]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/slab.c
mm: slab: remove ZONE_DMA_FLAG
[mirror_ubuntu-bionic-kernel.git] / mm / slab.c
1 /*
2 * linux/mm/slab.c
3 * Written by Mark Hemment, 1996/97.
4 * (markhe@nextd.demon.co.uk)
5 *
6 * kmem_cache_destroy() + some cleanup - 1999 Andrea Arcangeli
7 *
8 * Major cleanup, different bufctl logic, per-cpu arrays
9 * (c) 2000 Manfred Spraul
10 *
11 * Cleanup, make the head arrays unconditional, preparation for NUMA
12 * (c) 2002 Manfred Spraul
13 *
14 * An implementation of the Slab Allocator as described in outline in;
15 * UNIX Internals: The New Frontiers by Uresh Vahalia
16 * Pub: Prentice Hall ISBN 0-13-101908-2
17 * or with a little more detail in;
18 * The Slab Allocator: An Object-Caching Kernel Memory Allocator
19 * Jeff Bonwick (Sun Microsystems).
20 * Presented at: USENIX Summer 1994 Technical Conference
21 *
22 * The memory is organized in caches, one cache for each object type.
23 * (e.g. inode_cache, dentry_cache, buffer_head, vm_area_struct)
24 * Each cache consists out of many slabs (they are small (usually one
25 * page long) and always contiguous), and each slab contains multiple
26 * initialized objects.
27 *
28 * This means, that your constructor is used only for newly allocated
29 * slabs and you must pass objects with the same initializations to
30 * kmem_cache_free.
31 *
32 * Each cache can only support one memory type (GFP_DMA, GFP_HIGHMEM,
33 * normal). If you need a special memory type, then must create a new
34 * cache for that memory type.
35 *
36 * In order to reduce fragmentation, the slabs are sorted in 3 groups:
37 * full slabs with 0 free objects
38 * partial slabs
39 * empty slabs with no allocated objects
40 *
41 * If partial slabs exist, then new allocations come from these slabs,
42 * otherwise from empty slabs or new slabs are allocated.
43 *
44 * kmem_cache_destroy() CAN CRASH if you try to allocate from the cache
45 * during kmem_cache_destroy(). The caller must prevent concurrent allocs.
46 *
47 * Each cache has a short per-cpu head array, most allocs
48 * and frees go into that array, and if that array overflows, then 1/2
49 * of the entries in the array are given back into the global cache.
50 * The head array is strictly LIFO and should improve the cache hit rates.
51 * On SMP, it additionally reduces the spinlock operations.
52 *
53 * The c_cpuarray may not be read with enabled local interrupts -
54 * it's changed with a smp_call_function().
55 *
56 * SMP synchronization:
57 * constructors and destructors are called without any locking.
58 * Several members in struct kmem_cache and struct slab never change, they
59 * are accessed without any locking.
60 * The per-cpu arrays are never accessed from the wrong cpu, no locking,
61 * and local interrupts are disabled so slab code is preempt-safe.
62 * The non-constant members are protected with a per-cache irq spinlock.
63 *
64 * Many thanks to Mark Hemment, who wrote another per-cpu slab patch
65 * in 2000 - many ideas in the current implementation are derived from
66 * his patch.
67 *
68 * Further notes from the original documentation:
69 *
70 * 11 April '97. Started multi-threading - markhe
71 * The global cache-chain is protected by the mutex 'slab_mutex'.
72 * The sem is only needed when accessing/extending the cache-chain, which
73 * can never happen inside an interrupt (kmem_cache_create(),
74 * kmem_cache_shrink() and kmem_cache_reap()).
75 *
76 * At present, each engine can be growing a cache. This should be blocked.
77 *
78 * 15 March 2005. NUMA slab allocator.
79 * Shai Fultheim <shai@scalex86.org>.
80 * Shobhit Dayal <shobhit@calsoftinc.com>
81 * Alok N Kataria <alokk@calsoftinc.com>
82 * Christoph Lameter <christoph@lameter.com>
83 *
84 * Modified the slab allocator to be node aware on NUMA systems.
85 * Each node has its own list of partial, free and full slabs.
86 * All object allocations for a node occur from node specific slab lists.
87 */
88
89 #include <linux/slab.h>
90 #include <linux/mm.h>
91 #include <linux/poison.h>
92 #include <linux/swap.h>
93 #include <linux/cache.h>
94 #include <linux/interrupt.h>
95 #include <linux/init.h>
96 #include <linux/compiler.h>
97 #include <linux/cpuset.h>
98 #include <linux/proc_fs.h>
99 #include <linux/seq_file.h>
100 #include <linux/notifier.h>
101 #include <linux/kallsyms.h>
102 #include <linux/cpu.h>
103 #include <linux/sysctl.h>
104 #include <linux/module.h>
105 #include <linux/rcupdate.h>
106 #include <linux/string.h>
107 #include <linux/uaccess.h>
108 #include <linux/nodemask.h>
109 #include <linux/kmemleak.h>
110 #include <linux/mempolicy.h>
111 #include <linux/mutex.h>
112 #include <linux/fault-inject.h>
113 #include <linux/rtmutex.h>
114 #include <linux/reciprocal_div.h>
115 #include <linux/debugobjects.h>
116 #include <linux/kmemcheck.h>
117 #include <linux/memory.h>
118 #include <linux/prefetch.h>
119
120 #include <net/sock.h>
121
122 #include <asm/cacheflush.h>
123 #include <asm/tlbflush.h>
124 #include <asm/page.h>
125
126 #include <trace/events/kmem.h>
127
128 #include "internal.h"
129
130 #include "slab.h"
131
132 /*
133 * DEBUG - 1 for kmem_cache_create() to honour; SLAB_RED_ZONE & SLAB_POISON.
134 * 0 for faster, smaller code (especially in the critical paths).
135 *
136 * STATS - 1 to collect stats for /proc/slabinfo.
137 * 0 for faster, smaller code (especially in the critical paths).
138 *
139 * FORCED_DEBUG - 1 enables SLAB_RED_ZONE and SLAB_POISON (if possible)
140 */
141
142 #ifdef CONFIG_DEBUG_SLAB
143 #define DEBUG 1
144 #define STATS 1
145 #define FORCED_DEBUG 1
146 #else
147 #define DEBUG 0
148 #define STATS 0
149 #define FORCED_DEBUG 0
150 #endif
151
152 /* Shouldn't this be in a header file somewhere? */
153 #define BYTES_PER_WORD sizeof(void *)
154 #define REDZONE_ALIGN max(BYTES_PER_WORD, __alignof__(unsigned long long))
155
156 #ifndef ARCH_KMALLOC_FLAGS
157 #define ARCH_KMALLOC_FLAGS SLAB_HWCACHE_ALIGN
158 #endif
159
160 #define FREELIST_BYTE_INDEX (((PAGE_SIZE >> BITS_PER_BYTE) \
161 <= SLAB_OBJ_MIN_SIZE) ? 1 : 0)
162
163 #if FREELIST_BYTE_INDEX
164 typedef unsigned char freelist_idx_t;
165 #else
166 typedef unsigned short freelist_idx_t;
167 #endif
168
169 #define SLAB_OBJ_MAX_NUM ((1 << sizeof(freelist_idx_t) * BITS_PER_BYTE) - 1)
170
171 /*
172 * struct array_cache
173 *
174 * Purpose:
175 * - LIFO ordering, to hand out cache-warm objects from _alloc
176 * - reduce the number of linked list operations
177 * - reduce spinlock operations
178 *
179 * The limit is stored in the per-cpu structure to reduce the data cache
180 * footprint.
181 *
182 */
183 struct array_cache {
184 unsigned int avail;
185 unsigned int limit;
186 unsigned int batchcount;
187 unsigned int touched;
188 void *entry[]; /*
189 * Must have this definition in here for the proper
190 * alignment of array_cache. Also simplifies accessing
191 * the entries.
192 */
193 };
194
195 struct alien_cache {
196 spinlock_t lock;
197 struct array_cache ac;
198 };
199
200 /*
201 * Need this for bootstrapping a per node allocator.
202 */
203 #define NUM_INIT_LISTS (2 * MAX_NUMNODES)
204 static struct kmem_cache_node __initdata init_kmem_cache_node[NUM_INIT_LISTS];
205 #define CACHE_CACHE 0
206 #define SIZE_NODE (MAX_NUMNODES)
207
208 static int drain_freelist(struct kmem_cache *cache,
209 struct kmem_cache_node *n, int tofree);
210 static void free_block(struct kmem_cache *cachep, void **objpp, int len,
211 int node, struct list_head *list);
212 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list);
213 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp);
214 static void cache_reap(struct work_struct *unused);
215
216 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
217 void **list);
218 static inline void fixup_slab_list(struct kmem_cache *cachep,
219 struct kmem_cache_node *n, struct page *page,
220 void **list);
221 static int slab_early_init = 1;
222
223 #define INDEX_NODE kmalloc_index(sizeof(struct kmem_cache_node))
224
225 static void kmem_cache_node_init(struct kmem_cache_node *parent)
226 {
227 INIT_LIST_HEAD(&parent->slabs_full);
228 INIT_LIST_HEAD(&parent->slabs_partial);
229 INIT_LIST_HEAD(&parent->slabs_free);
230 parent->shared = NULL;
231 parent->alien = NULL;
232 parent->colour_next = 0;
233 spin_lock_init(&parent->list_lock);
234 parent->free_objects = 0;
235 parent->free_touched = 0;
236 }
237
238 #define MAKE_LIST(cachep, listp, slab, nodeid) \
239 do { \
240 INIT_LIST_HEAD(listp); \
241 list_splice(&get_node(cachep, nodeid)->slab, listp); \
242 } while (0)
243
244 #define MAKE_ALL_LISTS(cachep, ptr, nodeid) \
245 do { \
246 MAKE_LIST((cachep), (&(ptr)->slabs_full), slabs_full, nodeid); \
247 MAKE_LIST((cachep), (&(ptr)->slabs_partial), slabs_partial, nodeid); \
248 MAKE_LIST((cachep), (&(ptr)->slabs_free), slabs_free, nodeid); \
249 } while (0)
250
251 #define CFLGS_OBJFREELIST_SLAB (0x40000000UL)
252 #define CFLGS_OFF_SLAB (0x80000000UL)
253 #define OBJFREELIST_SLAB(x) ((x)->flags & CFLGS_OBJFREELIST_SLAB)
254 #define OFF_SLAB(x) ((x)->flags & CFLGS_OFF_SLAB)
255
256 #define BATCHREFILL_LIMIT 16
257 /*
258 * Optimization question: fewer reaps means less probability for unnessary
259 * cpucache drain/refill cycles.
260 *
261 * OTOH the cpuarrays can contain lots of objects,
262 * which could lock up otherwise freeable slabs.
263 */
264 #define REAPTIMEOUT_AC (2*HZ)
265 #define REAPTIMEOUT_NODE (4*HZ)
266
267 #if STATS
268 #define STATS_INC_ACTIVE(x) ((x)->num_active++)
269 #define STATS_DEC_ACTIVE(x) ((x)->num_active--)
270 #define STATS_INC_ALLOCED(x) ((x)->num_allocations++)
271 #define STATS_INC_GROWN(x) ((x)->grown++)
272 #define STATS_ADD_REAPED(x,y) ((x)->reaped += (y))
273 #define STATS_SET_HIGH(x) \
274 do { \
275 if ((x)->num_active > (x)->high_mark) \
276 (x)->high_mark = (x)->num_active; \
277 } while (0)
278 #define STATS_INC_ERR(x) ((x)->errors++)
279 #define STATS_INC_NODEALLOCS(x) ((x)->node_allocs++)
280 #define STATS_INC_NODEFREES(x) ((x)->node_frees++)
281 #define STATS_INC_ACOVERFLOW(x) ((x)->node_overflow++)
282 #define STATS_SET_FREEABLE(x, i) \
283 do { \
284 if ((x)->max_freeable < i) \
285 (x)->max_freeable = i; \
286 } while (0)
287 #define STATS_INC_ALLOCHIT(x) atomic_inc(&(x)->allochit)
288 #define STATS_INC_ALLOCMISS(x) atomic_inc(&(x)->allocmiss)
289 #define STATS_INC_FREEHIT(x) atomic_inc(&(x)->freehit)
290 #define STATS_INC_FREEMISS(x) atomic_inc(&(x)->freemiss)
291 #else
292 #define STATS_INC_ACTIVE(x) do { } while (0)
293 #define STATS_DEC_ACTIVE(x) do { } while (0)
294 #define STATS_INC_ALLOCED(x) do { } while (0)
295 #define STATS_INC_GROWN(x) do { } while (0)
296 #define STATS_ADD_REAPED(x,y) do { (void)(y); } while (0)
297 #define STATS_SET_HIGH(x) do { } while (0)
298 #define STATS_INC_ERR(x) do { } while (0)
299 #define STATS_INC_NODEALLOCS(x) do { } while (0)
300 #define STATS_INC_NODEFREES(x) do { } while (0)
301 #define STATS_INC_ACOVERFLOW(x) do { } while (0)
302 #define STATS_SET_FREEABLE(x, i) do { } while (0)
303 #define STATS_INC_ALLOCHIT(x) do { } while (0)
304 #define STATS_INC_ALLOCMISS(x) do { } while (0)
305 #define STATS_INC_FREEHIT(x) do { } while (0)
306 #define STATS_INC_FREEMISS(x) do { } while (0)
307 #endif
308
309 #if DEBUG
310
311 /*
312 * memory layout of objects:
313 * 0 : objp
314 * 0 .. cachep->obj_offset - BYTES_PER_WORD - 1: padding. This ensures that
315 * the end of an object is aligned with the end of the real
316 * allocation. Catches writes behind the end of the allocation.
317 * cachep->obj_offset - BYTES_PER_WORD .. cachep->obj_offset - 1:
318 * redzone word.
319 * cachep->obj_offset: The real object.
320 * cachep->size - 2* BYTES_PER_WORD: redzone word [BYTES_PER_WORD long]
321 * cachep->size - 1* BYTES_PER_WORD: last caller address
322 * [BYTES_PER_WORD long]
323 */
324 static int obj_offset(struct kmem_cache *cachep)
325 {
326 return cachep->obj_offset;
327 }
328
329 static unsigned long long *dbg_redzone1(struct kmem_cache *cachep, void *objp)
330 {
331 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
332 return (unsigned long long*) (objp + obj_offset(cachep) -
333 sizeof(unsigned long long));
334 }
335
336 static unsigned long long *dbg_redzone2(struct kmem_cache *cachep, void *objp)
337 {
338 BUG_ON(!(cachep->flags & SLAB_RED_ZONE));
339 if (cachep->flags & SLAB_STORE_USER)
340 return (unsigned long long *)(objp + cachep->size -
341 sizeof(unsigned long long) -
342 REDZONE_ALIGN);
343 return (unsigned long long *) (objp + cachep->size -
344 sizeof(unsigned long long));
345 }
346
347 static void **dbg_userword(struct kmem_cache *cachep, void *objp)
348 {
349 BUG_ON(!(cachep->flags & SLAB_STORE_USER));
350 return (void **)(objp + cachep->size - BYTES_PER_WORD);
351 }
352
353 #else
354
355 #define obj_offset(x) 0
356 #define dbg_redzone1(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
357 #define dbg_redzone2(cachep, objp) ({BUG(); (unsigned long long *)NULL;})
358 #define dbg_userword(cachep, objp) ({BUG(); (void **)NULL;})
359
360 #endif
361
362 #ifdef CONFIG_DEBUG_SLAB_LEAK
363
364 static inline bool is_store_user_clean(struct kmem_cache *cachep)
365 {
366 return atomic_read(&cachep->store_user_clean) == 1;
367 }
368
369 static inline void set_store_user_clean(struct kmem_cache *cachep)
370 {
371 atomic_set(&cachep->store_user_clean, 1);
372 }
373
374 static inline void set_store_user_dirty(struct kmem_cache *cachep)
375 {
376 if (is_store_user_clean(cachep))
377 atomic_set(&cachep->store_user_clean, 0);
378 }
379
380 #else
381 static inline void set_store_user_dirty(struct kmem_cache *cachep) {}
382
383 #endif
384
385 /*
386 * Do not go above this order unless 0 objects fit into the slab or
387 * overridden on the command line.
388 */
389 #define SLAB_MAX_ORDER_HI 1
390 #define SLAB_MAX_ORDER_LO 0
391 static int slab_max_order = SLAB_MAX_ORDER_LO;
392 static bool slab_max_order_set __initdata;
393
394 static inline struct kmem_cache *virt_to_cache(const void *obj)
395 {
396 struct page *page = virt_to_head_page(obj);
397 return page->slab_cache;
398 }
399
400 static inline void *index_to_obj(struct kmem_cache *cache, struct page *page,
401 unsigned int idx)
402 {
403 return page->s_mem + cache->size * idx;
404 }
405
406 /*
407 * We want to avoid an expensive divide : (offset / cache->size)
408 * Using the fact that size is a constant for a particular cache,
409 * we can replace (offset / cache->size) by
410 * reciprocal_divide(offset, cache->reciprocal_buffer_size)
411 */
412 static inline unsigned int obj_to_index(const struct kmem_cache *cache,
413 const struct page *page, void *obj)
414 {
415 u32 offset = (obj - page->s_mem);
416 return reciprocal_divide(offset, cache->reciprocal_buffer_size);
417 }
418
419 #define BOOT_CPUCACHE_ENTRIES 1
420 /* internal cache of cache description objs */
421 static struct kmem_cache kmem_cache_boot = {
422 .batchcount = 1,
423 .limit = BOOT_CPUCACHE_ENTRIES,
424 .shared = 1,
425 .size = sizeof(struct kmem_cache),
426 .name = "kmem_cache",
427 };
428
429 static DEFINE_PER_CPU(struct delayed_work, slab_reap_work);
430
431 static inline struct array_cache *cpu_cache_get(struct kmem_cache *cachep)
432 {
433 return this_cpu_ptr(cachep->cpu_cache);
434 }
435
436 /*
437 * Calculate the number of objects and left-over bytes for a given buffer size.
438 */
439 static unsigned int cache_estimate(unsigned long gfporder, size_t buffer_size,
440 unsigned long flags, size_t *left_over)
441 {
442 unsigned int num;
443 size_t slab_size = PAGE_SIZE << gfporder;
444
445 /*
446 * The slab management structure can be either off the slab or
447 * on it. For the latter case, the memory allocated for a
448 * slab is used for:
449 *
450 * - @buffer_size bytes for each object
451 * - One freelist_idx_t for each object
452 *
453 * We don't need to consider alignment of freelist because
454 * freelist will be at the end of slab page. The objects will be
455 * at the correct alignment.
456 *
457 * If the slab management structure is off the slab, then the
458 * alignment will already be calculated into the size. Because
459 * the slabs are all pages aligned, the objects will be at the
460 * correct alignment when allocated.
461 */
462 if (flags & (CFLGS_OBJFREELIST_SLAB | CFLGS_OFF_SLAB)) {
463 num = slab_size / buffer_size;
464 *left_over = slab_size % buffer_size;
465 } else {
466 num = slab_size / (buffer_size + sizeof(freelist_idx_t));
467 *left_over = slab_size %
468 (buffer_size + sizeof(freelist_idx_t));
469 }
470
471 return num;
472 }
473
474 #if DEBUG
475 #define slab_error(cachep, msg) __slab_error(__func__, cachep, msg)
476
477 static void __slab_error(const char *function, struct kmem_cache *cachep,
478 char *msg)
479 {
480 pr_err("slab error in %s(): cache `%s': %s\n",
481 function, cachep->name, msg);
482 dump_stack();
483 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
484 }
485 #endif
486
487 /*
488 * By default on NUMA we use alien caches to stage the freeing of
489 * objects allocated from other nodes. This causes massive memory
490 * inefficiencies when using fake NUMA setup to split memory into a
491 * large number of small nodes, so it can be disabled on the command
492 * line
493 */
494
495 static int use_alien_caches __read_mostly = 1;
496 static int __init noaliencache_setup(char *s)
497 {
498 use_alien_caches = 0;
499 return 1;
500 }
501 __setup("noaliencache", noaliencache_setup);
502
503 static int __init slab_max_order_setup(char *str)
504 {
505 get_option(&str, &slab_max_order);
506 slab_max_order = slab_max_order < 0 ? 0 :
507 min(slab_max_order, MAX_ORDER - 1);
508 slab_max_order_set = true;
509
510 return 1;
511 }
512 __setup("slab_max_order=", slab_max_order_setup);
513
514 #ifdef CONFIG_NUMA
515 /*
516 * Special reaping functions for NUMA systems called from cache_reap().
517 * These take care of doing round robin flushing of alien caches (containing
518 * objects freed on different nodes from which they were allocated) and the
519 * flushing of remote pcps by calling drain_node_pages.
520 */
521 static DEFINE_PER_CPU(unsigned long, slab_reap_node);
522
523 static void init_reap_node(int cpu)
524 {
525 int node;
526
527 node = next_node(cpu_to_mem(cpu), node_online_map);
528 if (node == MAX_NUMNODES)
529 node = first_node(node_online_map);
530
531 per_cpu(slab_reap_node, cpu) = node;
532 }
533
534 static void next_reap_node(void)
535 {
536 int node = __this_cpu_read(slab_reap_node);
537
538 node = next_node(node, node_online_map);
539 if (unlikely(node >= MAX_NUMNODES))
540 node = first_node(node_online_map);
541 __this_cpu_write(slab_reap_node, node);
542 }
543
544 #else
545 #define init_reap_node(cpu) do { } while (0)
546 #define next_reap_node(void) do { } while (0)
547 #endif
548
549 /*
550 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
551 * via the workqueue/eventd.
552 * Add the CPU number into the expiration time to minimize the possibility of
553 * the CPUs getting into lockstep and contending for the global cache chain
554 * lock.
555 */
556 static void start_cpu_timer(int cpu)
557 {
558 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
559
560 /*
561 * When this gets called from do_initcalls via cpucache_init(),
562 * init_workqueues() has already run, so keventd will be setup
563 * at that time.
564 */
565 if (keventd_up() && reap_work->work.func == NULL) {
566 init_reap_node(cpu);
567 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
568 schedule_delayed_work_on(cpu, reap_work,
569 __round_jiffies_relative(HZ, cpu));
570 }
571 }
572
573 static void init_arraycache(struct array_cache *ac, int limit, int batch)
574 {
575 /*
576 * The array_cache structures contain pointers to free object.
577 * However, when such objects are allocated or transferred to another
578 * cache the pointers are not cleared and they could be counted as
579 * valid references during a kmemleak scan. Therefore, kmemleak must
580 * not scan such objects.
581 */
582 kmemleak_no_scan(ac);
583 if (ac) {
584 ac->avail = 0;
585 ac->limit = limit;
586 ac->batchcount = batch;
587 ac->touched = 0;
588 }
589 }
590
591 static struct array_cache *alloc_arraycache(int node, int entries,
592 int batchcount, gfp_t gfp)
593 {
594 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
595 struct array_cache *ac = NULL;
596
597 ac = kmalloc_node(memsize, gfp, node);
598 init_arraycache(ac, entries, batchcount);
599 return ac;
600 }
601
602 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
603 struct page *page, void *objp)
604 {
605 struct kmem_cache_node *n;
606 int page_node;
607 LIST_HEAD(list);
608
609 page_node = page_to_nid(page);
610 n = get_node(cachep, page_node);
611
612 spin_lock(&n->list_lock);
613 free_block(cachep, &objp, 1, page_node, &list);
614 spin_unlock(&n->list_lock);
615
616 slabs_destroy(cachep, &list);
617 }
618
619 /*
620 * Transfer objects in one arraycache to another.
621 * Locking must be handled by the caller.
622 *
623 * Return the number of entries transferred.
624 */
625 static int transfer_objects(struct array_cache *to,
626 struct array_cache *from, unsigned int max)
627 {
628 /* Figure out how many entries to transfer */
629 int nr = min3(from->avail, max, to->limit - to->avail);
630
631 if (!nr)
632 return 0;
633
634 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
635 sizeof(void *) *nr);
636
637 from->avail -= nr;
638 to->avail += nr;
639 return nr;
640 }
641
642 #ifndef CONFIG_NUMA
643
644 #define drain_alien_cache(cachep, alien) do { } while (0)
645 #define reap_alien(cachep, n) do { } while (0)
646
647 static inline struct alien_cache **alloc_alien_cache(int node,
648 int limit, gfp_t gfp)
649 {
650 return NULL;
651 }
652
653 static inline void free_alien_cache(struct alien_cache **ac_ptr)
654 {
655 }
656
657 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
658 {
659 return 0;
660 }
661
662 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
663 gfp_t flags)
664 {
665 return NULL;
666 }
667
668 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
669 gfp_t flags, int nodeid)
670 {
671 return NULL;
672 }
673
674 static inline gfp_t gfp_exact_node(gfp_t flags)
675 {
676 return flags & ~__GFP_NOFAIL;
677 }
678
679 #else /* CONFIG_NUMA */
680
681 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
682 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
683
684 static struct alien_cache *__alloc_alien_cache(int node, int entries,
685 int batch, gfp_t gfp)
686 {
687 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
688 struct alien_cache *alc = NULL;
689
690 alc = kmalloc_node(memsize, gfp, node);
691 init_arraycache(&alc->ac, entries, batch);
692 spin_lock_init(&alc->lock);
693 return alc;
694 }
695
696 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
697 {
698 struct alien_cache **alc_ptr;
699 size_t memsize = sizeof(void *) * nr_node_ids;
700 int i;
701
702 if (limit > 1)
703 limit = 12;
704 alc_ptr = kzalloc_node(memsize, gfp, node);
705 if (!alc_ptr)
706 return NULL;
707
708 for_each_node(i) {
709 if (i == node || !node_online(i))
710 continue;
711 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
712 if (!alc_ptr[i]) {
713 for (i--; i >= 0; i--)
714 kfree(alc_ptr[i]);
715 kfree(alc_ptr);
716 return NULL;
717 }
718 }
719 return alc_ptr;
720 }
721
722 static void free_alien_cache(struct alien_cache **alc_ptr)
723 {
724 int i;
725
726 if (!alc_ptr)
727 return;
728 for_each_node(i)
729 kfree(alc_ptr[i]);
730 kfree(alc_ptr);
731 }
732
733 static void __drain_alien_cache(struct kmem_cache *cachep,
734 struct array_cache *ac, int node,
735 struct list_head *list)
736 {
737 struct kmem_cache_node *n = get_node(cachep, node);
738
739 if (ac->avail) {
740 spin_lock(&n->list_lock);
741 /*
742 * Stuff objects into the remote nodes shared array first.
743 * That way we could avoid the overhead of putting the objects
744 * into the free lists and getting them back later.
745 */
746 if (n->shared)
747 transfer_objects(n->shared, ac, ac->limit);
748
749 free_block(cachep, ac->entry, ac->avail, node, list);
750 ac->avail = 0;
751 spin_unlock(&n->list_lock);
752 }
753 }
754
755 /*
756 * Called from cache_reap() to regularly drain alien caches round robin.
757 */
758 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
759 {
760 int node = __this_cpu_read(slab_reap_node);
761
762 if (n->alien) {
763 struct alien_cache *alc = n->alien[node];
764 struct array_cache *ac;
765
766 if (alc) {
767 ac = &alc->ac;
768 if (ac->avail && spin_trylock_irq(&alc->lock)) {
769 LIST_HEAD(list);
770
771 __drain_alien_cache(cachep, ac, node, &list);
772 spin_unlock_irq(&alc->lock);
773 slabs_destroy(cachep, &list);
774 }
775 }
776 }
777 }
778
779 static void drain_alien_cache(struct kmem_cache *cachep,
780 struct alien_cache **alien)
781 {
782 int i = 0;
783 struct alien_cache *alc;
784 struct array_cache *ac;
785 unsigned long flags;
786
787 for_each_online_node(i) {
788 alc = alien[i];
789 if (alc) {
790 LIST_HEAD(list);
791
792 ac = &alc->ac;
793 spin_lock_irqsave(&alc->lock, flags);
794 __drain_alien_cache(cachep, ac, i, &list);
795 spin_unlock_irqrestore(&alc->lock, flags);
796 slabs_destroy(cachep, &list);
797 }
798 }
799 }
800
801 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
802 int node, int page_node)
803 {
804 struct kmem_cache_node *n;
805 struct alien_cache *alien = NULL;
806 struct array_cache *ac;
807 LIST_HEAD(list);
808
809 n = get_node(cachep, node);
810 STATS_INC_NODEFREES(cachep);
811 if (n->alien && n->alien[page_node]) {
812 alien = n->alien[page_node];
813 ac = &alien->ac;
814 spin_lock(&alien->lock);
815 if (unlikely(ac->avail == ac->limit)) {
816 STATS_INC_ACOVERFLOW(cachep);
817 __drain_alien_cache(cachep, ac, page_node, &list);
818 }
819 ac->entry[ac->avail++] = objp;
820 spin_unlock(&alien->lock);
821 slabs_destroy(cachep, &list);
822 } else {
823 n = get_node(cachep, page_node);
824 spin_lock(&n->list_lock);
825 free_block(cachep, &objp, 1, page_node, &list);
826 spin_unlock(&n->list_lock);
827 slabs_destroy(cachep, &list);
828 }
829 return 1;
830 }
831
832 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
833 {
834 int page_node = page_to_nid(virt_to_page(objp));
835 int node = numa_mem_id();
836 /*
837 * Make sure we are not freeing a object from another node to the array
838 * cache on this cpu.
839 */
840 if (likely(node == page_node))
841 return 0;
842
843 return __cache_free_alien(cachep, objp, node, page_node);
844 }
845
846 /*
847 * Construct gfp mask to allocate from a specific node but do not reclaim or
848 * warn about failures.
849 */
850 static inline gfp_t gfp_exact_node(gfp_t flags)
851 {
852 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
853 }
854 #endif
855
856 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
857 {
858 struct kmem_cache_node *n;
859
860 /*
861 * Set up the kmem_cache_node for cpu before we can
862 * begin anything. Make sure some other cpu on this
863 * node has not already allocated this
864 */
865 n = get_node(cachep, node);
866 if (n) {
867 spin_lock_irq(&n->list_lock);
868 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
869 cachep->num;
870 spin_unlock_irq(&n->list_lock);
871
872 return 0;
873 }
874
875 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
876 if (!n)
877 return -ENOMEM;
878
879 kmem_cache_node_init(n);
880 n->next_reap = jiffies + REAPTIMEOUT_NODE +
881 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
882
883 n->free_limit =
884 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
885
886 /*
887 * The kmem_cache_nodes don't come and go as CPUs
888 * come and go. slab_mutex is sufficient
889 * protection here.
890 */
891 cachep->node[node] = n;
892
893 return 0;
894 }
895
896 /*
897 * Allocates and initializes node for a node on each slab cache, used for
898 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
899 * will be allocated off-node since memory is not yet online for the new node.
900 * When hotplugging memory or a cpu, existing node are not replaced if
901 * already in use.
902 *
903 * Must hold slab_mutex.
904 */
905 static int init_cache_node_node(int node)
906 {
907 int ret;
908 struct kmem_cache *cachep;
909
910 list_for_each_entry(cachep, &slab_caches, list) {
911 ret = init_cache_node(cachep, node, GFP_KERNEL);
912 if (ret)
913 return ret;
914 }
915
916 return 0;
917 }
918
919 static int setup_kmem_cache_node(struct kmem_cache *cachep,
920 int node, gfp_t gfp, bool force_change)
921 {
922 int ret = -ENOMEM;
923 struct kmem_cache_node *n;
924 struct array_cache *old_shared = NULL;
925 struct array_cache *new_shared = NULL;
926 struct alien_cache **new_alien = NULL;
927 LIST_HEAD(list);
928
929 if (use_alien_caches) {
930 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
931 if (!new_alien)
932 goto fail;
933 }
934
935 if (cachep->shared) {
936 new_shared = alloc_arraycache(node,
937 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
938 if (!new_shared)
939 goto fail;
940 }
941
942 ret = init_cache_node(cachep, node, gfp);
943 if (ret)
944 goto fail;
945
946 n = get_node(cachep, node);
947 spin_lock_irq(&n->list_lock);
948 if (n->shared && force_change) {
949 free_block(cachep, n->shared->entry,
950 n->shared->avail, node, &list);
951 n->shared->avail = 0;
952 }
953
954 if (!n->shared || force_change) {
955 old_shared = n->shared;
956 n->shared = new_shared;
957 new_shared = NULL;
958 }
959
960 if (!n->alien) {
961 n->alien = new_alien;
962 new_alien = NULL;
963 }
964
965 spin_unlock_irq(&n->list_lock);
966 slabs_destroy(cachep, &list);
967
968 /*
969 * To protect lockless access to n->shared during irq disabled context.
970 * If n->shared isn't NULL in irq disabled context, accessing to it is
971 * guaranteed to be valid until irq is re-enabled, because it will be
972 * freed after synchronize_sched().
973 */
974 if (force_change)
975 synchronize_sched();
976
977 fail:
978 kfree(old_shared);
979 kfree(new_shared);
980 free_alien_cache(new_alien);
981
982 return ret;
983 }
984
985 static void cpuup_canceled(long cpu)
986 {
987 struct kmem_cache *cachep;
988 struct kmem_cache_node *n = NULL;
989 int node = cpu_to_mem(cpu);
990 const struct cpumask *mask = cpumask_of_node(node);
991
992 list_for_each_entry(cachep, &slab_caches, list) {
993 struct array_cache *nc;
994 struct array_cache *shared;
995 struct alien_cache **alien;
996 LIST_HEAD(list);
997
998 n = get_node(cachep, node);
999 if (!n)
1000 continue;
1001
1002 spin_lock_irq(&n->list_lock);
1003
1004 /* Free limit for this kmem_cache_node */
1005 n->free_limit -= cachep->batchcount;
1006
1007 /* cpu is dead; no one can alloc from it. */
1008 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1009 if (nc) {
1010 free_block(cachep, nc->entry, nc->avail, node, &list);
1011 nc->avail = 0;
1012 }
1013
1014 if (!cpumask_empty(mask)) {
1015 spin_unlock_irq(&n->list_lock);
1016 goto free_slab;
1017 }
1018
1019 shared = n->shared;
1020 if (shared) {
1021 free_block(cachep, shared->entry,
1022 shared->avail, node, &list);
1023 n->shared = NULL;
1024 }
1025
1026 alien = n->alien;
1027 n->alien = NULL;
1028
1029 spin_unlock_irq(&n->list_lock);
1030
1031 kfree(shared);
1032 if (alien) {
1033 drain_alien_cache(cachep, alien);
1034 free_alien_cache(alien);
1035 }
1036
1037 free_slab:
1038 slabs_destroy(cachep, &list);
1039 }
1040 /*
1041 * In the previous loop, all the objects were freed to
1042 * the respective cache's slabs, now we can go ahead and
1043 * shrink each nodelist to its limit.
1044 */
1045 list_for_each_entry(cachep, &slab_caches, list) {
1046 n = get_node(cachep, node);
1047 if (!n)
1048 continue;
1049 drain_freelist(cachep, n, INT_MAX);
1050 }
1051 }
1052
1053 static int cpuup_prepare(long cpu)
1054 {
1055 struct kmem_cache *cachep;
1056 int node = cpu_to_mem(cpu);
1057 int err;
1058
1059 /*
1060 * We need to do this right in the beginning since
1061 * alloc_arraycache's are going to use this list.
1062 * kmalloc_node allows us to add the slab to the right
1063 * kmem_cache_node and not this cpu's kmem_cache_node
1064 */
1065 err = init_cache_node_node(node);
1066 if (err < 0)
1067 goto bad;
1068
1069 /*
1070 * Now we can go ahead with allocating the shared arrays and
1071 * array caches
1072 */
1073 list_for_each_entry(cachep, &slab_caches, list) {
1074 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1075 if (err)
1076 goto bad;
1077 }
1078
1079 return 0;
1080 bad:
1081 cpuup_canceled(cpu);
1082 return -ENOMEM;
1083 }
1084
1085 static int cpuup_callback(struct notifier_block *nfb,
1086 unsigned long action, void *hcpu)
1087 {
1088 long cpu = (long)hcpu;
1089 int err = 0;
1090
1091 switch (action) {
1092 case CPU_UP_PREPARE:
1093 case CPU_UP_PREPARE_FROZEN:
1094 mutex_lock(&slab_mutex);
1095 err = cpuup_prepare(cpu);
1096 mutex_unlock(&slab_mutex);
1097 break;
1098 case CPU_ONLINE:
1099 case CPU_ONLINE_FROZEN:
1100 start_cpu_timer(cpu);
1101 break;
1102 #ifdef CONFIG_HOTPLUG_CPU
1103 case CPU_DOWN_PREPARE:
1104 case CPU_DOWN_PREPARE_FROZEN:
1105 /*
1106 * Shutdown cache reaper. Note that the slab_mutex is
1107 * held so that if cache_reap() is invoked it cannot do
1108 * anything expensive but will only modify reap_work
1109 * and reschedule the timer.
1110 */
1111 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1112 /* Now the cache_reaper is guaranteed to be not running. */
1113 per_cpu(slab_reap_work, cpu).work.func = NULL;
1114 break;
1115 case CPU_DOWN_FAILED:
1116 case CPU_DOWN_FAILED_FROZEN:
1117 start_cpu_timer(cpu);
1118 break;
1119 case CPU_DEAD:
1120 case CPU_DEAD_FROZEN:
1121 /*
1122 * Even if all the cpus of a node are down, we don't free the
1123 * kmem_cache_node of any cache. This to avoid a race between
1124 * cpu_down, and a kmalloc allocation from another cpu for
1125 * memory from the node of the cpu going down. The node
1126 * structure is usually allocated from kmem_cache_create() and
1127 * gets destroyed at kmem_cache_destroy().
1128 */
1129 /* fall through */
1130 #endif
1131 case CPU_UP_CANCELED:
1132 case CPU_UP_CANCELED_FROZEN:
1133 mutex_lock(&slab_mutex);
1134 cpuup_canceled(cpu);
1135 mutex_unlock(&slab_mutex);
1136 break;
1137 }
1138 return notifier_from_errno(err);
1139 }
1140
1141 static struct notifier_block cpucache_notifier = {
1142 &cpuup_callback, NULL, 0
1143 };
1144
1145 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1146 /*
1147 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1148 * Returns -EBUSY if all objects cannot be drained so that the node is not
1149 * removed.
1150 *
1151 * Must hold slab_mutex.
1152 */
1153 static int __meminit drain_cache_node_node(int node)
1154 {
1155 struct kmem_cache *cachep;
1156 int ret = 0;
1157
1158 list_for_each_entry(cachep, &slab_caches, list) {
1159 struct kmem_cache_node *n;
1160
1161 n = get_node(cachep, node);
1162 if (!n)
1163 continue;
1164
1165 drain_freelist(cachep, n, INT_MAX);
1166
1167 if (!list_empty(&n->slabs_full) ||
1168 !list_empty(&n->slabs_partial)) {
1169 ret = -EBUSY;
1170 break;
1171 }
1172 }
1173 return ret;
1174 }
1175
1176 static int __meminit slab_memory_callback(struct notifier_block *self,
1177 unsigned long action, void *arg)
1178 {
1179 struct memory_notify *mnb = arg;
1180 int ret = 0;
1181 int nid;
1182
1183 nid = mnb->status_change_nid;
1184 if (nid < 0)
1185 goto out;
1186
1187 switch (action) {
1188 case MEM_GOING_ONLINE:
1189 mutex_lock(&slab_mutex);
1190 ret = init_cache_node_node(nid);
1191 mutex_unlock(&slab_mutex);
1192 break;
1193 case MEM_GOING_OFFLINE:
1194 mutex_lock(&slab_mutex);
1195 ret = drain_cache_node_node(nid);
1196 mutex_unlock(&slab_mutex);
1197 break;
1198 case MEM_ONLINE:
1199 case MEM_OFFLINE:
1200 case MEM_CANCEL_ONLINE:
1201 case MEM_CANCEL_OFFLINE:
1202 break;
1203 }
1204 out:
1205 return notifier_from_errno(ret);
1206 }
1207 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1208
1209 /*
1210 * swap the static kmem_cache_node with kmalloced memory
1211 */
1212 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1213 int nodeid)
1214 {
1215 struct kmem_cache_node *ptr;
1216
1217 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1218 BUG_ON(!ptr);
1219
1220 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1221 /*
1222 * Do not assume that spinlocks can be initialized via memcpy:
1223 */
1224 spin_lock_init(&ptr->list_lock);
1225
1226 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1227 cachep->node[nodeid] = ptr;
1228 }
1229
1230 /*
1231 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1232 * size of kmem_cache_node.
1233 */
1234 static void __init set_up_node(struct kmem_cache *cachep, int index)
1235 {
1236 int node;
1237
1238 for_each_online_node(node) {
1239 cachep->node[node] = &init_kmem_cache_node[index + node];
1240 cachep->node[node]->next_reap = jiffies +
1241 REAPTIMEOUT_NODE +
1242 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1243 }
1244 }
1245
1246 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1247 static void freelist_randomize(struct rnd_state *state, freelist_idx_t *list,
1248 size_t count)
1249 {
1250 size_t i;
1251 unsigned int rand;
1252
1253 for (i = 0; i < count; i++)
1254 list[i] = i;
1255
1256 /* Fisher-Yates shuffle */
1257 for (i = count - 1; i > 0; i--) {
1258 rand = prandom_u32_state(state);
1259 rand %= (i + 1);
1260 swap(list[i], list[rand]);
1261 }
1262 }
1263
1264 /* Create a random sequence per cache */
1265 static int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1266 {
1267 unsigned int seed, count = cachep->num;
1268 struct rnd_state state;
1269
1270 if (count < 2)
1271 return 0;
1272
1273 /* If it fails, we will just use the global lists */
1274 cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), gfp);
1275 if (!cachep->random_seq)
1276 return -ENOMEM;
1277
1278 /* Get best entropy at this stage */
1279 get_random_bytes_arch(&seed, sizeof(seed));
1280 prandom_seed_state(&state, seed);
1281
1282 freelist_randomize(&state, cachep->random_seq, count);
1283 return 0;
1284 }
1285
1286 /* Destroy the per-cache random freelist sequence */
1287 static void cache_random_seq_destroy(struct kmem_cache *cachep)
1288 {
1289 kfree(cachep->random_seq);
1290 cachep->random_seq = NULL;
1291 }
1292 #else
1293 static inline int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1294 {
1295 return 0;
1296 }
1297 static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }
1298 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1299
1300
1301 /*
1302 * Initialisation. Called after the page allocator have been initialised and
1303 * before smp_init().
1304 */
1305 void __init kmem_cache_init(void)
1306 {
1307 int i;
1308
1309 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1310 sizeof(struct rcu_head));
1311 kmem_cache = &kmem_cache_boot;
1312
1313 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1314 use_alien_caches = 0;
1315
1316 for (i = 0; i < NUM_INIT_LISTS; i++)
1317 kmem_cache_node_init(&init_kmem_cache_node[i]);
1318
1319 /*
1320 * Fragmentation resistance on low memory - only use bigger
1321 * page orders on machines with more than 32MB of memory if
1322 * not overridden on the command line.
1323 */
1324 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1325 slab_max_order = SLAB_MAX_ORDER_HI;
1326
1327 /* Bootstrap is tricky, because several objects are allocated
1328 * from caches that do not exist yet:
1329 * 1) initialize the kmem_cache cache: it contains the struct
1330 * kmem_cache structures of all caches, except kmem_cache itself:
1331 * kmem_cache is statically allocated.
1332 * Initially an __init data area is used for the head array and the
1333 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1334 * array at the end of the bootstrap.
1335 * 2) Create the first kmalloc cache.
1336 * The struct kmem_cache for the new cache is allocated normally.
1337 * An __init data area is used for the head array.
1338 * 3) Create the remaining kmalloc caches, with minimally sized
1339 * head arrays.
1340 * 4) Replace the __init data head arrays for kmem_cache and the first
1341 * kmalloc cache with kmalloc allocated arrays.
1342 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1343 * the other cache's with kmalloc allocated memory.
1344 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1345 */
1346
1347 /* 1) create the kmem_cache */
1348
1349 /*
1350 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1351 */
1352 create_boot_cache(kmem_cache, "kmem_cache",
1353 offsetof(struct kmem_cache, node) +
1354 nr_node_ids * sizeof(struct kmem_cache_node *),
1355 SLAB_HWCACHE_ALIGN);
1356 list_add(&kmem_cache->list, &slab_caches);
1357 slab_state = PARTIAL;
1358
1359 /*
1360 * Initialize the caches that provide memory for the kmem_cache_node
1361 * structures first. Without this, further allocations will bug.
1362 */
1363 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1364 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1365 slab_state = PARTIAL_NODE;
1366 setup_kmalloc_cache_index_table();
1367
1368 slab_early_init = 0;
1369
1370 /* 5) Replace the bootstrap kmem_cache_node */
1371 {
1372 int nid;
1373
1374 for_each_online_node(nid) {
1375 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1376
1377 init_list(kmalloc_caches[INDEX_NODE],
1378 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1379 }
1380 }
1381
1382 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1383 }
1384
1385 void __init kmem_cache_init_late(void)
1386 {
1387 struct kmem_cache *cachep;
1388
1389 slab_state = UP;
1390
1391 /* 6) resize the head arrays to their final sizes */
1392 mutex_lock(&slab_mutex);
1393 list_for_each_entry(cachep, &slab_caches, list)
1394 if (enable_cpucache(cachep, GFP_NOWAIT))
1395 BUG();
1396 mutex_unlock(&slab_mutex);
1397
1398 /* Done! */
1399 slab_state = FULL;
1400
1401 /*
1402 * Register a cpu startup notifier callback that initializes
1403 * cpu_cache_get for all new cpus
1404 */
1405 register_cpu_notifier(&cpucache_notifier);
1406
1407 #ifdef CONFIG_NUMA
1408 /*
1409 * Register a memory hotplug callback that initializes and frees
1410 * node.
1411 */
1412 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1413 #endif
1414
1415 /*
1416 * The reap timers are started later, with a module init call: That part
1417 * of the kernel is not yet operational.
1418 */
1419 }
1420
1421 static int __init cpucache_init(void)
1422 {
1423 int cpu;
1424
1425 /*
1426 * Register the timers that return unneeded pages to the page allocator
1427 */
1428 for_each_online_cpu(cpu)
1429 start_cpu_timer(cpu);
1430
1431 /* Done! */
1432 slab_state = FULL;
1433 return 0;
1434 }
1435 __initcall(cpucache_init);
1436
1437 static noinline void
1438 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1439 {
1440 #if DEBUG
1441 struct kmem_cache_node *n;
1442 struct page *page;
1443 unsigned long flags;
1444 int node;
1445 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1446 DEFAULT_RATELIMIT_BURST);
1447
1448 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1449 return;
1450
1451 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1452 nodeid, gfpflags, &gfpflags);
1453 pr_warn(" cache: %s, object size: %d, order: %d\n",
1454 cachep->name, cachep->size, cachep->gfporder);
1455
1456 for_each_kmem_cache_node(cachep, node, n) {
1457 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1458 unsigned long active_slabs = 0, num_slabs = 0;
1459
1460 spin_lock_irqsave(&n->list_lock, flags);
1461 list_for_each_entry(page, &n->slabs_full, lru) {
1462 active_objs += cachep->num;
1463 active_slabs++;
1464 }
1465 list_for_each_entry(page, &n->slabs_partial, lru) {
1466 active_objs += page->active;
1467 active_slabs++;
1468 }
1469 list_for_each_entry(page, &n->slabs_free, lru)
1470 num_slabs++;
1471
1472 free_objects += n->free_objects;
1473 spin_unlock_irqrestore(&n->list_lock, flags);
1474
1475 num_slabs += active_slabs;
1476 num_objs = num_slabs * cachep->num;
1477 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1478 node, active_slabs, num_slabs, active_objs, num_objs,
1479 free_objects);
1480 }
1481 #endif
1482 }
1483
1484 /*
1485 * Interface to system's page allocator. No need to hold the
1486 * kmem_cache_node ->list_lock.
1487 *
1488 * If we requested dmaable memory, we will get it. Even if we
1489 * did not request dmaable memory, we might get it, but that
1490 * would be relatively rare and ignorable.
1491 */
1492 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1493 int nodeid)
1494 {
1495 struct page *page;
1496 int nr_pages;
1497
1498 flags |= cachep->allocflags;
1499 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1500 flags |= __GFP_RECLAIMABLE;
1501
1502 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1503 if (!page) {
1504 slab_out_of_memory(cachep, flags, nodeid);
1505 return NULL;
1506 }
1507
1508 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1509 __free_pages(page, cachep->gfporder);
1510 return NULL;
1511 }
1512
1513 nr_pages = (1 << cachep->gfporder);
1514 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1515 add_zone_page_state(page_zone(page),
1516 NR_SLAB_RECLAIMABLE, nr_pages);
1517 else
1518 add_zone_page_state(page_zone(page),
1519 NR_SLAB_UNRECLAIMABLE, nr_pages);
1520
1521 __SetPageSlab(page);
1522 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1523 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1524 SetPageSlabPfmemalloc(page);
1525
1526 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1527 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1528
1529 if (cachep->ctor)
1530 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1531 else
1532 kmemcheck_mark_unallocated_pages(page, nr_pages);
1533 }
1534
1535 return page;
1536 }
1537
1538 /*
1539 * Interface to system's page release.
1540 */
1541 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1542 {
1543 int order = cachep->gfporder;
1544 unsigned long nr_freed = (1 << order);
1545
1546 kmemcheck_free_shadow(page, order);
1547
1548 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1549 sub_zone_page_state(page_zone(page),
1550 NR_SLAB_RECLAIMABLE, nr_freed);
1551 else
1552 sub_zone_page_state(page_zone(page),
1553 NR_SLAB_UNRECLAIMABLE, nr_freed);
1554
1555 BUG_ON(!PageSlab(page));
1556 __ClearPageSlabPfmemalloc(page);
1557 __ClearPageSlab(page);
1558 page_mapcount_reset(page);
1559 page->mapping = NULL;
1560
1561 if (current->reclaim_state)
1562 current->reclaim_state->reclaimed_slab += nr_freed;
1563 memcg_uncharge_slab(page, order, cachep);
1564 __free_pages(page, order);
1565 }
1566
1567 static void kmem_rcu_free(struct rcu_head *head)
1568 {
1569 struct kmem_cache *cachep;
1570 struct page *page;
1571
1572 page = container_of(head, struct page, rcu_head);
1573 cachep = page->slab_cache;
1574
1575 kmem_freepages(cachep, page);
1576 }
1577
1578 #if DEBUG
1579 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1580 {
1581 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1582 (cachep->size % PAGE_SIZE) == 0)
1583 return true;
1584
1585 return false;
1586 }
1587
1588 #ifdef CONFIG_DEBUG_PAGEALLOC
1589 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1590 unsigned long caller)
1591 {
1592 int size = cachep->object_size;
1593
1594 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1595
1596 if (size < 5 * sizeof(unsigned long))
1597 return;
1598
1599 *addr++ = 0x12345678;
1600 *addr++ = caller;
1601 *addr++ = smp_processor_id();
1602 size -= 3 * sizeof(unsigned long);
1603 {
1604 unsigned long *sptr = &caller;
1605 unsigned long svalue;
1606
1607 while (!kstack_end(sptr)) {
1608 svalue = *sptr++;
1609 if (kernel_text_address(svalue)) {
1610 *addr++ = svalue;
1611 size -= sizeof(unsigned long);
1612 if (size <= sizeof(unsigned long))
1613 break;
1614 }
1615 }
1616
1617 }
1618 *addr++ = 0x87654321;
1619 }
1620
1621 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1622 int map, unsigned long caller)
1623 {
1624 if (!is_debug_pagealloc_cache(cachep))
1625 return;
1626
1627 if (caller)
1628 store_stackinfo(cachep, objp, caller);
1629
1630 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1631 }
1632
1633 #else
1634 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1635 int map, unsigned long caller) {}
1636
1637 #endif
1638
1639 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1640 {
1641 int size = cachep->object_size;
1642 addr = &((char *)addr)[obj_offset(cachep)];
1643
1644 memset(addr, val, size);
1645 *(unsigned char *)(addr + size - 1) = POISON_END;
1646 }
1647
1648 static void dump_line(char *data, int offset, int limit)
1649 {
1650 int i;
1651 unsigned char error = 0;
1652 int bad_count = 0;
1653
1654 pr_err("%03x: ", offset);
1655 for (i = 0; i < limit; i++) {
1656 if (data[offset + i] != POISON_FREE) {
1657 error = data[offset + i];
1658 bad_count++;
1659 }
1660 }
1661 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1662 &data[offset], limit, 1);
1663
1664 if (bad_count == 1) {
1665 error ^= POISON_FREE;
1666 if (!(error & (error - 1))) {
1667 pr_err("Single bit error detected. Probably bad RAM.\n");
1668 #ifdef CONFIG_X86
1669 pr_err("Run memtest86+ or a similar memory test tool.\n");
1670 #else
1671 pr_err("Run a memory test tool.\n");
1672 #endif
1673 }
1674 }
1675 }
1676 #endif
1677
1678 #if DEBUG
1679
1680 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1681 {
1682 int i, size;
1683 char *realobj;
1684
1685 if (cachep->flags & SLAB_RED_ZONE) {
1686 pr_err("Redzone: 0x%llx/0x%llx\n",
1687 *dbg_redzone1(cachep, objp),
1688 *dbg_redzone2(cachep, objp));
1689 }
1690
1691 if (cachep->flags & SLAB_STORE_USER) {
1692 pr_err("Last user: [<%p>](%pSR)\n",
1693 *dbg_userword(cachep, objp),
1694 *dbg_userword(cachep, objp));
1695 }
1696 realobj = (char *)objp + obj_offset(cachep);
1697 size = cachep->object_size;
1698 for (i = 0; i < size && lines; i += 16, lines--) {
1699 int limit;
1700 limit = 16;
1701 if (i + limit > size)
1702 limit = size - i;
1703 dump_line(realobj, i, limit);
1704 }
1705 }
1706
1707 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1708 {
1709 char *realobj;
1710 int size, i;
1711 int lines = 0;
1712
1713 if (is_debug_pagealloc_cache(cachep))
1714 return;
1715
1716 realobj = (char *)objp + obj_offset(cachep);
1717 size = cachep->object_size;
1718
1719 for (i = 0; i < size; i++) {
1720 char exp = POISON_FREE;
1721 if (i == size - 1)
1722 exp = POISON_END;
1723 if (realobj[i] != exp) {
1724 int limit;
1725 /* Mismatch ! */
1726 /* Print header */
1727 if (lines == 0) {
1728 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1729 print_tainted(), cachep->name,
1730 realobj, size);
1731 print_objinfo(cachep, objp, 0);
1732 }
1733 /* Hexdump the affected line */
1734 i = (i / 16) * 16;
1735 limit = 16;
1736 if (i + limit > size)
1737 limit = size - i;
1738 dump_line(realobj, i, limit);
1739 i += 16;
1740 lines++;
1741 /* Limit to 5 lines */
1742 if (lines > 5)
1743 break;
1744 }
1745 }
1746 if (lines != 0) {
1747 /* Print some data about the neighboring objects, if they
1748 * exist:
1749 */
1750 struct page *page = virt_to_head_page(objp);
1751 unsigned int objnr;
1752
1753 objnr = obj_to_index(cachep, page, objp);
1754 if (objnr) {
1755 objp = index_to_obj(cachep, page, objnr - 1);
1756 realobj = (char *)objp + obj_offset(cachep);
1757 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1758 print_objinfo(cachep, objp, 2);
1759 }
1760 if (objnr + 1 < cachep->num) {
1761 objp = index_to_obj(cachep, page, objnr + 1);
1762 realobj = (char *)objp + obj_offset(cachep);
1763 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1764 print_objinfo(cachep, objp, 2);
1765 }
1766 }
1767 }
1768 #endif
1769
1770 #if DEBUG
1771 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1772 struct page *page)
1773 {
1774 int i;
1775
1776 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1777 poison_obj(cachep, page->freelist - obj_offset(cachep),
1778 POISON_FREE);
1779 }
1780
1781 for (i = 0; i < cachep->num; i++) {
1782 void *objp = index_to_obj(cachep, page, i);
1783
1784 if (cachep->flags & SLAB_POISON) {
1785 check_poison_obj(cachep, objp);
1786 slab_kernel_map(cachep, objp, 1, 0);
1787 }
1788 if (cachep->flags & SLAB_RED_ZONE) {
1789 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1790 slab_error(cachep, "start of a freed object was overwritten");
1791 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1792 slab_error(cachep, "end of a freed object was overwritten");
1793 }
1794 }
1795 }
1796 #else
1797 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1798 struct page *page)
1799 {
1800 }
1801 #endif
1802
1803 /**
1804 * slab_destroy - destroy and release all objects in a slab
1805 * @cachep: cache pointer being destroyed
1806 * @page: page pointer being destroyed
1807 *
1808 * Destroy all the objs in a slab page, and release the mem back to the system.
1809 * Before calling the slab page must have been unlinked from the cache. The
1810 * kmem_cache_node ->list_lock is not held/needed.
1811 */
1812 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1813 {
1814 void *freelist;
1815
1816 freelist = page->freelist;
1817 slab_destroy_debugcheck(cachep, page);
1818 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1819 call_rcu(&page->rcu_head, kmem_rcu_free);
1820 else
1821 kmem_freepages(cachep, page);
1822
1823 /*
1824 * From now on, we don't use freelist
1825 * although actual page can be freed in rcu context
1826 */
1827 if (OFF_SLAB(cachep))
1828 kmem_cache_free(cachep->freelist_cache, freelist);
1829 }
1830
1831 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1832 {
1833 struct page *page, *n;
1834
1835 list_for_each_entry_safe(page, n, list, lru) {
1836 list_del(&page->lru);
1837 slab_destroy(cachep, page);
1838 }
1839 }
1840
1841 /**
1842 * calculate_slab_order - calculate size (page order) of slabs
1843 * @cachep: pointer to the cache that is being created
1844 * @size: size of objects to be created in this cache.
1845 * @flags: slab allocation flags
1846 *
1847 * Also calculates the number of objects per slab.
1848 *
1849 * This could be made much more intelligent. For now, try to avoid using
1850 * high order pages for slabs. When the gfp() functions are more friendly
1851 * towards high-order requests, this should be changed.
1852 */
1853 static size_t calculate_slab_order(struct kmem_cache *cachep,
1854 size_t size, unsigned long flags)
1855 {
1856 size_t left_over = 0;
1857 int gfporder;
1858
1859 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1860 unsigned int num;
1861 size_t remainder;
1862
1863 num = cache_estimate(gfporder, size, flags, &remainder);
1864 if (!num)
1865 continue;
1866
1867 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1868 if (num > SLAB_OBJ_MAX_NUM)
1869 break;
1870
1871 if (flags & CFLGS_OFF_SLAB) {
1872 struct kmem_cache *freelist_cache;
1873 size_t freelist_size;
1874
1875 freelist_size = num * sizeof(freelist_idx_t);
1876 freelist_cache = kmalloc_slab(freelist_size, 0u);
1877 if (!freelist_cache)
1878 continue;
1879
1880 /*
1881 * Needed to avoid possible looping condition
1882 * in cache_grow_begin()
1883 */
1884 if (OFF_SLAB(freelist_cache))
1885 continue;
1886
1887 /* check if off slab has enough benefit */
1888 if (freelist_cache->size > cachep->size / 2)
1889 continue;
1890 }
1891
1892 /* Found something acceptable - save it away */
1893 cachep->num = num;
1894 cachep->gfporder = gfporder;
1895 left_over = remainder;
1896
1897 /*
1898 * A VFS-reclaimable slab tends to have most allocations
1899 * as GFP_NOFS and we really don't want to have to be allocating
1900 * higher-order pages when we are unable to shrink dcache.
1901 */
1902 if (flags & SLAB_RECLAIM_ACCOUNT)
1903 break;
1904
1905 /*
1906 * Large number of objects is good, but very large slabs are
1907 * currently bad for the gfp()s.
1908 */
1909 if (gfporder >= slab_max_order)
1910 break;
1911
1912 /*
1913 * Acceptable internal fragmentation?
1914 */
1915 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1916 break;
1917 }
1918 return left_over;
1919 }
1920
1921 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1922 struct kmem_cache *cachep, int entries, int batchcount)
1923 {
1924 int cpu;
1925 size_t size;
1926 struct array_cache __percpu *cpu_cache;
1927
1928 size = sizeof(void *) * entries + sizeof(struct array_cache);
1929 cpu_cache = __alloc_percpu(size, sizeof(void *));
1930
1931 if (!cpu_cache)
1932 return NULL;
1933
1934 for_each_possible_cpu(cpu) {
1935 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1936 entries, batchcount);
1937 }
1938
1939 return cpu_cache;
1940 }
1941
1942 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1943 {
1944 if (slab_state >= FULL)
1945 return enable_cpucache(cachep, gfp);
1946
1947 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1948 if (!cachep->cpu_cache)
1949 return 1;
1950
1951 if (slab_state == DOWN) {
1952 /* Creation of first cache (kmem_cache). */
1953 set_up_node(kmem_cache, CACHE_CACHE);
1954 } else if (slab_state == PARTIAL) {
1955 /* For kmem_cache_node */
1956 set_up_node(cachep, SIZE_NODE);
1957 } else {
1958 int node;
1959
1960 for_each_online_node(node) {
1961 cachep->node[node] = kmalloc_node(
1962 sizeof(struct kmem_cache_node), gfp, node);
1963 BUG_ON(!cachep->node[node]);
1964 kmem_cache_node_init(cachep->node[node]);
1965 }
1966 }
1967
1968 cachep->node[numa_mem_id()]->next_reap =
1969 jiffies + REAPTIMEOUT_NODE +
1970 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1971
1972 cpu_cache_get(cachep)->avail = 0;
1973 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1974 cpu_cache_get(cachep)->batchcount = 1;
1975 cpu_cache_get(cachep)->touched = 0;
1976 cachep->batchcount = 1;
1977 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1978 return 0;
1979 }
1980
1981 unsigned long kmem_cache_flags(unsigned long object_size,
1982 unsigned long flags, const char *name,
1983 void (*ctor)(void *))
1984 {
1985 return flags;
1986 }
1987
1988 struct kmem_cache *
1989 __kmem_cache_alias(const char *name, size_t size, size_t align,
1990 unsigned long flags, void (*ctor)(void *))
1991 {
1992 struct kmem_cache *cachep;
1993
1994 cachep = find_mergeable(size, align, flags, name, ctor);
1995 if (cachep) {
1996 cachep->refcount++;
1997
1998 /*
1999 * Adjust the object sizes so that we clear
2000 * the complete object on kzalloc.
2001 */
2002 cachep->object_size = max_t(int, cachep->object_size, size);
2003 }
2004 return cachep;
2005 }
2006
2007 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
2008 size_t size, unsigned long flags)
2009 {
2010 size_t left;
2011
2012 cachep->num = 0;
2013
2014 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
2015 return false;
2016
2017 left = calculate_slab_order(cachep, size,
2018 flags | CFLGS_OBJFREELIST_SLAB);
2019 if (!cachep->num)
2020 return false;
2021
2022 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
2023 return false;
2024
2025 cachep->colour = left / cachep->colour_off;
2026
2027 return true;
2028 }
2029
2030 static bool set_off_slab_cache(struct kmem_cache *cachep,
2031 size_t size, unsigned long flags)
2032 {
2033 size_t left;
2034
2035 cachep->num = 0;
2036
2037 /*
2038 * Always use on-slab management when SLAB_NOLEAKTRACE
2039 * to avoid recursive calls into kmemleak.
2040 */
2041 if (flags & SLAB_NOLEAKTRACE)
2042 return false;
2043
2044 /*
2045 * Size is large, assume best to place the slab management obj
2046 * off-slab (should allow better packing of objs).
2047 */
2048 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
2049 if (!cachep->num)
2050 return false;
2051
2052 /*
2053 * If the slab has been placed off-slab, and we have enough space then
2054 * move it on-slab. This is at the expense of any extra colouring.
2055 */
2056 if (left >= cachep->num * sizeof(freelist_idx_t))
2057 return false;
2058
2059 cachep->colour = left / cachep->colour_off;
2060
2061 return true;
2062 }
2063
2064 static bool set_on_slab_cache(struct kmem_cache *cachep,
2065 size_t size, unsigned long flags)
2066 {
2067 size_t left;
2068
2069 cachep->num = 0;
2070
2071 left = calculate_slab_order(cachep, size, flags);
2072 if (!cachep->num)
2073 return false;
2074
2075 cachep->colour = left / cachep->colour_off;
2076
2077 return true;
2078 }
2079
2080 /**
2081 * __kmem_cache_create - Create a cache.
2082 * @cachep: cache management descriptor
2083 * @flags: SLAB flags
2084 *
2085 * Returns a ptr to the cache on success, NULL on failure.
2086 * Cannot be called within a int, but can be interrupted.
2087 * The @ctor is run when new pages are allocated by the cache.
2088 *
2089 * The flags are
2090 *
2091 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2092 * to catch references to uninitialised memory.
2093 *
2094 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2095 * for buffer overruns.
2096 *
2097 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2098 * cacheline. This can be beneficial if you're counting cycles as closely
2099 * as davem.
2100 */
2101 int
2102 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2103 {
2104 size_t ralign = BYTES_PER_WORD;
2105 gfp_t gfp;
2106 int err;
2107 size_t size = cachep->size;
2108
2109 #if DEBUG
2110 #if FORCED_DEBUG
2111 /*
2112 * Enable redzoning and last user accounting, except for caches with
2113 * large objects, if the increased size would increase the object size
2114 * above the next power of two: caches with object sizes just above a
2115 * power of two have a significant amount of internal fragmentation.
2116 */
2117 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2118 2 * sizeof(unsigned long long)))
2119 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2120 if (!(flags & SLAB_DESTROY_BY_RCU))
2121 flags |= SLAB_POISON;
2122 #endif
2123 #endif
2124
2125 /*
2126 * Check that size is in terms of words. This is needed to avoid
2127 * unaligned accesses for some archs when redzoning is used, and makes
2128 * sure any on-slab bufctl's are also correctly aligned.
2129 */
2130 if (size & (BYTES_PER_WORD - 1)) {
2131 size += (BYTES_PER_WORD - 1);
2132 size &= ~(BYTES_PER_WORD - 1);
2133 }
2134
2135 if (flags & SLAB_RED_ZONE) {
2136 ralign = REDZONE_ALIGN;
2137 /* If redzoning, ensure that the second redzone is suitably
2138 * aligned, by adjusting the object size accordingly. */
2139 size += REDZONE_ALIGN - 1;
2140 size &= ~(REDZONE_ALIGN - 1);
2141 }
2142
2143 /* 3) caller mandated alignment */
2144 if (ralign < cachep->align) {
2145 ralign = cachep->align;
2146 }
2147 /* disable debug if necessary */
2148 if (ralign > __alignof__(unsigned long long))
2149 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2150 /*
2151 * 4) Store it.
2152 */
2153 cachep->align = ralign;
2154 cachep->colour_off = cache_line_size();
2155 /* Offset must be a multiple of the alignment. */
2156 if (cachep->colour_off < cachep->align)
2157 cachep->colour_off = cachep->align;
2158
2159 if (slab_is_available())
2160 gfp = GFP_KERNEL;
2161 else
2162 gfp = GFP_NOWAIT;
2163
2164 #if DEBUG
2165
2166 /*
2167 * Both debugging options require word-alignment which is calculated
2168 * into align above.
2169 */
2170 if (flags & SLAB_RED_ZONE) {
2171 /* add space for red zone words */
2172 cachep->obj_offset += sizeof(unsigned long long);
2173 size += 2 * sizeof(unsigned long long);
2174 }
2175 if (flags & SLAB_STORE_USER) {
2176 /* user store requires one word storage behind the end of
2177 * the real object. But if the second red zone needs to be
2178 * aligned to 64 bits, we must allow that much space.
2179 */
2180 if (flags & SLAB_RED_ZONE)
2181 size += REDZONE_ALIGN;
2182 else
2183 size += BYTES_PER_WORD;
2184 }
2185 #endif
2186
2187 kasan_cache_create(cachep, &size, &flags);
2188
2189 size = ALIGN(size, cachep->align);
2190 /*
2191 * We should restrict the number of objects in a slab to implement
2192 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2193 */
2194 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2195 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2196
2197 #if DEBUG
2198 /*
2199 * To activate debug pagealloc, off-slab management is necessary
2200 * requirement. In early phase of initialization, small sized slab
2201 * doesn't get initialized so it would not be possible. So, we need
2202 * to check size >= 256. It guarantees that all necessary small
2203 * sized slab is initialized in current slab initialization sequence.
2204 */
2205 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2206 size >= 256 && cachep->object_size > cache_line_size()) {
2207 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2208 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2209
2210 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2211 flags |= CFLGS_OFF_SLAB;
2212 cachep->obj_offset += tmp_size - size;
2213 size = tmp_size;
2214 goto done;
2215 }
2216 }
2217 }
2218 #endif
2219
2220 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2221 flags |= CFLGS_OBJFREELIST_SLAB;
2222 goto done;
2223 }
2224
2225 if (set_off_slab_cache(cachep, size, flags)) {
2226 flags |= CFLGS_OFF_SLAB;
2227 goto done;
2228 }
2229
2230 if (set_on_slab_cache(cachep, size, flags))
2231 goto done;
2232
2233 return -E2BIG;
2234
2235 done:
2236 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2237 cachep->flags = flags;
2238 cachep->allocflags = __GFP_COMP;
2239 if (flags & SLAB_CACHE_DMA)
2240 cachep->allocflags |= GFP_DMA;
2241 cachep->size = size;
2242 cachep->reciprocal_buffer_size = reciprocal_value(size);
2243
2244 #if DEBUG
2245 /*
2246 * If we're going to use the generic kernel_map_pages()
2247 * poisoning, then it's going to smash the contents of
2248 * the redzone and userword anyhow, so switch them off.
2249 */
2250 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2251 (cachep->flags & SLAB_POISON) &&
2252 is_debug_pagealloc_cache(cachep))
2253 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2254 #endif
2255
2256 if (OFF_SLAB(cachep)) {
2257 cachep->freelist_cache =
2258 kmalloc_slab(cachep->freelist_size, 0u);
2259 }
2260
2261 err = setup_cpu_cache(cachep, gfp);
2262 if (err) {
2263 __kmem_cache_release(cachep);
2264 return err;
2265 }
2266
2267 return 0;
2268 }
2269
2270 #if DEBUG
2271 static void check_irq_off(void)
2272 {
2273 BUG_ON(!irqs_disabled());
2274 }
2275
2276 static void check_irq_on(void)
2277 {
2278 BUG_ON(irqs_disabled());
2279 }
2280
2281 static void check_mutex_acquired(void)
2282 {
2283 BUG_ON(!mutex_is_locked(&slab_mutex));
2284 }
2285
2286 static void check_spinlock_acquired(struct kmem_cache *cachep)
2287 {
2288 #ifdef CONFIG_SMP
2289 check_irq_off();
2290 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2291 #endif
2292 }
2293
2294 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2295 {
2296 #ifdef CONFIG_SMP
2297 check_irq_off();
2298 assert_spin_locked(&get_node(cachep, node)->list_lock);
2299 #endif
2300 }
2301
2302 #else
2303 #define check_irq_off() do { } while(0)
2304 #define check_irq_on() do { } while(0)
2305 #define check_mutex_acquired() do { } while(0)
2306 #define check_spinlock_acquired(x) do { } while(0)
2307 #define check_spinlock_acquired_node(x, y) do { } while(0)
2308 #endif
2309
2310 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2311 int node, bool free_all, struct list_head *list)
2312 {
2313 int tofree;
2314
2315 if (!ac || !ac->avail)
2316 return;
2317
2318 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2319 if (tofree > ac->avail)
2320 tofree = (ac->avail + 1) / 2;
2321
2322 free_block(cachep, ac->entry, tofree, node, list);
2323 ac->avail -= tofree;
2324 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2325 }
2326
2327 static void do_drain(void *arg)
2328 {
2329 struct kmem_cache *cachep = arg;
2330 struct array_cache *ac;
2331 int node = numa_mem_id();
2332 struct kmem_cache_node *n;
2333 LIST_HEAD(list);
2334
2335 check_irq_off();
2336 ac = cpu_cache_get(cachep);
2337 n = get_node(cachep, node);
2338 spin_lock(&n->list_lock);
2339 free_block(cachep, ac->entry, ac->avail, node, &list);
2340 spin_unlock(&n->list_lock);
2341 slabs_destroy(cachep, &list);
2342 ac->avail = 0;
2343 }
2344
2345 static void drain_cpu_caches(struct kmem_cache *cachep)
2346 {
2347 struct kmem_cache_node *n;
2348 int node;
2349 LIST_HEAD(list);
2350
2351 on_each_cpu(do_drain, cachep, 1);
2352 check_irq_on();
2353 for_each_kmem_cache_node(cachep, node, n)
2354 if (n->alien)
2355 drain_alien_cache(cachep, n->alien);
2356
2357 for_each_kmem_cache_node(cachep, node, n) {
2358 spin_lock_irq(&n->list_lock);
2359 drain_array_locked(cachep, n->shared, node, true, &list);
2360 spin_unlock_irq(&n->list_lock);
2361
2362 slabs_destroy(cachep, &list);
2363 }
2364 }
2365
2366 /*
2367 * Remove slabs from the list of free slabs.
2368 * Specify the number of slabs to drain in tofree.
2369 *
2370 * Returns the actual number of slabs released.
2371 */
2372 static int drain_freelist(struct kmem_cache *cache,
2373 struct kmem_cache_node *n, int tofree)
2374 {
2375 struct list_head *p;
2376 int nr_freed;
2377 struct page *page;
2378
2379 nr_freed = 0;
2380 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2381
2382 spin_lock_irq(&n->list_lock);
2383 p = n->slabs_free.prev;
2384 if (p == &n->slabs_free) {
2385 spin_unlock_irq(&n->list_lock);
2386 goto out;
2387 }
2388
2389 page = list_entry(p, struct page, lru);
2390 list_del(&page->lru);
2391 /*
2392 * Safe to drop the lock. The slab is no longer linked
2393 * to the cache.
2394 */
2395 n->free_objects -= cache->num;
2396 spin_unlock_irq(&n->list_lock);
2397 slab_destroy(cache, page);
2398 nr_freed++;
2399 }
2400 out:
2401 return nr_freed;
2402 }
2403
2404 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2405 {
2406 int ret = 0;
2407 int node;
2408 struct kmem_cache_node *n;
2409
2410 drain_cpu_caches(cachep);
2411
2412 check_irq_on();
2413 for_each_kmem_cache_node(cachep, node, n) {
2414 drain_freelist(cachep, n, INT_MAX);
2415
2416 ret += !list_empty(&n->slabs_full) ||
2417 !list_empty(&n->slabs_partial);
2418 }
2419 return (ret ? 1 : 0);
2420 }
2421
2422 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2423 {
2424 return __kmem_cache_shrink(cachep, false);
2425 }
2426
2427 void __kmem_cache_release(struct kmem_cache *cachep)
2428 {
2429 int i;
2430 struct kmem_cache_node *n;
2431
2432 cache_random_seq_destroy(cachep);
2433
2434 free_percpu(cachep->cpu_cache);
2435
2436 /* NUMA: free the node structures */
2437 for_each_kmem_cache_node(cachep, i, n) {
2438 kfree(n->shared);
2439 free_alien_cache(n->alien);
2440 kfree(n);
2441 cachep->node[i] = NULL;
2442 }
2443 }
2444
2445 /*
2446 * Get the memory for a slab management obj.
2447 *
2448 * For a slab cache when the slab descriptor is off-slab, the
2449 * slab descriptor can't come from the same cache which is being created,
2450 * Because if it is the case, that means we defer the creation of
2451 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2452 * And we eventually call down to __kmem_cache_create(), which
2453 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2454 * This is a "chicken-and-egg" problem.
2455 *
2456 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2457 * which are all initialized during kmem_cache_init().
2458 */
2459 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2460 struct page *page, int colour_off,
2461 gfp_t local_flags, int nodeid)
2462 {
2463 void *freelist;
2464 void *addr = page_address(page);
2465
2466 page->s_mem = addr + colour_off;
2467 page->active = 0;
2468
2469 if (OBJFREELIST_SLAB(cachep))
2470 freelist = NULL;
2471 else if (OFF_SLAB(cachep)) {
2472 /* Slab management obj is off-slab. */
2473 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2474 local_flags, nodeid);
2475 if (!freelist)
2476 return NULL;
2477 } else {
2478 /* We will use last bytes at the slab for freelist */
2479 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2480 cachep->freelist_size;
2481 }
2482
2483 return freelist;
2484 }
2485
2486 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2487 {
2488 return ((freelist_idx_t *)page->freelist)[idx];
2489 }
2490
2491 static inline void set_free_obj(struct page *page,
2492 unsigned int idx, freelist_idx_t val)
2493 {
2494 ((freelist_idx_t *)(page->freelist))[idx] = val;
2495 }
2496
2497 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2498 {
2499 #if DEBUG
2500 int i;
2501
2502 for (i = 0; i < cachep->num; i++) {
2503 void *objp = index_to_obj(cachep, page, i);
2504
2505 if (cachep->flags & SLAB_STORE_USER)
2506 *dbg_userword(cachep, objp) = NULL;
2507
2508 if (cachep->flags & SLAB_RED_ZONE) {
2509 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2510 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2511 }
2512 /*
2513 * Constructors are not allowed to allocate memory from the same
2514 * cache which they are a constructor for. Otherwise, deadlock.
2515 * They must also be threaded.
2516 */
2517 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2518 kasan_unpoison_object_data(cachep,
2519 objp + obj_offset(cachep));
2520 cachep->ctor(objp + obj_offset(cachep));
2521 kasan_poison_object_data(
2522 cachep, objp + obj_offset(cachep));
2523 }
2524
2525 if (cachep->flags & SLAB_RED_ZONE) {
2526 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2527 slab_error(cachep, "constructor overwrote the end of an object");
2528 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2529 slab_error(cachep, "constructor overwrote the start of an object");
2530 }
2531 /* need to poison the objs? */
2532 if (cachep->flags & SLAB_POISON) {
2533 poison_obj(cachep, objp, POISON_FREE);
2534 slab_kernel_map(cachep, objp, 0, 0);
2535 }
2536 }
2537 #endif
2538 }
2539
2540 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2541 /* Hold information during a freelist initialization */
2542 union freelist_init_state {
2543 struct {
2544 unsigned int pos;
2545 freelist_idx_t *list;
2546 unsigned int count;
2547 unsigned int rand;
2548 };
2549 struct rnd_state rnd_state;
2550 };
2551
2552 /*
2553 * Initialize the state based on the randomization methode available.
2554 * return true if the pre-computed list is available, false otherwize.
2555 */
2556 static bool freelist_state_initialize(union freelist_init_state *state,
2557 struct kmem_cache *cachep,
2558 unsigned int count)
2559 {
2560 bool ret;
2561 unsigned int rand;
2562
2563 /* Use best entropy available to define a random shift */
2564 get_random_bytes_arch(&rand, sizeof(rand));
2565
2566 /* Use a random state if the pre-computed list is not available */
2567 if (!cachep->random_seq) {
2568 prandom_seed_state(&state->rnd_state, rand);
2569 ret = false;
2570 } else {
2571 state->list = cachep->random_seq;
2572 state->count = count;
2573 state->pos = 0;
2574 state->rand = rand;
2575 ret = true;
2576 }
2577 return ret;
2578 }
2579
2580 /* Get the next entry on the list and randomize it using a random shift */
2581 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2582 {
2583 return (state->list[state->pos++] + state->rand) % state->count;
2584 }
2585
2586 /*
2587 * Shuffle the freelist initialization state based on pre-computed lists.
2588 * return true if the list was successfully shuffled, false otherwise.
2589 */
2590 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2591 {
2592 unsigned int objfreelist = 0, i, count = cachep->num;
2593 union freelist_init_state state;
2594 bool precomputed;
2595
2596 if (count < 2)
2597 return false;
2598
2599 precomputed = freelist_state_initialize(&state, cachep, count);
2600
2601 /* Take a random entry as the objfreelist */
2602 if (OBJFREELIST_SLAB(cachep)) {
2603 if (!precomputed)
2604 objfreelist = count - 1;
2605 else
2606 objfreelist = next_random_slot(&state);
2607 page->freelist = index_to_obj(cachep, page, objfreelist) +
2608 obj_offset(cachep);
2609 count--;
2610 }
2611
2612 /*
2613 * On early boot, generate the list dynamically.
2614 * Later use a pre-computed list for speed.
2615 */
2616 if (!precomputed) {
2617 freelist_randomize(&state.rnd_state, page->freelist, count);
2618 } else {
2619 for (i = 0; i < count; i++)
2620 set_free_obj(page, i, next_random_slot(&state));
2621 }
2622
2623 if (OBJFREELIST_SLAB(cachep))
2624 set_free_obj(page, cachep->num - 1, objfreelist);
2625
2626 return true;
2627 }
2628 #else
2629 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2630 struct page *page)
2631 {
2632 return false;
2633 }
2634 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2635
2636 static void cache_init_objs(struct kmem_cache *cachep,
2637 struct page *page)
2638 {
2639 int i;
2640 void *objp;
2641 bool shuffled;
2642
2643 cache_init_objs_debug(cachep, page);
2644
2645 /* Try to randomize the freelist if enabled */
2646 shuffled = shuffle_freelist(cachep, page);
2647
2648 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2649 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2650 obj_offset(cachep);
2651 }
2652
2653 for (i = 0; i < cachep->num; i++) {
2654 /* constructor could break poison info */
2655 if (DEBUG == 0 && cachep->ctor) {
2656 objp = index_to_obj(cachep, page, i);
2657 kasan_unpoison_object_data(cachep, objp);
2658 cachep->ctor(objp);
2659 kasan_poison_object_data(cachep, objp);
2660 }
2661
2662 if (!shuffled)
2663 set_free_obj(page, i, i);
2664 }
2665 }
2666
2667 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2668 {
2669 void *objp;
2670
2671 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2672 page->active++;
2673
2674 #if DEBUG
2675 if (cachep->flags & SLAB_STORE_USER)
2676 set_store_user_dirty(cachep);
2677 #endif
2678
2679 return objp;
2680 }
2681
2682 static void slab_put_obj(struct kmem_cache *cachep,
2683 struct page *page, void *objp)
2684 {
2685 unsigned int objnr = obj_to_index(cachep, page, objp);
2686 #if DEBUG
2687 unsigned int i;
2688
2689 /* Verify double free bug */
2690 for (i = page->active; i < cachep->num; i++) {
2691 if (get_free_obj(page, i) == objnr) {
2692 pr_err("slab: double free detected in cache '%s', objp %p\n",
2693 cachep->name, objp);
2694 BUG();
2695 }
2696 }
2697 #endif
2698 page->active--;
2699 if (!page->freelist)
2700 page->freelist = objp + obj_offset(cachep);
2701
2702 set_free_obj(page, page->active, objnr);
2703 }
2704
2705 /*
2706 * Map pages beginning at addr to the given cache and slab. This is required
2707 * for the slab allocator to be able to lookup the cache and slab of a
2708 * virtual address for kfree, ksize, and slab debugging.
2709 */
2710 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2711 void *freelist)
2712 {
2713 page->slab_cache = cache;
2714 page->freelist = freelist;
2715 }
2716
2717 /*
2718 * Grow (by 1) the number of slabs within a cache. This is called by
2719 * kmem_cache_alloc() when there are no active objs left in a cache.
2720 */
2721 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2722 gfp_t flags, int nodeid)
2723 {
2724 void *freelist;
2725 size_t offset;
2726 gfp_t local_flags;
2727 int page_node;
2728 struct kmem_cache_node *n;
2729 struct page *page;
2730
2731 /*
2732 * Be lazy and only check for valid flags here, keeping it out of the
2733 * critical path in kmem_cache_alloc().
2734 */
2735 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2736 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2737 BUG();
2738 }
2739 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2740
2741 check_irq_off();
2742 if (gfpflags_allow_blocking(local_flags))
2743 local_irq_enable();
2744
2745 /*
2746 * Get mem for the objs. Attempt to allocate a physical page from
2747 * 'nodeid'.
2748 */
2749 page = kmem_getpages(cachep, local_flags, nodeid);
2750 if (!page)
2751 goto failed;
2752
2753 page_node = page_to_nid(page);
2754 n = get_node(cachep, page_node);
2755
2756 /* Get colour for the slab, and cal the next value. */
2757 n->colour_next++;
2758 if (n->colour_next >= cachep->colour)
2759 n->colour_next = 0;
2760
2761 offset = n->colour_next;
2762 if (offset >= cachep->colour)
2763 offset = 0;
2764
2765 offset *= cachep->colour_off;
2766
2767 /* Get slab management. */
2768 freelist = alloc_slabmgmt(cachep, page, offset,
2769 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2770 if (OFF_SLAB(cachep) && !freelist)
2771 goto opps1;
2772
2773 slab_map_pages(cachep, page, freelist);
2774
2775 kasan_poison_slab(page);
2776 cache_init_objs(cachep, page);
2777
2778 if (gfpflags_allow_blocking(local_flags))
2779 local_irq_disable();
2780
2781 return page;
2782
2783 opps1:
2784 kmem_freepages(cachep, page);
2785 failed:
2786 if (gfpflags_allow_blocking(local_flags))
2787 local_irq_disable();
2788 return NULL;
2789 }
2790
2791 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2792 {
2793 struct kmem_cache_node *n;
2794 void *list = NULL;
2795
2796 check_irq_off();
2797
2798 if (!page)
2799 return;
2800
2801 INIT_LIST_HEAD(&page->lru);
2802 n = get_node(cachep, page_to_nid(page));
2803
2804 spin_lock(&n->list_lock);
2805 if (!page->active)
2806 list_add_tail(&page->lru, &(n->slabs_free));
2807 else
2808 fixup_slab_list(cachep, n, page, &list);
2809 STATS_INC_GROWN(cachep);
2810 n->free_objects += cachep->num - page->active;
2811 spin_unlock(&n->list_lock);
2812
2813 fixup_objfreelist_debug(cachep, &list);
2814 }
2815
2816 #if DEBUG
2817
2818 /*
2819 * Perform extra freeing checks:
2820 * - detect bad pointers.
2821 * - POISON/RED_ZONE checking
2822 */
2823 static void kfree_debugcheck(const void *objp)
2824 {
2825 if (!virt_addr_valid(objp)) {
2826 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2827 (unsigned long)objp);
2828 BUG();
2829 }
2830 }
2831
2832 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2833 {
2834 unsigned long long redzone1, redzone2;
2835
2836 redzone1 = *dbg_redzone1(cache, obj);
2837 redzone2 = *dbg_redzone2(cache, obj);
2838
2839 /*
2840 * Redzone is ok.
2841 */
2842 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2843 return;
2844
2845 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2846 slab_error(cache, "double free detected");
2847 else
2848 slab_error(cache, "memory outside object was overwritten");
2849
2850 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2851 obj, redzone1, redzone2);
2852 }
2853
2854 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2855 unsigned long caller)
2856 {
2857 unsigned int objnr;
2858 struct page *page;
2859
2860 BUG_ON(virt_to_cache(objp) != cachep);
2861
2862 objp -= obj_offset(cachep);
2863 kfree_debugcheck(objp);
2864 page = virt_to_head_page(objp);
2865
2866 if (cachep->flags & SLAB_RED_ZONE) {
2867 verify_redzone_free(cachep, objp);
2868 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2869 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2870 }
2871 if (cachep->flags & SLAB_STORE_USER) {
2872 set_store_user_dirty(cachep);
2873 *dbg_userword(cachep, objp) = (void *)caller;
2874 }
2875
2876 objnr = obj_to_index(cachep, page, objp);
2877
2878 BUG_ON(objnr >= cachep->num);
2879 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2880
2881 if (cachep->flags & SLAB_POISON) {
2882 poison_obj(cachep, objp, POISON_FREE);
2883 slab_kernel_map(cachep, objp, 0, caller);
2884 }
2885 return objp;
2886 }
2887
2888 #else
2889 #define kfree_debugcheck(x) do { } while(0)
2890 #define cache_free_debugcheck(x,objp,z) (objp)
2891 #endif
2892
2893 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2894 void **list)
2895 {
2896 #if DEBUG
2897 void *next = *list;
2898 void *objp;
2899
2900 while (next) {
2901 objp = next - obj_offset(cachep);
2902 next = *(void **)next;
2903 poison_obj(cachep, objp, POISON_FREE);
2904 }
2905 #endif
2906 }
2907
2908 static inline void fixup_slab_list(struct kmem_cache *cachep,
2909 struct kmem_cache_node *n, struct page *page,
2910 void **list)
2911 {
2912 /* move slabp to correct slabp list: */
2913 list_del(&page->lru);
2914 if (page->active == cachep->num) {
2915 list_add(&page->lru, &n->slabs_full);
2916 if (OBJFREELIST_SLAB(cachep)) {
2917 #if DEBUG
2918 /* Poisoning will be done without holding the lock */
2919 if (cachep->flags & SLAB_POISON) {
2920 void **objp = page->freelist;
2921
2922 *objp = *list;
2923 *list = objp;
2924 }
2925 #endif
2926 page->freelist = NULL;
2927 }
2928 } else
2929 list_add(&page->lru, &n->slabs_partial);
2930 }
2931
2932 /* Try to find non-pfmemalloc slab if needed */
2933 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2934 struct page *page, bool pfmemalloc)
2935 {
2936 if (!page)
2937 return NULL;
2938
2939 if (pfmemalloc)
2940 return page;
2941
2942 if (!PageSlabPfmemalloc(page))
2943 return page;
2944
2945 /* No need to keep pfmemalloc slab if we have enough free objects */
2946 if (n->free_objects > n->free_limit) {
2947 ClearPageSlabPfmemalloc(page);
2948 return page;
2949 }
2950
2951 /* Move pfmemalloc slab to the end of list to speed up next search */
2952 list_del(&page->lru);
2953 if (!page->active)
2954 list_add_tail(&page->lru, &n->slabs_free);
2955 else
2956 list_add_tail(&page->lru, &n->slabs_partial);
2957
2958 list_for_each_entry(page, &n->slabs_partial, lru) {
2959 if (!PageSlabPfmemalloc(page))
2960 return page;
2961 }
2962
2963 list_for_each_entry(page, &n->slabs_free, lru) {
2964 if (!PageSlabPfmemalloc(page))
2965 return page;
2966 }
2967
2968 return NULL;
2969 }
2970
2971 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2972 {
2973 struct page *page;
2974
2975 page = list_first_entry_or_null(&n->slabs_partial,
2976 struct page, lru);
2977 if (!page) {
2978 n->free_touched = 1;
2979 page = list_first_entry_or_null(&n->slabs_free,
2980 struct page, lru);
2981 }
2982
2983 if (sk_memalloc_socks())
2984 return get_valid_first_slab(n, page, pfmemalloc);
2985
2986 return page;
2987 }
2988
2989 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2990 struct kmem_cache_node *n, gfp_t flags)
2991 {
2992 struct page *page;
2993 void *obj;
2994 void *list = NULL;
2995
2996 if (!gfp_pfmemalloc_allowed(flags))
2997 return NULL;
2998
2999 spin_lock(&n->list_lock);
3000 page = get_first_slab(n, true);
3001 if (!page) {
3002 spin_unlock(&n->list_lock);
3003 return NULL;
3004 }
3005
3006 obj = slab_get_obj(cachep, page);
3007 n->free_objects--;
3008
3009 fixup_slab_list(cachep, n, page, &list);
3010
3011 spin_unlock(&n->list_lock);
3012 fixup_objfreelist_debug(cachep, &list);
3013
3014 return obj;
3015 }
3016
3017 /*
3018 * Slab list should be fixed up by fixup_slab_list() for existing slab
3019 * or cache_grow_end() for new slab
3020 */
3021 static __always_inline int alloc_block(struct kmem_cache *cachep,
3022 struct array_cache *ac, struct page *page, int batchcount)
3023 {
3024 /*
3025 * There must be at least one object available for
3026 * allocation.
3027 */
3028 BUG_ON(page->active >= cachep->num);
3029
3030 while (page->active < cachep->num && batchcount--) {
3031 STATS_INC_ALLOCED(cachep);
3032 STATS_INC_ACTIVE(cachep);
3033 STATS_SET_HIGH(cachep);
3034
3035 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
3036 }
3037
3038 return batchcount;
3039 }
3040
3041 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3042 {
3043 int batchcount;
3044 struct kmem_cache_node *n;
3045 struct array_cache *ac, *shared;
3046 int node;
3047 void *list = NULL;
3048 struct page *page;
3049
3050 check_irq_off();
3051 node = numa_mem_id();
3052
3053 ac = cpu_cache_get(cachep);
3054 batchcount = ac->batchcount;
3055 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3056 /*
3057 * If there was little recent activity on this cache, then
3058 * perform only a partial refill. Otherwise we could generate
3059 * refill bouncing.
3060 */
3061 batchcount = BATCHREFILL_LIMIT;
3062 }
3063 n = get_node(cachep, node);
3064
3065 BUG_ON(ac->avail > 0 || !n);
3066 shared = READ_ONCE(n->shared);
3067 if (!n->free_objects && (!shared || !shared->avail))
3068 goto direct_grow;
3069
3070 spin_lock(&n->list_lock);
3071 shared = READ_ONCE(n->shared);
3072
3073 /* See if we can refill from the shared array */
3074 if (shared && transfer_objects(ac, shared, batchcount)) {
3075 shared->touched = 1;
3076 goto alloc_done;
3077 }
3078
3079 while (batchcount > 0) {
3080 /* Get slab alloc is to come from. */
3081 page = get_first_slab(n, false);
3082 if (!page)
3083 goto must_grow;
3084
3085 check_spinlock_acquired(cachep);
3086
3087 batchcount = alloc_block(cachep, ac, page, batchcount);
3088 fixup_slab_list(cachep, n, page, &list);
3089 }
3090
3091 must_grow:
3092 n->free_objects -= ac->avail;
3093 alloc_done:
3094 spin_unlock(&n->list_lock);
3095 fixup_objfreelist_debug(cachep, &list);
3096
3097 direct_grow:
3098 if (unlikely(!ac->avail)) {
3099 /* Check if we can use obj in pfmemalloc slab */
3100 if (sk_memalloc_socks()) {
3101 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3102
3103 if (obj)
3104 return obj;
3105 }
3106
3107 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3108
3109 /*
3110 * cache_grow_begin() can reenable interrupts,
3111 * then ac could change.
3112 */
3113 ac = cpu_cache_get(cachep);
3114 if (!ac->avail && page)
3115 alloc_block(cachep, ac, page, batchcount);
3116 cache_grow_end(cachep, page);
3117
3118 if (!ac->avail)
3119 return NULL;
3120 }
3121 ac->touched = 1;
3122
3123 return ac->entry[--ac->avail];
3124 }
3125
3126 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3127 gfp_t flags)
3128 {
3129 might_sleep_if(gfpflags_allow_blocking(flags));
3130 }
3131
3132 #if DEBUG
3133 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3134 gfp_t flags, void *objp, unsigned long caller)
3135 {
3136 if (!objp)
3137 return objp;
3138 if (cachep->flags & SLAB_POISON) {
3139 check_poison_obj(cachep, objp);
3140 slab_kernel_map(cachep, objp, 1, 0);
3141 poison_obj(cachep, objp, POISON_INUSE);
3142 }
3143 if (cachep->flags & SLAB_STORE_USER)
3144 *dbg_userword(cachep, objp) = (void *)caller;
3145
3146 if (cachep->flags & SLAB_RED_ZONE) {
3147 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3148 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3149 slab_error(cachep, "double free, or memory outside object was overwritten");
3150 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3151 objp, *dbg_redzone1(cachep, objp),
3152 *dbg_redzone2(cachep, objp));
3153 }
3154 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3155 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3156 }
3157
3158 objp += obj_offset(cachep);
3159 if (cachep->ctor && cachep->flags & SLAB_POISON)
3160 cachep->ctor(objp);
3161 if (ARCH_SLAB_MINALIGN &&
3162 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3163 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3164 objp, (int)ARCH_SLAB_MINALIGN);
3165 }
3166 return objp;
3167 }
3168 #else
3169 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3170 #endif
3171
3172 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3173 {
3174 void *objp;
3175 struct array_cache *ac;
3176
3177 check_irq_off();
3178
3179 ac = cpu_cache_get(cachep);
3180 if (likely(ac->avail)) {
3181 ac->touched = 1;
3182 objp = ac->entry[--ac->avail];
3183
3184 STATS_INC_ALLOCHIT(cachep);
3185 goto out;
3186 }
3187
3188 STATS_INC_ALLOCMISS(cachep);
3189 objp = cache_alloc_refill(cachep, flags);
3190 /*
3191 * the 'ac' may be updated by cache_alloc_refill(),
3192 * and kmemleak_erase() requires its correct value.
3193 */
3194 ac = cpu_cache_get(cachep);
3195
3196 out:
3197 /*
3198 * To avoid a false negative, if an object that is in one of the
3199 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3200 * treat the array pointers as a reference to the object.
3201 */
3202 if (objp)
3203 kmemleak_erase(&ac->entry[ac->avail]);
3204 return objp;
3205 }
3206
3207 #ifdef CONFIG_NUMA
3208 /*
3209 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3210 *
3211 * If we are in_interrupt, then process context, including cpusets and
3212 * mempolicy, may not apply and should not be used for allocation policy.
3213 */
3214 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3215 {
3216 int nid_alloc, nid_here;
3217
3218 if (in_interrupt() || (flags & __GFP_THISNODE))
3219 return NULL;
3220 nid_alloc = nid_here = numa_mem_id();
3221 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3222 nid_alloc = cpuset_slab_spread_node();
3223 else if (current->mempolicy)
3224 nid_alloc = mempolicy_slab_node();
3225 if (nid_alloc != nid_here)
3226 return ____cache_alloc_node(cachep, flags, nid_alloc);
3227 return NULL;
3228 }
3229
3230 /*
3231 * Fallback function if there was no memory available and no objects on a
3232 * certain node and fall back is permitted. First we scan all the
3233 * available node for available objects. If that fails then we
3234 * perform an allocation without specifying a node. This allows the page
3235 * allocator to do its reclaim / fallback magic. We then insert the
3236 * slab into the proper nodelist and then allocate from it.
3237 */
3238 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3239 {
3240 struct zonelist *zonelist;
3241 struct zoneref *z;
3242 struct zone *zone;
3243 enum zone_type high_zoneidx = gfp_zone(flags);
3244 void *obj = NULL;
3245 struct page *page;
3246 int nid;
3247 unsigned int cpuset_mems_cookie;
3248
3249 if (flags & __GFP_THISNODE)
3250 return NULL;
3251
3252 retry_cpuset:
3253 cpuset_mems_cookie = read_mems_allowed_begin();
3254 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3255
3256 retry:
3257 /*
3258 * Look through allowed nodes for objects available
3259 * from existing per node queues.
3260 */
3261 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3262 nid = zone_to_nid(zone);
3263
3264 if (cpuset_zone_allowed(zone, flags) &&
3265 get_node(cache, nid) &&
3266 get_node(cache, nid)->free_objects) {
3267 obj = ____cache_alloc_node(cache,
3268 gfp_exact_node(flags), nid);
3269 if (obj)
3270 break;
3271 }
3272 }
3273
3274 if (!obj) {
3275 /*
3276 * This allocation will be performed within the constraints
3277 * of the current cpuset / memory policy requirements.
3278 * We may trigger various forms of reclaim on the allowed
3279 * set and go into memory reserves if necessary.
3280 */
3281 page = cache_grow_begin(cache, flags, numa_mem_id());
3282 cache_grow_end(cache, page);
3283 if (page) {
3284 nid = page_to_nid(page);
3285 obj = ____cache_alloc_node(cache,
3286 gfp_exact_node(flags), nid);
3287
3288 /*
3289 * Another processor may allocate the objects in
3290 * the slab since we are not holding any locks.
3291 */
3292 if (!obj)
3293 goto retry;
3294 }
3295 }
3296
3297 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3298 goto retry_cpuset;
3299 return obj;
3300 }
3301
3302 /*
3303 * A interface to enable slab creation on nodeid
3304 */
3305 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3306 int nodeid)
3307 {
3308 struct page *page;
3309 struct kmem_cache_node *n;
3310 void *obj = NULL;
3311 void *list = NULL;
3312
3313 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3314 n = get_node(cachep, nodeid);
3315 BUG_ON(!n);
3316
3317 check_irq_off();
3318 spin_lock(&n->list_lock);
3319 page = get_first_slab(n, false);
3320 if (!page)
3321 goto must_grow;
3322
3323 check_spinlock_acquired_node(cachep, nodeid);
3324
3325 STATS_INC_NODEALLOCS(cachep);
3326 STATS_INC_ACTIVE(cachep);
3327 STATS_SET_HIGH(cachep);
3328
3329 BUG_ON(page->active == cachep->num);
3330
3331 obj = slab_get_obj(cachep, page);
3332 n->free_objects--;
3333
3334 fixup_slab_list(cachep, n, page, &list);
3335
3336 spin_unlock(&n->list_lock);
3337 fixup_objfreelist_debug(cachep, &list);
3338 return obj;
3339
3340 must_grow:
3341 spin_unlock(&n->list_lock);
3342 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3343 if (page) {
3344 /* This slab isn't counted yet so don't update free_objects */
3345 obj = slab_get_obj(cachep, page);
3346 }
3347 cache_grow_end(cachep, page);
3348
3349 return obj ? obj : fallback_alloc(cachep, flags);
3350 }
3351
3352 static __always_inline void *
3353 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3354 unsigned long caller)
3355 {
3356 unsigned long save_flags;
3357 void *ptr;
3358 int slab_node = numa_mem_id();
3359
3360 flags &= gfp_allowed_mask;
3361 cachep = slab_pre_alloc_hook(cachep, flags);
3362 if (unlikely(!cachep))
3363 return NULL;
3364
3365 cache_alloc_debugcheck_before(cachep, flags);
3366 local_irq_save(save_flags);
3367
3368 if (nodeid == NUMA_NO_NODE)
3369 nodeid = slab_node;
3370
3371 if (unlikely(!get_node(cachep, nodeid))) {
3372 /* Node not bootstrapped yet */
3373 ptr = fallback_alloc(cachep, flags);
3374 goto out;
3375 }
3376
3377 if (nodeid == slab_node) {
3378 /*
3379 * Use the locally cached objects if possible.
3380 * However ____cache_alloc does not allow fallback
3381 * to other nodes. It may fail while we still have
3382 * objects on other nodes available.
3383 */
3384 ptr = ____cache_alloc(cachep, flags);
3385 if (ptr)
3386 goto out;
3387 }
3388 /* ___cache_alloc_node can fall back to other nodes */
3389 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3390 out:
3391 local_irq_restore(save_flags);
3392 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3393
3394 if (unlikely(flags & __GFP_ZERO) && ptr)
3395 memset(ptr, 0, cachep->object_size);
3396
3397 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3398 return ptr;
3399 }
3400
3401 static __always_inline void *
3402 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3403 {
3404 void *objp;
3405
3406 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3407 objp = alternate_node_alloc(cache, flags);
3408 if (objp)
3409 goto out;
3410 }
3411 objp = ____cache_alloc(cache, flags);
3412
3413 /*
3414 * We may just have run out of memory on the local node.
3415 * ____cache_alloc_node() knows how to locate memory on other nodes
3416 */
3417 if (!objp)
3418 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3419
3420 out:
3421 return objp;
3422 }
3423 #else
3424
3425 static __always_inline void *
3426 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3427 {
3428 return ____cache_alloc(cachep, flags);
3429 }
3430
3431 #endif /* CONFIG_NUMA */
3432
3433 static __always_inline void *
3434 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3435 {
3436 unsigned long save_flags;
3437 void *objp;
3438
3439 flags &= gfp_allowed_mask;
3440 cachep = slab_pre_alloc_hook(cachep, flags);
3441 if (unlikely(!cachep))
3442 return NULL;
3443
3444 cache_alloc_debugcheck_before(cachep, flags);
3445 local_irq_save(save_flags);
3446 objp = __do_cache_alloc(cachep, flags);
3447 local_irq_restore(save_flags);
3448 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3449 prefetchw(objp);
3450
3451 if (unlikely(flags & __GFP_ZERO) && objp)
3452 memset(objp, 0, cachep->object_size);
3453
3454 slab_post_alloc_hook(cachep, flags, 1, &objp);
3455 return objp;
3456 }
3457
3458 /*
3459 * Caller needs to acquire correct kmem_cache_node's list_lock
3460 * @list: List of detached free slabs should be freed by caller
3461 */
3462 static void free_block(struct kmem_cache *cachep, void **objpp,
3463 int nr_objects, int node, struct list_head *list)
3464 {
3465 int i;
3466 struct kmem_cache_node *n = get_node(cachep, node);
3467 struct page *page;
3468
3469 n->free_objects += nr_objects;
3470
3471 for (i = 0; i < nr_objects; i++) {
3472 void *objp;
3473 struct page *page;
3474
3475 objp = objpp[i];
3476
3477 page = virt_to_head_page(objp);
3478 list_del(&page->lru);
3479 check_spinlock_acquired_node(cachep, node);
3480 slab_put_obj(cachep, page, objp);
3481 STATS_DEC_ACTIVE(cachep);
3482
3483 /* fixup slab chains */
3484 if (page->active == 0)
3485 list_add(&page->lru, &n->slabs_free);
3486 else {
3487 /* Unconditionally move a slab to the end of the
3488 * partial list on free - maximum time for the
3489 * other objects to be freed, too.
3490 */
3491 list_add_tail(&page->lru, &n->slabs_partial);
3492 }
3493 }
3494
3495 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3496 n->free_objects -= cachep->num;
3497
3498 page = list_last_entry(&n->slabs_free, struct page, lru);
3499 list_del(&page->lru);
3500 list_add(&page->lru, list);
3501 }
3502 }
3503
3504 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3505 {
3506 int batchcount;
3507 struct kmem_cache_node *n;
3508 int node = numa_mem_id();
3509 LIST_HEAD(list);
3510
3511 batchcount = ac->batchcount;
3512
3513 check_irq_off();
3514 n = get_node(cachep, node);
3515 spin_lock(&n->list_lock);
3516 if (n->shared) {
3517 struct array_cache *shared_array = n->shared;
3518 int max = shared_array->limit - shared_array->avail;
3519 if (max) {
3520 if (batchcount > max)
3521 batchcount = max;
3522 memcpy(&(shared_array->entry[shared_array->avail]),
3523 ac->entry, sizeof(void *) * batchcount);
3524 shared_array->avail += batchcount;
3525 goto free_done;
3526 }
3527 }
3528
3529 free_block(cachep, ac->entry, batchcount, node, &list);
3530 free_done:
3531 #if STATS
3532 {
3533 int i = 0;
3534 struct page *page;
3535
3536 list_for_each_entry(page, &n->slabs_free, lru) {
3537 BUG_ON(page->active);
3538
3539 i++;
3540 }
3541 STATS_SET_FREEABLE(cachep, i);
3542 }
3543 #endif
3544 spin_unlock(&n->list_lock);
3545 slabs_destroy(cachep, &list);
3546 ac->avail -= batchcount;
3547 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3548 }
3549
3550 /*
3551 * Release an obj back to its cache. If the obj has a constructed state, it must
3552 * be in this state _before_ it is released. Called with disabled ints.
3553 */
3554 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3555 unsigned long caller)
3556 {
3557 struct array_cache *ac = cpu_cache_get(cachep);
3558
3559 kasan_slab_free(cachep, objp);
3560
3561 check_irq_off();
3562 kmemleak_free_recursive(objp, cachep->flags);
3563 objp = cache_free_debugcheck(cachep, objp, caller);
3564
3565 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3566
3567 /*
3568 * Skip calling cache_free_alien() when the platform is not numa.
3569 * This will avoid cache misses that happen while accessing slabp (which
3570 * is per page memory reference) to get nodeid. Instead use a global
3571 * variable to skip the call, which is mostly likely to be present in
3572 * the cache.
3573 */
3574 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3575 return;
3576
3577 if (ac->avail < ac->limit) {
3578 STATS_INC_FREEHIT(cachep);
3579 } else {
3580 STATS_INC_FREEMISS(cachep);
3581 cache_flusharray(cachep, ac);
3582 }
3583
3584 if (sk_memalloc_socks()) {
3585 struct page *page = virt_to_head_page(objp);
3586
3587 if (unlikely(PageSlabPfmemalloc(page))) {
3588 cache_free_pfmemalloc(cachep, page, objp);
3589 return;
3590 }
3591 }
3592
3593 ac->entry[ac->avail++] = objp;
3594 }
3595
3596 /**
3597 * kmem_cache_alloc - Allocate an object
3598 * @cachep: The cache to allocate from.
3599 * @flags: See kmalloc().
3600 *
3601 * Allocate an object from this cache. The flags are only relevant
3602 * if the cache has no available objects.
3603 */
3604 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3605 {
3606 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3607
3608 kasan_slab_alloc(cachep, ret, flags);
3609 trace_kmem_cache_alloc(_RET_IP_, ret,
3610 cachep->object_size, cachep->size, flags);
3611
3612 return ret;
3613 }
3614 EXPORT_SYMBOL(kmem_cache_alloc);
3615
3616 static __always_inline void
3617 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3618 size_t size, void **p, unsigned long caller)
3619 {
3620 size_t i;
3621
3622 for (i = 0; i < size; i++)
3623 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3624 }
3625
3626 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3627 void **p)
3628 {
3629 size_t i;
3630
3631 s = slab_pre_alloc_hook(s, flags);
3632 if (!s)
3633 return 0;
3634
3635 cache_alloc_debugcheck_before(s, flags);
3636
3637 local_irq_disable();
3638 for (i = 0; i < size; i++) {
3639 void *objp = __do_cache_alloc(s, flags);
3640
3641 if (unlikely(!objp))
3642 goto error;
3643 p[i] = objp;
3644 }
3645 local_irq_enable();
3646
3647 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3648
3649 /* Clear memory outside IRQ disabled section */
3650 if (unlikely(flags & __GFP_ZERO))
3651 for (i = 0; i < size; i++)
3652 memset(p[i], 0, s->object_size);
3653
3654 slab_post_alloc_hook(s, flags, size, p);
3655 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3656 return size;
3657 error:
3658 local_irq_enable();
3659 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3660 slab_post_alloc_hook(s, flags, i, p);
3661 __kmem_cache_free_bulk(s, i, p);
3662 return 0;
3663 }
3664 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3665
3666 #ifdef CONFIG_TRACING
3667 void *
3668 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3669 {
3670 void *ret;
3671
3672 ret = slab_alloc(cachep, flags, _RET_IP_);
3673
3674 kasan_kmalloc(cachep, ret, size, flags);
3675 trace_kmalloc(_RET_IP_, ret,
3676 size, cachep->size, flags);
3677 return ret;
3678 }
3679 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3680 #endif
3681
3682 #ifdef CONFIG_NUMA
3683 /**
3684 * kmem_cache_alloc_node - Allocate an object on the specified node
3685 * @cachep: The cache to allocate from.
3686 * @flags: See kmalloc().
3687 * @nodeid: node number of the target node.
3688 *
3689 * Identical to kmem_cache_alloc but it will allocate memory on the given
3690 * node, which can improve the performance for cpu bound structures.
3691 *
3692 * Fallback to other node is possible if __GFP_THISNODE is not set.
3693 */
3694 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3695 {
3696 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3697
3698 kasan_slab_alloc(cachep, ret, flags);
3699 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3700 cachep->object_size, cachep->size,
3701 flags, nodeid);
3702
3703 return ret;
3704 }
3705 EXPORT_SYMBOL(kmem_cache_alloc_node);
3706
3707 #ifdef CONFIG_TRACING
3708 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3709 gfp_t flags,
3710 int nodeid,
3711 size_t size)
3712 {
3713 void *ret;
3714
3715 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3716
3717 kasan_kmalloc(cachep, ret, size, flags);
3718 trace_kmalloc_node(_RET_IP_, ret,
3719 size, cachep->size,
3720 flags, nodeid);
3721 return ret;
3722 }
3723 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3724 #endif
3725
3726 static __always_inline void *
3727 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3728 {
3729 struct kmem_cache *cachep;
3730 void *ret;
3731
3732 cachep = kmalloc_slab(size, flags);
3733 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3734 return cachep;
3735 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3736 kasan_kmalloc(cachep, ret, size, flags);
3737
3738 return ret;
3739 }
3740
3741 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3742 {
3743 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3744 }
3745 EXPORT_SYMBOL(__kmalloc_node);
3746
3747 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3748 int node, unsigned long caller)
3749 {
3750 return __do_kmalloc_node(size, flags, node, caller);
3751 }
3752 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3753 #endif /* CONFIG_NUMA */
3754
3755 /**
3756 * __do_kmalloc - allocate memory
3757 * @size: how many bytes of memory are required.
3758 * @flags: the type of memory to allocate (see kmalloc).
3759 * @caller: function caller for debug tracking of the caller
3760 */
3761 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3762 unsigned long caller)
3763 {
3764 struct kmem_cache *cachep;
3765 void *ret;
3766
3767 cachep = kmalloc_slab(size, flags);
3768 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3769 return cachep;
3770 ret = slab_alloc(cachep, flags, caller);
3771
3772 kasan_kmalloc(cachep, ret, size, flags);
3773 trace_kmalloc(caller, ret,
3774 size, cachep->size, flags);
3775
3776 return ret;
3777 }
3778
3779 void *__kmalloc(size_t size, gfp_t flags)
3780 {
3781 return __do_kmalloc(size, flags, _RET_IP_);
3782 }
3783 EXPORT_SYMBOL(__kmalloc);
3784
3785 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3786 {
3787 return __do_kmalloc(size, flags, caller);
3788 }
3789 EXPORT_SYMBOL(__kmalloc_track_caller);
3790
3791 /**
3792 * kmem_cache_free - Deallocate an object
3793 * @cachep: The cache the allocation was from.
3794 * @objp: The previously allocated object.
3795 *
3796 * Free an object which was previously allocated from this
3797 * cache.
3798 */
3799 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3800 {
3801 unsigned long flags;
3802 cachep = cache_from_obj(cachep, objp);
3803 if (!cachep)
3804 return;
3805
3806 local_irq_save(flags);
3807 debug_check_no_locks_freed(objp, cachep->object_size);
3808 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3809 debug_check_no_obj_freed(objp, cachep->object_size);
3810 __cache_free(cachep, objp, _RET_IP_);
3811 local_irq_restore(flags);
3812
3813 trace_kmem_cache_free(_RET_IP_, objp);
3814 }
3815 EXPORT_SYMBOL(kmem_cache_free);
3816
3817 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3818 {
3819 struct kmem_cache *s;
3820 size_t i;
3821
3822 local_irq_disable();
3823 for (i = 0; i < size; i++) {
3824 void *objp = p[i];
3825
3826 if (!orig_s) /* called via kfree_bulk */
3827 s = virt_to_cache(objp);
3828 else
3829 s = cache_from_obj(orig_s, objp);
3830
3831 debug_check_no_locks_freed(objp, s->object_size);
3832 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3833 debug_check_no_obj_freed(objp, s->object_size);
3834
3835 __cache_free(s, objp, _RET_IP_);
3836 }
3837 local_irq_enable();
3838
3839 /* FIXME: add tracing */
3840 }
3841 EXPORT_SYMBOL(kmem_cache_free_bulk);
3842
3843 /**
3844 * kfree - free previously allocated memory
3845 * @objp: pointer returned by kmalloc.
3846 *
3847 * If @objp is NULL, no operation is performed.
3848 *
3849 * Don't free memory not originally allocated by kmalloc()
3850 * or you will run into trouble.
3851 */
3852 void kfree(const void *objp)
3853 {
3854 struct kmem_cache *c;
3855 unsigned long flags;
3856
3857 trace_kfree(_RET_IP_, objp);
3858
3859 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3860 return;
3861 local_irq_save(flags);
3862 kfree_debugcheck(objp);
3863 c = virt_to_cache(objp);
3864 debug_check_no_locks_freed(objp, c->object_size);
3865
3866 debug_check_no_obj_freed(objp, c->object_size);
3867 __cache_free(c, (void *)objp, _RET_IP_);
3868 local_irq_restore(flags);
3869 }
3870 EXPORT_SYMBOL(kfree);
3871
3872 /*
3873 * This initializes kmem_cache_node or resizes various caches for all nodes.
3874 */
3875 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3876 {
3877 int ret;
3878 int node;
3879 struct kmem_cache_node *n;
3880
3881 for_each_online_node(node) {
3882 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3883 if (ret)
3884 goto fail;
3885
3886 }
3887
3888 return 0;
3889
3890 fail:
3891 if (!cachep->list.next) {
3892 /* Cache is not active yet. Roll back what we did */
3893 node--;
3894 while (node >= 0) {
3895 n = get_node(cachep, node);
3896 if (n) {
3897 kfree(n->shared);
3898 free_alien_cache(n->alien);
3899 kfree(n);
3900 cachep->node[node] = NULL;
3901 }
3902 node--;
3903 }
3904 }
3905 return -ENOMEM;
3906 }
3907
3908 /* Always called with the slab_mutex held */
3909 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3910 int batchcount, int shared, gfp_t gfp)
3911 {
3912 struct array_cache __percpu *cpu_cache, *prev;
3913 int cpu;
3914
3915 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3916 if (!cpu_cache)
3917 return -ENOMEM;
3918
3919 prev = cachep->cpu_cache;
3920 cachep->cpu_cache = cpu_cache;
3921 kick_all_cpus_sync();
3922
3923 check_irq_on();
3924 cachep->batchcount = batchcount;
3925 cachep->limit = limit;
3926 cachep->shared = shared;
3927
3928 if (!prev)
3929 goto setup_node;
3930
3931 for_each_online_cpu(cpu) {
3932 LIST_HEAD(list);
3933 int node;
3934 struct kmem_cache_node *n;
3935 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3936
3937 node = cpu_to_mem(cpu);
3938 n = get_node(cachep, node);
3939 spin_lock_irq(&n->list_lock);
3940 free_block(cachep, ac->entry, ac->avail, node, &list);
3941 spin_unlock_irq(&n->list_lock);
3942 slabs_destroy(cachep, &list);
3943 }
3944 free_percpu(prev);
3945
3946 setup_node:
3947 return setup_kmem_cache_nodes(cachep, gfp);
3948 }
3949
3950 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3951 int batchcount, int shared, gfp_t gfp)
3952 {
3953 int ret;
3954 struct kmem_cache *c;
3955
3956 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3957
3958 if (slab_state < FULL)
3959 return ret;
3960
3961 if ((ret < 0) || !is_root_cache(cachep))
3962 return ret;
3963
3964 lockdep_assert_held(&slab_mutex);
3965 for_each_memcg_cache(c, cachep) {
3966 /* return value determined by the root cache only */
3967 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3968 }
3969
3970 return ret;
3971 }
3972
3973 /* Called with slab_mutex held always */
3974 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3975 {
3976 int err;
3977 int limit = 0;
3978 int shared = 0;
3979 int batchcount = 0;
3980
3981 err = cache_random_seq_create(cachep, gfp);
3982 if (err)
3983 goto end;
3984
3985 if (!is_root_cache(cachep)) {
3986 struct kmem_cache *root = memcg_root_cache(cachep);
3987 limit = root->limit;
3988 shared = root->shared;
3989 batchcount = root->batchcount;
3990 }
3991
3992 if (limit && shared && batchcount)
3993 goto skip_setup;
3994 /*
3995 * The head array serves three purposes:
3996 * - create a LIFO ordering, i.e. return objects that are cache-warm
3997 * - reduce the number of spinlock operations.
3998 * - reduce the number of linked list operations on the slab and
3999 * bufctl chains: array operations are cheaper.
4000 * The numbers are guessed, we should auto-tune as described by
4001 * Bonwick.
4002 */
4003 if (cachep->size > 131072)
4004 limit = 1;
4005 else if (cachep->size > PAGE_SIZE)
4006 limit = 8;
4007 else if (cachep->size > 1024)
4008 limit = 24;
4009 else if (cachep->size > 256)
4010 limit = 54;
4011 else
4012 limit = 120;
4013
4014 /*
4015 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4016 * allocation behaviour: Most allocs on one cpu, most free operations
4017 * on another cpu. For these cases, an efficient object passing between
4018 * cpus is necessary. This is provided by a shared array. The array
4019 * replaces Bonwick's magazine layer.
4020 * On uniprocessor, it's functionally equivalent (but less efficient)
4021 * to a larger limit. Thus disabled by default.
4022 */
4023 shared = 0;
4024 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4025 shared = 8;
4026
4027 #if DEBUG
4028 /*
4029 * With debugging enabled, large batchcount lead to excessively long
4030 * periods with disabled local interrupts. Limit the batchcount
4031 */
4032 if (limit > 32)
4033 limit = 32;
4034 #endif
4035 batchcount = (limit + 1) / 2;
4036 skip_setup:
4037 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4038 end:
4039 if (err)
4040 pr_err("enable_cpucache failed for %s, error %d\n",
4041 cachep->name, -err);
4042 return err;
4043 }
4044
4045 /*
4046 * Drain an array if it contains any elements taking the node lock only if
4047 * necessary. Note that the node listlock also protects the array_cache
4048 * if drain_array() is used on the shared array.
4049 */
4050 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4051 struct array_cache *ac, int node)
4052 {
4053 LIST_HEAD(list);
4054
4055 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4056 check_mutex_acquired();
4057
4058 if (!ac || !ac->avail)
4059 return;
4060
4061 if (ac->touched) {
4062 ac->touched = 0;
4063 return;
4064 }
4065
4066 spin_lock_irq(&n->list_lock);
4067 drain_array_locked(cachep, ac, node, false, &list);
4068 spin_unlock_irq(&n->list_lock);
4069
4070 slabs_destroy(cachep, &list);
4071 }
4072
4073 /**
4074 * cache_reap - Reclaim memory from caches.
4075 * @w: work descriptor
4076 *
4077 * Called from workqueue/eventd every few seconds.
4078 * Purpose:
4079 * - clear the per-cpu caches for this CPU.
4080 * - return freeable pages to the main free memory pool.
4081 *
4082 * If we cannot acquire the cache chain mutex then just give up - we'll try
4083 * again on the next iteration.
4084 */
4085 static void cache_reap(struct work_struct *w)
4086 {
4087 struct kmem_cache *searchp;
4088 struct kmem_cache_node *n;
4089 int node = numa_mem_id();
4090 struct delayed_work *work = to_delayed_work(w);
4091
4092 if (!mutex_trylock(&slab_mutex))
4093 /* Give up. Setup the next iteration. */
4094 goto out;
4095
4096 list_for_each_entry(searchp, &slab_caches, list) {
4097 check_irq_on();
4098
4099 /*
4100 * We only take the node lock if absolutely necessary and we
4101 * have established with reasonable certainty that
4102 * we can do some work if the lock was obtained.
4103 */
4104 n = get_node(searchp, node);
4105
4106 reap_alien(searchp, n);
4107
4108 drain_array(searchp, n, cpu_cache_get(searchp), node);
4109
4110 /*
4111 * These are racy checks but it does not matter
4112 * if we skip one check or scan twice.
4113 */
4114 if (time_after(n->next_reap, jiffies))
4115 goto next;
4116
4117 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4118
4119 drain_array(searchp, n, n->shared, node);
4120
4121 if (n->free_touched)
4122 n->free_touched = 0;
4123 else {
4124 int freed;
4125
4126 freed = drain_freelist(searchp, n, (n->free_limit +
4127 5 * searchp->num - 1) / (5 * searchp->num));
4128 STATS_ADD_REAPED(searchp, freed);
4129 }
4130 next:
4131 cond_resched();
4132 }
4133 check_irq_on();
4134 mutex_unlock(&slab_mutex);
4135 next_reap_node();
4136 out:
4137 /* Set up the next iteration */
4138 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4139 }
4140
4141 #ifdef CONFIG_SLABINFO
4142 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4143 {
4144 struct page *page;
4145 unsigned long active_objs;
4146 unsigned long num_objs;
4147 unsigned long active_slabs = 0;
4148 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4149 const char *name;
4150 char *error = NULL;
4151 int node;
4152 struct kmem_cache_node *n;
4153
4154 active_objs = 0;
4155 num_slabs = 0;
4156 for_each_kmem_cache_node(cachep, node, n) {
4157
4158 check_irq_on();
4159 spin_lock_irq(&n->list_lock);
4160
4161 list_for_each_entry(page, &n->slabs_full, lru) {
4162 if (page->active != cachep->num && !error)
4163 error = "slabs_full accounting error";
4164 active_objs += cachep->num;
4165 active_slabs++;
4166 }
4167 list_for_each_entry(page, &n->slabs_partial, lru) {
4168 if (page->active == cachep->num && !error)
4169 error = "slabs_partial accounting error";
4170 if (!page->active && !error)
4171 error = "slabs_partial accounting error";
4172 active_objs += page->active;
4173 active_slabs++;
4174 }
4175 list_for_each_entry(page, &n->slabs_free, lru) {
4176 if (page->active && !error)
4177 error = "slabs_free accounting error";
4178 num_slabs++;
4179 }
4180 free_objects += n->free_objects;
4181 if (n->shared)
4182 shared_avail += n->shared->avail;
4183
4184 spin_unlock_irq(&n->list_lock);
4185 }
4186 num_slabs += active_slabs;
4187 num_objs = num_slabs * cachep->num;
4188 if (num_objs - active_objs != free_objects && !error)
4189 error = "free_objects accounting error";
4190
4191 name = cachep->name;
4192 if (error)
4193 pr_err("slab: cache %s error: %s\n", name, error);
4194
4195 sinfo->active_objs = active_objs;
4196 sinfo->num_objs = num_objs;
4197 sinfo->active_slabs = active_slabs;
4198 sinfo->num_slabs = num_slabs;
4199 sinfo->shared_avail = shared_avail;
4200 sinfo->limit = cachep->limit;
4201 sinfo->batchcount = cachep->batchcount;
4202 sinfo->shared = cachep->shared;
4203 sinfo->objects_per_slab = cachep->num;
4204 sinfo->cache_order = cachep->gfporder;
4205 }
4206
4207 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4208 {
4209 #if STATS
4210 { /* node stats */
4211 unsigned long high = cachep->high_mark;
4212 unsigned long allocs = cachep->num_allocations;
4213 unsigned long grown = cachep->grown;
4214 unsigned long reaped = cachep->reaped;
4215 unsigned long errors = cachep->errors;
4216 unsigned long max_freeable = cachep->max_freeable;
4217 unsigned long node_allocs = cachep->node_allocs;
4218 unsigned long node_frees = cachep->node_frees;
4219 unsigned long overflows = cachep->node_overflow;
4220
4221 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4222 allocs, high, grown,
4223 reaped, errors, max_freeable, node_allocs,
4224 node_frees, overflows);
4225 }
4226 /* cpu stats */
4227 {
4228 unsigned long allochit = atomic_read(&cachep->allochit);
4229 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4230 unsigned long freehit = atomic_read(&cachep->freehit);
4231 unsigned long freemiss = atomic_read(&cachep->freemiss);
4232
4233 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4234 allochit, allocmiss, freehit, freemiss);
4235 }
4236 #endif
4237 }
4238
4239 #define MAX_SLABINFO_WRITE 128
4240 /**
4241 * slabinfo_write - Tuning for the slab allocator
4242 * @file: unused
4243 * @buffer: user buffer
4244 * @count: data length
4245 * @ppos: unused
4246 */
4247 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4248 size_t count, loff_t *ppos)
4249 {
4250 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4251 int limit, batchcount, shared, res;
4252 struct kmem_cache *cachep;
4253
4254 if (count > MAX_SLABINFO_WRITE)
4255 return -EINVAL;
4256 if (copy_from_user(&kbuf, buffer, count))
4257 return -EFAULT;
4258 kbuf[MAX_SLABINFO_WRITE] = '\0';
4259
4260 tmp = strchr(kbuf, ' ');
4261 if (!tmp)
4262 return -EINVAL;
4263 *tmp = '\0';
4264 tmp++;
4265 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4266 return -EINVAL;
4267
4268 /* Find the cache in the chain of caches. */
4269 mutex_lock(&slab_mutex);
4270 res = -EINVAL;
4271 list_for_each_entry(cachep, &slab_caches, list) {
4272 if (!strcmp(cachep->name, kbuf)) {
4273 if (limit < 1 || batchcount < 1 ||
4274 batchcount > limit || shared < 0) {
4275 res = 0;
4276 } else {
4277 res = do_tune_cpucache(cachep, limit,
4278 batchcount, shared,
4279 GFP_KERNEL);
4280 }
4281 break;
4282 }
4283 }
4284 mutex_unlock(&slab_mutex);
4285 if (res >= 0)
4286 res = count;
4287 return res;
4288 }
4289
4290 #ifdef CONFIG_DEBUG_SLAB_LEAK
4291
4292 static inline int add_caller(unsigned long *n, unsigned long v)
4293 {
4294 unsigned long *p;
4295 int l;
4296 if (!v)
4297 return 1;
4298 l = n[1];
4299 p = n + 2;
4300 while (l) {
4301 int i = l/2;
4302 unsigned long *q = p + 2 * i;
4303 if (*q == v) {
4304 q[1]++;
4305 return 1;
4306 }
4307 if (*q > v) {
4308 l = i;
4309 } else {
4310 p = q + 2;
4311 l -= i + 1;
4312 }
4313 }
4314 if (++n[1] == n[0])
4315 return 0;
4316 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4317 p[0] = v;
4318 p[1] = 1;
4319 return 1;
4320 }
4321
4322 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4323 struct page *page)
4324 {
4325 void *p;
4326 int i, j;
4327 unsigned long v;
4328
4329 if (n[0] == n[1])
4330 return;
4331 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4332 bool active = true;
4333
4334 for (j = page->active; j < c->num; j++) {
4335 if (get_free_obj(page, j) == i) {
4336 active = false;
4337 break;
4338 }
4339 }
4340
4341 if (!active)
4342 continue;
4343
4344 /*
4345 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4346 * mapping is established when actual object allocation and
4347 * we could mistakenly access the unmapped object in the cpu
4348 * cache.
4349 */
4350 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4351 continue;
4352
4353 if (!add_caller(n, v))
4354 return;
4355 }
4356 }
4357
4358 static void show_symbol(struct seq_file *m, unsigned long address)
4359 {
4360 #ifdef CONFIG_KALLSYMS
4361 unsigned long offset, size;
4362 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4363
4364 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4365 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4366 if (modname[0])
4367 seq_printf(m, " [%s]", modname);
4368 return;
4369 }
4370 #endif
4371 seq_printf(m, "%p", (void *)address);
4372 }
4373
4374 static int leaks_show(struct seq_file *m, void *p)
4375 {
4376 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4377 struct page *page;
4378 struct kmem_cache_node *n;
4379 const char *name;
4380 unsigned long *x = m->private;
4381 int node;
4382 int i;
4383
4384 if (!(cachep->flags & SLAB_STORE_USER))
4385 return 0;
4386 if (!(cachep->flags & SLAB_RED_ZONE))
4387 return 0;
4388
4389 /*
4390 * Set store_user_clean and start to grab stored user information
4391 * for all objects on this cache. If some alloc/free requests comes
4392 * during the processing, information would be wrong so restart
4393 * whole processing.
4394 */
4395 do {
4396 set_store_user_clean(cachep);
4397 drain_cpu_caches(cachep);
4398
4399 x[1] = 0;
4400
4401 for_each_kmem_cache_node(cachep, node, n) {
4402
4403 check_irq_on();
4404 spin_lock_irq(&n->list_lock);
4405
4406 list_for_each_entry(page, &n->slabs_full, lru)
4407 handle_slab(x, cachep, page);
4408 list_for_each_entry(page, &n->slabs_partial, lru)
4409 handle_slab(x, cachep, page);
4410 spin_unlock_irq(&n->list_lock);
4411 }
4412 } while (!is_store_user_clean(cachep));
4413
4414 name = cachep->name;
4415 if (x[0] == x[1]) {
4416 /* Increase the buffer size */
4417 mutex_unlock(&slab_mutex);
4418 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4419 if (!m->private) {
4420 /* Too bad, we are really out */
4421 m->private = x;
4422 mutex_lock(&slab_mutex);
4423 return -ENOMEM;
4424 }
4425 *(unsigned long *)m->private = x[0] * 2;
4426 kfree(x);
4427 mutex_lock(&slab_mutex);
4428 /* Now make sure this entry will be retried */
4429 m->count = m->size;
4430 return 0;
4431 }
4432 for (i = 0; i < x[1]; i++) {
4433 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4434 show_symbol(m, x[2*i+2]);
4435 seq_putc(m, '\n');
4436 }
4437
4438 return 0;
4439 }
4440
4441 static const struct seq_operations slabstats_op = {
4442 .start = slab_start,
4443 .next = slab_next,
4444 .stop = slab_stop,
4445 .show = leaks_show,
4446 };
4447
4448 static int slabstats_open(struct inode *inode, struct file *file)
4449 {
4450 unsigned long *n;
4451
4452 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4453 if (!n)
4454 return -ENOMEM;
4455
4456 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4457
4458 return 0;
4459 }
4460
4461 static const struct file_operations proc_slabstats_operations = {
4462 .open = slabstats_open,
4463 .read = seq_read,
4464 .llseek = seq_lseek,
4465 .release = seq_release_private,
4466 };
4467 #endif
4468
4469 static int __init slab_proc_init(void)
4470 {
4471 #ifdef CONFIG_DEBUG_SLAB_LEAK
4472 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4473 #endif
4474 return 0;
4475 }
4476 module_init(slab_proc_init);
4477 #endif
4478
4479 /**
4480 * ksize - get the actual amount of memory allocated for a given object
4481 * @objp: Pointer to the object
4482 *
4483 * kmalloc may internally round up allocations and return more memory
4484 * than requested. ksize() can be used to determine the actual amount of
4485 * memory allocated. The caller may use this additional memory, even though
4486 * a smaller amount of memory was initially specified with the kmalloc call.
4487 * The caller must guarantee that objp points to a valid object previously
4488 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4489 * must not be freed during the duration of the call.
4490 */
4491 size_t ksize(const void *objp)
4492 {
4493 size_t size;
4494
4495 BUG_ON(!objp);
4496 if (unlikely(objp == ZERO_SIZE_PTR))
4497 return 0;
4498
4499 size = virt_to_cache(objp)->object_size;
4500 /* We assume that ksize callers could use the whole allocated area,
4501 * so we need to unpoison this area.
4502 */
4503 kasan_krealloc(objp, size, GFP_NOWAIT);
4504
4505 return size;
4506 }
4507 EXPORT_SYMBOL(ksize);