefit */
1790 if (freelist_cache->size > cachep->size / 2)
1791 continue;
1792 }
1793
1794 /* Found something acceptable - save it away */
1795 cachep->num = num;
1796 cachep->gfporder = gfporder;
1797 left_over = remainder;
1798
1799 /*
1800 * A VFS-reclaimable slab tends to have most allocations
1801 * as GFP_NOFS and we really don't want to have to be allocating
1802 * higher-order pages when we are unable to shrink dcache.
1803 */
1804 if (flags & SLAB_RECLAIM_ACCOUNT)
1805 break;
1806
1807 /*
1808 * Large number of objects is good, but very large slabs are
1809 * currently bad for the gfp()s.
1810 */
1811 if (gfporder >= slab_max_order)
1812 break;
1813
1814 /*
1815 * Acceptable internal fragmentation?
1816 */
1817 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1818 break;
1819 }
1820 return left_over;
1821}
1822
1823static struct array_cache __percpu *alloc_kmem_cache_cpus(
1824 struct kmem_cache *cachep, int entries, int batchcount)
1825{
1826 int cpu;
1827 size_t size;
1828 struct array_cache __percpu *cpu_cache;
1829
1830 size = sizeof(void *) * entries + sizeof(struct array_cache);
1831 cpu_cache = __alloc_percpu(size, sizeof(void *));
1832
1833 if (!cpu_cache)
1834 return NULL;
1835
1836 for_each_possible_cpu(cpu) {
1837 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1838 entries, batchcount);
1839 }
1840
1841 return cpu_cache;
1842}
1843
1844static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1845{
1846 if (slab_state >= FULL)
1847 return enable_cpucache(cachep, gfp);
1848
1849 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1850 if (!cachep->cpu_cache)
1851 return 1;
1852
1853 if (slab_state == DOWN) {
1854 /* Creation of first cache (kmem_cache). */
1855 set_up_node(kmem_cache, CACHE_CACHE);
1856 } else if (slab_state == PARTIAL) {
1857 /* For kmem_cache_node */
1858 set_up_node(cachep, SIZE_NODE);
1859 } else {
1860 int node;
1861
1862 for_each_online_node(node) {
1863 cachep->node[node] = kmalloc_node(
1864 sizeof(struct kmem_cache_node), gfp, node);
1865 BUG_ON(!cachep->node[node]);
1866 kmem_cache_node_init(cachep->node[node]);
1867 }
1868 }
1869
1870 cachep->node[numa_mem_id()]->next_reap =
1871 jiffies + REAPTIMEOUT_NODE +
1872 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1873
1874 cpu_cache_get(cachep)->avail = 0;
1875 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1876 cpu_cache_get(cachep)->batchcount = 1;
1877 cpu_cache_get(cachep)->touched = 0;
1878 cachep->batchcount = 1;
1879 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1880 return 0;
1881}
1882
1883unsigned long kmem_cache_flags(unsigned long object_size,
1884 unsigned long flags, const char *name,
1885 void (*ctor)(void *))
1886{
1887 return flags;
1888}
1889
1890struct kmem_cache *
1891__kmem_cache_alias(const char *name, size_t size, size_t align,
1892 unsigned long flags, void (*ctor)(void *))
1893{
1894 struct kmem_cache *cachep;
1895
1896 cachep = find_mergeable(size, align, flags, name, ctor);
1897 if (cachep) {
1898 cachep->refcount++;
1899
1900 /*
1901 * Adjust the object sizes so that we clear
1902 * the complete object on kzalloc.
1903 */
1904 cachep->object_size = max_t(int, cachep->object_size, size);
1905 }
1906 return cachep;
1907}
1908
1909static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
1910 size_t size, unsigned long flags)
1911{
1912 size_t left;
1913
1914 cachep->num = 0;
1915
1916 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
1917 return false;
1918
1919 left = calculate_slab_order(cachep, size,
1920 flags | CFLGS_OBJFREELIST_SLAB);
1921 if (!cachep->num)
1922 return false;
1923
1924 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
1925 return false;
1926
1927 cachep->colour = left / cachep->colour_off;
1928
1929 return true;
1930}
1931
1932static bool set_off_slab_cache(struct kmem_cache *cachep,
1933 size_t size, unsigned long flags)
1934{
1935 size_t left;
1936
1937 cachep->num = 0;
1938
1939 /*
1940 * Always use on-slab management when SLAB_NOLEAKTRACE
1941 * to avoid recursive calls into kmemleak.
1942 */
1943 if (flags & SLAB_NOLEAKTRACE)
1944 return false;
1945
1946 /*
1947 * Size is large, assume best to place the slab management obj
1948 * off-slab (should allow better packing of objs).
1949 */
1950 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
1951 if (!cachep->num)
1952 return false;
1953
1954 /*
1955 * If the slab has been placed off-slab, and we have enough space then
1956 * move it on-slab. This is at the expense of any extra colouring.
1957 */
1958 if (left >= cachep->num * sizeof(freelist_idx_t))
1959 return false;
1960
1961 cachep->colour = left / cachep->colour_off;
1962
1963 return true;
1964}
1965
1966static bool set_on_slab_cache(struct kmem_cache *cachep,
1967 size_t size, unsigned long flags)
1968{
1969 size_t left;
1970
1971 cachep->num = 0;
1972
1973 left = calculate_slab_order(cachep, size, flags);
1974 if (!cachep->num)
1975 return false;
1976
1977 cachep->colour = left / cachep->colour_off;
1978
1979 return true;
1980}
1981
1982/**
1983 * __kmem_cache_create - Create a cache.
1984 * @cachep: cache management descriptor
1985 * @flags: SLAB flags
1986 *
1987 * Returns a ptr to the cache on success, NULL on failure.
1988 * Cannot be called within a int, but can be interrupted.
1989 * The @ctor is run when new pages are allocated by the cache.
1990 *
1991 * The flags are
1992 *
1993 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
1994 * to catch references to uninitialised memory.
1995 *
1996 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
1997 * for buffer overruns.
1998 *
1999 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2000 * cacheline. This can be beneficial if you're counting cycles as closely
2001 * as davem.
2002 */
2003int
2004__kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2005{
2006 size_t ralign = BYTES_PER_WORD;
2007 gfp_t gfp;
2008 int err;
2009 size_t size = cachep->size;
2010
2011#if DEBUG
2012#if FORCED_DEBUG
2013 /*
2014 * Enable redzoning and last user accounting, except for caches with
2015 * large objects, if the increased size would increase the object size
2016 * above the next power of two: caches with object sizes just above a
2017 * power of two have a significant amount of internal fragmentation.
2018 */
2019 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2020 2 * sizeof(unsigned long long)))
2021 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2022 if (!(flags & SLAB_DESTROY_BY_RCU))
2023 flags |= SLAB_POISON;
2024#endif
2025#endif
2026
2027 /*
2028 * Check that size is in terms of words. This is needed to avoid
2029 * unaligned accesses for some archs when redzoning is used, and makes
2030 * sure any on-slab bufctl's are also correctly aligned.
2031 */
2032 if (size & (BYTES_PER_WORD - 1)) {
2033 size += (BYTES_PER_WORD - 1);
2034 size &= ~(BYTES_PER_WORD - 1);
2035 }
2036
2037 if (flags & SLAB_RED_ZONE) {
2038 ralign = REDZONE_ALIGN;
2039 /* If redzoning, ensure that the second redzone is suitably
2040 * aligned, by adjusting the object size accordingly. */
2041 size += REDZONE_ALIGN - 1;
2042 size &= ~(REDZONE_ALIGN - 1);
2043 }
2044
2045 /* 3) caller mandated alignment */
2046 if (ralign < cachep->align) {
2047 ralign = cachep->align;
2048 }
2049 /* disable debug if necessary */
2050 if (ralign > __alignof__(unsigned long long))
2051 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2052 /*
2053 * 4) Store it.
2054 */
2055 cachep->align = ralign;
2056 cachep->colour_off = cache_line_size();
2057 /* Offset must be a multiple of the alignment. */
2058 if (cachep->colour_off < cachep->align)
2059 cachep->colour_off = cachep->align;
2060
2061 if (slab_is_available())
2062 gfp = GFP_KERNEL;
2063 else
2064 gfp = GFP_NOWAIT;
2065
2066#if DEBUG
2067
2068 /*
2069 * Both debugging options require word-alignment which is calculated
2070 * into align above.
2071 */
2072 if (flags & SLAB_RED_ZONE) {
2073 /* add space for red zone words */
2074 cachep->obj_offset += sizeof(unsigned long long);
2075 size += 2 * sizeof(unsigned long long);
2076 }
2077 if (flags & SLAB_STORE_USER) {
2078 /* user store requires one word storage behind the end of
2079 * the real object. But if the second red zone needs to be
2080 * aligned to 64 bits, we must allow that much space.
2081 */
2082 if (flags & SLAB_RED_ZONE)
2083 size += REDZONE_ALIGN;
2084 else
2085 size += BYTES_PER_WORD;
2086 }
2087#endif
2088
2089 kasan_cache_create(cachep, &size, &flags);
2090
2091 size = ALIGN(size, cachep->align);
2092 /*
2093 * We should restrict the number of objects in a slab to implement
2094 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2095 */
2096 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2097 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2098
2099#if DEBUG
2100 /*
2101 * To activate debug pagealloc, off-slab management is necessary
2102 * requirement. In early phase of initialization, small sized slab
2103 * doesn't get initialized so it would not be possible. So, we need
2104 * to check size >= 256. It guarantees that all necessary small
2105 * sized slab is initialized in current slab initialization sequence.
2106 */
2107 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2108 size >= 256 && cachep->object_size > cache_line_size()) {
2109 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2110 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2111
2112 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2113 flags |= CFLGS_OFF_SLAB;
2114 cachep->obj_offset += tmp_size - size;
2115 size = tmp_size;
2116 goto done;
2117 }
2118 }
2119 }
2120#endif
2121
2122 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2123 flags |= CFLGS_OBJFREELIST_SLAB;
2124 goto done;
2125 }
2126
2127 if (set_off_slab_cache(cachep, size, flags)) {
2128 flags |= CFLGS_OFF_SLAB;
2129 goto done;
2130 }
2131
2132 if (set_on_slab_cache(cachep, size, flags))
2133 goto done;
2134
2135 return -E2BIG;
2136
2137done:
2138 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2139 cachep->flags = flags;
2140 cachep->allocflags = __GFP_COMP;
2141 if (CONFIG_ZONE_DMA_FLAG && (flags & SLAB_CACHE_DMA))
2142 cachep->allocflags |= GFP_DMA;
2143 cachep->size = size;
2144 cachep->reciprocal_buffer_size = reciprocal_value(size);
2145
2146#if DEBUG
2147 /*
2148 * If we're going to use the generic kernel_map_pages()
2149 * poisoning, then it's going to smash the contents of
2150 * the redzone and userword anyhow, so switch them off.
2151 */
2152 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2153 (cachep->flags & SLAB_POISON) &&
2154 is_debug_pagealloc_cache(cachep))
2155 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2156#endif
2157
2158 if (OFF_SLAB(cachep)) {
2159 cachep->freelist_cache =
2160 kmalloc_slab(cachep->freelist_size, 0u);
2161 }
2162
2163 err = setup_cpu_cache(cachep, gfp);
2164 if (err) {
2165 __kmem_cache_release(cachep);
2166 return err;
2167 }
2168
2169 return 0;
2170}
2171
2172#if DEBUG
2173static void check_irq_off(void)
2174{
2175 BUG_ON(!irqs_disabled());
2176}
2177
2178static void check_irq_on(void)
2179{
2180 BUG_ON(irqs_disabled());
2181}
2182
2183static void check_spinlock_acquired(struct kmem_cache *cachep)
2184{
2185#ifdef CONFIG_SMP
2186 check_irq_off();
2187 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2188#endif
2189}
2190
2191static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2192{
2193#ifdef CONFIG_SMP
2194 check_irq_off();
2195 assert_spin_locked(&get_node(cachep, node)->list_lock);
2196#endif
2197}
2198
2199#else
2200#define check_irq_off() do { } while(0)
2201#define check_irq_on() do { } while(0)
2202#define check_spinlock_acquired(x) do { } while(0)
2203#define check_spinlock_acquired_node(x, y) do { } while(0)
2204#endif
2205
2206static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
2207 struct array_cache *ac,
2208 int force, int node);
2209
2210static void do_drain(void *arg)
2211{
2212 struct kmem_cache *cachep = arg;
2213 struct array_cache *ac;
2214 int node = numa_mem_id();
2215 struct kmem_cache_node *n;
2216 LIST_HEAD(list);
2217
2218 check_irq_off();
2219 ac = cpu_cache_get(cachep);
2220 n = get_node(cachep, node);
2221 spin_lock(&n->list_lock);
2222 free_block(cachep, ac->entry, ac->avail, node, &list);
2223 spin_unlock(&n->list_lock);
2224 slabs_destroy(cachep, &list);
2225 ac->avail = 0;
2226}
2227
2228static void drain_cpu_caches(struct kmem_cache *cachep)
2229{
2230 struct kmem_cache_node *n;
2231 int node;
2232
2233 on_each_cpu(do_drain, cachep, 1);
2234 check_irq_on();
2235 for_each_kmem_cache_node(cachep, node, n)
2236 if (n->alien)
2237 drain_alien_cache(cachep, n->alien);
2238
2239 for_each_kmem_cache_node(cachep, node, n)
2240 drain_array(cachep, n, n->shared, 1, node);
2241}
2242
2243/*
2244 * Remove slabs from the list of free slabs.
2245 * Specify the number of slabs to drain in tofree.
2246 *
2247 * Returns the actual number of slabs released.
2248 */
2249static int drain_freelist(struct kmem_cache *cache,
2250 struct kmem_cache_node *n, int tofree)
2251{
2252 struct list_head *p;
2253 int nr_freed;
2254 struct page *page;
2255
2256 nr_freed = 0;
2257 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2258
2259 spin_lock_irq(&n->list_lock);
2260 p = n->slabs_free.prev;
2261 if (p == &n->slabs_free) {
2262 spin_unlock_irq(&n->list_lock);
2263 goto out;
2264 }
2265
2266 page = list_entry(p, struct page, lru);
2267 list_del(&page->lru);
2268 /*
2269 * Safe to drop the lock. The slab is no longer linked
2270 * to the cache.
2271 */
2272 n->free_objects -= cache->num;
2273 spin_unlock_irq(&n->list_lock);
2274 slab_destroy(cache, page);
2275 nr_freed++;
2276 }
2277out:
2278 return nr_freed;
2279}
2280
2281int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2282{
2283 int ret = 0;
2284 int node;
2285 struct kmem_cache_node *n;
2286
2287 drain_cpu_caches(cachep);
2288
2289 check_irq_on();
2290 for_each_kmem_cache_node(cachep, node, n) {
2291 drain_freelist(cachep, n, slabs_tofree(cachep, n));
2292
2293 ret += !list_empty(&n->slabs_full) ||
2294 !list_empty(&n->slabs_partial);
2295 }
2296 return (ret ? 1 : 0);
2297}
2298
2299int __kmem_cache_shutdown(struct kmem_cache *cachep)
2300{
2301 return __kmem_cache_shrink(cachep, false);
2302}
2303
2304void __kmem_cache_release(struct kmem_cache *cachep)
2305{
2306 int i;
2307 struct kmem_cache_node *n;
2308
2309 free_percpu(cachep->cpu_cache);
2310
2311 /* NUMA: free the node structures */
2312 for_each_kmem_cache_node(cachep, i, n) {
2313 kfree(n->shared);
2314 free_alien_cache(n->alien);
2315 kfree(n);
2316 cachep->node[i] = NULL;
2317 }
2318}
2319
2320/*
2321 * Get the memory for a slab management obj.
2322 *
2323 * For a slab cache when the slab descriptor is off-slab, the
2324 * slab descriptor can't come from the same cache which is being created,
2325 * Because if it is the case, that means we defer the creation of
2326 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2327 * And we eventually call down to __kmem_cache_create(), which
2328 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2329 * This is a "chicken-and-egg" problem.
2330 *
2331 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2332 * which are all initialized during kmem_cache_init().
2333 */
2334static void *alloc_slabmgmt(struct kmem_cache *cachep,
2335 struct page *page, int colour_off,
2336 gfp_t local_flags, int nodeid)
2337{
2338 void *freelist;
2339 void *addr = page_address(page);
2340
2341 page->s_mem = addr + colour_off;
2342 page->active = 0;
2343
2344 if (OBJFREELIST_SLAB(cachep))
2345 freelist = NULL;
2346 else if (OFF_SLAB(cachep)) {
2347 /* Slab management obj is off-slab. */
2348 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2349 local_flags, nodeid);
2350 if (!freelist)
2351 return NULL;
2352 } else {
2353 /* We will use last bytes at the slab for freelist */
2354 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2355 cachep->freelist_size;
2356 }
2357
2358 return freelist;
2359}
2360
2361static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2362{
2363 return ((freelist_idx_t *)page->freelist)[idx];
2364}
2365
2366static inline void set_free_obj(struct page *page,
2367 unsigned int idx, freelist_idx_t val)
2368{
2369 ((freelist_idx_t *)(page->freelist))[idx] = val;
2370}
2371
2372static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2373{
2374#if DEBUG
2375 int i;
2376
2377 for (i = 0; i < cachep->num; i++) {
2378 void *objp = index_to_obj(cachep, page, i);
2379
2380 if (cachep->flags & SLAB_STORE_USER)
2381 *dbg_userword(cachep, objp) = NULL;
2382
2383 if (cachep->flags & SLAB_RED_ZONE) {
2384 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2385 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2386 }
2387 /*
2388 * Constructors are not allowed to allocate memory from the same
2389 * cache which they are a constructor for. Otherwise, deadlock.
2390 * They must also be threaded.
2391 */
2392 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2393 kasan_unpoison_object_data(cachep,
2394 objp + obj_offset(cachep));
2395 cachep->ctor(objp + obj_offset(cachep));
2396 kasan_poison_object_data(
2397 cachep, objp + obj_offset(cachep));
2398 }
2399
2400 if (cachep->flags & SLAB_RED_ZONE) {
2401 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2402 slab_error(cachep, "constructor overwrote the end of an object");
2403 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2404 slab_error(cachep, "constructor overwrote the start of an object");
2405 }
2406 /* need to poison the objs? */
2407 if (cachep->flags & SLAB_POISON) {
2408 poison_obj(cachep, objp, POISON_FREE);
2409 slab_kernel_map(cachep, objp, 0, 0);
2410 }
2411 }
2412#endif
2413}
2414
2415static void cache_init_objs(struct kmem_cache *cachep,
2416 struct page *page)
2417{
2418 int i;
2419 void *objp;
2420
2421 cache_init_objs_debug(cachep, page);
2422
2423 if (OBJFREELIST_SLAB(cachep)) {
2424 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2425 obj_offset(cachep);
2426 }
2427
2428 for (i = 0; i < cachep->num; i++) {
2429 /* constructor could break poison info */
2430 if (DEBUG == 0 && cachep->ctor) {
2431 objp = index_to_obj(cachep, page, i);
2432 kasan_unpoison_object_data(cachep, objp);
2433 cachep->ctor(objp);
2434 kasan_poison_object_data(cachep, objp);
2435 }
2436
2437 set_free_obj(page, i, i);
2438 }
2439}
2440
2441static void kmem_flagcheck(struct kmem_cache *cachep, gfp_t flags)
2442{
2443 if (CONFIG_ZONE_DMA_FLAG) {
2444 if (flags & GFP_DMA)
2445 BUG_ON(!(cachep->allocflags & GFP_DMA));
2446 else
2447 BUG_ON(cachep->allocflags & GFP_DMA);
2448 }
2449}
2450
2451static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2452{
2453 void *objp;
2454
2455 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2456 page->active++;
2457
2458#if DEBUG
2459 if (cachep->flags & SLAB_STORE_USER)
2460 set_store_user_dirty(cachep);
2461#endif
2462
2463 return objp;
2464}
2465
2466static void slab_put_obj(struct kmem_cache *cachep,
2467 struct page *page, void *objp)
2468{
2469 unsigned int objnr = obj_to_index(cachep, page, objp);
2470#if DEBUG
2471 unsigned int i;
2472
2473 /* Verify double free bug */
2474 for (i = page->active; i < cachep->num; i++) {
2475 if (get_free_obj(page, i) == objnr) {
2476 pr_err("slab: double free detected in cache '%s', objp %p\n",
2477 cachep->name, objp);
2478 BUG();
2479 }
2480 }
2481#endif
2482 page->active--;
2483 if (!page->freelist)
2484 page->freelist = objp + obj_offset(cachep);
2485
2486 set_free_obj(page, page->active, objnr);
2487}
2488
2489/*
2490 * Map pages beginning at addr to the given cache and slab. This is required
2491 * for the slab allocator to be able to lookup the cache and slab of a
2492 * virtual address for kfree, ksize, and slab debugging.
2493 */
2494static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2495 void *freelist)
2496{
2497 page->slab_cache = cache;
2498 page->freelist = freelist;
2499}
2500
2501/*
2502 * Grow (by 1) the number of slabs within a cache. This is called by
2503 * kmem_cache_alloc() when there are no active objs left in a cache.
2504 */
2505static int cache_grow(struct kmem_cache *cachep,
2506 gfp_t flags, int nodeid, struct page *page)
2507{
2508 void *freelist;
2509 size_t offset;
2510 gfp_t local_flags;
2511 struct kmem_cache_node *n;
2512
2513 /*
2514 * Be lazy and only check for valid flags here, keeping it out of the
2515 * critical path in kmem_cache_alloc().
2516 */
2517 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2518 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2519 BUG();
2520 }
2521 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2522
2523 /* Take the node list lock to change the colour_next on this node */
2524 check_irq_off();
2525 n = get_node(cachep, nodeid);
2526 spin_lock(&n->list_lock);
2527
2528 /* Get colour for the slab, and cal the next value. */
2529 offset = n->colour_next;
2530 n->colour_next++;
2531 if (n->colour_next >= cachep->colour)
2532 n->colour_next = 0;
2533 spin_unlock(&n->list_lock);
2534
2535 offset *= cachep->colour_off;
2536
2537 if (gfpflags_allow_blocking(local_flags))
2538 local_irq_enable();
2539
2540 /*
2541 * The test for missing atomic flag is performed here, rather than
2542 * the more obvious place, simply to reduce the critical path length
2543 * in kmem_cache_alloc(). If a caller is seriously mis-behaving they
2544 * will eventually be caught here (where it matters).
2545 */
2546 kmem_flagcheck(cachep, flags);
2547
2548 /*
2549 * Get mem for the objs. Attempt to allocate a physical page from
2550 * 'nodeid'.
2551 */
2552 if (!page)
2553 page = kmem_getpages(cachep, local_flags, nodeid);
2554 if (!page)
2555 goto failed;
2556
2557 /* Get slab management. */
2558 freelist = alloc_slabmgmt(cachep, page, offset,
2559 local_flags & ~GFP_CONSTRAINT_MASK, nodeid);
2560 if (OFF_SLAB(cachep) && !freelist)
2561 goto opps1;
2562
2563 slab_map_pages(cachep, page, freelist);
2564
2565 kasan_poison_slab(page);
2566 cache_init_objs(cachep, page);
2567
2568 if (gfpflags_allow_blocking(local_flags))
2569 local_irq_disable();
2570 check_irq_off();
2571 spin_lock(&n->list_lock);
2572
2573 /* Make slab active. */
2574 list_add_tail(&page->lru, &(n->slabs_free));
2575 STATS_INC_GROWN(cachep);
2576 n->free_objects += cachep->num;
2577 spin_unlock(&n->list_lock);
2578 return 1;
2579opps1:
2580 kmem_freepages(cachep, page);
2581failed:
2582 if (gfpflags_allow_blocking(local_flags))
2583 local_irq_disable();
2584 return 0;
2585}
2586
2587#if DEBUG
2588
2589/*
2590 * Perform extra freeing checks:
2591 * - detect bad pointers.
2592 * - POISON/RED_ZONE checking
2593 */
2594static void kfree_debugcheck(const void *objp)
2595{
2596 if (!virt_addr_valid(objp)) {
2597 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2598 (unsigned long)objp);
2599 BUG();
2600 }
2601}
2602
2603static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2604{
2605 unsigned long long redzone1, redzone2;
2606
2607 redzone1 = *dbg_redzone1(cache, obj);
2608 redzone2 = *dbg_redzone2(cache, obj);
2609
2610 /*
2611 * Redzone is ok.
2612 */
2613 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2614 return;
2615
2616 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2617 slab_error(cache, "double free detected");
2618 else
2619 slab_error(cache, "memory outside object was overwritten");
2620
2621 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2622 obj, redzone1, redzone2);
2623}
2624
2625static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2626 unsigned long caller)
2627{
2628 unsigned int objnr;
2629 struct page *page;
2630
2631 BUG_ON(virt_to_cache(objp) != cachep);
2632
2633 objp -= obj_offset(cachep);
2634 kfree_debugcheck(objp);
2635 page = virt_to_head_page(objp);
2636
2637 if (cachep->flags & SLAB_RED_ZONE) {
2638 verify_redzone_free(cachep, objp);
2639 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2640 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2641 }
2642 if (cachep->flags & SLAB_STORE_USER) {
2643 set_store_user_dirty(cachep);
2644 *dbg_userword(cachep, objp) = (void *)caller;
2645 }
2646
2647 objnr = obj_to_index(cachep, page, objp);
2648
2649 BUG_ON(objnr >= cachep->num);
2650 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2651
2652 if (cachep->flags & SLAB_POISON) {
2653 poison_obj(cachep, objp, POISON_FREE);
2654 slab_kernel_map(cachep, objp, 0, caller);
2655 }
2656 return objp;
2657}
2658
2659#else
2660#define kfree_debugcheck(x) do { } while(0)
2661#define cache_free_debugcheck(x,objp,z) (objp)
2662#endif
2663
2664static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2665 void **list)
2666{
2667#if DEBUG
2668 void *next = *list;
2669 void *objp;
2670
2671 while (next) {
2672 objp = next - obj_offset(cachep);
2673 next = *(void **)next;
2674 poison_obj(cachep, objp, POISON_FREE);
2675 }
2676#endif
2677}
2678
2679static inline void fixup_slab_list(struct kmem_cache *cachep,
2680 struct kmem_cache_node *n, struct page *page,
2681 void **list)
2682{
2683 /* move slabp to correct slabp list: */
2684 list_del(&page->lru);
2685 if (page->active == cachep->num) {
2686 list_add(&page->lru, &n->slabs_full);
2687 if (OBJFREELIST_SLAB(cachep)) {
2688#if DEBUG
2689 /* Poisoning will be done without holding the lock */
2690 if (cachep->flags & SLAB_POISON) {
2691 void **objp = page->freelist;
2692
2693 *objp = *list;
2694 *list = objp;
2695 }
2696#endif
2697 page->freelist = NULL;
2698 }
2699 } else
2700 list_add(&page->lru, &n->slabs_partial);
2701}
2702
2703/* Try to find non-pfmemalloc slab if needed */
2704static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2705 struct page *page, bool pfmemalloc)
2706{
2707 if (!page)
2708 return NULL;
2709
2710 if (pfmemalloc)
2711 return page;
2712
2713 if (!PageSlabPfmemalloc(page))
2714 return page;
2715
2716 /* No need to keep pfmemalloc slab if we have enough free objects */
2717 if (n->free_objects > n->free_limit) {
2718 ClearPageSlabPfmemalloc(page);
2719 return page;
2720 }
2721
2722 /* Move pfmemalloc slab to the end of list to speed up next search */
2723 list_del(&page->lru);
2724 if (!page->active)
2725 list_add_tail(&page->lru, &n->slabs_free);
2726 else
2727 list_add_tail(&page->lru, &n->slabs_partial);
2728
2729 list_for_each_entry(page, &n->slabs_partial, lru) {
2730 if (!PageSlabPfmemalloc(page))
2731 return page;
2732 }
2733
2734 list_for_each_entry(page, &n->slabs_free, lru) {
2735 if (!PageSlabPfmemalloc(page))
2736 return page;
2737 }
2738
2739 return NULL;
2740}
2741
2742static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2743{
2744 struct page *page;
2745
2746 page = list_first_entry_or_null(&n->slabs_partial,
2747 struct page, lru);
2748 if (!page) {
2749 n->free_touched = 1;
2750 page = list_first_entry_or_null(&n->slabs_free,
2751 struct page, lru);
2752 }
2753
2754 if (sk_memalloc_socks())
2755 return get_valid_first_slab(n, page, pfmemalloc);
2756
2757 return page;
2758}
2759
2760static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2761 struct kmem_cache_node *n, gfp_t flags)
2762{
2763 struct page *page;
2764 void *obj;
2765 void *list = NULL;
2766
2767 if (!gfp_pfmemalloc_allowed(flags))
2768 return NULL;
2769
2770 spin_lock(&n->list_lock);
2771 page = get_first_slab(n, true);
2772 if (!page) {
2773 spin_unlock(&n->list_lock);
2774 return NULL;
2775 }
2776
2777 obj = slab_get_obj(cachep, page);
2778 n->free_objects--;
2779
2780 fixup_slab_list(cachep, n, page, &list);
2781
2782 spin_unlock(&n->list_lock);
2783 fixup_objfreelist_debug(cachep, &list);
2784
2785 return obj;
2786}
2787
2788static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
2789{
2790 int batchcount;
2791 struct kmem_cache_node *n;
2792 struct array_cache *ac;
2793 int node;
2794 void *list = NULL;
2795
2796 check_irq_off();
2797 node = numa_mem_id();
2798
2799retry:
2800 ac = cpu_cache_get(cachep);
2801 batchcount = ac->batchcount;
2802 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
2803 /*
2804 * If there was little recent activity on this cache, then
2805 * perform only a partial refill. Otherwise we could generate
2806 * refill bouncing.
2807 */
2808 batchcount = BATCHREFILL_LIMIT;
2809 }
2810 n = get_node(cachep, node);
2811
2812 BUG_ON(ac->avail > 0 || !n);
2813 spin_lock(&n->list_lock);
2814
2815 /* See if we can refill from the shared array */
2816 if (n->shared && transfer_objects(ac, n->shared, batchcount)) {
2817 n->shared->touched = 1;
2818 goto alloc_done;
2819 }
2820
2821 while (batchcount > 0) {
2822 struct page *page;
2823 /* Get slab alloc is to come from. */
2824 page = get_first_slab(n, false);
2825 if (!page)
2826 goto must_grow;
2827
2828 check_spinlock_acquired(cachep);
2829
2830 /*
2831 * The slab was either on partial or free list so
2832 * there must be at least one object available for
2833 * allocation.
2834 */
2835 BUG_ON(page->active >= cachep->num);
2836
2837 while (page->active < cachep->num && batchcount--) {
2838 STATS_INC_ALLOCED(cachep);
2839 STATS_INC_ACTIVE(cachep);
2840 STATS_SET_HIGH(cachep);
2841
2842 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
2843 }
2844
2845 fixup_slab_list(cachep, n, page, &list);
2846 }
2847
2848must_grow:
2849 n->free_objects -= ac->avail;
2850alloc_done:
2851 spin_unlock(&n->list_lock);
2852 fixup_objfreelist_debug(cachep, &list);
2853
2854 if (unlikely(!ac->avail)) {
2855 int x;
2856
2857 /* Check if we can use obj in pfmemalloc slab */
2858 if (sk_memalloc_socks()) {
2859 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
2860
2861 if (obj)
2862 return obj;
2863 }
2864
2865 x = cache_grow(cachep, gfp_exact_node(flags), node, NULL);
2866
2867 /* cache_grow can reenable interrupts, then ac could change. */
2868 ac = cpu_cache_get(cachep);
2869 node = numa_mem_id();
2870
2871 /* no objects in sight? abort */
2872 if (!x && ac->avail == 0)
2873 return NULL;
2874
2875 if (!ac->avail) /* objects refilled by interrupt? */
2876 goto retry;
2877 }
2878 ac->touched = 1;
2879
2880 return ac->entry[--ac->avail];
2881}
2882
2883static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
2884 gfp_t flags)
2885{
2886 might_sleep_if(gfpflags_allow_blocking(flags));
2887#if DEBUG
2888 kmem_flagcheck(cachep, flags);
2889#endif
2890}
2891
2892#if DEBUG
2893static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
2894 gfp_t flags, void *objp, unsigned long caller)
2895{
2896 if (!objp)
2897 return objp;
2898 if (cachep->flags & SLAB_POISON) {
2899 check_poison_obj(cachep, objp);
2900 slab_kernel_map(cachep, objp, 1, 0);
2901 poison_obj(cachep, objp, POISON_INUSE);
2902 }
2903 if (cachep->flags & SLAB_STORE_USER)
2904 *dbg_userword(cachep, objp) = (void *)caller;
2905
2906 if (cachep->flags & SLAB_RED_ZONE) {
2907 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
2908 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
2909 slab_error(cachep, "double free, or memory outside object was overwritten");
2910 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2911 objp, *dbg_redzone1(cachep, objp),
2912 *dbg_redzone2(cachep, objp));
2913 }
2914 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
2915 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
2916 }
2917
2918 objp += obj_offset(cachep);
2919 if (cachep->ctor && cachep->flags & SLAB_POISON)
2920 cachep->ctor(objp);
2921 if (ARCH_SLAB_MINALIGN &&
2922 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
2923 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
2924 objp, (int)ARCH_SLAB_MINALIGN);
2925 }
2926 return objp;
2927}
2928#else
2929#define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
2930#endif
2931
2932static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
2933{
2934 void *objp;
2935 struct array_cache *ac;
2936
2937 check_irq_off();
2938
2939 ac = cpu_cache_get(cachep);
2940 if (likely(ac->avail)) {
2941 ac->touched = 1;
2942 objp = ac->entry[--ac->avail];
2943
2944 STATS_INC_ALLOCHIT(cachep);
2945 goto out;
2946 }
2947
2948 STATS_INC_ALLOCMISS(cachep);
2949 objp = cache_alloc_refill(cachep, flags);
2950 /*
2951 * the 'ac' may be updated by cache_alloc_refill(),
2952 * and kmemleak_erase() requires its correct value.
2953 */
2954 ac = cpu_cache_get(cachep);
2955
2956out:
2957 /*
2958 * To avoid a false negative, if an object that is in one of the
2959 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
2960 * treat the array pointers as a reference to the object.
2961 */
2962 if (objp)
2963 kmemleak_erase(&ac->entry[ac->avail]);
2964 return objp;
2965}
2966
2967#ifdef CONFIG_NUMA
2968/*
2969 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
2970 *
2971 * If we are in_interrupt, then process context, including cpusets and
2972 * mempolicy, may not apply and should not be used for allocation policy.
2973 */
2974static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
2975{
2976 int nid_alloc, nid_here;
2977
2978 if (in_interrupt() || (flags & __GFP_THISNODE))
2979 return NULL;
2980 nid_alloc = nid_here = numa_mem_id();
2981 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
2982 nid_alloc = cpuset_slab_spread_node();
2983 else if (current->mempolicy)
2984 nid_alloc = mempolicy_slab_node();
2985 if (nid_alloc != nid_here)
2986 return ____cache_alloc_node(cachep, flags, nid_alloc);
2987 return NULL;
2988}
2989
2990/*
2991 * Fallback function if there was no memory available and no objects on a
2992 * certain node and fall back is permitted. First we scan all the
2993 * available node for available objects. If that fails then we
2994 * perform an allocation without specifying a node. This allows the page
2995 * allocator to do its reclaim / fallback magic. We then insert the
2996 * slab into the proper nodelist and then allocate from it.
2997 */
2998static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
2999{
3000 struct zonelist *zonelist;
3001 gfp_t local_flags;
3002 struct zoneref *z;
3003 struct zone *zone;
3004 enum zone_type high_zoneidx = gfp_zone(flags);
3005 void *obj = NULL;
3006 int nid;
3007 unsigned int cpuset_mems_cookie;
3008
3009 if (flags & __GFP_THISNODE)
3010 return NULL;
3011
3012 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
3013
3014retry_cpuset:
3015 cpuset_mems_cookie = read_mems_allowed_begin();
3016 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3017
3018retry:
3019 /*
3020 * Look through allowed nodes for objects available
3021 * from existing per node queues.
3022 */
3023 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3024 nid = zone_to_nid(zone);
3025
3026 if (cpuset_zone_allowed(zone, flags) &&
3027 get_node(cache, nid) &&
3028 get_node(cache, nid)->free_objects) {
3029 obj = ____cache_alloc_node(cache,
3030 gfp_exact_node(flags), nid);
3031 if (obj)
3032 break;
3033 }
3034 }
3035
3036 if (!obj) {
3037 /*
3038 * This allocation will be performed within the constraints
3039 * of the current cpuset / memory policy requirements.
3040 * We may trigger various forms of reclaim on the allowed
3041 * set and go into memory reserves if necessary.
3042 */
3043 struct page *page;
3044
3045 if (gfpflags_allow_blocking(local_flags))
3046 local_irq_enable();
3047 kmem_flagcheck(cache, flags);
3048 page = kmem_getpages(cache, local_flags, numa_mem_id());
3049 if (gfpflags_allow_blocking(local_flags))
3050 local_irq_disable();
3051 if (page) {
3052 /*
3053 * Insert into the appropriate per node queues
3054 */
3055 nid = page_to_nid(page);
3056 if (cache_grow(cache, flags, nid, page)) {
3057 obj = ____cache_alloc_node(cache,
3058 gfp_exact_node(flags), nid);
3059 if (!obj)
3060 /*
3061 * Another processor may allocate the
3062 * objects in the slab since we are
3063 * not holding any locks.
3064 */
3065 goto retry;
3066 } else {
3067 /* cache_grow already freed obj */
3068 obj = NULL;
3069 }
3070 }
3071 }
3072
3073 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3074 goto retry_cpuset;
3075 return obj;
3076}
3077
3078/*
3079 * A interface to enable slab creation on nodeid
3080 */
3081static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3082 int nodeid)
3083{
3084 struct page *page;
3085 struct kmem_cache_node *n;
3086 void *obj;
3087 void *list = NULL;
3088 int x;
3089
3090 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3091 n = get_node(cachep, nodeid);
3092 BUG_ON(!n);
3093
3094retry:
3095 check_irq_off();
3096 spin_lock(&n->list_lock);
3097 page = get_first_slab(n, false);
3098 if (!page)
3099 goto must_grow;
3100
3101 check_spinlock_acquired_node(cachep, nodeid);
3102
3103 STATS_INC_NODEALLOCS(cachep);
3104 STATS_INC_ACTIVE(cachep);
3105 STATS_SET_HIGH(cachep);
3106
3107 BUG_ON(page->active == cachep->num);
3108
3109 obj = slab_get_obj(cachep, page);
3110 n->free_objects--;
3111
3112 fixup_slab_list(cachep, n, page, &list);
3113
3114 spin_unlock(&n->list_lock);
3115 fixup_objfreelist_debug(cachep, &list);
3116 goto done;
3117
3118must_grow:
3119 spin_unlock(&n->list_lock);
3120 x = cache_grow(cachep, gfp_exact_node(flags), nodeid, NULL);
3121 if (x)
3122 goto retry;
3123
3124 return fallback_alloc(cachep, flags);
3125
3126done:
3127 return obj;
3128}
3129
3130static __always_inline void *
3131slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3132 unsigned long caller)
3133{
3134 unsigned long save_flags;
3135 void *ptr;
3136 int slab_node = numa_mem_id();
3137
3138 flags &= gfp_allowed_mask;
3139 cachep = slab_pre_alloc_hook(cachep, flags);
3140 if (unlikely(!cachep))
3141 return NULL;
3142
3143 cache_alloc_debugcheck_before(cachep, flags);
3144 local_irq_save(save_flags);
3145
3146 if (nodeid == NUMA_NO_NODE)
3147 nodeid = slab_node;
3148
3149 if (unlikely(!get_node(cachep, nodeid))) {
3150 /* Node not bootstrapped yet */
3151 ptr = fallback_alloc(cachep, flags);
3152 goto out;
3153 }
3154
3155 if (nodeid == slab_node) {
3156 /*
3157 * Use the locally cached objects if possible.
3158 * However ____cache_alloc does not allow fallback
3159 * to other nodes. It may fail while we still have
3160 * objects on other nodes available.
3161 */
3162 ptr = ____cache_alloc(cachep, flags);
3163 if (ptr)
3164 goto out;
3165 }
3166 /* ___cache_alloc_node can fall back to other nodes */
3167 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3168 out:
3169 local_irq_restore(save_flags);
3170 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3171
3172 if (unlikely(flags & __GFP_ZERO) && ptr)
3173 memset(ptr, 0, cachep->object_size);
3174
3175 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3176 return ptr;
3177}
3178
3179static __always_inline void *
3180__do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3181{
3182 void *objp;
3183
3184 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3185 objp = alternate_node_alloc(cache, flags);
3186 if (objp)
3187 goto out;
3188 }
3189 objp = ____cache_alloc(cache, flags);
3190
3191 /*
3192 * We may just have run out of memory on the local node.
3193 * ____cache_alloc_node() knows how to locate memory on other nodes
3194 */
3195 if (!objp)
3196 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3197
3198 out:
3199 return objp;
3200}
3201#else
3202
3203static __always_inline void *
3204__do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3205{
3206 return ____cache_alloc(cachep, flags);
3207}
3208
3209#endif /* CONFIG_NUMA */
3210
3211static __always_inline void *
3212slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3213{
3214 unsigned long save_flags;
3215 void *objp;
3216
3217 flags &= gfp_allowed_mask;
3218 cachep = slab_pre_alloc_hook(cachep, flags);
3219 if (unlikely(!cachep))
3220 return NULL;
3221
3222 cache_alloc_debugcheck_before(cachep, flags);
3223 local_irq_save(save_flags);
3224 objp = __do_cache_alloc(cachep, flags);
3225 local_irq_restore(save_flags);
3226 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3227 prefetchw(objp);
3228
3229 if (unlikely(flags & __GFP_ZERO) && objp)
3230 memset(objp, 0, cachep->object_size);
3231
3232 slab_post_alloc_hook(cachep, flags, 1, &objp);
3233 return objp;
3234}
3235
3236/*
3237 * Caller needs to acquire correct kmem_cache_node's list_lock
3238 * @list: List of detached free slabs should be freed by caller
3239 */
3240static void free_block(struct kmem_cache *cachep, void **objpp,
3241 int nr_objects, int node, struct list_head *list)
3242{
3243 int i;
3244 struct kmem_cache_node *n = get_node(cachep, node);
3245
3246 for (i = 0; i < nr_objects; i++) {
3247 void *objp;
3248 struct page *page;
3249
3250 objp = objpp[i];
3251
3252 page = virt_to_head_page(objp);
3253 list_del(&page->lru);
3254 check_spinlock_acquired_node(cachep, node);
3255 slab_put_obj(cachep, page, objp);
3256 STATS_DEC_ACTIVE(cachep);
3257 n->free_objects++;
3258
3259 /* fixup slab chains */
3260 if (page->active == 0) {
3261 if (n->free_objects > n->free_limit) {
3262 n->free_objects -= cachep->num;
3263 list_add_tail(&page->lru, list);
3264 } else {
3265 list_add(&page->lru, &n->slabs_free);
3266 }
3267 } else {
3268 /* Unconditionally move a slab to the end of the
3269 * partial list on free - maximum time for the
3270 * other objects to be freed, too.
3271 */
3272 list_add_tail(&page->lru, &n->slabs_partial);
3273 }
3274 }
3275}
3276
3277static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3278{
3279 int batchcount;
3280 struct kmem_cache_node *n;
3281 int node = numa_mem_id();
3282 LIST_HEAD(list);
3283
3284 batchcount = ac->batchcount;
3285
3286 check_irq_off();
3287 n = get_node(cachep, node);
3288 spin_lock(&n->list_lock);
3289 if (n->shared) {
3290 struct array_cache *shared_array = n->shared;
3291 int max = shared_array->limit - shared_array->avail;
3292 if (max) {
3293 if (batchcount > max)
3294 batchcount = max;
3295 memcpy(&(shared_array->entry[shared_array->avail]),
3296 ac->entry, sizeof(void *) * batchcount);
3297 shared_array->avail += batchcount;
3298 goto free_done;
3299 }
3300 }
3301
3302 free_block(cachep, ac->entry, batchcount, node, &list);
3303free_done:
3304#if STATS
3305 {
3306 int i = 0;
3307 struct page *page;
3308
3309 list_for_each_entry(page, &n->slabs_free, lru) {
3310 BUG_ON(page->active);
3311
3312 i++;
3313 }
3314 STATS_SET_FREEABLE(cachep, i);
3315 }
3316#endif
3317 spin_unlock(&n->list_lock);
3318 slabs_destroy(cachep, &list);
3319 ac->avail -= batchcount;
3320 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3321}
3322
3323/*
3324 * Release an obj back to its cache. If the obj has a constructed state, it must
3325 * be in this state _before_ it is released. Called with disabled ints.
3326 */
3327static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3328 unsigned long caller)
3329{
3330 struct array_cache *ac = cpu_cache_get(cachep);
3331
3332 kasan_slab_free(cachep, objp);
3333
3334 check_irq_off();
3335 kmemleak_free_recursive(objp, cachep->flags);
3336 objp = cache_free_debugcheck(cachep, objp, caller);
3337
3338 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3339
3340 /*
3341 * Skip calling cache_free_alien() when the platform is not numa.
3342 * This will avoid cache misses that happen while accessing slabp (which
3343 * is per page memory reference) to get nodeid. Instead use a global
3344 * variable to skip the call, which is mostly likely to be present in
3345 * the cache.
3346 */
3347 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3348 return;
3349
3350 if (ac->avail < ac->limit) {
3351 STATS_INC_FREEHIT(cachep);
3352 } else {
3353 STATS_INC_FREEMISS(cachep);
3354 cache_flusharray(cachep, ac);
3355 }
3356
3357 if (sk_memalloc_socks()) {
3358 struct page *page = virt_to_head_page(objp);
3359
3360 if (unlikely(PageSlabPfmemalloc(page))) {
3361 cache_free_pfmemalloc(cachep, page, objp);
3362 return;
3363 }
3364 }
3365
3366 ac->entry[ac->avail++] = objp;
3367}
3368
3369/**
3370 * kmem_cache_alloc - Allocate an object
3371 * @cachep: The cache to allocate from.
3372 * @flags: See kmalloc().
3373 *
3374 * Allocate an object from this cache. The flags are only relevant
3375 * if the cache has no available objects.
3376 */
3377void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3378{
3379 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3380
3381 kasan_slab_alloc(cachep, ret, flags);
3382 trace_kmem_cache_alloc(_RET_IP_, ret,
3383 cachep->object_size, cachep->size, flags);
3384
3385 return ret;
3386}
3387EXPORT_SYMBOL(kmem_cache_alloc);
3388
3389static __always_inline void
3390cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3391 size_t size, void **p, unsigned long caller)
3392{
3393 size_t i;
3394
3395 for (i = 0; i < size; i++)
3396 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3397}
3398
3399int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3400 void **p)
3401{
3402 size_t i;
3403
3404 s = slab_pre_alloc_hook(s, flags);
3405 if (!s)
3406 return 0;
3407
3408 cache_alloc_debugcheck_before(s, flags);
3409
3410 local_irq_disable();
3411 for (i = 0; i < size; i++) {
3412 void *objp = __do_cache_alloc(s, flags);
3413
3414 if (unlikely(!objp))
3415 goto error;
3416 p[i] = objp;
3417 }
3418 local_irq_enable();
3419
3420 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3421
3422 /* Clear memory outside IRQ disabled section */
3423 if (unlikely(flags & __GFP_ZERO))
3424 for (i = 0; i < size; i++)
3425 memset(p[i], 0, s->object_size);
3426
3427 slab_post_alloc_hook(s, flags, size, p);
3428 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3429 return size;
3430error:
3431 local_irq_enable();
3432 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3433 slab_post_alloc_hook(s, flags, i, p);
3434 __kmem_cache_free_bulk(s, i, p);
3435 return 0;
3436}
3437EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3438
3439#ifdef CONFIG_TRACING
3440void *
3441kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3442{
3443 void *ret;
3444
3445 ret = slab_alloc(cachep, flags, _RET_IP_);
3446
3447 kasan_kmalloc(cachep, ret, size, flags);
3448 trace_kmalloc(_RET_IP_, ret,
3449 size, cachep->size, flags);
3450 return ret;
3451}
3452EXPORT_SYMBOL(kmem_cache_alloc_trace);
3453#endif
3454
3455#ifdef CONFIG_NUMA
3456/**
3457 * kmem_cache_alloc_node - Allocate an object on the specified node
3458 * @cachep: The cache to allocate from.
3459 * @flags: See kmalloc().
3460 * @nodeid: node number of the target node.
3461 *
3462 * Identical to kmem_cache_alloc but it will allocate memory on the given
3463 * node, which can improve the performance for cpu bound structures.
3464 *
3465 * Fallback to other node is possible if __GFP_THISNODE is not set.
3466 */
3467void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3468{
3469 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3470
3471 kasan_slab_alloc(cachep, ret, flags);
3472 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3473 cachep->object_size, cachep->size,
3474 flags, nodeid);
3475
3476 return ret;
3477}
3478EXPORT_SYMBOL(kmem_cache_alloc_node);
3479
3480#ifdef CONFIG_TRACING
3481void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3482 gfp_t flags,
3483 int nodeid,
3484 size_t size)
3485{
3486 void *ret;
3487
3488 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3489
3490 kasan_kmalloc(cachep, ret, size, flags);
3491 trace_kmalloc_node(_RET_IP_, ret,
3492 size, cachep->size,
3493 flags, nodeid);
3494 return ret;
3495}
3496EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3497#endif
3498
3499static __always_inline void *
3500__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3501{
3502 struct kmem_cache *cachep;
3503 void *ret;
3504
3505 cachep = kmalloc_slab(size, flags);
3506 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3507 return cachep;
3508 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3509 kasan_kmalloc(cachep, ret, size, flags);
3510
3511 return ret;
3512}
3513
3514void *__kmalloc_node(size_t size, gfp_t flags, int node)
3515{
3516 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3517}
3518EXPORT_SYMBOL(__kmalloc_node);
3519
3520void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3521 int node, unsigned long caller)
3522{
3523 return __do_kmalloc_node(size, flags, node, caller);
3524}
3525EXPORT_SYMBOL(__kmalloc_node_track_caller);
3526#endif /* CONFIG_NUMA */
3527
3528/**
3529 * __do_kmalloc - allocate memory
3530 * @size: how many bytes of memory are required.
3531 * @flags: the type of memory to allocate (see kmalloc).
3532 * @caller: function caller for debug tracking of the caller
3533 */
3534static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3535 unsigned long caller)
3536{
3537 struct kmem_cache *cachep;
3538 void *ret;
3539
3540 cachep = kmalloc_slab(size, flags);
3541 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3542 return cachep;
3543 ret = slab_alloc(cachep, flags, caller);
3544
3545 kasan_kmalloc(cachep, ret, size, flags);
3546 trace_kmalloc(caller, ret,
3547 size, cachep->size, flags);
3548
3549 return ret;
3550}
3551
3552void *__kmalloc(size_t size, gfp_t flags)
3553{
3554 return __do_kmalloc(size, flags, _RET_IP_);
3555}
3556EXPORT_SYMBOL(__kmalloc);
3557
3558void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3559{
3560 return __do_kmalloc(size, flags, caller);
3561}
3562EXPORT_SYMBOL(__kmalloc_track_caller);
3563
3564/**
3565 * kmem_cache_free - Deallocate an object
3566 * @cachep: The cache the allocation was from.
3567 * @objp: The previously allocated object.
3568 *
3569 * Free an object which was previously allocated from this
3570 * cache.
3571 */
3572void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3573{
3574 unsigned long flags;
3575 cachep = cache_from_obj(cachep, objp);
3576 if (!cachep)
3577 return;
3578
3579 local_irq_save(flags);
3580 debug_check_no_locks_freed(objp, cachep->object_size);
3581 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3582 debug_check_no_obj_freed(objp, cachep->object_size);
3583 __cache_free(cachep, objp, _RET_IP_);
3584 local_irq_restore(flags);
3585
3586 trace_kmem_cache_free(_RET_IP_, objp);
3587}
3588EXPORT_SYMBOL(kmem_cache_free);
3589
3590void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3591{
3592 struct kmem_cache *s;
3593 size_t i;
3594
3595 local_irq_disable();
3596 for (i = 0; i < size; i++) {
3597 void *objp = p[i];
3598
3599 if (!orig_s) /* called via kfree_bulk */
3600 s = virt_to_cache(objp);
3601 else
3602 s = cache_from_obj(orig_s, objp);
3603
3604 debug_check_no_locks_freed(objp, s->object_size);
3605 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3606 debug_check_no_obj_freed(objp, s->object_size);
3607
3608 __cache_free(s, objp, _RET_IP_);
3609 }
3610 local_irq_enable();
3611
3612 /* FIXME: add tracing */
3613}
3614EXPORT_SYMBOL(kmem_cache_free_bulk);
3615
3616/**
3617 * kfree - free previously allocated memory
3618 * @objp: pointer returned by kmalloc.
3619 *
3620 * If @objp is NULL, no operation is performed.
3621 *
3622 * Don't free memory not originally allocated by kmalloc()
3623 * or you will run into trouble.
3624 */
3625void kfree(const void *objp)
3626{
3627 struct kmem_cache *c;
3628 unsigned long flags;
3629
3630 trace_kfree(_RET_IP_, objp);
3631
3632 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3633 return;
3634 local_irq_save(flags);
3635 kfree_debugcheck(objp);
3636 c = virt_to_cache(objp);
3637 debug_check_no_locks_freed(objp, c->object_size);
3638
3639 debug_check_no_obj_freed(objp, c->object_size);
3640 __cache_free(c, (void *)objp, _RET_IP_);
3641 local_irq_restore(flags);
3642}
3643EXPORT_SYMBOL(kfree);
3644
3645/*
3646 * This initializes kmem_cache_node or resizes various caches for all nodes.
3647 */
3648static int alloc_kmem_cache_node(struct kmem_cache *cachep, gfp_t gfp)
3649{
3650 int node;
3651 struct kmem_cache_node *n;
3652 struct array_cache *new_shared;
3653 struct alien_cache **new_alien = NULL;
3654
3655 for_each_online_node(node) {
3656
3657 if (use_alien_caches) {
3658 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
3659 if (!new_alien)
3660 goto fail;
3661 }
3662
3663 new_shared = NULL;
3664 if (cachep->shared) {
3665 new_shared = alloc_arraycache(node,
3666 cachep->shared*cachep->batchcount,
3667 0xbaadf00d, gfp);
3668 if (!new_shared) {
3669 free_alien_cache(new_alien);
3670 goto fail;
3671 }
3672 }
3673
3674 n = get_node(cachep, node);
3675 if (n) {
3676 struct array_cache *shared = n->shared;
3677 LIST_HEAD(list);
3678
3679 spin_lock_irq(&n->list_lock);
3680
3681 if (shared)
3682 free_block(cachep, shared->entry,
3683 shared->avail, node, &list);
3684
3685 n->shared = new_shared;
3686 if (!n->alien) {
3687 n->alien = new_alien;
3688 new_alien = NULL;
3689 }
3690 n->free_limit = (1 + nr_cpus_node(node)) *
3691 cachep->batchcount + cachep->num;
3692 spin_unlock_irq(&n->list_lock);
3693 slabs_destroy(cachep, &list);
3694 kfree(shared);
3695 free_alien_cache(new_alien);
3696 continue;
3697 }
3698 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
3699 if (!n) {
3700 free_alien_cache(new_alien);
3701 kfree(new_shared);
3702 goto fail;
3703 }
3704
3705 kmem_cache_node_init(n);
3706 n->next_reap = jiffies + REAPTIMEOUT_NODE +
3707 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
3708 n->shared = new_shared;
3709 n->alien = new_alien;
3710 n->free_limit = (1 + nr_cpus_node(node)) *
3711 cachep->batchcount + cachep->num;
3712 cachep->node[node] = n;
3713 }
3714 return 0;
3715
3716fail:
3717 if (!cachep->list.next) {
3718 /* Cache is not active yet. Roll back what we did */
3719 node--;
3720 while (node >= 0) {
3721 n = get_node(cachep, node);
3722 if (n) {
3723 kfree(n->shared);
3724 free_alien_cache(n->alien);
3725 kfree(n);
3726 cachep->node[node] = NULL;
3727 }
3728 node--;
3729 }
3730 }
3731 return -ENOMEM;
3732}
3733
3734/* Always called with the slab_mutex held */
3735static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3736 int batchcount, int shared, gfp_t gfp)
3737{
3738 struct array_cache __percpu *cpu_cache, *prev;
3739 int cpu;
3740
3741 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3742 if (!cpu_cache)
3743 return -ENOMEM;
3744
3745 prev = cachep->cpu_cache;
3746 cachep->cpu_cache = cpu_cache;
3747 kick_all_cpus_sync();
3748
3749 check_irq_on();
3750 cachep->batchcount = batchcount;
3751 cachep->limit = limit;
3752 cachep->shared = shared;
3753
3754 if (!prev)
3755 goto alloc_node;
3756
3757 for_each_online_cpu(cpu) {
3758 LIST_HEAD(list);
3759 int node;
3760 struct kmem_cache_node *n;
3761 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3762
3763 node = cpu_to_mem(cpu);
3764 n = get_node(cachep, node);
3765 spin_lock_irq(&n->list_lock);
3766 free_block(cachep, ac->entry, ac->avail, node, &list);
3767 spin_unlock_irq(&n->list_lock);
3768 slabs_destroy(cachep, &list);
3769 }
3770 free_percpu(prev);
3771
3772alloc_node:
3773 return alloc_kmem_cache_node(cachep, gfp);
3774}
3775
3776static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3777 int batchcount, int shared, gfp_t gfp)
3778{
3779 int ret;
3780 struct kmem_cache *c;
3781
3782 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3783
3784 if (slab_state < FULL)
3785 return ret;
3786
3787 if ((ret < 0) || !is_root_cache(cachep))
3788 return ret;
3789
3790 lockdep_assert_held(&slab_mutex);
3791 for_each_memcg_cache(c, cachep) {
3792 /* return value determined by the root cache only */
3793 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3794 }
3795
3796 return ret;
3797}
3798
3799/* Called with slab_mutex held always */
3800static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3801{
3802 int err;
3803 int limit = 0;
3804 int shared = 0;
3805 int batchcount = 0;
3806
3807 if (!is_root_cache(cachep)) {
3808 struct kmem_cache *root = memcg_root_cache(cachep);
3809 limit = root->limit;
3810 shared = root->shared;
3811 batchcount = root->batchcount;
3812 }
3813
3814 if (limit && shared && batchcount)
3815 goto skip_setup;
3816 /*
3817 * The head array serves three purposes:
3818 * - create a LIFO ordering, i.e. return objects that are cache-warm
3819 * - reduce the number of spinlock operations.
3820 * - reduce the number of linked list operations on the slab and
3821 * bufctl chains: array operations are cheaper.
3822 * The numbers are guessed, we should auto-tune as described by
3823 * Bonwick.
3824 */
3825 if (cachep->size > 131072)
3826 limit = 1;
3827 else if (cachep->size > PAGE_SIZE)
3828 limit = 8;
3829 else if (cachep->size > 1024)
3830 limit = 24;
3831 else if (cachep->size > 256)
3832 limit = 54;
3833 else
3834 limit = 120;
3835
3836 /*
3837 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
3838 * allocation behaviour: Most allocs on one cpu, most free operations
3839 * on another cpu. For these cases, an efficient object passing between
3840 * cpus is necessary. This is provided by a shared array. The array
3841 * replaces Bonwick's magazine layer.
3842 * On uniprocessor, it's functionally equivalent (but less efficient)
3843 * to a larger limit. Thus disabled by default.
3844 */
3845 shared = 0;
3846 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
3847 shared = 8;
3848
3849#if DEBUG
3850 /*
3851 * With debugging enabled, large batchcount lead to excessively long
3852 * periods with disabled local interrupts. Limit the batchcount
3853 */
3854 if (limit > 32)
3855 limit = 32;
3856#endif
3857 batchcount = (limit + 1) / 2;
3858skip_setup:
3859 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3860 if (err)
3861 pr_err("enable_cpucache failed for %s, error %d\n",
3862 cachep->name, -err);
3863 return err;
3864}
3865
3866/*
3867 * Drain an array if it contains any elements taking the node lock only if
3868 * necessary. Note that the node listlock also protects the array_cache
3869 * if drain_array() is used on the shared array.
3870 */
3871static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
3872 struct array_cache *ac, int force, int node)
3873{
3874 LIST_HEAD(list);
3875 int tofree;
3876
3877 if (!ac || !ac->avail)
3878 return;
3879 if (ac->touched && !force) {
3880 ac->touched = 0;
3881 } else {
3882 spin_lock_irq(&n->list_lock);
3883 if (ac->avail) {
3884 tofree = force ? ac->avail : (ac->limit + 4) / 5;
3885 if (tofree > ac->avail)
3886 tofree = (ac->avail + 1) / 2;
3887 free_block(cachep, ac->entry, tofree, node, &list);
3888 ac->avail -= tofree;
3889 memmove(ac->entry, &(ac->entry[tofree]),
3890 sizeof(void *) * ac->avail);
3891 }
3892 spin_unlock_irq(&n->list_lock);
3893 slabs_destroy(cachep, &list);
3894 }
3895}
3896
3897/**
3898 * cache_reap - Reclaim memory from caches.
3899 * @w: work descriptor
3900 *
3901 * Called from workqueue/eventd every few seconds.
3902 * Purpose:
3903 * - clear the per-cpu caches for this CPU.
3904 * - return freeable pages to the main free memory pool.
3905 *
3906 * If we cannot acquire the cache chain mutex then just give up - we'll try
3907 * again on the next iteration.
3908 */
3909static void cache_reap(struct work_struct *w)
3910{
3911 struct kmem_cache *searchp;
3912 struct kmem_cache_node *n;
3913 int node = numa_mem_id();
3914 struct delayed_work *work = to_delayed_work(w);
3915
3916 if (!mutex_trylock(&slab_mutex))
3917 /* Give up. Setup the next iteration. */
3918 goto out;
3919
3920 list_for_each_entry(searchp, &slab_caches, list) {
3921 check_irq_on();
3922
3923 /*
3924 * We only take the node lock if absolutely necessary and we
3925 * have established with reasonable certainty that
3926 * we can do some work if the lock was obtained.
3927 */
3928 n = get_node(searchp, node);
3929
3930 reap_alien(searchp, n);
3931
3932 drain_array(searchp, n, cpu_cache_get(searchp), 0, node);
3933
3934 /*
3935 * These are racy checks but it does not matter
3936 * if we skip one check or scan twice.
3937 */
3938 if (time_after(n->next_reap, jiffies))
3939 goto next;
3940
3941 n->next_reap = jiffies + REAPTIMEOUT_NODE;
3942
3943 drain_array(searchp, n, n->shared, 0, node);
3944
3945 if (n->free_touched)
3946 n->free_touched = 0;
3947 else {
3948 int freed;
3949
3950 freed = drain_freelist(searchp, n, (n->free_limit +
3951 5 * searchp->num - 1) / (5 * searchp->num));
3952 STATS_ADD_REAPED(searchp, freed);
3953 }
3954next:
3955 cond_resched();
3956 }
3957 check_irq_on();
3958 mutex_unlock(&slab_mutex);
3959 next_reap_node();
3960out:
3961 /* Set up the next iteration */
3962 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
3963}
3964
3965#ifdef CONFIG_SLABINFO
3966void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
3967{
3968 struct page *page;
3969 unsigned long active_objs;
3970 unsigned long num_objs;
3971 unsigned long active_slabs = 0;
3972 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
3973 const char *name;
3974 char *error = NULL;
3975 int node;
3976 struct kmem_cache_node *n;
3977
3978 active_objs = 0;
3979 num_slabs = 0;
3980 for_each_kmem_cache_node(cachep, node, n) {
3981
3982 check_irq_on();
3983 spin_lock_irq(&n->list_lock);
3984
3985 list_for_each_entry(page, &n->slabs_full, lru) {
3986 if (page->active != cachep->num && !error)
3987 error = "slabs_full accounting error";
3988 active_objs += cachep->num;
3989 active_slabs++;
3990 }
3991 list_for_each_entry(page, &n->slabs_partial, lru) {
3992 if (page->active == cachep->num && !error)
3993 error = "slabs_partial accounting error";
3994 if (!page->active && !error)
3995 error = "slabs_partial accounting error";
3996 active_objs += page->active;
3997 active_slabs++;
3998 }
3999 list_for_each_entry(page, &n->slabs_free, lru) {
4000 if (page->active && !error)
4001 error = "slabs_free accounting error";
4002 num_slabs++;
4003 }
4004 free_objects += n->free_objects;
4005 if (n->shared)
4006 shared_avail += n->shared->avail;
4007
4008 spin_unlock_irq(&n->list_lock);
4009 }
4010 num_slabs += active_slabs;
4011 num_objs = num_slabs * cachep->num;
4012 if (num_objs - active_objs != free_objects && !error)
4013 error = "free_objects accounting error";
4014
4015 name = cachep->name;
4016 if (error)
4017 pr_err("slab: cache %s error: %s\n", name, error);
4018
4019 sinfo->active_objs = active_objs;
4020 sinfo->num_objs = num_objs;
4021 sinfo->active_slabs = active_slabs;
4022 sinfo->num_slabs = num_slabs;
4023 sinfo->shared_avail = shared_avail;
4024 sinfo->limit = cachep->limit;
4025 sinfo->batchcount = cachep->batchcount;
4026 sinfo->shared = cachep->shared;
4027 sinfo->objects_per_slab = cachep->num;
4028 sinfo->cache_order = cachep->gfporder;
4029}
4030
4031void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4032{
4033#if STATS
4034 { /* node stats */
4035 unsigned long high = cachep->high_mark;
4036 unsigned long allocs = cachep->num_allocations;
4037 unsigned long grown = cachep->grown;
4038 unsigned long reaped = cachep->reaped;
4039 unsigned long errors = cachep->errors;
4040 unsigned long max_freeable = cachep->max_freeable;
4041 unsigned long node_allocs = cachep->node_allocs;
4042 unsigned long node_frees = cachep->node_frees;
4043 unsigned long overflows = cachep->node_overflow;
4044
4045 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4046 allocs, high, grown,
4047 reaped, errors, max_freeable, node_allocs,
4048 node_frees, overflows);
4049 }
4050 /* cpu stats */
4051 {
4052 unsigned long allochit = atomic_read(&cachep->allochit);
4053 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4054 unsigned long freehit = atomic_read(&cachep->freehit);
4055 unsigned long freemiss = atomic_read(&cachep->freemiss);
4056
4057 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4058 allochit, allocmiss, freehit, freemiss);
4059 }
4060#endif
4061}
4062
4063#define MAX_SLABINFO_WRITE 128
4064/**
4065 * slabinfo_write - Tuning for the slab allocator
4066 * @file: unused
4067 * @buffer: user buffer
4068 * @count: data length
4069 * @ppos: unused
4070 */
4071ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4072 size_t count, loff_t *ppos)
4073{
4074 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4075 int limit, batchcount, shared, res;
4076 struct kmem_cache *cachep;
4077
4078 if (count > MAX_SLABINFO_WRITE)
4079 return -EINVAL;
4080 if (copy_from_user(&kbuf, buffer, count))
4081 return -EFAULT;
4082 kbuf[MAX_SLABINFO_WRITE] = '\0';
4083
4084 tmp = strchr(kbuf, ' ');
4085 if (!tmp)
4086 return -EINVAL;
4087 *tmp = '\0';
4088 tmp++;
4089 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4090 return -EINVAL;
4091
4092 /* Find the cache in the chain of caches. */
4093 mutex_lock(&slab_mutex);
4094 res = -EINVAL;
4095 list_for_each_entry(cachep, &slab_caches, list) {
4096 if (!strcmp(cachep->name, kbuf)) {
4097 if (limit < 1 || batchcount < 1 ||
4098 batchcount > limit || shared < 0) {
4099 res = 0;
4100 } else {
4101 res = do_tune_cpucache(cachep, limit,
4102 batchcount, shared,
4103 GFP_KERNEL);
4104 }
4105 break;
4106 }
4107 }
4108 mutex_unlock(&slab_mutex);
4109 if (res >= 0)
4110 res = count;
4111 return res;
4112}
4113
4114#ifdef CONFIG_DEBUG_SLAB_LEAK
4115
4116static inline int add_caller(unsigned long *n, unsigned long v)
4117{
4118 unsigned long *p;
4119 int l;
4120 if (!v)
4121 return 1;
4122 l = n[1];
4123 p = n + 2;
4124 while (l) {
4125 int i = l/2;
4126 unsigned long *q = p + 2 * i;
4127 if (*q == v) {
4128 q[1]++;
4129 return 1;
4130 }
4131 if (*q > v) {
4132 l = i;
4133 } else {
4134 p = q + 2;
4135 l -= i + 1;
4136 }
4137 }
4138 if (++n[1] == n[0])
4139 return 0;
4140 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4141 p[0] = v;
4142 p[1] = 1;
4143 return 1;
4144}
4145
4146static void handle_slab(unsigned long *n, struct kmem_cache *c,
4147 struct page *page)
4148{
4149 void *p;
4150 int i, j;
4151 unsigned long v;
4152
4153 if (n[0] == n[1])
4154 return;
4155 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4156 bool active = true;
4157
4158 for (j = page->active; j < c->num; j++) {
4159 if (get_free_obj(page, j) == i) {
4160 active = false;
4161 break;
4162 }
4163 }
4164
4165 if (!active)
4166 continue;
4167
4168 /*
4169 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4170 * mapping is established when actual object allocation and
4171 * we could mistakenly access the unmapped object in the cpu
4172 * cache.
4173 */
4174 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4175 continue;
4176
4177 if (!add_caller(n, v))
4178 return;
4179 }
4180}
4181
4182static void show_symbol(struct seq_file *m, unsigned long address)
4183{
4184#ifdef CONFIG_KALLSYMS
4185 unsigned long offset, size;
4186 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4187
4188 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4189 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4190 if (modname[0])
4191 seq_printf(m, " [%s]", modname);
4192 return;
4193 }
4194#endif
4195 seq_printf(m, "%p", (void *)address);
4196}
4197
4198static int leaks_show(struct seq_file *m, void *p)
4199{
4200 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4201 struct page *page;
4202 struct kmem_cache_node *n;
4203 const char *name;
4204 unsigned long *x = m->private;
4205 int node;
4206 int i;
4207
4208 if (!(cachep->flags & SLAB_STORE_USER))
4209 return 0;
4210 if (!(cachep->flags & SLAB_RED_ZONE))
4211 return 0;
4212
4213 /*
4214 * Set store_user_clean and start to grab stored user information
4215 * for all objects on this cache. If some alloc/free requests comes
4216 * during the processing, information would be wrong so restart
4217 * whole processing.
4218 */
4219 do {
4220 set_store_user_clean(cachep);
4221 drain_cpu_caches(cachep);
4222
4223 x[1] = 0;
4224
4225 for_each_kmem_cache_node(cachep, node, n) {
4226
4227 check_irq_on();
4228 spin_lock_irq(&n->list_lock);
4229
4230 list_for_each_entry(page, &n->slabs_full, lru)
4231 handle_slab(x, cachep, page);
4232 list_for_each_entry(page, &n->slabs_partial, lru)
4233 handle_slab(x, cachep, page);
4234 spin_unlock_irq(&n->list_lock);
4235 }
4236 } while (!is_store_user_clean(cachep));
4237
4238 name = cachep->name;
4239 if (x[0] == x[1]) {
4240 /* Increase the buffer size */
4241 mutex_unlock(&slab_mutex);
4242 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4243 if (!m->private) {
4244 /* Too bad, we are really out */
4245 m->private = x;
4246 mutex_lock(&slab_mutex);
4247 return -ENOMEM;
4248 }
4249 *(unsigned long *)m->private = x[0] * 2;
4250 kfree(x);
4251 mutex_lock(&slab_mutex);
4252 /* Now make sure this entry will be retried */
4253 m->count = m->size;
4254 return 0;
4255 }
4256 for (i = 0; i < x[1]; i++) {
4257 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4258 show_symbol(m, x[2*i+2]);
4259 seq_putc(m, '\n');
4260 }
4261
4262 return 0;
4263}
4264
4265static const struct seq_operations slabstats_op = {
4266 .start = slab_start,
4267 .next = slab_next,
4268 .stop = slab_stop,
4269 .show = leaks_show,
4270};
4271
4272static int slabstats_open(struct inode *inode, struct file *file)
4273{
4274 unsigned long *n;
4275
4276 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4277 if (!n)
4278 return -ENOMEM;
4279
4280 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4281
4282 return 0;
4283}
4284
4285static const struct file_operations proc_slabstats_operations = {
4286 .open = slabstats_open,
4287 .read = seq_read,
4288 .llseek = seq_lseek,
4289 .release = seq_release_private,
4290};
4291#endif
4292
4293static int __init slab_proc_init(void)
4294{
4295#ifdef CONFIG_DEBUG_SLAB_LEAK
4296 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4297#endif
4298 return 0;
4299}
4300module_init(slab_proc_init);
4301#endif
4302
4303/**
4304 * ksize - get the actual amount of memory allocated for a given object
4305 * @objp: Pointer to the object
4306 *
4307 * kmalloc may internally round up allocations and return more memory
4308 * than requested. ksize() can be used to determine the actual amount of
4309 * memory allocated. The caller may use this additional memory, even though
4310 * a smaller amount of memory was initially specified with the kmalloc call.
4311 * The caller must guarantee that objp points to a valid object previously
4312 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4313 * must not be freed during the duration of the call.
4314 */
4315size_t ksize(const void *objp)
4316{
4317 size_t size;
4318
4319 BUG_ON(!objp);
4320 if (unlikely(objp == ZERO_SIZE_PTR))
4321 return 0;
4322
4323 size = virt_to_cache(objp)->object_size;
4324 /* We assume that ksize callers could use the whole allocated area,
4325 * so we need to unpoison this area.
4326 */
4327 kasan_krealloc(objp, size, GFP_NOWAIT);
4328
4329 return size;
4330}
4331EXPORT_SYMBOL(ksize);