Loading...
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/kmsan.h>
26#include <linux/cpu.h>
27#include <linux/cpuset.h>
28#include <linux/mempolicy.h>
29#include <linux/ctype.h>
30#include <linux/stackdepot.h>
31#include <linux/debugobjects.h>
32#include <linux/kallsyms.h>
33#include <linux/kfence.h>
34#include <linux/memory.h>
35#include <linux/math64.h>
36#include <linux/fault-inject.h>
37#include <linux/stacktrace.h>
38#include <linux/prefetch.h>
39#include <linux/memcontrol.h>
40#include <linux/random.h>
41#include <kunit/test.h>
42#include <kunit/test-bug.h>
43#include <linux/sort.h>
44
45#include <linux/debugfs.h>
46#include <trace/events/kmem.h>
47
48#include "internal.h"
49
50/*
51 * Lock order:
52 * 1. slab_mutex (Global Mutex)
53 * 2. node->list_lock (Spinlock)
54 * 3. kmem_cache->cpu_slab->lock (Local lock)
55 * 4. slab_lock(slab) (Only on some arches)
56 * 5. object_map_lock (Only for debugging)
57 *
58 * slab_mutex
59 *
60 * The role of the slab_mutex is to protect the list of all the slabs
61 * and to synchronize major metadata changes to slab cache structures.
62 * Also synchronizes memory hotplug callbacks.
63 *
64 * slab_lock
65 *
66 * The slab_lock is a wrapper around the page lock, thus it is a bit
67 * spinlock.
68 *
69 * The slab_lock is only used on arches that do not have the ability
70 * to do a cmpxchg_double. It only protects:
71 *
72 * A. slab->freelist -> List of free objects in a slab
73 * B. slab->inuse -> Number of objects in use
74 * C. slab->objects -> Number of objects in slab
75 * D. slab->frozen -> frozen state
76 *
77 * Frozen slabs
78 *
79 * If a slab is frozen then it is exempt from list management. It is not
80 * on any list except per cpu partial list. The processor that froze the
81 * slab is the one who can perform list operations on the slab. Other
82 * processors may put objects onto the freelist but the processor that
83 * froze the slab is the only one that can retrieve the objects from the
84 * slab's freelist.
85 *
86 * list_lock
87 *
88 * The list_lock protects the partial and full list on each node and
89 * the partial slab counter. If taken then no new slabs may be added or
90 * removed from the lists nor make the number of partial slabs be modified.
91 * (Note that the total number of slabs is an atomic value that may be
92 * modified without taking the list lock).
93 *
94 * The list_lock is a centralized lock and thus we avoid taking it as
95 * much as possible. As long as SLUB does not have to handle partial
96 * slabs, operations can continue without any centralized lock. F.e.
97 * allocating a long series of objects that fill up slabs does not require
98 * the list lock.
99 *
100 * For debug caches, all allocations are forced to go through a list_lock
101 * protected region to serialize against concurrent validation.
102 *
103 * cpu_slab->lock local lock
104 *
105 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
106 * except the stat counters. This is a percpu structure manipulated only by
107 * the local cpu, so the lock protects against being preempted or interrupted
108 * by an irq. Fast path operations rely on lockless operations instead.
109 *
110 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
111 * which means the lockless fastpath cannot be used as it might interfere with
112 * an in-progress slow path operations. In this case the local lock is always
113 * taken but it still utilizes the freelist for the common operations.
114 *
115 * lockless fastpaths
116 *
117 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
118 * are fully lockless when satisfied from the percpu slab (and when
119 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
120 * They also don't disable preemption or migration or irqs. They rely on
121 * the transaction id (tid) field to detect being preempted or moved to
122 * another cpu.
123 *
124 * irq, preemption, migration considerations
125 *
126 * Interrupts are disabled as part of list_lock or local_lock operations, or
127 * around the slab_lock operation, in order to make the slab allocator safe
128 * to use in the context of an irq.
129 *
130 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
131 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
132 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
133 * doesn't have to be revalidated in each section protected by the local lock.
134 *
135 * SLUB assigns one slab for allocation to each processor.
136 * Allocations only occur from these slabs called cpu slabs.
137 *
138 * Slabs with free elements are kept on a partial list and during regular
139 * operations no list for full slabs is used. If an object in a full slab is
140 * freed then the slab will show up again on the partial lists.
141 * We track full slabs for debugging purposes though because otherwise we
142 * cannot scan all objects.
143 *
144 * Slabs are freed when they become empty. Teardown and setup is
145 * minimal so we rely on the page allocators per cpu caches for
146 * fast frees and allocs.
147 *
148 * slab->frozen The slab is frozen and exempt from list processing.
149 * This means that the slab is dedicated to a purpose
150 * such as satisfying allocations for a specific
151 * processor. Objects may be freed in the slab while
152 * it is frozen but slab_free will then skip the usual
153 * list operations. It is up to the processor holding
154 * the slab to integrate the slab into the slab lists
155 * when the slab is no longer needed.
156 *
157 * One use of this flag is to mark slabs that are
158 * used for allocations. Then such a slab becomes a cpu
159 * slab. The cpu slab may be equipped with an additional
160 * freelist that allows lockless access to
161 * free objects in addition to the regular freelist
162 * that requires the slab lock.
163 *
164 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
165 * options set. This moves slab handling out of
166 * the fast path and disables lockless freelists.
167 */
168
169/*
170 * We could simply use migrate_disable()/enable() but as long as it's a
171 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
172 */
173#ifndef CONFIG_PREEMPT_RT
174#define slub_get_cpu_ptr(var) get_cpu_ptr(var)
175#define slub_put_cpu_ptr(var) put_cpu_ptr(var)
176#define USE_LOCKLESS_FAST_PATH() (true)
177#else
178#define slub_get_cpu_ptr(var) \
179({ \
180 migrate_disable(); \
181 this_cpu_ptr(var); \
182})
183#define slub_put_cpu_ptr(var) \
184do { \
185 (void)(var); \
186 migrate_enable(); \
187} while (0)
188#define USE_LOCKLESS_FAST_PATH() (false)
189#endif
190
191#ifndef CONFIG_SLUB_TINY
192#define __fastpath_inline __always_inline
193#else
194#define __fastpath_inline
195#endif
196
197#ifdef CONFIG_SLUB_DEBUG
198#ifdef CONFIG_SLUB_DEBUG_ON
199DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
200#else
201DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
202#endif
203#endif /* CONFIG_SLUB_DEBUG */
204
205/* Structure holding parameters for get_partial() call chain */
206struct partial_context {
207 struct slab **slab;
208 gfp_t flags;
209 unsigned int orig_size;
210};
211
212static inline bool kmem_cache_debug(struct kmem_cache *s)
213{
214 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
215}
216
217static inline bool slub_debug_orig_size(struct kmem_cache *s)
218{
219 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
220 (s->flags & SLAB_KMALLOC));
221}
222
223void *fixup_red_left(struct kmem_cache *s, void *p)
224{
225 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
226 p += s->red_left_pad;
227
228 return p;
229}
230
231static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
232{
233#ifdef CONFIG_SLUB_CPU_PARTIAL
234 return !kmem_cache_debug(s);
235#else
236 return false;
237#endif
238}
239
240/*
241 * Issues still to be resolved:
242 *
243 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
244 *
245 * - Variable sizing of the per node arrays
246 */
247
248/* Enable to log cmpxchg failures */
249#undef SLUB_DEBUG_CMPXCHG
250
251#ifndef CONFIG_SLUB_TINY
252/*
253 * Minimum number of partial slabs. These will be left on the partial
254 * lists even if they are empty. kmem_cache_shrink may reclaim them.
255 */
256#define MIN_PARTIAL 5
257
258/*
259 * Maximum number of desirable partial slabs.
260 * The existence of more partial slabs makes kmem_cache_shrink
261 * sort the partial list by the number of objects in use.
262 */
263#define MAX_PARTIAL 10
264#else
265#define MIN_PARTIAL 0
266#define MAX_PARTIAL 0
267#endif
268
269#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
270 SLAB_POISON | SLAB_STORE_USER)
271
272/*
273 * These debug flags cannot use CMPXCHG because there might be consistency
274 * issues when checking or reading debug information
275 */
276#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
277 SLAB_TRACE)
278
279
280/*
281 * Debugging flags that require metadata to be stored in the slab. These get
282 * disabled when slub_debug=O is used and a cache's min order increases with
283 * metadata.
284 */
285#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
286
287#define OO_SHIFT 16
288#define OO_MASK ((1 << OO_SHIFT) - 1)
289#define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
290
291/* Internal SLUB flags */
292/* Poison object */
293#define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
294/* Use cmpxchg_double */
295#define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
296
297/*
298 * Tracking user of a slab.
299 */
300#define TRACK_ADDRS_COUNT 16
301struct track {
302 unsigned long addr; /* Called from address */
303#ifdef CONFIG_STACKDEPOT
304 depot_stack_handle_t handle;
305#endif
306 int cpu; /* Was running on cpu */
307 int pid; /* Pid context */
308 unsigned long when; /* When did the operation occur */
309};
310
311enum track_item { TRACK_ALLOC, TRACK_FREE };
312
313#ifdef SLAB_SUPPORTS_SYSFS
314static int sysfs_slab_add(struct kmem_cache *);
315static int sysfs_slab_alias(struct kmem_cache *, const char *);
316#else
317static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
318static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
319 { return 0; }
320#endif
321
322#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
323static void debugfs_slab_add(struct kmem_cache *);
324#else
325static inline void debugfs_slab_add(struct kmem_cache *s) { }
326#endif
327
328static inline void stat(const struct kmem_cache *s, enum stat_item si)
329{
330#ifdef CONFIG_SLUB_STATS
331 /*
332 * The rmw is racy on a preemptible kernel but this is acceptable, so
333 * avoid this_cpu_add()'s irq-disable overhead.
334 */
335 raw_cpu_inc(s->cpu_slab->stat[si]);
336#endif
337}
338
339/*
340 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
341 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
342 * differ during memory hotplug/hotremove operations.
343 * Protected by slab_mutex.
344 */
345static nodemask_t slab_nodes;
346
347#ifndef CONFIG_SLUB_TINY
348/*
349 * Workqueue used for flush_cpu_slab().
350 */
351static struct workqueue_struct *flushwq;
352#endif
353
354/********************************************************************
355 * Core slab cache functions
356 *******************************************************************/
357
358/*
359 * Returns freelist pointer (ptr). With hardening, this is obfuscated
360 * with an XOR of the address where the pointer is held and a per-cache
361 * random number.
362 */
363static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
364 unsigned long ptr_addr)
365{
366#ifdef CONFIG_SLAB_FREELIST_HARDENED
367 /*
368 * When CONFIG_KASAN_SW/HW_TAGS is enabled, ptr_addr might be tagged.
369 * Normally, this doesn't cause any issues, as both set_freepointer()
370 * and get_freepointer() are called with a pointer with the same tag.
371 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
372 * example, when __free_slub() iterates over objects in a cache, it
373 * passes untagged pointers to check_object(). check_object() in turns
374 * calls get_freepointer() with an untagged pointer, which causes the
375 * freepointer to be restored incorrectly.
376 */
377 return (void *)((unsigned long)ptr ^ s->random ^
378 swab((unsigned long)kasan_reset_tag((void *)ptr_addr)));
379#else
380 return ptr;
381#endif
382}
383
384/* Returns the freelist pointer recorded at location ptr_addr. */
385static inline void *freelist_dereference(const struct kmem_cache *s,
386 void *ptr_addr)
387{
388 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
389 (unsigned long)ptr_addr);
390}
391
392static inline void *get_freepointer(struct kmem_cache *s, void *object)
393{
394 object = kasan_reset_tag(object);
395 return freelist_dereference(s, object + s->offset);
396}
397
398#ifndef CONFIG_SLUB_TINY
399static void prefetch_freepointer(const struct kmem_cache *s, void *object)
400{
401 prefetchw(object + s->offset);
402}
403#endif
404
405/*
406 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
407 * pointer value in the case the current thread loses the race for the next
408 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
409 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
410 * KMSAN will still check all arguments of cmpxchg because of imperfect
411 * handling of inline assembly.
412 * To work around this problem, we apply __no_kmsan_checks to ensure that
413 * get_freepointer_safe() returns initialized memory.
414 */
415__no_kmsan_checks
416static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
417{
418 unsigned long freepointer_addr;
419 void *p;
420
421 if (!debug_pagealloc_enabled_static())
422 return get_freepointer(s, object);
423
424 object = kasan_reset_tag(object);
425 freepointer_addr = (unsigned long)object + s->offset;
426 copy_from_kernel_nofault(&p, (void **)freepointer_addr, sizeof(p));
427 return freelist_ptr(s, p, freepointer_addr);
428}
429
430static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
431{
432 unsigned long freeptr_addr = (unsigned long)object + s->offset;
433
434#ifdef CONFIG_SLAB_FREELIST_HARDENED
435 BUG_ON(object == fp); /* naive detection of double free or corruption */
436#endif
437
438 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
439 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
440}
441
442/* Loop over all objects in a slab */
443#define for_each_object(__p, __s, __addr, __objects) \
444 for (__p = fixup_red_left(__s, __addr); \
445 __p < (__addr) + (__objects) * (__s)->size; \
446 __p += (__s)->size)
447
448static inline unsigned int order_objects(unsigned int order, unsigned int size)
449{
450 return ((unsigned int)PAGE_SIZE << order) / size;
451}
452
453static inline struct kmem_cache_order_objects oo_make(unsigned int order,
454 unsigned int size)
455{
456 struct kmem_cache_order_objects x = {
457 (order << OO_SHIFT) + order_objects(order, size)
458 };
459
460 return x;
461}
462
463static inline unsigned int oo_order(struct kmem_cache_order_objects x)
464{
465 return x.x >> OO_SHIFT;
466}
467
468static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
469{
470 return x.x & OO_MASK;
471}
472
473#ifdef CONFIG_SLUB_CPU_PARTIAL
474static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
475{
476 unsigned int nr_slabs;
477
478 s->cpu_partial = nr_objects;
479
480 /*
481 * We take the number of objects but actually limit the number of
482 * slabs on the per cpu partial list, in order to limit excessive
483 * growth of the list. For simplicity we assume that the slabs will
484 * be half-full.
485 */
486 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
487 s->cpu_partial_slabs = nr_slabs;
488}
489#else
490static inline void
491slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
492{
493}
494#endif /* CONFIG_SLUB_CPU_PARTIAL */
495
496/*
497 * Per slab locking using the pagelock
498 */
499static __always_inline void slab_lock(struct slab *slab)
500{
501 struct page *page = slab_page(slab);
502
503 VM_BUG_ON_PAGE(PageTail(page), page);
504 bit_spin_lock(PG_locked, &page->flags);
505}
506
507static __always_inline void slab_unlock(struct slab *slab)
508{
509 struct page *page = slab_page(slab);
510
511 VM_BUG_ON_PAGE(PageTail(page), page);
512 __bit_spin_unlock(PG_locked, &page->flags);
513}
514
515/*
516 * Interrupts must be disabled (for the fallback code to work right), typically
517 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
518 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
519 * allocation/ free operation in hardirq context. Therefore nothing can
520 * interrupt the operation.
521 */
522static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
523 void *freelist_old, unsigned long counters_old,
524 void *freelist_new, unsigned long counters_new,
525 const char *n)
526{
527 if (USE_LOCKLESS_FAST_PATH())
528 lockdep_assert_irqs_disabled();
529#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
530 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
531 if (s->flags & __CMPXCHG_DOUBLE) {
532 if (cmpxchg_double(&slab->freelist, &slab->counters,
533 freelist_old, counters_old,
534 freelist_new, counters_new))
535 return true;
536 } else
537#endif
538 {
539 slab_lock(slab);
540 if (slab->freelist == freelist_old &&
541 slab->counters == counters_old) {
542 slab->freelist = freelist_new;
543 slab->counters = counters_new;
544 slab_unlock(slab);
545 return true;
546 }
547 slab_unlock(slab);
548 }
549
550 cpu_relax();
551 stat(s, CMPXCHG_DOUBLE_FAIL);
552
553#ifdef SLUB_DEBUG_CMPXCHG
554 pr_info("%s %s: cmpxchg double redo ", n, s->name);
555#endif
556
557 return false;
558}
559
560static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct slab *slab,
561 void *freelist_old, unsigned long counters_old,
562 void *freelist_new, unsigned long counters_new,
563 const char *n)
564{
565#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
566 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
567 if (s->flags & __CMPXCHG_DOUBLE) {
568 if (cmpxchg_double(&slab->freelist, &slab->counters,
569 freelist_old, counters_old,
570 freelist_new, counters_new))
571 return true;
572 } else
573#endif
574 {
575 unsigned long flags;
576
577 local_irq_save(flags);
578 slab_lock(slab);
579 if (slab->freelist == freelist_old &&
580 slab->counters == counters_old) {
581 slab->freelist = freelist_new;
582 slab->counters = counters_new;
583 slab_unlock(slab);
584 local_irq_restore(flags);
585 return true;
586 }
587 slab_unlock(slab);
588 local_irq_restore(flags);
589 }
590
591 cpu_relax();
592 stat(s, CMPXCHG_DOUBLE_FAIL);
593
594#ifdef SLUB_DEBUG_CMPXCHG
595 pr_info("%s %s: cmpxchg double redo ", n, s->name);
596#endif
597
598 return false;
599}
600
601#ifdef CONFIG_SLUB_DEBUG
602static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
603static DEFINE_SPINLOCK(object_map_lock);
604
605static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
606 struct slab *slab)
607{
608 void *addr = slab_address(slab);
609 void *p;
610
611 bitmap_zero(obj_map, slab->objects);
612
613 for (p = slab->freelist; p; p = get_freepointer(s, p))
614 set_bit(__obj_to_index(s, addr, p), obj_map);
615}
616
617#if IS_ENABLED(CONFIG_KUNIT)
618static bool slab_add_kunit_errors(void)
619{
620 struct kunit_resource *resource;
621
622 if (!kunit_get_current_test())
623 return false;
624
625 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
626 if (!resource)
627 return false;
628
629 (*(int *)resource->data)++;
630 kunit_put_resource(resource);
631 return true;
632}
633#else
634static inline bool slab_add_kunit_errors(void) { return false; }
635#endif
636
637static inline unsigned int size_from_object(struct kmem_cache *s)
638{
639 if (s->flags & SLAB_RED_ZONE)
640 return s->size - s->red_left_pad;
641
642 return s->size;
643}
644
645static inline void *restore_red_left(struct kmem_cache *s, void *p)
646{
647 if (s->flags & SLAB_RED_ZONE)
648 p -= s->red_left_pad;
649
650 return p;
651}
652
653/*
654 * Debug settings:
655 */
656#if defined(CONFIG_SLUB_DEBUG_ON)
657static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
658#else
659static slab_flags_t slub_debug;
660#endif
661
662static char *slub_debug_string;
663static int disable_higher_order_debug;
664
665/*
666 * slub is about to manipulate internal object metadata. This memory lies
667 * outside the range of the allocated object, so accessing it would normally
668 * be reported by kasan as a bounds error. metadata_access_enable() is used
669 * to tell kasan that these accesses are OK.
670 */
671static inline void metadata_access_enable(void)
672{
673 kasan_disable_current();
674}
675
676static inline void metadata_access_disable(void)
677{
678 kasan_enable_current();
679}
680
681/*
682 * Object debugging
683 */
684
685/* Verify that a pointer has an address that is valid within a slab page */
686static inline int check_valid_pointer(struct kmem_cache *s,
687 struct slab *slab, void *object)
688{
689 void *base;
690
691 if (!object)
692 return 1;
693
694 base = slab_address(slab);
695 object = kasan_reset_tag(object);
696 object = restore_red_left(s, object);
697 if (object < base || object >= base + slab->objects * s->size ||
698 (object - base) % s->size) {
699 return 0;
700 }
701
702 return 1;
703}
704
705static void print_section(char *level, char *text, u8 *addr,
706 unsigned int length)
707{
708 metadata_access_enable();
709 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
710 16, 1, kasan_reset_tag((void *)addr), length, 1);
711 metadata_access_disable();
712}
713
714/*
715 * See comment in calculate_sizes().
716 */
717static inline bool freeptr_outside_object(struct kmem_cache *s)
718{
719 return s->offset >= s->inuse;
720}
721
722/*
723 * Return offset of the end of info block which is inuse + free pointer if
724 * not overlapping with object.
725 */
726static inline unsigned int get_info_end(struct kmem_cache *s)
727{
728 if (freeptr_outside_object(s))
729 return s->inuse + sizeof(void *);
730 else
731 return s->inuse;
732}
733
734static struct track *get_track(struct kmem_cache *s, void *object,
735 enum track_item alloc)
736{
737 struct track *p;
738
739 p = object + get_info_end(s);
740
741 return kasan_reset_tag(p + alloc);
742}
743
744#ifdef CONFIG_STACKDEPOT
745static noinline depot_stack_handle_t set_track_prepare(void)
746{
747 depot_stack_handle_t handle;
748 unsigned long entries[TRACK_ADDRS_COUNT];
749 unsigned int nr_entries;
750
751 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
752 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
753
754 return handle;
755}
756#else
757static inline depot_stack_handle_t set_track_prepare(void)
758{
759 return 0;
760}
761#endif
762
763static void set_track_update(struct kmem_cache *s, void *object,
764 enum track_item alloc, unsigned long addr,
765 depot_stack_handle_t handle)
766{
767 struct track *p = get_track(s, object, alloc);
768
769#ifdef CONFIG_STACKDEPOT
770 p->handle = handle;
771#endif
772 p->addr = addr;
773 p->cpu = smp_processor_id();
774 p->pid = current->pid;
775 p->when = jiffies;
776}
777
778static __always_inline void set_track(struct kmem_cache *s, void *object,
779 enum track_item alloc, unsigned long addr)
780{
781 depot_stack_handle_t handle = set_track_prepare();
782
783 set_track_update(s, object, alloc, addr, handle);
784}
785
786static void init_tracking(struct kmem_cache *s, void *object)
787{
788 struct track *p;
789
790 if (!(s->flags & SLAB_STORE_USER))
791 return;
792
793 p = get_track(s, object, TRACK_ALLOC);
794 memset(p, 0, 2*sizeof(struct track));
795}
796
797static void print_track(const char *s, struct track *t, unsigned long pr_time)
798{
799 depot_stack_handle_t handle __maybe_unused;
800
801 if (!t->addr)
802 return;
803
804 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
805 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
806#ifdef CONFIG_STACKDEPOT
807 handle = READ_ONCE(t->handle);
808 if (handle)
809 stack_depot_print(handle);
810 else
811 pr_err("object allocation/free stack trace missing\n");
812#endif
813}
814
815void print_tracking(struct kmem_cache *s, void *object)
816{
817 unsigned long pr_time = jiffies;
818 if (!(s->flags & SLAB_STORE_USER))
819 return;
820
821 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
822 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
823}
824
825static void print_slab_info(const struct slab *slab)
826{
827 struct folio *folio = (struct folio *)slab_folio(slab);
828
829 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
830 slab, slab->objects, slab->inuse, slab->freelist,
831 folio_flags(folio, 0));
832}
833
834/*
835 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
836 * family will round up the real request size to these fixed ones, so
837 * there could be an extra area than what is requested. Save the original
838 * request size in the meta data area, for better debug and sanity check.
839 */
840static inline void set_orig_size(struct kmem_cache *s,
841 void *object, unsigned int orig_size)
842{
843 void *p = kasan_reset_tag(object);
844
845 if (!slub_debug_orig_size(s))
846 return;
847
848#ifdef CONFIG_KASAN_GENERIC
849 /*
850 * KASAN could save its free meta data in object's data area at
851 * offset 0, if the size is larger than 'orig_size', it will
852 * overlap the data redzone in [orig_size+1, object_size], and
853 * the check should be skipped.
854 */
855 if (kasan_metadata_size(s, true) > orig_size)
856 orig_size = s->object_size;
857#endif
858
859 p += get_info_end(s);
860 p += sizeof(struct track) * 2;
861
862 *(unsigned int *)p = orig_size;
863}
864
865static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
866{
867 void *p = kasan_reset_tag(object);
868
869 if (!slub_debug_orig_size(s))
870 return s->object_size;
871
872 p += get_info_end(s);
873 p += sizeof(struct track) * 2;
874
875 return *(unsigned int *)p;
876}
877
878void skip_orig_size_check(struct kmem_cache *s, const void *object)
879{
880 set_orig_size(s, (void *)object, s->object_size);
881}
882
883static void slab_bug(struct kmem_cache *s, char *fmt, ...)
884{
885 struct va_format vaf;
886 va_list args;
887
888 va_start(args, fmt);
889 vaf.fmt = fmt;
890 vaf.va = &args;
891 pr_err("=============================================================================\n");
892 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
893 pr_err("-----------------------------------------------------------------------------\n\n");
894 va_end(args);
895}
896
897__printf(2, 3)
898static void slab_fix(struct kmem_cache *s, char *fmt, ...)
899{
900 struct va_format vaf;
901 va_list args;
902
903 if (slab_add_kunit_errors())
904 return;
905
906 va_start(args, fmt);
907 vaf.fmt = fmt;
908 vaf.va = &args;
909 pr_err("FIX %s: %pV\n", s->name, &vaf);
910 va_end(args);
911}
912
913static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
914{
915 unsigned int off; /* Offset of last byte */
916 u8 *addr = slab_address(slab);
917
918 print_tracking(s, p);
919
920 print_slab_info(slab);
921
922 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
923 p, p - addr, get_freepointer(s, p));
924
925 if (s->flags & SLAB_RED_ZONE)
926 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
927 s->red_left_pad);
928 else if (p > addr + 16)
929 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
930
931 print_section(KERN_ERR, "Object ", p,
932 min_t(unsigned int, s->object_size, PAGE_SIZE));
933 if (s->flags & SLAB_RED_ZONE)
934 print_section(KERN_ERR, "Redzone ", p + s->object_size,
935 s->inuse - s->object_size);
936
937 off = get_info_end(s);
938
939 if (s->flags & SLAB_STORE_USER)
940 off += 2 * sizeof(struct track);
941
942 if (slub_debug_orig_size(s))
943 off += sizeof(unsigned int);
944
945 off += kasan_metadata_size(s, false);
946
947 if (off != size_from_object(s))
948 /* Beginning of the filler is the free pointer */
949 print_section(KERN_ERR, "Padding ", p + off,
950 size_from_object(s) - off);
951
952 dump_stack();
953}
954
955static void object_err(struct kmem_cache *s, struct slab *slab,
956 u8 *object, char *reason)
957{
958 if (slab_add_kunit_errors())
959 return;
960
961 slab_bug(s, "%s", reason);
962 print_trailer(s, slab, object);
963 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
964}
965
966static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
967 void **freelist, void *nextfree)
968{
969 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
970 !check_valid_pointer(s, slab, nextfree) && freelist) {
971 object_err(s, slab, *freelist, "Freechain corrupt");
972 *freelist = NULL;
973 slab_fix(s, "Isolate corrupted freechain");
974 return true;
975 }
976
977 return false;
978}
979
980static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
981 const char *fmt, ...)
982{
983 va_list args;
984 char buf[100];
985
986 if (slab_add_kunit_errors())
987 return;
988
989 va_start(args, fmt);
990 vsnprintf(buf, sizeof(buf), fmt, args);
991 va_end(args);
992 slab_bug(s, "%s", buf);
993 print_slab_info(slab);
994 dump_stack();
995 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
996}
997
998static void init_object(struct kmem_cache *s, void *object, u8 val)
999{
1000 u8 *p = kasan_reset_tag(object);
1001 unsigned int poison_size = s->object_size;
1002
1003 if (s->flags & SLAB_RED_ZONE) {
1004 memset(p - s->red_left_pad, val, s->red_left_pad);
1005
1006 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1007 /*
1008 * Redzone the extra allocated space by kmalloc than
1009 * requested, and the poison size will be limited to
1010 * the original request size accordingly.
1011 */
1012 poison_size = get_orig_size(s, object);
1013 }
1014 }
1015
1016 if (s->flags & __OBJECT_POISON) {
1017 memset(p, POISON_FREE, poison_size - 1);
1018 p[poison_size - 1] = POISON_END;
1019 }
1020
1021 if (s->flags & SLAB_RED_ZONE)
1022 memset(p + poison_size, val, s->inuse - poison_size);
1023}
1024
1025static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1026 void *from, void *to)
1027{
1028 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1029 memset(from, data, to - from);
1030}
1031
1032static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1033 u8 *object, char *what,
1034 u8 *start, unsigned int value, unsigned int bytes)
1035{
1036 u8 *fault;
1037 u8 *end;
1038 u8 *addr = slab_address(slab);
1039
1040 metadata_access_enable();
1041 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1042 metadata_access_disable();
1043 if (!fault)
1044 return 1;
1045
1046 end = start + bytes;
1047 while (end > fault && end[-1] == value)
1048 end--;
1049
1050 if (slab_add_kunit_errors())
1051 goto skip_bug_print;
1052
1053 slab_bug(s, "%s overwritten", what);
1054 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1055 fault, end - 1, fault - addr,
1056 fault[0], value);
1057 print_trailer(s, slab, object);
1058 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1059
1060skip_bug_print:
1061 restore_bytes(s, what, value, fault, end);
1062 return 0;
1063}
1064
1065/*
1066 * Object layout:
1067 *
1068 * object address
1069 * Bytes of the object to be managed.
1070 * If the freepointer may overlay the object then the free
1071 * pointer is at the middle of the object.
1072 *
1073 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1074 * 0xa5 (POISON_END)
1075 *
1076 * object + s->object_size
1077 * Padding to reach word boundary. This is also used for Redzoning.
1078 * Padding is extended by another word if Redzoning is enabled and
1079 * object_size == inuse.
1080 *
1081 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1082 * 0xcc (RED_ACTIVE) for objects in use.
1083 *
1084 * object + s->inuse
1085 * Meta data starts here.
1086 *
1087 * A. Free pointer (if we cannot overwrite object on free)
1088 * B. Tracking data for SLAB_STORE_USER
1089 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1090 * D. Padding to reach required alignment boundary or at minimum
1091 * one word if debugging is on to be able to detect writes
1092 * before the word boundary.
1093 *
1094 * Padding is done using 0x5a (POISON_INUSE)
1095 *
1096 * object + s->size
1097 * Nothing is used beyond s->size.
1098 *
1099 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1100 * ignored. And therefore no slab options that rely on these boundaries
1101 * may be used with merged slabcaches.
1102 */
1103
1104static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1105{
1106 unsigned long off = get_info_end(s); /* The end of info */
1107
1108 if (s->flags & SLAB_STORE_USER) {
1109 /* We also have user information there */
1110 off += 2 * sizeof(struct track);
1111
1112 if (s->flags & SLAB_KMALLOC)
1113 off += sizeof(unsigned int);
1114 }
1115
1116 off += kasan_metadata_size(s, false);
1117
1118 if (size_from_object(s) == off)
1119 return 1;
1120
1121 return check_bytes_and_report(s, slab, p, "Object padding",
1122 p + off, POISON_INUSE, size_from_object(s) - off);
1123}
1124
1125/* Check the pad bytes at the end of a slab page */
1126static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1127{
1128 u8 *start;
1129 u8 *fault;
1130 u8 *end;
1131 u8 *pad;
1132 int length;
1133 int remainder;
1134
1135 if (!(s->flags & SLAB_POISON))
1136 return;
1137
1138 start = slab_address(slab);
1139 length = slab_size(slab);
1140 end = start + length;
1141 remainder = length % s->size;
1142 if (!remainder)
1143 return;
1144
1145 pad = end - remainder;
1146 metadata_access_enable();
1147 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1148 metadata_access_disable();
1149 if (!fault)
1150 return;
1151 while (end > fault && end[-1] == POISON_INUSE)
1152 end--;
1153
1154 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1155 fault, end - 1, fault - start);
1156 print_section(KERN_ERR, "Padding ", pad, remainder);
1157
1158 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1159}
1160
1161static int check_object(struct kmem_cache *s, struct slab *slab,
1162 void *object, u8 val)
1163{
1164 u8 *p = object;
1165 u8 *endobject = object + s->object_size;
1166 unsigned int orig_size;
1167
1168 if (s->flags & SLAB_RED_ZONE) {
1169 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1170 object - s->red_left_pad, val, s->red_left_pad))
1171 return 0;
1172
1173 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1174 endobject, val, s->inuse - s->object_size))
1175 return 0;
1176
1177 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1178 orig_size = get_orig_size(s, object);
1179
1180 if (s->object_size > orig_size &&
1181 !check_bytes_and_report(s, slab, object,
1182 "kmalloc Redzone", p + orig_size,
1183 val, s->object_size - orig_size)) {
1184 return 0;
1185 }
1186 }
1187 } else {
1188 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1189 check_bytes_and_report(s, slab, p, "Alignment padding",
1190 endobject, POISON_INUSE,
1191 s->inuse - s->object_size);
1192 }
1193 }
1194
1195 if (s->flags & SLAB_POISON) {
1196 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
1197 (!check_bytes_and_report(s, slab, p, "Poison", p,
1198 POISON_FREE, s->object_size - 1) ||
1199 !check_bytes_and_report(s, slab, p, "End Poison",
1200 p + s->object_size - 1, POISON_END, 1)))
1201 return 0;
1202 /*
1203 * check_pad_bytes cleans up on its own.
1204 */
1205 check_pad_bytes(s, slab, p);
1206 }
1207
1208 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1209 /*
1210 * Object and freepointer overlap. Cannot check
1211 * freepointer while object is allocated.
1212 */
1213 return 1;
1214
1215 /* Check free pointer validity */
1216 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1217 object_err(s, slab, p, "Freepointer corrupt");
1218 /*
1219 * No choice but to zap it and thus lose the remainder
1220 * of the free objects in this slab. May cause
1221 * another error because the object count is now wrong.
1222 */
1223 set_freepointer(s, p, NULL);
1224 return 0;
1225 }
1226 return 1;
1227}
1228
1229static int check_slab(struct kmem_cache *s, struct slab *slab)
1230{
1231 int maxobj;
1232
1233 if (!folio_test_slab(slab_folio(slab))) {
1234 slab_err(s, slab, "Not a valid slab page");
1235 return 0;
1236 }
1237
1238 maxobj = order_objects(slab_order(slab), s->size);
1239 if (slab->objects > maxobj) {
1240 slab_err(s, slab, "objects %u > max %u",
1241 slab->objects, maxobj);
1242 return 0;
1243 }
1244 if (slab->inuse > slab->objects) {
1245 slab_err(s, slab, "inuse %u > max %u",
1246 slab->inuse, slab->objects);
1247 return 0;
1248 }
1249 /* Slab_pad_check fixes things up after itself */
1250 slab_pad_check(s, slab);
1251 return 1;
1252}
1253
1254/*
1255 * Determine if a certain object in a slab is on the freelist. Must hold the
1256 * slab lock to guarantee that the chains are in a consistent state.
1257 */
1258static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1259{
1260 int nr = 0;
1261 void *fp;
1262 void *object = NULL;
1263 int max_objects;
1264
1265 fp = slab->freelist;
1266 while (fp && nr <= slab->objects) {
1267 if (fp == search)
1268 return 1;
1269 if (!check_valid_pointer(s, slab, fp)) {
1270 if (object) {
1271 object_err(s, slab, object,
1272 "Freechain corrupt");
1273 set_freepointer(s, object, NULL);
1274 } else {
1275 slab_err(s, slab, "Freepointer corrupt");
1276 slab->freelist = NULL;
1277 slab->inuse = slab->objects;
1278 slab_fix(s, "Freelist cleared");
1279 return 0;
1280 }
1281 break;
1282 }
1283 object = fp;
1284 fp = get_freepointer(s, object);
1285 nr++;
1286 }
1287
1288 max_objects = order_objects(slab_order(slab), s->size);
1289 if (max_objects > MAX_OBJS_PER_PAGE)
1290 max_objects = MAX_OBJS_PER_PAGE;
1291
1292 if (slab->objects != max_objects) {
1293 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1294 slab->objects, max_objects);
1295 slab->objects = max_objects;
1296 slab_fix(s, "Number of objects adjusted");
1297 }
1298 if (slab->inuse != slab->objects - nr) {
1299 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1300 slab->inuse, slab->objects - nr);
1301 slab->inuse = slab->objects - nr;
1302 slab_fix(s, "Object count adjusted");
1303 }
1304 return search == NULL;
1305}
1306
1307static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1308 int alloc)
1309{
1310 if (s->flags & SLAB_TRACE) {
1311 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1312 s->name,
1313 alloc ? "alloc" : "free",
1314 object, slab->inuse,
1315 slab->freelist);
1316
1317 if (!alloc)
1318 print_section(KERN_INFO, "Object ", (void *)object,
1319 s->object_size);
1320
1321 dump_stack();
1322 }
1323}
1324
1325/*
1326 * Tracking of fully allocated slabs for debugging purposes.
1327 */
1328static void add_full(struct kmem_cache *s,
1329 struct kmem_cache_node *n, struct slab *slab)
1330{
1331 if (!(s->flags & SLAB_STORE_USER))
1332 return;
1333
1334 lockdep_assert_held(&n->list_lock);
1335 list_add(&slab->slab_list, &n->full);
1336}
1337
1338static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1339{
1340 if (!(s->flags & SLAB_STORE_USER))
1341 return;
1342
1343 lockdep_assert_held(&n->list_lock);
1344 list_del(&slab->slab_list);
1345}
1346
1347/* Tracking of the number of slabs for debugging purposes */
1348static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1349{
1350 struct kmem_cache_node *n = get_node(s, node);
1351
1352 return atomic_long_read(&n->nr_slabs);
1353}
1354
1355static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1356{
1357 return atomic_long_read(&n->nr_slabs);
1358}
1359
1360static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1361{
1362 struct kmem_cache_node *n = get_node(s, node);
1363
1364 /*
1365 * May be called early in order to allocate a slab for the
1366 * kmem_cache_node structure. Solve the chicken-egg
1367 * dilemma by deferring the increment of the count during
1368 * bootstrap (see early_kmem_cache_node_alloc).
1369 */
1370 if (likely(n)) {
1371 atomic_long_inc(&n->nr_slabs);
1372 atomic_long_add(objects, &n->total_objects);
1373 }
1374}
1375static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1376{
1377 struct kmem_cache_node *n = get_node(s, node);
1378
1379 atomic_long_dec(&n->nr_slabs);
1380 atomic_long_sub(objects, &n->total_objects);
1381}
1382
1383/* Object debug checks for alloc/free paths */
1384static void setup_object_debug(struct kmem_cache *s, void *object)
1385{
1386 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1387 return;
1388
1389 init_object(s, object, SLUB_RED_INACTIVE);
1390 init_tracking(s, object);
1391}
1392
1393static
1394void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1395{
1396 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1397 return;
1398
1399 metadata_access_enable();
1400 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1401 metadata_access_disable();
1402}
1403
1404static inline int alloc_consistency_checks(struct kmem_cache *s,
1405 struct slab *slab, void *object)
1406{
1407 if (!check_slab(s, slab))
1408 return 0;
1409
1410 if (!check_valid_pointer(s, slab, object)) {
1411 object_err(s, slab, object, "Freelist Pointer check fails");
1412 return 0;
1413 }
1414
1415 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1416 return 0;
1417
1418 return 1;
1419}
1420
1421static noinline bool alloc_debug_processing(struct kmem_cache *s,
1422 struct slab *slab, void *object, int orig_size)
1423{
1424 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1425 if (!alloc_consistency_checks(s, slab, object))
1426 goto bad;
1427 }
1428
1429 /* Success. Perform special debug activities for allocs */
1430 trace(s, slab, object, 1);
1431 set_orig_size(s, object, orig_size);
1432 init_object(s, object, SLUB_RED_ACTIVE);
1433 return true;
1434
1435bad:
1436 if (folio_test_slab(slab_folio(slab))) {
1437 /*
1438 * If this is a slab page then lets do the best we can
1439 * to avoid issues in the future. Marking all objects
1440 * as used avoids touching the remaining objects.
1441 */
1442 slab_fix(s, "Marking all objects used");
1443 slab->inuse = slab->objects;
1444 slab->freelist = NULL;
1445 }
1446 return false;
1447}
1448
1449static inline int free_consistency_checks(struct kmem_cache *s,
1450 struct slab *slab, void *object, unsigned long addr)
1451{
1452 if (!check_valid_pointer(s, slab, object)) {
1453 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1454 return 0;
1455 }
1456
1457 if (on_freelist(s, slab, object)) {
1458 object_err(s, slab, object, "Object already free");
1459 return 0;
1460 }
1461
1462 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1463 return 0;
1464
1465 if (unlikely(s != slab->slab_cache)) {
1466 if (!folio_test_slab(slab_folio(slab))) {
1467 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1468 object);
1469 } else if (!slab->slab_cache) {
1470 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1471 object);
1472 dump_stack();
1473 } else
1474 object_err(s, slab, object,
1475 "page slab pointer corrupt.");
1476 return 0;
1477 }
1478 return 1;
1479}
1480
1481/*
1482 * Parse a block of slub_debug options. Blocks are delimited by ';'
1483 *
1484 * @str: start of block
1485 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1486 * @slabs: return start of list of slabs, or NULL when there's no list
1487 * @init: assume this is initial parsing and not per-kmem-create parsing
1488 *
1489 * returns the start of next block if there's any, or NULL
1490 */
1491static char *
1492parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1493{
1494 bool higher_order_disable = false;
1495
1496 /* Skip any completely empty blocks */
1497 while (*str && *str == ';')
1498 str++;
1499
1500 if (*str == ',') {
1501 /*
1502 * No options but restriction on slabs. This means full
1503 * debugging for slabs matching a pattern.
1504 */
1505 *flags = DEBUG_DEFAULT_FLAGS;
1506 goto check_slabs;
1507 }
1508 *flags = 0;
1509
1510 /* Determine which debug features should be switched on */
1511 for (; *str && *str != ',' && *str != ';'; str++) {
1512 switch (tolower(*str)) {
1513 case '-':
1514 *flags = 0;
1515 break;
1516 case 'f':
1517 *flags |= SLAB_CONSISTENCY_CHECKS;
1518 break;
1519 case 'z':
1520 *flags |= SLAB_RED_ZONE;
1521 break;
1522 case 'p':
1523 *flags |= SLAB_POISON;
1524 break;
1525 case 'u':
1526 *flags |= SLAB_STORE_USER;
1527 break;
1528 case 't':
1529 *flags |= SLAB_TRACE;
1530 break;
1531 case 'a':
1532 *flags |= SLAB_FAILSLAB;
1533 break;
1534 case 'o':
1535 /*
1536 * Avoid enabling debugging on caches if its minimum
1537 * order would increase as a result.
1538 */
1539 higher_order_disable = true;
1540 break;
1541 default:
1542 if (init)
1543 pr_err("slub_debug option '%c' unknown. skipped\n", *str);
1544 }
1545 }
1546check_slabs:
1547 if (*str == ',')
1548 *slabs = ++str;
1549 else
1550 *slabs = NULL;
1551
1552 /* Skip over the slab list */
1553 while (*str && *str != ';')
1554 str++;
1555
1556 /* Skip any completely empty blocks */
1557 while (*str && *str == ';')
1558 str++;
1559
1560 if (init && higher_order_disable)
1561 disable_higher_order_debug = 1;
1562
1563 if (*str)
1564 return str;
1565 else
1566 return NULL;
1567}
1568
1569static int __init setup_slub_debug(char *str)
1570{
1571 slab_flags_t flags;
1572 slab_flags_t global_flags;
1573 char *saved_str;
1574 char *slab_list;
1575 bool global_slub_debug_changed = false;
1576 bool slab_list_specified = false;
1577
1578 global_flags = DEBUG_DEFAULT_FLAGS;
1579 if (*str++ != '=' || !*str)
1580 /*
1581 * No options specified. Switch on full debugging.
1582 */
1583 goto out;
1584
1585 saved_str = str;
1586 while (str) {
1587 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1588
1589 if (!slab_list) {
1590 global_flags = flags;
1591 global_slub_debug_changed = true;
1592 } else {
1593 slab_list_specified = true;
1594 if (flags & SLAB_STORE_USER)
1595 stack_depot_want_early_init();
1596 }
1597 }
1598
1599 /*
1600 * For backwards compatibility, a single list of flags with list of
1601 * slabs means debugging is only changed for those slabs, so the global
1602 * slub_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1603 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1604 * long as there is no option specifying flags without a slab list.
1605 */
1606 if (slab_list_specified) {
1607 if (!global_slub_debug_changed)
1608 global_flags = slub_debug;
1609 slub_debug_string = saved_str;
1610 }
1611out:
1612 slub_debug = global_flags;
1613 if (slub_debug & SLAB_STORE_USER)
1614 stack_depot_want_early_init();
1615 if (slub_debug != 0 || slub_debug_string)
1616 static_branch_enable(&slub_debug_enabled);
1617 else
1618 static_branch_disable(&slub_debug_enabled);
1619 if ((static_branch_unlikely(&init_on_alloc) ||
1620 static_branch_unlikely(&init_on_free)) &&
1621 (slub_debug & SLAB_POISON))
1622 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1623 return 1;
1624}
1625
1626__setup("slub_debug", setup_slub_debug);
1627
1628/*
1629 * kmem_cache_flags - apply debugging options to the cache
1630 * @object_size: the size of an object without meta data
1631 * @flags: flags to set
1632 * @name: name of the cache
1633 *
1634 * Debug option(s) are applied to @flags. In addition to the debug
1635 * option(s), if a slab name (or multiple) is specified i.e.
1636 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1637 * then only the select slabs will receive the debug option(s).
1638 */
1639slab_flags_t kmem_cache_flags(unsigned int object_size,
1640 slab_flags_t flags, const char *name)
1641{
1642 char *iter;
1643 size_t len;
1644 char *next_block;
1645 slab_flags_t block_flags;
1646 slab_flags_t slub_debug_local = slub_debug;
1647
1648 if (flags & SLAB_NO_USER_FLAGS)
1649 return flags;
1650
1651 /*
1652 * If the slab cache is for debugging (e.g. kmemleak) then
1653 * don't store user (stack trace) information by default,
1654 * but let the user enable it via the command line below.
1655 */
1656 if (flags & SLAB_NOLEAKTRACE)
1657 slub_debug_local &= ~SLAB_STORE_USER;
1658
1659 len = strlen(name);
1660 next_block = slub_debug_string;
1661 /* Go through all blocks of debug options, see if any matches our slab's name */
1662 while (next_block) {
1663 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1664 if (!iter)
1665 continue;
1666 /* Found a block that has a slab list, search it */
1667 while (*iter) {
1668 char *end, *glob;
1669 size_t cmplen;
1670
1671 end = strchrnul(iter, ',');
1672 if (next_block && next_block < end)
1673 end = next_block - 1;
1674
1675 glob = strnchr(iter, end - iter, '*');
1676 if (glob)
1677 cmplen = glob - iter;
1678 else
1679 cmplen = max_t(size_t, len, (end - iter));
1680
1681 if (!strncmp(name, iter, cmplen)) {
1682 flags |= block_flags;
1683 return flags;
1684 }
1685
1686 if (!*end || *end == ';')
1687 break;
1688 iter = end + 1;
1689 }
1690 }
1691
1692 return flags | slub_debug_local;
1693}
1694#else /* !CONFIG_SLUB_DEBUG */
1695static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1696static inline
1697void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1698
1699static inline bool alloc_debug_processing(struct kmem_cache *s,
1700 struct slab *slab, void *object, int orig_size) { return true; }
1701
1702static inline bool free_debug_processing(struct kmem_cache *s,
1703 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1704 unsigned long addr, depot_stack_handle_t handle) { return true; }
1705
1706static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1707static inline int check_object(struct kmem_cache *s, struct slab *slab,
1708 void *object, u8 val) { return 1; }
1709static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1710static inline void set_track(struct kmem_cache *s, void *object,
1711 enum track_item alloc, unsigned long addr) {}
1712static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1713 struct slab *slab) {}
1714static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1715 struct slab *slab) {}
1716slab_flags_t kmem_cache_flags(unsigned int object_size,
1717 slab_flags_t flags, const char *name)
1718{
1719 return flags;
1720}
1721#define slub_debug 0
1722
1723#define disable_higher_order_debug 0
1724
1725static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1726 { return 0; }
1727static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1728 { return 0; }
1729static inline void inc_slabs_node(struct kmem_cache *s, int node,
1730 int objects) {}
1731static inline void dec_slabs_node(struct kmem_cache *s, int node,
1732 int objects) {}
1733
1734#ifndef CONFIG_SLUB_TINY
1735static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1736 void **freelist, void *nextfree)
1737{
1738 return false;
1739}
1740#endif
1741#endif /* CONFIG_SLUB_DEBUG */
1742
1743/*
1744 * Hooks for other subsystems that check memory allocations. In a typical
1745 * production configuration these hooks all should produce no code at all.
1746 */
1747static __always_inline bool slab_free_hook(struct kmem_cache *s,
1748 void *x, bool init)
1749{
1750 kmemleak_free_recursive(x, s->flags);
1751 kmsan_slab_free(s, x);
1752
1753 debug_check_no_locks_freed(x, s->object_size);
1754
1755 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1756 debug_check_no_obj_freed(x, s->object_size);
1757
1758 /* Use KCSAN to help debug racy use-after-free. */
1759 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
1760 __kcsan_check_access(x, s->object_size,
1761 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
1762
1763 /*
1764 * As memory initialization might be integrated into KASAN,
1765 * kasan_slab_free and initialization memset's must be
1766 * kept together to avoid discrepancies in behavior.
1767 *
1768 * The initialization memset's clear the object and the metadata,
1769 * but don't touch the SLAB redzone.
1770 */
1771 if (init) {
1772 int rsize;
1773
1774 if (!kasan_has_integrated_init())
1775 memset(kasan_reset_tag(x), 0, s->object_size);
1776 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
1777 memset((char *)kasan_reset_tag(x) + s->inuse, 0,
1778 s->size - s->inuse - rsize);
1779 }
1780 /* KASAN might put x into memory quarantine, delaying its reuse. */
1781 return kasan_slab_free(s, x, init);
1782}
1783
1784static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1785 void **head, void **tail,
1786 int *cnt)
1787{
1788
1789 void *object;
1790 void *next = *head;
1791 void *old_tail = *tail ? *tail : *head;
1792
1793 if (is_kfence_address(next)) {
1794 slab_free_hook(s, next, false);
1795 return true;
1796 }
1797
1798 /* Head and tail of the reconstructed freelist */
1799 *head = NULL;
1800 *tail = NULL;
1801
1802 do {
1803 object = next;
1804 next = get_freepointer(s, object);
1805
1806 /* If object's reuse doesn't have to be delayed */
1807 if (!slab_free_hook(s, object, slab_want_init_on_free(s))) {
1808 /* Move object to the new freelist */
1809 set_freepointer(s, object, *head);
1810 *head = object;
1811 if (!*tail)
1812 *tail = object;
1813 } else {
1814 /*
1815 * Adjust the reconstructed freelist depth
1816 * accordingly if object's reuse is delayed.
1817 */
1818 --(*cnt);
1819 }
1820 } while (object != old_tail);
1821
1822 if (*head == *tail)
1823 *tail = NULL;
1824
1825 return *head != NULL;
1826}
1827
1828static void *setup_object(struct kmem_cache *s, void *object)
1829{
1830 setup_object_debug(s, object);
1831 object = kasan_init_slab_obj(s, object);
1832 if (unlikely(s->ctor)) {
1833 kasan_unpoison_object_data(s, object);
1834 s->ctor(object);
1835 kasan_poison_object_data(s, object);
1836 }
1837 return object;
1838}
1839
1840/*
1841 * Slab allocation and freeing
1842 */
1843static inline struct slab *alloc_slab_page(gfp_t flags, int node,
1844 struct kmem_cache_order_objects oo)
1845{
1846 struct folio *folio;
1847 struct slab *slab;
1848 unsigned int order = oo_order(oo);
1849
1850 if (node == NUMA_NO_NODE)
1851 folio = (struct folio *)alloc_pages(flags, order);
1852 else
1853 folio = (struct folio *)__alloc_pages_node(node, flags, order);
1854
1855 if (!folio)
1856 return NULL;
1857
1858 slab = folio_slab(folio);
1859 __folio_set_slab(folio);
1860 /* Make the flag visible before any changes to folio->mapping */
1861 smp_wmb();
1862 if (page_is_pfmemalloc(folio_page(folio, 0)))
1863 slab_set_pfmemalloc(slab);
1864
1865 return slab;
1866}
1867
1868#ifdef CONFIG_SLAB_FREELIST_RANDOM
1869/* Pre-initialize the random sequence cache */
1870static int init_cache_random_seq(struct kmem_cache *s)
1871{
1872 unsigned int count = oo_objects(s->oo);
1873 int err;
1874
1875 /* Bailout if already initialised */
1876 if (s->random_seq)
1877 return 0;
1878
1879 err = cache_random_seq_create(s, count, GFP_KERNEL);
1880 if (err) {
1881 pr_err("SLUB: Unable to initialize free list for %s\n",
1882 s->name);
1883 return err;
1884 }
1885
1886 /* Transform to an offset on the set of pages */
1887 if (s->random_seq) {
1888 unsigned int i;
1889
1890 for (i = 0; i < count; i++)
1891 s->random_seq[i] *= s->size;
1892 }
1893 return 0;
1894}
1895
1896/* Initialize each random sequence freelist per cache */
1897static void __init init_freelist_randomization(void)
1898{
1899 struct kmem_cache *s;
1900
1901 mutex_lock(&slab_mutex);
1902
1903 list_for_each_entry(s, &slab_caches, list)
1904 init_cache_random_seq(s);
1905
1906 mutex_unlock(&slab_mutex);
1907}
1908
1909/* Get the next entry on the pre-computed freelist randomized */
1910static void *next_freelist_entry(struct kmem_cache *s, struct slab *slab,
1911 unsigned long *pos, void *start,
1912 unsigned long page_limit,
1913 unsigned long freelist_count)
1914{
1915 unsigned int idx;
1916
1917 /*
1918 * If the target page allocation failed, the number of objects on the
1919 * page might be smaller than the usual size defined by the cache.
1920 */
1921 do {
1922 idx = s->random_seq[*pos];
1923 *pos += 1;
1924 if (*pos >= freelist_count)
1925 *pos = 0;
1926 } while (unlikely(idx >= page_limit));
1927
1928 return (char *)start + idx;
1929}
1930
1931/* Shuffle the single linked freelist based on a random pre-computed sequence */
1932static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1933{
1934 void *start;
1935 void *cur;
1936 void *next;
1937 unsigned long idx, pos, page_limit, freelist_count;
1938
1939 if (slab->objects < 2 || !s->random_seq)
1940 return false;
1941
1942 freelist_count = oo_objects(s->oo);
1943 pos = get_random_u32_below(freelist_count);
1944
1945 page_limit = slab->objects * s->size;
1946 start = fixup_red_left(s, slab_address(slab));
1947
1948 /* First entry is used as the base of the freelist */
1949 cur = next_freelist_entry(s, slab, &pos, start, page_limit,
1950 freelist_count);
1951 cur = setup_object(s, cur);
1952 slab->freelist = cur;
1953
1954 for (idx = 1; idx < slab->objects; idx++) {
1955 next = next_freelist_entry(s, slab, &pos, start, page_limit,
1956 freelist_count);
1957 next = setup_object(s, next);
1958 set_freepointer(s, cur, next);
1959 cur = next;
1960 }
1961 set_freepointer(s, cur, NULL);
1962
1963 return true;
1964}
1965#else
1966static inline int init_cache_random_seq(struct kmem_cache *s)
1967{
1968 return 0;
1969}
1970static inline void init_freelist_randomization(void) { }
1971static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
1972{
1973 return false;
1974}
1975#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1976
1977static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1978{
1979 struct slab *slab;
1980 struct kmem_cache_order_objects oo = s->oo;
1981 gfp_t alloc_gfp;
1982 void *start, *p, *next;
1983 int idx;
1984 bool shuffle;
1985
1986 flags &= gfp_allowed_mask;
1987
1988 flags |= s->allocflags;
1989
1990 /*
1991 * Let the initial higher-order allocation fail under memory pressure
1992 * so we fall-back to the minimum order allocation.
1993 */
1994 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1995 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1996 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
1997
1998 slab = alloc_slab_page(alloc_gfp, node, oo);
1999 if (unlikely(!slab)) {
2000 oo = s->min;
2001 alloc_gfp = flags;
2002 /*
2003 * Allocation may have failed due to fragmentation.
2004 * Try a lower order alloc if possible
2005 */
2006 slab = alloc_slab_page(alloc_gfp, node, oo);
2007 if (unlikely(!slab))
2008 return NULL;
2009 stat(s, ORDER_FALLBACK);
2010 }
2011
2012 slab->objects = oo_objects(oo);
2013 slab->inuse = 0;
2014 slab->frozen = 0;
2015
2016 account_slab(slab, oo_order(oo), s, flags);
2017
2018 slab->slab_cache = s;
2019
2020 kasan_poison_slab(slab);
2021
2022 start = slab_address(slab);
2023
2024 setup_slab_debug(s, slab, start);
2025
2026 shuffle = shuffle_freelist(s, slab);
2027
2028 if (!shuffle) {
2029 start = fixup_red_left(s, start);
2030 start = setup_object(s, start);
2031 slab->freelist = start;
2032 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2033 next = p + s->size;
2034 next = setup_object(s, next);
2035 set_freepointer(s, p, next);
2036 p = next;
2037 }
2038 set_freepointer(s, p, NULL);
2039 }
2040
2041 return slab;
2042}
2043
2044static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2045{
2046 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2047 flags = kmalloc_fix_flags(flags);
2048
2049 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2050
2051 return allocate_slab(s,
2052 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2053}
2054
2055static void __free_slab(struct kmem_cache *s, struct slab *slab)
2056{
2057 struct folio *folio = slab_folio(slab);
2058 int order = folio_order(folio);
2059 int pages = 1 << order;
2060
2061 __slab_clear_pfmemalloc(slab);
2062 folio->mapping = NULL;
2063 /* Make the mapping reset visible before clearing the flag */
2064 smp_wmb();
2065 __folio_clear_slab(folio);
2066 if (current->reclaim_state)
2067 current->reclaim_state->reclaimed_slab += pages;
2068 unaccount_slab(slab, order, s);
2069 __free_pages(folio_page(folio, 0), order);
2070}
2071
2072static void rcu_free_slab(struct rcu_head *h)
2073{
2074 struct slab *slab = container_of(h, struct slab, rcu_head);
2075
2076 __free_slab(slab->slab_cache, slab);
2077}
2078
2079static void free_slab(struct kmem_cache *s, struct slab *slab)
2080{
2081 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2082 void *p;
2083
2084 slab_pad_check(s, slab);
2085 for_each_object(p, s, slab_address(slab), slab->objects)
2086 check_object(s, slab, p, SLUB_RED_INACTIVE);
2087 }
2088
2089 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2090 call_rcu(&slab->rcu_head, rcu_free_slab);
2091 else
2092 __free_slab(s, slab);
2093}
2094
2095static void discard_slab(struct kmem_cache *s, struct slab *slab)
2096{
2097 dec_slabs_node(s, slab_nid(slab), slab->objects);
2098 free_slab(s, slab);
2099}
2100
2101/*
2102 * Management of partially allocated slabs.
2103 */
2104static inline void
2105__add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2106{
2107 n->nr_partial++;
2108 if (tail == DEACTIVATE_TO_TAIL)
2109 list_add_tail(&slab->slab_list, &n->partial);
2110 else
2111 list_add(&slab->slab_list, &n->partial);
2112}
2113
2114static inline void add_partial(struct kmem_cache_node *n,
2115 struct slab *slab, int tail)
2116{
2117 lockdep_assert_held(&n->list_lock);
2118 __add_partial(n, slab, tail);
2119}
2120
2121static inline void remove_partial(struct kmem_cache_node *n,
2122 struct slab *slab)
2123{
2124 lockdep_assert_held(&n->list_lock);
2125 list_del(&slab->slab_list);
2126 n->nr_partial--;
2127}
2128
2129/*
2130 * Called only for kmem_cache_debug() caches instead of acquire_slab(), with a
2131 * slab from the n->partial list. Remove only a single object from the slab, do
2132 * the alloc_debug_processing() checks and leave the slab on the list, or move
2133 * it to full list if it was the last free object.
2134 */
2135static void *alloc_single_from_partial(struct kmem_cache *s,
2136 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2137{
2138 void *object;
2139
2140 lockdep_assert_held(&n->list_lock);
2141
2142 object = slab->freelist;
2143 slab->freelist = get_freepointer(s, object);
2144 slab->inuse++;
2145
2146 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2147 remove_partial(n, slab);
2148 return NULL;
2149 }
2150
2151 if (slab->inuse == slab->objects) {
2152 remove_partial(n, slab);
2153 add_full(s, n, slab);
2154 }
2155
2156 return object;
2157}
2158
2159/*
2160 * Called only for kmem_cache_debug() caches to allocate from a freshly
2161 * allocated slab. Allocate a single object instead of whole freelist
2162 * and put the slab to the partial (or full) list.
2163 */
2164static void *alloc_single_from_new_slab(struct kmem_cache *s,
2165 struct slab *slab, int orig_size)
2166{
2167 int nid = slab_nid(slab);
2168 struct kmem_cache_node *n = get_node(s, nid);
2169 unsigned long flags;
2170 void *object;
2171
2172
2173 object = slab->freelist;
2174 slab->freelist = get_freepointer(s, object);
2175 slab->inuse = 1;
2176
2177 if (!alloc_debug_processing(s, slab, object, orig_size))
2178 /*
2179 * It's not really expected that this would fail on a
2180 * freshly allocated slab, but a concurrent memory
2181 * corruption in theory could cause that.
2182 */
2183 return NULL;
2184
2185 spin_lock_irqsave(&n->list_lock, flags);
2186
2187 if (slab->inuse == slab->objects)
2188 add_full(s, n, slab);
2189 else
2190 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2191
2192 inc_slabs_node(s, nid, slab->objects);
2193 spin_unlock_irqrestore(&n->list_lock, flags);
2194
2195 return object;
2196}
2197
2198/*
2199 * Remove slab from the partial list, freeze it and
2200 * return the pointer to the freelist.
2201 *
2202 * Returns a list of objects or NULL if it fails.
2203 */
2204static inline void *acquire_slab(struct kmem_cache *s,
2205 struct kmem_cache_node *n, struct slab *slab,
2206 int mode)
2207{
2208 void *freelist;
2209 unsigned long counters;
2210 struct slab new;
2211
2212 lockdep_assert_held(&n->list_lock);
2213
2214 /*
2215 * Zap the freelist and set the frozen bit.
2216 * The old freelist is the list of objects for the
2217 * per cpu allocation list.
2218 */
2219 freelist = slab->freelist;
2220 counters = slab->counters;
2221 new.counters = counters;
2222 if (mode) {
2223 new.inuse = slab->objects;
2224 new.freelist = NULL;
2225 } else {
2226 new.freelist = freelist;
2227 }
2228
2229 VM_BUG_ON(new.frozen);
2230 new.frozen = 1;
2231
2232 if (!__cmpxchg_double_slab(s, slab,
2233 freelist, counters,
2234 new.freelist, new.counters,
2235 "acquire_slab"))
2236 return NULL;
2237
2238 remove_partial(n, slab);
2239 WARN_ON(!freelist);
2240 return freelist;
2241}
2242
2243#ifdef CONFIG_SLUB_CPU_PARTIAL
2244static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2245#else
2246static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2247 int drain) { }
2248#endif
2249static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2250
2251/*
2252 * Try to allocate a partial slab from a specific node.
2253 */
2254static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
2255 struct partial_context *pc)
2256{
2257 struct slab *slab, *slab2;
2258 void *object = NULL;
2259 unsigned long flags;
2260 unsigned int partial_slabs = 0;
2261
2262 /*
2263 * Racy check. If we mistakenly see no partial slabs then we
2264 * just allocate an empty slab. If we mistakenly try to get a
2265 * partial slab and there is none available then get_partial()
2266 * will return NULL.
2267 */
2268 if (!n || !n->nr_partial)
2269 return NULL;
2270
2271 spin_lock_irqsave(&n->list_lock, flags);
2272 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2273 void *t;
2274
2275 if (!pfmemalloc_match(slab, pc->flags))
2276 continue;
2277
2278 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2279 object = alloc_single_from_partial(s, n, slab,
2280 pc->orig_size);
2281 if (object)
2282 break;
2283 continue;
2284 }
2285
2286 t = acquire_slab(s, n, slab, object == NULL);
2287 if (!t)
2288 break;
2289
2290 if (!object) {
2291 *pc->slab = slab;
2292 stat(s, ALLOC_FROM_PARTIAL);
2293 object = t;
2294 } else {
2295 put_cpu_partial(s, slab, 0);
2296 stat(s, CPU_PARTIAL_NODE);
2297 partial_slabs++;
2298 }
2299#ifdef CONFIG_SLUB_CPU_PARTIAL
2300 if (!kmem_cache_has_cpu_partial(s)
2301 || partial_slabs > s->cpu_partial_slabs / 2)
2302 break;
2303#else
2304 break;
2305#endif
2306
2307 }
2308 spin_unlock_irqrestore(&n->list_lock, flags);
2309 return object;
2310}
2311
2312/*
2313 * Get a slab from somewhere. Search in increasing NUMA distances.
2314 */
2315static void *get_any_partial(struct kmem_cache *s, struct partial_context *pc)
2316{
2317#ifdef CONFIG_NUMA
2318 struct zonelist *zonelist;
2319 struct zoneref *z;
2320 struct zone *zone;
2321 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2322 void *object;
2323 unsigned int cpuset_mems_cookie;
2324
2325 /*
2326 * The defrag ratio allows a configuration of the tradeoffs between
2327 * inter node defragmentation and node local allocations. A lower
2328 * defrag_ratio increases the tendency to do local allocations
2329 * instead of attempting to obtain partial slabs from other nodes.
2330 *
2331 * If the defrag_ratio is set to 0 then kmalloc() always
2332 * returns node local objects. If the ratio is higher then kmalloc()
2333 * may return off node objects because partial slabs are obtained
2334 * from other nodes and filled up.
2335 *
2336 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2337 * (which makes defrag_ratio = 1000) then every (well almost)
2338 * allocation will first attempt to defrag slab caches on other nodes.
2339 * This means scanning over all nodes to look for partial slabs which
2340 * may be expensive if we do it every time we are trying to find a slab
2341 * with available objects.
2342 */
2343 if (!s->remote_node_defrag_ratio ||
2344 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2345 return NULL;
2346
2347 do {
2348 cpuset_mems_cookie = read_mems_allowed_begin();
2349 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2350 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2351 struct kmem_cache_node *n;
2352
2353 n = get_node(s, zone_to_nid(zone));
2354
2355 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2356 n->nr_partial > s->min_partial) {
2357 object = get_partial_node(s, n, pc);
2358 if (object) {
2359 /*
2360 * Don't check read_mems_allowed_retry()
2361 * here - if mems_allowed was updated in
2362 * parallel, that was a harmless race
2363 * between allocation and the cpuset
2364 * update
2365 */
2366 return object;
2367 }
2368 }
2369 }
2370 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2371#endif /* CONFIG_NUMA */
2372 return NULL;
2373}
2374
2375/*
2376 * Get a partial slab, lock it and return it.
2377 */
2378static void *get_partial(struct kmem_cache *s, int node, struct partial_context *pc)
2379{
2380 void *object;
2381 int searchnode = node;
2382
2383 if (node == NUMA_NO_NODE)
2384 searchnode = numa_mem_id();
2385
2386 object = get_partial_node(s, get_node(s, searchnode), pc);
2387 if (object || node != NUMA_NO_NODE)
2388 return object;
2389
2390 return get_any_partial(s, pc);
2391}
2392
2393#ifndef CONFIG_SLUB_TINY
2394
2395#ifdef CONFIG_PREEMPTION
2396/*
2397 * Calculate the next globally unique transaction for disambiguation
2398 * during cmpxchg. The transactions start with the cpu number and are then
2399 * incremented by CONFIG_NR_CPUS.
2400 */
2401#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2402#else
2403/*
2404 * No preemption supported therefore also no need to check for
2405 * different cpus.
2406 */
2407#define TID_STEP 1
2408#endif /* CONFIG_PREEMPTION */
2409
2410static inline unsigned long next_tid(unsigned long tid)
2411{
2412 return tid + TID_STEP;
2413}
2414
2415#ifdef SLUB_DEBUG_CMPXCHG
2416static inline unsigned int tid_to_cpu(unsigned long tid)
2417{
2418 return tid % TID_STEP;
2419}
2420
2421static inline unsigned long tid_to_event(unsigned long tid)
2422{
2423 return tid / TID_STEP;
2424}
2425#endif
2426
2427static inline unsigned int init_tid(int cpu)
2428{
2429 return cpu;
2430}
2431
2432static inline void note_cmpxchg_failure(const char *n,
2433 const struct kmem_cache *s, unsigned long tid)
2434{
2435#ifdef SLUB_DEBUG_CMPXCHG
2436 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2437
2438 pr_info("%s %s: cmpxchg redo ", n, s->name);
2439
2440#ifdef CONFIG_PREEMPTION
2441 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2442 pr_warn("due to cpu change %d -> %d\n",
2443 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2444 else
2445#endif
2446 if (tid_to_event(tid) != tid_to_event(actual_tid))
2447 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2448 tid_to_event(tid), tid_to_event(actual_tid));
2449 else
2450 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2451 actual_tid, tid, next_tid(tid));
2452#endif
2453 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2454}
2455
2456static void init_kmem_cache_cpus(struct kmem_cache *s)
2457{
2458 int cpu;
2459 struct kmem_cache_cpu *c;
2460
2461 for_each_possible_cpu(cpu) {
2462 c = per_cpu_ptr(s->cpu_slab, cpu);
2463 local_lock_init(&c->lock);
2464 c->tid = init_tid(cpu);
2465 }
2466}
2467
2468/*
2469 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2470 * unfreezes the slabs and puts it on the proper list.
2471 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2472 * by the caller.
2473 */
2474static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2475 void *freelist)
2476{
2477 enum slab_modes { M_NONE, M_PARTIAL, M_FREE, M_FULL_NOLIST };
2478 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2479 int free_delta = 0;
2480 enum slab_modes mode = M_NONE;
2481 void *nextfree, *freelist_iter, *freelist_tail;
2482 int tail = DEACTIVATE_TO_HEAD;
2483 unsigned long flags = 0;
2484 struct slab new;
2485 struct slab old;
2486
2487 if (slab->freelist) {
2488 stat(s, DEACTIVATE_REMOTE_FREES);
2489 tail = DEACTIVATE_TO_TAIL;
2490 }
2491
2492 /*
2493 * Stage one: Count the objects on cpu's freelist as free_delta and
2494 * remember the last object in freelist_tail for later splicing.
2495 */
2496 freelist_tail = NULL;
2497 freelist_iter = freelist;
2498 while (freelist_iter) {
2499 nextfree = get_freepointer(s, freelist_iter);
2500
2501 /*
2502 * If 'nextfree' is invalid, it is possible that the object at
2503 * 'freelist_iter' is already corrupted. So isolate all objects
2504 * starting at 'freelist_iter' by skipping them.
2505 */
2506 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2507 break;
2508
2509 freelist_tail = freelist_iter;
2510 free_delta++;
2511
2512 freelist_iter = nextfree;
2513 }
2514
2515 /*
2516 * Stage two: Unfreeze the slab while splicing the per-cpu
2517 * freelist to the head of slab's freelist.
2518 *
2519 * Ensure that the slab is unfrozen while the list presence
2520 * reflects the actual number of objects during unfreeze.
2521 *
2522 * We first perform cmpxchg holding lock and insert to list
2523 * when it succeed. If there is mismatch then the slab is not
2524 * unfrozen and number of objects in the slab may have changed.
2525 * Then release lock and retry cmpxchg again.
2526 */
2527redo:
2528
2529 old.freelist = READ_ONCE(slab->freelist);
2530 old.counters = READ_ONCE(slab->counters);
2531 VM_BUG_ON(!old.frozen);
2532
2533 /* Determine target state of the slab */
2534 new.counters = old.counters;
2535 if (freelist_tail) {
2536 new.inuse -= free_delta;
2537 set_freepointer(s, freelist_tail, old.freelist);
2538 new.freelist = freelist;
2539 } else
2540 new.freelist = old.freelist;
2541
2542 new.frozen = 0;
2543
2544 if (!new.inuse && n->nr_partial >= s->min_partial) {
2545 mode = M_FREE;
2546 } else if (new.freelist) {
2547 mode = M_PARTIAL;
2548 /*
2549 * Taking the spinlock removes the possibility that
2550 * acquire_slab() will see a slab that is frozen
2551 */
2552 spin_lock_irqsave(&n->list_lock, flags);
2553 } else {
2554 mode = M_FULL_NOLIST;
2555 }
2556
2557
2558 if (!cmpxchg_double_slab(s, slab,
2559 old.freelist, old.counters,
2560 new.freelist, new.counters,
2561 "unfreezing slab")) {
2562 if (mode == M_PARTIAL)
2563 spin_unlock_irqrestore(&n->list_lock, flags);
2564 goto redo;
2565 }
2566
2567
2568 if (mode == M_PARTIAL) {
2569 add_partial(n, slab, tail);
2570 spin_unlock_irqrestore(&n->list_lock, flags);
2571 stat(s, tail);
2572 } else if (mode == M_FREE) {
2573 stat(s, DEACTIVATE_EMPTY);
2574 discard_slab(s, slab);
2575 stat(s, FREE_SLAB);
2576 } else if (mode == M_FULL_NOLIST) {
2577 stat(s, DEACTIVATE_FULL);
2578 }
2579}
2580
2581#ifdef CONFIG_SLUB_CPU_PARTIAL
2582static void __unfreeze_partials(struct kmem_cache *s, struct slab *partial_slab)
2583{
2584 struct kmem_cache_node *n = NULL, *n2 = NULL;
2585 struct slab *slab, *slab_to_discard = NULL;
2586 unsigned long flags = 0;
2587
2588 while (partial_slab) {
2589 struct slab new;
2590 struct slab old;
2591
2592 slab = partial_slab;
2593 partial_slab = slab->next;
2594
2595 n2 = get_node(s, slab_nid(slab));
2596 if (n != n2) {
2597 if (n)
2598 spin_unlock_irqrestore(&n->list_lock, flags);
2599
2600 n = n2;
2601 spin_lock_irqsave(&n->list_lock, flags);
2602 }
2603
2604 do {
2605
2606 old.freelist = slab->freelist;
2607 old.counters = slab->counters;
2608 VM_BUG_ON(!old.frozen);
2609
2610 new.counters = old.counters;
2611 new.freelist = old.freelist;
2612
2613 new.frozen = 0;
2614
2615 } while (!__cmpxchg_double_slab(s, slab,
2616 old.freelist, old.counters,
2617 new.freelist, new.counters,
2618 "unfreezing slab"));
2619
2620 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2621 slab->next = slab_to_discard;
2622 slab_to_discard = slab;
2623 } else {
2624 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2625 stat(s, FREE_ADD_PARTIAL);
2626 }
2627 }
2628
2629 if (n)
2630 spin_unlock_irqrestore(&n->list_lock, flags);
2631
2632 while (slab_to_discard) {
2633 slab = slab_to_discard;
2634 slab_to_discard = slab_to_discard->next;
2635
2636 stat(s, DEACTIVATE_EMPTY);
2637 discard_slab(s, slab);
2638 stat(s, FREE_SLAB);
2639 }
2640}
2641
2642/*
2643 * Unfreeze all the cpu partial slabs.
2644 */
2645static void unfreeze_partials(struct kmem_cache *s)
2646{
2647 struct slab *partial_slab;
2648 unsigned long flags;
2649
2650 local_lock_irqsave(&s->cpu_slab->lock, flags);
2651 partial_slab = this_cpu_read(s->cpu_slab->partial);
2652 this_cpu_write(s->cpu_slab->partial, NULL);
2653 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2654
2655 if (partial_slab)
2656 __unfreeze_partials(s, partial_slab);
2657}
2658
2659static void unfreeze_partials_cpu(struct kmem_cache *s,
2660 struct kmem_cache_cpu *c)
2661{
2662 struct slab *partial_slab;
2663
2664 partial_slab = slub_percpu_partial(c);
2665 c->partial = NULL;
2666
2667 if (partial_slab)
2668 __unfreeze_partials(s, partial_slab);
2669}
2670
2671/*
2672 * Put a slab that was just frozen (in __slab_free|get_partial_node) into a
2673 * partial slab slot if available.
2674 *
2675 * If we did not find a slot then simply move all the partials to the
2676 * per node partial list.
2677 */
2678static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2679{
2680 struct slab *oldslab;
2681 struct slab *slab_to_unfreeze = NULL;
2682 unsigned long flags;
2683 int slabs = 0;
2684
2685 local_lock_irqsave(&s->cpu_slab->lock, flags);
2686
2687 oldslab = this_cpu_read(s->cpu_slab->partial);
2688
2689 if (oldslab) {
2690 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2691 /*
2692 * Partial array is full. Move the existing set to the
2693 * per node partial list. Postpone the actual unfreezing
2694 * outside of the critical section.
2695 */
2696 slab_to_unfreeze = oldslab;
2697 oldslab = NULL;
2698 } else {
2699 slabs = oldslab->slabs;
2700 }
2701 }
2702
2703 slabs++;
2704
2705 slab->slabs = slabs;
2706 slab->next = oldslab;
2707
2708 this_cpu_write(s->cpu_slab->partial, slab);
2709
2710 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2711
2712 if (slab_to_unfreeze) {
2713 __unfreeze_partials(s, slab_to_unfreeze);
2714 stat(s, CPU_PARTIAL_DRAIN);
2715 }
2716}
2717
2718#else /* CONFIG_SLUB_CPU_PARTIAL */
2719
2720static inline void unfreeze_partials(struct kmem_cache *s) { }
2721static inline void unfreeze_partials_cpu(struct kmem_cache *s,
2722 struct kmem_cache_cpu *c) { }
2723
2724#endif /* CONFIG_SLUB_CPU_PARTIAL */
2725
2726static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2727{
2728 unsigned long flags;
2729 struct slab *slab;
2730 void *freelist;
2731
2732 local_lock_irqsave(&s->cpu_slab->lock, flags);
2733
2734 slab = c->slab;
2735 freelist = c->freelist;
2736
2737 c->slab = NULL;
2738 c->freelist = NULL;
2739 c->tid = next_tid(c->tid);
2740
2741 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2742
2743 if (slab) {
2744 deactivate_slab(s, slab, freelist);
2745 stat(s, CPUSLAB_FLUSH);
2746 }
2747}
2748
2749static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2750{
2751 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2752 void *freelist = c->freelist;
2753 struct slab *slab = c->slab;
2754
2755 c->slab = NULL;
2756 c->freelist = NULL;
2757 c->tid = next_tid(c->tid);
2758
2759 if (slab) {
2760 deactivate_slab(s, slab, freelist);
2761 stat(s, CPUSLAB_FLUSH);
2762 }
2763
2764 unfreeze_partials_cpu(s, c);
2765}
2766
2767struct slub_flush_work {
2768 struct work_struct work;
2769 struct kmem_cache *s;
2770 bool skip;
2771};
2772
2773/*
2774 * Flush cpu slab.
2775 *
2776 * Called from CPU work handler with migration disabled.
2777 */
2778static void flush_cpu_slab(struct work_struct *w)
2779{
2780 struct kmem_cache *s;
2781 struct kmem_cache_cpu *c;
2782 struct slub_flush_work *sfw;
2783
2784 sfw = container_of(w, struct slub_flush_work, work);
2785
2786 s = sfw->s;
2787 c = this_cpu_ptr(s->cpu_slab);
2788
2789 if (c->slab)
2790 flush_slab(s, c);
2791
2792 unfreeze_partials(s);
2793}
2794
2795static bool has_cpu_slab(int cpu, struct kmem_cache *s)
2796{
2797 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2798
2799 return c->slab || slub_percpu_partial(c);
2800}
2801
2802static DEFINE_MUTEX(flush_lock);
2803static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
2804
2805static void flush_all_cpus_locked(struct kmem_cache *s)
2806{
2807 struct slub_flush_work *sfw;
2808 unsigned int cpu;
2809
2810 lockdep_assert_cpus_held();
2811 mutex_lock(&flush_lock);
2812
2813 for_each_online_cpu(cpu) {
2814 sfw = &per_cpu(slub_flush, cpu);
2815 if (!has_cpu_slab(cpu, s)) {
2816 sfw->skip = true;
2817 continue;
2818 }
2819 INIT_WORK(&sfw->work, flush_cpu_slab);
2820 sfw->skip = false;
2821 sfw->s = s;
2822 queue_work_on(cpu, flushwq, &sfw->work);
2823 }
2824
2825 for_each_online_cpu(cpu) {
2826 sfw = &per_cpu(slub_flush, cpu);
2827 if (sfw->skip)
2828 continue;
2829 flush_work(&sfw->work);
2830 }
2831
2832 mutex_unlock(&flush_lock);
2833}
2834
2835static void flush_all(struct kmem_cache *s)
2836{
2837 cpus_read_lock();
2838 flush_all_cpus_locked(s);
2839 cpus_read_unlock();
2840}
2841
2842/*
2843 * Use the cpu notifier to insure that the cpu slabs are flushed when
2844 * necessary.
2845 */
2846static int slub_cpu_dead(unsigned int cpu)
2847{
2848 struct kmem_cache *s;
2849
2850 mutex_lock(&slab_mutex);
2851 list_for_each_entry(s, &slab_caches, list)
2852 __flush_cpu_slab(s, cpu);
2853 mutex_unlock(&slab_mutex);
2854 return 0;
2855}
2856
2857#else /* CONFIG_SLUB_TINY */
2858static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
2859static inline void flush_all(struct kmem_cache *s) { }
2860static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
2861static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
2862#endif /* CONFIG_SLUB_TINY */
2863
2864/*
2865 * Check if the objects in a per cpu structure fit numa
2866 * locality expectations.
2867 */
2868static inline int node_match(struct slab *slab, int node)
2869{
2870#ifdef CONFIG_NUMA
2871 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
2872 return 0;
2873#endif
2874 return 1;
2875}
2876
2877#ifdef CONFIG_SLUB_DEBUG
2878static int count_free(struct slab *slab)
2879{
2880 return slab->objects - slab->inuse;
2881}
2882
2883static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2884{
2885 return atomic_long_read(&n->total_objects);
2886}
2887
2888/* Supports checking bulk free of a constructed freelist */
2889static inline bool free_debug_processing(struct kmem_cache *s,
2890 struct slab *slab, void *head, void *tail, int *bulk_cnt,
2891 unsigned long addr, depot_stack_handle_t handle)
2892{
2893 bool checks_ok = false;
2894 void *object = head;
2895 int cnt = 0;
2896
2897 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2898 if (!check_slab(s, slab))
2899 goto out;
2900 }
2901
2902 if (slab->inuse < *bulk_cnt) {
2903 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
2904 slab->inuse, *bulk_cnt);
2905 goto out;
2906 }
2907
2908next_object:
2909
2910 if (++cnt > *bulk_cnt)
2911 goto out_cnt;
2912
2913 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
2914 if (!free_consistency_checks(s, slab, object, addr))
2915 goto out;
2916 }
2917
2918 if (s->flags & SLAB_STORE_USER)
2919 set_track_update(s, object, TRACK_FREE, addr, handle);
2920 trace(s, slab, object, 0);
2921 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
2922 init_object(s, object, SLUB_RED_INACTIVE);
2923
2924 /* Reached end of constructed freelist yet? */
2925 if (object != tail) {
2926 object = get_freepointer(s, object);
2927 goto next_object;
2928 }
2929 checks_ok = true;
2930
2931out_cnt:
2932 if (cnt != *bulk_cnt) {
2933 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
2934 *bulk_cnt, cnt);
2935 *bulk_cnt = cnt;
2936 }
2937
2938out:
2939
2940 if (!checks_ok)
2941 slab_fix(s, "Object at 0x%p not freed", object);
2942
2943 return checks_ok;
2944}
2945#endif /* CONFIG_SLUB_DEBUG */
2946
2947#if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
2948static unsigned long count_partial(struct kmem_cache_node *n,
2949 int (*get_count)(struct slab *))
2950{
2951 unsigned long flags;
2952 unsigned long x = 0;
2953 struct slab *slab;
2954
2955 spin_lock_irqsave(&n->list_lock, flags);
2956 list_for_each_entry(slab, &n->partial, slab_list)
2957 x += get_count(slab);
2958 spin_unlock_irqrestore(&n->list_lock, flags);
2959 return x;
2960}
2961#endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
2962
2963#ifdef CONFIG_SLUB_DEBUG
2964static noinline void
2965slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2966{
2967 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2968 DEFAULT_RATELIMIT_BURST);
2969 int node;
2970 struct kmem_cache_node *n;
2971
2972 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2973 return;
2974
2975 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2976 nid, gfpflags, &gfpflags);
2977 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2978 s->name, s->object_size, s->size, oo_order(s->oo),
2979 oo_order(s->min));
2980
2981 if (oo_order(s->min) > get_order(s->object_size))
2982 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2983 s->name);
2984
2985 for_each_kmem_cache_node(s, node, n) {
2986 unsigned long nr_slabs;
2987 unsigned long nr_objs;
2988 unsigned long nr_free;
2989
2990 nr_free = count_partial(n, count_free);
2991 nr_slabs = node_nr_slabs(n);
2992 nr_objs = node_nr_objs(n);
2993
2994 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2995 node, nr_slabs, nr_objs, nr_free);
2996 }
2997}
2998#else /* CONFIG_SLUB_DEBUG */
2999static inline void
3000slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3001#endif
3002
3003static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3004{
3005 if (unlikely(slab_test_pfmemalloc(slab)))
3006 return gfp_pfmemalloc_allowed(gfpflags);
3007
3008 return true;
3009}
3010
3011#ifndef CONFIG_SLUB_TINY
3012/*
3013 * Check the slab->freelist and either transfer the freelist to the
3014 * per cpu freelist or deactivate the slab.
3015 *
3016 * The slab is still frozen if the return value is not NULL.
3017 *
3018 * If this function returns NULL then the slab has been unfrozen.
3019 */
3020static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3021{
3022 struct slab new;
3023 unsigned long counters;
3024 void *freelist;
3025
3026 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3027
3028 do {
3029 freelist = slab->freelist;
3030 counters = slab->counters;
3031
3032 new.counters = counters;
3033 VM_BUG_ON(!new.frozen);
3034
3035 new.inuse = slab->objects;
3036 new.frozen = freelist != NULL;
3037
3038 } while (!__cmpxchg_double_slab(s, slab,
3039 freelist, counters,
3040 NULL, new.counters,
3041 "get_freelist"));
3042
3043 return freelist;
3044}
3045
3046/*
3047 * Slow path. The lockless freelist is empty or we need to perform
3048 * debugging duties.
3049 *
3050 * Processing is still very fast if new objects have been freed to the
3051 * regular freelist. In that case we simply take over the regular freelist
3052 * as the lockless freelist and zap the regular freelist.
3053 *
3054 * If that is not working then we fall back to the partial lists. We take the
3055 * first element of the freelist as the object to allocate now and move the
3056 * rest of the freelist to the lockless freelist.
3057 *
3058 * And if we were unable to get a new slab from the partial slab lists then
3059 * we need to allocate a new slab. This is the slowest path since it involves
3060 * a call to the page allocator and the setup of a new slab.
3061 *
3062 * Version of __slab_alloc to use when we know that preemption is
3063 * already disabled (which is the case for bulk allocation).
3064 */
3065static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3066 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3067{
3068 void *freelist;
3069 struct slab *slab;
3070 unsigned long flags;
3071 struct partial_context pc;
3072
3073 stat(s, ALLOC_SLOWPATH);
3074
3075reread_slab:
3076
3077 slab = READ_ONCE(c->slab);
3078 if (!slab) {
3079 /*
3080 * if the node is not online or has no normal memory, just
3081 * ignore the node constraint
3082 */
3083 if (unlikely(node != NUMA_NO_NODE &&
3084 !node_isset(node, slab_nodes)))
3085 node = NUMA_NO_NODE;
3086 goto new_slab;
3087 }
3088redo:
3089
3090 if (unlikely(!node_match(slab, node))) {
3091 /*
3092 * same as above but node_match() being false already
3093 * implies node != NUMA_NO_NODE
3094 */
3095 if (!node_isset(node, slab_nodes)) {
3096 node = NUMA_NO_NODE;
3097 } else {
3098 stat(s, ALLOC_NODE_MISMATCH);
3099 goto deactivate_slab;
3100 }
3101 }
3102
3103 /*
3104 * By rights, we should be searching for a slab page that was
3105 * PFMEMALLOC but right now, we are losing the pfmemalloc
3106 * information when the page leaves the per-cpu allocator
3107 */
3108 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3109 goto deactivate_slab;
3110
3111 /* must check again c->slab in case we got preempted and it changed */
3112 local_lock_irqsave(&s->cpu_slab->lock, flags);
3113 if (unlikely(slab != c->slab)) {
3114 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3115 goto reread_slab;
3116 }
3117 freelist = c->freelist;
3118 if (freelist)
3119 goto load_freelist;
3120
3121 freelist = get_freelist(s, slab);
3122
3123 if (!freelist) {
3124 c->slab = NULL;
3125 c->tid = next_tid(c->tid);
3126 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3127 stat(s, DEACTIVATE_BYPASS);
3128 goto new_slab;
3129 }
3130
3131 stat(s, ALLOC_REFILL);
3132
3133load_freelist:
3134
3135 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3136
3137 /*
3138 * freelist is pointing to the list of objects to be used.
3139 * slab is pointing to the slab from which the objects are obtained.
3140 * That slab must be frozen for per cpu allocations to work.
3141 */
3142 VM_BUG_ON(!c->slab->frozen);
3143 c->freelist = get_freepointer(s, freelist);
3144 c->tid = next_tid(c->tid);
3145 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3146 return freelist;
3147
3148deactivate_slab:
3149
3150 local_lock_irqsave(&s->cpu_slab->lock, flags);
3151 if (slab != c->slab) {
3152 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3153 goto reread_slab;
3154 }
3155 freelist = c->freelist;
3156 c->slab = NULL;
3157 c->freelist = NULL;
3158 c->tid = next_tid(c->tid);
3159 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3160 deactivate_slab(s, slab, freelist);
3161
3162new_slab:
3163
3164 if (slub_percpu_partial(c)) {
3165 local_lock_irqsave(&s->cpu_slab->lock, flags);
3166 if (unlikely(c->slab)) {
3167 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3168 goto reread_slab;
3169 }
3170 if (unlikely(!slub_percpu_partial(c))) {
3171 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3172 /* we were preempted and partial list got empty */
3173 goto new_objects;
3174 }
3175
3176 slab = c->slab = slub_percpu_partial(c);
3177 slub_set_percpu_partial(c, slab);
3178 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3179 stat(s, CPU_PARTIAL_ALLOC);
3180 goto redo;
3181 }
3182
3183new_objects:
3184
3185 pc.flags = gfpflags;
3186 pc.slab = &slab;
3187 pc.orig_size = orig_size;
3188 freelist = get_partial(s, node, &pc);
3189 if (freelist)
3190 goto check_new_slab;
3191
3192 slub_put_cpu_ptr(s->cpu_slab);
3193 slab = new_slab(s, gfpflags, node);
3194 c = slub_get_cpu_ptr(s->cpu_slab);
3195
3196 if (unlikely(!slab)) {
3197 slab_out_of_memory(s, gfpflags, node);
3198 return NULL;
3199 }
3200
3201 stat(s, ALLOC_SLAB);
3202
3203 if (kmem_cache_debug(s)) {
3204 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3205
3206 if (unlikely(!freelist))
3207 goto new_objects;
3208
3209 if (s->flags & SLAB_STORE_USER)
3210 set_track(s, freelist, TRACK_ALLOC, addr);
3211
3212 return freelist;
3213 }
3214
3215 /*
3216 * No other reference to the slab yet so we can
3217 * muck around with it freely without cmpxchg
3218 */
3219 freelist = slab->freelist;
3220 slab->freelist = NULL;
3221 slab->inuse = slab->objects;
3222 slab->frozen = 1;
3223
3224 inc_slabs_node(s, slab_nid(slab), slab->objects);
3225
3226check_new_slab:
3227
3228 if (kmem_cache_debug(s)) {
3229 /*
3230 * For debug caches here we had to go through
3231 * alloc_single_from_partial() so just store the tracking info
3232 * and return the object
3233 */
3234 if (s->flags & SLAB_STORE_USER)
3235 set_track(s, freelist, TRACK_ALLOC, addr);
3236
3237 return freelist;
3238 }
3239
3240 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3241 /*
3242 * For !pfmemalloc_match() case we don't load freelist so that
3243 * we don't make further mismatched allocations easier.
3244 */
3245 deactivate_slab(s, slab, get_freepointer(s, freelist));
3246 return freelist;
3247 }
3248
3249retry_load_slab:
3250
3251 local_lock_irqsave(&s->cpu_slab->lock, flags);
3252 if (unlikely(c->slab)) {
3253 void *flush_freelist = c->freelist;
3254 struct slab *flush_slab = c->slab;
3255
3256 c->slab = NULL;
3257 c->freelist = NULL;
3258 c->tid = next_tid(c->tid);
3259
3260 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3261
3262 deactivate_slab(s, flush_slab, flush_freelist);
3263
3264 stat(s, CPUSLAB_FLUSH);
3265
3266 goto retry_load_slab;
3267 }
3268 c->slab = slab;
3269
3270 goto load_freelist;
3271}
3272
3273/*
3274 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3275 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3276 * pointer.
3277 */
3278static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3279 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3280{
3281 void *p;
3282
3283#ifdef CONFIG_PREEMPT_COUNT
3284 /*
3285 * We may have been preempted and rescheduled on a different
3286 * cpu before disabling preemption. Need to reload cpu area
3287 * pointer.
3288 */
3289 c = slub_get_cpu_ptr(s->cpu_slab);
3290#endif
3291
3292 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3293#ifdef CONFIG_PREEMPT_COUNT
3294 slub_put_cpu_ptr(s->cpu_slab);
3295#endif
3296 return p;
3297}
3298
3299static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3300 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3301{
3302 struct kmem_cache_cpu *c;
3303 struct slab *slab;
3304 unsigned long tid;
3305 void *object;
3306
3307redo:
3308 /*
3309 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3310 * enabled. We may switch back and forth between cpus while
3311 * reading from one cpu area. That does not matter as long
3312 * as we end up on the original cpu again when doing the cmpxchg.
3313 *
3314 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3315 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3316 * the tid. If we are preempted and switched to another cpu between the
3317 * two reads, it's OK as the two are still associated with the same cpu
3318 * and cmpxchg later will validate the cpu.
3319 */
3320 c = raw_cpu_ptr(s->cpu_slab);
3321 tid = READ_ONCE(c->tid);
3322
3323 /*
3324 * Irqless object alloc/free algorithm used here depends on sequence
3325 * of fetching cpu_slab's data. tid should be fetched before anything
3326 * on c to guarantee that object and slab associated with previous tid
3327 * won't be used with current tid. If we fetch tid first, object and
3328 * slab could be one associated with next tid and our alloc/free
3329 * request will be failed. In this case, we will retry. So, no problem.
3330 */
3331 barrier();
3332
3333 /*
3334 * The transaction ids are globally unique per cpu and per operation on
3335 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3336 * occurs on the right processor and that there was no operation on the
3337 * linked list in between.
3338 */
3339
3340 object = c->freelist;
3341 slab = c->slab;
3342
3343 if (!USE_LOCKLESS_FAST_PATH() ||
3344 unlikely(!object || !slab || !node_match(slab, node))) {
3345 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3346 } else {
3347 void *next_object = get_freepointer_safe(s, object);
3348
3349 /*
3350 * The cmpxchg will only match if there was no additional
3351 * operation and if we are on the right processor.
3352 *
3353 * The cmpxchg does the following atomically (without lock
3354 * semantics!)
3355 * 1. Relocate first pointer to the current per cpu area.
3356 * 2. Verify that tid and freelist have not been changed
3357 * 3. If they were not changed replace tid and freelist
3358 *
3359 * Since this is without lock semantics the protection is only
3360 * against code executing on this cpu *not* from access by
3361 * other cpus.
3362 */
3363 if (unlikely(!this_cpu_cmpxchg_double(
3364 s->cpu_slab->freelist, s->cpu_slab->tid,
3365 object, tid,
3366 next_object, next_tid(tid)))) {
3367
3368 note_cmpxchg_failure("slab_alloc", s, tid);
3369 goto redo;
3370 }
3371 prefetch_freepointer(s, next_object);
3372 stat(s, ALLOC_FASTPATH);
3373 }
3374
3375 return object;
3376}
3377#else /* CONFIG_SLUB_TINY */
3378static void *__slab_alloc_node(struct kmem_cache *s,
3379 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3380{
3381 struct partial_context pc;
3382 struct slab *slab;
3383 void *object;
3384
3385 pc.flags = gfpflags;
3386 pc.slab = &slab;
3387 pc.orig_size = orig_size;
3388 object = get_partial(s, node, &pc);
3389
3390 if (object)
3391 return object;
3392
3393 slab = new_slab(s, gfpflags, node);
3394 if (unlikely(!slab)) {
3395 slab_out_of_memory(s, gfpflags, node);
3396 return NULL;
3397 }
3398
3399 object = alloc_single_from_new_slab(s, slab, orig_size);
3400
3401 return object;
3402}
3403#endif /* CONFIG_SLUB_TINY */
3404
3405/*
3406 * If the object has been wiped upon free, make sure it's fully initialized by
3407 * zeroing out freelist pointer.
3408 */
3409static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3410 void *obj)
3411{
3412 if (unlikely(slab_want_init_on_free(s)) && obj)
3413 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3414 0, sizeof(void *));
3415}
3416
3417/*
3418 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3419 * have the fastpath folded into their functions. So no function call
3420 * overhead for requests that can be satisfied on the fastpath.
3421 *
3422 * The fastpath works by first checking if the lockless freelist can be used.
3423 * If not then __slab_alloc is called for slow processing.
3424 *
3425 * Otherwise we can simply pick the next object from the lockless free list.
3426 */
3427static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3428 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3429{
3430 void *object;
3431 struct obj_cgroup *objcg = NULL;
3432 bool init = false;
3433
3434 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3435 if (!s)
3436 return NULL;
3437
3438 object = kfence_alloc(s, orig_size, gfpflags);
3439 if (unlikely(object))
3440 goto out;
3441
3442 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3443
3444 maybe_wipe_obj_freeptr(s, object);
3445 init = slab_want_init_on_alloc(gfpflags, s);
3446
3447out:
3448 /*
3449 * When init equals 'true', like for kzalloc() family, only
3450 * @orig_size bytes might be zeroed instead of s->object_size
3451 */
3452 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3453
3454 return object;
3455}
3456
3457static __fastpath_inline void *slab_alloc(struct kmem_cache *s, struct list_lru *lru,
3458 gfp_t gfpflags, unsigned long addr, size_t orig_size)
3459{
3460 return slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, addr, orig_size);
3461}
3462
3463static __fastpath_inline
3464void *__kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3465 gfp_t gfpflags)
3466{
3467 void *ret = slab_alloc(s, lru, gfpflags, _RET_IP_, s->object_size);
3468
3469 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3470
3471 return ret;
3472}
3473
3474void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3475{
3476 return __kmem_cache_alloc_lru(s, NULL, gfpflags);
3477}
3478EXPORT_SYMBOL(kmem_cache_alloc);
3479
3480void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3481 gfp_t gfpflags)
3482{
3483 return __kmem_cache_alloc_lru(s, lru, gfpflags);
3484}
3485EXPORT_SYMBOL(kmem_cache_alloc_lru);
3486
3487void *__kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags,
3488 int node, size_t orig_size,
3489 unsigned long caller)
3490{
3491 return slab_alloc_node(s, NULL, gfpflags, node,
3492 caller, orig_size);
3493}
3494
3495void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3496{
3497 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3498
3499 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3500
3501 return ret;
3502}
3503EXPORT_SYMBOL(kmem_cache_alloc_node);
3504
3505static noinline void free_to_partial_list(
3506 struct kmem_cache *s, struct slab *slab,
3507 void *head, void *tail, int bulk_cnt,
3508 unsigned long addr)
3509{
3510 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
3511 struct slab *slab_free = NULL;
3512 int cnt = bulk_cnt;
3513 unsigned long flags;
3514 depot_stack_handle_t handle = 0;
3515
3516 if (s->flags & SLAB_STORE_USER)
3517 handle = set_track_prepare();
3518
3519 spin_lock_irqsave(&n->list_lock, flags);
3520
3521 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
3522 void *prior = slab->freelist;
3523
3524 /* Perform the actual freeing while we still hold the locks */
3525 slab->inuse -= cnt;
3526 set_freepointer(s, tail, prior);
3527 slab->freelist = head;
3528
3529 /*
3530 * If the slab is empty, and node's partial list is full,
3531 * it should be discarded anyway no matter it's on full or
3532 * partial list.
3533 */
3534 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
3535 slab_free = slab;
3536
3537 if (!prior) {
3538 /* was on full list */
3539 remove_full(s, n, slab);
3540 if (!slab_free) {
3541 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3542 stat(s, FREE_ADD_PARTIAL);
3543 }
3544 } else if (slab_free) {
3545 remove_partial(n, slab);
3546 stat(s, FREE_REMOVE_PARTIAL);
3547 }
3548 }
3549
3550 if (slab_free) {
3551 /*
3552 * Update the counters while still holding n->list_lock to
3553 * prevent spurious validation warnings
3554 */
3555 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
3556 }
3557
3558 spin_unlock_irqrestore(&n->list_lock, flags);
3559
3560 if (slab_free) {
3561 stat(s, FREE_SLAB);
3562 free_slab(s, slab_free);
3563 }
3564}
3565
3566/*
3567 * Slow path handling. This may still be called frequently since objects
3568 * have a longer lifetime than the cpu slabs in most processing loads.
3569 *
3570 * So we still attempt to reduce cache line usage. Just take the slab
3571 * lock and free the item. If there is no additional partial slab
3572 * handling required then we can return immediately.
3573 */
3574static void __slab_free(struct kmem_cache *s, struct slab *slab,
3575 void *head, void *tail, int cnt,
3576 unsigned long addr)
3577
3578{
3579 void *prior;
3580 int was_frozen;
3581 struct slab new;
3582 unsigned long counters;
3583 struct kmem_cache_node *n = NULL;
3584 unsigned long flags;
3585
3586 stat(s, FREE_SLOWPATH);
3587
3588 if (kfence_free(head))
3589 return;
3590
3591 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
3592 free_to_partial_list(s, slab, head, tail, cnt, addr);
3593 return;
3594 }
3595
3596 do {
3597 if (unlikely(n)) {
3598 spin_unlock_irqrestore(&n->list_lock, flags);
3599 n = NULL;
3600 }
3601 prior = slab->freelist;
3602 counters = slab->counters;
3603 set_freepointer(s, tail, prior);
3604 new.counters = counters;
3605 was_frozen = new.frozen;
3606 new.inuse -= cnt;
3607 if ((!new.inuse || !prior) && !was_frozen) {
3608
3609 if (kmem_cache_has_cpu_partial(s) && !prior) {
3610
3611 /*
3612 * Slab was on no list before and will be
3613 * partially empty
3614 * We can defer the list move and instead
3615 * freeze it.
3616 */
3617 new.frozen = 1;
3618
3619 } else { /* Needs to be taken off a list */
3620
3621 n = get_node(s, slab_nid(slab));
3622 /*
3623 * Speculatively acquire the list_lock.
3624 * If the cmpxchg does not succeed then we may
3625 * drop the list_lock without any processing.
3626 *
3627 * Otherwise the list_lock will synchronize with
3628 * other processors updating the list of slabs.
3629 */
3630 spin_lock_irqsave(&n->list_lock, flags);
3631
3632 }
3633 }
3634
3635 } while (!cmpxchg_double_slab(s, slab,
3636 prior, counters,
3637 head, new.counters,
3638 "__slab_free"));
3639
3640 if (likely(!n)) {
3641
3642 if (likely(was_frozen)) {
3643 /*
3644 * The list lock was not taken therefore no list
3645 * activity can be necessary.
3646 */
3647 stat(s, FREE_FROZEN);
3648 } else if (new.frozen) {
3649 /*
3650 * If we just froze the slab then put it onto the
3651 * per cpu partial list.
3652 */
3653 put_cpu_partial(s, slab, 1);
3654 stat(s, CPU_PARTIAL_FREE);
3655 }
3656
3657 return;
3658 }
3659
3660 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
3661 goto slab_empty;
3662
3663 /*
3664 * Objects left in the slab. If it was not on the partial list before
3665 * then add it.
3666 */
3667 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
3668 remove_full(s, n, slab);
3669 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3670 stat(s, FREE_ADD_PARTIAL);
3671 }
3672 spin_unlock_irqrestore(&n->list_lock, flags);
3673 return;
3674
3675slab_empty:
3676 if (prior) {
3677 /*
3678 * Slab on the partial list.
3679 */
3680 remove_partial(n, slab);
3681 stat(s, FREE_REMOVE_PARTIAL);
3682 } else {
3683 /* Slab must be on the full list */
3684 remove_full(s, n, slab);
3685 }
3686
3687 spin_unlock_irqrestore(&n->list_lock, flags);
3688 stat(s, FREE_SLAB);
3689 discard_slab(s, slab);
3690}
3691
3692#ifndef CONFIG_SLUB_TINY
3693/*
3694 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
3695 * can perform fastpath freeing without additional function calls.
3696 *
3697 * The fastpath is only possible if we are freeing to the current cpu slab
3698 * of this processor. This typically the case if we have just allocated
3699 * the item before.
3700 *
3701 * If fastpath is not possible then fall back to __slab_free where we deal
3702 * with all sorts of special processing.
3703 *
3704 * Bulk free of a freelist with several objects (all pointing to the
3705 * same slab) possible by specifying head and tail ptr, plus objects
3706 * count (cnt). Bulk free indicated by tail pointer being set.
3707 */
3708static __always_inline void do_slab_free(struct kmem_cache *s,
3709 struct slab *slab, void *head, void *tail,
3710 int cnt, unsigned long addr)
3711{
3712 void *tail_obj = tail ? : head;
3713 struct kmem_cache_cpu *c;
3714 unsigned long tid;
3715 void **freelist;
3716
3717redo:
3718 /*
3719 * Determine the currently cpus per cpu slab.
3720 * The cpu may change afterward. However that does not matter since
3721 * data is retrieved via this pointer. If we are on the same cpu
3722 * during the cmpxchg then the free will succeed.
3723 */
3724 c = raw_cpu_ptr(s->cpu_slab);
3725 tid = READ_ONCE(c->tid);
3726
3727 /* Same with comment on barrier() in slab_alloc_node() */
3728 barrier();
3729
3730 if (unlikely(slab != c->slab)) {
3731 __slab_free(s, slab, head, tail_obj, cnt, addr);
3732 return;
3733 }
3734
3735 if (USE_LOCKLESS_FAST_PATH()) {
3736 freelist = READ_ONCE(c->freelist);
3737
3738 set_freepointer(s, tail_obj, freelist);
3739
3740 if (unlikely(!this_cpu_cmpxchg_double(
3741 s->cpu_slab->freelist, s->cpu_slab->tid,
3742 freelist, tid,
3743 head, next_tid(tid)))) {
3744
3745 note_cmpxchg_failure("slab_free", s, tid);
3746 goto redo;
3747 }
3748 } else {
3749 /* Update the free list under the local lock */
3750 local_lock(&s->cpu_slab->lock);
3751 c = this_cpu_ptr(s->cpu_slab);
3752 if (unlikely(slab != c->slab)) {
3753 local_unlock(&s->cpu_slab->lock);
3754 goto redo;
3755 }
3756 tid = c->tid;
3757 freelist = c->freelist;
3758
3759 set_freepointer(s, tail_obj, freelist);
3760 c->freelist = head;
3761 c->tid = next_tid(tid);
3762
3763 local_unlock(&s->cpu_slab->lock);
3764 }
3765 stat(s, FREE_FASTPATH);
3766}
3767#else /* CONFIG_SLUB_TINY */
3768static void do_slab_free(struct kmem_cache *s,
3769 struct slab *slab, void *head, void *tail,
3770 int cnt, unsigned long addr)
3771{
3772 void *tail_obj = tail ? : head;
3773
3774 __slab_free(s, slab, head, tail_obj, cnt, addr);
3775}
3776#endif /* CONFIG_SLUB_TINY */
3777
3778static __fastpath_inline void slab_free(struct kmem_cache *s, struct slab *slab,
3779 void *head, void *tail, void **p, int cnt,
3780 unsigned long addr)
3781{
3782 memcg_slab_free_hook(s, slab, p, cnt);
3783 /*
3784 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3785 * to remove objects, whose reuse must be delayed.
3786 */
3787 if (slab_free_freelist_hook(s, &head, &tail, &cnt))
3788 do_slab_free(s, slab, head, tail, cnt, addr);
3789}
3790
3791#ifdef CONFIG_KASAN_GENERIC
3792void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3793{
3794 do_slab_free(cache, virt_to_slab(x), x, NULL, 1, addr);
3795}
3796#endif
3797
3798void __kmem_cache_free(struct kmem_cache *s, void *x, unsigned long caller)
3799{
3800 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, caller);
3801}
3802
3803void kmem_cache_free(struct kmem_cache *s, void *x)
3804{
3805 s = cache_from_obj(s, x);
3806 if (!s)
3807 return;
3808 trace_kmem_cache_free(_RET_IP_, x, s);
3809 slab_free(s, virt_to_slab(x), x, NULL, &x, 1, _RET_IP_);
3810}
3811EXPORT_SYMBOL(kmem_cache_free);
3812
3813struct detached_freelist {
3814 struct slab *slab;
3815 void *tail;
3816 void *freelist;
3817 int cnt;
3818 struct kmem_cache *s;
3819};
3820
3821/*
3822 * This function progressively scans the array with free objects (with
3823 * a limited look ahead) and extract objects belonging to the same
3824 * slab. It builds a detached freelist directly within the given
3825 * slab/objects. This can happen without any need for
3826 * synchronization, because the objects are owned by running process.
3827 * The freelist is build up as a single linked list in the objects.
3828 * The idea is, that this detached freelist can then be bulk
3829 * transferred to the real freelist(s), but only requiring a single
3830 * synchronization primitive. Look ahead in the array is limited due
3831 * to performance reasons.
3832 */
3833static inline
3834int build_detached_freelist(struct kmem_cache *s, size_t size,
3835 void **p, struct detached_freelist *df)
3836{
3837 int lookahead = 3;
3838 void *object;
3839 struct folio *folio;
3840 size_t same;
3841
3842 object = p[--size];
3843 folio = virt_to_folio(object);
3844 if (!s) {
3845 /* Handle kalloc'ed objects */
3846 if (unlikely(!folio_test_slab(folio))) {
3847 free_large_kmalloc(folio, object);
3848 df->slab = NULL;
3849 return size;
3850 }
3851 /* Derive kmem_cache from object */
3852 df->slab = folio_slab(folio);
3853 df->s = df->slab->slab_cache;
3854 } else {
3855 df->slab = folio_slab(folio);
3856 df->s = cache_from_obj(s, object); /* Support for memcg */
3857 }
3858
3859 /* Start new detached freelist */
3860 df->tail = object;
3861 df->freelist = object;
3862 df->cnt = 1;
3863
3864 if (is_kfence_address(object))
3865 return size;
3866
3867 set_freepointer(df->s, object, NULL);
3868
3869 same = size;
3870 while (size) {
3871 object = p[--size];
3872 /* df->slab is always set at this point */
3873 if (df->slab == virt_to_slab(object)) {
3874 /* Opportunity build freelist */
3875 set_freepointer(df->s, object, df->freelist);
3876 df->freelist = object;
3877 df->cnt++;
3878 same--;
3879 if (size != same)
3880 swap(p[size], p[same]);
3881 continue;
3882 }
3883
3884 /* Limit look ahead search */
3885 if (!--lookahead)
3886 break;
3887 }
3888
3889 return same;
3890}
3891
3892/* Note that interrupts must be enabled when calling this function. */
3893void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3894{
3895 if (!size)
3896 return;
3897
3898 do {
3899 struct detached_freelist df;
3900
3901 size = build_detached_freelist(s, size, p, &df);
3902 if (!df.slab)
3903 continue;
3904
3905 slab_free(df.s, df.slab, df.freelist, df.tail, &p[size], df.cnt,
3906 _RET_IP_);
3907 } while (likely(size));
3908}
3909EXPORT_SYMBOL(kmem_cache_free_bulk);
3910
3911#ifndef CONFIG_SLUB_TINY
3912static inline int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
3913 size_t size, void **p, struct obj_cgroup *objcg)
3914{
3915 struct kmem_cache_cpu *c;
3916 int i;
3917
3918 /*
3919 * Drain objects in the per cpu slab, while disabling local
3920 * IRQs, which protects against PREEMPT and interrupts
3921 * handlers invoking normal fastpath.
3922 */
3923 c = slub_get_cpu_ptr(s->cpu_slab);
3924 local_lock_irq(&s->cpu_slab->lock);
3925
3926 for (i = 0; i < size; i++) {
3927 void *object = kfence_alloc(s, s->object_size, flags);
3928
3929 if (unlikely(object)) {
3930 p[i] = object;
3931 continue;
3932 }
3933
3934 object = c->freelist;
3935 if (unlikely(!object)) {
3936 /*
3937 * We may have removed an object from c->freelist using
3938 * the fastpath in the previous iteration; in that case,
3939 * c->tid has not been bumped yet.
3940 * Since ___slab_alloc() may reenable interrupts while
3941 * allocating memory, we should bump c->tid now.
3942 */
3943 c->tid = next_tid(c->tid);
3944
3945 local_unlock_irq(&s->cpu_slab->lock);
3946
3947 /*
3948 * Invoking slow path likely have side-effect
3949 * of re-populating per CPU c->freelist
3950 */
3951 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3952 _RET_IP_, c, s->object_size);
3953 if (unlikely(!p[i]))
3954 goto error;
3955
3956 c = this_cpu_ptr(s->cpu_slab);
3957 maybe_wipe_obj_freeptr(s, p[i]);
3958
3959 local_lock_irq(&s->cpu_slab->lock);
3960
3961 continue; /* goto for-loop */
3962 }
3963 c->freelist = get_freepointer(s, object);
3964 p[i] = object;
3965 maybe_wipe_obj_freeptr(s, p[i]);
3966 }
3967 c->tid = next_tid(c->tid);
3968 local_unlock_irq(&s->cpu_slab->lock);
3969 slub_put_cpu_ptr(s->cpu_slab);
3970
3971 return i;
3972
3973error:
3974 slub_put_cpu_ptr(s->cpu_slab);
3975 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
3976 kmem_cache_free_bulk(s, i, p);
3977 return 0;
3978
3979}
3980#else /* CONFIG_SLUB_TINY */
3981static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
3982 size_t size, void **p, struct obj_cgroup *objcg)
3983{
3984 int i;
3985
3986 for (i = 0; i < size; i++) {
3987 void *object = kfence_alloc(s, s->object_size, flags);
3988
3989 if (unlikely(object)) {
3990 p[i] = object;
3991 continue;
3992 }
3993
3994 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
3995 _RET_IP_, s->object_size);
3996 if (unlikely(!p[i]))
3997 goto error;
3998
3999 maybe_wipe_obj_freeptr(s, p[i]);
4000 }
4001
4002 return i;
4003
4004error:
4005 slab_post_alloc_hook(s, objcg, flags, i, p, false, s->object_size);
4006 kmem_cache_free_bulk(s, i, p);
4007 return 0;
4008}
4009#endif /* CONFIG_SLUB_TINY */
4010
4011/* Note that interrupts must be enabled when calling this function. */
4012int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4013 void **p)
4014{
4015 int i;
4016 struct obj_cgroup *objcg = NULL;
4017
4018 if (!size)
4019 return 0;
4020
4021 /* memcg and kmem_cache debug support */
4022 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4023 if (unlikely(!s))
4024 return 0;
4025
4026 i = __kmem_cache_alloc_bulk(s, flags, size, p, objcg);
4027
4028 /*
4029 * memcg and kmem_cache debug support and memory initialization.
4030 * Done outside of the IRQ disabled fastpath loop.
4031 */
4032 if (i != 0)
4033 slab_post_alloc_hook(s, objcg, flags, size, p,
4034 slab_want_init_on_alloc(flags, s), s->object_size);
4035 return i;
4036}
4037EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4038
4039
4040/*
4041 * Object placement in a slab is made very easy because we always start at
4042 * offset 0. If we tune the size of the object to the alignment then we can
4043 * get the required alignment by putting one properly sized object after
4044 * another.
4045 *
4046 * Notice that the allocation order determines the sizes of the per cpu
4047 * caches. Each processor has always one slab available for allocations.
4048 * Increasing the allocation order reduces the number of times that slabs
4049 * must be moved on and off the partial lists and is therefore a factor in
4050 * locking overhead.
4051 */
4052
4053/*
4054 * Minimum / Maximum order of slab pages. This influences locking overhead
4055 * and slab fragmentation. A higher order reduces the number of partial slabs
4056 * and increases the number of allocations possible without having to
4057 * take the list_lock.
4058 */
4059static unsigned int slub_min_order;
4060static unsigned int slub_max_order =
4061 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4062static unsigned int slub_min_objects;
4063
4064/*
4065 * Calculate the order of allocation given an slab object size.
4066 *
4067 * The order of allocation has significant impact on performance and other
4068 * system components. Generally order 0 allocations should be preferred since
4069 * order 0 does not cause fragmentation in the page allocator. Larger objects
4070 * be problematic to put into order 0 slabs because there may be too much
4071 * unused space left. We go to a higher order if more than 1/16th of the slab
4072 * would be wasted.
4073 *
4074 * In order to reach satisfactory performance we must ensure that a minimum
4075 * number of objects is in one slab. Otherwise we may generate too much
4076 * activity on the partial lists which requires taking the list_lock. This is
4077 * less a concern for large slabs though which are rarely used.
4078 *
4079 * slub_max_order specifies the order where we begin to stop considering the
4080 * number of objects in a slab as critical. If we reach slub_max_order then
4081 * we try to keep the page order as low as possible. So we accept more waste
4082 * of space in favor of a small page order.
4083 *
4084 * Higher order allocations also allow the placement of more objects in a
4085 * slab and thereby reduce object handling overhead. If the user has
4086 * requested a higher minimum order then we start with that one instead of
4087 * the smallest order which will fit the object.
4088 */
4089static inline unsigned int calc_slab_order(unsigned int size,
4090 unsigned int min_objects, unsigned int max_order,
4091 unsigned int fract_leftover)
4092{
4093 unsigned int min_order = slub_min_order;
4094 unsigned int order;
4095
4096 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4097 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4098
4099 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
4100 order <= max_order; order++) {
4101
4102 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4103 unsigned int rem;
4104
4105 rem = slab_size % size;
4106
4107 if (rem <= slab_size / fract_leftover)
4108 break;
4109 }
4110
4111 return order;
4112}
4113
4114static inline int calculate_order(unsigned int size)
4115{
4116 unsigned int order;
4117 unsigned int min_objects;
4118 unsigned int max_objects;
4119 unsigned int nr_cpus;
4120
4121 /*
4122 * Attempt to find best configuration for a slab. This
4123 * works by first attempting to generate a layout with
4124 * the best configuration and backing off gradually.
4125 *
4126 * First we increase the acceptable waste in a slab. Then
4127 * we reduce the minimum objects required in a slab.
4128 */
4129 min_objects = slub_min_objects;
4130 if (!min_objects) {
4131 /*
4132 * Some architectures will only update present cpus when
4133 * onlining them, so don't trust the number if it's just 1. But
4134 * we also don't want to use nr_cpu_ids always, as on some other
4135 * architectures, there can be many possible cpus, but never
4136 * onlined. Here we compromise between trying to avoid too high
4137 * order on systems that appear larger than they are, and too
4138 * low order on systems that appear smaller than they are.
4139 */
4140 nr_cpus = num_present_cpus();
4141 if (nr_cpus <= 1)
4142 nr_cpus = nr_cpu_ids;
4143 min_objects = 4 * (fls(nr_cpus) + 1);
4144 }
4145 max_objects = order_objects(slub_max_order, size);
4146 min_objects = min(min_objects, max_objects);
4147
4148 while (min_objects > 1) {
4149 unsigned int fraction;
4150
4151 fraction = 16;
4152 while (fraction >= 4) {
4153 order = calc_slab_order(size, min_objects,
4154 slub_max_order, fraction);
4155 if (order <= slub_max_order)
4156 return order;
4157 fraction /= 2;
4158 }
4159 min_objects--;
4160 }
4161
4162 /*
4163 * We were unable to place multiple objects in a slab. Now
4164 * lets see if we can place a single object there.
4165 */
4166 order = calc_slab_order(size, 1, slub_max_order, 1);
4167 if (order <= slub_max_order)
4168 return order;
4169
4170 /*
4171 * Doh this slab cannot be placed using slub_max_order.
4172 */
4173 order = calc_slab_order(size, 1, MAX_ORDER, 1);
4174 if (order < MAX_ORDER)
4175 return order;
4176 return -ENOSYS;
4177}
4178
4179static void
4180init_kmem_cache_node(struct kmem_cache_node *n)
4181{
4182 n->nr_partial = 0;
4183 spin_lock_init(&n->list_lock);
4184 INIT_LIST_HEAD(&n->partial);
4185#ifdef CONFIG_SLUB_DEBUG
4186 atomic_long_set(&n->nr_slabs, 0);
4187 atomic_long_set(&n->total_objects, 0);
4188 INIT_LIST_HEAD(&n->full);
4189#endif
4190}
4191
4192#ifndef CONFIG_SLUB_TINY
4193static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4194{
4195 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4196 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4197 sizeof(struct kmem_cache_cpu));
4198
4199 /*
4200 * Must align to double word boundary for the double cmpxchg
4201 * instructions to work; see __pcpu_double_call_return_bool().
4202 */
4203 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4204 2 * sizeof(void *));
4205
4206 if (!s->cpu_slab)
4207 return 0;
4208
4209 init_kmem_cache_cpus(s);
4210
4211 return 1;
4212}
4213#else
4214static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4215{
4216 return 1;
4217}
4218#endif /* CONFIG_SLUB_TINY */
4219
4220static struct kmem_cache *kmem_cache_node;
4221
4222/*
4223 * No kmalloc_node yet so do it by hand. We know that this is the first
4224 * slab on the node for this slabcache. There are no concurrent accesses
4225 * possible.
4226 *
4227 * Note that this function only works on the kmem_cache_node
4228 * when allocating for the kmem_cache_node. This is used for bootstrapping
4229 * memory on a fresh node that has no slab structures yet.
4230 */
4231static void early_kmem_cache_node_alloc(int node)
4232{
4233 struct slab *slab;
4234 struct kmem_cache_node *n;
4235
4236 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4237
4238 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4239
4240 BUG_ON(!slab);
4241 inc_slabs_node(kmem_cache_node, slab_nid(slab), slab->objects);
4242 if (slab_nid(slab) != node) {
4243 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4244 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4245 }
4246
4247 n = slab->freelist;
4248 BUG_ON(!n);
4249#ifdef CONFIG_SLUB_DEBUG
4250 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4251 init_tracking(kmem_cache_node, n);
4252#endif
4253 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4254 slab->freelist = get_freepointer(kmem_cache_node, n);
4255 slab->inuse = 1;
4256 kmem_cache_node->node[node] = n;
4257 init_kmem_cache_node(n);
4258 inc_slabs_node(kmem_cache_node, node, slab->objects);
4259
4260 /*
4261 * No locks need to be taken here as it has just been
4262 * initialized and there is no concurrent access.
4263 */
4264 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4265}
4266
4267static void free_kmem_cache_nodes(struct kmem_cache *s)
4268{
4269 int node;
4270 struct kmem_cache_node *n;
4271
4272 for_each_kmem_cache_node(s, node, n) {
4273 s->node[node] = NULL;
4274 kmem_cache_free(kmem_cache_node, n);
4275 }
4276}
4277
4278void __kmem_cache_release(struct kmem_cache *s)
4279{
4280 cache_random_seq_destroy(s);
4281#ifndef CONFIG_SLUB_TINY
4282 free_percpu(s->cpu_slab);
4283#endif
4284 free_kmem_cache_nodes(s);
4285}
4286
4287static int init_kmem_cache_nodes(struct kmem_cache *s)
4288{
4289 int node;
4290
4291 for_each_node_mask(node, slab_nodes) {
4292 struct kmem_cache_node *n;
4293
4294 if (slab_state == DOWN) {
4295 early_kmem_cache_node_alloc(node);
4296 continue;
4297 }
4298 n = kmem_cache_alloc_node(kmem_cache_node,
4299 GFP_KERNEL, node);
4300
4301 if (!n) {
4302 free_kmem_cache_nodes(s);
4303 return 0;
4304 }
4305
4306 init_kmem_cache_node(n);
4307 s->node[node] = n;
4308 }
4309 return 1;
4310}
4311
4312static void set_cpu_partial(struct kmem_cache *s)
4313{
4314#ifdef CONFIG_SLUB_CPU_PARTIAL
4315 unsigned int nr_objects;
4316
4317 /*
4318 * cpu_partial determined the maximum number of objects kept in the
4319 * per cpu partial lists of a processor.
4320 *
4321 * Per cpu partial lists mainly contain slabs that just have one
4322 * object freed. If they are used for allocation then they can be
4323 * filled up again with minimal effort. The slab will never hit the
4324 * per node partial lists and therefore no locking will be required.
4325 *
4326 * For backwards compatibility reasons, this is determined as number
4327 * of objects, even though we now limit maximum number of pages, see
4328 * slub_set_cpu_partial()
4329 */
4330 if (!kmem_cache_has_cpu_partial(s))
4331 nr_objects = 0;
4332 else if (s->size >= PAGE_SIZE)
4333 nr_objects = 6;
4334 else if (s->size >= 1024)
4335 nr_objects = 24;
4336 else if (s->size >= 256)
4337 nr_objects = 52;
4338 else
4339 nr_objects = 120;
4340
4341 slub_set_cpu_partial(s, nr_objects);
4342#endif
4343}
4344
4345/*
4346 * calculate_sizes() determines the order and the distribution of data within
4347 * a slab object.
4348 */
4349static int calculate_sizes(struct kmem_cache *s)
4350{
4351 slab_flags_t flags = s->flags;
4352 unsigned int size = s->object_size;
4353 unsigned int order;
4354
4355 /*
4356 * Round up object size to the next word boundary. We can only
4357 * place the free pointer at word boundaries and this determines
4358 * the possible location of the free pointer.
4359 */
4360 size = ALIGN(size, sizeof(void *));
4361
4362#ifdef CONFIG_SLUB_DEBUG
4363 /*
4364 * Determine if we can poison the object itself. If the user of
4365 * the slab may touch the object after free or before allocation
4366 * then we should never poison the object itself.
4367 */
4368 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4369 !s->ctor)
4370 s->flags |= __OBJECT_POISON;
4371 else
4372 s->flags &= ~__OBJECT_POISON;
4373
4374
4375 /*
4376 * If we are Redzoning then check if there is some space between the
4377 * end of the object and the free pointer. If not then add an
4378 * additional word to have some bytes to store Redzone information.
4379 */
4380 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4381 size += sizeof(void *);
4382#endif
4383
4384 /*
4385 * With that we have determined the number of bytes in actual use
4386 * by the object and redzoning.
4387 */
4388 s->inuse = size;
4389
4390 if (slub_debug_orig_size(s) ||
4391 (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4392 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4393 s->ctor) {
4394 /*
4395 * Relocate free pointer after the object if it is not
4396 * permitted to overwrite the first word of the object on
4397 * kmem_cache_free.
4398 *
4399 * This is the case if we do RCU, have a constructor or
4400 * destructor, are poisoning the objects, or are
4401 * redzoning an object smaller than sizeof(void *).
4402 *
4403 * The assumption that s->offset >= s->inuse means free
4404 * pointer is outside of the object is used in the
4405 * freeptr_outside_object() function. If that is no
4406 * longer true, the function needs to be modified.
4407 */
4408 s->offset = size;
4409 size += sizeof(void *);
4410 } else {
4411 /*
4412 * Store freelist pointer near middle of object to keep
4413 * it away from the edges of the object to avoid small
4414 * sized over/underflows from neighboring allocations.
4415 */
4416 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
4417 }
4418
4419#ifdef CONFIG_SLUB_DEBUG
4420 if (flags & SLAB_STORE_USER) {
4421 /*
4422 * Need to store information about allocs and frees after
4423 * the object.
4424 */
4425 size += 2 * sizeof(struct track);
4426
4427 /* Save the original kmalloc request size */
4428 if (flags & SLAB_KMALLOC)
4429 size += sizeof(unsigned int);
4430 }
4431#endif
4432
4433 kasan_cache_create(s, &size, &s->flags);
4434#ifdef CONFIG_SLUB_DEBUG
4435 if (flags & SLAB_RED_ZONE) {
4436 /*
4437 * Add some empty padding so that we can catch
4438 * overwrites from earlier objects rather than let
4439 * tracking information or the free pointer be
4440 * corrupted if a user writes before the start
4441 * of the object.
4442 */
4443 size += sizeof(void *);
4444
4445 s->red_left_pad = sizeof(void *);
4446 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
4447 size += s->red_left_pad;
4448 }
4449#endif
4450
4451 /*
4452 * SLUB stores one object immediately after another beginning from
4453 * offset 0. In order to align the objects we have to simply size
4454 * each object to conform to the alignment.
4455 */
4456 size = ALIGN(size, s->align);
4457 s->size = size;
4458 s->reciprocal_size = reciprocal_value(size);
4459 order = calculate_order(size);
4460
4461 if ((int)order < 0)
4462 return 0;
4463
4464 s->allocflags = 0;
4465 if (order)
4466 s->allocflags |= __GFP_COMP;
4467
4468 if (s->flags & SLAB_CACHE_DMA)
4469 s->allocflags |= GFP_DMA;
4470
4471 if (s->flags & SLAB_CACHE_DMA32)
4472 s->allocflags |= GFP_DMA32;
4473
4474 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4475 s->allocflags |= __GFP_RECLAIMABLE;
4476
4477 /*
4478 * Determine the number of objects per slab
4479 */
4480 s->oo = oo_make(order, size);
4481 s->min = oo_make(get_order(size), size);
4482
4483 return !!oo_objects(s->oo);
4484}
4485
4486static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
4487{
4488 s->flags = kmem_cache_flags(s->size, flags, s->name);
4489#ifdef CONFIG_SLAB_FREELIST_HARDENED
4490 s->random = get_random_long();
4491#endif
4492
4493 if (!calculate_sizes(s))
4494 goto error;
4495 if (disable_higher_order_debug) {
4496 /*
4497 * Disable debugging flags that store metadata if the min slab
4498 * order increased.
4499 */
4500 if (get_order(s->size) > get_order(s->object_size)) {
4501 s->flags &= ~DEBUG_METADATA_FLAGS;
4502 s->offset = 0;
4503 if (!calculate_sizes(s))
4504 goto error;
4505 }
4506 }
4507
4508#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
4509 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
4510 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
4511 /* Enable fast mode */
4512 s->flags |= __CMPXCHG_DOUBLE;
4513#endif
4514
4515 /*
4516 * The larger the object size is, the more slabs we want on the partial
4517 * list to avoid pounding the page allocator excessively.
4518 */
4519 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
4520 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
4521
4522 set_cpu_partial(s);
4523
4524#ifdef CONFIG_NUMA
4525 s->remote_node_defrag_ratio = 1000;
4526#endif
4527
4528 /* Initialize the pre-computed randomized freelist if slab is up */
4529 if (slab_state >= UP) {
4530 if (init_cache_random_seq(s))
4531 goto error;
4532 }
4533
4534 if (!init_kmem_cache_nodes(s))
4535 goto error;
4536
4537 if (alloc_kmem_cache_cpus(s))
4538 return 0;
4539
4540error:
4541 __kmem_cache_release(s);
4542 return -EINVAL;
4543}
4544
4545static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
4546 const char *text)
4547{
4548#ifdef CONFIG_SLUB_DEBUG
4549 void *addr = slab_address(slab);
4550 void *p;
4551
4552 slab_err(s, slab, text, s->name);
4553
4554 spin_lock(&object_map_lock);
4555 __fill_map(object_map, s, slab);
4556
4557 for_each_object(p, s, addr, slab->objects) {
4558
4559 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
4560 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
4561 print_tracking(s, p);
4562 }
4563 }
4564 spin_unlock(&object_map_lock);
4565#endif
4566}
4567
4568/*
4569 * Attempt to free all partial slabs on a node.
4570 * This is called from __kmem_cache_shutdown(). We must take list_lock
4571 * because sysfs file might still access partial list after the shutdowning.
4572 */
4573static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
4574{
4575 LIST_HEAD(discard);
4576 struct slab *slab, *h;
4577
4578 BUG_ON(irqs_disabled());
4579 spin_lock_irq(&n->list_lock);
4580 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
4581 if (!slab->inuse) {
4582 remove_partial(n, slab);
4583 list_add(&slab->slab_list, &discard);
4584 } else {
4585 list_slab_objects(s, slab,
4586 "Objects remaining in %s on __kmem_cache_shutdown()");
4587 }
4588 }
4589 spin_unlock_irq(&n->list_lock);
4590
4591 list_for_each_entry_safe(slab, h, &discard, slab_list)
4592 discard_slab(s, slab);
4593}
4594
4595bool __kmem_cache_empty(struct kmem_cache *s)
4596{
4597 int node;
4598 struct kmem_cache_node *n;
4599
4600 for_each_kmem_cache_node(s, node, n)
4601 if (n->nr_partial || slabs_node(s, node))
4602 return false;
4603 return true;
4604}
4605
4606/*
4607 * Release all resources used by a slab cache.
4608 */
4609int __kmem_cache_shutdown(struct kmem_cache *s)
4610{
4611 int node;
4612 struct kmem_cache_node *n;
4613
4614 flush_all_cpus_locked(s);
4615 /* Attempt to free all objects */
4616 for_each_kmem_cache_node(s, node, n) {
4617 free_partial(s, n);
4618 if (n->nr_partial || slabs_node(s, node))
4619 return 1;
4620 }
4621 return 0;
4622}
4623
4624#ifdef CONFIG_PRINTK
4625void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
4626{
4627 void *base;
4628 int __maybe_unused i;
4629 unsigned int objnr;
4630 void *objp;
4631 void *objp0;
4632 struct kmem_cache *s = slab->slab_cache;
4633 struct track __maybe_unused *trackp;
4634
4635 kpp->kp_ptr = object;
4636 kpp->kp_slab = slab;
4637 kpp->kp_slab_cache = s;
4638 base = slab_address(slab);
4639 objp0 = kasan_reset_tag(object);
4640#ifdef CONFIG_SLUB_DEBUG
4641 objp = restore_red_left(s, objp0);
4642#else
4643 objp = objp0;
4644#endif
4645 objnr = obj_to_index(s, slab, objp);
4646 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
4647 objp = base + s->size * objnr;
4648 kpp->kp_objp = objp;
4649 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
4650 || (objp - base) % s->size) ||
4651 !(s->flags & SLAB_STORE_USER))
4652 return;
4653#ifdef CONFIG_SLUB_DEBUG
4654 objp = fixup_red_left(s, objp);
4655 trackp = get_track(s, objp, TRACK_ALLOC);
4656 kpp->kp_ret = (void *)trackp->addr;
4657#ifdef CONFIG_STACKDEPOT
4658 {
4659 depot_stack_handle_t handle;
4660 unsigned long *entries;
4661 unsigned int nr_entries;
4662
4663 handle = READ_ONCE(trackp->handle);
4664 if (handle) {
4665 nr_entries = stack_depot_fetch(handle, &entries);
4666 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4667 kpp->kp_stack[i] = (void *)entries[i];
4668 }
4669
4670 trackp = get_track(s, objp, TRACK_FREE);
4671 handle = READ_ONCE(trackp->handle);
4672 if (handle) {
4673 nr_entries = stack_depot_fetch(handle, &entries);
4674 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
4675 kpp->kp_free_stack[i] = (void *)entries[i];
4676 }
4677 }
4678#endif
4679#endif
4680}
4681#endif
4682
4683/********************************************************************
4684 * Kmalloc subsystem
4685 *******************************************************************/
4686
4687static int __init setup_slub_min_order(char *str)
4688{
4689 get_option(&str, (int *)&slub_min_order);
4690
4691 return 1;
4692}
4693
4694__setup("slub_min_order=", setup_slub_min_order);
4695
4696static int __init setup_slub_max_order(char *str)
4697{
4698 get_option(&str, (int *)&slub_max_order);
4699 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
4700
4701 return 1;
4702}
4703
4704__setup("slub_max_order=", setup_slub_max_order);
4705
4706static int __init setup_slub_min_objects(char *str)
4707{
4708 get_option(&str, (int *)&slub_min_objects);
4709
4710 return 1;
4711}
4712
4713__setup("slub_min_objects=", setup_slub_min_objects);
4714
4715#ifdef CONFIG_HARDENED_USERCOPY
4716/*
4717 * Rejects incorrectly sized objects and objects that are to be copied
4718 * to/from userspace but do not fall entirely within the containing slab
4719 * cache's usercopy region.
4720 *
4721 * Returns NULL if check passes, otherwise const char * to name of cache
4722 * to indicate an error.
4723 */
4724void __check_heap_object(const void *ptr, unsigned long n,
4725 const struct slab *slab, bool to_user)
4726{
4727 struct kmem_cache *s;
4728 unsigned int offset;
4729 bool is_kfence = is_kfence_address(ptr);
4730
4731 ptr = kasan_reset_tag(ptr);
4732
4733 /* Find object and usable object size. */
4734 s = slab->slab_cache;
4735
4736 /* Reject impossible pointers. */
4737 if (ptr < slab_address(slab))
4738 usercopy_abort("SLUB object not in SLUB page?!", NULL,
4739 to_user, 0, n);
4740
4741 /* Find offset within object. */
4742 if (is_kfence)
4743 offset = ptr - kfence_object_start(ptr);
4744 else
4745 offset = (ptr - slab_address(slab)) % s->size;
4746
4747 /* Adjust for redzone and reject if within the redzone. */
4748 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
4749 if (offset < s->red_left_pad)
4750 usercopy_abort("SLUB object in left red zone",
4751 s->name, to_user, offset, n);
4752 offset -= s->red_left_pad;
4753 }
4754
4755 /* Allow address range falling entirely within usercopy region. */
4756 if (offset >= s->useroffset &&
4757 offset - s->useroffset <= s->usersize &&
4758 n <= s->useroffset - offset + s->usersize)
4759 return;
4760
4761 usercopy_abort("SLUB object", s->name, to_user, offset, n);
4762}
4763#endif /* CONFIG_HARDENED_USERCOPY */
4764
4765#define SHRINK_PROMOTE_MAX 32
4766
4767/*
4768 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
4769 * up most to the head of the partial lists. New allocations will then
4770 * fill those up and thus they can be removed from the partial lists.
4771 *
4772 * The slabs with the least items are placed last. This results in them
4773 * being allocated from last increasing the chance that the last objects
4774 * are freed in them.
4775 */
4776static int __kmem_cache_do_shrink(struct kmem_cache *s)
4777{
4778 int node;
4779 int i;
4780 struct kmem_cache_node *n;
4781 struct slab *slab;
4782 struct slab *t;
4783 struct list_head discard;
4784 struct list_head promote[SHRINK_PROMOTE_MAX];
4785 unsigned long flags;
4786 int ret = 0;
4787
4788 for_each_kmem_cache_node(s, node, n) {
4789 INIT_LIST_HEAD(&discard);
4790 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
4791 INIT_LIST_HEAD(promote + i);
4792
4793 spin_lock_irqsave(&n->list_lock, flags);
4794
4795 /*
4796 * Build lists of slabs to discard or promote.
4797 *
4798 * Note that concurrent frees may occur while we hold the
4799 * list_lock. slab->inuse here is the upper limit.
4800 */
4801 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
4802 int free = slab->objects - slab->inuse;
4803
4804 /* Do not reread slab->inuse */
4805 barrier();
4806
4807 /* We do not keep full slabs on the list */
4808 BUG_ON(free <= 0);
4809
4810 if (free == slab->objects) {
4811 list_move(&slab->slab_list, &discard);
4812 n->nr_partial--;
4813 dec_slabs_node(s, node, slab->objects);
4814 } else if (free <= SHRINK_PROMOTE_MAX)
4815 list_move(&slab->slab_list, promote + free - 1);
4816 }
4817
4818 /*
4819 * Promote the slabs filled up most to the head of the
4820 * partial list.
4821 */
4822 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4823 list_splice(promote + i, &n->partial);
4824
4825 spin_unlock_irqrestore(&n->list_lock, flags);
4826
4827 /* Release empty slabs */
4828 list_for_each_entry_safe(slab, t, &discard, slab_list)
4829 free_slab(s, slab);
4830
4831 if (slabs_node(s, node))
4832 ret = 1;
4833 }
4834
4835 return ret;
4836}
4837
4838int __kmem_cache_shrink(struct kmem_cache *s)
4839{
4840 flush_all(s);
4841 return __kmem_cache_do_shrink(s);
4842}
4843
4844static int slab_mem_going_offline_callback(void *arg)
4845{
4846 struct kmem_cache *s;
4847
4848 mutex_lock(&slab_mutex);
4849 list_for_each_entry(s, &slab_caches, list) {
4850 flush_all_cpus_locked(s);
4851 __kmem_cache_do_shrink(s);
4852 }
4853 mutex_unlock(&slab_mutex);
4854
4855 return 0;
4856}
4857
4858static void slab_mem_offline_callback(void *arg)
4859{
4860 struct memory_notify *marg = arg;
4861 int offline_node;
4862
4863 offline_node = marg->status_change_nid_normal;
4864
4865 /*
4866 * If the node still has available memory. we need kmem_cache_node
4867 * for it yet.
4868 */
4869 if (offline_node < 0)
4870 return;
4871
4872 mutex_lock(&slab_mutex);
4873 node_clear(offline_node, slab_nodes);
4874 /*
4875 * We no longer free kmem_cache_node structures here, as it would be
4876 * racy with all get_node() users, and infeasible to protect them with
4877 * slab_mutex.
4878 */
4879 mutex_unlock(&slab_mutex);
4880}
4881
4882static int slab_mem_going_online_callback(void *arg)
4883{
4884 struct kmem_cache_node *n;
4885 struct kmem_cache *s;
4886 struct memory_notify *marg = arg;
4887 int nid = marg->status_change_nid_normal;
4888 int ret = 0;
4889
4890 /*
4891 * If the node's memory is already available, then kmem_cache_node is
4892 * already created. Nothing to do.
4893 */
4894 if (nid < 0)
4895 return 0;
4896
4897 /*
4898 * We are bringing a node online. No memory is available yet. We must
4899 * allocate a kmem_cache_node structure in order to bring the node
4900 * online.
4901 */
4902 mutex_lock(&slab_mutex);
4903 list_for_each_entry(s, &slab_caches, list) {
4904 /*
4905 * The structure may already exist if the node was previously
4906 * onlined and offlined.
4907 */
4908 if (get_node(s, nid))
4909 continue;
4910 /*
4911 * XXX: kmem_cache_alloc_node will fallback to other nodes
4912 * since memory is not yet available from the node that
4913 * is brought up.
4914 */
4915 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4916 if (!n) {
4917 ret = -ENOMEM;
4918 goto out;
4919 }
4920 init_kmem_cache_node(n);
4921 s->node[nid] = n;
4922 }
4923 /*
4924 * Any cache created after this point will also have kmem_cache_node
4925 * initialized for the new node.
4926 */
4927 node_set(nid, slab_nodes);
4928out:
4929 mutex_unlock(&slab_mutex);
4930 return ret;
4931}
4932
4933static int slab_memory_callback(struct notifier_block *self,
4934 unsigned long action, void *arg)
4935{
4936 int ret = 0;
4937
4938 switch (action) {
4939 case MEM_GOING_ONLINE:
4940 ret = slab_mem_going_online_callback(arg);
4941 break;
4942 case MEM_GOING_OFFLINE:
4943 ret = slab_mem_going_offline_callback(arg);
4944 break;
4945 case MEM_OFFLINE:
4946 case MEM_CANCEL_ONLINE:
4947 slab_mem_offline_callback(arg);
4948 break;
4949 case MEM_ONLINE:
4950 case MEM_CANCEL_OFFLINE:
4951 break;
4952 }
4953 if (ret)
4954 ret = notifier_from_errno(ret);
4955 else
4956 ret = NOTIFY_OK;
4957 return ret;
4958}
4959
4960/********************************************************************
4961 * Basic setup of slabs
4962 *******************************************************************/
4963
4964/*
4965 * Used for early kmem_cache structures that were allocated using
4966 * the page allocator. Allocate them properly then fix up the pointers
4967 * that may be pointing to the wrong kmem_cache structure.
4968 */
4969
4970static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4971{
4972 int node;
4973 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4974 struct kmem_cache_node *n;
4975
4976 memcpy(s, static_cache, kmem_cache->object_size);
4977
4978 /*
4979 * This runs very early, and only the boot processor is supposed to be
4980 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4981 * IPIs around.
4982 */
4983 __flush_cpu_slab(s, smp_processor_id());
4984 for_each_kmem_cache_node(s, node, n) {
4985 struct slab *p;
4986
4987 list_for_each_entry(p, &n->partial, slab_list)
4988 p->slab_cache = s;
4989
4990#ifdef CONFIG_SLUB_DEBUG
4991 list_for_each_entry(p, &n->full, slab_list)
4992 p->slab_cache = s;
4993#endif
4994 }
4995 list_add(&s->list, &slab_caches);
4996 return s;
4997}
4998
4999void __init kmem_cache_init(void)
5000{
5001 static __initdata struct kmem_cache boot_kmem_cache,
5002 boot_kmem_cache_node;
5003 int node;
5004
5005 if (debug_guardpage_minorder())
5006 slub_max_order = 0;
5007
5008 /* Print slub debugging pointers without hashing */
5009 if (__slub_debug_enabled())
5010 no_hash_pointers_enable(NULL);
5011
5012 kmem_cache_node = &boot_kmem_cache_node;
5013 kmem_cache = &boot_kmem_cache;
5014
5015 /*
5016 * Initialize the nodemask for which we will allocate per node
5017 * structures. Here we don't need taking slab_mutex yet.
5018 */
5019 for_each_node_state(node, N_NORMAL_MEMORY)
5020 node_set(node, slab_nodes);
5021
5022 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5023 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5024
5025 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5026
5027 /* Able to allocate the per node structures */
5028 slab_state = PARTIAL;
5029
5030 create_boot_cache(kmem_cache, "kmem_cache",
5031 offsetof(struct kmem_cache, node) +
5032 nr_node_ids * sizeof(struct kmem_cache_node *),
5033 SLAB_HWCACHE_ALIGN, 0, 0);
5034
5035 kmem_cache = bootstrap(&boot_kmem_cache);
5036 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5037
5038 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5039 setup_kmalloc_cache_index_table();
5040 create_kmalloc_caches(0);
5041
5042 /* Setup random freelists for each cache */
5043 init_freelist_randomization();
5044
5045 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5046 slub_cpu_dead);
5047
5048 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5049 cache_line_size(),
5050 slub_min_order, slub_max_order, slub_min_objects,
5051 nr_cpu_ids, nr_node_ids);
5052}
5053
5054void __init kmem_cache_init_late(void)
5055{
5056#ifndef CONFIG_SLUB_TINY
5057 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5058 WARN_ON(!flushwq);
5059#endif
5060}
5061
5062struct kmem_cache *
5063__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5064 slab_flags_t flags, void (*ctor)(void *))
5065{
5066 struct kmem_cache *s;
5067
5068 s = find_mergeable(size, align, flags, name, ctor);
5069 if (s) {
5070 if (sysfs_slab_alias(s, name))
5071 return NULL;
5072
5073 s->refcount++;
5074
5075 /*
5076 * Adjust the object sizes so that we clear
5077 * the complete object on kzalloc.
5078 */
5079 s->object_size = max(s->object_size, size);
5080 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5081 }
5082
5083 return s;
5084}
5085
5086int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5087{
5088 int err;
5089
5090 err = kmem_cache_open(s, flags);
5091 if (err)
5092 return err;
5093
5094 /* Mutex is not taken during early boot */
5095 if (slab_state <= UP)
5096 return 0;
5097
5098 err = sysfs_slab_add(s);
5099 if (err) {
5100 __kmem_cache_release(s);
5101 return err;
5102 }
5103
5104 if (s->flags & SLAB_STORE_USER)
5105 debugfs_slab_add(s);
5106
5107 return 0;
5108}
5109
5110#ifdef SLAB_SUPPORTS_SYSFS
5111static int count_inuse(struct slab *slab)
5112{
5113 return slab->inuse;
5114}
5115
5116static int count_total(struct slab *slab)
5117{
5118 return slab->objects;
5119}
5120#endif
5121
5122#ifdef CONFIG_SLUB_DEBUG
5123static void validate_slab(struct kmem_cache *s, struct slab *slab,
5124 unsigned long *obj_map)
5125{
5126 void *p;
5127 void *addr = slab_address(slab);
5128
5129 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5130 return;
5131
5132 /* Now we know that a valid freelist exists */
5133 __fill_map(obj_map, s, slab);
5134 for_each_object(p, s, addr, slab->objects) {
5135 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5136 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5137
5138 if (!check_object(s, slab, p, val))
5139 break;
5140 }
5141}
5142
5143static int validate_slab_node(struct kmem_cache *s,
5144 struct kmem_cache_node *n, unsigned long *obj_map)
5145{
5146 unsigned long count = 0;
5147 struct slab *slab;
5148 unsigned long flags;
5149
5150 spin_lock_irqsave(&n->list_lock, flags);
5151
5152 list_for_each_entry(slab, &n->partial, slab_list) {
5153 validate_slab(s, slab, obj_map);
5154 count++;
5155 }
5156 if (count != n->nr_partial) {
5157 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5158 s->name, count, n->nr_partial);
5159 slab_add_kunit_errors();
5160 }
5161
5162 if (!(s->flags & SLAB_STORE_USER))
5163 goto out;
5164
5165 list_for_each_entry(slab, &n->full, slab_list) {
5166 validate_slab(s, slab, obj_map);
5167 count++;
5168 }
5169 if (count != atomic_long_read(&n->nr_slabs)) {
5170 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5171 s->name, count, atomic_long_read(&n->nr_slabs));
5172 slab_add_kunit_errors();
5173 }
5174
5175out:
5176 spin_unlock_irqrestore(&n->list_lock, flags);
5177 return count;
5178}
5179
5180long validate_slab_cache(struct kmem_cache *s)
5181{
5182 int node;
5183 unsigned long count = 0;
5184 struct kmem_cache_node *n;
5185 unsigned long *obj_map;
5186
5187 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5188 if (!obj_map)
5189 return -ENOMEM;
5190
5191 flush_all(s);
5192 for_each_kmem_cache_node(s, node, n)
5193 count += validate_slab_node(s, n, obj_map);
5194
5195 bitmap_free(obj_map);
5196
5197 return count;
5198}
5199EXPORT_SYMBOL(validate_slab_cache);
5200
5201#ifdef CONFIG_DEBUG_FS
5202/*
5203 * Generate lists of code addresses where slabcache objects are allocated
5204 * and freed.
5205 */
5206
5207struct location {
5208 depot_stack_handle_t handle;
5209 unsigned long count;
5210 unsigned long addr;
5211 unsigned long waste;
5212 long long sum_time;
5213 long min_time;
5214 long max_time;
5215 long min_pid;
5216 long max_pid;
5217 DECLARE_BITMAP(cpus, NR_CPUS);
5218 nodemask_t nodes;
5219};
5220
5221struct loc_track {
5222 unsigned long max;
5223 unsigned long count;
5224 struct location *loc;
5225 loff_t idx;
5226};
5227
5228static struct dentry *slab_debugfs_root;
5229
5230static void free_loc_track(struct loc_track *t)
5231{
5232 if (t->max)
5233 free_pages((unsigned long)t->loc,
5234 get_order(sizeof(struct location) * t->max));
5235}
5236
5237static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5238{
5239 struct location *l;
5240 int order;
5241
5242 order = get_order(sizeof(struct location) * max);
5243
5244 l = (void *)__get_free_pages(flags, order);
5245 if (!l)
5246 return 0;
5247
5248 if (t->count) {
5249 memcpy(l, t->loc, sizeof(struct location) * t->count);
5250 free_loc_track(t);
5251 }
5252 t->max = max;
5253 t->loc = l;
5254 return 1;
5255}
5256
5257static int add_location(struct loc_track *t, struct kmem_cache *s,
5258 const struct track *track,
5259 unsigned int orig_size)
5260{
5261 long start, end, pos;
5262 struct location *l;
5263 unsigned long caddr, chandle, cwaste;
5264 unsigned long age = jiffies - track->when;
5265 depot_stack_handle_t handle = 0;
5266 unsigned int waste = s->object_size - orig_size;
5267
5268#ifdef CONFIG_STACKDEPOT
5269 handle = READ_ONCE(track->handle);
5270#endif
5271 start = -1;
5272 end = t->count;
5273
5274 for ( ; ; ) {
5275 pos = start + (end - start + 1) / 2;
5276
5277 /*
5278 * There is nothing at "end". If we end up there
5279 * we need to add something to before end.
5280 */
5281 if (pos == end)
5282 break;
5283
5284 l = &t->loc[pos];
5285 caddr = l->addr;
5286 chandle = l->handle;
5287 cwaste = l->waste;
5288 if ((track->addr == caddr) && (handle == chandle) &&
5289 (waste == cwaste)) {
5290
5291 l->count++;
5292 if (track->when) {
5293 l->sum_time += age;
5294 if (age < l->min_time)
5295 l->min_time = age;
5296 if (age > l->max_time)
5297 l->max_time = age;
5298
5299 if (track->pid < l->min_pid)
5300 l->min_pid = track->pid;
5301 if (track->pid > l->max_pid)
5302 l->max_pid = track->pid;
5303
5304 cpumask_set_cpu(track->cpu,
5305 to_cpumask(l->cpus));
5306 }
5307 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5308 return 1;
5309 }
5310
5311 if (track->addr < caddr)
5312 end = pos;
5313 else if (track->addr == caddr && handle < chandle)
5314 end = pos;
5315 else if (track->addr == caddr && handle == chandle &&
5316 waste < cwaste)
5317 end = pos;
5318 else
5319 start = pos;
5320 }
5321
5322 /*
5323 * Not found. Insert new tracking element.
5324 */
5325 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5326 return 0;
5327
5328 l = t->loc + pos;
5329 if (pos < t->count)
5330 memmove(l + 1, l,
5331 (t->count - pos) * sizeof(struct location));
5332 t->count++;
5333 l->count = 1;
5334 l->addr = track->addr;
5335 l->sum_time = age;
5336 l->min_time = age;
5337 l->max_time = age;
5338 l->min_pid = track->pid;
5339 l->max_pid = track->pid;
5340 l->handle = handle;
5341 l->waste = waste;
5342 cpumask_clear(to_cpumask(l->cpus));
5343 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5344 nodes_clear(l->nodes);
5345 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5346 return 1;
5347}
5348
5349static void process_slab(struct loc_track *t, struct kmem_cache *s,
5350 struct slab *slab, enum track_item alloc,
5351 unsigned long *obj_map)
5352{
5353 void *addr = slab_address(slab);
5354 bool is_alloc = (alloc == TRACK_ALLOC);
5355 void *p;
5356
5357 __fill_map(obj_map, s, slab);
5358
5359 for_each_object(p, s, addr, slab->objects)
5360 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5361 add_location(t, s, get_track(s, p, alloc),
5362 is_alloc ? get_orig_size(s, p) :
5363 s->object_size);
5364}
5365#endif /* CONFIG_DEBUG_FS */
5366#endif /* CONFIG_SLUB_DEBUG */
5367
5368#ifdef SLAB_SUPPORTS_SYSFS
5369enum slab_stat_type {
5370 SL_ALL, /* All slabs */
5371 SL_PARTIAL, /* Only partially allocated slabs */
5372 SL_CPU, /* Only slabs used for cpu caches */
5373 SL_OBJECTS, /* Determine allocated objects not slabs */
5374 SL_TOTAL /* Determine object capacity not slabs */
5375};
5376
5377#define SO_ALL (1 << SL_ALL)
5378#define SO_PARTIAL (1 << SL_PARTIAL)
5379#define SO_CPU (1 << SL_CPU)
5380#define SO_OBJECTS (1 << SL_OBJECTS)
5381#define SO_TOTAL (1 << SL_TOTAL)
5382
5383static ssize_t show_slab_objects(struct kmem_cache *s,
5384 char *buf, unsigned long flags)
5385{
5386 unsigned long total = 0;
5387 int node;
5388 int x;
5389 unsigned long *nodes;
5390 int len = 0;
5391
5392 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
5393 if (!nodes)
5394 return -ENOMEM;
5395
5396 if (flags & SO_CPU) {
5397 int cpu;
5398
5399 for_each_possible_cpu(cpu) {
5400 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
5401 cpu);
5402 int node;
5403 struct slab *slab;
5404
5405 slab = READ_ONCE(c->slab);
5406 if (!slab)
5407 continue;
5408
5409 node = slab_nid(slab);
5410 if (flags & SO_TOTAL)
5411 x = slab->objects;
5412 else if (flags & SO_OBJECTS)
5413 x = slab->inuse;
5414 else
5415 x = 1;
5416
5417 total += x;
5418 nodes[node] += x;
5419
5420#ifdef CONFIG_SLUB_CPU_PARTIAL
5421 slab = slub_percpu_partial_read_once(c);
5422 if (slab) {
5423 node = slab_nid(slab);
5424 if (flags & SO_TOTAL)
5425 WARN_ON_ONCE(1);
5426 else if (flags & SO_OBJECTS)
5427 WARN_ON_ONCE(1);
5428 else
5429 x = slab->slabs;
5430 total += x;
5431 nodes[node] += x;
5432 }
5433#endif
5434 }
5435 }
5436
5437 /*
5438 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
5439 * already held which will conflict with an existing lock order:
5440 *
5441 * mem_hotplug_lock->slab_mutex->kernfs_mutex
5442 *
5443 * We don't really need mem_hotplug_lock (to hold off
5444 * slab_mem_going_offline_callback) here because slab's memory hot
5445 * unplug code doesn't destroy the kmem_cache->node[] data.
5446 */
5447
5448#ifdef CONFIG_SLUB_DEBUG
5449 if (flags & SO_ALL) {
5450 struct kmem_cache_node *n;
5451
5452 for_each_kmem_cache_node(s, node, n) {
5453
5454 if (flags & SO_TOTAL)
5455 x = atomic_long_read(&n->total_objects);
5456 else if (flags & SO_OBJECTS)
5457 x = atomic_long_read(&n->total_objects) -
5458 count_partial(n, count_free);
5459 else
5460 x = atomic_long_read(&n->nr_slabs);
5461 total += x;
5462 nodes[node] += x;
5463 }
5464
5465 } else
5466#endif
5467 if (flags & SO_PARTIAL) {
5468 struct kmem_cache_node *n;
5469
5470 for_each_kmem_cache_node(s, node, n) {
5471 if (flags & SO_TOTAL)
5472 x = count_partial(n, count_total);
5473 else if (flags & SO_OBJECTS)
5474 x = count_partial(n, count_inuse);
5475 else
5476 x = n->nr_partial;
5477 total += x;
5478 nodes[node] += x;
5479 }
5480 }
5481
5482 len += sysfs_emit_at(buf, len, "%lu", total);
5483#ifdef CONFIG_NUMA
5484 for (node = 0; node < nr_node_ids; node++) {
5485 if (nodes[node])
5486 len += sysfs_emit_at(buf, len, " N%d=%lu",
5487 node, nodes[node]);
5488 }
5489#endif
5490 len += sysfs_emit_at(buf, len, "\n");
5491 kfree(nodes);
5492
5493 return len;
5494}
5495
5496#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
5497#define to_slab(n) container_of(n, struct kmem_cache, kobj)
5498
5499struct slab_attribute {
5500 struct attribute attr;
5501 ssize_t (*show)(struct kmem_cache *s, char *buf);
5502 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
5503};
5504
5505#define SLAB_ATTR_RO(_name) \
5506 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
5507
5508#define SLAB_ATTR(_name) \
5509 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
5510
5511static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
5512{
5513 return sysfs_emit(buf, "%u\n", s->size);
5514}
5515SLAB_ATTR_RO(slab_size);
5516
5517static ssize_t align_show(struct kmem_cache *s, char *buf)
5518{
5519 return sysfs_emit(buf, "%u\n", s->align);
5520}
5521SLAB_ATTR_RO(align);
5522
5523static ssize_t object_size_show(struct kmem_cache *s, char *buf)
5524{
5525 return sysfs_emit(buf, "%u\n", s->object_size);
5526}
5527SLAB_ATTR_RO(object_size);
5528
5529static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
5530{
5531 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
5532}
5533SLAB_ATTR_RO(objs_per_slab);
5534
5535static ssize_t order_show(struct kmem_cache *s, char *buf)
5536{
5537 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
5538}
5539SLAB_ATTR_RO(order);
5540
5541static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
5542{
5543 return sysfs_emit(buf, "%lu\n", s->min_partial);
5544}
5545
5546static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
5547 size_t length)
5548{
5549 unsigned long min;
5550 int err;
5551
5552 err = kstrtoul(buf, 10, &min);
5553 if (err)
5554 return err;
5555
5556 s->min_partial = min;
5557 return length;
5558}
5559SLAB_ATTR(min_partial);
5560
5561static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
5562{
5563 unsigned int nr_partial = 0;
5564#ifdef CONFIG_SLUB_CPU_PARTIAL
5565 nr_partial = s->cpu_partial;
5566#endif
5567
5568 return sysfs_emit(buf, "%u\n", nr_partial);
5569}
5570
5571static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
5572 size_t length)
5573{
5574 unsigned int objects;
5575 int err;
5576
5577 err = kstrtouint(buf, 10, &objects);
5578 if (err)
5579 return err;
5580 if (objects && !kmem_cache_has_cpu_partial(s))
5581 return -EINVAL;
5582
5583 slub_set_cpu_partial(s, objects);
5584 flush_all(s);
5585 return length;
5586}
5587SLAB_ATTR(cpu_partial);
5588
5589static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5590{
5591 if (!s->ctor)
5592 return 0;
5593 return sysfs_emit(buf, "%pS\n", s->ctor);
5594}
5595SLAB_ATTR_RO(ctor);
5596
5597static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5598{
5599 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5600}
5601SLAB_ATTR_RO(aliases);
5602
5603static ssize_t partial_show(struct kmem_cache *s, char *buf)
5604{
5605 return show_slab_objects(s, buf, SO_PARTIAL);
5606}
5607SLAB_ATTR_RO(partial);
5608
5609static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5610{
5611 return show_slab_objects(s, buf, SO_CPU);
5612}
5613SLAB_ATTR_RO(cpu_slabs);
5614
5615static ssize_t objects_show(struct kmem_cache *s, char *buf)
5616{
5617 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5618}
5619SLAB_ATTR_RO(objects);
5620
5621static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5622{
5623 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5624}
5625SLAB_ATTR_RO(objects_partial);
5626
5627static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5628{
5629 int objects = 0;
5630 int slabs = 0;
5631 int cpu __maybe_unused;
5632 int len = 0;
5633
5634#ifdef CONFIG_SLUB_CPU_PARTIAL
5635 for_each_online_cpu(cpu) {
5636 struct slab *slab;
5637
5638 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5639
5640 if (slab)
5641 slabs += slab->slabs;
5642 }
5643#endif
5644
5645 /* Approximate half-full slabs, see slub_set_cpu_partial() */
5646 objects = (slabs * oo_objects(s->oo)) / 2;
5647 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
5648
5649#if defined(CONFIG_SLUB_CPU_PARTIAL) && defined(CONFIG_SMP)
5650 for_each_online_cpu(cpu) {
5651 struct slab *slab;
5652
5653 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5654 if (slab) {
5655 slabs = READ_ONCE(slab->slabs);
5656 objects = (slabs * oo_objects(s->oo)) / 2;
5657 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
5658 cpu, objects, slabs);
5659 }
5660 }
5661#endif
5662 len += sysfs_emit_at(buf, len, "\n");
5663
5664 return len;
5665}
5666SLAB_ATTR_RO(slabs_cpu_partial);
5667
5668static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5669{
5670 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5671}
5672SLAB_ATTR_RO(reclaim_account);
5673
5674static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5675{
5676 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5677}
5678SLAB_ATTR_RO(hwcache_align);
5679
5680#ifdef CONFIG_ZONE_DMA
5681static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5682{
5683 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5684}
5685SLAB_ATTR_RO(cache_dma);
5686#endif
5687
5688#ifdef CONFIG_HARDENED_USERCOPY
5689static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5690{
5691 return sysfs_emit(buf, "%u\n", s->usersize);
5692}
5693SLAB_ATTR_RO(usersize);
5694#endif
5695
5696static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5697{
5698 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5699}
5700SLAB_ATTR_RO(destroy_by_rcu);
5701
5702#ifdef CONFIG_SLUB_DEBUG
5703static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5704{
5705 return show_slab_objects(s, buf, SO_ALL);
5706}
5707SLAB_ATTR_RO(slabs);
5708
5709static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5710{
5711 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5712}
5713SLAB_ATTR_RO(total_objects);
5714
5715static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5716{
5717 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5718}
5719SLAB_ATTR_RO(sanity_checks);
5720
5721static ssize_t trace_show(struct kmem_cache *s, char *buf)
5722{
5723 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5724}
5725SLAB_ATTR_RO(trace);
5726
5727static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5728{
5729 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5730}
5731
5732SLAB_ATTR_RO(red_zone);
5733
5734static ssize_t poison_show(struct kmem_cache *s, char *buf)
5735{
5736 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
5737}
5738
5739SLAB_ATTR_RO(poison);
5740
5741static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5742{
5743 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5744}
5745
5746SLAB_ATTR_RO(store_user);
5747
5748static ssize_t validate_show(struct kmem_cache *s, char *buf)
5749{
5750 return 0;
5751}
5752
5753static ssize_t validate_store(struct kmem_cache *s,
5754 const char *buf, size_t length)
5755{
5756 int ret = -EINVAL;
5757
5758 if (buf[0] == '1' && kmem_cache_debug(s)) {
5759 ret = validate_slab_cache(s);
5760 if (ret >= 0)
5761 ret = length;
5762 }
5763 return ret;
5764}
5765SLAB_ATTR(validate);
5766
5767#endif /* CONFIG_SLUB_DEBUG */
5768
5769#ifdef CONFIG_FAILSLAB
5770static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5771{
5772 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5773}
5774
5775static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5776 size_t length)
5777{
5778 if (s->refcount > 1)
5779 return -EINVAL;
5780
5781 if (buf[0] == '1')
5782 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
5783 else
5784 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
5785
5786 return length;
5787}
5788SLAB_ATTR(failslab);
5789#endif
5790
5791static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5792{
5793 return 0;
5794}
5795
5796static ssize_t shrink_store(struct kmem_cache *s,
5797 const char *buf, size_t length)
5798{
5799 if (buf[0] == '1')
5800 kmem_cache_shrink(s);
5801 else
5802 return -EINVAL;
5803 return length;
5804}
5805SLAB_ATTR(shrink);
5806
5807#ifdef CONFIG_NUMA
5808static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5809{
5810 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5811}
5812
5813static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5814 const char *buf, size_t length)
5815{
5816 unsigned int ratio;
5817 int err;
5818
5819 err = kstrtouint(buf, 10, &ratio);
5820 if (err)
5821 return err;
5822 if (ratio > 100)
5823 return -ERANGE;
5824
5825 s->remote_node_defrag_ratio = ratio * 10;
5826
5827 return length;
5828}
5829SLAB_ATTR(remote_node_defrag_ratio);
5830#endif
5831
5832#ifdef CONFIG_SLUB_STATS
5833static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5834{
5835 unsigned long sum = 0;
5836 int cpu;
5837 int len = 0;
5838 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5839
5840 if (!data)
5841 return -ENOMEM;
5842
5843 for_each_online_cpu(cpu) {
5844 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5845
5846 data[cpu] = x;
5847 sum += x;
5848 }
5849
5850 len += sysfs_emit_at(buf, len, "%lu", sum);
5851
5852#ifdef CONFIG_SMP
5853 for_each_online_cpu(cpu) {
5854 if (data[cpu])
5855 len += sysfs_emit_at(buf, len, " C%d=%u",
5856 cpu, data[cpu]);
5857 }
5858#endif
5859 kfree(data);
5860 len += sysfs_emit_at(buf, len, "\n");
5861
5862 return len;
5863}
5864
5865static void clear_stat(struct kmem_cache *s, enum stat_item si)
5866{
5867 int cpu;
5868
5869 for_each_online_cpu(cpu)
5870 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5871}
5872
5873#define STAT_ATTR(si, text) \
5874static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5875{ \
5876 return show_stat(s, buf, si); \
5877} \
5878static ssize_t text##_store(struct kmem_cache *s, \
5879 const char *buf, size_t length) \
5880{ \
5881 if (buf[0] != '0') \
5882 return -EINVAL; \
5883 clear_stat(s, si); \
5884 return length; \
5885} \
5886SLAB_ATTR(text); \
5887
5888STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5889STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5890STAT_ATTR(FREE_FASTPATH, free_fastpath);
5891STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5892STAT_ATTR(FREE_FROZEN, free_frozen);
5893STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5894STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5895STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5896STAT_ATTR(ALLOC_SLAB, alloc_slab);
5897STAT_ATTR(ALLOC_REFILL, alloc_refill);
5898STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5899STAT_ATTR(FREE_SLAB, free_slab);
5900STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5901STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5902STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5903STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5904STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5905STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5906STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5907STAT_ATTR(ORDER_FALLBACK, order_fallback);
5908STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5909STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5910STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5911STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5912STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5913STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5914#endif /* CONFIG_SLUB_STATS */
5915
5916#ifdef CONFIG_KFENCE
5917static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
5918{
5919 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
5920}
5921
5922static ssize_t skip_kfence_store(struct kmem_cache *s,
5923 const char *buf, size_t length)
5924{
5925 int ret = length;
5926
5927 if (buf[0] == '0')
5928 s->flags &= ~SLAB_SKIP_KFENCE;
5929 else if (buf[0] == '1')
5930 s->flags |= SLAB_SKIP_KFENCE;
5931 else
5932 ret = -EINVAL;
5933
5934 return ret;
5935}
5936SLAB_ATTR(skip_kfence);
5937#endif
5938
5939static struct attribute *slab_attrs[] = {
5940 &slab_size_attr.attr,
5941 &object_size_attr.attr,
5942 &objs_per_slab_attr.attr,
5943 &order_attr.attr,
5944 &min_partial_attr.attr,
5945 &cpu_partial_attr.attr,
5946 &objects_attr.attr,
5947 &objects_partial_attr.attr,
5948 &partial_attr.attr,
5949 &cpu_slabs_attr.attr,
5950 &ctor_attr.attr,
5951 &aliases_attr.attr,
5952 &align_attr.attr,
5953 &hwcache_align_attr.attr,
5954 &reclaim_account_attr.attr,
5955 &destroy_by_rcu_attr.attr,
5956 &shrink_attr.attr,
5957 &slabs_cpu_partial_attr.attr,
5958#ifdef CONFIG_SLUB_DEBUG
5959 &total_objects_attr.attr,
5960 &slabs_attr.attr,
5961 &sanity_checks_attr.attr,
5962 &trace_attr.attr,
5963 &red_zone_attr.attr,
5964 &poison_attr.attr,
5965 &store_user_attr.attr,
5966 &validate_attr.attr,
5967#endif
5968#ifdef CONFIG_ZONE_DMA
5969 &cache_dma_attr.attr,
5970#endif
5971#ifdef CONFIG_NUMA
5972 &remote_node_defrag_ratio_attr.attr,
5973#endif
5974#ifdef CONFIG_SLUB_STATS
5975 &alloc_fastpath_attr.attr,
5976 &alloc_slowpath_attr.attr,
5977 &free_fastpath_attr.attr,
5978 &free_slowpath_attr.attr,
5979 &free_frozen_attr.attr,
5980 &free_add_partial_attr.attr,
5981 &free_remove_partial_attr.attr,
5982 &alloc_from_partial_attr.attr,
5983 &alloc_slab_attr.attr,
5984 &alloc_refill_attr.attr,
5985 &alloc_node_mismatch_attr.attr,
5986 &free_slab_attr.attr,
5987 &cpuslab_flush_attr.attr,
5988 &deactivate_full_attr.attr,
5989 &deactivate_empty_attr.attr,
5990 &deactivate_to_head_attr.attr,
5991 &deactivate_to_tail_attr.attr,
5992 &deactivate_remote_frees_attr.attr,
5993 &deactivate_bypass_attr.attr,
5994 &order_fallback_attr.attr,
5995 &cmpxchg_double_fail_attr.attr,
5996 &cmpxchg_double_cpu_fail_attr.attr,
5997 &cpu_partial_alloc_attr.attr,
5998 &cpu_partial_free_attr.attr,
5999 &cpu_partial_node_attr.attr,
6000 &cpu_partial_drain_attr.attr,
6001#endif
6002#ifdef CONFIG_FAILSLAB
6003 &failslab_attr.attr,
6004#endif
6005#ifdef CONFIG_HARDENED_USERCOPY
6006 &usersize_attr.attr,
6007#endif
6008#ifdef CONFIG_KFENCE
6009 &skip_kfence_attr.attr,
6010#endif
6011
6012 NULL
6013};
6014
6015static const struct attribute_group slab_attr_group = {
6016 .attrs = slab_attrs,
6017};
6018
6019static ssize_t slab_attr_show(struct kobject *kobj,
6020 struct attribute *attr,
6021 char *buf)
6022{
6023 struct slab_attribute *attribute;
6024 struct kmem_cache *s;
6025
6026 attribute = to_slab_attr(attr);
6027 s = to_slab(kobj);
6028
6029 if (!attribute->show)
6030 return -EIO;
6031
6032 return attribute->show(s, buf);
6033}
6034
6035static ssize_t slab_attr_store(struct kobject *kobj,
6036 struct attribute *attr,
6037 const char *buf, size_t len)
6038{
6039 struct slab_attribute *attribute;
6040 struct kmem_cache *s;
6041
6042 attribute = to_slab_attr(attr);
6043 s = to_slab(kobj);
6044
6045 if (!attribute->store)
6046 return -EIO;
6047
6048 return attribute->store(s, buf, len);
6049}
6050
6051static void kmem_cache_release(struct kobject *k)
6052{
6053 slab_kmem_cache_release(to_slab(k));
6054}
6055
6056static const struct sysfs_ops slab_sysfs_ops = {
6057 .show = slab_attr_show,
6058 .store = slab_attr_store,
6059};
6060
6061static struct kobj_type slab_ktype = {
6062 .sysfs_ops = &slab_sysfs_ops,
6063 .release = kmem_cache_release,
6064};
6065
6066static struct kset *slab_kset;
6067
6068static inline struct kset *cache_kset(struct kmem_cache *s)
6069{
6070 return slab_kset;
6071}
6072
6073#define ID_STR_LENGTH 32
6074
6075/* Create a unique string id for a slab cache:
6076 *
6077 * Format :[flags-]size
6078 */
6079static char *create_unique_id(struct kmem_cache *s)
6080{
6081 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6082 char *p = name;
6083
6084 if (!name)
6085 return ERR_PTR(-ENOMEM);
6086
6087 *p++ = ':';
6088 /*
6089 * First flags affecting slabcache operations. We will only
6090 * get here for aliasable slabs so we do not need to support
6091 * too many flags. The flags here must cover all flags that
6092 * are matched during merging to guarantee that the id is
6093 * unique.
6094 */
6095 if (s->flags & SLAB_CACHE_DMA)
6096 *p++ = 'd';
6097 if (s->flags & SLAB_CACHE_DMA32)
6098 *p++ = 'D';
6099 if (s->flags & SLAB_RECLAIM_ACCOUNT)
6100 *p++ = 'a';
6101 if (s->flags & SLAB_CONSISTENCY_CHECKS)
6102 *p++ = 'F';
6103 if (s->flags & SLAB_ACCOUNT)
6104 *p++ = 'A';
6105 if (p != name + 1)
6106 *p++ = '-';
6107 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6108
6109 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6110 kfree(name);
6111 return ERR_PTR(-EINVAL);
6112 }
6113 kmsan_unpoison_memory(name, p - name);
6114 return name;
6115}
6116
6117static int sysfs_slab_add(struct kmem_cache *s)
6118{
6119 int err;
6120 const char *name;
6121 struct kset *kset = cache_kset(s);
6122 int unmergeable = slab_unmergeable(s);
6123
6124 if (!unmergeable && disable_higher_order_debug &&
6125 (slub_debug & DEBUG_METADATA_FLAGS))
6126 unmergeable = 1;
6127
6128 if (unmergeable) {
6129 /*
6130 * Slabcache can never be merged so we can use the name proper.
6131 * This is typically the case for debug situations. In that
6132 * case we can catch duplicate names easily.
6133 */
6134 sysfs_remove_link(&slab_kset->kobj, s->name);
6135 name = s->name;
6136 } else {
6137 /*
6138 * Create a unique name for the slab as a target
6139 * for the symlinks.
6140 */
6141 name = create_unique_id(s);
6142 if (IS_ERR(name))
6143 return PTR_ERR(name);
6144 }
6145
6146 s->kobj.kset = kset;
6147 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6148 if (err)
6149 goto out;
6150
6151 err = sysfs_create_group(&s->kobj, &slab_attr_group);
6152 if (err)
6153 goto out_del_kobj;
6154
6155 if (!unmergeable) {
6156 /* Setup first alias */
6157 sysfs_slab_alias(s, s->name);
6158 }
6159out:
6160 if (!unmergeable)
6161 kfree(name);
6162 return err;
6163out_del_kobj:
6164 kobject_del(&s->kobj);
6165 goto out;
6166}
6167
6168void sysfs_slab_unlink(struct kmem_cache *s)
6169{
6170 if (slab_state >= FULL)
6171 kobject_del(&s->kobj);
6172}
6173
6174void sysfs_slab_release(struct kmem_cache *s)
6175{
6176 if (slab_state >= FULL)
6177 kobject_put(&s->kobj);
6178}
6179
6180/*
6181 * Need to buffer aliases during bootup until sysfs becomes
6182 * available lest we lose that information.
6183 */
6184struct saved_alias {
6185 struct kmem_cache *s;
6186 const char *name;
6187 struct saved_alias *next;
6188};
6189
6190static struct saved_alias *alias_list;
6191
6192static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6193{
6194 struct saved_alias *al;
6195
6196 if (slab_state == FULL) {
6197 /*
6198 * If we have a leftover link then remove it.
6199 */
6200 sysfs_remove_link(&slab_kset->kobj, name);
6201 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6202 }
6203
6204 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6205 if (!al)
6206 return -ENOMEM;
6207
6208 al->s = s;
6209 al->name = name;
6210 al->next = alias_list;
6211 alias_list = al;
6212 kmsan_unpoison_memory(al, sizeof(*al));
6213 return 0;
6214}
6215
6216static int __init slab_sysfs_init(void)
6217{
6218 struct kmem_cache *s;
6219 int err;
6220
6221 mutex_lock(&slab_mutex);
6222
6223 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6224 if (!slab_kset) {
6225 mutex_unlock(&slab_mutex);
6226 pr_err("Cannot register slab subsystem.\n");
6227 return -ENOSYS;
6228 }
6229
6230 slab_state = FULL;
6231
6232 list_for_each_entry(s, &slab_caches, list) {
6233 err = sysfs_slab_add(s);
6234 if (err)
6235 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6236 s->name);
6237 }
6238
6239 while (alias_list) {
6240 struct saved_alias *al = alias_list;
6241
6242 alias_list = alias_list->next;
6243 err = sysfs_slab_alias(al->s, al->name);
6244 if (err)
6245 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6246 al->name);
6247 kfree(al);
6248 }
6249
6250 mutex_unlock(&slab_mutex);
6251 return 0;
6252}
6253late_initcall(slab_sysfs_init);
6254#endif /* SLAB_SUPPORTS_SYSFS */
6255
6256#if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6257static int slab_debugfs_show(struct seq_file *seq, void *v)
6258{
6259 struct loc_track *t = seq->private;
6260 struct location *l;
6261 unsigned long idx;
6262
6263 idx = (unsigned long) t->idx;
6264 if (idx < t->count) {
6265 l = &t->loc[idx];
6266
6267 seq_printf(seq, "%7ld ", l->count);
6268
6269 if (l->addr)
6270 seq_printf(seq, "%pS", (void *)l->addr);
6271 else
6272 seq_puts(seq, "<not-available>");
6273
6274 if (l->waste)
6275 seq_printf(seq, " waste=%lu/%lu",
6276 l->count * l->waste, l->waste);
6277
6278 if (l->sum_time != l->min_time) {
6279 seq_printf(seq, " age=%ld/%llu/%ld",
6280 l->min_time, div_u64(l->sum_time, l->count),
6281 l->max_time);
6282 } else
6283 seq_printf(seq, " age=%ld", l->min_time);
6284
6285 if (l->min_pid != l->max_pid)
6286 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6287 else
6288 seq_printf(seq, " pid=%ld",
6289 l->min_pid);
6290
6291 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6292 seq_printf(seq, " cpus=%*pbl",
6293 cpumask_pr_args(to_cpumask(l->cpus)));
6294
6295 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6296 seq_printf(seq, " nodes=%*pbl",
6297 nodemask_pr_args(&l->nodes));
6298
6299#ifdef CONFIG_STACKDEPOT
6300 {
6301 depot_stack_handle_t handle;
6302 unsigned long *entries;
6303 unsigned int nr_entries, j;
6304
6305 handle = READ_ONCE(l->handle);
6306 if (handle) {
6307 nr_entries = stack_depot_fetch(handle, &entries);
6308 seq_puts(seq, "\n");
6309 for (j = 0; j < nr_entries; j++)
6310 seq_printf(seq, " %pS\n", (void *)entries[j]);
6311 }
6312 }
6313#endif
6314 seq_puts(seq, "\n");
6315 }
6316
6317 if (!idx && !t->count)
6318 seq_puts(seq, "No data\n");
6319
6320 return 0;
6321}
6322
6323static void slab_debugfs_stop(struct seq_file *seq, void *v)
6324{
6325}
6326
6327static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6328{
6329 struct loc_track *t = seq->private;
6330
6331 t->idx = ++(*ppos);
6332 if (*ppos <= t->count)
6333 return ppos;
6334
6335 return NULL;
6336}
6337
6338static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6339{
6340 struct location *loc1 = (struct location *)a;
6341 struct location *loc2 = (struct location *)b;
6342
6343 if (loc1->count > loc2->count)
6344 return -1;
6345 else
6346 return 1;
6347}
6348
6349static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6350{
6351 struct loc_track *t = seq->private;
6352
6353 t->idx = *ppos;
6354 return ppos;
6355}
6356
6357static const struct seq_operations slab_debugfs_sops = {
6358 .start = slab_debugfs_start,
6359 .next = slab_debugfs_next,
6360 .stop = slab_debugfs_stop,
6361 .show = slab_debugfs_show,
6362};
6363
6364static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6365{
6366
6367 struct kmem_cache_node *n;
6368 enum track_item alloc;
6369 int node;
6370 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6371 sizeof(struct loc_track));
6372 struct kmem_cache *s = file_inode(filep)->i_private;
6373 unsigned long *obj_map;
6374
6375 if (!t)
6376 return -ENOMEM;
6377
6378 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6379 if (!obj_map) {
6380 seq_release_private(inode, filep);
6381 return -ENOMEM;
6382 }
6383
6384 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6385 alloc = TRACK_ALLOC;
6386 else
6387 alloc = TRACK_FREE;
6388
6389 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
6390 bitmap_free(obj_map);
6391 seq_release_private(inode, filep);
6392 return -ENOMEM;
6393 }
6394
6395 for_each_kmem_cache_node(s, node, n) {
6396 unsigned long flags;
6397 struct slab *slab;
6398
6399 if (!atomic_long_read(&n->nr_slabs))
6400 continue;
6401
6402 spin_lock_irqsave(&n->list_lock, flags);
6403 list_for_each_entry(slab, &n->partial, slab_list)
6404 process_slab(t, s, slab, alloc, obj_map);
6405 list_for_each_entry(slab, &n->full, slab_list)
6406 process_slab(t, s, slab, alloc, obj_map);
6407 spin_unlock_irqrestore(&n->list_lock, flags);
6408 }
6409
6410 /* Sort locations by count */
6411 sort_r(t->loc, t->count, sizeof(struct location),
6412 cmp_loc_by_count, NULL, NULL);
6413
6414 bitmap_free(obj_map);
6415 return 0;
6416}
6417
6418static int slab_debug_trace_release(struct inode *inode, struct file *file)
6419{
6420 struct seq_file *seq = file->private_data;
6421 struct loc_track *t = seq->private;
6422
6423 free_loc_track(t);
6424 return seq_release_private(inode, file);
6425}
6426
6427static const struct file_operations slab_debugfs_fops = {
6428 .open = slab_debug_trace_open,
6429 .read = seq_read,
6430 .llseek = seq_lseek,
6431 .release = slab_debug_trace_release,
6432};
6433
6434static void debugfs_slab_add(struct kmem_cache *s)
6435{
6436 struct dentry *slab_cache_dir;
6437
6438 if (unlikely(!slab_debugfs_root))
6439 return;
6440
6441 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
6442
6443 debugfs_create_file("alloc_traces", 0400,
6444 slab_cache_dir, s, &slab_debugfs_fops);
6445
6446 debugfs_create_file("free_traces", 0400,
6447 slab_cache_dir, s, &slab_debugfs_fops);
6448}
6449
6450void debugfs_slab_release(struct kmem_cache *s)
6451{
6452 debugfs_remove_recursive(debugfs_lookup(s->name, slab_debugfs_root));
6453}
6454
6455static int __init slab_debugfs_init(void)
6456{
6457 struct kmem_cache *s;
6458
6459 slab_debugfs_root = debugfs_create_dir("slab", NULL);
6460
6461 list_for_each_entry(s, &slab_caches, list)
6462 if (s->flags & SLAB_STORE_USER)
6463 debugfs_slab_add(s);
6464
6465 return 0;
6466
6467}
6468__initcall(slab_debugfs_init);
6469#endif
6470/*
6471 * The /proc/slabinfo ABI
6472 */
6473#ifdef CONFIG_SLUB_DEBUG
6474void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
6475{
6476 unsigned long nr_slabs = 0;
6477 unsigned long nr_objs = 0;
6478 unsigned long nr_free = 0;
6479 int node;
6480 struct kmem_cache_node *n;
6481
6482 for_each_kmem_cache_node(s, node, n) {
6483 nr_slabs += node_nr_slabs(n);
6484 nr_objs += node_nr_objs(n);
6485 nr_free += count_partial(n, count_free);
6486 }
6487
6488 sinfo->active_objs = nr_objs - nr_free;
6489 sinfo->num_objs = nr_objs;
6490 sinfo->active_slabs = nr_slabs;
6491 sinfo->num_slabs = nr_slabs;
6492 sinfo->objects_per_slab = oo_objects(s->oo);
6493 sinfo->cache_order = oo_order(s->oo);
6494}
6495
6496void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
6497{
6498}
6499
6500ssize_t slabinfo_write(struct file *file, const char __user *buffer,
6501 size_t count, loff_t *ppos)
6502{
6503 return -EIO;
6504}
6505#endif /* CONFIG_SLUB_DEBUG */
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> /* mm_account_reclaimed_pages() */
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/kmsan.h>
26#include <linux/cpu.h>
27#include <linux/cpuset.h>
28#include <linux/mempolicy.h>
29#include <linux/ctype.h>
30#include <linux/stackdepot.h>
31#include <linux/debugobjects.h>
32#include <linux/kallsyms.h>
33#include <linux/kfence.h>
34#include <linux/memory.h>
35#include <linux/math64.h>
36#include <linux/fault-inject.h>
37#include <linux/kmemleak.h>
38#include <linux/stacktrace.h>
39#include <linux/prefetch.h>
40#include <linux/memcontrol.h>
41#include <linux/random.h>
42#include <kunit/test.h>
43#include <kunit/test-bug.h>
44#include <linux/sort.h>
45
46#include <linux/debugfs.h>
47#include <trace/events/kmem.h>
48
49#include "internal.h"
50
51/*
52 * Lock order:
53 * 1. slab_mutex (Global Mutex)
54 * 2. node->list_lock (Spinlock)
55 * 3. kmem_cache->cpu_slab->lock (Local lock)
56 * 4. slab_lock(slab) (Only on some arches)
57 * 5. object_map_lock (Only for debugging)
58 *
59 * slab_mutex
60 *
61 * The role of the slab_mutex is to protect the list of all the slabs
62 * and to synchronize major metadata changes to slab cache structures.
63 * Also synchronizes memory hotplug callbacks.
64 *
65 * slab_lock
66 *
67 * The slab_lock is a wrapper around the page lock, thus it is a bit
68 * spinlock.
69 *
70 * The slab_lock is only used on arches that do not have the ability
71 * to do a cmpxchg_double. It only protects:
72 *
73 * A. slab->freelist -> List of free objects in a slab
74 * B. slab->inuse -> Number of objects in use
75 * C. slab->objects -> Number of objects in slab
76 * D. slab->frozen -> frozen state
77 *
78 * Frozen slabs
79 *
80 * If a slab is frozen then it is exempt from list management. It is
81 * the cpu slab which is actively allocated from by the processor that
82 * froze it and it is not on any list. The processor that froze the
83 * slab is the one who can perform list operations on the slab. Other
84 * processors may put objects onto the freelist but the processor that
85 * froze the slab is the only one that can retrieve the objects from the
86 * slab's freelist.
87 *
88 * CPU partial slabs
89 *
90 * The partially empty slabs cached on the CPU partial list are used
91 * for performance reasons, which speeds up the allocation process.
92 * These slabs are not frozen, but are also exempt from list management,
93 * by clearing the PG_workingset flag when moving out of the node
94 * partial list. Please see __slab_free() for more details.
95 *
96 * To sum up, the current scheme is:
97 * - node partial slab: PG_Workingset && !frozen
98 * - cpu partial slab: !PG_Workingset && !frozen
99 * - cpu slab: !PG_Workingset && frozen
100 * - full slab: !PG_Workingset && !frozen
101 *
102 * list_lock
103 *
104 * The list_lock protects the partial and full list on each node and
105 * the partial slab counter. If taken then no new slabs may be added or
106 * removed from the lists nor make the number of partial slabs be modified.
107 * (Note that the total number of slabs is an atomic value that may be
108 * modified without taking the list lock).
109 *
110 * The list_lock is a centralized lock and thus we avoid taking it as
111 * much as possible. As long as SLUB does not have to handle partial
112 * slabs, operations can continue without any centralized lock. F.e.
113 * allocating a long series of objects that fill up slabs does not require
114 * the list lock.
115 *
116 * For debug caches, all allocations are forced to go through a list_lock
117 * protected region to serialize against concurrent validation.
118 *
119 * cpu_slab->lock local lock
120 *
121 * This locks protect slowpath manipulation of all kmem_cache_cpu fields
122 * except the stat counters. This is a percpu structure manipulated only by
123 * the local cpu, so the lock protects against being preempted or interrupted
124 * by an irq. Fast path operations rely on lockless operations instead.
125 *
126 * On PREEMPT_RT, the local lock neither disables interrupts nor preemption
127 * which means the lockless fastpath cannot be used as it might interfere with
128 * an in-progress slow path operations. In this case the local lock is always
129 * taken but it still utilizes the freelist for the common operations.
130 *
131 * lockless fastpaths
132 *
133 * The fast path allocation (slab_alloc_node()) and freeing (do_slab_free())
134 * are fully lockless when satisfied from the percpu slab (and when
135 * cmpxchg_double is possible to use, otherwise slab_lock is taken).
136 * They also don't disable preemption or migration or irqs. They rely on
137 * the transaction id (tid) field to detect being preempted or moved to
138 * another cpu.
139 *
140 * irq, preemption, migration considerations
141 *
142 * Interrupts are disabled as part of list_lock or local_lock operations, or
143 * around the slab_lock operation, in order to make the slab allocator safe
144 * to use in the context of an irq.
145 *
146 * In addition, preemption (or migration on PREEMPT_RT) is disabled in the
147 * allocation slowpath, bulk allocation, and put_cpu_partial(), so that the
148 * local cpu doesn't change in the process and e.g. the kmem_cache_cpu pointer
149 * doesn't have to be revalidated in each section protected by the local lock.
150 *
151 * SLUB assigns one slab for allocation to each processor.
152 * Allocations only occur from these slabs called cpu slabs.
153 *
154 * Slabs with free elements are kept on a partial list and during regular
155 * operations no list for full slabs is used. If an object in a full slab is
156 * freed then the slab will show up again on the partial lists.
157 * We track full slabs for debugging purposes though because otherwise we
158 * cannot scan all objects.
159 *
160 * Slabs are freed when they become empty. Teardown and setup is
161 * minimal so we rely on the page allocators per cpu caches for
162 * fast frees and allocs.
163 *
164 * slab->frozen The slab is frozen and exempt from list processing.
165 * This means that the slab is dedicated to a purpose
166 * such as satisfying allocations for a specific
167 * processor. Objects may be freed in the slab while
168 * it is frozen but slab_free will then skip the usual
169 * list operations. It is up to the processor holding
170 * the slab to integrate the slab into the slab lists
171 * when the slab is no longer needed.
172 *
173 * One use of this flag is to mark slabs that are
174 * used for allocations. Then such a slab becomes a cpu
175 * slab. The cpu slab may be equipped with an additional
176 * freelist that allows lockless access to
177 * free objects in addition to the regular freelist
178 * that requires the slab lock.
179 *
180 * SLAB_DEBUG_FLAGS Slab requires special handling due to debug
181 * options set. This moves slab handling out of
182 * the fast path and disables lockless freelists.
183 */
184
185/*
186 * We could simply use migrate_disable()/enable() but as long as it's a
187 * function call even on !PREEMPT_RT, use inline preempt_disable() there.
188 */
189#ifndef CONFIG_PREEMPT_RT
190#define slub_get_cpu_ptr(var) get_cpu_ptr(var)
191#define slub_put_cpu_ptr(var) put_cpu_ptr(var)
192#define USE_LOCKLESS_FAST_PATH() (true)
193#else
194#define slub_get_cpu_ptr(var) \
195({ \
196 migrate_disable(); \
197 this_cpu_ptr(var); \
198})
199#define slub_put_cpu_ptr(var) \
200do { \
201 (void)(var); \
202 migrate_enable(); \
203} while (0)
204#define USE_LOCKLESS_FAST_PATH() (false)
205#endif
206
207#ifndef CONFIG_SLUB_TINY
208#define __fastpath_inline __always_inline
209#else
210#define __fastpath_inline
211#endif
212
213#ifdef CONFIG_SLUB_DEBUG
214#ifdef CONFIG_SLUB_DEBUG_ON
215DEFINE_STATIC_KEY_TRUE(slub_debug_enabled);
216#else
217DEFINE_STATIC_KEY_FALSE(slub_debug_enabled);
218#endif
219#endif /* CONFIG_SLUB_DEBUG */
220
221/* Structure holding parameters for get_partial() call chain */
222struct partial_context {
223 gfp_t flags;
224 unsigned int orig_size;
225 void *object;
226};
227
228static inline bool kmem_cache_debug(struct kmem_cache *s)
229{
230 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
231}
232
233static inline bool slub_debug_orig_size(struct kmem_cache *s)
234{
235 return (kmem_cache_debug_flags(s, SLAB_STORE_USER) &&
236 (s->flags & SLAB_KMALLOC));
237}
238
239void *fixup_red_left(struct kmem_cache *s, void *p)
240{
241 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
242 p += s->red_left_pad;
243
244 return p;
245}
246
247static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
248{
249#ifdef CONFIG_SLUB_CPU_PARTIAL
250 return !kmem_cache_debug(s);
251#else
252 return false;
253#endif
254}
255
256/*
257 * Issues still to be resolved:
258 *
259 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
260 *
261 * - Variable sizing of the per node arrays
262 */
263
264/* Enable to log cmpxchg failures */
265#undef SLUB_DEBUG_CMPXCHG
266
267#ifndef CONFIG_SLUB_TINY
268/*
269 * Minimum number of partial slabs. These will be left on the partial
270 * lists even if they are empty. kmem_cache_shrink may reclaim them.
271 */
272#define MIN_PARTIAL 5
273
274/*
275 * Maximum number of desirable partial slabs.
276 * The existence of more partial slabs makes kmem_cache_shrink
277 * sort the partial list by the number of objects in use.
278 */
279#define MAX_PARTIAL 10
280#else
281#define MIN_PARTIAL 0
282#define MAX_PARTIAL 0
283#endif
284
285#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
286 SLAB_POISON | SLAB_STORE_USER)
287
288/*
289 * These debug flags cannot use CMPXCHG because there might be consistency
290 * issues when checking or reading debug information
291 */
292#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
293 SLAB_TRACE)
294
295
296/*
297 * Debugging flags that require metadata to be stored in the slab. These get
298 * disabled when slab_debug=O is used and a cache's min order increases with
299 * metadata.
300 */
301#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
302
303#define OO_SHIFT 16
304#define OO_MASK ((1 << OO_SHIFT) - 1)
305#define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
306
307/* Internal SLUB flags */
308/* Poison object */
309#define __OBJECT_POISON __SLAB_FLAG_BIT(_SLAB_OBJECT_POISON)
310/* Use cmpxchg_double */
311
312#ifdef system_has_freelist_aba
313#define __CMPXCHG_DOUBLE __SLAB_FLAG_BIT(_SLAB_CMPXCHG_DOUBLE)
314#else
315#define __CMPXCHG_DOUBLE __SLAB_FLAG_UNUSED
316#endif
317
318/*
319 * Tracking user of a slab.
320 */
321#define TRACK_ADDRS_COUNT 16
322struct track {
323 unsigned long addr; /* Called from address */
324#ifdef CONFIG_STACKDEPOT
325 depot_stack_handle_t handle;
326#endif
327 int cpu; /* Was running on cpu */
328 int pid; /* Pid context */
329 unsigned long when; /* When did the operation occur */
330};
331
332enum track_item { TRACK_ALLOC, TRACK_FREE };
333
334#ifdef SLAB_SUPPORTS_SYSFS
335static int sysfs_slab_add(struct kmem_cache *);
336static int sysfs_slab_alias(struct kmem_cache *, const char *);
337#else
338static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
339static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
340 { return 0; }
341#endif
342
343#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
344static void debugfs_slab_add(struct kmem_cache *);
345#else
346static inline void debugfs_slab_add(struct kmem_cache *s) { }
347#endif
348
349enum stat_item {
350 ALLOC_FASTPATH, /* Allocation from cpu slab */
351 ALLOC_SLOWPATH, /* Allocation by getting a new cpu slab */
352 FREE_FASTPATH, /* Free to cpu slab */
353 FREE_SLOWPATH, /* Freeing not to cpu slab */
354 FREE_FROZEN, /* Freeing to frozen slab */
355 FREE_ADD_PARTIAL, /* Freeing moves slab to partial list */
356 FREE_REMOVE_PARTIAL, /* Freeing removes last object */
357 ALLOC_FROM_PARTIAL, /* Cpu slab acquired from node partial list */
358 ALLOC_SLAB, /* Cpu slab acquired from page allocator */
359 ALLOC_REFILL, /* Refill cpu slab from slab freelist */
360 ALLOC_NODE_MISMATCH, /* Switching cpu slab */
361 FREE_SLAB, /* Slab freed to the page allocator */
362 CPUSLAB_FLUSH, /* Abandoning of the cpu slab */
363 DEACTIVATE_FULL, /* Cpu slab was full when deactivated */
364 DEACTIVATE_EMPTY, /* Cpu slab was empty when deactivated */
365 DEACTIVATE_TO_HEAD, /* Cpu slab was moved to the head of partials */
366 DEACTIVATE_TO_TAIL, /* Cpu slab was moved to the tail of partials */
367 DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */
368 DEACTIVATE_BYPASS, /* Implicit deactivation */
369 ORDER_FALLBACK, /* Number of times fallback was necessary */
370 CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */
371 CMPXCHG_DOUBLE_FAIL, /* Failures of slab freelist update */
372 CPU_PARTIAL_ALLOC, /* Used cpu partial on alloc */
373 CPU_PARTIAL_FREE, /* Refill cpu partial on free */
374 CPU_PARTIAL_NODE, /* Refill cpu partial from node partial */
375 CPU_PARTIAL_DRAIN, /* Drain cpu partial to node partial */
376 NR_SLUB_STAT_ITEMS
377};
378
379#ifndef CONFIG_SLUB_TINY
380/*
381 * When changing the layout, make sure freelist and tid are still compatible
382 * with this_cpu_cmpxchg_double() alignment requirements.
383 */
384struct kmem_cache_cpu {
385 union {
386 struct {
387 void **freelist; /* Pointer to next available object */
388 unsigned long tid; /* Globally unique transaction id */
389 };
390 freelist_aba_t freelist_tid;
391 };
392 struct slab *slab; /* The slab from which we are allocating */
393#ifdef CONFIG_SLUB_CPU_PARTIAL
394 struct slab *partial; /* Partially allocated slabs */
395#endif
396 local_lock_t lock; /* Protects the fields above */
397#ifdef CONFIG_SLUB_STATS
398 unsigned int stat[NR_SLUB_STAT_ITEMS];
399#endif
400};
401#endif /* CONFIG_SLUB_TINY */
402
403static inline void stat(const struct kmem_cache *s, enum stat_item si)
404{
405#ifdef CONFIG_SLUB_STATS
406 /*
407 * The rmw is racy on a preemptible kernel but this is acceptable, so
408 * avoid this_cpu_add()'s irq-disable overhead.
409 */
410 raw_cpu_inc(s->cpu_slab->stat[si]);
411#endif
412}
413
414static inline
415void stat_add(const struct kmem_cache *s, enum stat_item si, int v)
416{
417#ifdef CONFIG_SLUB_STATS
418 raw_cpu_add(s->cpu_slab->stat[si], v);
419#endif
420}
421
422/*
423 * The slab lists for all objects.
424 */
425struct kmem_cache_node {
426 spinlock_t list_lock;
427 unsigned long nr_partial;
428 struct list_head partial;
429#ifdef CONFIG_SLUB_DEBUG
430 atomic_long_t nr_slabs;
431 atomic_long_t total_objects;
432 struct list_head full;
433#endif
434};
435
436static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
437{
438 return s->node[node];
439}
440
441/*
442 * Iterator over all nodes. The body will be executed for each node that has
443 * a kmem_cache_node structure allocated (which is true for all online nodes)
444 */
445#define for_each_kmem_cache_node(__s, __node, __n) \
446 for (__node = 0; __node < nr_node_ids; __node++) \
447 if ((__n = get_node(__s, __node)))
448
449/*
450 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
451 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
452 * differ during memory hotplug/hotremove operations.
453 * Protected by slab_mutex.
454 */
455static nodemask_t slab_nodes;
456
457#ifndef CONFIG_SLUB_TINY
458/*
459 * Workqueue used for flush_cpu_slab().
460 */
461static struct workqueue_struct *flushwq;
462#endif
463
464/********************************************************************
465 * Core slab cache functions
466 *******************************************************************/
467
468/*
469 * freeptr_t represents a SLUB freelist pointer, which might be encoded
470 * and not dereferenceable if CONFIG_SLAB_FREELIST_HARDENED is enabled.
471 */
472typedef struct { unsigned long v; } freeptr_t;
473
474/*
475 * Returns freelist pointer (ptr). With hardening, this is obfuscated
476 * with an XOR of the address where the pointer is held and a per-cache
477 * random number.
478 */
479static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s,
480 void *ptr, unsigned long ptr_addr)
481{
482 unsigned long encoded;
483
484#ifdef CONFIG_SLAB_FREELIST_HARDENED
485 encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr);
486#else
487 encoded = (unsigned long)ptr;
488#endif
489 return (freeptr_t){.v = encoded};
490}
491
492static inline void *freelist_ptr_decode(const struct kmem_cache *s,
493 freeptr_t ptr, unsigned long ptr_addr)
494{
495 void *decoded;
496
497#ifdef CONFIG_SLAB_FREELIST_HARDENED
498 decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr));
499#else
500 decoded = (void *)ptr.v;
501#endif
502 return decoded;
503}
504
505static inline void *get_freepointer(struct kmem_cache *s, void *object)
506{
507 unsigned long ptr_addr;
508 freeptr_t p;
509
510 object = kasan_reset_tag(object);
511 ptr_addr = (unsigned long)object + s->offset;
512 p = *(freeptr_t *)(ptr_addr);
513 return freelist_ptr_decode(s, p, ptr_addr);
514}
515
516#ifndef CONFIG_SLUB_TINY
517static void prefetch_freepointer(const struct kmem_cache *s, void *object)
518{
519 prefetchw(object + s->offset);
520}
521#endif
522
523/*
524 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
525 * pointer value in the case the current thread loses the race for the next
526 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
527 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
528 * KMSAN will still check all arguments of cmpxchg because of imperfect
529 * handling of inline assembly.
530 * To work around this problem, we apply __no_kmsan_checks to ensure that
531 * get_freepointer_safe() returns initialized memory.
532 */
533__no_kmsan_checks
534static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
535{
536 unsigned long freepointer_addr;
537 freeptr_t p;
538
539 if (!debug_pagealloc_enabled_static())
540 return get_freepointer(s, object);
541
542 object = kasan_reset_tag(object);
543 freepointer_addr = (unsigned long)object + s->offset;
544 copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p));
545 return freelist_ptr_decode(s, p, freepointer_addr);
546}
547
548static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
549{
550 unsigned long freeptr_addr = (unsigned long)object + s->offset;
551
552#ifdef CONFIG_SLAB_FREELIST_HARDENED
553 BUG_ON(object == fp); /* naive detection of double free or corruption */
554#endif
555
556 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
557 *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr);
558}
559
560/*
561 * See comment in calculate_sizes().
562 */
563static inline bool freeptr_outside_object(struct kmem_cache *s)
564{
565 return s->offset >= s->inuse;
566}
567
568/*
569 * Return offset of the end of info block which is inuse + free pointer if
570 * not overlapping with object.
571 */
572static inline unsigned int get_info_end(struct kmem_cache *s)
573{
574 if (freeptr_outside_object(s))
575 return s->inuse + sizeof(void *);
576 else
577 return s->inuse;
578}
579
580/* Loop over all objects in a slab */
581#define for_each_object(__p, __s, __addr, __objects) \
582 for (__p = fixup_red_left(__s, __addr); \
583 __p < (__addr) + (__objects) * (__s)->size; \
584 __p += (__s)->size)
585
586static inline unsigned int order_objects(unsigned int order, unsigned int size)
587{
588 return ((unsigned int)PAGE_SIZE << order) / size;
589}
590
591static inline struct kmem_cache_order_objects oo_make(unsigned int order,
592 unsigned int size)
593{
594 struct kmem_cache_order_objects x = {
595 (order << OO_SHIFT) + order_objects(order, size)
596 };
597
598 return x;
599}
600
601static inline unsigned int oo_order(struct kmem_cache_order_objects x)
602{
603 return x.x >> OO_SHIFT;
604}
605
606static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
607{
608 return x.x & OO_MASK;
609}
610
611#ifdef CONFIG_SLUB_CPU_PARTIAL
612static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
613{
614 unsigned int nr_slabs;
615
616 s->cpu_partial = nr_objects;
617
618 /*
619 * We take the number of objects but actually limit the number of
620 * slabs on the per cpu partial list, in order to limit excessive
621 * growth of the list. For simplicity we assume that the slabs will
622 * be half-full.
623 */
624 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
625 s->cpu_partial_slabs = nr_slabs;
626}
627#else
628static inline void
629slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
630{
631}
632#endif /* CONFIG_SLUB_CPU_PARTIAL */
633
634/*
635 * Per slab locking using the pagelock
636 */
637static __always_inline void slab_lock(struct slab *slab)
638{
639 struct page *page = slab_page(slab);
640
641 VM_BUG_ON_PAGE(PageTail(page), page);
642 bit_spin_lock(PG_locked, &page->flags);
643}
644
645static __always_inline void slab_unlock(struct slab *slab)
646{
647 struct page *page = slab_page(slab);
648
649 VM_BUG_ON_PAGE(PageTail(page), page);
650 bit_spin_unlock(PG_locked, &page->flags);
651}
652
653static inline bool
654__update_freelist_fast(struct slab *slab,
655 void *freelist_old, unsigned long counters_old,
656 void *freelist_new, unsigned long counters_new)
657{
658#ifdef system_has_freelist_aba
659 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
660 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
661
662 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
663#else
664 return false;
665#endif
666}
667
668static inline bool
669__update_freelist_slow(struct slab *slab,
670 void *freelist_old, unsigned long counters_old,
671 void *freelist_new, unsigned long counters_new)
672{
673 bool ret = false;
674
675 slab_lock(slab);
676 if (slab->freelist == freelist_old &&
677 slab->counters == counters_old) {
678 slab->freelist = freelist_new;
679 slab->counters = counters_new;
680 ret = true;
681 }
682 slab_unlock(slab);
683
684 return ret;
685}
686
687/*
688 * Interrupts must be disabled (for the fallback code to work right), typically
689 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
690 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
691 * allocation/ free operation in hardirq context. Therefore nothing can
692 * interrupt the operation.
693 */
694static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
695 void *freelist_old, unsigned long counters_old,
696 void *freelist_new, unsigned long counters_new,
697 const char *n)
698{
699 bool ret;
700
701 if (USE_LOCKLESS_FAST_PATH())
702 lockdep_assert_irqs_disabled();
703
704 if (s->flags & __CMPXCHG_DOUBLE) {
705 ret = __update_freelist_fast(slab, freelist_old, counters_old,
706 freelist_new, counters_new);
707 } else {
708 ret = __update_freelist_slow(slab, freelist_old, counters_old,
709 freelist_new, counters_new);
710 }
711 if (likely(ret))
712 return true;
713
714 cpu_relax();
715 stat(s, CMPXCHG_DOUBLE_FAIL);
716
717#ifdef SLUB_DEBUG_CMPXCHG
718 pr_info("%s %s: cmpxchg double redo ", n, s->name);
719#endif
720
721 return false;
722}
723
724static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
725 void *freelist_old, unsigned long counters_old,
726 void *freelist_new, unsigned long counters_new,
727 const char *n)
728{
729 bool ret;
730
731 if (s->flags & __CMPXCHG_DOUBLE) {
732 ret = __update_freelist_fast(slab, freelist_old, counters_old,
733 freelist_new, counters_new);
734 } else {
735 unsigned long flags;
736
737 local_irq_save(flags);
738 ret = __update_freelist_slow(slab, freelist_old, counters_old,
739 freelist_new, counters_new);
740 local_irq_restore(flags);
741 }
742 if (likely(ret))
743 return true;
744
745 cpu_relax();
746 stat(s, CMPXCHG_DOUBLE_FAIL);
747
748#ifdef SLUB_DEBUG_CMPXCHG
749 pr_info("%s %s: cmpxchg double redo ", n, s->name);
750#endif
751
752 return false;
753}
754
755#ifdef CONFIG_SLUB_DEBUG
756static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
757static DEFINE_SPINLOCK(object_map_lock);
758
759static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
760 struct slab *slab)
761{
762 void *addr = slab_address(slab);
763 void *p;
764
765 bitmap_zero(obj_map, slab->objects);
766
767 for (p = slab->freelist; p; p = get_freepointer(s, p))
768 set_bit(__obj_to_index(s, addr, p), obj_map);
769}
770
771#if IS_ENABLED(CONFIG_KUNIT)
772static bool slab_add_kunit_errors(void)
773{
774 struct kunit_resource *resource;
775
776 if (!kunit_get_current_test())
777 return false;
778
779 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
780 if (!resource)
781 return false;
782
783 (*(int *)resource->data)++;
784 kunit_put_resource(resource);
785 return true;
786}
787#else
788static inline bool slab_add_kunit_errors(void) { return false; }
789#endif
790
791static inline unsigned int size_from_object(struct kmem_cache *s)
792{
793 if (s->flags & SLAB_RED_ZONE)
794 return s->size - s->red_left_pad;
795
796 return s->size;
797}
798
799static inline void *restore_red_left(struct kmem_cache *s, void *p)
800{
801 if (s->flags & SLAB_RED_ZONE)
802 p -= s->red_left_pad;
803
804 return p;
805}
806
807/*
808 * Debug settings:
809 */
810#if defined(CONFIG_SLUB_DEBUG_ON)
811static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
812#else
813static slab_flags_t slub_debug;
814#endif
815
816static char *slub_debug_string;
817static int disable_higher_order_debug;
818
819/*
820 * slub is about to manipulate internal object metadata. This memory lies
821 * outside the range of the allocated object, so accessing it would normally
822 * be reported by kasan as a bounds error. metadata_access_enable() is used
823 * to tell kasan that these accesses are OK.
824 */
825static inline void metadata_access_enable(void)
826{
827 kasan_disable_current();
828}
829
830static inline void metadata_access_disable(void)
831{
832 kasan_enable_current();
833}
834
835/*
836 * Object debugging
837 */
838
839/* Verify that a pointer has an address that is valid within a slab page */
840static inline int check_valid_pointer(struct kmem_cache *s,
841 struct slab *slab, void *object)
842{
843 void *base;
844
845 if (!object)
846 return 1;
847
848 base = slab_address(slab);
849 object = kasan_reset_tag(object);
850 object = restore_red_left(s, object);
851 if (object < base || object >= base + slab->objects * s->size ||
852 (object - base) % s->size) {
853 return 0;
854 }
855
856 return 1;
857}
858
859static void print_section(char *level, char *text, u8 *addr,
860 unsigned int length)
861{
862 metadata_access_enable();
863 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
864 16, 1, kasan_reset_tag((void *)addr), length, 1);
865 metadata_access_disable();
866}
867
868static struct track *get_track(struct kmem_cache *s, void *object,
869 enum track_item alloc)
870{
871 struct track *p;
872
873 p = object + get_info_end(s);
874
875 return kasan_reset_tag(p + alloc);
876}
877
878#ifdef CONFIG_STACKDEPOT
879static noinline depot_stack_handle_t set_track_prepare(void)
880{
881 depot_stack_handle_t handle;
882 unsigned long entries[TRACK_ADDRS_COUNT];
883 unsigned int nr_entries;
884
885 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
886 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
887
888 return handle;
889}
890#else
891static inline depot_stack_handle_t set_track_prepare(void)
892{
893 return 0;
894}
895#endif
896
897static void set_track_update(struct kmem_cache *s, void *object,
898 enum track_item alloc, unsigned long addr,
899 depot_stack_handle_t handle)
900{
901 struct track *p = get_track(s, object, alloc);
902
903#ifdef CONFIG_STACKDEPOT
904 p->handle = handle;
905#endif
906 p->addr = addr;
907 p->cpu = smp_processor_id();
908 p->pid = current->pid;
909 p->when = jiffies;
910}
911
912static __always_inline void set_track(struct kmem_cache *s, void *object,
913 enum track_item alloc, unsigned long addr)
914{
915 depot_stack_handle_t handle = set_track_prepare();
916
917 set_track_update(s, object, alloc, addr, handle);
918}
919
920static void init_tracking(struct kmem_cache *s, void *object)
921{
922 struct track *p;
923
924 if (!(s->flags & SLAB_STORE_USER))
925 return;
926
927 p = get_track(s, object, TRACK_ALLOC);
928 memset(p, 0, 2*sizeof(struct track));
929}
930
931static void print_track(const char *s, struct track *t, unsigned long pr_time)
932{
933 depot_stack_handle_t handle __maybe_unused;
934
935 if (!t->addr)
936 return;
937
938 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
939 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
940#ifdef CONFIG_STACKDEPOT
941 handle = READ_ONCE(t->handle);
942 if (handle)
943 stack_depot_print(handle);
944 else
945 pr_err("object allocation/free stack trace missing\n");
946#endif
947}
948
949void print_tracking(struct kmem_cache *s, void *object)
950{
951 unsigned long pr_time = jiffies;
952 if (!(s->flags & SLAB_STORE_USER))
953 return;
954
955 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
956 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
957}
958
959static void print_slab_info(const struct slab *slab)
960{
961 struct folio *folio = (struct folio *)slab_folio(slab);
962
963 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
964 slab, slab->objects, slab->inuse, slab->freelist,
965 folio_flags(folio, 0));
966}
967
968/*
969 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
970 * family will round up the real request size to these fixed ones, so
971 * there could be an extra area than what is requested. Save the original
972 * request size in the meta data area, for better debug and sanity check.
973 */
974static inline void set_orig_size(struct kmem_cache *s,
975 void *object, unsigned int orig_size)
976{
977 void *p = kasan_reset_tag(object);
978 unsigned int kasan_meta_size;
979
980 if (!slub_debug_orig_size(s))
981 return;
982
983 /*
984 * KASAN can save its free meta data inside of the object at offset 0.
985 * If this meta data size is larger than 'orig_size', it will overlap
986 * the data redzone in [orig_size+1, object_size]. Thus, we adjust
987 * 'orig_size' to be as at least as big as KASAN's meta data.
988 */
989 kasan_meta_size = kasan_metadata_size(s, true);
990 if (kasan_meta_size > orig_size)
991 orig_size = kasan_meta_size;
992
993 p += get_info_end(s);
994 p += sizeof(struct track) * 2;
995
996 *(unsigned int *)p = orig_size;
997}
998
999static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
1000{
1001 void *p = kasan_reset_tag(object);
1002
1003 if (!slub_debug_orig_size(s))
1004 return s->object_size;
1005
1006 p += get_info_end(s);
1007 p += sizeof(struct track) * 2;
1008
1009 return *(unsigned int *)p;
1010}
1011
1012void skip_orig_size_check(struct kmem_cache *s, const void *object)
1013{
1014 set_orig_size(s, (void *)object, s->object_size);
1015}
1016
1017static void slab_bug(struct kmem_cache *s, char *fmt, ...)
1018{
1019 struct va_format vaf;
1020 va_list args;
1021
1022 va_start(args, fmt);
1023 vaf.fmt = fmt;
1024 vaf.va = &args;
1025 pr_err("=============================================================================\n");
1026 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
1027 pr_err("-----------------------------------------------------------------------------\n\n");
1028 va_end(args);
1029}
1030
1031__printf(2, 3)
1032static void slab_fix(struct kmem_cache *s, char *fmt, ...)
1033{
1034 struct va_format vaf;
1035 va_list args;
1036
1037 if (slab_add_kunit_errors())
1038 return;
1039
1040 va_start(args, fmt);
1041 vaf.fmt = fmt;
1042 vaf.va = &args;
1043 pr_err("FIX %s: %pV\n", s->name, &vaf);
1044 va_end(args);
1045}
1046
1047static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
1048{
1049 unsigned int off; /* Offset of last byte */
1050 u8 *addr = slab_address(slab);
1051
1052 print_tracking(s, p);
1053
1054 print_slab_info(slab);
1055
1056 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
1057 p, p - addr, get_freepointer(s, p));
1058
1059 if (s->flags & SLAB_RED_ZONE)
1060 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
1061 s->red_left_pad);
1062 else if (p > addr + 16)
1063 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
1064
1065 print_section(KERN_ERR, "Object ", p,
1066 min_t(unsigned int, s->object_size, PAGE_SIZE));
1067 if (s->flags & SLAB_RED_ZONE)
1068 print_section(KERN_ERR, "Redzone ", p + s->object_size,
1069 s->inuse - s->object_size);
1070
1071 off = get_info_end(s);
1072
1073 if (s->flags & SLAB_STORE_USER)
1074 off += 2 * sizeof(struct track);
1075
1076 if (slub_debug_orig_size(s))
1077 off += sizeof(unsigned int);
1078
1079 off += kasan_metadata_size(s, false);
1080
1081 if (off != size_from_object(s))
1082 /* Beginning of the filler is the free pointer */
1083 print_section(KERN_ERR, "Padding ", p + off,
1084 size_from_object(s) - off);
1085
1086 dump_stack();
1087}
1088
1089static void object_err(struct kmem_cache *s, struct slab *slab,
1090 u8 *object, char *reason)
1091{
1092 if (slab_add_kunit_errors())
1093 return;
1094
1095 slab_bug(s, "%s", reason);
1096 print_trailer(s, slab, object);
1097 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1098}
1099
1100static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1101 void **freelist, void *nextfree)
1102{
1103 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
1104 !check_valid_pointer(s, slab, nextfree) && freelist) {
1105 object_err(s, slab, *freelist, "Freechain corrupt");
1106 *freelist = NULL;
1107 slab_fix(s, "Isolate corrupted freechain");
1108 return true;
1109 }
1110
1111 return false;
1112}
1113
1114static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1115 const char *fmt, ...)
1116{
1117 va_list args;
1118 char buf[100];
1119
1120 if (slab_add_kunit_errors())
1121 return;
1122
1123 va_start(args, fmt);
1124 vsnprintf(buf, sizeof(buf), fmt, args);
1125 va_end(args);
1126 slab_bug(s, "%s", buf);
1127 print_slab_info(slab);
1128 dump_stack();
1129 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1130}
1131
1132static void init_object(struct kmem_cache *s, void *object, u8 val)
1133{
1134 u8 *p = kasan_reset_tag(object);
1135 unsigned int poison_size = s->object_size;
1136
1137 if (s->flags & SLAB_RED_ZONE) {
1138 memset(p - s->red_left_pad, val, s->red_left_pad);
1139
1140 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1141 /*
1142 * Redzone the extra allocated space by kmalloc than
1143 * requested, and the poison size will be limited to
1144 * the original request size accordingly.
1145 */
1146 poison_size = get_orig_size(s, object);
1147 }
1148 }
1149
1150 if (s->flags & __OBJECT_POISON) {
1151 memset(p, POISON_FREE, poison_size - 1);
1152 p[poison_size - 1] = POISON_END;
1153 }
1154
1155 if (s->flags & SLAB_RED_ZONE)
1156 memset(p + poison_size, val, s->inuse - poison_size);
1157}
1158
1159static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1160 void *from, void *to)
1161{
1162 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1163 memset(from, data, to - from);
1164}
1165
1166static int check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1167 u8 *object, char *what,
1168 u8 *start, unsigned int value, unsigned int bytes)
1169{
1170 u8 *fault;
1171 u8 *end;
1172 u8 *addr = slab_address(slab);
1173
1174 metadata_access_enable();
1175 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1176 metadata_access_disable();
1177 if (!fault)
1178 return 1;
1179
1180 end = start + bytes;
1181 while (end > fault && end[-1] == value)
1182 end--;
1183
1184 if (slab_add_kunit_errors())
1185 goto skip_bug_print;
1186
1187 slab_bug(s, "%s overwritten", what);
1188 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1189 fault, end - 1, fault - addr,
1190 fault[0], value);
1191 print_trailer(s, slab, object);
1192 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1193
1194skip_bug_print:
1195 restore_bytes(s, what, value, fault, end);
1196 return 0;
1197}
1198
1199/*
1200 * Object layout:
1201 *
1202 * object address
1203 * Bytes of the object to be managed.
1204 * If the freepointer may overlay the object then the free
1205 * pointer is at the middle of the object.
1206 *
1207 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1208 * 0xa5 (POISON_END)
1209 *
1210 * object + s->object_size
1211 * Padding to reach word boundary. This is also used for Redzoning.
1212 * Padding is extended by another word if Redzoning is enabled and
1213 * object_size == inuse.
1214 *
1215 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
1216 * 0xcc (RED_ACTIVE) for objects in use.
1217 *
1218 * object + s->inuse
1219 * Meta data starts here.
1220 *
1221 * A. Free pointer (if we cannot overwrite object on free)
1222 * B. Tracking data for SLAB_STORE_USER
1223 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1224 * D. Padding to reach required alignment boundary or at minimum
1225 * one word if debugging is on to be able to detect writes
1226 * before the word boundary.
1227 *
1228 * Padding is done using 0x5a (POISON_INUSE)
1229 *
1230 * object + s->size
1231 * Nothing is used beyond s->size.
1232 *
1233 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1234 * ignored. And therefore no slab options that rely on these boundaries
1235 * may be used with merged slabcaches.
1236 */
1237
1238static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1239{
1240 unsigned long off = get_info_end(s); /* The end of info */
1241
1242 if (s->flags & SLAB_STORE_USER) {
1243 /* We also have user information there */
1244 off += 2 * sizeof(struct track);
1245
1246 if (s->flags & SLAB_KMALLOC)
1247 off += sizeof(unsigned int);
1248 }
1249
1250 off += kasan_metadata_size(s, false);
1251
1252 if (size_from_object(s) == off)
1253 return 1;
1254
1255 return check_bytes_and_report(s, slab, p, "Object padding",
1256 p + off, POISON_INUSE, size_from_object(s) - off);
1257}
1258
1259/* Check the pad bytes at the end of a slab page */
1260static void slab_pad_check(struct kmem_cache *s, struct slab *slab)
1261{
1262 u8 *start;
1263 u8 *fault;
1264 u8 *end;
1265 u8 *pad;
1266 int length;
1267 int remainder;
1268
1269 if (!(s->flags & SLAB_POISON))
1270 return;
1271
1272 start = slab_address(slab);
1273 length = slab_size(slab);
1274 end = start + length;
1275 remainder = length % s->size;
1276 if (!remainder)
1277 return;
1278
1279 pad = end - remainder;
1280 metadata_access_enable();
1281 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1282 metadata_access_disable();
1283 if (!fault)
1284 return;
1285 while (end > fault && end[-1] == POISON_INUSE)
1286 end--;
1287
1288 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1289 fault, end - 1, fault - start);
1290 print_section(KERN_ERR, "Padding ", pad, remainder);
1291
1292 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1293}
1294
1295static int check_object(struct kmem_cache *s, struct slab *slab,
1296 void *object, u8 val)
1297{
1298 u8 *p = object;
1299 u8 *endobject = object + s->object_size;
1300 unsigned int orig_size, kasan_meta_size;
1301
1302 if (s->flags & SLAB_RED_ZONE) {
1303 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1304 object - s->red_left_pad, val, s->red_left_pad))
1305 return 0;
1306
1307 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1308 endobject, val, s->inuse - s->object_size))
1309 return 0;
1310
1311 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1312 orig_size = get_orig_size(s, object);
1313
1314 if (s->object_size > orig_size &&
1315 !check_bytes_and_report(s, slab, object,
1316 "kmalloc Redzone", p + orig_size,
1317 val, s->object_size - orig_size)) {
1318 return 0;
1319 }
1320 }
1321 } else {
1322 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1323 check_bytes_and_report(s, slab, p, "Alignment padding",
1324 endobject, POISON_INUSE,
1325 s->inuse - s->object_size);
1326 }
1327 }
1328
1329 if (s->flags & SLAB_POISON) {
1330 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) {
1331 /*
1332 * KASAN can save its free meta data inside of the
1333 * object at offset 0. Thus, skip checking the part of
1334 * the redzone that overlaps with the meta data.
1335 */
1336 kasan_meta_size = kasan_metadata_size(s, true);
1337 if (kasan_meta_size < s->object_size - 1 &&
1338 !check_bytes_and_report(s, slab, p, "Poison",
1339 p + kasan_meta_size, POISON_FREE,
1340 s->object_size - kasan_meta_size - 1))
1341 return 0;
1342 if (kasan_meta_size < s->object_size &&
1343 !check_bytes_and_report(s, slab, p, "End Poison",
1344 p + s->object_size - 1, POISON_END, 1))
1345 return 0;
1346 }
1347 /*
1348 * check_pad_bytes cleans up on its own.
1349 */
1350 check_pad_bytes(s, slab, p);
1351 }
1352
1353 if (!freeptr_outside_object(s) && val == SLUB_RED_ACTIVE)
1354 /*
1355 * Object and freepointer overlap. Cannot check
1356 * freepointer while object is allocated.
1357 */
1358 return 1;
1359
1360 /* Check free pointer validity */
1361 if (!check_valid_pointer(s, slab, get_freepointer(s, p))) {
1362 object_err(s, slab, p, "Freepointer corrupt");
1363 /*
1364 * No choice but to zap it and thus lose the remainder
1365 * of the free objects in this slab. May cause
1366 * another error because the object count is now wrong.
1367 */
1368 set_freepointer(s, p, NULL);
1369 return 0;
1370 }
1371 return 1;
1372}
1373
1374static int check_slab(struct kmem_cache *s, struct slab *slab)
1375{
1376 int maxobj;
1377
1378 if (!folio_test_slab(slab_folio(slab))) {
1379 slab_err(s, slab, "Not a valid slab page");
1380 return 0;
1381 }
1382
1383 maxobj = order_objects(slab_order(slab), s->size);
1384 if (slab->objects > maxobj) {
1385 slab_err(s, slab, "objects %u > max %u",
1386 slab->objects, maxobj);
1387 return 0;
1388 }
1389 if (slab->inuse > slab->objects) {
1390 slab_err(s, slab, "inuse %u > max %u",
1391 slab->inuse, slab->objects);
1392 return 0;
1393 }
1394 /* Slab_pad_check fixes things up after itself */
1395 slab_pad_check(s, slab);
1396 return 1;
1397}
1398
1399/*
1400 * Determine if a certain object in a slab is on the freelist. Must hold the
1401 * slab lock to guarantee that the chains are in a consistent state.
1402 */
1403static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1404{
1405 int nr = 0;
1406 void *fp;
1407 void *object = NULL;
1408 int max_objects;
1409
1410 fp = slab->freelist;
1411 while (fp && nr <= slab->objects) {
1412 if (fp == search)
1413 return 1;
1414 if (!check_valid_pointer(s, slab, fp)) {
1415 if (object) {
1416 object_err(s, slab, object,
1417 "Freechain corrupt");
1418 set_freepointer(s, object, NULL);
1419 } else {
1420 slab_err(s, slab, "Freepointer corrupt");
1421 slab->freelist = NULL;
1422 slab->inuse = slab->objects;
1423 slab_fix(s, "Freelist cleared");
1424 return 0;
1425 }
1426 break;
1427 }
1428 object = fp;
1429 fp = get_freepointer(s, object);
1430 nr++;
1431 }
1432
1433 max_objects = order_objects(slab_order(slab), s->size);
1434 if (max_objects > MAX_OBJS_PER_PAGE)
1435 max_objects = MAX_OBJS_PER_PAGE;
1436
1437 if (slab->objects != max_objects) {
1438 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1439 slab->objects, max_objects);
1440 slab->objects = max_objects;
1441 slab_fix(s, "Number of objects adjusted");
1442 }
1443 if (slab->inuse != slab->objects - nr) {
1444 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1445 slab->inuse, slab->objects - nr);
1446 slab->inuse = slab->objects - nr;
1447 slab_fix(s, "Object count adjusted");
1448 }
1449 return search == NULL;
1450}
1451
1452static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1453 int alloc)
1454{
1455 if (s->flags & SLAB_TRACE) {
1456 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1457 s->name,
1458 alloc ? "alloc" : "free",
1459 object, slab->inuse,
1460 slab->freelist);
1461
1462 if (!alloc)
1463 print_section(KERN_INFO, "Object ", (void *)object,
1464 s->object_size);
1465
1466 dump_stack();
1467 }
1468}
1469
1470/*
1471 * Tracking of fully allocated slabs for debugging purposes.
1472 */
1473static void add_full(struct kmem_cache *s,
1474 struct kmem_cache_node *n, struct slab *slab)
1475{
1476 if (!(s->flags & SLAB_STORE_USER))
1477 return;
1478
1479 lockdep_assert_held(&n->list_lock);
1480 list_add(&slab->slab_list, &n->full);
1481}
1482
1483static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1484{
1485 if (!(s->flags & SLAB_STORE_USER))
1486 return;
1487
1488 lockdep_assert_held(&n->list_lock);
1489 list_del(&slab->slab_list);
1490}
1491
1492static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1493{
1494 return atomic_long_read(&n->nr_slabs);
1495}
1496
1497static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1498{
1499 struct kmem_cache_node *n = get_node(s, node);
1500
1501 atomic_long_inc(&n->nr_slabs);
1502 atomic_long_add(objects, &n->total_objects);
1503}
1504static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1505{
1506 struct kmem_cache_node *n = get_node(s, node);
1507
1508 atomic_long_dec(&n->nr_slabs);
1509 atomic_long_sub(objects, &n->total_objects);
1510}
1511
1512/* Object debug checks for alloc/free paths */
1513static void setup_object_debug(struct kmem_cache *s, void *object)
1514{
1515 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1516 return;
1517
1518 init_object(s, object, SLUB_RED_INACTIVE);
1519 init_tracking(s, object);
1520}
1521
1522static
1523void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1524{
1525 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1526 return;
1527
1528 metadata_access_enable();
1529 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1530 metadata_access_disable();
1531}
1532
1533static inline int alloc_consistency_checks(struct kmem_cache *s,
1534 struct slab *slab, void *object)
1535{
1536 if (!check_slab(s, slab))
1537 return 0;
1538
1539 if (!check_valid_pointer(s, slab, object)) {
1540 object_err(s, slab, object, "Freelist Pointer check fails");
1541 return 0;
1542 }
1543
1544 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1545 return 0;
1546
1547 return 1;
1548}
1549
1550static noinline bool alloc_debug_processing(struct kmem_cache *s,
1551 struct slab *slab, void *object, int orig_size)
1552{
1553 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1554 if (!alloc_consistency_checks(s, slab, object))
1555 goto bad;
1556 }
1557
1558 /* Success. Perform special debug activities for allocs */
1559 trace(s, slab, object, 1);
1560 set_orig_size(s, object, orig_size);
1561 init_object(s, object, SLUB_RED_ACTIVE);
1562 return true;
1563
1564bad:
1565 if (folio_test_slab(slab_folio(slab))) {
1566 /*
1567 * If this is a slab page then lets do the best we can
1568 * to avoid issues in the future. Marking all objects
1569 * as used avoids touching the remaining objects.
1570 */
1571 slab_fix(s, "Marking all objects used");
1572 slab->inuse = slab->objects;
1573 slab->freelist = NULL;
1574 }
1575 return false;
1576}
1577
1578static inline int free_consistency_checks(struct kmem_cache *s,
1579 struct slab *slab, void *object, unsigned long addr)
1580{
1581 if (!check_valid_pointer(s, slab, object)) {
1582 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1583 return 0;
1584 }
1585
1586 if (on_freelist(s, slab, object)) {
1587 object_err(s, slab, object, "Object already free");
1588 return 0;
1589 }
1590
1591 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1592 return 0;
1593
1594 if (unlikely(s != slab->slab_cache)) {
1595 if (!folio_test_slab(slab_folio(slab))) {
1596 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1597 object);
1598 } else if (!slab->slab_cache) {
1599 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1600 object);
1601 dump_stack();
1602 } else
1603 object_err(s, slab, object,
1604 "page slab pointer corrupt.");
1605 return 0;
1606 }
1607 return 1;
1608}
1609
1610/*
1611 * Parse a block of slab_debug options. Blocks are delimited by ';'
1612 *
1613 * @str: start of block
1614 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1615 * @slabs: return start of list of slabs, or NULL when there's no list
1616 * @init: assume this is initial parsing and not per-kmem-create parsing
1617 *
1618 * returns the start of next block if there's any, or NULL
1619 */
1620static char *
1621parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1622{
1623 bool higher_order_disable = false;
1624
1625 /* Skip any completely empty blocks */
1626 while (*str && *str == ';')
1627 str++;
1628
1629 if (*str == ',') {
1630 /*
1631 * No options but restriction on slabs. This means full
1632 * debugging for slabs matching a pattern.
1633 */
1634 *flags = DEBUG_DEFAULT_FLAGS;
1635 goto check_slabs;
1636 }
1637 *flags = 0;
1638
1639 /* Determine which debug features should be switched on */
1640 for (; *str && *str != ',' && *str != ';'; str++) {
1641 switch (tolower(*str)) {
1642 case '-':
1643 *flags = 0;
1644 break;
1645 case 'f':
1646 *flags |= SLAB_CONSISTENCY_CHECKS;
1647 break;
1648 case 'z':
1649 *flags |= SLAB_RED_ZONE;
1650 break;
1651 case 'p':
1652 *flags |= SLAB_POISON;
1653 break;
1654 case 'u':
1655 *flags |= SLAB_STORE_USER;
1656 break;
1657 case 't':
1658 *flags |= SLAB_TRACE;
1659 break;
1660 case 'a':
1661 *flags |= SLAB_FAILSLAB;
1662 break;
1663 case 'o':
1664 /*
1665 * Avoid enabling debugging on caches if its minimum
1666 * order would increase as a result.
1667 */
1668 higher_order_disable = true;
1669 break;
1670 default:
1671 if (init)
1672 pr_err("slab_debug option '%c' unknown. skipped\n", *str);
1673 }
1674 }
1675check_slabs:
1676 if (*str == ',')
1677 *slabs = ++str;
1678 else
1679 *slabs = NULL;
1680
1681 /* Skip over the slab list */
1682 while (*str && *str != ';')
1683 str++;
1684
1685 /* Skip any completely empty blocks */
1686 while (*str && *str == ';')
1687 str++;
1688
1689 if (init && higher_order_disable)
1690 disable_higher_order_debug = 1;
1691
1692 if (*str)
1693 return str;
1694 else
1695 return NULL;
1696}
1697
1698static int __init setup_slub_debug(char *str)
1699{
1700 slab_flags_t flags;
1701 slab_flags_t global_flags;
1702 char *saved_str;
1703 char *slab_list;
1704 bool global_slub_debug_changed = false;
1705 bool slab_list_specified = false;
1706
1707 global_flags = DEBUG_DEFAULT_FLAGS;
1708 if (*str++ != '=' || !*str)
1709 /*
1710 * No options specified. Switch on full debugging.
1711 */
1712 goto out;
1713
1714 saved_str = str;
1715 while (str) {
1716 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1717
1718 if (!slab_list) {
1719 global_flags = flags;
1720 global_slub_debug_changed = true;
1721 } else {
1722 slab_list_specified = true;
1723 if (flags & SLAB_STORE_USER)
1724 stack_depot_request_early_init();
1725 }
1726 }
1727
1728 /*
1729 * For backwards compatibility, a single list of flags with list of
1730 * slabs means debugging is only changed for those slabs, so the global
1731 * slab_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1732 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1733 * long as there is no option specifying flags without a slab list.
1734 */
1735 if (slab_list_specified) {
1736 if (!global_slub_debug_changed)
1737 global_flags = slub_debug;
1738 slub_debug_string = saved_str;
1739 }
1740out:
1741 slub_debug = global_flags;
1742 if (slub_debug & SLAB_STORE_USER)
1743 stack_depot_request_early_init();
1744 if (slub_debug != 0 || slub_debug_string)
1745 static_branch_enable(&slub_debug_enabled);
1746 else
1747 static_branch_disable(&slub_debug_enabled);
1748 if ((static_branch_unlikely(&init_on_alloc) ||
1749 static_branch_unlikely(&init_on_free)) &&
1750 (slub_debug & SLAB_POISON))
1751 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1752 return 1;
1753}
1754
1755__setup("slab_debug", setup_slub_debug);
1756__setup_param("slub_debug", slub_debug, setup_slub_debug, 0);
1757
1758/*
1759 * kmem_cache_flags - apply debugging options to the cache
1760 * @flags: flags to set
1761 * @name: name of the cache
1762 *
1763 * Debug option(s) are applied to @flags. In addition to the debug
1764 * option(s), if a slab name (or multiple) is specified i.e.
1765 * slab_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1766 * then only the select slabs will receive the debug option(s).
1767 */
1768slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1769{
1770 char *iter;
1771 size_t len;
1772 char *next_block;
1773 slab_flags_t block_flags;
1774 slab_flags_t slub_debug_local = slub_debug;
1775
1776 if (flags & SLAB_NO_USER_FLAGS)
1777 return flags;
1778
1779 /*
1780 * If the slab cache is for debugging (e.g. kmemleak) then
1781 * don't store user (stack trace) information by default,
1782 * but let the user enable it via the command line below.
1783 */
1784 if (flags & SLAB_NOLEAKTRACE)
1785 slub_debug_local &= ~SLAB_STORE_USER;
1786
1787 len = strlen(name);
1788 next_block = slub_debug_string;
1789 /* Go through all blocks of debug options, see if any matches our slab's name */
1790 while (next_block) {
1791 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1792 if (!iter)
1793 continue;
1794 /* Found a block that has a slab list, search it */
1795 while (*iter) {
1796 char *end, *glob;
1797 size_t cmplen;
1798
1799 end = strchrnul(iter, ',');
1800 if (next_block && next_block < end)
1801 end = next_block - 1;
1802
1803 glob = strnchr(iter, end - iter, '*');
1804 if (glob)
1805 cmplen = glob - iter;
1806 else
1807 cmplen = max_t(size_t, len, (end - iter));
1808
1809 if (!strncmp(name, iter, cmplen)) {
1810 flags |= block_flags;
1811 return flags;
1812 }
1813
1814 if (!*end || *end == ';')
1815 break;
1816 iter = end + 1;
1817 }
1818 }
1819
1820 return flags | slub_debug_local;
1821}
1822#else /* !CONFIG_SLUB_DEBUG */
1823static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1824static inline
1825void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1826
1827static inline bool alloc_debug_processing(struct kmem_cache *s,
1828 struct slab *slab, void *object, int orig_size) { return true; }
1829
1830static inline bool free_debug_processing(struct kmem_cache *s,
1831 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1832 unsigned long addr, depot_stack_handle_t handle) { return true; }
1833
1834static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1835static inline int check_object(struct kmem_cache *s, struct slab *slab,
1836 void *object, u8 val) { return 1; }
1837static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1838static inline void set_track(struct kmem_cache *s, void *object,
1839 enum track_item alloc, unsigned long addr) {}
1840static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1841 struct slab *slab) {}
1842static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1843 struct slab *slab) {}
1844slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1845{
1846 return flags;
1847}
1848#define slub_debug 0
1849
1850#define disable_higher_order_debug 0
1851
1852static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1853 { return 0; }
1854static inline void inc_slabs_node(struct kmem_cache *s, int node,
1855 int objects) {}
1856static inline void dec_slabs_node(struct kmem_cache *s, int node,
1857 int objects) {}
1858
1859#ifndef CONFIG_SLUB_TINY
1860static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1861 void **freelist, void *nextfree)
1862{
1863 return false;
1864}
1865#endif
1866#endif /* CONFIG_SLUB_DEBUG */
1867
1868static inline enum node_stat_item cache_vmstat_idx(struct kmem_cache *s)
1869{
1870 return (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1871 NR_SLAB_RECLAIMABLE_B : NR_SLAB_UNRECLAIMABLE_B;
1872}
1873
1874#ifdef CONFIG_MEMCG_KMEM
1875static inline void memcg_free_slab_cgroups(struct slab *slab)
1876{
1877 kfree(slab_objcgs(slab));
1878 slab->memcg_data = 0;
1879}
1880
1881static inline size_t obj_full_size(struct kmem_cache *s)
1882{
1883 /*
1884 * For each accounted object there is an extra space which is used
1885 * to store obj_cgroup membership. Charge it too.
1886 */
1887 return s->size + sizeof(struct obj_cgroup *);
1888}
1889
1890/*
1891 * Returns false if the allocation should fail.
1892 */
1893static bool __memcg_slab_pre_alloc_hook(struct kmem_cache *s,
1894 struct list_lru *lru,
1895 struct obj_cgroup **objcgp,
1896 size_t objects, gfp_t flags)
1897{
1898 /*
1899 * The obtained objcg pointer is safe to use within the current scope,
1900 * defined by current task or set_active_memcg() pair.
1901 * obj_cgroup_get() is used to get a permanent reference.
1902 */
1903 struct obj_cgroup *objcg = current_obj_cgroup();
1904 if (!objcg)
1905 return true;
1906
1907 if (lru) {
1908 int ret;
1909 struct mem_cgroup *memcg;
1910
1911 memcg = get_mem_cgroup_from_objcg(objcg);
1912 ret = memcg_list_lru_alloc(memcg, lru, flags);
1913 css_put(&memcg->css);
1914
1915 if (ret)
1916 return false;
1917 }
1918
1919 if (obj_cgroup_charge(objcg, flags, objects * obj_full_size(s)))
1920 return false;
1921
1922 *objcgp = objcg;
1923 return true;
1924}
1925
1926/*
1927 * Returns false if the allocation should fail.
1928 */
1929static __fastpath_inline
1930bool memcg_slab_pre_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
1931 struct obj_cgroup **objcgp, size_t objects,
1932 gfp_t flags)
1933{
1934 if (!memcg_kmem_online())
1935 return true;
1936
1937 if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
1938 return true;
1939
1940 return likely(__memcg_slab_pre_alloc_hook(s, lru, objcgp, objects,
1941 flags));
1942}
1943
1944static void __memcg_slab_post_alloc_hook(struct kmem_cache *s,
1945 struct obj_cgroup *objcg,
1946 gfp_t flags, size_t size,
1947 void **p)
1948{
1949 struct slab *slab;
1950 unsigned long off;
1951 size_t i;
1952
1953 flags &= gfp_allowed_mask;
1954
1955 for (i = 0; i < size; i++) {
1956 if (likely(p[i])) {
1957 slab = virt_to_slab(p[i]);
1958
1959 if (!slab_objcgs(slab) &&
1960 memcg_alloc_slab_cgroups(slab, s, flags, false)) {
1961 obj_cgroup_uncharge(objcg, obj_full_size(s));
1962 continue;
1963 }
1964
1965 off = obj_to_index(s, slab, p[i]);
1966 obj_cgroup_get(objcg);
1967 slab_objcgs(slab)[off] = objcg;
1968 mod_objcg_state(objcg, slab_pgdat(slab),
1969 cache_vmstat_idx(s), obj_full_size(s));
1970 } else {
1971 obj_cgroup_uncharge(objcg, obj_full_size(s));
1972 }
1973 }
1974}
1975
1976static __fastpath_inline
1977void memcg_slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
1978 gfp_t flags, size_t size, void **p)
1979{
1980 if (likely(!memcg_kmem_online() || !objcg))
1981 return;
1982
1983 return __memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
1984}
1985
1986static void __memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
1987 void **p, int objects,
1988 struct obj_cgroup **objcgs)
1989{
1990 for (int i = 0; i < objects; i++) {
1991 struct obj_cgroup *objcg;
1992 unsigned int off;
1993
1994 off = obj_to_index(s, slab, p[i]);
1995 objcg = objcgs[off];
1996 if (!objcg)
1997 continue;
1998
1999 objcgs[off] = NULL;
2000 obj_cgroup_uncharge(objcg, obj_full_size(s));
2001 mod_objcg_state(objcg, slab_pgdat(slab), cache_vmstat_idx(s),
2002 -obj_full_size(s));
2003 obj_cgroup_put(objcg);
2004 }
2005}
2006
2007static __fastpath_inline
2008void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2009 int objects)
2010{
2011 struct obj_cgroup **objcgs;
2012
2013 if (!memcg_kmem_online())
2014 return;
2015
2016 objcgs = slab_objcgs(slab);
2017 if (likely(!objcgs))
2018 return;
2019
2020 __memcg_slab_free_hook(s, slab, p, objects, objcgs);
2021}
2022
2023static inline
2024void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2025 struct obj_cgroup *objcg)
2026{
2027 if (objcg)
2028 obj_cgroup_uncharge(objcg, objects * obj_full_size(s));
2029}
2030#else /* CONFIG_MEMCG_KMEM */
2031static inline void memcg_free_slab_cgroups(struct slab *slab)
2032{
2033}
2034
2035static inline bool memcg_slab_pre_alloc_hook(struct kmem_cache *s,
2036 struct list_lru *lru,
2037 struct obj_cgroup **objcgp,
2038 size_t objects, gfp_t flags)
2039{
2040 return true;
2041}
2042
2043static inline void memcg_slab_post_alloc_hook(struct kmem_cache *s,
2044 struct obj_cgroup *objcg,
2045 gfp_t flags, size_t size,
2046 void **p)
2047{
2048}
2049
2050static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2051 void **p, int objects)
2052{
2053}
2054
2055static inline
2056void memcg_slab_alloc_error_hook(struct kmem_cache *s, int objects,
2057 struct obj_cgroup *objcg)
2058{
2059}
2060#endif /* CONFIG_MEMCG_KMEM */
2061
2062/*
2063 * Hooks for other subsystems that check memory allocations. In a typical
2064 * production configuration these hooks all should produce no code at all.
2065 *
2066 * Returns true if freeing of the object can proceed, false if its reuse
2067 * was delayed by KASAN quarantine, or it was returned to KFENCE.
2068 */
2069static __always_inline
2070bool slab_free_hook(struct kmem_cache *s, void *x, bool init)
2071{
2072 kmemleak_free_recursive(x, s->flags);
2073 kmsan_slab_free(s, x);
2074
2075 debug_check_no_locks_freed(x, s->object_size);
2076
2077 if (!(s->flags & SLAB_DEBUG_OBJECTS))
2078 debug_check_no_obj_freed(x, s->object_size);
2079
2080 /* Use KCSAN to help debug racy use-after-free. */
2081 if (!(s->flags & SLAB_TYPESAFE_BY_RCU))
2082 __kcsan_check_access(x, s->object_size,
2083 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2084
2085 if (kfence_free(x))
2086 return false;
2087
2088 /*
2089 * As memory initialization might be integrated into KASAN,
2090 * kasan_slab_free and initialization memset's must be
2091 * kept together to avoid discrepancies in behavior.
2092 *
2093 * The initialization memset's clear the object and the metadata,
2094 * but don't touch the SLAB redzone.
2095 *
2096 * The object's freepointer is also avoided if stored outside the
2097 * object.
2098 */
2099 if (unlikely(init)) {
2100 int rsize;
2101 unsigned int inuse;
2102
2103 inuse = get_info_end(s);
2104 if (!kasan_has_integrated_init())
2105 memset(kasan_reset_tag(x), 0, s->object_size);
2106 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2107 memset((char *)kasan_reset_tag(x) + inuse, 0,
2108 s->size - inuse - rsize);
2109 }
2110 /* KASAN might put x into memory quarantine, delaying its reuse. */
2111 return !kasan_slab_free(s, x, init);
2112}
2113
2114static inline bool slab_free_freelist_hook(struct kmem_cache *s,
2115 void **head, void **tail,
2116 int *cnt)
2117{
2118
2119 void *object;
2120 void *next = *head;
2121 void *old_tail = *tail;
2122 bool init;
2123
2124 if (is_kfence_address(next)) {
2125 slab_free_hook(s, next, false);
2126 return false;
2127 }
2128
2129 /* Head and tail of the reconstructed freelist */
2130 *head = NULL;
2131 *tail = NULL;
2132
2133 init = slab_want_init_on_free(s);
2134
2135 do {
2136 object = next;
2137 next = get_freepointer(s, object);
2138
2139 /* If object's reuse doesn't have to be delayed */
2140 if (likely(slab_free_hook(s, object, init))) {
2141 /* Move object to the new freelist */
2142 set_freepointer(s, object, *head);
2143 *head = object;
2144 if (!*tail)
2145 *tail = object;
2146 } else {
2147 /*
2148 * Adjust the reconstructed freelist depth
2149 * accordingly if object's reuse is delayed.
2150 */
2151 --(*cnt);
2152 }
2153 } while (object != old_tail);
2154
2155 return *head != NULL;
2156}
2157
2158static void *setup_object(struct kmem_cache *s, void *object)
2159{
2160 setup_object_debug(s, object);
2161 object = kasan_init_slab_obj(s, object);
2162 if (unlikely(s->ctor)) {
2163 kasan_unpoison_new_object(s, object);
2164 s->ctor(object);
2165 kasan_poison_new_object(s, object);
2166 }
2167 return object;
2168}
2169
2170/*
2171 * Slab allocation and freeing
2172 */
2173static inline struct slab *alloc_slab_page(gfp_t flags, int node,
2174 struct kmem_cache_order_objects oo)
2175{
2176 struct folio *folio;
2177 struct slab *slab;
2178 unsigned int order = oo_order(oo);
2179
2180 folio = (struct folio *)alloc_pages_node(node, flags, order);
2181 if (!folio)
2182 return NULL;
2183
2184 slab = folio_slab(folio);
2185 __folio_set_slab(folio);
2186 /* Make the flag visible before any changes to folio->mapping */
2187 smp_wmb();
2188 if (folio_is_pfmemalloc(folio))
2189 slab_set_pfmemalloc(slab);
2190
2191 return slab;
2192}
2193
2194#ifdef CONFIG_SLAB_FREELIST_RANDOM
2195/* Pre-initialize the random sequence cache */
2196static int init_cache_random_seq(struct kmem_cache *s)
2197{
2198 unsigned int count = oo_objects(s->oo);
2199 int err;
2200
2201 /* Bailout if already initialised */
2202 if (s->random_seq)
2203 return 0;
2204
2205 err = cache_random_seq_create(s, count, GFP_KERNEL);
2206 if (err) {
2207 pr_err("SLUB: Unable to initialize free list for %s\n",
2208 s->name);
2209 return err;
2210 }
2211
2212 /* Transform to an offset on the set of pages */
2213 if (s->random_seq) {
2214 unsigned int i;
2215
2216 for (i = 0; i < count; i++)
2217 s->random_seq[i] *= s->size;
2218 }
2219 return 0;
2220}
2221
2222/* Initialize each random sequence freelist per cache */
2223static void __init init_freelist_randomization(void)
2224{
2225 struct kmem_cache *s;
2226
2227 mutex_lock(&slab_mutex);
2228
2229 list_for_each_entry(s, &slab_caches, list)
2230 init_cache_random_seq(s);
2231
2232 mutex_unlock(&slab_mutex);
2233}
2234
2235/* Get the next entry on the pre-computed freelist randomized */
2236static void *next_freelist_entry(struct kmem_cache *s,
2237 unsigned long *pos, void *start,
2238 unsigned long page_limit,
2239 unsigned long freelist_count)
2240{
2241 unsigned int idx;
2242
2243 /*
2244 * If the target page allocation failed, the number of objects on the
2245 * page might be smaller than the usual size defined by the cache.
2246 */
2247 do {
2248 idx = s->random_seq[*pos];
2249 *pos += 1;
2250 if (*pos >= freelist_count)
2251 *pos = 0;
2252 } while (unlikely(idx >= page_limit));
2253
2254 return (char *)start + idx;
2255}
2256
2257/* Shuffle the single linked freelist based on a random pre-computed sequence */
2258static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2259{
2260 void *start;
2261 void *cur;
2262 void *next;
2263 unsigned long idx, pos, page_limit, freelist_count;
2264
2265 if (slab->objects < 2 || !s->random_seq)
2266 return false;
2267
2268 freelist_count = oo_objects(s->oo);
2269 pos = get_random_u32_below(freelist_count);
2270
2271 page_limit = slab->objects * s->size;
2272 start = fixup_red_left(s, slab_address(slab));
2273
2274 /* First entry is used as the base of the freelist */
2275 cur = next_freelist_entry(s, &pos, start, page_limit, freelist_count);
2276 cur = setup_object(s, cur);
2277 slab->freelist = cur;
2278
2279 for (idx = 1; idx < slab->objects; idx++) {
2280 next = next_freelist_entry(s, &pos, start, page_limit,
2281 freelist_count);
2282 next = setup_object(s, next);
2283 set_freepointer(s, cur, next);
2284 cur = next;
2285 }
2286 set_freepointer(s, cur, NULL);
2287
2288 return true;
2289}
2290#else
2291static inline int init_cache_random_seq(struct kmem_cache *s)
2292{
2293 return 0;
2294}
2295static inline void init_freelist_randomization(void) { }
2296static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2297{
2298 return false;
2299}
2300#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2301
2302static __always_inline void account_slab(struct slab *slab, int order,
2303 struct kmem_cache *s, gfp_t gfp)
2304{
2305 if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
2306 memcg_alloc_slab_cgroups(slab, s, gfp, true);
2307
2308 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2309 PAGE_SIZE << order);
2310}
2311
2312static __always_inline void unaccount_slab(struct slab *slab, int order,
2313 struct kmem_cache *s)
2314{
2315 if (memcg_kmem_online())
2316 memcg_free_slab_cgroups(slab);
2317
2318 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2319 -(PAGE_SIZE << order));
2320}
2321
2322static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2323{
2324 struct slab *slab;
2325 struct kmem_cache_order_objects oo = s->oo;
2326 gfp_t alloc_gfp;
2327 void *start, *p, *next;
2328 int idx;
2329 bool shuffle;
2330
2331 flags &= gfp_allowed_mask;
2332
2333 flags |= s->allocflags;
2334
2335 /*
2336 * Let the initial higher-order allocation fail under memory pressure
2337 * so we fall-back to the minimum order allocation.
2338 */
2339 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2340 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2341 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2342
2343 slab = alloc_slab_page(alloc_gfp, node, oo);
2344 if (unlikely(!slab)) {
2345 oo = s->min;
2346 alloc_gfp = flags;
2347 /*
2348 * Allocation may have failed due to fragmentation.
2349 * Try a lower order alloc if possible
2350 */
2351 slab = alloc_slab_page(alloc_gfp, node, oo);
2352 if (unlikely(!slab))
2353 return NULL;
2354 stat(s, ORDER_FALLBACK);
2355 }
2356
2357 slab->objects = oo_objects(oo);
2358 slab->inuse = 0;
2359 slab->frozen = 0;
2360
2361 account_slab(slab, oo_order(oo), s, flags);
2362
2363 slab->slab_cache = s;
2364
2365 kasan_poison_slab(slab);
2366
2367 start = slab_address(slab);
2368
2369 setup_slab_debug(s, slab, start);
2370
2371 shuffle = shuffle_freelist(s, slab);
2372
2373 if (!shuffle) {
2374 start = fixup_red_left(s, start);
2375 start = setup_object(s, start);
2376 slab->freelist = start;
2377 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2378 next = p + s->size;
2379 next = setup_object(s, next);
2380 set_freepointer(s, p, next);
2381 p = next;
2382 }
2383 set_freepointer(s, p, NULL);
2384 }
2385
2386 return slab;
2387}
2388
2389static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2390{
2391 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2392 flags = kmalloc_fix_flags(flags);
2393
2394 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2395
2396 return allocate_slab(s,
2397 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2398}
2399
2400static void __free_slab(struct kmem_cache *s, struct slab *slab)
2401{
2402 struct folio *folio = slab_folio(slab);
2403 int order = folio_order(folio);
2404 int pages = 1 << order;
2405
2406 __slab_clear_pfmemalloc(slab);
2407 folio->mapping = NULL;
2408 /* Make the mapping reset visible before clearing the flag */
2409 smp_wmb();
2410 __folio_clear_slab(folio);
2411 mm_account_reclaimed_pages(pages);
2412 unaccount_slab(slab, order, s);
2413 __free_pages(&folio->page, order);
2414}
2415
2416static void rcu_free_slab(struct rcu_head *h)
2417{
2418 struct slab *slab = container_of(h, struct slab, rcu_head);
2419
2420 __free_slab(slab->slab_cache, slab);
2421}
2422
2423static void free_slab(struct kmem_cache *s, struct slab *slab)
2424{
2425 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2426 void *p;
2427
2428 slab_pad_check(s, slab);
2429 for_each_object(p, s, slab_address(slab), slab->objects)
2430 check_object(s, slab, p, SLUB_RED_INACTIVE);
2431 }
2432
2433 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2434 call_rcu(&slab->rcu_head, rcu_free_slab);
2435 else
2436 __free_slab(s, slab);
2437}
2438
2439static void discard_slab(struct kmem_cache *s, struct slab *slab)
2440{
2441 dec_slabs_node(s, slab_nid(slab), slab->objects);
2442 free_slab(s, slab);
2443}
2444
2445/*
2446 * SLUB reuses PG_workingset bit to keep track of whether it's on
2447 * the per-node partial list.
2448 */
2449static inline bool slab_test_node_partial(const struct slab *slab)
2450{
2451 return folio_test_workingset((struct folio *)slab_folio(slab));
2452}
2453
2454static inline void slab_set_node_partial(struct slab *slab)
2455{
2456 set_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2457}
2458
2459static inline void slab_clear_node_partial(struct slab *slab)
2460{
2461 clear_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2462}
2463
2464/*
2465 * Management of partially allocated slabs.
2466 */
2467static inline void
2468__add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2469{
2470 n->nr_partial++;
2471 if (tail == DEACTIVATE_TO_TAIL)
2472 list_add_tail(&slab->slab_list, &n->partial);
2473 else
2474 list_add(&slab->slab_list, &n->partial);
2475 slab_set_node_partial(slab);
2476}
2477
2478static inline void add_partial(struct kmem_cache_node *n,
2479 struct slab *slab, int tail)
2480{
2481 lockdep_assert_held(&n->list_lock);
2482 __add_partial(n, slab, tail);
2483}
2484
2485static inline void remove_partial(struct kmem_cache_node *n,
2486 struct slab *slab)
2487{
2488 lockdep_assert_held(&n->list_lock);
2489 list_del(&slab->slab_list);
2490 slab_clear_node_partial(slab);
2491 n->nr_partial--;
2492}
2493
2494/*
2495 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
2496 * slab from the n->partial list. Remove only a single object from the slab, do
2497 * the alloc_debug_processing() checks and leave the slab on the list, or move
2498 * it to full list if it was the last free object.
2499 */
2500static void *alloc_single_from_partial(struct kmem_cache *s,
2501 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2502{
2503 void *object;
2504
2505 lockdep_assert_held(&n->list_lock);
2506
2507 object = slab->freelist;
2508 slab->freelist = get_freepointer(s, object);
2509 slab->inuse++;
2510
2511 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2512 remove_partial(n, slab);
2513 return NULL;
2514 }
2515
2516 if (slab->inuse == slab->objects) {
2517 remove_partial(n, slab);
2518 add_full(s, n, slab);
2519 }
2520
2521 return object;
2522}
2523
2524/*
2525 * Called only for kmem_cache_debug() caches to allocate from a freshly
2526 * allocated slab. Allocate a single object instead of whole freelist
2527 * and put the slab to the partial (or full) list.
2528 */
2529static void *alloc_single_from_new_slab(struct kmem_cache *s,
2530 struct slab *slab, int orig_size)
2531{
2532 int nid = slab_nid(slab);
2533 struct kmem_cache_node *n = get_node(s, nid);
2534 unsigned long flags;
2535 void *object;
2536
2537
2538 object = slab->freelist;
2539 slab->freelist = get_freepointer(s, object);
2540 slab->inuse = 1;
2541
2542 if (!alloc_debug_processing(s, slab, object, orig_size))
2543 /*
2544 * It's not really expected that this would fail on a
2545 * freshly allocated slab, but a concurrent memory
2546 * corruption in theory could cause that.
2547 */
2548 return NULL;
2549
2550 spin_lock_irqsave(&n->list_lock, flags);
2551
2552 if (slab->inuse == slab->objects)
2553 add_full(s, n, slab);
2554 else
2555 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2556
2557 inc_slabs_node(s, nid, slab->objects);
2558 spin_unlock_irqrestore(&n->list_lock, flags);
2559
2560 return object;
2561}
2562
2563#ifdef CONFIG_SLUB_CPU_PARTIAL
2564static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2565#else
2566static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2567 int drain) { }
2568#endif
2569static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2570
2571/*
2572 * Try to allocate a partial slab from a specific node.
2573 */
2574static struct slab *get_partial_node(struct kmem_cache *s,
2575 struct kmem_cache_node *n,
2576 struct partial_context *pc)
2577{
2578 struct slab *slab, *slab2, *partial = NULL;
2579 unsigned long flags;
2580 unsigned int partial_slabs = 0;
2581
2582 /*
2583 * Racy check. If we mistakenly see no partial slabs then we
2584 * just allocate an empty slab. If we mistakenly try to get a
2585 * partial slab and there is none available then get_partial()
2586 * will return NULL.
2587 */
2588 if (!n || !n->nr_partial)
2589 return NULL;
2590
2591 spin_lock_irqsave(&n->list_lock, flags);
2592 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2593 if (!pfmemalloc_match(slab, pc->flags))
2594 continue;
2595
2596 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2597 void *object = alloc_single_from_partial(s, n, slab,
2598 pc->orig_size);
2599 if (object) {
2600 partial = slab;
2601 pc->object = object;
2602 break;
2603 }
2604 continue;
2605 }
2606
2607 remove_partial(n, slab);
2608
2609 if (!partial) {
2610 partial = slab;
2611 stat(s, ALLOC_FROM_PARTIAL);
2612 } else {
2613 put_cpu_partial(s, slab, 0);
2614 stat(s, CPU_PARTIAL_NODE);
2615 partial_slabs++;
2616 }
2617#ifdef CONFIG_SLUB_CPU_PARTIAL
2618 if (!kmem_cache_has_cpu_partial(s)
2619 || partial_slabs > s->cpu_partial_slabs / 2)
2620 break;
2621#else
2622 break;
2623#endif
2624
2625 }
2626 spin_unlock_irqrestore(&n->list_lock, flags);
2627 return partial;
2628}
2629
2630/*
2631 * Get a slab from somewhere. Search in increasing NUMA distances.
2632 */
2633static struct slab *get_any_partial(struct kmem_cache *s,
2634 struct partial_context *pc)
2635{
2636#ifdef CONFIG_NUMA
2637 struct zonelist *zonelist;
2638 struct zoneref *z;
2639 struct zone *zone;
2640 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2641 struct slab *slab;
2642 unsigned int cpuset_mems_cookie;
2643
2644 /*
2645 * The defrag ratio allows a configuration of the tradeoffs between
2646 * inter node defragmentation and node local allocations. A lower
2647 * defrag_ratio increases the tendency to do local allocations
2648 * instead of attempting to obtain partial slabs from other nodes.
2649 *
2650 * If the defrag_ratio is set to 0 then kmalloc() always
2651 * returns node local objects. If the ratio is higher then kmalloc()
2652 * may return off node objects because partial slabs are obtained
2653 * from other nodes and filled up.
2654 *
2655 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2656 * (which makes defrag_ratio = 1000) then every (well almost)
2657 * allocation will first attempt to defrag slab caches on other nodes.
2658 * This means scanning over all nodes to look for partial slabs which
2659 * may be expensive if we do it every time we are trying to find a slab
2660 * with available objects.
2661 */
2662 if (!s->remote_node_defrag_ratio ||
2663 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2664 return NULL;
2665
2666 do {
2667 cpuset_mems_cookie = read_mems_allowed_begin();
2668 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2669 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2670 struct kmem_cache_node *n;
2671
2672 n = get_node(s, zone_to_nid(zone));
2673
2674 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2675 n->nr_partial > s->min_partial) {
2676 slab = get_partial_node(s, n, pc);
2677 if (slab) {
2678 /*
2679 * Don't check read_mems_allowed_retry()
2680 * here - if mems_allowed was updated in
2681 * parallel, that was a harmless race
2682 * between allocation and the cpuset
2683 * update
2684 */
2685 return slab;
2686 }
2687 }
2688 }
2689 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2690#endif /* CONFIG_NUMA */
2691 return NULL;
2692}
2693
2694/*
2695 * Get a partial slab, lock it and return it.
2696 */
2697static struct slab *get_partial(struct kmem_cache *s, int node,
2698 struct partial_context *pc)
2699{
2700 struct slab *slab;
2701 int searchnode = node;
2702
2703 if (node == NUMA_NO_NODE)
2704 searchnode = numa_mem_id();
2705
2706 slab = get_partial_node(s, get_node(s, searchnode), pc);
2707 if (slab || node != NUMA_NO_NODE)
2708 return slab;
2709
2710 return get_any_partial(s, pc);
2711}
2712
2713#ifndef CONFIG_SLUB_TINY
2714
2715#ifdef CONFIG_PREEMPTION
2716/*
2717 * Calculate the next globally unique transaction for disambiguation
2718 * during cmpxchg. The transactions start with the cpu number and are then
2719 * incremented by CONFIG_NR_CPUS.
2720 */
2721#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2722#else
2723/*
2724 * No preemption supported therefore also no need to check for
2725 * different cpus.
2726 */
2727#define TID_STEP 1
2728#endif /* CONFIG_PREEMPTION */
2729
2730static inline unsigned long next_tid(unsigned long tid)
2731{
2732 return tid + TID_STEP;
2733}
2734
2735#ifdef SLUB_DEBUG_CMPXCHG
2736static inline unsigned int tid_to_cpu(unsigned long tid)
2737{
2738 return tid % TID_STEP;
2739}
2740
2741static inline unsigned long tid_to_event(unsigned long tid)
2742{
2743 return tid / TID_STEP;
2744}
2745#endif
2746
2747static inline unsigned int init_tid(int cpu)
2748{
2749 return cpu;
2750}
2751
2752static inline void note_cmpxchg_failure(const char *n,
2753 const struct kmem_cache *s, unsigned long tid)
2754{
2755#ifdef SLUB_DEBUG_CMPXCHG
2756 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2757
2758 pr_info("%s %s: cmpxchg redo ", n, s->name);
2759
2760#ifdef CONFIG_PREEMPTION
2761 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2762 pr_warn("due to cpu change %d -> %d\n",
2763 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2764 else
2765#endif
2766 if (tid_to_event(tid) != tid_to_event(actual_tid))
2767 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2768 tid_to_event(tid), tid_to_event(actual_tid));
2769 else
2770 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2771 actual_tid, tid, next_tid(tid));
2772#endif
2773 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2774}
2775
2776static void init_kmem_cache_cpus(struct kmem_cache *s)
2777{
2778 int cpu;
2779 struct kmem_cache_cpu *c;
2780
2781 for_each_possible_cpu(cpu) {
2782 c = per_cpu_ptr(s->cpu_slab, cpu);
2783 local_lock_init(&c->lock);
2784 c->tid = init_tid(cpu);
2785 }
2786}
2787
2788/*
2789 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
2790 * unfreezes the slabs and puts it on the proper list.
2791 * Assumes the slab has been already safely taken away from kmem_cache_cpu
2792 * by the caller.
2793 */
2794static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
2795 void *freelist)
2796{
2797 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
2798 int free_delta = 0;
2799 void *nextfree, *freelist_iter, *freelist_tail;
2800 int tail = DEACTIVATE_TO_HEAD;
2801 unsigned long flags = 0;
2802 struct slab new;
2803 struct slab old;
2804
2805 if (slab->freelist) {
2806 stat(s, DEACTIVATE_REMOTE_FREES);
2807 tail = DEACTIVATE_TO_TAIL;
2808 }
2809
2810 /*
2811 * Stage one: Count the objects on cpu's freelist as free_delta and
2812 * remember the last object in freelist_tail for later splicing.
2813 */
2814 freelist_tail = NULL;
2815 freelist_iter = freelist;
2816 while (freelist_iter) {
2817 nextfree = get_freepointer(s, freelist_iter);
2818
2819 /*
2820 * If 'nextfree' is invalid, it is possible that the object at
2821 * 'freelist_iter' is already corrupted. So isolate all objects
2822 * starting at 'freelist_iter' by skipping them.
2823 */
2824 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
2825 break;
2826
2827 freelist_tail = freelist_iter;
2828 free_delta++;
2829
2830 freelist_iter = nextfree;
2831 }
2832
2833 /*
2834 * Stage two: Unfreeze the slab while splicing the per-cpu
2835 * freelist to the head of slab's freelist.
2836 */
2837 do {
2838 old.freelist = READ_ONCE(slab->freelist);
2839 old.counters = READ_ONCE(slab->counters);
2840 VM_BUG_ON(!old.frozen);
2841
2842 /* Determine target state of the slab */
2843 new.counters = old.counters;
2844 new.frozen = 0;
2845 if (freelist_tail) {
2846 new.inuse -= free_delta;
2847 set_freepointer(s, freelist_tail, old.freelist);
2848 new.freelist = freelist;
2849 } else {
2850 new.freelist = old.freelist;
2851 }
2852 } while (!slab_update_freelist(s, slab,
2853 old.freelist, old.counters,
2854 new.freelist, new.counters,
2855 "unfreezing slab"));
2856
2857 /*
2858 * Stage three: Manipulate the slab list based on the updated state.
2859 */
2860 if (!new.inuse && n->nr_partial >= s->min_partial) {
2861 stat(s, DEACTIVATE_EMPTY);
2862 discard_slab(s, slab);
2863 stat(s, FREE_SLAB);
2864 } else if (new.freelist) {
2865 spin_lock_irqsave(&n->list_lock, flags);
2866 add_partial(n, slab, tail);
2867 spin_unlock_irqrestore(&n->list_lock, flags);
2868 stat(s, tail);
2869 } else {
2870 stat(s, DEACTIVATE_FULL);
2871 }
2872}
2873
2874#ifdef CONFIG_SLUB_CPU_PARTIAL
2875static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
2876{
2877 struct kmem_cache_node *n = NULL, *n2 = NULL;
2878 struct slab *slab, *slab_to_discard = NULL;
2879 unsigned long flags = 0;
2880
2881 while (partial_slab) {
2882 slab = partial_slab;
2883 partial_slab = slab->next;
2884
2885 n2 = get_node(s, slab_nid(slab));
2886 if (n != n2) {
2887 if (n)
2888 spin_unlock_irqrestore(&n->list_lock, flags);
2889
2890 n = n2;
2891 spin_lock_irqsave(&n->list_lock, flags);
2892 }
2893
2894 if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
2895 slab->next = slab_to_discard;
2896 slab_to_discard = slab;
2897 } else {
2898 add_partial(n, slab, DEACTIVATE_TO_TAIL);
2899 stat(s, FREE_ADD_PARTIAL);
2900 }
2901 }
2902
2903 if (n)
2904 spin_unlock_irqrestore(&n->list_lock, flags);
2905
2906 while (slab_to_discard) {
2907 slab = slab_to_discard;
2908 slab_to_discard = slab_to_discard->next;
2909
2910 stat(s, DEACTIVATE_EMPTY);
2911 discard_slab(s, slab);
2912 stat(s, FREE_SLAB);
2913 }
2914}
2915
2916/*
2917 * Put all the cpu partial slabs to the node partial list.
2918 */
2919static void put_partials(struct kmem_cache *s)
2920{
2921 struct slab *partial_slab;
2922 unsigned long flags;
2923
2924 local_lock_irqsave(&s->cpu_slab->lock, flags);
2925 partial_slab = this_cpu_read(s->cpu_slab->partial);
2926 this_cpu_write(s->cpu_slab->partial, NULL);
2927 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2928
2929 if (partial_slab)
2930 __put_partials(s, partial_slab);
2931}
2932
2933static void put_partials_cpu(struct kmem_cache *s,
2934 struct kmem_cache_cpu *c)
2935{
2936 struct slab *partial_slab;
2937
2938 partial_slab = slub_percpu_partial(c);
2939 c->partial = NULL;
2940
2941 if (partial_slab)
2942 __put_partials(s, partial_slab);
2943}
2944
2945/*
2946 * Put a slab into a partial slab slot if available.
2947 *
2948 * If we did not find a slot then simply move all the partials to the
2949 * per node partial list.
2950 */
2951static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
2952{
2953 struct slab *oldslab;
2954 struct slab *slab_to_put = NULL;
2955 unsigned long flags;
2956 int slabs = 0;
2957
2958 local_lock_irqsave(&s->cpu_slab->lock, flags);
2959
2960 oldslab = this_cpu_read(s->cpu_slab->partial);
2961
2962 if (oldslab) {
2963 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
2964 /*
2965 * Partial array is full. Move the existing set to the
2966 * per node partial list. Postpone the actual unfreezing
2967 * outside of the critical section.
2968 */
2969 slab_to_put = oldslab;
2970 oldslab = NULL;
2971 } else {
2972 slabs = oldslab->slabs;
2973 }
2974 }
2975
2976 slabs++;
2977
2978 slab->slabs = slabs;
2979 slab->next = oldslab;
2980
2981 this_cpu_write(s->cpu_slab->partial, slab);
2982
2983 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
2984
2985 if (slab_to_put) {
2986 __put_partials(s, slab_to_put);
2987 stat(s, CPU_PARTIAL_DRAIN);
2988 }
2989}
2990
2991#else /* CONFIG_SLUB_CPU_PARTIAL */
2992
2993static inline void put_partials(struct kmem_cache *s) { }
2994static inline void put_partials_cpu(struct kmem_cache *s,
2995 struct kmem_cache_cpu *c) { }
2996
2997#endif /* CONFIG_SLUB_CPU_PARTIAL */
2998
2999static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
3000{
3001 unsigned long flags;
3002 struct slab *slab;
3003 void *freelist;
3004
3005 local_lock_irqsave(&s->cpu_slab->lock, flags);
3006
3007 slab = c->slab;
3008 freelist = c->freelist;
3009
3010 c->slab = NULL;
3011 c->freelist = NULL;
3012 c->tid = next_tid(c->tid);
3013
3014 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3015
3016 if (slab) {
3017 deactivate_slab(s, slab, freelist);
3018 stat(s, CPUSLAB_FLUSH);
3019 }
3020}
3021
3022static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
3023{
3024 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3025 void *freelist = c->freelist;
3026 struct slab *slab = c->slab;
3027
3028 c->slab = NULL;
3029 c->freelist = NULL;
3030 c->tid = next_tid(c->tid);
3031
3032 if (slab) {
3033 deactivate_slab(s, slab, freelist);
3034 stat(s, CPUSLAB_FLUSH);
3035 }
3036
3037 put_partials_cpu(s, c);
3038}
3039
3040struct slub_flush_work {
3041 struct work_struct work;
3042 struct kmem_cache *s;
3043 bool skip;
3044};
3045
3046/*
3047 * Flush cpu slab.
3048 *
3049 * Called from CPU work handler with migration disabled.
3050 */
3051static void flush_cpu_slab(struct work_struct *w)
3052{
3053 struct kmem_cache *s;
3054 struct kmem_cache_cpu *c;
3055 struct slub_flush_work *sfw;
3056
3057 sfw = container_of(w, struct slub_flush_work, work);
3058
3059 s = sfw->s;
3060 c = this_cpu_ptr(s->cpu_slab);
3061
3062 if (c->slab)
3063 flush_slab(s, c);
3064
3065 put_partials(s);
3066}
3067
3068static bool has_cpu_slab(int cpu, struct kmem_cache *s)
3069{
3070 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3071
3072 return c->slab || slub_percpu_partial(c);
3073}
3074
3075static DEFINE_MUTEX(flush_lock);
3076static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
3077
3078static void flush_all_cpus_locked(struct kmem_cache *s)
3079{
3080 struct slub_flush_work *sfw;
3081 unsigned int cpu;
3082
3083 lockdep_assert_cpus_held();
3084 mutex_lock(&flush_lock);
3085
3086 for_each_online_cpu(cpu) {
3087 sfw = &per_cpu(slub_flush, cpu);
3088 if (!has_cpu_slab(cpu, s)) {
3089 sfw->skip = true;
3090 continue;
3091 }
3092 INIT_WORK(&sfw->work, flush_cpu_slab);
3093 sfw->skip = false;
3094 sfw->s = s;
3095 queue_work_on(cpu, flushwq, &sfw->work);
3096 }
3097
3098 for_each_online_cpu(cpu) {
3099 sfw = &per_cpu(slub_flush, cpu);
3100 if (sfw->skip)
3101 continue;
3102 flush_work(&sfw->work);
3103 }
3104
3105 mutex_unlock(&flush_lock);
3106}
3107
3108static void flush_all(struct kmem_cache *s)
3109{
3110 cpus_read_lock();
3111 flush_all_cpus_locked(s);
3112 cpus_read_unlock();
3113}
3114
3115/*
3116 * Use the cpu notifier to insure that the cpu slabs are flushed when
3117 * necessary.
3118 */
3119static int slub_cpu_dead(unsigned int cpu)
3120{
3121 struct kmem_cache *s;
3122
3123 mutex_lock(&slab_mutex);
3124 list_for_each_entry(s, &slab_caches, list)
3125 __flush_cpu_slab(s, cpu);
3126 mutex_unlock(&slab_mutex);
3127 return 0;
3128}
3129
3130#else /* CONFIG_SLUB_TINY */
3131static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
3132static inline void flush_all(struct kmem_cache *s) { }
3133static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
3134static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
3135#endif /* CONFIG_SLUB_TINY */
3136
3137/*
3138 * Check if the objects in a per cpu structure fit numa
3139 * locality expectations.
3140 */
3141static inline int node_match(struct slab *slab, int node)
3142{
3143#ifdef CONFIG_NUMA
3144 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
3145 return 0;
3146#endif
3147 return 1;
3148}
3149
3150#ifdef CONFIG_SLUB_DEBUG
3151static int count_free(struct slab *slab)
3152{
3153 return slab->objects - slab->inuse;
3154}
3155
3156static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
3157{
3158 return atomic_long_read(&n->total_objects);
3159}
3160
3161/* Supports checking bulk free of a constructed freelist */
3162static inline bool free_debug_processing(struct kmem_cache *s,
3163 struct slab *slab, void *head, void *tail, int *bulk_cnt,
3164 unsigned long addr, depot_stack_handle_t handle)
3165{
3166 bool checks_ok = false;
3167 void *object = head;
3168 int cnt = 0;
3169
3170 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3171 if (!check_slab(s, slab))
3172 goto out;
3173 }
3174
3175 if (slab->inuse < *bulk_cnt) {
3176 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
3177 slab->inuse, *bulk_cnt);
3178 goto out;
3179 }
3180
3181next_object:
3182
3183 if (++cnt > *bulk_cnt)
3184 goto out_cnt;
3185
3186 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3187 if (!free_consistency_checks(s, slab, object, addr))
3188 goto out;
3189 }
3190
3191 if (s->flags & SLAB_STORE_USER)
3192 set_track_update(s, object, TRACK_FREE, addr, handle);
3193 trace(s, slab, object, 0);
3194 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
3195 init_object(s, object, SLUB_RED_INACTIVE);
3196
3197 /* Reached end of constructed freelist yet? */
3198 if (object != tail) {
3199 object = get_freepointer(s, object);
3200 goto next_object;
3201 }
3202 checks_ok = true;
3203
3204out_cnt:
3205 if (cnt != *bulk_cnt) {
3206 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
3207 *bulk_cnt, cnt);
3208 *bulk_cnt = cnt;
3209 }
3210
3211out:
3212
3213 if (!checks_ok)
3214 slab_fix(s, "Object at 0x%p not freed", object);
3215
3216 return checks_ok;
3217}
3218#endif /* CONFIG_SLUB_DEBUG */
3219
3220#if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
3221static unsigned long count_partial(struct kmem_cache_node *n,
3222 int (*get_count)(struct slab *))
3223{
3224 unsigned long flags;
3225 unsigned long x = 0;
3226 struct slab *slab;
3227
3228 spin_lock_irqsave(&n->list_lock, flags);
3229 list_for_each_entry(slab, &n->partial, slab_list)
3230 x += get_count(slab);
3231 spin_unlock_irqrestore(&n->list_lock, flags);
3232 return x;
3233}
3234#endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
3235
3236#ifdef CONFIG_SLUB_DEBUG
3237static noinline void
3238slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
3239{
3240 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
3241 DEFAULT_RATELIMIT_BURST);
3242 int node;
3243 struct kmem_cache_node *n;
3244
3245 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
3246 return;
3247
3248 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
3249 nid, gfpflags, &gfpflags);
3250 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
3251 s->name, s->object_size, s->size, oo_order(s->oo),
3252 oo_order(s->min));
3253
3254 if (oo_order(s->min) > get_order(s->object_size))
3255 pr_warn(" %s debugging increased min order, use slab_debug=O to disable.\n",
3256 s->name);
3257
3258 for_each_kmem_cache_node(s, node, n) {
3259 unsigned long nr_slabs;
3260 unsigned long nr_objs;
3261 unsigned long nr_free;
3262
3263 nr_free = count_partial(n, count_free);
3264 nr_slabs = node_nr_slabs(n);
3265 nr_objs = node_nr_objs(n);
3266
3267 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
3268 node, nr_slabs, nr_objs, nr_free);
3269 }
3270}
3271#else /* CONFIG_SLUB_DEBUG */
3272static inline void
3273slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3274#endif
3275
3276static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3277{
3278 if (unlikely(slab_test_pfmemalloc(slab)))
3279 return gfp_pfmemalloc_allowed(gfpflags);
3280
3281 return true;
3282}
3283
3284#ifndef CONFIG_SLUB_TINY
3285static inline bool
3286__update_cpu_freelist_fast(struct kmem_cache *s,
3287 void *freelist_old, void *freelist_new,
3288 unsigned long tid)
3289{
3290 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3291 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3292
3293 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3294 &old.full, new.full);
3295}
3296
3297/*
3298 * Check the slab->freelist and either transfer the freelist to the
3299 * per cpu freelist or deactivate the slab.
3300 *
3301 * The slab is still frozen if the return value is not NULL.
3302 *
3303 * If this function returns NULL then the slab has been unfrozen.
3304 */
3305static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3306{
3307 struct slab new;
3308 unsigned long counters;
3309 void *freelist;
3310
3311 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3312
3313 do {
3314 freelist = slab->freelist;
3315 counters = slab->counters;
3316
3317 new.counters = counters;
3318
3319 new.inuse = slab->objects;
3320 new.frozen = freelist != NULL;
3321
3322 } while (!__slab_update_freelist(s, slab,
3323 freelist, counters,
3324 NULL, new.counters,
3325 "get_freelist"));
3326
3327 return freelist;
3328}
3329
3330/*
3331 * Freeze the partial slab and return the pointer to the freelist.
3332 */
3333static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
3334{
3335 struct slab new;
3336 unsigned long counters;
3337 void *freelist;
3338
3339 do {
3340 freelist = slab->freelist;
3341 counters = slab->counters;
3342
3343 new.counters = counters;
3344 VM_BUG_ON(new.frozen);
3345
3346 new.inuse = slab->objects;
3347 new.frozen = 1;
3348
3349 } while (!slab_update_freelist(s, slab,
3350 freelist, counters,
3351 NULL, new.counters,
3352 "freeze_slab"));
3353
3354 return freelist;
3355}
3356
3357/*
3358 * Slow path. The lockless freelist is empty or we need to perform
3359 * debugging duties.
3360 *
3361 * Processing is still very fast if new objects have been freed to the
3362 * regular freelist. In that case we simply take over the regular freelist
3363 * as the lockless freelist and zap the regular freelist.
3364 *
3365 * If that is not working then we fall back to the partial lists. We take the
3366 * first element of the freelist as the object to allocate now and move the
3367 * rest of the freelist to the lockless freelist.
3368 *
3369 * And if we were unable to get a new slab from the partial slab lists then
3370 * we need to allocate a new slab. This is the slowest path since it involves
3371 * a call to the page allocator and the setup of a new slab.
3372 *
3373 * Version of __slab_alloc to use when we know that preemption is
3374 * already disabled (which is the case for bulk allocation).
3375 */
3376static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3377 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3378{
3379 void *freelist;
3380 struct slab *slab;
3381 unsigned long flags;
3382 struct partial_context pc;
3383
3384 stat(s, ALLOC_SLOWPATH);
3385
3386reread_slab:
3387
3388 slab = READ_ONCE(c->slab);
3389 if (!slab) {
3390 /*
3391 * if the node is not online or has no normal memory, just
3392 * ignore the node constraint
3393 */
3394 if (unlikely(node != NUMA_NO_NODE &&
3395 !node_isset(node, slab_nodes)))
3396 node = NUMA_NO_NODE;
3397 goto new_slab;
3398 }
3399
3400 if (unlikely(!node_match(slab, node))) {
3401 /*
3402 * same as above but node_match() being false already
3403 * implies node != NUMA_NO_NODE
3404 */
3405 if (!node_isset(node, slab_nodes)) {
3406 node = NUMA_NO_NODE;
3407 } else {
3408 stat(s, ALLOC_NODE_MISMATCH);
3409 goto deactivate_slab;
3410 }
3411 }
3412
3413 /*
3414 * By rights, we should be searching for a slab page that was
3415 * PFMEMALLOC but right now, we are losing the pfmemalloc
3416 * information when the page leaves the per-cpu allocator
3417 */
3418 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3419 goto deactivate_slab;
3420
3421 /* must check again c->slab in case we got preempted and it changed */
3422 local_lock_irqsave(&s->cpu_slab->lock, flags);
3423 if (unlikely(slab != c->slab)) {
3424 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3425 goto reread_slab;
3426 }
3427 freelist = c->freelist;
3428 if (freelist)
3429 goto load_freelist;
3430
3431 freelist = get_freelist(s, slab);
3432
3433 if (!freelist) {
3434 c->slab = NULL;
3435 c->tid = next_tid(c->tid);
3436 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3437 stat(s, DEACTIVATE_BYPASS);
3438 goto new_slab;
3439 }
3440
3441 stat(s, ALLOC_REFILL);
3442
3443load_freelist:
3444
3445 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3446
3447 /*
3448 * freelist is pointing to the list of objects to be used.
3449 * slab is pointing to the slab from which the objects are obtained.
3450 * That slab must be frozen for per cpu allocations to work.
3451 */
3452 VM_BUG_ON(!c->slab->frozen);
3453 c->freelist = get_freepointer(s, freelist);
3454 c->tid = next_tid(c->tid);
3455 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3456 return freelist;
3457
3458deactivate_slab:
3459
3460 local_lock_irqsave(&s->cpu_slab->lock, flags);
3461 if (slab != c->slab) {
3462 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3463 goto reread_slab;
3464 }
3465 freelist = c->freelist;
3466 c->slab = NULL;
3467 c->freelist = NULL;
3468 c->tid = next_tid(c->tid);
3469 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3470 deactivate_slab(s, slab, freelist);
3471
3472new_slab:
3473
3474#ifdef CONFIG_SLUB_CPU_PARTIAL
3475 while (slub_percpu_partial(c)) {
3476 local_lock_irqsave(&s->cpu_slab->lock, flags);
3477 if (unlikely(c->slab)) {
3478 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3479 goto reread_slab;
3480 }
3481 if (unlikely(!slub_percpu_partial(c))) {
3482 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3483 /* we were preempted and partial list got empty */
3484 goto new_objects;
3485 }
3486
3487 slab = slub_percpu_partial(c);
3488 slub_set_percpu_partial(c, slab);
3489
3490 if (likely(node_match(slab, node) &&
3491 pfmemalloc_match(slab, gfpflags))) {
3492 c->slab = slab;
3493 freelist = get_freelist(s, slab);
3494 VM_BUG_ON(!freelist);
3495 stat(s, CPU_PARTIAL_ALLOC);
3496 goto load_freelist;
3497 }
3498
3499 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3500
3501 slab->next = NULL;
3502 __put_partials(s, slab);
3503 }
3504#endif
3505
3506new_objects:
3507
3508 pc.flags = gfpflags;
3509 pc.orig_size = orig_size;
3510 slab = get_partial(s, node, &pc);
3511 if (slab) {
3512 if (kmem_cache_debug(s)) {
3513 freelist = pc.object;
3514 /*
3515 * For debug caches here we had to go through
3516 * alloc_single_from_partial() so just store the
3517 * tracking info and return the object.
3518 */
3519 if (s->flags & SLAB_STORE_USER)
3520 set_track(s, freelist, TRACK_ALLOC, addr);
3521
3522 return freelist;
3523 }
3524
3525 freelist = freeze_slab(s, slab);
3526 goto retry_load_slab;
3527 }
3528
3529 slub_put_cpu_ptr(s->cpu_slab);
3530 slab = new_slab(s, gfpflags, node);
3531 c = slub_get_cpu_ptr(s->cpu_slab);
3532
3533 if (unlikely(!slab)) {
3534 slab_out_of_memory(s, gfpflags, node);
3535 return NULL;
3536 }
3537
3538 stat(s, ALLOC_SLAB);
3539
3540 if (kmem_cache_debug(s)) {
3541 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3542
3543 if (unlikely(!freelist))
3544 goto new_objects;
3545
3546 if (s->flags & SLAB_STORE_USER)
3547 set_track(s, freelist, TRACK_ALLOC, addr);
3548
3549 return freelist;
3550 }
3551
3552 /*
3553 * No other reference to the slab yet so we can
3554 * muck around with it freely without cmpxchg
3555 */
3556 freelist = slab->freelist;
3557 slab->freelist = NULL;
3558 slab->inuse = slab->objects;
3559 slab->frozen = 1;
3560
3561 inc_slabs_node(s, slab_nid(slab), slab->objects);
3562
3563 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3564 /*
3565 * For !pfmemalloc_match() case we don't load freelist so that
3566 * we don't make further mismatched allocations easier.
3567 */
3568 deactivate_slab(s, slab, get_freepointer(s, freelist));
3569 return freelist;
3570 }
3571
3572retry_load_slab:
3573
3574 local_lock_irqsave(&s->cpu_slab->lock, flags);
3575 if (unlikely(c->slab)) {
3576 void *flush_freelist = c->freelist;
3577 struct slab *flush_slab = c->slab;
3578
3579 c->slab = NULL;
3580 c->freelist = NULL;
3581 c->tid = next_tid(c->tid);
3582
3583 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3584
3585 deactivate_slab(s, flush_slab, flush_freelist);
3586
3587 stat(s, CPUSLAB_FLUSH);
3588
3589 goto retry_load_slab;
3590 }
3591 c->slab = slab;
3592
3593 goto load_freelist;
3594}
3595
3596/*
3597 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3598 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3599 * pointer.
3600 */
3601static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3602 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3603{
3604 void *p;
3605
3606#ifdef CONFIG_PREEMPT_COUNT
3607 /*
3608 * We may have been preempted and rescheduled on a different
3609 * cpu before disabling preemption. Need to reload cpu area
3610 * pointer.
3611 */
3612 c = slub_get_cpu_ptr(s->cpu_slab);
3613#endif
3614
3615 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3616#ifdef CONFIG_PREEMPT_COUNT
3617 slub_put_cpu_ptr(s->cpu_slab);
3618#endif
3619 return p;
3620}
3621
3622static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3623 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3624{
3625 struct kmem_cache_cpu *c;
3626 struct slab *slab;
3627 unsigned long tid;
3628 void *object;
3629
3630redo:
3631 /*
3632 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3633 * enabled. We may switch back and forth between cpus while
3634 * reading from one cpu area. That does not matter as long
3635 * as we end up on the original cpu again when doing the cmpxchg.
3636 *
3637 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3638 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3639 * the tid. If we are preempted and switched to another cpu between the
3640 * two reads, it's OK as the two are still associated with the same cpu
3641 * and cmpxchg later will validate the cpu.
3642 */
3643 c = raw_cpu_ptr(s->cpu_slab);
3644 tid = READ_ONCE(c->tid);
3645
3646 /*
3647 * Irqless object alloc/free algorithm used here depends on sequence
3648 * of fetching cpu_slab's data. tid should be fetched before anything
3649 * on c to guarantee that object and slab associated with previous tid
3650 * won't be used with current tid. If we fetch tid first, object and
3651 * slab could be one associated with next tid and our alloc/free
3652 * request will be failed. In this case, we will retry. So, no problem.
3653 */
3654 barrier();
3655
3656 /*
3657 * The transaction ids are globally unique per cpu and per operation on
3658 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3659 * occurs on the right processor and that there was no operation on the
3660 * linked list in between.
3661 */
3662
3663 object = c->freelist;
3664 slab = c->slab;
3665
3666 if (!USE_LOCKLESS_FAST_PATH() ||
3667 unlikely(!object || !slab || !node_match(slab, node))) {
3668 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3669 } else {
3670 void *next_object = get_freepointer_safe(s, object);
3671
3672 /*
3673 * The cmpxchg will only match if there was no additional
3674 * operation and if we are on the right processor.
3675 *
3676 * The cmpxchg does the following atomically (without lock
3677 * semantics!)
3678 * 1. Relocate first pointer to the current per cpu area.
3679 * 2. Verify that tid and freelist have not been changed
3680 * 3. If they were not changed replace tid and freelist
3681 *
3682 * Since this is without lock semantics the protection is only
3683 * against code executing on this cpu *not* from access by
3684 * other cpus.
3685 */
3686 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
3687 note_cmpxchg_failure("slab_alloc", s, tid);
3688 goto redo;
3689 }
3690 prefetch_freepointer(s, next_object);
3691 stat(s, ALLOC_FASTPATH);
3692 }
3693
3694 return object;
3695}
3696#else /* CONFIG_SLUB_TINY */
3697static void *__slab_alloc_node(struct kmem_cache *s,
3698 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3699{
3700 struct partial_context pc;
3701 struct slab *slab;
3702 void *object;
3703
3704 pc.flags = gfpflags;
3705 pc.orig_size = orig_size;
3706 slab = get_partial(s, node, &pc);
3707
3708 if (slab)
3709 return pc.object;
3710
3711 slab = new_slab(s, gfpflags, node);
3712 if (unlikely(!slab)) {
3713 slab_out_of_memory(s, gfpflags, node);
3714 return NULL;
3715 }
3716
3717 object = alloc_single_from_new_slab(s, slab, orig_size);
3718
3719 return object;
3720}
3721#endif /* CONFIG_SLUB_TINY */
3722
3723/*
3724 * If the object has been wiped upon free, make sure it's fully initialized by
3725 * zeroing out freelist pointer.
3726 */
3727static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
3728 void *obj)
3729{
3730 if (unlikely(slab_want_init_on_free(s)) && obj &&
3731 !freeptr_outside_object(s))
3732 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
3733 0, sizeof(void *));
3734}
3735
3736noinline int should_failslab(struct kmem_cache *s, gfp_t gfpflags)
3737{
3738 if (__should_failslab(s, gfpflags))
3739 return -ENOMEM;
3740 return 0;
3741}
3742ALLOW_ERROR_INJECTION(should_failslab, ERRNO);
3743
3744static __fastpath_inline
3745struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s,
3746 struct list_lru *lru,
3747 struct obj_cgroup **objcgp,
3748 size_t size, gfp_t flags)
3749{
3750 flags &= gfp_allowed_mask;
3751
3752 might_alloc(flags);
3753
3754 if (unlikely(should_failslab(s, flags)))
3755 return NULL;
3756
3757 if (unlikely(!memcg_slab_pre_alloc_hook(s, lru, objcgp, size, flags)))
3758 return NULL;
3759
3760 return s;
3761}
3762
3763static __fastpath_inline
3764void slab_post_alloc_hook(struct kmem_cache *s, struct obj_cgroup *objcg,
3765 gfp_t flags, size_t size, void **p, bool init,
3766 unsigned int orig_size)
3767{
3768 unsigned int zero_size = s->object_size;
3769 bool kasan_init = init;
3770 size_t i;
3771 gfp_t init_flags = flags & gfp_allowed_mask;
3772
3773 /*
3774 * For kmalloc object, the allocated memory size(object_size) is likely
3775 * larger than the requested size(orig_size). If redzone check is
3776 * enabled for the extra space, don't zero it, as it will be redzoned
3777 * soon. The redzone operation for this extra space could be seen as a
3778 * replacement of current poisoning under certain debug option, and
3779 * won't break other sanity checks.
3780 */
3781 if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
3782 (s->flags & SLAB_KMALLOC))
3783 zero_size = orig_size;
3784
3785 /*
3786 * When slab_debug is enabled, avoid memory initialization integrated
3787 * into KASAN and instead zero out the memory via the memset below with
3788 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
3789 * cause false-positive reports. This does not lead to a performance
3790 * penalty on production builds, as slab_debug is not intended to be
3791 * enabled there.
3792 */
3793 if (__slub_debug_enabled())
3794 kasan_init = false;
3795
3796 /*
3797 * As memory initialization might be integrated into KASAN,
3798 * kasan_slab_alloc and initialization memset must be
3799 * kept together to avoid discrepancies in behavior.
3800 *
3801 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
3802 */
3803 for (i = 0; i < size; i++) {
3804 p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
3805 if (p[i] && init && (!kasan_init ||
3806 !kasan_has_integrated_init()))
3807 memset(p[i], 0, zero_size);
3808 kmemleak_alloc_recursive(p[i], s->object_size, 1,
3809 s->flags, init_flags);
3810 kmsan_slab_alloc(s, p[i], init_flags);
3811 }
3812
3813 memcg_slab_post_alloc_hook(s, objcg, flags, size, p);
3814}
3815
3816/*
3817 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
3818 * have the fastpath folded into their functions. So no function call
3819 * overhead for requests that can be satisfied on the fastpath.
3820 *
3821 * The fastpath works by first checking if the lockless freelist can be used.
3822 * If not then __slab_alloc is called for slow processing.
3823 *
3824 * Otherwise we can simply pick the next object from the lockless free list.
3825 */
3826static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
3827 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3828{
3829 void *object;
3830 struct obj_cgroup *objcg = NULL;
3831 bool init = false;
3832
3833 s = slab_pre_alloc_hook(s, lru, &objcg, 1, gfpflags);
3834 if (unlikely(!s))
3835 return NULL;
3836
3837 object = kfence_alloc(s, orig_size, gfpflags);
3838 if (unlikely(object))
3839 goto out;
3840
3841 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
3842
3843 maybe_wipe_obj_freeptr(s, object);
3844 init = slab_want_init_on_alloc(gfpflags, s);
3845
3846out:
3847 /*
3848 * When init equals 'true', like for kzalloc() family, only
3849 * @orig_size bytes might be zeroed instead of s->object_size
3850 */
3851 slab_post_alloc_hook(s, objcg, gfpflags, 1, &object, init, orig_size);
3852
3853 return object;
3854}
3855
3856void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
3857{
3858 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
3859 s->object_size);
3860
3861 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3862
3863 return ret;
3864}
3865EXPORT_SYMBOL(kmem_cache_alloc);
3866
3867void *kmem_cache_alloc_lru(struct kmem_cache *s, struct list_lru *lru,
3868 gfp_t gfpflags)
3869{
3870 void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
3871 s->object_size);
3872
3873 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
3874
3875 return ret;
3876}
3877EXPORT_SYMBOL(kmem_cache_alloc_lru);
3878
3879/**
3880 * kmem_cache_alloc_node - Allocate an object on the specified node
3881 * @s: The cache to allocate from.
3882 * @gfpflags: See kmalloc().
3883 * @node: node number of the target node.
3884 *
3885 * Identical to kmem_cache_alloc but it will allocate memory on the given
3886 * node, which can improve the performance for cpu bound structures.
3887 *
3888 * Fallback to other node is possible if __GFP_THISNODE is not set.
3889 *
3890 * Return: pointer to the new object or %NULL in case of error
3891 */
3892void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
3893{
3894 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
3895
3896 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
3897
3898 return ret;
3899}
3900EXPORT_SYMBOL(kmem_cache_alloc_node);
3901
3902/*
3903 * To avoid unnecessary overhead, we pass through large allocation requests
3904 * directly to the page allocator. We use __GFP_COMP, because we will need to
3905 * know the allocation order to free the pages properly in kfree.
3906 */
3907static void *__kmalloc_large_node(size_t size, gfp_t flags, int node)
3908{
3909 struct folio *folio;
3910 void *ptr = NULL;
3911 unsigned int order = get_order(size);
3912
3913 if (unlikely(flags & GFP_SLAB_BUG_MASK))
3914 flags = kmalloc_fix_flags(flags);
3915
3916 flags |= __GFP_COMP;
3917 folio = (struct folio *)alloc_pages_node(node, flags, order);
3918 if (folio) {
3919 ptr = folio_address(folio);
3920 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
3921 PAGE_SIZE << order);
3922 }
3923
3924 ptr = kasan_kmalloc_large(ptr, size, flags);
3925 /* As ptr might get tagged, call kmemleak hook after KASAN. */
3926 kmemleak_alloc(ptr, size, 1, flags);
3927 kmsan_kmalloc_large(ptr, size, flags);
3928
3929 return ptr;
3930}
3931
3932void *kmalloc_large(size_t size, gfp_t flags)
3933{
3934 void *ret = __kmalloc_large_node(size, flags, NUMA_NO_NODE);
3935
3936 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3937 flags, NUMA_NO_NODE);
3938 return ret;
3939}
3940EXPORT_SYMBOL(kmalloc_large);
3941
3942void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3943{
3944 void *ret = __kmalloc_large_node(size, flags, node);
3945
3946 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
3947 flags, node);
3948 return ret;
3949}
3950EXPORT_SYMBOL(kmalloc_large_node);
3951
3952static __always_inline
3953void *__do_kmalloc_node(size_t size, gfp_t flags, int node,
3954 unsigned long caller)
3955{
3956 struct kmem_cache *s;
3957 void *ret;
3958
3959 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3960 ret = __kmalloc_large_node(size, flags, node);
3961 trace_kmalloc(caller, ret, size,
3962 PAGE_SIZE << get_order(size), flags, node);
3963 return ret;
3964 }
3965
3966 if (unlikely(!size))
3967 return ZERO_SIZE_PTR;
3968
3969 s = kmalloc_slab(size, flags, caller);
3970
3971 ret = slab_alloc_node(s, NULL, flags, node, caller, size);
3972 ret = kasan_kmalloc(s, ret, size, flags);
3973 trace_kmalloc(caller, ret, size, s->size, flags, node);
3974 return ret;
3975}
3976
3977void *__kmalloc_node(size_t size, gfp_t flags, int node)
3978{
3979 return __do_kmalloc_node(size, flags, node, _RET_IP_);
3980}
3981EXPORT_SYMBOL(__kmalloc_node);
3982
3983void *__kmalloc(size_t size, gfp_t flags)
3984{
3985 return __do_kmalloc_node(size, flags, NUMA_NO_NODE, _RET_IP_);
3986}
3987EXPORT_SYMBOL(__kmalloc);
3988
3989void *__kmalloc_node_track_caller(size_t size, gfp_t flags,
3990 int node, unsigned long caller)
3991{
3992 return __do_kmalloc_node(size, flags, node, caller);
3993}
3994EXPORT_SYMBOL(__kmalloc_node_track_caller);
3995
3996void *kmalloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
3997{
3998 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
3999 _RET_IP_, size);
4000
4001 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
4002
4003 ret = kasan_kmalloc(s, ret, size, gfpflags);
4004 return ret;
4005}
4006EXPORT_SYMBOL(kmalloc_trace);
4007
4008void *kmalloc_node_trace(struct kmem_cache *s, gfp_t gfpflags,
4009 int node, size_t size)
4010{
4011 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
4012
4013 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
4014
4015 ret = kasan_kmalloc(s, ret, size, gfpflags);
4016 return ret;
4017}
4018EXPORT_SYMBOL(kmalloc_node_trace);
4019
4020static noinline void free_to_partial_list(
4021 struct kmem_cache *s, struct slab *slab,
4022 void *head, void *tail, int bulk_cnt,
4023 unsigned long addr)
4024{
4025 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
4026 struct slab *slab_free = NULL;
4027 int cnt = bulk_cnt;
4028 unsigned long flags;
4029 depot_stack_handle_t handle = 0;
4030
4031 if (s->flags & SLAB_STORE_USER)
4032 handle = set_track_prepare();
4033
4034 spin_lock_irqsave(&n->list_lock, flags);
4035
4036 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
4037 void *prior = slab->freelist;
4038
4039 /* Perform the actual freeing while we still hold the locks */
4040 slab->inuse -= cnt;
4041 set_freepointer(s, tail, prior);
4042 slab->freelist = head;
4043
4044 /*
4045 * If the slab is empty, and node's partial list is full,
4046 * it should be discarded anyway no matter it's on full or
4047 * partial list.
4048 */
4049 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
4050 slab_free = slab;
4051
4052 if (!prior) {
4053 /* was on full list */
4054 remove_full(s, n, slab);
4055 if (!slab_free) {
4056 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4057 stat(s, FREE_ADD_PARTIAL);
4058 }
4059 } else if (slab_free) {
4060 remove_partial(n, slab);
4061 stat(s, FREE_REMOVE_PARTIAL);
4062 }
4063 }
4064
4065 if (slab_free) {
4066 /*
4067 * Update the counters while still holding n->list_lock to
4068 * prevent spurious validation warnings
4069 */
4070 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
4071 }
4072
4073 spin_unlock_irqrestore(&n->list_lock, flags);
4074
4075 if (slab_free) {
4076 stat(s, FREE_SLAB);
4077 free_slab(s, slab_free);
4078 }
4079}
4080
4081/*
4082 * Slow path handling. This may still be called frequently since objects
4083 * have a longer lifetime than the cpu slabs in most processing loads.
4084 *
4085 * So we still attempt to reduce cache line usage. Just take the slab
4086 * lock and free the item. If there is no additional partial slab
4087 * handling required then we can return immediately.
4088 */
4089static void __slab_free(struct kmem_cache *s, struct slab *slab,
4090 void *head, void *tail, int cnt,
4091 unsigned long addr)
4092
4093{
4094 void *prior;
4095 int was_frozen;
4096 struct slab new;
4097 unsigned long counters;
4098 struct kmem_cache_node *n = NULL;
4099 unsigned long flags;
4100 bool on_node_partial;
4101
4102 stat(s, FREE_SLOWPATH);
4103
4104 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4105 free_to_partial_list(s, slab, head, tail, cnt, addr);
4106 return;
4107 }
4108
4109 do {
4110 if (unlikely(n)) {
4111 spin_unlock_irqrestore(&n->list_lock, flags);
4112 n = NULL;
4113 }
4114 prior = slab->freelist;
4115 counters = slab->counters;
4116 set_freepointer(s, tail, prior);
4117 new.counters = counters;
4118 was_frozen = new.frozen;
4119 new.inuse -= cnt;
4120 if ((!new.inuse || !prior) && !was_frozen) {
4121 /* Needs to be taken off a list */
4122 if (!kmem_cache_has_cpu_partial(s) || prior) {
4123
4124 n = get_node(s, slab_nid(slab));
4125 /*
4126 * Speculatively acquire the list_lock.
4127 * If the cmpxchg does not succeed then we may
4128 * drop the list_lock without any processing.
4129 *
4130 * Otherwise the list_lock will synchronize with
4131 * other processors updating the list of slabs.
4132 */
4133 spin_lock_irqsave(&n->list_lock, flags);
4134
4135 on_node_partial = slab_test_node_partial(slab);
4136 }
4137 }
4138
4139 } while (!slab_update_freelist(s, slab,
4140 prior, counters,
4141 head, new.counters,
4142 "__slab_free"));
4143
4144 if (likely(!n)) {
4145
4146 if (likely(was_frozen)) {
4147 /*
4148 * The list lock was not taken therefore no list
4149 * activity can be necessary.
4150 */
4151 stat(s, FREE_FROZEN);
4152 } else if (kmem_cache_has_cpu_partial(s) && !prior) {
4153 /*
4154 * If we started with a full slab then put it onto the
4155 * per cpu partial list.
4156 */
4157 put_cpu_partial(s, slab, 1);
4158 stat(s, CPU_PARTIAL_FREE);
4159 }
4160
4161 return;
4162 }
4163
4164 /*
4165 * This slab was partially empty but not on the per-node partial list,
4166 * in which case we shouldn't manipulate its list, just return.
4167 */
4168 if (prior && !on_node_partial) {
4169 spin_unlock_irqrestore(&n->list_lock, flags);
4170 return;
4171 }
4172
4173 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
4174 goto slab_empty;
4175
4176 /*
4177 * Objects left in the slab. If it was not on the partial list before
4178 * then add it.
4179 */
4180 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
4181 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4182 stat(s, FREE_ADD_PARTIAL);
4183 }
4184 spin_unlock_irqrestore(&n->list_lock, flags);
4185 return;
4186
4187slab_empty:
4188 if (prior) {
4189 /*
4190 * Slab on the partial list.
4191 */
4192 remove_partial(n, slab);
4193 stat(s, FREE_REMOVE_PARTIAL);
4194 }
4195
4196 spin_unlock_irqrestore(&n->list_lock, flags);
4197 stat(s, FREE_SLAB);
4198 discard_slab(s, slab);
4199}
4200
4201#ifndef CONFIG_SLUB_TINY
4202/*
4203 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
4204 * can perform fastpath freeing without additional function calls.
4205 *
4206 * The fastpath is only possible if we are freeing to the current cpu slab
4207 * of this processor. This typically the case if we have just allocated
4208 * the item before.
4209 *
4210 * If fastpath is not possible then fall back to __slab_free where we deal
4211 * with all sorts of special processing.
4212 *
4213 * Bulk free of a freelist with several objects (all pointing to the
4214 * same slab) possible by specifying head and tail ptr, plus objects
4215 * count (cnt). Bulk free indicated by tail pointer being set.
4216 */
4217static __always_inline void do_slab_free(struct kmem_cache *s,
4218 struct slab *slab, void *head, void *tail,
4219 int cnt, unsigned long addr)
4220{
4221 struct kmem_cache_cpu *c;
4222 unsigned long tid;
4223 void **freelist;
4224
4225redo:
4226 /*
4227 * Determine the currently cpus per cpu slab.
4228 * The cpu may change afterward. However that does not matter since
4229 * data is retrieved via this pointer. If we are on the same cpu
4230 * during the cmpxchg then the free will succeed.
4231 */
4232 c = raw_cpu_ptr(s->cpu_slab);
4233 tid = READ_ONCE(c->tid);
4234
4235 /* Same with comment on barrier() in slab_alloc_node() */
4236 barrier();
4237
4238 if (unlikely(slab != c->slab)) {
4239 __slab_free(s, slab, head, tail, cnt, addr);
4240 return;
4241 }
4242
4243 if (USE_LOCKLESS_FAST_PATH()) {
4244 freelist = READ_ONCE(c->freelist);
4245
4246 set_freepointer(s, tail, freelist);
4247
4248 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
4249 note_cmpxchg_failure("slab_free", s, tid);
4250 goto redo;
4251 }
4252 } else {
4253 /* Update the free list under the local lock */
4254 local_lock(&s->cpu_slab->lock);
4255 c = this_cpu_ptr(s->cpu_slab);
4256 if (unlikely(slab != c->slab)) {
4257 local_unlock(&s->cpu_slab->lock);
4258 goto redo;
4259 }
4260 tid = c->tid;
4261 freelist = c->freelist;
4262
4263 set_freepointer(s, tail, freelist);
4264 c->freelist = head;
4265 c->tid = next_tid(tid);
4266
4267 local_unlock(&s->cpu_slab->lock);
4268 }
4269 stat_add(s, FREE_FASTPATH, cnt);
4270}
4271#else /* CONFIG_SLUB_TINY */
4272static void do_slab_free(struct kmem_cache *s,
4273 struct slab *slab, void *head, void *tail,
4274 int cnt, unsigned long addr)
4275{
4276 __slab_free(s, slab, head, tail, cnt, addr);
4277}
4278#endif /* CONFIG_SLUB_TINY */
4279
4280static __fastpath_inline
4281void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
4282 unsigned long addr)
4283{
4284 memcg_slab_free_hook(s, slab, &object, 1);
4285
4286 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s))))
4287 do_slab_free(s, slab, object, object, 1, addr);
4288}
4289
4290static __fastpath_inline
4291void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
4292 void *tail, void **p, int cnt, unsigned long addr)
4293{
4294 memcg_slab_free_hook(s, slab, p, cnt);
4295 /*
4296 * With KASAN enabled slab_free_freelist_hook modifies the freelist
4297 * to remove objects, whose reuse must be delayed.
4298 */
4299 if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
4300 do_slab_free(s, slab, head, tail, cnt, addr);
4301}
4302
4303#ifdef CONFIG_KASAN_GENERIC
4304void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
4305{
4306 do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
4307}
4308#endif
4309
4310static inline struct kmem_cache *virt_to_cache(const void *obj)
4311{
4312 struct slab *slab;
4313
4314 slab = virt_to_slab(obj);
4315 if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
4316 return NULL;
4317 return slab->slab_cache;
4318}
4319
4320static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
4321{
4322 struct kmem_cache *cachep;
4323
4324 if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
4325 !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
4326 return s;
4327
4328 cachep = virt_to_cache(x);
4329 if (WARN(cachep && cachep != s,
4330 "%s: Wrong slab cache. %s but object is from %s\n",
4331 __func__, s->name, cachep->name))
4332 print_tracking(cachep, x);
4333 return cachep;
4334}
4335
4336/**
4337 * kmem_cache_free - Deallocate an object
4338 * @s: The cache the allocation was from.
4339 * @x: The previously allocated object.
4340 *
4341 * Free an object which was previously allocated from this
4342 * cache.
4343 */
4344void kmem_cache_free(struct kmem_cache *s, void *x)
4345{
4346 s = cache_from_obj(s, x);
4347 if (!s)
4348 return;
4349 trace_kmem_cache_free(_RET_IP_, x, s);
4350 slab_free(s, virt_to_slab(x), x, _RET_IP_);
4351}
4352EXPORT_SYMBOL(kmem_cache_free);
4353
4354static void free_large_kmalloc(struct folio *folio, void *object)
4355{
4356 unsigned int order = folio_order(folio);
4357
4358 if (WARN_ON_ONCE(order == 0))
4359 pr_warn_once("object pointer: 0x%p\n", object);
4360
4361 kmemleak_free(object);
4362 kasan_kfree_large(object);
4363 kmsan_kfree_large(object);
4364
4365 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4366 -(PAGE_SIZE << order));
4367 folio_put(folio);
4368}
4369
4370/**
4371 * kfree - free previously allocated memory
4372 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
4373 *
4374 * If @object is NULL, no operation is performed.
4375 */
4376void kfree(const void *object)
4377{
4378 struct folio *folio;
4379 struct slab *slab;
4380 struct kmem_cache *s;
4381 void *x = (void *)object;
4382
4383 trace_kfree(_RET_IP_, object);
4384
4385 if (unlikely(ZERO_OR_NULL_PTR(object)))
4386 return;
4387
4388 folio = virt_to_folio(object);
4389 if (unlikely(!folio_test_slab(folio))) {
4390 free_large_kmalloc(folio, (void *)object);
4391 return;
4392 }
4393
4394 slab = folio_slab(folio);
4395 s = slab->slab_cache;
4396 slab_free(s, slab, x, _RET_IP_);
4397}
4398EXPORT_SYMBOL(kfree);
4399
4400struct detached_freelist {
4401 struct slab *slab;
4402 void *tail;
4403 void *freelist;
4404 int cnt;
4405 struct kmem_cache *s;
4406};
4407
4408/*
4409 * This function progressively scans the array with free objects (with
4410 * a limited look ahead) and extract objects belonging to the same
4411 * slab. It builds a detached freelist directly within the given
4412 * slab/objects. This can happen without any need for
4413 * synchronization, because the objects are owned by running process.
4414 * The freelist is build up as a single linked list in the objects.
4415 * The idea is, that this detached freelist can then be bulk
4416 * transferred to the real freelist(s), but only requiring a single
4417 * synchronization primitive. Look ahead in the array is limited due
4418 * to performance reasons.
4419 */
4420static inline
4421int build_detached_freelist(struct kmem_cache *s, size_t size,
4422 void **p, struct detached_freelist *df)
4423{
4424 int lookahead = 3;
4425 void *object;
4426 struct folio *folio;
4427 size_t same;
4428
4429 object = p[--size];
4430 folio = virt_to_folio(object);
4431 if (!s) {
4432 /* Handle kalloc'ed objects */
4433 if (unlikely(!folio_test_slab(folio))) {
4434 free_large_kmalloc(folio, object);
4435 df->slab = NULL;
4436 return size;
4437 }
4438 /* Derive kmem_cache from object */
4439 df->slab = folio_slab(folio);
4440 df->s = df->slab->slab_cache;
4441 } else {
4442 df->slab = folio_slab(folio);
4443 df->s = cache_from_obj(s, object); /* Support for memcg */
4444 }
4445
4446 /* Start new detached freelist */
4447 df->tail = object;
4448 df->freelist = object;
4449 df->cnt = 1;
4450
4451 if (is_kfence_address(object))
4452 return size;
4453
4454 set_freepointer(df->s, object, NULL);
4455
4456 same = size;
4457 while (size) {
4458 object = p[--size];
4459 /* df->slab is always set at this point */
4460 if (df->slab == virt_to_slab(object)) {
4461 /* Opportunity build freelist */
4462 set_freepointer(df->s, object, df->freelist);
4463 df->freelist = object;
4464 df->cnt++;
4465 same--;
4466 if (size != same)
4467 swap(p[size], p[same]);
4468 continue;
4469 }
4470
4471 /* Limit look ahead search */
4472 if (!--lookahead)
4473 break;
4474 }
4475
4476 return same;
4477}
4478
4479/*
4480 * Internal bulk free of objects that were not initialised by the post alloc
4481 * hooks and thus should not be processed by the free hooks
4482 */
4483static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4484{
4485 if (!size)
4486 return;
4487
4488 do {
4489 struct detached_freelist df;
4490
4491 size = build_detached_freelist(s, size, p, &df);
4492 if (!df.slab)
4493 continue;
4494
4495 do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
4496 _RET_IP_);
4497 } while (likely(size));
4498}
4499
4500/* Note that interrupts must be enabled when calling this function. */
4501void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4502{
4503 if (!size)
4504 return;
4505
4506 do {
4507 struct detached_freelist df;
4508
4509 size = build_detached_freelist(s, size, p, &df);
4510 if (!df.slab)
4511 continue;
4512
4513 slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
4514 df.cnt, _RET_IP_);
4515 } while (likely(size));
4516}
4517EXPORT_SYMBOL(kmem_cache_free_bulk);
4518
4519#ifndef CONFIG_SLUB_TINY
4520static inline
4521int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4522 void **p)
4523{
4524 struct kmem_cache_cpu *c;
4525 unsigned long irqflags;
4526 int i;
4527
4528 /*
4529 * Drain objects in the per cpu slab, while disabling local
4530 * IRQs, which protects against PREEMPT and interrupts
4531 * handlers invoking normal fastpath.
4532 */
4533 c = slub_get_cpu_ptr(s->cpu_slab);
4534 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4535
4536 for (i = 0; i < size; i++) {
4537 void *object = kfence_alloc(s, s->object_size, flags);
4538
4539 if (unlikely(object)) {
4540 p[i] = object;
4541 continue;
4542 }
4543
4544 object = c->freelist;
4545 if (unlikely(!object)) {
4546 /*
4547 * We may have removed an object from c->freelist using
4548 * the fastpath in the previous iteration; in that case,
4549 * c->tid has not been bumped yet.
4550 * Since ___slab_alloc() may reenable interrupts while
4551 * allocating memory, we should bump c->tid now.
4552 */
4553 c->tid = next_tid(c->tid);
4554
4555 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4556
4557 /*
4558 * Invoking slow path likely have side-effect
4559 * of re-populating per CPU c->freelist
4560 */
4561 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
4562 _RET_IP_, c, s->object_size);
4563 if (unlikely(!p[i]))
4564 goto error;
4565
4566 c = this_cpu_ptr(s->cpu_slab);
4567 maybe_wipe_obj_freeptr(s, p[i]);
4568
4569 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
4570
4571 continue; /* goto for-loop */
4572 }
4573 c->freelist = get_freepointer(s, object);
4574 p[i] = object;
4575 maybe_wipe_obj_freeptr(s, p[i]);
4576 stat(s, ALLOC_FASTPATH);
4577 }
4578 c->tid = next_tid(c->tid);
4579 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
4580 slub_put_cpu_ptr(s->cpu_slab);
4581
4582 return i;
4583
4584error:
4585 slub_put_cpu_ptr(s->cpu_slab);
4586 __kmem_cache_free_bulk(s, i, p);
4587 return 0;
4588
4589}
4590#else /* CONFIG_SLUB_TINY */
4591static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
4592 size_t size, void **p)
4593{
4594 int i;
4595
4596 for (i = 0; i < size; i++) {
4597 void *object = kfence_alloc(s, s->object_size, flags);
4598
4599 if (unlikely(object)) {
4600 p[i] = object;
4601 continue;
4602 }
4603
4604 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
4605 _RET_IP_, s->object_size);
4606 if (unlikely(!p[i]))
4607 goto error;
4608
4609 maybe_wipe_obj_freeptr(s, p[i]);
4610 }
4611
4612 return i;
4613
4614error:
4615 __kmem_cache_free_bulk(s, i, p);
4616 return 0;
4617}
4618#endif /* CONFIG_SLUB_TINY */
4619
4620/* Note that interrupts must be enabled when calling this function. */
4621int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
4622 void **p)
4623{
4624 int i;
4625 struct obj_cgroup *objcg = NULL;
4626
4627 if (!size)
4628 return 0;
4629
4630 /* memcg and kmem_cache debug support */
4631 s = slab_pre_alloc_hook(s, NULL, &objcg, size, flags);
4632 if (unlikely(!s))
4633 return 0;
4634
4635 i = __kmem_cache_alloc_bulk(s, flags, size, p);
4636
4637 /*
4638 * memcg and kmem_cache debug support and memory initialization.
4639 * Done outside of the IRQ disabled fastpath loop.
4640 */
4641 if (likely(i != 0)) {
4642 slab_post_alloc_hook(s, objcg, flags, size, p,
4643 slab_want_init_on_alloc(flags, s), s->object_size);
4644 } else {
4645 memcg_slab_alloc_error_hook(s, size, objcg);
4646 }
4647
4648 return i;
4649}
4650EXPORT_SYMBOL(kmem_cache_alloc_bulk);
4651
4652
4653/*
4654 * Object placement in a slab is made very easy because we always start at
4655 * offset 0. If we tune the size of the object to the alignment then we can
4656 * get the required alignment by putting one properly sized object after
4657 * another.
4658 *
4659 * Notice that the allocation order determines the sizes of the per cpu
4660 * caches. Each processor has always one slab available for allocations.
4661 * Increasing the allocation order reduces the number of times that slabs
4662 * must be moved on and off the partial lists and is therefore a factor in
4663 * locking overhead.
4664 */
4665
4666/*
4667 * Minimum / Maximum order of slab pages. This influences locking overhead
4668 * and slab fragmentation. A higher order reduces the number of partial slabs
4669 * and increases the number of allocations possible without having to
4670 * take the list_lock.
4671 */
4672static unsigned int slub_min_order;
4673static unsigned int slub_max_order =
4674 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
4675static unsigned int slub_min_objects;
4676
4677/*
4678 * Calculate the order of allocation given an slab object size.
4679 *
4680 * The order of allocation has significant impact on performance and other
4681 * system components. Generally order 0 allocations should be preferred since
4682 * order 0 does not cause fragmentation in the page allocator. Larger objects
4683 * be problematic to put into order 0 slabs because there may be too much
4684 * unused space left. We go to a higher order if more than 1/16th of the slab
4685 * would be wasted.
4686 *
4687 * In order to reach satisfactory performance we must ensure that a minimum
4688 * number of objects is in one slab. Otherwise we may generate too much
4689 * activity on the partial lists which requires taking the list_lock. This is
4690 * less a concern for large slabs though which are rarely used.
4691 *
4692 * slab_max_order specifies the order where we begin to stop considering the
4693 * number of objects in a slab as critical. If we reach slab_max_order then
4694 * we try to keep the page order as low as possible. So we accept more waste
4695 * of space in favor of a small page order.
4696 *
4697 * Higher order allocations also allow the placement of more objects in a
4698 * slab and thereby reduce object handling overhead. If the user has
4699 * requested a higher minimum order then we start with that one instead of
4700 * the smallest order which will fit the object.
4701 */
4702static inline unsigned int calc_slab_order(unsigned int size,
4703 unsigned int min_order, unsigned int max_order,
4704 unsigned int fract_leftover)
4705{
4706 unsigned int order;
4707
4708 for (order = min_order; order <= max_order; order++) {
4709
4710 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
4711 unsigned int rem;
4712
4713 rem = slab_size % size;
4714
4715 if (rem <= slab_size / fract_leftover)
4716 break;
4717 }
4718
4719 return order;
4720}
4721
4722static inline int calculate_order(unsigned int size)
4723{
4724 unsigned int order;
4725 unsigned int min_objects;
4726 unsigned int max_objects;
4727 unsigned int min_order;
4728
4729 min_objects = slub_min_objects;
4730 if (!min_objects) {
4731 /*
4732 * Some architectures will only update present cpus when
4733 * onlining them, so don't trust the number if it's just 1. But
4734 * we also don't want to use nr_cpu_ids always, as on some other
4735 * architectures, there can be many possible cpus, but never
4736 * onlined. Here we compromise between trying to avoid too high
4737 * order on systems that appear larger than they are, and too
4738 * low order on systems that appear smaller than they are.
4739 */
4740 unsigned int nr_cpus = num_present_cpus();
4741 if (nr_cpus <= 1)
4742 nr_cpus = nr_cpu_ids;
4743 min_objects = 4 * (fls(nr_cpus) + 1);
4744 }
4745 /* min_objects can't be 0 because get_order(0) is undefined */
4746 max_objects = max(order_objects(slub_max_order, size), 1U);
4747 min_objects = min(min_objects, max_objects);
4748
4749 min_order = max_t(unsigned int, slub_min_order,
4750 get_order(min_objects * size));
4751 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
4752 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
4753
4754 /*
4755 * Attempt to find best configuration for a slab. This works by first
4756 * attempting to generate a layout with the best possible configuration
4757 * and backing off gradually.
4758 *
4759 * We start with accepting at most 1/16 waste and try to find the
4760 * smallest order from min_objects-derived/slab_min_order up to
4761 * slab_max_order that will satisfy the constraint. Note that increasing
4762 * the order can only result in same or less fractional waste, not more.
4763 *
4764 * If that fails, we increase the acceptable fraction of waste and try
4765 * again. The last iteration with fraction of 1/2 would effectively
4766 * accept any waste and give us the order determined by min_objects, as
4767 * long as at least single object fits within slab_max_order.
4768 */
4769 for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
4770 order = calc_slab_order(size, min_order, slub_max_order,
4771 fraction);
4772 if (order <= slub_max_order)
4773 return order;
4774 }
4775
4776 /*
4777 * Doh this slab cannot be placed using slab_max_order.
4778 */
4779 order = get_order(size);
4780 if (order <= MAX_PAGE_ORDER)
4781 return order;
4782 return -ENOSYS;
4783}
4784
4785static void
4786init_kmem_cache_node(struct kmem_cache_node *n)
4787{
4788 n->nr_partial = 0;
4789 spin_lock_init(&n->list_lock);
4790 INIT_LIST_HEAD(&n->partial);
4791#ifdef CONFIG_SLUB_DEBUG
4792 atomic_long_set(&n->nr_slabs, 0);
4793 atomic_long_set(&n->total_objects, 0);
4794 INIT_LIST_HEAD(&n->full);
4795#endif
4796}
4797
4798#ifndef CONFIG_SLUB_TINY
4799static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4800{
4801 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
4802 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
4803 sizeof(struct kmem_cache_cpu));
4804
4805 /*
4806 * Must align to double word boundary for the double cmpxchg
4807 * instructions to work; see __pcpu_double_call_return_bool().
4808 */
4809 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
4810 2 * sizeof(void *));
4811
4812 if (!s->cpu_slab)
4813 return 0;
4814
4815 init_kmem_cache_cpus(s);
4816
4817 return 1;
4818}
4819#else
4820static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
4821{
4822 return 1;
4823}
4824#endif /* CONFIG_SLUB_TINY */
4825
4826static struct kmem_cache *kmem_cache_node;
4827
4828/*
4829 * No kmalloc_node yet so do it by hand. We know that this is the first
4830 * slab on the node for this slabcache. There are no concurrent accesses
4831 * possible.
4832 *
4833 * Note that this function only works on the kmem_cache_node
4834 * when allocating for the kmem_cache_node. This is used for bootstrapping
4835 * memory on a fresh node that has no slab structures yet.
4836 */
4837static void early_kmem_cache_node_alloc(int node)
4838{
4839 struct slab *slab;
4840 struct kmem_cache_node *n;
4841
4842 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
4843
4844 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
4845
4846 BUG_ON(!slab);
4847 if (slab_nid(slab) != node) {
4848 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
4849 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
4850 }
4851
4852 n = slab->freelist;
4853 BUG_ON(!n);
4854#ifdef CONFIG_SLUB_DEBUG
4855 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
4856 init_tracking(kmem_cache_node, n);
4857#endif
4858 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
4859 slab->freelist = get_freepointer(kmem_cache_node, n);
4860 slab->inuse = 1;
4861 kmem_cache_node->node[node] = n;
4862 init_kmem_cache_node(n);
4863 inc_slabs_node(kmem_cache_node, node, slab->objects);
4864
4865 /*
4866 * No locks need to be taken here as it has just been
4867 * initialized and there is no concurrent access.
4868 */
4869 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
4870}
4871
4872static void free_kmem_cache_nodes(struct kmem_cache *s)
4873{
4874 int node;
4875 struct kmem_cache_node *n;
4876
4877 for_each_kmem_cache_node(s, node, n) {
4878 s->node[node] = NULL;
4879 kmem_cache_free(kmem_cache_node, n);
4880 }
4881}
4882
4883void __kmem_cache_release(struct kmem_cache *s)
4884{
4885 cache_random_seq_destroy(s);
4886#ifndef CONFIG_SLUB_TINY
4887 free_percpu(s->cpu_slab);
4888#endif
4889 free_kmem_cache_nodes(s);
4890}
4891
4892static int init_kmem_cache_nodes(struct kmem_cache *s)
4893{
4894 int node;
4895
4896 for_each_node_mask(node, slab_nodes) {
4897 struct kmem_cache_node *n;
4898
4899 if (slab_state == DOWN) {
4900 early_kmem_cache_node_alloc(node);
4901 continue;
4902 }
4903 n = kmem_cache_alloc_node(kmem_cache_node,
4904 GFP_KERNEL, node);
4905
4906 if (!n) {
4907 free_kmem_cache_nodes(s);
4908 return 0;
4909 }
4910
4911 init_kmem_cache_node(n);
4912 s->node[node] = n;
4913 }
4914 return 1;
4915}
4916
4917static void set_cpu_partial(struct kmem_cache *s)
4918{
4919#ifdef CONFIG_SLUB_CPU_PARTIAL
4920 unsigned int nr_objects;
4921
4922 /*
4923 * cpu_partial determined the maximum number of objects kept in the
4924 * per cpu partial lists of a processor.
4925 *
4926 * Per cpu partial lists mainly contain slabs that just have one
4927 * object freed. If they are used for allocation then they can be
4928 * filled up again with minimal effort. The slab will never hit the
4929 * per node partial lists and therefore no locking will be required.
4930 *
4931 * For backwards compatibility reasons, this is determined as number
4932 * of objects, even though we now limit maximum number of pages, see
4933 * slub_set_cpu_partial()
4934 */
4935 if (!kmem_cache_has_cpu_partial(s))
4936 nr_objects = 0;
4937 else if (s->size >= PAGE_SIZE)
4938 nr_objects = 6;
4939 else if (s->size >= 1024)
4940 nr_objects = 24;
4941 else if (s->size >= 256)
4942 nr_objects = 52;
4943 else
4944 nr_objects = 120;
4945
4946 slub_set_cpu_partial(s, nr_objects);
4947#endif
4948}
4949
4950/*
4951 * calculate_sizes() determines the order and the distribution of data within
4952 * a slab object.
4953 */
4954static int calculate_sizes(struct kmem_cache *s)
4955{
4956 slab_flags_t flags = s->flags;
4957 unsigned int size = s->object_size;
4958 unsigned int order;
4959
4960 /*
4961 * Round up object size to the next word boundary. We can only
4962 * place the free pointer at word boundaries and this determines
4963 * the possible location of the free pointer.
4964 */
4965 size = ALIGN(size, sizeof(void *));
4966
4967#ifdef CONFIG_SLUB_DEBUG
4968 /*
4969 * Determine if we can poison the object itself. If the user of
4970 * the slab may touch the object after free or before allocation
4971 * then we should never poison the object itself.
4972 */
4973 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
4974 !s->ctor)
4975 s->flags |= __OBJECT_POISON;
4976 else
4977 s->flags &= ~__OBJECT_POISON;
4978
4979
4980 /*
4981 * If we are Redzoning then check if there is some space between the
4982 * end of the object and the free pointer. If not then add an
4983 * additional word to have some bytes to store Redzone information.
4984 */
4985 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
4986 size += sizeof(void *);
4987#endif
4988
4989 /*
4990 * With that we have determined the number of bytes in actual use
4991 * by the object and redzoning.
4992 */
4993 s->inuse = size;
4994
4995 if (slub_debug_orig_size(s) ||
4996 (flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
4997 ((flags & SLAB_RED_ZONE) && s->object_size < sizeof(void *)) ||
4998 s->ctor) {
4999 /*
5000 * Relocate free pointer after the object if it is not
5001 * permitted to overwrite the first word of the object on
5002 * kmem_cache_free.
5003 *
5004 * This is the case if we do RCU, have a constructor or
5005 * destructor, are poisoning the objects, or are
5006 * redzoning an object smaller than sizeof(void *).
5007 *
5008 * The assumption that s->offset >= s->inuse means free
5009 * pointer is outside of the object is used in the
5010 * freeptr_outside_object() function. If that is no
5011 * longer true, the function needs to be modified.
5012 */
5013 s->offset = size;
5014 size += sizeof(void *);
5015 } else {
5016 /*
5017 * Store freelist pointer near middle of object to keep
5018 * it away from the edges of the object to avoid small
5019 * sized over/underflows from neighboring allocations.
5020 */
5021 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
5022 }
5023
5024#ifdef CONFIG_SLUB_DEBUG
5025 if (flags & SLAB_STORE_USER) {
5026 /*
5027 * Need to store information about allocs and frees after
5028 * the object.
5029 */
5030 size += 2 * sizeof(struct track);
5031
5032 /* Save the original kmalloc request size */
5033 if (flags & SLAB_KMALLOC)
5034 size += sizeof(unsigned int);
5035 }
5036#endif
5037
5038 kasan_cache_create(s, &size, &s->flags);
5039#ifdef CONFIG_SLUB_DEBUG
5040 if (flags & SLAB_RED_ZONE) {
5041 /*
5042 * Add some empty padding so that we can catch
5043 * overwrites from earlier objects rather than let
5044 * tracking information or the free pointer be
5045 * corrupted if a user writes before the start
5046 * of the object.
5047 */
5048 size += sizeof(void *);
5049
5050 s->red_left_pad = sizeof(void *);
5051 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
5052 size += s->red_left_pad;
5053 }
5054#endif
5055
5056 /*
5057 * SLUB stores one object immediately after another beginning from
5058 * offset 0. In order to align the objects we have to simply size
5059 * each object to conform to the alignment.
5060 */
5061 size = ALIGN(size, s->align);
5062 s->size = size;
5063 s->reciprocal_size = reciprocal_value(size);
5064 order = calculate_order(size);
5065
5066 if ((int)order < 0)
5067 return 0;
5068
5069 s->allocflags = 0;
5070 if (order)
5071 s->allocflags |= __GFP_COMP;
5072
5073 if (s->flags & SLAB_CACHE_DMA)
5074 s->allocflags |= GFP_DMA;
5075
5076 if (s->flags & SLAB_CACHE_DMA32)
5077 s->allocflags |= GFP_DMA32;
5078
5079 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5080 s->allocflags |= __GFP_RECLAIMABLE;
5081
5082 /*
5083 * Determine the number of objects per slab
5084 */
5085 s->oo = oo_make(order, size);
5086 s->min = oo_make(get_order(size), size);
5087
5088 return !!oo_objects(s->oo);
5089}
5090
5091static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
5092{
5093 s->flags = kmem_cache_flags(flags, s->name);
5094#ifdef CONFIG_SLAB_FREELIST_HARDENED
5095 s->random = get_random_long();
5096#endif
5097
5098 if (!calculate_sizes(s))
5099 goto error;
5100 if (disable_higher_order_debug) {
5101 /*
5102 * Disable debugging flags that store metadata if the min slab
5103 * order increased.
5104 */
5105 if (get_order(s->size) > get_order(s->object_size)) {
5106 s->flags &= ~DEBUG_METADATA_FLAGS;
5107 s->offset = 0;
5108 if (!calculate_sizes(s))
5109 goto error;
5110 }
5111 }
5112
5113#ifdef system_has_freelist_aba
5114 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
5115 /* Enable fast mode */
5116 s->flags |= __CMPXCHG_DOUBLE;
5117 }
5118#endif
5119
5120 /*
5121 * The larger the object size is, the more slabs we want on the partial
5122 * list to avoid pounding the page allocator excessively.
5123 */
5124 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
5125 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
5126
5127 set_cpu_partial(s);
5128
5129#ifdef CONFIG_NUMA
5130 s->remote_node_defrag_ratio = 1000;
5131#endif
5132
5133 /* Initialize the pre-computed randomized freelist if slab is up */
5134 if (slab_state >= UP) {
5135 if (init_cache_random_seq(s))
5136 goto error;
5137 }
5138
5139 if (!init_kmem_cache_nodes(s))
5140 goto error;
5141
5142 if (alloc_kmem_cache_cpus(s))
5143 return 0;
5144
5145error:
5146 __kmem_cache_release(s);
5147 return -EINVAL;
5148}
5149
5150static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
5151 const char *text)
5152{
5153#ifdef CONFIG_SLUB_DEBUG
5154 void *addr = slab_address(slab);
5155 void *p;
5156
5157 slab_err(s, slab, text, s->name);
5158
5159 spin_lock(&object_map_lock);
5160 __fill_map(object_map, s, slab);
5161
5162 for_each_object(p, s, addr, slab->objects) {
5163
5164 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
5165 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
5166 print_tracking(s, p);
5167 }
5168 }
5169 spin_unlock(&object_map_lock);
5170#endif
5171}
5172
5173/*
5174 * Attempt to free all partial slabs on a node.
5175 * This is called from __kmem_cache_shutdown(). We must take list_lock
5176 * because sysfs file might still access partial list after the shutdowning.
5177 */
5178static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
5179{
5180 LIST_HEAD(discard);
5181 struct slab *slab, *h;
5182
5183 BUG_ON(irqs_disabled());
5184 spin_lock_irq(&n->list_lock);
5185 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
5186 if (!slab->inuse) {
5187 remove_partial(n, slab);
5188 list_add(&slab->slab_list, &discard);
5189 } else {
5190 list_slab_objects(s, slab,
5191 "Objects remaining in %s on __kmem_cache_shutdown()");
5192 }
5193 }
5194 spin_unlock_irq(&n->list_lock);
5195
5196 list_for_each_entry_safe(slab, h, &discard, slab_list)
5197 discard_slab(s, slab);
5198}
5199
5200bool __kmem_cache_empty(struct kmem_cache *s)
5201{
5202 int node;
5203 struct kmem_cache_node *n;
5204
5205 for_each_kmem_cache_node(s, node, n)
5206 if (n->nr_partial || node_nr_slabs(n))
5207 return false;
5208 return true;
5209}
5210
5211/*
5212 * Release all resources used by a slab cache.
5213 */
5214int __kmem_cache_shutdown(struct kmem_cache *s)
5215{
5216 int node;
5217 struct kmem_cache_node *n;
5218
5219 flush_all_cpus_locked(s);
5220 /* Attempt to free all objects */
5221 for_each_kmem_cache_node(s, node, n) {
5222 free_partial(s, n);
5223 if (n->nr_partial || node_nr_slabs(n))
5224 return 1;
5225 }
5226 return 0;
5227}
5228
5229#ifdef CONFIG_PRINTK
5230void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
5231{
5232 void *base;
5233 int __maybe_unused i;
5234 unsigned int objnr;
5235 void *objp;
5236 void *objp0;
5237 struct kmem_cache *s = slab->slab_cache;
5238 struct track __maybe_unused *trackp;
5239
5240 kpp->kp_ptr = object;
5241 kpp->kp_slab = slab;
5242 kpp->kp_slab_cache = s;
5243 base = slab_address(slab);
5244 objp0 = kasan_reset_tag(object);
5245#ifdef CONFIG_SLUB_DEBUG
5246 objp = restore_red_left(s, objp0);
5247#else
5248 objp = objp0;
5249#endif
5250 objnr = obj_to_index(s, slab, objp);
5251 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
5252 objp = base + s->size * objnr;
5253 kpp->kp_objp = objp;
5254 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
5255 || (objp - base) % s->size) ||
5256 !(s->flags & SLAB_STORE_USER))
5257 return;
5258#ifdef CONFIG_SLUB_DEBUG
5259 objp = fixup_red_left(s, objp);
5260 trackp = get_track(s, objp, TRACK_ALLOC);
5261 kpp->kp_ret = (void *)trackp->addr;
5262#ifdef CONFIG_STACKDEPOT
5263 {
5264 depot_stack_handle_t handle;
5265 unsigned long *entries;
5266 unsigned int nr_entries;
5267
5268 handle = READ_ONCE(trackp->handle);
5269 if (handle) {
5270 nr_entries = stack_depot_fetch(handle, &entries);
5271 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5272 kpp->kp_stack[i] = (void *)entries[i];
5273 }
5274
5275 trackp = get_track(s, objp, TRACK_FREE);
5276 handle = READ_ONCE(trackp->handle);
5277 if (handle) {
5278 nr_entries = stack_depot_fetch(handle, &entries);
5279 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5280 kpp->kp_free_stack[i] = (void *)entries[i];
5281 }
5282 }
5283#endif
5284#endif
5285}
5286#endif
5287
5288/********************************************************************
5289 * Kmalloc subsystem
5290 *******************************************************************/
5291
5292static int __init setup_slub_min_order(char *str)
5293{
5294 get_option(&str, (int *)&slub_min_order);
5295
5296 if (slub_min_order > slub_max_order)
5297 slub_max_order = slub_min_order;
5298
5299 return 1;
5300}
5301
5302__setup("slab_min_order=", setup_slub_min_order);
5303__setup_param("slub_min_order=", slub_min_order, setup_slub_min_order, 0);
5304
5305
5306static int __init setup_slub_max_order(char *str)
5307{
5308 get_option(&str, (int *)&slub_max_order);
5309 slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
5310
5311 if (slub_min_order > slub_max_order)
5312 slub_min_order = slub_max_order;
5313
5314 return 1;
5315}
5316
5317__setup("slab_max_order=", setup_slub_max_order);
5318__setup_param("slub_max_order=", slub_max_order, setup_slub_max_order, 0);
5319
5320static int __init setup_slub_min_objects(char *str)
5321{
5322 get_option(&str, (int *)&slub_min_objects);
5323
5324 return 1;
5325}
5326
5327__setup("slab_min_objects=", setup_slub_min_objects);
5328__setup_param("slub_min_objects=", slub_min_objects, setup_slub_min_objects, 0);
5329
5330#ifdef CONFIG_HARDENED_USERCOPY
5331/*
5332 * Rejects incorrectly sized objects and objects that are to be copied
5333 * to/from userspace but do not fall entirely within the containing slab
5334 * cache's usercopy region.
5335 *
5336 * Returns NULL if check passes, otherwise const char * to name of cache
5337 * to indicate an error.
5338 */
5339void __check_heap_object(const void *ptr, unsigned long n,
5340 const struct slab *slab, bool to_user)
5341{
5342 struct kmem_cache *s;
5343 unsigned int offset;
5344 bool is_kfence = is_kfence_address(ptr);
5345
5346 ptr = kasan_reset_tag(ptr);
5347
5348 /* Find object and usable object size. */
5349 s = slab->slab_cache;
5350
5351 /* Reject impossible pointers. */
5352 if (ptr < slab_address(slab))
5353 usercopy_abort("SLUB object not in SLUB page?!", NULL,
5354 to_user, 0, n);
5355
5356 /* Find offset within object. */
5357 if (is_kfence)
5358 offset = ptr - kfence_object_start(ptr);
5359 else
5360 offset = (ptr - slab_address(slab)) % s->size;
5361
5362 /* Adjust for redzone and reject if within the redzone. */
5363 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
5364 if (offset < s->red_left_pad)
5365 usercopy_abort("SLUB object in left red zone",
5366 s->name, to_user, offset, n);
5367 offset -= s->red_left_pad;
5368 }
5369
5370 /* Allow address range falling entirely within usercopy region. */
5371 if (offset >= s->useroffset &&
5372 offset - s->useroffset <= s->usersize &&
5373 n <= s->useroffset - offset + s->usersize)
5374 return;
5375
5376 usercopy_abort("SLUB object", s->name, to_user, offset, n);
5377}
5378#endif /* CONFIG_HARDENED_USERCOPY */
5379
5380#define SHRINK_PROMOTE_MAX 32
5381
5382/*
5383 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
5384 * up most to the head of the partial lists. New allocations will then
5385 * fill those up and thus they can be removed from the partial lists.
5386 *
5387 * The slabs with the least items are placed last. This results in them
5388 * being allocated from last increasing the chance that the last objects
5389 * are freed in them.
5390 */
5391static int __kmem_cache_do_shrink(struct kmem_cache *s)
5392{
5393 int node;
5394 int i;
5395 struct kmem_cache_node *n;
5396 struct slab *slab;
5397 struct slab *t;
5398 struct list_head discard;
5399 struct list_head promote[SHRINK_PROMOTE_MAX];
5400 unsigned long flags;
5401 int ret = 0;
5402
5403 for_each_kmem_cache_node(s, node, n) {
5404 INIT_LIST_HEAD(&discard);
5405 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
5406 INIT_LIST_HEAD(promote + i);
5407
5408 spin_lock_irqsave(&n->list_lock, flags);
5409
5410 /*
5411 * Build lists of slabs to discard or promote.
5412 *
5413 * Note that concurrent frees may occur while we hold the
5414 * list_lock. slab->inuse here is the upper limit.
5415 */
5416 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
5417 int free = slab->objects - slab->inuse;
5418
5419 /* Do not reread slab->inuse */
5420 barrier();
5421
5422 /* We do not keep full slabs on the list */
5423 BUG_ON(free <= 0);
5424
5425 if (free == slab->objects) {
5426 list_move(&slab->slab_list, &discard);
5427 slab_clear_node_partial(slab);
5428 n->nr_partial--;
5429 dec_slabs_node(s, node, slab->objects);
5430 } else if (free <= SHRINK_PROMOTE_MAX)
5431 list_move(&slab->slab_list, promote + free - 1);
5432 }
5433
5434 /*
5435 * Promote the slabs filled up most to the head of the
5436 * partial list.
5437 */
5438 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
5439 list_splice(promote + i, &n->partial);
5440
5441 spin_unlock_irqrestore(&n->list_lock, flags);
5442
5443 /* Release empty slabs */
5444 list_for_each_entry_safe(slab, t, &discard, slab_list)
5445 free_slab(s, slab);
5446
5447 if (node_nr_slabs(n))
5448 ret = 1;
5449 }
5450
5451 return ret;
5452}
5453
5454int __kmem_cache_shrink(struct kmem_cache *s)
5455{
5456 flush_all(s);
5457 return __kmem_cache_do_shrink(s);
5458}
5459
5460static int slab_mem_going_offline_callback(void *arg)
5461{
5462 struct kmem_cache *s;
5463
5464 mutex_lock(&slab_mutex);
5465 list_for_each_entry(s, &slab_caches, list) {
5466 flush_all_cpus_locked(s);
5467 __kmem_cache_do_shrink(s);
5468 }
5469 mutex_unlock(&slab_mutex);
5470
5471 return 0;
5472}
5473
5474static void slab_mem_offline_callback(void *arg)
5475{
5476 struct memory_notify *marg = arg;
5477 int offline_node;
5478
5479 offline_node = marg->status_change_nid_normal;
5480
5481 /*
5482 * If the node still has available memory. we need kmem_cache_node
5483 * for it yet.
5484 */
5485 if (offline_node < 0)
5486 return;
5487
5488 mutex_lock(&slab_mutex);
5489 node_clear(offline_node, slab_nodes);
5490 /*
5491 * We no longer free kmem_cache_node structures here, as it would be
5492 * racy with all get_node() users, and infeasible to protect them with
5493 * slab_mutex.
5494 */
5495 mutex_unlock(&slab_mutex);
5496}
5497
5498static int slab_mem_going_online_callback(void *arg)
5499{
5500 struct kmem_cache_node *n;
5501 struct kmem_cache *s;
5502 struct memory_notify *marg = arg;
5503 int nid = marg->status_change_nid_normal;
5504 int ret = 0;
5505
5506 /*
5507 * If the node's memory is already available, then kmem_cache_node is
5508 * already created. Nothing to do.
5509 */
5510 if (nid < 0)
5511 return 0;
5512
5513 /*
5514 * We are bringing a node online. No memory is available yet. We must
5515 * allocate a kmem_cache_node structure in order to bring the node
5516 * online.
5517 */
5518 mutex_lock(&slab_mutex);
5519 list_for_each_entry(s, &slab_caches, list) {
5520 /*
5521 * The structure may already exist if the node was previously
5522 * onlined and offlined.
5523 */
5524 if (get_node(s, nid))
5525 continue;
5526 /*
5527 * XXX: kmem_cache_alloc_node will fallback to other nodes
5528 * since memory is not yet available from the node that
5529 * is brought up.
5530 */
5531 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
5532 if (!n) {
5533 ret = -ENOMEM;
5534 goto out;
5535 }
5536 init_kmem_cache_node(n);
5537 s->node[nid] = n;
5538 }
5539 /*
5540 * Any cache created after this point will also have kmem_cache_node
5541 * initialized for the new node.
5542 */
5543 node_set(nid, slab_nodes);
5544out:
5545 mutex_unlock(&slab_mutex);
5546 return ret;
5547}
5548
5549static int slab_memory_callback(struct notifier_block *self,
5550 unsigned long action, void *arg)
5551{
5552 int ret = 0;
5553
5554 switch (action) {
5555 case MEM_GOING_ONLINE:
5556 ret = slab_mem_going_online_callback(arg);
5557 break;
5558 case MEM_GOING_OFFLINE:
5559 ret = slab_mem_going_offline_callback(arg);
5560 break;
5561 case MEM_OFFLINE:
5562 case MEM_CANCEL_ONLINE:
5563 slab_mem_offline_callback(arg);
5564 break;
5565 case MEM_ONLINE:
5566 case MEM_CANCEL_OFFLINE:
5567 break;
5568 }
5569 if (ret)
5570 ret = notifier_from_errno(ret);
5571 else
5572 ret = NOTIFY_OK;
5573 return ret;
5574}
5575
5576/********************************************************************
5577 * Basic setup of slabs
5578 *******************************************************************/
5579
5580/*
5581 * Used for early kmem_cache structures that were allocated using
5582 * the page allocator. Allocate them properly then fix up the pointers
5583 * that may be pointing to the wrong kmem_cache structure.
5584 */
5585
5586static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
5587{
5588 int node;
5589 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
5590 struct kmem_cache_node *n;
5591
5592 memcpy(s, static_cache, kmem_cache->object_size);
5593
5594 /*
5595 * This runs very early, and only the boot processor is supposed to be
5596 * up. Even if it weren't true, IRQs are not up so we couldn't fire
5597 * IPIs around.
5598 */
5599 __flush_cpu_slab(s, smp_processor_id());
5600 for_each_kmem_cache_node(s, node, n) {
5601 struct slab *p;
5602
5603 list_for_each_entry(p, &n->partial, slab_list)
5604 p->slab_cache = s;
5605
5606#ifdef CONFIG_SLUB_DEBUG
5607 list_for_each_entry(p, &n->full, slab_list)
5608 p->slab_cache = s;
5609#endif
5610 }
5611 list_add(&s->list, &slab_caches);
5612 return s;
5613}
5614
5615void __init kmem_cache_init(void)
5616{
5617 static __initdata struct kmem_cache boot_kmem_cache,
5618 boot_kmem_cache_node;
5619 int node;
5620
5621 if (debug_guardpage_minorder())
5622 slub_max_order = 0;
5623
5624 /* Print slub debugging pointers without hashing */
5625 if (__slub_debug_enabled())
5626 no_hash_pointers_enable(NULL);
5627
5628 kmem_cache_node = &boot_kmem_cache_node;
5629 kmem_cache = &boot_kmem_cache;
5630
5631 /*
5632 * Initialize the nodemask for which we will allocate per node
5633 * structures. Here we don't need taking slab_mutex yet.
5634 */
5635 for_each_node_state(node, N_NORMAL_MEMORY)
5636 node_set(node, slab_nodes);
5637
5638 create_boot_cache(kmem_cache_node, "kmem_cache_node",
5639 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
5640
5641 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
5642
5643 /* Able to allocate the per node structures */
5644 slab_state = PARTIAL;
5645
5646 create_boot_cache(kmem_cache, "kmem_cache",
5647 offsetof(struct kmem_cache, node) +
5648 nr_node_ids * sizeof(struct kmem_cache_node *),
5649 SLAB_HWCACHE_ALIGN, 0, 0);
5650
5651 kmem_cache = bootstrap(&boot_kmem_cache);
5652 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
5653
5654 /* Now we can use the kmem_cache to allocate kmalloc slabs */
5655 setup_kmalloc_cache_index_table();
5656 create_kmalloc_caches();
5657
5658 /* Setup random freelists for each cache */
5659 init_freelist_randomization();
5660
5661 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
5662 slub_cpu_dead);
5663
5664 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
5665 cache_line_size(),
5666 slub_min_order, slub_max_order, slub_min_objects,
5667 nr_cpu_ids, nr_node_ids);
5668}
5669
5670void __init kmem_cache_init_late(void)
5671{
5672#ifndef CONFIG_SLUB_TINY
5673 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
5674 WARN_ON(!flushwq);
5675#endif
5676}
5677
5678struct kmem_cache *
5679__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
5680 slab_flags_t flags, void (*ctor)(void *))
5681{
5682 struct kmem_cache *s;
5683
5684 s = find_mergeable(size, align, flags, name, ctor);
5685 if (s) {
5686 if (sysfs_slab_alias(s, name))
5687 return NULL;
5688
5689 s->refcount++;
5690
5691 /*
5692 * Adjust the object sizes so that we clear
5693 * the complete object on kzalloc.
5694 */
5695 s->object_size = max(s->object_size, size);
5696 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
5697 }
5698
5699 return s;
5700}
5701
5702int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
5703{
5704 int err;
5705
5706 err = kmem_cache_open(s, flags);
5707 if (err)
5708 return err;
5709
5710 /* Mutex is not taken during early boot */
5711 if (slab_state <= UP)
5712 return 0;
5713
5714 err = sysfs_slab_add(s);
5715 if (err) {
5716 __kmem_cache_release(s);
5717 return err;
5718 }
5719
5720 if (s->flags & SLAB_STORE_USER)
5721 debugfs_slab_add(s);
5722
5723 return 0;
5724}
5725
5726#ifdef SLAB_SUPPORTS_SYSFS
5727static int count_inuse(struct slab *slab)
5728{
5729 return slab->inuse;
5730}
5731
5732static int count_total(struct slab *slab)
5733{
5734 return slab->objects;
5735}
5736#endif
5737
5738#ifdef CONFIG_SLUB_DEBUG
5739static void validate_slab(struct kmem_cache *s, struct slab *slab,
5740 unsigned long *obj_map)
5741{
5742 void *p;
5743 void *addr = slab_address(slab);
5744
5745 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
5746 return;
5747
5748 /* Now we know that a valid freelist exists */
5749 __fill_map(obj_map, s, slab);
5750 for_each_object(p, s, addr, slab->objects) {
5751 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
5752 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
5753
5754 if (!check_object(s, slab, p, val))
5755 break;
5756 }
5757}
5758
5759static int validate_slab_node(struct kmem_cache *s,
5760 struct kmem_cache_node *n, unsigned long *obj_map)
5761{
5762 unsigned long count = 0;
5763 struct slab *slab;
5764 unsigned long flags;
5765
5766 spin_lock_irqsave(&n->list_lock, flags);
5767
5768 list_for_each_entry(slab, &n->partial, slab_list) {
5769 validate_slab(s, slab, obj_map);
5770 count++;
5771 }
5772 if (count != n->nr_partial) {
5773 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
5774 s->name, count, n->nr_partial);
5775 slab_add_kunit_errors();
5776 }
5777
5778 if (!(s->flags & SLAB_STORE_USER))
5779 goto out;
5780
5781 list_for_each_entry(slab, &n->full, slab_list) {
5782 validate_slab(s, slab, obj_map);
5783 count++;
5784 }
5785 if (count != node_nr_slabs(n)) {
5786 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
5787 s->name, count, node_nr_slabs(n));
5788 slab_add_kunit_errors();
5789 }
5790
5791out:
5792 spin_unlock_irqrestore(&n->list_lock, flags);
5793 return count;
5794}
5795
5796long validate_slab_cache(struct kmem_cache *s)
5797{
5798 int node;
5799 unsigned long count = 0;
5800 struct kmem_cache_node *n;
5801 unsigned long *obj_map;
5802
5803 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
5804 if (!obj_map)
5805 return -ENOMEM;
5806
5807 flush_all(s);
5808 for_each_kmem_cache_node(s, node, n)
5809 count += validate_slab_node(s, n, obj_map);
5810
5811 bitmap_free(obj_map);
5812
5813 return count;
5814}
5815EXPORT_SYMBOL(validate_slab_cache);
5816
5817#ifdef CONFIG_DEBUG_FS
5818/*
5819 * Generate lists of code addresses where slabcache objects are allocated
5820 * and freed.
5821 */
5822
5823struct location {
5824 depot_stack_handle_t handle;
5825 unsigned long count;
5826 unsigned long addr;
5827 unsigned long waste;
5828 long long sum_time;
5829 long min_time;
5830 long max_time;
5831 long min_pid;
5832 long max_pid;
5833 DECLARE_BITMAP(cpus, NR_CPUS);
5834 nodemask_t nodes;
5835};
5836
5837struct loc_track {
5838 unsigned long max;
5839 unsigned long count;
5840 struct location *loc;
5841 loff_t idx;
5842};
5843
5844static struct dentry *slab_debugfs_root;
5845
5846static void free_loc_track(struct loc_track *t)
5847{
5848 if (t->max)
5849 free_pages((unsigned long)t->loc,
5850 get_order(sizeof(struct location) * t->max));
5851}
5852
5853static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
5854{
5855 struct location *l;
5856 int order;
5857
5858 order = get_order(sizeof(struct location) * max);
5859
5860 l = (void *)__get_free_pages(flags, order);
5861 if (!l)
5862 return 0;
5863
5864 if (t->count) {
5865 memcpy(l, t->loc, sizeof(struct location) * t->count);
5866 free_loc_track(t);
5867 }
5868 t->max = max;
5869 t->loc = l;
5870 return 1;
5871}
5872
5873static int add_location(struct loc_track *t, struct kmem_cache *s,
5874 const struct track *track,
5875 unsigned int orig_size)
5876{
5877 long start, end, pos;
5878 struct location *l;
5879 unsigned long caddr, chandle, cwaste;
5880 unsigned long age = jiffies - track->when;
5881 depot_stack_handle_t handle = 0;
5882 unsigned int waste = s->object_size - orig_size;
5883
5884#ifdef CONFIG_STACKDEPOT
5885 handle = READ_ONCE(track->handle);
5886#endif
5887 start = -1;
5888 end = t->count;
5889
5890 for ( ; ; ) {
5891 pos = start + (end - start + 1) / 2;
5892
5893 /*
5894 * There is nothing at "end". If we end up there
5895 * we need to add something to before end.
5896 */
5897 if (pos == end)
5898 break;
5899
5900 l = &t->loc[pos];
5901 caddr = l->addr;
5902 chandle = l->handle;
5903 cwaste = l->waste;
5904 if ((track->addr == caddr) && (handle == chandle) &&
5905 (waste == cwaste)) {
5906
5907 l->count++;
5908 if (track->when) {
5909 l->sum_time += age;
5910 if (age < l->min_time)
5911 l->min_time = age;
5912 if (age > l->max_time)
5913 l->max_time = age;
5914
5915 if (track->pid < l->min_pid)
5916 l->min_pid = track->pid;
5917 if (track->pid > l->max_pid)
5918 l->max_pid = track->pid;
5919
5920 cpumask_set_cpu(track->cpu,
5921 to_cpumask(l->cpus));
5922 }
5923 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5924 return 1;
5925 }
5926
5927 if (track->addr < caddr)
5928 end = pos;
5929 else if (track->addr == caddr && handle < chandle)
5930 end = pos;
5931 else if (track->addr == caddr && handle == chandle &&
5932 waste < cwaste)
5933 end = pos;
5934 else
5935 start = pos;
5936 }
5937
5938 /*
5939 * Not found. Insert new tracking element.
5940 */
5941 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
5942 return 0;
5943
5944 l = t->loc + pos;
5945 if (pos < t->count)
5946 memmove(l + 1, l,
5947 (t->count - pos) * sizeof(struct location));
5948 t->count++;
5949 l->count = 1;
5950 l->addr = track->addr;
5951 l->sum_time = age;
5952 l->min_time = age;
5953 l->max_time = age;
5954 l->min_pid = track->pid;
5955 l->max_pid = track->pid;
5956 l->handle = handle;
5957 l->waste = waste;
5958 cpumask_clear(to_cpumask(l->cpus));
5959 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
5960 nodes_clear(l->nodes);
5961 node_set(page_to_nid(virt_to_page(track)), l->nodes);
5962 return 1;
5963}
5964
5965static void process_slab(struct loc_track *t, struct kmem_cache *s,
5966 struct slab *slab, enum track_item alloc,
5967 unsigned long *obj_map)
5968{
5969 void *addr = slab_address(slab);
5970 bool is_alloc = (alloc == TRACK_ALLOC);
5971 void *p;
5972
5973 __fill_map(obj_map, s, slab);
5974
5975 for_each_object(p, s, addr, slab->objects)
5976 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
5977 add_location(t, s, get_track(s, p, alloc),
5978 is_alloc ? get_orig_size(s, p) :
5979 s->object_size);
5980}
5981#endif /* CONFIG_DEBUG_FS */
5982#endif /* CONFIG_SLUB_DEBUG */
5983
5984#ifdef SLAB_SUPPORTS_SYSFS
5985enum slab_stat_type {
5986 SL_ALL, /* All slabs */
5987 SL_PARTIAL, /* Only partially allocated slabs */
5988 SL_CPU, /* Only slabs used for cpu caches */
5989 SL_OBJECTS, /* Determine allocated objects not slabs */
5990 SL_TOTAL /* Determine object capacity not slabs */
5991};
5992
5993#define SO_ALL (1 << SL_ALL)
5994#define SO_PARTIAL (1 << SL_PARTIAL)
5995#define SO_CPU (1 << SL_CPU)
5996#define SO_OBJECTS (1 << SL_OBJECTS)
5997#define SO_TOTAL (1 << SL_TOTAL)
5998
5999static ssize_t show_slab_objects(struct kmem_cache *s,
6000 char *buf, unsigned long flags)
6001{
6002 unsigned long total = 0;
6003 int node;
6004 int x;
6005 unsigned long *nodes;
6006 int len = 0;
6007
6008 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
6009 if (!nodes)
6010 return -ENOMEM;
6011
6012 if (flags & SO_CPU) {
6013 int cpu;
6014
6015 for_each_possible_cpu(cpu) {
6016 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
6017 cpu);
6018 int node;
6019 struct slab *slab;
6020
6021 slab = READ_ONCE(c->slab);
6022 if (!slab)
6023 continue;
6024
6025 node = slab_nid(slab);
6026 if (flags & SO_TOTAL)
6027 x = slab->objects;
6028 else if (flags & SO_OBJECTS)
6029 x = slab->inuse;
6030 else
6031 x = 1;
6032
6033 total += x;
6034 nodes[node] += x;
6035
6036#ifdef CONFIG_SLUB_CPU_PARTIAL
6037 slab = slub_percpu_partial_read_once(c);
6038 if (slab) {
6039 node = slab_nid(slab);
6040 if (flags & SO_TOTAL)
6041 WARN_ON_ONCE(1);
6042 else if (flags & SO_OBJECTS)
6043 WARN_ON_ONCE(1);
6044 else
6045 x = slab->slabs;
6046 total += x;
6047 nodes[node] += x;
6048 }
6049#endif
6050 }
6051 }
6052
6053 /*
6054 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
6055 * already held which will conflict with an existing lock order:
6056 *
6057 * mem_hotplug_lock->slab_mutex->kernfs_mutex
6058 *
6059 * We don't really need mem_hotplug_lock (to hold off
6060 * slab_mem_going_offline_callback) here because slab's memory hot
6061 * unplug code doesn't destroy the kmem_cache->node[] data.
6062 */
6063
6064#ifdef CONFIG_SLUB_DEBUG
6065 if (flags & SO_ALL) {
6066 struct kmem_cache_node *n;
6067
6068 for_each_kmem_cache_node(s, node, n) {
6069
6070 if (flags & SO_TOTAL)
6071 x = node_nr_objs(n);
6072 else if (flags & SO_OBJECTS)
6073 x = node_nr_objs(n) - count_partial(n, count_free);
6074 else
6075 x = node_nr_slabs(n);
6076 total += x;
6077 nodes[node] += x;
6078 }
6079
6080 } else
6081#endif
6082 if (flags & SO_PARTIAL) {
6083 struct kmem_cache_node *n;
6084
6085 for_each_kmem_cache_node(s, node, n) {
6086 if (flags & SO_TOTAL)
6087 x = count_partial(n, count_total);
6088 else if (flags & SO_OBJECTS)
6089 x = count_partial(n, count_inuse);
6090 else
6091 x = n->nr_partial;
6092 total += x;
6093 nodes[node] += x;
6094 }
6095 }
6096
6097 len += sysfs_emit_at(buf, len, "%lu", total);
6098#ifdef CONFIG_NUMA
6099 for (node = 0; node < nr_node_ids; node++) {
6100 if (nodes[node])
6101 len += sysfs_emit_at(buf, len, " N%d=%lu",
6102 node, nodes[node]);
6103 }
6104#endif
6105 len += sysfs_emit_at(buf, len, "\n");
6106 kfree(nodes);
6107
6108 return len;
6109}
6110
6111#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
6112#define to_slab(n) container_of(n, struct kmem_cache, kobj)
6113
6114struct slab_attribute {
6115 struct attribute attr;
6116 ssize_t (*show)(struct kmem_cache *s, char *buf);
6117 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
6118};
6119
6120#define SLAB_ATTR_RO(_name) \
6121 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
6122
6123#define SLAB_ATTR(_name) \
6124 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
6125
6126static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
6127{
6128 return sysfs_emit(buf, "%u\n", s->size);
6129}
6130SLAB_ATTR_RO(slab_size);
6131
6132static ssize_t align_show(struct kmem_cache *s, char *buf)
6133{
6134 return sysfs_emit(buf, "%u\n", s->align);
6135}
6136SLAB_ATTR_RO(align);
6137
6138static ssize_t object_size_show(struct kmem_cache *s, char *buf)
6139{
6140 return sysfs_emit(buf, "%u\n", s->object_size);
6141}
6142SLAB_ATTR_RO(object_size);
6143
6144static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
6145{
6146 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
6147}
6148SLAB_ATTR_RO(objs_per_slab);
6149
6150static ssize_t order_show(struct kmem_cache *s, char *buf)
6151{
6152 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
6153}
6154SLAB_ATTR_RO(order);
6155
6156static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
6157{
6158 return sysfs_emit(buf, "%lu\n", s->min_partial);
6159}
6160
6161static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
6162 size_t length)
6163{
6164 unsigned long min;
6165 int err;
6166
6167 err = kstrtoul(buf, 10, &min);
6168 if (err)
6169 return err;
6170
6171 s->min_partial = min;
6172 return length;
6173}
6174SLAB_ATTR(min_partial);
6175
6176static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
6177{
6178 unsigned int nr_partial = 0;
6179#ifdef CONFIG_SLUB_CPU_PARTIAL
6180 nr_partial = s->cpu_partial;
6181#endif
6182
6183 return sysfs_emit(buf, "%u\n", nr_partial);
6184}
6185
6186static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
6187 size_t length)
6188{
6189 unsigned int objects;
6190 int err;
6191
6192 err = kstrtouint(buf, 10, &objects);
6193 if (err)
6194 return err;
6195 if (objects && !kmem_cache_has_cpu_partial(s))
6196 return -EINVAL;
6197
6198 slub_set_cpu_partial(s, objects);
6199 flush_all(s);
6200 return length;
6201}
6202SLAB_ATTR(cpu_partial);
6203
6204static ssize_t ctor_show(struct kmem_cache *s, char *buf)
6205{
6206 if (!s->ctor)
6207 return 0;
6208 return sysfs_emit(buf, "%pS\n", s->ctor);
6209}
6210SLAB_ATTR_RO(ctor);
6211
6212static ssize_t aliases_show(struct kmem_cache *s, char *buf)
6213{
6214 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
6215}
6216SLAB_ATTR_RO(aliases);
6217
6218static ssize_t partial_show(struct kmem_cache *s, char *buf)
6219{
6220 return show_slab_objects(s, buf, SO_PARTIAL);
6221}
6222SLAB_ATTR_RO(partial);
6223
6224static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
6225{
6226 return show_slab_objects(s, buf, SO_CPU);
6227}
6228SLAB_ATTR_RO(cpu_slabs);
6229
6230static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
6231{
6232 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
6233}
6234SLAB_ATTR_RO(objects_partial);
6235
6236static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
6237{
6238 int objects = 0;
6239 int slabs = 0;
6240 int cpu __maybe_unused;
6241 int len = 0;
6242
6243#ifdef CONFIG_SLUB_CPU_PARTIAL
6244 for_each_online_cpu(cpu) {
6245 struct slab *slab;
6246
6247 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6248
6249 if (slab)
6250 slabs += slab->slabs;
6251 }
6252#endif
6253
6254 /* Approximate half-full slabs, see slub_set_cpu_partial() */
6255 objects = (slabs * oo_objects(s->oo)) / 2;
6256 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
6257
6258#ifdef CONFIG_SLUB_CPU_PARTIAL
6259 for_each_online_cpu(cpu) {
6260 struct slab *slab;
6261
6262 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6263 if (slab) {
6264 slabs = READ_ONCE(slab->slabs);
6265 objects = (slabs * oo_objects(s->oo)) / 2;
6266 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
6267 cpu, objects, slabs);
6268 }
6269 }
6270#endif
6271 len += sysfs_emit_at(buf, len, "\n");
6272
6273 return len;
6274}
6275SLAB_ATTR_RO(slabs_cpu_partial);
6276
6277static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
6278{
6279 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
6280}
6281SLAB_ATTR_RO(reclaim_account);
6282
6283static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
6284{
6285 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
6286}
6287SLAB_ATTR_RO(hwcache_align);
6288
6289#ifdef CONFIG_ZONE_DMA
6290static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
6291{
6292 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
6293}
6294SLAB_ATTR_RO(cache_dma);
6295#endif
6296
6297#ifdef CONFIG_HARDENED_USERCOPY
6298static ssize_t usersize_show(struct kmem_cache *s, char *buf)
6299{
6300 return sysfs_emit(buf, "%u\n", s->usersize);
6301}
6302SLAB_ATTR_RO(usersize);
6303#endif
6304
6305static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
6306{
6307 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
6308}
6309SLAB_ATTR_RO(destroy_by_rcu);
6310
6311#ifdef CONFIG_SLUB_DEBUG
6312static ssize_t slabs_show(struct kmem_cache *s, char *buf)
6313{
6314 return show_slab_objects(s, buf, SO_ALL);
6315}
6316SLAB_ATTR_RO(slabs);
6317
6318static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
6319{
6320 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
6321}
6322SLAB_ATTR_RO(total_objects);
6323
6324static ssize_t objects_show(struct kmem_cache *s, char *buf)
6325{
6326 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
6327}
6328SLAB_ATTR_RO(objects);
6329
6330static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
6331{
6332 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
6333}
6334SLAB_ATTR_RO(sanity_checks);
6335
6336static ssize_t trace_show(struct kmem_cache *s, char *buf)
6337{
6338 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
6339}
6340SLAB_ATTR_RO(trace);
6341
6342static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
6343{
6344 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
6345}
6346
6347SLAB_ATTR_RO(red_zone);
6348
6349static ssize_t poison_show(struct kmem_cache *s, char *buf)
6350{
6351 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
6352}
6353
6354SLAB_ATTR_RO(poison);
6355
6356static ssize_t store_user_show(struct kmem_cache *s, char *buf)
6357{
6358 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
6359}
6360
6361SLAB_ATTR_RO(store_user);
6362
6363static ssize_t validate_show(struct kmem_cache *s, char *buf)
6364{
6365 return 0;
6366}
6367
6368static ssize_t validate_store(struct kmem_cache *s,
6369 const char *buf, size_t length)
6370{
6371 int ret = -EINVAL;
6372
6373 if (buf[0] == '1' && kmem_cache_debug(s)) {
6374 ret = validate_slab_cache(s);
6375 if (ret >= 0)
6376 ret = length;
6377 }
6378 return ret;
6379}
6380SLAB_ATTR(validate);
6381
6382#endif /* CONFIG_SLUB_DEBUG */
6383
6384#ifdef CONFIG_FAILSLAB
6385static ssize_t failslab_show(struct kmem_cache *s, char *buf)
6386{
6387 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
6388}
6389
6390static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
6391 size_t length)
6392{
6393 if (s->refcount > 1)
6394 return -EINVAL;
6395
6396 if (buf[0] == '1')
6397 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
6398 else
6399 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
6400
6401 return length;
6402}
6403SLAB_ATTR(failslab);
6404#endif
6405
6406static ssize_t shrink_show(struct kmem_cache *s, char *buf)
6407{
6408 return 0;
6409}
6410
6411static ssize_t shrink_store(struct kmem_cache *s,
6412 const char *buf, size_t length)
6413{
6414 if (buf[0] == '1')
6415 kmem_cache_shrink(s);
6416 else
6417 return -EINVAL;
6418 return length;
6419}
6420SLAB_ATTR(shrink);
6421
6422#ifdef CONFIG_NUMA
6423static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
6424{
6425 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
6426}
6427
6428static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
6429 const char *buf, size_t length)
6430{
6431 unsigned int ratio;
6432 int err;
6433
6434 err = kstrtouint(buf, 10, &ratio);
6435 if (err)
6436 return err;
6437 if (ratio > 100)
6438 return -ERANGE;
6439
6440 s->remote_node_defrag_ratio = ratio * 10;
6441
6442 return length;
6443}
6444SLAB_ATTR(remote_node_defrag_ratio);
6445#endif
6446
6447#ifdef CONFIG_SLUB_STATS
6448static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
6449{
6450 unsigned long sum = 0;
6451 int cpu;
6452 int len = 0;
6453 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
6454
6455 if (!data)
6456 return -ENOMEM;
6457
6458 for_each_online_cpu(cpu) {
6459 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
6460
6461 data[cpu] = x;
6462 sum += x;
6463 }
6464
6465 len += sysfs_emit_at(buf, len, "%lu", sum);
6466
6467#ifdef CONFIG_SMP
6468 for_each_online_cpu(cpu) {
6469 if (data[cpu])
6470 len += sysfs_emit_at(buf, len, " C%d=%u",
6471 cpu, data[cpu]);
6472 }
6473#endif
6474 kfree(data);
6475 len += sysfs_emit_at(buf, len, "\n");
6476
6477 return len;
6478}
6479
6480static void clear_stat(struct kmem_cache *s, enum stat_item si)
6481{
6482 int cpu;
6483
6484 for_each_online_cpu(cpu)
6485 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
6486}
6487
6488#define STAT_ATTR(si, text) \
6489static ssize_t text##_show(struct kmem_cache *s, char *buf) \
6490{ \
6491 return show_stat(s, buf, si); \
6492} \
6493static ssize_t text##_store(struct kmem_cache *s, \
6494 const char *buf, size_t length) \
6495{ \
6496 if (buf[0] != '0') \
6497 return -EINVAL; \
6498 clear_stat(s, si); \
6499 return length; \
6500} \
6501SLAB_ATTR(text); \
6502
6503STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
6504STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
6505STAT_ATTR(FREE_FASTPATH, free_fastpath);
6506STAT_ATTR(FREE_SLOWPATH, free_slowpath);
6507STAT_ATTR(FREE_FROZEN, free_frozen);
6508STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
6509STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
6510STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
6511STAT_ATTR(ALLOC_SLAB, alloc_slab);
6512STAT_ATTR(ALLOC_REFILL, alloc_refill);
6513STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
6514STAT_ATTR(FREE_SLAB, free_slab);
6515STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
6516STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
6517STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
6518STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
6519STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
6520STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
6521STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
6522STAT_ATTR(ORDER_FALLBACK, order_fallback);
6523STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
6524STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
6525STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
6526STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
6527STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
6528STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
6529#endif /* CONFIG_SLUB_STATS */
6530
6531#ifdef CONFIG_KFENCE
6532static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
6533{
6534 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
6535}
6536
6537static ssize_t skip_kfence_store(struct kmem_cache *s,
6538 const char *buf, size_t length)
6539{
6540 int ret = length;
6541
6542 if (buf[0] == '0')
6543 s->flags &= ~SLAB_SKIP_KFENCE;
6544 else if (buf[0] == '1')
6545 s->flags |= SLAB_SKIP_KFENCE;
6546 else
6547 ret = -EINVAL;
6548
6549 return ret;
6550}
6551SLAB_ATTR(skip_kfence);
6552#endif
6553
6554static struct attribute *slab_attrs[] = {
6555 &slab_size_attr.attr,
6556 &object_size_attr.attr,
6557 &objs_per_slab_attr.attr,
6558 &order_attr.attr,
6559 &min_partial_attr.attr,
6560 &cpu_partial_attr.attr,
6561 &objects_partial_attr.attr,
6562 &partial_attr.attr,
6563 &cpu_slabs_attr.attr,
6564 &ctor_attr.attr,
6565 &aliases_attr.attr,
6566 &align_attr.attr,
6567 &hwcache_align_attr.attr,
6568 &reclaim_account_attr.attr,
6569 &destroy_by_rcu_attr.attr,
6570 &shrink_attr.attr,
6571 &slabs_cpu_partial_attr.attr,
6572#ifdef CONFIG_SLUB_DEBUG
6573 &total_objects_attr.attr,
6574 &objects_attr.attr,
6575 &slabs_attr.attr,
6576 &sanity_checks_attr.attr,
6577 &trace_attr.attr,
6578 &red_zone_attr.attr,
6579 &poison_attr.attr,
6580 &store_user_attr.attr,
6581 &validate_attr.attr,
6582#endif
6583#ifdef CONFIG_ZONE_DMA
6584 &cache_dma_attr.attr,
6585#endif
6586#ifdef CONFIG_NUMA
6587 &remote_node_defrag_ratio_attr.attr,
6588#endif
6589#ifdef CONFIG_SLUB_STATS
6590 &alloc_fastpath_attr.attr,
6591 &alloc_slowpath_attr.attr,
6592 &free_fastpath_attr.attr,
6593 &free_slowpath_attr.attr,
6594 &free_frozen_attr.attr,
6595 &free_add_partial_attr.attr,
6596 &free_remove_partial_attr.attr,
6597 &alloc_from_partial_attr.attr,
6598 &alloc_slab_attr.attr,
6599 &alloc_refill_attr.attr,
6600 &alloc_node_mismatch_attr.attr,
6601 &free_slab_attr.attr,
6602 &cpuslab_flush_attr.attr,
6603 &deactivate_full_attr.attr,
6604 &deactivate_empty_attr.attr,
6605 &deactivate_to_head_attr.attr,
6606 &deactivate_to_tail_attr.attr,
6607 &deactivate_remote_frees_attr.attr,
6608 &deactivate_bypass_attr.attr,
6609 &order_fallback_attr.attr,
6610 &cmpxchg_double_fail_attr.attr,
6611 &cmpxchg_double_cpu_fail_attr.attr,
6612 &cpu_partial_alloc_attr.attr,
6613 &cpu_partial_free_attr.attr,
6614 &cpu_partial_node_attr.attr,
6615 &cpu_partial_drain_attr.attr,
6616#endif
6617#ifdef CONFIG_FAILSLAB
6618 &failslab_attr.attr,
6619#endif
6620#ifdef CONFIG_HARDENED_USERCOPY
6621 &usersize_attr.attr,
6622#endif
6623#ifdef CONFIG_KFENCE
6624 &skip_kfence_attr.attr,
6625#endif
6626
6627 NULL
6628};
6629
6630static const struct attribute_group slab_attr_group = {
6631 .attrs = slab_attrs,
6632};
6633
6634static ssize_t slab_attr_show(struct kobject *kobj,
6635 struct attribute *attr,
6636 char *buf)
6637{
6638 struct slab_attribute *attribute;
6639 struct kmem_cache *s;
6640
6641 attribute = to_slab_attr(attr);
6642 s = to_slab(kobj);
6643
6644 if (!attribute->show)
6645 return -EIO;
6646
6647 return attribute->show(s, buf);
6648}
6649
6650static ssize_t slab_attr_store(struct kobject *kobj,
6651 struct attribute *attr,
6652 const char *buf, size_t len)
6653{
6654 struct slab_attribute *attribute;
6655 struct kmem_cache *s;
6656
6657 attribute = to_slab_attr(attr);
6658 s = to_slab(kobj);
6659
6660 if (!attribute->store)
6661 return -EIO;
6662
6663 return attribute->store(s, buf, len);
6664}
6665
6666static void kmem_cache_release(struct kobject *k)
6667{
6668 slab_kmem_cache_release(to_slab(k));
6669}
6670
6671static const struct sysfs_ops slab_sysfs_ops = {
6672 .show = slab_attr_show,
6673 .store = slab_attr_store,
6674};
6675
6676static const struct kobj_type slab_ktype = {
6677 .sysfs_ops = &slab_sysfs_ops,
6678 .release = kmem_cache_release,
6679};
6680
6681static struct kset *slab_kset;
6682
6683static inline struct kset *cache_kset(struct kmem_cache *s)
6684{
6685 return slab_kset;
6686}
6687
6688#define ID_STR_LENGTH 32
6689
6690/* Create a unique string id for a slab cache:
6691 *
6692 * Format :[flags-]size
6693 */
6694static char *create_unique_id(struct kmem_cache *s)
6695{
6696 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
6697 char *p = name;
6698
6699 if (!name)
6700 return ERR_PTR(-ENOMEM);
6701
6702 *p++ = ':';
6703 /*
6704 * First flags affecting slabcache operations. We will only
6705 * get here for aliasable slabs so we do not need to support
6706 * too many flags. The flags here must cover all flags that
6707 * are matched during merging to guarantee that the id is
6708 * unique.
6709 */
6710 if (s->flags & SLAB_CACHE_DMA)
6711 *p++ = 'd';
6712 if (s->flags & SLAB_CACHE_DMA32)
6713 *p++ = 'D';
6714 if (s->flags & SLAB_RECLAIM_ACCOUNT)
6715 *p++ = 'a';
6716 if (s->flags & SLAB_CONSISTENCY_CHECKS)
6717 *p++ = 'F';
6718 if (s->flags & SLAB_ACCOUNT)
6719 *p++ = 'A';
6720 if (p != name + 1)
6721 *p++ = '-';
6722 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
6723
6724 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
6725 kfree(name);
6726 return ERR_PTR(-EINVAL);
6727 }
6728 kmsan_unpoison_memory(name, p - name);
6729 return name;
6730}
6731
6732static int sysfs_slab_add(struct kmem_cache *s)
6733{
6734 int err;
6735 const char *name;
6736 struct kset *kset = cache_kset(s);
6737 int unmergeable = slab_unmergeable(s);
6738
6739 if (!unmergeable && disable_higher_order_debug &&
6740 (slub_debug & DEBUG_METADATA_FLAGS))
6741 unmergeable = 1;
6742
6743 if (unmergeable) {
6744 /*
6745 * Slabcache can never be merged so we can use the name proper.
6746 * This is typically the case for debug situations. In that
6747 * case we can catch duplicate names easily.
6748 */
6749 sysfs_remove_link(&slab_kset->kobj, s->name);
6750 name = s->name;
6751 } else {
6752 /*
6753 * Create a unique name for the slab as a target
6754 * for the symlinks.
6755 */
6756 name = create_unique_id(s);
6757 if (IS_ERR(name))
6758 return PTR_ERR(name);
6759 }
6760
6761 s->kobj.kset = kset;
6762 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
6763 if (err)
6764 goto out;
6765
6766 err = sysfs_create_group(&s->kobj, &slab_attr_group);
6767 if (err)
6768 goto out_del_kobj;
6769
6770 if (!unmergeable) {
6771 /* Setup first alias */
6772 sysfs_slab_alias(s, s->name);
6773 }
6774out:
6775 if (!unmergeable)
6776 kfree(name);
6777 return err;
6778out_del_kobj:
6779 kobject_del(&s->kobj);
6780 goto out;
6781}
6782
6783void sysfs_slab_unlink(struct kmem_cache *s)
6784{
6785 kobject_del(&s->kobj);
6786}
6787
6788void sysfs_slab_release(struct kmem_cache *s)
6789{
6790 kobject_put(&s->kobj);
6791}
6792
6793/*
6794 * Need to buffer aliases during bootup until sysfs becomes
6795 * available lest we lose that information.
6796 */
6797struct saved_alias {
6798 struct kmem_cache *s;
6799 const char *name;
6800 struct saved_alias *next;
6801};
6802
6803static struct saved_alias *alias_list;
6804
6805static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
6806{
6807 struct saved_alias *al;
6808
6809 if (slab_state == FULL) {
6810 /*
6811 * If we have a leftover link then remove it.
6812 */
6813 sysfs_remove_link(&slab_kset->kobj, name);
6814 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
6815 }
6816
6817 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
6818 if (!al)
6819 return -ENOMEM;
6820
6821 al->s = s;
6822 al->name = name;
6823 al->next = alias_list;
6824 alias_list = al;
6825 kmsan_unpoison_memory(al, sizeof(*al));
6826 return 0;
6827}
6828
6829static int __init slab_sysfs_init(void)
6830{
6831 struct kmem_cache *s;
6832 int err;
6833
6834 mutex_lock(&slab_mutex);
6835
6836 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
6837 if (!slab_kset) {
6838 mutex_unlock(&slab_mutex);
6839 pr_err("Cannot register slab subsystem.\n");
6840 return -ENOMEM;
6841 }
6842
6843 slab_state = FULL;
6844
6845 list_for_each_entry(s, &slab_caches, list) {
6846 err = sysfs_slab_add(s);
6847 if (err)
6848 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
6849 s->name);
6850 }
6851
6852 while (alias_list) {
6853 struct saved_alias *al = alias_list;
6854
6855 alias_list = alias_list->next;
6856 err = sysfs_slab_alias(al->s, al->name);
6857 if (err)
6858 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
6859 al->name);
6860 kfree(al);
6861 }
6862
6863 mutex_unlock(&slab_mutex);
6864 return 0;
6865}
6866late_initcall(slab_sysfs_init);
6867#endif /* SLAB_SUPPORTS_SYSFS */
6868
6869#if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
6870static int slab_debugfs_show(struct seq_file *seq, void *v)
6871{
6872 struct loc_track *t = seq->private;
6873 struct location *l;
6874 unsigned long idx;
6875
6876 idx = (unsigned long) t->idx;
6877 if (idx < t->count) {
6878 l = &t->loc[idx];
6879
6880 seq_printf(seq, "%7ld ", l->count);
6881
6882 if (l->addr)
6883 seq_printf(seq, "%pS", (void *)l->addr);
6884 else
6885 seq_puts(seq, "<not-available>");
6886
6887 if (l->waste)
6888 seq_printf(seq, " waste=%lu/%lu",
6889 l->count * l->waste, l->waste);
6890
6891 if (l->sum_time != l->min_time) {
6892 seq_printf(seq, " age=%ld/%llu/%ld",
6893 l->min_time, div_u64(l->sum_time, l->count),
6894 l->max_time);
6895 } else
6896 seq_printf(seq, " age=%ld", l->min_time);
6897
6898 if (l->min_pid != l->max_pid)
6899 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
6900 else
6901 seq_printf(seq, " pid=%ld",
6902 l->min_pid);
6903
6904 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
6905 seq_printf(seq, " cpus=%*pbl",
6906 cpumask_pr_args(to_cpumask(l->cpus)));
6907
6908 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
6909 seq_printf(seq, " nodes=%*pbl",
6910 nodemask_pr_args(&l->nodes));
6911
6912#ifdef CONFIG_STACKDEPOT
6913 {
6914 depot_stack_handle_t handle;
6915 unsigned long *entries;
6916 unsigned int nr_entries, j;
6917
6918 handle = READ_ONCE(l->handle);
6919 if (handle) {
6920 nr_entries = stack_depot_fetch(handle, &entries);
6921 seq_puts(seq, "\n");
6922 for (j = 0; j < nr_entries; j++)
6923 seq_printf(seq, " %pS\n", (void *)entries[j]);
6924 }
6925 }
6926#endif
6927 seq_puts(seq, "\n");
6928 }
6929
6930 if (!idx && !t->count)
6931 seq_puts(seq, "No data\n");
6932
6933 return 0;
6934}
6935
6936static void slab_debugfs_stop(struct seq_file *seq, void *v)
6937{
6938}
6939
6940static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
6941{
6942 struct loc_track *t = seq->private;
6943
6944 t->idx = ++(*ppos);
6945 if (*ppos <= t->count)
6946 return ppos;
6947
6948 return NULL;
6949}
6950
6951static int cmp_loc_by_count(const void *a, const void *b, const void *data)
6952{
6953 struct location *loc1 = (struct location *)a;
6954 struct location *loc2 = (struct location *)b;
6955
6956 if (loc1->count > loc2->count)
6957 return -1;
6958 else
6959 return 1;
6960}
6961
6962static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
6963{
6964 struct loc_track *t = seq->private;
6965
6966 t->idx = *ppos;
6967 return ppos;
6968}
6969
6970static const struct seq_operations slab_debugfs_sops = {
6971 .start = slab_debugfs_start,
6972 .next = slab_debugfs_next,
6973 .stop = slab_debugfs_stop,
6974 .show = slab_debugfs_show,
6975};
6976
6977static int slab_debug_trace_open(struct inode *inode, struct file *filep)
6978{
6979
6980 struct kmem_cache_node *n;
6981 enum track_item alloc;
6982 int node;
6983 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
6984 sizeof(struct loc_track));
6985 struct kmem_cache *s = file_inode(filep)->i_private;
6986 unsigned long *obj_map;
6987
6988 if (!t)
6989 return -ENOMEM;
6990
6991 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6992 if (!obj_map) {
6993 seq_release_private(inode, filep);
6994 return -ENOMEM;
6995 }
6996
6997 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
6998 alloc = TRACK_ALLOC;
6999 else
7000 alloc = TRACK_FREE;
7001
7002 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
7003 bitmap_free(obj_map);
7004 seq_release_private(inode, filep);
7005 return -ENOMEM;
7006 }
7007
7008 for_each_kmem_cache_node(s, node, n) {
7009 unsigned long flags;
7010 struct slab *slab;
7011
7012 if (!node_nr_slabs(n))
7013 continue;
7014
7015 spin_lock_irqsave(&n->list_lock, flags);
7016 list_for_each_entry(slab, &n->partial, slab_list)
7017 process_slab(t, s, slab, alloc, obj_map);
7018 list_for_each_entry(slab, &n->full, slab_list)
7019 process_slab(t, s, slab, alloc, obj_map);
7020 spin_unlock_irqrestore(&n->list_lock, flags);
7021 }
7022
7023 /* Sort locations by count */
7024 sort_r(t->loc, t->count, sizeof(struct location),
7025 cmp_loc_by_count, NULL, NULL);
7026
7027 bitmap_free(obj_map);
7028 return 0;
7029}
7030
7031static int slab_debug_trace_release(struct inode *inode, struct file *file)
7032{
7033 struct seq_file *seq = file->private_data;
7034 struct loc_track *t = seq->private;
7035
7036 free_loc_track(t);
7037 return seq_release_private(inode, file);
7038}
7039
7040static const struct file_operations slab_debugfs_fops = {
7041 .open = slab_debug_trace_open,
7042 .read = seq_read,
7043 .llseek = seq_lseek,
7044 .release = slab_debug_trace_release,
7045};
7046
7047static void debugfs_slab_add(struct kmem_cache *s)
7048{
7049 struct dentry *slab_cache_dir;
7050
7051 if (unlikely(!slab_debugfs_root))
7052 return;
7053
7054 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
7055
7056 debugfs_create_file("alloc_traces", 0400,
7057 slab_cache_dir, s, &slab_debugfs_fops);
7058
7059 debugfs_create_file("free_traces", 0400,
7060 slab_cache_dir, s, &slab_debugfs_fops);
7061}
7062
7063void debugfs_slab_release(struct kmem_cache *s)
7064{
7065 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
7066}
7067
7068static int __init slab_debugfs_init(void)
7069{
7070 struct kmem_cache *s;
7071
7072 slab_debugfs_root = debugfs_create_dir("slab", NULL);
7073
7074 list_for_each_entry(s, &slab_caches, list)
7075 if (s->flags & SLAB_STORE_USER)
7076 debugfs_slab_add(s);
7077
7078 return 0;
7079
7080}
7081__initcall(slab_debugfs_init);
7082#endif
7083/*
7084 * The /proc/slabinfo ABI
7085 */
7086#ifdef CONFIG_SLUB_DEBUG
7087void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
7088{
7089 unsigned long nr_slabs = 0;
7090 unsigned long nr_objs = 0;
7091 unsigned long nr_free = 0;
7092 int node;
7093 struct kmem_cache_node *n;
7094
7095 for_each_kmem_cache_node(s, node, n) {
7096 nr_slabs += node_nr_slabs(n);
7097 nr_objs += node_nr_objs(n);
7098 nr_free += count_partial(n, count_free);
7099 }
7100
7101 sinfo->active_objs = nr_objs - nr_free;
7102 sinfo->num_objs = nr_objs;
7103 sinfo->active_slabs = nr_slabs;
7104 sinfo->num_slabs = nr_slabs;
7105 sinfo->objects_per_slab = oo_objects(s->oo);
7106 sinfo->cache_order = oo_order(s->oo);
7107}
7108
7109void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
7110{
7111}
7112
7113ssize_t slabinfo_write(struct file *file, const char __user *buffer,
7114 size_t count, loff_t *ppos)
7115{
7116 return -EIO;
7117}
7118#endif /* CONFIG_SLUB_DEBUG */