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