]> git.proxmox.com Git - mirror_ubuntu-bionic-kernel.git/blob - mm/slab.c
mm: kasan: initial memory quarantine implementation
[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 per_cpu(slab_reap_node, cpu) = next_node_in(cpu_to_mem(cpu),
526 node_online_map);
527 }
528
529 static void next_reap_node(void)
530 {
531 int node = __this_cpu_read(slab_reap_node);
532
533 node = next_node_in(node, node_online_map);
534 __this_cpu_write(slab_reap_node, node);
535 }
536
537 #else
538 #define init_reap_node(cpu) do { } while (0)
539 #define next_reap_node(void) do { } while (0)
540 #endif
541
542 /*
543 * Initiate the reap timer running on the target CPU. We run at around 1 to 2Hz
544 * via the workqueue/eventd.
545 * Add the CPU number into the expiration time to minimize the possibility of
546 * the CPUs getting into lockstep and contending for the global cache chain
547 * lock.
548 */
549 static void start_cpu_timer(int cpu)
550 {
551 struct delayed_work *reap_work = &per_cpu(slab_reap_work, cpu);
552
553 /*
554 * When this gets called from do_initcalls via cpucache_init(),
555 * init_workqueues() has already run, so keventd will be setup
556 * at that time.
557 */
558 if (keventd_up() && reap_work->work.func == NULL) {
559 init_reap_node(cpu);
560 INIT_DEFERRABLE_WORK(reap_work, cache_reap);
561 schedule_delayed_work_on(cpu, reap_work,
562 __round_jiffies_relative(HZ, cpu));
563 }
564 }
565
566 static void init_arraycache(struct array_cache *ac, int limit, int batch)
567 {
568 /*
569 * The array_cache structures contain pointers to free object.
570 * However, when such objects are allocated or transferred to another
571 * cache the pointers are not cleared and they could be counted as
572 * valid references during a kmemleak scan. Therefore, kmemleak must
573 * not scan such objects.
574 */
575 kmemleak_no_scan(ac);
576 if (ac) {
577 ac->avail = 0;
578 ac->limit = limit;
579 ac->batchcount = batch;
580 ac->touched = 0;
581 }
582 }
583
584 static struct array_cache *alloc_arraycache(int node, int entries,
585 int batchcount, gfp_t gfp)
586 {
587 size_t memsize = sizeof(void *) * entries + sizeof(struct array_cache);
588 struct array_cache *ac = NULL;
589
590 ac = kmalloc_node(memsize, gfp, node);
591 init_arraycache(ac, entries, batchcount);
592 return ac;
593 }
594
595 static noinline void cache_free_pfmemalloc(struct kmem_cache *cachep,
596 struct page *page, void *objp)
597 {
598 struct kmem_cache_node *n;
599 int page_node;
600 LIST_HEAD(list);
601
602 page_node = page_to_nid(page);
603 n = get_node(cachep, page_node);
604
605 spin_lock(&n->list_lock);
606 free_block(cachep, &objp, 1, page_node, &list);
607 spin_unlock(&n->list_lock);
608
609 slabs_destroy(cachep, &list);
610 }
611
612 /*
613 * Transfer objects in one arraycache to another.
614 * Locking must be handled by the caller.
615 *
616 * Return the number of entries transferred.
617 */
618 static int transfer_objects(struct array_cache *to,
619 struct array_cache *from, unsigned int max)
620 {
621 /* Figure out how many entries to transfer */
622 int nr = min3(from->avail, max, to->limit - to->avail);
623
624 if (!nr)
625 return 0;
626
627 memcpy(to->entry + to->avail, from->entry + from->avail -nr,
628 sizeof(void *) *nr);
629
630 from->avail -= nr;
631 to->avail += nr;
632 return nr;
633 }
634
635 #ifndef CONFIG_NUMA
636
637 #define drain_alien_cache(cachep, alien) do { } while (0)
638 #define reap_alien(cachep, n) do { } while (0)
639
640 static inline struct alien_cache **alloc_alien_cache(int node,
641 int limit, gfp_t gfp)
642 {
643 return NULL;
644 }
645
646 static inline void free_alien_cache(struct alien_cache **ac_ptr)
647 {
648 }
649
650 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
651 {
652 return 0;
653 }
654
655 static inline void *alternate_node_alloc(struct kmem_cache *cachep,
656 gfp_t flags)
657 {
658 return NULL;
659 }
660
661 static inline void *____cache_alloc_node(struct kmem_cache *cachep,
662 gfp_t flags, int nodeid)
663 {
664 return NULL;
665 }
666
667 static inline gfp_t gfp_exact_node(gfp_t flags)
668 {
669 return flags & ~__GFP_NOFAIL;
670 }
671
672 #else /* CONFIG_NUMA */
673
674 static void *____cache_alloc_node(struct kmem_cache *, gfp_t, int);
675 static void *alternate_node_alloc(struct kmem_cache *, gfp_t);
676
677 static struct alien_cache *__alloc_alien_cache(int node, int entries,
678 int batch, gfp_t gfp)
679 {
680 size_t memsize = sizeof(void *) * entries + sizeof(struct alien_cache);
681 struct alien_cache *alc = NULL;
682
683 alc = kmalloc_node(memsize, gfp, node);
684 init_arraycache(&alc->ac, entries, batch);
685 spin_lock_init(&alc->lock);
686 return alc;
687 }
688
689 static struct alien_cache **alloc_alien_cache(int node, int limit, gfp_t gfp)
690 {
691 struct alien_cache **alc_ptr;
692 size_t memsize = sizeof(void *) * nr_node_ids;
693 int i;
694
695 if (limit > 1)
696 limit = 12;
697 alc_ptr = kzalloc_node(memsize, gfp, node);
698 if (!alc_ptr)
699 return NULL;
700
701 for_each_node(i) {
702 if (i == node || !node_online(i))
703 continue;
704 alc_ptr[i] = __alloc_alien_cache(node, limit, 0xbaadf00d, gfp);
705 if (!alc_ptr[i]) {
706 for (i--; i >= 0; i--)
707 kfree(alc_ptr[i]);
708 kfree(alc_ptr);
709 return NULL;
710 }
711 }
712 return alc_ptr;
713 }
714
715 static void free_alien_cache(struct alien_cache **alc_ptr)
716 {
717 int i;
718
719 if (!alc_ptr)
720 return;
721 for_each_node(i)
722 kfree(alc_ptr[i]);
723 kfree(alc_ptr);
724 }
725
726 static void __drain_alien_cache(struct kmem_cache *cachep,
727 struct array_cache *ac, int node,
728 struct list_head *list)
729 {
730 struct kmem_cache_node *n = get_node(cachep, node);
731
732 if (ac->avail) {
733 spin_lock(&n->list_lock);
734 /*
735 * Stuff objects into the remote nodes shared array first.
736 * That way we could avoid the overhead of putting the objects
737 * into the free lists and getting them back later.
738 */
739 if (n->shared)
740 transfer_objects(n->shared, ac, ac->limit);
741
742 free_block(cachep, ac->entry, ac->avail, node, list);
743 ac->avail = 0;
744 spin_unlock(&n->list_lock);
745 }
746 }
747
748 /*
749 * Called from cache_reap() to regularly drain alien caches round robin.
750 */
751 static void reap_alien(struct kmem_cache *cachep, struct kmem_cache_node *n)
752 {
753 int node = __this_cpu_read(slab_reap_node);
754
755 if (n->alien) {
756 struct alien_cache *alc = n->alien[node];
757 struct array_cache *ac;
758
759 if (alc) {
760 ac = &alc->ac;
761 if (ac->avail && spin_trylock_irq(&alc->lock)) {
762 LIST_HEAD(list);
763
764 __drain_alien_cache(cachep, ac, node, &list);
765 spin_unlock_irq(&alc->lock);
766 slabs_destroy(cachep, &list);
767 }
768 }
769 }
770 }
771
772 static void drain_alien_cache(struct kmem_cache *cachep,
773 struct alien_cache **alien)
774 {
775 int i = 0;
776 struct alien_cache *alc;
777 struct array_cache *ac;
778 unsigned long flags;
779
780 for_each_online_node(i) {
781 alc = alien[i];
782 if (alc) {
783 LIST_HEAD(list);
784
785 ac = &alc->ac;
786 spin_lock_irqsave(&alc->lock, flags);
787 __drain_alien_cache(cachep, ac, i, &list);
788 spin_unlock_irqrestore(&alc->lock, flags);
789 slabs_destroy(cachep, &list);
790 }
791 }
792 }
793
794 static int __cache_free_alien(struct kmem_cache *cachep, void *objp,
795 int node, int page_node)
796 {
797 struct kmem_cache_node *n;
798 struct alien_cache *alien = NULL;
799 struct array_cache *ac;
800 LIST_HEAD(list);
801
802 n = get_node(cachep, node);
803 STATS_INC_NODEFREES(cachep);
804 if (n->alien && n->alien[page_node]) {
805 alien = n->alien[page_node];
806 ac = &alien->ac;
807 spin_lock(&alien->lock);
808 if (unlikely(ac->avail == ac->limit)) {
809 STATS_INC_ACOVERFLOW(cachep);
810 __drain_alien_cache(cachep, ac, page_node, &list);
811 }
812 ac->entry[ac->avail++] = objp;
813 spin_unlock(&alien->lock);
814 slabs_destroy(cachep, &list);
815 } else {
816 n = get_node(cachep, page_node);
817 spin_lock(&n->list_lock);
818 free_block(cachep, &objp, 1, page_node, &list);
819 spin_unlock(&n->list_lock);
820 slabs_destroy(cachep, &list);
821 }
822 return 1;
823 }
824
825 static inline int cache_free_alien(struct kmem_cache *cachep, void *objp)
826 {
827 int page_node = page_to_nid(virt_to_page(objp));
828 int node = numa_mem_id();
829 /*
830 * Make sure we are not freeing a object from another node to the array
831 * cache on this cpu.
832 */
833 if (likely(node == page_node))
834 return 0;
835
836 return __cache_free_alien(cachep, objp, node, page_node);
837 }
838
839 /*
840 * Construct gfp mask to allocate from a specific node but do not reclaim or
841 * warn about failures.
842 */
843 static inline gfp_t gfp_exact_node(gfp_t flags)
844 {
845 return (flags | __GFP_THISNODE | __GFP_NOWARN) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
846 }
847 #endif
848
849 static int init_cache_node(struct kmem_cache *cachep, int node, gfp_t gfp)
850 {
851 struct kmem_cache_node *n;
852
853 /*
854 * Set up the kmem_cache_node for cpu before we can
855 * begin anything. Make sure some other cpu on this
856 * node has not already allocated this
857 */
858 n = get_node(cachep, node);
859 if (n) {
860 spin_lock_irq(&n->list_lock);
861 n->free_limit = (1 + nr_cpus_node(node)) * cachep->batchcount +
862 cachep->num;
863 spin_unlock_irq(&n->list_lock);
864
865 return 0;
866 }
867
868 n = kmalloc_node(sizeof(struct kmem_cache_node), gfp, node);
869 if (!n)
870 return -ENOMEM;
871
872 kmem_cache_node_init(n);
873 n->next_reap = jiffies + REAPTIMEOUT_NODE +
874 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
875
876 n->free_limit =
877 (1 + nr_cpus_node(node)) * cachep->batchcount + cachep->num;
878
879 /*
880 * The kmem_cache_nodes don't come and go as CPUs
881 * come and go. slab_mutex is sufficient
882 * protection here.
883 */
884 cachep->node[node] = n;
885
886 return 0;
887 }
888
889 /*
890 * Allocates and initializes node for a node on each slab cache, used for
891 * either memory or cpu hotplug. If memory is being hot-added, the kmem_cache_node
892 * will be allocated off-node since memory is not yet online for the new node.
893 * When hotplugging memory or a cpu, existing node are not replaced if
894 * already in use.
895 *
896 * Must hold slab_mutex.
897 */
898 static int init_cache_node_node(int node)
899 {
900 int ret;
901 struct kmem_cache *cachep;
902
903 list_for_each_entry(cachep, &slab_caches, list) {
904 ret = init_cache_node(cachep, node, GFP_KERNEL);
905 if (ret)
906 return ret;
907 }
908
909 return 0;
910 }
911
912 static int setup_kmem_cache_node(struct kmem_cache *cachep,
913 int node, gfp_t gfp, bool force_change)
914 {
915 int ret = -ENOMEM;
916 struct kmem_cache_node *n;
917 struct array_cache *old_shared = NULL;
918 struct array_cache *new_shared = NULL;
919 struct alien_cache **new_alien = NULL;
920 LIST_HEAD(list);
921
922 if (use_alien_caches) {
923 new_alien = alloc_alien_cache(node, cachep->limit, gfp);
924 if (!new_alien)
925 goto fail;
926 }
927
928 if (cachep->shared) {
929 new_shared = alloc_arraycache(node,
930 cachep->shared * cachep->batchcount, 0xbaadf00d, gfp);
931 if (!new_shared)
932 goto fail;
933 }
934
935 ret = init_cache_node(cachep, node, gfp);
936 if (ret)
937 goto fail;
938
939 n = get_node(cachep, node);
940 spin_lock_irq(&n->list_lock);
941 if (n->shared && force_change) {
942 free_block(cachep, n->shared->entry,
943 n->shared->avail, node, &list);
944 n->shared->avail = 0;
945 }
946
947 if (!n->shared || force_change) {
948 old_shared = n->shared;
949 n->shared = new_shared;
950 new_shared = NULL;
951 }
952
953 if (!n->alien) {
954 n->alien = new_alien;
955 new_alien = NULL;
956 }
957
958 spin_unlock_irq(&n->list_lock);
959 slabs_destroy(cachep, &list);
960
961 /*
962 * To protect lockless access to n->shared during irq disabled context.
963 * If n->shared isn't NULL in irq disabled context, accessing to it is
964 * guaranteed to be valid until irq is re-enabled, because it will be
965 * freed after synchronize_sched().
966 */
967 if (force_change)
968 synchronize_sched();
969
970 fail:
971 kfree(old_shared);
972 kfree(new_shared);
973 free_alien_cache(new_alien);
974
975 return ret;
976 }
977
978 static void cpuup_canceled(long cpu)
979 {
980 struct kmem_cache *cachep;
981 struct kmem_cache_node *n = NULL;
982 int node = cpu_to_mem(cpu);
983 const struct cpumask *mask = cpumask_of_node(node);
984
985 list_for_each_entry(cachep, &slab_caches, list) {
986 struct array_cache *nc;
987 struct array_cache *shared;
988 struct alien_cache **alien;
989 LIST_HEAD(list);
990
991 n = get_node(cachep, node);
992 if (!n)
993 continue;
994
995 spin_lock_irq(&n->list_lock);
996
997 /* Free limit for this kmem_cache_node */
998 n->free_limit -= cachep->batchcount;
999
1000 /* cpu is dead; no one can alloc from it. */
1001 nc = per_cpu_ptr(cachep->cpu_cache, cpu);
1002 if (nc) {
1003 free_block(cachep, nc->entry, nc->avail, node, &list);
1004 nc->avail = 0;
1005 }
1006
1007 if (!cpumask_empty(mask)) {
1008 spin_unlock_irq(&n->list_lock);
1009 goto free_slab;
1010 }
1011
1012 shared = n->shared;
1013 if (shared) {
1014 free_block(cachep, shared->entry,
1015 shared->avail, node, &list);
1016 n->shared = NULL;
1017 }
1018
1019 alien = n->alien;
1020 n->alien = NULL;
1021
1022 spin_unlock_irq(&n->list_lock);
1023
1024 kfree(shared);
1025 if (alien) {
1026 drain_alien_cache(cachep, alien);
1027 free_alien_cache(alien);
1028 }
1029
1030 free_slab:
1031 slabs_destroy(cachep, &list);
1032 }
1033 /*
1034 * In the previous loop, all the objects were freed to
1035 * the respective cache's slabs, now we can go ahead and
1036 * shrink each nodelist to its limit.
1037 */
1038 list_for_each_entry(cachep, &slab_caches, list) {
1039 n = get_node(cachep, node);
1040 if (!n)
1041 continue;
1042 drain_freelist(cachep, n, INT_MAX);
1043 }
1044 }
1045
1046 static int cpuup_prepare(long cpu)
1047 {
1048 struct kmem_cache *cachep;
1049 int node = cpu_to_mem(cpu);
1050 int err;
1051
1052 /*
1053 * We need to do this right in the beginning since
1054 * alloc_arraycache's are going to use this list.
1055 * kmalloc_node allows us to add the slab to the right
1056 * kmem_cache_node and not this cpu's kmem_cache_node
1057 */
1058 err = init_cache_node_node(node);
1059 if (err < 0)
1060 goto bad;
1061
1062 /*
1063 * Now we can go ahead with allocating the shared arrays and
1064 * array caches
1065 */
1066 list_for_each_entry(cachep, &slab_caches, list) {
1067 err = setup_kmem_cache_node(cachep, node, GFP_KERNEL, false);
1068 if (err)
1069 goto bad;
1070 }
1071
1072 return 0;
1073 bad:
1074 cpuup_canceled(cpu);
1075 return -ENOMEM;
1076 }
1077
1078 static int cpuup_callback(struct notifier_block *nfb,
1079 unsigned long action, void *hcpu)
1080 {
1081 long cpu = (long)hcpu;
1082 int err = 0;
1083
1084 switch (action) {
1085 case CPU_UP_PREPARE:
1086 case CPU_UP_PREPARE_FROZEN:
1087 mutex_lock(&slab_mutex);
1088 err = cpuup_prepare(cpu);
1089 mutex_unlock(&slab_mutex);
1090 break;
1091 case CPU_ONLINE:
1092 case CPU_ONLINE_FROZEN:
1093 start_cpu_timer(cpu);
1094 break;
1095 #ifdef CONFIG_HOTPLUG_CPU
1096 case CPU_DOWN_PREPARE:
1097 case CPU_DOWN_PREPARE_FROZEN:
1098 /*
1099 * Shutdown cache reaper. Note that the slab_mutex is
1100 * held so that if cache_reap() is invoked it cannot do
1101 * anything expensive but will only modify reap_work
1102 * and reschedule the timer.
1103 */
1104 cancel_delayed_work_sync(&per_cpu(slab_reap_work, cpu));
1105 /* Now the cache_reaper is guaranteed to be not running. */
1106 per_cpu(slab_reap_work, cpu).work.func = NULL;
1107 break;
1108 case CPU_DOWN_FAILED:
1109 case CPU_DOWN_FAILED_FROZEN:
1110 start_cpu_timer(cpu);
1111 break;
1112 case CPU_DEAD:
1113 case CPU_DEAD_FROZEN:
1114 /*
1115 * Even if all the cpus of a node are down, we don't free the
1116 * kmem_cache_node of any cache. This to avoid a race between
1117 * cpu_down, and a kmalloc allocation from another cpu for
1118 * memory from the node of the cpu going down. The node
1119 * structure is usually allocated from kmem_cache_create() and
1120 * gets destroyed at kmem_cache_destroy().
1121 */
1122 /* fall through */
1123 #endif
1124 case CPU_UP_CANCELED:
1125 case CPU_UP_CANCELED_FROZEN:
1126 mutex_lock(&slab_mutex);
1127 cpuup_canceled(cpu);
1128 mutex_unlock(&slab_mutex);
1129 break;
1130 }
1131 return notifier_from_errno(err);
1132 }
1133
1134 static struct notifier_block cpucache_notifier = {
1135 &cpuup_callback, NULL, 0
1136 };
1137
1138 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
1139 /*
1140 * Drains freelist for a node on each slab cache, used for memory hot-remove.
1141 * Returns -EBUSY if all objects cannot be drained so that the node is not
1142 * removed.
1143 *
1144 * Must hold slab_mutex.
1145 */
1146 static int __meminit drain_cache_node_node(int node)
1147 {
1148 struct kmem_cache *cachep;
1149 int ret = 0;
1150
1151 list_for_each_entry(cachep, &slab_caches, list) {
1152 struct kmem_cache_node *n;
1153
1154 n = get_node(cachep, node);
1155 if (!n)
1156 continue;
1157
1158 drain_freelist(cachep, n, INT_MAX);
1159
1160 if (!list_empty(&n->slabs_full) ||
1161 !list_empty(&n->slabs_partial)) {
1162 ret = -EBUSY;
1163 break;
1164 }
1165 }
1166 return ret;
1167 }
1168
1169 static int __meminit slab_memory_callback(struct notifier_block *self,
1170 unsigned long action, void *arg)
1171 {
1172 struct memory_notify *mnb = arg;
1173 int ret = 0;
1174 int nid;
1175
1176 nid = mnb->status_change_nid;
1177 if (nid < 0)
1178 goto out;
1179
1180 switch (action) {
1181 case MEM_GOING_ONLINE:
1182 mutex_lock(&slab_mutex);
1183 ret = init_cache_node_node(nid);
1184 mutex_unlock(&slab_mutex);
1185 break;
1186 case MEM_GOING_OFFLINE:
1187 mutex_lock(&slab_mutex);
1188 ret = drain_cache_node_node(nid);
1189 mutex_unlock(&slab_mutex);
1190 break;
1191 case MEM_ONLINE:
1192 case MEM_OFFLINE:
1193 case MEM_CANCEL_ONLINE:
1194 case MEM_CANCEL_OFFLINE:
1195 break;
1196 }
1197 out:
1198 return notifier_from_errno(ret);
1199 }
1200 #endif /* CONFIG_NUMA && CONFIG_MEMORY_HOTPLUG */
1201
1202 /*
1203 * swap the static kmem_cache_node with kmalloced memory
1204 */
1205 static void __init init_list(struct kmem_cache *cachep, struct kmem_cache_node *list,
1206 int nodeid)
1207 {
1208 struct kmem_cache_node *ptr;
1209
1210 ptr = kmalloc_node(sizeof(struct kmem_cache_node), GFP_NOWAIT, nodeid);
1211 BUG_ON(!ptr);
1212
1213 memcpy(ptr, list, sizeof(struct kmem_cache_node));
1214 /*
1215 * Do not assume that spinlocks can be initialized via memcpy:
1216 */
1217 spin_lock_init(&ptr->list_lock);
1218
1219 MAKE_ALL_LISTS(cachep, ptr, nodeid);
1220 cachep->node[nodeid] = ptr;
1221 }
1222
1223 /*
1224 * For setting up all the kmem_cache_node for cache whose buffer_size is same as
1225 * size of kmem_cache_node.
1226 */
1227 static void __init set_up_node(struct kmem_cache *cachep, int index)
1228 {
1229 int node;
1230
1231 for_each_online_node(node) {
1232 cachep->node[node] = &init_kmem_cache_node[index + node];
1233 cachep->node[node]->next_reap = jiffies +
1234 REAPTIMEOUT_NODE +
1235 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1236 }
1237 }
1238
1239 #ifdef CONFIG_SLAB_FREELIST_RANDOM
1240 static void freelist_randomize(struct rnd_state *state, freelist_idx_t *list,
1241 size_t count)
1242 {
1243 size_t i;
1244 unsigned int rand;
1245
1246 for (i = 0; i < count; i++)
1247 list[i] = i;
1248
1249 /* Fisher-Yates shuffle */
1250 for (i = count - 1; i > 0; i--) {
1251 rand = prandom_u32_state(state);
1252 rand %= (i + 1);
1253 swap(list[i], list[rand]);
1254 }
1255 }
1256
1257 /* Create a random sequence per cache */
1258 static int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1259 {
1260 unsigned int seed, count = cachep->num;
1261 struct rnd_state state;
1262
1263 if (count < 2)
1264 return 0;
1265
1266 /* If it fails, we will just use the global lists */
1267 cachep->random_seq = kcalloc(count, sizeof(freelist_idx_t), gfp);
1268 if (!cachep->random_seq)
1269 return -ENOMEM;
1270
1271 /* Get best entropy at this stage */
1272 get_random_bytes_arch(&seed, sizeof(seed));
1273 prandom_seed_state(&state, seed);
1274
1275 freelist_randomize(&state, cachep->random_seq, count);
1276 return 0;
1277 }
1278
1279 /* Destroy the per-cache random freelist sequence */
1280 static void cache_random_seq_destroy(struct kmem_cache *cachep)
1281 {
1282 kfree(cachep->random_seq);
1283 cachep->random_seq = NULL;
1284 }
1285 #else
1286 static inline int cache_random_seq_create(struct kmem_cache *cachep, gfp_t gfp)
1287 {
1288 return 0;
1289 }
1290 static inline void cache_random_seq_destroy(struct kmem_cache *cachep) { }
1291 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
1292
1293
1294 /*
1295 * Initialisation. Called after the page allocator have been initialised and
1296 * before smp_init().
1297 */
1298 void __init kmem_cache_init(void)
1299 {
1300 int i;
1301
1302 BUILD_BUG_ON(sizeof(((struct page *)NULL)->lru) <
1303 sizeof(struct rcu_head));
1304 kmem_cache = &kmem_cache_boot;
1305
1306 if (!IS_ENABLED(CONFIG_NUMA) || num_possible_nodes() == 1)
1307 use_alien_caches = 0;
1308
1309 for (i = 0; i < NUM_INIT_LISTS; i++)
1310 kmem_cache_node_init(&init_kmem_cache_node[i]);
1311
1312 /*
1313 * Fragmentation resistance on low memory - only use bigger
1314 * page orders on machines with more than 32MB of memory if
1315 * not overridden on the command line.
1316 */
1317 if (!slab_max_order_set && totalram_pages > (32 << 20) >> PAGE_SHIFT)
1318 slab_max_order = SLAB_MAX_ORDER_HI;
1319
1320 /* Bootstrap is tricky, because several objects are allocated
1321 * from caches that do not exist yet:
1322 * 1) initialize the kmem_cache cache: it contains the struct
1323 * kmem_cache structures of all caches, except kmem_cache itself:
1324 * kmem_cache is statically allocated.
1325 * Initially an __init data area is used for the head array and the
1326 * kmem_cache_node structures, it's replaced with a kmalloc allocated
1327 * array at the end of the bootstrap.
1328 * 2) Create the first kmalloc cache.
1329 * The struct kmem_cache for the new cache is allocated normally.
1330 * An __init data area is used for the head array.
1331 * 3) Create the remaining kmalloc caches, with minimally sized
1332 * head arrays.
1333 * 4) Replace the __init data head arrays for kmem_cache and the first
1334 * kmalloc cache with kmalloc allocated arrays.
1335 * 5) Replace the __init data for kmem_cache_node for kmem_cache and
1336 * the other cache's with kmalloc allocated memory.
1337 * 6) Resize the head arrays of the kmalloc caches to their final sizes.
1338 */
1339
1340 /* 1) create the kmem_cache */
1341
1342 /*
1343 * struct kmem_cache size depends on nr_node_ids & nr_cpu_ids
1344 */
1345 create_boot_cache(kmem_cache, "kmem_cache",
1346 offsetof(struct kmem_cache, node) +
1347 nr_node_ids * sizeof(struct kmem_cache_node *),
1348 SLAB_HWCACHE_ALIGN);
1349 list_add(&kmem_cache->list, &slab_caches);
1350 slab_state = PARTIAL;
1351
1352 /*
1353 * Initialize the caches that provide memory for the kmem_cache_node
1354 * structures first. Without this, further allocations will bug.
1355 */
1356 kmalloc_caches[INDEX_NODE] = create_kmalloc_cache("kmalloc-node",
1357 kmalloc_size(INDEX_NODE), ARCH_KMALLOC_FLAGS);
1358 slab_state = PARTIAL_NODE;
1359 setup_kmalloc_cache_index_table();
1360
1361 slab_early_init = 0;
1362
1363 /* 5) Replace the bootstrap kmem_cache_node */
1364 {
1365 int nid;
1366
1367 for_each_online_node(nid) {
1368 init_list(kmem_cache, &init_kmem_cache_node[CACHE_CACHE + nid], nid);
1369
1370 init_list(kmalloc_caches[INDEX_NODE],
1371 &init_kmem_cache_node[SIZE_NODE + nid], nid);
1372 }
1373 }
1374
1375 create_kmalloc_caches(ARCH_KMALLOC_FLAGS);
1376 }
1377
1378 void __init kmem_cache_init_late(void)
1379 {
1380 struct kmem_cache *cachep;
1381
1382 slab_state = UP;
1383
1384 /* 6) resize the head arrays to their final sizes */
1385 mutex_lock(&slab_mutex);
1386 list_for_each_entry(cachep, &slab_caches, list)
1387 if (enable_cpucache(cachep, GFP_NOWAIT))
1388 BUG();
1389 mutex_unlock(&slab_mutex);
1390
1391 /* Done! */
1392 slab_state = FULL;
1393
1394 /*
1395 * Register a cpu startup notifier callback that initializes
1396 * cpu_cache_get for all new cpus
1397 */
1398 register_cpu_notifier(&cpucache_notifier);
1399
1400 #ifdef CONFIG_NUMA
1401 /*
1402 * Register a memory hotplug callback that initializes and frees
1403 * node.
1404 */
1405 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
1406 #endif
1407
1408 /*
1409 * The reap timers are started later, with a module init call: That part
1410 * of the kernel is not yet operational.
1411 */
1412 }
1413
1414 static int __init cpucache_init(void)
1415 {
1416 int cpu;
1417
1418 /*
1419 * Register the timers that return unneeded pages to the page allocator
1420 */
1421 for_each_online_cpu(cpu)
1422 start_cpu_timer(cpu);
1423
1424 /* Done! */
1425 slab_state = FULL;
1426 return 0;
1427 }
1428 __initcall(cpucache_init);
1429
1430 static noinline void
1431 slab_out_of_memory(struct kmem_cache *cachep, gfp_t gfpflags, int nodeid)
1432 {
1433 #if DEBUG
1434 struct kmem_cache_node *n;
1435 struct page *page;
1436 unsigned long flags;
1437 int node;
1438 static DEFINE_RATELIMIT_STATE(slab_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
1439 DEFAULT_RATELIMIT_BURST);
1440
1441 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slab_oom_rs))
1442 return;
1443
1444 pr_warn("SLAB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
1445 nodeid, gfpflags, &gfpflags);
1446 pr_warn(" cache: %s, object size: %d, order: %d\n",
1447 cachep->name, cachep->size, cachep->gfporder);
1448
1449 for_each_kmem_cache_node(cachep, node, n) {
1450 unsigned long active_objs = 0, num_objs = 0, free_objects = 0;
1451 unsigned long active_slabs = 0, num_slabs = 0;
1452
1453 spin_lock_irqsave(&n->list_lock, flags);
1454 list_for_each_entry(page, &n->slabs_full, lru) {
1455 active_objs += cachep->num;
1456 active_slabs++;
1457 }
1458 list_for_each_entry(page, &n->slabs_partial, lru) {
1459 active_objs += page->active;
1460 active_slabs++;
1461 }
1462 list_for_each_entry(page, &n->slabs_free, lru)
1463 num_slabs++;
1464
1465 free_objects += n->free_objects;
1466 spin_unlock_irqrestore(&n->list_lock, flags);
1467
1468 num_slabs += active_slabs;
1469 num_objs = num_slabs * cachep->num;
1470 pr_warn(" node %d: slabs: %ld/%ld, objs: %ld/%ld, free: %ld\n",
1471 node, active_slabs, num_slabs, active_objs, num_objs,
1472 free_objects);
1473 }
1474 #endif
1475 }
1476
1477 /*
1478 * Interface to system's page allocator. No need to hold the
1479 * kmem_cache_node ->list_lock.
1480 *
1481 * If we requested dmaable memory, we will get it. Even if we
1482 * did not request dmaable memory, we might get it, but that
1483 * would be relatively rare and ignorable.
1484 */
1485 static struct page *kmem_getpages(struct kmem_cache *cachep, gfp_t flags,
1486 int nodeid)
1487 {
1488 struct page *page;
1489 int nr_pages;
1490
1491 flags |= cachep->allocflags;
1492 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1493 flags |= __GFP_RECLAIMABLE;
1494
1495 page = __alloc_pages_node(nodeid, flags | __GFP_NOTRACK, cachep->gfporder);
1496 if (!page) {
1497 slab_out_of_memory(cachep, flags, nodeid);
1498 return NULL;
1499 }
1500
1501 if (memcg_charge_slab(page, flags, cachep->gfporder, cachep)) {
1502 __free_pages(page, cachep->gfporder);
1503 return NULL;
1504 }
1505
1506 nr_pages = (1 << cachep->gfporder);
1507 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1508 add_zone_page_state(page_zone(page),
1509 NR_SLAB_RECLAIMABLE, nr_pages);
1510 else
1511 add_zone_page_state(page_zone(page),
1512 NR_SLAB_UNRECLAIMABLE, nr_pages);
1513
1514 __SetPageSlab(page);
1515 /* Record if ALLOC_NO_WATERMARKS was set when allocating the slab */
1516 if (sk_memalloc_socks() && page_is_pfmemalloc(page))
1517 SetPageSlabPfmemalloc(page);
1518
1519 if (kmemcheck_enabled && !(cachep->flags & SLAB_NOTRACK)) {
1520 kmemcheck_alloc_shadow(page, cachep->gfporder, flags, nodeid);
1521
1522 if (cachep->ctor)
1523 kmemcheck_mark_uninitialized_pages(page, nr_pages);
1524 else
1525 kmemcheck_mark_unallocated_pages(page, nr_pages);
1526 }
1527
1528 return page;
1529 }
1530
1531 /*
1532 * Interface to system's page release.
1533 */
1534 static void kmem_freepages(struct kmem_cache *cachep, struct page *page)
1535 {
1536 int order = cachep->gfporder;
1537 unsigned long nr_freed = (1 << order);
1538
1539 kmemcheck_free_shadow(page, order);
1540
1541 if (cachep->flags & SLAB_RECLAIM_ACCOUNT)
1542 sub_zone_page_state(page_zone(page),
1543 NR_SLAB_RECLAIMABLE, nr_freed);
1544 else
1545 sub_zone_page_state(page_zone(page),
1546 NR_SLAB_UNRECLAIMABLE, nr_freed);
1547
1548 BUG_ON(!PageSlab(page));
1549 __ClearPageSlabPfmemalloc(page);
1550 __ClearPageSlab(page);
1551 page_mapcount_reset(page);
1552 page->mapping = NULL;
1553
1554 if (current->reclaim_state)
1555 current->reclaim_state->reclaimed_slab += nr_freed;
1556 memcg_uncharge_slab(page, order, cachep);
1557 __free_pages(page, order);
1558 }
1559
1560 static void kmem_rcu_free(struct rcu_head *head)
1561 {
1562 struct kmem_cache *cachep;
1563 struct page *page;
1564
1565 page = container_of(head, struct page, rcu_head);
1566 cachep = page->slab_cache;
1567
1568 kmem_freepages(cachep, page);
1569 }
1570
1571 #if DEBUG
1572 static bool is_debug_pagealloc_cache(struct kmem_cache *cachep)
1573 {
1574 if (debug_pagealloc_enabled() && OFF_SLAB(cachep) &&
1575 (cachep->size % PAGE_SIZE) == 0)
1576 return true;
1577
1578 return false;
1579 }
1580
1581 #ifdef CONFIG_DEBUG_PAGEALLOC
1582 static void store_stackinfo(struct kmem_cache *cachep, unsigned long *addr,
1583 unsigned long caller)
1584 {
1585 int size = cachep->object_size;
1586
1587 addr = (unsigned long *)&((char *)addr)[obj_offset(cachep)];
1588
1589 if (size < 5 * sizeof(unsigned long))
1590 return;
1591
1592 *addr++ = 0x12345678;
1593 *addr++ = caller;
1594 *addr++ = smp_processor_id();
1595 size -= 3 * sizeof(unsigned long);
1596 {
1597 unsigned long *sptr = &caller;
1598 unsigned long svalue;
1599
1600 while (!kstack_end(sptr)) {
1601 svalue = *sptr++;
1602 if (kernel_text_address(svalue)) {
1603 *addr++ = svalue;
1604 size -= sizeof(unsigned long);
1605 if (size <= sizeof(unsigned long))
1606 break;
1607 }
1608 }
1609
1610 }
1611 *addr++ = 0x87654321;
1612 }
1613
1614 static void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1615 int map, unsigned long caller)
1616 {
1617 if (!is_debug_pagealloc_cache(cachep))
1618 return;
1619
1620 if (caller)
1621 store_stackinfo(cachep, objp, caller);
1622
1623 kernel_map_pages(virt_to_page(objp), cachep->size / PAGE_SIZE, map);
1624 }
1625
1626 #else
1627 static inline void slab_kernel_map(struct kmem_cache *cachep, void *objp,
1628 int map, unsigned long caller) {}
1629
1630 #endif
1631
1632 static void poison_obj(struct kmem_cache *cachep, void *addr, unsigned char val)
1633 {
1634 int size = cachep->object_size;
1635 addr = &((char *)addr)[obj_offset(cachep)];
1636
1637 memset(addr, val, size);
1638 *(unsigned char *)(addr + size - 1) = POISON_END;
1639 }
1640
1641 static void dump_line(char *data, int offset, int limit)
1642 {
1643 int i;
1644 unsigned char error = 0;
1645 int bad_count = 0;
1646
1647 pr_err("%03x: ", offset);
1648 for (i = 0; i < limit; i++) {
1649 if (data[offset + i] != POISON_FREE) {
1650 error = data[offset + i];
1651 bad_count++;
1652 }
1653 }
1654 print_hex_dump(KERN_CONT, "", 0, 16, 1,
1655 &data[offset], limit, 1);
1656
1657 if (bad_count == 1) {
1658 error ^= POISON_FREE;
1659 if (!(error & (error - 1))) {
1660 pr_err("Single bit error detected. Probably bad RAM.\n");
1661 #ifdef CONFIG_X86
1662 pr_err("Run memtest86+ or a similar memory test tool.\n");
1663 #else
1664 pr_err("Run a memory test tool.\n");
1665 #endif
1666 }
1667 }
1668 }
1669 #endif
1670
1671 #if DEBUG
1672
1673 static void print_objinfo(struct kmem_cache *cachep, void *objp, int lines)
1674 {
1675 int i, size;
1676 char *realobj;
1677
1678 if (cachep->flags & SLAB_RED_ZONE) {
1679 pr_err("Redzone: 0x%llx/0x%llx\n",
1680 *dbg_redzone1(cachep, objp),
1681 *dbg_redzone2(cachep, objp));
1682 }
1683
1684 if (cachep->flags & SLAB_STORE_USER) {
1685 pr_err("Last user: [<%p>](%pSR)\n",
1686 *dbg_userword(cachep, objp),
1687 *dbg_userword(cachep, objp));
1688 }
1689 realobj = (char *)objp + obj_offset(cachep);
1690 size = cachep->object_size;
1691 for (i = 0; i < size && lines; i += 16, lines--) {
1692 int limit;
1693 limit = 16;
1694 if (i + limit > size)
1695 limit = size - i;
1696 dump_line(realobj, i, limit);
1697 }
1698 }
1699
1700 static void check_poison_obj(struct kmem_cache *cachep, void *objp)
1701 {
1702 char *realobj;
1703 int size, i;
1704 int lines = 0;
1705
1706 if (is_debug_pagealloc_cache(cachep))
1707 return;
1708
1709 realobj = (char *)objp + obj_offset(cachep);
1710 size = cachep->object_size;
1711
1712 for (i = 0; i < size; i++) {
1713 char exp = POISON_FREE;
1714 if (i == size - 1)
1715 exp = POISON_END;
1716 if (realobj[i] != exp) {
1717 int limit;
1718 /* Mismatch ! */
1719 /* Print header */
1720 if (lines == 0) {
1721 pr_err("Slab corruption (%s): %s start=%p, len=%d\n",
1722 print_tainted(), cachep->name,
1723 realobj, size);
1724 print_objinfo(cachep, objp, 0);
1725 }
1726 /* Hexdump the affected line */
1727 i = (i / 16) * 16;
1728 limit = 16;
1729 if (i + limit > size)
1730 limit = size - i;
1731 dump_line(realobj, i, limit);
1732 i += 16;
1733 lines++;
1734 /* Limit to 5 lines */
1735 if (lines > 5)
1736 break;
1737 }
1738 }
1739 if (lines != 0) {
1740 /* Print some data about the neighboring objects, if they
1741 * exist:
1742 */
1743 struct page *page = virt_to_head_page(objp);
1744 unsigned int objnr;
1745
1746 objnr = obj_to_index(cachep, page, objp);
1747 if (objnr) {
1748 objp = index_to_obj(cachep, page, objnr - 1);
1749 realobj = (char *)objp + obj_offset(cachep);
1750 pr_err("Prev obj: start=%p, len=%d\n", realobj, size);
1751 print_objinfo(cachep, objp, 2);
1752 }
1753 if (objnr + 1 < cachep->num) {
1754 objp = index_to_obj(cachep, page, objnr + 1);
1755 realobj = (char *)objp + obj_offset(cachep);
1756 pr_err("Next obj: start=%p, len=%d\n", realobj, size);
1757 print_objinfo(cachep, objp, 2);
1758 }
1759 }
1760 }
1761 #endif
1762
1763 #if DEBUG
1764 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1765 struct page *page)
1766 {
1767 int i;
1768
1769 if (OBJFREELIST_SLAB(cachep) && cachep->flags & SLAB_POISON) {
1770 poison_obj(cachep, page->freelist - obj_offset(cachep),
1771 POISON_FREE);
1772 }
1773
1774 for (i = 0; i < cachep->num; i++) {
1775 void *objp = index_to_obj(cachep, page, i);
1776
1777 if (cachep->flags & SLAB_POISON) {
1778 check_poison_obj(cachep, objp);
1779 slab_kernel_map(cachep, objp, 1, 0);
1780 }
1781 if (cachep->flags & SLAB_RED_ZONE) {
1782 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
1783 slab_error(cachep, "start of a freed object was overwritten");
1784 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
1785 slab_error(cachep, "end of a freed object was overwritten");
1786 }
1787 }
1788 }
1789 #else
1790 static void slab_destroy_debugcheck(struct kmem_cache *cachep,
1791 struct page *page)
1792 {
1793 }
1794 #endif
1795
1796 /**
1797 * slab_destroy - destroy and release all objects in a slab
1798 * @cachep: cache pointer being destroyed
1799 * @page: page pointer being destroyed
1800 *
1801 * Destroy all the objs in a slab page, and release the mem back to the system.
1802 * Before calling the slab page must have been unlinked from the cache. The
1803 * kmem_cache_node ->list_lock is not held/needed.
1804 */
1805 static void slab_destroy(struct kmem_cache *cachep, struct page *page)
1806 {
1807 void *freelist;
1808
1809 freelist = page->freelist;
1810 slab_destroy_debugcheck(cachep, page);
1811 if (unlikely(cachep->flags & SLAB_DESTROY_BY_RCU))
1812 call_rcu(&page->rcu_head, kmem_rcu_free);
1813 else
1814 kmem_freepages(cachep, page);
1815
1816 /*
1817 * From now on, we don't use freelist
1818 * although actual page can be freed in rcu context
1819 */
1820 if (OFF_SLAB(cachep))
1821 kmem_cache_free(cachep->freelist_cache, freelist);
1822 }
1823
1824 static void slabs_destroy(struct kmem_cache *cachep, struct list_head *list)
1825 {
1826 struct page *page, *n;
1827
1828 list_for_each_entry_safe(page, n, list, lru) {
1829 list_del(&page->lru);
1830 slab_destroy(cachep, page);
1831 }
1832 }
1833
1834 /**
1835 * calculate_slab_order - calculate size (page order) of slabs
1836 * @cachep: pointer to the cache that is being created
1837 * @size: size of objects to be created in this cache.
1838 * @flags: slab allocation flags
1839 *
1840 * Also calculates the number of objects per slab.
1841 *
1842 * This could be made much more intelligent. For now, try to avoid using
1843 * high order pages for slabs. When the gfp() functions are more friendly
1844 * towards high-order requests, this should be changed.
1845 */
1846 static size_t calculate_slab_order(struct kmem_cache *cachep,
1847 size_t size, unsigned long flags)
1848 {
1849 size_t left_over = 0;
1850 int gfporder;
1851
1852 for (gfporder = 0; gfporder <= KMALLOC_MAX_ORDER; gfporder++) {
1853 unsigned int num;
1854 size_t remainder;
1855
1856 num = cache_estimate(gfporder, size, flags, &remainder);
1857 if (!num)
1858 continue;
1859
1860 /* Can't handle number of objects more than SLAB_OBJ_MAX_NUM */
1861 if (num > SLAB_OBJ_MAX_NUM)
1862 break;
1863
1864 if (flags & CFLGS_OFF_SLAB) {
1865 struct kmem_cache *freelist_cache;
1866 size_t freelist_size;
1867
1868 freelist_size = num * sizeof(freelist_idx_t);
1869 freelist_cache = kmalloc_slab(freelist_size, 0u);
1870 if (!freelist_cache)
1871 continue;
1872
1873 /*
1874 * Needed to avoid possible looping condition
1875 * in cache_grow_begin()
1876 */
1877 if (OFF_SLAB(freelist_cache))
1878 continue;
1879
1880 /* check if off slab has enough benefit */
1881 if (freelist_cache->size > cachep->size / 2)
1882 continue;
1883 }
1884
1885 /* Found something acceptable - save it away */
1886 cachep->num = num;
1887 cachep->gfporder = gfporder;
1888 left_over = remainder;
1889
1890 /*
1891 * A VFS-reclaimable slab tends to have most allocations
1892 * as GFP_NOFS and we really don't want to have to be allocating
1893 * higher-order pages when we are unable to shrink dcache.
1894 */
1895 if (flags & SLAB_RECLAIM_ACCOUNT)
1896 break;
1897
1898 /*
1899 * Large number of objects is good, but very large slabs are
1900 * currently bad for the gfp()s.
1901 */
1902 if (gfporder >= slab_max_order)
1903 break;
1904
1905 /*
1906 * Acceptable internal fragmentation?
1907 */
1908 if (left_over * 8 <= (PAGE_SIZE << gfporder))
1909 break;
1910 }
1911 return left_over;
1912 }
1913
1914 static struct array_cache __percpu *alloc_kmem_cache_cpus(
1915 struct kmem_cache *cachep, int entries, int batchcount)
1916 {
1917 int cpu;
1918 size_t size;
1919 struct array_cache __percpu *cpu_cache;
1920
1921 size = sizeof(void *) * entries + sizeof(struct array_cache);
1922 cpu_cache = __alloc_percpu(size, sizeof(void *));
1923
1924 if (!cpu_cache)
1925 return NULL;
1926
1927 for_each_possible_cpu(cpu) {
1928 init_arraycache(per_cpu_ptr(cpu_cache, cpu),
1929 entries, batchcount);
1930 }
1931
1932 return cpu_cache;
1933 }
1934
1935 static int __init_refok setup_cpu_cache(struct kmem_cache *cachep, gfp_t gfp)
1936 {
1937 if (slab_state >= FULL)
1938 return enable_cpucache(cachep, gfp);
1939
1940 cachep->cpu_cache = alloc_kmem_cache_cpus(cachep, 1, 1);
1941 if (!cachep->cpu_cache)
1942 return 1;
1943
1944 if (slab_state == DOWN) {
1945 /* Creation of first cache (kmem_cache). */
1946 set_up_node(kmem_cache, CACHE_CACHE);
1947 } else if (slab_state == PARTIAL) {
1948 /* For kmem_cache_node */
1949 set_up_node(cachep, SIZE_NODE);
1950 } else {
1951 int node;
1952
1953 for_each_online_node(node) {
1954 cachep->node[node] = kmalloc_node(
1955 sizeof(struct kmem_cache_node), gfp, node);
1956 BUG_ON(!cachep->node[node]);
1957 kmem_cache_node_init(cachep->node[node]);
1958 }
1959 }
1960
1961 cachep->node[numa_mem_id()]->next_reap =
1962 jiffies + REAPTIMEOUT_NODE +
1963 ((unsigned long)cachep) % REAPTIMEOUT_NODE;
1964
1965 cpu_cache_get(cachep)->avail = 0;
1966 cpu_cache_get(cachep)->limit = BOOT_CPUCACHE_ENTRIES;
1967 cpu_cache_get(cachep)->batchcount = 1;
1968 cpu_cache_get(cachep)->touched = 0;
1969 cachep->batchcount = 1;
1970 cachep->limit = BOOT_CPUCACHE_ENTRIES;
1971 return 0;
1972 }
1973
1974 unsigned long kmem_cache_flags(unsigned long object_size,
1975 unsigned long flags, const char *name,
1976 void (*ctor)(void *))
1977 {
1978 return flags;
1979 }
1980
1981 struct kmem_cache *
1982 __kmem_cache_alias(const char *name, size_t size, size_t align,
1983 unsigned long flags, void (*ctor)(void *))
1984 {
1985 struct kmem_cache *cachep;
1986
1987 cachep = find_mergeable(size, align, flags, name, ctor);
1988 if (cachep) {
1989 cachep->refcount++;
1990
1991 /*
1992 * Adjust the object sizes so that we clear
1993 * the complete object on kzalloc.
1994 */
1995 cachep->object_size = max_t(int, cachep->object_size, size);
1996 }
1997 return cachep;
1998 }
1999
2000 static bool set_objfreelist_slab_cache(struct kmem_cache *cachep,
2001 size_t size, unsigned long flags)
2002 {
2003 size_t left;
2004
2005 cachep->num = 0;
2006
2007 if (cachep->ctor || flags & SLAB_DESTROY_BY_RCU)
2008 return false;
2009
2010 left = calculate_slab_order(cachep, size,
2011 flags | CFLGS_OBJFREELIST_SLAB);
2012 if (!cachep->num)
2013 return false;
2014
2015 if (cachep->num * sizeof(freelist_idx_t) > cachep->object_size)
2016 return false;
2017
2018 cachep->colour = left / cachep->colour_off;
2019
2020 return true;
2021 }
2022
2023 static bool set_off_slab_cache(struct kmem_cache *cachep,
2024 size_t size, unsigned long flags)
2025 {
2026 size_t left;
2027
2028 cachep->num = 0;
2029
2030 /*
2031 * Always use on-slab management when SLAB_NOLEAKTRACE
2032 * to avoid recursive calls into kmemleak.
2033 */
2034 if (flags & SLAB_NOLEAKTRACE)
2035 return false;
2036
2037 /*
2038 * Size is large, assume best to place the slab management obj
2039 * off-slab (should allow better packing of objs).
2040 */
2041 left = calculate_slab_order(cachep, size, flags | CFLGS_OFF_SLAB);
2042 if (!cachep->num)
2043 return false;
2044
2045 /*
2046 * If the slab has been placed off-slab, and we have enough space then
2047 * move it on-slab. This is at the expense of any extra colouring.
2048 */
2049 if (left >= cachep->num * sizeof(freelist_idx_t))
2050 return false;
2051
2052 cachep->colour = left / cachep->colour_off;
2053
2054 return true;
2055 }
2056
2057 static bool set_on_slab_cache(struct kmem_cache *cachep,
2058 size_t size, unsigned long flags)
2059 {
2060 size_t left;
2061
2062 cachep->num = 0;
2063
2064 left = calculate_slab_order(cachep, size, flags);
2065 if (!cachep->num)
2066 return false;
2067
2068 cachep->colour = left / cachep->colour_off;
2069
2070 return true;
2071 }
2072
2073 /**
2074 * __kmem_cache_create - Create a cache.
2075 * @cachep: cache management descriptor
2076 * @flags: SLAB flags
2077 *
2078 * Returns a ptr to the cache on success, NULL on failure.
2079 * Cannot be called within a int, but can be interrupted.
2080 * The @ctor is run when new pages are allocated by the cache.
2081 *
2082 * The flags are
2083 *
2084 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
2085 * to catch references to uninitialised memory.
2086 *
2087 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
2088 * for buffer overruns.
2089 *
2090 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
2091 * cacheline. This can be beneficial if you're counting cycles as closely
2092 * as davem.
2093 */
2094 int
2095 __kmem_cache_create (struct kmem_cache *cachep, unsigned long flags)
2096 {
2097 size_t ralign = BYTES_PER_WORD;
2098 gfp_t gfp;
2099 int err;
2100 size_t size = cachep->size;
2101
2102 #if DEBUG
2103 #if FORCED_DEBUG
2104 /*
2105 * Enable redzoning and last user accounting, except for caches with
2106 * large objects, if the increased size would increase the object size
2107 * above the next power of two: caches with object sizes just above a
2108 * power of two have a significant amount of internal fragmentation.
2109 */
2110 if (size < 4096 || fls(size - 1) == fls(size-1 + REDZONE_ALIGN +
2111 2 * sizeof(unsigned long long)))
2112 flags |= SLAB_RED_ZONE | SLAB_STORE_USER;
2113 if (!(flags & SLAB_DESTROY_BY_RCU))
2114 flags |= SLAB_POISON;
2115 #endif
2116 #endif
2117
2118 /*
2119 * Check that size is in terms of words. This is needed to avoid
2120 * unaligned accesses for some archs when redzoning is used, and makes
2121 * sure any on-slab bufctl's are also correctly aligned.
2122 */
2123 if (size & (BYTES_PER_WORD - 1)) {
2124 size += (BYTES_PER_WORD - 1);
2125 size &= ~(BYTES_PER_WORD - 1);
2126 }
2127
2128 if (flags & SLAB_RED_ZONE) {
2129 ralign = REDZONE_ALIGN;
2130 /* If redzoning, ensure that the second redzone is suitably
2131 * aligned, by adjusting the object size accordingly. */
2132 size += REDZONE_ALIGN - 1;
2133 size &= ~(REDZONE_ALIGN - 1);
2134 }
2135
2136 /* 3) caller mandated alignment */
2137 if (ralign < cachep->align) {
2138 ralign = cachep->align;
2139 }
2140 /* disable debug if necessary */
2141 if (ralign > __alignof__(unsigned long long))
2142 flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2143 /*
2144 * 4) Store it.
2145 */
2146 cachep->align = ralign;
2147 cachep->colour_off = cache_line_size();
2148 /* Offset must be a multiple of the alignment. */
2149 if (cachep->colour_off < cachep->align)
2150 cachep->colour_off = cachep->align;
2151
2152 if (slab_is_available())
2153 gfp = GFP_KERNEL;
2154 else
2155 gfp = GFP_NOWAIT;
2156
2157 #if DEBUG
2158
2159 /*
2160 * Both debugging options require word-alignment which is calculated
2161 * into align above.
2162 */
2163 if (flags & SLAB_RED_ZONE) {
2164 /* add space for red zone words */
2165 cachep->obj_offset += sizeof(unsigned long long);
2166 size += 2 * sizeof(unsigned long long);
2167 }
2168 if (flags & SLAB_STORE_USER) {
2169 /* user store requires one word storage behind the end of
2170 * the real object. But if the second red zone needs to be
2171 * aligned to 64 bits, we must allow that much space.
2172 */
2173 if (flags & SLAB_RED_ZONE)
2174 size += REDZONE_ALIGN;
2175 else
2176 size += BYTES_PER_WORD;
2177 }
2178 #endif
2179
2180 kasan_cache_create(cachep, &size, &flags);
2181
2182 size = ALIGN(size, cachep->align);
2183 /*
2184 * We should restrict the number of objects in a slab to implement
2185 * byte sized index. Refer comment on SLAB_OBJ_MIN_SIZE definition.
2186 */
2187 if (FREELIST_BYTE_INDEX && size < SLAB_OBJ_MIN_SIZE)
2188 size = ALIGN(SLAB_OBJ_MIN_SIZE, cachep->align);
2189
2190 #if DEBUG
2191 /*
2192 * To activate debug pagealloc, off-slab management is necessary
2193 * requirement. In early phase of initialization, small sized slab
2194 * doesn't get initialized so it would not be possible. So, we need
2195 * to check size >= 256. It guarantees that all necessary small
2196 * sized slab is initialized in current slab initialization sequence.
2197 */
2198 if (debug_pagealloc_enabled() && (flags & SLAB_POISON) &&
2199 size >= 256 && cachep->object_size > cache_line_size()) {
2200 if (size < PAGE_SIZE || size % PAGE_SIZE == 0) {
2201 size_t tmp_size = ALIGN(size, PAGE_SIZE);
2202
2203 if (set_off_slab_cache(cachep, tmp_size, flags)) {
2204 flags |= CFLGS_OFF_SLAB;
2205 cachep->obj_offset += tmp_size - size;
2206 size = tmp_size;
2207 goto done;
2208 }
2209 }
2210 }
2211 #endif
2212
2213 if (set_objfreelist_slab_cache(cachep, size, flags)) {
2214 flags |= CFLGS_OBJFREELIST_SLAB;
2215 goto done;
2216 }
2217
2218 if (set_off_slab_cache(cachep, size, flags)) {
2219 flags |= CFLGS_OFF_SLAB;
2220 goto done;
2221 }
2222
2223 if (set_on_slab_cache(cachep, size, flags))
2224 goto done;
2225
2226 return -E2BIG;
2227
2228 done:
2229 cachep->freelist_size = cachep->num * sizeof(freelist_idx_t);
2230 cachep->flags = flags;
2231 cachep->allocflags = __GFP_COMP;
2232 if (flags & SLAB_CACHE_DMA)
2233 cachep->allocflags |= GFP_DMA;
2234 cachep->size = size;
2235 cachep->reciprocal_buffer_size = reciprocal_value(size);
2236
2237 #if DEBUG
2238 /*
2239 * If we're going to use the generic kernel_map_pages()
2240 * poisoning, then it's going to smash the contents of
2241 * the redzone and userword anyhow, so switch them off.
2242 */
2243 if (IS_ENABLED(CONFIG_PAGE_POISONING) &&
2244 (cachep->flags & SLAB_POISON) &&
2245 is_debug_pagealloc_cache(cachep))
2246 cachep->flags &= ~(SLAB_RED_ZONE | SLAB_STORE_USER);
2247 #endif
2248
2249 if (OFF_SLAB(cachep)) {
2250 cachep->freelist_cache =
2251 kmalloc_slab(cachep->freelist_size, 0u);
2252 }
2253
2254 err = setup_cpu_cache(cachep, gfp);
2255 if (err) {
2256 __kmem_cache_release(cachep);
2257 return err;
2258 }
2259
2260 return 0;
2261 }
2262
2263 #if DEBUG
2264 static void check_irq_off(void)
2265 {
2266 BUG_ON(!irqs_disabled());
2267 }
2268
2269 static void check_irq_on(void)
2270 {
2271 BUG_ON(irqs_disabled());
2272 }
2273
2274 static void check_mutex_acquired(void)
2275 {
2276 BUG_ON(!mutex_is_locked(&slab_mutex));
2277 }
2278
2279 static void check_spinlock_acquired(struct kmem_cache *cachep)
2280 {
2281 #ifdef CONFIG_SMP
2282 check_irq_off();
2283 assert_spin_locked(&get_node(cachep, numa_mem_id())->list_lock);
2284 #endif
2285 }
2286
2287 static void check_spinlock_acquired_node(struct kmem_cache *cachep, int node)
2288 {
2289 #ifdef CONFIG_SMP
2290 check_irq_off();
2291 assert_spin_locked(&get_node(cachep, node)->list_lock);
2292 #endif
2293 }
2294
2295 #else
2296 #define check_irq_off() do { } while(0)
2297 #define check_irq_on() do { } while(0)
2298 #define check_mutex_acquired() do { } while(0)
2299 #define check_spinlock_acquired(x) do { } while(0)
2300 #define check_spinlock_acquired_node(x, y) do { } while(0)
2301 #endif
2302
2303 static void drain_array_locked(struct kmem_cache *cachep, struct array_cache *ac,
2304 int node, bool free_all, struct list_head *list)
2305 {
2306 int tofree;
2307
2308 if (!ac || !ac->avail)
2309 return;
2310
2311 tofree = free_all ? ac->avail : (ac->limit + 4) / 5;
2312 if (tofree > ac->avail)
2313 tofree = (ac->avail + 1) / 2;
2314
2315 free_block(cachep, ac->entry, tofree, node, list);
2316 ac->avail -= tofree;
2317 memmove(ac->entry, &(ac->entry[tofree]), sizeof(void *) * ac->avail);
2318 }
2319
2320 static void do_drain(void *arg)
2321 {
2322 struct kmem_cache *cachep = arg;
2323 struct array_cache *ac;
2324 int node = numa_mem_id();
2325 struct kmem_cache_node *n;
2326 LIST_HEAD(list);
2327
2328 check_irq_off();
2329 ac = cpu_cache_get(cachep);
2330 n = get_node(cachep, node);
2331 spin_lock(&n->list_lock);
2332 free_block(cachep, ac->entry, ac->avail, node, &list);
2333 spin_unlock(&n->list_lock);
2334 slabs_destroy(cachep, &list);
2335 ac->avail = 0;
2336 }
2337
2338 static void drain_cpu_caches(struct kmem_cache *cachep)
2339 {
2340 struct kmem_cache_node *n;
2341 int node;
2342 LIST_HEAD(list);
2343
2344 on_each_cpu(do_drain, cachep, 1);
2345 check_irq_on();
2346 for_each_kmem_cache_node(cachep, node, n)
2347 if (n->alien)
2348 drain_alien_cache(cachep, n->alien);
2349
2350 for_each_kmem_cache_node(cachep, node, n) {
2351 spin_lock_irq(&n->list_lock);
2352 drain_array_locked(cachep, n->shared, node, true, &list);
2353 spin_unlock_irq(&n->list_lock);
2354
2355 slabs_destroy(cachep, &list);
2356 }
2357 }
2358
2359 /*
2360 * Remove slabs from the list of free slabs.
2361 * Specify the number of slabs to drain in tofree.
2362 *
2363 * Returns the actual number of slabs released.
2364 */
2365 static int drain_freelist(struct kmem_cache *cache,
2366 struct kmem_cache_node *n, int tofree)
2367 {
2368 struct list_head *p;
2369 int nr_freed;
2370 struct page *page;
2371
2372 nr_freed = 0;
2373 while (nr_freed < tofree && !list_empty(&n->slabs_free)) {
2374
2375 spin_lock_irq(&n->list_lock);
2376 p = n->slabs_free.prev;
2377 if (p == &n->slabs_free) {
2378 spin_unlock_irq(&n->list_lock);
2379 goto out;
2380 }
2381
2382 page = list_entry(p, struct page, lru);
2383 list_del(&page->lru);
2384 /*
2385 * Safe to drop the lock. The slab is no longer linked
2386 * to the cache.
2387 */
2388 n->free_objects -= cache->num;
2389 spin_unlock_irq(&n->list_lock);
2390 slab_destroy(cache, page);
2391 nr_freed++;
2392 }
2393 out:
2394 return nr_freed;
2395 }
2396
2397 int __kmem_cache_shrink(struct kmem_cache *cachep, bool deactivate)
2398 {
2399 int ret = 0;
2400 int node;
2401 struct kmem_cache_node *n;
2402
2403 drain_cpu_caches(cachep);
2404
2405 check_irq_on();
2406 for_each_kmem_cache_node(cachep, node, n) {
2407 drain_freelist(cachep, n, INT_MAX);
2408
2409 ret += !list_empty(&n->slabs_full) ||
2410 !list_empty(&n->slabs_partial);
2411 }
2412 return (ret ? 1 : 0);
2413 }
2414
2415 int __kmem_cache_shutdown(struct kmem_cache *cachep)
2416 {
2417 return __kmem_cache_shrink(cachep, false);
2418 }
2419
2420 void __kmem_cache_release(struct kmem_cache *cachep)
2421 {
2422 int i;
2423 struct kmem_cache_node *n;
2424
2425 cache_random_seq_destroy(cachep);
2426
2427 free_percpu(cachep->cpu_cache);
2428
2429 /* NUMA: free the node structures */
2430 for_each_kmem_cache_node(cachep, i, n) {
2431 kfree(n->shared);
2432 free_alien_cache(n->alien);
2433 kfree(n);
2434 cachep->node[i] = NULL;
2435 }
2436 }
2437
2438 /*
2439 * Get the memory for a slab management obj.
2440 *
2441 * For a slab cache when the slab descriptor is off-slab, the
2442 * slab descriptor can't come from the same cache which is being created,
2443 * Because if it is the case, that means we defer the creation of
2444 * the kmalloc_{dma,}_cache of size sizeof(slab descriptor) to this point.
2445 * And we eventually call down to __kmem_cache_create(), which
2446 * in turn looks up in the kmalloc_{dma,}_caches for the disired-size one.
2447 * This is a "chicken-and-egg" problem.
2448 *
2449 * So the off-slab slab descriptor shall come from the kmalloc_{dma,}_caches,
2450 * which are all initialized during kmem_cache_init().
2451 */
2452 static void *alloc_slabmgmt(struct kmem_cache *cachep,
2453 struct page *page, int colour_off,
2454 gfp_t local_flags, int nodeid)
2455 {
2456 void *freelist;
2457 void *addr = page_address(page);
2458
2459 page->s_mem = addr + colour_off;
2460 page->active = 0;
2461
2462 if (OBJFREELIST_SLAB(cachep))
2463 freelist = NULL;
2464 else if (OFF_SLAB(cachep)) {
2465 /* Slab management obj is off-slab. */
2466 freelist = kmem_cache_alloc_node(cachep->freelist_cache,
2467 local_flags, nodeid);
2468 if (!freelist)
2469 return NULL;
2470 } else {
2471 /* We will use last bytes at the slab for freelist */
2472 freelist = addr + (PAGE_SIZE << cachep->gfporder) -
2473 cachep->freelist_size;
2474 }
2475
2476 return freelist;
2477 }
2478
2479 static inline freelist_idx_t get_free_obj(struct page *page, unsigned int idx)
2480 {
2481 return ((freelist_idx_t *)page->freelist)[idx];
2482 }
2483
2484 static inline void set_free_obj(struct page *page,
2485 unsigned int idx, freelist_idx_t val)
2486 {
2487 ((freelist_idx_t *)(page->freelist))[idx] = val;
2488 }
2489
2490 static void cache_init_objs_debug(struct kmem_cache *cachep, struct page *page)
2491 {
2492 #if DEBUG
2493 int i;
2494
2495 for (i = 0; i < cachep->num; i++) {
2496 void *objp = index_to_obj(cachep, page, i);
2497
2498 if (cachep->flags & SLAB_STORE_USER)
2499 *dbg_userword(cachep, objp) = NULL;
2500
2501 if (cachep->flags & SLAB_RED_ZONE) {
2502 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2503 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2504 }
2505 /*
2506 * Constructors are not allowed to allocate memory from the same
2507 * cache which they are a constructor for. Otherwise, deadlock.
2508 * They must also be threaded.
2509 */
2510 if (cachep->ctor && !(cachep->flags & SLAB_POISON)) {
2511 kasan_unpoison_object_data(cachep,
2512 objp + obj_offset(cachep));
2513 cachep->ctor(objp + obj_offset(cachep));
2514 kasan_poison_object_data(
2515 cachep, objp + obj_offset(cachep));
2516 }
2517
2518 if (cachep->flags & SLAB_RED_ZONE) {
2519 if (*dbg_redzone2(cachep, objp) != RED_INACTIVE)
2520 slab_error(cachep, "constructor overwrote the end of an object");
2521 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE)
2522 slab_error(cachep, "constructor overwrote the start of an object");
2523 }
2524 /* need to poison the objs? */
2525 if (cachep->flags & SLAB_POISON) {
2526 poison_obj(cachep, objp, POISON_FREE);
2527 slab_kernel_map(cachep, objp, 0, 0);
2528 }
2529 }
2530 #endif
2531 }
2532
2533 #ifdef CONFIG_SLAB_FREELIST_RANDOM
2534 /* Hold information during a freelist initialization */
2535 union freelist_init_state {
2536 struct {
2537 unsigned int pos;
2538 freelist_idx_t *list;
2539 unsigned int count;
2540 unsigned int rand;
2541 };
2542 struct rnd_state rnd_state;
2543 };
2544
2545 /*
2546 * Initialize the state based on the randomization methode available.
2547 * return true if the pre-computed list is available, false otherwize.
2548 */
2549 static bool freelist_state_initialize(union freelist_init_state *state,
2550 struct kmem_cache *cachep,
2551 unsigned int count)
2552 {
2553 bool ret;
2554 unsigned int rand;
2555
2556 /* Use best entropy available to define a random shift */
2557 get_random_bytes_arch(&rand, sizeof(rand));
2558
2559 /* Use a random state if the pre-computed list is not available */
2560 if (!cachep->random_seq) {
2561 prandom_seed_state(&state->rnd_state, rand);
2562 ret = false;
2563 } else {
2564 state->list = cachep->random_seq;
2565 state->count = count;
2566 state->pos = 0;
2567 state->rand = rand;
2568 ret = true;
2569 }
2570 return ret;
2571 }
2572
2573 /* Get the next entry on the list and randomize it using a random shift */
2574 static freelist_idx_t next_random_slot(union freelist_init_state *state)
2575 {
2576 return (state->list[state->pos++] + state->rand) % state->count;
2577 }
2578
2579 /*
2580 * Shuffle the freelist initialization state based on pre-computed lists.
2581 * return true if the list was successfully shuffled, false otherwise.
2582 */
2583 static bool shuffle_freelist(struct kmem_cache *cachep, struct page *page)
2584 {
2585 unsigned int objfreelist = 0, i, count = cachep->num;
2586 union freelist_init_state state;
2587 bool precomputed;
2588
2589 if (count < 2)
2590 return false;
2591
2592 precomputed = freelist_state_initialize(&state, cachep, count);
2593
2594 /* Take a random entry as the objfreelist */
2595 if (OBJFREELIST_SLAB(cachep)) {
2596 if (!precomputed)
2597 objfreelist = count - 1;
2598 else
2599 objfreelist = next_random_slot(&state);
2600 page->freelist = index_to_obj(cachep, page, objfreelist) +
2601 obj_offset(cachep);
2602 count--;
2603 }
2604
2605 /*
2606 * On early boot, generate the list dynamically.
2607 * Later use a pre-computed list for speed.
2608 */
2609 if (!precomputed) {
2610 freelist_randomize(&state.rnd_state, page->freelist, count);
2611 } else {
2612 for (i = 0; i < count; i++)
2613 set_free_obj(page, i, next_random_slot(&state));
2614 }
2615
2616 if (OBJFREELIST_SLAB(cachep))
2617 set_free_obj(page, cachep->num - 1, objfreelist);
2618
2619 return true;
2620 }
2621 #else
2622 static inline bool shuffle_freelist(struct kmem_cache *cachep,
2623 struct page *page)
2624 {
2625 return false;
2626 }
2627 #endif /* CONFIG_SLAB_FREELIST_RANDOM */
2628
2629 static void cache_init_objs(struct kmem_cache *cachep,
2630 struct page *page)
2631 {
2632 int i;
2633 void *objp;
2634 bool shuffled;
2635
2636 cache_init_objs_debug(cachep, page);
2637
2638 /* Try to randomize the freelist if enabled */
2639 shuffled = shuffle_freelist(cachep, page);
2640
2641 if (!shuffled && OBJFREELIST_SLAB(cachep)) {
2642 page->freelist = index_to_obj(cachep, page, cachep->num - 1) +
2643 obj_offset(cachep);
2644 }
2645
2646 for (i = 0; i < cachep->num; i++) {
2647 /* constructor could break poison info */
2648 if (DEBUG == 0 && cachep->ctor) {
2649 objp = index_to_obj(cachep, page, i);
2650 kasan_unpoison_object_data(cachep, objp);
2651 cachep->ctor(objp);
2652 kasan_poison_object_data(cachep, objp);
2653 }
2654
2655 if (!shuffled)
2656 set_free_obj(page, i, i);
2657 }
2658 }
2659
2660 static void *slab_get_obj(struct kmem_cache *cachep, struct page *page)
2661 {
2662 void *objp;
2663
2664 objp = index_to_obj(cachep, page, get_free_obj(page, page->active));
2665 page->active++;
2666
2667 #if DEBUG
2668 if (cachep->flags & SLAB_STORE_USER)
2669 set_store_user_dirty(cachep);
2670 #endif
2671
2672 return objp;
2673 }
2674
2675 static void slab_put_obj(struct kmem_cache *cachep,
2676 struct page *page, void *objp)
2677 {
2678 unsigned int objnr = obj_to_index(cachep, page, objp);
2679 #if DEBUG
2680 unsigned int i;
2681
2682 /* Verify double free bug */
2683 for (i = page->active; i < cachep->num; i++) {
2684 if (get_free_obj(page, i) == objnr) {
2685 pr_err("slab: double free detected in cache '%s', objp %p\n",
2686 cachep->name, objp);
2687 BUG();
2688 }
2689 }
2690 #endif
2691 page->active--;
2692 if (!page->freelist)
2693 page->freelist = objp + obj_offset(cachep);
2694
2695 set_free_obj(page, page->active, objnr);
2696 }
2697
2698 /*
2699 * Map pages beginning at addr to the given cache and slab. This is required
2700 * for the slab allocator to be able to lookup the cache and slab of a
2701 * virtual address for kfree, ksize, and slab debugging.
2702 */
2703 static void slab_map_pages(struct kmem_cache *cache, struct page *page,
2704 void *freelist)
2705 {
2706 page->slab_cache = cache;
2707 page->freelist = freelist;
2708 }
2709
2710 /*
2711 * Grow (by 1) the number of slabs within a cache. This is called by
2712 * kmem_cache_alloc() when there are no active objs left in a cache.
2713 */
2714 static struct page *cache_grow_begin(struct kmem_cache *cachep,
2715 gfp_t flags, int nodeid)
2716 {
2717 void *freelist;
2718 size_t offset;
2719 gfp_t local_flags;
2720 int page_node;
2721 struct kmem_cache_node *n;
2722 struct page *page;
2723
2724 /*
2725 * Be lazy and only check for valid flags here, keeping it out of the
2726 * critical path in kmem_cache_alloc().
2727 */
2728 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
2729 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
2730 BUG();
2731 }
2732 local_flags = flags & (GFP_CONSTRAINT_MASK|GFP_RECLAIM_MASK);
2733
2734 check_irq_off();
2735 if (gfpflags_allow_blocking(local_flags))
2736 local_irq_enable();
2737
2738 /*
2739 * Get mem for the objs. Attempt to allocate a physical page from
2740 * 'nodeid'.
2741 */
2742 page = kmem_getpages(cachep, local_flags, nodeid);
2743 if (!page)
2744 goto failed;
2745
2746 page_node = page_to_nid(page);
2747 n = get_node(cachep, page_node);
2748
2749 /* Get colour for the slab, and cal the next value. */
2750 n->colour_next++;
2751 if (n->colour_next >= cachep->colour)
2752 n->colour_next = 0;
2753
2754 offset = n->colour_next;
2755 if (offset >= cachep->colour)
2756 offset = 0;
2757
2758 offset *= cachep->colour_off;
2759
2760 /* Get slab management. */
2761 freelist = alloc_slabmgmt(cachep, page, offset,
2762 local_flags & ~GFP_CONSTRAINT_MASK, page_node);
2763 if (OFF_SLAB(cachep) && !freelist)
2764 goto opps1;
2765
2766 slab_map_pages(cachep, page, freelist);
2767
2768 kasan_poison_slab(page);
2769 cache_init_objs(cachep, page);
2770
2771 if (gfpflags_allow_blocking(local_flags))
2772 local_irq_disable();
2773
2774 return page;
2775
2776 opps1:
2777 kmem_freepages(cachep, page);
2778 failed:
2779 if (gfpflags_allow_blocking(local_flags))
2780 local_irq_disable();
2781 return NULL;
2782 }
2783
2784 static void cache_grow_end(struct kmem_cache *cachep, struct page *page)
2785 {
2786 struct kmem_cache_node *n;
2787 void *list = NULL;
2788
2789 check_irq_off();
2790
2791 if (!page)
2792 return;
2793
2794 INIT_LIST_HEAD(&page->lru);
2795 n = get_node(cachep, page_to_nid(page));
2796
2797 spin_lock(&n->list_lock);
2798 if (!page->active)
2799 list_add_tail(&page->lru, &(n->slabs_free));
2800 else
2801 fixup_slab_list(cachep, n, page, &list);
2802 STATS_INC_GROWN(cachep);
2803 n->free_objects += cachep->num - page->active;
2804 spin_unlock(&n->list_lock);
2805
2806 fixup_objfreelist_debug(cachep, &list);
2807 }
2808
2809 #if DEBUG
2810
2811 /*
2812 * Perform extra freeing checks:
2813 * - detect bad pointers.
2814 * - POISON/RED_ZONE checking
2815 */
2816 static void kfree_debugcheck(const void *objp)
2817 {
2818 if (!virt_addr_valid(objp)) {
2819 pr_err("kfree_debugcheck: out of range ptr %lxh\n",
2820 (unsigned long)objp);
2821 BUG();
2822 }
2823 }
2824
2825 static inline void verify_redzone_free(struct kmem_cache *cache, void *obj)
2826 {
2827 unsigned long long redzone1, redzone2;
2828
2829 redzone1 = *dbg_redzone1(cache, obj);
2830 redzone2 = *dbg_redzone2(cache, obj);
2831
2832 /*
2833 * Redzone is ok.
2834 */
2835 if (redzone1 == RED_ACTIVE && redzone2 == RED_ACTIVE)
2836 return;
2837
2838 if (redzone1 == RED_INACTIVE && redzone2 == RED_INACTIVE)
2839 slab_error(cache, "double free detected");
2840 else
2841 slab_error(cache, "memory outside object was overwritten");
2842
2843 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
2844 obj, redzone1, redzone2);
2845 }
2846
2847 static void *cache_free_debugcheck(struct kmem_cache *cachep, void *objp,
2848 unsigned long caller)
2849 {
2850 unsigned int objnr;
2851 struct page *page;
2852
2853 BUG_ON(virt_to_cache(objp) != cachep);
2854
2855 objp -= obj_offset(cachep);
2856 kfree_debugcheck(objp);
2857 page = virt_to_head_page(objp);
2858
2859 if (cachep->flags & SLAB_RED_ZONE) {
2860 verify_redzone_free(cachep, objp);
2861 *dbg_redzone1(cachep, objp) = RED_INACTIVE;
2862 *dbg_redzone2(cachep, objp) = RED_INACTIVE;
2863 }
2864 if (cachep->flags & SLAB_STORE_USER) {
2865 set_store_user_dirty(cachep);
2866 *dbg_userword(cachep, objp) = (void *)caller;
2867 }
2868
2869 objnr = obj_to_index(cachep, page, objp);
2870
2871 BUG_ON(objnr >= cachep->num);
2872 BUG_ON(objp != index_to_obj(cachep, page, objnr));
2873
2874 if (cachep->flags & SLAB_POISON) {
2875 poison_obj(cachep, objp, POISON_FREE);
2876 slab_kernel_map(cachep, objp, 0, caller);
2877 }
2878 return objp;
2879 }
2880
2881 #else
2882 #define kfree_debugcheck(x) do { } while(0)
2883 #define cache_free_debugcheck(x,objp,z) (objp)
2884 #endif
2885
2886 static inline void fixup_objfreelist_debug(struct kmem_cache *cachep,
2887 void **list)
2888 {
2889 #if DEBUG
2890 void *next = *list;
2891 void *objp;
2892
2893 while (next) {
2894 objp = next - obj_offset(cachep);
2895 next = *(void **)next;
2896 poison_obj(cachep, objp, POISON_FREE);
2897 }
2898 #endif
2899 }
2900
2901 static inline void fixup_slab_list(struct kmem_cache *cachep,
2902 struct kmem_cache_node *n, struct page *page,
2903 void **list)
2904 {
2905 /* move slabp to correct slabp list: */
2906 list_del(&page->lru);
2907 if (page->active == cachep->num) {
2908 list_add(&page->lru, &n->slabs_full);
2909 if (OBJFREELIST_SLAB(cachep)) {
2910 #if DEBUG
2911 /* Poisoning will be done without holding the lock */
2912 if (cachep->flags & SLAB_POISON) {
2913 void **objp = page->freelist;
2914
2915 *objp = *list;
2916 *list = objp;
2917 }
2918 #endif
2919 page->freelist = NULL;
2920 }
2921 } else
2922 list_add(&page->lru, &n->slabs_partial);
2923 }
2924
2925 /* Try to find non-pfmemalloc slab if needed */
2926 static noinline struct page *get_valid_first_slab(struct kmem_cache_node *n,
2927 struct page *page, bool pfmemalloc)
2928 {
2929 if (!page)
2930 return NULL;
2931
2932 if (pfmemalloc)
2933 return page;
2934
2935 if (!PageSlabPfmemalloc(page))
2936 return page;
2937
2938 /* No need to keep pfmemalloc slab if we have enough free objects */
2939 if (n->free_objects > n->free_limit) {
2940 ClearPageSlabPfmemalloc(page);
2941 return page;
2942 }
2943
2944 /* Move pfmemalloc slab to the end of list to speed up next search */
2945 list_del(&page->lru);
2946 if (!page->active)
2947 list_add_tail(&page->lru, &n->slabs_free);
2948 else
2949 list_add_tail(&page->lru, &n->slabs_partial);
2950
2951 list_for_each_entry(page, &n->slabs_partial, lru) {
2952 if (!PageSlabPfmemalloc(page))
2953 return page;
2954 }
2955
2956 list_for_each_entry(page, &n->slabs_free, lru) {
2957 if (!PageSlabPfmemalloc(page))
2958 return page;
2959 }
2960
2961 return NULL;
2962 }
2963
2964 static struct page *get_first_slab(struct kmem_cache_node *n, bool pfmemalloc)
2965 {
2966 struct page *page;
2967
2968 page = list_first_entry_or_null(&n->slabs_partial,
2969 struct page, lru);
2970 if (!page) {
2971 n->free_touched = 1;
2972 page = list_first_entry_or_null(&n->slabs_free,
2973 struct page, lru);
2974 }
2975
2976 if (sk_memalloc_socks())
2977 return get_valid_first_slab(n, page, pfmemalloc);
2978
2979 return page;
2980 }
2981
2982 static noinline void *cache_alloc_pfmemalloc(struct kmem_cache *cachep,
2983 struct kmem_cache_node *n, gfp_t flags)
2984 {
2985 struct page *page;
2986 void *obj;
2987 void *list = NULL;
2988
2989 if (!gfp_pfmemalloc_allowed(flags))
2990 return NULL;
2991
2992 spin_lock(&n->list_lock);
2993 page = get_first_slab(n, true);
2994 if (!page) {
2995 spin_unlock(&n->list_lock);
2996 return NULL;
2997 }
2998
2999 obj = slab_get_obj(cachep, page);
3000 n->free_objects--;
3001
3002 fixup_slab_list(cachep, n, page, &list);
3003
3004 spin_unlock(&n->list_lock);
3005 fixup_objfreelist_debug(cachep, &list);
3006
3007 return obj;
3008 }
3009
3010 /*
3011 * Slab list should be fixed up by fixup_slab_list() for existing slab
3012 * or cache_grow_end() for new slab
3013 */
3014 static __always_inline int alloc_block(struct kmem_cache *cachep,
3015 struct array_cache *ac, struct page *page, int batchcount)
3016 {
3017 /*
3018 * There must be at least one object available for
3019 * allocation.
3020 */
3021 BUG_ON(page->active >= cachep->num);
3022
3023 while (page->active < cachep->num && batchcount--) {
3024 STATS_INC_ALLOCED(cachep);
3025 STATS_INC_ACTIVE(cachep);
3026 STATS_SET_HIGH(cachep);
3027
3028 ac->entry[ac->avail++] = slab_get_obj(cachep, page);
3029 }
3030
3031 return batchcount;
3032 }
3033
3034 static void *cache_alloc_refill(struct kmem_cache *cachep, gfp_t flags)
3035 {
3036 int batchcount;
3037 struct kmem_cache_node *n;
3038 struct array_cache *ac, *shared;
3039 int node;
3040 void *list = NULL;
3041 struct page *page;
3042
3043 check_irq_off();
3044 node = numa_mem_id();
3045
3046 ac = cpu_cache_get(cachep);
3047 batchcount = ac->batchcount;
3048 if (!ac->touched && batchcount > BATCHREFILL_LIMIT) {
3049 /*
3050 * If there was little recent activity on this cache, then
3051 * perform only a partial refill. Otherwise we could generate
3052 * refill bouncing.
3053 */
3054 batchcount = BATCHREFILL_LIMIT;
3055 }
3056 n = get_node(cachep, node);
3057
3058 BUG_ON(ac->avail > 0 || !n);
3059 shared = READ_ONCE(n->shared);
3060 if (!n->free_objects && (!shared || !shared->avail))
3061 goto direct_grow;
3062
3063 spin_lock(&n->list_lock);
3064 shared = READ_ONCE(n->shared);
3065
3066 /* See if we can refill from the shared array */
3067 if (shared && transfer_objects(ac, shared, batchcount)) {
3068 shared->touched = 1;
3069 goto alloc_done;
3070 }
3071
3072 while (batchcount > 0) {
3073 /* Get slab alloc is to come from. */
3074 page = get_first_slab(n, false);
3075 if (!page)
3076 goto must_grow;
3077
3078 check_spinlock_acquired(cachep);
3079
3080 batchcount = alloc_block(cachep, ac, page, batchcount);
3081 fixup_slab_list(cachep, n, page, &list);
3082 }
3083
3084 must_grow:
3085 n->free_objects -= ac->avail;
3086 alloc_done:
3087 spin_unlock(&n->list_lock);
3088 fixup_objfreelist_debug(cachep, &list);
3089
3090 direct_grow:
3091 if (unlikely(!ac->avail)) {
3092 /* Check if we can use obj in pfmemalloc slab */
3093 if (sk_memalloc_socks()) {
3094 void *obj = cache_alloc_pfmemalloc(cachep, n, flags);
3095
3096 if (obj)
3097 return obj;
3098 }
3099
3100 page = cache_grow_begin(cachep, gfp_exact_node(flags), node);
3101
3102 /*
3103 * cache_grow_begin() can reenable interrupts,
3104 * then ac could change.
3105 */
3106 ac = cpu_cache_get(cachep);
3107 if (!ac->avail && page)
3108 alloc_block(cachep, ac, page, batchcount);
3109 cache_grow_end(cachep, page);
3110
3111 if (!ac->avail)
3112 return NULL;
3113 }
3114 ac->touched = 1;
3115
3116 return ac->entry[--ac->avail];
3117 }
3118
3119 static inline void cache_alloc_debugcheck_before(struct kmem_cache *cachep,
3120 gfp_t flags)
3121 {
3122 might_sleep_if(gfpflags_allow_blocking(flags));
3123 }
3124
3125 #if DEBUG
3126 static void *cache_alloc_debugcheck_after(struct kmem_cache *cachep,
3127 gfp_t flags, void *objp, unsigned long caller)
3128 {
3129 if (!objp)
3130 return objp;
3131 if (cachep->flags & SLAB_POISON) {
3132 check_poison_obj(cachep, objp);
3133 slab_kernel_map(cachep, objp, 1, 0);
3134 poison_obj(cachep, objp, POISON_INUSE);
3135 }
3136 if (cachep->flags & SLAB_STORE_USER)
3137 *dbg_userword(cachep, objp) = (void *)caller;
3138
3139 if (cachep->flags & SLAB_RED_ZONE) {
3140 if (*dbg_redzone1(cachep, objp) != RED_INACTIVE ||
3141 *dbg_redzone2(cachep, objp) != RED_INACTIVE) {
3142 slab_error(cachep, "double free, or memory outside object was overwritten");
3143 pr_err("%p: redzone 1:0x%llx, redzone 2:0x%llx\n",
3144 objp, *dbg_redzone1(cachep, objp),
3145 *dbg_redzone2(cachep, objp));
3146 }
3147 *dbg_redzone1(cachep, objp) = RED_ACTIVE;
3148 *dbg_redzone2(cachep, objp) = RED_ACTIVE;
3149 }
3150
3151 objp += obj_offset(cachep);
3152 if (cachep->ctor && cachep->flags & SLAB_POISON)
3153 cachep->ctor(objp);
3154 if (ARCH_SLAB_MINALIGN &&
3155 ((unsigned long)objp & (ARCH_SLAB_MINALIGN-1))) {
3156 pr_err("0x%p: not aligned to ARCH_SLAB_MINALIGN=%d\n",
3157 objp, (int)ARCH_SLAB_MINALIGN);
3158 }
3159 return objp;
3160 }
3161 #else
3162 #define cache_alloc_debugcheck_after(a,b,objp,d) (objp)
3163 #endif
3164
3165 static inline void *____cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3166 {
3167 void *objp;
3168 struct array_cache *ac;
3169
3170 check_irq_off();
3171
3172 ac = cpu_cache_get(cachep);
3173 if (likely(ac->avail)) {
3174 ac->touched = 1;
3175 objp = ac->entry[--ac->avail];
3176
3177 STATS_INC_ALLOCHIT(cachep);
3178 goto out;
3179 }
3180
3181 STATS_INC_ALLOCMISS(cachep);
3182 objp = cache_alloc_refill(cachep, flags);
3183 /*
3184 * the 'ac' may be updated by cache_alloc_refill(),
3185 * and kmemleak_erase() requires its correct value.
3186 */
3187 ac = cpu_cache_get(cachep);
3188
3189 out:
3190 /*
3191 * To avoid a false negative, if an object that is in one of the
3192 * per-CPU caches is leaked, we need to make sure kmemleak doesn't
3193 * treat the array pointers as a reference to the object.
3194 */
3195 if (objp)
3196 kmemleak_erase(&ac->entry[ac->avail]);
3197 return objp;
3198 }
3199
3200 #ifdef CONFIG_NUMA
3201 /*
3202 * Try allocating on another node if PFA_SPREAD_SLAB is a mempolicy is set.
3203 *
3204 * If we are in_interrupt, then process context, including cpusets and
3205 * mempolicy, may not apply and should not be used for allocation policy.
3206 */
3207 static void *alternate_node_alloc(struct kmem_cache *cachep, gfp_t flags)
3208 {
3209 int nid_alloc, nid_here;
3210
3211 if (in_interrupt() || (flags & __GFP_THISNODE))
3212 return NULL;
3213 nid_alloc = nid_here = numa_mem_id();
3214 if (cpuset_do_slab_mem_spread() && (cachep->flags & SLAB_MEM_SPREAD))
3215 nid_alloc = cpuset_slab_spread_node();
3216 else if (current->mempolicy)
3217 nid_alloc = mempolicy_slab_node();
3218 if (nid_alloc != nid_here)
3219 return ____cache_alloc_node(cachep, flags, nid_alloc);
3220 return NULL;
3221 }
3222
3223 /*
3224 * Fallback function if there was no memory available and no objects on a
3225 * certain node and fall back is permitted. First we scan all the
3226 * available node for available objects. If that fails then we
3227 * perform an allocation without specifying a node. This allows the page
3228 * allocator to do its reclaim / fallback magic. We then insert the
3229 * slab into the proper nodelist and then allocate from it.
3230 */
3231 static void *fallback_alloc(struct kmem_cache *cache, gfp_t flags)
3232 {
3233 struct zonelist *zonelist;
3234 struct zoneref *z;
3235 struct zone *zone;
3236 enum zone_type high_zoneidx = gfp_zone(flags);
3237 void *obj = NULL;
3238 struct page *page;
3239 int nid;
3240 unsigned int cpuset_mems_cookie;
3241
3242 if (flags & __GFP_THISNODE)
3243 return NULL;
3244
3245 retry_cpuset:
3246 cpuset_mems_cookie = read_mems_allowed_begin();
3247 zonelist = node_zonelist(mempolicy_slab_node(), flags);
3248
3249 retry:
3250 /*
3251 * Look through allowed nodes for objects available
3252 * from existing per node queues.
3253 */
3254 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
3255 nid = zone_to_nid(zone);
3256
3257 if (cpuset_zone_allowed(zone, flags) &&
3258 get_node(cache, nid) &&
3259 get_node(cache, nid)->free_objects) {
3260 obj = ____cache_alloc_node(cache,
3261 gfp_exact_node(flags), nid);
3262 if (obj)
3263 break;
3264 }
3265 }
3266
3267 if (!obj) {
3268 /*
3269 * This allocation will be performed within the constraints
3270 * of the current cpuset / memory policy requirements.
3271 * We may trigger various forms of reclaim on the allowed
3272 * set and go into memory reserves if necessary.
3273 */
3274 page = cache_grow_begin(cache, flags, numa_mem_id());
3275 cache_grow_end(cache, page);
3276 if (page) {
3277 nid = page_to_nid(page);
3278 obj = ____cache_alloc_node(cache,
3279 gfp_exact_node(flags), nid);
3280
3281 /*
3282 * Another processor may allocate the objects in
3283 * the slab since we are not holding any locks.
3284 */
3285 if (!obj)
3286 goto retry;
3287 }
3288 }
3289
3290 if (unlikely(!obj && read_mems_allowed_retry(cpuset_mems_cookie)))
3291 goto retry_cpuset;
3292 return obj;
3293 }
3294
3295 /*
3296 * A interface to enable slab creation on nodeid
3297 */
3298 static void *____cache_alloc_node(struct kmem_cache *cachep, gfp_t flags,
3299 int nodeid)
3300 {
3301 struct page *page;
3302 struct kmem_cache_node *n;
3303 void *obj = NULL;
3304 void *list = NULL;
3305
3306 VM_BUG_ON(nodeid < 0 || nodeid >= MAX_NUMNODES);
3307 n = get_node(cachep, nodeid);
3308 BUG_ON(!n);
3309
3310 check_irq_off();
3311 spin_lock(&n->list_lock);
3312 page = get_first_slab(n, false);
3313 if (!page)
3314 goto must_grow;
3315
3316 check_spinlock_acquired_node(cachep, nodeid);
3317
3318 STATS_INC_NODEALLOCS(cachep);
3319 STATS_INC_ACTIVE(cachep);
3320 STATS_SET_HIGH(cachep);
3321
3322 BUG_ON(page->active == cachep->num);
3323
3324 obj = slab_get_obj(cachep, page);
3325 n->free_objects--;
3326
3327 fixup_slab_list(cachep, n, page, &list);
3328
3329 spin_unlock(&n->list_lock);
3330 fixup_objfreelist_debug(cachep, &list);
3331 return obj;
3332
3333 must_grow:
3334 spin_unlock(&n->list_lock);
3335 page = cache_grow_begin(cachep, gfp_exact_node(flags), nodeid);
3336 if (page) {
3337 /* This slab isn't counted yet so don't update free_objects */
3338 obj = slab_get_obj(cachep, page);
3339 }
3340 cache_grow_end(cachep, page);
3341
3342 return obj ? obj : fallback_alloc(cachep, flags);
3343 }
3344
3345 static __always_inline void *
3346 slab_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid,
3347 unsigned long caller)
3348 {
3349 unsigned long save_flags;
3350 void *ptr;
3351 int slab_node = numa_mem_id();
3352
3353 flags &= gfp_allowed_mask;
3354 cachep = slab_pre_alloc_hook(cachep, flags);
3355 if (unlikely(!cachep))
3356 return NULL;
3357
3358 cache_alloc_debugcheck_before(cachep, flags);
3359 local_irq_save(save_flags);
3360
3361 if (nodeid == NUMA_NO_NODE)
3362 nodeid = slab_node;
3363
3364 if (unlikely(!get_node(cachep, nodeid))) {
3365 /* Node not bootstrapped yet */
3366 ptr = fallback_alloc(cachep, flags);
3367 goto out;
3368 }
3369
3370 if (nodeid == slab_node) {
3371 /*
3372 * Use the locally cached objects if possible.
3373 * However ____cache_alloc does not allow fallback
3374 * to other nodes. It may fail while we still have
3375 * objects on other nodes available.
3376 */
3377 ptr = ____cache_alloc(cachep, flags);
3378 if (ptr)
3379 goto out;
3380 }
3381 /* ___cache_alloc_node can fall back to other nodes */
3382 ptr = ____cache_alloc_node(cachep, flags, nodeid);
3383 out:
3384 local_irq_restore(save_flags);
3385 ptr = cache_alloc_debugcheck_after(cachep, flags, ptr, caller);
3386
3387 if (unlikely(flags & __GFP_ZERO) && ptr)
3388 memset(ptr, 0, cachep->object_size);
3389
3390 slab_post_alloc_hook(cachep, flags, 1, &ptr);
3391 return ptr;
3392 }
3393
3394 static __always_inline void *
3395 __do_cache_alloc(struct kmem_cache *cache, gfp_t flags)
3396 {
3397 void *objp;
3398
3399 if (current->mempolicy || cpuset_do_slab_mem_spread()) {
3400 objp = alternate_node_alloc(cache, flags);
3401 if (objp)
3402 goto out;
3403 }
3404 objp = ____cache_alloc(cache, flags);
3405
3406 /*
3407 * We may just have run out of memory on the local node.
3408 * ____cache_alloc_node() knows how to locate memory on other nodes
3409 */
3410 if (!objp)
3411 objp = ____cache_alloc_node(cache, flags, numa_mem_id());
3412
3413 out:
3414 return objp;
3415 }
3416 #else
3417
3418 static __always_inline void *
3419 __do_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3420 {
3421 return ____cache_alloc(cachep, flags);
3422 }
3423
3424 #endif /* CONFIG_NUMA */
3425
3426 static __always_inline void *
3427 slab_alloc(struct kmem_cache *cachep, gfp_t flags, unsigned long caller)
3428 {
3429 unsigned long save_flags;
3430 void *objp;
3431
3432 flags &= gfp_allowed_mask;
3433 cachep = slab_pre_alloc_hook(cachep, flags);
3434 if (unlikely(!cachep))
3435 return NULL;
3436
3437 cache_alloc_debugcheck_before(cachep, flags);
3438 local_irq_save(save_flags);
3439 objp = __do_cache_alloc(cachep, flags);
3440 local_irq_restore(save_flags);
3441 objp = cache_alloc_debugcheck_after(cachep, flags, objp, caller);
3442 prefetchw(objp);
3443
3444 if (unlikely(flags & __GFP_ZERO) && objp)
3445 memset(objp, 0, cachep->object_size);
3446
3447 slab_post_alloc_hook(cachep, flags, 1, &objp);
3448 return objp;
3449 }
3450
3451 /*
3452 * Caller needs to acquire correct kmem_cache_node's list_lock
3453 * @list: List of detached free slabs should be freed by caller
3454 */
3455 static void free_block(struct kmem_cache *cachep, void **objpp,
3456 int nr_objects, int node, struct list_head *list)
3457 {
3458 int i;
3459 struct kmem_cache_node *n = get_node(cachep, node);
3460 struct page *page;
3461
3462 n->free_objects += nr_objects;
3463
3464 for (i = 0; i < nr_objects; i++) {
3465 void *objp;
3466 struct page *page;
3467
3468 objp = objpp[i];
3469
3470 page = virt_to_head_page(objp);
3471 list_del(&page->lru);
3472 check_spinlock_acquired_node(cachep, node);
3473 slab_put_obj(cachep, page, objp);
3474 STATS_DEC_ACTIVE(cachep);
3475
3476 /* fixup slab chains */
3477 if (page->active == 0)
3478 list_add(&page->lru, &n->slabs_free);
3479 else {
3480 /* Unconditionally move a slab to the end of the
3481 * partial list on free - maximum time for the
3482 * other objects to be freed, too.
3483 */
3484 list_add_tail(&page->lru, &n->slabs_partial);
3485 }
3486 }
3487
3488 while (n->free_objects > n->free_limit && !list_empty(&n->slabs_free)) {
3489 n->free_objects -= cachep->num;
3490
3491 page = list_last_entry(&n->slabs_free, struct page, lru);
3492 list_del(&page->lru);
3493 list_add(&page->lru, list);
3494 }
3495 }
3496
3497 static void cache_flusharray(struct kmem_cache *cachep, struct array_cache *ac)
3498 {
3499 int batchcount;
3500 struct kmem_cache_node *n;
3501 int node = numa_mem_id();
3502 LIST_HEAD(list);
3503
3504 batchcount = ac->batchcount;
3505
3506 check_irq_off();
3507 n = get_node(cachep, node);
3508 spin_lock(&n->list_lock);
3509 if (n->shared) {
3510 struct array_cache *shared_array = n->shared;
3511 int max = shared_array->limit - shared_array->avail;
3512 if (max) {
3513 if (batchcount > max)
3514 batchcount = max;
3515 memcpy(&(shared_array->entry[shared_array->avail]),
3516 ac->entry, sizeof(void *) * batchcount);
3517 shared_array->avail += batchcount;
3518 goto free_done;
3519 }
3520 }
3521
3522 free_block(cachep, ac->entry, batchcount, node, &list);
3523 free_done:
3524 #if STATS
3525 {
3526 int i = 0;
3527 struct page *page;
3528
3529 list_for_each_entry(page, &n->slabs_free, lru) {
3530 BUG_ON(page->active);
3531
3532 i++;
3533 }
3534 STATS_SET_FREEABLE(cachep, i);
3535 }
3536 #endif
3537 spin_unlock(&n->list_lock);
3538 slabs_destroy(cachep, &list);
3539 ac->avail -= batchcount;
3540 memmove(ac->entry, &(ac->entry[batchcount]), sizeof(void *)*ac->avail);
3541 }
3542
3543 /*
3544 * Release an obj back to its cache. If the obj has a constructed state, it must
3545 * be in this state _before_ it is released. Called with disabled ints.
3546 */
3547 static inline void __cache_free(struct kmem_cache *cachep, void *objp,
3548 unsigned long caller)
3549 {
3550 /* Put the object into the quarantine, don't touch it for now. */
3551 if (kasan_slab_free(cachep, objp))
3552 return;
3553
3554 ___cache_free(cachep, objp, caller);
3555 }
3556
3557 void ___cache_free(struct kmem_cache *cachep, void *objp,
3558 unsigned long caller)
3559 {
3560 struct array_cache *ac = cpu_cache_get(cachep);
3561
3562 check_irq_off();
3563 kmemleak_free_recursive(objp, cachep->flags);
3564 objp = cache_free_debugcheck(cachep, objp, caller);
3565
3566 kmemcheck_slab_free(cachep, objp, cachep->object_size);
3567
3568 /*
3569 * Skip calling cache_free_alien() when the platform is not numa.
3570 * This will avoid cache misses that happen while accessing slabp (which
3571 * is per page memory reference) to get nodeid. Instead use a global
3572 * variable to skip the call, which is mostly likely to be present in
3573 * the cache.
3574 */
3575 if (nr_online_nodes > 1 && cache_free_alien(cachep, objp))
3576 return;
3577
3578 if (ac->avail < ac->limit) {
3579 STATS_INC_FREEHIT(cachep);
3580 } else {
3581 STATS_INC_FREEMISS(cachep);
3582 cache_flusharray(cachep, ac);
3583 }
3584
3585 if (sk_memalloc_socks()) {
3586 struct page *page = virt_to_head_page(objp);
3587
3588 if (unlikely(PageSlabPfmemalloc(page))) {
3589 cache_free_pfmemalloc(cachep, page, objp);
3590 return;
3591 }
3592 }
3593
3594 ac->entry[ac->avail++] = objp;
3595 }
3596
3597 /**
3598 * kmem_cache_alloc - Allocate an object
3599 * @cachep: The cache to allocate from.
3600 * @flags: See kmalloc().
3601 *
3602 * Allocate an object from this cache. The flags are only relevant
3603 * if the cache has no available objects.
3604 */
3605 void *kmem_cache_alloc(struct kmem_cache *cachep, gfp_t flags)
3606 {
3607 void *ret = slab_alloc(cachep, flags, _RET_IP_);
3608
3609 kasan_slab_alloc(cachep, ret, flags);
3610 trace_kmem_cache_alloc(_RET_IP_, ret,
3611 cachep->object_size, cachep->size, flags);
3612
3613 return ret;
3614 }
3615 EXPORT_SYMBOL(kmem_cache_alloc);
3616
3617 static __always_inline void
3618 cache_alloc_debugcheck_after_bulk(struct kmem_cache *s, gfp_t flags,
3619 size_t size, void **p, unsigned long caller)
3620 {
3621 size_t i;
3622
3623 for (i = 0; i < size; i++)
3624 p[i] = cache_alloc_debugcheck_after(s, flags, p[i], caller);
3625 }
3626
3627 int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3628 void **p)
3629 {
3630 size_t i;
3631
3632 s = slab_pre_alloc_hook(s, flags);
3633 if (!s)
3634 return 0;
3635
3636 cache_alloc_debugcheck_before(s, flags);
3637
3638 local_irq_disable();
3639 for (i = 0; i < size; i++) {
3640 void *objp = __do_cache_alloc(s, flags);
3641
3642 if (unlikely(!objp))
3643 goto error;
3644 p[i] = objp;
3645 }
3646 local_irq_enable();
3647
3648 cache_alloc_debugcheck_after_bulk(s, flags, size, p, _RET_IP_);
3649
3650 /* Clear memory outside IRQ disabled section */
3651 if (unlikely(flags & __GFP_ZERO))
3652 for (i = 0; i < size; i++)
3653 memset(p[i], 0, s->object_size);
3654
3655 slab_post_alloc_hook(s, flags, size, p);
3656 /* FIXME: Trace call missing. Christoph would like a bulk variant */
3657 return size;
3658 error:
3659 local_irq_enable();
3660 cache_alloc_debugcheck_after_bulk(s, flags, i, p, _RET_IP_);
3661 slab_post_alloc_hook(s, flags, i, p);
3662 __kmem_cache_free_bulk(s, i, p);
3663 return 0;
3664 }
3665 EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3666
3667 #ifdef CONFIG_TRACING
3668 void *
3669 kmem_cache_alloc_trace(struct kmem_cache *cachep, gfp_t flags, size_t size)
3670 {
3671 void *ret;
3672
3673 ret = slab_alloc(cachep, flags, _RET_IP_);
3674
3675 kasan_kmalloc(cachep, ret, size, flags);
3676 trace_kmalloc(_RET_IP_, ret,
3677 size, cachep->size, flags);
3678 return ret;
3679 }
3680 EXPORT_SYMBOL(kmem_cache_alloc_trace);
3681 #endif
3682
3683 #ifdef CONFIG_NUMA
3684 /**
3685 * kmem_cache_alloc_node - Allocate an object on the specified node
3686 * @cachep: The cache to allocate from.
3687 * @flags: See kmalloc().
3688 * @nodeid: node number of the target node.
3689 *
3690 * Identical to kmem_cache_alloc but it will allocate memory on the given
3691 * node, which can improve the performance for cpu bound structures.
3692 *
3693 * Fallback to other node is possible if __GFP_THISNODE is not set.
3694 */
3695 void *kmem_cache_alloc_node(struct kmem_cache *cachep, gfp_t flags, int nodeid)
3696 {
3697 void *ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3698
3699 kasan_slab_alloc(cachep, ret, flags);
3700 trace_kmem_cache_alloc_node(_RET_IP_, ret,
3701 cachep->object_size, cachep->size,
3702 flags, nodeid);
3703
3704 return ret;
3705 }
3706 EXPORT_SYMBOL(kmem_cache_alloc_node);
3707
3708 #ifdef CONFIG_TRACING
3709 void *kmem_cache_alloc_node_trace(struct kmem_cache *cachep,
3710 gfp_t flags,
3711 int nodeid,
3712 size_t size)
3713 {
3714 void *ret;
3715
3716 ret = slab_alloc_node(cachep, flags, nodeid, _RET_IP_);
3717
3718 kasan_kmalloc(cachep, ret, size, flags);
3719 trace_kmalloc_node(_RET_IP_, ret,
3720 size, cachep->size,
3721 flags, nodeid);
3722 return ret;
3723 }
3724 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3725 #endif
3726
3727 static __always_inline void *
3728 __do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
3729 {
3730 struct kmem_cache *cachep;
3731 void *ret;
3732
3733 cachep = kmalloc_slab(size, flags);
3734 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3735 return cachep;
3736 ret = kmem_cache_alloc_node_trace(cachep, flags, node, size);
3737 kasan_kmalloc(cachep, ret, size, flags);
3738
3739 return ret;
3740 }
3741
3742 void *__kmalloc_node(size_t size, gfp_t flags, int node)
3743 {
3744 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3745 }
3746 EXPORT_SYMBOL(__kmalloc_node);
3747
3748 void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3749 int node, unsigned long caller)
3750 {
3751 return __do_kmalloc_node(size, flags, node, caller);
3752 }
3753 EXPORT_SYMBOL(__kmalloc_node_track_caller);
3754 #endif /* CONFIG_NUMA */
3755
3756 /**
3757 * __do_kmalloc - allocate memory
3758 * @size: how many bytes of memory are required.
3759 * @flags: the type of memory to allocate (see kmalloc).
3760 * @caller: function caller for debug tracking of the caller
3761 */
3762 static __always_inline void *__do_kmalloc(size_t size, gfp_t flags,
3763 unsigned long caller)
3764 {
3765 struct kmem_cache *cachep;
3766 void *ret;
3767
3768 cachep = kmalloc_slab(size, flags);
3769 if (unlikely(ZERO_OR_NULL_PTR(cachep)))
3770 return cachep;
3771 ret = slab_alloc(cachep, flags, caller);
3772
3773 kasan_kmalloc(cachep, ret, size, flags);
3774 trace_kmalloc(caller, ret,
3775 size, cachep->size, flags);
3776
3777 return ret;
3778 }
3779
3780 void *__kmalloc(size_t size, gfp_t flags)
3781 {
3782 return __do_kmalloc(size, flags, _RET_IP_);
3783 }
3784 EXPORT_SYMBOL(__kmalloc);
3785
3786 void *__kmalloc_track_caller(size_t size, gfp_t flags, unsigned long caller)
3787 {
3788 return __do_kmalloc(size, flags, caller);
3789 }
3790 EXPORT_SYMBOL(__kmalloc_track_caller);
3791
3792 /**
3793 * kmem_cache_free - Deallocate an object
3794 * @cachep: The cache the allocation was from.
3795 * @objp: The previously allocated object.
3796 *
3797 * Free an object which was previously allocated from this
3798 * cache.
3799 */
3800 void kmem_cache_free(struct kmem_cache *cachep, void *objp)
3801 {
3802 unsigned long flags;
3803 cachep = cache_from_obj(cachep, objp);
3804 if (!cachep)
3805 return;
3806
3807 local_irq_save(flags);
3808 debug_check_no_locks_freed(objp, cachep->object_size);
3809 if (!(cachep->flags & SLAB_DEBUG_OBJECTS))
3810 debug_check_no_obj_freed(objp, cachep->object_size);
3811 __cache_free(cachep, objp, _RET_IP_);
3812 local_irq_restore(flags);
3813
3814 trace_kmem_cache_free(_RET_IP_, objp);
3815 }
3816 EXPORT_SYMBOL(kmem_cache_free);
3817
3818 void kmem_cache_free_bulk(struct kmem_cache *orig_s, size_t size, void **p)
3819 {
3820 struct kmem_cache *s;
3821 size_t i;
3822
3823 local_irq_disable();
3824 for (i = 0; i < size; i++) {
3825 void *objp = p[i];
3826
3827 if (!orig_s) /* called via kfree_bulk */
3828 s = virt_to_cache(objp);
3829 else
3830 s = cache_from_obj(orig_s, objp);
3831
3832 debug_check_no_locks_freed(objp, s->object_size);
3833 if (!(s->flags & SLAB_DEBUG_OBJECTS))
3834 debug_check_no_obj_freed(objp, s->object_size);
3835
3836 __cache_free(s, objp, _RET_IP_);
3837 }
3838 local_irq_enable();
3839
3840 /* FIXME: add tracing */
3841 }
3842 EXPORT_SYMBOL(kmem_cache_free_bulk);
3843
3844 /**
3845 * kfree - free previously allocated memory
3846 * @objp: pointer returned by kmalloc.
3847 *
3848 * If @objp is NULL, no operation is performed.
3849 *
3850 * Don't free memory not originally allocated by kmalloc()
3851 * or you will run into trouble.
3852 */
3853 void kfree(const void *objp)
3854 {
3855 struct kmem_cache *c;
3856 unsigned long flags;
3857
3858 trace_kfree(_RET_IP_, objp);
3859
3860 if (unlikely(ZERO_OR_NULL_PTR(objp)))
3861 return;
3862 local_irq_save(flags);
3863 kfree_debugcheck(objp);
3864 c = virt_to_cache(objp);
3865 debug_check_no_locks_freed(objp, c->object_size);
3866
3867 debug_check_no_obj_freed(objp, c->object_size);
3868 __cache_free(c, (void *)objp, _RET_IP_);
3869 local_irq_restore(flags);
3870 }
3871 EXPORT_SYMBOL(kfree);
3872
3873 /*
3874 * This initializes kmem_cache_node or resizes various caches for all nodes.
3875 */
3876 static int setup_kmem_cache_nodes(struct kmem_cache *cachep, gfp_t gfp)
3877 {
3878 int ret;
3879 int node;
3880 struct kmem_cache_node *n;
3881
3882 for_each_online_node(node) {
3883 ret = setup_kmem_cache_node(cachep, node, gfp, true);
3884 if (ret)
3885 goto fail;
3886
3887 }
3888
3889 return 0;
3890
3891 fail:
3892 if (!cachep->list.next) {
3893 /* Cache is not active yet. Roll back what we did */
3894 node--;
3895 while (node >= 0) {
3896 n = get_node(cachep, node);
3897 if (n) {
3898 kfree(n->shared);
3899 free_alien_cache(n->alien);
3900 kfree(n);
3901 cachep->node[node] = NULL;
3902 }
3903 node--;
3904 }
3905 }
3906 return -ENOMEM;
3907 }
3908
3909 /* Always called with the slab_mutex held */
3910 static int __do_tune_cpucache(struct kmem_cache *cachep, int limit,
3911 int batchcount, int shared, gfp_t gfp)
3912 {
3913 struct array_cache __percpu *cpu_cache, *prev;
3914 int cpu;
3915
3916 cpu_cache = alloc_kmem_cache_cpus(cachep, limit, batchcount);
3917 if (!cpu_cache)
3918 return -ENOMEM;
3919
3920 prev = cachep->cpu_cache;
3921 cachep->cpu_cache = cpu_cache;
3922 kick_all_cpus_sync();
3923
3924 check_irq_on();
3925 cachep->batchcount = batchcount;
3926 cachep->limit = limit;
3927 cachep->shared = shared;
3928
3929 if (!prev)
3930 goto setup_node;
3931
3932 for_each_online_cpu(cpu) {
3933 LIST_HEAD(list);
3934 int node;
3935 struct kmem_cache_node *n;
3936 struct array_cache *ac = per_cpu_ptr(prev, cpu);
3937
3938 node = cpu_to_mem(cpu);
3939 n = get_node(cachep, node);
3940 spin_lock_irq(&n->list_lock);
3941 free_block(cachep, ac->entry, ac->avail, node, &list);
3942 spin_unlock_irq(&n->list_lock);
3943 slabs_destroy(cachep, &list);
3944 }
3945 free_percpu(prev);
3946
3947 setup_node:
3948 return setup_kmem_cache_nodes(cachep, gfp);
3949 }
3950
3951 static int do_tune_cpucache(struct kmem_cache *cachep, int limit,
3952 int batchcount, int shared, gfp_t gfp)
3953 {
3954 int ret;
3955 struct kmem_cache *c;
3956
3957 ret = __do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
3958
3959 if (slab_state < FULL)
3960 return ret;
3961
3962 if ((ret < 0) || !is_root_cache(cachep))
3963 return ret;
3964
3965 lockdep_assert_held(&slab_mutex);
3966 for_each_memcg_cache(c, cachep) {
3967 /* return value determined by the root cache only */
3968 __do_tune_cpucache(c, limit, batchcount, shared, gfp);
3969 }
3970
3971 return ret;
3972 }
3973
3974 /* Called with slab_mutex held always */
3975 static int enable_cpucache(struct kmem_cache *cachep, gfp_t gfp)
3976 {
3977 int err;
3978 int limit = 0;
3979 int shared = 0;
3980 int batchcount = 0;
3981
3982 err = cache_random_seq_create(cachep, gfp);
3983 if (err)
3984 goto end;
3985
3986 if (!is_root_cache(cachep)) {
3987 struct kmem_cache *root = memcg_root_cache(cachep);
3988 limit = root->limit;
3989 shared = root->shared;
3990 batchcount = root->batchcount;
3991 }
3992
3993 if (limit && shared && batchcount)
3994 goto skip_setup;
3995 /*
3996 * The head array serves three purposes:
3997 * - create a LIFO ordering, i.e. return objects that are cache-warm
3998 * - reduce the number of spinlock operations.
3999 * - reduce the number of linked list operations on the slab and
4000 * bufctl chains: array operations are cheaper.
4001 * The numbers are guessed, we should auto-tune as described by
4002 * Bonwick.
4003 */
4004 if (cachep->size > 131072)
4005 limit = 1;
4006 else if (cachep->size > PAGE_SIZE)
4007 limit = 8;
4008 else if (cachep->size > 1024)
4009 limit = 24;
4010 else if (cachep->size > 256)
4011 limit = 54;
4012 else
4013 limit = 120;
4014
4015 /*
4016 * CPU bound tasks (e.g. network routing) can exhibit cpu bound
4017 * allocation behaviour: Most allocs on one cpu, most free operations
4018 * on another cpu. For these cases, an efficient object passing between
4019 * cpus is necessary. This is provided by a shared array. The array
4020 * replaces Bonwick's magazine layer.
4021 * On uniprocessor, it's functionally equivalent (but less efficient)
4022 * to a larger limit. Thus disabled by default.
4023 */
4024 shared = 0;
4025 if (cachep->size <= PAGE_SIZE && num_possible_cpus() > 1)
4026 shared = 8;
4027
4028 #if DEBUG
4029 /*
4030 * With debugging enabled, large batchcount lead to excessively long
4031 * periods with disabled local interrupts. Limit the batchcount
4032 */
4033 if (limit > 32)
4034 limit = 32;
4035 #endif
4036 batchcount = (limit + 1) / 2;
4037 skip_setup:
4038 err = do_tune_cpucache(cachep, limit, batchcount, shared, gfp);
4039 end:
4040 if (err)
4041 pr_err("enable_cpucache failed for %s, error %d\n",
4042 cachep->name, -err);
4043 return err;
4044 }
4045
4046 /*
4047 * Drain an array if it contains any elements taking the node lock only if
4048 * necessary. Note that the node listlock also protects the array_cache
4049 * if drain_array() is used on the shared array.
4050 */
4051 static void drain_array(struct kmem_cache *cachep, struct kmem_cache_node *n,
4052 struct array_cache *ac, int node)
4053 {
4054 LIST_HEAD(list);
4055
4056 /* ac from n->shared can be freed if we don't hold the slab_mutex. */
4057 check_mutex_acquired();
4058
4059 if (!ac || !ac->avail)
4060 return;
4061
4062 if (ac->touched) {
4063 ac->touched = 0;
4064 return;
4065 }
4066
4067 spin_lock_irq(&n->list_lock);
4068 drain_array_locked(cachep, ac, node, false, &list);
4069 spin_unlock_irq(&n->list_lock);
4070
4071 slabs_destroy(cachep, &list);
4072 }
4073
4074 /**
4075 * cache_reap - Reclaim memory from caches.
4076 * @w: work descriptor
4077 *
4078 * Called from workqueue/eventd every few seconds.
4079 * Purpose:
4080 * - clear the per-cpu caches for this CPU.
4081 * - return freeable pages to the main free memory pool.
4082 *
4083 * If we cannot acquire the cache chain mutex then just give up - we'll try
4084 * again on the next iteration.
4085 */
4086 static void cache_reap(struct work_struct *w)
4087 {
4088 struct kmem_cache *searchp;
4089 struct kmem_cache_node *n;
4090 int node = numa_mem_id();
4091 struct delayed_work *work = to_delayed_work(w);
4092
4093 if (!mutex_trylock(&slab_mutex))
4094 /* Give up. Setup the next iteration. */
4095 goto out;
4096
4097 list_for_each_entry(searchp, &slab_caches, list) {
4098 check_irq_on();
4099
4100 /*
4101 * We only take the node lock if absolutely necessary and we
4102 * have established with reasonable certainty that
4103 * we can do some work if the lock was obtained.
4104 */
4105 n = get_node(searchp, node);
4106
4107 reap_alien(searchp, n);
4108
4109 drain_array(searchp, n, cpu_cache_get(searchp), node);
4110
4111 /*
4112 * These are racy checks but it does not matter
4113 * if we skip one check or scan twice.
4114 */
4115 if (time_after(n->next_reap, jiffies))
4116 goto next;
4117
4118 n->next_reap = jiffies + REAPTIMEOUT_NODE;
4119
4120 drain_array(searchp, n, n->shared, node);
4121
4122 if (n->free_touched)
4123 n->free_touched = 0;
4124 else {
4125 int freed;
4126
4127 freed = drain_freelist(searchp, n, (n->free_limit +
4128 5 * searchp->num - 1) / (5 * searchp->num));
4129 STATS_ADD_REAPED(searchp, freed);
4130 }
4131 next:
4132 cond_resched();
4133 }
4134 check_irq_on();
4135 mutex_unlock(&slab_mutex);
4136 next_reap_node();
4137 out:
4138 /* Set up the next iteration */
4139 schedule_delayed_work(work, round_jiffies_relative(REAPTIMEOUT_AC));
4140 }
4141
4142 #ifdef CONFIG_SLABINFO
4143 void get_slabinfo(struct kmem_cache *cachep, struct slabinfo *sinfo)
4144 {
4145 struct page *page;
4146 unsigned long active_objs;
4147 unsigned long num_objs;
4148 unsigned long active_slabs = 0;
4149 unsigned long num_slabs, free_objects = 0, shared_avail = 0;
4150 const char *name;
4151 char *error = NULL;
4152 int node;
4153 struct kmem_cache_node *n;
4154
4155 active_objs = 0;
4156 num_slabs = 0;
4157 for_each_kmem_cache_node(cachep, node, n) {
4158
4159 check_irq_on();
4160 spin_lock_irq(&n->list_lock);
4161
4162 list_for_each_entry(page, &n->slabs_full, lru) {
4163 if (page->active != cachep->num && !error)
4164 error = "slabs_full accounting error";
4165 active_objs += cachep->num;
4166 active_slabs++;
4167 }
4168 list_for_each_entry(page, &n->slabs_partial, lru) {
4169 if (page->active == cachep->num && !error)
4170 error = "slabs_partial accounting error";
4171 if (!page->active && !error)
4172 error = "slabs_partial accounting error";
4173 active_objs += page->active;
4174 active_slabs++;
4175 }
4176 list_for_each_entry(page, &n->slabs_free, lru) {
4177 if (page->active && !error)
4178 error = "slabs_free accounting error";
4179 num_slabs++;
4180 }
4181 free_objects += n->free_objects;
4182 if (n->shared)
4183 shared_avail += n->shared->avail;
4184
4185 spin_unlock_irq(&n->list_lock);
4186 }
4187 num_slabs += active_slabs;
4188 num_objs = num_slabs * cachep->num;
4189 if (num_objs - active_objs != free_objects && !error)
4190 error = "free_objects accounting error";
4191
4192 name = cachep->name;
4193 if (error)
4194 pr_err("slab: cache %s error: %s\n", name, error);
4195
4196 sinfo->active_objs = active_objs;
4197 sinfo->num_objs = num_objs;
4198 sinfo->active_slabs = active_slabs;
4199 sinfo->num_slabs = num_slabs;
4200 sinfo->shared_avail = shared_avail;
4201 sinfo->limit = cachep->limit;
4202 sinfo->batchcount = cachep->batchcount;
4203 sinfo->shared = cachep->shared;
4204 sinfo->objects_per_slab = cachep->num;
4205 sinfo->cache_order = cachep->gfporder;
4206 }
4207
4208 void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *cachep)
4209 {
4210 #if STATS
4211 { /* node stats */
4212 unsigned long high = cachep->high_mark;
4213 unsigned long allocs = cachep->num_allocations;
4214 unsigned long grown = cachep->grown;
4215 unsigned long reaped = cachep->reaped;
4216 unsigned long errors = cachep->errors;
4217 unsigned long max_freeable = cachep->max_freeable;
4218 unsigned long node_allocs = cachep->node_allocs;
4219 unsigned long node_frees = cachep->node_frees;
4220 unsigned long overflows = cachep->node_overflow;
4221
4222 seq_printf(m, " : globalstat %7lu %6lu %5lu %4lu %4lu %4lu %4lu %4lu %4lu",
4223 allocs, high, grown,
4224 reaped, errors, max_freeable, node_allocs,
4225 node_frees, overflows);
4226 }
4227 /* cpu stats */
4228 {
4229 unsigned long allochit = atomic_read(&cachep->allochit);
4230 unsigned long allocmiss = atomic_read(&cachep->allocmiss);
4231 unsigned long freehit = atomic_read(&cachep->freehit);
4232 unsigned long freemiss = atomic_read(&cachep->freemiss);
4233
4234 seq_printf(m, " : cpustat %6lu %6lu %6lu %6lu",
4235 allochit, allocmiss, freehit, freemiss);
4236 }
4237 #endif
4238 }
4239
4240 #define MAX_SLABINFO_WRITE 128
4241 /**
4242 * slabinfo_write - Tuning for the slab allocator
4243 * @file: unused
4244 * @buffer: user buffer
4245 * @count: data length
4246 * @ppos: unused
4247 */
4248 ssize_t slabinfo_write(struct file *file, const char __user *buffer,
4249 size_t count, loff_t *ppos)
4250 {
4251 char kbuf[MAX_SLABINFO_WRITE + 1], *tmp;
4252 int limit, batchcount, shared, res;
4253 struct kmem_cache *cachep;
4254
4255 if (count > MAX_SLABINFO_WRITE)
4256 return -EINVAL;
4257 if (copy_from_user(&kbuf, buffer, count))
4258 return -EFAULT;
4259 kbuf[MAX_SLABINFO_WRITE] = '\0';
4260
4261 tmp = strchr(kbuf, ' ');
4262 if (!tmp)
4263 return -EINVAL;
4264 *tmp = '\0';
4265 tmp++;
4266 if (sscanf(tmp, " %d %d %d", &limit, &batchcount, &shared) != 3)
4267 return -EINVAL;
4268
4269 /* Find the cache in the chain of caches. */
4270 mutex_lock(&slab_mutex);
4271 res = -EINVAL;
4272 list_for_each_entry(cachep, &slab_caches, list) {
4273 if (!strcmp(cachep->name, kbuf)) {
4274 if (limit < 1 || batchcount < 1 ||
4275 batchcount > limit || shared < 0) {
4276 res = 0;
4277 } else {
4278 res = do_tune_cpucache(cachep, limit,
4279 batchcount, shared,
4280 GFP_KERNEL);
4281 }
4282 break;
4283 }
4284 }
4285 mutex_unlock(&slab_mutex);
4286 if (res >= 0)
4287 res = count;
4288 return res;
4289 }
4290
4291 #ifdef CONFIG_DEBUG_SLAB_LEAK
4292
4293 static inline int add_caller(unsigned long *n, unsigned long v)
4294 {
4295 unsigned long *p;
4296 int l;
4297 if (!v)
4298 return 1;
4299 l = n[1];
4300 p = n + 2;
4301 while (l) {
4302 int i = l/2;
4303 unsigned long *q = p + 2 * i;
4304 if (*q == v) {
4305 q[1]++;
4306 return 1;
4307 }
4308 if (*q > v) {
4309 l = i;
4310 } else {
4311 p = q + 2;
4312 l -= i + 1;
4313 }
4314 }
4315 if (++n[1] == n[0])
4316 return 0;
4317 memmove(p + 2, p, n[1] * 2 * sizeof(unsigned long) - ((void *)p - (void *)n));
4318 p[0] = v;
4319 p[1] = 1;
4320 return 1;
4321 }
4322
4323 static void handle_slab(unsigned long *n, struct kmem_cache *c,
4324 struct page *page)
4325 {
4326 void *p;
4327 int i, j;
4328 unsigned long v;
4329
4330 if (n[0] == n[1])
4331 return;
4332 for (i = 0, p = page->s_mem; i < c->num; i++, p += c->size) {
4333 bool active = true;
4334
4335 for (j = page->active; j < c->num; j++) {
4336 if (get_free_obj(page, j) == i) {
4337 active = false;
4338 break;
4339 }
4340 }
4341
4342 if (!active)
4343 continue;
4344
4345 /*
4346 * probe_kernel_read() is used for DEBUG_PAGEALLOC. page table
4347 * mapping is established when actual object allocation and
4348 * we could mistakenly access the unmapped object in the cpu
4349 * cache.
4350 */
4351 if (probe_kernel_read(&v, dbg_userword(c, p), sizeof(v)))
4352 continue;
4353
4354 if (!add_caller(n, v))
4355 return;
4356 }
4357 }
4358
4359 static void show_symbol(struct seq_file *m, unsigned long address)
4360 {
4361 #ifdef CONFIG_KALLSYMS
4362 unsigned long offset, size;
4363 char modname[MODULE_NAME_LEN], name[KSYM_NAME_LEN];
4364
4365 if (lookup_symbol_attrs(address, &size, &offset, modname, name) == 0) {
4366 seq_printf(m, "%s+%#lx/%#lx", name, offset, size);
4367 if (modname[0])
4368 seq_printf(m, " [%s]", modname);
4369 return;
4370 }
4371 #endif
4372 seq_printf(m, "%p", (void *)address);
4373 }
4374
4375 static int leaks_show(struct seq_file *m, void *p)
4376 {
4377 struct kmem_cache *cachep = list_entry(p, struct kmem_cache, list);
4378 struct page *page;
4379 struct kmem_cache_node *n;
4380 const char *name;
4381 unsigned long *x = m->private;
4382 int node;
4383 int i;
4384
4385 if (!(cachep->flags & SLAB_STORE_USER))
4386 return 0;
4387 if (!(cachep->flags & SLAB_RED_ZONE))
4388 return 0;
4389
4390 /*
4391 * Set store_user_clean and start to grab stored user information
4392 * for all objects on this cache. If some alloc/free requests comes
4393 * during the processing, information would be wrong so restart
4394 * whole processing.
4395 */
4396 do {
4397 set_store_user_clean(cachep);
4398 drain_cpu_caches(cachep);
4399
4400 x[1] = 0;
4401
4402 for_each_kmem_cache_node(cachep, node, n) {
4403
4404 check_irq_on();
4405 spin_lock_irq(&n->list_lock);
4406
4407 list_for_each_entry(page, &n->slabs_full, lru)
4408 handle_slab(x, cachep, page);
4409 list_for_each_entry(page, &n->slabs_partial, lru)
4410 handle_slab(x, cachep, page);
4411 spin_unlock_irq(&n->list_lock);
4412 }
4413 } while (!is_store_user_clean(cachep));
4414
4415 name = cachep->name;
4416 if (x[0] == x[1]) {
4417 /* Increase the buffer size */
4418 mutex_unlock(&slab_mutex);
4419 m->private = kzalloc(x[0] * 4 * sizeof(unsigned long), GFP_KERNEL);
4420 if (!m->private) {
4421 /* Too bad, we are really out */
4422 m->private = x;
4423 mutex_lock(&slab_mutex);
4424 return -ENOMEM;
4425 }
4426 *(unsigned long *)m->private = x[0] * 2;
4427 kfree(x);
4428 mutex_lock(&slab_mutex);
4429 /* Now make sure this entry will be retried */
4430 m->count = m->size;
4431 return 0;
4432 }
4433 for (i = 0; i < x[1]; i++) {
4434 seq_printf(m, "%s: %lu ", name, x[2*i+3]);
4435 show_symbol(m, x[2*i+2]);
4436 seq_putc(m, '\n');
4437 }
4438
4439 return 0;
4440 }
4441
4442 static const struct seq_operations slabstats_op = {
4443 .start = slab_start,
4444 .next = slab_next,
4445 .stop = slab_stop,
4446 .show = leaks_show,
4447 };
4448
4449 static int slabstats_open(struct inode *inode, struct file *file)
4450 {
4451 unsigned long *n;
4452
4453 n = __seq_open_private(file, &slabstats_op, PAGE_SIZE);
4454 if (!n)
4455 return -ENOMEM;
4456
4457 *n = PAGE_SIZE / (2 * sizeof(unsigned long));
4458
4459 return 0;
4460 }
4461
4462 static const struct file_operations proc_slabstats_operations = {
4463 .open = slabstats_open,
4464 .read = seq_read,
4465 .llseek = seq_lseek,
4466 .release = seq_release_private,
4467 };
4468 #endif
4469
4470 static int __init slab_proc_init(void)
4471 {
4472 #ifdef CONFIG_DEBUG_SLAB_LEAK
4473 proc_create("slab_allocators", 0, NULL, &proc_slabstats_operations);
4474 #endif
4475 return 0;
4476 }
4477 module_init(slab_proc_init);
4478 #endif
4479
4480 /**
4481 * ksize - get the actual amount of memory allocated for a given object
4482 * @objp: Pointer to the object
4483 *
4484 * kmalloc may internally round up allocations and return more memory
4485 * than requested. ksize() can be used to determine the actual amount of
4486 * memory allocated. The caller may use this additional memory, even though
4487 * a smaller amount of memory was initially specified with the kmalloc call.
4488 * The caller must guarantee that objp points to a valid object previously
4489 * allocated with either kmalloc() or kmem_cache_alloc(). The object
4490 * must not be freed during the duration of the call.
4491 */
4492 size_t ksize(const void *objp)
4493 {
4494 size_t size;
4495
4496 BUG_ON(!objp);
4497 if (unlikely(objp == ZERO_SIZE_PTR))
4498 return 0;
4499
4500 size = virt_to_cache(objp)->object_size;
4501 /* We assume that ksize callers could use the whole allocated area,
4502 * so we need to unpoison this area.
4503 */
4504 kasan_krealloc(objp, size, GFP_NOWAIT);
4505
4506 return size;
4507 }
4508 EXPORT_SYMBOL(ksize);