1// SPDX-License-Identifier: GPL-2.0
2/*
3 * SLUB: A slab allocator that limits cache line use instead of queuing
4 * objects in per cpu and per node lists.
5 *
6 * The allocator synchronizes using per slab locks or atomic operations
7 * and only uses a centralized lock to manage a pool of partial slabs.
8 *
9 * (C) 2007 SGI, Christoph Lameter
10 * (C) 2011 Linux Foundation, Christoph Lameter
11 */
12
13#include <linux/mm.h>
14#include <linux/swap.h> /* struct reclaim_state */
15#include <linux/module.h>
16#include <linux/bit_spinlock.h>
17#include <linux/interrupt.h>
18#include <linux/swab.h>
19#include <linux/bitops.h>
20#include <linux/slab.h>
21#include "slab.h"
22#include <linux/proc_fs.h>
23#include <linux/seq_file.h>
24#include <linux/kasan.h>
25#include <linux/cpu.h>
26#include <linux/cpuset.h>
27#include <linux/mempolicy.h>
28#include <linux/ctype.h>
29#include <linux/stackdepot.h>
30#include <linux/debugobjects.h>
31#include <linux/kallsyms.h>
32#include <linux/kfence.h>
33#include <linux/memory.h>
34#include <linux/math64.h>
35#include <linux/fault-inject.h>
36#include <linux/stacktrace.h>
37#include <linux/prefetch.h>
38#include <linux/memcontrol.h>
39#include <linux/random.h>
40#include <kunit/test.h>
41#include <linux/sort.h>
42
43#include <linux/debugfs.h>
44#include <trace/events/kmem.h>
45
46#include "internal.h"
47
48/*
49 * Lock order:
50 *   1. slab_mutex (Global Mutex)
51 *   2. node->list_lock (Spinlock)
52 *   3. kmem_cache->cpu_slab->lock (Local lock)
53 *   4. slab_lock(slab) (Only on some arches or for debugging)
54 *   5. object_map_lock (Only for debugging)
55 *
56 *   slab_mutex
57 *
58 *   The role of the slab_mutex is to protect the list of all the slabs
59 *   and to synchronize major metadata changes to slab cache structures.
60 *   Also synchronizes memory hotplug callbacks.
61 *
62 *   slab_lock
63 *
64 *   The slab_lock is a wrapper around the page lock, thus it is a bit
65 *   spinlock.
66 *
67 *   The slab_lock is only used for debugging and on arches that do not
68 *   have the ability to do a cmpxchg_double. It only protects:
69 *	A. slab->freelist	-> List of free objects in a slab
70 *	B. slab->inuse		-> Number of objects in use
71 *	C. slab->objects	-> Number of objects in slab
72 *	D. slab->frozen		-> frozen state
73 *
74 *   Frozen slabs
75 *
76 *   If a slab is frozen then it is exempt from list management. It is not
77 *   on any list except per cpu partial list. The processor that froze the
78 *   slab is the one who can perform list operations on the slab. Other
79 *   processors may put objects onto the freelist but the processor that
80 *   froze the slab is the only one that can retrieve the objects from the
81 *   slab's freelist.
82 *
83 *   list_lock
84 *
85 *   The list_lock protects the partial and full list on each node and
86 *   the partial slab counter. If taken then no new slabs may be added or
87 *   removed from the lists nor make the number of partial slabs be modified.
88 *   (Note that the total number of slabs is an atomic value that may be
89 *   modified without taking the list lock).
90 *
91 *   The list_lock is a centralized lock and thus we avoid taking it as
92 *   much as possible. As long as SLUB does not have to handle partial
93 *   slabs, operations can continue without any centralized lock. F.e.
94 *   allocating a long series of objects that fill up slabs does not require
95 *   the list lock.
96 *
97 *   cpu_slab->lock local lock
98 *
99 *   This locks protect slowpath manipulation of all kmem_cache_cpu fields
100 *   except the stat counters. This is a percpu structure manipulated only by
101 *   the local cpu, so the lock protects against being preempted or interrupted
102 *   by an irq. Fast path operations rely on lockless operations instead.
103 *   On PREEMPT_RT, the local lock does not actually disable irqs (and thus
104 *   prevent the lockless operations), so fastpath operations also need to take
105 *   the lock and are no longer lockless.
106 *
107 *   lockless fastpaths
108 *
109 *   The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
110 *   are fully lockless when satisfied from the percpu slab (and when
111 *   cmpxchg_double is possible to use, otherwise slab_lock is taken).
112 *   They also don't disable preemption or migration or irqs. They rely on
113 *   the transaction id (tid) field to detect being preempted or moved to
114 *   another cpu.
115 *
116 *   irq, preemption, migration considerations
117 *
118 *   Interrupts are disabled as part of list_lock or local_lock operations, or
119 *   around the slab_lock operation, in order to make the slab allocator safe
120 *   to use in the context of an irq.
121 *
122 *   In addition, preemption (or migration on PREEMPT_RT) is disabled in the
123 *   allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
124 *   local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
125 *   doesn't have to be revalidated in each section protected by the local lock.
126 *
127 * SLUB assigns one slab for allocation to each processor.
128 * Allocations only occur from these slabs called cpu slabs.
129 *
130 * Slabs with free elements are kept on a partial list and during regular
131 * operations no list for full slabs is used. If an object in a full slab is
132 * freed then the slab will show up again on the partial lists.
133 * We track full slabs for debugging purposes though because otherwise we
134 * cannot scan all objects.
135 *
136 * Slabs are freed when they become empty. Teardown and setup is
137 * minimal so we rely on the page allocators per cpu caches for
138 * fast frees and allocs.
139 *
140 * slab->frozen		The slab is frozen and exempt from list processing.
141 * 			This means that the slab is dedicated to a purpose
142 * 			such as satisfying allocations for a specific
143 * 			processor. Objects may be freed in the slab while
144 * 			it is frozen but slab_free will then skip the usual
145 * 			list operations. It is up to the processor holding
146 * 			the slab to integrate the slab into the slab lists
147 * 			when the slab is no longer needed.
148 *
149 * 			One use of this flag is to mark slabs that are
150 * 			used for allocations. Then such a slab becomes a cpu
151 * 			slab. The cpu slab may be equipped with an additional
152 * 			freelist that allows lockless access to
153 * 			free objects in addition to the regular freelist
154 * 			that requires the slab lock.
155 *
156 * SLAB_DEBUG_FLAGS	Slab requires special handling due to debug
157 * 			options set. This moves	slab handling out of
158 * 			the fast path and disables lockless freelists.
159 */
160
161/*
162 * We could simply use migrate_disable()/enable() but as long as it's a
163 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
164 */
165#ifndef CONFIG_PREEMPT_RT
166#define slub_get_cpu_ptr(var)	get_cpu_ptr(var)
167#define slub_put_cpu_ptr(var)	put_cpu_ptr(var)
168#else
169#define slub_get_cpu_ptr(var)		\
170({					\
171	migrate_disable();		\
172	this_cpu_ptr(var);		\
173})
174#define slub_put_cpu_ptr(var)		\
175do {					\
176	(void)(var);			\
177	migrate_enable();		\
178} while (0)
179#endif
180
181#ifdef CONFIG_SLUB_DEBUG
182#ifdef CONFIG_SLUB_DEBUG_ON
183DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
184#else
185DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
186#endif
187#endif		/* CONFIG_SLUB_DEBUG */
188
189static inline bool kmem_cache_debug(struct kmem_cache *s)
190{
191	return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
192}
193
194void *fixup_red_left(struct kmem_cache *s, void *p)
195{
196	if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
197		p += s->red_left_pad;
198
199	return p;
200}
201
202static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
203{
204#ifdef CONFIG_SLUB_CPU_PARTIAL
205	return !kmem_cache_debug(s);
206#else
207	return false;
208#endif
209}
210
211/*
212 * Issues still to be resolved:
213 *
214 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
215 *
216 * - Variable sizing of the per node arrays
217 */
218
219/* Enable to log cmpxchg failures */
220#undef SLUB_DEBUG_CMPXCHG
221
222/*
223 * Minimum number of partial slabs. These will be left on the partial
224 * lists even if they are empty. kmem_cache_shrink may reclaim them.
225 */
226#define MIN_PARTIAL 5
227
228/*
229 * Maximum number of desirable partial slabs.
230 * The existence of more partial slabs makes kmem_cache_shrink
231 * sort the partial list by the number of objects in use.
232 */
233#define MAX_PARTIAL 10
234
235#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
236				SLAB_POISON | SLAB_STORE_USER)
237
238/*
239 * These debug flags cannot use CMPXCHG because there might be consistency
240 * issues when checking or reading debug information
241 */
242#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
243				SLAB_TRACE)
244
245
246/*
247 * Debugging flags that require metadata to be stored in the slab.  These get
248 * disabled when slub_debug=O is used and a cache's min order increases with
249 * metadata.
250 */
251#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
252
253#define OO_SHIFT	16
254#define OO_MASK		((1 << OO_SHIFT) - 1)
255#define MAX_OBJS_PER_PAGE	32767 /* since slab.objects is u15 */
256
257/* Internal SLUB flags */
258/* Poison object */
259#define __OBJECT_POISON		((slab_flags_t __force)0x80000000U)
260/* Use cmpxchg_double */
261#define __CMPXCHG_DOUBLE	((slab_flags_t __force)0x40000000U)
262
263/*
264 * Tracking user of a slab.
265 */
266#define TRACK_ADDRS_COUNT 16
267struct track {
268	unsigned long addr;	/* Called from address */
269#ifdef CONFIG_STACKDEPOT
270	depot_stack_handle_t handle;
271#endif
272	int cpu;		/* Was running on cpu */
273	int pid;		/* Pid context */
274	unsigned long when;	/* When did the operation occur */
275};
276
277enum track_item { TRACK_ALLOC, TRACK_FREE };
278
279#ifdef CONFIG_SYSFS
280static int sysfs_slab_add(struct kmem_cache *);
281static int sysfs_slab_alias(struct kmem_cache *, const char *);
282#else
283static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
284static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
285							{ return 0; }
286#endif
287
288#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
289static void debugfs_slab_add(struct kmem_cache *);
290#else
291static inline void debugfs_slab_add(struct kmem_cache *s) { }
292#endif
293
294static inline void stat(const struct kmem_cache *s, enum stat_item si)
295{
296#ifdef CONFIG_SLUB_STATS
297	/*
298	 * The rmw is racy on a preemptible kernel but this is acceptable, so
299	 * avoid this_cpu_add()'s irq-disable overhead.
300	 */
301	raw_cpu_inc(s->cpu_slab->stat[si]);
302#endif
303}
304
305/*
306 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
307 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
308 * differ during memory hotplug/hotremove operations.
309 * Protected by slab_mutex.
310 */
311static nodemask_t slab_nodes;
312
313/*
314 * Workqueue used for flush_cpu_slab().
315 */
316static struct workqueue_struct *flushwq;
317
318/********************************************************************
319 * 			Core slab cache functions
320 *******************************************************************/
321
322/*
323 * Returns freelist pointer (ptr). With hardening, this is obfuscated
324 * with an XOR of the address where the pointer is held and a per-cache
325 * random number.
326 */
327static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
328				 unsigned long ptr_addr)
329{
330#ifdef CONFIG_SLAB_FREELIST_HARDENED
331	/*
332	 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
333	 * Normally, this doesn't cause any issues, as both set_freepointer()
334	 * and get_freepointer() are called with a pointer with the same tag.
335	 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
336	 * example, when __free_slub() iterates over objects in a cache, it
337	 * passes untagged pointers to check_object(). check_object() in turns
338	 * calls get_freepointer() with an untagged pointer, which causes the
339	 * freepointer to be restored incorrectly.
340	 */
341	return (void *)((unsigned long)ptr ^ s->random ^
342			swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
343#else
344	return ptr;
345#endif
346}
347
348/* Returns the freelist pointer recorded at location ptr_addr. */
349static inline void *freelist_dereference(const struct kmem_cache *s,
350					 void *ptr_addr)
351{
352	return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
353			    (unsigned long)ptr_addr);
354}
355
356static inline void *get_freepointer(struct kmem_cache *s, void *object)
357{
358	object = kasan_reset_tag(object);
359	return freelist_dereference(s, object + s->offset);
360}
361
362static void prefetch_freepointer(const struct kmem_cache *s, void *object)
363{
364	prefetchw(object + s->offset);
365}
366
367static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
368{
369	unsigned long freepointer_addr;
370	void *p;
371
372	if (!debug_pagealloc_enabled_static())
373		return get_freepointer(s, object);
374
375	object = kasan_reset_tag(object);
376	freepointer_addr = (unsigned long)object + s->offset;
377	copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
378	return freelist_ptr(s, p, freepointer_addr);
379}
380
381static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
382{
383	unsigned long freeptr_addr = (unsigned long)object + s->offset;
384
385#ifdef CONFIG_SLAB_FREELIST_HARDENED
386	BUG_ON(object == fp); /* naive detection of double free or corruption */
387#endif
388
389	freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
390	*(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
391}
392
393/* Loop over all objects in a slab */
394#define for_each_object(__p, __s, __addr, __objects) \
395	for (__p = fixup_red_left(__s, __addr); \
396		__p < (__addr) + (__objects) * (__s)->size; \
397		__p += (__s)->size)
398
399static inline unsigned int order_objects(unsigned int order, unsigned int size)
400{
401	return ((unsigned int)PAGE_SIZE << order) / size;
402}
403
404static inline struct kmem_cache_order_objects oo_make(unsigned int order,
405		unsigned int size)
406{
407	struct kmem_cache_order_objects x = {
408		(order << OO_SHIFT) + order_objects(order, size)
409	};
410
411	return x;
412}
413
414static inline unsigned int oo_order(struct kmem_cache_order_objects x)
415{
416	return x.x >> OO_SHIFT;
417}
418
419static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
420{
421	return x.x & OO_MASK;
422}
423
424#ifdef CONFIG_SLUB_CPU_PARTIAL
425static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
426{
427	unsigned int nr_slabs;
428
429	s->cpu_partial = nr_objects;
430
431	/*
432	 * We take the number of objects but actually limit the number of
433	 * slabs on the per cpu partial list, in order to limit excessive
434	 * growth of the list. For simplicity we assume that the slabs will
435	 * be half-full.
436	 */
437	nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
438	s->cpu_partial_slabs = nr_slabs;
439}
440#else
441static inline void
442slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
443{
444}
445#endif /* CONFIG_SLUB_CPU_PARTIAL */
446
447/*
448 * Per slab locking using the pagelock
449 */
450static __always_inline void __slab_lock(struct slab *slab)
451{
452	struct page *page = slab_page(slab);
453
454	VM_BUG_ON_PAGE(PageTail(page), page);
455	bit_spin_lock(PG_locked, &page->flags);
456}
457
458static __always_inline void __slab_unlock(struct slab *slab)
459{
460	struct page *page = slab_page(slab);
461
462	VM_BUG_ON_PAGE(PageTail(page), page);
463	__bit_spin_unlock(PG_locked, &page->flags);
464}
465
466static __always_inline void slab_lock(struct slab *slab, unsigned long *flags)
467{
468	if (IS_ENABLED(CONFIG_PREEMPT_RT))
469		local_irq_save(*flags);
470	__slab_lock(slab);
471}
472
473static __always_inline void slab_unlock(struct slab *slab, unsigned long *flags)
474{
475	__slab_unlock(slab);
476	if (IS_ENABLED(CONFIG_PREEMPT_RT))
477		local_irq_restore(*flags);
478}
479
480/*
481 * Interrupts must be disabled (for the fallback code to work right), typically
482 * by an _irqsave() lock variant. Except on PREEMPT_RT where locks are different
483 * so we disable interrupts as part of slab_[un]lock().
484 */
485static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
486		void *freelist_old, unsigned long counters_old,
487		void *freelist_new, unsigned long counters_new,
488		const char *n)
489{
490	if (!IS_ENABLED(CONFIG_PREEMPT_RT))
491		lockdep_assert_irqs_disabled();
492#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
493    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
494	if (s->flags & __CMPXCHG_DOUBLE) {
495		if (cmpxchg_double(&slab->freelist, &slab->counters,
496				   freelist_old, counters_old,
497				   freelist_new, counters_new))
498			return true;
499	} else
500#endif
501	{
502		/* init to 0 to prevent spurious warnings */
503		unsigned long flags = 0;
504
505		slab_lock(slab, &flags);
506		if (slab->freelist == freelist_old &&
507					slab->counters == counters_old) {
508			slab->freelist = freelist_new;
509			slab->counters = counters_new;
510			slab_unlock(slab, &flags);
511			return true;
512		}
513		slab_unlock(slab, &flags);
514	}
515
516	cpu_relax();
517	stat(s, CMPXCHG_DOUBLE_FAIL);
518
519#ifdef SLUB_DEBUG_CMPXCHG
520	pr_info("%s %s: cmpxchg double redo ", n, s->name);
521#endif
522
523	return false;
524}
525
526static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
527		void *freelist_old, unsigned long counters_old,
528		void *freelist_new, unsigned long counters_new,
529		const char *n)
530{
531#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
532    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
533	if (s->flags & __CMPXCHG_DOUBLE) {
534		if (cmpxchg_double(&slab->freelist, &slab->counters,
535				   freelist_old, counters_old,
536				   freelist_new, counters_new))
537			return true;
538	} else
539#endif
540	{
541		unsigned long flags;
542
543		local_irq_save(flags);
544		__slab_lock(slab);
545		if (slab->freelist == freelist_old &&
546					slab->counters == counters_old) {
547			slab->freelist = freelist_new;
548			slab->counters = counters_new;
549			__slab_unlock(slab);
550			local_irq_restore(flags);
551			return true;
552		}
553		__slab_unlock(slab);
554		local_irq_restore(flags);
555	}
556
557	cpu_relax();
558	stat(s, CMPXCHG_DOUBLE_FAIL);
559
560#ifdef SLUB_DEBUG_CMPXCHG
561	pr_info("%s %s: cmpxchg double redo ", n, s->name);
562#endif
563
564	return false;
565}
566
567#ifdef CONFIG_SLUB_DEBUG
568static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
569static DEFINE_RAW_SPINLOCK(object_map_lock);
570
571static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
572		       struct slab *slab)
573{
574	void *addr = slab_address(slab);
575	void *p;
576
577	bitmap_zero(obj_map, slab->objects);
578
579	for (p = slab->freelist; p; p = get_freepointer(s, p))
580		set_bit(__obj_to_index(s, addr, p), obj_map);
581}
582
583#if IS_ENABLED(CONFIG_KUNIT)
584static bool slab_add_kunit_errors(void)
585{
586	struct kunit_resource *resource;
587
588	if (likely(!current->kunit_test))
589		return false;
590
591	resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
592	if (!resource)
593		return false;
594
595	(*(int *)resource->data)++;
596	kunit_put_resource(resource);
597	return true;
598}
599#else
600static inline bool slab_add_kunit_errors(void) { return false; }
601#endif
602
603/*
604 * Determine a map of objects in use in a slab.
605 *
606 * Node listlock must be held to guarantee that the slab does
607 * not vanish from under us.
608 */
609static unsigned long *get_map(struct kmem_cache *s, struct slab *slab)
610	__acquires(&object_map_lock)
611{
612	VM_BUG_ON(!irqs_disabled());
613
614	raw_spin_lock(&object_map_lock);
615
616	__fill_map(object_map, s, slab);
617
618	return object_map;
619}
620
621static void put_map(unsigned long *map) __releases(&object_map_lock)
622{
623	VM_BUG_ON(map != object_map);
624	raw_spin_unlock(&object_map_lock);
625}
626
627static inline unsigned int size_from_object(struct kmem_cache *s)
628{
629	if (s->flags & SLAB_RED_ZONE)
630		return s->size - s->red_left_pad;
631
632	return s->size;
633}
634
635static inline void *restore_red_left(struct kmem_cache *s, void *p)
636{
637	if (s->flags & SLAB_RED_ZONE)
638		p -= s->red_left_pad;
639
640	return p;
641}
642
643/*
644 * Debug settings:
645 */
646#if defined(CONFIG_SLUB_DEBUG_ON)
647static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
648#else
649static slab_flags_t slub_debug;
650#endif
651
652static char *slub_debug_string;
653static int disable_higher_order_debug;
654
655/*
656 * slub is about to manipulate internal object metadata.  This memory lies
657 * outside the range of the allocated object, so accessing it would normally
658 * be reported by kasan as a bounds error.  metadata_access_enable() is used
659 * to tell kasan that these accesses are OK.
660 */
661static inline void metadata_access_enable(void)
662{
663	kasan_disable_current();
664}
665
666static inline void metadata_access_disable(void)
667{
668	kasan_enable_current();
669}
670
671/*
672 * Object debugging
673 */
674
675/* Verify that a pointer has an address that is valid within a slab page */
676static inline int check_valid_pointer(struct kmem_cache *s,
677				struct slab *slab, void *object)
678{
679	void *base;
680
681	if (!object)
682		return 1;
683
684	base = slab_address(slab);
685	object = kasan_reset_tag(object);
686	object = restore_red_left(s, object);
687	if (object < base || object >= base + slab->objects * s->size ||
688		(object - base) % s->size) {
689		return 0;
690	}
691
692	return 1;
693}
694
695static void print_section(char *level, char *text, u8 *addr,
696			  unsigned int length)
697{
698	metadata_access_enable();
699	print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
700			16, 1, kasan_reset_tag((void *)addr), length, 1);
701	metadata_access_disable();
702}
703
704/*
705 * See comment in calculate_sizes().
706 */
707static inline bool freeptr_outside_object(struct kmem_cache *s)
708{
709	return s->offset >= s->inuse;
710}
711
712/*
713 * Return offset of the end of info block which is inuse + free pointer if
714 * not overlapping with object.
715 */
716static inline unsigned int get_info_end(struct kmem_cache *s)
717{
718	if (freeptr_outside_object(s))
719		return s->inuse + sizeof(void *);
720	else
721		return s->inuse;
722}
723
724static struct track *get_track(struct kmem_cache *s, void *object,
725	enum track_item alloc)
726{
727	struct track *p;
728
729	p = object + get_info_end(s);
730
731	return kasan_reset_tag(p + alloc);
732}
733
734#ifdef CONFIG_STACKDEPOT
735static noinline depot_stack_handle_t set_track_prepare(void)
736{
737	depot_stack_handle_t handle;
738	unsigned long entries[TRACK_ADDRS_COUNT];
739	unsigned int nr_entries;
740
741	nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
742	handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
743
744	return handle;
745}
746#else
747static inline depot_stack_handle_t set_track_prepare(void)
748{
749	return 0;
750}
751#endif
752
753static void set_track_update(struct kmem_cache *s, void *object,
754			     enum track_item alloc, unsigned long addr,
755			     depot_stack_handle_t handle)
756{
757	struct track *p = get_track(s, object, alloc);
758
759#ifdef CONFIG_STACKDEPOT
760	p->handle = handle;
761#endif
762	p->addr = addr;
763	p->cpu = smp_processor_id();
764	p->pid = current->pid;
765	p->when = jiffies;
766}
767
768static __always_inline void set_track(struct kmem_cache *s, void *object,
769				      enum track_item alloc, unsigned long addr)
770{
771	depot_stack_handle_t handle = set_track_prepare();
772
773	set_track_update(s, object, alloc, addr, handle);
774}
775
776static void init_tracking(struct kmem_cache *s, void *object)
777{
778	struct track *p;
779
780	if (!(s->flags & SLAB_STORE_USER))
781		return;
782
783	p = get_track(s, object, TRACK_ALLOC);
784	memset(p, 0, 2*sizeof(struct track));
785}
786
787static void print_track(const char *s, struct track *t, unsigned long pr_time)
788{
789	depot_stack_handle_t handle __maybe_unused;
790
791	if (!t->addr)
792		return;
793
794	pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
795	       s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
796#ifdef CONFIG_STACKDEPOT
797	handle = READ_ONCE(t->handle);
798	if (handle)
799		stack_depot_print(handle);
800	else
801		pr_err("object allocation/free stack trace missing\n");
802#endif
803}
804
805void print_tracking(struct kmem_cache *s, void *object)
806{
807	unsigned long pr_time = jiffies;
808	if (!(s->flags & SLAB_STORE_USER))
809		return;
810
811	print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
812	print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
813}
814
815static void print_slab_info(const struct slab *slab)
816{
817	struct folio *folio = (struct folio *)slab_folio(slab);
818
819	pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
820	       slab, slab->objects, slab->inuse, slab->freelist,
821	       folio_flags(folio, 0));
822}
823
824static void slab_bug(struct kmem_cache *s, char *fmt, ...)
825{
826	struct va_format vaf;
827	va_list args;
828
829	va_start(args, fmt);
830	vaf.fmt = fmt;
831	vaf.va = &args;
832	pr_err("=============================================================================\n");
833	pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
834	pr_err("-----------------------------------------------------------------------------\n\n");
835	va_end(args);
836}
837
838__printf(2, 3)
839static void slab_fix(struct kmem_cache *s, char *fmt, ...)
840{
841	struct va_format vaf;
842	va_list args;
843
844	if (slab_add_kunit_errors())
845		return;
846
847	va_start(args, fmt);
848	vaf.fmt = fmt;
849	vaf.va = &args;
850	pr_err("FIX %s: %pV\n", s->name, &vaf);
851	va_end(args);
852}
853
854static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
855{
856	unsigned int off;	/* Offset of last byte */
857	u8 *addr = slab_address(slab);
858
859	print_tracking(s, p);
860
861	print_slab_info(slab);
862
863	pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
864	       p, p - addr, get_freepointer(s, p));
865
866	if (s->flags & SLAB_RED_ZONE)
867		print_section(KERN_ERR, "Redzone  ", p - s->red_left_pad,
868			      s->red_left_pad);
869	else if (p > addr + 16)
870		print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
871
872	print_section(KERN_ERR,         "Object   ", p,
873		      min_t(unsigned int, s->object_size, PAGE_SIZE));
874	if (s->flags & SLAB_RED_ZONE)
875		print_section(KERN_ERR, "Redzone  ", p + s->object_size,
876			s->inuse - s->object_size);
877
878	off = get_info_end(s);
879
880	if (s->flags & SLAB_STORE_USER)
881		off += 2 * sizeof(struct track);
882
883	off += kasan_metadata_size(s);
884
885	if (off != size_from_object(s))
886		/* Beginning of the filler is the free pointer */
887		print_section(KERN_ERR, "Padding  ", p + off,
888			      size_from_object(s) - off);
889
890	dump_stack();
891}
892
893static void object_err(struct kmem_cache *s, struct slab *slab,
894			u8 *object, char *reason)
895{
896	if (slab_add_kunit_errors())
897		return;
898
899	slab_bug(s, "%s", reason);
900	print_trailer(s, slab, object);
901	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
902}
903
904static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
905			       void **freelist, void *nextfree)
906{
907	if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
908	    !check_valid_pointer(s, slab, nextfree) && freelist) {
909		object_err(s, slab, *freelist, "Freechain corrupt");
910		*freelist = NULL;
911		slab_fix(s, "Isolate corrupted freechain");
912		return true;
913	}
914
915	return false;
916}
917
918static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
919			const char *fmt, ...)
920{
921	va_list args;
922	char buf[100];
923
924	if (slab_add_kunit_errors())
925		return;
926
927	va_start(args, fmt);
928	vsnprintf(buf, sizeof(buf), fmt, args);
929	va_end(args);
930	slab_bug(s, "%s", buf);
931	print_slab_info(slab);
932	dump_stack();
933	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
934}
935
936static void init_object(struct kmem_cache *s, void *object, u8 val)
937{
938	u8 *p = kasan_reset_tag(object);
939
940	if (s->flags & SLAB_RED_ZONE)
941		memset(p - s->red_left_pad, val, s->red_left_pad);
942
943	if (s->flags & __OBJECT_POISON) {
944		memset(p, POISON_FREE, s->object_size - 1);
945		p[s->object_size - 1] = POISON_END;
946	}
947
948	if (s->flags & SLAB_RED_ZONE)
949		memset(p + s->object_size, val, s->inuse - s->object_size);
950}
951
952static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
953						void *from, void *to)
954{
955	slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
956	memset(from, data, to - from);
957}
958
959static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
960			u8 *object, char *what,
961			u8 *start, unsigned int value, unsigned int bytes)
962{
963	u8 *fault;
964	u8 *end;
965	u8 *addr = slab_address(slab);
966
967	metadata_access_enable();
968	fault = memchr_inv(kasan_reset_tag(start), value, bytes);
969	metadata_access_disable();
970	if (!fault)
971		return 1;
972
973	end = start + bytes;
974	while (end > fault && end[-1] == value)
975		end--;
976
977	if (slab_add_kunit_errors())
978		goto skip_bug_print;
979
980	slab_bug(s, "%s overwritten", what);
981	pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
982					fault, end - 1, fault - addr,
983					fault[0], value);
984	print_trailer(s, slab, object);
985	add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
986
987skip_bug_print:
988	restore_bytes(s, what, value, fault, end);
989	return 0;
990}
991
992/*
993 * Object layout:
994 *
995 * object address
996 * 	Bytes of the object to be managed.
997 * 	If the freepointer may overlay the object then the free
998 *	pointer is at the middle of the object.
999 *
1000 * 	Poisoning uses 0x6b (POISON_FREE) and the last byte is
1001 * 	0xa5 (POISON_END)
1002 *
1003 * object + s->object_size
1004 * 	Padding to reach word boundary. This is also used for Redzoning.
1005 * 	Padding is extended by another word if Redzoning is enabled and
1006 * 	object_size == inuse.
1007 *
1008 * 	We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1009 * 	0xcc (RED_ACTIVE) for objects in use.
1010 *
1011 * object + s->inuse
1012 * 	Meta data starts here.
1013 *
1014 * 	A. Free pointer (if we cannot overwrite object on free)
1015 * 	B. Tracking data for SLAB_STORE_USER
1016 *	C. Padding to reach required alignment boundary or at minimum
1017 * 		one word if debugging is on to be able to detect writes
1018 * 		before the word boundary.
1019 *
1020 *	Padding is done using 0x5a (POISON_INUSE)
1021 *
1022 * object + s->size
1023 * 	Nothing is used beyond s->size.
1024 *
1025 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1026 * ignored. And therefore no slab options that rely on these boundaries
1027 * may be used with merged slabcaches.
1028 */
1029
1030static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1031{
1032	unsigned long off = get_info_end(s);	/* The end of info */
1033
1034	if (s->flags & SLAB_STORE_USER)
1035		/* We also have user information there */
1036		off += 2 * sizeof(struct track);
1037
1038	off += kasan_metadata_size(s);
1039
1040	if (size_from_object(s) == off)
1041		return 1;
1042
1043	return check_bytes_and_report(s, slab, p, "Object padding",
1044			p + off, POISON_INUSE, size_from_object(s) - off);
1045}
1046
1047/* Check the pad bytes at the end of a slab page */
1048static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1049{
1050	u8 *start;
1051	u8 *fault;
1052	u8 *end;
1053	u8 *pad;
1054	int length;
1055	int remainder;
1056
1057	if (!(s->flags & SLAB_POISON))
1058		return;
1059
1060	start = slab_address(slab);
1061	length = slab_size(slab);
1062	end = start + length;
1063	remainder = length % s->size;
1064	if (!remainder)
1065		return;
1066
1067	pad = end - remainder;
1068	metadata_access_enable();
1069	fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1070	metadata_access_disable();
1071	if (!fault)
1072		return;
1073	while (end > fault && end[-1] == POISON_INUSE)
1074		end--;
1075
1076	slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1077			fault, end - 1, fault - start);
1078	print_section(KERN_ERR, "Padding ", pad, remainder);
1079
1080	restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1081}
1082
1083static int check_object(struct kmem_cache *s, struct slab *slab,
1084					void *object, u8 val)
1085{
1086	u8 *p = object;
1087	u8 *endobject = object + s->object_size;
1088
1089	if (s->flags & SLAB_RED_ZONE) {
1090		if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1091			object - s->red_left_pad, val, s->red_left_pad))
1092			return 0;
1093
1094		if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1095			endobject, val, s->inuse - s->object_size))
1096			return 0;
1097	} else {
1098		if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1099			check_bytes_and_report(s, slab, p, "Alignment padding",
1100				endobject, POISON_INUSE,
1101				s->inuse - s->object_size);
1102		}
1103	}
1104
1105	if (s->flags & SLAB_POISON) {
1106		if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1107			(!check_bytes_and_report(s, slab, p, "Poison", p,
1108					POISON_FREE, s->object_size - 1) ||
1109			 !check_bytes_and_report(s, slab, p, "End Poison",
1110				p + s->object_size - 1, POISON_END, 1)))
1111			return 0;
1112		/*
1113		 * check_pad_bytes cleans up on its own.
1114		 */
1115		check_pad_bytes(s, slab, p);
1116	}
1117
1118	if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1119		/*
1120		 * Object and freepointer overlap. Cannot check
1121		 * freepointer while object is allocated.
1122		 */
1123		return 1;
1124
1125	/* Check free pointer validity */
1126	if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1127		object_err(s, slab, p, "Freepointer corrupt");
1128		/*
1129		 * No choice but to zap it and thus lose the remainder
1130		 * of the free objects in this slab. May cause
1131		 * another error because the object count is now wrong.
1132		 */
1133		set_freepointer(s, p, NULL);
1134		return 0;
1135	}
1136	return 1;
1137}
1138
1139static int check_slab(struct kmem_cache *s, struct slab *slab)
1140{
1141	int maxobj;
1142
1143	if (!folio_test_slab(slab_folio(slab))) {
1144		slab_err(s, slab, "Not a valid slab page");
1145		return 0;
1146	}
1147
1148	maxobj = order_objects(slab_order(slab), s->size);
1149	if (slab->objects > maxobj) {
1150		slab_err(s, slab, "objects %u > max %u",
1151			slab->objects, maxobj);
1152		return 0;
1153	}
1154	if (slab->inuse > slab->objects) {
1155		slab_err(s, slab, "inuse %u > max %u",
1156			slab->inuse, slab->objects);
1157		return 0;
1158	}
1159	/* Slab_pad_check fixes things up after itself */
1160	slab_pad_check(s, slab);
1161	return 1;
1162}
1163
1164/*
1165 * Determine if a certain object in a slab is on the freelist. Must hold the
1166 * slab lock to guarantee that the chains are in a consistent state.
1167 */
1168static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1169{
1170	int nr = 0;
1171	void *fp;
1172	void *object = NULL;
1173	int max_objects;
1174
1175	fp = slab->freelist;
1176	while (fp && nr <= slab->objects) {
1177		if (fp == search)
1178			return 1;
1179		if (!check_valid_pointer(s, slab, fp)) {
1180			if (object) {
1181				object_err(s, slab, object,
1182					"Freechain corrupt");
1183				set_freepointer(s, object, NULL);
1184			} else {
1185				slab_err(s, slab, "Freepointer corrupt");
1186				slab->freelist = NULL;
1187				slab->inuse = slab->objects;
1188				slab_fix(s, "Freelist cleared");
1189				return 0;
1190			}
1191			break;
1192		}
1193		object = fp;
1194		fp = get_freepointer(s, object);
1195		nr++;
1196	}
1197
1198	max_objects = order_objects(slab_order(slab), s->size);
1199	if (max_objects > MAX_OBJS_PER_PAGE)
1200		max_objects = MAX_OBJS_PER_PAGE;
1201
1202	if (slab->objects != max_objects) {
1203		slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1204			 slab->objects, max_objects);
1205		slab->objects = max_objects;
1206		slab_fix(s, "Number of objects adjusted");
1207	}
1208	if (slab->inuse != slab->objects - nr) {
1209		slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1210			 slab->inuse, slab->objects - nr);
1211		slab->inuse = slab->objects - nr;
1212		slab_fix(s, "Object count adjusted");
1213	}
1214	return search == NULL;
1215}
1216
1217static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1218								int alloc)
1219{
1220	if (s->flags & SLAB_TRACE) {
1221		pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1222			s->name,
1223			alloc ? "alloc" : "free",
1224			object, slab->inuse,
1225			slab->freelist);
1226
1227		if (!alloc)
1228			print_section(KERN_INFO, "Object ", (void *)object,
1229					s->object_size);
1230
1231		dump_stack();
1232	}
1233}
1234
1235/*
1236 * Tracking of fully allocated slabs for debugging purposes.
1237 */
1238static void add_full(struct kmem_cache *s,
1239	struct kmem_cache_node *n, struct slab *slab)
1240{
1241	if (!(s->flags & SLAB_STORE_USER))
1242		return;
1243
1244	lockdep_assert_held(&n->list_lock);
1245	list_add(&slab->slab_list, &n->full);
1246}
1247
1248static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1249{
1250	if (!(s->flags & SLAB_STORE_USER))
1251		return;
1252
1253	lockdep_assert_held(&n->list_lock);
1254	list_del(&slab->slab_list);
1255}
1256
1257/* Tracking of the number of slabs for debugging purposes */
1258static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1259{
1260	struct kmem_cache_node *n = get_node(s, node);
1261
1262	return atomic_long_read(&n->nr_slabs);
1263}
1264
1265static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1266{
1267	return atomic_long_read(&n->nr_slabs);
1268}
1269
1270static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1271{
1272	struct kmem_cache_node *n = get_node(s, node);
1273
1274	/*
1275	 * May be called early in order to allocate a slab for the
1276	 * kmem_cache_node structure. Solve the chicken-egg
1277	 * dilemma by deferring the increment of the count during
1278	 * bootstrap (see early_kmem_cache_node_alloc).
1279	 */
1280	if (likely(n)) {
1281		atomic_long_inc(&n->nr_slabs);
1282		atomic_long_add(objects, &n->total_objects);
1283	}
1284}
1285static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1286{
1287	struct kmem_cache_node *n = get_node(s, node);
1288
1289	atomic_long_dec(&n->nr_slabs);
1290	atomic_long_sub(objects, &n->total_objects);
1291}
1292
1293/* Object debug checks for alloc/free paths */
1294static void setup_object_debug(struct kmem_cache *s, void *object)
1295{
1296	if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1297		return;
1298
1299	init_object(s, object, SLUB_RED_INACTIVE);
1300	init_tracking(s, object);
1301}
1302
1303static
1304void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1305{
1306	if (!kmem_cache_debug_flags(s, SLAB_POISON))
1307		return;
1308
1309	metadata_access_enable();
1310	memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1311	metadata_access_disable();
1312}
1313
1314static inline int alloc_consistency_checks(struct kmem_cache *s,
1315					struct slab *slab, void *object)
1316{
1317	if (!check_slab(s, slab))
1318		return 0;
1319
1320	if (!check_valid_pointer(s, slab, object)) {
1321		object_err(s, slab, object, "Freelist Pointer check fails");
1322		return 0;
1323	}
1324
1325	if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1326		return 0;
1327
1328	return 1;
1329}
1330
1331static noinline int alloc_debug_processing(struct kmem_cache *s,
1332					struct slab *slab,
1333					void *object, unsigned long addr)
1334{
1335	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1336		if (!alloc_consistency_checks(s, slab, object))
1337			goto bad;
1338	}
1339
1340	/* Success perform special debug activities for allocs */
1341	if (s->flags & SLAB_STORE_USER)
1342		set_track(s, object, TRACK_ALLOC, addr);
1343	trace(s, slab, object, 1);
1344	init_object(s, object, SLUB_RED_ACTIVE);
1345	return 1;
1346
1347bad:
1348	if (folio_test_slab(slab_folio(slab))) {
1349		/*
1350		 * If this is a slab page then lets do the best we can
1351		 * to avoid issues in the future. Marking all objects
1352		 * as used avoids touching the remaining objects.
1353		 */
1354		slab_fix(s, "Marking all objects used");
1355		slab->inuse = slab->objects;
1356		slab->freelist = NULL;
1357	}
1358	return 0;
1359}
1360
1361static inline int free_consistency_checks(struct kmem_cache *s,
1362		struct slab *slab, void *object, unsigned long addr)
1363{
1364	if (!check_valid_pointer(s, slab, object)) {
1365		slab_err(s, slab, "Invalid object pointer 0x%p", object);
1366		return 0;
1367	}
1368
1369	if (on_freelist(s, slab, object)) {
1370		object_err(s, slab, object, "Object already free");
1371		return 0;
1372	}
1373
1374	if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1375		return 0;
1376
1377	if (unlikely(s != slab->slab_cache)) {
1378		if (!folio_test_slab(slab_folio(slab))) {
1379			slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1380				 object);
1381		} else if (!slab->slab_cache) {
1382			pr_err("SLUB <none>: no slab for object 0x%p.\n",
1383			       object);
1384			dump_stack();
1385		} else
1386			object_err(s, slab, object,
1387					"page slab pointer corrupt.");
1388		return 0;
1389	}
1390	return 1;
1391}
1392
1393/* Supports checking bulk free of a constructed freelist */
1394static noinline int free_debug_processing(
1395	struct kmem_cache *s, struct slab *slab,
1396	void *head, void *tail, int bulk_cnt,
1397	unsigned long addr)
1398{
1399	struct kmem_cache_node *n = get_node(s, slab_nid(slab));
1400	void *object = head;
1401	int cnt = 0;
1402	unsigned long flags, flags2;
1403	int ret = 0;
1404	depot_stack_handle_t handle = 0;
1405
1406	if (s->flags & SLAB_STORE_USER)
1407		handle = set_track_prepare();
1408
1409	spin_lock_irqsave(&n->list_lock, flags);
1410	slab_lock(slab, &flags2);
1411
1412	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1413		if (!check_slab(s, slab))
1414			goto out;
1415	}
1416
1417next_object:
1418	cnt++;
1419
1420	if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1421		if (!free_consistency_checks(s, slab, object, addr))
1422			goto out;
1423	}
1424
1425	if (s->flags & SLAB_STORE_USER)
1426		set_track_update(s, object, TRACK_FREE, addr, handle);
1427	trace(s, slab, object, 0);
1428	/* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1429	init_object(s, object, SLUB_RED_INACTIVE);
1430
1431	/* Reached end of constructed freelist yet? */
1432	if (object != tail) {
1433		object = get_freepointer(s, object);
1434		goto next_object;
1435	}
1436	ret = 1;
1437
1438out:
1439	if (cnt != bulk_cnt)
1440		slab_err(s, slab, "Bulk freelist count(%d) invalid(%d)\n",
1441			 bulk_cnt, cnt);
1442
1443	slab_unlock(slab, &flags2);
1444	spin_unlock_irqrestore(&n->list_lock, flags);
1445	if (!ret)
1446		slab_fix(s, "Object at 0x%p not freed", object);
1447	return ret;
1448}
1449
1450/*
1451 * Parse a block of slub_debug options. Blocks are delimited by ';'
1452 *
1453 * @str:    start of block
1454 * @flags:  returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1455 * @slabs:  return start of list of slabs, or NULL when there's no list
1456 * @init:   assume this is initial parsing and not per-kmem-create parsing
1457 *
1458 * returns the start of next block if there's any, or NULL
1459 */
1460static char *
1461parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1462{
1463	bool higher_order_disable = false;
1464
1465	/* Skip any completely empty blocks */
1466	while (*str && *str == ';')
1467		str++;
1468
1469	if (*str == ',') {
1470		/*
1471		 * No options but restriction on slabs. This means full
1472		 * debugging for slabs matching a pattern.
1473		 */
1474		*flags = DEBUG_DEFAULT_FLAGS;
1475		goto check_slabs;
1476	}
1477	*flags = 0;
1478
1479	/* Determine which debug features should be switched on */
1480	for (; *str && *str != ',' && *str != ';'; str++) {
1481		switch (tolower(*str)) {
1482		case '-':
1483			*flags = 0;
1484			break;
1485		case 'f':
1486			*flags |= SLAB_CONSISTENCY_CHECKS;
1487			break;
1488		case 'z':
1489			*flags |= SLAB_RED_ZONE;
1490			break;
1491		case 'p':
1492			*flags |= SLAB_POISON;
1493			break;
1494		case 'u':
1495			*flags |= SLAB_STORE_USER;
1496			break;
1497		case 't':
1498			*flags |= SLAB_TRACE;
1499			break;
1500		case 'a':
1501			*flags |= SLAB_FAILSLAB;
1502			break;
1503		case 'o':
1504			/*
1505			 * Avoid enabling debugging on caches if its minimum
1506			 * order would increase as a result.
1507			 */
1508			higher_order_disable = true;
1509			break;
1510		default:
1511			if (init)
1512				pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1513		}
1514	}
1515check_slabs:
1516	if (*str == ',')
1517		*slabs = ++str;
1518	else
1519		*slabs = NULL;
1520
1521	/* Skip over the slab list */
1522	while (*str && *str != ';')
1523		str++;
1524
1525	/* Skip any completely empty blocks */
1526	while (*str && *str == ';')
1527		str++;
1528
1529	if (init && higher_order_disable)
1530		disable_higher_order_debug = 1;
1531
1532	if (*str)
1533		return str;
1534	else
1535		return NULL;
1536}
1537
1538static int __init setup_slub_debug(char *str)
1539{
1540	slab_flags_t flags;
1541	slab_flags_t global_flags;
1542	char *saved_str;
1543	char *slab_list;
1544	bool global_slub_debug_changed = false;
1545	bool slab_list_specified = false;
1546
1547	global_flags = DEBUG_DEFAULT_FLAGS;
1548	if (*str++ != '=' || !*str)
1549		/*
1550		 * No options specified. Switch on full debugging.
1551		 */
1552		goto out;
1553
1554	saved_str = str;
1555	while (str) {
1556		str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1557
1558		if (!slab_list) {
1559			global_flags = flags;
1560			global_slub_debug_changed = true;
1561		} else {
1562			slab_list_specified = true;
1563			if (flags & SLAB_STORE_USER)
1564				stack_depot_want_early_init();
1565		}
1566	}
1567
1568	/*
1569	 * For backwards compatibility, a single list of flags with list of
1570	 * slabs means debugging is only changed for those slabs, so the global
1571	 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1572	 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1573	 * long as there is no option specifying flags without a slab list.
1574	 */
1575	if (slab_list_specified) {
1576		if (!global_slub_debug_changed)
1577			global_flags = slub_debug;
1578		slub_debug_string = saved_str;
1579	}
1580out:
1581	slub_debug = global_flags;
1582	if (slub_debug & SLAB_STORE_USER)
1583		stack_depot_want_early_init();
1584	if (slub_debug != 0 || slub_debug_string)
1585		static_branch_enable(&slub_debug_enabled);
1586	else
1587		static_branch_disable(&slub_debug_enabled);
1588	if ((static_branch_unlikely(&init_on_alloc) ||
1589	     static_branch_unlikely(&init_on_free)) &&
1590	    (slub_debug & SLAB_POISON))
1591		pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1592	return 1;
1593}
1594
1595__setup("slub_debug", setup_slub_debug);
1596
1597/*
1598 * kmem_cache_flags - apply debugging options to the cache
1599 * @object_size:	the size of an object without meta data
1600 * @flags:		flags to set
1601 * @name:		name of the cache
1602 *
1603 * Debug option(s) are applied to @flags. In addition to the debug
1604 * option(s), if a slab name (or multiple) is specified i.e.
1605 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1606 * then only the select slabs will receive the debug option(s).
1607 */
1608slab_flags_t kmem_cache_flags(unsigned int object_size,
1609	slab_flags_t flags, const char *name)
1610{
1611	char *iter;
1612	size_t len;
1613	char *next_block;
1614	slab_flags_t block_flags;
1615	slab_flags_t slub_debug_local = slub_debug;
1616
1617	if (flags & SLAB_NO_USER_FLAGS)
1618		return flags;
1619
1620	/*
1621	 * If the slab cache is for debugging (e.g. kmemleak) then
1622	 * don't store user (stack trace) information by default,
1623	 * but let the user enable it via the command line below.
1624	 */
1625	if (flags & SLAB_NOLEAKTRACE)
1626		slub_debug_local &= ~SLAB_STORE_USER;
1627
1628	len = strlen(name);
1629	next_block = slub_debug_string;
1630	/* Go through all blocks of debug options, see if any matches our slab's name */
1631	while (next_block) {
1632		next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1633		if (!iter)
1634			continue;
1635		/* Found a block that has a slab list, search it */
1636		while (*iter) {
1637			char *end, *glob;
1638			size_t cmplen;
1639
1640			end = strchrnul(iter, ',');
1641			if (next_block && next_block < end)
1642				end = next_block - 1;
1643
1644			glob = strnchr(iter, end - iter, '*');
1645			if (glob)
1646				cmplen = glob - iter;
1647			else
1648				cmplen = max_t(size_t, len, (end - iter));
1649
1650			if (!strncmp(name, iter, cmplen)) {
1651				flags |= block_flags;
1652				return flags;
1653			}
1654
1655			if (!*end || *end == ';')
1656				break;
1657			iter = end + 1;
1658		}
1659	}
1660
1661	return flags | slub_debug_local;
1662}
1663#else /* !CONFIG_SLUB_DEBUG */
1664static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1665static inline
1666void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1667
1668static inline int alloc_debug_processing(struct kmem_cache *s,
1669	struct slab *slab, void *object, unsigned long addr) { return 0; }
1670
1671static inline int free_debug_processing(
1672	struct kmem_cache *s, struct slab *slab,
1673	void *head, void *tail, int bulk_cnt,
1674	unsigned long addr) { return 0; }
1675
1676static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1677static inline int check_object(struct kmem_cache *s, struct slab *slab,
1678			void *object, u8 val) { return 1; }
1679static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1680					struct slab *slab) {}
1681static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1682					struct slab *slab) {}
1683slab_flags_t kmem_cache_flags(unsigned int object_size,
1684	slab_flags_t flags, const char *name)
1685{
1686	return flags;
1687}
1688#define slub_debug 0
1689
1690#define disable_higher_order_debug 0
1691
1692static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1693							{ return 0; }
1694static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1695							{ return 0; }
1696static inline void inc_slabs_node(struct kmem_cache *s, int node,
1697							int objects) {}
1698static inline void dec_slabs_node(struct kmem_cache *s, int node,
1699							int objects) {}
1700
1701static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1702			       void **freelist, void *nextfree)
1703{
1704	return false;
1705}
1706#endif /* CONFIG_SLUB_DEBUG */
1707
1708/*
1709 * Hooks for other subsystems that check memory allocations. In a typical
1710 * production configuration these hooks all should produce no code at all.
1711 */
1712static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1713{
1714	ptr = kasan_kmalloc_large(ptr, size, flags);
1715	/* As ptr might get tagged, call kmemleak hook after KASAN. */
1716	kmemleak_alloc(ptr, size, 1, flags);
1717	return ptr;
1718}
1719
1720static __always_inline void kfree_hook(void *x)
1721{
1722	kmemleak_free(x);
1723	kasan_kfree_large(x);
1724}
1725
1726static __always_inline bool slab_free_hook(struct kmem_cache *s,
1727						void *x, bool init)
1728{
1729	kmemleak_free_recursive(x, s->flags);
1730
1731	debug_check_no_locks_freed(x, s->object_size);
1732
1733	if (!(s->flags & SLAB_DEBUG_OBJECTS))
1734		debug_check_no_obj_freed(x, s->object_size);
1735
1736	/* Use KCSAN to help debug racy use-after-free. */
1737	if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1738		__kcsan_check_access(x, s->object_size,
1739				     KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1740
1741	/*
1742	 * As memory initialization might be integrated into KASAN,
1743	 * kasan_slab_free and initialization memset's must be
1744	 * kept together to avoid discrepancies in behavior.
1745	 *
1746	 * The initialization memset's clear the object and the metadata,
1747	 * but don't touch the SLAB redzone.
1748	 */
1749	if (init) {
1750		int rsize;
1751
1752		if (!kasan_has_integrated_init())
1753			memset(kasan_reset_tag(x), 0, s->object_size);
1754		rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1755		memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1756		       s->size - s->inuse - rsize);
1757	}
1758	/* KASAN might put x into memory quarantine, delaying its reuse. */
1759	return kasan_slab_free(s, x, init);
1760}
1761
1762static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1763					   void **head, void **tail,
1764					   int *cnt)
1765{
1766
1767	void *object;
1768	void *next = *head;
1769	void *old_tail = *tail ? *tail : *head;
1770
1771	if (is_kfence_address(next)) {
1772		slab_free_hook(s, next, false);
1773		return true;
1774	}
1775
1776	/* Head and tail of the reconstructed freelist */
1777	*head = NULL;
1778	*tail = NULL;
1779
1780	do {
1781		object = next;
1782		next = get_freepointer(s, object);
1783
1784		/* If object's reuse doesn't have to be delayed */
1785		if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1786			/* Move object to the new freelist */
1787			set_freepointer(s, object, *head);
1788			*head = object;
1789			if (!*tail)
1790				*tail = object;
1791		} else {
1792			/*
1793			 * Adjust the reconstructed freelist depth
1794			 * accordingly if object's reuse is delayed.
1795			 */
1796			--(*cnt);
1797		}
1798	} while (object != old_tail);
1799
1800	if (*head == *tail)
1801		*tail = NULL;
1802
1803	return *head != NULL;
1804}
1805
1806static void *setup_object(struct kmem_cache *s, void *object)
1807{
1808	setup_object_debug(s, object);
1809	object = kasan_init_slab_obj(s, object);
1810	if (unlikely(s->ctor)) {
1811		kasan_unpoison_object_data(s, object);
1812		s->ctor(object);
1813		kasan_poison_object_data(s, object);
1814	}
1815	return object;
1816}
1817
1818/*
1819 * Slab allocation and freeing
1820 */
1821static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1822		struct kmem_cache_order_objects oo)
1823{
1824	struct folio *folio;
1825	struct slab *slab;
1826	unsigned int order = oo_order(oo);
1827
1828	if (node == NUMA_NO_NODE)
1829		folio = (struct folio *)alloc_pages(flags, order);
1830	else
1831		folio = (struct folio *)__alloc_pages_node(node, flags, order);
1832
1833	if (!folio)
1834		return NULL;
1835
1836	slab = folio_slab(folio);
1837	__folio_set_slab(folio);
1838	if (page_is_pfmemalloc(folio_page(folio, 0)))
1839		slab_set_pfmemalloc(slab);
1840
1841	return slab;
1842}
1843
1844#ifdef CONFIG_SLAB_FREELIST_RANDOM
1845/* Pre-initialize the random sequence cache */
1846static int init_cache_random_seq(struct kmem_cache *s)
1847{
1848	unsigned int count = oo_objects(s->oo);
1849	int err;
1850
1851	/* Bailout if already initialised */
1852	if (s->random_seq)
1853		return 0;
1854
1855	err = cache_random_seq_create(s, count, GFP_KERNEL);
1856	if (err) {
1857		pr_err("SLUB: Unable to initialize free list for %s\n",
1858			s->name);
1859		return err;
1860	}
1861
1862	/* Transform to an offset on the set of pages */
1863	if (s->random_seq) {
1864		unsigned int i;
1865
1866		for (i = 0; i < count; i++)
1867			s->random_seq[i] *= s->size;
1868	}
1869	return 0;
1870}
1871
1872/* Initialize each random sequence freelist per cache */
1873static void __init init_freelist_randomization(void)
1874{
1875	struct kmem_cache *s;
1876
1877	mutex_lock(&slab_mutex);
1878
1879	list_for_each_entry(s, &slab_caches, list)
1880		init_cache_random_seq(s);
1881
1882	mutex_unlock(&slab_mutex);
1883}
1884
1885/* Get the next entry on the pre-computed freelist randomized */
1886static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1887				unsigned long *pos, void *start,
1888				unsigned long page_limit,
1889				unsigned long freelist_count)
1890{
1891	unsigned int idx;
1892
1893	/*
1894	 * If the target page allocation failed, the number of objects on the
1895	 * page might be smaller than the usual size defined by the cache.
1896	 */
1897	do {
1898		idx = s->random_seq[*pos];
1899		*pos += 1;
1900		if (*pos >= freelist_count)
1901			*pos = 0;
1902	} while (unlikely(idx >= page_limit));
1903
1904	return (char *)start + idx;
1905}
1906
1907/* Shuffle the single linked freelist based on a random pre-computed sequence */
1908static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1909{
1910	void *start;
1911	void *cur;
1912	void *next;
1913	unsigned long idx, pos, page_limit, freelist_count;
1914
1915	if (slab->objects < 2 || !s->random_seq)
1916		return false;
1917
1918	freelist_count = oo_objects(s->oo);
1919	pos = get_random_int() % freelist_count;
1920
1921	page_limit = slab->objects * s->size;
1922	start = fixup_red_left(s, slab_address(slab));
1923
1924	/* First entry is used as the base of the freelist */
1925	cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1926				freelist_count);
1927	cur = setup_object(s, cur);
1928	slab->freelist = cur;
1929
1930	for (idx = 1; idx < slab->objects; idx++) {
1931		next = next_freelist_entry(s, slab, &pos, start, page_limit,
1932			freelist_count);
1933		next = setup_object(s, next);
1934		set_freepointer(s, cur, next);
1935		cur = next;
1936	}
1937	set_freepointer(s, cur, NULL);
1938
1939	return true;
1940}
1941#else
1942static inline int init_cache_random_seq(struct kmem_cache *s)
1943{
1944	return 0;
1945}
1946static inline void init_freelist_randomization(void) { }
1947static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1948{
1949	return false;
1950}
1951#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1952
1953static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1954{
1955	struct slab *slab;
1956	struct kmem_cache_order_objects oo = s->oo;
1957	gfp_t alloc_gfp;
1958	void *start, *p, *next;
1959	int idx;
1960	bool shuffle;
1961
1962	flags &= gfp_allowed_mask;
1963
1964	flags |= s->allocflags;
1965
1966	/*
1967	 * Let the initial higher-order allocation fail under memory pressure
1968	 * so we fall-back to the minimum order allocation.
1969	 */
1970	alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1971	if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1972		alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
1973
1974	slab = alloc_slab_page(alloc_gfp, node, oo);
1975	if (unlikely(!slab)) {
1976		oo = s->min;
1977		alloc_gfp = flags;
1978		/*
1979		 * Allocation may have failed due to fragmentation.
1980		 * Try a lower order alloc if possible
1981		 */
1982		slab = alloc_slab_page(alloc_gfp, node, oo);
1983		if (unlikely(!slab))
1984			goto out;
1985		stat(s, ORDER_FALLBACK);
1986	}
1987
1988	slab->objects = oo_objects(oo);
1989
1990	account_slab(slab, oo_order(oo), s, flags);
1991
1992	slab->slab_cache = s;
1993
1994	kasan_poison_slab(slab);
1995
1996	start = slab_address(slab);
1997
1998	setup_slab_debug(s, slab, start);
1999
2000	shuffle = shuffle_freelist(s, slab);
2001
2002	if (!shuffle) {
2003		start = fixup_red_left(s, start);
2004		start = setup_object(s, start);
2005		slab->freelist = start;
2006		for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2007			next = p + s->size;
2008			next = setup_object(s, next);
2009			set_freepointer(s, p, next);
2010			p = next;
2011		}
2012		set_freepointer(s, p, NULL);
2013	}
2014
2015	slab->inuse = slab->objects;
2016	slab->frozen = 1;
2017
2018out:
2019	if (!slab)
2020		return NULL;
2021
2022	inc_slabs_node(s, slab_nid(slab), slab->objects);
2023
2024	return slab;
2025}
2026
2027static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2028{
2029	if (unlikely(flags & GFP_SLAB_BUG_MASK))
2030		flags = kmalloc_fix_flags(flags);
2031
2032	WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2033
2034	return allocate_slab(s,
2035		flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2036}
2037
2038static void __free_slab(struct kmem_cache *s, struct slab *slab)
2039{
2040	struct folio *folio = slab_folio(slab);
2041	int order = folio_order(folio);
2042	int pages = 1 << order;
2043
2044	if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2045		void *p;
2046
2047		slab_pad_check(s, slab);
2048		for_each_object(p, s, slab_address(slab), slab->objects)
2049			check_object(s, slab, p, SLUB_RED_INACTIVE);
2050	}
2051
2052	__slab_clear_pfmemalloc(slab);
2053	__folio_clear_slab(folio);
2054	folio->mapping = NULL;
2055	if (current->reclaim_state)
2056		current->reclaim_state->reclaimed_slab += pages;
2057	unaccount_slab(slab, order, s);
2058	__free_pages(folio_page(folio, 0), order);
2059}
2060
2061static void rcu_free_slab(struct rcu_head *h)
2062{
2063	struct slab *slab = container_of(h, struct slab, rcu_head);
2064
2065	__free_slab(slab->slab_cache, slab);
2066}
2067
2068static void free_slab(struct kmem_cache *s, struct slab *slab)
2069{
2070	if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
2071		call_rcu(&slab->rcu_head, rcu_free_slab);
2072	} else
2073		__free_slab(s, slab);
2074}
2075
2076static void discard_slab(struct kmem_cache *s, struct slab *slab)
2077{
2078	dec_slabs_node(s, slab_nid(slab), slab->objects);
2079	free_slab(s, slab);
2080}
2081
2082/*
2083 * Management of partially allocated slabs.
2084 */
2085static inline void
2086__add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2087{
2088	n->nr_partial++;
2089	if (tail == DEACTIVATE_TO_TAIL)
2090		list_add_tail(&slab->slab_list, &n->partial);
2091	else
2092		list_add(&slab->slab_list, &n->partial);
2093}
2094
2095static inline void add_partial(struct kmem_cache_node *n,
2096				struct slab *slab, int tail)
2097{
2098	lockdep_assert_held(&n->list_lock);
2099	__add_partial(n, slab, tail);
2100}
2101
2102static inline void remove_partial(struct kmem_cache_node *n,
2103					struct slab *slab)
2104{
2105	lockdep_assert_held(&n->list_lock);
2106	list_del(&slab->slab_list);
2107	n->nr_partial--;
2108}
2109
2110/*
2111 * Remove slab from the partial list, freeze it and
2112 * return the pointer to the freelist.
2113 *
2114 * Returns a list of objects or NULL if it fails.
2115 */
2116static inline void *acquire_slab(struct kmem_cache *s,
2117		struct kmem_cache_node *n, struct slab *slab,
2118		int mode)
2119{
2120	void *freelist;
2121	unsigned long counters;
2122	struct slab new;
2123
2124	lockdep_assert_held(&n->list_lock);
2125
2126	/*
2127	 * Zap the freelist and set the frozen bit.
2128	 * The old freelist is the list of objects for the
2129	 * per cpu allocation list.
2130	 */
2131	freelist = slab->freelist;
2132	counters = slab->counters;
2133	new.counters = counters;
2134	if (mode) {
2135		new.inuse = slab->objects;
2136		new.freelist = NULL;
2137	} else {
2138		new.freelist = freelist;
2139	}
2140
2141	VM_BUG_ON(new.frozen);
2142	new.frozen = 1;
2143
2144	if (!__cmpxchg_double_slab(s, slab,
2145			freelist, counters,
2146			new.freelist, new.counters,
2147			"acquire_slab"))
2148		return NULL;
2149
2150	remove_partial(n, slab);
2151	WARN_ON(!freelist);
2152	return freelist;
2153}
2154
2155#ifdef CONFIG_SLUB_CPU_PARTIAL
2156static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2157#else
2158static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2159				   int drain) { }
2160#endif
2161static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2162
2163/*
2164 * Try to allocate a partial slab from a specific node.
2165 */
2166static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2167			      struct slab **ret_slab, gfp_t gfpflags)
2168{
2169	struct slab *slab, *slab2;
2170	void *object = NULL;
2171	unsigned long flags;
2172	unsigned int partial_slabs = 0;
2173
2174	/*
2175	 * Racy check. If we mistakenly see no partial slabs then we
2176	 * just allocate an empty slab. If we mistakenly try to get a
2177	 * partial slab and there is none available then get_partial()
2178	 * will return NULL.
2179	 */
2180	if (!n || !n->nr_partial)
2181		return NULL;
2182
2183	spin_lock_irqsave(&n->list_lock, flags);
2184	list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2185		void *t;
2186
2187		if (!pfmemalloc_match(slab, gfpflags))
2188			continue;
2189
2190		t = acquire_slab(s, n, slab, object == NULL);
2191		if (!t)
2192			break;
2193
2194		if (!object) {
2195			*ret_slab = slab;
2196			stat(s, ALLOC_FROM_PARTIAL);
2197			object = t;
2198		} else {
2199			put_cpu_partial(s, slab, 0);
2200			stat(s, CPU_PARTIAL_NODE);
2201			partial_slabs++;
2202		}
2203#ifdef CONFIG_SLUB_CPU_PARTIAL
2204		if (!kmem_cache_has_cpu_partial(s)
2205			|| partial_slabs > s->cpu_partial_slabs / 2)
2206			break;
2207#else
2208		break;
2209#endif
2210
2211	}
2212	spin_unlock_irqrestore(&n->list_lock, flags);
2213	return object;
2214}
2215
2216/*
2217 * Get a slab from somewhere. Search in increasing NUMA distances.
2218 */
2219static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
2220			     struct slab **ret_slab)
2221{
2222#ifdef CONFIG_NUMA
2223	struct zonelist *zonelist;
2224	struct zoneref *z;
2225	struct zone *zone;
2226	enum zone_type highest_zoneidx = gfp_zone(flags);
2227	void *object;
2228	unsigned int cpuset_mems_cookie;
2229
2230	/*
2231	 * The defrag ratio allows a configuration of the tradeoffs between
2232	 * inter node defragmentation and node local allocations. A lower
2233	 * defrag_ratio increases the tendency to do local allocations
2234	 * instead of attempting to obtain partial slabs from other nodes.
2235	 *
2236	 * If the defrag_ratio is set to 0 then kmalloc() always
2237	 * returns node local objects. If the ratio is higher then kmalloc()
2238	 * may return off node objects because partial slabs are obtained
2239	 * from other nodes and filled up.
2240	 *
2241	 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2242	 * (which makes defrag_ratio = 1000) then every (well almost)
2243	 * allocation will first attempt to defrag slab caches on other nodes.
2244	 * This means scanning over all nodes to look for partial slabs which
2245	 * may be expensive if we do it every time we are trying to find a slab
2246	 * with available objects.
2247	 */
2248	if (!s->remote_node_defrag_ratio ||
2249			get_cycles() % 1024 > s->remote_node_defrag_ratio)
2250		return NULL;
2251
2252	do {
2253		cpuset_mems_cookie = read_mems_allowed_begin();
2254		zonelist = node_zonelist(mempolicy_slab_node(), flags);
2255		for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2256			struct kmem_cache_node *n;
2257
2258			n = get_node(s, zone_to_nid(zone));
2259
2260			if (n && cpuset_zone_allowed(zone, flags) &&
2261					n->nr_partial > s->min_partial) {
2262				object = get_partial_node(s, n, ret_slab, flags);
2263				if (object) {
2264					/*
2265					 * Don't check read_mems_allowed_retry()
2266					 * here - if mems_allowed was updated in
2267					 * parallel, that was a harmless race
2268					 * between allocation and the cpuset
2269					 * update
2270					 */
2271					return object;
2272				}
2273			}
2274		}
2275	} while (read_mems_allowed_retry(cpuset_mems_cookie));
2276#endif	/* CONFIG_NUMA */
2277	return NULL;
2278}
2279
2280/*
2281 * Get a partial slab, lock it and return it.
2282 */
2283static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
2284			 struct slab **ret_slab)
2285{
2286	void *object;
2287	int searchnode = node;
2288
2289	if (node == NUMA_NO_NODE)
2290		searchnode = numa_mem_id();
2291
2292	object = get_partial_node(s, get_node(s, searchnode), ret_slab, flags);
2293	if (object || node != NUMA_NO_NODE)
2294		return object;
2295
2296	return get_any_partial(s, flags, ret_slab);
2297}
2298
2299#ifdef CONFIG_PREEMPTION
2300/*
2301 * Calculate the next globally unique transaction for disambiguation
2302 * during cmpxchg. The transactions start with the cpu number and are then
2303 * incremented by CONFIG_NR_CPUS.
2304 */
2305#define TID_STEP  roundup_pow_of_two(CONFIG_NR_CPUS)
2306#else
2307/*
2308 * No preemption supported therefore also no need to check for
2309 * different cpus.
2310 */
2311#define TID_STEP 1
2312#endif
2313
2314static inline unsigned long next_tid(unsigned long tid)
2315{
2316	return tid + TID_STEP;
2317}
2318
2319#ifdef SLUB_DEBUG_CMPXCHG
2320static inline unsigned int tid_to_cpu(unsigned long tid)
2321{
2322	return tid % TID_STEP;
2323}
2324
2325static inline unsigned long tid_to_event(unsigned long tid)
2326{
2327	return tid / TID_STEP;
2328}
2329#endif
2330
2331static inline unsigned int init_tid(int cpu)
2332{
2333	return cpu;
2334}
2335
2336static inline void note_cmpxchg_failure(const char *n,
2337		const struct kmem_cache *s, unsigned long tid)
2338{
2339#ifdef SLUB_DEBUG_CMPXCHG
2340	unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2341
2342	pr_info("%s %s: cmpxchg redo ", n, s->name);
2343
2344#ifdef CONFIG_PREEMPTION
2345	if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2346		pr_warn("due to cpu change %d -> %d\n",
2347			tid_to_cpu(tid), tid_to_cpu(actual_tid));
2348	else
2349#endif
2350	if (tid_to_event(tid) != tid_to_event(actual_tid))
2351		pr_warn("due to cpu running other code. Event %ld->%ld\n",
2352			tid_to_event(tid), tid_to_event(actual_tid));
2353	else
2354		pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2355			actual_tid, tid, next_tid(tid));
2356#endif
2357	stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2358}
2359
2360static void init_kmem_cache_cpus(struct kmem_cache *s)
2361{
2362	int cpu;
2363	struct kmem_cache_cpu *c;
2364
2365	for_each_possible_cpu(cpu) {
2366		c = per_cpu_ptr(s->cpu_slab, cpu);
2367		local_lock_init(&c->lock);
2368		c->tid = init_tid(cpu);
2369	}
2370}
2371
2372/*
2373 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2374 * unfreezes the slabs and puts it on the proper list.
2375 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2376 * by the caller.
2377 */
2378static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2379			    void *freelist)
2380{
2381	enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE, M_FULL_NOLIST };
2382	struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2383	int free_delta = 0;
2384	enum slab_modes mode = M_NONE;
2385	void *nextfree, *freelist_iter, *freelist_tail;
2386	int tail = DEACTIVATE_TO_HEAD;
2387	unsigned long flags = 0;
2388	struct slab new;
2389	struct slab old;
2390
2391	if (slab->freelist) {
2392		stat(s, DEACTIVATE_REMOTE_FREES);
2393		tail = DEACTIVATE_TO_TAIL;
2394	}
2395
2396	/*
2397	 * Stage one: Count the objects on cpu's freelist as free_delta and
2398	 * remember the last object in freelist_tail for later splicing.
2399	 */
2400	freelist_tail = NULL;
2401	freelist_iter = freelist;
2402	while (freelist_iter) {
2403		nextfree = get_freepointer(s, freelist_iter);
2404
2405		/*
2406		 * If 'nextfree' is invalid, it is possible that the object at
2407		 * 'freelist_iter' is already corrupted.  So isolate all objects
2408		 * starting at 'freelist_iter' by skipping them.
2409		 */
2410		if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2411			break;
2412
2413		freelist_tail = freelist_iter;
2414		free_delta++;
2415
2416		freelist_iter = nextfree;
2417	}
2418
2419	/*
2420	 * Stage two: Unfreeze the slab while splicing the per-cpu
2421	 * freelist to the head of slab's freelist.
2422	 *
2423	 * Ensure that the slab is unfrozen while the list presence
2424	 * reflects the actual number of objects during unfreeze.
2425	 *
2426	 * We first perform cmpxchg holding lock and insert to list
2427	 * when it succeed. If there is mismatch then the slab is not
2428	 * unfrozen and number of objects in the slab may have changed.
2429	 * Then release lock and retry cmpxchg again.
2430	 */
2431redo:
2432
2433	old.freelist = READ_ONCE(slab->freelist);
2434	old.counters = READ_ONCE(slab->counters);
2435	VM_BUG_ON(!old.frozen);
2436
2437	/* Determine target state of the slab */
2438	new.counters = old.counters;
2439	if (freelist_tail) {
2440		new.inuse -= free_delta;
2441		set_freepointer(s, freelist_tail, old.freelist);
2442		new.freelist = freelist;
2443	} else
2444		new.freelist = old.freelist;
2445
2446	new.frozen = 0;
2447
2448	if (!new.inuse && n->nr_partial >= s->min_partial) {
2449		mode = M_FREE;
2450	} else if (new.freelist) {
2451		mode = M_PARTIAL;
2452		/*
2453		 * Taking the spinlock removes the possibility that
2454		 * acquire_slab() will see a slab that is frozen
2455		 */
2456		spin_lock_irqsave(&n->list_lock, flags);
2457	} else if (kmem_cache_debug_flags(s, SLAB_STORE_USER)) {
2458		mode = M_FULL;
2459		/*
2460		 * This also ensures that the scanning of full
2461		 * slabs from diagnostic functions will not see
2462		 * any frozen slabs.
2463		 */
2464		spin_lock_irqsave(&n->list_lock, flags);
2465	} else {
2466		mode = M_FULL_NOLIST;
2467	}
2468
2469
2470	if (!cmpxchg_double_slab(s, slab,
2471				old.freelist, old.counters,
2472				new.freelist, new.counters,
2473				"unfreezing slab")) {
2474		if (mode == M_PARTIAL || mode == M_FULL)
2475			spin_unlock_irqrestore(&n->list_lock, flags);
2476		goto redo;
2477	}
2478
2479
2480	if (mode == M_PARTIAL) {
2481		add_partial(n, slab, tail);
2482		spin_unlock_irqrestore(&n->list_lock, flags);
2483		stat(s, tail);
2484	} else if (mode == M_FREE) {
2485		stat(s, DEACTIVATE_EMPTY);
2486		discard_slab(s, slab);
2487		stat(s, FREE_SLAB);
2488	} else if (mode == M_FULL) {
2489		add_full(s, n, slab);
2490		spin_unlock_irqrestore(&n->list_lock, flags);
2491		stat(s, DEACTIVATE_FULL);
2492	} else if (mode == M_FULL_NOLIST) {
2493		stat(s, DEACTIVATE_FULL);
2494	}
2495}
2496
2497#ifdef CONFIG_SLUB_CPU_PARTIAL
2498static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2499{
2500	struct kmem_cache_node *n = NULL, *n2 = NULL;
2501	struct slab *slab, *slab_to_discard = NULL;
2502	unsigned long flags = 0;
2503
2504	while (partial_slab) {
2505		struct slab new;
2506		struct slab old;
2507
2508		slab = partial_slab;
2509		partial_slab = slab->next;
2510
2511		n2 = get_node(s, slab_nid(slab));
2512		if (n != n2) {
2513			if (n)
2514				spin_unlock_irqrestore(&n->list_lock, flags);
2515
2516			n = n2;
2517			spin_lock_irqsave(&n->list_lock, flags);
2518		}
2519
2520		do {
2521
2522			old.freelist = slab->freelist;
2523			old.counters = slab->counters;
2524			VM_BUG_ON(!old.frozen);
2525
2526			new.counters = old.counters;
2527			new.freelist = old.freelist;
2528
2529			new.frozen = 0;
2530
2531		} while (!__cmpxchg_double_slab(s, slab,
2532				old.freelist, old.counters,
2533				new.freelist, new.counters,
2534				"unfreezing slab"));
2535
2536		if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2537			slab->next = slab_to_discard;
2538			slab_to_discard = slab;
2539		} else {
2540			add_partial(n, slab, DEACTIVATE_TO_TAIL);
2541			stat(s, FREE_ADD_PARTIAL);
2542		}
2543	}
2544
2545	if (n)
2546		spin_unlock_irqrestore(&n->list_lock, flags);
2547
2548	while (slab_to_discard) {
2549		slab = slab_to_discard;
2550		slab_to_discard = slab_to_discard->next;
2551
2552		stat(s, DEACTIVATE_EMPTY);
2553		discard_slab(s, slab);
2554		stat(s, FREE_SLAB);
2555	}
2556}
2557
2558/*
2559 * Unfreeze all the cpu partial slabs.
2560 */
2561static void unfreeze_partials(struct kmem_cache *s)
2562{
2563	struct slab *partial_slab;
2564	unsigned long flags;
2565
2566	local_lock_irqsave(&s->cpu_slab->lock, flags);
2567	partial_slab = this_cpu_read(s->cpu_slab->partial);
2568	this_cpu_write(s->cpu_slab->partial, NULL);
2569	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2570
2571	if (partial_slab)
2572		__unfreeze_partials(s, partial_slab);
2573}
2574
2575static void unfreeze_partials_cpu(struct kmem_cache *s,
2576				  struct kmem_cache_cpu *c)
2577{
2578	struct slab *partial_slab;
2579
2580	partial_slab = slub_percpu_partial(c);
2581	c->partial = NULL;
2582
2583	if (partial_slab)
2584		__unfreeze_partials(s, partial_slab);
2585}
2586
2587/*
2588 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2589 * partial slab slot if available.
2590 *
2591 * If we did not find a slot then simply move all the partials to the
2592 * per node partial list.
2593 */
2594static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2595{
2596	struct slab *oldslab;
2597	struct slab *slab_to_unfreeze = NULL;
2598	unsigned long flags;
2599	int slabs = 0;
2600
2601	local_lock_irqsave(&s->cpu_slab->lock, flags);
2602
2603	oldslab = this_cpu_read(s->cpu_slab->partial);
2604
2605	if (oldslab) {
2606		if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2607			/*
2608			 * Partial array is full. Move the existing set to the
2609			 * per node partial list. Postpone the actual unfreezing
2610			 * outside of the critical section.
2611			 */
2612			slab_to_unfreeze = oldslab;
2613			oldslab = NULL;
2614		} else {
2615			slabs = oldslab->slabs;
2616		}
2617	}
2618
2619	slabs++;
2620
2621	slab->slabs = slabs;
2622	slab->next = oldslab;
2623
2624	this_cpu_write(s->cpu_slab->partial, slab);
2625
2626	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2627
2628	if (slab_to_unfreeze) {
2629		__unfreeze_partials(s, slab_to_unfreeze);
2630		stat(s, CPU_PARTIAL_DRAIN);
2631	}
2632}
2633
2634#else	/* CONFIG_SLUB_CPU_PARTIAL */
2635
2636static inline void unfreeze_partials(struct kmem_cache *s) { }
2637static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2638				  struct kmem_cache_cpu *c) { }
2639
2640#endif	/* CONFIG_SLUB_CPU_PARTIAL */
2641
2642static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2643{
2644	unsigned long flags;
2645	struct slab *slab;
2646	void *freelist;
2647
2648	local_lock_irqsave(&s->cpu_slab->lock, flags);
2649
2650	slab = c->slab;
2651	freelist = c->freelist;
2652
2653	c->slab = NULL;
2654	c->freelist = NULL;
2655	c->tid = next_tid(c->tid);
2656
2657	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2658
2659	if (slab) {
2660		deactivate_slab(s, slab, freelist);
2661		stat(s, CPUSLAB_FLUSH);
2662	}
2663}
2664
2665static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2666{
2667	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2668	void *freelist = c->freelist;
2669	struct slab *slab = c->slab;
2670
2671	c->slab = NULL;
2672	c->freelist = NULL;
2673	c->tid = next_tid(c->tid);
2674
2675	if (slab) {
2676		deactivate_slab(s, slab, freelist);
2677		stat(s, CPUSLAB_FLUSH);
2678	}
2679
2680	unfreeze_partials_cpu(s, c);
2681}
2682
2683struct slub_flush_work {
2684	struct work_struct work;
2685	struct kmem_cache *s;
2686	bool skip;
2687};
2688
2689/*
2690 * Flush cpu slab.
2691 *
2692 * Called from CPU work handler with migration disabled.
2693 */
2694static void flush_cpu_slab(struct work_struct *w)
2695{
2696	struct kmem_cache *s;
2697	struct kmem_cache_cpu *c;
2698	struct slub_flush_work *sfw;
2699
2700	sfw = container_of(w, struct slub_flush_work, work);
2701
2702	s = sfw->s;
2703	c = this_cpu_ptr(s->cpu_slab);
2704
2705	if (c->slab)
2706		flush_slab(s, c);
2707
2708	unfreeze_partials(s);
2709}
2710
2711static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2712{
2713	struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2714
2715	return c->slab || slub_percpu_partial(c);
2716}
2717
2718static DEFINE_MUTEX(flush_lock);
2719static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2720
2721static void flush_all_cpus_locked(struct kmem_cache *s)
2722{
2723	struct slub_flush_work *sfw;
2724	unsigned int cpu;
2725
2726	lockdep_assert_cpus_held();
2727	mutex_lock(&flush_lock);
2728
2729	for_each_online_cpu(cpu) {
2730		sfw = &per_cpu(slub_flush, cpu);
2731		if (!has_cpu_slab(cpu, s)) {
2732			sfw->skip = true;
2733			continue;
2734		}
2735		INIT_WORK(&sfw->work, flush_cpu_slab);
2736		sfw->skip = false;
2737		sfw->s = s;
2738		queue_work_on(cpu, flushwq, &sfw->work);
2739	}
2740
2741	for_each_online_cpu(cpu) {
2742		sfw = &per_cpu(slub_flush, cpu);
2743		if (sfw->skip)
2744			continue;
2745		flush_work(&sfw->work);
2746	}
2747
2748	mutex_unlock(&flush_lock);
2749}
2750
2751static void flush_all(struct kmem_cache *s)
2752{
2753	cpus_read_lock();
2754	flush_all_cpus_locked(s);
2755	cpus_read_unlock();
2756}
2757
2758/*
2759 * Use the cpu notifier to insure that the cpu slabs are flushed when
2760 * necessary.
2761 */
2762static int slub_cpu_dead(unsigned int cpu)
2763{
2764	struct kmem_cache *s;
2765
2766	mutex_lock(&slab_mutex);
2767	list_for_each_entry(s, &slab_caches, list)
2768		__flush_cpu_slab(s, cpu);
2769	mutex_unlock(&slab_mutex);
2770	return 0;
2771}
2772
2773/*
2774 * Check if the objects in a per cpu structure fit numa
2775 * locality expectations.
2776 */
2777static inline int node_match(struct slab *slab, int node)
2778{
2779#ifdef CONFIG_NUMA
2780	if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2781		return 0;
2782#endif
2783	return 1;
2784}
2785
2786#ifdef CONFIG_SLUB_DEBUG
2787static int count_free(struct slab *slab)
2788{
2789	return slab->objects - slab->inuse;
2790}
2791
2792static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2793{
2794	return atomic_long_read(&n->total_objects);
2795}
2796#endif /* CONFIG_SLUB_DEBUG */
2797
2798#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2799static unsigned long count_partial(struct kmem_cache_node *n,
2800					int (*get_count)(struct slab *))
2801{
2802	unsigned long flags;
2803	unsigned long x = 0;
2804	struct slab *slab;
2805
2806	spin_lock_irqsave(&n->list_lock, flags);
2807	list_for_each_entry(slab, &n->partial, slab_list)
2808		x += get_count(slab);
2809	spin_unlock_irqrestore(&n->list_lock, flags);
2810	return x;
2811}
2812#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2813
2814static noinline void
2815slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2816{
2817#ifdef CONFIG_SLUB_DEBUG
2818	static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2819				      DEFAULT_RATELIMIT_BURST);
2820	int node;
2821	struct kmem_cache_node *n;
2822
2823	if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2824		return;
2825
2826	pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2827		nid, gfpflags, &gfpflags);
2828	pr_warn("  cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2829		s->name, s->object_size, s->size, oo_order(s->oo),
2830		oo_order(s->min));
2831
2832	if (oo_order(s->min) > get_order(s->object_size))
2833		pr_warn("  %s debugging increased min order, use slub_debug=O to disable.\n",
2834			s->name);
2835
2836	for_each_kmem_cache_node(s, node, n) {
2837		unsigned long nr_slabs;
2838		unsigned long nr_objs;
2839		unsigned long nr_free;
2840
2841		nr_free  = count_partial(n, count_free);
2842		nr_slabs = node_nr_slabs(n);
2843		nr_objs  = node_nr_objs(n);
2844
2845		pr_warn("  node %d: slabs: %ld, objs: %ld, free: %ld\n",
2846			node, nr_slabs, nr_objs, nr_free);
2847	}
2848#endif
2849}
2850
2851static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
2852{
2853	if (unlikely(slab_test_pfmemalloc(slab)))
2854		return gfp_pfmemalloc_allowed(gfpflags);
2855
2856	return true;
2857}
2858
2859/*
2860 * Check the slab->freelist and either transfer the freelist to the
2861 * per cpu freelist or deactivate the slab.
2862 *
2863 * The slab is still frozen if the return value is not NULL.
2864 *
2865 * If this function returns NULL then the slab has been unfrozen.
2866 */
2867static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
2868{
2869	struct slab new;
2870	unsigned long counters;
2871	void *freelist;
2872
2873	lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
2874
2875	do {
2876		freelist = slab->freelist;
2877		counters = slab->counters;
2878
2879		new.counters = counters;
2880		VM_BUG_ON(!new.frozen);
2881
2882		new.inuse = slab->objects;
2883		new.frozen = freelist != NULL;
2884
2885	} while (!__cmpxchg_double_slab(s, slab,
2886		freelist, counters,
2887		NULL, new.counters,
2888		"get_freelist"));
2889
2890	return freelist;
2891}
2892
2893/*
2894 * Slow path. The lockless freelist is empty or we need to perform
2895 * debugging duties.
2896 *
2897 * Processing is still very fast if new objects have been freed to the
2898 * regular freelist. In that case we simply take over the regular freelist
2899 * as the lockless freelist and zap the regular freelist.
2900 *
2901 * If that is not working then we fall back to the partial lists. We take the
2902 * first element of the freelist as the object to allocate now and move the
2903 * rest of the freelist to the lockless freelist.
2904 *
2905 * And if we were unable to get a new slab from the partial slab lists then
2906 * we need to allocate a new slab. This is the slowest path since it involves
2907 * a call to the page allocator and the setup of a new slab.
2908 *
2909 * Version of __slab_alloc to use when we know that preemption is
2910 * already disabled (which is the case for bulk allocation).
2911 */
2912static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2913			  unsigned long addr, struct kmem_cache_cpu *c)
2914{
2915	void *freelist;
2916	struct slab *slab;
2917	unsigned long flags;
2918
2919	stat(s, ALLOC_SLOWPATH);
2920
2921reread_slab:
2922
2923	slab = READ_ONCE(c->slab);
2924	if (!slab) {
2925		/*
2926		 * if the node is not online or has no normal memory, just
2927		 * ignore the node constraint
2928		 */
2929		if (unlikely(node != NUMA_NO_NODE &&
2930			     !node_isset(node, slab_nodes)))
2931			node = NUMA_NO_NODE;
2932		goto new_slab;
2933	}
2934redo:
2935
2936	if (unlikely(!node_match(slab, node))) {
2937		/*
2938		 * same as above but node_match() being false already
2939		 * implies node != NUMA_NO_NODE
2940		 */
2941		if (!node_isset(node, slab_nodes)) {
2942			node = NUMA_NO_NODE;
2943		} else {
2944			stat(s, ALLOC_NODE_MISMATCH);
2945			goto deactivate_slab;
2946		}
2947	}
2948
2949	/*
2950	 * By rights, we should be searching for a slab page that was
2951	 * PFMEMALLOC but right now, we are losing the pfmemalloc
2952	 * information when the page leaves the per-cpu allocator
2953	 */
2954	if (unlikely(!pfmemalloc_match(slab, gfpflags)))
2955		goto deactivate_slab;
2956
2957	/* must check again c->slab in case we got preempted and it changed */
2958	local_lock_irqsave(&s->cpu_slab->lock, flags);
2959	if (unlikely(slab != c->slab)) {
2960		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2961		goto reread_slab;
2962	}
2963	freelist = c->freelist;
2964	if (freelist)
2965		goto load_freelist;
2966
2967	freelist = get_freelist(s, slab);
2968
2969	if (!freelist) {
2970		c->slab = NULL;
2971		c->tid = next_tid(c->tid);
2972		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2973		stat(s, DEACTIVATE_BYPASS);
2974		goto new_slab;
2975	}
2976
2977	stat(s, ALLOC_REFILL);
2978
2979load_freelist:
2980
2981	lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
2982
2983	/*
2984	 * freelist is pointing to the list of objects to be used.
2985	 * slab is pointing to the slab from which the objects are obtained.
2986	 * That slab must be frozen for per cpu allocations to work.
2987	 */
2988	VM_BUG_ON(!c->slab->frozen);
2989	c->freelist = get_freepointer(s, freelist);
2990	c->tid = next_tid(c->tid);
2991	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2992	return freelist;
2993
2994deactivate_slab:
2995
2996	local_lock_irqsave(&s->cpu_slab->lock, flags);
2997	if (slab != c->slab) {
2998		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2999		goto reread_slab;
3000	}
3001	freelist = c->freelist;
3002	c->slab = NULL;
3003	c->freelist = NULL;
3004	c->tid = next_tid(c->tid);
3005	local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3006	deactivate_slab(s, slab, freelist);
3007
3008new_slab:
3009
3010	if (slub_percpu_partial(c)) {
3011		local_lock_irqsave(&s->cpu_slab->lock, flags);
3012		if (unlikely(c->slab)) {
3013			local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3014			goto reread_slab;
3015		}
3016		if (unlikely(!slub_percpu_partial(c))) {
3017			local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3018			/* we were preempted and partial list got empty */
3019			goto new_objects;
3020		}
3021
3022		slab = c->slab = slub_percpu_partial(c);
3023		slub_set_percpu_partial(c, slab);
3024		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3025		stat(s, CPU_PARTIAL_ALLOC);
3026		goto redo;
3027	}
3028
3029new_objects:
3030
3031	freelist = get_partial(s, gfpflags, node, &slab);
3032	if (freelist)
3033		goto check_new_slab;
3034
3035	slub_put_cpu_ptr(s->cpu_slab);
3036	slab = new_slab(s, gfpflags, node);
3037	c = slub_get_cpu_ptr(s->cpu_slab);
3038
3039	if (unlikely(!slab)) {
3040		slab_out_of_memory(s, gfpflags, node);
3041		return NULL;
3042	}
3043
3044	/*
3045	 * No other reference to the slab yet so we can
3046	 * muck around with it freely without cmpxchg
3047	 */
3048	freelist = slab->freelist;
3049	slab->freelist = NULL;
3050
3051	stat(s, ALLOC_SLAB);
3052
3053check_new_slab:
3054
3055	if (kmem_cache_debug(s)) {
3056		if (!alloc_debug_processing(s, slab, freelist, addr)) {
3057			/* Slab failed checks. Next slab needed */
3058			goto new_slab;
3059		} else {
3060			/*
3061			 * For debug case, we don't load freelist so that all
3062			 * allocations go through alloc_debug_processing()
3063			 */
3064			goto return_single;
3065		}
3066	}
3067
3068	if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3069		/*
3070		 * For !pfmemalloc_match() case we don't load freelist so that
3071		 * we don't make further mismatched allocations easier.
3072		 */
3073		goto return_single;
3074
3075retry_load_slab:
3076
3077	local_lock_irqsave(&s->cpu_slab->lock, flags);
3078	if (unlikely(c->slab)) {
3079		void *flush_freelist = c->freelist;
3080		struct slab *flush_slab = c->slab;
3081
3082		c->slab = NULL;
3083		c->freelist = NULL;
3084		c->tid = next_tid(c->tid);
3085
3086		local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3087
3088		deactivate_slab(s, flush_slab, flush_freelist);
3089
3090		stat(s, CPUSLAB_FLUSH);
3091
3092		goto retry_load_slab;
3093	}
3094	c->slab = slab;
3095
3096	goto load_freelist;
3097
3098return_single:
3099
3100	deactivate_slab(s, slab, get_freepointer(s, freelist));
3101	return freelist;
3102}
3103
3104/*
3105 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3106 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3107 * pointer.
3108 */
3109static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3110			  unsigned long addr, struct kmem_cache_cpu *c)
3111{
3112	void *p;
3113
3114#ifdef CONFIG_PREEMPT_COUNT
3115	/*
3116	 * We may have been preempted and rescheduled on a different
3117	 * cpu before disabling preemption. Need to reload cpu area
3118	 * pointer.
3119	 */
3120	c = slub_get_cpu_ptr(s->cpu_slab);
3121#endif
3122
3123	p = ___slab_alloc(s, gfpflags, node, addr, c);
3124#ifdef CONFIG_PREEMPT_COUNT
3125	slub_put_cpu_ptr(s->cpu_slab);
3126#endif
3127	return p;
3128}
3129
3130/*
3131 * If the object has been wiped upon free, make sure it's fully initialized by
3132 * zeroing out freelist pointer.
3133 */
3134static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3135						   void *obj)
3136{
3137	if (unlikely(slab_want_init_on_free(s)) && obj)
3138		memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3139			0, sizeof(void *));
3140}
3141
3142/*
3143 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3144 * have the fastpath folded into their functions. So no function call
3145 * overhead for requests that can be satisfied on the fastpath.
3146 *
3147 * The fastpath works by first checking if the lockless freelist can be used.
3148 * If not then __slab_alloc is called for slow processing.
3149 *
3150 * Otherwise we can simply pick the next object from the lockless free list.
3151 */
3152static __always_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3153		gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3154{
3155	void *object;
3156	struct kmem_cache_cpu *c;
3157	struct slab *slab;
3158	unsigned long tid;
3159	struct obj_cgroup *objcg = NULL;
3160	bool init = false;
3161
3162	s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3163	if (!s)
3164		return NULL;
3165
3166	object = kfence_alloc(s, orig_size, gfpflags);
3167	if (unlikely(object))
3168		goto out;
3169
3170redo:
3171	/*
3172	 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3173	 * enabled. We may switch back and forth between cpus while
3174	 * reading from one cpu area. That does not matter as long
3175	 * as we end up on the original cpu again when doing the cmpxchg.
3176	 *
3177	 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3178	 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3179	 * the tid. If we are preempted and switched to another cpu between the
3180	 * two reads, it's OK as the two are still associated with the same cpu
3181	 * and cmpxchg later will validate the cpu.
3182	 */
3183	c = raw_cpu_ptr(s->cpu_slab);
3184	tid = READ_ONCE(c->tid);
3185
3186	/*
3187	 * Irqless object alloc/free algorithm used here depends on sequence
3188	 * of fetching cpu_slab's data. tid should be fetched before anything
3189	 * on c to guarantee that object and slab associated with previous tid
3190	 * won't be used with current tid. If we fetch tid first, object and
3191	 * slab could be one associated with next tid and our alloc/free
3192	 * request will be failed. In this case, we will retry. So, no problem.
3193	 */
3194	barrier();
3195
3196	/*
3197	 * The transaction ids are globally unique per cpu and per operation on
3198	 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3199	 * occurs on the right processor and that there was no operation on the
3200	 * linked list in between.
3201	 */
3202
3203	object = c->freelist;
3204	slab = c->slab;
3205	/*
3206	 * We cannot use the lockless fastpath on PREEMPT_RT because if a
3207	 * slowpath has taken the local_lock_irqsave(), it is not protected
3208	 * against a fast path operation in an irq handler. So we need to take
3209	 * the slow path which uses local_lock. It is still relatively fast if
3210	 * there is a suitable cpu freelist.
3211	 */
3212	if (IS_ENABLED(CONFIG_PREEMPT_RT) ||
3213	    unlikely(!object || !slab || !node_match(slab, node))) {
3214		object = __slab_alloc(s, gfpflags, node, addr, c);
3215	} else {
3216		void *next_object = get_freepointer_safe(s, object);
3217
3218		/*
3219		 * The cmpxchg will only match if there was no additional
3220		 * operation and if we are on the right processor.
3221		 *
3222		 * The cmpxchg does the following atomically (without lock
3223		 * semantics!)
3224		 * 1. Relocate first pointer to the current per cpu area.
3225		 * 2. Verify that tid and freelist have not been changed
3226		 * 3. If they were not changed replace tid and freelist
3227		 *
3228		 * Since this is without lock semantics the protection is only
3229		 * against code executing on this cpu *not* from access by
3230		 * other cpus.
3231		 */
3232		if (unlikely(!this_cpu_cmpxchg_double(
3233				s->cpu_slab->freelist, s->cpu_slab->tid,
3234				object, tid,
3235				next_object, next_tid(tid)))) {
3236
3237			note_cmpxchg_failure("slab_alloc", s, tid);
3238			goto redo;
3239		}
3240		prefetch_freepointer(s, next_object);
3241		stat(s, ALLOC_FASTPATH);
3242	}
3243
3244	maybe_wipe_obj_freeptr(s, object);
3245	init = slab_want_init_on_alloc(gfpflags, s);
3246
3247out:
3248	slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init);
3249
3250	return object;
3251}
3252
3253static __always_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3254		gfp_t gfpflags, unsigned long addr, size_t orig_size)
3255{
3256	return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3257}
3258
3259static __always_inline
3260void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3261			     gfp_t gfpflags)
3262{
3263	void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3264
3265	trace_kmem_cache_alloc(_RET_IP_, ret, s, s->object_size,
3266				s->size, gfpflags);
3267
3268	return ret;
3269}
3270
3271void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3272{
3273	return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3274}
3275EXPORT_SYMBOL(kmem_cache_alloc);
3276
3277void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3278			   gfp_t gfpflags)
3279{
3280	return __kmem_cache_alloc_lru(s, lru, gfpflags);
3281}
3282EXPORT_SYMBOL(kmem_cache_alloc_lru);
3283
3284#ifdef CONFIG_TRACING
3285void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3286{
3287	void *ret = slab_alloc(s, NULL, gfpflags, _RET_IP_, size);
3288	trace_kmalloc(_RET_IP_, ret, s, size, s->size, gfpflags);
3289	ret = kasan_kmalloc(s, ret, size, gfpflags);
3290	return ret;
3291}
3292EXPORT_SYMBOL(kmem_cache_alloc_trace);
3293#endif
3294
3295#ifdef CONFIG_NUMA
3296void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3297{
3298	void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3299
3300	trace_kmem_cache_alloc_node(_RET_IP_, ret, s,
3301				    s->object_size, s->size, gfpflags, node);
3302
3303	return ret;
3304}
3305EXPORT_SYMBOL(kmem_cache_alloc_node);
3306
3307#ifdef CONFIG_TRACING
3308void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
3309				    gfp_t gfpflags,
3310				    int node, size_t size)
3311{
3312	void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
3313
3314	trace_kmalloc_node(_RET_IP_, ret, s,
3315			   size, s->size, gfpflags, node);
3316
3317	ret = kasan_kmalloc(s, ret, size, gfpflags);
3318	return ret;
3319}
3320EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
3321#endif
3322#endif	/* CONFIG_NUMA */
3323
3324/*
3325 * Slow path handling. This may still be called frequently since objects
3326 * have a longer lifetime than the cpu slabs in most processing loads.
3327 *
3328 * So we still attempt to reduce cache line usage. Just take the slab
3329 * lock and free the item. If there is no additional partial slab
3330 * handling required then we can return immediately.
3331 */
3332static void __slab_free(struct kmem_cache *s, struct slab *slab,
3333			void *head, void *tail, int cnt,
3334			unsigned long addr)
3335
3336{
3337	void *prior;
3338	int was_frozen;
3339	struct slab new;
3340	unsigned long counters;
3341	struct kmem_cache_node *n = NULL;
3342	unsigned long flags;
3343
3344	stat(s, FREE_SLOWPATH);
3345
3346	if (kfence_free(head))
3347		return;
3348
3349	if (kmem_cache_debug(s) &&
3350	    !free_debug_processing(s, slab, head, tail, cnt, addr))
3351		return;
3352
3353	do {
3354		if (unlikely(n)) {
3355			spin_unlock_irqrestore(&n->list_lock, flags);
3356			n = NULL;
3357		}
3358		prior = slab->freelist;
3359		counters = slab->counters;
3360		set_freepointer(s, tail, prior);
3361		new.counters = counters;
3362		was_frozen = new.frozen;
3363		new.inuse -= cnt;
3364		if ((!new.inuse || !prior) && !was_frozen) {
3365
3366			if (kmem_cache_has_cpu_partial(s) && !prior) {
3367
3368				/*
3369				 * Slab was on no list before and will be
3370				 * partially empty
3371				 * We can defer the list move and instead
3372				 * freeze it.
3373				 */
3374				new.frozen = 1;
3375
3376			} else { /* Needs to be taken off a list */
3377
3378				n = get_node(s, slab_nid(slab));
3379				/*
3380				 * Speculatively acquire the list_lock.
3381				 * If the cmpxchg does not succeed then we may
3382				 * drop the list_lock without any processing.
3383				 *
3384				 * Otherwise the list_lock will synchronize with
3385				 * other processors updating the list of slabs.
3386				 */
3387				spin_lock_irqsave(&n->list_lock, flags);
3388
3389			}
3390		}
3391
3392	} while (!cmpxchg_double_slab(s, slab,
3393		prior, counters,
3394		head, new.counters,
3395		"__slab_free"));
3396
3397	if (likely(!n)) {
3398
3399		if (likely(was_frozen)) {
3400			/*
3401			 * The list lock was not taken therefore no list
3402			 * activity can be necessary.
3403			 */
3404			stat(s, FREE_FROZEN);
3405		} else if (new.frozen) {
3406			/*
3407			 * If we just froze the slab then put it onto the
3408			 * per cpu partial list.
3409			 */
3410			put_cpu_partial(s, slab, 1);
3411			stat(s, CPU_PARTIAL_FREE);
3412		}
3413
3414		return;
3415	}
3416
3417	if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3418		goto slab_empty;
3419
3420	/*
3421	 * Objects left in the slab. If it was not on the partial list before
3422	 * then add it.
3423	 */
3424	if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3425		remove_full(s, n, slab);
3426		add_partial(n, slab, DEACTIVATE_TO_TAIL);
3427		stat(s, FREE_ADD_PARTIAL);
3428	}
3429	spin_unlock_irqrestore(&n->list_lock, flags);
3430	return;
3431
3432slab_empty:
3433	if (prior) {
3434		/*
3435		 * Slab on the partial list.
3436		 */
3437		remove_partial(n, slab);
3438		stat(s, FREE_REMOVE_PARTIAL);
3439	} else {
3440		/* Slab must be on the full list */
3441		remove_full(s, n, slab);
3442	}
3443
3444	spin_unlock_irqrestore(&n->list_lock, flags);
3445	stat(s, FREE_SLAB);
3446	discard_slab(s, slab);
3447}
3448
3449/*
3450 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3451 * can perform fastpath freeing without additional function calls.
3452 *
3453 * The fastpath is only possible if we are freeing to the current cpu slab
3454 * of this processor. This typically the case if we have just allocated
3455 * the item before.
3456 *
3457 * If fastpath is not possible then fall back to __slab_free where we deal
3458 * with all sorts of special processing.
3459 *
3460 * Bulk free of a freelist with several objects (all pointing to the
3461 * same slab) possible by specifying head and tail ptr, plus objects
3462 * count (cnt). Bulk free indicated by tail pointer being set.
3463 */
3464static __always_inline void do_slab_free(struct kmem_cache *s,
3465				struct slab *slab, void *head, void *tail,
3466				int cnt, unsigned long addr)
3467{
3468	void *tail_obj = tail ? : head;
3469	struct kmem_cache_cpu *c;
3470	unsigned long tid;
3471
3472redo:
3473	/*
3474	 * Determine the currently cpus per cpu slab.
3475	 * The cpu may change afterward. However that does not matter since
3476	 * data is retrieved via this pointer. If we are on the same cpu
3477	 * during the cmpxchg then the free will succeed.
3478	 */
3479	c = raw_cpu_ptr(s->cpu_slab);
3480	tid = READ_ONCE(c->tid);
3481
3482	/* Same with comment on barrier() in slab_alloc_node() */
3483	barrier();
3484
3485	if (likely(slab == c->slab)) {
3486#ifndef CONFIG_PREEMPT_RT
3487		void **freelist = READ_ONCE(c->freelist);
3488
3489		set_freepointer(s, tail_obj, freelist);
3490
3491		if (unlikely(!this_cpu_cmpxchg_double(
3492				s->cpu_slab->freelist, s->cpu_slab->tid,
3493				freelist, tid,
3494				head, next_tid(tid)))) {
3495
3496			note_cmpxchg_failure("slab_free", s, tid);
3497			goto redo;
3498		}
3499#else /* CONFIG_PREEMPT_RT */
3500		/*
3501		 * We cannot use the lockless fastpath on PREEMPT_RT because if
3502		 * a slowpath has taken the local_lock_irqsave(), it is not
3503		 * protected against a fast path operation in an irq handler. So
3504		 * we need to take the local_lock. We shouldn't simply defer to
3505		 * __slab_free() as that wouldn't use the cpu freelist at all.
3506		 */
3507		void **freelist;
3508
3509		local_lock(&s->cpu_slab->lock);
3510		c = this_cpu_ptr(s->cpu_slab);
3511		if (unlikely(slab != c->slab)) {
3512			local_unlock(&s->cpu_slab->lock);
3513			goto redo;
3514		}
3515		tid = c->tid;
3516		freelist = c->freelist;
3517
3518		set_freepointer(s, tail_obj, freelist);
3519		c->freelist = head;
3520		c->tid = next_tid(tid);
3521
3522		local_unlock(&s->cpu_slab->lock);
3523#endif
3524		stat(s, FREE_FASTPATH);
3525	} else
3526		__slab_free(s, slab, head, tail_obj, cnt, addr);
3527
3528}
3529
3530static __always_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3531				      void *head, void *tail, void **p, int cnt,
3532				      unsigned long addr)
3533{
3534	memcg_slab_free_hook(s, slab, p, cnt);
3535	/*
3536	 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3537	 * to remove objects, whose reuse must be delayed.
3538	 */
3539	if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3540		do_slab_free(s, slab, head, tail, cnt, addr);
3541}
3542
3543#ifdef CONFIG_KASAN_GENERIC
3544void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3545{
3546	do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3547}
3548#endif
3549
3550void kmem_cache_free(struct kmem_cache *s, void *x)
3551{
3552	s = cache_from_obj(s, x);
3553	if (!s)
3554		return;
3555	trace_kmem_cache_free(_RET_IP_, x, s->name);
3556	slab_free(s, virt_to_slab(x), x, NULL, &x, 1, _RET_IP_);
3557}
3558EXPORT_SYMBOL(kmem_cache_free);
3559
3560struct detached_freelist {
3561	struct slab *slab;
3562	void *tail;
3563	void *freelist;
3564	int cnt;
3565	struct kmem_cache *s;
3566};
3567
3568static inline void free_large_kmalloc(struct folio *folio, void *object)
3569{
3570	unsigned int order = folio_order(folio);
3571
3572	if (WARN_ON_ONCE(order == 0))
3573		pr_warn_once("object pointer: 0x%p\n", object);
3574
3575	kfree_hook(object);
3576	mod_lruvec_page_state(folio_page(folio, 0), NR_SLAB_UNRECLAIMABLE_B,
3577			      -(PAGE_SIZE << order));
3578	__free_pages(folio_page(folio, 0), order);
3579}
3580
3581/*
3582 * This function progressively scans the array with free objects (with
3583 * a limited look ahead) and extract objects belonging to the same
3584 * slab.  It builds a detached freelist directly within the given
3585 * slab/objects.  This can happen without any need for
3586 * synchronization, because the objects are owned by running process.
3587 * The freelist is build up as a single linked list in the objects.
3588 * The idea is, that this detached freelist can then be bulk
3589 * transferred to the real freelist(s), but only requiring a single
3590 * synchronization primitive.  Look ahead in the array is limited due
3591 * to performance reasons.
3592 */
3593static inline
3594int build_detached_freelist(struct kmem_cache *s, size_t size,
3595			    void **p, struct detached_freelist *df)
3596{
3597	int lookahead = 3;
3598	void *object;
3599	struct folio *folio;
3600	size_t same;
3601
3602	object = p[--size];
3603	folio = virt_to_folio(object);
3604	if (!s) {
3605		/* Handle kalloc'ed objects */
3606		if (unlikely(!folio_test_slab(folio))) {
3607			free_large_kmalloc(folio, object);
3608			df->slab = NULL;
3609			return size;
3610		}
3611		/* Derive kmem_cache from object */
3612		df->slab = folio_slab(folio);
3613		df->s = df->slab->slab_cache;
3614	} else {
3615		df->slab = folio_slab(folio);
3616		df->s = cache_from_obj(s, object); /* Support for memcg */
3617	}
3618
3619	/* Start new detached freelist */
3620	df->tail = object;
3621	df->freelist = object;
3622	df->cnt = 1;
3623
3624	if (is_kfence_address(object))
3625		return size;
3626
3627	set_freepointer(df->s, object, NULL);
3628
3629	same = size;
3630	while (size) {
3631		object = p[--size];
3632		/* df->slab is always set at this point */
3633		if (df->slab == virt_to_slab(object)) {
3634			/* Opportunity build freelist */
3635			set_freepointer(df->s, object, df->freelist);
3636			df->freelist = object;
3637			df->cnt++;
3638			same--;
3639			if (size != same)
3640				swap(p[size], p[same]);
3641			continue;
3642		}
3643
3644		/* Limit look ahead search */
3645		if (!--lookahead)
3646			break;
3647	}
3648
3649	return same;
3650}
3651
3652/* Note that interrupts must be enabled when calling this function. */
3653void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3654{
3655	if (!size)
3656		return;
3657
3658	do {
3659		struct detached_freelist df;
3660
3661		size = build_detached_freelist(s, size, p, &df);
3662		if (!df.slab)
3663			continue;
3664
3665		slab_free(df.s, df.slab, df.freelist, df.tail, &p[size], df.cnt,
3666			  _RET_IP_);
3667	} while (likely(size));
3668}
3669EXPORT_SYMBOL(kmem_cache_free_bulk);
3670
3671/* Note that interrupts must be enabled when calling this function. */
3672int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3673			  void **p)
3674{
3675	struct kmem_cache_cpu *c;
3676	int i;
3677	struct obj_cgroup *objcg = NULL;
3678
3679	/* memcg and kmem_cache debug support */
3680	s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
3681	if (unlikely(!s))
3682		return false;
3683	/*
3684	 * Drain objects in the per cpu slab, while disabling local
3685	 * IRQs, which protects against PREEMPT and interrupts
3686	 * handlers invoking normal fastpath.
3687	 */
3688	c = slub_get_cpu_ptr(s->cpu_slab);
3689	local_lock_irq(&s->cpu_slab->lock);
3690
3691	for (i = 0; i < size; i++) {
3692		void *object = kfence_alloc(s, s->object_size, flags);
3693
3694		if (unlikely(object)) {
3695			p[i] = object;
3696			continue;
3697		}
3698
3699		object = c->freelist;
3700		if (unlikely(!object)) {
3701			/*
3702			 * We may have removed an object from c->freelist using
3703			 * the fastpath in the previous iteration; in that case,
3704			 * c->tid has not been bumped yet.
3705			 * Since ___slab_alloc() may reenable interrupts while
3706			 * allocating memory, we should bump c->tid now.
3707			 */
3708			c->tid = next_tid(c->tid);
3709
3710			local_unlock_irq(&s->cpu_slab->lock);
3711
3712			/*
3713			 * Invoking slow path likely have side-effect
3714			 * of re-populating per CPU c->freelist
3715			 */
3716			p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3717					    _RET_IP_, c);
3718			if (unlikely(!p[i]))
3719				goto error;
3720
3721			c = this_cpu_ptr(s->cpu_slab);
3722			maybe_wipe_obj_freeptr(s, p[i]);
3723
3724			local_lock_irq(&s->cpu_slab->lock);
3725
3726			continue; /* goto for-loop */
3727		}
3728		c->freelist = get_freepointer(s, object);
3729		p[i] = object;
3730		maybe_wipe_obj_freeptr(s, p[i]);
3731	}
3732	c->tid = next_tid(c->tid);
3733	local_unlock_irq(&s->cpu_slab->lock);
3734	slub_put_cpu_ptr(s->cpu_slab);
3735
3736	/*
3737	 * memcg and kmem_cache debug support and memory initialization.
3738	 * Done outside of the IRQ disabled fastpath loop.
3739	 */
3740	slab_post_alloc_hook(s, objcg, flags, size, p,
3741				slab_want_init_on_alloc(flags, s));
3742	return i;
3743error:
3744	slub_put_cpu_ptr(s->cpu_slab);
3745	slab_post_alloc_hook(s, objcg, flags, i, p, false);
3746	kmem_cache_free_bulk(s, i, p);
3747	return 0;
3748}
3749EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3750
3751
3752/*
3753 * Object placement in a slab is made very easy because we always start at
3754 * offset 0. If we tune the size of the object to the alignment then we can
3755 * get the required alignment by putting one properly sized object after
3756 * another.
3757 *
3758 * Notice that the allocation order determines the sizes of the per cpu
3759 * caches. Each processor has always one slab available for allocations.
3760 * Increasing the allocation order reduces the number of times that slabs
3761 * must be moved on and off the partial lists and is therefore a factor in
3762 * locking overhead.
3763 */
3764
3765/*
3766 * Minimum / Maximum order of slab pages. This influences locking overhead
3767 * and slab fragmentation. A higher order reduces the number of partial slabs
3768 * and increases the number of allocations possible without having to
3769 * take the list_lock.
3770 */
3771static unsigned int slub_min_order;
3772static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3773static unsigned int slub_min_objects;
3774
3775/*
3776 * Calculate the order of allocation given an slab object size.
3777 *
3778 * The order of allocation has significant impact on performance and other
3779 * system components. Generally order 0 allocations should be preferred since
3780 * order 0 does not cause fragmentation in the page allocator. Larger objects
3781 * be problematic to put into order 0 slabs because there may be too much
3782 * unused space left. We go to a higher order if more than 1/16th of the slab
3783 * would be wasted.
3784 *
3785 * In order to reach satisfactory performance we must ensure that a minimum
3786 * number of objects is in one slab. Otherwise we may generate too much
3787 * activity on the partial lists which requires taking the list_lock. This is
3788 * less a concern for large slabs though which are rarely used.
3789 *
3790 * slub_max_order specifies the order where we begin to stop considering the
3791 * number of objects in a slab as critical. If we reach slub_max_order then
3792 * we try to keep the page order as low as possible. So we accept more waste
3793 * of space in favor of a small page order.
3794 *
3795 * Higher order allocations also allow the placement of more objects in a
3796 * slab and thereby reduce object handling overhead. If the user has
3797 * requested a higher minimum order then we start with that one instead of
3798 * the smallest order which will fit the object.
3799 */
3800static inline unsigned int calc_slab_order(unsigned int size,
3801		unsigned int min_objects, unsigned int max_order,
3802		unsigned int fract_leftover)
3803{
3804	unsigned int min_order = slub_min_order;
3805	unsigned int order;
3806
3807	if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3808		return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3809
3810	for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3811			order <= max_order; order++) {
3812
3813		unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3814		unsigned int rem;
3815
3816		rem = slab_size % size;
3817
3818		if (rem <= slab_size / fract_leftover)
3819			break;
3820	}
3821
3822	return order;
3823}
3824
3825static inline int calculate_order(unsigned int size)
3826{
3827	unsigned int order;
3828	unsigned int min_objects;
3829	unsigned int max_objects;
3830	unsigned int nr_cpus;
3831
3832	/*
3833	 * Attempt to find best configuration for a slab. This
3834	 * works by first attempting to generate a layout with
3835	 * the best configuration and backing off gradually.
3836	 *
3837	 * First we increase the acceptable waste in a slab. Then
3838	 * we reduce the minimum objects required in a slab.
3839	 */
3840	min_objects = slub_min_objects;
3841	if (!min_objects) {
3842		/*
3843		 * Some architectures will only update present cpus when
3844		 * onlining them, so don't trust the number if it's just 1. But
3845		 * we also don't want to use nr_cpu_ids always, as on some other
3846		 * architectures, there can be many possible cpus, but never
3847		 * onlined. Here we compromise between trying to avoid too high
3848		 * order on systems that appear larger than they are, and too
3849		 * low order on systems that appear smaller than they are.
3850		 */
3851		nr_cpus = num_present_cpus();
3852		if (nr_cpus <= 1)
3853			nr_cpus = nr_cpu_ids;
3854		min_objects = 4 * (fls(nr_cpus) + 1);
3855	}
3856	max_objects = order_objects(slub_max_order, size);
3857	min_objects = min(min_objects, max_objects);
3858
3859	while (min_objects > 1) {
3860		unsigned int fraction;
3861
3862		fraction = 16;
3863		while (fraction >= 4) {
3864			order = calc_slab_order(size, min_objects,
3865					slub_max_order, fraction);
3866			if (order <= slub_max_order)
3867				return order;
3868			fraction /= 2;
3869		}
3870		min_objects--;
3871	}
3872
3873	/*
3874	 * We were unable to place multiple objects in a slab. Now
3875	 * lets see if we can place a single object there.
3876	 */
3877	order = calc_slab_order(size, 1, slub_max_order, 1);
3878	if (order <= slub_max_order)
3879		return order;
3880
3881	/*
3882	 * Doh this slab cannot be placed using slub_max_order.
3883	 */
3884	order = calc_slab_order(size, 1, MAX_ORDER, 1);
3885	if (order < MAX_ORDER)
3886		return order;
3887	return -ENOSYS;
3888}
3889
3890static void
3891init_kmem_cache_node(struct kmem_cache_node *n)
3892{
3893	n->nr_partial = 0;
3894	spin_lock_init(&n->list_lock);
3895	INIT_LIST_HEAD(&n->partial);
3896#ifdef CONFIG_SLUB_DEBUG
3897	atomic_long_set(&n->nr_slabs, 0);
3898	atomic_long_set(&n->total_objects, 0);
3899	INIT_LIST_HEAD(&n->full);
3900#endif
3901}
3902
3903static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3904{
3905	BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3906			KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3907
3908	/*
3909	 * Must align to double word boundary for the double cmpxchg
3910	 * instructions to work; see __pcpu_double_call_return_bool().
3911	 */
3912	s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3913				     2 * sizeof(void *));
3914
3915	if (!s->cpu_slab)
3916		return 0;
3917
3918	init_kmem_cache_cpus(s);
3919
3920	return 1;
3921}
3922
3923static struct kmem_cache *kmem_cache_node;
3924
3925/*
3926 * No kmalloc_node yet so do it by hand. We know that this is the first
3927 * slab on the node for this slabcache. There are no concurrent accesses
3928 * possible.
3929 *
3930 * Note that this function only works on the kmem_cache_node
3931 * when allocating for the kmem_cache_node. This is used for bootstrapping
3932 * memory on a fresh node that has no slab structures yet.
3933 */
3934static void early_kmem_cache_node_alloc(int node)
3935{
3936	struct slab *slab;
3937	struct kmem_cache_node *n;
3938
3939	BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3940
3941	slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3942
3943	BUG_ON(!slab);
3944	if (slab_nid(slab) != node) {
3945		pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3946		pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3947	}
3948
3949	n = slab->freelist;
3950	BUG_ON(!n);
3951#ifdef CONFIG_SLUB_DEBUG
3952	init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3953	init_tracking(kmem_cache_node, n);
3954#endif
3955	n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
3956	slab->freelist = get_freepointer(kmem_cache_node, n);
3957	slab->inuse = 1;
3958	slab->frozen = 0;
3959	kmem_cache_node->node[node] = n;
3960	init_kmem_cache_node(n);
3961	inc_slabs_node(kmem_cache_node, node, slab->objects);
3962
3963	/*
3964	 * No locks need to be taken here as it has just been
3965	 * initialized and there is no concurrent access.
3966	 */
3967	__add_partial(n, slab, DEACTIVATE_TO_HEAD);
3968}
3969
3970static void free_kmem_cache_nodes(struct kmem_cache *s)
3971{
3972	int node;
3973	struct kmem_cache_node *n;
3974
3975	for_each_kmem_cache_node(s, node, n) {
3976		s->node[node] = NULL;
3977		kmem_cache_free(kmem_cache_node, n);
3978	}
3979}
3980
3981void __kmem_cache_release(struct kmem_cache *s)
3982{
3983	cache_random_seq_destroy(s);
3984	free_percpu(s->cpu_slab);
3985	free_kmem_cache_nodes(s);
3986}
3987
3988static int init_kmem_cache_nodes(struct kmem_cache *s)
3989{
3990	int node;
3991
3992	for_each_node_mask(node, slab_nodes) {
3993		struct kmem_cache_node *n;
3994
3995		if (slab_state == DOWN) {
3996			early_kmem_cache_node_alloc(node);
3997			continue;
3998		}
3999		n = kmem_cache_alloc_node(kmem_cache_node,
4000						GFP_KERNEL, node);
4001
4002		if (!n) {
4003			free_kmem_cache_nodes(s);
4004			return 0;
4005		}
4006
4007		init_kmem_cache_node(n);
4008		s->node[node] = n;
4009	}
4010	return 1;
4011}
4012
4013static void set_cpu_partial(struct kmem_cache *s)
4014{
4015#ifdef CONFIG_SLUB_CPU_PARTIAL
4016	unsigned int nr_objects;
4017
4018	/*
4019	 * cpu_partial determined the maximum number of objects kept in the
4020	 * per cpu partial lists of a processor.
4021	 *
4022	 * Per cpu partial lists mainly contain slabs that just have one
4023	 * object freed. If they are used for allocation then they can be
4024	 * filled up again with minimal effort. The slab will never hit the
4025	 * per node partial lists and therefore no locking will be required.
4026	 *
4027	 * For backwards compatibility reasons, this is determined as number
4028	 * of objects, even though we now limit maximum number of pages, see
4029	 * slub_set_cpu_partial()
4030	 */
4031	if (!kmem_cache_has_cpu_partial(s))
4032		nr_objects = 0;
4033	else if (s->size >= PAGE_SIZE)
4034		nr_objects = 6;
4035	else if (s->size >= 1024)
4036		nr_objects = 24;
4037	else if (s->size >= 256)
4038		nr_objects = 52;
4039	else
4040		nr_objects = 120;
4041
4042	slub_set_cpu_partial(s, nr_objects);
4043#endif
4044}
4045
4046/*
4047 * calculate_sizes() determines the order and the distribution of data within
4048 * a slab object.
4049 */
4050static int calculate_sizes(struct kmem_cache *s)
4051{
4052	slab_flags_t flags = s->flags;
4053	unsigned int size = s->object_size;
4054	unsigned int order;
4055
4056	/*
4057	 * Round up object size to the next word boundary. We can only
4058	 * place the free pointer at word boundaries and this determines
4059	 * the possible location of the free pointer.
4060	 */
4061	size = ALIGN(size, sizeof(void *));
4062
4063#ifdef CONFIG_SLUB_DEBUG
4064	/*
4065	 * Determine if we can poison the object itself. If the user of
4066	 * the slab may touch the object after free or before allocation
4067	 * then we should never poison the object itself.
4068	 */
4069	if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4070			!s->ctor)
4071		s->flags |= __OBJECT_POISON;
4072	else
4073		s->flags &= ~__OBJECT_POISON;
4074
4075
4076	/*
4077	 * If we are Redzoning then check if there is some space between the
4078	 * end of the object and the free pointer. If not then add an
4079	 * additional word to have some bytes to store Redzone information.
4080	 */
4081	if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4082		size += sizeof(void *);
4083#endif
4084
4085	/*
4086	 * With that we have determined the number of bytes in actual use
4087	 * by the object and redzoning.
4088	 */
4089	s->inuse = size;
4090
4091	if ((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4092	    ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4093	    s->ctor) {
4094		/*
4095		 * Relocate free pointer after the object if it is not
4096		 * permitted to overwrite the first word of the object on
4097		 * kmem_cache_free.
4098		 *
4099		 * This is the case if we do RCU, have a constructor or
4100		 * destructor, are poisoning the objects, or are
4101		 * redzoning an object smaller than sizeof(void *).
4102		 *
4103		 * The assumption that s->offset >= s->inuse means free
4104		 * pointer is outside of the object is used in the
4105		 * freeptr_outside_object() function. If that is no
4106		 * longer true, the function needs to be modified.
4107		 */
4108		s->offset = size;
4109		size += sizeof(void *);
4110	} else {
4111		/*
4112		 * Store freelist pointer near middle of object to keep
4113		 * it away from the edges of the object to avoid small
4114		 * sized over/underflows from neighboring allocations.
4115		 */
4116		s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4117	}
4118
4119#ifdef CONFIG_SLUB_DEBUG
4120	if (flags & SLAB_STORE_USER)
4121		/*
4122		 * Need to store information about allocs and frees after
4123		 * the object.
4124		 */
4125		size += 2 * sizeof(struct track);
4126#endif
4127
4128	kasan_cache_create(s, &size, &s->flags);
4129#ifdef CONFIG_SLUB_DEBUG
4130	if (flags & SLAB_RED_ZONE) {
4131		/*
4132		 * Add some empty padding so that we can catch
4133		 * overwrites from earlier objects rather than let
4134		 * tracking information or the free pointer be
4135		 * corrupted if a user writes before the start
4136		 * of the object.
4137		 */
4138		size += sizeof(void *);
4139
4140		s->red_left_pad = sizeof(void *);
4141		s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4142		size += s->red_left_pad;
4143	}
4144#endif
4145
4146	/*
4147	 * SLUB stores one object immediately after another beginning from
4148	 * offset 0. In order to align the objects we have to simply size
4149	 * each object to conform to the alignment.
4150	 */
4151	size = ALIGN(size, s->align);
4152	s->size = size;
4153	s->reciprocal_size = reciprocal_value(size);
4154	order = calculate_order(size);
4155
4156	if ((int)order < 0)
4157		return 0;
4158
4159	s->allocflags = 0;
4160	if (order)
4161		s->allocflags |= __GFP_COMP;
4162
4163	if (s->flags & SLAB_CACHE_DMA)
4164		s->allocflags |= GFP_DMA;
4165
4166	if (s->flags & SLAB_CACHE_DMA32)
4167		s->allocflags |= GFP_DMA32;
4168
4169	if (s->flags & SLAB_RECLAIM_ACCOUNT)
4170		s->allocflags |= __GFP_RECLAIMABLE;
4171
4172	/*
4173	 * Determine the number of objects per slab
4174	 */
4175	s->oo = oo_make(order, size);
4176	s->min = oo_make(get_order(size), size);
4177
4178	return !!oo_objects(s->oo);
4179}
4180
4181static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4182{
4183	s->flags = kmem_cache_flags(s->size, flags, s->name);
4184#ifdef CONFIG_SLAB_FREELIST_HARDENED
4185	s->random = get_random_long();
4186#endif
4187
4188	if (!calculate_sizes(s))
4189		goto error;
4190	if (disable_higher_order_debug) {
4191		/*
4192		 * Disable debugging flags that store metadata if the min slab
4193		 * order increased.
4194		 */
4195		if (get_order(s->size) > get_order(s->object_size)) {
4196			s->flags &= ~DEBUG_METADATA_FLAGS;
4197			s->offset = 0;
4198			if (!calculate_sizes(s))
4199				goto error;
4200		}
4201	}
4202
4203#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4204    defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4205	if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4206		/* Enable fast mode */
4207		s->flags |= __CMPXCHG_DOUBLE;
4208#endif
4209
4210	/*
4211	 * The larger the object size is, the more slabs we want on the partial
4212	 * list to avoid pounding the page allocator excessively.
4213	 */
4214	s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4215	s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4216
4217	set_cpu_partial(s);
4218
4219#ifdef CONFIG_NUMA
4220	s->remote_node_defrag_ratio = 1000;
4221#endif
4222
4223	/* Initialize the pre-computed randomized freelist if slab is up */
4224	if (slab_state >= UP) {
4225		if (init_cache_random_seq(s))
4226			goto error;
4227	}
4228
4229	if (!init_kmem_cache_nodes(s))
4230		goto error;
4231
4232	if (alloc_kmem_cache_cpus(s))
4233		return 0;
4234
4235error:
4236	__kmem_cache_release(s);
4237	return -EINVAL;
4238}
4239
4240static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4241			      const char *text)
4242{
4243#ifdef CONFIG_SLUB_DEBUG
4244	void *addr = slab_address(slab);
4245	unsigned long flags;
4246	unsigned long *map;
4247	void *p;
4248
4249	slab_err(s, slab, text, s->name);
4250	slab_lock(slab, &flags);
4251
4252	map = get_map(s, slab);
4253	for_each_object(p, s, addr, slab->objects) {
4254
4255		if (!test_bit(__obj_to_index(s, addr, p), map)) {
4256			pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4257			print_tracking(s, p);
4258		}
4259	}
4260	put_map(map);
4261	slab_unlock(slab, &flags);
4262#endif
4263}
4264
4265/*
4266 * Attempt to free all partial slabs on a node.
4267 * This is called from __kmem_cache_shutdown(). We must take list_lock
4268 * because sysfs file might still access partial list after the shutdowning.
4269 */
4270static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4271{
4272	LIST_HEAD(discard);
4273	struct slab *slab, *h;
4274
4275	BUG_ON(irqs_disabled());
4276	spin_lock_irq(&n->list_lock);
4277	list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4278		if (!slab->inuse) {
4279			remove_partial(n, slab);
4280			list_add(&slab->slab_list, &discard);
4281		} else {
4282			list_slab_objects(s, slab,
4283			  "Objects remaining in %s on __kmem_cache_shutdown()");
4284		}
4285	}
4286	spin_unlock_irq(&n->list_lock);
4287
4288	list_for_each_entry_safe(slab, h, &discard, slab_list)
4289		discard_slab(s, slab);
4290}
4291
4292bool __kmem_cache_empty(struct kmem_cache *s)
4293{
4294	int node;
4295	struct kmem_cache_node *n;
4296
4297	for_each_kmem_cache_node(s, node, n)
4298		if (n->nr_partial || slabs_node(s, node))
4299			return false;
4300	return true;
4301}
4302
4303/*
4304 * Release all resources used by a slab cache.
4305 */
4306int __kmem_cache_shutdown(struct kmem_cache *s)
4307{
4308	int node;
4309	struct kmem_cache_node *n;
4310
4311	flush_all_cpus_locked(s);
4312	/* Attempt to free all objects */
4313	for_each_kmem_cache_node(s, node, n) {
4314		free_partial(s, n);
4315		if (n->nr_partial || slabs_node(s, node))
4316			return 1;
4317	}
4318	return 0;
4319}
4320
4321#ifdef CONFIG_PRINTK
4322void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4323{
4324	void *base;
4325	int __maybe_unused i;
4326	unsigned int objnr;
4327	void *objp;
4328	void *objp0;
4329	struct kmem_cache *s = slab->slab_cache;
4330	struct track __maybe_unused *trackp;
4331
4332	kpp->kp_ptr = object;
4333	kpp->kp_slab = slab;
4334	kpp->kp_slab_cache = s;
4335	base = slab_address(slab);
4336	objp0 = kasan_reset_tag(object);
4337#ifdef CONFIG_SLUB_DEBUG
4338	objp = restore_red_left(s, objp0);
4339#else
4340	objp = objp0;
4341#endif
4342	objnr = obj_to_index(s, slab, objp);
4343	kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4344	objp = base + s->size * objnr;
4345	kpp->kp_objp = objp;
4346	if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4347			 || (objp - base) % s->size) ||
4348	    !(s->flags & SLAB_STORE_USER))
4349		return;
4350#ifdef CONFIG_SLUB_DEBUG
4351	objp = fixup_red_left(s, objp);
4352	trackp = get_track(s, objp, TRACK_ALLOC);
4353	kpp->kp_ret = (void *)trackp->addr;
4354#ifdef CONFIG_STACKDEPOT
4355	{
4356		depot_stack_handle_t handle;
4357		unsigned long *entries;
4358		unsigned int nr_entries;
4359
4360		handle = READ_ONCE(trackp->handle);
4361		if (handle) {
4362			nr_entries = stack_depot_fetch(handle, &entries);
4363			for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4364				kpp->kp_stack[i] = (void *)entries[i];
4365		}
4366
4367		trackp = get_track(s, objp, TRACK_FREE);
4368		handle = READ_ONCE(trackp->handle);
4369		if (handle) {
4370			nr_entries = stack_depot_fetch(handle, &entries);
4371			for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4372				kpp->kp_free_stack[i] = (void *)entries[i];
4373		}
4374	}
4375#endif
4376#endif
4377}
4378#endif
4379
4380/********************************************************************
4381 *		Kmalloc subsystem
4382 *******************************************************************/
4383
4384static int __init setup_slub_min_order(char *str)
4385{
4386	get_option(&str, (int *)&slub_min_order);
4387
4388	return 1;
4389}
4390
4391__setup("slub_min_order=", setup_slub_min_order);
4392
4393static int __init setup_slub_max_order(char *str)
4394{
4395	get_option(&str, (int *)&slub_max_order);
4396	slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4397
4398	return 1;
4399}
4400
4401__setup("slub_max_order=", setup_slub_max_order);
4402
4403static int __init setup_slub_min_objects(char *str)
4404{
4405	get_option(&str, (int *)&slub_min_objects);
4406
4407	return 1;
4408}
4409
4410__setup("slub_min_objects=", setup_slub_min_objects);
4411
4412void *__kmalloc(size_t size, gfp_t flags)
4413{
4414	struct kmem_cache *s;
4415	void *ret;
4416
4417	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4418		return kmalloc_large(size, flags);
4419
4420	s = kmalloc_slab(size, flags);
4421
4422	if (unlikely(ZERO_OR_NULL_PTR(s)))
4423		return s;
4424
4425	ret = slab_alloc(s, NULL, flags, _RET_IP_, size);
4426
4427	trace_kmalloc(_RET_IP_, ret, s, size, s->size, flags);
4428
4429	ret = kasan_kmalloc(s, ret, size, flags);
4430
4431	return ret;
4432}
4433EXPORT_SYMBOL(__kmalloc);
4434
4435#ifdef CONFIG_NUMA
4436static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
4437{
4438	struct page *page;
4439	void *ptr = NULL;
4440	unsigned int order = get_order(size);
4441
4442	flags |= __GFP_COMP;
4443	page = alloc_pages_node(node, flags, order);
4444	if (page) {
4445		ptr = page_address(page);
4446		mod_lruvec_page_state(page, NR_SLAB_UNRECLAIMABLE_B,
4447				      PAGE_SIZE << order);
4448	}
4449
4450	return kmalloc_large_node_hook(ptr, size, flags);
4451}
4452
4453void *__kmalloc_node(size_t size, gfp_t flags, int node)
4454{
4455	struct kmem_cache *s;
4456	void *ret;
4457
4458	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4459		ret = kmalloc_large_node(size, flags, node);
4460
4461		trace_kmalloc_node(_RET_IP_, ret, NULL,
4462				   size, PAGE_SIZE << get_order(size),
4463				   flags, node);
4464
4465		return ret;
4466	}
4467
4468	s = kmalloc_slab(size, flags);
4469
4470	if (unlikely(ZERO_OR_NULL_PTR(s)))
4471		return s;
4472
4473	ret = slab_alloc_node(s, NULL, flags, node, _RET_IP_, size);
4474
4475	trace_kmalloc_node(_RET_IP_, ret, s, size, s->size, flags, node);
4476
4477	ret = kasan_kmalloc(s, ret, size, flags);
4478
4479	return ret;
4480}
4481EXPORT_SYMBOL(__kmalloc_node);
4482#endif	/* CONFIG_NUMA */
4483
4484#ifdef CONFIG_HARDENED_USERCOPY
4485/*
4486 * Rejects incorrectly sized objects and objects that are to be copied
4487 * to/from userspace but do not fall entirely within the containing slab
4488 * cache's usercopy region.
4489 *
4490 * Returns NULL if check passes, otherwise const char * to name of cache
4491 * to indicate an error.
4492 */
4493void __check_heap_object(const void *ptr, unsigned long n,
4494			 const struct slab *slab, bool to_user)
4495{
4496	struct kmem_cache *s;
4497	unsigned int offset;
4498	bool is_kfence = is_kfence_address(ptr);
4499
4500	ptr = kasan_reset_tag(ptr);
4501
4502	/* Find object and usable object size. */
4503	s = slab->slab_cache;
4504
4505	/* Reject impossible pointers. */
4506	if (ptr < slab_address(slab))
4507		usercopy_abort("SLUB object not in SLUB page?!", NULL,
4508			       to_user, 0, n);
4509
4510	/* Find offset within object. */
4511	if (is_kfence)
4512		offset = ptr - kfence_object_start(ptr);
4513	else
4514		offset = (ptr - slab_address(slab)) % s->size;
4515
4516	/* Adjust for redzone and reject if within the redzone. */
4517	if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4518		if (offset < s->red_left_pad)
4519			usercopy_abort("SLUB object in left red zone",
4520				       s->name, to_user, offset, n);
4521		offset -= s->red_left_pad;
4522	}
4523
4524	/* Allow address range falling entirely within usercopy region. */
4525	if (offset >= s->useroffset &&
4526	    offset - s->useroffset <= s->usersize &&
4527	    n <= s->useroffset - offset + s->usersize)
4528		return;
4529
4530	usercopy_abort("SLUB object", s->name, to_user, offset, n);
4531}
4532#endif /* CONFIG_HARDENED_USERCOPY */
4533
4534size_t __ksize(const void *object)
4535{
4536	struct folio *folio;
4537
4538	if (unlikely(object == ZERO_SIZE_PTR))
4539		return 0;
4540
4541	folio = virt_to_folio(object);
4542
4543	if (unlikely(!folio_test_slab(folio)))
4544		return folio_size(folio);
4545
4546	return slab_ksize(folio_slab(folio)->slab_cache);
4547}
4548EXPORT_SYMBOL(__ksize);
4549
4550void kfree(const void *x)
4551{
4552	struct folio *folio;
4553	struct slab *slab;
4554	void *object = (void *)x;
4555
4556	trace_kfree(_RET_IP_, x);
4557
4558	if (unlikely(ZERO_OR_NULL_PTR(x)))
4559		return;
4560
4561	folio = virt_to_folio(x);
4562	if (unlikely(!folio_test_slab(folio))) {
4563		free_large_kmalloc(folio, object);
4564		return;
4565	}
4566	slab = folio_slab(folio);
4567	slab_free(slab->slab_cache, slab, object, NULL, &object, 1, _RET_IP_);
4568}
4569EXPORT_SYMBOL(kfree);
4570
4571#define SHRINK_PROMOTE_MAX 32
4572
4573/*
4574 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4575 * up most to the head of the partial lists. New allocations will then
4576 * fill those up and thus they can be removed from the partial lists.
4577 *
4578 * The slabs with the least items are placed last. This results in them
4579 * being allocated from last increasing the chance that the last objects
4580 * are freed in them.
4581 */
4582static int __kmem_cache_do_shrink(struct kmem_cache *s)
4583{
4584	int node;
4585	int i;
4586	struct kmem_cache_node *n;
4587	struct slab *slab;
4588	struct slab *t;
4589	struct list_head discard;
4590	struct list_head promote[SHRINK_PROMOTE_MAX];
4591	unsigned long flags;
4592	int ret = 0;
4593
4594	for_each_kmem_cache_node(s, node, n) {
4595		INIT_LIST_HEAD(&discard);
4596		for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4597			INIT_LIST_HEAD(promote + i);
4598
4599		spin_lock_irqsave(&n->list_lock, flags);
4600
4601		/*
4602		 * Build lists of slabs to discard or promote.
4603		 *
4604		 * Note that concurrent frees may occur while we hold the
4605		 * list_lock. slab->inuse here is the upper limit.
4606		 */
4607		list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4608			int free = slab->objects - slab->inuse;
4609
4610			/* Do not reread slab->inuse */
4611			barrier();
4612
4613			/* We do not keep full slabs on the list */
4614			BUG_ON(free <= 0);
4615
4616			if (free == slab->objects) {
4617				list_move(&slab->slab_list, &discard);
4618				n->nr_partial--;
4619			} else if (free <= SHRINK_PROMOTE_MAX)
4620				list_move(&slab->slab_list, promote + free - 1);
4621		}
4622
4623		/*
4624		 * Promote the slabs filled up most to the head of the
4625		 * partial list.
4626		 */
4627		for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4628			list_splice(promote + i, &n->partial);
4629
4630		spin_unlock_irqrestore(&n->list_lock, flags);
4631
4632		/* Release empty slabs */
4633		list_for_each_entry_safe(slab, t, &discard, slab_list)
4634			discard_slab(s, slab);
4635
4636		if (slabs_node(s, node))
4637			ret = 1;
4638	}
4639
4640	return ret;
4641}
4642
4643int __kmem_cache_shrink(struct kmem_cache *s)
4644{
4645	flush_all(s);
4646	return __kmem_cache_do_shrink(s);
4647}
4648
4649static int slab_mem_going_offline_callback(void *arg)
4650{
4651	struct kmem_cache *s;
4652
4653	mutex_lock(&slab_mutex);
4654	list_for_each_entry(s, &slab_caches, list) {
4655		flush_all_cpus_locked(s);
4656		__kmem_cache_do_shrink(s);
4657	}
4658	mutex_unlock(&slab_mutex);
4659
4660	return 0;
4661}
4662
4663static void slab_mem_offline_callback(void *arg)
4664{
4665	struct memory_notify *marg = arg;
4666	int offline_node;
4667
4668	offline_node = marg->status_change_nid_normal;
4669
4670	/*
4671	 * If the node still has available memory. we need kmem_cache_node
4672	 * for it yet.
4673	 */
4674	if (offline_node < 0)
4675		return;
4676
4677	mutex_lock(&slab_mutex);
4678	node_clear(offline_node, slab_nodes);
4679	/*
4680	 * We no longer free kmem_cache_node structures here, as it would be
4681	 * racy with all get_node() users, and infeasible to protect them with
4682	 * slab_mutex.
4683	 */
4684	mutex_unlock(&slab_mutex);
4685}
4686
4687static int slab_mem_going_online_callback(void *arg)
4688{
4689	struct kmem_cache_node *n;
4690	struct kmem_cache *s;
4691	struct memory_notify *marg = arg;
4692	int nid = marg->status_change_nid_normal;
4693	int ret = 0;
4694
4695	/*
4696	 * If the node's memory is already available, then kmem_cache_node is
4697	 * already created. Nothing to do.
4698	 */
4699	if (nid < 0)
4700		return 0;
4701
4702	/*
4703	 * We are bringing a node online. No memory is available yet. We must
4704	 * allocate a kmem_cache_node structure in order to bring the node
4705	 * online.
4706	 */
4707	mutex_lock(&slab_mutex);
4708	list_for_each_entry(s, &slab_caches, list) {
4709		/*
4710		 * The structure may already exist if the node was previously
4711		 * onlined and offlined.
4712		 */
4713		if (get_node(s, nid))
4714			continue;
4715		/*
4716		 * XXX: kmem_cache_alloc_node will fallback to other nodes
4717		 *      since memory is not yet available from the node that
4718		 *      is brought up.
4719		 */
4720		n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4721		if (!n) {
4722			ret = -ENOMEM;
4723			goto out;
4724		}
4725		init_kmem_cache_node(n);
4726		s->node[nid] = n;
4727	}
4728	/*
4729	 * Any cache created after this point will also have kmem_cache_node
4730	 * initialized for the new node.
4731	 */
4732	node_set(nid, slab_nodes);
4733out:
4734	mutex_unlock(&slab_mutex);
4735	return ret;
4736}
4737
4738static int slab_memory_callback(struct notifier_block *self,
4739				unsigned long action, void *arg)
4740{
4741	int ret = 0;
4742
4743	switch (action) {
4744	case MEM_GOING_ONLINE:
4745		ret = slab_mem_going_online_callback(arg);
4746		break;
4747	case MEM_GOING_OFFLINE:
4748		ret = slab_mem_going_offline_callback(arg);
4749		break;
4750	case MEM_OFFLINE:
4751	case MEM_CANCEL_ONLINE:
4752		slab_mem_offline_callback(arg);
4753		break;
4754	case MEM_ONLINE:
4755	case MEM_CANCEL_OFFLINE:
4756		break;
4757	}
4758	if (ret)
4759		ret = notifier_from_errno(ret);
4760	else
4761		ret = NOTIFY_OK;
4762	return ret;
4763}
4764
4765static struct notifier_block slab_memory_callback_nb = {
4766	.notifier_call = slab_memory_callback,
4767	.priority = SLAB_CALLBACK_PRI,
4768};
4769
4770/********************************************************************
4771 *			Basic setup of slabs
4772 *******************************************************************/
4773
4774/*
4775 * Used for early kmem_cache structures that were allocated using
4776 * the page allocator. Allocate them properly then fix up the pointers
4777 * that may be pointing to the wrong kmem_cache structure.
4778 */
4779
4780static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4781{
4782	int node;
4783	struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4784	struct kmem_cache_node *n;
4785
4786	memcpy(s, static_cache, kmem_cache->object_size);
4787
4788	/*
4789	 * This runs very early, and only the boot processor is supposed to be
4790	 * up.  Even if it weren't true, IRQs are not up so we couldn't fire
4791	 * IPIs around.
4792	 */
4793	__flush_cpu_slab(s, smp_processor_id());
4794	for_each_kmem_cache_node(s, node, n) {
4795		struct slab *p;
4796
4797		list_for_each_entry(p, &n->partial, slab_list)
4798			p->slab_cache = s;
4799
4800#ifdef CONFIG_SLUB_DEBUG
4801		list_for_each_entry(p, &n->full, slab_list)
4802			p->slab_cache = s;
4803#endif
4804	}
4805	list_add(&s->list, &slab_caches);
4806	return s;
4807}
4808
4809void __init kmem_cache_init(void)
4810{
4811	static __initdata struct kmem_cache boot_kmem_cache,
4812		boot_kmem_cache_node;
4813	int node;
4814
4815	if (debug_guardpage_minorder())
4816		slub_max_order = 0;
4817
4818	/* Print slub debugging pointers without hashing */
4819	if (__slub_debug_enabled())
4820		no_hash_pointers_enable(NULL);
4821
4822	kmem_cache_node = &boot_kmem_cache_node;
4823	kmem_cache = &boot_kmem_cache;
4824
4825	/*
4826	 * Initialize the nodemask for which we will allocate per node
4827	 * structures. Here we don't need taking slab_mutex yet.
4828	 */
4829	for_each_node_state(node, N_NORMAL_MEMORY)
4830		node_set(node, slab_nodes);
4831
4832	create_boot_cache(kmem_cache_node, "kmem_cache_node",
4833		sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4834
4835	register_hotmemory_notifier(&slab_memory_callback_nb);
4836
4837	/* Able to allocate the per node structures */
4838	slab_state = PARTIAL;
4839
4840	create_boot_cache(kmem_cache, "kmem_cache",
4841			offsetof(struct kmem_cache, node) +
4842				nr_node_ids * sizeof(struct kmem_cache_node *),
4843		       SLAB_HWCACHE_ALIGN, 0, 0);
4844
4845	kmem_cache = bootstrap(&boot_kmem_cache);
4846	kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4847
4848	/* Now we can use the kmem_cache to allocate kmalloc slabs */
4849	setup_kmalloc_cache_index_table();
4850	create_kmalloc_caches(0);
4851
4852	/* Setup random freelists for each cache */
4853	init_freelist_randomization();
4854
4855	cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4856				  slub_cpu_dead);
4857
4858	pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4859		cache_line_size(),
4860		slub_min_order, slub_max_order, slub_min_objects,
4861		nr_cpu_ids, nr_node_ids);
4862}
4863
4864void __init kmem_cache_init_late(void)
4865{
4866	flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
4867	WARN_ON(!flushwq);
4868}
4869
4870struct kmem_cache *
4871__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4872		   slab_flags_t flags, void (*ctor)(void *))
4873{
4874	struct kmem_cache *s;
4875
4876	s = find_mergeable(size, align, flags, name, ctor);
4877	if (s) {
4878		if (sysfs_slab_alias(s, name))
4879			return NULL;
4880
4881		s->refcount++;
4882
4883		/*
4884		 * Adjust the object sizes so that we clear
4885		 * the complete object on kzalloc.
4886		 */
4887		s->object_size = max(s->object_size, size);
4888		s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4889	}
4890
4891	return s;
4892}
4893
4894int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4895{
4896	int err;
4897
4898	err = kmem_cache_open(s, flags);
4899	if (err)
4900		return err;
4901
4902	/* Mutex is not taken during early boot */
4903	if (slab_state <= UP)
4904		return 0;
4905
4906	err = sysfs_slab_add(s);
4907	if (err) {
4908		__kmem_cache_release(s);
4909		return err;
4910	}
4911
4912	if (s->flags & SLAB_STORE_USER)
4913		debugfs_slab_add(s);
4914
4915	return 0;
4916}
4917
4918void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4919{
4920	struct kmem_cache *s;
4921	void *ret;
4922
4923	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4924		return kmalloc_large(size, gfpflags);
4925
4926	s = kmalloc_slab(size, gfpflags);
4927
4928	if (unlikely(ZERO_OR_NULL_PTR(s)))
4929		return s;
4930
4931	ret = slab_alloc(s, NULL, gfpflags, caller, size);
4932
4933	/* Honor the call site pointer we received. */
4934	trace_kmalloc(caller, ret, s, size, s->size, gfpflags);
4935
4936	ret = kasan_kmalloc(s, ret, size, gfpflags);
4937
4938	return ret;
4939}
4940EXPORT_SYMBOL(__kmalloc_track_caller);
4941
4942#ifdef CONFIG_NUMA
4943void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4944					int node, unsigned long caller)
4945{
4946	struct kmem_cache *s;
4947	void *ret;
4948
4949	if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4950		ret = kmalloc_large_node(size, gfpflags, node);
4951
4952		trace_kmalloc_node(caller, ret, NULL,
4953				   size, PAGE_SIZE << get_order(size),
4954				   gfpflags, node);
4955
4956		return ret;
4957	}
4958
4959	s = kmalloc_slab(size, gfpflags);
4960
4961	if (unlikely(ZERO_OR_NULL_PTR(s)))
4962		return s;
4963
4964	ret = slab_alloc_node(s, NULL, gfpflags, node, caller, size);
4965
4966	/* Honor the call site pointer we received. */
4967	trace_kmalloc_node(caller, ret, s, size, s->size, gfpflags, node);
4968
4969	ret = kasan_kmalloc(s, ret, size, gfpflags);
4970
4971	return ret;
4972}
4973EXPORT_SYMBOL(__kmalloc_node_track_caller);
4974#endif
4975
4976#ifdef CONFIG_SYSFS
4977static int count_inuse(struct slab *slab)
4978{
4979	return slab->inuse;
4980}
4981
4982static int count_total(struct slab *slab)
4983{
4984	return slab->objects;
4985}
4986#endif
4987
4988#ifdef CONFIG_SLUB_DEBUG
4989static void validate_slab(struct kmem_cache *s, struct slab *slab,
4990			  unsigned long *obj_map)
4991{
4992	void *p;
4993	void *addr = slab_address(slab);
4994	unsigned long flags;
4995
4996	slab_lock(slab, &flags);
4997
4998	if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
4999		goto unlock;
5000
5001	/* Now we know that a valid freelist exists */
5002	__fill_map(obj_map, s, slab);
5003	for_each_object(p, s, addr, slab->objects) {
5004		u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5005			 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5006
5007		if (!check_object(s, slab, p, val))
5008			break;
5009	}
5010unlock:
5011	slab_unlock(slab, &flags);
5012}
5013
5014static int validate_slab_node(struct kmem_cache *s,
5015		struct kmem_cache_node *n, unsigned long *obj_map)
5016{
5017	unsigned long count = 0;
5018	struct slab *slab;
5019	unsigned long flags;
5020
5021	spin_lock_irqsave(&n->list_lock, flags);
5022
5023	list_for_each_entry(slab, &n->partial, slab_list) {
5024		validate_slab(s, slab, obj_map);
5025		count++;
5026	}
5027	if (count != n->nr_partial) {
5028		pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5029		       s->name, count, n->nr_partial);
5030		slab_add_kunit_errors();
5031	}
5032
5033	if (!(s->flags & SLAB_STORE_USER))
5034		goto out;
5035
5036	list_for_each_entry(slab, &n->full, slab_list) {
5037		validate_slab(s, slab, obj_map);
5038		count++;
5039	}
5040	if (count != atomic_long_read(&n->nr_slabs)) {
5041		pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5042		       s->name, count, atomic_long_read(&n->nr_slabs));
5043		slab_add_kunit_errors();
5044	}
5045
5046out:
5047	spin_unlock_irqrestore(&n->list_lock, flags);
5048	return count;
5049}
5050
5051long validate_slab_cache(struct kmem_cache *s)
5052{
5053	int node;
5054	unsigned long count = 0;
5055	struct kmem_cache_node *n;
5056	unsigned long *obj_map;
5057
5058	obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5059	if (!obj_map)
5060		return -ENOMEM;
5061
5062	flush_all(s);
5063	for_each_kmem_cache_node(s, node, n)
5064		count += validate_slab_node(s, n, obj_map);
5065
5066	bitmap_free(obj_map);
5067
5068	return count;
5069}
5070EXPORT_SYMBOL(validate_slab_cache);
5071
5072#ifdef CONFIG_DEBUG_FS
5073/*
5074 * Generate lists of code addresses where slabcache objects are allocated
5075 * and freed.
5076 */
5077
5078struct location {
5079	depot_stack_handle_t handle;
5080	unsigned long count;
5081	unsigned long addr;
5082	long long sum_time;
5083	long min_time;
5084	long max_time;
5085	long min_pid;
5086	long max_pid;
5087	DECLARE_BITMAP(cpus, NR_CPUS);
5088	nodemask_t nodes;
5089};
5090
5091struct loc_track {
5092	unsigned long max;
5093	unsigned long count;
5094	struct location *loc;
5095	loff_t idx;
5096};
5097
5098static struct dentry *slab_debugfs_root;
5099
5100static void free_loc_track(struct loc_track *t)
5101{
5102	if (t->max)
5103		free_pages((unsigned long)t->loc,
5104			get_order(sizeof(struct location) * t->max));
5105}
5106
5107static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5108{
5109	struct location *l;
5110	int order;
5111
5112	order = get_order(sizeof(struct location) * max);
5113
5114	l = (void *)__get_free_pages(flags, order);
5115	if (!l)
5116		return 0;
5117
5118	if (t->count) {
5119		memcpy(l, t->loc, sizeof(struct location) * t->count);
5120		free_loc_track(t);
5121	}
5122	t->max = max;
5123	t->loc = l;
5124	return 1;
5125}
5126
5127static int add_location(struct loc_track *t, struct kmem_cache *s,
5128				const struct track *track)
5129{
5130	long start, end, pos;
5131	struct location *l;
5132	unsigned long caddr, chandle;
5133	unsigned long age = jiffies - track->when;
5134	depot_stack_handle_t handle = 0;
5135
5136#ifdef CONFIG_STACKDEPOT
5137	handle = READ_ONCE(track->handle);
5138#endif
5139	start = -1;
5140	end = t->count;
5141
5142	for ( ; ; ) {
5143		pos = start + (end - start + 1) / 2;
5144
5145		/*
5146		 * There is nothing at "end". If we end up there
5147		 * we need to add something to before end.
5148		 */
5149		if (pos == end)
5150			break;
5151
5152		caddr = t->loc[pos].addr;
5153		chandle = t->loc[pos].handle;
5154		if ((track->addr == caddr) && (handle == chandle)) {
5155
5156			l = &t->loc[pos];
5157			l->count++;
5158			if (track->when) {
5159				l->sum_time += age;
5160				if (age < l->min_time)
5161					l->min_time = age;
5162				if (age > l->max_time)
5163					l->max_time = age;
5164
5165				if (track->pid < l->min_pid)
5166					l->min_pid = track->pid;
5167				if (track->pid > l->max_pid)
5168					l->max_pid = track->pid;
5169
5170				cpumask_set_cpu(track->cpu,
5171						to_cpumask(l->cpus));
5172			}
5173			node_set(page_to_nid(virt_to_page(track)), l->nodes);
5174			return 1;
5175		}
5176
5177		if (track->addr < caddr)
5178			end = pos;
5179		else if (track->addr == caddr && handle < chandle)
5180			end = pos;
5181		else
5182			start = pos;
5183	}
5184
5185	/*
5186	 * Not found. Insert new tracking element.
5187	 */
5188	if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5189		return 0;
5190
5191	l = t->loc + pos;
5192	if (pos < t->count)
5193		memmove(l + 1, l,
5194			(t->count - pos) * sizeof(struct location));
5195	t->count++;
5196	l->count = 1;
5197	l->addr = track->addr;
5198	l->sum_time = age;
5199	l->min_time = age;
5200	l->max_time = age;
5201	l->min_pid = track->pid;
5202	l->max_pid = track->pid;
5203	l->handle = handle;
5204	cpumask_clear(to_cpumask(l->cpus));
5205	cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5206	nodes_clear(l->nodes);
5207	node_set(page_to_nid(virt_to_page(track)), l->nodes);
5208	return 1;
5209}
5210
5211static void process_slab(struct loc_track *t, struct kmem_cache *s,
5212		struct slab *slab, enum track_item alloc,
5213		unsigned long *obj_map)
5214{
5215	void *addr = slab_address(slab);
5216	void *p;
5217
5218	__fill_map(obj_map, s, slab);
5219
5220	for_each_object(p, s, addr, slab->objects)
5221		if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5222			add_location(t, s, get_track(s, p, alloc));
5223}
5224#endif  /* CONFIG_DEBUG_FS   */
5225#endif	/* CONFIG_SLUB_DEBUG */
5226
5227#ifdef CONFIG_SYSFS
5228enum slab_stat_type {
5229	SL_ALL,			/* All slabs */
5230	SL_PARTIAL,		/* Only partially allocated slabs */
5231	SL_CPU,			/* Only slabs used for cpu caches */
5232	SL_OBJECTS,		/* Determine allocated objects not slabs */
5233	SL_TOTAL		/* Determine object capacity not slabs */
5234};
5235
5236#define SO_ALL		(1 << SL_ALL)
5237#define SO_PARTIAL	(1 << SL_PARTIAL)
5238#define SO_CPU		(1 << SL_CPU)
5239#define SO_OBJECTS	(1 << SL_OBJECTS)
5240#define SO_TOTAL	(1 << SL_TOTAL)
5241
5242static ssize_t show_slab_objects(struct kmem_cache *s,
5243				 char *buf, unsigned long flags)
5244{
5245	unsigned long total = 0;
5246	int node;
5247	int x;
5248	unsigned long *nodes;
5249	int len = 0;
5250
5251	nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5252	if (!nodes)
5253		return -ENOMEM;
5254
5255	if (flags & SO_CPU) {
5256		int cpu;
5257
5258		for_each_possible_cpu(cpu) {
5259			struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5260							       cpu);
5261			int node;
5262			struct slab *slab;
5263
5264			slab = READ_ONCE(c->slab);
5265			if (!slab)
5266				continue;
5267
5268			node = slab_nid(slab);
5269			if (flags & SO_TOTAL)
5270				x = slab->objects;
5271			else if (flags & SO_OBJECTS)
5272				x = slab->inuse;
5273			else
5274				x = 1;
5275
5276			total += x;
5277			nodes[node] += x;
5278
5279#ifdef CONFIG_SLUB_CPU_PARTIAL
5280			slab = slub_percpu_partial_read_once(c);
5281			if (slab) {
5282				node = slab_nid(slab);
5283				if (flags & SO_TOTAL)
5284					WARN_ON_ONCE(1);
5285				else if (flags & SO_OBJECTS)
5286					WARN_ON_ONCE(1);
5287				else
5288					x = slab->slabs;
5289				total += x;
5290				nodes[node] += x;
5291			}
5292#endif
5293		}
5294	}
5295
5296	/*
5297	 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5298	 * already held which will conflict with an existing lock order:
5299	 *
5300	 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5301	 *
5302	 * We don't really need mem_hotplug_lock (to hold off
5303	 * slab_mem_going_offline_callback) here because slab's memory hot
5304	 * unplug code doesn't destroy the kmem_cache->node[] data.
5305	 */
5306
5307#ifdef CONFIG_SLUB_DEBUG
5308	if (flags & SO_ALL) {
5309		struct kmem_cache_node *n;
5310
5311		for_each_kmem_cache_node(s, node, n) {
5312
5313			if (flags & SO_TOTAL)
5314				x = atomic_long_read(&n->total_objects);
5315			else if (flags & SO_OBJECTS)
5316				x = atomic_long_read(&n->total_objects) -
5317					count_partial(n, count_free);
5318			else
5319				x = atomic_long_read(&n->nr_slabs);
5320			total += x;
5321			nodes[node] += x;
5322		}
5323
5324	} else
5325#endif
5326	if (flags & SO_PARTIAL) {
5327		struct kmem_cache_node *n;
5328
5329		for_each_kmem_cache_node(s, node, n) {
5330			if (flags & SO_TOTAL)
5331				x = count_partial(n, count_total);
5332			else if (flags & SO_OBJECTS)
5333				x = count_partial(n, count_inuse);
5334			else
5335				x = n->nr_partial;
5336			total += x;
5337			nodes[node] += x;
5338		}
5339	}
5340
5341	len += sysfs_emit_at(buf, len, "%lu", total);
5342#ifdef CONFIG_NUMA
5343	for (node = 0; node < nr_node_ids; node++) {
5344		if (nodes[node])
5345			len += sysfs_emit_at(buf, len, " N%d=%lu",
5346					     node, nodes[node]);
5347	}
5348#endif
5349	len += sysfs_emit_at(buf, len, "\n");
5350	kfree(nodes);
5351
5352	return len;
5353}
5354
5355#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5356#define to_slab(n) container_of(n, struct kmem_cache, kobj)
5357
5358struct slab_attribute {
5359	struct attribute attr;
5360	ssize_t (*show)(struct kmem_cache *s, char *buf);
5361	ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5362};
5363
5364#define SLAB_ATTR_RO(_name) \
5365	static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5366
5367#define SLAB_ATTR(_name) \
5368	static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5369
5370static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5371{
5372	return sysfs_emit(buf, "%u\n", s->size);
5373}
5374SLAB_ATTR_RO(slab_size);
5375
5376static ssize_t align_show(struct kmem_cache *s, char *buf)
5377{
5378	return sysfs_emit(buf, "%u\n", s->align);
5379}
5380SLAB_ATTR_RO(align);
5381
5382static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5383{
5384	return sysfs_emit(buf, "%u\n", s->object_size);
5385}
5386SLAB_ATTR_RO(object_size);
5387
5388static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5389{
5390	return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5391}
5392SLAB_ATTR_RO(objs_per_slab);
5393
5394static ssize_t order_show(struct kmem_cache *s, char *buf)
5395{
5396	return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5397}
5398SLAB_ATTR_RO(order);
5399
5400static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5401{
5402	return sysfs_emit(buf, "%lu\n", s->min_partial);
5403}
5404
5405static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5406				 size_t length)
5407{
5408	unsigned long min;
5409	int err;
5410
5411	err = kstrtoul(buf, 10, &min);
5412	if (err)
5413		return err;
5414
5415	s->min_partial = min;
5416	return length;
5417}
5418SLAB_ATTR(min_partial);
5419
5420static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5421{
5422	unsigned int nr_partial = 0;
5423#ifdef CONFIG_SLUB_CPU_PARTIAL
5424	nr_partial = s->cpu_partial;
5425#endif
5426
5427	return sysfs_emit(buf, "%u\n", nr_partial);
5428}
5429
5430static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5431				 size_t length)
5432{
5433	unsigned int objects;
5434	int err;
5435
5436	err = kstrtouint(buf, 10, &objects);
5437	if (err)
5438		return err;
5439	if (objects && !kmem_cache_has_cpu_partial(s))
5440		return -EINVAL;
5441
5442	slub_set_cpu_partial(s, objects);
5443	flush_all(s);
5444	return length;
5445}
5446SLAB_ATTR(cpu_partial);
5447
5448static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5449{
5450	if (!s->ctor)
5451		return 0;
5452	return sysfs_emit(buf, "%pS\n", s->ctor);
5453}
5454SLAB_ATTR_RO(ctor);
5455
5456static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5457{
5458	return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5459}
5460SLAB_ATTR_RO(aliases);
5461
5462static ssize_t partial_show(struct kmem_cache *s, char *buf)
5463{
5464	return show_slab_objects(s, buf, SO_PARTIAL);
5465}
5466SLAB_ATTR_RO(partial);
5467
5468static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5469{
5470	return show_slab_objects(s, buf, SO_CPU);
5471}
5472SLAB_ATTR_RO(cpu_slabs);
5473
5474static ssize_t objects_show(struct kmem_cache *s, char *buf)
5475{
5476	return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5477}
5478SLAB_ATTR_RO(objects);
5479
5480static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5481{
5482	return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5483}
5484SLAB_ATTR_RO(objects_partial);
5485
5486static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5487{
5488	int objects = 0;
5489	int slabs = 0;
5490	int cpu __maybe_unused;
5491	int len = 0;
5492
5493#ifdef CONFIG_SLUB_CPU_PARTIAL
5494	for_each_online_cpu(cpu) {
5495		struct slab *slab;
5496
5497		slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5498
5499		if (slab)
5500			slabs += slab->slabs;
5501	}
5502#endif
5503
5504	/* Approximate half-full slabs, see slub_set_cpu_partial() */
5505	objects = (slabs * oo_objects(s->oo)) / 2;
5506	len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5507
5508#if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5509	for_each_online_cpu(cpu) {
5510		struct slab *slab;
5511
5512		slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5513		if (slab) {
5514			slabs = READ_ONCE(slab->slabs);
5515			objects = (slabs * oo_objects(s->oo)) / 2;
5516			len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5517					     cpu, objects, slabs);
5518		}
5519	}
5520#endif
5521	len += sysfs_emit_at(buf, len, "\n");
5522
5523	return len;
5524}
5525SLAB_ATTR_RO(slabs_cpu_partial);
5526
5527static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5528{
5529	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5530}
5531SLAB_ATTR_RO(reclaim_account);
5532
5533static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5534{
5535	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5536}
5537SLAB_ATTR_RO(hwcache_align);
5538
5539#ifdef CONFIG_ZONE_DMA
5540static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5541{
5542	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5543}
5544SLAB_ATTR_RO(cache_dma);
5545#endif
5546
5547static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5548{
5549	return sysfs_emit(buf, "%u\n", s->usersize);
5550}
5551SLAB_ATTR_RO(usersize);
5552
5553static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5554{
5555	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5556}
5557SLAB_ATTR_RO(destroy_by_rcu);
5558
5559#ifdef CONFIG_SLUB_DEBUG
5560static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5561{
5562	return show_slab_objects(s, buf, SO_ALL);
5563}
5564SLAB_ATTR_RO(slabs);
5565
5566static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5567{
5568	return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5569}
5570SLAB_ATTR_RO(total_objects);
5571
5572static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5573{
5574	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5575}
5576SLAB_ATTR_RO(sanity_checks);
5577
5578static ssize_t trace_show(struct kmem_cache *s, char *buf)
5579{
5580	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5581}
5582SLAB_ATTR_RO(trace);
5583
5584static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5585{
5586	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5587}
5588
5589SLAB_ATTR_RO(red_zone);
5590
5591static ssize_t poison_show(struct kmem_cache *s, char *buf)
5592{
5593	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5594}
5595
5596SLAB_ATTR_RO(poison);
5597
5598static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5599{
5600	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5601}
5602
5603SLAB_ATTR_RO(store_user);
5604
5605static ssize_t validate_show(struct kmem_cache *s, char *buf)
5606{
5607	return 0;
5608}
5609
5610static ssize_t validate_store(struct kmem_cache *s,
5611			const char *buf, size_t length)
5612{
5613	int ret = -EINVAL;
5614
5615	if (buf[0] == '1') {
5616		ret = validate_slab_cache(s);
5617		if (ret >= 0)
5618			ret = length;
5619	}
5620	return ret;
5621}
5622SLAB_ATTR(validate);
5623
5624#endif /* CONFIG_SLUB_DEBUG */
5625
5626#ifdef CONFIG_FAILSLAB
5627static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5628{
5629	return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5630}
5631SLAB_ATTR_RO(failslab);
5632#endif
5633
5634static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5635{
5636	return 0;
5637}
5638
5639static ssize_t shrink_store(struct kmem_cache *s,
5640			const char *buf, size_t length)
5641{
5642	if (buf[0] == '1')
5643		kmem_cache_shrink(s);
5644	else
5645		return -EINVAL;
5646	return length;
5647}
5648SLAB_ATTR(shrink);
5649
5650#ifdef CONFIG_NUMA
5651static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5652{
5653	return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5654}
5655
5656static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5657				const char *buf, size_t length)
5658{
5659	unsigned int ratio;
5660	int err;
5661
5662	err = kstrtouint(buf, 10, &ratio);
5663	if (err)
5664		return err;
5665	if (ratio > 100)
5666		return -ERANGE;
5667
5668	s->remote_node_defrag_ratio = ratio * 10;
5669
5670	return length;
5671}
5672SLAB_ATTR(remote_node_defrag_ratio);
5673#endif
5674
5675#ifdef CONFIG_SLUB_STATS
5676static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5677{
5678	unsigned long sum  = 0;
5679	int cpu;
5680	int len = 0;
5681	int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5682
5683	if (!data)
5684		return -ENOMEM;
5685
5686	for_each_online_cpu(cpu) {
5687		unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5688
5689		data[cpu] = x;
5690		sum += x;
5691	}
5692
5693	len += sysfs_emit_at(buf, len, "%lu", sum);
5694
5695#ifdef CONFIG_SMP
5696	for_each_online_cpu(cpu) {
5697		if (data[cpu])
5698			len += sysfs_emit_at(buf, len, " C%d=%u",
5699					     cpu, data[cpu]);
5700	}
5701#endif
5702	kfree(data);
5703	len += sysfs_emit_at(buf, len, "\n");
5704
5705	return len;
5706}
5707
5708static void clear_stat(struct kmem_cache *s, enum stat_item si)
5709{
5710	int cpu;
5711
5712	for_each_online_cpu(cpu)
5713		per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5714}
5715
5716#define STAT_ATTR(si, text) 					\
5717static ssize_t text##_show(struct kmem_cache *s, char *buf)	\
5718{								\
5719	return show_stat(s, buf, si);				\
5720}								\
5721static ssize_t text##_store(struct kmem_cache *s,		\
5722				const char *buf, size_t length)	\
5723{								\
5724	if (buf[0] != '0')					\
5725		return -EINVAL;					\
5726	clear_stat(s, si);					\
5727	return length;						\
5728}								\
5729SLAB_ATTR(text);						\
5730
5731STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5732STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5733STAT_ATTR(FREE_FASTPATH, free_fastpath);
5734STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5735STAT_ATTR(FREE_FROZEN, free_frozen);
5736STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5737STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5738STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5739STAT_ATTR(ALLOC_SLAB, alloc_slab);
5740STAT_ATTR(ALLOC_REFILL, alloc_refill);
5741STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5742STAT_ATTR(FREE_SLAB, free_slab);
5743STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5744STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5745STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5746STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5747STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5748STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5749STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5750STAT_ATTR(ORDER_FALLBACK, order_fallback);
5751STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5752STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5753STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5754STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5755STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5756STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5757#endif	/* CONFIG_SLUB_STATS */
5758
5759static struct attribute *slab_attrs[] = {
5760	&slab_size_attr.attr,
5761	&object_size_attr.attr,
5762	&objs_per_slab_attr.attr,
5763	&order_attr.attr,
5764	&min_partial_attr.attr,
5765	&cpu_partial_attr.attr,
5766	&objects_attr.attr,
5767	&objects_partial_attr.attr,
5768	&partial_attr.attr,
5769	&cpu_slabs_attr.attr,
5770	&ctor_attr.attr,
5771	&aliases_attr.attr,
5772	&align_attr.attr,
5773	&hwcache_align_attr.attr,
5774	&reclaim_account_attr.attr,
5775	&destroy_by_rcu_attr.attr,
5776	&shrink_attr.attr,
5777	&slabs_cpu_partial_attr.attr,
5778#ifdef CONFIG_SLUB_DEBUG
5779	&total_objects_attr.attr,
5780	&slabs_attr.attr,
5781	&sanity_checks_attr.attr,
5782	&trace_attr.attr,
5783	&red_zone_attr.attr,
5784	&poison_attr.attr,
5785	&store_user_attr.attr,
5786	&validate_attr.attr,
5787#endif
5788#ifdef CONFIG_ZONE_DMA
5789	&cache_dma_attr.attr,
5790#endif
5791#ifdef CONFIG_NUMA
5792	&remote_node_defrag_ratio_attr.attr,
5793#endif
5794#ifdef CONFIG_SLUB_STATS
5795	&alloc_fastpath_attr.attr,
5796	&alloc_slowpath_attr.attr,
5797	&free_fastpath_attr.attr,
5798	&free_slowpath_attr.attr,
5799	&free_frozen_attr.attr,
5800	&free_add_partial_attr.attr,
5801	&free_remove_partial_attr.attr,
5802	&alloc_from_partial_attr.attr,
5803	&alloc_slab_attr.attr,
5804	&alloc_refill_attr.attr,
5805	&alloc_node_mismatch_attr.attr,
5806	&free_slab_attr.attr,
5807	&cpuslab_flush_attr.attr,
5808	&deactivate_full_attr.attr,
5809	&deactivate_empty_attr.attr,
5810	&deactivate_to_head_attr.attr,
5811	&deactivate_to_tail_attr.attr,
5812	&deactivate_remote_frees_attr.attr,
5813	&deactivate_bypass_attr.attr,
5814	&order_fallback_attr.attr,
5815	&cmpxchg_double_fail_attr.attr,
5816	&cmpxchg_double_cpu_fail_attr.attr,
5817	&cpu_partial_alloc_attr.attr,
5818	&cpu_partial_free_attr.attr,
5819	&cpu_partial_node_attr.attr,
5820	&cpu_partial_drain_attr.attr,
5821#endif
5822#ifdef CONFIG_FAILSLAB
5823	&failslab_attr.attr,
5824#endif
5825	&usersize_attr.attr,
5826
5827	NULL
5828};
5829
5830static const struct attribute_group slab_attr_group = {
5831	.attrs = slab_attrs,
5832};
5833
5834static ssize_t slab_attr_show(struct kobject *kobj,
5835				struct attribute *attr,
5836				char *buf)
5837{
5838	struct slab_attribute *attribute;
5839	struct kmem_cache *s;
5840	int err;
5841
5842	attribute = to_slab_attr(attr);
5843	s = to_slab(kobj);
5844
5845	if (!attribute->show)
5846		return -EIO;
5847
5848	err = attribute->show(s, buf);
5849
5850	return err;
5851}
5852
5853static ssize_t slab_attr_store(struct kobject *kobj,
5854				struct attribute *attr,
5855				const char *buf, size_t len)
5856{
5857	struct slab_attribute *attribute;
5858	struct kmem_cache *s;
5859	int err;
5860
5861	attribute = to_slab_attr(attr);
5862	s = to_slab(kobj);
5863
5864	if (!attribute->store)
5865		return -EIO;
5866
5867	err = attribute->store(s, buf, len);
5868	return err;
5869}
5870
5871static void kmem_cache_release(struct kobject *k)
5872{
5873	slab_kmem_cache_release(to_slab(k));
5874}
5875
5876static const struct sysfs_ops slab_sysfs_ops = {
5877	.show = slab_attr_show,
5878	.store = slab_attr_store,
5879};
5880
5881static struct kobj_type slab_ktype = {
5882	.sysfs_ops = &slab_sysfs_ops,
5883	.release = kmem_cache_release,
5884};
5885
5886static struct kset *slab_kset;
5887
5888static inline struct kset *cache_kset(struct kmem_cache *s)
5889{
5890	return slab_kset;
5891}
5892
5893#define ID_STR_LENGTH 64
5894
5895/* Create a unique string id for a slab cache:
5896 *
5897 * Format	:[flags-]size
5898 */
5899static char *create_unique_id(struct kmem_cache *s)
5900{
5901	char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5902	char *p = name;
5903
5904	if (!name)
5905		return ERR_PTR(-ENOMEM);
5906
5907	*p++ = ':';
5908	/*
5909	 * First flags affecting slabcache operations. We will only
5910	 * get here for aliasable slabs so we do not need to support
5911	 * too many flags. The flags here must cover all flags that
5912	 * are matched during merging to guarantee that the id is
5913	 * unique.
5914	 */
5915	if (s->flags & SLAB_CACHE_DMA)
5916		*p++ = 'd';
5917	if (s->flags & SLAB_CACHE_DMA32)
5918		*p++ = 'D';
5919	if (s->flags & SLAB_RECLAIM_ACCOUNT)
5920		*p++ = 'a';
5921	if (s->flags & SLAB_CONSISTENCY_CHECKS)
5922		*p++ = 'F';
5923	if (s->flags & SLAB_ACCOUNT)
5924		*p++ = 'A';
5925	if (p != name + 1)
5926		*p++ = '-';
5927	p += sprintf(p, "%07u", s->size);
5928
5929	BUG_ON(p > name + ID_STR_LENGTH - 1);
5930	return name;
5931}
5932
5933static int sysfs_slab_add(struct kmem_cache *s)
5934{
5935	int err;
5936	const char *name;
5937	struct kset *kset = cache_kset(s);
5938	int unmergeable = slab_unmergeable(s);
5939
5940	if (!kset) {
5941		kobject_init(&s->kobj, &slab_ktype);
5942		return 0;
5943	}
5944
5945	if (!unmergeable && disable_higher_order_debug &&
5946			(slub_debug & DEBUG_METADATA_FLAGS))
5947		unmergeable = 1;
5948
5949	if (unmergeable) {
5950		/*
5951		 * Slabcache can never be merged so we can use the name proper.
5952		 * This is typically the case for debug situations. In that
5953		 * case we can catch duplicate names easily.
5954		 */
5955		sysfs_remove_link(&slab_kset->kobj, s->name);
5956		name = s->name;
5957	} else {
5958		/*
5959		 * Create a unique name for the slab as a target
5960		 * for the symlinks.
5961		 */
5962		name = create_unique_id(s);
5963		if (IS_ERR(name))
5964			return PTR_ERR(name);
5965	}
5966
5967	s->kobj.kset = kset;
5968	err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5969	if (err)
5970		goto out;
5971
5972	err = sysfs_create_group(&s->kobj, &slab_attr_group);
5973	if (err)
5974		goto out_del_kobj;
5975
5976	if (!unmergeable) {
5977		/* Setup first alias */
5978		sysfs_slab_alias(s, s->name);
5979	}
5980out:
5981	if (!unmergeable)
5982		kfree(name);
5983	return err;
5984out_del_kobj:
5985	kobject_del(&s->kobj);
5986	goto out;
5987}
5988
5989void sysfs_slab_unlink(struct kmem_cache *s)
5990{
5991	if (slab_state >= FULL)
5992		kobject_del(&s->kobj);
5993}
5994
5995void sysfs_slab_release(struct kmem_cache *s)
5996{
5997	if (slab_state >= FULL)
5998		kobject_put(&s->kobj);
5999}
6000
6001/*
6002 * Need to buffer aliases during bootup until sysfs becomes
6003 * available lest we lose that information.
6004 */
6005struct saved_alias {
6006	struct kmem_cache *s;
6007	const char *name;
6008	struct saved_alias *next;
6009};
6010
6011static struct saved_alias *alias_list;
6012
6013static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6014{
6015	struct saved_alias *al;
6016
6017	if (slab_state == FULL) {
6018		/*
6019		 * If we have a leftover link then remove it.
6020		 */
6021		sysfs_remove_link(&slab_kset->kobj, name);
6022		return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6023	}
6024
6025	al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6026	if (!al)
6027		return -ENOMEM;
6028
6029	al->s = s;
6030	al->name = name;
6031	al->next = alias_list;
6032	alias_list = al;
6033	return 0;
6034}
6035
6036static int __init slab_sysfs_init(void)
6037{
6038	struct kmem_cache *s;
6039	int err;
6040
6041	mutex_lock(&slab_mutex);
6042
6043	slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6044	if (!slab_kset) {
6045		mutex_unlock(&slab_mutex);
6046		pr_err("Cannot register slab subsystem.\n");
6047		return -ENOSYS;
6048	}
6049
6050	slab_state = FULL;
6051
6052	list_for_each_entry(s, &slab_caches, list) {
6053		err = sysfs_slab_add(s);
6054		if (err)
6055			pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6056			       s->name);
6057	}
6058
6059	while (alias_list) {
6060		struct saved_alias *al = alias_list;
6061
6062		alias_list = alias_list->next;
6063		err = sysfs_slab_alias(al->s, al->name);
6064		if (err)
6065			pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6066			       al->name);
6067		kfree(al);
6068	}
6069
6070	mutex_unlock(&slab_mutex);
6071	return 0;
6072}
6073
6074__initcall(slab_sysfs_init);
6075#endif /* CONFIG_SYSFS */
6076
6077#if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6078static int slab_debugfs_show(struct seq_file *seq, void *v)
6079{
6080	struct loc_track *t = seq->private;
6081	struct location *l;
6082	unsigned long idx;
6083
6084	idx = (unsigned long) t->idx;
6085	if (idx < t->count) {
6086		l = &t->loc[idx];
6087
6088		seq_printf(seq, "%7ld ", l->count);
6089
6090		if (l->addr)
6091			seq_printf(seq, "%pS", (void *)l->addr);
6092		else
6093			seq_puts(seq, "<not-available>");
6094
6095		if (l->sum_time != l->min_time) {
6096			seq_printf(seq, " age=%ld/%llu/%ld",
6097				l->min_time, div_u64(l->sum_time, l->count),
6098				l->max_time);
6099		} else
6100			seq_printf(seq, " age=%ld", l->min_time);
6101
6102		if (l->min_pid != l->max_pid)
6103			seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6104		else
6105			seq_printf(seq, " pid=%ld",
6106				l->min_pid);
6107
6108		if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6109			seq_printf(seq, " cpus=%*pbl",
6110				 cpumask_pr_args(to_cpumask(l->cpus)));
6111
6112		if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6113			seq_printf(seq, " nodes=%*pbl",
6114				 nodemask_pr_args(&l->nodes));
6115
6116#ifdef CONFIG_STACKDEPOT
6117		{
6118			depot_stack_handle_t handle;
6119			unsigned long *entries;
6120			unsigned int nr_entries, j;
6121
6122			handle = READ_ONCE(l->handle);
6123			if (handle) {
6124				nr_entries = stack_depot_fetch(handle, &entries);
6125				seq_puts(seq, "\n");
6126				for (j = 0; j < nr_entries; j++)
6127					seq_printf(seq, "        %pS\n", (void *)entries[j]);
6128			}
6129		}
6130#endif
6131		seq_puts(seq, "\n");
6132	}
6133
6134	if (!idx && !t->count)
6135		seq_puts(seq, "No data\n");
6136
6137	return 0;
6138}
6139
6140static void slab_debugfs_stop(struct seq_file *seq, void *v)
6141{
6142}
6143
6144static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6145{
6146	struct loc_track *t = seq->private;
6147
6148	t->idx = ++(*ppos);
6149	if (*ppos <= t->count)
6150		return ppos;
6151
6152	return NULL;
6153}
6154
6155static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6156{
6157	struct location *loc1 = (struct location *)a;
6158	struct location *loc2 = (struct location *)b;
6159
6160	if (loc1->count > loc2->count)
6161		return -1;
6162	else
6163		return 1;
6164}
6165
6166static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6167{
6168	struct loc_track *t = seq->private;
6169
6170	t->idx = *ppos;
6171	return ppos;
6172}
6173
6174static const struct seq_operations slab_debugfs_sops = {
6175	.start  = slab_debugfs_start,
6176	.next   = slab_debugfs_next,
6177	.stop   = slab_debugfs_stop,
6178	.show   = slab_debugfs_show,
6179};
6180
6181static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6182{
6183
6184	struct kmem_cache_node *n;
6185	enum track_item alloc;
6186	int node;
6187	struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6188						sizeof(struct loc_track));
6189	struct kmem_cache *s = file_inode(filep)->i_private;
6190	unsigned long *obj_map;
6191
6192	if (!t)
6193		return -ENOMEM;
6194
6195	obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6196	if (!obj_map) {
6197		seq_release_private(inode, filep);
6198		return -ENOMEM;
6199	}
6200
6201	if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6202		alloc = TRACK_ALLOC;
6203	else
6204		alloc = TRACK_FREE;
6205
6206	if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6207		bitmap_free(obj_map);
6208		seq_release_private(inode, filep);
6209		return -ENOMEM;
6210	}
6211
6212	for_each_kmem_cache_node(s, node, n) {
6213		unsigned long flags;
6214		struct slab *slab;
6215
6216		if (!atomic_long_read(&n->nr_slabs))
6217			continue;
6218
6219		spin_lock_irqsave(&n->list_lock, flags);
6220		list_for_each_entry(slab, &n->partial, slab_list)
6221			process_slab(t, s, slab, alloc, obj_map);
6222		list_for_each_entry(slab, &n->full, slab_list)
6223			process_slab(t, s, slab, alloc, obj_map);
6224		spin_unlock_irqrestore(&n->list_lock, flags);
6225	}
6226
6227	/* Sort locations by count */
6228	sort_r(t->loc, t->count, sizeof(struct location),
6229		cmp_loc_by_count, NULL, NULL);
6230
6231	bitmap_free(obj_map);
6232	return 0;
6233}
6234
6235static int slab_debug_trace_release(struct inode *inode, struct file *file)
6236{
6237	struct seq_file *seq = file->private_data;
6238	struct loc_track *t = seq->private;
6239
6240	free_loc_track(t);
6241	return seq_release_private(inode, file);
6242}
6243
6244static const struct file_operations slab_debugfs_fops = {
6245	.open    = slab_debug_trace_open,
6246	.read    = seq_read,
6247	.llseek  = seq_lseek,
6248	.release = slab_debug_trace_release,
6249};
6250
6251static void debugfs_slab_add(struct kmem_cache *s)
6252{
6253	struct dentry *slab_cache_dir;
6254
6255	if (unlikely(!slab_debugfs_root))
6256		return;
6257
6258	slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6259
6260	debugfs_create_file("alloc_traces", 0400,
6261		slab_cache_dir, s, &slab_debugfs_fops);
6262
6263	debugfs_create_file("free_traces", 0400,
6264		slab_cache_dir, s, &slab_debugfs_fops);
6265}
6266
6267void debugfs_slab_release(struct kmem_cache *s)
6268{
6269	debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6270}
6271
6272static int __init slab_debugfs_init(void)
6273{
6274	struct kmem_cache *s;
6275
6276	slab_debugfs_root = debugfs_create_dir("slab", NULL);
6277
6278	list_for_each_entry(s, &slab_caches, list)
6279		if (s->flags & SLAB_STORE_USER)
6280			debugfs_slab_add(s);
6281
6282	return 0;
6283
6284}
6285__initcall(slab_debugfs_init);
6286#endif
6287/*
6288 * The /proc/slabinfo ABI
6289 */
6290#ifdef CONFIG_SLUB_DEBUG
6291void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6292{
6293	unsigned long nr_slabs = 0;
6294	unsigned long nr_objs = 0;
6295	unsigned long nr_free = 0;
6296	int node;
6297	struct kmem_cache_node *n;
6298
6299	for_each_kmem_cache_node(s, node, n) {
6300		nr_slabs += node_nr_slabs(n);
6301		nr_objs += node_nr_objs(n);
6302		nr_free += count_partial(n, count_free);
6303	}
6304
6305	sinfo->active_objs = nr_objs - nr_free;
6306	sinfo->num_objs = nr_objs;
6307	sinfo->active_slabs = nr_slabs;
6308	sinfo->num_slabs = nr_slabs;
6309	sinfo->objects_per_slab = oo_objects(s->oo);
6310	sinfo->cache_order = oo_order(s->oo);
6311}
6312
6313void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6314{
6315}
6316
6317ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6318		       size_t count, loff_t *ppos)
6319{
6320	return -EIO;
6321}
6322#endif /* CONFIG_SLUB_DEBUG */
6323