Loading...
Note: File does not exist in v3.1.
1/*
2 * Slab allocator functions that are independent of the allocator strategy
3 *
4 * (C) 2012 Christoph Lameter <cl@linux.com>
5 */
6#include <linux/slab.h>
7
8#include <linux/mm.h>
9#include <linux/poison.h>
10#include <linux/interrupt.h>
11#include <linux/memory.h>
12#include <linux/compiler.h>
13#include <linux/module.h>
14#include <linux/cpu.h>
15#include <linux/uaccess.h>
16#include <linux/seq_file.h>
17#include <linux/proc_fs.h>
18#include <asm/cacheflush.h>
19#include <asm/tlbflush.h>
20#include <asm/page.h>
21#include <linux/memcontrol.h>
22
23#define CREATE_TRACE_POINTS
24#include <trace/events/kmem.h>
25
26#include "slab.h"
27
28enum slab_state slab_state;
29LIST_HEAD(slab_caches);
30DEFINE_MUTEX(slab_mutex);
31struct kmem_cache *kmem_cache;
32
33/*
34 * Set of flags that will prevent slab merging
35 */
36#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
37 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
38 SLAB_FAILSLAB | SLAB_KASAN)
39
40#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
41 SLAB_NOTRACK | SLAB_ACCOUNT)
42
43/*
44 * Merge control. If this is set then no merging of slab caches will occur.
45 * (Could be removed. This was introduced to pacify the merge skeptics.)
46 */
47static int slab_nomerge;
48
49static int __init setup_slab_nomerge(char *str)
50{
51 slab_nomerge = 1;
52 return 1;
53}
54
55#ifdef CONFIG_SLUB
56__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
57#endif
58
59__setup("slab_nomerge", setup_slab_nomerge);
60
61/*
62 * Determine the size of a slab object
63 */
64unsigned int kmem_cache_size(struct kmem_cache *s)
65{
66 return s->object_size;
67}
68EXPORT_SYMBOL(kmem_cache_size);
69
70#ifdef CONFIG_DEBUG_VM
71static int kmem_cache_sanity_check(const char *name, size_t size)
72{
73 struct kmem_cache *s = NULL;
74
75 if (!name || in_interrupt() || size < sizeof(void *) ||
76 size > KMALLOC_MAX_SIZE) {
77 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
78 return -EINVAL;
79 }
80
81 list_for_each_entry(s, &slab_caches, list) {
82 char tmp;
83 int res;
84
85 /*
86 * This happens when the module gets unloaded and doesn't
87 * destroy its slab cache and no-one else reuses the vmalloc
88 * area of the module. Print a warning.
89 */
90 res = probe_kernel_address(s->name, tmp);
91 if (res) {
92 pr_err("Slab cache with size %d has lost its name\n",
93 s->object_size);
94 continue;
95 }
96 }
97
98 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
99 return 0;
100}
101#else
102static inline int kmem_cache_sanity_check(const char *name, size_t size)
103{
104 return 0;
105}
106#endif
107
108void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
109{
110 size_t i;
111
112 for (i = 0; i < nr; i++) {
113 if (s)
114 kmem_cache_free(s, p[i]);
115 else
116 kfree(p[i]);
117 }
118}
119
120int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
121 void **p)
122{
123 size_t i;
124
125 for (i = 0; i < nr; i++) {
126 void *x = p[i] = kmem_cache_alloc(s, flags);
127 if (!x) {
128 __kmem_cache_free_bulk(s, i, p);
129 return 0;
130 }
131 }
132 return i;
133}
134
135#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
136void slab_init_memcg_params(struct kmem_cache *s)
137{
138 s->memcg_params.is_root_cache = true;
139 INIT_LIST_HEAD(&s->memcg_params.list);
140 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
141}
142
143static int init_memcg_params(struct kmem_cache *s,
144 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
145{
146 struct memcg_cache_array *arr;
147
148 if (memcg) {
149 s->memcg_params.is_root_cache = false;
150 s->memcg_params.memcg = memcg;
151 s->memcg_params.root_cache = root_cache;
152 return 0;
153 }
154
155 slab_init_memcg_params(s);
156
157 if (!memcg_nr_cache_ids)
158 return 0;
159
160 arr = kzalloc(sizeof(struct memcg_cache_array) +
161 memcg_nr_cache_ids * sizeof(void *),
162 GFP_KERNEL);
163 if (!arr)
164 return -ENOMEM;
165
166 RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
167 return 0;
168}
169
170static void destroy_memcg_params(struct kmem_cache *s)
171{
172 if (is_root_cache(s))
173 kfree(rcu_access_pointer(s->memcg_params.memcg_caches));
174}
175
176static int update_memcg_params(struct kmem_cache *s, int new_array_size)
177{
178 struct memcg_cache_array *old, *new;
179
180 if (!is_root_cache(s))
181 return 0;
182
183 new = kzalloc(sizeof(struct memcg_cache_array) +
184 new_array_size * sizeof(void *), GFP_KERNEL);
185 if (!new)
186 return -ENOMEM;
187
188 old = rcu_dereference_protected(s->memcg_params.memcg_caches,
189 lockdep_is_held(&slab_mutex));
190 if (old)
191 memcpy(new->entries, old->entries,
192 memcg_nr_cache_ids * sizeof(void *));
193
194 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
195 if (old)
196 kfree_rcu(old, rcu);
197 return 0;
198}
199
200int memcg_update_all_caches(int num_memcgs)
201{
202 struct kmem_cache *s;
203 int ret = 0;
204
205 mutex_lock(&slab_mutex);
206 list_for_each_entry(s, &slab_caches, list) {
207 ret = update_memcg_params(s, num_memcgs);
208 /*
209 * Instead of freeing the memory, we'll just leave the caches
210 * up to this point in an updated state.
211 */
212 if (ret)
213 break;
214 }
215 mutex_unlock(&slab_mutex);
216 return ret;
217}
218#else
219static inline int init_memcg_params(struct kmem_cache *s,
220 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
221{
222 return 0;
223}
224
225static inline void destroy_memcg_params(struct kmem_cache *s)
226{
227}
228#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
229
230/*
231 * Find a mergeable slab cache
232 */
233int slab_unmergeable(struct kmem_cache *s)
234{
235 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
236 return 1;
237
238 if (!is_root_cache(s))
239 return 1;
240
241 if (s->ctor)
242 return 1;
243
244 /*
245 * We may have set a slab to be unmergeable during bootstrap.
246 */
247 if (s->refcount < 0)
248 return 1;
249
250 return 0;
251}
252
253struct kmem_cache *find_mergeable(size_t size, size_t align,
254 unsigned long flags, const char *name, void (*ctor)(void *))
255{
256 struct kmem_cache *s;
257
258 if (slab_nomerge || (flags & SLAB_NEVER_MERGE))
259 return NULL;
260
261 if (ctor)
262 return NULL;
263
264 size = ALIGN(size, sizeof(void *));
265 align = calculate_alignment(flags, align, size);
266 size = ALIGN(size, align);
267 flags = kmem_cache_flags(size, flags, name, NULL);
268
269 list_for_each_entry_reverse(s, &slab_caches, list) {
270 if (slab_unmergeable(s))
271 continue;
272
273 if (size > s->size)
274 continue;
275
276 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
277 continue;
278 /*
279 * Check if alignment is compatible.
280 * Courtesy of Adrian Drzewiecki
281 */
282 if ((s->size & ~(align - 1)) != s->size)
283 continue;
284
285 if (s->size - size >= sizeof(void *))
286 continue;
287
288 if (IS_ENABLED(CONFIG_SLAB) && align &&
289 (align > s->align || s->align % align))
290 continue;
291
292 return s;
293 }
294 return NULL;
295}
296
297/*
298 * Figure out what the alignment of the objects will be given a set of
299 * flags, a user specified alignment and the size of the objects.
300 */
301unsigned long calculate_alignment(unsigned long flags,
302 unsigned long align, unsigned long size)
303{
304 /*
305 * If the user wants hardware cache aligned objects then follow that
306 * suggestion if the object is sufficiently large.
307 *
308 * The hardware cache alignment cannot override the specified
309 * alignment though. If that is greater then use it.
310 */
311 if (flags & SLAB_HWCACHE_ALIGN) {
312 unsigned long ralign = cache_line_size();
313 while (size <= ralign / 2)
314 ralign /= 2;
315 align = max(align, ralign);
316 }
317
318 if (align < ARCH_SLAB_MINALIGN)
319 align = ARCH_SLAB_MINALIGN;
320
321 return ALIGN(align, sizeof(void *));
322}
323
324static struct kmem_cache *create_cache(const char *name,
325 size_t object_size, size_t size, size_t align,
326 unsigned long flags, void (*ctor)(void *),
327 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
328{
329 struct kmem_cache *s;
330 int err;
331
332 err = -ENOMEM;
333 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
334 if (!s)
335 goto out;
336
337 s->name = name;
338 s->object_size = object_size;
339 s->size = size;
340 s->align = align;
341 s->ctor = ctor;
342
343 err = init_memcg_params(s, memcg, root_cache);
344 if (err)
345 goto out_free_cache;
346
347 err = __kmem_cache_create(s, flags);
348 if (err)
349 goto out_free_cache;
350
351 s->refcount = 1;
352 list_add(&s->list, &slab_caches);
353out:
354 if (err)
355 return ERR_PTR(err);
356 return s;
357
358out_free_cache:
359 destroy_memcg_params(s);
360 kmem_cache_free(kmem_cache, s);
361 goto out;
362}
363
364/*
365 * kmem_cache_create - Create a cache.
366 * @name: A string which is used in /proc/slabinfo to identify this cache.
367 * @size: The size of objects to be created in this cache.
368 * @align: The required alignment for the objects.
369 * @flags: SLAB flags
370 * @ctor: A constructor for the objects.
371 *
372 * Returns a ptr to the cache on success, NULL on failure.
373 * Cannot be called within a interrupt, but can be interrupted.
374 * The @ctor is run when new pages are allocated by the cache.
375 *
376 * The flags are
377 *
378 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
379 * to catch references to uninitialised memory.
380 *
381 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
382 * for buffer overruns.
383 *
384 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
385 * cacheline. This can be beneficial if you're counting cycles as closely
386 * as davem.
387 */
388struct kmem_cache *
389kmem_cache_create(const char *name, size_t size, size_t align,
390 unsigned long flags, void (*ctor)(void *))
391{
392 struct kmem_cache *s = NULL;
393 const char *cache_name;
394 int err;
395
396 get_online_cpus();
397 get_online_mems();
398 memcg_get_cache_ids();
399
400 mutex_lock(&slab_mutex);
401
402 err = kmem_cache_sanity_check(name, size);
403 if (err) {
404 goto out_unlock;
405 }
406
407 /* Refuse requests with allocator specific flags */
408 if (flags & ~SLAB_FLAGS_PERMITTED) {
409 err = -EINVAL;
410 goto out_unlock;
411 }
412
413 /*
414 * Some allocators will constraint the set of valid flags to a subset
415 * of all flags. We expect them to define CACHE_CREATE_MASK in this
416 * case, and we'll just provide them with a sanitized version of the
417 * passed flags.
418 */
419 flags &= CACHE_CREATE_MASK;
420
421 s = __kmem_cache_alias(name, size, align, flags, ctor);
422 if (s)
423 goto out_unlock;
424
425 cache_name = kstrdup_const(name, GFP_KERNEL);
426 if (!cache_name) {
427 err = -ENOMEM;
428 goto out_unlock;
429 }
430
431 s = create_cache(cache_name, size, size,
432 calculate_alignment(flags, align, size),
433 flags, ctor, NULL, NULL);
434 if (IS_ERR(s)) {
435 err = PTR_ERR(s);
436 kfree_const(cache_name);
437 }
438
439out_unlock:
440 mutex_unlock(&slab_mutex);
441
442 memcg_put_cache_ids();
443 put_online_mems();
444 put_online_cpus();
445
446 if (err) {
447 if (flags & SLAB_PANIC)
448 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
449 name, err);
450 else {
451 pr_warn("kmem_cache_create(%s) failed with error %d\n",
452 name, err);
453 dump_stack();
454 }
455 return NULL;
456 }
457 return s;
458}
459EXPORT_SYMBOL(kmem_cache_create);
460
461static int shutdown_cache(struct kmem_cache *s,
462 struct list_head *release, bool *need_rcu_barrier)
463{
464 if (__kmem_cache_shutdown(s) != 0)
465 return -EBUSY;
466
467 if (s->flags & SLAB_DESTROY_BY_RCU)
468 *need_rcu_barrier = true;
469
470 list_move(&s->list, release);
471 return 0;
472}
473
474static void release_caches(struct list_head *release, bool need_rcu_barrier)
475{
476 struct kmem_cache *s, *s2;
477
478 if (need_rcu_barrier)
479 rcu_barrier();
480
481 list_for_each_entry_safe(s, s2, release, list) {
482#ifdef SLAB_SUPPORTS_SYSFS
483 sysfs_slab_remove(s);
484#else
485 slab_kmem_cache_release(s);
486#endif
487 }
488}
489
490#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
491/*
492 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
493 * @memcg: The memory cgroup the new cache is for.
494 * @root_cache: The parent of the new cache.
495 *
496 * This function attempts to create a kmem cache that will serve allocation
497 * requests going from @memcg to @root_cache. The new cache inherits properties
498 * from its parent.
499 */
500void memcg_create_kmem_cache(struct mem_cgroup *memcg,
501 struct kmem_cache *root_cache)
502{
503 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
504 struct cgroup_subsys_state *css = &memcg->css;
505 struct memcg_cache_array *arr;
506 struct kmem_cache *s = NULL;
507 char *cache_name;
508 int idx;
509
510 get_online_cpus();
511 get_online_mems();
512
513 mutex_lock(&slab_mutex);
514
515 /*
516 * The memory cgroup could have been offlined while the cache
517 * creation work was pending.
518 */
519 if (memcg->kmem_state != KMEM_ONLINE)
520 goto out_unlock;
521
522 idx = memcg_cache_id(memcg);
523 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
524 lockdep_is_held(&slab_mutex));
525
526 /*
527 * Since per-memcg caches are created asynchronously on first
528 * allocation (see memcg_kmem_get_cache()), several threads can try to
529 * create the same cache, but only one of them may succeed.
530 */
531 if (arr->entries[idx])
532 goto out_unlock;
533
534 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
535 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
536 css->serial_nr, memcg_name_buf);
537 if (!cache_name)
538 goto out_unlock;
539
540 s = create_cache(cache_name, root_cache->object_size,
541 root_cache->size, root_cache->align,
542 root_cache->flags & CACHE_CREATE_MASK,
543 root_cache->ctor, memcg, root_cache);
544 /*
545 * If we could not create a memcg cache, do not complain, because
546 * that's not critical at all as we can always proceed with the root
547 * cache.
548 */
549 if (IS_ERR(s)) {
550 kfree(cache_name);
551 goto out_unlock;
552 }
553
554 list_add(&s->memcg_params.list, &root_cache->memcg_params.list);
555
556 /*
557 * Since readers won't lock (see cache_from_memcg_idx()), we need a
558 * barrier here to ensure nobody will see the kmem_cache partially
559 * initialized.
560 */
561 smp_wmb();
562 arr->entries[idx] = s;
563
564out_unlock:
565 mutex_unlock(&slab_mutex);
566
567 put_online_mems();
568 put_online_cpus();
569}
570
571void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
572{
573 int idx;
574 struct memcg_cache_array *arr;
575 struct kmem_cache *s, *c;
576
577 idx = memcg_cache_id(memcg);
578
579 get_online_cpus();
580 get_online_mems();
581
582#ifdef CONFIG_SLUB
583 /*
584 * In case of SLUB, we need to disable empty slab caching to
585 * avoid pinning the offline memory cgroup by freeable kmem
586 * pages charged to it. SLAB doesn't need this, as it
587 * periodically purges unused slabs.
588 */
589 mutex_lock(&slab_mutex);
590 list_for_each_entry(s, &slab_caches, list) {
591 c = is_root_cache(s) ? cache_from_memcg_idx(s, idx) : NULL;
592 if (c) {
593 c->cpu_partial = 0;
594 c->min_partial = 0;
595 }
596 }
597 mutex_unlock(&slab_mutex);
598 /*
599 * kmem_cache->cpu_partial is checked locklessly (see
600 * put_cpu_partial()). Make sure the change is visible.
601 */
602 synchronize_sched();
603#endif
604
605 mutex_lock(&slab_mutex);
606 list_for_each_entry(s, &slab_caches, list) {
607 if (!is_root_cache(s))
608 continue;
609
610 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
611 lockdep_is_held(&slab_mutex));
612 c = arr->entries[idx];
613 if (!c)
614 continue;
615
616 __kmem_cache_shrink(c);
617 arr->entries[idx] = NULL;
618 }
619 mutex_unlock(&slab_mutex);
620
621 put_online_mems();
622 put_online_cpus();
623}
624
625static int __shutdown_memcg_cache(struct kmem_cache *s,
626 struct list_head *release, bool *need_rcu_barrier)
627{
628 BUG_ON(is_root_cache(s));
629
630 if (shutdown_cache(s, release, need_rcu_barrier))
631 return -EBUSY;
632
633 list_del(&s->memcg_params.list);
634 return 0;
635}
636
637void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
638{
639 LIST_HEAD(release);
640 bool need_rcu_barrier = false;
641 struct kmem_cache *s, *s2;
642
643 get_online_cpus();
644 get_online_mems();
645
646 mutex_lock(&slab_mutex);
647 list_for_each_entry_safe(s, s2, &slab_caches, list) {
648 if (is_root_cache(s) || s->memcg_params.memcg != memcg)
649 continue;
650 /*
651 * The cgroup is about to be freed and therefore has no charges
652 * left. Hence, all its caches must be empty by now.
653 */
654 BUG_ON(__shutdown_memcg_cache(s, &release, &need_rcu_barrier));
655 }
656 mutex_unlock(&slab_mutex);
657
658 put_online_mems();
659 put_online_cpus();
660
661 release_caches(&release, need_rcu_barrier);
662}
663
664static int shutdown_memcg_caches(struct kmem_cache *s,
665 struct list_head *release, bool *need_rcu_barrier)
666{
667 struct memcg_cache_array *arr;
668 struct kmem_cache *c, *c2;
669 LIST_HEAD(busy);
670 int i;
671
672 BUG_ON(!is_root_cache(s));
673
674 /*
675 * First, shutdown active caches, i.e. caches that belong to online
676 * memory cgroups.
677 */
678 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
679 lockdep_is_held(&slab_mutex));
680 for_each_memcg_cache_index(i) {
681 c = arr->entries[i];
682 if (!c)
683 continue;
684 if (__shutdown_memcg_cache(c, release, need_rcu_barrier))
685 /*
686 * The cache still has objects. Move it to a temporary
687 * list so as not to try to destroy it for a second
688 * time while iterating over inactive caches below.
689 */
690 list_move(&c->memcg_params.list, &busy);
691 else
692 /*
693 * The cache is empty and will be destroyed soon. Clear
694 * the pointer to it in the memcg_caches array so that
695 * it will never be accessed even if the root cache
696 * stays alive.
697 */
698 arr->entries[i] = NULL;
699 }
700
701 /*
702 * Second, shutdown all caches left from memory cgroups that are now
703 * offline.
704 */
705 list_for_each_entry_safe(c, c2, &s->memcg_params.list,
706 memcg_params.list)
707 __shutdown_memcg_cache(c, release, need_rcu_barrier);
708
709 list_splice(&busy, &s->memcg_params.list);
710
711 /*
712 * A cache being destroyed must be empty. In particular, this means
713 * that all per memcg caches attached to it must be empty too.
714 */
715 if (!list_empty(&s->memcg_params.list))
716 return -EBUSY;
717 return 0;
718}
719#else
720static inline int shutdown_memcg_caches(struct kmem_cache *s,
721 struct list_head *release, bool *need_rcu_barrier)
722{
723 return 0;
724}
725#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
726
727void slab_kmem_cache_release(struct kmem_cache *s)
728{
729 __kmem_cache_release(s);
730 destroy_memcg_params(s);
731 kfree_const(s->name);
732 kmem_cache_free(kmem_cache, s);
733}
734
735void kmem_cache_destroy(struct kmem_cache *s)
736{
737 LIST_HEAD(release);
738 bool need_rcu_barrier = false;
739 int err;
740
741 if (unlikely(!s))
742 return;
743
744 get_online_cpus();
745 get_online_mems();
746
747 kasan_cache_destroy(s);
748 mutex_lock(&slab_mutex);
749
750 s->refcount--;
751 if (s->refcount)
752 goto out_unlock;
753
754 err = shutdown_memcg_caches(s, &release, &need_rcu_barrier);
755 if (!err)
756 err = shutdown_cache(s, &release, &need_rcu_barrier);
757
758 if (err) {
759 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
760 s->name);
761 dump_stack();
762 }
763out_unlock:
764 mutex_unlock(&slab_mutex);
765
766 put_online_mems();
767 put_online_cpus();
768
769 release_caches(&release, need_rcu_barrier);
770}
771EXPORT_SYMBOL(kmem_cache_destroy);
772
773/**
774 * kmem_cache_shrink - Shrink a cache.
775 * @cachep: The cache to shrink.
776 *
777 * Releases as many slabs as possible for a cache.
778 * To help debugging, a zero exit status indicates all slabs were released.
779 */
780int kmem_cache_shrink(struct kmem_cache *cachep)
781{
782 int ret;
783
784 get_online_cpus();
785 get_online_mems();
786 kasan_cache_shrink(cachep);
787 ret = __kmem_cache_shrink(cachep);
788 put_online_mems();
789 put_online_cpus();
790 return ret;
791}
792EXPORT_SYMBOL(kmem_cache_shrink);
793
794bool slab_is_available(void)
795{
796 return slab_state >= UP;
797}
798
799#ifndef CONFIG_SLOB
800/* Create a cache during boot when no slab services are available yet */
801void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
802 unsigned long flags)
803{
804 int err;
805
806 s->name = name;
807 s->size = s->object_size = size;
808 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
809
810 slab_init_memcg_params(s);
811
812 err = __kmem_cache_create(s, flags);
813
814 if (err)
815 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
816 name, size, err);
817
818 s->refcount = -1; /* Exempt from merging for now */
819}
820
821struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
822 unsigned long flags)
823{
824 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
825
826 if (!s)
827 panic("Out of memory when creating slab %s\n", name);
828
829 create_boot_cache(s, name, size, flags);
830 list_add(&s->list, &slab_caches);
831 s->refcount = 1;
832 return s;
833}
834
835struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
836EXPORT_SYMBOL(kmalloc_caches);
837
838#ifdef CONFIG_ZONE_DMA
839struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
840EXPORT_SYMBOL(kmalloc_dma_caches);
841#endif
842
843/*
844 * Conversion table for small slabs sizes / 8 to the index in the
845 * kmalloc array. This is necessary for slabs < 192 since we have non power
846 * of two cache sizes there. The size of larger slabs can be determined using
847 * fls.
848 */
849static s8 size_index[24] = {
850 3, /* 8 */
851 4, /* 16 */
852 5, /* 24 */
853 5, /* 32 */
854 6, /* 40 */
855 6, /* 48 */
856 6, /* 56 */
857 6, /* 64 */
858 1, /* 72 */
859 1, /* 80 */
860 1, /* 88 */
861 1, /* 96 */
862 7, /* 104 */
863 7, /* 112 */
864 7, /* 120 */
865 7, /* 128 */
866 2, /* 136 */
867 2, /* 144 */
868 2, /* 152 */
869 2, /* 160 */
870 2, /* 168 */
871 2, /* 176 */
872 2, /* 184 */
873 2 /* 192 */
874};
875
876static inline int size_index_elem(size_t bytes)
877{
878 return (bytes - 1) / 8;
879}
880
881/*
882 * Find the kmem_cache structure that serves a given size of
883 * allocation
884 */
885struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
886{
887 int index;
888
889 if (unlikely(size > KMALLOC_MAX_SIZE)) {
890 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
891 return NULL;
892 }
893
894 if (size <= 192) {
895 if (!size)
896 return ZERO_SIZE_PTR;
897
898 index = size_index[size_index_elem(size)];
899 } else
900 index = fls(size - 1);
901
902#ifdef CONFIG_ZONE_DMA
903 if (unlikely((flags & GFP_DMA)))
904 return kmalloc_dma_caches[index];
905
906#endif
907 return kmalloc_caches[index];
908}
909
910/*
911 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
912 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
913 * kmalloc-67108864.
914 */
915static struct {
916 const char *name;
917 unsigned long size;
918} const kmalloc_info[] __initconst = {
919 {NULL, 0}, {"kmalloc-96", 96},
920 {"kmalloc-192", 192}, {"kmalloc-8", 8},
921 {"kmalloc-16", 16}, {"kmalloc-32", 32},
922 {"kmalloc-64", 64}, {"kmalloc-128", 128},
923 {"kmalloc-256", 256}, {"kmalloc-512", 512},
924 {"kmalloc-1024", 1024}, {"kmalloc-2048", 2048},
925 {"kmalloc-4096", 4096}, {"kmalloc-8192", 8192},
926 {"kmalloc-16384", 16384}, {"kmalloc-32768", 32768},
927 {"kmalloc-65536", 65536}, {"kmalloc-131072", 131072},
928 {"kmalloc-262144", 262144}, {"kmalloc-524288", 524288},
929 {"kmalloc-1048576", 1048576}, {"kmalloc-2097152", 2097152},
930 {"kmalloc-4194304", 4194304}, {"kmalloc-8388608", 8388608},
931 {"kmalloc-16777216", 16777216}, {"kmalloc-33554432", 33554432},
932 {"kmalloc-67108864", 67108864}
933};
934
935/*
936 * Patch up the size_index table if we have strange large alignment
937 * requirements for the kmalloc array. This is only the case for
938 * MIPS it seems. The standard arches will not generate any code here.
939 *
940 * Largest permitted alignment is 256 bytes due to the way we
941 * handle the index determination for the smaller caches.
942 *
943 * Make sure that nothing crazy happens if someone starts tinkering
944 * around with ARCH_KMALLOC_MINALIGN
945 */
946void __init setup_kmalloc_cache_index_table(void)
947{
948 int i;
949
950 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
951 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
952
953 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
954 int elem = size_index_elem(i);
955
956 if (elem >= ARRAY_SIZE(size_index))
957 break;
958 size_index[elem] = KMALLOC_SHIFT_LOW;
959 }
960
961 if (KMALLOC_MIN_SIZE >= 64) {
962 /*
963 * The 96 byte size cache is not used if the alignment
964 * is 64 byte.
965 */
966 for (i = 64 + 8; i <= 96; i += 8)
967 size_index[size_index_elem(i)] = 7;
968
969 }
970
971 if (KMALLOC_MIN_SIZE >= 128) {
972 /*
973 * The 192 byte sized cache is not used if the alignment
974 * is 128 byte. Redirect kmalloc to use the 256 byte cache
975 * instead.
976 */
977 for (i = 128 + 8; i <= 192; i += 8)
978 size_index[size_index_elem(i)] = 8;
979 }
980}
981
982static void __init new_kmalloc_cache(int idx, unsigned long flags)
983{
984 kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
985 kmalloc_info[idx].size, flags);
986}
987
988/*
989 * Create the kmalloc array. Some of the regular kmalloc arrays
990 * may already have been created because they were needed to
991 * enable allocations for slab creation.
992 */
993void __init create_kmalloc_caches(unsigned long flags)
994{
995 int i;
996
997 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
998 if (!kmalloc_caches[i])
999 new_kmalloc_cache(i, flags);
1000
1001 /*
1002 * Caches that are not of the two-to-the-power-of size.
1003 * These have to be created immediately after the
1004 * earlier power of two caches
1005 */
1006 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
1007 new_kmalloc_cache(1, flags);
1008 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
1009 new_kmalloc_cache(2, flags);
1010 }
1011
1012 /* Kmalloc array is now usable */
1013 slab_state = UP;
1014
1015#ifdef CONFIG_ZONE_DMA
1016 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1017 struct kmem_cache *s = kmalloc_caches[i];
1018
1019 if (s) {
1020 int size = kmalloc_size(i);
1021 char *n = kasprintf(GFP_NOWAIT,
1022 "dma-kmalloc-%d", size);
1023
1024 BUG_ON(!n);
1025 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
1026 size, SLAB_CACHE_DMA | flags);
1027 }
1028 }
1029#endif
1030}
1031#endif /* !CONFIG_SLOB */
1032
1033/*
1034 * To avoid unnecessary overhead, we pass through large allocation requests
1035 * directly to the page allocator. We use __GFP_COMP, because we will need to
1036 * know the allocation order to free the pages properly in kfree.
1037 */
1038void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1039{
1040 void *ret;
1041 struct page *page;
1042
1043 flags |= __GFP_COMP;
1044 page = alloc_pages(flags, order);
1045 ret = page ? page_address(page) : NULL;
1046 kmemleak_alloc(ret, size, 1, flags);
1047 kasan_kmalloc_large(ret, size, flags);
1048 return ret;
1049}
1050EXPORT_SYMBOL(kmalloc_order);
1051
1052#ifdef CONFIG_TRACING
1053void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1054{
1055 void *ret = kmalloc_order(size, flags, order);
1056 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1057 return ret;
1058}
1059EXPORT_SYMBOL(kmalloc_order_trace);
1060#endif
1061
1062#ifdef CONFIG_SLAB_FREELIST_RANDOM
1063/* Randomize a generic freelist */
1064static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1065 size_t count)
1066{
1067 size_t i;
1068 unsigned int rand;
1069
1070 for (i = 0; i < count; i++)
1071 list[i] = i;
1072
1073 /* Fisher-Yates shuffle */
1074 for (i = count - 1; i > 0; i--) {
1075 rand = prandom_u32_state(state);
1076 rand %= (i + 1);
1077 swap(list[i], list[rand]);
1078 }
1079}
1080
1081/* Create a random sequence per cache */
1082int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1083 gfp_t gfp)
1084{
1085 struct rnd_state state;
1086
1087 if (count < 2 || cachep->random_seq)
1088 return 0;
1089
1090 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1091 if (!cachep->random_seq)
1092 return -ENOMEM;
1093
1094 /* Get best entropy at this stage of boot */
1095 prandom_seed_state(&state, get_random_long());
1096
1097 freelist_randomize(&state, cachep->random_seq, count);
1098 return 0;
1099}
1100
1101/* Destroy the per-cache random freelist sequence */
1102void cache_random_seq_destroy(struct kmem_cache *cachep)
1103{
1104 kfree(cachep->random_seq);
1105 cachep->random_seq = NULL;
1106}
1107#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1108
1109#ifdef CONFIG_SLABINFO
1110
1111#ifdef CONFIG_SLAB
1112#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
1113#else
1114#define SLABINFO_RIGHTS S_IRUSR
1115#endif
1116
1117static void print_slabinfo_header(struct seq_file *m)
1118{
1119 /*
1120 * Output format version, so at least we can change it
1121 * without _too_ many complaints.
1122 */
1123#ifdef CONFIG_DEBUG_SLAB
1124 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1125#else
1126 seq_puts(m, "slabinfo - version: 2.1\n");
1127#endif
1128 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1129 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1130 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1131#ifdef CONFIG_DEBUG_SLAB
1132 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1133 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1134#endif
1135 seq_putc(m, '\n');
1136}
1137
1138void *slab_start(struct seq_file *m, loff_t *pos)
1139{
1140 mutex_lock(&slab_mutex);
1141 return seq_list_start(&slab_caches, *pos);
1142}
1143
1144void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1145{
1146 return seq_list_next(p, &slab_caches, pos);
1147}
1148
1149void slab_stop(struct seq_file *m, void *p)
1150{
1151 mutex_unlock(&slab_mutex);
1152}
1153
1154static void
1155memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1156{
1157 struct kmem_cache *c;
1158 struct slabinfo sinfo;
1159
1160 if (!is_root_cache(s))
1161 return;
1162
1163 for_each_memcg_cache(c, s) {
1164 memset(&sinfo, 0, sizeof(sinfo));
1165 get_slabinfo(c, &sinfo);
1166
1167 info->active_slabs += sinfo.active_slabs;
1168 info->num_slabs += sinfo.num_slabs;
1169 info->shared_avail += sinfo.shared_avail;
1170 info->active_objs += sinfo.active_objs;
1171 info->num_objs += sinfo.num_objs;
1172 }
1173}
1174
1175static void cache_show(struct kmem_cache *s, struct seq_file *m)
1176{
1177 struct slabinfo sinfo;
1178
1179 memset(&sinfo, 0, sizeof(sinfo));
1180 get_slabinfo(s, &sinfo);
1181
1182 memcg_accumulate_slabinfo(s, &sinfo);
1183
1184 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1185 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1186 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1187
1188 seq_printf(m, " : tunables %4u %4u %4u",
1189 sinfo.limit, sinfo.batchcount, sinfo.shared);
1190 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1191 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1192 slabinfo_show_stats(m, s);
1193 seq_putc(m, '\n');
1194}
1195
1196static int slab_show(struct seq_file *m, void *p)
1197{
1198 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1199
1200 if (p == slab_caches.next)
1201 print_slabinfo_header(m);
1202 if (is_root_cache(s))
1203 cache_show(s, m);
1204 return 0;
1205}
1206
1207#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
1208int memcg_slab_show(struct seq_file *m, void *p)
1209{
1210 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1211 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1212
1213 if (p == slab_caches.next)
1214 print_slabinfo_header(m);
1215 if (!is_root_cache(s) && s->memcg_params.memcg == memcg)
1216 cache_show(s, m);
1217 return 0;
1218}
1219#endif
1220
1221/*
1222 * slabinfo_op - iterator that generates /proc/slabinfo
1223 *
1224 * Output layout:
1225 * cache-name
1226 * num-active-objs
1227 * total-objs
1228 * object size
1229 * num-active-slabs
1230 * total-slabs
1231 * num-pages-per-slab
1232 * + further values on SMP and with statistics enabled
1233 */
1234static const struct seq_operations slabinfo_op = {
1235 .start = slab_start,
1236 .next = slab_next,
1237 .stop = slab_stop,
1238 .show = slab_show,
1239};
1240
1241static int slabinfo_open(struct inode *inode, struct file *file)
1242{
1243 return seq_open(file, &slabinfo_op);
1244}
1245
1246static const struct file_operations proc_slabinfo_operations = {
1247 .open = slabinfo_open,
1248 .read = seq_read,
1249 .write = slabinfo_write,
1250 .llseek = seq_lseek,
1251 .release = seq_release,
1252};
1253
1254static int __init slab_proc_init(void)
1255{
1256 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1257 &proc_slabinfo_operations);
1258 return 0;
1259}
1260module_init(slab_proc_init);
1261#endif /* CONFIG_SLABINFO */
1262
1263static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1264 gfp_t flags)
1265{
1266 void *ret;
1267 size_t ks = 0;
1268
1269 if (p)
1270 ks = ksize(p);
1271
1272 if (ks >= new_size) {
1273 kasan_krealloc((void *)p, new_size, flags);
1274 return (void *)p;
1275 }
1276
1277 ret = kmalloc_track_caller(new_size, flags);
1278 if (ret && p)
1279 memcpy(ret, p, ks);
1280
1281 return ret;
1282}
1283
1284/**
1285 * __krealloc - like krealloc() but don't free @p.
1286 * @p: object to reallocate memory for.
1287 * @new_size: how many bytes of memory are required.
1288 * @flags: the type of memory to allocate.
1289 *
1290 * This function is like krealloc() except it never frees the originally
1291 * allocated buffer. Use this if you don't want to free the buffer immediately
1292 * like, for example, with RCU.
1293 */
1294void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1295{
1296 if (unlikely(!new_size))
1297 return ZERO_SIZE_PTR;
1298
1299 return __do_krealloc(p, new_size, flags);
1300
1301}
1302EXPORT_SYMBOL(__krealloc);
1303
1304/**
1305 * krealloc - reallocate memory. The contents will remain unchanged.
1306 * @p: object to reallocate memory for.
1307 * @new_size: how many bytes of memory are required.
1308 * @flags: the type of memory to allocate.
1309 *
1310 * The contents of the object pointed to are preserved up to the
1311 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1312 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1313 * %NULL pointer, the object pointed to is freed.
1314 */
1315void *krealloc(const void *p, size_t new_size, gfp_t flags)
1316{
1317 void *ret;
1318
1319 if (unlikely(!new_size)) {
1320 kfree(p);
1321 return ZERO_SIZE_PTR;
1322 }
1323
1324 ret = __do_krealloc(p, new_size, flags);
1325 if (ret && p != ret)
1326 kfree(p);
1327
1328 return ret;
1329}
1330EXPORT_SYMBOL(krealloc);
1331
1332/**
1333 * kzfree - like kfree but zero memory
1334 * @p: object to free memory of
1335 *
1336 * The memory of the object @p points to is zeroed before freed.
1337 * If @p is %NULL, kzfree() does nothing.
1338 *
1339 * Note: this function zeroes the whole allocated buffer which can be a good
1340 * deal bigger than the requested buffer size passed to kmalloc(). So be
1341 * careful when using this function in performance sensitive code.
1342 */
1343void kzfree(const void *p)
1344{
1345 size_t ks;
1346 void *mem = (void *)p;
1347
1348 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1349 return;
1350 ks = ksize(mem);
1351 memset(mem, 0, ks);
1352 kfree(mem);
1353}
1354EXPORT_SYMBOL(kzfree);
1355
1356/* Tracepoints definitions. */
1357EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1358EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1359EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1360EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1361EXPORT_TRACEPOINT_SYMBOL(kfree);
1362EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);