Loading...
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);
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/kfence.h>
16#include <linux/module.h>
17#include <linux/cpu.h>
18#include <linux/uaccess.h>
19#include <linux/seq_file.h>
20#include <linux/proc_fs.h>
21#include <linux/debugfs.h>
22#include <linux/kasan.h>
23#include <asm/cacheflush.h>
24#include <asm/tlbflush.h>
25#include <asm/page.h>
26#include <linux/memcontrol.h>
27
28#define CREATE_TRACE_POINTS
29#include <trace/events/kmem.h>
30
31#include "internal.h"
32
33#include "slab.h"
34
35enum slab_state slab_state;
36LIST_HEAD(slab_caches);
37DEFINE_MUTEX(slab_mutex);
38struct kmem_cache *kmem_cache;
39
40#ifdef CONFIG_HARDENED_USERCOPY
41bool usercopy_fallback __ro_after_init =
42 IS_ENABLED(CONFIG_HARDENED_USERCOPY_FALLBACK);
43module_param(usercopy_fallback, bool, 0400);
44MODULE_PARM_DESC(usercopy_fallback,
45 "WARN instead of reject usercopy whitelist violations");
46#endif
47
48static LIST_HEAD(slab_caches_to_rcu_destroy);
49static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
50static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
51 slab_caches_to_rcu_destroy_workfn);
52
53/*
54 * Set of flags that will prevent slab merging
55 */
56#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
57 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
58 SLAB_FAILSLAB | kasan_never_merge())
59
60#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
61 SLAB_CACHE_DMA32 | SLAB_ACCOUNT)
62
63/*
64 * Merge control. If this is set then no merging of slab caches will occur.
65 */
66static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
67
68static int __init setup_slab_nomerge(char *str)
69{
70 slab_nomerge = true;
71 return 1;
72}
73
74static int __init setup_slab_merge(char *str)
75{
76 slab_nomerge = false;
77 return 1;
78}
79
80#ifdef CONFIG_SLUB
81__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
82__setup_param("slub_merge", slub_merge, setup_slab_merge, 0);
83#endif
84
85__setup("slab_nomerge", setup_slab_nomerge);
86__setup("slab_merge", setup_slab_merge);
87
88/*
89 * Determine the size of a slab object
90 */
91unsigned int kmem_cache_size(struct kmem_cache *s)
92{
93 return s->object_size;
94}
95EXPORT_SYMBOL(kmem_cache_size);
96
97#ifdef CONFIG_DEBUG_VM
98static int kmem_cache_sanity_check(const char *name, unsigned int size)
99{
100 if (!name || in_interrupt() || size > KMALLOC_MAX_SIZE) {
101 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
102 return -EINVAL;
103 }
104
105 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
106 return 0;
107}
108#else
109static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
110{
111 return 0;
112}
113#endif
114
115void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
116{
117 size_t i;
118
119 for (i = 0; i < nr; i++) {
120 if (s)
121 kmem_cache_free(s, p[i]);
122 else
123 kfree(p[i]);
124 }
125}
126
127int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
128 void **p)
129{
130 size_t i;
131
132 for (i = 0; i < nr; i++) {
133 void *x = p[i] = kmem_cache_alloc(s, flags);
134 if (!x) {
135 __kmem_cache_free_bulk(s, i, p);
136 return 0;
137 }
138 }
139 return i;
140}
141
142/*
143 * Figure out what the alignment of the objects will be given a set of
144 * flags, a user specified alignment and the size of the objects.
145 */
146static unsigned int calculate_alignment(slab_flags_t flags,
147 unsigned int align, unsigned int size)
148{
149 /*
150 * If the user wants hardware cache aligned objects then follow that
151 * suggestion if the object is sufficiently large.
152 *
153 * The hardware cache alignment cannot override the specified
154 * alignment though. If that is greater then use it.
155 */
156 if (flags & SLAB_HWCACHE_ALIGN) {
157 unsigned int ralign;
158
159 ralign = cache_line_size();
160 while (size <= ralign / 2)
161 ralign /= 2;
162 align = max(align, ralign);
163 }
164
165 if (align < ARCH_SLAB_MINALIGN)
166 align = ARCH_SLAB_MINALIGN;
167
168 return ALIGN(align, sizeof(void *));
169}
170
171/*
172 * Find a mergeable slab cache
173 */
174int slab_unmergeable(struct kmem_cache *s)
175{
176 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
177 return 1;
178
179 if (s->ctor)
180 return 1;
181
182 if (s->usersize)
183 return 1;
184
185 /*
186 * We may have set a slab to be unmergeable during bootstrap.
187 */
188 if (s->refcount < 0)
189 return 1;
190
191 return 0;
192}
193
194struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
195 slab_flags_t flags, const char *name, void (*ctor)(void *))
196{
197 struct kmem_cache *s;
198
199 if (slab_nomerge)
200 return NULL;
201
202 if (ctor)
203 return NULL;
204
205 size = ALIGN(size, sizeof(void *));
206 align = calculate_alignment(flags, align, size);
207 size = ALIGN(size, align);
208 flags = kmem_cache_flags(size, flags, name);
209
210 if (flags & SLAB_NEVER_MERGE)
211 return NULL;
212
213 list_for_each_entry_reverse(s, &slab_caches, list) {
214 if (slab_unmergeable(s))
215 continue;
216
217 if (size > s->size)
218 continue;
219
220 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
221 continue;
222 /*
223 * Check if alignment is compatible.
224 * Courtesy of Adrian Drzewiecki
225 */
226 if ((s->size & ~(align - 1)) != s->size)
227 continue;
228
229 if (s->size - size >= sizeof(void *))
230 continue;
231
232 if (IS_ENABLED(CONFIG_SLAB) && align &&
233 (align > s->align || s->align % align))
234 continue;
235
236 return s;
237 }
238 return NULL;
239}
240
241static struct kmem_cache *create_cache(const char *name,
242 unsigned int object_size, unsigned int align,
243 slab_flags_t flags, unsigned int useroffset,
244 unsigned int usersize, void (*ctor)(void *),
245 struct kmem_cache *root_cache)
246{
247 struct kmem_cache *s;
248 int err;
249
250 if (WARN_ON(useroffset + usersize > object_size))
251 useroffset = usersize = 0;
252
253 err = -ENOMEM;
254 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
255 if (!s)
256 goto out;
257
258 s->name = name;
259 s->size = s->object_size = object_size;
260 s->align = align;
261 s->ctor = ctor;
262 s->useroffset = useroffset;
263 s->usersize = usersize;
264
265 err = __kmem_cache_create(s, flags);
266 if (err)
267 goto out_free_cache;
268
269 s->refcount = 1;
270 list_add(&s->list, &slab_caches);
271out:
272 if (err)
273 return ERR_PTR(err);
274 return s;
275
276out_free_cache:
277 kmem_cache_free(kmem_cache, s);
278 goto out;
279}
280
281/**
282 * kmem_cache_create_usercopy - Create a cache with a region suitable
283 * for copying to userspace
284 * @name: A string which is used in /proc/slabinfo to identify this cache.
285 * @size: The size of objects to be created in this cache.
286 * @align: The required alignment for the objects.
287 * @flags: SLAB flags
288 * @useroffset: Usercopy region offset
289 * @usersize: Usercopy region size
290 * @ctor: A constructor for the objects.
291 *
292 * Cannot be called within a interrupt, but can be interrupted.
293 * The @ctor is run when new pages are allocated by the cache.
294 *
295 * The flags are
296 *
297 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
298 * to catch references to uninitialised memory.
299 *
300 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
301 * for buffer overruns.
302 *
303 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
304 * cacheline. This can be beneficial if you're counting cycles as closely
305 * as davem.
306 *
307 * Return: a pointer to the cache on success, NULL on failure.
308 */
309struct kmem_cache *
310kmem_cache_create_usercopy(const char *name,
311 unsigned int size, unsigned int align,
312 slab_flags_t flags,
313 unsigned int useroffset, unsigned int usersize,
314 void (*ctor)(void *))
315{
316 struct kmem_cache *s = NULL;
317 const char *cache_name;
318 int err;
319
320#ifdef CONFIG_SLUB_DEBUG
321 /*
322 * If no slub_debug was enabled globally, the static key is not yet
323 * enabled by setup_slub_debug(). Enable it if the cache is being
324 * created with any of the debugging flags passed explicitly.
325 */
326 if (flags & SLAB_DEBUG_FLAGS)
327 static_branch_enable(&slub_debug_enabled);
328#endif
329
330 mutex_lock(&slab_mutex);
331
332 err = kmem_cache_sanity_check(name, size);
333 if (err) {
334 goto out_unlock;
335 }
336
337 /* Refuse requests with allocator specific flags */
338 if (flags & ~SLAB_FLAGS_PERMITTED) {
339 err = -EINVAL;
340 goto out_unlock;
341 }
342
343 /*
344 * Some allocators will constraint the set of valid flags to a subset
345 * of all flags. We expect them to define CACHE_CREATE_MASK in this
346 * case, and we'll just provide them with a sanitized version of the
347 * passed flags.
348 */
349 flags &= CACHE_CREATE_MASK;
350
351 /* Fail closed on bad usersize of useroffset values. */
352 if (WARN_ON(!usersize && useroffset) ||
353 WARN_ON(size < usersize || size - usersize < useroffset))
354 usersize = useroffset = 0;
355
356 if (!usersize)
357 s = __kmem_cache_alias(name, size, align, flags, ctor);
358 if (s)
359 goto out_unlock;
360
361 cache_name = kstrdup_const(name, GFP_KERNEL);
362 if (!cache_name) {
363 err = -ENOMEM;
364 goto out_unlock;
365 }
366
367 s = create_cache(cache_name, size,
368 calculate_alignment(flags, align, size),
369 flags, useroffset, usersize, ctor, NULL);
370 if (IS_ERR(s)) {
371 err = PTR_ERR(s);
372 kfree_const(cache_name);
373 }
374
375out_unlock:
376 mutex_unlock(&slab_mutex);
377
378 if (err) {
379 if (flags & SLAB_PANIC)
380 panic("%s: Failed to create slab '%s'. Error %d\n",
381 __func__, name, err);
382 else {
383 pr_warn("%s(%s) failed with error %d\n",
384 __func__, name, err);
385 dump_stack();
386 }
387 return NULL;
388 }
389 return s;
390}
391EXPORT_SYMBOL(kmem_cache_create_usercopy);
392
393/**
394 * kmem_cache_create - Create a cache.
395 * @name: A string which is used in /proc/slabinfo to identify this cache.
396 * @size: The size of objects to be created in this cache.
397 * @align: The required alignment for the objects.
398 * @flags: SLAB flags
399 * @ctor: A constructor for the objects.
400 *
401 * Cannot be called within a interrupt, but can be interrupted.
402 * The @ctor is run when new pages are allocated by the cache.
403 *
404 * The flags are
405 *
406 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
407 * to catch references to uninitialised memory.
408 *
409 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
410 * for buffer overruns.
411 *
412 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
413 * cacheline. This can be beneficial if you're counting cycles as closely
414 * as davem.
415 *
416 * Return: a pointer to the cache on success, NULL on failure.
417 */
418struct kmem_cache *
419kmem_cache_create(const char *name, unsigned int size, unsigned int align,
420 slab_flags_t flags, void (*ctor)(void *))
421{
422 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
423 ctor);
424}
425EXPORT_SYMBOL(kmem_cache_create);
426
427static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
428{
429 LIST_HEAD(to_destroy);
430 struct kmem_cache *s, *s2;
431
432 /*
433 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
434 * @slab_caches_to_rcu_destroy list. The slab pages are freed
435 * through RCU and the associated kmem_cache are dereferenced
436 * while freeing the pages, so the kmem_caches should be freed only
437 * after the pending RCU operations are finished. As rcu_barrier()
438 * is a pretty slow operation, we batch all pending destructions
439 * asynchronously.
440 */
441 mutex_lock(&slab_mutex);
442 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
443 mutex_unlock(&slab_mutex);
444
445 if (list_empty(&to_destroy))
446 return;
447
448 rcu_barrier();
449
450 list_for_each_entry_safe(s, s2, &to_destroy, list) {
451 debugfs_slab_release(s);
452 kfence_shutdown_cache(s);
453#ifdef SLAB_SUPPORTS_SYSFS
454 sysfs_slab_release(s);
455#else
456 slab_kmem_cache_release(s);
457#endif
458 }
459}
460
461static int shutdown_cache(struct kmem_cache *s)
462{
463 /* free asan quarantined objects */
464 kasan_cache_shutdown(s);
465
466 if (__kmem_cache_shutdown(s) != 0)
467 return -EBUSY;
468
469 list_del(&s->list);
470
471 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
472#ifdef SLAB_SUPPORTS_SYSFS
473 sysfs_slab_unlink(s);
474#endif
475 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
476 schedule_work(&slab_caches_to_rcu_destroy_work);
477 } else {
478 kfence_shutdown_cache(s);
479 debugfs_slab_release(s);
480#ifdef SLAB_SUPPORTS_SYSFS
481 sysfs_slab_unlink(s);
482 sysfs_slab_release(s);
483#else
484 slab_kmem_cache_release(s);
485#endif
486 }
487
488 return 0;
489}
490
491void slab_kmem_cache_release(struct kmem_cache *s)
492{
493 __kmem_cache_release(s);
494 kfree_const(s->name);
495 kmem_cache_free(kmem_cache, s);
496}
497
498void kmem_cache_destroy(struct kmem_cache *s)
499{
500 int err;
501
502 if (unlikely(!s))
503 return;
504
505 mutex_lock(&slab_mutex);
506
507 s->refcount--;
508 if (s->refcount)
509 goto out_unlock;
510
511 err = shutdown_cache(s);
512 if (err) {
513 pr_err("%s %s: Slab cache still has objects\n",
514 __func__, s->name);
515 dump_stack();
516 }
517out_unlock:
518 mutex_unlock(&slab_mutex);
519}
520EXPORT_SYMBOL(kmem_cache_destroy);
521
522/**
523 * kmem_cache_shrink - Shrink a cache.
524 * @cachep: The cache to shrink.
525 *
526 * Releases as many slabs as possible for a cache.
527 * To help debugging, a zero exit status indicates all slabs were released.
528 *
529 * Return: %0 if all slabs were released, non-zero otherwise
530 */
531int kmem_cache_shrink(struct kmem_cache *cachep)
532{
533 int ret;
534
535
536 kasan_cache_shrink(cachep);
537 ret = __kmem_cache_shrink(cachep);
538
539 return ret;
540}
541EXPORT_SYMBOL(kmem_cache_shrink);
542
543bool slab_is_available(void)
544{
545 return slab_state >= UP;
546}
547
548#ifdef CONFIG_PRINTK
549/**
550 * kmem_valid_obj - does the pointer reference a valid slab object?
551 * @object: pointer to query.
552 *
553 * Return: %true if the pointer is to a not-yet-freed object from
554 * kmalloc() or kmem_cache_alloc(), either %true or %false if the pointer
555 * is to an already-freed object, and %false otherwise.
556 */
557bool kmem_valid_obj(void *object)
558{
559 struct page *page;
560
561 /* Some arches consider ZERO_SIZE_PTR to be a valid address. */
562 if (object < (void *)PAGE_SIZE || !virt_addr_valid(object))
563 return false;
564 page = virt_to_head_page(object);
565 return PageSlab(page);
566}
567EXPORT_SYMBOL_GPL(kmem_valid_obj);
568
569/**
570 * kmem_dump_obj - Print available slab provenance information
571 * @object: slab object for which to find provenance information.
572 *
573 * This function uses pr_cont(), so that the caller is expected to have
574 * printed out whatever preamble is appropriate. The provenance information
575 * depends on the type of object and on how much debugging is enabled.
576 * For a slab-cache object, the fact that it is a slab object is printed,
577 * and, if available, the slab name, return address, and stack trace from
578 * the allocation and last free path of that object.
579 *
580 * This function will splat if passed a pointer to a non-slab object.
581 * If you are not sure what type of object you have, you should instead
582 * use mem_dump_obj().
583 */
584void kmem_dump_obj(void *object)
585{
586 char *cp = IS_ENABLED(CONFIG_MMU) ? "" : "/vmalloc";
587 int i;
588 struct page *page;
589 unsigned long ptroffset;
590 struct kmem_obj_info kp = { };
591
592 if (WARN_ON_ONCE(!virt_addr_valid(object)))
593 return;
594 page = virt_to_head_page(object);
595 if (WARN_ON_ONCE(!PageSlab(page))) {
596 pr_cont(" non-slab memory.\n");
597 return;
598 }
599 kmem_obj_info(&kp, object, page);
600 if (kp.kp_slab_cache)
601 pr_cont(" slab%s %s", cp, kp.kp_slab_cache->name);
602 else
603 pr_cont(" slab%s", cp);
604 if (kp.kp_objp)
605 pr_cont(" start %px", kp.kp_objp);
606 if (kp.kp_data_offset)
607 pr_cont(" data offset %lu", kp.kp_data_offset);
608 if (kp.kp_objp) {
609 ptroffset = ((char *)object - (char *)kp.kp_objp) - kp.kp_data_offset;
610 pr_cont(" pointer offset %lu", ptroffset);
611 }
612 if (kp.kp_slab_cache && kp.kp_slab_cache->usersize)
613 pr_cont(" size %u", kp.kp_slab_cache->usersize);
614 if (kp.kp_ret)
615 pr_cont(" allocated at %pS\n", kp.kp_ret);
616 else
617 pr_cont("\n");
618 for (i = 0; i < ARRAY_SIZE(kp.kp_stack); i++) {
619 if (!kp.kp_stack[i])
620 break;
621 pr_info(" %pS\n", kp.kp_stack[i]);
622 }
623
624 if (kp.kp_free_stack[0])
625 pr_cont(" Free path:\n");
626
627 for (i = 0; i < ARRAY_SIZE(kp.kp_free_stack); i++) {
628 if (!kp.kp_free_stack[i])
629 break;
630 pr_info(" %pS\n", kp.kp_free_stack[i]);
631 }
632
633}
634EXPORT_SYMBOL_GPL(kmem_dump_obj);
635#endif
636
637#ifndef CONFIG_SLOB
638/* Create a cache during boot when no slab services are available yet */
639void __init create_boot_cache(struct kmem_cache *s, const char *name,
640 unsigned int size, slab_flags_t flags,
641 unsigned int useroffset, unsigned int usersize)
642{
643 int err;
644 unsigned int align = ARCH_KMALLOC_MINALIGN;
645
646 s->name = name;
647 s->size = s->object_size = size;
648
649 /*
650 * For power of two sizes, guarantee natural alignment for kmalloc
651 * caches, regardless of SL*B debugging options.
652 */
653 if (is_power_of_2(size))
654 align = max(align, size);
655 s->align = calculate_alignment(flags, align, size);
656
657 s->useroffset = useroffset;
658 s->usersize = usersize;
659
660 err = __kmem_cache_create(s, flags);
661
662 if (err)
663 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
664 name, size, err);
665
666 s->refcount = -1; /* Exempt from merging for now */
667}
668
669struct kmem_cache *__init create_kmalloc_cache(const char *name,
670 unsigned int size, slab_flags_t flags,
671 unsigned int useroffset, unsigned int usersize)
672{
673 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
674
675 if (!s)
676 panic("Out of memory when creating slab %s\n", name);
677
678 create_boot_cache(s, name, size, flags, useroffset, usersize);
679 kasan_cache_create_kmalloc(s);
680 list_add(&s->list, &slab_caches);
681 s->refcount = 1;
682 return s;
683}
684
685struct kmem_cache *
686kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] __ro_after_init =
687{ /* initialization for https://bugs.llvm.org/show_bug.cgi?id=42570 */ };
688EXPORT_SYMBOL(kmalloc_caches);
689
690/*
691 * Conversion table for small slabs sizes / 8 to the index in the
692 * kmalloc array. This is necessary for slabs < 192 since we have non power
693 * of two cache sizes there. The size of larger slabs can be determined using
694 * fls.
695 */
696static u8 size_index[24] __ro_after_init = {
697 3, /* 8 */
698 4, /* 16 */
699 5, /* 24 */
700 5, /* 32 */
701 6, /* 40 */
702 6, /* 48 */
703 6, /* 56 */
704 6, /* 64 */
705 1, /* 72 */
706 1, /* 80 */
707 1, /* 88 */
708 1, /* 96 */
709 7, /* 104 */
710 7, /* 112 */
711 7, /* 120 */
712 7, /* 128 */
713 2, /* 136 */
714 2, /* 144 */
715 2, /* 152 */
716 2, /* 160 */
717 2, /* 168 */
718 2, /* 176 */
719 2, /* 184 */
720 2 /* 192 */
721};
722
723static inline unsigned int size_index_elem(unsigned int bytes)
724{
725 return (bytes - 1) / 8;
726}
727
728/*
729 * Find the kmem_cache structure that serves a given size of
730 * allocation
731 */
732struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
733{
734 unsigned int index;
735
736 if (size <= 192) {
737 if (!size)
738 return ZERO_SIZE_PTR;
739
740 index = size_index[size_index_elem(size)];
741 } else {
742 if (WARN_ON_ONCE(size > KMALLOC_MAX_CACHE_SIZE))
743 return NULL;
744 index = fls(size - 1);
745 }
746
747 return kmalloc_caches[kmalloc_type(flags)][index];
748}
749
750#ifdef CONFIG_ZONE_DMA
751#define KMALLOC_DMA_NAME(sz) .name[KMALLOC_DMA] = "dma-kmalloc-" #sz,
752#else
753#define KMALLOC_DMA_NAME(sz)
754#endif
755
756#ifdef CONFIG_MEMCG_KMEM
757#define KMALLOC_CGROUP_NAME(sz) .name[KMALLOC_CGROUP] = "kmalloc-cg-" #sz,
758#else
759#define KMALLOC_CGROUP_NAME(sz)
760#endif
761
762#define INIT_KMALLOC_INFO(__size, __short_size) \
763{ \
764 .name[KMALLOC_NORMAL] = "kmalloc-" #__short_size, \
765 .name[KMALLOC_RECLAIM] = "kmalloc-rcl-" #__short_size, \
766 KMALLOC_CGROUP_NAME(__short_size) \
767 KMALLOC_DMA_NAME(__short_size) \
768 .size = __size, \
769}
770
771/*
772 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
773 * kmalloc_index() supports up to 2^25=32MB, so the final entry of the table is
774 * kmalloc-32M.
775 */
776const struct kmalloc_info_struct kmalloc_info[] __initconst = {
777 INIT_KMALLOC_INFO(0, 0),
778 INIT_KMALLOC_INFO(96, 96),
779 INIT_KMALLOC_INFO(192, 192),
780 INIT_KMALLOC_INFO(8, 8),
781 INIT_KMALLOC_INFO(16, 16),
782 INIT_KMALLOC_INFO(32, 32),
783 INIT_KMALLOC_INFO(64, 64),
784 INIT_KMALLOC_INFO(128, 128),
785 INIT_KMALLOC_INFO(256, 256),
786 INIT_KMALLOC_INFO(512, 512),
787 INIT_KMALLOC_INFO(1024, 1k),
788 INIT_KMALLOC_INFO(2048, 2k),
789 INIT_KMALLOC_INFO(4096, 4k),
790 INIT_KMALLOC_INFO(8192, 8k),
791 INIT_KMALLOC_INFO(16384, 16k),
792 INIT_KMALLOC_INFO(32768, 32k),
793 INIT_KMALLOC_INFO(65536, 64k),
794 INIT_KMALLOC_INFO(131072, 128k),
795 INIT_KMALLOC_INFO(262144, 256k),
796 INIT_KMALLOC_INFO(524288, 512k),
797 INIT_KMALLOC_INFO(1048576, 1M),
798 INIT_KMALLOC_INFO(2097152, 2M),
799 INIT_KMALLOC_INFO(4194304, 4M),
800 INIT_KMALLOC_INFO(8388608, 8M),
801 INIT_KMALLOC_INFO(16777216, 16M),
802 INIT_KMALLOC_INFO(33554432, 32M)
803};
804
805/*
806 * Patch up the size_index table if we have strange large alignment
807 * requirements for the kmalloc array. This is only the case for
808 * MIPS it seems. The standard arches will not generate any code here.
809 *
810 * Largest permitted alignment is 256 bytes due to the way we
811 * handle the index determination for the smaller caches.
812 *
813 * Make sure that nothing crazy happens if someone starts tinkering
814 * around with ARCH_KMALLOC_MINALIGN
815 */
816void __init setup_kmalloc_cache_index_table(void)
817{
818 unsigned int i;
819
820 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
821 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
822
823 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
824 unsigned int elem = size_index_elem(i);
825
826 if (elem >= ARRAY_SIZE(size_index))
827 break;
828 size_index[elem] = KMALLOC_SHIFT_LOW;
829 }
830
831 if (KMALLOC_MIN_SIZE >= 64) {
832 /*
833 * The 96 byte size cache is not used if the alignment
834 * is 64 byte.
835 */
836 for (i = 64 + 8; i <= 96; i += 8)
837 size_index[size_index_elem(i)] = 7;
838
839 }
840
841 if (KMALLOC_MIN_SIZE >= 128) {
842 /*
843 * The 192 byte sized cache is not used if the alignment
844 * is 128 byte. Redirect kmalloc to use the 256 byte cache
845 * instead.
846 */
847 for (i = 128 + 8; i <= 192; i += 8)
848 size_index[size_index_elem(i)] = 8;
849 }
850}
851
852static void __init
853new_kmalloc_cache(int idx, enum kmalloc_cache_type type, slab_flags_t flags)
854{
855 if (type == KMALLOC_RECLAIM) {
856 flags |= SLAB_RECLAIM_ACCOUNT;
857 } else if (IS_ENABLED(CONFIG_MEMCG_KMEM) && (type == KMALLOC_CGROUP)) {
858 if (cgroup_memory_nokmem) {
859 kmalloc_caches[type][idx] = kmalloc_caches[KMALLOC_NORMAL][idx];
860 return;
861 }
862 flags |= SLAB_ACCOUNT;
863 }
864
865 kmalloc_caches[type][idx] = create_kmalloc_cache(
866 kmalloc_info[idx].name[type],
867 kmalloc_info[idx].size, flags, 0,
868 kmalloc_info[idx].size);
869
870 /*
871 * If CONFIG_MEMCG_KMEM is enabled, disable cache merging for
872 * KMALLOC_NORMAL caches.
873 */
874 if (IS_ENABLED(CONFIG_MEMCG_KMEM) && (type == KMALLOC_NORMAL))
875 kmalloc_caches[type][idx]->refcount = -1;
876}
877
878/*
879 * Create the kmalloc array. Some of the regular kmalloc arrays
880 * may already have been created because they were needed to
881 * enable allocations for slab creation.
882 */
883void __init create_kmalloc_caches(slab_flags_t flags)
884{
885 int i;
886 enum kmalloc_cache_type type;
887
888 /*
889 * Including KMALLOC_CGROUP if CONFIG_MEMCG_KMEM defined
890 */
891 for (type = KMALLOC_NORMAL; type <= KMALLOC_RECLAIM; type++) {
892 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
893 if (!kmalloc_caches[type][i])
894 new_kmalloc_cache(i, type, flags);
895
896 /*
897 * Caches that are not of the two-to-the-power-of size.
898 * These have to be created immediately after the
899 * earlier power of two caches
900 */
901 if (KMALLOC_MIN_SIZE <= 32 && i == 6 &&
902 !kmalloc_caches[type][1])
903 new_kmalloc_cache(1, type, flags);
904 if (KMALLOC_MIN_SIZE <= 64 && i == 7 &&
905 !kmalloc_caches[type][2])
906 new_kmalloc_cache(2, type, flags);
907 }
908 }
909
910 /* Kmalloc array is now usable */
911 slab_state = UP;
912
913#ifdef CONFIG_ZONE_DMA
914 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
915 struct kmem_cache *s = kmalloc_caches[KMALLOC_NORMAL][i];
916
917 if (s) {
918 kmalloc_caches[KMALLOC_DMA][i] = create_kmalloc_cache(
919 kmalloc_info[i].name[KMALLOC_DMA],
920 kmalloc_info[i].size,
921 SLAB_CACHE_DMA | flags, 0,
922 kmalloc_info[i].size);
923 }
924 }
925#endif
926}
927#endif /* !CONFIG_SLOB */
928
929gfp_t kmalloc_fix_flags(gfp_t flags)
930{
931 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
932
933 flags &= ~GFP_SLAB_BUG_MASK;
934 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
935 invalid_mask, &invalid_mask, flags, &flags);
936 dump_stack();
937
938 return flags;
939}
940
941/*
942 * To avoid unnecessary overhead, we pass through large allocation requests
943 * directly to the page allocator. We use __GFP_COMP, because we will need to
944 * know the allocation order to free the pages properly in kfree.
945 */
946void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
947{
948 void *ret = NULL;
949 struct page *page;
950
951 if (unlikely(flags & GFP_SLAB_BUG_MASK))
952 flags = kmalloc_fix_flags(flags);
953
954 flags |= __GFP_COMP;
955 page = alloc_pages(flags, order);
956 if (likely(page)) {
957 ret = page_address(page);
958 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
959 PAGE_SIZE << order);
960 }
961 ret = kasan_kmalloc_large(ret, size, flags);
962 /* As ret might get tagged, call kmemleak hook after KASAN. */
963 kmemleak_alloc(ret, size, 1, flags);
964 return ret;
965}
966EXPORT_SYMBOL(kmalloc_order);
967
968#ifdef CONFIG_TRACING
969void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
970{
971 void *ret = kmalloc_order(size, flags, order);
972 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
973 return ret;
974}
975EXPORT_SYMBOL(kmalloc_order_trace);
976#endif
977
978#ifdef CONFIG_SLAB_FREELIST_RANDOM
979/* Randomize a generic freelist */
980static void freelist_randomize(struct rnd_state *state, unsigned int *list,
981 unsigned int count)
982{
983 unsigned int rand;
984 unsigned int i;
985
986 for (i = 0; i < count; i++)
987 list[i] = i;
988
989 /* Fisher-Yates shuffle */
990 for (i = count - 1; i > 0; i--) {
991 rand = prandom_u32_state(state);
992 rand %= (i + 1);
993 swap(list[i], list[rand]);
994 }
995}
996
997/* Create a random sequence per cache */
998int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
999 gfp_t gfp)
1000{
1001 struct rnd_state state;
1002
1003 if (count < 2 || cachep->random_seq)
1004 return 0;
1005
1006 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1007 if (!cachep->random_seq)
1008 return -ENOMEM;
1009
1010 /* Get best entropy at this stage of boot */
1011 prandom_seed_state(&state, get_random_long());
1012
1013 freelist_randomize(&state, cachep->random_seq, count);
1014 return 0;
1015}
1016
1017/* Destroy the per-cache random freelist sequence */
1018void cache_random_seq_destroy(struct kmem_cache *cachep)
1019{
1020 kfree(cachep->random_seq);
1021 cachep->random_seq = NULL;
1022}
1023#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1024
1025#if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1026#ifdef CONFIG_SLAB
1027#define SLABINFO_RIGHTS (0600)
1028#else
1029#define SLABINFO_RIGHTS (0400)
1030#endif
1031
1032static void print_slabinfo_header(struct seq_file *m)
1033{
1034 /*
1035 * Output format version, so at least we can change it
1036 * without _too_ many complaints.
1037 */
1038#ifdef CONFIG_DEBUG_SLAB
1039 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1040#else
1041 seq_puts(m, "slabinfo - version: 2.1\n");
1042#endif
1043 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1044 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1045 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1046#ifdef CONFIG_DEBUG_SLAB
1047 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1048 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1049#endif
1050 seq_putc(m, '\n');
1051}
1052
1053void *slab_start(struct seq_file *m, loff_t *pos)
1054{
1055 mutex_lock(&slab_mutex);
1056 return seq_list_start(&slab_caches, *pos);
1057}
1058
1059void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1060{
1061 return seq_list_next(p, &slab_caches, pos);
1062}
1063
1064void slab_stop(struct seq_file *m, void *p)
1065{
1066 mutex_unlock(&slab_mutex);
1067}
1068
1069static void cache_show(struct kmem_cache *s, struct seq_file *m)
1070{
1071 struct slabinfo sinfo;
1072
1073 memset(&sinfo, 0, sizeof(sinfo));
1074 get_slabinfo(s, &sinfo);
1075
1076 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1077 s->name, sinfo.active_objs, sinfo.num_objs, s->size,
1078 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1079
1080 seq_printf(m, " : tunables %4u %4u %4u",
1081 sinfo.limit, sinfo.batchcount, sinfo.shared);
1082 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1083 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1084 slabinfo_show_stats(m, s);
1085 seq_putc(m, '\n');
1086}
1087
1088static int slab_show(struct seq_file *m, void *p)
1089{
1090 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1091
1092 if (p == slab_caches.next)
1093 print_slabinfo_header(m);
1094 cache_show(s, m);
1095 return 0;
1096}
1097
1098void dump_unreclaimable_slab(void)
1099{
1100 struct kmem_cache *s;
1101 struct slabinfo sinfo;
1102
1103 /*
1104 * Here acquiring slab_mutex is risky since we don't prefer to get
1105 * sleep in oom path. But, without mutex hold, it may introduce a
1106 * risk of crash.
1107 * Use mutex_trylock to protect the list traverse, dump nothing
1108 * without acquiring the mutex.
1109 */
1110 if (!mutex_trylock(&slab_mutex)) {
1111 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1112 return;
1113 }
1114
1115 pr_info("Unreclaimable slab info:\n");
1116 pr_info("Name Used Total\n");
1117
1118 list_for_each_entry(s, &slab_caches, list) {
1119 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1120 continue;
1121
1122 get_slabinfo(s, &sinfo);
1123
1124 if (sinfo.num_objs > 0)
1125 pr_info("%-17s %10luKB %10luKB\n", s->name,
1126 (sinfo.active_objs * s->size) / 1024,
1127 (sinfo.num_objs * s->size) / 1024);
1128 }
1129 mutex_unlock(&slab_mutex);
1130}
1131
1132#if defined(CONFIG_MEMCG_KMEM)
1133int memcg_slab_show(struct seq_file *m, void *p)
1134{
1135 /*
1136 * Deprecated.
1137 * Please, take a look at tools/cgroup/slabinfo.py .
1138 */
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 proc_ops slabinfo_proc_ops = {
1169 .proc_flags = PROC_ENTRY_PERMANENT,
1170 .proc_open = slabinfo_open,
1171 .proc_read = seq_read,
1172 .proc_write = slabinfo_write,
1173 .proc_lseek = seq_lseek,
1174 .proc_release = seq_release,
1175};
1176
1177static int __init slab_proc_init(void)
1178{
1179 proc_create("slabinfo", SLABINFO_RIGHTS, NULL, &slabinfo_proc_ops);
1180 return 0;
1181}
1182module_init(slab_proc_init);
1183
1184#endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1185
1186static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1187 gfp_t flags)
1188{
1189 void *ret;
1190 size_t ks;
1191
1192 /* Don't use instrumented ksize to allow precise KASAN poisoning. */
1193 if (likely(!ZERO_OR_NULL_PTR(p))) {
1194 if (!kasan_check_byte(p))
1195 return NULL;
1196 ks = kfence_ksize(p) ?: __ksize(p);
1197 } else
1198 ks = 0;
1199
1200 /* If the object still fits, repoison it precisely. */
1201 if (ks >= new_size) {
1202 p = kasan_krealloc((void *)p, new_size, flags);
1203 return (void *)p;
1204 }
1205
1206 ret = kmalloc_track_caller(new_size, flags);
1207 if (ret && p) {
1208 /* Disable KASAN checks as the object's redzone is accessed. */
1209 kasan_disable_current();
1210 memcpy(ret, kasan_reset_tag(p), ks);
1211 kasan_enable_current();
1212 }
1213
1214 return ret;
1215}
1216
1217/**
1218 * krealloc - reallocate memory. The contents will remain unchanged.
1219 * @p: object to reallocate memory for.
1220 * @new_size: how many bytes of memory are required.
1221 * @flags: the type of memory to allocate.
1222 *
1223 * The contents of the object pointed to are preserved up to the
1224 * lesser of the new and old sizes (__GFP_ZERO flag is effectively ignored).
1225 * If @p is %NULL, krealloc() behaves exactly like kmalloc(). If @new_size
1226 * is 0 and @p is not a %NULL pointer, the object pointed to is freed.
1227 *
1228 * Return: pointer to the allocated memory or %NULL in case of error
1229 */
1230void *krealloc(const void *p, size_t new_size, gfp_t flags)
1231{
1232 void *ret;
1233
1234 if (unlikely(!new_size)) {
1235 kfree(p);
1236 return ZERO_SIZE_PTR;
1237 }
1238
1239 ret = __do_krealloc(p, new_size, flags);
1240 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
1241 kfree(p);
1242
1243 return ret;
1244}
1245EXPORT_SYMBOL(krealloc);
1246
1247/**
1248 * kfree_sensitive - Clear sensitive information in memory before freeing
1249 * @p: object to free memory of
1250 *
1251 * The memory of the object @p points to is zeroed before freed.
1252 * If @p is %NULL, kfree_sensitive() does nothing.
1253 *
1254 * Note: this function zeroes the whole allocated buffer which can be a good
1255 * deal bigger than the requested buffer size passed to kmalloc(). So be
1256 * careful when using this function in performance sensitive code.
1257 */
1258void kfree_sensitive(const void *p)
1259{
1260 size_t ks;
1261 void *mem = (void *)p;
1262
1263 ks = ksize(mem);
1264 if (ks)
1265 memzero_explicit(mem, ks);
1266 kfree(mem);
1267}
1268EXPORT_SYMBOL(kfree_sensitive);
1269
1270/**
1271 * ksize - get the actual amount of memory allocated for a given object
1272 * @objp: Pointer to the object
1273 *
1274 * kmalloc may internally round up allocations and return more memory
1275 * than requested. ksize() can be used to determine the actual amount of
1276 * memory allocated. The caller may use this additional memory, even though
1277 * a smaller amount of memory was initially specified with the kmalloc call.
1278 * The caller must guarantee that objp points to a valid object previously
1279 * allocated with either kmalloc() or kmem_cache_alloc(). The object
1280 * must not be freed during the duration of the call.
1281 *
1282 * Return: size of the actual memory used by @objp in bytes
1283 */
1284size_t ksize(const void *objp)
1285{
1286 size_t size;
1287
1288 /*
1289 * We need to first check that the pointer to the object is valid, and
1290 * only then unpoison the memory. The report printed from ksize() is
1291 * more useful, then when it's printed later when the behaviour could
1292 * be undefined due to a potential use-after-free or double-free.
1293 *
1294 * We use kasan_check_byte(), which is supported for the hardware
1295 * tag-based KASAN mode, unlike kasan_check_read/write().
1296 *
1297 * If the pointed to memory is invalid, we return 0 to avoid users of
1298 * ksize() writing to and potentially corrupting the memory region.
1299 *
1300 * We want to perform the check before __ksize(), to avoid potentially
1301 * crashing in __ksize() due to accessing invalid metadata.
1302 */
1303 if (unlikely(ZERO_OR_NULL_PTR(objp)) || !kasan_check_byte(objp))
1304 return 0;
1305
1306 size = kfence_ksize(objp) ?: __ksize(objp);
1307 /*
1308 * We assume that ksize callers could use whole allocated area,
1309 * so we need to unpoison this area.
1310 */
1311 kasan_unpoison_range(objp, size);
1312 return size;
1313}
1314EXPORT_SYMBOL(ksize);
1315
1316/* Tracepoints definitions. */
1317EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1318EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1319EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1320EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1321EXPORT_TRACEPOINT_SYMBOL(kfree);
1322EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1323
1324int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1325{
1326 if (__should_failslab(s, gfpflags))
1327 return -ENOMEM;
1328 return 0;
1329}
1330ALLOW_ERROR_INJECTION(should_failslab, ERRNO);