Loading...
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 /*
408 * Some allocators will constraint the set of valid flags to a subset
409 * of all flags. We expect them to define CACHE_CREATE_MASK in this
410 * case, and we'll just provide them with a sanitized version of the
411 * passed flags.
412 */
413 flags &= CACHE_CREATE_MASK;
414
415 s = __kmem_cache_alias(name, size, align, flags, ctor);
416 if (s)
417 goto out_unlock;
418
419 cache_name = kstrdup_const(name, GFP_KERNEL);
420 if (!cache_name) {
421 err = -ENOMEM;
422 goto out_unlock;
423 }
424
425 s = create_cache(cache_name, size, size,
426 calculate_alignment(flags, align, size),
427 flags, ctor, NULL, NULL);
428 if (IS_ERR(s)) {
429 err = PTR_ERR(s);
430 kfree_const(cache_name);
431 }
432
433out_unlock:
434 mutex_unlock(&slab_mutex);
435
436 memcg_put_cache_ids();
437 put_online_mems();
438 put_online_cpus();
439
440 if (err) {
441 if (flags & SLAB_PANIC)
442 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
443 name, err);
444 else {
445 pr_warn("kmem_cache_create(%s) failed with error %d\n",
446 name, err);
447 dump_stack();
448 }
449 return NULL;
450 }
451 return s;
452}
453EXPORT_SYMBOL(kmem_cache_create);
454
455static int shutdown_cache(struct kmem_cache *s,
456 struct list_head *release, bool *need_rcu_barrier)
457{
458 if (__kmem_cache_shutdown(s) != 0)
459 return -EBUSY;
460
461 if (s->flags & SLAB_DESTROY_BY_RCU)
462 *need_rcu_barrier = true;
463
464 list_move(&s->list, release);
465 return 0;
466}
467
468static void release_caches(struct list_head *release, bool need_rcu_barrier)
469{
470 struct kmem_cache *s, *s2;
471
472 if (need_rcu_barrier)
473 rcu_barrier();
474
475 list_for_each_entry_safe(s, s2, release, list) {
476#ifdef SLAB_SUPPORTS_SYSFS
477 sysfs_slab_remove(s);
478#else
479 slab_kmem_cache_release(s);
480#endif
481 }
482}
483
484#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
485/*
486 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
487 * @memcg: The memory cgroup the new cache is for.
488 * @root_cache: The parent of the new cache.
489 *
490 * This function attempts to create a kmem cache that will serve allocation
491 * requests going from @memcg to @root_cache. The new cache inherits properties
492 * from its parent.
493 */
494void memcg_create_kmem_cache(struct mem_cgroup *memcg,
495 struct kmem_cache *root_cache)
496{
497 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
498 struct cgroup_subsys_state *css = &memcg->css;
499 struct memcg_cache_array *arr;
500 struct kmem_cache *s = NULL;
501 char *cache_name;
502 int idx;
503
504 get_online_cpus();
505 get_online_mems();
506
507 mutex_lock(&slab_mutex);
508
509 /*
510 * The memory cgroup could have been offlined while the cache
511 * creation work was pending.
512 */
513 if (memcg->kmem_state != KMEM_ONLINE)
514 goto out_unlock;
515
516 idx = memcg_cache_id(memcg);
517 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
518 lockdep_is_held(&slab_mutex));
519
520 /*
521 * Since per-memcg caches are created asynchronously on first
522 * allocation (see memcg_kmem_get_cache()), several threads can try to
523 * create the same cache, but only one of them may succeed.
524 */
525 if (arr->entries[idx])
526 goto out_unlock;
527
528 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
529 cache_name = kasprintf(GFP_KERNEL, "%s(%d:%s)", root_cache->name,
530 css->id, memcg_name_buf);
531 if (!cache_name)
532 goto out_unlock;
533
534 s = create_cache(cache_name, root_cache->object_size,
535 root_cache->size, root_cache->align,
536 root_cache->flags, root_cache->ctor,
537 memcg, root_cache);
538 /*
539 * If we could not create a memcg cache, do not complain, because
540 * that's not critical at all as we can always proceed with the root
541 * cache.
542 */
543 if (IS_ERR(s)) {
544 kfree(cache_name);
545 goto out_unlock;
546 }
547
548 list_add(&s->memcg_params.list, &root_cache->memcg_params.list);
549
550 /*
551 * Since readers won't lock (see cache_from_memcg_idx()), we need a
552 * barrier here to ensure nobody will see the kmem_cache partially
553 * initialized.
554 */
555 smp_wmb();
556 arr->entries[idx] = s;
557
558out_unlock:
559 mutex_unlock(&slab_mutex);
560
561 put_online_mems();
562 put_online_cpus();
563}
564
565void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
566{
567 int idx;
568 struct memcg_cache_array *arr;
569 struct kmem_cache *s, *c;
570
571 idx = memcg_cache_id(memcg);
572
573 get_online_cpus();
574 get_online_mems();
575
576 mutex_lock(&slab_mutex);
577 list_for_each_entry(s, &slab_caches, list) {
578 if (!is_root_cache(s))
579 continue;
580
581 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
582 lockdep_is_held(&slab_mutex));
583 c = arr->entries[idx];
584 if (!c)
585 continue;
586
587 __kmem_cache_shrink(c, true);
588 arr->entries[idx] = NULL;
589 }
590 mutex_unlock(&slab_mutex);
591
592 put_online_mems();
593 put_online_cpus();
594}
595
596static int __shutdown_memcg_cache(struct kmem_cache *s,
597 struct list_head *release, bool *need_rcu_barrier)
598{
599 BUG_ON(is_root_cache(s));
600
601 if (shutdown_cache(s, release, need_rcu_barrier))
602 return -EBUSY;
603
604 list_del(&s->memcg_params.list);
605 return 0;
606}
607
608void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
609{
610 LIST_HEAD(release);
611 bool need_rcu_barrier = false;
612 struct kmem_cache *s, *s2;
613
614 get_online_cpus();
615 get_online_mems();
616
617 mutex_lock(&slab_mutex);
618 list_for_each_entry_safe(s, s2, &slab_caches, list) {
619 if (is_root_cache(s) || s->memcg_params.memcg != memcg)
620 continue;
621 /*
622 * The cgroup is about to be freed and therefore has no charges
623 * left. Hence, all its caches must be empty by now.
624 */
625 BUG_ON(__shutdown_memcg_cache(s, &release, &need_rcu_barrier));
626 }
627 mutex_unlock(&slab_mutex);
628
629 put_online_mems();
630 put_online_cpus();
631
632 release_caches(&release, need_rcu_barrier);
633}
634
635static int shutdown_memcg_caches(struct kmem_cache *s,
636 struct list_head *release, bool *need_rcu_barrier)
637{
638 struct memcg_cache_array *arr;
639 struct kmem_cache *c, *c2;
640 LIST_HEAD(busy);
641 int i;
642
643 BUG_ON(!is_root_cache(s));
644
645 /*
646 * First, shutdown active caches, i.e. caches that belong to online
647 * memory cgroups.
648 */
649 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
650 lockdep_is_held(&slab_mutex));
651 for_each_memcg_cache_index(i) {
652 c = arr->entries[i];
653 if (!c)
654 continue;
655 if (__shutdown_memcg_cache(c, release, need_rcu_barrier))
656 /*
657 * The cache still has objects. Move it to a temporary
658 * list so as not to try to destroy it for a second
659 * time while iterating over inactive caches below.
660 */
661 list_move(&c->memcg_params.list, &busy);
662 else
663 /*
664 * The cache is empty and will be destroyed soon. Clear
665 * the pointer to it in the memcg_caches array so that
666 * it will never be accessed even if the root cache
667 * stays alive.
668 */
669 arr->entries[i] = NULL;
670 }
671
672 /*
673 * Second, shutdown all caches left from memory cgroups that are now
674 * offline.
675 */
676 list_for_each_entry_safe(c, c2, &s->memcg_params.list,
677 memcg_params.list)
678 __shutdown_memcg_cache(c, release, need_rcu_barrier);
679
680 list_splice(&busy, &s->memcg_params.list);
681
682 /*
683 * A cache being destroyed must be empty. In particular, this means
684 * that all per memcg caches attached to it must be empty too.
685 */
686 if (!list_empty(&s->memcg_params.list))
687 return -EBUSY;
688 return 0;
689}
690#else
691static inline int shutdown_memcg_caches(struct kmem_cache *s,
692 struct list_head *release, bool *need_rcu_barrier)
693{
694 return 0;
695}
696#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
697
698void slab_kmem_cache_release(struct kmem_cache *s)
699{
700 __kmem_cache_release(s);
701 destroy_memcg_params(s);
702 kfree_const(s->name);
703 kmem_cache_free(kmem_cache, s);
704}
705
706void kmem_cache_destroy(struct kmem_cache *s)
707{
708 LIST_HEAD(release);
709 bool need_rcu_barrier = false;
710 int err;
711
712 if (unlikely(!s))
713 return;
714
715 get_online_cpus();
716 get_online_mems();
717
718 mutex_lock(&slab_mutex);
719
720 s->refcount--;
721 if (s->refcount)
722 goto out_unlock;
723
724 err = shutdown_memcg_caches(s, &release, &need_rcu_barrier);
725 if (!err)
726 err = shutdown_cache(s, &release, &need_rcu_barrier);
727
728 if (err) {
729 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
730 s->name);
731 dump_stack();
732 }
733out_unlock:
734 mutex_unlock(&slab_mutex);
735
736 put_online_mems();
737 put_online_cpus();
738
739 release_caches(&release, need_rcu_barrier);
740}
741EXPORT_SYMBOL(kmem_cache_destroy);
742
743/**
744 * kmem_cache_shrink - Shrink a cache.
745 * @cachep: The cache to shrink.
746 *
747 * Releases as many slabs as possible for a cache.
748 * To help debugging, a zero exit status indicates all slabs were released.
749 */
750int kmem_cache_shrink(struct kmem_cache *cachep)
751{
752 int ret;
753
754 get_online_cpus();
755 get_online_mems();
756 ret = __kmem_cache_shrink(cachep, false);
757 put_online_mems();
758 put_online_cpus();
759 return ret;
760}
761EXPORT_SYMBOL(kmem_cache_shrink);
762
763bool slab_is_available(void)
764{
765 return slab_state >= UP;
766}
767
768#ifndef CONFIG_SLOB
769/* Create a cache during boot when no slab services are available yet */
770void __init create_boot_cache(struct kmem_cache *s, const char *name, size_t size,
771 unsigned long flags)
772{
773 int err;
774
775 s->name = name;
776 s->size = s->object_size = size;
777 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
778
779 slab_init_memcg_params(s);
780
781 err = __kmem_cache_create(s, flags);
782
783 if (err)
784 panic("Creation of kmalloc slab %s size=%zu failed. Reason %d\n",
785 name, size, err);
786
787 s->refcount = -1; /* Exempt from merging for now */
788}
789
790struct kmem_cache *__init create_kmalloc_cache(const char *name, size_t size,
791 unsigned long flags)
792{
793 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
794
795 if (!s)
796 panic("Out of memory when creating slab %s\n", name);
797
798 create_boot_cache(s, name, size, flags);
799 list_add(&s->list, &slab_caches);
800 s->refcount = 1;
801 return s;
802}
803
804struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1];
805EXPORT_SYMBOL(kmalloc_caches);
806
807#ifdef CONFIG_ZONE_DMA
808struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1];
809EXPORT_SYMBOL(kmalloc_dma_caches);
810#endif
811
812/*
813 * Conversion table for small slabs sizes / 8 to the index in the
814 * kmalloc array. This is necessary for slabs < 192 since we have non power
815 * of two cache sizes there. The size of larger slabs can be determined using
816 * fls.
817 */
818static s8 size_index[24] = {
819 3, /* 8 */
820 4, /* 16 */
821 5, /* 24 */
822 5, /* 32 */
823 6, /* 40 */
824 6, /* 48 */
825 6, /* 56 */
826 6, /* 64 */
827 1, /* 72 */
828 1, /* 80 */
829 1, /* 88 */
830 1, /* 96 */
831 7, /* 104 */
832 7, /* 112 */
833 7, /* 120 */
834 7, /* 128 */
835 2, /* 136 */
836 2, /* 144 */
837 2, /* 152 */
838 2, /* 160 */
839 2, /* 168 */
840 2, /* 176 */
841 2, /* 184 */
842 2 /* 192 */
843};
844
845static inline int size_index_elem(size_t bytes)
846{
847 return (bytes - 1) / 8;
848}
849
850/*
851 * Find the kmem_cache structure that serves a given size of
852 * allocation
853 */
854struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
855{
856 int index;
857
858 if (unlikely(size > KMALLOC_MAX_SIZE)) {
859 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
860 return NULL;
861 }
862
863 if (size <= 192) {
864 if (!size)
865 return ZERO_SIZE_PTR;
866
867 index = size_index[size_index_elem(size)];
868 } else
869 index = fls(size - 1);
870
871#ifdef CONFIG_ZONE_DMA
872 if (unlikely((flags & GFP_DMA)))
873 return kmalloc_dma_caches[index];
874
875#endif
876 return kmalloc_caches[index];
877}
878
879/*
880 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
881 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
882 * kmalloc-67108864.
883 */
884static struct {
885 const char *name;
886 unsigned long size;
887} const kmalloc_info[] __initconst = {
888 {NULL, 0}, {"kmalloc-96", 96},
889 {"kmalloc-192", 192}, {"kmalloc-8", 8},
890 {"kmalloc-16", 16}, {"kmalloc-32", 32},
891 {"kmalloc-64", 64}, {"kmalloc-128", 128},
892 {"kmalloc-256", 256}, {"kmalloc-512", 512},
893 {"kmalloc-1024", 1024}, {"kmalloc-2048", 2048},
894 {"kmalloc-4096", 4096}, {"kmalloc-8192", 8192},
895 {"kmalloc-16384", 16384}, {"kmalloc-32768", 32768},
896 {"kmalloc-65536", 65536}, {"kmalloc-131072", 131072},
897 {"kmalloc-262144", 262144}, {"kmalloc-524288", 524288},
898 {"kmalloc-1048576", 1048576}, {"kmalloc-2097152", 2097152},
899 {"kmalloc-4194304", 4194304}, {"kmalloc-8388608", 8388608},
900 {"kmalloc-16777216", 16777216}, {"kmalloc-33554432", 33554432},
901 {"kmalloc-67108864", 67108864}
902};
903
904/*
905 * Patch up the size_index table if we have strange large alignment
906 * requirements for the kmalloc array. This is only the case for
907 * MIPS it seems. The standard arches will not generate any code here.
908 *
909 * Largest permitted alignment is 256 bytes due to the way we
910 * handle the index determination for the smaller caches.
911 *
912 * Make sure that nothing crazy happens if someone starts tinkering
913 * around with ARCH_KMALLOC_MINALIGN
914 */
915void __init setup_kmalloc_cache_index_table(void)
916{
917 int i;
918
919 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
920 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
921
922 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
923 int elem = size_index_elem(i);
924
925 if (elem >= ARRAY_SIZE(size_index))
926 break;
927 size_index[elem] = KMALLOC_SHIFT_LOW;
928 }
929
930 if (KMALLOC_MIN_SIZE >= 64) {
931 /*
932 * The 96 byte size cache is not used if the alignment
933 * is 64 byte.
934 */
935 for (i = 64 + 8; i <= 96; i += 8)
936 size_index[size_index_elem(i)] = 7;
937
938 }
939
940 if (KMALLOC_MIN_SIZE >= 128) {
941 /*
942 * The 192 byte sized cache is not used if the alignment
943 * is 128 byte. Redirect kmalloc to use the 256 byte cache
944 * instead.
945 */
946 for (i = 128 + 8; i <= 192; i += 8)
947 size_index[size_index_elem(i)] = 8;
948 }
949}
950
951static void __init new_kmalloc_cache(int idx, unsigned long flags)
952{
953 kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
954 kmalloc_info[idx].size, flags);
955}
956
957/*
958 * Create the kmalloc array. Some of the regular kmalloc arrays
959 * may already have been created because they were needed to
960 * enable allocations for slab creation.
961 */
962void __init create_kmalloc_caches(unsigned long flags)
963{
964 int i;
965
966 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
967 if (!kmalloc_caches[i])
968 new_kmalloc_cache(i, flags);
969
970 /*
971 * Caches that are not of the two-to-the-power-of size.
972 * These have to be created immediately after the
973 * earlier power of two caches
974 */
975 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
976 new_kmalloc_cache(1, flags);
977 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
978 new_kmalloc_cache(2, flags);
979 }
980
981 /* Kmalloc array is now usable */
982 slab_state = UP;
983
984#ifdef CONFIG_ZONE_DMA
985 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
986 struct kmem_cache *s = kmalloc_caches[i];
987
988 if (s) {
989 int size = kmalloc_size(i);
990 char *n = kasprintf(GFP_NOWAIT,
991 "dma-kmalloc-%d", size);
992
993 BUG_ON(!n);
994 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
995 size, SLAB_CACHE_DMA | flags);
996 }
997 }
998#endif
999}
1000#endif /* !CONFIG_SLOB */
1001
1002/*
1003 * To avoid unnecessary overhead, we pass through large allocation requests
1004 * directly to the page allocator. We use __GFP_COMP, because we will need to
1005 * know the allocation order to free the pages properly in kfree.
1006 */
1007void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1008{
1009 void *ret;
1010 struct page *page;
1011
1012 flags |= __GFP_COMP;
1013 page = alloc_kmem_pages(flags, order);
1014 ret = page ? page_address(page) : NULL;
1015 kmemleak_alloc(ret, size, 1, flags);
1016 kasan_kmalloc_large(ret, size, flags);
1017 return ret;
1018}
1019EXPORT_SYMBOL(kmalloc_order);
1020
1021#ifdef CONFIG_TRACING
1022void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1023{
1024 void *ret = kmalloc_order(size, flags, order);
1025 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1026 return ret;
1027}
1028EXPORT_SYMBOL(kmalloc_order_trace);
1029#endif
1030
1031#ifdef CONFIG_SLABINFO
1032
1033#ifdef CONFIG_SLAB
1034#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
1035#else
1036#define SLABINFO_RIGHTS S_IRUSR
1037#endif
1038
1039static void print_slabinfo_header(struct seq_file *m)
1040{
1041 /*
1042 * Output format version, so at least we can change it
1043 * without _too_ many complaints.
1044 */
1045#ifdef CONFIG_DEBUG_SLAB
1046 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1047#else
1048 seq_puts(m, "slabinfo - version: 2.1\n");
1049#endif
1050 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1051 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1052 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1053#ifdef CONFIG_DEBUG_SLAB
1054 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1055 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1056#endif
1057 seq_putc(m, '\n');
1058}
1059
1060void *slab_start(struct seq_file *m, loff_t *pos)
1061{
1062 mutex_lock(&slab_mutex);
1063 return seq_list_start(&slab_caches, *pos);
1064}
1065
1066void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1067{
1068 return seq_list_next(p, &slab_caches, pos);
1069}
1070
1071void slab_stop(struct seq_file *m, void *p)
1072{
1073 mutex_unlock(&slab_mutex);
1074}
1075
1076static void
1077memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1078{
1079 struct kmem_cache *c;
1080 struct slabinfo sinfo;
1081
1082 if (!is_root_cache(s))
1083 return;
1084
1085 for_each_memcg_cache(c, s) {
1086 memset(&sinfo, 0, sizeof(sinfo));
1087 get_slabinfo(c, &sinfo);
1088
1089 info->active_slabs += sinfo.active_slabs;
1090 info->num_slabs += sinfo.num_slabs;
1091 info->shared_avail += sinfo.shared_avail;
1092 info->active_objs += sinfo.active_objs;
1093 info->num_objs += sinfo.num_objs;
1094 }
1095}
1096
1097static void cache_show(struct kmem_cache *s, struct seq_file *m)
1098{
1099 struct slabinfo sinfo;
1100
1101 memset(&sinfo, 0, sizeof(sinfo));
1102 get_slabinfo(s, &sinfo);
1103
1104 memcg_accumulate_slabinfo(s, &sinfo);
1105
1106 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1107 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1108 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1109
1110 seq_printf(m, " : tunables %4u %4u %4u",
1111 sinfo.limit, sinfo.batchcount, sinfo.shared);
1112 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1113 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1114 slabinfo_show_stats(m, s);
1115 seq_putc(m, '\n');
1116}
1117
1118static int slab_show(struct seq_file *m, void *p)
1119{
1120 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1121
1122 if (p == slab_caches.next)
1123 print_slabinfo_header(m);
1124 if (is_root_cache(s))
1125 cache_show(s, m);
1126 return 0;
1127}
1128
1129#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
1130int memcg_slab_show(struct seq_file *m, void *p)
1131{
1132 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1133 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1134
1135 if (p == slab_caches.next)
1136 print_slabinfo_header(m);
1137 if (!is_root_cache(s) && s->memcg_params.memcg == memcg)
1138 cache_show(s, m);
1139 return 0;
1140}
1141#endif
1142
1143/*
1144 * slabinfo_op - iterator that generates /proc/slabinfo
1145 *
1146 * Output layout:
1147 * cache-name
1148 * num-active-objs
1149 * total-objs
1150 * object size
1151 * num-active-slabs
1152 * total-slabs
1153 * num-pages-per-slab
1154 * + further values on SMP and with statistics enabled
1155 */
1156static const struct seq_operations slabinfo_op = {
1157 .start = slab_start,
1158 .next = slab_next,
1159 .stop = slab_stop,
1160 .show = slab_show,
1161};
1162
1163static int slabinfo_open(struct inode *inode, struct file *file)
1164{
1165 return seq_open(file, &slabinfo_op);
1166}
1167
1168static const struct file_operations proc_slabinfo_operations = {
1169 .open = slabinfo_open,
1170 .read = seq_read,
1171 .write = slabinfo_write,
1172 .llseek = seq_lseek,
1173 .release = seq_release,
1174};
1175
1176static int __init slab_proc_init(void)
1177{
1178 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1179 &proc_slabinfo_operations);
1180 return 0;
1181}
1182module_init(slab_proc_init);
1183#endif /* CONFIG_SLABINFO */
1184
1185static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1186 gfp_t flags)
1187{
1188 void *ret;
1189 size_t ks = 0;
1190
1191 if (p)
1192 ks = ksize(p);
1193
1194 if (ks >= new_size) {
1195 kasan_krealloc((void *)p, new_size, flags);
1196 return (void *)p;
1197 }
1198
1199 ret = kmalloc_track_caller(new_size, flags);
1200 if (ret && p)
1201 memcpy(ret, p, ks);
1202
1203 return ret;
1204}
1205
1206/**
1207 * __krealloc - like krealloc() but don't free @p.
1208 * @p: object to reallocate memory for.
1209 * @new_size: how many bytes of memory are required.
1210 * @flags: the type of memory to allocate.
1211 *
1212 * This function is like krealloc() except it never frees the originally
1213 * allocated buffer. Use this if you don't want to free the buffer immediately
1214 * like, for example, with RCU.
1215 */
1216void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1217{
1218 if (unlikely(!new_size))
1219 return ZERO_SIZE_PTR;
1220
1221 return __do_krealloc(p, new_size, flags);
1222
1223}
1224EXPORT_SYMBOL(__krealloc);
1225
1226/**
1227 * krealloc - reallocate memory. The contents will remain unchanged.
1228 * @p: object to reallocate memory for.
1229 * @new_size: how many bytes of memory are required.
1230 * @flags: the type of memory to allocate.
1231 *
1232 * The contents of the object pointed to are preserved up to the
1233 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1234 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1235 * %NULL pointer, the object pointed to is freed.
1236 */
1237void *krealloc(const void *p, size_t new_size, gfp_t flags)
1238{
1239 void *ret;
1240
1241 if (unlikely(!new_size)) {
1242 kfree(p);
1243 return ZERO_SIZE_PTR;
1244 }
1245
1246 ret = __do_krealloc(p, new_size, flags);
1247 if (ret && p != ret)
1248 kfree(p);
1249
1250 return ret;
1251}
1252EXPORT_SYMBOL(krealloc);
1253
1254/**
1255 * kzfree - like kfree but zero memory
1256 * @p: object to free memory of
1257 *
1258 * The memory of the object @p points to is zeroed before freed.
1259 * If @p is %NULL, kzfree() does nothing.
1260 *
1261 * Note: this function zeroes the whole allocated buffer which can be a good
1262 * deal bigger than the requested buffer size passed to kmalloc(). So be
1263 * careful when using this function in performance sensitive code.
1264 */
1265void kzfree(const void *p)
1266{
1267 size_t ks;
1268 void *mem = (void *)p;
1269
1270 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1271 return;
1272 ks = ksize(mem);
1273 memset(mem, 0, ks);
1274 kfree(mem);
1275}
1276EXPORT_SYMBOL(kzfree);
1277
1278/* Tracepoints definitions. */
1279EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1280EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1281EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1282EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1283EXPORT_TRACEPOINT_SYMBOL(kfree);
1284EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1// SPDX-License-Identifier: GPL-2.0
2/*
3 * Slab allocator functions that are independent of the allocator strategy
4 *
5 * (C) 2012 Christoph Lameter <cl@linux.com>
6 */
7#include <linux/slab.h>
8
9#include <linux/mm.h>
10#include <linux/poison.h>
11#include <linux/interrupt.h>
12#include <linux/memory.h>
13#include <linux/cache.h>
14#include <linux/compiler.h>
15#include <linux/module.h>
16#include <linux/cpu.h>
17#include <linux/uaccess.h>
18#include <linux/seq_file.h>
19#include <linux/proc_fs.h>
20#include <asm/cacheflush.h>
21#include <asm/tlbflush.h>
22#include <asm/page.h>
23#include <linux/memcontrol.h>
24
25#define CREATE_TRACE_POINTS
26#include <trace/events/kmem.h>
27
28#include "slab.h"
29
30enum slab_state slab_state;
31LIST_HEAD(slab_caches);
32DEFINE_MUTEX(slab_mutex);
33struct kmem_cache *kmem_cache;
34
35#ifdef CONFIG_HARDENED_USERCOPY
36bool usercopy_fallback __ro_after_init =
37 IS_ENABLED(CONFIG_HARDENED_USERCOPY_FALLBACK);
38module_param(usercopy_fallback, bool, 0400);
39MODULE_PARM_DESC(usercopy_fallback,
40 "WARN instead of reject usercopy whitelist violations");
41#endif
42
43static LIST_HEAD(slab_caches_to_rcu_destroy);
44static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
45static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
46 slab_caches_to_rcu_destroy_workfn);
47
48/*
49 * Set of flags that will prevent slab merging
50 */
51#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
52 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
53 SLAB_FAILSLAB | SLAB_KASAN)
54
55#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
56 SLAB_ACCOUNT)
57
58/*
59 * Merge control. If this is set then no merging of slab caches will occur.
60 */
61static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
62
63static int __init setup_slab_nomerge(char *str)
64{
65 slab_nomerge = true;
66 return 1;
67}
68
69#ifdef CONFIG_SLUB
70__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
71#endif
72
73__setup("slab_nomerge", setup_slab_nomerge);
74
75/*
76 * Determine the size of a slab object
77 */
78unsigned int kmem_cache_size(struct kmem_cache *s)
79{
80 return s->object_size;
81}
82EXPORT_SYMBOL(kmem_cache_size);
83
84#ifdef CONFIG_DEBUG_VM
85static int kmem_cache_sanity_check(const char *name, unsigned int size)
86{
87 if (!name || in_interrupt() || size < sizeof(void *) ||
88 size > KMALLOC_MAX_SIZE) {
89 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
90 return -EINVAL;
91 }
92
93 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
94 return 0;
95}
96#else
97static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
98{
99 return 0;
100}
101#endif
102
103void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
104{
105 size_t i;
106
107 for (i = 0; i < nr; i++) {
108 if (s)
109 kmem_cache_free(s, p[i]);
110 else
111 kfree(p[i]);
112 }
113}
114
115int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
116 void **p)
117{
118 size_t i;
119
120 for (i = 0; i < nr; i++) {
121 void *x = p[i] = kmem_cache_alloc(s, flags);
122 if (!x) {
123 __kmem_cache_free_bulk(s, i, p);
124 return 0;
125 }
126 }
127 return i;
128}
129
130#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
131
132LIST_HEAD(slab_root_caches);
133
134void slab_init_memcg_params(struct kmem_cache *s)
135{
136 s->memcg_params.root_cache = NULL;
137 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
138 INIT_LIST_HEAD(&s->memcg_params.children);
139}
140
141static int init_memcg_params(struct kmem_cache *s,
142 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
143{
144 struct memcg_cache_array *arr;
145
146 if (root_cache) {
147 s->memcg_params.root_cache = root_cache;
148 s->memcg_params.memcg = memcg;
149 INIT_LIST_HEAD(&s->memcg_params.children_node);
150 INIT_LIST_HEAD(&s->memcg_params.kmem_caches_node);
151 return 0;
152 }
153
154 slab_init_memcg_params(s);
155
156 if (!memcg_nr_cache_ids)
157 return 0;
158
159 arr = kvzalloc(sizeof(struct memcg_cache_array) +
160 memcg_nr_cache_ids * sizeof(void *),
161 GFP_KERNEL);
162 if (!arr)
163 return -ENOMEM;
164
165 RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
166 return 0;
167}
168
169static void destroy_memcg_params(struct kmem_cache *s)
170{
171 if (is_root_cache(s))
172 kvfree(rcu_access_pointer(s->memcg_params.memcg_caches));
173}
174
175static void free_memcg_params(struct rcu_head *rcu)
176{
177 struct memcg_cache_array *old;
178
179 old = container_of(rcu, struct memcg_cache_array, rcu);
180 kvfree(old);
181}
182
183static int update_memcg_params(struct kmem_cache *s, int new_array_size)
184{
185 struct memcg_cache_array *old, *new;
186
187 new = kvzalloc(sizeof(struct memcg_cache_array) +
188 new_array_size * sizeof(void *), GFP_KERNEL);
189 if (!new)
190 return -ENOMEM;
191
192 old = rcu_dereference_protected(s->memcg_params.memcg_caches,
193 lockdep_is_held(&slab_mutex));
194 if (old)
195 memcpy(new->entries, old->entries,
196 memcg_nr_cache_ids * sizeof(void *));
197
198 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
199 if (old)
200 call_rcu(&old->rcu, free_memcg_params);
201 return 0;
202}
203
204int memcg_update_all_caches(int num_memcgs)
205{
206 struct kmem_cache *s;
207 int ret = 0;
208
209 mutex_lock(&slab_mutex);
210 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
211 ret = update_memcg_params(s, num_memcgs);
212 /*
213 * Instead of freeing the memory, we'll just leave the caches
214 * up to this point in an updated state.
215 */
216 if (ret)
217 break;
218 }
219 mutex_unlock(&slab_mutex);
220 return ret;
221}
222
223void memcg_link_cache(struct kmem_cache *s)
224{
225 if (is_root_cache(s)) {
226 list_add(&s->root_caches_node, &slab_root_caches);
227 } else {
228 list_add(&s->memcg_params.children_node,
229 &s->memcg_params.root_cache->memcg_params.children);
230 list_add(&s->memcg_params.kmem_caches_node,
231 &s->memcg_params.memcg->kmem_caches);
232 }
233}
234
235static void memcg_unlink_cache(struct kmem_cache *s)
236{
237 if (is_root_cache(s)) {
238 list_del(&s->root_caches_node);
239 } else {
240 list_del(&s->memcg_params.children_node);
241 list_del(&s->memcg_params.kmem_caches_node);
242 }
243}
244#else
245static inline int init_memcg_params(struct kmem_cache *s,
246 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
247{
248 return 0;
249}
250
251static inline void destroy_memcg_params(struct kmem_cache *s)
252{
253}
254
255static inline void memcg_unlink_cache(struct kmem_cache *s)
256{
257}
258#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
259
260/*
261 * Figure out what the alignment of the objects will be given a set of
262 * flags, a user specified alignment and the size of the objects.
263 */
264static unsigned int calculate_alignment(slab_flags_t flags,
265 unsigned int align, unsigned int size)
266{
267 /*
268 * If the user wants hardware cache aligned objects then follow that
269 * suggestion if the object is sufficiently large.
270 *
271 * The hardware cache alignment cannot override the specified
272 * alignment though. If that is greater then use it.
273 */
274 if (flags & SLAB_HWCACHE_ALIGN) {
275 unsigned int ralign;
276
277 ralign = cache_line_size();
278 while (size <= ralign / 2)
279 ralign /= 2;
280 align = max(align, ralign);
281 }
282
283 if (align < ARCH_SLAB_MINALIGN)
284 align = ARCH_SLAB_MINALIGN;
285
286 return ALIGN(align, sizeof(void *));
287}
288
289/*
290 * Find a mergeable slab cache
291 */
292int slab_unmergeable(struct kmem_cache *s)
293{
294 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
295 return 1;
296
297 if (!is_root_cache(s))
298 return 1;
299
300 if (s->ctor)
301 return 1;
302
303 if (s->usersize)
304 return 1;
305
306 /*
307 * We may have set a slab to be unmergeable during bootstrap.
308 */
309 if (s->refcount < 0)
310 return 1;
311
312 return 0;
313}
314
315struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
316 slab_flags_t flags, const char *name, void (*ctor)(void *))
317{
318 struct kmem_cache *s;
319
320 if (slab_nomerge)
321 return NULL;
322
323 if (ctor)
324 return NULL;
325
326 size = ALIGN(size, sizeof(void *));
327 align = calculate_alignment(flags, align, size);
328 size = ALIGN(size, align);
329 flags = kmem_cache_flags(size, flags, name, NULL);
330
331 if (flags & SLAB_NEVER_MERGE)
332 return NULL;
333
334 list_for_each_entry_reverse(s, &slab_root_caches, root_caches_node) {
335 if (slab_unmergeable(s))
336 continue;
337
338 if (size > s->size)
339 continue;
340
341 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
342 continue;
343 /*
344 * Check if alignment is compatible.
345 * Courtesy of Adrian Drzewiecki
346 */
347 if ((s->size & ~(align - 1)) != s->size)
348 continue;
349
350 if (s->size - size >= sizeof(void *))
351 continue;
352
353 if (IS_ENABLED(CONFIG_SLAB) && align &&
354 (align > s->align || s->align % align))
355 continue;
356
357 return s;
358 }
359 return NULL;
360}
361
362static struct kmem_cache *create_cache(const char *name,
363 unsigned int object_size, unsigned int align,
364 slab_flags_t flags, unsigned int useroffset,
365 unsigned int usersize, void (*ctor)(void *),
366 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
367{
368 struct kmem_cache *s;
369 int err;
370
371 if (WARN_ON(useroffset + usersize > object_size))
372 useroffset = usersize = 0;
373
374 err = -ENOMEM;
375 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
376 if (!s)
377 goto out;
378
379 s->name = name;
380 s->size = s->object_size = object_size;
381 s->align = align;
382 s->ctor = ctor;
383 s->useroffset = useroffset;
384 s->usersize = usersize;
385
386 err = init_memcg_params(s, memcg, root_cache);
387 if (err)
388 goto out_free_cache;
389
390 err = __kmem_cache_create(s, flags);
391 if (err)
392 goto out_free_cache;
393
394 s->refcount = 1;
395 list_add(&s->list, &slab_caches);
396 memcg_link_cache(s);
397out:
398 if (err)
399 return ERR_PTR(err);
400 return s;
401
402out_free_cache:
403 destroy_memcg_params(s);
404 kmem_cache_free(kmem_cache, s);
405 goto out;
406}
407
408/*
409 * kmem_cache_create_usercopy - Create a cache.
410 * @name: A string which is used in /proc/slabinfo to identify this cache.
411 * @size: The size of objects to be created in this cache.
412 * @align: The required alignment for the objects.
413 * @flags: SLAB flags
414 * @useroffset: Usercopy region offset
415 * @usersize: Usercopy region size
416 * @ctor: A constructor for the objects.
417 *
418 * Returns a ptr to the cache on success, NULL on failure.
419 * Cannot be called within a interrupt, but can be interrupted.
420 * The @ctor is run when new pages are allocated by the cache.
421 *
422 * The flags are
423 *
424 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
425 * to catch references to uninitialised memory.
426 *
427 * %SLAB_RED_ZONE - Insert `Red' zones around the allocated memory to check
428 * for buffer overruns.
429 *
430 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
431 * cacheline. This can be beneficial if you're counting cycles as closely
432 * as davem.
433 */
434struct kmem_cache *
435kmem_cache_create_usercopy(const char *name,
436 unsigned int size, unsigned int align,
437 slab_flags_t flags,
438 unsigned int useroffset, unsigned int usersize,
439 void (*ctor)(void *))
440{
441 struct kmem_cache *s = NULL;
442 const char *cache_name;
443 int err;
444
445 get_online_cpus();
446 get_online_mems();
447 memcg_get_cache_ids();
448
449 mutex_lock(&slab_mutex);
450
451 err = kmem_cache_sanity_check(name, size);
452 if (err) {
453 goto out_unlock;
454 }
455
456 /* Refuse requests with allocator specific flags */
457 if (flags & ~SLAB_FLAGS_PERMITTED) {
458 err = -EINVAL;
459 goto out_unlock;
460 }
461
462 /*
463 * Some allocators will constraint the set of valid flags to a subset
464 * of all flags. We expect them to define CACHE_CREATE_MASK in this
465 * case, and we'll just provide them with a sanitized version of the
466 * passed flags.
467 */
468 flags &= CACHE_CREATE_MASK;
469
470 /* Fail closed on bad usersize of useroffset values. */
471 if (WARN_ON(!usersize && useroffset) ||
472 WARN_ON(size < usersize || size - usersize < useroffset))
473 usersize = useroffset = 0;
474
475 if (!usersize)
476 s = __kmem_cache_alias(name, size, align, flags, ctor);
477 if (s)
478 goto out_unlock;
479
480 cache_name = kstrdup_const(name, GFP_KERNEL);
481 if (!cache_name) {
482 err = -ENOMEM;
483 goto out_unlock;
484 }
485
486 s = create_cache(cache_name, size,
487 calculate_alignment(flags, align, size),
488 flags, useroffset, usersize, ctor, NULL, NULL);
489 if (IS_ERR(s)) {
490 err = PTR_ERR(s);
491 kfree_const(cache_name);
492 }
493
494out_unlock:
495 mutex_unlock(&slab_mutex);
496
497 memcg_put_cache_ids();
498 put_online_mems();
499 put_online_cpus();
500
501 if (err) {
502 if (flags & SLAB_PANIC)
503 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
504 name, err);
505 else {
506 pr_warn("kmem_cache_create(%s) failed with error %d\n",
507 name, err);
508 dump_stack();
509 }
510 return NULL;
511 }
512 return s;
513}
514EXPORT_SYMBOL(kmem_cache_create_usercopy);
515
516struct kmem_cache *
517kmem_cache_create(const char *name, unsigned int size, unsigned int align,
518 slab_flags_t flags, void (*ctor)(void *))
519{
520 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
521 ctor);
522}
523EXPORT_SYMBOL(kmem_cache_create);
524
525static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
526{
527 LIST_HEAD(to_destroy);
528 struct kmem_cache *s, *s2;
529
530 /*
531 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
532 * @slab_caches_to_rcu_destroy list. The slab pages are freed
533 * through RCU and and the associated kmem_cache are dereferenced
534 * while freeing the pages, so the kmem_caches should be freed only
535 * after the pending RCU operations are finished. As rcu_barrier()
536 * is a pretty slow operation, we batch all pending destructions
537 * asynchronously.
538 */
539 mutex_lock(&slab_mutex);
540 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
541 mutex_unlock(&slab_mutex);
542
543 if (list_empty(&to_destroy))
544 return;
545
546 rcu_barrier();
547
548 list_for_each_entry_safe(s, s2, &to_destroy, list) {
549#ifdef SLAB_SUPPORTS_SYSFS
550 sysfs_slab_release(s);
551#else
552 slab_kmem_cache_release(s);
553#endif
554 }
555}
556
557static int shutdown_cache(struct kmem_cache *s)
558{
559 /* free asan quarantined objects */
560 kasan_cache_shutdown(s);
561
562 if (__kmem_cache_shutdown(s) != 0)
563 return -EBUSY;
564
565 memcg_unlink_cache(s);
566 list_del(&s->list);
567
568 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
569 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
570 schedule_work(&slab_caches_to_rcu_destroy_work);
571 } else {
572#ifdef SLAB_SUPPORTS_SYSFS
573 sysfs_slab_release(s);
574#else
575 slab_kmem_cache_release(s);
576#endif
577 }
578
579 return 0;
580}
581
582#if defined(CONFIG_MEMCG) && !defined(CONFIG_SLOB)
583/*
584 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
585 * @memcg: The memory cgroup the new cache is for.
586 * @root_cache: The parent of the new cache.
587 *
588 * This function attempts to create a kmem cache that will serve allocation
589 * requests going from @memcg to @root_cache. The new cache inherits properties
590 * from its parent.
591 */
592void memcg_create_kmem_cache(struct mem_cgroup *memcg,
593 struct kmem_cache *root_cache)
594{
595 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
596 struct cgroup_subsys_state *css = &memcg->css;
597 struct memcg_cache_array *arr;
598 struct kmem_cache *s = NULL;
599 char *cache_name;
600 int idx;
601
602 get_online_cpus();
603 get_online_mems();
604
605 mutex_lock(&slab_mutex);
606
607 /*
608 * The memory cgroup could have been offlined while the cache
609 * creation work was pending.
610 */
611 if (memcg->kmem_state != KMEM_ONLINE)
612 goto out_unlock;
613
614 idx = memcg_cache_id(memcg);
615 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
616 lockdep_is_held(&slab_mutex));
617
618 /*
619 * Since per-memcg caches are created asynchronously on first
620 * allocation (see memcg_kmem_get_cache()), several threads can try to
621 * create the same cache, but only one of them may succeed.
622 */
623 if (arr->entries[idx])
624 goto out_unlock;
625
626 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
627 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
628 css->serial_nr, memcg_name_buf);
629 if (!cache_name)
630 goto out_unlock;
631
632 s = create_cache(cache_name, root_cache->object_size,
633 root_cache->align,
634 root_cache->flags & CACHE_CREATE_MASK,
635 root_cache->useroffset, root_cache->usersize,
636 root_cache->ctor, memcg, root_cache);
637 /*
638 * If we could not create a memcg cache, do not complain, because
639 * that's not critical at all as we can always proceed with the root
640 * cache.
641 */
642 if (IS_ERR(s)) {
643 kfree(cache_name);
644 goto out_unlock;
645 }
646
647 /*
648 * Since readers won't lock (see cache_from_memcg_idx()), we need a
649 * barrier here to ensure nobody will see the kmem_cache partially
650 * initialized.
651 */
652 smp_wmb();
653 arr->entries[idx] = s;
654
655out_unlock:
656 mutex_unlock(&slab_mutex);
657
658 put_online_mems();
659 put_online_cpus();
660}
661
662static void kmemcg_deactivate_workfn(struct work_struct *work)
663{
664 struct kmem_cache *s = container_of(work, struct kmem_cache,
665 memcg_params.deact_work);
666
667 get_online_cpus();
668 get_online_mems();
669
670 mutex_lock(&slab_mutex);
671
672 s->memcg_params.deact_fn(s);
673
674 mutex_unlock(&slab_mutex);
675
676 put_online_mems();
677 put_online_cpus();
678
679 /* done, put the ref from slab_deactivate_memcg_cache_rcu_sched() */
680 css_put(&s->memcg_params.memcg->css);
681}
682
683static void kmemcg_deactivate_rcufn(struct rcu_head *head)
684{
685 struct kmem_cache *s = container_of(head, struct kmem_cache,
686 memcg_params.deact_rcu_head);
687
688 /*
689 * We need to grab blocking locks. Bounce to ->deact_work. The
690 * work item shares the space with the RCU head and can't be
691 * initialized eariler.
692 */
693 INIT_WORK(&s->memcg_params.deact_work, kmemcg_deactivate_workfn);
694 queue_work(memcg_kmem_cache_wq, &s->memcg_params.deact_work);
695}
696
697/**
698 * slab_deactivate_memcg_cache_rcu_sched - schedule deactivation after a
699 * sched RCU grace period
700 * @s: target kmem_cache
701 * @deact_fn: deactivation function to call
702 *
703 * Schedule @deact_fn to be invoked with online cpus, mems and slab_mutex
704 * held after a sched RCU grace period. The slab is guaranteed to stay
705 * alive until @deact_fn is finished. This is to be used from
706 * __kmemcg_cache_deactivate().
707 */
708void slab_deactivate_memcg_cache_rcu_sched(struct kmem_cache *s,
709 void (*deact_fn)(struct kmem_cache *))
710{
711 if (WARN_ON_ONCE(is_root_cache(s)) ||
712 WARN_ON_ONCE(s->memcg_params.deact_fn))
713 return;
714
715 /* pin memcg so that @s doesn't get destroyed in the middle */
716 css_get(&s->memcg_params.memcg->css);
717
718 s->memcg_params.deact_fn = deact_fn;
719 call_rcu_sched(&s->memcg_params.deact_rcu_head, kmemcg_deactivate_rcufn);
720}
721
722void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg)
723{
724 int idx;
725 struct memcg_cache_array *arr;
726 struct kmem_cache *s, *c;
727
728 idx = memcg_cache_id(memcg);
729
730 get_online_cpus();
731 get_online_mems();
732
733 mutex_lock(&slab_mutex);
734 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
735 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
736 lockdep_is_held(&slab_mutex));
737 c = arr->entries[idx];
738 if (!c)
739 continue;
740
741 __kmemcg_cache_deactivate(c);
742 arr->entries[idx] = NULL;
743 }
744 mutex_unlock(&slab_mutex);
745
746 put_online_mems();
747 put_online_cpus();
748}
749
750void memcg_destroy_kmem_caches(struct mem_cgroup *memcg)
751{
752 struct kmem_cache *s, *s2;
753
754 get_online_cpus();
755 get_online_mems();
756
757 mutex_lock(&slab_mutex);
758 list_for_each_entry_safe(s, s2, &memcg->kmem_caches,
759 memcg_params.kmem_caches_node) {
760 /*
761 * The cgroup is about to be freed and therefore has no charges
762 * left. Hence, all its caches must be empty by now.
763 */
764 BUG_ON(shutdown_cache(s));
765 }
766 mutex_unlock(&slab_mutex);
767
768 put_online_mems();
769 put_online_cpus();
770}
771
772static int shutdown_memcg_caches(struct kmem_cache *s)
773{
774 struct memcg_cache_array *arr;
775 struct kmem_cache *c, *c2;
776 LIST_HEAD(busy);
777 int i;
778
779 BUG_ON(!is_root_cache(s));
780
781 /*
782 * First, shutdown active caches, i.e. caches that belong to online
783 * memory cgroups.
784 */
785 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
786 lockdep_is_held(&slab_mutex));
787 for_each_memcg_cache_index(i) {
788 c = arr->entries[i];
789 if (!c)
790 continue;
791 if (shutdown_cache(c))
792 /*
793 * The cache still has objects. Move it to a temporary
794 * list so as not to try to destroy it for a second
795 * time while iterating over inactive caches below.
796 */
797 list_move(&c->memcg_params.children_node, &busy);
798 else
799 /*
800 * The cache is empty and will be destroyed soon. Clear
801 * the pointer to it in the memcg_caches array so that
802 * it will never be accessed even if the root cache
803 * stays alive.
804 */
805 arr->entries[i] = NULL;
806 }
807
808 /*
809 * Second, shutdown all caches left from memory cgroups that are now
810 * offline.
811 */
812 list_for_each_entry_safe(c, c2, &s->memcg_params.children,
813 memcg_params.children_node)
814 shutdown_cache(c);
815
816 list_splice(&busy, &s->memcg_params.children);
817
818 /*
819 * A cache being destroyed must be empty. In particular, this means
820 * that all per memcg caches attached to it must be empty too.
821 */
822 if (!list_empty(&s->memcg_params.children))
823 return -EBUSY;
824 return 0;
825}
826#else
827static inline int shutdown_memcg_caches(struct kmem_cache *s)
828{
829 return 0;
830}
831#endif /* CONFIG_MEMCG && !CONFIG_SLOB */
832
833void slab_kmem_cache_release(struct kmem_cache *s)
834{
835 __kmem_cache_release(s);
836 destroy_memcg_params(s);
837 kfree_const(s->name);
838 kmem_cache_free(kmem_cache, s);
839}
840
841void kmem_cache_destroy(struct kmem_cache *s)
842{
843 int err;
844
845 if (unlikely(!s))
846 return;
847
848 get_online_cpus();
849 get_online_mems();
850
851 mutex_lock(&slab_mutex);
852
853 s->refcount--;
854 if (s->refcount)
855 goto out_unlock;
856
857 err = shutdown_memcg_caches(s);
858 if (!err)
859 err = shutdown_cache(s);
860
861 if (err) {
862 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
863 s->name);
864 dump_stack();
865 }
866out_unlock:
867 mutex_unlock(&slab_mutex);
868
869 put_online_mems();
870 put_online_cpus();
871}
872EXPORT_SYMBOL(kmem_cache_destroy);
873
874/**
875 * kmem_cache_shrink - Shrink a cache.
876 * @cachep: The cache to shrink.
877 *
878 * Releases as many slabs as possible for a cache.
879 * To help debugging, a zero exit status indicates all slabs were released.
880 */
881int kmem_cache_shrink(struct kmem_cache *cachep)
882{
883 int ret;
884
885 get_online_cpus();
886 get_online_mems();
887 kasan_cache_shrink(cachep);
888 ret = __kmem_cache_shrink(cachep);
889 put_online_mems();
890 put_online_cpus();
891 return ret;
892}
893EXPORT_SYMBOL(kmem_cache_shrink);
894
895bool slab_is_available(void)
896{
897 return slab_state >= UP;
898}
899
900#ifndef CONFIG_SLOB
901/* Create a cache during boot when no slab services are available yet */
902void __init create_boot_cache(struct kmem_cache *s, const char *name,
903 unsigned int size, slab_flags_t flags,
904 unsigned int useroffset, unsigned int usersize)
905{
906 int err;
907
908 s->name = name;
909 s->size = s->object_size = size;
910 s->align = calculate_alignment(flags, ARCH_KMALLOC_MINALIGN, size);
911 s->useroffset = useroffset;
912 s->usersize = usersize;
913
914 slab_init_memcg_params(s);
915
916 err = __kmem_cache_create(s, flags);
917
918 if (err)
919 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
920 name, size, err);
921
922 s->refcount = -1; /* Exempt from merging for now */
923}
924
925struct kmem_cache *__init create_kmalloc_cache(const char *name,
926 unsigned int size, slab_flags_t flags,
927 unsigned int useroffset, unsigned int usersize)
928{
929 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
930
931 if (!s)
932 panic("Out of memory when creating slab %s\n", name);
933
934 create_boot_cache(s, name, size, flags, useroffset, usersize);
935 list_add(&s->list, &slab_caches);
936 memcg_link_cache(s);
937 s->refcount = 1;
938 return s;
939}
940
941struct kmem_cache *kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __ro_after_init;
942EXPORT_SYMBOL(kmalloc_caches);
943
944#ifdef CONFIG_ZONE_DMA
945struct kmem_cache *kmalloc_dma_caches[KMALLOC_SHIFT_HIGH + 1] __ro_after_init;
946EXPORT_SYMBOL(kmalloc_dma_caches);
947#endif
948
949/*
950 * Conversion table for small slabs sizes / 8 to the index in the
951 * kmalloc array. This is necessary for slabs < 192 since we have non power
952 * of two cache sizes there. The size of larger slabs can be determined using
953 * fls.
954 */
955static u8 size_index[24] __ro_after_init = {
956 3, /* 8 */
957 4, /* 16 */
958 5, /* 24 */
959 5, /* 32 */
960 6, /* 40 */
961 6, /* 48 */
962 6, /* 56 */
963 6, /* 64 */
964 1, /* 72 */
965 1, /* 80 */
966 1, /* 88 */
967 1, /* 96 */
968 7, /* 104 */
969 7, /* 112 */
970 7, /* 120 */
971 7, /* 128 */
972 2, /* 136 */
973 2, /* 144 */
974 2, /* 152 */
975 2, /* 160 */
976 2, /* 168 */
977 2, /* 176 */
978 2, /* 184 */
979 2 /* 192 */
980};
981
982static inline unsigned int size_index_elem(unsigned int bytes)
983{
984 return (bytes - 1) / 8;
985}
986
987/*
988 * Find the kmem_cache structure that serves a given size of
989 * allocation
990 */
991struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
992{
993 unsigned int index;
994
995 if (unlikely(size > KMALLOC_MAX_SIZE)) {
996 WARN_ON_ONCE(!(flags & __GFP_NOWARN));
997 return NULL;
998 }
999
1000 if (size <= 192) {
1001 if (!size)
1002 return ZERO_SIZE_PTR;
1003
1004 index = size_index[size_index_elem(size)];
1005 } else
1006 index = fls(size - 1);
1007
1008#ifdef CONFIG_ZONE_DMA
1009 if (unlikely((flags & GFP_DMA)))
1010 return kmalloc_dma_caches[index];
1011
1012#endif
1013 return kmalloc_caches[index];
1014}
1015
1016/*
1017 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
1018 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
1019 * kmalloc-67108864.
1020 */
1021const struct kmalloc_info_struct kmalloc_info[] __initconst = {
1022 {NULL, 0}, {"kmalloc-96", 96},
1023 {"kmalloc-192", 192}, {"kmalloc-8", 8},
1024 {"kmalloc-16", 16}, {"kmalloc-32", 32},
1025 {"kmalloc-64", 64}, {"kmalloc-128", 128},
1026 {"kmalloc-256", 256}, {"kmalloc-512", 512},
1027 {"kmalloc-1024", 1024}, {"kmalloc-2048", 2048},
1028 {"kmalloc-4096", 4096}, {"kmalloc-8192", 8192},
1029 {"kmalloc-16384", 16384}, {"kmalloc-32768", 32768},
1030 {"kmalloc-65536", 65536}, {"kmalloc-131072", 131072},
1031 {"kmalloc-262144", 262144}, {"kmalloc-524288", 524288},
1032 {"kmalloc-1048576", 1048576}, {"kmalloc-2097152", 2097152},
1033 {"kmalloc-4194304", 4194304}, {"kmalloc-8388608", 8388608},
1034 {"kmalloc-16777216", 16777216}, {"kmalloc-33554432", 33554432},
1035 {"kmalloc-67108864", 67108864}
1036};
1037
1038/*
1039 * Patch up the size_index table if we have strange large alignment
1040 * requirements for the kmalloc array. This is only the case for
1041 * MIPS it seems. The standard arches will not generate any code here.
1042 *
1043 * Largest permitted alignment is 256 bytes due to the way we
1044 * handle the index determination for the smaller caches.
1045 *
1046 * Make sure that nothing crazy happens if someone starts tinkering
1047 * around with ARCH_KMALLOC_MINALIGN
1048 */
1049void __init setup_kmalloc_cache_index_table(void)
1050{
1051 unsigned int i;
1052
1053 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
1054 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
1055
1056 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
1057 unsigned int elem = size_index_elem(i);
1058
1059 if (elem >= ARRAY_SIZE(size_index))
1060 break;
1061 size_index[elem] = KMALLOC_SHIFT_LOW;
1062 }
1063
1064 if (KMALLOC_MIN_SIZE >= 64) {
1065 /*
1066 * The 96 byte size cache is not used if the alignment
1067 * is 64 byte.
1068 */
1069 for (i = 64 + 8; i <= 96; i += 8)
1070 size_index[size_index_elem(i)] = 7;
1071
1072 }
1073
1074 if (KMALLOC_MIN_SIZE >= 128) {
1075 /*
1076 * The 192 byte sized cache is not used if the alignment
1077 * is 128 byte. Redirect kmalloc to use the 256 byte cache
1078 * instead.
1079 */
1080 for (i = 128 + 8; i <= 192; i += 8)
1081 size_index[size_index_elem(i)] = 8;
1082 }
1083}
1084
1085static void __init new_kmalloc_cache(int idx, slab_flags_t flags)
1086{
1087 kmalloc_caches[idx] = create_kmalloc_cache(kmalloc_info[idx].name,
1088 kmalloc_info[idx].size, flags, 0,
1089 kmalloc_info[idx].size);
1090}
1091
1092/*
1093 * Create the kmalloc array. Some of the regular kmalloc arrays
1094 * may already have been created because they were needed to
1095 * enable allocations for slab creation.
1096 */
1097void __init create_kmalloc_caches(slab_flags_t flags)
1098{
1099 int i;
1100
1101 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
1102 if (!kmalloc_caches[i])
1103 new_kmalloc_cache(i, flags);
1104
1105 /*
1106 * Caches that are not of the two-to-the-power-of size.
1107 * These have to be created immediately after the
1108 * earlier power of two caches
1109 */
1110 if (KMALLOC_MIN_SIZE <= 32 && !kmalloc_caches[1] && i == 6)
1111 new_kmalloc_cache(1, flags);
1112 if (KMALLOC_MIN_SIZE <= 64 && !kmalloc_caches[2] && i == 7)
1113 new_kmalloc_cache(2, flags);
1114 }
1115
1116 /* Kmalloc array is now usable */
1117 slab_state = UP;
1118
1119#ifdef CONFIG_ZONE_DMA
1120 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1121 struct kmem_cache *s = kmalloc_caches[i];
1122
1123 if (s) {
1124 unsigned int size = kmalloc_size(i);
1125 char *n = kasprintf(GFP_NOWAIT,
1126 "dma-kmalloc-%u", size);
1127
1128 BUG_ON(!n);
1129 kmalloc_dma_caches[i] = create_kmalloc_cache(n,
1130 size, SLAB_CACHE_DMA | flags, 0, 0);
1131 }
1132 }
1133#endif
1134}
1135#endif /* !CONFIG_SLOB */
1136
1137/*
1138 * To avoid unnecessary overhead, we pass through large allocation requests
1139 * directly to the page allocator. We use __GFP_COMP, because we will need to
1140 * know the allocation order to free the pages properly in kfree.
1141 */
1142void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1143{
1144 void *ret;
1145 struct page *page;
1146
1147 flags |= __GFP_COMP;
1148 page = alloc_pages(flags, order);
1149 ret = page ? page_address(page) : NULL;
1150 kmemleak_alloc(ret, size, 1, flags);
1151 kasan_kmalloc_large(ret, size, flags);
1152 return ret;
1153}
1154EXPORT_SYMBOL(kmalloc_order);
1155
1156#ifdef CONFIG_TRACING
1157void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1158{
1159 void *ret = kmalloc_order(size, flags, order);
1160 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1161 return ret;
1162}
1163EXPORT_SYMBOL(kmalloc_order_trace);
1164#endif
1165
1166#ifdef CONFIG_SLAB_FREELIST_RANDOM
1167/* Randomize a generic freelist */
1168static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1169 unsigned int count)
1170{
1171 unsigned int rand;
1172 unsigned int i;
1173
1174 for (i = 0; i < count; i++)
1175 list[i] = i;
1176
1177 /* Fisher-Yates shuffle */
1178 for (i = count - 1; i > 0; i--) {
1179 rand = prandom_u32_state(state);
1180 rand %= (i + 1);
1181 swap(list[i], list[rand]);
1182 }
1183}
1184
1185/* Create a random sequence per cache */
1186int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1187 gfp_t gfp)
1188{
1189 struct rnd_state state;
1190
1191 if (count < 2 || cachep->random_seq)
1192 return 0;
1193
1194 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1195 if (!cachep->random_seq)
1196 return -ENOMEM;
1197
1198 /* Get best entropy at this stage of boot */
1199 prandom_seed_state(&state, get_random_long());
1200
1201 freelist_randomize(&state, cachep->random_seq, count);
1202 return 0;
1203}
1204
1205/* Destroy the per-cache random freelist sequence */
1206void cache_random_seq_destroy(struct kmem_cache *cachep)
1207{
1208 kfree(cachep->random_seq);
1209 cachep->random_seq = NULL;
1210}
1211#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1212
1213#if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1214#ifdef CONFIG_SLAB
1215#define SLABINFO_RIGHTS (S_IWUSR | S_IRUSR)
1216#else
1217#define SLABINFO_RIGHTS S_IRUSR
1218#endif
1219
1220static void print_slabinfo_header(struct seq_file *m)
1221{
1222 /*
1223 * Output format version, so at least we can change it
1224 * without _too_ many complaints.
1225 */
1226#ifdef CONFIG_DEBUG_SLAB
1227 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1228#else
1229 seq_puts(m, "slabinfo - version: 2.1\n");
1230#endif
1231 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1232 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1233 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1234#ifdef CONFIG_DEBUG_SLAB
1235 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1236 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1237#endif
1238 seq_putc(m, '\n');
1239}
1240
1241void *slab_start(struct seq_file *m, loff_t *pos)
1242{
1243 mutex_lock(&slab_mutex);
1244 return seq_list_start(&slab_root_caches, *pos);
1245}
1246
1247void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1248{
1249 return seq_list_next(p, &slab_root_caches, pos);
1250}
1251
1252void slab_stop(struct seq_file *m, void *p)
1253{
1254 mutex_unlock(&slab_mutex);
1255}
1256
1257static void
1258memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1259{
1260 struct kmem_cache *c;
1261 struct slabinfo sinfo;
1262
1263 if (!is_root_cache(s))
1264 return;
1265
1266 for_each_memcg_cache(c, s) {
1267 memset(&sinfo, 0, sizeof(sinfo));
1268 get_slabinfo(c, &sinfo);
1269
1270 info->active_slabs += sinfo.active_slabs;
1271 info->num_slabs += sinfo.num_slabs;
1272 info->shared_avail += sinfo.shared_avail;
1273 info->active_objs += sinfo.active_objs;
1274 info->num_objs += sinfo.num_objs;
1275 }
1276}
1277
1278static void cache_show(struct kmem_cache *s, struct seq_file *m)
1279{
1280 struct slabinfo sinfo;
1281
1282 memset(&sinfo, 0, sizeof(sinfo));
1283 get_slabinfo(s, &sinfo);
1284
1285 memcg_accumulate_slabinfo(s, &sinfo);
1286
1287 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1288 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1289 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1290
1291 seq_printf(m, " : tunables %4u %4u %4u",
1292 sinfo.limit, sinfo.batchcount, sinfo.shared);
1293 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1294 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1295 slabinfo_show_stats(m, s);
1296 seq_putc(m, '\n');
1297}
1298
1299static int slab_show(struct seq_file *m, void *p)
1300{
1301 struct kmem_cache *s = list_entry(p, struct kmem_cache, root_caches_node);
1302
1303 if (p == slab_root_caches.next)
1304 print_slabinfo_header(m);
1305 cache_show(s, m);
1306 return 0;
1307}
1308
1309void dump_unreclaimable_slab(void)
1310{
1311 struct kmem_cache *s, *s2;
1312 struct slabinfo sinfo;
1313
1314 /*
1315 * Here acquiring slab_mutex is risky since we don't prefer to get
1316 * sleep in oom path. But, without mutex hold, it may introduce a
1317 * risk of crash.
1318 * Use mutex_trylock to protect the list traverse, dump nothing
1319 * without acquiring the mutex.
1320 */
1321 if (!mutex_trylock(&slab_mutex)) {
1322 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1323 return;
1324 }
1325
1326 pr_info("Unreclaimable slab info:\n");
1327 pr_info("Name Used Total\n");
1328
1329 list_for_each_entry_safe(s, s2, &slab_caches, list) {
1330 if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
1331 continue;
1332
1333 get_slabinfo(s, &sinfo);
1334
1335 if (sinfo.num_objs > 0)
1336 pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
1337 (sinfo.active_objs * s->size) / 1024,
1338 (sinfo.num_objs * s->size) / 1024);
1339 }
1340 mutex_unlock(&slab_mutex);
1341}
1342
1343#if defined(CONFIG_MEMCG)
1344void *memcg_slab_start(struct seq_file *m, loff_t *pos)
1345{
1346 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1347
1348 mutex_lock(&slab_mutex);
1349 return seq_list_start(&memcg->kmem_caches, *pos);
1350}
1351
1352void *memcg_slab_next(struct seq_file *m, void *p, loff_t *pos)
1353{
1354 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1355
1356 return seq_list_next(p, &memcg->kmem_caches, pos);
1357}
1358
1359void memcg_slab_stop(struct seq_file *m, void *p)
1360{
1361 mutex_unlock(&slab_mutex);
1362}
1363
1364int memcg_slab_show(struct seq_file *m, void *p)
1365{
1366 struct kmem_cache *s = list_entry(p, struct kmem_cache,
1367 memcg_params.kmem_caches_node);
1368 struct mem_cgroup *memcg = mem_cgroup_from_css(seq_css(m));
1369
1370 if (p == memcg->kmem_caches.next)
1371 print_slabinfo_header(m);
1372 cache_show(s, m);
1373 return 0;
1374}
1375#endif
1376
1377/*
1378 * slabinfo_op - iterator that generates /proc/slabinfo
1379 *
1380 * Output layout:
1381 * cache-name
1382 * num-active-objs
1383 * total-objs
1384 * object size
1385 * num-active-slabs
1386 * total-slabs
1387 * num-pages-per-slab
1388 * + further values on SMP and with statistics enabled
1389 */
1390static const struct seq_operations slabinfo_op = {
1391 .start = slab_start,
1392 .next = slab_next,
1393 .stop = slab_stop,
1394 .show = slab_show,
1395};
1396
1397static int slabinfo_open(struct inode *inode, struct file *file)
1398{
1399 return seq_open(file, &slabinfo_op);
1400}
1401
1402static const struct file_operations proc_slabinfo_operations = {
1403 .open = slabinfo_open,
1404 .read = seq_read,
1405 .write = slabinfo_write,
1406 .llseek = seq_lseek,
1407 .release = seq_release,
1408};
1409
1410static int __init slab_proc_init(void)
1411{
1412 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1413 &proc_slabinfo_operations);
1414 return 0;
1415}
1416module_init(slab_proc_init);
1417#endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1418
1419static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1420 gfp_t flags)
1421{
1422 void *ret;
1423 size_t ks = 0;
1424
1425 if (p)
1426 ks = ksize(p);
1427
1428 if (ks >= new_size) {
1429 kasan_krealloc((void *)p, new_size, flags);
1430 return (void *)p;
1431 }
1432
1433 ret = kmalloc_track_caller(new_size, flags);
1434 if (ret && p)
1435 memcpy(ret, p, ks);
1436
1437 return ret;
1438}
1439
1440/**
1441 * __krealloc - like krealloc() but don't free @p.
1442 * @p: object to reallocate memory for.
1443 * @new_size: how many bytes of memory are required.
1444 * @flags: the type of memory to allocate.
1445 *
1446 * This function is like krealloc() except it never frees the originally
1447 * allocated buffer. Use this if you don't want to free the buffer immediately
1448 * like, for example, with RCU.
1449 */
1450void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1451{
1452 if (unlikely(!new_size))
1453 return ZERO_SIZE_PTR;
1454
1455 return __do_krealloc(p, new_size, flags);
1456
1457}
1458EXPORT_SYMBOL(__krealloc);
1459
1460/**
1461 * krealloc - reallocate memory. The contents will remain unchanged.
1462 * @p: object to reallocate memory for.
1463 * @new_size: how many bytes of memory are required.
1464 * @flags: the type of memory to allocate.
1465 *
1466 * The contents of the object pointed to are preserved up to the
1467 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1468 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1469 * %NULL pointer, the object pointed to is freed.
1470 */
1471void *krealloc(const void *p, size_t new_size, gfp_t flags)
1472{
1473 void *ret;
1474
1475 if (unlikely(!new_size)) {
1476 kfree(p);
1477 return ZERO_SIZE_PTR;
1478 }
1479
1480 ret = __do_krealloc(p, new_size, flags);
1481 if (ret && p != ret)
1482 kfree(p);
1483
1484 return ret;
1485}
1486EXPORT_SYMBOL(krealloc);
1487
1488/**
1489 * kzfree - like kfree but zero memory
1490 * @p: object to free memory of
1491 *
1492 * The memory of the object @p points to is zeroed before freed.
1493 * If @p is %NULL, kzfree() does nothing.
1494 *
1495 * Note: this function zeroes the whole allocated buffer which can be a good
1496 * deal bigger than the requested buffer size passed to kmalloc(). So be
1497 * careful when using this function in performance sensitive code.
1498 */
1499void kzfree(const void *p)
1500{
1501 size_t ks;
1502 void *mem = (void *)p;
1503
1504 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1505 return;
1506 ks = ksize(mem);
1507 memset(mem, 0, ks);
1508 kfree(mem);
1509}
1510EXPORT_SYMBOL(kzfree);
1511
1512/* Tracepoints definitions. */
1513EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1514EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1515EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1516EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1517EXPORT_TRACEPOINT_SYMBOL(kfree);
1518EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1519
1520int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1521{
1522 if (__should_failslab(s, gfpflags))
1523 return -ENOMEM;
1524 return 0;
1525}
1526ALLOW_ERROR_INJECTION(should_failslab, ERRNO);