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 <linux/debugfs.h>
21#include <asm/cacheflush.h>
22#include <asm/tlbflush.h>
23#include <asm/page.h>
24#include <linux/memcontrol.h>
25
26#define CREATE_TRACE_POINTS
27#include <trace/events/kmem.h>
28
29#include "slab.h"
30
31enum slab_state slab_state;
32LIST_HEAD(slab_caches);
33DEFINE_MUTEX(slab_mutex);
34struct kmem_cache *kmem_cache;
35
36#ifdef CONFIG_HARDENED_USERCOPY
37bool usercopy_fallback __ro_after_init =
38 IS_ENABLED(CONFIG_HARDENED_USERCOPY_FALLBACK);
39module_param(usercopy_fallback, bool, 0400);
40MODULE_PARM_DESC(usercopy_fallback,
41 "WARN instead of reject usercopy whitelist violations");
42#endif
43
44static LIST_HEAD(slab_caches_to_rcu_destroy);
45static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
46static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
47 slab_caches_to_rcu_destroy_workfn);
48
49/*
50 * Set of flags that will prevent slab merging
51 */
52#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
53 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
54 SLAB_FAILSLAB | SLAB_KASAN)
55
56#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
57 SLAB_CACHE_DMA32 | SLAB_ACCOUNT)
58
59/*
60 * Merge control. If this is set then no merging of slab caches will occur.
61 */
62static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
63
64static int __init setup_slab_nomerge(char *str)
65{
66 slab_nomerge = true;
67 return 1;
68}
69
70#ifdef CONFIG_SLUB
71__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
72#endif
73
74__setup("slab_nomerge", setup_slab_nomerge);
75
76/*
77 * Determine the size of a slab object
78 */
79unsigned int kmem_cache_size(struct kmem_cache *s)
80{
81 return s->object_size;
82}
83EXPORT_SYMBOL(kmem_cache_size);
84
85#ifdef CONFIG_DEBUG_VM
86static int kmem_cache_sanity_check(const char *name, unsigned int size)
87{
88 if (!name || in_interrupt() || size < sizeof(void *) ||
89 size > KMALLOC_MAX_SIZE) {
90 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
91 return -EINVAL;
92 }
93
94 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
95 return 0;
96}
97#else
98static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
99{
100 return 0;
101}
102#endif
103
104void __kmem_cache_free_bulk(struct kmem_cache *s, size_t nr, void **p)
105{
106 size_t i;
107
108 for (i = 0; i < nr; i++) {
109 if (s)
110 kmem_cache_free(s, p[i]);
111 else
112 kfree(p[i]);
113 }
114}
115
116int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t nr,
117 void **p)
118{
119 size_t i;
120
121 for (i = 0; i < nr; i++) {
122 void *x = p[i] = kmem_cache_alloc(s, flags);
123 if (!x) {
124 __kmem_cache_free_bulk(s, i, p);
125 return 0;
126 }
127 }
128 return i;
129}
130
131#ifdef CONFIG_MEMCG_KMEM
132
133LIST_HEAD(slab_root_caches);
134static DEFINE_SPINLOCK(memcg_kmem_wq_lock);
135
136static void kmemcg_cache_shutdown(struct percpu_ref *percpu_ref);
137
138void slab_init_memcg_params(struct kmem_cache *s)
139{
140 s->memcg_params.root_cache = NULL;
141 RCU_INIT_POINTER(s->memcg_params.memcg_caches, NULL);
142 INIT_LIST_HEAD(&s->memcg_params.children);
143 s->memcg_params.dying = false;
144}
145
146static int init_memcg_params(struct kmem_cache *s,
147 struct kmem_cache *root_cache)
148{
149 struct memcg_cache_array *arr;
150
151 if (root_cache) {
152 int ret = percpu_ref_init(&s->memcg_params.refcnt,
153 kmemcg_cache_shutdown,
154 0, GFP_KERNEL);
155 if (ret)
156 return ret;
157
158 s->memcg_params.root_cache = root_cache;
159 INIT_LIST_HEAD(&s->memcg_params.children_node);
160 INIT_LIST_HEAD(&s->memcg_params.kmem_caches_node);
161 return 0;
162 }
163
164 slab_init_memcg_params(s);
165
166 if (!memcg_nr_cache_ids)
167 return 0;
168
169 arr = kvzalloc(sizeof(struct memcg_cache_array) +
170 memcg_nr_cache_ids * sizeof(void *),
171 GFP_KERNEL);
172 if (!arr)
173 return -ENOMEM;
174
175 RCU_INIT_POINTER(s->memcg_params.memcg_caches, arr);
176 return 0;
177}
178
179static void destroy_memcg_params(struct kmem_cache *s)
180{
181 if (is_root_cache(s)) {
182 kvfree(rcu_access_pointer(s->memcg_params.memcg_caches));
183 } else {
184 mem_cgroup_put(s->memcg_params.memcg);
185 WRITE_ONCE(s->memcg_params.memcg, NULL);
186 percpu_ref_exit(&s->memcg_params.refcnt);
187 }
188}
189
190static void free_memcg_params(struct rcu_head *rcu)
191{
192 struct memcg_cache_array *old;
193
194 old = container_of(rcu, struct memcg_cache_array, rcu);
195 kvfree(old);
196}
197
198static int update_memcg_params(struct kmem_cache *s, int new_array_size)
199{
200 struct memcg_cache_array *old, *new;
201
202 new = kvzalloc(sizeof(struct memcg_cache_array) +
203 new_array_size * sizeof(void *), GFP_KERNEL);
204 if (!new)
205 return -ENOMEM;
206
207 old = rcu_dereference_protected(s->memcg_params.memcg_caches,
208 lockdep_is_held(&slab_mutex));
209 if (old)
210 memcpy(new->entries, old->entries,
211 memcg_nr_cache_ids * sizeof(void *));
212
213 rcu_assign_pointer(s->memcg_params.memcg_caches, new);
214 if (old)
215 call_rcu(&old->rcu, free_memcg_params);
216 return 0;
217}
218
219int memcg_update_all_caches(int num_memcgs)
220{
221 struct kmem_cache *s;
222 int ret = 0;
223
224 mutex_lock(&slab_mutex);
225 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
226 ret = update_memcg_params(s, num_memcgs);
227 /*
228 * Instead of freeing the memory, we'll just leave the caches
229 * up to this point in an updated state.
230 */
231 if (ret)
232 break;
233 }
234 mutex_unlock(&slab_mutex);
235 return ret;
236}
237
238void memcg_link_cache(struct kmem_cache *s, struct mem_cgroup *memcg)
239{
240 if (is_root_cache(s)) {
241 list_add(&s->root_caches_node, &slab_root_caches);
242 } else {
243 css_get(&memcg->css);
244 s->memcg_params.memcg = memcg;
245 list_add(&s->memcg_params.children_node,
246 &s->memcg_params.root_cache->memcg_params.children);
247 list_add(&s->memcg_params.kmem_caches_node,
248 &s->memcg_params.memcg->kmem_caches);
249 }
250}
251
252static void memcg_unlink_cache(struct kmem_cache *s)
253{
254 if (is_root_cache(s)) {
255 list_del(&s->root_caches_node);
256 } else {
257 list_del(&s->memcg_params.children_node);
258 list_del(&s->memcg_params.kmem_caches_node);
259 }
260}
261#else
262static inline int init_memcg_params(struct kmem_cache *s,
263 struct kmem_cache *root_cache)
264{
265 return 0;
266}
267
268static inline void destroy_memcg_params(struct kmem_cache *s)
269{
270}
271
272static inline void memcg_unlink_cache(struct kmem_cache *s)
273{
274}
275#endif /* CONFIG_MEMCG_KMEM */
276
277/*
278 * Figure out what the alignment of the objects will be given a set of
279 * flags, a user specified alignment and the size of the objects.
280 */
281static unsigned int calculate_alignment(slab_flags_t flags,
282 unsigned int align, unsigned int size)
283{
284 /*
285 * If the user wants hardware cache aligned objects then follow that
286 * suggestion if the object is sufficiently large.
287 *
288 * The hardware cache alignment cannot override the specified
289 * alignment though. If that is greater then use it.
290 */
291 if (flags & SLAB_HWCACHE_ALIGN) {
292 unsigned int ralign;
293
294 ralign = cache_line_size();
295 while (size <= ralign / 2)
296 ralign /= 2;
297 align = max(align, ralign);
298 }
299
300 if (align < ARCH_SLAB_MINALIGN)
301 align = ARCH_SLAB_MINALIGN;
302
303 return ALIGN(align, sizeof(void *));
304}
305
306/*
307 * Find a mergeable slab cache
308 */
309int slab_unmergeable(struct kmem_cache *s)
310{
311 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
312 return 1;
313
314 if (!is_root_cache(s))
315 return 1;
316
317 if (s->ctor)
318 return 1;
319
320 if (s->usersize)
321 return 1;
322
323 /*
324 * We may have set a slab to be unmergeable during bootstrap.
325 */
326 if (s->refcount < 0)
327 return 1;
328
329 return 0;
330}
331
332struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
333 slab_flags_t flags, const char *name, void (*ctor)(void *))
334{
335 struct kmem_cache *s;
336
337 if (slab_nomerge)
338 return NULL;
339
340 if (ctor)
341 return NULL;
342
343 size = ALIGN(size, sizeof(void *));
344 align = calculate_alignment(flags, align, size);
345 size = ALIGN(size, align);
346 flags = kmem_cache_flags(size, flags, name, NULL);
347
348 if (flags & SLAB_NEVER_MERGE)
349 return NULL;
350
351 list_for_each_entry_reverse(s, &slab_root_caches, root_caches_node) {
352 if (slab_unmergeable(s))
353 continue;
354
355 if (size > s->size)
356 continue;
357
358 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
359 continue;
360 /*
361 * Check if alignment is compatible.
362 * Courtesy of Adrian Drzewiecki
363 */
364 if ((s->size & ~(align - 1)) != s->size)
365 continue;
366
367 if (s->size - size >= sizeof(void *))
368 continue;
369
370 if (IS_ENABLED(CONFIG_SLAB) && align &&
371 (align > s->align || s->align % align))
372 continue;
373
374 return s;
375 }
376 return NULL;
377}
378
379static struct kmem_cache *create_cache(const char *name,
380 unsigned int object_size, unsigned int align,
381 slab_flags_t flags, unsigned int useroffset,
382 unsigned int usersize, void (*ctor)(void *),
383 struct mem_cgroup *memcg, struct kmem_cache *root_cache)
384{
385 struct kmem_cache *s;
386 int err;
387
388 if (WARN_ON(useroffset + usersize > object_size))
389 useroffset = usersize = 0;
390
391 err = -ENOMEM;
392 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
393 if (!s)
394 goto out;
395
396 s->name = name;
397 s->size = s->object_size = object_size;
398 s->align = align;
399 s->ctor = ctor;
400 s->useroffset = useroffset;
401 s->usersize = usersize;
402
403 err = init_memcg_params(s, root_cache);
404 if (err)
405 goto out_free_cache;
406
407 err = __kmem_cache_create(s, flags);
408 if (err)
409 goto out_free_cache;
410
411 s->refcount = 1;
412 list_add(&s->list, &slab_caches);
413 memcg_link_cache(s, memcg);
414out:
415 if (err)
416 return ERR_PTR(err);
417 return s;
418
419out_free_cache:
420 destroy_memcg_params(s);
421 kmem_cache_free(kmem_cache, s);
422 goto out;
423}
424
425/**
426 * kmem_cache_create_usercopy - Create a cache with a region suitable
427 * for copying to userspace
428 * @name: A string which is used in /proc/slabinfo to identify this cache.
429 * @size: The size of objects to be created in this cache.
430 * @align: The required alignment for the objects.
431 * @flags: SLAB flags
432 * @useroffset: Usercopy region offset
433 * @usersize: Usercopy region size
434 * @ctor: A constructor for the objects.
435 *
436 * Cannot be called within a interrupt, but can be interrupted.
437 * The @ctor is run when new pages are allocated by the cache.
438 *
439 * The flags are
440 *
441 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
442 * to catch references to uninitialised memory.
443 *
444 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
445 * for buffer overruns.
446 *
447 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
448 * cacheline. This can be beneficial if you're counting cycles as closely
449 * as davem.
450 *
451 * Return: a pointer to the cache on success, NULL on failure.
452 */
453struct kmem_cache *
454kmem_cache_create_usercopy(const char *name,
455 unsigned int size, unsigned int align,
456 slab_flags_t flags,
457 unsigned int useroffset, unsigned int usersize,
458 void (*ctor)(void *))
459{
460 struct kmem_cache *s = NULL;
461 const char *cache_name;
462 int err;
463
464 get_online_cpus();
465 get_online_mems();
466 memcg_get_cache_ids();
467
468 mutex_lock(&slab_mutex);
469
470 err = kmem_cache_sanity_check(name, size);
471 if (err) {
472 goto out_unlock;
473 }
474
475 /* Refuse requests with allocator specific flags */
476 if (flags & ~SLAB_FLAGS_PERMITTED) {
477 err = -EINVAL;
478 goto out_unlock;
479 }
480
481 /*
482 * Some allocators will constraint the set of valid flags to a subset
483 * of all flags. We expect them to define CACHE_CREATE_MASK in this
484 * case, and we'll just provide them with a sanitized version of the
485 * passed flags.
486 */
487 flags &= CACHE_CREATE_MASK;
488
489 /* Fail closed on bad usersize of useroffset values. */
490 if (WARN_ON(!usersize && useroffset) ||
491 WARN_ON(size < usersize || size - usersize < useroffset))
492 usersize = useroffset = 0;
493
494 if (!usersize)
495 s = __kmem_cache_alias(name, size, align, flags, ctor);
496 if (s)
497 goto out_unlock;
498
499 cache_name = kstrdup_const(name, GFP_KERNEL);
500 if (!cache_name) {
501 err = -ENOMEM;
502 goto out_unlock;
503 }
504
505 s = create_cache(cache_name, size,
506 calculate_alignment(flags, align, size),
507 flags, useroffset, usersize, ctor, NULL, NULL);
508 if (IS_ERR(s)) {
509 err = PTR_ERR(s);
510 kfree_const(cache_name);
511 }
512
513out_unlock:
514 mutex_unlock(&slab_mutex);
515
516 memcg_put_cache_ids();
517 put_online_mems();
518 put_online_cpus();
519
520 if (err) {
521 if (flags & SLAB_PANIC)
522 panic("kmem_cache_create: Failed to create slab '%s'. Error %d\n",
523 name, err);
524 else {
525 pr_warn("kmem_cache_create(%s) failed with error %d\n",
526 name, err);
527 dump_stack();
528 }
529 return NULL;
530 }
531 return s;
532}
533EXPORT_SYMBOL(kmem_cache_create_usercopy);
534
535/**
536 * kmem_cache_create - Create a cache.
537 * @name: A string which is used in /proc/slabinfo to identify this cache.
538 * @size: The size of objects to be created in this cache.
539 * @align: The required alignment for the objects.
540 * @flags: SLAB flags
541 * @ctor: A constructor for the objects.
542 *
543 * Cannot be called within a interrupt, but can be interrupted.
544 * The @ctor is run when new pages are allocated by the cache.
545 *
546 * The flags are
547 *
548 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
549 * to catch references to uninitialised memory.
550 *
551 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
552 * for buffer overruns.
553 *
554 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
555 * cacheline. This can be beneficial if you're counting cycles as closely
556 * as davem.
557 *
558 * Return: a pointer to the cache on success, NULL on failure.
559 */
560struct kmem_cache *
561kmem_cache_create(const char *name, unsigned int size, unsigned int align,
562 slab_flags_t flags, void (*ctor)(void *))
563{
564 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
565 ctor);
566}
567EXPORT_SYMBOL(kmem_cache_create);
568
569static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
570{
571 LIST_HEAD(to_destroy);
572 struct kmem_cache *s, *s2;
573
574 /*
575 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
576 * @slab_caches_to_rcu_destroy list. The slab pages are freed
577 * through RCU and and the associated kmem_cache are dereferenced
578 * while freeing the pages, so the kmem_caches should be freed only
579 * after the pending RCU operations are finished. As rcu_barrier()
580 * is a pretty slow operation, we batch all pending destructions
581 * asynchronously.
582 */
583 mutex_lock(&slab_mutex);
584 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
585 mutex_unlock(&slab_mutex);
586
587 if (list_empty(&to_destroy))
588 return;
589
590 rcu_barrier();
591
592 list_for_each_entry_safe(s, s2, &to_destroy, list) {
593#ifdef SLAB_SUPPORTS_SYSFS
594 sysfs_slab_release(s);
595#else
596 slab_kmem_cache_release(s);
597#endif
598 }
599}
600
601static int shutdown_cache(struct kmem_cache *s)
602{
603 /* free asan quarantined objects */
604 kasan_cache_shutdown(s);
605
606 if (__kmem_cache_shutdown(s) != 0)
607 return -EBUSY;
608
609 memcg_unlink_cache(s);
610 list_del(&s->list);
611
612 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
613#ifdef SLAB_SUPPORTS_SYSFS
614 sysfs_slab_unlink(s);
615#endif
616 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
617 schedule_work(&slab_caches_to_rcu_destroy_work);
618 } else {
619#ifdef SLAB_SUPPORTS_SYSFS
620 sysfs_slab_unlink(s);
621 sysfs_slab_release(s);
622#else
623 slab_kmem_cache_release(s);
624#endif
625 }
626
627 return 0;
628}
629
630#ifdef CONFIG_MEMCG_KMEM
631/*
632 * memcg_create_kmem_cache - Create a cache for a memory cgroup.
633 * @memcg: The memory cgroup the new cache is for.
634 * @root_cache: The parent of the new cache.
635 *
636 * This function attempts to create a kmem cache that will serve allocation
637 * requests going from @memcg to @root_cache. The new cache inherits properties
638 * from its parent.
639 */
640void memcg_create_kmem_cache(struct mem_cgroup *memcg,
641 struct kmem_cache *root_cache)
642{
643 static char memcg_name_buf[NAME_MAX + 1]; /* protected by slab_mutex */
644 struct cgroup_subsys_state *css = &memcg->css;
645 struct memcg_cache_array *arr;
646 struct kmem_cache *s = NULL;
647 char *cache_name;
648 int idx;
649
650 get_online_cpus();
651 get_online_mems();
652
653 mutex_lock(&slab_mutex);
654
655 /*
656 * The memory cgroup could have been offlined while the cache
657 * creation work was pending.
658 */
659 if (memcg->kmem_state != KMEM_ONLINE)
660 goto out_unlock;
661
662 idx = memcg_cache_id(memcg);
663 arr = rcu_dereference_protected(root_cache->memcg_params.memcg_caches,
664 lockdep_is_held(&slab_mutex));
665
666 /*
667 * Since per-memcg caches are created asynchronously on first
668 * allocation (see memcg_kmem_get_cache()), several threads can try to
669 * create the same cache, but only one of them may succeed.
670 */
671 if (arr->entries[idx])
672 goto out_unlock;
673
674 cgroup_name(css->cgroup, memcg_name_buf, sizeof(memcg_name_buf));
675 cache_name = kasprintf(GFP_KERNEL, "%s(%llu:%s)", root_cache->name,
676 css->serial_nr, memcg_name_buf);
677 if (!cache_name)
678 goto out_unlock;
679
680 s = create_cache(cache_name, root_cache->object_size,
681 root_cache->align,
682 root_cache->flags & CACHE_CREATE_MASK,
683 root_cache->useroffset, root_cache->usersize,
684 root_cache->ctor, memcg, root_cache);
685 /*
686 * If we could not create a memcg cache, do not complain, because
687 * that's not critical at all as we can always proceed with the root
688 * cache.
689 */
690 if (IS_ERR(s)) {
691 kfree(cache_name);
692 goto out_unlock;
693 }
694
695 /*
696 * Since readers won't lock (see memcg_kmem_get_cache()), we need a
697 * barrier here to ensure nobody will see the kmem_cache partially
698 * initialized.
699 */
700 smp_wmb();
701 arr->entries[idx] = s;
702
703out_unlock:
704 mutex_unlock(&slab_mutex);
705
706 put_online_mems();
707 put_online_cpus();
708}
709
710static void kmemcg_workfn(struct work_struct *work)
711{
712 struct kmem_cache *s = container_of(work, struct kmem_cache,
713 memcg_params.work);
714
715 get_online_cpus();
716 get_online_mems();
717
718 mutex_lock(&slab_mutex);
719 s->memcg_params.work_fn(s);
720 mutex_unlock(&slab_mutex);
721
722 put_online_mems();
723 put_online_cpus();
724}
725
726static void kmemcg_rcufn(struct rcu_head *head)
727{
728 struct kmem_cache *s = container_of(head, struct kmem_cache,
729 memcg_params.rcu_head);
730
731 /*
732 * We need to grab blocking locks. Bounce to ->work. The
733 * work item shares the space with the RCU head and can't be
734 * initialized eariler.
735 */
736 INIT_WORK(&s->memcg_params.work, kmemcg_workfn);
737 queue_work(memcg_kmem_cache_wq, &s->memcg_params.work);
738}
739
740static void kmemcg_cache_shutdown_fn(struct kmem_cache *s)
741{
742 WARN_ON(shutdown_cache(s));
743}
744
745static void kmemcg_cache_shutdown(struct percpu_ref *percpu_ref)
746{
747 struct kmem_cache *s = container_of(percpu_ref, struct kmem_cache,
748 memcg_params.refcnt);
749 unsigned long flags;
750
751 spin_lock_irqsave(&memcg_kmem_wq_lock, flags);
752 if (s->memcg_params.root_cache->memcg_params.dying)
753 goto unlock;
754
755 s->memcg_params.work_fn = kmemcg_cache_shutdown_fn;
756 INIT_WORK(&s->memcg_params.work, kmemcg_workfn);
757 queue_work(memcg_kmem_cache_wq, &s->memcg_params.work);
758
759unlock:
760 spin_unlock_irqrestore(&memcg_kmem_wq_lock, flags);
761}
762
763static void kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
764{
765 __kmemcg_cache_deactivate_after_rcu(s);
766 percpu_ref_kill(&s->memcg_params.refcnt);
767}
768
769static void kmemcg_cache_deactivate(struct kmem_cache *s)
770{
771 if (WARN_ON_ONCE(is_root_cache(s)))
772 return;
773
774 __kmemcg_cache_deactivate(s);
775 s->flags |= SLAB_DEACTIVATED;
776
777 /*
778 * memcg_kmem_wq_lock is used to synchronize memcg_params.dying
779 * flag and make sure that no new kmem_cache deactivation tasks
780 * are queued (see flush_memcg_workqueue() ).
781 */
782 spin_lock_irq(&memcg_kmem_wq_lock);
783 if (s->memcg_params.root_cache->memcg_params.dying)
784 goto unlock;
785
786 s->memcg_params.work_fn = kmemcg_cache_deactivate_after_rcu;
787 call_rcu(&s->memcg_params.rcu_head, kmemcg_rcufn);
788unlock:
789 spin_unlock_irq(&memcg_kmem_wq_lock);
790}
791
792void memcg_deactivate_kmem_caches(struct mem_cgroup *memcg,
793 struct mem_cgroup *parent)
794{
795 int idx;
796 struct memcg_cache_array *arr;
797 struct kmem_cache *s, *c;
798 unsigned int nr_reparented;
799
800 idx = memcg_cache_id(memcg);
801
802 get_online_cpus();
803 get_online_mems();
804
805 mutex_lock(&slab_mutex);
806 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
807 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
808 lockdep_is_held(&slab_mutex));
809 c = arr->entries[idx];
810 if (!c)
811 continue;
812
813 kmemcg_cache_deactivate(c);
814 arr->entries[idx] = NULL;
815 }
816 nr_reparented = 0;
817 list_for_each_entry(s, &memcg->kmem_caches,
818 memcg_params.kmem_caches_node) {
819 WRITE_ONCE(s->memcg_params.memcg, parent);
820 css_put(&memcg->css);
821 nr_reparented++;
822 }
823 if (nr_reparented) {
824 list_splice_init(&memcg->kmem_caches,
825 &parent->kmem_caches);
826 css_get_many(&parent->css, nr_reparented);
827 }
828 mutex_unlock(&slab_mutex);
829
830 put_online_mems();
831 put_online_cpus();
832}
833
834static int shutdown_memcg_caches(struct kmem_cache *s)
835{
836 struct memcg_cache_array *arr;
837 struct kmem_cache *c, *c2;
838 LIST_HEAD(busy);
839 int i;
840
841 BUG_ON(!is_root_cache(s));
842
843 /*
844 * First, shutdown active caches, i.e. caches that belong to online
845 * memory cgroups.
846 */
847 arr = rcu_dereference_protected(s->memcg_params.memcg_caches,
848 lockdep_is_held(&slab_mutex));
849 for_each_memcg_cache_index(i) {
850 c = arr->entries[i];
851 if (!c)
852 continue;
853 if (shutdown_cache(c))
854 /*
855 * The cache still has objects. Move it to a temporary
856 * list so as not to try to destroy it for a second
857 * time while iterating over inactive caches below.
858 */
859 list_move(&c->memcg_params.children_node, &busy);
860 else
861 /*
862 * The cache is empty and will be destroyed soon. Clear
863 * the pointer to it in the memcg_caches array so that
864 * it will never be accessed even if the root cache
865 * stays alive.
866 */
867 arr->entries[i] = NULL;
868 }
869
870 /*
871 * Second, shutdown all caches left from memory cgroups that are now
872 * offline.
873 */
874 list_for_each_entry_safe(c, c2, &s->memcg_params.children,
875 memcg_params.children_node)
876 shutdown_cache(c);
877
878 list_splice(&busy, &s->memcg_params.children);
879
880 /*
881 * A cache being destroyed must be empty. In particular, this means
882 * that all per memcg caches attached to it must be empty too.
883 */
884 if (!list_empty(&s->memcg_params.children))
885 return -EBUSY;
886 return 0;
887}
888
889static void flush_memcg_workqueue(struct kmem_cache *s)
890{
891 spin_lock_irq(&memcg_kmem_wq_lock);
892 s->memcg_params.dying = true;
893 spin_unlock_irq(&memcg_kmem_wq_lock);
894
895 /*
896 * SLAB and SLUB deactivate the kmem_caches through call_rcu. Make
897 * sure all registered rcu callbacks have been invoked.
898 */
899 rcu_barrier();
900
901 /*
902 * SLAB and SLUB create memcg kmem_caches through workqueue and SLUB
903 * deactivates the memcg kmem_caches through workqueue. Make sure all
904 * previous workitems on workqueue are processed.
905 */
906 flush_workqueue(memcg_kmem_cache_wq);
907}
908#else
909static inline int shutdown_memcg_caches(struct kmem_cache *s)
910{
911 return 0;
912}
913
914static inline void flush_memcg_workqueue(struct kmem_cache *s)
915{
916}
917#endif /* CONFIG_MEMCG_KMEM */
918
919void slab_kmem_cache_release(struct kmem_cache *s)
920{
921 __kmem_cache_release(s);
922 destroy_memcg_params(s);
923 kfree_const(s->name);
924 kmem_cache_free(kmem_cache, s);
925}
926
927void kmem_cache_destroy(struct kmem_cache *s)
928{
929 int err;
930
931 if (unlikely(!s))
932 return;
933
934 flush_memcg_workqueue(s);
935
936 get_online_cpus();
937 get_online_mems();
938
939 mutex_lock(&slab_mutex);
940
941 s->refcount--;
942 if (s->refcount)
943 goto out_unlock;
944
945 err = shutdown_memcg_caches(s);
946 if (!err)
947 err = shutdown_cache(s);
948
949 if (err) {
950 pr_err("kmem_cache_destroy %s: Slab cache still has objects\n",
951 s->name);
952 dump_stack();
953 }
954out_unlock:
955 mutex_unlock(&slab_mutex);
956
957 put_online_mems();
958 put_online_cpus();
959}
960EXPORT_SYMBOL(kmem_cache_destroy);
961
962/**
963 * kmem_cache_shrink - Shrink a cache.
964 * @cachep: The cache to shrink.
965 *
966 * Releases as many slabs as possible for a cache.
967 * To help debugging, a zero exit status indicates all slabs were released.
968 *
969 * Return: %0 if all slabs were released, non-zero otherwise
970 */
971int kmem_cache_shrink(struct kmem_cache *cachep)
972{
973 int ret;
974
975 get_online_cpus();
976 get_online_mems();
977 kasan_cache_shrink(cachep);
978 ret = __kmem_cache_shrink(cachep);
979 put_online_mems();
980 put_online_cpus();
981 return ret;
982}
983EXPORT_SYMBOL(kmem_cache_shrink);
984
985/**
986 * kmem_cache_shrink_all - shrink a cache and all memcg caches for root cache
987 * @s: The cache pointer
988 */
989void kmem_cache_shrink_all(struct kmem_cache *s)
990{
991 struct kmem_cache *c;
992
993 if (!IS_ENABLED(CONFIG_MEMCG_KMEM) || !is_root_cache(s)) {
994 kmem_cache_shrink(s);
995 return;
996 }
997
998 get_online_cpus();
999 get_online_mems();
1000 kasan_cache_shrink(s);
1001 __kmem_cache_shrink(s);
1002
1003 /*
1004 * We have to take the slab_mutex to protect from the memcg list
1005 * modification.
1006 */
1007 mutex_lock(&slab_mutex);
1008 for_each_memcg_cache(c, s) {
1009 /*
1010 * Don't need to shrink deactivated memcg caches.
1011 */
1012 if (s->flags & SLAB_DEACTIVATED)
1013 continue;
1014 kasan_cache_shrink(c);
1015 __kmem_cache_shrink(c);
1016 }
1017 mutex_unlock(&slab_mutex);
1018 put_online_mems();
1019 put_online_cpus();
1020}
1021
1022bool slab_is_available(void)
1023{
1024 return slab_state >= UP;
1025}
1026
1027#ifndef CONFIG_SLOB
1028/* Create a cache during boot when no slab services are available yet */
1029void __init create_boot_cache(struct kmem_cache *s, const char *name,
1030 unsigned int size, slab_flags_t flags,
1031 unsigned int useroffset, unsigned int usersize)
1032{
1033 int err;
1034 unsigned int align = ARCH_KMALLOC_MINALIGN;
1035
1036 s->name = name;
1037 s->size = s->object_size = size;
1038
1039 /*
1040 * For power of two sizes, guarantee natural alignment for kmalloc
1041 * caches, regardless of SL*B debugging options.
1042 */
1043 if (is_power_of_2(size))
1044 align = max(align, size);
1045 s->align = calculate_alignment(flags, align, size);
1046
1047 s->useroffset = useroffset;
1048 s->usersize = usersize;
1049
1050 slab_init_memcg_params(s);
1051
1052 err = __kmem_cache_create(s, flags);
1053
1054 if (err)
1055 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
1056 name, size, err);
1057
1058 s->refcount = -1; /* Exempt from merging for now */
1059}
1060
1061struct kmem_cache *__init create_kmalloc_cache(const char *name,
1062 unsigned int size, slab_flags_t flags,
1063 unsigned int useroffset, unsigned int usersize)
1064{
1065 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
1066
1067 if (!s)
1068 panic("Out of memory when creating slab %s\n", name);
1069
1070 create_boot_cache(s, name, size, flags, useroffset, usersize);
1071 list_add(&s->list, &slab_caches);
1072 memcg_link_cache(s, NULL);
1073 s->refcount = 1;
1074 return s;
1075}
1076
1077struct kmem_cache *
1078kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] __ro_after_init =
1079{ /* initialization for https://bugs.llvm.org/show_bug.cgi?id=42570 */ };
1080EXPORT_SYMBOL(kmalloc_caches);
1081
1082/*
1083 * Conversion table for small slabs sizes / 8 to the index in the
1084 * kmalloc array. This is necessary for slabs < 192 since we have non power
1085 * of two cache sizes there. The size of larger slabs can be determined using
1086 * fls.
1087 */
1088static u8 size_index[24] __ro_after_init = {
1089 3, /* 8 */
1090 4, /* 16 */
1091 5, /* 24 */
1092 5, /* 32 */
1093 6, /* 40 */
1094 6, /* 48 */
1095 6, /* 56 */
1096 6, /* 64 */
1097 1, /* 72 */
1098 1, /* 80 */
1099 1, /* 88 */
1100 1, /* 96 */
1101 7, /* 104 */
1102 7, /* 112 */
1103 7, /* 120 */
1104 7, /* 128 */
1105 2, /* 136 */
1106 2, /* 144 */
1107 2, /* 152 */
1108 2, /* 160 */
1109 2, /* 168 */
1110 2, /* 176 */
1111 2, /* 184 */
1112 2 /* 192 */
1113};
1114
1115static inline unsigned int size_index_elem(unsigned int bytes)
1116{
1117 return (bytes - 1) / 8;
1118}
1119
1120/*
1121 * Find the kmem_cache structure that serves a given size of
1122 * allocation
1123 */
1124struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
1125{
1126 unsigned int index;
1127
1128 if (size <= 192) {
1129 if (!size)
1130 return ZERO_SIZE_PTR;
1131
1132 index = size_index[size_index_elem(size)];
1133 } else {
1134 if (WARN_ON_ONCE(size > KMALLOC_MAX_CACHE_SIZE))
1135 return NULL;
1136 index = fls(size - 1);
1137 }
1138
1139 return kmalloc_caches[kmalloc_type(flags)][index];
1140}
1141
1142/*
1143 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
1144 * kmalloc_index() supports up to 2^26=64MB, so the final entry of the table is
1145 * kmalloc-67108864.
1146 */
1147const struct kmalloc_info_struct kmalloc_info[] __initconst = {
1148 {NULL, 0}, {"kmalloc-96", 96},
1149 {"kmalloc-192", 192}, {"kmalloc-8", 8},
1150 {"kmalloc-16", 16}, {"kmalloc-32", 32},
1151 {"kmalloc-64", 64}, {"kmalloc-128", 128},
1152 {"kmalloc-256", 256}, {"kmalloc-512", 512},
1153 {"kmalloc-1k", 1024}, {"kmalloc-2k", 2048},
1154 {"kmalloc-4k", 4096}, {"kmalloc-8k", 8192},
1155 {"kmalloc-16k", 16384}, {"kmalloc-32k", 32768},
1156 {"kmalloc-64k", 65536}, {"kmalloc-128k", 131072},
1157 {"kmalloc-256k", 262144}, {"kmalloc-512k", 524288},
1158 {"kmalloc-1M", 1048576}, {"kmalloc-2M", 2097152},
1159 {"kmalloc-4M", 4194304}, {"kmalloc-8M", 8388608},
1160 {"kmalloc-16M", 16777216}, {"kmalloc-32M", 33554432},
1161 {"kmalloc-64M", 67108864}
1162};
1163
1164/*
1165 * Patch up the size_index table if we have strange large alignment
1166 * requirements for the kmalloc array. This is only the case for
1167 * MIPS it seems. The standard arches will not generate any code here.
1168 *
1169 * Largest permitted alignment is 256 bytes due to the way we
1170 * handle the index determination for the smaller caches.
1171 *
1172 * Make sure that nothing crazy happens if someone starts tinkering
1173 * around with ARCH_KMALLOC_MINALIGN
1174 */
1175void __init setup_kmalloc_cache_index_table(void)
1176{
1177 unsigned int i;
1178
1179 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
1180 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
1181
1182 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
1183 unsigned int elem = size_index_elem(i);
1184
1185 if (elem >= ARRAY_SIZE(size_index))
1186 break;
1187 size_index[elem] = KMALLOC_SHIFT_LOW;
1188 }
1189
1190 if (KMALLOC_MIN_SIZE >= 64) {
1191 /*
1192 * The 96 byte size cache is not used if the alignment
1193 * is 64 byte.
1194 */
1195 for (i = 64 + 8; i <= 96; i += 8)
1196 size_index[size_index_elem(i)] = 7;
1197
1198 }
1199
1200 if (KMALLOC_MIN_SIZE >= 128) {
1201 /*
1202 * The 192 byte sized cache is not used if the alignment
1203 * is 128 byte. Redirect kmalloc to use the 256 byte cache
1204 * instead.
1205 */
1206 for (i = 128 + 8; i <= 192; i += 8)
1207 size_index[size_index_elem(i)] = 8;
1208 }
1209}
1210
1211static const char *
1212kmalloc_cache_name(const char *prefix, unsigned int size)
1213{
1214
1215 static const char units[3] = "\0kM";
1216 int idx = 0;
1217
1218 while (size >= 1024 && (size % 1024 == 0)) {
1219 size /= 1024;
1220 idx++;
1221 }
1222
1223 return kasprintf(GFP_NOWAIT, "%s-%u%c", prefix, size, units[idx]);
1224}
1225
1226static void __init
1227new_kmalloc_cache(int idx, int type, slab_flags_t flags)
1228{
1229 const char *name;
1230
1231 if (type == KMALLOC_RECLAIM) {
1232 flags |= SLAB_RECLAIM_ACCOUNT;
1233 name = kmalloc_cache_name("kmalloc-rcl",
1234 kmalloc_info[idx].size);
1235 BUG_ON(!name);
1236 } else {
1237 name = kmalloc_info[idx].name;
1238 }
1239
1240 kmalloc_caches[type][idx] = create_kmalloc_cache(name,
1241 kmalloc_info[idx].size, flags, 0,
1242 kmalloc_info[idx].size);
1243}
1244
1245/*
1246 * Create the kmalloc array. Some of the regular kmalloc arrays
1247 * may already have been created because they were needed to
1248 * enable allocations for slab creation.
1249 */
1250void __init create_kmalloc_caches(slab_flags_t flags)
1251{
1252 int i, type;
1253
1254 for (type = KMALLOC_NORMAL; type <= KMALLOC_RECLAIM; type++) {
1255 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
1256 if (!kmalloc_caches[type][i])
1257 new_kmalloc_cache(i, type, flags);
1258
1259 /*
1260 * Caches that are not of the two-to-the-power-of size.
1261 * These have to be created immediately after the
1262 * earlier power of two caches
1263 */
1264 if (KMALLOC_MIN_SIZE <= 32 && i == 6 &&
1265 !kmalloc_caches[type][1])
1266 new_kmalloc_cache(1, type, flags);
1267 if (KMALLOC_MIN_SIZE <= 64 && i == 7 &&
1268 !kmalloc_caches[type][2])
1269 new_kmalloc_cache(2, type, flags);
1270 }
1271 }
1272
1273 /* Kmalloc array is now usable */
1274 slab_state = UP;
1275
1276#ifdef CONFIG_ZONE_DMA
1277 for (i = 0; i <= KMALLOC_SHIFT_HIGH; i++) {
1278 struct kmem_cache *s = kmalloc_caches[KMALLOC_NORMAL][i];
1279
1280 if (s) {
1281 unsigned int size = kmalloc_size(i);
1282 const char *n = kmalloc_cache_name("dma-kmalloc", size);
1283
1284 BUG_ON(!n);
1285 kmalloc_caches[KMALLOC_DMA][i] = create_kmalloc_cache(
1286 n, size, SLAB_CACHE_DMA | flags, 0, 0);
1287 }
1288 }
1289#endif
1290}
1291#endif /* !CONFIG_SLOB */
1292
1293/*
1294 * To avoid unnecessary overhead, we pass through large allocation requests
1295 * directly to the page allocator. We use __GFP_COMP, because we will need to
1296 * know the allocation order to free the pages properly in kfree.
1297 */
1298void *kmalloc_order(size_t size, gfp_t flags, unsigned int order)
1299{
1300 void *ret = NULL;
1301 struct page *page;
1302
1303 flags |= __GFP_COMP;
1304 page = alloc_pages(flags, order);
1305 if (likely(page)) {
1306 ret = page_address(page);
1307 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
1308 1 << order);
1309 }
1310 ret = kasan_kmalloc_large(ret, size, flags);
1311 /* As ret might get tagged, call kmemleak hook after KASAN. */
1312 kmemleak_alloc(ret, size, 1, flags);
1313 return ret;
1314}
1315EXPORT_SYMBOL(kmalloc_order);
1316
1317#ifdef CONFIG_TRACING
1318void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
1319{
1320 void *ret = kmalloc_order(size, flags, order);
1321 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
1322 return ret;
1323}
1324EXPORT_SYMBOL(kmalloc_order_trace);
1325#endif
1326
1327#ifdef CONFIG_SLAB_FREELIST_RANDOM
1328/* Randomize a generic freelist */
1329static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1330 unsigned int count)
1331{
1332 unsigned int rand;
1333 unsigned int i;
1334
1335 for (i = 0; i < count; i++)
1336 list[i] = i;
1337
1338 /* Fisher-Yates shuffle */
1339 for (i = count - 1; i > 0; i--) {
1340 rand = prandom_u32_state(state);
1341 rand %= (i + 1);
1342 swap(list[i], list[rand]);
1343 }
1344}
1345
1346/* Create a random sequence per cache */
1347int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1348 gfp_t gfp)
1349{
1350 struct rnd_state state;
1351
1352 if (count < 2 || cachep->random_seq)
1353 return 0;
1354
1355 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1356 if (!cachep->random_seq)
1357 return -ENOMEM;
1358
1359 /* Get best entropy at this stage of boot */
1360 prandom_seed_state(&state, get_random_long());
1361
1362 freelist_randomize(&state, cachep->random_seq, count);
1363 return 0;
1364}
1365
1366/* Destroy the per-cache random freelist sequence */
1367void cache_random_seq_destroy(struct kmem_cache *cachep)
1368{
1369 kfree(cachep->random_seq);
1370 cachep->random_seq = NULL;
1371}
1372#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1373
1374#if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1375#ifdef CONFIG_SLAB
1376#define SLABINFO_RIGHTS (0600)
1377#else
1378#define SLABINFO_RIGHTS (0400)
1379#endif
1380
1381static void print_slabinfo_header(struct seq_file *m)
1382{
1383 /*
1384 * Output format version, so at least we can change it
1385 * without _too_ many complaints.
1386 */
1387#ifdef CONFIG_DEBUG_SLAB
1388 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1389#else
1390 seq_puts(m, "slabinfo - version: 2.1\n");
1391#endif
1392 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1393 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1394 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1395#ifdef CONFIG_DEBUG_SLAB
1396 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1397 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1398#endif
1399 seq_putc(m, '\n');
1400}
1401
1402void *slab_start(struct seq_file *m, loff_t *pos)
1403{
1404 mutex_lock(&slab_mutex);
1405 return seq_list_start(&slab_root_caches, *pos);
1406}
1407
1408void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1409{
1410 return seq_list_next(p, &slab_root_caches, pos);
1411}
1412
1413void slab_stop(struct seq_file *m, void *p)
1414{
1415 mutex_unlock(&slab_mutex);
1416}
1417
1418static void
1419memcg_accumulate_slabinfo(struct kmem_cache *s, struct slabinfo *info)
1420{
1421 struct kmem_cache *c;
1422 struct slabinfo sinfo;
1423
1424 if (!is_root_cache(s))
1425 return;
1426
1427 for_each_memcg_cache(c, s) {
1428 memset(&sinfo, 0, sizeof(sinfo));
1429 get_slabinfo(c, &sinfo);
1430
1431 info->active_slabs += sinfo.active_slabs;
1432 info->num_slabs += sinfo.num_slabs;
1433 info->shared_avail += sinfo.shared_avail;
1434 info->active_objs += sinfo.active_objs;
1435 info->num_objs += sinfo.num_objs;
1436 }
1437}
1438
1439static void cache_show(struct kmem_cache *s, struct seq_file *m)
1440{
1441 struct slabinfo sinfo;
1442
1443 memset(&sinfo, 0, sizeof(sinfo));
1444 get_slabinfo(s, &sinfo);
1445
1446 memcg_accumulate_slabinfo(s, &sinfo);
1447
1448 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1449 cache_name(s), sinfo.active_objs, sinfo.num_objs, s->size,
1450 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1451
1452 seq_printf(m, " : tunables %4u %4u %4u",
1453 sinfo.limit, sinfo.batchcount, sinfo.shared);
1454 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1455 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1456 slabinfo_show_stats(m, s);
1457 seq_putc(m, '\n');
1458}
1459
1460static int slab_show(struct seq_file *m, void *p)
1461{
1462 struct kmem_cache *s = list_entry(p, struct kmem_cache, root_caches_node);
1463
1464 if (p == slab_root_caches.next)
1465 print_slabinfo_header(m);
1466 cache_show(s, m);
1467 return 0;
1468}
1469
1470void dump_unreclaimable_slab(void)
1471{
1472 struct kmem_cache *s, *s2;
1473 struct slabinfo sinfo;
1474
1475 /*
1476 * Here acquiring slab_mutex is risky since we don't prefer to get
1477 * sleep in oom path. But, without mutex hold, it may introduce a
1478 * risk of crash.
1479 * Use mutex_trylock to protect the list traverse, dump nothing
1480 * without acquiring the mutex.
1481 */
1482 if (!mutex_trylock(&slab_mutex)) {
1483 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1484 return;
1485 }
1486
1487 pr_info("Unreclaimable slab info:\n");
1488 pr_info("Name Used Total\n");
1489
1490 list_for_each_entry_safe(s, s2, &slab_caches, list) {
1491 if (!is_root_cache(s) || (s->flags & SLAB_RECLAIM_ACCOUNT))
1492 continue;
1493
1494 get_slabinfo(s, &sinfo);
1495
1496 if (sinfo.num_objs > 0)
1497 pr_info("%-17s %10luKB %10luKB\n", cache_name(s),
1498 (sinfo.active_objs * s->size) / 1024,
1499 (sinfo.num_objs * s->size) / 1024);
1500 }
1501 mutex_unlock(&slab_mutex);
1502}
1503
1504#if defined(CONFIG_MEMCG)
1505void *memcg_slab_start(struct seq_file *m, loff_t *pos)
1506{
1507 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1508
1509 mutex_lock(&slab_mutex);
1510 return seq_list_start(&memcg->kmem_caches, *pos);
1511}
1512
1513void *memcg_slab_next(struct seq_file *m, void *p, loff_t *pos)
1514{
1515 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1516
1517 return seq_list_next(p, &memcg->kmem_caches, pos);
1518}
1519
1520void memcg_slab_stop(struct seq_file *m, void *p)
1521{
1522 mutex_unlock(&slab_mutex);
1523}
1524
1525int memcg_slab_show(struct seq_file *m, void *p)
1526{
1527 struct kmem_cache *s = list_entry(p, struct kmem_cache,
1528 memcg_params.kmem_caches_node);
1529 struct mem_cgroup *memcg = mem_cgroup_from_seq(m);
1530
1531 if (p == memcg->kmem_caches.next)
1532 print_slabinfo_header(m);
1533 cache_show(s, m);
1534 return 0;
1535}
1536#endif
1537
1538/*
1539 * slabinfo_op - iterator that generates /proc/slabinfo
1540 *
1541 * Output layout:
1542 * cache-name
1543 * num-active-objs
1544 * total-objs
1545 * object size
1546 * num-active-slabs
1547 * total-slabs
1548 * num-pages-per-slab
1549 * + further values on SMP and with statistics enabled
1550 */
1551static const struct seq_operations slabinfo_op = {
1552 .start = slab_start,
1553 .next = slab_next,
1554 .stop = slab_stop,
1555 .show = slab_show,
1556};
1557
1558static int slabinfo_open(struct inode *inode, struct file *file)
1559{
1560 return seq_open(file, &slabinfo_op);
1561}
1562
1563static const struct file_operations proc_slabinfo_operations = {
1564 .open = slabinfo_open,
1565 .read = seq_read,
1566 .write = slabinfo_write,
1567 .llseek = seq_lseek,
1568 .release = seq_release,
1569};
1570
1571static int __init slab_proc_init(void)
1572{
1573 proc_create("slabinfo", SLABINFO_RIGHTS, NULL,
1574 &proc_slabinfo_operations);
1575 return 0;
1576}
1577module_init(slab_proc_init);
1578
1579#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_MEMCG_KMEM)
1580/*
1581 * Display information about kmem caches that have child memcg caches.
1582 */
1583static int memcg_slabinfo_show(struct seq_file *m, void *unused)
1584{
1585 struct kmem_cache *s, *c;
1586 struct slabinfo sinfo;
1587
1588 mutex_lock(&slab_mutex);
1589 seq_puts(m, "# <name> <css_id[:dead|deact]> <active_objs> <num_objs>");
1590 seq_puts(m, " <active_slabs> <num_slabs>\n");
1591 list_for_each_entry(s, &slab_root_caches, root_caches_node) {
1592 /*
1593 * Skip kmem caches that don't have any memcg children.
1594 */
1595 if (list_empty(&s->memcg_params.children))
1596 continue;
1597
1598 memset(&sinfo, 0, sizeof(sinfo));
1599 get_slabinfo(s, &sinfo);
1600 seq_printf(m, "%-17s root %6lu %6lu %6lu %6lu\n",
1601 cache_name(s), sinfo.active_objs, sinfo.num_objs,
1602 sinfo.active_slabs, sinfo.num_slabs);
1603
1604 for_each_memcg_cache(c, s) {
1605 struct cgroup_subsys_state *css;
1606 char *status = "";
1607
1608 css = &c->memcg_params.memcg->css;
1609 if (!(css->flags & CSS_ONLINE))
1610 status = ":dead";
1611 else if (c->flags & SLAB_DEACTIVATED)
1612 status = ":deact";
1613
1614 memset(&sinfo, 0, sizeof(sinfo));
1615 get_slabinfo(c, &sinfo);
1616 seq_printf(m, "%-17s %4d%-6s %6lu %6lu %6lu %6lu\n",
1617 cache_name(c), css->id, status,
1618 sinfo.active_objs, sinfo.num_objs,
1619 sinfo.active_slabs, sinfo.num_slabs);
1620 }
1621 }
1622 mutex_unlock(&slab_mutex);
1623 return 0;
1624}
1625DEFINE_SHOW_ATTRIBUTE(memcg_slabinfo);
1626
1627static int __init memcg_slabinfo_init(void)
1628{
1629 debugfs_create_file("memcg_slabinfo", S_IFREG | S_IRUGO,
1630 NULL, NULL, &memcg_slabinfo_fops);
1631 return 0;
1632}
1633
1634late_initcall(memcg_slabinfo_init);
1635#endif /* CONFIG_DEBUG_FS && CONFIG_MEMCG_KMEM */
1636#endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1637
1638static __always_inline void *__do_krealloc(const void *p, size_t new_size,
1639 gfp_t flags)
1640{
1641 void *ret;
1642 size_t ks = 0;
1643
1644 if (p)
1645 ks = ksize(p);
1646
1647 if (ks >= new_size) {
1648 p = kasan_krealloc((void *)p, new_size, flags);
1649 return (void *)p;
1650 }
1651
1652 ret = kmalloc_track_caller(new_size, flags);
1653 if (ret && p)
1654 memcpy(ret, p, ks);
1655
1656 return ret;
1657}
1658
1659/**
1660 * __krealloc - like krealloc() but don't free @p.
1661 * @p: object to reallocate memory for.
1662 * @new_size: how many bytes of memory are required.
1663 * @flags: the type of memory to allocate.
1664 *
1665 * This function is like krealloc() except it never frees the originally
1666 * allocated buffer. Use this if you don't want to free the buffer immediately
1667 * like, for example, with RCU.
1668 *
1669 * Return: pointer to the allocated memory or %NULL in case of error
1670 */
1671void *__krealloc(const void *p, size_t new_size, gfp_t flags)
1672{
1673 if (unlikely(!new_size))
1674 return ZERO_SIZE_PTR;
1675
1676 return __do_krealloc(p, new_size, flags);
1677
1678}
1679EXPORT_SYMBOL(__krealloc);
1680
1681/**
1682 * krealloc - reallocate memory. The contents will remain unchanged.
1683 * @p: object to reallocate memory for.
1684 * @new_size: how many bytes of memory are required.
1685 * @flags: the type of memory to allocate.
1686 *
1687 * The contents of the object pointed to are preserved up to the
1688 * lesser of the new and old sizes. If @p is %NULL, krealloc()
1689 * behaves exactly like kmalloc(). If @new_size is 0 and @p is not a
1690 * %NULL pointer, the object pointed to is freed.
1691 *
1692 * Return: pointer to the allocated memory or %NULL in case of error
1693 */
1694void *krealloc(const void *p, size_t new_size, gfp_t flags)
1695{
1696 void *ret;
1697
1698 if (unlikely(!new_size)) {
1699 kfree(p);
1700 return ZERO_SIZE_PTR;
1701 }
1702
1703 ret = __do_krealloc(p, new_size, flags);
1704 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
1705 kfree(p);
1706
1707 return ret;
1708}
1709EXPORT_SYMBOL(krealloc);
1710
1711/**
1712 * kzfree - like kfree but zero memory
1713 * @p: object to free memory of
1714 *
1715 * The memory of the object @p points to is zeroed before freed.
1716 * If @p is %NULL, kzfree() does nothing.
1717 *
1718 * Note: this function zeroes the whole allocated buffer which can be a good
1719 * deal bigger than the requested buffer size passed to kmalloc(). So be
1720 * careful when using this function in performance sensitive code.
1721 */
1722void kzfree(const void *p)
1723{
1724 size_t ks;
1725 void *mem = (void *)p;
1726
1727 if (unlikely(ZERO_OR_NULL_PTR(mem)))
1728 return;
1729 ks = ksize(mem);
1730 memset(mem, 0, ks);
1731 kfree(mem);
1732}
1733EXPORT_SYMBOL(kzfree);
1734
1735/**
1736 * ksize - get the actual amount of memory allocated for a given object
1737 * @objp: Pointer to the object
1738 *
1739 * kmalloc may internally round up allocations and return more memory
1740 * than requested. ksize() can be used to determine the actual amount of
1741 * memory allocated. The caller may use this additional memory, even though
1742 * a smaller amount of memory was initially specified with the kmalloc call.
1743 * The caller must guarantee that objp points to a valid object previously
1744 * allocated with either kmalloc() or kmem_cache_alloc(). The object
1745 * must not be freed during the duration of the call.
1746 *
1747 * Return: size of the actual memory used by @objp in bytes
1748 */
1749size_t ksize(const void *objp)
1750{
1751 size_t size;
1752
1753 if (WARN_ON_ONCE(!objp))
1754 return 0;
1755 /*
1756 * We need to check that the pointed to object is valid, and only then
1757 * unpoison the shadow memory below. We use __kasan_check_read(), to
1758 * generate a more useful report at the time ksize() is called (rather
1759 * than later where behaviour is undefined due to potential
1760 * use-after-free or double-free).
1761 *
1762 * If the pointed to memory is invalid we return 0, to avoid users of
1763 * ksize() writing to and potentially corrupting the memory region.
1764 *
1765 * We want to perform the check before __ksize(), to avoid potentially
1766 * crashing in __ksize() due to accessing invalid metadata.
1767 */
1768 if (unlikely(objp == ZERO_SIZE_PTR) || !__kasan_check_read(objp, 1))
1769 return 0;
1770
1771 size = __ksize(objp);
1772 /*
1773 * We assume that ksize callers could use whole allocated area,
1774 * so we need to unpoison this area.
1775 */
1776 kasan_unpoison_shadow(objp, size);
1777 return size;
1778}
1779EXPORT_SYMBOL(ksize);
1780
1781/* Tracepoints definitions. */
1782EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1783EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1784EXPORT_TRACEPOINT_SYMBOL(kmalloc_node);
1785EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc_node);
1786EXPORT_TRACEPOINT_SYMBOL(kfree);
1787EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1788
1789int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1790{
1791 if (__should_failslab(s, gfpflags))
1792 return -ENOMEM;
1793 return 0;
1794}
1795ALLOW_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#include <linux/stackdepot.h>
28
29#include "internal.h"
30#include "slab.h"
31
32#define CREATE_TRACE_POINTS
33#include <trace/events/kmem.h>
34
35enum slab_state slab_state;
36LIST_HEAD(slab_caches);
37DEFINE_MUTEX(slab_mutex);
38struct kmem_cache *kmem_cache;
39
40static LIST_HEAD(slab_caches_to_rcu_destroy);
41static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work);
42static DECLARE_WORK(slab_caches_to_rcu_destroy_work,
43 slab_caches_to_rcu_destroy_workfn);
44
45/*
46 * Set of flags that will prevent slab merging
47 */
48#define SLAB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
49 SLAB_TRACE | SLAB_TYPESAFE_BY_RCU | SLAB_NOLEAKTRACE | \
50 SLAB_FAILSLAB | kasan_never_merge())
51
52#define SLAB_MERGE_SAME (SLAB_RECLAIM_ACCOUNT | SLAB_CACHE_DMA | \
53 SLAB_CACHE_DMA32 | SLAB_ACCOUNT)
54
55/*
56 * Merge control. If this is set then no merging of slab caches will occur.
57 */
58static bool slab_nomerge = !IS_ENABLED(CONFIG_SLAB_MERGE_DEFAULT);
59
60static int __init setup_slab_nomerge(char *str)
61{
62 slab_nomerge = true;
63 return 1;
64}
65
66static int __init setup_slab_merge(char *str)
67{
68 slab_nomerge = false;
69 return 1;
70}
71
72#ifdef CONFIG_SLUB
73__setup_param("slub_nomerge", slub_nomerge, setup_slab_nomerge, 0);
74__setup_param("slub_merge", slub_merge, setup_slab_merge, 0);
75#endif
76
77__setup("slab_nomerge", setup_slab_nomerge);
78__setup("slab_merge", setup_slab_merge);
79
80/*
81 * Determine the size of a slab object
82 */
83unsigned int kmem_cache_size(struct kmem_cache *s)
84{
85 return s->object_size;
86}
87EXPORT_SYMBOL(kmem_cache_size);
88
89#ifdef CONFIG_DEBUG_VM
90static int kmem_cache_sanity_check(const char *name, unsigned int size)
91{
92 if (!name || in_interrupt() || size > KMALLOC_MAX_SIZE) {
93 pr_err("kmem_cache_create(%s) integrity check failed\n", name);
94 return -EINVAL;
95 }
96
97 WARN_ON(strchr(name, ' ')); /* It confuses parsers */
98 return 0;
99}
100#else
101static inline int kmem_cache_sanity_check(const char *name, unsigned int size)
102{
103 return 0;
104}
105#endif
106
107/*
108 * Figure out what the alignment of the objects will be given a set of
109 * flags, a user specified alignment and the size of the objects.
110 */
111static unsigned int calculate_alignment(slab_flags_t flags,
112 unsigned int align, unsigned int size)
113{
114 /*
115 * If the user wants hardware cache aligned objects then follow that
116 * suggestion if the object is sufficiently large.
117 *
118 * The hardware cache alignment cannot override the specified
119 * alignment though. If that is greater then use it.
120 */
121 if (flags & SLAB_HWCACHE_ALIGN) {
122 unsigned int ralign;
123
124 ralign = cache_line_size();
125 while (size <= ralign / 2)
126 ralign /= 2;
127 align = max(align, ralign);
128 }
129
130 align = max(align, arch_slab_minalign());
131
132 return ALIGN(align, sizeof(void *));
133}
134
135/*
136 * Find a mergeable slab cache
137 */
138int slab_unmergeable(struct kmem_cache *s)
139{
140 if (slab_nomerge || (s->flags & SLAB_NEVER_MERGE))
141 return 1;
142
143 if (s->ctor)
144 return 1;
145
146#ifdef CONFIG_HARDENED_USERCOPY
147 if (s->usersize)
148 return 1;
149#endif
150
151 /*
152 * We may have set a slab to be unmergeable during bootstrap.
153 */
154 if (s->refcount < 0)
155 return 1;
156
157 return 0;
158}
159
160struct kmem_cache *find_mergeable(unsigned int size, unsigned int align,
161 slab_flags_t flags, const char *name, void (*ctor)(void *))
162{
163 struct kmem_cache *s;
164
165 if (slab_nomerge)
166 return NULL;
167
168 if (ctor)
169 return NULL;
170
171 size = ALIGN(size, sizeof(void *));
172 align = calculate_alignment(flags, align, size);
173 size = ALIGN(size, align);
174 flags = kmem_cache_flags(size, flags, name);
175
176 if (flags & SLAB_NEVER_MERGE)
177 return NULL;
178
179 list_for_each_entry_reverse(s, &slab_caches, list) {
180 if (slab_unmergeable(s))
181 continue;
182
183 if (size > s->size)
184 continue;
185
186 if ((flags & SLAB_MERGE_SAME) != (s->flags & SLAB_MERGE_SAME))
187 continue;
188 /*
189 * Check if alignment is compatible.
190 * Courtesy of Adrian Drzewiecki
191 */
192 if ((s->size & ~(align - 1)) != s->size)
193 continue;
194
195 if (s->size - size >= sizeof(void *))
196 continue;
197
198 if (IS_ENABLED(CONFIG_SLAB) && align &&
199 (align > s->align || s->align % align))
200 continue;
201
202 return s;
203 }
204 return NULL;
205}
206
207static struct kmem_cache *create_cache(const char *name,
208 unsigned int object_size, unsigned int align,
209 slab_flags_t flags, unsigned int useroffset,
210 unsigned int usersize, void (*ctor)(void *),
211 struct kmem_cache *root_cache)
212{
213 struct kmem_cache *s;
214 int err;
215
216 if (WARN_ON(useroffset + usersize > object_size))
217 useroffset = usersize = 0;
218
219 err = -ENOMEM;
220 s = kmem_cache_zalloc(kmem_cache, GFP_KERNEL);
221 if (!s)
222 goto out;
223
224 s->name = name;
225 s->size = s->object_size = object_size;
226 s->align = align;
227 s->ctor = ctor;
228#ifdef CONFIG_HARDENED_USERCOPY
229 s->useroffset = useroffset;
230 s->usersize = usersize;
231#endif
232
233 err = __kmem_cache_create(s, flags);
234 if (err)
235 goto out_free_cache;
236
237 s->refcount = 1;
238 list_add(&s->list, &slab_caches);
239out:
240 if (err)
241 return ERR_PTR(err);
242 return s;
243
244out_free_cache:
245 kmem_cache_free(kmem_cache, s);
246 goto out;
247}
248
249/**
250 * kmem_cache_create_usercopy - Create a cache with a region suitable
251 * for copying to userspace
252 * @name: A string which is used in /proc/slabinfo to identify this cache.
253 * @size: The size of objects to be created in this cache.
254 * @align: The required alignment for the objects.
255 * @flags: SLAB flags
256 * @useroffset: Usercopy region offset
257 * @usersize: Usercopy region size
258 * @ctor: A constructor for the objects.
259 *
260 * Cannot be called within a interrupt, but can be interrupted.
261 * The @ctor is run when new pages are allocated by the cache.
262 *
263 * The flags are
264 *
265 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
266 * to catch references to uninitialised memory.
267 *
268 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
269 * for buffer overruns.
270 *
271 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
272 * cacheline. This can be beneficial if you're counting cycles as closely
273 * as davem.
274 *
275 * Return: a pointer to the cache on success, NULL on failure.
276 */
277struct kmem_cache *
278kmem_cache_create_usercopy(const char *name,
279 unsigned int size, unsigned int align,
280 slab_flags_t flags,
281 unsigned int useroffset, unsigned int usersize,
282 void (*ctor)(void *))
283{
284 struct kmem_cache *s = NULL;
285 const char *cache_name;
286 int err;
287
288#ifdef CONFIG_SLUB_DEBUG
289 /*
290 * If no slub_debug was enabled globally, the static key is not yet
291 * enabled by setup_slub_debug(). Enable it if the cache is being
292 * created with any of the debugging flags passed explicitly.
293 * It's also possible that this is the first cache created with
294 * SLAB_STORE_USER and we should init stack_depot for it.
295 */
296 if (flags & SLAB_DEBUG_FLAGS)
297 static_branch_enable(&slub_debug_enabled);
298 if (flags & SLAB_STORE_USER)
299 stack_depot_init();
300#endif
301
302 mutex_lock(&slab_mutex);
303
304 err = kmem_cache_sanity_check(name, size);
305 if (err) {
306 goto out_unlock;
307 }
308
309 /* Refuse requests with allocator specific flags */
310 if (flags & ~SLAB_FLAGS_PERMITTED) {
311 err = -EINVAL;
312 goto out_unlock;
313 }
314
315 /*
316 * Some allocators will constraint the set of valid flags to a subset
317 * of all flags. We expect them to define CACHE_CREATE_MASK in this
318 * case, and we'll just provide them with a sanitized version of the
319 * passed flags.
320 */
321 flags &= CACHE_CREATE_MASK;
322
323 /* Fail closed on bad usersize of useroffset values. */
324 if (!IS_ENABLED(CONFIG_HARDENED_USERCOPY) ||
325 WARN_ON(!usersize && useroffset) ||
326 WARN_ON(size < usersize || size - usersize < useroffset))
327 usersize = useroffset = 0;
328
329 if (!usersize)
330 s = __kmem_cache_alias(name, size, align, flags, ctor);
331 if (s)
332 goto out_unlock;
333
334 cache_name = kstrdup_const(name, GFP_KERNEL);
335 if (!cache_name) {
336 err = -ENOMEM;
337 goto out_unlock;
338 }
339
340 s = create_cache(cache_name, size,
341 calculate_alignment(flags, align, size),
342 flags, useroffset, usersize, ctor, NULL);
343 if (IS_ERR(s)) {
344 err = PTR_ERR(s);
345 kfree_const(cache_name);
346 }
347
348out_unlock:
349 mutex_unlock(&slab_mutex);
350
351 if (err) {
352 if (flags & SLAB_PANIC)
353 panic("%s: Failed to create slab '%s'. Error %d\n",
354 __func__, name, err);
355 else {
356 pr_warn("%s(%s) failed with error %d\n",
357 __func__, name, err);
358 dump_stack();
359 }
360 return NULL;
361 }
362 return s;
363}
364EXPORT_SYMBOL(kmem_cache_create_usercopy);
365
366/**
367 * kmem_cache_create - Create a cache.
368 * @name: A string which is used in /proc/slabinfo to identify this cache.
369 * @size: The size of objects to be created in this cache.
370 * @align: The required alignment for the objects.
371 * @flags: SLAB flags
372 * @ctor: A constructor for the objects.
373 *
374 * Cannot be called within a interrupt, but can be interrupted.
375 * The @ctor is run when new pages are allocated by the cache.
376 *
377 * The flags are
378 *
379 * %SLAB_POISON - Poison the slab with a known test pattern (a5a5a5a5)
380 * to catch references to uninitialised memory.
381 *
382 * %SLAB_RED_ZONE - Insert `Red` zones around the allocated memory to check
383 * for buffer overruns.
384 *
385 * %SLAB_HWCACHE_ALIGN - Align the objects in this cache to a hardware
386 * cacheline. This can be beneficial if you're counting cycles as closely
387 * as davem.
388 *
389 * Return: a pointer to the cache on success, NULL on failure.
390 */
391struct kmem_cache *
392kmem_cache_create(const char *name, unsigned int size, unsigned int align,
393 slab_flags_t flags, void (*ctor)(void *))
394{
395 return kmem_cache_create_usercopy(name, size, align, flags, 0, 0,
396 ctor);
397}
398EXPORT_SYMBOL(kmem_cache_create);
399
400#ifdef SLAB_SUPPORTS_SYSFS
401/*
402 * For a given kmem_cache, kmem_cache_destroy() should only be called
403 * once or there will be a use-after-free problem. The actual deletion
404 * and release of the kobject does not need slab_mutex or cpu_hotplug_lock
405 * protection. So they are now done without holding those locks.
406 *
407 * Note that there will be a slight delay in the deletion of sysfs files
408 * if kmem_cache_release() is called indrectly from a work function.
409 */
410static void kmem_cache_release(struct kmem_cache *s)
411{
412 sysfs_slab_unlink(s);
413 sysfs_slab_release(s);
414}
415#else
416static void kmem_cache_release(struct kmem_cache *s)
417{
418 slab_kmem_cache_release(s);
419}
420#endif
421
422static void slab_caches_to_rcu_destroy_workfn(struct work_struct *work)
423{
424 LIST_HEAD(to_destroy);
425 struct kmem_cache *s, *s2;
426
427 /*
428 * On destruction, SLAB_TYPESAFE_BY_RCU kmem_caches are put on the
429 * @slab_caches_to_rcu_destroy list. The slab pages are freed
430 * through RCU and the associated kmem_cache are dereferenced
431 * while freeing the pages, so the kmem_caches should be freed only
432 * after the pending RCU operations are finished. As rcu_barrier()
433 * is a pretty slow operation, we batch all pending destructions
434 * asynchronously.
435 */
436 mutex_lock(&slab_mutex);
437 list_splice_init(&slab_caches_to_rcu_destroy, &to_destroy);
438 mutex_unlock(&slab_mutex);
439
440 if (list_empty(&to_destroy))
441 return;
442
443 rcu_barrier();
444
445 list_for_each_entry_safe(s, s2, &to_destroy, list) {
446 debugfs_slab_release(s);
447 kfence_shutdown_cache(s);
448 kmem_cache_release(s);
449 }
450}
451
452static int shutdown_cache(struct kmem_cache *s)
453{
454 /* free asan quarantined objects */
455 kasan_cache_shutdown(s);
456
457 if (__kmem_cache_shutdown(s) != 0)
458 return -EBUSY;
459
460 list_del(&s->list);
461
462 if (s->flags & SLAB_TYPESAFE_BY_RCU) {
463 list_add_tail(&s->list, &slab_caches_to_rcu_destroy);
464 schedule_work(&slab_caches_to_rcu_destroy_work);
465 } else {
466 kfence_shutdown_cache(s);
467 debugfs_slab_release(s);
468 }
469
470 return 0;
471}
472
473void slab_kmem_cache_release(struct kmem_cache *s)
474{
475 __kmem_cache_release(s);
476 kfree_const(s->name);
477 kmem_cache_free(kmem_cache, s);
478}
479
480void kmem_cache_destroy(struct kmem_cache *s)
481{
482 int refcnt;
483 bool rcu_set;
484
485 if (unlikely(!s) || !kasan_check_byte(s))
486 return;
487
488 cpus_read_lock();
489 mutex_lock(&slab_mutex);
490
491 rcu_set = s->flags & SLAB_TYPESAFE_BY_RCU;
492
493 refcnt = --s->refcount;
494 if (refcnt)
495 goto out_unlock;
496
497 WARN(shutdown_cache(s),
498 "%s %s: Slab cache still has objects when called from %pS",
499 __func__, s->name, (void *)_RET_IP_);
500out_unlock:
501 mutex_unlock(&slab_mutex);
502 cpus_read_unlock();
503 if (!refcnt && !rcu_set)
504 kmem_cache_release(s);
505}
506EXPORT_SYMBOL(kmem_cache_destroy);
507
508/**
509 * kmem_cache_shrink - Shrink a cache.
510 * @cachep: The cache to shrink.
511 *
512 * Releases as many slabs as possible for a cache.
513 * To help debugging, a zero exit status indicates all slabs were released.
514 *
515 * Return: %0 if all slabs were released, non-zero otherwise
516 */
517int kmem_cache_shrink(struct kmem_cache *cachep)
518{
519 kasan_cache_shrink(cachep);
520
521 return __kmem_cache_shrink(cachep);
522}
523EXPORT_SYMBOL(kmem_cache_shrink);
524
525bool slab_is_available(void)
526{
527 return slab_state >= UP;
528}
529
530#ifdef CONFIG_PRINTK
531/**
532 * kmem_valid_obj - does the pointer reference a valid slab object?
533 * @object: pointer to query.
534 *
535 * Return: %true if the pointer is to a not-yet-freed object from
536 * kmalloc() or kmem_cache_alloc(), either %true or %false if the pointer
537 * is to an already-freed object, and %false otherwise.
538 */
539bool kmem_valid_obj(void *object)
540{
541 struct folio *folio;
542
543 /* Some arches consider ZERO_SIZE_PTR to be a valid address. */
544 if (object < (void *)PAGE_SIZE || !virt_addr_valid(object))
545 return false;
546 folio = virt_to_folio(object);
547 return folio_test_slab(folio);
548}
549EXPORT_SYMBOL_GPL(kmem_valid_obj);
550
551static void kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
552{
553 if (__kfence_obj_info(kpp, object, slab))
554 return;
555 __kmem_obj_info(kpp, object, slab);
556}
557
558/**
559 * kmem_dump_obj - Print available slab provenance information
560 * @object: slab object for which to find provenance information.
561 *
562 * This function uses pr_cont(), so that the caller is expected to have
563 * printed out whatever preamble is appropriate. The provenance information
564 * depends on the type of object and on how much debugging is enabled.
565 * For a slab-cache object, the fact that it is a slab object is printed,
566 * and, if available, the slab name, return address, and stack trace from
567 * the allocation and last free path of that object.
568 *
569 * This function will splat if passed a pointer to a non-slab object.
570 * If you are not sure what type of object you have, you should instead
571 * use mem_dump_obj().
572 */
573void kmem_dump_obj(void *object)
574{
575 char *cp = IS_ENABLED(CONFIG_MMU) ? "" : "/vmalloc";
576 int i;
577 struct slab *slab;
578 unsigned long ptroffset;
579 struct kmem_obj_info kp = { };
580
581 if (WARN_ON_ONCE(!virt_addr_valid(object)))
582 return;
583 slab = virt_to_slab(object);
584 if (WARN_ON_ONCE(!slab)) {
585 pr_cont(" non-slab memory.\n");
586 return;
587 }
588 kmem_obj_info(&kp, object, slab);
589 if (kp.kp_slab_cache)
590 pr_cont(" slab%s %s", cp, kp.kp_slab_cache->name);
591 else
592 pr_cont(" slab%s", cp);
593 if (is_kfence_address(object))
594 pr_cont(" (kfence)");
595 if (kp.kp_objp)
596 pr_cont(" start %px", kp.kp_objp);
597 if (kp.kp_data_offset)
598 pr_cont(" data offset %lu", kp.kp_data_offset);
599 if (kp.kp_objp) {
600 ptroffset = ((char *)object - (char *)kp.kp_objp) - kp.kp_data_offset;
601 pr_cont(" pointer offset %lu", ptroffset);
602 }
603 if (kp.kp_slab_cache && kp.kp_slab_cache->object_size)
604 pr_cont(" size %u", kp.kp_slab_cache->object_size);
605 if (kp.kp_ret)
606 pr_cont(" allocated at %pS\n", kp.kp_ret);
607 else
608 pr_cont("\n");
609 for (i = 0; i < ARRAY_SIZE(kp.kp_stack); i++) {
610 if (!kp.kp_stack[i])
611 break;
612 pr_info(" %pS\n", kp.kp_stack[i]);
613 }
614
615 if (kp.kp_free_stack[0])
616 pr_cont(" Free path:\n");
617
618 for (i = 0; i < ARRAY_SIZE(kp.kp_free_stack); i++) {
619 if (!kp.kp_free_stack[i])
620 break;
621 pr_info(" %pS\n", kp.kp_free_stack[i]);
622 }
623
624}
625EXPORT_SYMBOL_GPL(kmem_dump_obj);
626#endif
627
628#ifndef CONFIG_SLOB
629/* Create a cache during boot when no slab services are available yet */
630void __init create_boot_cache(struct kmem_cache *s, const char *name,
631 unsigned int size, slab_flags_t flags,
632 unsigned int useroffset, unsigned int usersize)
633{
634 int err;
635 unsigned int align = ARCH_KMALLOC_MINALIGN;
636
637 s->name = name;
638 s->size = s->object_size = size;
639
640 /*
641 * For power of two sizes, guarantee natural alignment for kmalloc
642 * caches, regardless of SL*B debugging options.
643 */
644 if (is_power_of_2(size))
645 align = max(align, size);
646 s->align = calculate_alignment(flags, align, size);
647
648#ifdef CONFIG_HARDENED_USERCOPY
649 s->useroffset = useroffset;
650 s->usersize = usersize;
651#endif
652
653 err = __kmem_cache_create(s, flags);
654
655 if (err)
656 panic("Creation of kmalloc slab %s size=%u failed. Reason %d\n",
657 name, size, err);
658
659 s->refcount = -1; /* Exempt from merging for now */
660}
661
662struct kmem_cache *__init create_kmalloc_cache(const char *name,
663 unsigned int size, slab_flags_t flags,
664 unsigned int useroffset, unsigned int usersize)
665{
666 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
667
668 if (!s)
669 panic("Out of memory when creating slab %s\n", name);
670
671 create_boot_cache(s, name, size, flags | SLAB_KMALLOC, useroffset,
672 usersize);
673 kasan_cache_create_kmalloc(s);
674 list_add(&s->list, &slab_caches);
675 s->refcount = 1;
676 return s;
677}
678
679struct kmem_cache *
680kmalloc_caches[NR_KMALLOC_TYPES][KMALLOC_SHIFT_HIGH + 1] __ro_after_init =
681{ /* initialization for https://bugs.llvm.org/show_bug.cgi?id=42570 */ };
682EXPORT_SYMBOL(kmalloc_caches);
683
684/*
685 * Conversion table for small slabs sizes / 8 to the index in the
686 * kmalloc array. This is necessary for slabs < 192 since we have non power
687 * of two cache sizes there. The size of larger slabs can be determined using
688 * fls.
689 */
690static u8 size_index[24] __ro_after_init = {
691 3, /* 8 */
692 4, /* 16 */
693 5, /* 24 */
694 5, /* 32 */
695 6, /* 40 */
696 6, /* 48 */
697 6, /* 56 */
698 6, /* 64 */
699 1, /* 72 */
700 1, /* 80 */
701 1, /* 88 */
702 1, /* 96 */
703 7, /* 104 */
704 7, /* 112 */
705 7, /* 120 */
706 7, /* 128 */
707 2, /* 136 */
708 2, /* 144 */
709 2, /* 152 */
710 2, /* 160 */
711 2, /* 168 */
712 2, /* 176 */
713 2, /* 184 */
714 2 /* 192 */
715};
716
717static inline unsigned int size_index_elem(unsigned int bytes)
718{
719 return (bytes - 1) / 8;
720}
721
722/*
723 * Find the kmem_cache structure that serves a given size of
724 * allocation
725 */
726struct kmem_cache *kmalloc_slab(size_t size, gfp_t flags)
727{
728 unsigned int index;
729
730 if (size <= 192) {
731 if (!size)
732 return ZERO_SIZE_PTR;
733
734 index = size_index[size_index_elem(size)];
735 } else {
736 if (WARN_ON_ONCE(size > KMALLOC_MAX_CACHE_SIZE))
737 return NULL;
738 index = fls(size - 1);
739 }
740
741 return kmalloc_caches[kmalloc_type(flags)][index];
742}
743
744size_t kmalloc_size_roundup(size_t size)
745{
746 struct kmem_cache *c;
747
748 /* Short-circuit the 0 size case. */
749 if (unlikely(size == 0))
750 return 0;
751 /* Short-circuit saturated "too-large" case. */
752 if (unlikely(size == SIZE_MAX))
753 return SIZE_MAX;
754 /* Above the smaller buckets, size is a multiple of page size. */
755 if (size > KMALLOC_MAX_CACHE_SIZE)
756 return PAGE_SIZE << get_order(size);
757
758 /* The flags don't matter since size_index is common to all. */
759 c = kmalloc_slab(size, GFP_KERNEL);
760 return c ? c->object_size : 0;
761}
762EXPORT_SYMBOL(kmalloc_size_roundup);
763
764#ifdef CONFIG_ZONE_DMA
765#define KMALLOC_DMA_NAME(sz) .name[KMALLOC_DMA] = "dma-kmalloc-" #sz,
766#else
767#define KMALLOC_DMA_NAME(sz)
768#endif
769
770#ifdef CONFIG_MEMCG_KMEM
771#define KMALLOC_CGROUP_NAME(sz) .name[KMALLOC_CGROUP] = "kmalloc-cg-" #sz,
772#else
773#define KMALLOC_CGROUP_NAME(sz)
774#endif
775
776#ifndef CONFIG_SLUB_TINY
777#define KMALLOC_RCL_NAME(sz) .name[KMALLOC_RECLAIM] = "kmalloc-rcl-" #sz,
778#else
779#define KMALLOC_RCL_NAME(sz)
780#endif
781
782#define INIT_KMALLOC_INFO(__size, __short_size) \
783{ \
784 .name[KMALLOC_NORMAL] = "kmalloc-" #__short_size, \
785 KMALLOC_RCL_NAME(__short_size) \
786 KMALLOC_CGROUP_NAME(__short_size) \
787 KMALLOC_DMA_NAME(__short_size) \
788 .size = __size, \
789}
790
791/*
792 * kmalloc_info[] is to make slub_debug=,kmalloc-xx option work at boot time.
793 * kmalloc_index() supports up to 2^21=2MB, so the final entry of the table is
794 * kmalloc-2M.
795 */
796const struct kmalloc_info_struct kmalloc_info[] __initconst = {
797 INIT_KMALLOC_INFO(0, 0),
798 INIT_KMALLOC_INFO(96, 96),
799 INIT_KMALLOC_INFO(192, 192),
800 INIT_KMALLOC_INFO(8, 8),
801 INIT_KMALLOC_INFO(16, 16),
802 INIT_KMALLOC_INFO(32, 32),
803 INIT_KMALLOC_INFO(64, 64),
804 INIT_KMALLOC_INFO(128, 128),
805 INIT_KMALLOC_INFO(256, 256),
806 INIT_KMALLOC_INFO(512, 512),
807 INIT_KMALLOC_INFO(1024, 1k),
808 INIT_KMALLOC_INFO(2048, 2k),
809 INIT_KMALLOC_INFO(4096, 4k),
810 INIT_KMALLOC_INFO(8192, 8k),
811 INIT_KMALLOC_INFO(16384, 16k),
812 INIT_KMALLOC_INFO(32768, 32k),
813 INIT_KMALLOC_INFO(65536, 64k),
814 INIT_KMALLOC_INFO(131072, 128k),
815 INIT_KMALLOC_INFO(262144, 256k),
816 INIT_KMALLOC_INFO(524288, 512k),
817 INIT_KMALLOC_INFO(1048576, 1M),
818 INIT_KMALLOC_INFO(2097152, 2M)
819};
820
821/*
822 * Patch up the size_index table if we have strange large alignment
823 * requirements for the kmalloc array. This is only the case for
824 * MIPS it seems. The standard arches will not generate any code here.
825 *
826 * Largest permitted alignment is 256 bytes due to the way we
827 * handle the index determination for the smaller caches.
828 *
829 * Make sure that nothing crazy happens if someone starts tinkering
830 * around with ARCH_KMALLOC_MINALIGN
831 */
832void __init setup_kmalloc_cache_index_table(void)
833{
834 unsigned int i;
835
836 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
837 !is_power_of_2(KMALLOC_MIN_SIZE));
838
839 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
840 unsigned int elem = size_index_elem(i);
841
842 if (elem >= ARRAY_SIZE(size_index))
843 break;
844 size_index[elem] = KMALLOC_SHIFT_LOW;
845 }
846
847 if (KMALLOC_MIN_SIZE >= 64) {
848 /*
849 * The 96 byte sized cache is not used if the alignment
850 * is 64 byte.
851 */
852 for (i = 64 + 8; i <= 96; i += 8)
853 size_index[size_index_elem(i)] = 7;
854
855 }
856
857 if (KMALLOC_MIN_SIZE >= 128) {
858 /*
859 * The 192 byte sized cache is not used if the alignment
860 * is 128 byte. Redirect kmalloc to use the 256 byte cache
861 * instead.
862 */
863 for (i = 128 + 8; i <= 192; i += 8)
864 size_index[size_index_elem(i)] = 8;
865 }
866}
867
868static void __init
869new_kmalloc_cache(int idx, enum kmalloc_cache_type type, slab_flags_t flags)
870{
871 if ((KMALLOC_RECLAIM != KMALLOC_NORMAL) && (type == KMALLOC_RECLAIM)) {
872 flags |= SLAB_RECLAIM_ACCOUNT;
873 } else if (IS_ENABLED(CONFIG_MEMCG_KMEM) && (type == KMALLOC_CGROUP)) {
874 if (mem_cgroup_kmem_disabled()) {
875 kmalloc_caches[type][idx] = kmalloc_caches[KMALLOC_NORMAL][idx];
876 return;
877 }
878 flags |= SLAB_ACCOUNT;
879 } else if (IS_ENABLED(CONFIG_ZONE_DMA) && (type == KMALLOC_DMA)) {
880 flags |= SLAB_CACHE_DMA;
881 }
882
883 kmalloc_caches[type][idx] = create_kmalloc_cache(
884 kmalloc_info[idx].name[type],
885 kmalloc_info[idx].size, flags, 0,
886 kmalloc_info[idx].size);
887
888 /*
889 * If CONFIG_MEMCG_KMEM is enabled, disable cache merging for
890 * KMALLOC_NORMAL caches.
891 */
892 if (IS_ENABLED(CONFIG_MEMCG_KMEM) && (type == KMALLOC_NORMAL))
893 kmalloc_caches[type][idx]->refcount = -1;
894}
895
896/*
897 * Create the kmalloc array. Some of the regular kmalloc arrays
898 * may already have been created because they were needed to
899 * enable allocations for slab creation.
900 */
901void __init create_kmalloc_caches(slab_flags_t flags)
902{
903 int i;
904 enum kmalloc_cache_type type;
905
906 /*
907 * Including KMALLOC_CGROUP if CONFIG_MEMCG_KMEM defined
908 */
909 for (type = KMALLOC_NORMAL; type < NR_KMALLOC_TYPES; type++) {
910 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
911 if (!kmalloc_caches[type][i])
912 new_kmalloc_cache(i, type, flags);
913
914 /*
915 * Caches that are not of the two-to-the-power-of size.
916 * These have to be created immediately after the
917 * earlier power of two caches
918 */
919 if (KMALLOC_MIN_SIZE <= 32 && i == 6 &&
920 !kmalloc_caches[type][1])
921 new_kmalloc_cache(1, type, flags);
922 if (KMALLOC_MIN_SIZE <= 64 && i == 7 &&
923 !kmalloc_caches[type][2])
924 new_kmalloc_cache(2, type, flags);
925 }
926 }
927
928 /* Kmalloc array is now usable */
929 slab_state = UP;
930}
931
932void free_large_kmalloc(struct folio *folio, void *object)
933{
934 unsigned int order = folio_order(folio);
935
936 if (WARN_ON_ONCE(order == 0))
937 pr_warn_once("object pointer: 0x%p\n", object);
938
939 kmemleak_free(object);
940 kasan_kfree_large(object);
941 kmsan_kfree_large(object);
942
943 mod_lruvec_page_state(folio_page(folio, 0), NR_SLAB_UNRECLAIMABLE_B,
944 -(PAGE_SIZE << order));
945 __free_pages(folio_page(folio, 0), order);
946}
947
948static void *__kmalloc_large_node(size_t size, gfp_t flags, int node);
949static __always_inline
950void *__do_kmalloc_node(size_t size, gfp_t flags, int node, unsigned long caller)
951{
952 struct kmem_cache *s;
953 void *ret;
954
955 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
956 ret = __kmalloc_large_node(size, flags, node);
957 trace_kmalloc(caller, ret, size,
958 PAGE_SIZE << get_order(size), flags, node);
959 return ret;
960 }
961
962 s = kmalloc_slab(size, flags);
963
964 if (unlikely(ZERO_OR_NULL_PTR(s)))
965 return s;
966
967 ret = __kmem_cache_alloc_node(s, flags, node, size, caller);
968 ret = kasan_kmalloc(s, ret, size, flags);
969 trace_kmalloc(caller, ret, size, s->size, flags, node);
970 return ret;
971}
972
973void *__kmalloc_node(size_t size, gfp_t flags, int node)
974{
975 return __do_kmalloc_node(size, flags, node, _RET_IP_);
976}
977EXPORT_SYMBOL(__kmalloc_node);
978
979void *__kmalloc(size_t size, gfp_t flags)
980{
981 return __do_kmalloc_node(size, flags, NUMA_NO_NODE, _RET_IP_);
982}
983EXPORT_SYMBOL(__kmalloc);
984
985void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
986 int node, unsigned long caller)
987{
988 return __do_kmalloc_node(size, flags, node, caller);
989}
990EXPORT_SYMBOL(__kmalloc_node_track_caller);
991
992/**
993 * kfree - free previously allocated memory
994 * @object: pointer returned by kmalloc.
995 *
996 * If @object is NULL, no operation is performed.
997 *
998 * Don't free memory not originally allocated by kmalloc()
999 * or you will run into trouble.
1000 */
1001void kfree(const void *object)
1002{
1003 struct folio *folio;
1004 struct slab *slab;
1005 struct kmem_cache *s;
1006
1007 trace_kfree(_RET_IP_, object);
1008
1009 if (unlikely(ZERO_OR_NULL_PTR(object)))
1010 return;
1011
1012 folio = virt_to_folio(object);
1013 if (unlikely(!folio_test_slab(folio))) {
1014 free_large_kmalloc(folio, (void *)object);
1015 return;
1016 }
1017
1018 slab = folio_slab(folio);
1019 s = slab->slab_cache;
1020 __kmem_cache_free(s, (void *)object, _RET_IP_);
1021}
1022EXPORT_SYMBOL(kfree);
1023
1024/**
1025 * __ksize -- Report full size of underlying allocation
1026 * @object: pointer to the object
1027 *
1028 * This should only be used internally to query the true size of allocations.
1029 * It is not meant to be a way to discover the usable size of an allocation
1030 * after the fact. Instead, use kmalloc_size_roundup(). Using memory beyond
1031 * the originally requested allocation size may trigger KASAN, UBSAN_BOUNDS,
1032 * and/or FORTIFY_SOURCE.
1033 *
1034 * Return: size of the actual memory used by @object in bytes
1035 */
1036size_t __ksize(const void *object)
1037{
1038 struct folio *folio;
1039
1040 if (unlikely(object == ZERO_SIZE_PTR))
1041 return 0;
1042
1043 folio = virt_to_folio(object);
1044
1045 if (unlikely(!folio_test_slab(folio))) {
1046 if (WARN_ON(folio_size(folio) <= KMALLOC_MAX_CACHE_SIZE))
1047 return 0;
1048 if (WARN_ON(object != folio_address(folio)))
1049 return 0;
1050 return folio_size(folio);
1051 }
1052
1053#ifdef CONFIG_SLUB_DEBUG
1054 skip_orig_size_check(folio_slab(folio)->slab_cache, object);
1055#endif
1056
1057 return slab_ksize(folio_slab(folio)->slab_cache);
1058}
1059
1060void *kmalloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
1061{
1062 void *ret = __kmem_cache_alloc_node(s, gfpflags, NUMA_NO_NODE,
1063 size, _RET_IP_);
1064
1065 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
1066
1067 ret = kasan_kmalloc(s, ret, size, gfpflags);
1068 return ret;
1069}
1070EXPORT_SYMBOL(kmalloc_trace);
1071
1072void *kmalloc_node_trace(struct kmem_cache *s, gfp_t gfpflags,
1073 int node, size_t size)
1074{
1075 void *ret = __kmem_cache_alloc_node(s, gfpflags, node, size, _RET_IP_);
1076
1077 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
1078
1079 ret = kasan_kmalloc(s, ret, size, gfpflags);
1080 return ret;
1081}
1082EXPORT_SYMBOL(kmalloc_node_trace);
1083#endif /* !CONFIG_SLOB */
1084
1085gfp_t kmalloc_fix_flags(gfp_t flags)
1086{
1087 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1088
1089 flags &= ~GFP_SLAB_BUG_MASK;
1090 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1091 invalid_mask, &invalid_mask, flags, &flags);
1092 dump_stack();
1093
1094 return flags;
1095}
1096
1097/*
1098 * To avoid unnecessary overhead, we pass through large allocation requests
1099 * directly to the page allocator. We use __GFP_COMP, because we will need to
1100 * know the allocation order to free the pages properly in kfree.
1101 */
1102
1103static void *__kmalloc_large_node(size_t size, gfp_t flags, int node)
1104{
1105 struct page *page;
1106 void *ptr = NULL;
1107 unsigned int order = get_order(size);
1108
1109 if (unlikely(flags & GFP_SLAB_BUG_MASK))
1110 flags = kmalloc_fix_flags(flags);
1111
1112 flags |= __GFP_COMP;
1113 page = alloc_pages_node(node, flags, order);
1114 if (page) {
1115 ptr = page_address(page);
1116 mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
1117 PAGE_SIZE << order);
1118 }
1119
1120 ptr = kasan_kmalloc_large(ptr, size, flags);
1121 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1122 kmemleak_alloc(ptr, size, 1, flags);
1123 kmsan_kmalloc_large(ptr, size, flags);
1124
1125 return ptr;
1126}
1127
1128void *kmalloc_large(size_t size, gfp_t flags)
1129{
1130 void *ret = __kmalloc_large_node(size, flags, NUMA_NO_NODE);
1131
1132 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
1133 flags, NUMA_NO_NODE);
1134 return ret;
1135}
1136EXPORT_SYMBOL(kmalloc_large);
1137
1138void *kmalloc_large_node(size_t size, gfp_t flags, int node)
1139{
1140 void *ret = __kmalloc_large_node(size, flags, node);
1141
1142 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
1143 flags, node);
1144 return ret;
1145}
1146EXPORT_SYMBOL(kmalloc_large_node);
1147
1148#ifdef CONFIG_SLAB_FREELIST_RANDOM
1149/* Randomize a generic freelist */
1150static void freelist_randomize(struct rnd_state *state, unsigned int *list,
1151 unsigned int count)
1152{
1153 unsigned int rand;
1154 unsigned int i;
1155
1156 for (i = 0; i < count; i++)
1157 list[i] = i;
1158
1159 /* Fisher-Yates shuffle */
1160 for (i = count - 1; i > 0; i--) {
1161 rand = prandom_u32_state(state);
1162 rand %= (i + 1);
1163 swap(list[i], list[rand]);
1164 }
1165}
1166
1167/* Create a random sequence per cache */
1168int cache_random_seq_create(struct kmem_cache *cachep, unsigned int count,
1169 gfp_t gfp)
1170{
1171 struct rnd_state state;
1172
1173 if (count < 2 || cachep->random_seq)
1174 return 0;
1175
1176 cachep->random_seq = kcalloc(count, sizeof(unsigned int), gfp);
1177 if (!cachep->random_seq)
1178 return -ENOMEM;
1179
1180 /* Get best entropy at this stage of boot */
1181 prandom_seed_state(&state, get_random_long());
1182
1183 freelist_randomize(&state, cachep->random_seq, count);
1184 return 0;
1185}
1186
1187/* Destroy the per-cache random freelist sequence */
1188void cache_random_seq_destroy(struct kmem_cache *cachep)
1189{
1190 kfree(cachep->random_seq);
1191 cachep->random_seq = NULL;
1192}
1193#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1194
1195#if defined(CONFIG_SLAB) || defined(CONFIG_SLUB_DEBUG)
1196#ifdef CONFIG_SLAB
1197#define SLABINFO_RIGHTS (0600)
1198#else
1199#define SLABINFO_RIGHTS (0400)
1200#endif
1201
1202static void print_slabinfo_header(struct seq_file *m)
1203{
1204 /*
1205 * Output format version, so at least we can change it
1206 * without _too_ many complaints.
1207 */
1208#ifdef CONFIG_DEBUG_SLAB
1209 seq_puts(m, "slabinfo - version: 2.1 (statistics)\n");
1210#else
1211 seq_puts(m, "slabinfo - version: 2.1\n");
1212#endif
1213 seq_puts(m, "# name <active_objs> <num_objs> <objsize> <objperslab> <pagesperslab>");
1214 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
1215 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
1216#ifdef CONFIG_DEBUG_SLAB
1217 seq_puts(m, " : globalstat <listallocs> <maxobjs> <grown> <reaped> <error> <maxfreeable> <nodeallocs> <remotefrees> <alienoverflow>");
1218 seq_puts(m, " : cpustat <allochit> <allocmiss> <freehit> <freemiss>");
1219#endif
1220 seq_putc(m, '\n');
1221}
1222
1223static void *slab_start(struct seq_file *m, loff_t *pos)
1224{
1225 mutex_lock(&slab_mutex);
1226 return seq_list_start(&slab_caches, *pos);
1227}
1228
1229static void *slab_next(struct seq_file *m, void *p, loff_t *pos)
1230{
1231 return seq_list_next(p, &slab_caches, pos);
1232}
1233
1234static void slab_stop(struct seq_file *m, void *p)
1235{
1236 mutex_unlock(&slab_mutex);
1237}
1238
1239static void cache_show(struct kmem_cache *s, struct seq_file *m)
1240{
1241 struct slabinfo sinfo;
1242
1243 memset(&sinfo, 0, sizeof(sinfo));
1244 get_slabinfo(s, &sinfo);
1245
1246 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d",
1247 s->name, sinfo.active_objs, sinfo.num_objs, s->size,
1248 sinfo.objects_per_slab, (1 << sinfo.cache_order));
1249
1250 seq_printf(m, " : tunables %4u %4u %4u",
1251 sinfo.limit, sinfo.batchcount, sinfo.shared);
1252 seq_printf(m, " : slabdata %6lu %6lu %6lu",
1253 sinfo.active_slabs, sinfo.num_slabs, sinfo.shared_avail);
1254 slabinfo_show_stats(m, s);
1255 seq_putc(m, '\n');
1256}
1257
1258static int slab_show(struct seq_file *m, void *p)
1259{
1260 struct kmem_cache *s = list_entry(p, struct kmem_cache, list);
1261
1262 if (p == slab_caches.next)
1263 print_slabinfo_header(m);
1264 cache_show(s, m);
1265 return 0;
1266}
1267
1268void dump_unreclaimable_slab(void)
1269{
1270 struct kmem_cache *s;
1271 struct slabinfo sinfo;
1272
1273 /*
1274 * Here acquiring slab_mutex is risky since we don't prefer to get
1275 * sleep in oom path. But, without mutex hold, it may introduce a
1276 * risk of crash.
1277 * Use mutex_trylock to protect the list traverse, dump nothing
1278 * without acquiring the mutex.
1279 */
1280 if (!mutex_trylock(&slab_mutex)) {
1281 pr_warn("excessive unreclaimable slab but cannot dump stats\n");
1282 return;
1283 }
1284
1285 pr_info("Unreclaimable slab info:\n");
1286 pr_info("Name Used Total\n");
1287
1288 list_for_each_entry(s, &slab_caches, list) {
1289 if (s->flags & SLAB_RECLAIM_ACCOUNT)
1290 continue;
1291
1292 get_slabinfo(s, &sinfo);
1293
1294 if (sinfo.num_objs > 0)
1295 pr_info("%-17s %10luKB %10luKB\n", s->name,
1296 (sinfo.active_objs * s->size) / 1024,
1297 (sinfo.num_objs * s->size) / 1024);
1298 }
1299 mutex_unlock(&slab_mutex);
1300}
1301
1302/*
1303 * slabinfo_op - iterator that generates /proc/slabinfo
1304 *
1305 * Output layout:
1306 * cache-name
1307 * num-active-objs
1308 * total-objs
1309 * object size
1310 * num-active-slabs
1311 * total-slabs
1312 * num-pages-per-slab
1313 * + further values on SMP and with statistics enabled
1314 */
1315static const struct seq_operations slabinfo_op = {
1316 .start = slab_start,
1317 .next = slab_next,
1318 .stop = slab_stop,
1319 .show = slab_show,
1320};
1321
1322static int slabinfo_open(struct inode *inode, struct file *file)
1323{
1324 return seq_open(file, &slabinfo_op);
1325}
1326
1327static const struct proc_ops slabinfo_proc_ops = {
1328 .proc_flags = PROC_ENTRY_PERMANENT,
1329 .proc_open = slabinfo_open,
1330 .proc_read = seq_read,
1331 .proc_write = slabinfo_write,
1332 .proc_lseek = seq_lseek,
1333 .proc_release = seq_release,
1334};
1335
1336static int __init slab_proc_init(void)
1337{
1338 proc_create("slabinfo", SLABINFO_RIGHTS, NULL, &slabinfo_proc_ops);
1339 return 0;
1340}
1341module_init(slab_proc_init);
1342
1343#endif /* CONFIG_SLAB || CONFIG_SLUB_DEBUG */
1344
1345static __always_inline __realloc_size(2) void *
1346__do_krealloc(const void *p, size_t new_size, gfp_t flags)
1347{
1348 void *ret;
1349 size_t ks;
1350
1351 /* Check for double-free before calling ksize. */
1352 if (likely(!ZERO_OR_NULL_PTR(p))) {
1353 if (!kasan_check_byte(p))
1354 return NULL;
1355 ks = ksize(p);
1356 } else
1357 ks = 0;
1358
1359 /* If the object still fits, repoison it precisely. */
1360 if (ks >= new_size) {
1361 p = kasan_krealloc((void *)p, new_size, flags);
1362 return (void *)p;
1363 }
1364
1365 ret = kmalloc_track_caller(new_size, flags);
1366 if (ret && p) {
1367 /* Disable KASAN checks as the object's redzone is accessed. */
1368 kasan_disable_current();
1369 memcpy(ret, kasan_reset_tag(p), ks);
1370 kasan_enable_current();
1371 }
1372
1373 return ret;
1374}
1375
1376/**
1377 * krealloc - reallocate memory. The contents will remain unchanged.
1378 * @p: object to reallocate memory for.
1379 * @new_size: how many bytes of memory are required.
1380 * @flags: the type of memory to allocate.
1381 *
1382 * The contents of the object pointed to are preserved up to the
1383 * lesser of the new and old sizes (__GFP_ZERO flag is effectively ignored).
1384 * If @p is %NULL, krealloc() behaves exactly like kmalloc(). If @new_size
1385 * is 0 and @p is not a %NULL pointer, the object pointed to is freed.
1386 *
1387 * Return: pointer to the allocated memory or %NULL in case of error
1388 */
1389void *krealloc(const void *p, size_t new_size, gfp_t flags)
1390{
1391 void *ret;
1392
1393 if (unlikely(!new_size)) {
1394 kfree(p);
1395 return ZERO_SIZE_PTR;
1396 }
1397
1398 ret = __do_krealloc(p, new_size, flags);
1399 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
1400 kfree(p);
1401
1402 return ret;
1403}
1404EXPORT_SYMBOL(krealloc);
1405
1406/**
1407 * kfree_sensitive - Clear sensitive information in memory before freeing
1408 * @p: object to free memory of
1409 *
1410 * The memory of the object @p points to is zeroed before freed.
1411 * If @p is %NULL, kfree_sensitive() does nothing.
1412 *
1413 * Note: this function zeroes the whole allocated buffer which can be a good
1414 * deal bigger than the requested buffer size passed to kmalloc(). So be
1415 * careful when using this function in performance sensitive code.
1416 */
1417void kfree_sensitive(const void *p)
1418{
1419 size_t ks;
1420 void *mem = (void *)p;
1421
1422 ks = ksize(mem);
1423 if (ks) {
1424 kasan_unpoison_range(mem, ks);
1425 memzero_explicit(mem, ks);
1426 }
1427 kfree(mem);
1428}
1429EXPORT_SYMBOL(kfree_sensitive);
1430
1431size_t ksize(const void *objp)
1432{
1433 /*
1434 * We need to first check that the pointer to the object is valid.
1435 * The KASAN report printed from ksize() is more useful, then when
1436 * it's printed later when the behaviour could be undefined due to
1437 * a potential use-after-free or double-free.
1438 *
1439 * We use kasan_check_byte(), which is supported for the hardware
1440 * tag-based KASAN mode, unlike kasan_check_read/write().
1441 *
1442 * If the pointed to memory is invalid, we return 0 to avoid users of
1443 * ksize() writing to and potentially corrupting the memory region.
1444 *
1445 * We want to perform the check before __ksize(), to avoid potentially
1446 * crashing in __ksize() due to accessing invalid metadata.
1447 */
1448 if (unlikely(ZERO_OR_NULL_PTR(objp)) || !kasan_check_byte(objp))
1449 return 0;
1450
1451 return kfence_ksize(objp) ?: __ksize(objp);
1452}
1453EXPORT_SYMBOL(ksize);
1454
1455/* Tracepoints definitions. */
1456EXPORT_TRACEPOINT_SYMBOL(kmalloc);
1457EXPORT_TRACEPOINT_SYMBOL(kmem_cache_alloc);
1458EXPORT_TRACEPOINT_SYMBOL(kfree);
1459EXPORT_TRACEPOINT_SYMBOL(kmem_cache_free);
1460
1461int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
1462{
1463 if (__should_failslab(s, gfpflags))
1464 return -ENOMEM;
1465 return 0;
1466}
1467ALLOW_ERROR_INJECTION(should_failslab, ERRNO);