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 operatios
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/bitops.h>
19#include <linux/slab.h>
20#include "slab.h"
21#include <linux/proc_fs.h>
22#include <linux/seq_file.h>
23#include <linux/kasan.h>
24#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
28#include <linux/debugobjects.h>
29#include <linux/kallsyms.h>
30#include <linux/memory.h>
31#include <linux/math64.h>
32#include <linux/fault-inject.h>
33#include <linux/stacktrace.h>
34#include <linux/prefetch.h>
35#include <linux/memcontrol.h>
36#include <linux/random.h>
37
38#include <trace/events/kmem.h>
39
40#include "internal.h"
41
42/*
43 * Lock order:
44 * 1. slab_mutex (Global Mutex)
45 * 2. node->list_lock
46 * 3. slab_lock(page) (Only on some arches and for debugging)
47 *
48 * slab_mutex
49 *
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
52 *
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects:
55 * A. page->freelist -> List of object free in a page
56 * B. page->inuse -> Number of objects in use
57 * C. page->objects -> Number of objects in page
58 * D. page->frozen -> frozen state
59 *
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list except per cpu partial list. The processor that froze the
62 * slab is the one who can perform list operations on the page. Other
63 * processors may put objects onto the freelist but the processor that
64 * froze the slab is the only one that can retrieve the objects from the
65 * page's freelist.
66 *
67 * The list_lock protects the partial and full list on each node and
68 * the partial slab counter. If taken then no new slabs may be added or
69 * removed from the lists nor make the number of partial slabs be modified.
70 * (Note that the total number of slabs is an atomic value that may be
71 * modified without taking the list lock).
72 *
73 * The list_lock is a centralized lock and thus we avoid taking it as
74 * much as possible. As long as SLUB does not have to handle partial
75 * slabs, operations can continue without any centralized lock. F.e.
76 * allocating a long series of objects that fill up slabs does not require
77 * the list lock.
78 * Interrupts are disabled during allocation and deallocation in order to
79 * make the slab allocator safe to use in the context of an irq. In addition
80 * interrupts are disabled to ensure that the processor does not change
81 * while handling per_cpu slabs, due to kernel preemption.
82 *
83 * SLUB assigns one slab for allocation to each processor.
84 * Allocations only occur from these slabs called cpu slabs.
85 *
86 * Slabs with free elements are kept on a partial list and during regular
87 * operations no list for full slabs is used. If an object in a full slab is
88 * freed then the slab will show up again on the partial lists.
89 * We track full slabs for debugging purposes though because otherwise we
90 * cannot scan all objects.
91 *
92 * Slabs are freed when they become empty. Teardown and setup is
93 * minimal so we rely on the page allocators per cpu caches for
94 * fast frees and allocs.
95 *
96 * Overloading of page flags that are otherwise used for LRU management.
97 *
98 * PageActive The slab is frozen and exempt from list processing.
99 * This means that the slab is dedicated to a purpose
100 * such as satisfying allocations for a specific
101 * processor. Objects may be freed in the slab while
102 * it is frozen but slab_free will then skip the usual
103 * list operations. It is up to the processor holding
104 * the slab to integrate the slab into the slab lists
105 * when the slab is no longer needed.
106 *
107 * One use of this flag is to mark slabs that are
108 * used for allocations. Then such a slab becomes a cpu
109 * slab. The cpu slab may be equipped with an additional
110 * freelist that allows lockless access to
111 * free objects in addition to the regular freelist
112 * that requires the slab lock.
113 *
114 * PageError Slab requires special handling due to debug
115 * options set. This moves slab handling out of
116 * the fast path and disables lockless freelists.
117 */
118
119static inline int kmem_cache_debug(struct kmem_cache *s)
120{
121#ifdef CONFIG_SLUB_DEBUG
122 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
123#else
124 return 0;
125#endif
126}
127
128void *fixup_red_left(struct kmem_cache *s, void *p)
129{
130 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
131 p += s->red_left_pad;
132
133 return p;
134}
135
136static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
137{
138#ifdef CONFIG_SLUB_CPU_PARTIAL
139 return !kmem_cache_debug(s);
140#else
141 return false;
142#endif
143}
144
145/*
146 * Issues still to be resolved:
147 *
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
149 *
150 * - Variable sizing of the per node arrays
151 */
152
153/* Enable to test recovery from slab corruption on boot */
154#undef SLUB_RESILIENCY_TEST
155
156/* Enable to log cmpxchg failures */
157#undef SLUB_DEBUG_CMPXCHG
158
159/*
160 * Mininum number of partial slabs. These will be left on the partial
161 * lists even if they are empty. kmem_cache_shrink may reclaim them.
162 */
163#define MIN_PARTIAL 5
164
165/*
166 * Maximum number of desirable partial slabs.
167 * The existence of more partial slabs makes kmem_cache_shrink
168 * sort the partial list by the number of objects in use.
169 */
170#define MAX_PARTIAL 10
171
172#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
173 SLAB_POISON | SLAB_STORE_USER)
174
175/*
176 * These debug flags cannot use CMPXCHG because there might be consistency
177 * issues when checking or reading debug information
178 */
179#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
180 SLAB_TRACE)
181
182
183/*
184 * Debugging flags that require metadata to be stored in the slab. These get
185 * disabled when slub_debug=O is used and a cache's min order increases with
186 * metadata.
187 */
188#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
189
190#define OO_SHIFT 16
191#define OO_MASK ((1 << OO_SHIFT) - 1)
192#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
193
194/* Internal SLUB flags */
195/* Poison object */
196#define __OBJECT_POISON ((slab_flags_t __force)0x80000000U)
197/* Use cmpxchg_double */
198#define __CMPXCHG_DOUBLE ((slab_flags_t __force)0x40000000U)
199
200/*
201 * Tracking user of a slab.
202 */
203#define TRACK_ADDRS_COUNT 16
204struct track {
205 unsigned long addr; /* Called from address */
206#ifdef CONFIG_STACKTRACE
207 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
208#endif
209 int cpu; /* Was running on cpu */
210 int pid; /* Pid context */
211 unsigned long when; /* When did the operation occur */
212};
213
214enum track_item { TRACK_ALLOC, TRACK_FREE };
215
216#ifdef CONFIG_SYSFS
217static int sysfs_slab_add(struct kmem_cache *);
218static int sysfs_slab_alias(struct kmem_cache *, const char *);
219static void memcg_propagate_slab_attrs(struct kmem_cache *s);
220static void sysfs_slab_remove(struct kmem_cache *s);
221#else
222static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
223static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
224 { return 0; }
225static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
226static inline void sysfs_slab_remove(struct kmem_cache *s) { }
227#endif
228
229static inline void stat(const struct kmem_cache *s, enum stat_item si)
230{
231#ifdef CONFIG_SLUB_STATS
232 /*
233 * The rmw is racy on a preemptible kernel but this is acceptable, so
234 * avoid this_cpu_add()'s irq-disable overhead.
235 */
236 raw_cpu_inc(s->cpu_slab->stat[si]);
237#endif
238}
239
240/********************************************************************
241 * Core slab cache functions
242 *******************************************************************/
243
244/*
245 * Returns freelist pointer (ptr). With hardening, this is obfuscated
246 * with an XOR of the address where the pointer is held and a per-cache
247 * random number.
248 */
249static inline void *freelist_ptr(const struct kmem_cache *s, void *ptr,
250 unsigned long ptr_addr)
251{
252#ifdef CONFIG_SLAB_FREELIST_HARDENED
253 /*
254 * When CONFIG_KASAN_SW_TAGS is enabled, ptr_addr might be tagged.
255 * Normally, this doesn't cause any issues, as both set_freepointer()
256 * and get_freepointer() are called with a pointer with the same tag.
257 * However, there are some issues with CONFIG_SLUB_DEBUG code. For
258 * example, when __free_slub() iterates over objects in a cache, it
259 * passes untagged pointers to check_object(). check_object() in turns
260 * calls get_freepointer() with an untagged pointer, which causes the
261 * freepointer to be restored incorrectly.
262 */
263 return (void *)((unsigned long)ptr ^ s->random ^
264 (unsigned long)kasan_reset_tag((void *)ptr_addr));
265#else
266 return ptr;
267#endif
268}
269
270/* Returns the freelist pointer recorded at location ptr_addr. */
271static inline void *freelist_dereference(const struct kmem_cache *s,
272 void *ptr_addr)
273{
274 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
275 (unsigned long)ptr_addr);
276}
277
278static inline void *get_freepointer(struct kmem_cache *s, void *object)
279{
280 return freelist_dereference(s, object + s->offset);
281}
282
283static void prefetch_freepointer(const struct kmem_cache *s, void *object)
284{
285 prefetch(object + s->offset);
286}
287
288static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
289{
290 unsigned long freepointer_addr;
291 void *p;
292
293 if (!debug_pagealloc_enabled())
294 return get_freepointer(s, object);
295
296 freepointer_addr = (unsigned long)object + s->offset;
297 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
298 return freelist_ptr(s, p, freepointer_addr);
299}
300
301static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
302{
303 unsigned long freeptr_addr = (unsigned long)object + s->offset;
304
305#ifdef CONFIG_SLAB_FREELIST_HARDENED
306 BUG_ON(object == fp); /* naive detection of double free or corruption */
307#endif
308
309 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
310}
311
312/* Loop over all objects in a slab */
313#define for_each_object(__p, __s, __addr, __objects) \
314 for (__p = fixup_red_left(__s, __addr); \
315 __p < (__addr) + (__objects) * (__s)->size; \
316 __p += (__s)->size)
317
318/* Determine object index from a given position */
319static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
320{
321 return (kasan_reset_tag(p) - addr) / s->size;
322}
323
324static inline unsigned int order_objects(unsigned int order, unsigned int size)
325{
326 return ((unsigned int)PAGE_SIZE << order) / size;
327}
328
329static inline struct kmem_cache_order_objects oo_make(unsigned int order,
330 unsigned int size)
331{
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size)
334 };
335
336 return x;
337}
338
339static inline unsigned int oo_order(struct kmem_cache_order_objects x)
340{
341 return x.x >> OO_SHIFT;
342}
343
344static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
345{
346 return x.x & OO_MASK;
347}
348
349/*
350 * Per slab locking using the pagelock
351 */
352static __always_inline void slab_lock(struct page *page)
353{
354 VM_BUG_ON_PAGE(PageTail(page), page);
355 bit_spin_lock(PG_locked, &page->flags);
356}
357
358static __always_inline void slab_unlock(struct page *page)
359{
360 VM_BUG_ON_PAGE(PageTail(page), page);
361 __bit_spin_unlock(PG_locked, &page->flags);
362}
363
364/* Interrupts must be disabled (for the fallback code to work right) */
365static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
366 void *freelist_old, unsigned long counters_old,
367 void *freelist_new, unsigned long counters_new,
368 const char *n)
369{
370 VM_BUG_ON(!irqs_disabled());
371#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
372 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
373 if (s->flags & __CMPXCHG_DOUBLE) {
374 if (cmpxchg_double(&page->freelist, &page->counters,
375 freelist_old, counters_old,
376 freelist_new, counters_new))
377 return true;
378 } else
379#endif
380 {
381 slab_lock(page);
382 if (page->freelist == freelist_old &&
383 page->counters == counters_old) {
384 page->freelist = freelist_new;
385 page->counters = counters_new;
386 slab_unlock(page);
387 return true;
388 }
389 slab_unlock(page);
390 }
391
392 cpu_relax();
393 stat(s, CMPXCHG_DOUBLE_FAIL);
394
395#ifdef SLUB_DEBUG_CMPXCHG
396 pr_info("%s %s: cmpxchg double redo ", n, s->name);
397#endif
398
399 return false;
400}
401
402static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
403 void *freelist_old, unsigned long counters_old,
404 void *freelist_new, unsigned long counters_new,
405 const char *n)
406{
407#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
408 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
409 if (s->flags & __CMPXCHG_DOUBLE) {
410 if (cmpxchg_double(&page->freelist, &page->counters,
411 freelist_old, counters_old,
412 freelist_new, counters_new))
413 return true;
414 } else
415#endif
416 {
417 unsigned long flags;
418
419 local_irq_save(flags);
420 slab_lock(page);
421 if (page->freelist == freelist_old &&
422 page->counters == counters_old) {
423 page->freelist = freelist_new;
424 page->counters = counters_new;
425 slab_unlock(page);
426 local_irq_restore(flags);
427 return true;
428 }
429 slab_unlock(page);
430 local_irq_restore(flags);
431 }
432
433 cpu_relax();
434 stat(s, CMPXCHG_DOUBLE_FAIL);
435
436#ifdef SLUB_DEBUG_CMPXCHG
437 pr_info("%s %s: cmpxchg double redo ", n, s->name);
438#endif
439
440 return false;
441}
442
443#ifdef CONFIG_SLUB_DEBUG
444/*
445 * Determine a map of object in use on a page.
446 *
447 * Node listlock must be held to guarantee that the page does
448 * not vanish from under us.
449 */
450static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
451{
452 void *p;
453 void *addr = page_address(page);
454
455 for (p = page->freelist; p; p = get_freepointer(s, p))
456 set_bit(slab_index(p, s, addr), map);
457}
458
459static inline unsigned int size_from_object(struct kmem_cache *s)
460{
461 if (s->flags & SLAB_RED_ZONE)
462 return s->size - s->red_left_pad;
463
464 return s->size;
465}
466
467static inline void *restore_red_left(struct kmem_cache *s, void *p)
468{
469 if (s->flags & SLAB_RED_ZONE)
470 p -= s->red_left_pad;
471
472 return p;
473}
474
475/*
476 * Debug settings:
477 */
478#if defined(CONFIG_SLUB_DEBUG_ON)
479static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
480#else
481static slab_flags_t slub_debug;
482#endif
483
484static char *slub_debug_slabs;
485static int disable_higher_order_debug;
486
487/*
488 * slub is about to manipulate internal object metadata. This memory lies
489 * outside the range of the allocated object, so accessing it would normally
490 * be reported by kasan as a bounds error. metadata_access_enable() is used
491 * to tell kasan that these accesses are OK.
492 */
493static inline void metadata_access_enable(void)
494{
495 kasan_disable_current();
496}
497
498static inline void metadata_access_disable(void)
499{
500 kasan_enable_current();
501}
502
503/*
504 * Object debugging
505 */
506
507/* Verify that a pointer has an address that is valid within a slab page */
508static inline int check_valid_pointer(struct kmem_cache *s,
509 struct page *page, void *object)
510{
511 void *base;
512
513 if (!object)
514 return 1;
515
516 base = page_address(page);
517 object = kasan_reset_tag(object);
518 object = restore_red_left(s, object);
519 if (object < base || object >= base + page->objects * s->size ||
520 (object - base) % s->size) {
521 return 0;
522 }
523
524 return 1;
525}
526
527static void print_section(char *level, char *text, u8 *addr,
528 unsigned int length)
529{
530 metadata_access_enable();
531 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
532 length, 1);
533 metadata_access_disable();
534}
535
536static struct track *get_track(struct kmem_cache *s, void *object,
537 enum track_item alloc)
538{
539 struct track *p;
540
541 if (s->offset)
542 p = object + s->offset + sizeof(void *);
543 else
544 p = object + s->inuse;
545
546 return p + alloc;
547}
548
549static void set_track(struct kmem_cache *s, void *object,
550 enum track_item alloc, unsigned long addr)
551{
552 struct track *p = get_track(s, object, alloc);
553
554 if (addr) {
555#ifdef CONFIG_STACKTRACE
556 unsigned int nr_entries;
557
558 metadata_access_enable();
559 nr_entries = stack_trace_save(p->addrs, TRACK_ADDRS_COUNT, 3);
560 metadata_access_disable();
561
562 if (nr_entries < TRACK_ADDRS_COUNT)
563 p->addrs[nr_entries] = 0;
564#endif
565 p->addr = addr;
566 p->cpu = smp_processor_id();
567 p->pid = current->pid;
568 p->when = jiffies;
569 } else {
570 memset(p, 0, sizeof(struct track));
571 }
572}
573
574static void init_tracking(struct kmem_cache *s, void *object)
575{
576 if (!(s->flags & SLAB_STORE_USER))
577 return;
578
579 set_track(s, object, TRACK_FREE, 0UL);
580 set_track(s, object, TRACK_ALLOC, 0UL);
581}
582
583static void print_track(const char *s, struct track *t, unsigned long pr_time)
584{
585 if (!t->addr)
586 return;
587
588 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
589 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
590#ifdef CONFIG_STACKTRACE
591 {
592 int i;
593 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
594 if (t->addrs[i])
595 pr_err("\t%pS\n", (void *)t->addrs[i]);
596 else
597 break;
598 }
599#endif
600}
601
602static void print_tracking(struct kmem_cache *s, void *object)
603{
604 unsigned long pr_time = jiffies;
605 if (!(s->flags & SLAB_STORE_USER))
606 return;
607
608 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
609 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
610}
611
612static void print_page_info(struct page *page)
613{
614 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
615 page, page->objects, page->inuse, page->freelist, page->flags);
616
617}
618
619static void slab_bug(struct kmem_cache *s, char *fmt, ...)
620{
621 struct va_format vaf;
622 va_list args;
623
624 va_start(args, fmt);
625 vaf.fmt = fmt;
626 vaf.va = &args;
627 pr_err("=============================================================================\n");
628 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
629 pr_err("-----------------------------------------------------------------------------\n\n");
630
631 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
632 va_end(args);
633}
634
635static void slab_fix(struct kmem_cache *s, char *fmt, ...)
636{
637 struct va_format vaf;
638 va_list args;
639
640 va_start(args, fmt);
641 vaf.fmt = fmt;
642 vaf.va = &args;
643 pr_err("FIX %s: %pV\n", s->name, &vaf);
644 va_end(args);
645}
646
647static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
648{
649 unsigned int off; /* Offset of last byte */
650 u8 *addr = page_address(page);
651
652 print_tracking(s, p);
653
654 print_page_info(page);
655
656 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
657 p, p - addr, get_freepointer(s, p));
658
659 if (s->flags & SLAB_RED_ZONE)
660 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
661 s->red_left_pad);
662 else if (p > addr + 16)
663 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
664
665 print_section(KERN_ERR, "Object ", p,
666 min_t(unsigned int, s->object_size, PAGE_SIZE));
667 if (s->flags & SLAB_RED_ZONE)
668 print_section(KERN_ERR, "Redzone ", p + s->object_size,
669 s->inuse - s->object_size);
670
671 if (s->offset)
672 off = s->offset + sizeof(void *);
673 else
674 off = s->inuse;
675
676 if (s->flags & SLAB_STORE_USER)
677 off += 2 * sizeof(struct track);
678
679 off += kasan_metadata_size(s);
680
681 if (off != size_from_object(s))
682 /* Beginning of the filler is the free pointer */
683 print_section(KERN_ERR, "Padding ", p + off,
684 size_from_object(s) - off);
685
686 dump_stack();
687}
688
689void object_err(struct kmem_cache *s, struct page *page,
690 u8 *object, char *reason)
691{
692 slab_bug(s, "%s", reason);
693 print_trailer(s, page, object);
694}
695
696static __printf(3, 4) void slab_err(struct kmem_cache *s, struct page *page,
697 const char *fmt, ...)
698{
699 va_list args;
700 char buf[100];
701
702 va_start(args, fmt);
703 vsnprintf(buf, sizeof(buf), fmt, args);
704 va_end(args);
705 slab_bug(s, "%s", buf);
706 print_page_info(page);
707 dump_stack();
708}
709
710static void init_object(struct kmem_cache *s, void *object, u8 val)
711{
712 u8 *p = object;
713
714 if (s->flags & SLAB_RED_ZONE)
715 memset(p - s->red_left_pad, val, s->red_left_pad);
716
717 if (s->flags & __OBJECT_POISON) {
718 memset(p, POISON_FREE, s->object_size - 1);
719 p[s->object_size - 1] = POISON_END;
720 }
721
722 if (s->flags & SLAB_RED_ZONE)
723 memset(p + s->object_size, val, s->inuse - s->object_size);
724}
725
726static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
727 void *from, void *to)
728{
729 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
730 memset(from, data, to - from);
731}
732
733static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
734 u8 *object, char *what,
735 u8 *start, unsigned int value, unsigned int bytes)
736{
737 u8 *fault;
738 u8 *end;
739
740 metadata_access_enable();
741 fault = memchr_inv(start, value, bytes);
742 metadata_access_disable();
743 if (!fault)
744 return 1;
745
746 end = start + bytes;
747 while (end > fault && end[-1] == value)
748 end--;
749
750 slab_bug(s, "%s overwritten", what);
751 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
752 fault, end - 1, fault[0], value);
753 print_trailer(s, page, object);
754
755 restore_bytes(s, what, value, fault, end);
756 return 0;
757}
758
759/*
760 * Object layout:
761 *
762 * object address
763 * Bytes of the object to be managed.
764 * If the freepointer may overlay the object then the free
765 * pointer is the first word of the object.
766 *
767 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
768 * 0xa5 (POISON_END)
769 *
770 * object + s->object_size
771 * Padding to reach word boundary. This is also used for Redzoning.
772 * Padding is extended by another word if Redzoning is enabled and
773 * object_size == inuse.
774 *
775 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
776 * 0xcc (RED_ACTIVE) for objects in use.
777 *
778 * object + s->inuse
779 * Meta data starts here.
780 *
781 * A. Free pointer (if we cannot overwrite object on free)
782 * B. Tracking data for SLAB_STORE_USER
783 * C. Padding to reach required alignment boundary or at mininum
784 * one word if debugging is on to be able to detect writes
785 * before the word boundary.
786 *
787 * Padding is done using 0x5a (POISON_INUSE)
788 *
789 * object + s->size
790 * Nothing is used beyond s->size.
791 *
792 * If slabcaches are merged then the object_size and inuse boundaries are mostly
793 * ignored. And therefore no slab options that rely on these boundaries
794 * may be used with merged slabcaches.
795 */
796
797static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
798{
799 unsigned long off = s->inuse; /* The end of info */
800
801 if (s->offset)
802 /* Freepointer is placed after the object. */
803 off += sizeof(void *);
804
805 if (s->flags & SLAB_STORE_USER)
806 /* We also have user information there */
807 off += 2 * sizeof(struct track);
808
809 off += kasan_metadata_size(s);
810
811 if (size_from_object(s) == off)
812 return 1;
813
814 return check_bytes_and_report(s, page, p, "Object padding",
815 p + off, POISON_INUSE, size_from_object(s) - off);
816}
817
818/* Check the pad bytes at the end of a slab page */
819static int slab_pad_check(struct kmem_cache *s, struct page *page)
820{
821 u8 *start;
822 u8 *fault;
823 u8 *end;
824 u8 *pad;
825 int length;
826 int remainder;
827
828 if (!(s->flags & SLAB_POISON))
829 return 1;
830
831 start = page_address(page);
832 length = page_size(page);
833 end = start + length;
834 remainder = length % s->size;
835 if (!remainder)
836 return 1;
837
838 pad = end - remainder;
839 metadata_access_enable();
840 fault = memchr_inv(pad, POISON_INUSE, remainder);
841 metadata_access_disable();
842 if (!fault)
843 return 1;
844 while (end > fault && end[-1] == POISON_INUSE)
845 end--;
846
847 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
848 print_section(KERN_ERR, "Padding ", pad, remainder);
849
850 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
851 return 0;
852}
853
854static int check_object(struct kmem_cache *s, struct page *page,
855 void *object, u8 val)
856{
857 u8 *p = object;
858 u8 *endobject = object + s->object_size;
859
860 if (s->flags & SLAB_RED_ZONE) {
861 if (!check_bytes_and_report(s, page, object, "Redzone",
862 object - s->red_left_pad, val, s->red_left_pad))
863 return 0;
864
865 if (!check_bytes_and_report(s, page, object, "Redzone",
866 endobject, val, s->inuse - s->object_size))
867 return 0;
868 } else {
869 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
870 check_bytes_and_report(s, page, p, "Alignment padding",
871 endobject, POISON_INUSE,
872 s->inuse - s->object_size);
873 }
874 }
875
876 if (s->flags & SLAB_POISON) {
877 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
878 (!check_bytes_and_report(s, page, p, "Poison", p,
879 POISON_FREE, s->object_size - 1) ||
880 !check_bytes_and_report(s, page, p, "Poison",
881 p + s->object_size - 1, POISON_END, 1)))
882 return 0;
883 /*
884 * check_pad_bytes cleans up on its own.
885 */
886 check_pad_bytes(s, page, p);
887 }
888
889 if (!s->offset && val == SLUB_RED_ACTIVE)
890 /*
891 * Object and freepointer overlap. Cannot check
892 * freepointer while object is allocated.
893 */
894 return 1;
895
896 /* Check free pointer validity */
897 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
898 object_err(s, page, p, "Freepointer corrupt");
899 /*
900 * No choice but to zap it and thus lose the remainder
901 * of the free objects in this slab. May cause
902 * another error because the object count is now wrong.
903 */
904 set_freepointer(s, p, NULL);
905 return 0;
906 }
907 return 1;
908}
909
910static int check_slab(struct kmem_cache *s, struct page *page)
911{
912 int maxobj;
913
914 VM_BUG_ON(!irqs_disabled());
915
916 if (!PageSlab(page)) {
917 slab_err(s, page, "Not a valid slab page");
918 return 0;
919 }
920
921 maxobj = order_objects(compound_order(page), s->size);
922 if (page->objects > maxobj) {
923 slab_err(s, page, "objects %u > max %u",
924 page->objects, maxobj);
925 return 0;
926 }
927 if (page->inuse > page->objects) {
928 slab_err(s, page, "inuse %u > max %u",
929 page->inuse, page->objects);
930 return 0;
931 }
932 /* Slab_pad_check fixes things up after itself */
933 slab_pad_check(s, page);
934 return 1;
935}
936
937/*
938 * Determine if a certain object on a page is on the freelist. Must hold the
939 * slab lock to guarantee that the chains are in a consistent state.
940 */
941static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
942{
943 int nr = 0;
944 void *fp;
945 void *object = NULL;
946 int max_objects;
947
948 fp = page->freelist;
949 while (fp && nr <= page->objects) {
950 if (fp == search)
951 return 1;
952 if (!check_valid_pointer(s, page, fp)) {
953 if (object) {
954 object_err(s, page, object,
955 "Freechain corrupt");
956 set_freepointer(s, object, NULL);
957 } else {
958 slab_err(s, page, "Freepointer corrupt");
959 page->freelist = NULL;
960 page->inuse = page->objects;
961 slab_fix(s, "Freelist cleared");
962 return 0;
963 }
964 break;
965 }
966 object = fp;
967 fp = get_freepointer(s, object);
968 nr++;
969 }
970
971 max_objects = order_objects(compound_order(page), s->size);
972 if (max_objects > MAX_OBJS_PER_PAGE)
973 max_objects = MAX_OBJS_PER_PAGE;
974
975 if (page->objects != max_objects) {
976 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
977 page->objects, max_objects);
978 page->objects = max_objects;
979 slab_fix(s, "Number of objects adjusted.");
980 }
981 if (page->inuse != page->objects - nr) {
982 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
983 page->inuse, page->objects - nr);
984 page->inuse = page->objects - nr;
985 slab_fix(s, "Object count adjusted.");
986 }
987 return search == NULL;
988}
989
990static void trace(struct kmem_cache *s, struct page *page, void *object,
991 int alloc)
992{
993 if (s->flags & SLAB_TRACE) {
994 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
995 s->name,
996 alloc ? "alloc" : "free",
997 object, page->inuse,
998 page->freelist);
999
1000 if (!alloc)
1001 print_section(KERN_INFO, "Object ", (void *)object,
1002 s->object_size);
1003
1004 dump_stack();
1005 }
1006}
1007
1008/*
1009 * Tracking of fully allocated slabs for debugging purposes.
1010 */
1011static void add_full(struct kmem_cache *s,
1012 struct kmem_cache_node *n, struct page *page)
1013{
1014 if (!(s->flags & SLAB_STORE_USER))
1015 return;
1016
1017 lockdep_assert_held(&n->list_lock);
1018 list_add(&page->slab_list, &n->full);
1019}
1020
1021static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1022{
1023 if (!(s->flags & SLAB_STORE_USER))
1024 return;
1025
1026 lockdep_assert_held(&n->list_lock);
1027 list_del(&page->slab_list);
1028}
1029
1030/* Tracking of the number of slabs for debugging purposes */
1031static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1032{
1033 struct kmem_cache_node *n = get_node(s, node);
1034
1035 return atomic_long_read(&n->nr_slabs);
1036}
1037
1038static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1039{
1040 return atomic_long_read(&n->nr_slabs);
1041}
1042
1043static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1044{
1045 struct kmem_cache_node *n = get_node(s, node);
1046
1047 /*
1048 * May be called early in order to allocate a slab for the
1049 * kmem_cache_node structure. Solve the chicken-egg
1050 * dilemma by deferring the increment of the count during
1051 * bootstrap (see early_kmem_cache_node_alloc).
1052 */
1053 if (likely(n)) {
1054 atomic_long_inc(&n->nr_slabs);
1055 atomic_long_add(objects, &n->total_objects);
1056 }
1057}
1058static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1059{
1060 struct kmem_cache_node *n = get_node(s, node);
1061
1062 atomic_long_dec(&n->nr_slabs);
1063 atomic_long_sub(objects, &n->total_objects);
1064}
1065
1066/* Object debug checks for alloc/free paths */
1067static void setup_object_debug(struct kmem_cache *s, struct page *page,
1068 void *object)
1069{
1070 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1071 return;
1072
1073 init_object(s, object, SLUB_RED_INACTIVE);
1074 init_tracking(s, object);
1075}
1076
1077static
1078void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr)
1079{
1080 if (!(s->flags & SLAB_POISON))
1081 return;
1082
1083 metadata_access_enable();
1084 memset(addr, POISON_INUSE, page_size(page));
1085 metadata_access_disable();
1086}
1087
1088static inline int alloc_consistency_checks(struct kmem_cache *s,
1089 struct page *page, void *object)
1090{
1091 if (!check_slab(s, page))
1092 return 0;
1093
1094 if (!check_valid_pointer(s, page, object)) {
1095 object_err(s, page, object, "Freelist Pointer check fails");
1096 return 0;
1097 }
1098
1099 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1100 return 0;
1101
1102 return 1;
1103}
1104
1105static noinline int alloc_debug_processing(struct kmem_cache *s,
1106 struct page *page,
1107 void *object, unsigned long addr)
1108{
1109 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1110 if (!alloc_consistency_checks(s, page, object))
1111 goto bad;
1112 }
1113
1114 /* Success perform special debug activities for allocs */
1115 if (s->flags & SLAB_STORE_USER)
1116 set_track(s, object, TRACK_ALLOC, addr);
1117 trace(s, page, object, 1);
1118 init_object(s, object, SLUB_RED_ACTIVE);
1119 return 1;
1120
1121bad:
1122 if (PageSlab(page)) {
1123 /*
1124 * If this is a slab page then lets do the best we can
1125 * to avoid issues in the future. Marking all objects
1126 * as used avoids touching the remaining objects.
1127 */
1128 slab_fix(s, "Marking all objects used");
1129 page->inuse = page->objects;
1130 page->freelist = NULL;
1131 }
1132 return 0;
1133}
1134
1135static inline int free_consistency_checks(struct kmem_cache *s,
1136 struct page *page, void *object, unsigned long addr)
1137{
1138 if (!check_valid_pointer(s, page, object)) {
1139 slab_err(s, page, "Invalid object pointer 0x%p", object);
1140 return 0;
1141 }
1142
1143 if (on_freelist(s, page, object)) {
1144 object_err(s, page, object, "Object already free");
1145 return 0;
1146 }
1147
1148 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1149 return 0;
1150
1151 if (unlikely(s != page->slab_cache)) {
1152 if (!PageSlab(page)) {
1153 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1154 object);
1155 } else if (!page->slab_cache) {
1156 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1157 object);
1158 dump_stack();
1159 } else
1160 object_err(s, page, object,
1161 "page slab pointer corrupt.");
1162 return 0;
1163 }
1164 return 1;
1165}
1166
1167/* Supports checking bulk free of a constructed freelist */
1168static noinline int free_debug_processing(
1169 struct kmem_cache *s, struct page *page,
1170 void *head, void *tail, int bulk_cnt,
1171 unsigned long addr)
1172{
1173 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1174 void *object = head;
1175 int cnt = 0;
1176 unsigned long uninitialized_var(flags);
1177 int ret = 0;
1178
1179 spin_lock_irqsave(&n->list_lock, flags);
1180 slab_lock(page);
1181
1182 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1183 if (!check_slab(s, page))
1184 goto out;
1185 }
1186
1187next_object:
1188 cnt++;
1189
1190 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1191 if (!free_consistency_checks(s, page, object, addr))
1192 goto out;
1193 }
1194
1195 if (s->flags & SLAB_STORE_USER)
1196 set_track(s, object, TRACK_FREE, addr);
1197 trace(s, page, object, 0);
1198 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1199 init_object(s, object, SLUB_RED_INACTIVE);
1200
1201 /* Reached end of constructed freelist yet? */
1202 if (object != tail) {
1203 object = get_freepointer(s, object);
1204 goto next_object;
1205 }
1206 ret = 1;
1207
1208out:
1209 if (cnt != bulk_cnt)
1210 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1211 bulk_cnt, cnt);
1212
1213 slab_unlock(page);
1214 spin_unlock_irqrestore(&n->list_lock, flags);
1215 if (!ret)
1216 slab_fix(s, "Object at 0x%p not freed", object);
1217 return ret;
1218}
1219
1220static int __init setup_slub_debug(char *str)
1221{
1222 slub_debug = DEBUG_DEFAULT_FLAGS;
1223 if (*str++ != '=' || !*str)
1224 /*
1225 * No options specified. Switch on full debugging.
1226 */
1227 goto out;
1228
1229 if (*str == ',')
1230 /*
1231 * No options but restriction on slabs. This means full
1232 * debugging for slabs matching a pattern.
1233 */
1234 goto check_slabs;
1235
1236 slub_debug = 0;
1237 if (*str == '-')
1238 /*
1239 * Switch off all debugging measures.
1240 */
1241 goto out;
1242
1243 /*
1244 * Determine which debug features should be switched on
1245 */
1246 for (; *str && *str != ','; str++) {
1247 switch (tolower(*str)) {
1248 case 'f':
1249 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1250 break;
1251 case 'z':
1252 slub_debug |= SLAB_RED_ZONE;
1253 break;
1254 case 'p':
1255 slub_debug |= SLAB_POISON;
1256 break;
1257 case 'u':
1258 slub_debug |= SLAB_STORE_USER;
1259 break;
1260 case 't':
1261 slub_debug |= SLAB_TRACE;
1262 break;
1263 case 'a':
1264 slub_debug |= SLAB_FAILSLAB;
1265 break;
1266 case 'o':
1267 /*
1268 * Avoid enabling debugging on caches if its minimum
1269 * order would increase as a result.
1270 */
1271 disable_higher_order_debug = 1;
1272 break;
1273 default:
1274 pr_err("slub_debug option '%c' unknown. skipped\n",
1275 *str);
1276 }
1277 }
1278
1279check_slabs:
1280 if (*str == ',')
1281 slub_debug_slabs = str + 1;
1282out:
1283 if ((static_branch_unlikely(&init_on_alloc) ||
1284 static_branch_unlikely(&init_on_free)) &&
1285 (slub_debug & SLAB_POISON))
1286 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1287 return 1;
1288}
1289
1290__setup("slub_debug", setup_slub_debug);
1291
1292/*
1293 * kmem_cache_flags - apply debugging options to the cache
1294 * @object_size: the size of an object without meta data
1295 * @flags: flags to set
1296 * @name: name of the cache
1297 * @ctor: constructor function
1298 *
1299 * Debug option(s) are applied to @flags. In addition to the debug
1300 * option(s), if a slab name (or multiple) is specified i.e.
1301 * slub_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1302 * then only the select slabs will receive the debug option(s).
1303 */
1304slab_flags_t kmem_cache_flags(unsigned int object_size,
1305 slab_flags_t flags, const char *name,
1306 void (*ctor)(void *))
1307{
1308 char *iter;
1309 size_t len;
1310
1311 /* If slub_debug = 0, it folds into the if conditional. */
1312 if (!slub_debug_slabs)
1313 return flags | slub_debug;
1314
1315 len = strlen(name);
1316 iter = slub_debug_slabs;
1317 while (*iter) {
1318 char *end, *glob;
1319 size_t cmplen;
1320
1321 end = strchrnul(iter, ',');
1322
1323 glob = strnchr(iter, end - iter, '*');
1324 if (glob)
1325 cmplen = glob - iter;
1326 else
1327 cmplen = max_t(size_t, len, (end - iter));
1328
1329 if (!strncmp(name, iter, cmplen)) {
1330 flags |= slub_debug;
1331 break;
1332 }
1333
1334 if (!*end)
1335 break;
1336 iter = end + 1;
1337 }
1338
1339 return flags;
1340}
1341#else /* !CONFIG_SLUB_DEBUG */
1342static inline void setup_object_debug(struct kmem_cache *s,
1343 struct page *page, void *object) {}
1344static inline
1345void setup_page_debug(struct kmem_cache *s, struct page *page, void *addr) {}
1346
1347static inline int alloc_debug_processing(struct kmem_cache *s,
1348 struct page *page, void *object, unsigned long addr) { return 0; }
1349
1350static inline int free_debug_processing(
1351 struct kmem_cache *s, struct page *page,
1352 void *head, void *tail, int bulk_cnt,
1353 unsigned long addr) { return 0; }
1354
1355static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1356 { return 1; }
1357static inline int check_object(struct kmem_cache *s, struct page *page,
1358 void *object, u8 val) { return 1; }
1359static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1360 struct page *page) {}
1361static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1362 struct page *page) {}
1363slab_flags_t kmem_cache_flags(unsigned int object_size,
1364 slab_flags_t flags, const char *name,
1365 void (*ctor)(void *))
1366{
1367 return flags;
1368}
1369#define slub_debug 0
1370
1371#define disable_higher_order_debug 0
1372
1373static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1374 { return 0; }
1375static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1376 { return 0; }
1377static inline void inc_slabs_node(struct kmem_cache *s, int node,
1378 int objects) {}
1379static inline void dec_slabs_node(struct kmem_cache *s, int node,
1380 int objects) {}
1381
1382#endif /* CONFIG_SLUB_DEBUG */
1383
1384/*
1385 * Hooks for other subsystems that check memory allocations. In a typical
1386 * production configuration these hooks all should produce no code at all.
1387 */
1388static inline void *kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1389{
1390 ptr = kasan_kmalloc_large(ptr, size, flags);
1391 /* As ptr might get tagged, call kmemleak hook after KASAN. */
1392 kmemleak_alloc(ptr, size, 1, flags);
1393 return ptr;
1394}
1395
1396static __always_inline void kfree_hook(void *x)
1397{
1398 kmemleak_free(x);
1399 kasan_kfree_large(x, _RET_IP_);
1400}
1401
1402static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1403{
1404 kmemleak_free_recursive(x, s->flags);
1405
1406 /*
1407 * Trouble is that we may no longer disable interrupts in the fast path
1408 * So in order to make the debug calls that expect irqs to be
1409 * disabled we need to disable interrupts temporarily.
1410 */
1411#ifdef CONFIG_LOCKDEP
1412 {
1413 unsigned long flags;
1414
1415 local_irq_save(flags);
1416 debug_check_no_locks_freed(x, s->object_size);
1417 local_irq_restore(flags);
1418 }
1419#endif
1420 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1421 debug_check_no_obj_freed(x, s->object_size);
1422
1423 /* KASAN might put x into memory quarantine, delaying its reuse */
1424 return kasan_slab_free(s, x, _RET_IP_);
1425}
1426
1427static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1428 void **head, void **tail)
1429{
1430
1431 void *object;
1432 void *next = *head;
1433 void *old_tail = *tail ? *tail : *head;
1434 int rsize;
1435
1436 /* Head and tail of the reconstructed freelist */
1437 *head = NULL;
1438 *tail = NULL;
1439
1440 do {
1441 object = next;
1442 next = get_freepointer(s, object);
1443
1444 if (slab_want_init_on_free(s)) {
1445 /*
1446 * Clear the object and the metadata, but don't touch
1447 * the redzone.
1448 */
1449 memset(object, 0, s->object_size);
1450 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad
1451 : 0;
1452 memset((char *)object + s->inuse, 0,
1453 s->size - s->inuse - rsize);
1454
1455 }
1456 /* If object's reuse doesn't have to be delayed */
1457 if (!slab_free_hook(s, object)) {
1458 /* Move object to the new freelist */
1459 set_freepointer(s, object, *head);
1460 *head = object;
1461 if (!*tail)
1462 *tail = object;
1463 }
1464 } while (object != old_tail);
1465
1466 if (*head == *tail)
1467 *tail = NULL;
1468
1469 return *head != NULL;
1470}
1471
1472static void *setup_object(struct kmem_cache *s, struct page *page,
1473 void *object)
1474{
1475 setup_object_debug(s, page, object);
1476 object = kasan_init_slab_obj(s, object);
1477 if (unlikely(s->ctor)) {
1478 kasan_unpoison_object_data(s, object);
1479 s->ctor(object);
1480 kasan_poison_object_data(s, object);
1481 }
1482 return object;
1483}
1484
1485/*
1486 * Slab allocation and freeing
1487 */
1488static inline struct page *alloc_slab_page(struct kmem_cache *s,
1489 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1490{
1491 struct page *page;
1492 unsigned int order = oo_order(oo);
1493
1494 if (node == NUMA_NO_NODE)
1495 page = alloc_pages(flags, order);
1496 else
1497 page = __alloc_pages_node(node, flags, order);
1498
1499 if (page && charge_slab_page(page, flags, order, s)) {
1500 __free_pages(page, order);
1501 page = NULL;
1502 }
1503
1504 return page;
1505}
1506
1507#ifdef CONFIG_SLAB_FREELIST_RANDOM
1508/* Pre-initialize the random sequence cache */
1509static int init_cache_random_seq(struct kmem_cache *s)
1510{
1511 unsigned int count = oo_objects(s->oo);
1512 int err;
1513
1514 /* Bailout if already initialised */
1515 if (s->random_seq)
1516 return 0;
1517
1518 err = cache_random_seq_create(s, count, GFP_KERNEL);
1519 if (err) {
1520 pr_err("SLUB: Unable to initialize free list for %s\n",
1521 s->name);
1522 return err;
1523 }
1524
1525 /* Transform to an offset on the set of pages */
1526 if (s->random_seq) {
1527 unsigned int i;
1528
1529 for (i = 0; i < count; i++)
1530 s->random_seq[i] *= s->size;
1531 }
1532 return 0;
1533}
1534
1535/* Initialize each random sequence freelist per cache */
1536static void __init init_freelist_randomization(void)
1537{
1538 struct kmem_cache *s;
1539
1540 mutex_lock(&slab_mutex);
1541
1542 list_for_each_entry(s, &slab_caches, list)
1543 init_cache_random_seq(s);
1544
1545 mutex_unlock(&slab_mutex);
1546}
1547
1548/* Get the next entry on the pre-computed freelist randomized */
1549static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1550 unsigned long *pos, void *start,
1551 unsigned long page_limit,
1552 unsigned long freelist_count)
1553{
1554 unsigned int idx;
1555
1556 /*
1557 * If the target page allocation failed, the number of objects on the
1558 * page might be smaller than the usual size defined by the cache.
1559 */
1560 do {
1561 idx = s->random_seq[*pos];
1562 *pos += 1;
1563 if (*pos >= freelist_count)
1564 *pos = 0;
1565 } while (unlikely(idx >= page_limit));
1566
1567 return (char *)start + idx;
1568}
1569
1570/* Shuffle the single linked freelist based on a random pre-computed sequence */
1571static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1572{
1573 void *start;
1574 void *cur;
1575 void *next;
1576 unsigned long idx, pos, page_limit, freelist_count;
1577
1578 if (page->objects < 2 || !s->random_seq)
1579 return false;
1580
1581 freelist_count = oo_objects(s->oo);
1582 pos = get_random_int() % freelist_count;
1583
1584 page_limit = page->objects * s->size;
1585 start = fixup_red_left(s, page_address(page));
1586
1587 /* First entry is used as the base of the freelist */
1588 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1589 freelist_count);
1590 cur = setup_object(s, page, cur);
1591 page->freelist = cur;
1592
1593 for (idx = 1; idx < page->objects; idx++) {
1594 next = next_freelist_entry(s, page, &pos, start, page_limit,
1595 freelist_count);
1596 next = setup_object(s, page, next);
1597 set_freepointer(s, cur, next);
1598 cur = next;
1599 }
1600 set_freepointer(s, cur, NULL);
1601
1602 return true;
1603}
1604#else
1605static inline int init_cache_random_seq(struct kmem_cache *s)
1606{
1607 return 0;
1608}
1609static inline void init_freelist_randomization(void) { }
1610static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1611{
1612 return false;
1613}
1614#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1615
1616static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1617{
1618 struct page *page;
1619 struct kmem_cache_order_objects oo = s->oo;
1620 gfp_t alloc_gfp;
1621 void *start, *p, *next;
1622 int idx;
1623 bool shuffle;
1624
1625 flags &= gfp_allowed_mask;
1626
1627 if (gfpflags_allow_blocking(flags))
1628 local_irq_enable();
1629
1630 flags |= s->allocflags;
1631
1632 /*
1633 * Let the initial higher-order allocation fail under memory pressure
1634 * so we fall-back to the minimum order allocation.
1635 */
1636 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1637 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1638 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1639
1640 page = alloc_slab_page(s, alloc_gfp, node, oo);
1641 if (unlikely(!page)) {
1642 oo = s->min;
1643 alloc_gfp = flags;
1644 /*
1645 * Allocation may have failed due to fragmentation.
1646 * Try a lower order alloc if possible
1647 */
1648 page = alloc_slab_page(s, alloc_gfp, node, oo);
1649 if (unlikely(!page))
1650 goto out;
1651 stat(s, ORDER_FALLBACK);
1652 }
1653
1654 page->objects = oo_objects(oo);
1655
1656 page->slab_cache = s;
1657 __SetPageSlab(page);
1658 if (page_is_pfmemalloc(page))
1659 SetPageSlabPfmemalloc(page);
1660
1661 kasan_poison_slab(page);
1662
1663 start = page_address(page);
1664
1665 setup_page_debug(s, page, start);
1666
1667 shuffle = shuffle_freelist(s, page);
1668
1669 if (!shuffle) {
1670 start = fixup_red_left(s, start);
1671 start = setup_object(s, page, start);
1672 page->freelist = start;
1673 for (idx = 0, p = start; idx < page->objects - 1; idx++) {
1674 next = p + s->size;
1675 next = setup_object(s, page, next);
1676 set_freepointer(s, p, next);
1677 p = next;
1678 }
1679 set_freepointer(s, p, NULL);
1680 }
1681
1682 page->inuse = page->objects;
1683 page->frozen = 1;
1684
1685out:
1686 if (gfpflags_allow_blocking(flags))
1687 local_irq_disable();
1688 if (!page)
1689 return NULL;
1690
1691 inc_slabs_node(s, page_to_nid(page), page->objects);
1692
1693 return page;
1694}
1695
1696static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1697{
1698 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1699 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1700 flags &= ~GFP_SLAB_BUG_MASK;
1701 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1702 invalid_mask, &invalid_mask, flags, &flags);
1703 dump_stack();
1704 }
1705
1706 return allocate_slab(s,
1707 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1708}
1709
1710static void __free_slab(struct kmem_cache *s, struct page *page)
1711{
1712 int order = compound_order(page);
1713 int pages = 1 << order;
1714
1715 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1716 void *p;
1717
1718 slab_pad_check(s, page);
1719 for_each_object(p, s, page_address(page),
1720 page->objects)
1721 check_object(s, page, p, SLUB_RED_INACTIVE);
1722 }
1723
1724 __ClearPageSlabPfmemalloc(page);
1725 __ClearPageSlab(page);
1726
1727 page->mapping = NULL;
1728 if (current->reclaim_state)
1729 current->reclaim_state->reclaimed_slab += pages;
1730 uncharge_slab_page(page, order, s);
1731 __free_pages(page, order);
1732}
1733
1734static void rcu_free_slab(struct rcu_head *h)
1735{
1736 struct page *page = container_of(h, struct page, rcu_head);
1737
1738 __free_slab(page->slab_cache, page);
1739}
1740
1741static void free_slab(struct kmem_cache *s, struct page *page)
1742{
1743 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1744 call_rcu(&page->rcu_head, rcu_free_slab);
1745 } else
1746 __free_slab(s, page);
1747}
1748
1749static void discard_slab(struct kmem_cache *s, struct page *page)
1750{
1751 dec_slabs_node(s, page_to_nid(page), page->objects);
1752 free_slab(s, page);
1753}
1754
1755/*
1756 * Management of partially allocated slabs.
1757 */
1758static inline void
1759__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1760{
1761 n->nr_partial++;
1762 if (tail == DEACTIVATE_TO_TAIL)
1763 list_add_tail(&page->slab_list, &n->partial);
1764 else
1765 list_add(&page->slab_list, &n->partial);
1766}
1767
1768static inline void add_partial(struct kmem_cache_node *n,
1769 struct page *page, int tail)
1770{
1771 lockdep_assert_held(&n->list_lock);
1772 __add_partial(n, page, tail);
1773}
1774
1775static inline void remove_partial(struct kmem_cache_node *n,
1776 struct page *page)
1777{
1778 lockdep_assert_held(&n->list_lock);
1779 list_del(&page->slab_list);
1780 n->nr_partial--;
1781}
1782
1783/*
1784 * Remove slab from the partial list, freeze it and
1785 * return the pointer to the freelist.
1786 *
1787 * Returns a list of objects or NULL if it fails.
1788 */
1789static inline void *acquire_slab(struct kmem_cache *s,
1790 struct kmem_cache_node *n, struct page *page,
1791 int mode, int *objects)
1792{
1793 void *freelist;
1794 unsigned long counters;
1795 struct page new;
1796
1797 lockdep_assert_held(&n->list_lock);
1798
1799 /*
1800 * Zap the freelist and set the frozen bit.
1801 * The old freelist is the list of objects for the
1802 * per cpu allocation list.
1803 */
1804 freelist = page->freelist;
1805 counters = page->counters;
1806 new.counters = counters;
1807 *objects = new.objects - new.inuse;
1808 if (mode) {
1809 new.inuse = page->objects;
1810 new.freelist = NULL;
1811 } else {
1812 new.freelist = freelist;
1813 }
1814
1815 VM_BUG_ON(new.frozen);
1816 new.frozen = 1;
1817
1818 if (!__cmpxchg_double_slab(s, page,
1819 freelist, counters,
1820 new.freelist, new.counters,
1821 "acquire_slab"))
1822 return NULL;
1823
1824 remove_partial(n, page);
1825 WARN_ON(!freelist);
1826 return freelist;
1827}
1828
1829static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1830static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1831
1832/*
1833 * Try to allocate a partial slab from a specific node.
1834 */
1835static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1836 struct kmem_cache_cpu *c, gfp_t flags)
1837{
1838 struct page *page, *page2;
1839 void *object = NULL;
1840 unsigned int available = 0;
1841 int objects;
1842
1843 /*
1844 * Racy check. If we mistakenly see no partial slabs then we
1845 * just allocate an empty slab. If we mistakenly try to get a
1846 * partial slab and there is none available then get_partials()
1847 * will return NULL.
1848 */
1849 if (!n || !n->nr_partial)
1850 return NULL;
1851
1852 spin_lock(&n->list_lock);
1853 list_for_each_entry_safe(page, page2, &n->partial, slab_list) {
1854 void *t;
1855
1856 if (!pfmemalloc_match(page, flags))
1857 continue;
1858
1859 t = acquire_slab(s, n, page, object == NULL, &objects);
1860 if (!t)
1861 break;
1862
1863 available += objects;
1864 if (!object) {
1865 c->page = page;
1866 stat(s, ALLOC_FROM_PARTIAL);
1867 object = t;
1868 } else {
1869 put_cpu_partial(s, page, 0);
1870 stat(s, CPU_PARTIAL_NODE);
1871 }
1872 if (!kmem_cache_has_cpu_partial(s)
1873 || available > slub_cpu_partial(s) / 2)
1874 break;
1875
1876 }
1877 spin_unlock(&n->list_lock);
1878 return object;
1879}
1880
1881/*
1882 * Get a page from somewhere. Search in increasing NUMA distances.
1883 */
1884static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1885 struct kmem_cache_cpu *c)
1886{
1887#ifdef CONFIG_NUMA
1888 struct zonelist *zonelist;
1889 struct zoneref *z;
1890 struct zone *zone;
1891 enum zone_type high_zoneidx = gfp_zone(flags);
1892 void *object;
1893 unsigned int cpuset_mems_cookie;
1894
1895 /*
1896 * The defrag ratio allows a configuration of the tradeoffs between
1897 * inter node defragmentation and node local allocations. A lower
1898 * defrag_ratio increases the tendency to do local allocations
1899 * instead of attempting to obtain partial slabs from other nodes.
1900 *
1901 * If the defrag_ratio is set to 0 then kmalloc() always
1902 * returns node local objects. If the ratio is higher then kmalloc()
1903 * may return off node objects because partial slabs are obtained
1904 * from other nodes and filled up.
1905 *
1906 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1907 * (which makes defrag_ratio = 1000) then every (well almost)
1908 * allocation will first attempt to defrag slab caches on other nodes.
1909 * This means scanning over all nodes to look for partial slabs which
1910 * may be expensive if we do it every time we are trying to find a slab
1911 * with available objects.
1912 */
1913 if (!s->remote_node_defrag_ratio ||
1914 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1915 return NULL;
1916
1917 do {
1918 cpuset_mems_cookie = read_mems_allowed_begin();
1919 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1920 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1921 struct kmem_cache_node *n;
1922
1923 n = get_node(s, zone_to_nid(zone));
1924
1925 if (n && cpuset_zone_allowed(zone, flags) &&
1926 n->nr_partial > s->min_partial) {
1927 object = get_partial_node(s, n, c, flags);
1928 if (object) {
1929 /*
1930 * Don't check read_mems_allowed_retry()
1931 * here - if mems_allowed was updated in
1932 * parallel, that was a harmless race
1933 * between allocation and the cpuset
1934 * update
1935 */
1936 return object;
1937 }
1938 }
1939 }
1940 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1941#endif /* CONFIG_NUMA */
1942 return NULL;
1943}
1944
1945/*
1946 * Get a partial page, lock it and return it.
1947 */
1948static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1949 struct kmem_cache_cpu *c)
1950{
1951 void *object;
1952 int searchnode = node;
1953
1954 if (node == NUMA_NO_NODE)
1955 searchnode = numa_mem_id();
1956 else if (!node_present_pages(node))
1957 searchnode = node_to_mem_node(node);
1958
1959 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1960 if (object || node != NUMA_NO_NODE)
1961 return object;
1962
1963 return get_any_partial(s, flags, c);
1964}
1965
1966#ifdef CONFIG_PREEMPT
1967/*
1968 * Calculate the next globally unique transaction for disambiguiation
1969 * during cmpxchg. The transactions start with the cpu number and are then
1970 * incremented by CONFIG_NR_CPUS.
1971 */
1972#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1973#else
1974/*
1975 * No preemption supported therefore also no need to check for
1976 * different cpus.
1977 */
1978#define TID_STEP 1
1979#endif
1980
1981static inline unsigned long next_tid(unsigned long tid)
1982{
1983 return tid + TID_STEP;
1984}
1985
1986#ifdef SLUB_DEBUG_CMPXCHG
1987static inline unsigned int tid_to_cpu(unsigned long tid)
1988{
1989 return tid % TID_STEP;
1990}
1991
1992static inline unsigned long tid_to_event(unsigned long tid)
1993{
1994 return tid / TID_STEP;
1995}
1996#endif
1997
1998static inline unsigned int init_tid(int cpu)
1999{
2000 return cpu;
2001}
2002
2003static inline void note_cmpxchg_failure(const char *n,
2004 const struct kmem_cache *s, unsigned long tid)
2005{
2006#ifdef SLUB_DEBUG_CMPXCHG
2007 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
2008
2009 pr_info("%s %s: cmpxchg redo ", n, s->name);
2010
2011#ifdef CONFIG_PREEMPT
2012 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2013 pr_warn("due to cpu change %d -> %d\n",
2014 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2015 else
2016#endif
2017 if (tid_to_event(tid) != tid_to_event(actual_tid))
2018 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2019 tid_to_event(tid), tid_to_event(actual_tid));
2020 else
2021 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2022 actual_tid, tid, next_tid(tid));
2023#endif
2024 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2025}
2026
2027static void init_kmem_cache_cpus(struct kmem_cache *s)
2028{
2029 int cpu;
2030
2031 for_each_possible_cpu(cpu)
2032 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2033}
2034
2035/*
2036 * Remove the cpu slab
2037 */
2038static void deactivate_slab(struct kmem_cache *s, struct page *page,
2039 void *freelist, struct kmem_cache_cpu *c)
2040{
2041 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2042 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2043 int lock = 0;
2044 enum slab_modes l = M_NONE, m = M_NONE;
2045 void *nextfree;
2046 int tail = DEACTIVATE_TO_HEAD;
2047 struct page new;
2048 struct page old;
2049
2050 if (page->freelist) {
2051 stat(s, DEACTIVATE_REMOTE_FREES);
2052 tail = DEACTIVATE_TO_TAIL;
2053 }
2054
2055 /*
2056 * Stage one: Free all available per cpu objects back
2057 * to the page freelist while it is still frozen. Leave the
2058 * last one.
2059 *
2060 * There is no need to take the list->lock because the page
2061 * is still frozen.
2062 */
2063 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2064 void *prior;
2065 unsigned long counters;
2066
2067 do {
2068 prior = page->freelist;
2069 counters = page->counters;
2070 set_freepointer(s, freelist, prior);
2071 new.counters = counters;
2072 new.inuse--;
2073 VM_BUG_ON(!new.frozen);
2074
2075 } while (!__cmpxchg_double_slab(s, page,
2076 prior, counters,
2077 freelist, new.counters,
2078 "drain percpu freelist"));
2079
2080 freelist = nextfree;
2081 }
2082
2083 /*
2084 * Stage two: Ensure that the page is unfrozen while the
2085 * list presence reflects the actual number of objects
2086 * during unfreeze.
2087 *
2088 * We setup the list membership and then perform a cmpxchg
2089 * with the count. If there is a mismatch then the page
2090 * is not unfrozen but the page is on the wrong list.
2091 *
2092 * Then we restart the process which may have to remove
2093 * the page from the list that we just put it on again
2094 * because the number of objects in the slab may have
2095 * changed.
2096 */
2097redo:
2098
2099 old.freelist = page->freelist;
2100 old.counters = page->counters;
2101 VM_BUG_ON(!old.frozen);
2102
2103 /* Determine target state of the slab */
2104 new.counters = old.counters;
2105 if (freelist) {
2106 new.inuse--;
2107 set_freepointer(s, freelist, old.freelist);
2108 new.freelist = freelist;
2109 } else
2110 new.freelist = old.freelist;
2111
2112 new.frozen = 0;
2113
2114 if (!new.inuse && n->nr_partial >= s->min_partial)
2115 m = M_FREE;
2116 else if (new.freelist) {
2117 m = M_PARTIAL;
2118 if (!lock) {
2119 lock = 1;
2120 /*
2121 * Taking the spinlock removes the possibility
2122 * that acquire_slab() will see a slab page that
2123 * is frozen
2124 */
2125 spin_lock(&n->list_lock);
2126 }
2127 } else {
2128 m = M_FULL;
2129 if (kmem_cache_debug(s) && !lock) {
2130 lock = 1;
2131 /*
2132 * This also ensures that the scanning of full
2133 * slabs from diagnostic functions will not see
2134 * any frozen slabs.
2135 */
2136 spin_lock(&n->list_lock);
2137 }
2138 }
2139
2140 if (l != m) {
2141 if (l == M_PARTIAL)
2142 remove_partial(n, page);
2143 else if (l == M_FULL)
2144 remove_full(s, n, page);
2145
2146 if (m == M_PARTIAL)
2147 add_partial(n, page, tail);
2148 else if (m == M_FULL)
2149 add_full(s, n, page);
2150 }
2151
2152 l = m;
2153 if (!__cmpxchg_double_slab(s, page,
2154 old.freelist, old.counters,
2155 new.freelist, new.counters,
2156 "unfreezing slab"))
2157 goto redo;
2158
2159 if (lock)
2160 spin_unlock(&n->list_lock);
2161
2162 if (m == M_PARTIAL)
2163 stat(s, tail);
2164 else if (m == M_FULL)
2165 stat(s, DEACTIVATE_FULL);
2166 else if (m == M_FREE) {
2167 stat(s, DEACTIVATE_EMPTY);
2168 discard_slab(s, page);
2169 stat(s, FREE_SLAB);
2170 }
2171
2172 c->page = NULL;
2173 c->freelist = NULL;
2174}
2175
2176/*
2177 * Unfreeze all the cpu partial slabs.
2178 *
2179 * This function must be called with interrupts disabled
2180 * for the cpu using c (or some other guarantee must be there
2181 * to guarantee no concurrent accesses).
2182 */
2183static void unfreeze_partials(struct kmem_cache *s,
2184 struct kmem_cache_cpu *c)
2185{
2186#ifdef CONFIG_SLUB_CPU_PARTIAL
2187 struct kmem_cache_node *n = NULL, *n2 = NULL;
2188 struct page *page, *discard_page = NULL;
2189
2190 while ((page = c->partial)) {
2191 struct page new;
2192 struct page old;
2193
2194 c->partial = page->next;
2195
2196 n2 = get_node(s, page_to_nid(page));
2197 if (n != n2) {
2198 if (n)
2199 spin_unlock(&n->list_lock);
2200
2201 n = n2;
2202 spin_lock(&n->list_lock);
2203 }
2204
2205 do {
2206
2207 old.freelist = page->freelist;
2208 old.counters = page->counters;
2209 VM_BUG_ON(!old.frozen);
2210
2211 new.counters = old.counters;
2212 new.freelist = old.freelist;
2213
2214 new.frozen = 0;
2215
2216 } while (!__cmpxchg_double_slab(s, page,
2217 old.freelist, old.counters,
2218 new.freelist, new.counters,
2219 "unfreezing slab"));
2220
2221 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2222 page->next = discard_page;
2223 discard_page = page;
2224 } else {
2225 add_partial(n, page, DEACTIVATE_TO_TAIL);
2226 stat(s, FREE_ADD_PARTIAL);
2227 }
2228 }
2229
2230 if (n)
2231 spin_unlock(&n->list_lock);
2232
2233 while (discard_page) {
2234 page = discard_page;
2235 discard_page = discard_page->next;
2236
2237 stat(s, DEACTIVATE_EMPTY);
2238 discard_slab(s, page);
2239 stat(s, FREE_SLAB);
2240 }
2241#endif /* CONFIG_SLUB_CPU_PARTIAL */
2242}
2243
2244/*
2245 * Put a page that was just frozen (in __slab_free|get_partial_node) into a
2246 * partial page slot if available.
2247 *
2248 * If we did not find a slot then simply move all the partials to the
2249 * per node partial list.
2250 */
2251static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2252{
2253#ifdef CONFIG_SLUB_CPU_PARTIAL
2254 struct page *oldpage;
2255 int pages;
2256 int pobjects;
2257
2258 preempt_disable();
2259 do {
2260 pages = 0;
2261 pobjects = 0;
2262 oldpage = this_cpu_read(s->cpu_slab->partial);
2263
2264 if (oldpage) {
2265 pobjects = oldpage->pobjects;
2266 pages = oldpage->pages;
2267 if (drain && pobjects > s->cpu_partial) {
2268 unsigned long flags;
2269 /*
2270 * partial array is full. Move the existing
2271 * set to the per node partial list.
2272 */
2273 local_irq_save(flags);
2274 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2275 local_irq_restore(flags);
2276 oldpage = NULL;
2277 pobjects = 0;
2278 pages = 0;
2279 stat(s, CPU_PARTIAL_DRAIN);
2280 }
2281 }
2282
2283 pages++;
2284 pobjects += page->objects - page->inuse;
2285
2286 page->pages = pages;
2287 page->pobjects = pobjects;
2288 page->next = oldpage;
2289
2290 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2291 != oldpage);
2292 if (unlikely(!s->cpu_partial)) {
2293 unsigned long flags;
2294
2295 local_irq_save(flags);
2296 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2297 local_irq_restore(flags);
2298 }
2299 preempt_enable();
2300#endif /* CONFIG_SLUB_CPU_PARTIAL */
2301}
2302
2303static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2304{
2305 stat(s, CPUSLAB_FLUSH);
2306 deactivate_slab(s, c->page, c->freelist, c);
2307
2308 c->tid = next_tid(c->tid);
2309}
2310
2311/*
2312 * Flush cpu slab.
2313 *
2314 * Called from IPI handler with interrupts disabled.
2315 */
2316static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2317{
2318 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2319
2320 if (c->page)
2321 flush_slab(s, c);
2322
2323 unfreeze_partials(s, c);
2324}
2325
2326static void flush_cpu_slab(void *d)
2327{
2328 struct kmem_cache *s = d;
2329
2330 __flush_cpu_slab(s, smp_processor_id());
2331}
2332
2333static bool has_cpu_slab(int cpu, void *info)
2334{
2335 struct kmem_cache *s = info;
2336 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2337
2338 return c->page || slub_percpu_partial(c);
2339}
2340
2341static void flush_all(struct kmem_cache *s)
2342{
2343 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2344}
2345
2346/*
2347 * Use the cpu notifier to insure that the cpu slabs are flushed when
2348 * necessary.
2349 */
2350static int slub_cpu_dead(unsigned int cpu)
2351{
2352 struct kmem_cache *s;
2353 unsigned long flags;
2354
2355 mutex_lock(&slab_mutex);
2356 list_for_each_entry(s, &slab_caches, list) {
2357 local_irq_save(flags);
2358 __flush_cpu_slab(s, cpu);
2359 local_irq_restore(flags);
2360 }
2361 mutex_unlock(&slab_mutex);
2362 return 0;
2363}
2364
2365/*
2366 * Check if the objects in a per cpu structure fit numa
2367 * locality expectations.
2368 */
2369static inline int node_match(struct page *page, int node)
2370{
2371#ifdef CONFIG_NUMA
2372 if (node != NUMA_NO_NODE && page_to_nid(page) != node)
2373 return 0;
2374#endif
2375 return 1;
2376}
2377
2378#ifdef CONFIG_SLUB_DEBUG
2379static int count_free(struct page *page)
2380{
2381 return page->objects - page->inuse;
2382}
2383
2384static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2385{
2386 return atomic_long_read(&n->total_objects);
2387}
2388#endif /* CONFIG_SLUB_DEBUG */
2389
2390#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2391static unsigned long count_partial(struct kmem_cache_node *n,
2392 int (*get_count)(struct page *))
2393{
2394 unsigned long flags;
2395 unsigned long x = 0;
2396 struct page *page;
2397
2398 spin_lock_irqsave(&n->list_lock, flags);
2399 list_for_each_entry(page, &n->partial, slab_list)
2400 x += get_count(page);
2401 spin_unlock_irqrestore(&n->list_lock, flags);
2402 return x;
2403}
2404#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2405
2406static noinline void
2407slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2408{
2409#ifdef CONFIG_SLUB_DEBUG
2410 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2411 DEFAULT_RATELIMIT_BURST);
2412 int node;
2413 struct kmem_cache_node *n;
2414
2415 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2416 return;
2417
2418 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2419 nid, gfpflags, &gfpflags);
2420 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2421 s->name, s->object_size, s->size, oo_order(s->oo),
2422 oo_order(s->min));
2423
2424 if (oo_order(s->min) > get_order(s->object_size))
2425 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2426 s->name);
2427
2428 for_each_kmem_cache_node(s, node, n) {
2429 unsigned long nr_slabs;
2430 unsigned long nr_objs;
2431 unsigned long nr_free;
2432
2433 nr_free = count_partial(n, count_free);
2434 nr_slabs = node_nr_slabs(n);
2435 nr_objs = node_nr_objs(n);
2436
2437 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2438 node, nr_slabs, nr_objs, nr_free);
2439 }
2440#endif
2441}
2442
2443static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2444 int node, struct kmem_cache_cpu **pc)
2445{
2446 void *freelist;
2447 struct kmem_cache_cpu *c = *pc;
2448 struct page *page;
2449
2450 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2451
2452 freelist = get_partial(s, flags, node, c);
2453
2454 if (freelist)
2455 return freelist;
2456
2457 page = new_slab(s, flags, node);
2458 if (page) {
2459 c = raw_cpu_ptr(s->cpu_slab);
2460 if (c->page)
2461 flush_slab(s, c);
2462
2463 /*
2464 * No other reference to the page yet so we can
2465 * muck around with it freely without cmpxchg
2466 */
2467 freelist = page->freelist;
2468 page->freelist = NULL;
2469
2470 stat(s, ALLOC_SLAB);
2471 c->page = page;
2472 *pc = c;
2473 }
2474
2475 return freelist;
2476}
2477
2478static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2479{
2480 if (unlikely(PageSlabPfmemalloc(page)))
2481 return gfp_pfmemalloc_allowed(gfpflags);
2482
2483 return true;
2484}
2485
2486/*
2487 * Check the page->freelist of a page and either transfer the freelist to the
2488 * per cpu freelist or deactivate the page.
2489 *
2490 * The page is still frozen if the return value is not NULL.
2491 *
2492 * If this function returns NULL then the page has been unfrozen.
2493 *
2494 * This function must be called with interrupt disabled.
2495 */
2496static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2497{
2498 struct page new;
2499 unsigned long counters;
2500 void *freelist;
2501
2502 do {
2503 freelist = page->freelist;
2504 counters = page->counters;
2505
2506 new.counters = counters;
2507 VM_BUG_ON(!new.frozen);
2508
2509 new.inuse = page->objects;
2510 new.frozen = freelist != NULL;
2511
2512 } while (!__cmpxchg_double_slab(s, page,
2513 freelist, counters,
2514 NULL, new.counters,
2515 "get_freelist"));
2516
2517 return freelist;
2518}
2519
2520/*
2521 * Slow path. The lockless freelist is empty or we need to perform
2522 * debugging duties.
2523 *
2524 * Processing is still very fast if new objects have been freed to the
2525 * regular freelist. In that case we simply take over the regular freelist
2526 * as the lockless freelist and zap the regular freelist.
2527 *
2528 * If that is not working then we fall back to the partial lists. We take the
2529 * first element of the freelist as the object to allocate now and move the
2530 * rest of the freelist to the lockless freelist.
2531 *
2532 * And if we were unable to get a new slab from the partial slab lists then
2533 * we need to allocate a new slab. This is the slowest path since it involves
2534 * a call to the page allocator and the setup of a new slab.
2535 *
2536 * Version of __slab_alloc to use when we know that interrupts are
2537 * already disabled (which is the case for bulk allocation).
2538 */
2539static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2540 unsigned long addr, struct kmem_cache_cpu *c)
2541{
2542 void *freelist;
2543 struct page *page;
2544
2545 page = c->page;
2546 if (!page)
2547 goto new_slab;
2548redo:
2549
2550 if (unlikely(!node_match(page, node))) {
2551 int searchnode = node;
2552
2553 if (node != NUMA_NO_NODE && !node_present_pages(node))
2554 searchnode = node_to_mem_node(node);
2555
2556 if (unlikely(!node_match(page, searchnode))) {
2557 stat(s, ALLOC_NODE_MISMATCH);
2558 deactivate_slab(s, page, c->freelist, c);
2559 goto new_slab;
2560 }
2561 }
2562
2563 /*
2564 * By rights, we should be searching for a slab page that was
2565 * PFMEMALLOC but right now, we are losing the pfmemalloc
2566 * information when the page leaves the per-cpu allocator
2567 */
2568 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2569 deactivate_slab(s, page, c->freelist, c);
2570 goto new_slab;
2571 }
2572
2573 /* must check again c->freelist in case of cpu migration or IRQ */
2574 freelist = c->freelist;
2575 if (freelist)
2576 goto load_freelist;
2577
2578 freelist = get_freelist(s, page);
2579
2580 if (!freelist) {
2581 c->page = NULL;
2582 stat(s, DEACTIVATE_BYPASS);
2583 goto new_slab;
2584 }
2585
2586 stat(s, ALLOC_REFILL);
2587
2588load_freelist:
2589 /*
2590 * freelist is pointing to the list of objects to be used.
2591 * page is pointing to the page from which the objects are obtained.
2592 * That page must be frozen for per cpu allocations to work.
2593 */
2594 VM_BUG_ON(!c->page->frozen);
2595 c->freelist = get_freepointer(s, freelist);
2596 c->tid = next_tid(c->tid);
2597 return freelist;
2598
2599new_slab:
2600
2601 if (slub_percpu_partial(c)) {
2602 page = c->page = slub_percpu_partial(c);
2603 slub_set_percpu_partial(c, page);
2604 stat(s, CPU_PARTIAL_ALLOC);
2605 goto redo;
2606 }
2607
2608 freelist = new_slab_objects(s, gfpflags, node, &c);
2609
2610 if (unlikely(!freelist)) {
2611 slab_out_of_memory(s, gfpflags, node);
2612 return NULL;
2613 }
2614
2615 page = c->page;
2616 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2617 goto load_freelist;
2618
2619 /* Only entered in the debug case */
2620 if (kmem_cache_debug(s) &&
2621 !alloc_debug_processing(s, page, freelist, addr))
2622 goto new_slab; /* Slab failed checks. Next slab needed */
2623
2624 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2625 return freelist;
2626}
2627
2628/*
2629 * Another one that disabled interrupt and compensates for possible
2630 * cpu changes by refetching the per cpu area pointer.
2631 */
2632static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2633 unsigned long addr, struct kmem_cache_cpu *c)
2634{
2635 void *p;
2636 unsigned long flags;
2637
2638 local_irq_save(flags);
2639#ifdef CONFIG_PREEMPT
2640 /*
2641 * We may have been preempted and rescheduled on a different
2642 * cpu before disabling interrupts. Need to reload cpu area
2643 * pointer.
2644 */
2645 c = this_cpu_ptr(s->cpu_slab);
2646#endif
2647
2648 p = ___slab_alloc(s, gfpflags, node, addr, c);
2649 local_irq_restore(flags);
2650 return p;
2651}
2652
2653/*
2654 * If the object has been wiped upon free, make sure it's fully initialized by
2655 * zeroing out freelist pointer.
2656 */
2657static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
2658 void *obj)
2659{
2660 if (unlikely(slab_want_init_on_free(s)) && obj)
2661 memset((void *)((char *)obj + s->offset), 0, sizeof(void *));
2662}
2663
2664/*
2665 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2666 * have the fastpath folded into their functions. So no function call
2667 * overhead for requests that can be satisfied on the fastpath.
2668 *
2669 * The fastpath works by first checking if the lockless freelist can be used.
2670 * If not then __slab_alloc is called for slow processing.
2671 *
2672 * Otherwise we can simply pick the next object from the lockless free list.
2673 */
2674static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2675 gfp_t gfpflags, int node, unsigned long addr)
2676{
2677 void *object;
2678 struct kmem_cache_cpu *c;
2679 struct page *page;
2680 unsigned long tid;
2681
2682 s = slab_pre_alloc_hook(s, gfpflags);
2683 if (!s)
2684 return NULL;
2685redo:
2686 /*
2687 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2688 * enabled. We may switch back and forth between cpus while
2689 * reading from one cpu area. That does not matter as long
2690 * as we end up on the original cpu again when doing the cmpxchg.
2691 *
2692 * We should guarantee that tid and kmem_cache are retrieved on
2693 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2694 * to check if it is matched or not.
2695 */
2696 do {
2697 tid = this_cpu_read(s->cpu_slab->tid);
2698 c = raw_cpu_ptr(s->cpu_slab);
2699 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2700 unlikely(tid != READ_ONCE(c->tid)));
2701
2702 /*
2703 * Irqless object alloc/free algorithm used here depends on sequence
2704 * of fetching cpu_slab's data. tid should be fetched before anything
2705 * on c to guarantee that object and page associated with previous tid
2706 * won't be used with current tid. If we fetch tid first, object and
2707 * page could be one associated with next tid and our alloc/free
2708 * request will be failed. In this case, we will retry. So, no problem.
2709 */
2710 barrier();
2711
2712 /*
2713 * The transaction ids are globally unique per cpu and per operation on
2714 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2715 * occurs on the right processor and that there was no operation on the
2716 * linked list in between.
2717 */
2718
2719 object = c->freelist;
2720 page = c->page;
2721 if (unlikely(!object || !node_match(page, node))) {
2722 object = __slab_alloc(s, gfpflags, node, addr, c);
2723 stat(s, ALLOC_SLOWPATH);
2724 } else {
2725 void *next_object = get_freepointer_safe(s, object);
2726
2727 /*
2728 * The cmpxchg will only match if there was no additional
2729 * operation and if we are on the right processor.
2730 *
2731 * The cmpxchg does the following atomically (without lock
2732 * semantics!)
2733 * 1. Relocate first pointer to the current per cpu area.
2734 * 2. Verify that tid and freelist have not been changed
2735 * 3. If they were not changed replace tid and freelist
2736 *
2737 * Since this is without lock semantics the protection is only
2738 * against code executing on this cpu *not* from access by
2739 * other cpus.
2740 */
2741 if (unlikely(!this_cpu_cmpxchg_double(
2742 s->cpu_slab->freelist, s->cpu_slab->tid,
2743 object, tid,
2744 next_object, next_tid(tid)))) {
2745
2746 note_cmpxchg_failure("slab_alloc", s, tid);
2747 goto redo;
2748 }
2749 prefetch_freepointer(s, next_object);
2750 stat(s, ALLOC_FASTPATH);
2751 }
2752
2753 maybe_wipe_obj_freeptr(s, object);
2754
2755 if (unlikely(slab_want_init_on_alloc(gfpflags, s)) && object)
2756 memset(object, 0, s->object_size);
2757
2758 slab_post_alloc_hook(s, gfpflags, 1, &object);
2759
2760 return object;
2761}
2762
2763static __always_inline void *slab_alloc(struct kmem_cache *s,
2764 gfp_t gfpflags, unsigned long addr)
2765{
2766 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2767}
2768
2769void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2770{
2771 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2772
2773 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2774 s->size, gfpflags);
2775
2776 return ret;
2777}
2778EXPORT_SYMBOL(kmem_cache_alloc);
2779
2780#ifdef CONFIG_TRACING
2781void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2782{
2783 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2784 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2785 ret = kasan_kmalloc(s, ret, size, gfpflags);
2786 return ret;
2787}
2788EXPORT_SYMBOL(kmem_cache_alloc_trace);
2789#endif
2790
2791#ifdef CONFIG_NUMA
2792void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2793{
2794 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2795
2796 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2797 s->object_size, s->size, gfpflags, node);
2798
2799 return ret;
2800}
2801EXPORT_SYMBOL(kmem_cache_alloc_node);
2802
2803#ifdef CONFIG_TRACING
2804void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2805 gfp_t gfpflags,
2806 int node, size_t size)
2807{
2808 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2809
2810 trace_kmalloc_node(_RET_IP_, ret,
2811 size, s->size, gfpflags, node);
2812
2813 ret = kasan_kmalloc(s, ret, size, gfpflags);
2814 return ret;
2815}
2816EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2817#endif
2818#endif /* CONFIG_NUMA */
2819
2820/*
2821 * Slow path handling. This may still be called frequently since objects
2822 * have a longer lifetime than the cpu slabs in most processing loads.
2823 *
2824 * So we still attempt to reduce cache line usage. Just take the slab
2825 * lock and free the item. If there is no additional partial page
2826 * handling required then we can return immediately.
2827 */
2828static void __slab_free(struct kmem_cache *s, struct page *page,
2829 void *head, void *tail, int cnt,
2830 unsigned long addr)
2831
2832{
2833 void *prior;
2834 int was_frozen;
2835 struct page new;
2836 unsigned long counters;
2837 struct kmem_cache_node *n = NULL;
2838 unsigned long uninitialized_var(flags);
2839
2840 stat(s, FREE_SLOWPATH);
2841
2842 if (kmem_cache_debug(s) &&
2843 !free_debug_processing(s, page, head, tail, cnt, addr))
2844 return;
2845
2846 do {
2847 if (unlikely(n)) {
2848 spin_unlock_irqrestore(&n->list_lock, flags);
2849 n = NULL;
2850 }
2851 prior = page->freelist;
2852 counters = page->counters;
2853 set_freepointer(s, tail, prior);
2854 new.counters = counters;
2855 was_frozen = new.frozen;
2856 new.inuse -= cnt;
2857 if ((!new.inuse || !prior) && !was_frozen) {
2858
2859 if (kmem_cache_has_cpu_partial(s) && !prior) {
2860
2861 /*
2862 * Slab was on no list before and will be
2863 * partially empty
2864 * We can defer the list move and instead
2865 * freeze it.
2866 */
2867 new.frozen = 1;
2868
2869 } else { /* Needs to be taken off a list */
2870
2871 n = get_node(s, page_to_nid(page));
2872 /*
2873 * Speculatively acquire the list_lock.
2874 * If the cmpxchg does not succeed then we may
2875 * drop the list_lock without any processing.
2876 *
2877 * Otherwise the list_lock will synchronize with
2878 * other processors updating the list of slabs.
2879 */
2880 spin_lock_irqsave(&n->list_lock, flags);
2881
2882 }
2883 }
2884
2885 } while (!cmpxchg_double_slab(s, page,
2886 prior, counters,
2887 head, new.counters,
2888 "__slab_free"));
2889
2890 if (likely(!n)) {
2891
2892 /*
2893 * If we just froze the page then put it onto the
2894 * per cpu partial list.
2895 */
2896 if (new.frozen && !was_frozen) {
2897 put_cpu_partial(s, page, 1);
2898 stat(s, CPU_PARTIAL_FREE);
2899 }
2900 /*
2901 * The list lock was not taken therefore no list
2902 * activity can be necessary.
2903 */
2904 if (was_frozen)
2905 stat(s, FREE_FROZEN);
2906 return;
2907 }
2908
2909 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2910 goto slab_empty;
2911
2912 /*
2913 * Objects left in the slab. If it was not on the partial list before
2914 * then add it.
2915 */
2916 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2917 remove_full(s, n, page);
2918 add_partial(n, page, DEACTIVATE_TO_TAIL);
2919 stat(s, FREE_ADD_PARTIAL);
2920 }
2921 spin_unlock_irqrestore(&n->list_lock, flags);
2922 return;
2923
2924slab_empty:
2925 if (prior) {
2926 /*
2927 * Slab on the partial list.
2928 */
2929 remove_partial(n, page);
2930 stat(s, FREE_REMOVE_PARTIAL);
2931 } else {
2932 /* Slab must be on the full list */
2933 remove_full(s, n, page);
2934 }
2935
2936 spin_unlock_irqrestore(&n->list_lock, flags);
2937 stat(s, FREE_SLAB);
2938 discard_slab(s, page);
2939}
2940
2941/*
2942 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2943 * can perform fastpath freeing without additional function calls.
2944 *
2945 * The fastpath is only possible if we are freeing to the current cpu slab
2946 * of this processor. This typically the case if we have just allocated
2947 * the item before.
2948 *
2949 * If fastpath is not possible then fall back to __slab_free where we deal
2950 * with all sorts of special processing.
2951 *
2952 * Bulk free of a freelist with several objects (all pointing to the
2953 * same page) possible by specifying head and tail ptr, plus objects
2954 * count (cnt). Bulk free indicated by tail pointer being set.
2955 */
2956static __always_inline void do_slab_free(struct kmem_cache *s,
2957 struct page *page, void *head, void *tail,
2958 int cnt, unsigned long addr)
2959{
2960 void *tail_obj = tail ? : head;
2961 struct kmem_cache_cpu *c;
2962 unsigned long tid;
2963redo:
2964 /*
2965 * Determine the currently cpus per cpu slab.
2966 * The cpu may change afterward. However that does not matter since
2967 * data is retrieved via this pointer. If we are on the same cpu
2968 * during the cmpxchg then the free will succeed.
2969 */
2970 do {
2971 tid = this_cpu_read(s->cpu_slab->tid);
2972 c = raw_cpu_ptr(s->cpu_slab);
2973 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2974 unlikely(tid != READ_ONCE(c->tid)));
2975
2976 /* Same with comment on barrier() in slab_alloc_node() */
2977 barrier();
2978
2979 if (likely(page == c->page)) {
2980 set_freepointer(s, tail_obj, c->freelist);
2981
2982 if (unlikely(!this_cpu_cmpxchg_double(
2983 s->cpu_slab->freelist, s->cpu_slab->tid,
2984 c->freelist, tid,
2985 head, next_tid(tid)))) {
2986
2987 note_cmpxchg_failure("slab_free", s, tid);
2988 goto redo;
2989 }
2990 stat(s, FREE_FASTPATH);
2991 } else
2992 __slab_free(s, page, head, tail_obj, cnt, addr);
2993
2994}
2995
2996static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2997 void *head, void *tail, int cnt,
2998 unsigned long addr)
2999{
3000 /*
3001 * With KASAN enabled slab_free_freelist_hook modifies the freelist
3002 * to remove objects, whose reuse must be delayed.
3003 */
3004 if (slab_free_freelist_hook(s, &head, &tail))
3005 do_slab_free(s, page, head, tail, cnt, addr);
3006}
3007
3008#ifdef CONFIG_KASAN_GENERIC
3009void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
3010{
3011 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
3012}
3013#endif
3014
3015void kmem_cache_free(struct kmem_cache *s, void *x)
3016{
3017 s = cache_from_obj(s, x);
3018 if (!s)
3019 return;
3020 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3021 trace_kmem_cache_free(_RET_IP_, x);
3022}
3023EXPORT_SYMBOL(kmem_cache_free);
3024
3025struct detached_freelist {
3026 struct page *page;
3027 void *tail;
3028 void *freelist;
3029 int cnt;
3030 struct kmem_cache *s;
3031};
3032
3033/*
3034 * This function progressively scans the array with free objects (with
3035 * a limited look ahead) and extract objects belonging to the same
3036 * page. It builds a detached freelist directly within the given
3037 * page/objects. This can happen without any need for
3038 * synchronization, because the objects are owned by running process.
3039 * The freelist is build up as a single linked list in the objects.
3040 * The idea is, that this detached freelist can then be bulk
3041 * transferred to the real freelist(s), but only requiring a single
3042 * synchronization primitive. Look ahead in the array is limited due
3043 * to performance reasons.
3044 */
3045static inline
3046int build_detached_freelist(struct kmem_cache *s, size_t size,
3047 void **p, struct detached_freelist *df)
3048{
3049 size_t first_skipped_index = 0;
3050 int lookahead = 3;
3051 void *object;
3052 struct page *page;
3053
3054 /* Always re-init detached_freelist */
3055 df->page = NULL;
3056
3057 do {
3058 object = p[--size];
3059 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3060 } while (!object && size);
3061
3062 if (!object)
3063 return 0;
3064
3065 page = virt_to_head_page(object);
3066 if (!s) {
3067 /* Handle kalloc'ed objects */
3068 if (unlikely(!PageSlab(page))) {
3069 BUG_ON(!PageCompound(page));
3070 kfree_hook(object);
3071 __free_pages(page, compound_order(page));
3072 p[size] = NULL; /* mark object processed */
3073 return size;
3074 }
3075 /* Derive kmem_cache from object */
3076 df->s = page->slab_cache;
3077 } else {
3078 df->s = cache_from_obj(s, object); /* Support for memcg */
3079 }
3080
3081 /* Start new detached freelist */
3082 df->page = page;
3083 set_freepointer(df->s, object, NULL);
3084 df->tail = object;
3085 df->freelist = object;
3086 p[size] = NULL; /* mark object processed */
3087 df->cnt = 1;
3088
3089 while (size) {
3090 object = p[--size];
3091 if (!object)
3092 continue; /* Skip processed objects */
3093
3094 /* df->page is always set at this point */
3095 if (df->page == virt_to_head_page(object)) {
3096 /* Opportunity build freelist */
3097 set_freepointer(df->s, object, df->freelist);
3098 df->freelist = object;
3099 df->cnt++;
3100 p[size] = NULL; /* mark object processed */
3101
3102 continue;
3103 }
3104
3105 /* Limit look ahead search */
3106 if (!--lookahead)
3107 break;
3108
3109 if (!first_skipped_index)
3110 first_skipped_index = size + 1;
3111 }
3112
3113 return first_skipped_index;
3114}
3115
3116/* Note that interrupts must be enabled when calling this function. */
3117void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3118{
3119 if (WARN_ON(!size))
3120 return;
3121
3122 do {
3123 struct detached_freelist df;
3124
3125 size = build_detached_freelist(s, size, p, &df);
3126 if (!df.page)
3127 continue;
3128
3129 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3130 } while (likely(size));
3131}
3132EXPORT_SYMBOL(kmem_cache_free_bulk);
3133
3134/* Note that interrupts must be enabled when calling this function. */
3135int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3136 void **p)
3137{
3138 struct kmem_cache_cpu *c;
3139 int i;
3140
3141 /* memcg and kmem_cache debug support */
3142 s = slab_pre_alloc_hook(s, flags);
3143 if (unlikely(!s))
3144 return false;
3145 /*
3146 * Drain objects in the per cpu slab, while disabling local
3147 * IRQs, which protects against PREEMPT and interrupts
3148 * handlers invoking normal fastpath.
3149 */
3150 local_irq_disable();
3151 c = this_cpu_ptr(s->cpu_slab);
3152
3153 for (i = 0; i < size; i++) {
3154 void *object = c->freelist;
3155
3156 if (unlikely(!object)) {
3157 /*
3158 * Invoking slow path likely have side-effect
3159 * of re-populating per CPU c->freelist
3160 */
3161 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3162 _RET_IP_, c);
3163 if (unlikely(!p[i]))
3164 goto error;
3165
3166 c = this_cpu_ptr(s->cpu_slab);
3167 maybe_wipe_obj_freeptr(s, p[i]);
3168
3169 continue; /* goto for-loop */
3170 }
3171 c->freelist = get_freepointer(s, object);
3172 p[i] = object;
3173 maybe_wipe_obj_freeptr(s, p[i]);
3174 }
3175 c->tid = next_tid(c->tid);
3176 local_irq_enable();
3177
3178 /* Clear memory outside IRQ disabled fastpath loop */
3179 if (unlikely(slab_want_init_on_alloc(flags, s))) {
3180 int j;
3181
3182 for (j = 0; j < i; j++)
3183 memset(p[j], 0, s->object_size);
3184 }
3185
3186 /* memcg and kmem_cache debug support */
3187 slab_post_alloc_hook(s, flags, size, p);
3188 return i;
3189error:
3190 local_irq_enable();
3191 slab_post_alloc_hook(s, flags, i, p);
3192 __kmem_cache_free_bulk(s, i, p);
3193 return 0;
3194}
3195EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3196
3197
3198/*
3199 * Object placement in a slab is made very easy because we always start at
3200 * offset 0. If we tune the size of the object to the alignment then we can
3201 * get the required alignment by putting one properly sized object after
3202 * another.
3203 *
3204 * Notice that the allocation order determines the sizes of the per cpu
3205 * caches. Each processor has always one slab available for allocations.
3206 * Increasing the allocation order reduces the number of times that slabs
3207 * must be moved on and off the partial lists and is therefore a factor in
3208 * locking overhead.
3209 */
3210
3211/*
3212 * Mininum / Maximum order of slab pages. This influences locking overhead
3213 * and slab fragmentation. A higher order reduces the number of partial slabs
3214 * and increases the number of allocations possible without having to
3215 * take the list_lock.
3216 */
3217static unsigned int slub_min_order;
3218static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3219static unsigned int slub_min_objects;
3220
3221/*
3222 * Calculate the order of allocation given an slab object size.
3223 *
3224 * The order of allocation has significant impact on performance and other
3225 * system components. Generally order 0 allocations should be preferred since
3226 * order 0 does not cause fragmentation in the page allocator. Larger objects
3227 * be problematic to put into order 0 slabs because there may be too much
3228 * unused space left. We go to a higher order if more than 1/16th of the slab
3229 * would be wasted.
3230 *
3231 * In order to reach satisfactory performance we must ensure that a minimum
3232 * number of objects is in one slab. Otherwise we may generate too much
3233 * activity on the partial lists which requires taking the list_lock. This is
3234 * less a concern for large slabs though which are rarely used.
3235 *
3236 * slub_max_order specifies the order where we begin to stop considering the
3237 * number of objects in a slab as critical. If we reach slub_max_order then
3238 * we try to keep the page order as low as possible. So we accept more waste
3239 * of space in favor of a small page order.
3240 *
3241 * Higher order allocations also allow the placement of more objects in a
3242 * slab and thereby reduce object handling overhead. If the user has
3243 * requested a higher mininum order then we start with that one instead of
3244 * the smallest order which will fit the object.
3245 */
3246static inline unsigned int slab_order(unsigned int size,
3247 unsigned int min_objects, unsigned int max_order,
3248 unsigned int fract_leftover)
3249{
3250 unsigned int min_order = slub_min_order;
3251 unsigned int order;
3252
3253 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
3254 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3255
3256 for (order = max(min_order, (unsigned int)get_order(min_objects * size));
3257 order <= max_order; order++) {
3258
3259 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3260 unsigned int rem;
3261
3262 rem = slab_size % size;
3263
3264 if (rem <= slab_size / fract_leftover)
3265 break;
3266 }
3267
3268 return order;
3269}
3270
3271static inline int calculate_order(unsigned int size)
3272{
3273 unsigned int order;
3274 unsigned int min_objects;
3275 unsigned int max_objects;
3276
3277 /*
3278 * Attempt to find best configuration for a slab. This
3279 * works by first attempting to generate a layout with
3280 * the best configuration and backing off gradually.
3281 *
3282 * First we increase the acceptable waste in a slab. Then
3283 * we reduce the minimum objects required in a slab.
3284 */
3285 min_objects = slub_min_objects;
3286 if (!min_objects)
3287 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3288 max_objects = order_objects(slub_max_order, size);
3289 min_objects = min(min_objects, max_objects);
3290
3291 while (min_objects > 1) {
3292 unsigned int fraction;
3293
3294 fraction = 16;
3295 while (fraction >= 4) {
3296 order = slab_order(size, min_objects,
3297 slub_max_order, fraction);
3298 if (order <= slub_max_order)
3299 return order;
3300 fraction /= 2;
3301 }
3302 min_objects--;
3303 }
3304
3305 /*
3306 * We were unable to place multiple objects in a slab. Now
3307 * lets see if we can place a single object there.
3308 */
3309 order = slab_order(size, 1, slub_max_order, 1);
3310 if (order <= slub_max_order)
3311 return order;
3312
3313 /*
3314 * Doh this slab cannot be placed using slub_max_order.
3315 */
3316 order = slab_order(size, 1, MAX_ORDER, 1);
3317 if (order < MAX_ORDER)
3318 return order;
3319 return -ENOSYS;
3320}
3321
3322static void
3323init_kmem_cache_node(struct kmem_cache_node *n)
3324{
3325 n->nr_partial = 0;
3326 spin_lock_init(&n->list_lock);
3327 INIT_LIST_HEAD(&n->partial);
3328#ifdef CONFIG_SLUB_DEBUG
3329 atomic_long_set(&n->nr_slabs, 0);
3330 atomic_long_set(&n->total_objects, 0);
3331 INIT_LIST_HEAD(&n->full);
3332#endif
3333}
3334
3335static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3336{
3337 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3338 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3339
3340 /*
3341 * Must align to double word boundary for the double cmpxchg
3342 * instructions to work; see __pcpu_double_call_return_bool().
3343 */
3344 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3345 2 * sizeof(void *));
3346
3347 if (!s->cpu_slab)
3348 return 0;
3349
3350 init_kmem_cache_cpus(s);
3351
3352 return 1;
3353}
3354
3355static struct kmem_cache *kmem_cache_node;
3356
3357/*
3358 * No kmalloc_node yet so do it by hand. We know that this is the first
3359 * slab on the node for this slabcache. There are no concurrent accesses
3360 * possible.
3361 *
3362 * Note that this function only works on the kmem_cache_node
3363 * when allocating for the kmem_cache_node. This is used for bootstrapping
3364 * memory on a fresh node that has no slab structures yet.
3365 */
3366static void early_kmem_cache_node_alloc(int node)
3367{
3368 struct page *page;
3369 struct kmem_cache_node *n;
3370
3371 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3372
3373 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3374
3375 BUG_ON(!page);
3376 if (page_to_nid(page) != node) {
3377 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3378 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3379 }
3380
3381 n = page->freelist;
3382 BUG_ON(!n);
3383#ifdef CONFIG_SLUB_DEBUG
3384 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3385 init_tracking(kmem_cache_node, n);
3386#endif
3387 n = kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3388 GFP_KERNEL);
3389 page->freelist = get_freepointer(kmem_cache_node, n);
3390 page->inuse = 1;
3391 page->frozen = 0;
3392 kmem_cache_node->node[node] = n;
3393 init_kmem_cache_node(n);
3394 inc_slabs_node(kmem_cache_node, node, page->objects);
3395
3396 /*
3397 * No locks need to be taken here as it has just been
3398 * initialized and there is no concurrent access.
3399 */
3400 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3401}
3402
3403static void free_kmem_cache_nodes(struct kmem_cache *s)
3404{
3405 int node;
3406 struct kmem_cache_node *n;
3407
3408 for_each_kmem_cache_node(s, node, n) {
3409 s->node[node] = NULL;
3410 kmem_cache_free(kmem_cache_node, n);
3411 }
3412}
3413
3414void __kmem_cache_release(struct kmem_cache *s)
3415{
3416 cache_random_seq_destroy(s);
3417 free_percpu(s->cpu_slab);
3418 free_kmem_cache_nodes(s);
3419}
3420
3421static int init_kmem_cache_nodes(struct kmem_cache *s)
3422{
3423 int node;
3424
3425 for_each_node_state(node, N_NORMAL_MEMORY) {
3426 struct kmem_cache_node *n;
3427
3428 if (slab_state == DOWN) {
3429 early_kmem_cache_node_alloc(node);
3430 continue;
3431 }
3432 n = kmem_cache_alloc_node(kmem_cache_node,
3433 GFP_KERNEL, node);
3434
3435 if (!n) {
3436 free_kmem_cache_nodes(s);
3437 return 0;
3438 }
3439
3440 init_kmem_cache_node(n);
3441 s->node[node] = n;
3442 }
3443 return 1;
3444}
3445
3446static void set_min_partial(struct kmem_cache *s, unsigned long min)
3447{
3448 if (min < MIN_PARTIAL)
3449 min = MIN_PARTIAL;
3450 else if (min > MAX_PARTIAL)
3451 min = MAX_PARTIAL;
3452 s->min_partial = min;
3453}
3454
3455static void set_cpu_partial(struct kmem_cache *s)
3456{
3457#ifdef CONFIG_SLUB_CPU_PARTIAL
3458 /*
3459 * cpu_partial determined the maximum number of objects kept in the
3460 * per cpu partial lists of a processor.
3461 *
3462 * Per cpu partial lists mainly contain slabs that just have one
3463 * object freed. If they are used for allocation then they can be
3464 * filled up again with minimal effort. The slab will never hit the
3465 * per node partial lists and therefore no locking will be required.
3466 *
3467 * This setting also determines
3468 *
3469 * A) The number of objects from per cpu partial slabs dumped to the
3470 * per node list when we reach the limit.
3471 * B) The number of objects in cpu partial slabs to extract from the
3472 * per node list when we run out of per cpu objects. We only fetch
3473 * 50% to keep some capacity around for frees.
3474 */
3475 if (!kmem_cache_has_cpu_partial(s))
3476 s->cpu_partial = 0;
3477 else if (s->size >= PAGE_SIZE)
3478 s->cpu_partial = 2;
3479 else if (s->size >= 1024)
3480 s->cpu_partial = 6;
3481 else if (s->size >= 256)
3482 s->cpu_partial = 13;
3483 else
3484 s->cpu_partial = 30;
3485#endif
3486}
3487
3488/*
3489 * calculate_sizes() determines the order and the distribution of data within
3490 * a slab object.
3491 */
3492static int calculate_sizes(struct kmem_cache *s, int forced_order)
3493{
3494 slab_flags_t flags = s->flags;
3495 unsigned int size = s->object_size;
3496 unsigned int order;
3497
3498 /*
3499 * Round up object size to the next word boundary. We can only
3500 * place the free pointer at word boundaries and this determines
3501 * the possible location of the free pointer.
3502 */
3503 size = ALIGN(size, sizeof(void *));
3504
3505#ifdef CONFIG_SLUB_DEBUG
3506 /*
3507 * Determine if we can poison the object itself. If the user of
3508 * the slab may touch the object after free or before allocation
3509 * then we should never poison the object itself.
3510 */
3511 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3512 !s->ctor)
3513 s->flags |= __OBJECT_POISON;
3514 else
3515 s->flags &= ~__OBJECT_POISON;
3516
3517
3518 /*
3519 * If we are Redzoning then check if there is some space between the
3520 * end of the object and the free pointer. If not then add an
3521 * additional word to have some bytes to store Redzone information.
3522 */
3523 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3524 size += sizeof(void *);
3525#endif
3526
3527 /*
3528 * With that we have determined the number of bytes in actual use
3529 * by the object. This is the potential offset to the free pointer.
3530 */
3531 s->inuse = size;
3532
3533 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3534 s->ctor)) {
3535 /*
3536 * Relocate free pointer after the object if it is not
3537 * permitted to overwrite the first word of the object on
3538 * kmem_cache_free.
3539 *
3540 * This is the case if we do RCU, have a constructor or
3541 * destructor or are poisoning the objects.
3542 */
3543 s->offset = size;
3544 size += sizeof(void *);
3545 }
3546
3547#ifdef CONFIG_SLUB_DEBUG
3548 if (flags & SLAB_STORE_USER)
3549 /*
3550 * Need to store information about allocs and frees after
3551 * the object.
3552 */
3553 size += 2 * sizeof(struct track);
3554#endif
3555
3556 kasan_cache_create(s, &size, &s->flags);
3557#ifdef CONFIG_SLUB_DEBUG
3558 if (flags & SLAB_RED_ZONE) {
3559 /*
3560 * Add some empty padding so that we can catch
3561 * overwrites from earlier objects rather than let
3562 * tracking information or the free pointer be
3563 * corrupted if a user writes before the start
3564 * of the object.
3565 */
3566 size += sizeof(void *);
3567
3568 s->red_left_pad = sizeof(void *);
3569 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3570 size += s->red_left_pad;
3571 }
3572#endif
3573
3574 /*
3575 * SLUB stores one object immediately after another beginning from
3576 * offset 0. In order to align the objects we have to simply size
3577 * each object to conform to the alignment.
3578 */
3579 size = ALIGN(size, s->align);
3580 s->size = size;
3581 if (forced_order >= 0)
3582 order = forced_order;
3583 else
3584 order = calculate_order(size);
3585
3586 if ((int)order < 0)
3587 return 0;
3588
3589 s->allocflags = 0;
3590 if (order)
3591 s->allocflags |= __GFP_COMP;
3592
3593 if (s->flags & SLAB_CACHE_DMA)
3594 s->allocflags |= GFP_DMA;
3595
3596 if (s->flags & SLAB_CACHE_DMA32)
3597 s->allocflags |= GFP_DMA32;
3598
3599 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3600 s->allocflags |= __GFP_RECLAIMABLE;
3601
3602 /*
3603 * Determine the number of objects per slab
3604 */
3605 s->oo = oo_make(order, size);
3606 s->min = oo_make(get_order(size), size);
3607 if (oo_objects(s->oo) > oo_objects(s->max))
3608 s->max = s->oo;
3609
3610 return !!oo_objects(s->oo);
3611}
3612
3613static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3614{
3615 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3616#ifdef CONFIG_SLAB_FREELIST_HARDENED
3617 s->random = get_random_long();
3618#endif
3619
3620 if (!calculate_sizes(s, -1))
3621 goto error;
3622 if (disable_higher_order_debug) {
3623 /*
3624 * Disable debugging flags that store metadata if the min slab
3625 * order increased.
3626 */
3627 if (get_order(s->size) > get_order(s->object_size)) {
3628 s->flags &= ~DEBUG_METADATA_FLAGS;
3629 s->offset = 0;
3630 if (!calculate_sizes(s, -1))
3631 goto error;
3632 }
3633 }
3634
3635#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3636 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3637 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3638 /* Enable fast mode */
3639 s->flags |= __CMPXCHG_DOUBLE;
3640#endif
3641
3642 /*
3643 * The larger the object size is, the more pages we want on the partial
3644 * list to avoid pounding the page allocator excessively.
3645 */
3646 set_min_partial(s, ilog2(s->size) / 2);
3647
3648 set_cpu_partial(s);
3649
3650#ifdef CONFIG_NUMA
3651 s->remote_node_defrag_ratio = 1000;
3652#endif
3653
3654 /* Initialize the pre-computed randomized freelist if slab is up */
3655 if (slab_state >= UP) {
3656 if (init_cache_random_seq(s))
3657 goto error;
3658 }
3659
3660 if (!init_kmem_cache_nodes(s))
3661 goto error;
3662
3663 if (alloc_kmem_cache_cpus(s))
3664 return 0;
3665
3666 free_kmem_cache_nodes(s);
3667error:
3668 return -EINVAL;
3669}
3670
3671static void list_slab_objects(struct kmem_cache *s, struct page *page,
3672 const char *text)
3673{
3674#ifdef CONFIG_SLUB_DEBUG
3675 void *addr = page_address(page);
3676 void *p;
3677 unsigned long *map = bitmap_zalloc(page->objects, GFP_ATOMIC);
3678 if (!map)
3679 return;
3680 slab_err(s, page, text, s->name);
3681 slab_lock(page);
3682
3683 get_map(s, page, map);
3684 for_each_object(p, s, addr, page->objects) {
3685
3686 if (!test_bit(slab_index(p, s, addr), map)) {
3687 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3688 print_tracking(s, p);
3689 }
3690 }
3691 slab_unlock(page);
3692 bitmap_free(map);
3693#endif
3694}
3695
3696/*
3697 * Attempt to free all partial slabs on a node.
3698 * This is called from __kmem_cache_shutdown(). We must take list_lock
3699 * because sysfs file might still access partial list after the shutdowning.
3700 */
3701static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3702{
3703 LIST_HEAD(discard);
3704 struct page *page, *h;
3705
3706 BUG_ON(irqs_disabled());
3707 spin_lock_irq(&n->list_lock);
3708 list_for_each_entry_safe(page, h, &n->partial, slab_list) {
3709 if (!page->inuse) {
3710 remove_partial(n, page);
3711 list_add(&page->slab_list, &discard);
3712 } else {
3713 list_slab_objects(s, page,
3714 "Objects remaining in %s on __kmem_cache_shutdown()");
3715 }
3716 }
3717 spin_unlock_irq(&n->list_lock);
3718
3719 list_for_each_entry_safe(page, h, &discard, slab_list)
3720 discard_slab(s, page);
3721}
3722
3723bool __kmem_cache_empty(struct kmem_cache *s)
3724{
3725 int node;
3726 struct kmem_cache_node *n;
3727
3728 for_each_kmem_cache_node(s, node, n)
3729 if (n->nr_partial || slabs_node(s, node))
3730 return false;
3731 return true;
3732}
3733
3734/*
3735 * Release all resources used by a slab cache.
3736 */
3737int __kmem_cache_shutdown(struct kmem_cache *s)
3738{
3739 int node;
3740 struct kmem_cache_node *n;
3741
3742 flush_all(s);
3743 /* Attempt to free all objects */
3744 for_each_kmem_cache_node(s, node, n) {
3745 free_partial(s, n);
3746 if (n->nr_partial || slabs_node(s, node))
3747 return 1;
3748 }
3749 sysfs_slab_remove(s);
3750 return 0;
3751}
3752
3753/********************************************************************
3754 * Kmalloc subsystem
3755 *******************************************************************/
3756
3757static int __init setup_slub_min_order(char *str)
3758{
3759 get_option(&str, (int *)&slub_min_order);
3760
3761 return 1;
3762}
3763
3764__setup("slub_min_order=", setup_slub_min_order);
3765
3766static int __init setup_slub_max_order(char *str)
3767{
3768 get_option(&str, (int *)&slub_max_order);
3769 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3770
3771 return 1;
3772}
3773
3774__setup("slub_max_order=", setup_slub_max_order);
3775
3776static int __init setup_slub_min_objects(char *str)
3777{
3778 get_option(&str, (int *)&slub_min_objects);
3779
3780 return 1;
3781}
3782
3783__setup("slub_min_objects=", setup_slub_min_objects);
3784
3785void *__kmalloc(size_t size, gfp_t flags)
3786{
3787 struct kmem_cache *s;
3788 void *ret;
3789
3790 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3791 return kmalloc_large(size, flags);
3792
3793 s = kmalloc_slab(size, flags);
3794
3795 if (unlikely(ZERO_OR_NULL_PTR(s)))
3796 return s;
3797
3798 ret = slab_alloc(s, flags, _RET_IP_);
3799
3800 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3801
3802 ret = kasan_kmalloc(s, ret, size, flags);
3803
3804 return ret;
3805}
3806EXPORT_SYMBOL(__kmalloc);
3807
3808#ifdef CONFIG_NUMA
3809static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3810{
3811 struct page *page;
3812 void *ptr = NULL;
3813 unsigned int order = get_order(size);
3814
3815 flags |= __GFP_COMP;
3816 page = alloc_pages_node(node, flags, order);
3817 if (page) {
3818 ptr = page_address(page);
3819 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3820 1 << order);
3821 }
3822
3823 return kmalloc_large_node_hook(ptr, size, flags);
3824}
3825
3826void *__kmalloc_node(size_t size, gfp_t flags, int node)
3827{
3828 struct kmem_cache *s;
3829 void *ret;
3830
3831 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3832 ret = kmalloc_large_node(size, flags, node);
3833
3834 trace_kmalloc_node(_RET_IP_, ret,
3835 size, PAGE_SIZE << get_order(size),
3836 flags, node);
3837
3838 return ret;
3839 }
3840
3841 s = kmalloc_slab(size, flags);
3842
3843 if (unlikely(ZERO_OR_NULL_PTR(s)))
3844 return s;
3845
3846 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3847
3848 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3849
3850 ret = kasan_kmalloc(s, ret, size, flags);
3851
3852 return ret;
3853}
3854EXPORT_SYMBOL(__kmalloc_node);
3855#endif /* CONFIG_NUMA */
3856
3857#ifdef CONFIG_HARDENED_USERCOPY
3858/*
3859 * Rejects incorrectly sized objects and objects that are to be copied
3860 * to/from userspace but do not fall entirely within the containing slab
3861 * cache's usercopy region.
3862 *
3863 * Returns NULL if check passes, otherwise const char * to name of cache
3864 * to indicate an error.
3865 */
3866void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3867 bool to_user)
3868{
3869 struct kmem_cache *s;
3870 unsigned int offset;
3871 size_t object_size;
3872
3873 ptr = kasan_reset_tag(ptr);
3874
3875 /* Find object and usable object size. */
3876 s = page->slab_cache;
3877
3878 /* Reject impossible pointers. */
3879 if (ptr < page_address(page))
3880 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3881 to_user, 0, n);
3882
3883 /* Find offset within object. */
3884 offset = (ptr - page_address(page)) % s->size;
3885
3886 /* Adjust for redzone and reject if within the redzone. */
3887 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3888 if (offset < s->red_left_pad)
3889 usercopy_abort("SLUB object in left red zone",
3890 s->name, to_user, offset, n);
3891 offset -= s->red_left_pad;
3892 }
3893
3894 /* Allow address range falling entirely within usercopy region. */
3895 if (offset >= s->useroffset &&
3896 offset - s->useroffset <= s->usersize &&
3897 n <= s->useroffset - offset + s->usersize)
3898 return;
3899
3900 /*
3901 * If the copy is still within the allocated object, produce
3902 * a warning instead of rejecting the copy. This is intended
3903 * to be a temporary method to find any missing usercopy
3904 * whitelists.
3905 */
3906 object_size = slab_ksize(s);
3907 if (usercopy_fallback &&
3908 offset <= object_size && n <= object_size - offset) {
3909 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3910 return;
3911 }
3912
3913 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3914}
3915#endif /* CONFIG_HARDENED_USERCOPY */
3916
3917size_t __ksize(const void *object)
3918{
3919 struct page *page;
3920
3921 if (unlikely(object == ZERO_SIZE_PTR))
3922 return 0;
3923
3924 page = virt_to_head_page(object);
3925
3926 if (unlikely(!PageSlab(page))) {
3927 WARN_ON(!PageCompound(page));
3928 return page_size(page);
3929 }
3930
3931 return slab_ksize(page->slab_cache);
3932}
3933EXPORT_SYMBOL(__ksize);
3934
3935void kfree(const void *x)
3936{
3937 struct page *page;
3938 void *object = (void *)x;
3939
3940 trace_kfree(_RET_IP_, x);
3941
3942 if (unlikely(ZERO_OR_NULL_PTR(x)))
3943 return;
3944
3945 page = virt_to_head_page(x);
3946 if (unlikely(!PageSlab(page))) {
3947 unsigned int order = compound_order(page);
3948
3949 BUG_ON(!PageCompound(page));
3950 kfree_hook(object);
3951 mod_node_page_state(page_pgdat(page), NR_SLAB_UNRECLAIMABLE,
3952 -(1 << order));
3953 __free_pages(page, order);
3954 return;
3955 }
3956 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3957}
3958EXPORT_SYMBOL(kfree);
3959
3960#define SHRINK_PROMOTE_MAX 32
3961
3962/*
3963 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3964 * up most to the head of the partial lists. New allocations will then
3965 * fill those up and thus they can be removed from the partial lists.
3966 *
3967 * The slabs with the least items are placed last. This results in them
3968 * being allocated from last increasing the chance that the last objects
3969 * are freed in them.
3970 */
3971int __kmem_cache_shrink(struct kmem_cache *s)
3972{
3973 int node;
3974 int i;
3975 struct kmem_cache_node *n;
3976 struct page *page;
3977 struct page *t;
3978 struct list_head discard;
3979 struct list_head promote[SHRINK_PROMOTE_MAX];
3980 unsigned long flags;
3981 int ret = 0;
3982
3983 flush_all(s);
3984 for_each_kmem_cache_node(s, node, n) {
3985 INIT_LIST_HEAD(&discard);
3986 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3987 INIT_LIST_HEAD(promote + i);
3988
3989 spin_lock_irqsave(&n->list_lock, flags);
3990
3991 /*
3992 * Build lists of slabs to discard or promote.
3993 *
3994 * Note that concurrent frees may occur while we hold the
3995 * list_lock. page->inuse here is the upper limit.
3996 */
3997 list_for_each_entry_safe(page, t, &n->partial, slab_list) {
3998 int free = page->objects - page->inuse;
3999
4000 /* Do not reread page->inuse */
4001 barrier();
4002
4003 /* We do not keep full slabs on the list */
4004 BUG_ON(free <= 0);
4005
4006 if (free == page->objects) {
4007 list_move(&page->slab_list, &discard);
4008 n->nr_partial--;
4009 } else if (free <= SHRINK_PROMOTE_MAX)
4010 list_move(&page->slab_list, promote + free - 1);
4011 }
4012
4013 /*
4014 * Promote the slabs filled up most to the head of the
4015 * partial list.
4016 */
4017 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4018 list_splice(promote + i, &n->partial);
4019
4020 spin_unlock_irqrestore(&n->list_lock, flags);
4021
4022 /* Release empty slabs */
4023 list_for_each_entry_safe(page, t, &discard, slab_list)
4024 discard_slab(s, page);
4025
4026 if (slabs_node(s, node))
4027 ret = 1;
4028 }
4029
4030 return ret;
4031}
4032
4033#ifdef CONFIG_MEMCG
4034void __kmemcg_cache_deactivate_after_rcu(struct kmem_cache *s)
4035{
4036 /*
4037 * Called with all the locks held after a sched RCU grace period.
4038 * Even if @s becomes empty after shrinking, we can't know that @s
4039 * doesn't have allocations already in-flight and thus can't
4040 * destroy @s until the associated memcg is released.
4041 *
4042 * However, let's remove the sysfs files for empty caches here.
4043 * Each cache has a lot of interface files which aren't
4044 * particularly useful for empty draining caches; otherwise, we can
4045 * easily end up with millions of unnecessary sysfs files on
4046 * systems which have a lot of memory and transient cgroups.
4047 */
4048 if (!__kmem_cache_shrink(s))
4049 sysfs_slab_remove(s);
4050}
4051
4052void __kmemcg_cache_deactivate(struct kmem_cache *s)
4053{
4054 /*
4055 * Disable empty slabs caching. Used to avoid pinning offline
4056 * memory cgroups by kmem pages that can be freed.
4057 */
4058 slub_set_cpu_partial(s, 0);
4059 s->min_partial = 0;
4060}
4061#endif /* CONFIG_MEMCG */
4062
4063static int slab_mem_going_offline_callback(void *arg)
4064{
4065 struct kmem_cache *s;
4066
4067 mutex_lock(&slab_mutex);
4068 list_for_each_entry(s, &slab_caches, list)
4069 __kmem_cache_shrink(s);
4070 mutex_unlock(&slab_mutex);
4071
4072 return 0;
4073}
4074
4075static void slab_mem_offline_callback(void *arg)
4076{
4077 struct kmem_cache_node *n;
4078 struct kmem_cache *s;
4079 struct memory_notify *marg = arg;
4080 int offline_node;
4081
4082 offline_node = marg->status_change_nid_normal;
4083
4084 /*
4085 * If the node still has available memory. we need kmem_cache_node
4086 * for it yet.
4087 */
4088 if (offline_node < 0)
4089 return;
4090
4091 mutex_lock(&slab_mutex);
4092 list_for_each_entry(s, &slab_caches, list) {
4093 n = get_node(s, offline_node);
4094 if (n) {
4095 /*
4096 * if n->nr_slabs > 0, slabs still exist on the node
4097 * that is going down. We were unable to free them,
4098 * and offline_pages() function shouldn't call this
4099 * callback. So, we must fail.
4100 */
4101 BUG_ON(slabs_node(s, offline_node));
4102
4103 s->node[offline_node] = NULL;
4104 kmem_cache_free(kmem_cache_node, n);
4105 }
4106 }
4107 mutex_unlock(&slab_mutex);
4108}
4109
4110static int slab_mem_going_online_callback(void *arg)
4111{
4112 struct kmem_cache_node *n;
4113 struct kmem_cache *s;
4114 struct memory_notify *marg = arg;
4115 int nid = marg->status_change_nid_normal;
4116 int ret = 0;
4117
4118 /*
4119 * If the node's memory is already available, then kmem_cache_node is
4120 * already created. Nothing to do.
4121 */
4122 if (nid < 0)
4123 return 0;
4124
4125 /*
4126 * We are bringing a node online. No memory is available yet. We must
4127 * allocate a kmem_cache_node structure in order to bring the node
4128 * online.
4129 */
4130 mutex_lock(&slab_mutex);
4131 list_for_each_entry(s, &slab_caches, list) {
4132 /*
4133 * XXX: kmem_cache_alloc_node will fallback to other nodes
4134 * since memory is not yet available from the node that
4135 * is brought up.
4136 */
4137 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4138 if (!n) {
4139 ret = -ENOMEM;
4140 goto out;
4141 }
4142 init_kmem_cache_node(n);
4143 s->node[nid] = n;
4144 }
4145out:
4146 mutex_unlock(&slab_mutex);
4147 return ret;
4148}
4149
4150static int slab_memory_callback(struct notifier_block *self,
4151 unsigned long action, void *arg)
4152{
4153 int ret = 0;
4154
4155 switch (action) {
4156 case MEM_GOING_ONLINE:
4157 ret = slab_mem_going_online_callback(arg);
4158 break;
4159 case MEM_GOING_OFFLINE:
4160 ret = slab_mem_going_offline_callback(arg);
4161 break;
4162 case MEM_OFFLINE:
4163 case MEM_CANCEL_ONLINE:
4164 slab_mem_offline_callback(arg);
4165 break;
4166 case MEM_ONLINE:
4167 case MEM_CANCEL_OFFLINE:
4168 break;
4169 }
4170 if (ret)
4171 ret = notifier_from_errno(ret);
4172 else
4173 ret = NOTIFY_OK;
4174 return ret;
4175}
4176
4177static struct notifier_block slab_memory_callback_nb = {
4178 .notifier_call = slab_memory_callback,
4179 .priority = SLAB_CALLBACK_PRI,
4180};
4181
4182/********************************************************************
4183 * Basic setup of slabs
4184 *******************************************************************/
4185
4186/*
4187 * Used for early kmem_cache structures that were allocated using
4188 * the page allocator. Allocate them properly then fix up the pointers
4189 * that may be pointing to the wrong kmem_cache structure.
4190 */
4191
4192static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4193{
4194 int node;
4195 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4196 struct kmem_cache_node *n;
4197
4198 memcpy(s, static_cache, kmem_cache->object_size);
4199
4200 /*
4201 * This runs very early, and only the boot processor is supposed to be
4202 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4203 * IPIs around.
4204 */
4205 __flush_cpu_slab(s, smp_processor_id());
4206 for_each_kmem_cache_node(s, node, n) {
4207 struct page *p;
4208
4209 list_for_each_entry(p, &n->partial, slab_list)
4210 p->slab_cache = s;
4211
4212#ifdef CONFIG_SLUB_DEBUG
4213 list_for_each_entry(p, &n->full, slab_list)
4214 p->slab_cache = s;
4215#endif
4216 }
4217 slab_init_memcg_params(s);
4218 list_add(&s->list, &slab_caches);
4219 memcg_link_cache(s, NULL);
4220 return s;
4221}
4222
4223void __init kmem_cache_init(void)
4224{
4225 static __initdata struct kmem_cache boot_kmem_cache,
4226 boot_kmem_cache_node;
4227
4228 if (debug_guardpage_minorder())
4229 slub_max_order = 0;
4230
4231 kmem_cache_node = &boot_kmem_cache_node;
4232 kmem_cache = &boot_kmem_cache;
4233
4234 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4235 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4236
4237 register_hotmemory_notifier(&slab_memory_callback_nb);
4238
4239 /* Able to allocate the per node structures */
4240 slab_state = PARTIAL;
4241
4242 create_boot_cache(kmem_cache, "kmem_cache",
4243 offsetof(struct kmem_cache, node) +
4244 nr_node_ids * sizeof(struct kmem_cache_node *),
4245 SLAB_HWCACHE_ALIGN, 0, 0);
4246
4247 kmem_cache = bootstrap(&boot_kmem_cache);
4248 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
4249
4250 /* Now we can use the kmem_cache to allocate kmalloc slabs */
4251 setup_kmalloc_cache_index_table();
4252 create_kmalloc_caches(0);
4253
4254 /* Setup random freelists for each cache */
4255 init_freelist_randomization();
4256
4257 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
4258 slub_cpu_dead);
4259
4260 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
4261 cache_line_size(),
4262 slub_min_order, slub_max_order, slub_min_objects,
4263 nr_cpu_ids, nr_node_ids);
4264}
4265
4266void __init kmem_cache_init_late(void)
4267{
4268}
4269
4270struct kmem_cache *
4271__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
4272 slab_flags_t flags, void (*ctor)(void *))
4273{
4274 struct kmem_cache *s, *c;
4275
4276 s = find_mergeable(size, align, flags, name, ctor);
4277 if (s) {
4278 s->refcount++;
4279
4280 /*
4281 * Adjust the object sizes so that we clear
4282 * the complete object on kzalloc.
4283 */
4284 s->object_size = max(s->object_size, size);
4285 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
4286
4287 for_each_memcg_cache(c, s) {
4288 c->object_size = s->object_size;
4289 c->inuse = max(c->inuse, ALIGN(size, sizeof(void *)));
4290 }
4291
4292 if (sysfs_slab_alias(s, name)) {
4293 s->refcount--;
4294 s = NULL;
4295 }
4296 }
4297
4298 return s;
4299}
4300
4301int __kmem_cache_create(struct kmem_cache *s, slab_flags_t flags)
4302{
4303 int err;
4304
4305 err = kmem_cache_open(s, flags);
4306 if (err)
4307 return err;
4308
4309 /* Mutex is not taken during early boot */
4310 if (slab_state <= UP)
4311 return 0;
4312
4313 memcg_propagate_slab_attrs(s);
4314 err = sysfs_slab_add(s);
4315 if (err)
4316 __kmem_cache_release(s);
4317
4318 return err;
4319}
4320
4321void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4322{
4323 struct kmem_cache *s;
4324 void *ret;
4325
4326 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4327 return kmalloc_large(size, gfpflags);
4328
4329 s = kmalloc_slab(size, gfpflags);
4330
4331 if (unlikely(ZERO_OR_NULL_PTR(s)))
4332 return s;
4333
4334 ret = slab_alloc(s, gfpflags, caller);
4335
4336 /* Honor the call site pointer we received. */
4337 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4338
4339 return ret;
4340}
4341
4342#ifdef CONFIG_NUMA
4343void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4344 int node, unsigned long caller)
4345{
4346 struct kmem_cache *s;
4347 void *ret;
4348
4349 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4350 ret = kmalloc_large_node(size, gfpflags, node);
4351
4352 trace_kmalloc_node(caller, ret,
4353 size, PAGE_SIZE << get_order(size),
4354 gfpflags, node);
4355
4356 return ret;
4357 }
4358
4359 s = kmalloc_slab(size, gfpflags);
4360
4361 if (unlikely(ZERO_OR_NULL_PTR(s)))
4362 return s;
4363
4364 ret = slab_alloc_node(s, gfpflags, node, caller);
4365
4366 /* Honor the call site pointer we received. */
4367 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4368
4369 return ret;
4370}
4371#endif
4372
4373#ifdef CONFIG_SYSFS
4374static int count_inuse(struct page *page)
4375{
4376 return page->inuse;
4377}
4378
4379static int count_total(struct page *page)
4380{
4381 return page->objects;
4382}
4383#endif
4384
4385#ifdef CONFIG_SLUB_DEBUG
4386static int validate_slab(struct kmem_cache *s, struct page *page,
4387 unsigned long *map)
4388{
4389 void *p;
4390 void *addr = page_address(page);
4391
4392 if (!check_slab(s, page) ||
4393 !on_freelist(s, page, NULL))
4394 return 0;
4395
4396 /* Now we know that a valid freelist exists */
4397 bitmap_zero(map, page->objects);
4398
4399 get_map(s, page, map);
4400 for_each_object(p, s, addr, page->objects) {
4401 if (test_bit(slab_index(p, s, addr), map))
4402 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4403 return 0;
4404 }
4405
4406 for_each_object(p, s, addr, page->objects)
4407 if (!test_bit(slab_index(p, s, addr), map))
4408 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4409 return 0;
4410 return 1;
4411}
4412
4413static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4414 unsigned long *map)
4415{
4416 slab_lock(page);
4417 validate_slab(s, page, map);
4418 slab_unlock(page);
4419}
4420
4421static int validate_slab_node(struct kmem_cache *s,
4422 struct kmem_cache_node *n, unsigned long *map)
4423{
4424 unsigned long count = 0;
4425 struct page *page;
4426 unsigned long flags;
4427
4428 spin_lock_irqsave(&n->list_lock, flags);
4429
4430 list_for_each_entry(page, &n->partial, slab_list) {
4431 validate_slab_slab(s, page, map);
4432 count++;
4433 }
4434 if (count != n->nr_partial)
4435 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4436 s->name, count, n->nr_partial);
4437
4438 if (!(s->flags & SLAB_STORE_USER))
4439 goto out;
4440
4441 list_for_each_entry(page, &n->full, slab_list) {
4442 validate_slab_slab(s, page, map);
4443 count++;
4444 }
4445 if (count != atomic_long_read(&n->nr_slabs))
4446 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4447 s->name, count, atomic_long_read(&n->nr_slabs));
4448
4449out:
4450 spin_unlock_irqrestore(&n->list_lock, flags);
4451 return count;
4452}
4453
4454static long validate_slab_cache(struct kmem_cache *s)
4455{
4456 int node;
4457 unsigned long count = 0;
4458 struct kmem_cache_node *n;
4459 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4460
4461 if (!map)
4462 return -ENOMEM;
4463
4464 flush_all(s);
4465 for_each_kmem_cache_node(s, node, n)
4466 count += validate_slab_node(s, n, map);
4467 bitmap_free(map);
4468 return count;
4469}
4470/*
4471 * Generate lists of code addresses where slabcache objects are allocated
4472 * and freed.
4473 */
4474
4475struct location {
4476 unsigned long count;
4477 unsigned long addr;
4478 long long sum_time;
4479 long min_time;
4480 long max_time;
4481 long min_pid;
4482 long max_pid;
4483 DECLARE_BITMAP(cpus, NR_CPUS);
4484 nodemask_t nodes;
4485};
4486
4487struct loc_track {
4488 unsigned long max;
4489 unsigned long count;
4490 struct location *loc;
4491};
4492
4493static void free_loc_track(struct loc_track *t)
4494{
4495 if (t->max)
4496 free_pages((unsigned long)t->loc,
4497 get_order(sizeof(struct location) * t->max));
4498}
4499
4500static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4501{
4502 struct location *l;
4503 int order;
4504
4505 order = get_order(sizeof(struct location) * max);
4506
4507 l = (void *)__get_free_pages(flags, order);
4508 if (!l)
4509 return 0;
4510
4511 if (t->count) {
4512 memcpy(l, t->loc, sizeof(struct location) * t->count);
4513 free_loc_track(t);
4514 }
4515 t->max = max;
4516 t->loc = l;
4517 return 1;
4518}
4519
4520static int add_location(struct loc_track *t, struct kmem_cache *s,
4521 const struct track *track)
4522{
4523 long start, end, pos;
4524 struct location *l;
4525 unsigned long caddr;
4526 unsigned long age = jiffies - track->when;
4527
4528 start = -1;
4529 end = t->count;
4530
4531 for ( ; ; ) {
4532 pos = start + (end - start + 1) / 2;
4533
4534 /*
4535 * There is nothing at "end". If we end up there
4536 * we need to add something to before end.
4537 */
4538 if (pos == end)
4539 break;
4540
4541 caddr = t->loc[pos].addr;
4542 if (track->addr == caddr) {
4543
4544 l = &t->loc[pos];
4545 l->count++;
4546 if (track->when) {
4547 l->sum_time += age;
4548 if (age < l->min_time)
4549 l->min_time = age;
4550 if (age > l->max_time)
4551 l->max_time = age;
4552
4553 if (track->pid < l->min_pid)
4554 l->min_pid = track->pid;
4555 if (track->pid > l->max_pid)
4556 l->max_pid = track->pid;
4557
4558 cpumask_set_cpu(track->cpu,
4559 to_cpumask(l->cpus));
4560 }
4561 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4562 return 1;
4563 }
4564
4565 if (track->addr < caddr)
4566 end = pos;
4567 else
4568 start = pos;
4569 }
4570
4571 /*
4572 * Not found. Insert new tracking element.
4573 */
4574 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4575 return 0;
4576
4577 l = t->loc + pos;
4578 if (pos < t->count)
4579 memmove(l + 1, l,
4580 (t->count - pos) * sizeof(struct location));
4581 t->count++;
4582 l->count = 1;
4583 l->addr = track->addr;
4584 l->sum_time = age;
4585 l->min_time = age;
4586 l->max_time = age;
4587 l->min_pid = track->pid;
4588 l->max_pid = track->pid;
4589 cpumask_clear(to_cpumask(l->cpus));
4590 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4591 nodes_clear(l->nodes);
4592 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4593 return 1;
4594}
4595
4596static void process_slab(struct loc_track *t, struct kmem_cache *s,
4597 struct page *page, enum track_item alloc,
4598 unsigned long *map)
4599{
4600 void *addr = page_address(page);
4601 void *p;
4602
4603 bitmap_zero(map, page->objects);
4604 get_map(s, page, map);
4605
4606 for_each_object(p, s, addr, page->objects)
4607 if (!test_bit(slab_index(p, s, addr), map))
4608 add_location(t, s, get_track(s, p, alloc));
4609}
4610
4611static int list_locations(struct kmem_cache *s, char *buf,
4612 enum track_item alloc)
4613{
4614 int len = 0;
4615 unsigned long i;
4616 struct loc_track t = { 0, 0, NULL };
4617 int node;
4618 struct kmem_cache_node *n;
4619 unsigned long *map = bitmap_alloc(oo_objects(s->max), GFP_KERNEL);
4620
4621 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4622 GFP_KERNEL)) {
4623 bitmap_free(map);
4624 return sprintf(buf, "Out of memory\n");
4625 }
4626 /* Push back cpu slabs */
4627 flush_all(s);
4628
4629 for_each_kmem_cache_node(s, node, n) {
4630 unsigned long flags;
4631 struct page *page;
4632
4633 if (!atomic_long_read(&n->nr_slabs))
4634 continue;
4635
4636 spin_lock_irqsave(&n->list_lock, flags);
4637 list_for_each_entry(page, &n->partial, slab_list)
4638 process_slab(&t, s, page, alloc, map);
4639 list_for_each_entry(page, &n->full, slab_list)
4640 process_slab(&t, s, page, alloc, map);
4641 spin_unlock_irqrestore(&n->list_lock, flags);
4642 }
4643
4644 for (i = 0; i < t.count; i++) {
4645 struct location *l = &t.loc[i];
4646
4647 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4648 break;
4649 len += sprintf(buf + len, "%7ld ", l->count);
4650
4651 if (l->addr)
4652 len += sprintf(buf + len, "%pS", (void *)l->addr);
4653 else
4654 len += sprintf(buf + len, "<not-available>");
4655
4656 if (l->sum_time != l->min_time) {
4657 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4658 l->min_time,
4659 (long)div_u64(l->sum_time, l->count),
4660 l->max_time);
4661 } else
4662 len += sprintf(buf + len, " age=%ld",
4663 l->min_time);
4664
4665 if (l->min_pid != l->max_pid)
4666 len += sprintf(buf + len, " pid=%ld-%ld",
4667 l->min_pid, l->max_pid);
4668 else
4669 len += sprintf(buf + len, " pid=%ld",
4670 l->min_pid);
4671
4672 if (num_online_cpus() > 1 &&
4673 !cpumask_empty(to_cpumask(l->cpus)) &&
4674 len < PAGE_SIZE - 60)
4675 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4676 " cpus=%*pbl",
4677 cpumask_pr_args(to_cpumask(l->cpus)));
4678
4679 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4680 len < PAGE_SIZE - 60)
4681 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4682 " nodes=%*pbl",
4683 nodemask_pr_args(&l->nodes));
4684
4685 len += sprintf(buf + len, "\n");
4686 }
4687
4688 free_loc_track(&t);
4689 bitmap_free(map);
4690 if (!t.count)
4691 len += sprintf(buf, "No data\n");
4692 return len;
4693}
4694#endif /* CONFIG_SLUB_DEBUG */
4695
4696#ifdef SLUB_RESILIENCY_TEST
4697static void __init resiliency_test(void)
4698{
4699 u8 *p;
4700 int type = KMALLOC_NORMAL;
4701
4702 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4703
4704 pr_err("SLUB resiliency testing\n");
4705 pr_err("-----------------------\n");
4706 pr_err("A. Corruption after allocation\n");
4707
4708 p = kzalloc(16, GFP_KERNEL);
4709 p[16] = 0x12;
4710 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4711 p + 16);
4712
4713 validate_slab_cache(kmalloc_caches[type][4]);
4714
4715 /* Hmmm... The next two are dangerous */
4716 p = kzalloc(32, GFP_KERNEL);
4717 p[32 + sizeof(void *)] = 0x34;
4718 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4719 p);
4720 pr_err("If allocated object is overwritten then not detectable\n\n");
4721
4722 validate_slab_cache(kmalloc_caches[type][5]);
4723 p = kzalloc(64, GFP_KERNEL);
4724 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4725 *p = 0x56;
4726 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4727 p);
4728 pr_err("If allocated object is overwritten then not detectable\n\n");
4729 validate_slab_cache(kmalloc_caches[type][6]);
4730
4731 pr_err("\nB. Corruption after free\n");
4732 p = kzalloc(128, GFP_KERNEL);
4733 kfree(p);
4734 *p = 0x78;
4735 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4736 validate_slab_cache(kmalloc_caches[type][7]);
4737
4738 p = kzalloc(256, GFP_KERNEL);
4739 kfree(p);
4740 p[50] = 0x9a;
4741 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4742 validate_slab_cache(kmalloc_caches[type][8]);
4743
4744 p = kzalloc(512, GFP_KERNEL);
4745 kfree(p);
4746 p[512] = 0xab;
4747 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4748 validate_slab_cache(kmalloc_caches[type][9]);
4749}
4750#else
4751#ifdef CONFIG_SYSFS
4752static void resiliency_test(void) {};
4753#endif
4754#endif /* SLUB_RESILIENCY_TEST */
4755
4756#ifdef CONFIG_SYSFS
4757enum slab_stat_type {
4758 SL_ALL, /* All slabs */
4759 SL_PARTIAL, /* Only partially allocated slabs */
4760 SL_CPU, /* Only slabs used for cpu caches */
4761 SL_OBJECTS, /* Determine allocated objects not slabs */
4762 SL_TOTAL /* Determine object capacity not slabs */
4763};
4764
4765#define SO_ALL (1 << SL_ALL)
4766#define SO_PARTIAL (1 << SL_PARTIAL)
4767#define SO_CPU (1 << SL_CPU)
4768#define SO_OBJECTS (1 << SL_OBJECTS)
4769#define SO_TOTAL (1 << SL_TOTAL)
4770
4771#ifdef CONFIG_MEMCG
4772static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4773
4774static int __init setup_slub_memcg_sysfs(char *str)
4775{
4776 int v;
4777
4778 if (get_option(&str, &v) > 0)
4779 memcg_sysfs_enabled = v;
4780
4781 return 1;
4782}
4783
4784__setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4785#endif
4786
4787static ssize_t show_slab_objects(struct kmem_cache *s,
4788 char *buf, unsigned long flags)
4789{
4790 unsigned long total = 0;
4791 int node;
4792 int x;
4793 unsigned long *nodes;
4794
4795 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
4796 if (!nodes)
4797 return -ENOMEM;
4798
4799 if (flags & SO_CPU) {
4800 int cpu;
4801
4802 for_each_possible_cpu(cpu) {
4803 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4804 cpu);
4805 int node;
4806 struct page *page;
4807
4808 page = READ_ONCE(c->page);
4809 if (!page)
4810 continue;
4811
4812 node = page_to_nid(page);
4813 if (flags & SO_TOTAL)
4814 x = page->objects;
4815 else if (flags & SO_OBJECTS)
4816 x = page->inuse;
4817 else
4818 x = 1;
4819
4820 total += x;
4821 nodes[node] += x;
4822
4823 page = slub_percpu_partial_read_once(c);
4824 if (page) {
4825 node = page_to_nid(page);
4826 if (flags & SO_TOTAL)
4827 WARN_ON_ONCE(1);
4828 else if (flags & SO_OBJECTS)
4829 WARN_ON_ONCE(1);
4830 else
4831 x = page->pages;
4832 total += x;
4833 nodes[node] += x;
4834 }
4835 }
4836 }
4837
4838 /*
4839 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
4840 * already held which will conflict with an existing lock order:
4841 *
4842 * mem_hotplug_lock->slab_mutex->kernfs_mutex
4843 *
4844 * We don't really need mem_hotplug_lock (to hold off
4845 * slab_mem_going_offline_callback) here because slab's memory hot
4846 * unplug code doesn't destroy the kmem_cache->node[] data.
4847 */
4848
4849#ifdef CONFIG_SLUB_DEBUG
4850 if (flags & SO_ALL) {
4851 struct kmem_cache_node *n;
4852
4853 for_each_kmem_cache_node(s, node, n) {
4854
4855 if (flags & SO_TOTAL)
4856 x = atomic_long_read(&n->total_objects);
4857 else if (flags & SO_OBJECTS)
4858 x = atomic_long_read(&n->total_objects) -
4859 count_partial(n, count_free);
4860 else
4861 x = atomic_long_read(&n->nr_slabs);
4862 total += x;
4863 nodes[node] += x;
4864 }
4865
4866 } else
4867#endif
4868 if (flags & SO_PARTIAL) {
4869 struct kmem_cache_node *n;
4870
4871 for_each_kmem_cache_node(s, node, n) {
4872 if (flags & SO_TOTAL)
4873 x = count_partial(n, count_total);
4874 else if (flags & SO_OBJECTS)
4875 x = count_partial(n, count_inuse);
4876 else
4877 x = n->nr_partial;
4878 total += x;
4879 nodes[node] += x;
4880 }
4881 }
4882 x = sprintf(buf, "%lu", total);
4883#ifdef CONFIG_NUMA
4884 for (node = 0; node < nr_node_ids; node++)
4885 if (nodes[node])
4886 x += sprintf(buf + x, " N%d=%lu",
4887 node, nodes[node]);
4888#endif
4889 kfree(nodes);
4890 return x + sprintf(buf + x, "\n");
4891}
4892
4893#ifdef CONFIG_SLUB_DEBUG
4894static int any_slab_objects(struct kmem_cache *s)
4895{
4896 int node;
4897 struct kmem_cache_node *n;
4898
4899 for_each_kmem_cache_node(s, node, n)
4900 if (atomic_long_read(&n->total_objects))
4901 return 1;
4902
4903 return 0;
4904}
4905#endif
4906
4907#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4908#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4909
4910struct slab_attribute {
4911 struct attribute attr;
4912 ssize_t (*show)(struct kmem_cache *s, char *buf);
4913 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4914};
4915
4916#define SLAB_ATTR_RO(_name) \
4917 static struct slab_attribute _name##_attr = \
4918 __ATTR(_name, 0400, _name##_show, NULL)
4919
4920#define SLAB_ATTR(_name) \
4921 static struct slab_attribute _name##_attr = \
4922 __ATTR(_name, 0600, _name##_show, _name##_store)
4923
4924static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4925{
4926 return sprintf(buf, "%u\n", s->size);
4927}
4928SLAB_ATTR_RO(slab_size);
4929
4930static ssize_t align_show(struct kmem_cache *s, char *buf)
4931{
4932 return sprintf(buf, "%u\n", s->align);
4933}
4934SLAB_ATTR_RO(align);
4935
4936static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4937{
4938 return sprintf(buf, "%u\n", s->object_size);
4939}
4940SLAB_ATTR_RO(object_size);
4941
4942static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4943{
4944 return sprintf(buf, "%u\n", oo_objects(s->oo));
4945}
4946SLAB_ATTR_RO(objs_per_slab);
4947
4948static ssize_t order_store(struct kmem_cache *s,
4949 const char *buf, size_t length)
4950{
4951 unsigned int order;
4952 int err;
4953
4954 err = kstrtouint(buf, 10, &order);
4955 if (err)
4956 return err;
4957
4958 if (order > slub_max_order || order < slub_min_order)
4959 return -EINVAL;
4960
4961 calculate_sizes(s, order);
4962 return length;
4963}
4964
4965static ssize_t order_show(struct kmem_cache *s, char *buf)
4966{
4967 return sprintf(buf, "%u\n", oo_order(s->oo));
4968}
4969SLAB_ATTR(order);
4970
4971static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4972{
4973 return sprintf(buf, "%lu\n", s->min_partial);
4974}
4975
4976static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4977 size_t length)
4978{
4979 unsigned long min;
4980 int err;
4981
4982 err = kstrtoul(buf, 10, &min);
4983 if (err)
4984 return err;
4985
4986 set_min_partial(s, min);
4987 return length;
4988}
4989SLAB_ATTR(min_partial);
4990
4991static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4992{
4993 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4994}
4995
4996static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4997 size_t length)
4998{
4999 unsigned int objects;
5000 int err;
5001
5002 err = kstrtouint(buf, 10, &objects);
5003 if (err)
5004 return err;
5005 if (objects && !kmem_cache_has_cpu_partial(s))
5006 return -EINVAL;
5007
5008 slub_set_cpu_partial(s, objects);
5009 flush_all(s);
5010 return length;
5011}
5012SLAB_ATTR(cpu_partial);
5013
5014static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5015{
5016 if (!s->ctor)
5017 return 0;
5018 return sprintf(buf, "%pS\n", s->ctor);
5019}
5020SLAB_ATTR_RO(ctor);
5021
5022static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5023{
5024 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5025}
5026SLAB_ATTR_RO(aliases);
5027
5028static ssize_t partial_show(struct kmem_cache *s, char *buf)
5029{
5030 return show_slab_objects(s, buf, SO_PARTIAL);
5031}
5032SLAB_ATTR_RO(partial);
5033
5034static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5035{
5036 return show_slab_objects(s, buf, SO_CPU);
5037}
5038SLAB_ATTR_RO(cpu_slabs);
5039
5040static ssize_t objects_show(struct kmem_cache *s, char *buf)
5041{
5042 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5043}
5044SLAB_ATTR_RO(objects);
5045
5046static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5047{
5048 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5049}
5050SLAB_ATTR_RO(objects_partial);
5051
5052static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5053{
5054 int objects = 0;
5055 int pages = 0;
5056 int cpu;
5057 int len;
5058
5059 for_each_online_cpu(cpu) {
5060 struct page *page;
5061
5062 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5063
5064 if (page) {
5065 pages += page->pages;
5066 objects += page->pobjects;
5067 }
5068 }
5069
5070 len = sprintf(buf, "%d(%d)", objects, pages);
5071
5072#ifdef CONFIG_SMP
5073 for_each_online_cpu(cpu) {
5074 struct page *page;
5075
5076 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5077
5078 if (page && len < PAGE_SIZE - 20)
5079 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5080 page->pobjects, page->pages);
5081 }
5082#endif
5083 return len + sprintf(buf + len, "\n");
5084}
5085SLAB_ATTR_RO(slabs_cpu_partial);
5086
5087static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5088{
5089 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5090}
5091
5092static ssize_t reclaim_account_store(struct kmem_cache *s,
5093 const char *buf, size_t length)
5094{
5095 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5096 if (buf[0] == '1')
5097 s->flags |= SLAB_RECLAIM_ACCOUNT;
5098 return length;
5099}
5100SLAB_ATTR(reclaim_account);
5101
5102static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5103{
5104 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5105}
5106SLAB_ATTR_RO(hwcache_align);
5107
5108#ifdef CONFIG_ZONE_DMA
5109static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5110{
5111 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5112}
5113SLAB_ATTR_RO(cache_dma);
5114#endif
5115
5116static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5117{
5118 return sprintf(buf, "%u\n", s->usersize);
5119}
5120SLAB_ATTR_RO(usersize);
5121
5122static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5123{
5124 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5125}
5126SLAB_ATTR_RO(destroy_by_rcu);
5127
5128#ifdef CONFIG_SLUB_DEBUG
5129static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5130{
5131 return show_slab_objects(s, buf, SO_ALL);
5132}
5133SLAB_ATTR_RO(slabs);
5134
5135static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5136{
5137 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5138}
5139SLAB_ATTR_RO(total_objects);
5140
5141static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5142{
5143 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5144}
5145
5146static ssize_t sanity_checks_store(struct kmem_cache *s,
5147 const char *buf, size_t length)
5148{
5149 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5150 if (buf[0] == '1') {
5151 s->flags &= ~__CMPXCHG_DOUBLE;
5152 s->flags |= SLAB_CONSISTENCY_CHECKS;
5153 }
5154 return length;
5155}
5156SLAB_ATTR(sanity_checks);
5157
5158static ssize_t trace_show(struct kmem_cache *s, char *buf)
5159{
5160 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5161}
5162
5163static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5164 size_t length)
5165{
5166 /*
5167 * Tracing a merged cache is going to give confusing results
5168 * as well as cause other issues like converting a mergeable
5169 * cache into an umergeable one.
5170 */
5171 if (s->refcount > 1)
5172 return -EINVAL;
5173
5174 s->flags &= ~SLAB_TRACE;
5175 if (buf[0] == '1') {
5176 s->flags &= ~__CMPXCHG_DOUBLE;
5177 s->flags |= SLAB_TRACE;
5178 }
5179 return length;
5180}
5181SLAB_ATTR(trace);
5182
5183static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5184{
5185 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5186}
5187
5188static ssize_t red_zone_store(struct kmem_cache *s,
5189 const char *buf, size_t length)
5190{
5191 if (any_slab_objects(s))
5192 return -EBUSY;
5193
5194 s->flags &= ~SLAB_RED_ZONE;
5195 if (buf[0] == '1') {
5196 s->flags |= SLAB_RED_ZONE;
5197 }
5198 calculate_sizes(s, -1);
5199 return length;
5200}
5201SLAB_ATTR(red_zone);
5202
5203static ssize_t poison_show(struct kmem_cache *s, char *buf)
5204{
5205 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5206}
5207
5208static ssize_t poison_store(struct kmem_cache *s,
5209 const char *buf, size_t length)
5210{
5211 if (any_slab_objects(s))
5212 return -EBUSY;
5213
5214 s->flags &= ~SLAB_POISON;
5215 if (buf[0] == '1') {
5216 s->flags |= SLAB_POISON;
5217 }
5218 calculate_sizes(s, -1);
5219 return length;
5220}
5221SLAB_ATTR(poison);
5222
5223static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5224{
5225 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5226}
5227
5228static ssize_t store_user_store(struct kmem_cache *s,
5229 const char *buf, size_t length)
5230{
5231 if (any_slab_objects(s))
5232 return -EBUSY;
5233
5234 s->flags &= ~SLAB_STORE_USER;
5235 if (buf[0] == '1') {
5236 s->flags &= ~__CMPXCHG_DOUBLE;
5237 s->flags |= SLAB_STORE_USER;
5238 }
5239 calculate_sizes(s, -1);
5240 return length;
5241}
5242SLAB_ATTR(store_user);
5243
5244static ssize_t validate_show(struct kmem_cache *s, char *buf)
5245{
5246 return 0;
5247}
5248
5249static ssize_t validate_store(struct kmem_cache *s,
5250 const char *buf, size_t length)
5251{
5252 int ret = -EINVAL;
5253
5254 if (buf[0] == '1') {
5255 ret = validate_slab_cache(s);
5256 if (ret >= 0)
5257 ret = length;
5258 }
5259 return ret;
5260}
5261SLAB_ATTR(validate);
5262
5263static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5264{
5265 if (!(s->flags & SLAB_STORE_USER))
5266 return -ENOSYS;
5267 return list_locations(s, buf, TRACK_ALLOC);
5268}
5269SLAB_ATTR_RO(alloc_calls);
5270
5271static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5272{
5273 if (!(s->flags & SLAB_STORE_USER))
5274 return -ENOSYS;
5275 return list_locations(s, buf, TRACK_FREE);
5276}
5277SLAB_ATTR_RO(free_calls);
5278#endif /* CONFIG_SLUB_DEBUG */
5279
5280#ifdef CONFIG_FAILSLAB
5281static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5282{
5283 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5284}
5285
5286static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5287 size_t length)
5288{
5289 if (s->refcount > 1)
5290 return -EINVAL;
5291
5292 s->flags &= ~SLAB_FAILSLAB;
5293 if (buf[0] == '1')
5294 s->flags |= SLAB_FAILSLAB;
5295 return length;
5296}
5297SLAB_ATTR(failslab);
5298#endif
5299
5300static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5301{
5302 return 0;
5303}
5304
5305static ssize_t shrink_store(struct kmem_cache *s,
5306 const char *buf, size_t length)
5307{
5308 if (buf[0] == '1')
5309 kmem_cache_shrink_all(s);
5310 else
5311 return -EINVAL;
5312 return length;
5313}
5314SLAB_ATTR(shrink);
5315
5316#ifdef CONFIG_NUMA
5317static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5318{
5319 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5320}
5321
5322static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5323 const char *buf, size_t length)
5324{
5325 unsigned int ratio;
5326 int err;
5327
5328 err = kstrtouint(buf, 10, &ratio);
5329 if (err)
5330 return err;
5331 if (ratio > 100)
5332 return -ERANGE;
5333
5334 s->remote_node_defrag_ratio = ratio * 10;
5335
5336 return length;
5337}
5338SLAB_ATTR(remote_node_defrag_ratio);
5339#endif
5340
5341#ifdef CONFIG_SLUB_STATS
5342static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5343{
5344 unsigned long sum = 0;
5345 int cpu;
5346 int len;
5347 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
5348
5349 if (!data)
5350 return -ENOMEM;
5351
5352 for_each_online_cpu(cpu) {
5353 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5354
5355 data[cpu] = x;
5356 sum += x;
5357 }
5358
5359 len = sprintf(buf, "%lu", sum);
5360
5361#ifdef CONFIG_SMP
5362 for_each_online_cpu(cpu) {
5363 if (data[cpu] && len < PAGE_SIZE - 20)
5364 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5365 }
5366#endif
5367 kfree(data);
5368 return len + sprintf(buf + len, "\n");
5369}
5370
5371static void clear_stat(struct kmem_cache *s, enum stat_item si)
5372{
5373 int cpu;
5374
5375 for_each_online_cpu(cpu)
5376 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5377}
5378
5379#define STAT_ATTR(si, text) \
5380static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5381{ \
5382 return show_stat(s, buf, si); \
5383} \
5384static ssize_t text##_store(struct kmem_cache *s, \
5385 const char *buf, size_t length) \
5386{ \
5387 if (buf[0] != '0') \
5388 return -EINVAL; \
5389 clear_stat(s, si); \
5390 return length; \
5391} \
5392SLAB_ATTR(text); \
5393
5394STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5395STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5396STAT_ATTR(FREE_FASTPATH, free_fastpath);
5397STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5398STAT_ATTR(FREE_FROZEN, free_frozen);
5399STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5400STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5401STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5402STAT_ATTR(ALLOC_SLAB, alloc_slab);
5403STAT_ATTR(ALLOC_REFILL, alloc_refill);
5404STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5405STAT_ATTR(FREE_SLAB, free_slab);
5406STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5407STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5408STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5409STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5410STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5411STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5412STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5413STAT_ATTR(ORDER_FALLBACK, order_fallback);
5414STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5415STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5416STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5417STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5418STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5419STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5420#endif /* CONFIG_SLUB_STATS */
5421
5422static struct attribute *slab_attrs[] = {
5423 &slab_size_attr.attr,
5424 &object_size_attr.attr,
5425 &objs_per_slab_attr.attr,
5426 &order_attr.attr,
5427 &min_partial_attr.attr,
5428 &cpu_partial_attr.attr,
5429 &objects_attr.attr,
5430 &objects_partial_attr.attr,
5431 &partial_attr.attr,
5432 &cpu_slabs_attr.attr,
5433 &ctor_attr.attr,
5434 &aliases_attr.attr,
5435 &align_attr.attr,
5436 &hwcache_align_attr.attr,
5437 &reclaim_account_attr.attr,
5438 &destroy_by_rcu_attr.attr,
5439 &shrink_attr.attr,
5440 &slabs_cpu_partial_attr.attr,
5441#ifdef CONFIG_SLUB_DEBUG
5442 &total_objects_attr.attr,
5443 &slabs_attr.attr,
5444 &sanity_checks_attr.attr,
5445 &trace_attr.attr,
5446 &red_zone_attr.attr,
5447 &poison_attr.attr,
5448 &store_user_attr.attr,
5449 &validate_attr.attr,
5450 &alloc_calls_attr.attr,
5451 &free_calls_attr.attr,
5452#endif
5453#ifdef CONFIG_ZONE_DMA
5454 &cache_dma_attr.attr,
5455#endif
5456#ifdef CONFIG_NUMA
5457 &remote_node_defrag_ratio_attr.attr,
5458#endif
5459#ifdef CONFIG_SLUB_STATS
5460 &alloc_fastpath_attr.attr,
5461 &alloc_slowpath_attr.attr,
5462 &free_fastpath_attr.attr,
5463 &free_slowpath_attr.attr,
5464 &free_frozen_attr.attr,
5465 &free_add_partial_attr.attr,
5466 &free_remove_partial_attr.attr,
5467 &alloc_from_partial_attr.attr,
5468 &alloc_slab_attr.attr,
5469 &alloc_refill_attr.attr,
5470 &alloc_node_mismatch_attr.attr,
5471 &free_slab_attr.attr,
5472 &cpuslab_flush_attr.attr,
5473 &deactivate_full_attr.attr,
5474 &deactivate_empty_attr.attr,
5475 &deactivate_to_head_attr.attr,
5476 &deactivate_to_tail_attr.attr,
5477 &deactivate_remote_frees_attr.attr,
5478 &deactivate_bypass_attr.attr,
5479 &order_fallback_attr.attr,
5480 &cmpxchg_double_fail_attr.attr,
5481 &cmpxchg_double_cpu_fail_attr.attr,
5482 &cpu_partial_alloc_attr.attr,
5483 &cpu_partial_free_attr.attr,
5484 &cpu_partial_node_attr.attr,
5485 &cpu_partial_drain_attr.attr,
5486#endif
5487#ifdef CONFIG_FAILSLAB
5488 &failslab_attr.attr,
5489#endif
5490 &usersize_attr.attr,
5491
5492 NULL
5493};
5494
5495static const struct attribute_group slab_attr_group = {
5496 .attrs = slab_attrs,
5497};
5498
5499static ssize_t slab_attr_show(struct kobject *kobj,
5500 struct attribute *attr,
5501 char *buf)
5502{
5503 struct slab_attribute *attribute;
5504 struct kmem_cache *s;
5505 int err;
5506
5507 attribute = to_slab_attr(attr);
5508 s = to_slab(kobj);
5509
5510 if (!attribute->show)
5511 return -EIO;
5512
5513 err = attribute->show(s, buf);
5514
5515 return err;
5516}
5517
5518static ssize_t slab_attr_store(struct kobject *kobj,
5519 struct attribute *attr,
5520 const char *buf, size_t len)
5521{
5522 struct slab_attribute *attribute;
5523 struct kmem_cache *s;
5524 int err;
5525
5526 attribute = to_slab_attr(attr);
5527 s = to_slab(kobj);
5528
5529 if (!attribute->store)
5530 return -EIO;
5531
5532 err = attribute->store(s, buf, len);
5533#ifdef CONFIG_MEMCG
5534 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5535 struct kmem_cache *c;
5536
5537 mutex_lock(&slab_mutex);
5538 if (s->max_attr_size < len)
5539 s->max_attr_size = len;
5540
5541 /*
5542 * This is a best effort propagation, so this function's return
5543 * value will be determined by the parent cache only. This is
5544 * basically because not all attributes will have a well
5545 * defined semantics for rollbacks - most of the actions will
5546 * have permanent effects.
5547 *
5548 * Returning the error value of any of the children that fail
5549 * is not 100 % defined, in the sense that users seeing the
5550 * error code won't be able to know anything about the state of
5551 * the cache.
5552 *
5553 * Only returning the error code for the parent cache at least
5554 * has well defined semantics. The cache being written to
5555 * directly either failed or succeeded, in which case we loop
5556 * through the descendants with best-effort propagation.
5557 */
5558 for_each_memcg_cache(c, s)
5559 attribute->store(c, buf, len);
5560 mutex_unlock(&slab_mutex);
5561 }
5562#endif
5563 return err;
5564}
5565
5566static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5567{
5568#ifdef CONFIG_MEMCG
5569 int i;
5570 char *buffer = NULL;
5571 struct kmem_cache *root_cache;
5572
5573 if (is_root_cache(s))
5574 return;
5575
5576 root_cache = s->memcg_params.root_cache;
5577
5578 /*
5579 * This mean this cache had no attribute written. Therefore, no point
5580 * in copying default values around
5581 */
5582 if (!root_cache->max_attr_size)
5583 return;
5584
5585 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5586 char mbuf[64];
5587 char *buf;
5588 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5589 ssize_t len;
5590
5591 if (!attr || !attr->store || !attr->show)
5592 continue;
5593
5594 /*
5595 * It is really bad that we have to allocate here, so we will
5596 * do it only as a fallback. If we actually allocate, though,
5597 * we can just use the allocated buffer until the end.
5598 *
5599 * Most of the slub attributes will tend to be very small in
5600 * size, but sysfs allows buffers up to a page, so they can
5601 * theoretically happen.
5602 */
5603 if (buffer)
5604 buf = buffer;
5605 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5606 buf = mbuf;
5607 else {
5608 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5609 if (WARN_ON(!buffer))
5610 continue;
5611 buf = buffer;
5612 }
5613
5614 len = attr->show(root_cache, buf);
5615 if (len > 0)
5616 attr->store(s, buf, len);
5617 }
5618
5619 if (buffer)
5620 free_page((unsigned long)buffer);
5621#endif /* CONFIG_MEMCG */
5622}
5623
5624static void kmem_cache_release(struct kobject *k)
5625{
5626 slab_kmem_cache_release(to_slab(k));
5627}
5628
5629static const struct sysfs_ops slab_sysfs_ops = {
5630 .show = slab_attr_show,
5631 .store = slab_attr_store,
5632};
5633
5634static struct kobj_type slab_ktype = {
5635 .sysfs_ops = &slab_sysfs_ops,
5636 .release = kmem_cache_release,
5637};
5638
5639static int uevent_filter(struct kset *kset, struct kobject *kobj)
5640{
5641 struct kobj_type *ktype = get_ktype(kobj);
5642
5643 if (ktype == &slab_ktype)
5644 return 1;
5645 return 0;
5646}
5647
5648static const struct kset_uevent_ops slab_uevent_ops = {
5649 .filter = uevent_filter,
5650};
5651
5652static struct kset *slab_kset;
5653
5654static inline struct kset *cache_kset(struct kmem_cache *s)
5655{
5656#ifdef CONFIG_MEMCG
5657 if (!is_root_cache(s))
5658 return s->memcg_params.root_cache->memcg_kset;
5659#endif
5660 return slab_kset;
5661}
5662
5663#define ID_STR_LENGTH 64
5664
5665/* Create a unique string id for a slab cache:
5666 *
5667 * Format :[flags-]size
5668 */
5669static char *create_unique_id(struct kmem_cache *s)
5670{
5671 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5672 char *p = name;
5673
5674 BUG_ON(!name);
5675
5676 *p++ = ':';
5677 /*
5678 * First flags affecting slabcache operations. We will only
5679 * get here for aliasable slabs so we do not need to support
5680 * too many flags. The flags here must cover all flags that
5681 * are matched during merging to guarantee that the id is
5682 * unique.
5683 */
5684 if (s->flags & SLAB_CACHE_DMA)
5685 *p++ = 'd';
5686 if (s->flags & SLAB_CACHE_DMA32)
5687 *p++ = 'D';
5688 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5689 *p++ = 'a';
5690 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5691 *p++ = 'F';
5692 if (s->flags & SLAB_ACCOUNT)
5693 *p++ = 'A';
5694 if (p != name + 1)
5695 *p++ = '-';
5696 p += sprintf(p, "%07u", s->size);
5697
5698 BUG_ON(p > name + ID_STR_LENGTH - 1);
5699 return name;
5700}
5701
5702static void sysfs_slab_remove_workfn(struct work_struct *work)
5703{
5704 struct kmem_cache *s =
5705 container_of(work, struct kmem_cache, kobj_remove_work);
5706
5707 if (!s->kobj.state_in_sysfs)
5708 /*
5709 * For a memcg cache, this may be called during
5710 * deactivation and again on shutdown. Remove only once.
5711 * A cache is never shut down before deactivation is
5712 * complete, so no need to worry about synchronization.
5713 */
5714 goto out;
5715
5716#ifdef CONFIG_MEMCG
5717 kset_unregister(s->memcg_kset);
5718#endif
5719 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5720out:
5721 kobject_put(&s->kobj);
5722}
5723
5724static int sysfs_slab_add(struct kmem_cache *s)
5725{
5726 int err;
5727 const char *name;
5728 struct kset *kset = cache_kset(s);
5729 int unmergeable = slab_unmergeable(s);
5730
5731 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5732
5733 if (!kset) {
5734 kobject_init(&s->kobj, &slab_ktype);
5735 return 0;
5736 }
5737
5738 if (!unmergeable && disable_higher_order_debug &&
5739 (slub_debug & DEBUG_METADATA_FLAGS))
5740 unmergeable = 1;
5741
5742 if (unmergeable) {
5743 /*
5744 * Slabcache can never be merged so we can use the name proper.
5745 * This is typically the case for debug situations. In that
5746 * case we can catch duplicate names easily.
5747 */
5748 sysfs_remove_link(&slab_kset->kobj, s->name);
5749 name = s->name;
5750 } else {
5751 /*
5752 * Create a unique name for the slab as a target
5753 * for the symlinks.
5754 */
5755 name = create_unique_id(s);
5756 }
5757
5758 s->kobj.kset = kset;
5759 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5760 if (err)
5761 goto out;
5762
5763 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5764 if (err)
5765 goto out_del_kobj;
5766
5767#ifdef CONFIG_MEMCG
5768 if (is_root_cache(s) && memcg_sysfs_enabled) {
5769 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5770 if (!s->memcg_kset) {
5771 err = -ENOMEM;
5772 goto out_del_kobj;
5773 }
5774 }
5775#endif
5776
5777 kobject_uevent(&s->kobj, KOBJ_ADD);
5778 if (!unmergeable) {
5779 /* Setup first alias */
5780 sysfs_slab_alias(s, s->name);
5781 }
5782out:
5783 if (!unmergeable)
5784 kfree(name);
5785 return err;
5786out_del_kobj:
5787 kobject_del(&s->kobj);
5788 goto out;
5789}
5790
5791static void sysfs_slab_remove(struct kmem_cache *s)
5792{
5793 if (slab_state < FULL)
5794 /*
5795 * Sysfs has not been setup yet so no need to remove the
5796 * cache from sysfs.
5797 */
5798 return;
5799
5800 kobject_get(&s->kobj);
5801 schedule_work(&s->kobj_remove_work);
5802}
5803
5804void sysfs_slab_unlink(struct kmem_cache *s)
5805{
5806 if (slab_state >= FULL)
5807 kobject_del(&s->kobj);
5808}
5809
5810void sysfs_slab_release(struct kmem_cache *s)
5811{
5812 if (slab_state >= FULL)
5813 kobject_put(&s->kobj);
5814}
5815
5816/*
5817 * Need to buffer aliases during bootup until sysfs becomes
5818 * available lest we lose that information.
5819 */
5820struct saved_alias {
5821 struct kmem_cache *s;
5822 const char *name;
5823 struct saved_alias *next;
5824};
5825
5826static struct saved_alias *alias_list;
5827
5828static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5829{
5830 struct saved_alias *al;
5831
5832 if (slab_state == FULL) {
5833 /*
5834 * If we have a leftover link then remove it.
5835 */
5836 sysfs_remove_link(&slab_kset->kobj, name);
5837 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5838 }
5839
5840 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5841 if (!al)
5842 return -ENOMEM;
5843
5844 al->s = s;
5845 al->name = name;
5846 al->next = alias_list;
5847 alias_list = al;
5848 return 0;
5849}
5850
5851static int __init slab_sysfs_init(void)
5852{
5853 struct kmem_cache *s;
5854 int err;
5855
5856 mutex_lock(&slab_mutex);
5857
5858 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5859 if (!slab_kset) {
5860 mutex_unlock(&slab_mutex);
5861 pr_err("Cannot register slab subsystem.\n");
5862 return -ENOSYS;
5863 }
5864
5865 slab_state = FULL;
5866
5867 list_for_each_entry(s, &slab_caches, list) {
5868 err = sysfs_slab_add(s);
5869 if (err)
5870 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5871 s->name);
5872 }
5873
5874 while (alias_list) {
5875 struct saved_alias *al = alias_list;
5876
5877 alias_list = alias_list->next;
5878 err = sysfs_slab_alias(al->s, al->name);
5879 if (err)
5880 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5881 al->name);
5882 kfree(al);
5883 }
5884
5885 mutex_unlock(&slab_mutex);
5886 resiliency_test();
5887 return 0;
5888}
5889
5890__initcall(slab_sysfs_init);
5891#endif /* CONFIG_SYSFS */
5892
5893/*
5894 * The /proc/slabinfo ABI
5895 */
5896#ifdef CONFIG_SLUB_DEBUG
5897void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5898{
5899 unsigned long nr_slabs = 0;
5900 unsigned long nr_objs = 0;
5901 unsigned long nr_free = 0;
5902 int node;
5903 struct kmem_cache_node *n;
5904
5905 for_each_kmem_cache_node(s, node, n) {
5906 nr_slabs += node_nr_slabs(n);
5907 nr_objs += node_nr_objs(n);
5908 nr_free += count_partial(n, count_free);
5909 }
5910
5911 sinfo->active_objs = nr_objs - nr_free;
5912 sinfo->num_objs = nr_objs;
5913 sinfo->active_slabs = nr_slabs;
5914 sinfo->num_slabs = nr_slabs;
5915 sinfo->objects_per_slab = oo_objects(s->oo);
5916 sinfo->cache_order = oo_order(s->oo);
5917}
5918
5919void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5920{
5921}
5922
5923ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5924 size_t count, loff_t *ppos)
5925{
5926 return -EIO;
5927}
5928#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 */