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