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/notifier.h>
23#include <linux/seq_file.h>
24#include <linux/kasan.h>
25#include <linux/cpu.h>
26#include <linux/cpuset.h>
27#include <linux/mempolicy.h>
28#include <linux/ctype.h>
29#include <linux/debugobjects.h>
30#include <linux/kallsyms.h>
31#include <linux/memory.h>
32#include <linux/math64.h>
33#include <linux/fault-inject.h>
34#include <linux/stacktrace.h>
35#include <linux/prefetch.h>
36#include <linux/memcontrol.h>
37#include <linux/random.h>
38
39#include <trace/events/kmem.h>
40
41#include "internal.h"
42
43/*
44 * Lock order:
45 * 1. slab_mutex (Global Mutex)
46 * 2. node->list_lock
47 * 3. slab_lock(page) (Only on some arches and for debugging)
48 *
49 * slab_mutex
50 *
51 * The role of the slab_mutex is to protect the list of all the slabs
52 * and to synchronize major metadata changes to slab cache structures.
53 *
54 * The slab_lock is only used for debugging and on arches that do not
55 * have the ability to do a cmpxchg_double. It only protects the second
56 * double word in the page struct. Meaning
57 * A. page->freelist -> List of object free in a page
58 * B. page->counters -> Counters of objects
59 * C. page->frozen -> frozen state
60 *
61 * If a slab is frozen then it is exempt from list management. It is not
62 * on any list. The processor that froze the slab is the one who can
63 * perform list operations on the page. Other processors may put objects
64 * onto the freelist but the processor that froze the slab is the only
65 * one that can retrieve the objects from the 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 return (void *)((unsigned long)ptr ^ s->random ^ ptr_addr);
254#else
255 return ptr;
256#endif
257}
258
259/* Returns the freelist pointer recorded at location ptr_addr. */
260static inline void *freelist_dereference(const struct kmem_cache *s,
261 void *ptr_addr)
262{
263 return freelist_ptr(s, (void *)*(unsigned long *)(ptr_addr),
264 (unsigned long)ptr_addr);
265}
266
267static inline void *get_freepointer(struct kmem_cache *s, void *object)
268{
269 return freelist_dereference(s, object + s->offset);
270}
271
272static void prefetch_freepointer(const struct kmem_cache *s, void *object)
273{
274 if (object)
275 prefetch(freelist_dereference(s, object + s->offset));
276}
277
278static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
279{
280 unsigned long freepointer_addr;
281 void *p;
282
283 if (!debug_pagealloc_enabled())
284 return get_freepointer(s, object);
285
286 freepointer_addr = (unsigned long)object + s->offset;
287 probe_kernel_read(&p, (void **)freepointer_addr, sizeof(p));
288 return freelist_ptr(s, p, freepointer_addr);
289}
290
291static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
292{
293 unsigned long freeptr_addr = (unsigned long)object + s->offset;
294
295#ifdef CONFIG_SLAB_FREELIST_HARDENED
296 BUG_ON(object == fp); /* naive detection of double free or corruption */
297#endif
298
299 *(void **)freeptr_addr = freelist_ptr(s, fp, freeptr_addr);
300}
301
302/* Loop over all objects in a slab */
303#define for_each_object(__p, __s, __addr, __objects) \
304 for (__p = fixup_red_left(__s, __addr); \
305 __p < (__addr) + (__objects) * (__s)->size; \
306 __p += (__s)->size)
307
308#define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
309 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
310 __idx <= __objects; \
311 __p += (__s)->size, __idx++)
312
313/* Determine object index from a given position */
314static inline unsigned int slab_index(void *p, struct kmem_cache *s, void *addr)
315{
316 return (p - addr) / s->size;
317}
318
319static inline unsigned int order_objects(unsigned int order, unsigned int size, unsigned int reserved)
320{
321 return (((unsigned int)PAGE_SIZE << order) - reserved) / size;
322}
323
324static inline struct kmem_cache_order_objects oo_make(unsigned int order,
325 unsigned int size, unsigned int reserved)
326{
327 struct kmem_cache_order_objects x = {
328 (order << OO_SHIFT) + order_objects(order, size, reserved)
329 };
330
331 return x;
332}
333
334static inline unsigned int oo_order(struct kmem_cache_order_objects x)
335{
336 return x.x >> OO_SHIFT;
337}
338
339static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
340{
341 return x.x & OO_MASK;
342}
343
344/*
345 * Per slab locking using the pagelock
346 */
347static __always_inline void slab_lock(struct page *page)
348{
349 VM_BUG_ON_PAGE(PageTail(page), page);
350 bit_spin_lock(PG_locked, &page->flags);
351}
352
353static __always_inline void slab_unlock(struct page *page)
354{
355 VM_BUG_ON_PAGE(PageTail(page), page);
356 __bit_spin_unlock(PG_locked, &page->flags);
357}
358
359static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
360{
361 struct page tmp;
362 tmp.counters = counters_new;
363 /*
364 * page->counters can cover frozen/inuse/objects as well
365 * as page->_refcount. If we assign to ->counters directly
366 * we run the risk of losing updates to page->_refcount, so
367 * be careful and only assign to the fields we need.
368 */
369 page->frozen = tmp.frozen;
370 page->inuse = tmp.inuse;
371 page->objects = tmp.objects;
372}
373
374/* Interrupts must be disabled (for the fallback code to work right) */
375static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
378 const char *n)
379{
380 VM_BUG_ON(!irqs_disabled());
381#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
387 return true;
388 } else
389#endif
390 {
391 slab_lock(page);
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 set_page_slub_counters(page, counters_new);
396 slab_unlock(page);
397 return true;
398 }
399 slab_unlock(page);
400 }
401
402 cpu_relax();
403 stat(s, CMPXCHG_DOUBLE_FAIL);
404
405#ifdef SLUB_DEBUG_CMPXCHG
406 pr_info("%s %s: cmpxchg double redo ", n, s->name);
407#endif
408
409 return false;
410}
411
412static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
415 const char *n)
416{
417#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
423 return true;
424 } else
425#endif
426 {
427 unsigned long flags;
428
429 local_irq_save(flags);
430 slab_lock(page);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 set_page_slub_counters(page, counters_new);
435 slab_unlock(page);
436 local_irq_restore(flags);
437 return true;
438 }
439 slab_unlock(page);
440 local_irq_restore(flags);
441 }
442
443 cpu_relax();
444 stat(s, CMPXCHG_DOUBLE_FAIL);
445
446#ifdef SLUB_DEBUG_CMPXCHG
447 pr_info("%s %s: cmpxchg double redo ", n, s->name);
448#endif
449
450 return false;
451}
452
453#ifdef CONFIG_SLUB_DEBUG
454/*
455 * Determine a map of object in use on a page.
456 *
457 * Node listlock must be held to guarantee that the page does
458 * not vanish from under us.
459 */
460static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
461{
462 void *p;
463 void *addr = page_address(page);
464
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(slab_index(p, s, addr), map);
467}
468
469static inline unsigned int size_from_object(struct kmem_cache *s)
470{
471 if (s->flags & SLAB_RED_ZONE)
472 return s->size - s->red_left_pad;
473
474 return s->size;
475}
476
477static inline void *restore_red_left(struct kmem_cache *s, void *p)
478{
479 if (s->flags & SLAB_RED_ZONE)
480 p -= s->red_left_pad;
481
482 return p;
483}
484
485/*
486 * Debug settings:
487 */
488#if defined(CONFIG_SLUB_DEBUG_ON)
489static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
490#else
491static slab_flags_t slub_debug;
492#endif
493
494static char *slub_debug_slabs;
495static int disable_higher_order_debug;
496
497/*
498 * slub is about to manipulate internal object metadata. This memory lies
499 * outside the range of the allocated object, so accessing it would normally
500 * be reported by kasan as a bounds error. metadata_access_enable() is used
501 * to tell kasan that these accesses are OK.
502 */
503static inline void metadata_access_enable(void)
504{
505 kasan_disable_current();
506}
507
508static inline void metadata_access_disable(void)
509{
510 kasan_enable_current();
511}
512
513/*
514 * Object debugging
515 */
516
517/* Verify that a pointer has an address that is valid within a slab page */
518static inline int check_valid_pointer(struct kmem_cache *s,
519 struct page *page, void *object)
520{
521 void *base;
522
523 if (!object)
524 return 1;
525
526 base = page_address(page);
527 object = restore_red_left(s, object);
528 if (object < base || object >= base + page->objects * s->size ||
529 (object - base) % s->size) {
530 return 0;
531 }
532
533 return 1;
534}
535
536static void print_section(char *level, char *text, u8 *addr,
537 unsigned int length)
538{
539 metadata_access_enable();
540 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
541 length, 1);
542 metadata_access_disable();
543}
544
545static struct track *get_track(struct kmem_cache *s, void *object,
546 enum track_item alloc)
547{
548 struct track *p;
549
550 if (s->offset)
551 p = object + s->offset + sizeof(void *);
552 else
553 p = object + s->inuse;
554
555 return p + alloc;
556}
557
558static void set_track(struct kmem_cache *s, void *object,
559 enum track_item alloc, unsigned long addr)
560{
561 struct track *p = get_track(s, object, alloc);
562
563 if (addr) {
564#ifdef CONFIG_STACKTRACE
565 struct stack_trace trace;
566 int i;
567
568 trace.nr_entries = 0;
569 trace.max_entries = TRACK_ADDRS_COUNT;
570 trace.entries = p->addrs;
571 trace.skip = 3;
572 metadata_access_enable();
573 save_stack_trace(&trace);
574 metadata_access_disable();
575
576 /* See rant in lockdep.c */
577 if (trace.nr_entries != 0 &&
578 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
579 trace.nr_entries--;
580
581 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
582 p->addrs[i] = 0;
583#endif
584 p->addr = addr;
585 p->cpu = smp_processor_id();
586 p->pid = current->pid;
587 p->when = jiffies;
588 } else
589 memset(p, 0, sizeof(struct track));
590}
591
592static void init_tracking(struct kmem_cache *s, void *object)
593{
594 if (!(s->flags & SLAB_STORE_USER))
595 return;
596
597 set_track(s, object, TRACK_FREE, 0UL);
598 set_track(s, object, TRACK_ALLOC, 0UL);
599}
600
601static void print_track(const char *s, struct track *t, unsigned long pr_time)
602{
603 if (!t->addr)
604 return;
605
606 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
607 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
608#ifdef CONFIG_STACKTRACE
609 {
610 int i;
611 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
612 if (t->addrs[i])
613 pr_err("\t%pS\n", (void *)t->addrs[i]);
614 else
615 break;
616 }
617#endif
618}
619
620static void print_tracking(struct kmem_cache *s, void *object)
621{
622 unsigned long pr_time = jiffies;
623 if (!(s->flags & SLAB_STORE_USER))
624 return;
625
626 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
627 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
628}
629
630static void print_page_info(struct page *page)
631{
632 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
633 page, page->objects, page->inuse, page->freelist, page->flags);
634
635}
636
637static void slab_bug(struct kmem_cache *s, char *fmt, ...)
638{
639 struct va_format vaf;
640 va_list args;
641
642 va_start(args, fmt);
643 vaf.fmt = fmt;
644 vaf.va = &args;
645 pr_err("=============================================================================\n");
646 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
647 pr_err("-----------------------------------------------------------------------------\n\n");
648
649 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
650 va_end(args);
651}
652
653static void slab_fix(struct kmem_cache *s, char *fmt, ...)
654{
655 struct va_format vaf;
656 va_list args;
657
658 va_start(args, fmt);
659 vaf.fmt = fmt;
660 vaf.va = &args;
661 pr_err("FIX %s: %pV\n", s->name, &vaf);
662 va_end(args);
663}
664
665static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
666{
667 unsigned int off; /* Offset of last byte */
668 u8 *addr = page_address(page);
669
670 print_tracking(s, p);
671
672 print_page_info(page);
673
674 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
675 p, p - addr, get_freepointer(s, p));
676
677 if (s->flags & SLAB_RED_ZONE)
678 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
679 s->red_left_pad);
680 else if (p > addr + 16)
681 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
682
683 print_section(KERN_ERR, "Object ", p,
684 min_t(unsigned int, s->object_size, PAGE_SIZE));
685 if (s->flags & SLAB_RED_ZONE)
686 print_section(KERN_ERR, "Redzone ", p + s->object_size,
687 s->inuse - s->object_size);
688
689 if (s->offset)
690 off = s->offset + sizeof(void *);
691 else
692 off = s->inuse;
693
694 if (s->flags & SLAB_STORE_USER)
695 off += 2 * sizeof(struct track);
696
697 off += kasan_metadata_size(s);
698
699 if (off != size_from_object(s))
700 /* Beginning of the filler is the free pointer */
701 print_section(KERN_ERR, "Padding ", p + off,
702 size_from_object(s) - off);
703
704 dump_stack();
705}
706
707void object_err(struct kmem_cache *s, struct page *page,
708 u8 *object, char *reason)
709{
710 slab_bug(s, "%s", reason);
711 print_trailer(s, page, object);
712}
713
714static void slab_err(struct kmem_cache *s, struct page *page,
715 const char *fmt, ...)
716{
717 va_list args;
718 char buf[100];
719
720 va_start(args, fmt);
721 vsnprintf(buf, sizeof(buf), fmt, args);
722 va_end(args);
723 slab_bug(s, "%s", buf);
724 print_page_info(page);
725 dump_stack();
726}
727
728static void init_object(struct kmem_cache *s, void *object, u8 val)
729{
730 u8 *p = object;
731
732 if (s->flags & SLAB_RED_ZONE)
733 memset(p - s->red_left_pad, val, s->red_left_pad);
734
735 if (s->flags & __OBJECT_POISON) {
736 memset(p, POISON_FREE, s->object_size - 1);
737 p[s->object_size - 1] = POISON_END;
738 }
739
740 if (s->flags & SLAB_RED_ZONE)
741 memset(p + s->object_size, val, s->inuse - s->object_size);
742}
743
744static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
745 void *from, void *to)
746{
747 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
748 memset(from, data, to - from);
749}
750
751static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
752 u8 *object, char *what,
753 u8 *start, unsigned int value, unsigned int bytes)
754{
755 u8 *fault;
756 u8 *end;
757
758 metadata_access_enable();
759 fault = memchr_inv(start, value, bytes);
760 metadata_access_disable();
761 if (!fault)
762 return 1;
763
764 end = start + bytes;
765 while (end > fault && end[-1] == value)
766 end--;
767
768 slab_bug(s, "%s overwritten", what);
769 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
770 fault, end - 1, fault[0], value);
771 print_trailer(s, page, object);
772
773 restore_bytes(s, what, value, fault, end);
774 return 0;
775}
776
777/*
778 * Object layout:
779 *
780 * object address
781 * Bytes of the object to be managed.
782 * If the freepointer may overlay the object then the free
783 * pointer is the first word of the object.
784 *
785 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
786 * 0xa5 (POISON_END)
787 *
788 * object + s->object_size
789 * Padding to reach word boundary. This is also used for Redzoning.
790 * Padding is extended by another word if Redzoning is enabled and
791 * object_size == inuse.
792 *
793 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
794 * 0xcc (RED_ACTIVE) for objects in use.
795 *
796 * object + s->inuse
797 * Meta data starts here.
798 *
799 * A. Free pointer (if we cannot overwrite object on free)
800 * B. Tracking data for SLAB_STORE_USER
801 * C. Padding to reach required alignment boundary or at mininum
802 * one word if debugging is on to be able to detect writes
803 * before the word boundary.
804 *
805 * Padding is done using 0x5a (POISON_INUSE)
806 *
807 * object + s->size
808 * Nothing is used beyond s->size.
809 *
810 * If slabcaches are merged then the object_size and inuse boundaries are mostly
811 * ignored. And therefore no slab options that rely on these boundaries
812 * may be used with merged slabcaches.
813 */
814
815static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
816{
817 unsigned long off = s->inuse; /* The end of info */
818
819 if (s->offset)
820 /* Freepointer is placed after the object. */
821 off += sizeof(void *);
822
823 if (s->flags & SLAB_STORE_USER)
824 /* We also have user information there */
825 off += 2 * sizeof(struct track);
826
827 off += kasan_metadata_size(s);
828
829 if (size_from_object(s) == off)
830 return 1;
831
832 return check_bytes_and_report(s, page, p, "Object padding",
833 p + off, POISON_INUSE, size_from_object(s) - off);
834}
835
836/* Check the pad bytes at the end of a slab page */
837static int slab_pad_check(struct kmem_cache *s, struct page *page)
838{
839 u8 *start;
840 u8 *fault;
841 u8 *end;
842 u8 *pad;
843 int length;
844 int remainder;
845
846 if (!(s->flags & SLAB_POISON))
847 return 1;
848
849 start = page_address(page);
850 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
851 end = start + length;
852 remainder = length % s->size;
853 if (!remainder)
854 return 1;
855
856 pad = end - remainder;
857 metadata_access_enable();
858 fault = memchr_inv(pad, POISON_INUSE, remainder);
859 metadata_access_disable();
860 if (!fault)
861 return 1;
862 while (end > fault && end[-1] == POISON_INUSE)
863 end--;
864
865 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
866 print_section(KERN_ERR, "Padding ", pad, remainder);
867
868 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
869 return 0;
870}
871
872static int check_object(struct kmem_cache *s, struct page *page,
873 void *object, u8 val)
874{
875 u8 *p = object;
876 u8 *endobject = object + s->object_size;
877
878 if (s->flags & SLAB_RED_ZONE) {
879 if (!check_bytes_and_report(s, page, object, "Redzone",
880 object - s->red_left_pad, val, s->red_left_pad))
881 return 0;
882
883 if (!check_bytes_and_report(s, page, object, "Redzone",
884 endobject, val, s->inuse - s->object_size))
885 return 0;
886 } else {
887 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
888 check_bytes_and_report(s, page, p, "Alignment padding",
889 endobject, POISON_INUSE,
890 s->inuse - s->object_size);
891 }
892 }
893
894 if (s->flags & SLAB_POISON) {
895 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
896 (!check_bytes_and_report(s, page, p, "Poison", p,
897 POISON_FREE, s->object_size - 1) ||
898 !check_bytes_and_report(s, page, p, "Poison",
899 p + s->object_size - 1, POISON_END, 1)))
900 return 0;
901 /*
902 * check_pad_bytes cleans up on its own.
903 */
904 check_pad_bytes(s, page, p);
905 }
906
907 if (!s->offset && val == SLUB_RED_ACTIVE)
908 /*
909 * Object and freepointer overlap. Cannot check
910 * freepointer while object is allocated.
911 */
912 return 1;
913
914 /* Check free pointer validity */
915 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
916 object_err(s, page, p, "Freepointer corrupt");
917 /*
918 * No choice but to zap it and thus lose the remainder
919 * of the free objects in this slab. May cause
920 * another error because the object count is now wrong.
921 */
922 set_freepointer(s, p, NULL);
923 return 0;
924 }
925 return 1;
926}
927
928static int check_slab(struct kmem_cache *s, struct page *page)
929{
930 int maxobj;
931
932 VM_BUG_ON(!irqs_disabled());
933
934 if (!PageSlab(page)) {
935 slab_err(s, page, "Not a valid slab page");
936 return 0;
937 }
938
939 maxobj = order_objects(compound_order(page), s->size, s->reserved);
940 if (page->objects > maxobj) {
941 slab_err(s, page, "objects %u > max %u",
942 page->objects, maxobj);
943 return 0;
944 }
945 if (page->inuse > page->objects) {
946 slab_err(s, page, "inuse %u > max %u",
947 page->inuse, page->objects);
948 return 0;
949 }
950 /* Slab_pad_check fixes things up after itself */
951 slab_pad_check(s, page);
952 return 1;
953}
954
955/*
956 * Determine if a certain object on a page is on the freelist. Must hold the
957 * slab lock to guarantee that the chains are in a consistent state.
958 */
959static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
960{
961 int nr = 0;
962 void *fp;
963 void *object = NULL;
964 int max_objects;
965
966 fp = page->freelist;
967 while (fp && nr <= page->objects) {
968 if (fp == search)
969 return 1;
970 if (!check_valid_pointer(s, page, fp)) {
971 if (object) {
972 object_err(s, page, object,
973 "Freechain corrupt");
974 set_freepointer(s, object, NULL);
975 } else {
976 slab_err(s, page, "Freepointer corrupt");
977 page->freelist = NULL;
978 page->inuse = page->objects;
979 slab_fix(s, "Freelist cleared");
980 return 0;
981 }
982 break;
983 }
984 object = fp;
985 fp = get_freepointer(s, object);
986 nr++;
987 }
988
989 max_objects = order_objects(compound_order(page), s->size, s->reserved);
990 if (max_objects > MAX_OBJS_PER_PAGE)
991 max_objects = MAX_OBJS_PER_PAGE;
992
993 if (page->objects != max_objects) {
994 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
995 page->objects, max_objects);
996 page->objects = max_objects;
997 slab_fix(s, "Number of objects adjusted.");
998 }
999 if (page->inuse != page->objects - nr) {
1000 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
1001 page->inuse, page->objects - nr);
1002 page->inuse = page->objects - nr;
1003 slab_fix(s, "Object count adjusted.");
1004 }
1005 return search == NULL;
1006}
1007
1008static void trace(struct kmem_cache *s, struct page *page, void *object,
1009 int alloc)
1010{
1011 if (s->flags & SLAB_TRACE) {
1012 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1013 s->name,
1014 alloc ? "alloc" : "free",
1015 object, page->inuse,
1016 page->freelist);
1017
1018 if (!alloc)
1019 print_section(KERN_INFO, "Object ", (void *)object,
1020 s->object_size);
1021
1022 dump_stack();
1023 }
1024}
1025
1026/*
1027 * Tracking of fully allocated slabs for debugging purposes.
1028 */
1029static void add_full(struct kmem_cache *s,
1030 struct kmem_cache_node *n, struct page *page)
1031{
1032 if (!(s->flags & SLAB_STORE_USER))
1033 return;
1034
1035 lockdep_assert_held(&n->list_lock);
1036 list_add(&page->lru, &n->full);
1037}
1038
1039static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1040{
1041 if (!(s->flags & SLAB_STORE_USER))
1042 return;
1043
1044 lockdep_assert_held(&n->list_lock);
1045 list_del(&page->lru);
1046}
1047
1048/* Tracking of the number of slabs for debugging purposes */
1049static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1050{
1051 struct kmem_cache_node *n = get_node(s, node);
1052
1053 return atomic_long_read(&n->nr_slabs);
1054}
1055
1056static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1057{
1058 return atomic_long_read(&n->nr_slabs);
1059}
1060
1061static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1062{
1063 struct kmem_cache_node *n = get_node(s, node);
1064
1065 /*
1066 * May be called early in order to allocate a slab for the
1067 * kmem_cache_node structure. Solve the chicken-egg
1068 * dilemma by deferring the increment of the count during
1069 * bootstrap (see early_kmem_cache_node_alloc).
1070 */
1071 if (likely(n)) {
1072 atomic_long_inc(&n->nr_slabs);
1073 atomic_long_add(objects, &n->total_objects);
1074 }
1075}
1076static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1077{
1078 struct kmem_cache_node *n = get_node(s, node);
1079
1080 atomic_long_dec(&n->nr_slabs);
1081 atomic_long_sub(objects, &n->total_objects);
1082}
1083
1084/* Object debug checks for alloc/free paths */
1085static void setup_object_debug(struct kmem_cache *s, struct page *page,
1086 void *object)
1087{
1088 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1089 return;
1090
1091 init_object(s, object, SLUB_RED_INACTIVE);
1092 init_tracking(s, object);
1093}
1094
1095static inline int alloc_consistency_checks(struct kmem_cache *s,
1096 struct page *page,
1097 void *object, unsigned long addr)
1098{
1099 if (!check_slab(s, page))
1100 return 0;
1101
1102 if (!check_valid_pointer(s, page, object)) {
1103 object_err(s, page, object, "Freelist Pointer check fails");
1104 return 0;
1105 }
1106
1107 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1108 return 0;
1109
1110 return 1;
1111}
1112
1113static noinline int alloc_debug_processing(struct kmem_cache *s,
1114 struct page *page,
1115 void *object, unsigned long addr)
1116{
1117 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1118 if (!alloc_consistency_checks(s, page, object, addr))
1119 goto bad;
1120 }
1121
1122 /* Success perform special debug activities for allocs */
1123 if (s->flags & SLAB_STORE_USER)
1124 set_track(s, object, TRACK_ALLOC, addr);
1125 trace(s, page, object, 1);
1126 init_object(s, object, SLUB_RED_ACTIVE);
1127 return 1;
1128
1129bad:
1130 if (PageSlab(page)) {
1131 /*
1132 * If this is a slab page then lets do the best we can
1133 * to avoid issues in the future. Marking all objects
1134 * as used avoids touching the remaining objects.
1135 */
1136 slab_fix(s, "Marking all objects used");
1137 page->inuse = page->objects;
1138 page->freelist = NULL;
1139 }
1140 return 0;
1141}
1142
1143static inline int free_consistency_checks(struct kmem_cache *s,
1144 struct page *page, void *object, unsigned long addr)
1145{
1146 if (!check_valid_pointer(s, page, object)) {
1147 slab_err(s, page, "Invalid object pointer 0x%p", object);
1148 return 0;
1149 }
1150
1151 if (on_freelist(s, page, object)) {
1152 object_err(s, page, object, "Object already free");
1153 return 0;
1154 }
1155
1156 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1157 return 0;
1158
1159 if (unlikely(s != page->slab_cache)) {
1160 if (!PageSlab(page)) {
1161 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1162 object);
1163 } else if (!page->slab_cache) {
1164 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1165 object);
1166 dump_stack();
1167 } else
1168 object_err(s, page, object,
1169 "page slab pointer corrupt.");
1170 return 0;
1171 }
1172 return 1;
1173}
1174
1175/* Supports checking bulk free of a constructed freelist */
1176static noinline int free_debug_processing(
1177 struct kmem_cache *s, struct page *page,
1178 void *head, void *tail, int bulk_cnt,
1179 unsigned long addr)
1180{
1181 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1182 void *object = head;
1183 int cnt = 0;
1184 unsigned long uninitialized_var(flags);
1185 int ret = 0;
1186
1187 spin_lock_irqsave(&n->list_lock, flags);
1188 slab_lock(page);
1189
1190 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1191 if (!check_slab(s, page))
1192 goto out;
1193 }
1194
1195next_object:
1196 cnt++;
1197
1198 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1199 if (!free_consistency_checks(s, page, object, addr))
1200 goto out;
1201 }
1202
1203 if (s->flags & SLAB_STORE_USER)
1204 set_track(s, object, TRACK_FREE, addr);
1205 trace(s, page, object, 0);
1206 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1207 init_object(s, object, SLUB_RED_INACTIVE);
1208
1209 /* Reached end of constructed freelist yet? */
1210 if (object != tail) {
1211 object = get_freepointer(s, object);
1212 goto next_object;
1213 }
1214 ret = 1;
1215
1216out:
1217 if (cnt != bulk_cnt)
1218 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1219 bulk_cnt, cnt);
1220
1221 slab_unlock(page);
1222 spin_unlock_irqrestore(&n->list_lock, flags);
1223 if (!ret)
1224 slab_fix(s, "Object at 0x%p not freed", object);
1225 return ret;
1226}
1227
1228static int __init setup_slub_debug(char *str)
1229{
1230 slub_debug = DEBUG_DEFAULT_FLAGS;
1231 if (*str++ != '=' || !*str)
1232 /*
1233 * No options specified. Switch on full debugging.
1234 */
1235 goto out;
1236
1237 if (*str == ',')
1238 /*
1239 * No options but restriction on slabs. This means full
1240 * debugging for slabs matching a pattern.
1241 */
1242 goto check_slabs;
1243
1244 slub_debug = 0;
1245 if (*str == '-')
1246 /*
1247 * Switch off all debugging measures.
1248 */
1249 goto out;
1250
1251 /*
1252 * Determine which debug features should be switched on
1253 */
1254 for (; *str && *str != ','; str++) {
1255 switch (tolower(*str)) {
1256 case 'f':
1257 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1258 break;
1259 case 'z':
1260 slub_debug |= SLAB_RED_ZONE;
1261 break;
1262 case 'p':
1263 slub_debug |= SLAB_POISON;
1264 break;
1265 case 'u':
1266 slub_debug |= SLAB_STORE_USER;
1267 break;
1268 case 't':
1269 slub_debug |= SLAB_TRACE;
1270 break;
1271 case 'a':
1272 slub_debug |= SLAB_FAILSLAB;
1273 break;
1274 case 'o':
1275 /*
1276 * Avoid enabling debugging on caches if its minimum
1277 * order would increase as a result.
1278 */
1279 disable_higher_order_debug = 1;
1280 break;
1281 default:
1282 pr_err("slub_debug option '%c' unknown. skipped\n",
1283 *str);
1284 }
1285 }
1286
1287check_slabs:
1288 if (*str == ',')
1289 slub_debug_slabs = str + 1;
1290out:
1291 return 1;
1292}
1293
1294__setup("slub_debug", setup_slub_debug);
1295
1296slab_flags_t kmem_cache_flags(unsigned int object_size,
1297 slab_flags_t flags, const char *name,
1298 void (*ctor)(void *))
1299{
1300 /*
1301 * Enable debugging if selected on the kernel commandline.
1302 */
1303 if (slub_debug && (!slub_debug_slabs || (name &&
1304 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1305 flags |= slub_debug;
1306
1307 return flags;
1308}
1309#else /* !CONFIG_SLUB_DEBUG */
1310static inline void setup_object_debug(struct kmem_cache *s,
1311 struct page *page, void *object) {}
1312
1313static inline int alloc_debug_processing(struct kmem_cache *s,
1314 struct page *page, void *object, unsigned long addr) { return 0; }
1315
1316static inline int free_debug_processing(
1317 struct kmem_cache *s, struct page *page,
1318 void *head, void *tail, int bulk_cnt,
1319 unsigned long addr) { return 0; }
1320
1321static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1322 { return 1; }
1323static inline int check_object(struct kmem_cache *s, struct page *page,
1324 void *object, u8 val) { return 1; }
1325static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1326 struct page *page) {}
1327static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1328 struct page *page) {}
1329slab_flags_t kmem_cache_flags(unsigned int object_size,
1330 slab_flags_t flags, const char *name,
1331 void (*ctor)(void *))
1332{
1333 return flags;
1334}
1335#define slub_debug 0
1336
1337#define disable_higher_order_debug 0
1338
1339static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1340 { return 0; }
1341static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1342 { return 0; }
1343static inline void inc_slabs_node(struct kmem_cache *s, int node,
1344 int objects) {}
1345static inline void dec_slabs_node(struct kmem_cache *s, int node,
1346 int objects) {}
1347
1348#endif /* CONFIG_SLUB_DEBUG */
1349
1350/*
1351 * Hooks for other subsystems that check memory allocations. In a typical
1352 * production configuration these hooks all should produce no code at all.
1353 */
1354static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1355{
1356 kmemleak_alloc(ptr, size, 1, flags);
1357 kasan_kmalloc_large(ptr, size, flags);
1358}
1359
1360static __always_inline void kfree_hook(void *x)
1361{
1362 kmemleak_free(x);
1363 kasan_kfree_large(x, _RET_IP_);
1364}
1365
1366static __always_inline bool slab_free_hook(struct kmem_cache *s, void *x)
1367{
1368 kmemleak_free_recursive(x, s->flags);
1369
1370 /*
1371 * Trouble is that we may no longer disable interrupts in the fast path
1372 * So in order to make the debug calls that expect irqs to be
1373 * disabled we need to disable interrupts temporarily.
1374 */
1375#ifdef CONFIG_LOCKDEP
1376 {
1377 unsigned long flags;
1378
1379 local_irq_save(flags);
1380 debug_check_no_locks_freed(x, s->object_size);
1381 local_irq_restore(flags);
1382 }
1383#endif
1384 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1385 debug_check_no_obj_freed(x, s->object_size);
1386
1387 /* KASAN might put x into memory quarantine, delaying its reuse */
1388 return kasan_slab_free(s, x, _RET_IP_);
1389}
1390
1391static inline bool slab_free_freelist_hook(struct kmem_cache *s,
1392 void **head, void **tail)
1393{
1394/*
1395 * Compiler cannot detect this function can be removed if slab_free_hook()
1396 * evaluates to nothing. Thus, catch all relevant config debug options here.
1397 */
1398#if defined(CONFIG_LOCKDEP) || \
1399 defined(CONFIG_DEBUG_KMEMLEAK) || \
1400 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1401 defined(CONFIG_KASAN)
1402
1403 void *object;
1404 void *next = *head;
1405 void *old_tail = *tail ? *tail : *head;
1406
1407 /* Head and tail of the reconstructed freelist */
1408 *head = NULL;
1409 *tail = NULL;
1410
1411 do {
1412 object = next;
1413 next = get_freepointer(s, object);
1414 /* If object's reuse doesn't have to be delayed */
1415 if (!slab_free_hook(s, object)) {
1416 /* Move object to the new freelist */
1417 set_freepointer(s, object, *head);
1418 *head = object;
1419 if (!*tail)
1420 *tail = object;
1421 }
1422 } while (object != old_tail);
1423
1424 if (*head == *tail)
1425 *tail = NULL;
1426
1427 return *head != NULL;
1428#else
1429 return true;
1430#endif
1431}
1432
1433static void setup_object(struct kmem_cache *s, struct page *page,
1434 void *object)
1435{
1436 setup_object_debug(s, page, object);
1437 kasan_init_slab_obj(s, object);
1438 if (unlikely(s->ctor)) {
1439 kasan_unpoison_object_data(s, object);
1440 s->ctor(object);
1441 kasan_poison_object_data(s, object);
1442 }
1443}
1444
1445/*
1446 * Slab allocation and freeing
1447 */
1448static inline struct page *alloc_slab_page(struct kmem_cache *s,
1449 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1450{
1451 struct page *page;
1452 unsigned int order = oo_order(oo);
1453
1454 if (node == NUMA_NO_NODE)
1455 page = alloc_pages(flags, order);
1456 else
1457 page = __alloc_pages_node(node, flags, order);
1458
1459 if (page && memcg_charge_slab(page, flags, order, s)) {
1460 __free_pages(page, order);
1461 page = NULL;
1462 }
1463
1464 return page;
1465}
1466
1467#ifdef CONFIG_SLAB_FREELIST_RANDOM
1468/* Pre-initialize the random sequence cache */
1469static int init_cache_random_seq(struct kmem_cache *s)
1470{
1471 unsigned int count = oo_objects(s->oo);
1472 int err;
1473
1474 /* Bailout if already initialised */
1475 if (s->random_seq)
1476 return 0;
1477
1478 err = cache_random_seq_create(s, count, GFP_KERNEL);
1479 if (err) {
1480 pr_err("SLUB: Unable to initialize free list for %s\n",
1481 s->name);
1482 return err;
1483 }
1484
1485 /* Transform to an offset on the set of pages */
1486 if (s->random_seq) {
1487 unsigned int i;
1488
1489 for (i = 0; i < count; i++)
1490 s->random_seq[i] *= s->size;
1491 }
1492 return 0;
1493}
1494
1495/* Initialize each random sequence freelist per cache */
1496static void __init init_freelist_randomization(void)
1497{
1498 struct kmem_cache *s;
1499
1500 mutex_lock(&slab_mutex);
1501
1502 list_for_each_entry(s, &slab_caches, list)
1503 init_cache_random_seq(s);
1504
1505 mutex_unlock(&slab_mutex);
1506}
1507
1508/* Get the next entry on the pre-computed freelist randomized */
1509static void *next_freelist_entry(struct kmem_cache *s, struct page *page,
1510 unsigned long *pos, void *start,
1511 unsigned long page_limit,
1512 unsigned long freelist_count)
1513{
1514 unsigned int idx;
1515
1516 /*
1517 * If the target page allocation failed, the number of objects on the
1518 * page might be smaller than the usual size defined by the cache.
1519 */
1520 do {
1521 idx = s->random_seq[*pos];
1522 *pos += 1;
1523 if (*pos >= freelist_count)
1524 *pos = 0;
1525 } while (unlikely(idx >= page_limit));
1526
1527 return (char *)start + idx;
1528}
1529
1530/* Shuffle the single linked freelist based on a random pre-computed sequence */
1531static bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1532{
1533 void *start;
1534 void *cur;
1535 void *next;
1536 unsigned long idx, pos, page_limit, freelist_count;
1537
1538 if (page->objects < 2 || !s->random_seq)
1539 return false;
1540
1541 freelist_count = oo_objects(s->oo);
1542 pos = get_random_int() % freelist_count;
1543
1544 page_limit = page->objects * s->size;
1545 start = fixup_red_left(s, page_address(page));
1546
1547 /* First entry is used as the base of the freelist */
1548 cur = next_freelist_entry(s, page, &pos, start, page_limit,
1549 freelist_count);
1550 page->freelist = cur;
1551
1552 for (idx = 1; idx < page->objects; idx++) {
1553 setup_object(s, page, cur);
1554 next = next_freelist_entry(s, page, &pos, start, page_limit,
1555 freelist_count);
1556 set_freepointer(s, cur, next);
1557 cur = next;
1558 }
1559 setup_object(s, page, cur);
1560 set_freepointer(s, cur, NULL);
1561
1562 return true;
1563}
1564#else
1565static inline int init_cache_random_seq(struct kmem_cache *s)
1566{
1567 return 0;
1568}
1569static inline void init_freelist_randomization(void) { }
1570static inline bool shuffle_freelist(struct kmem_cache *s, struct page *page)
1571{
1572 return false;
1573}
1574#endif /* CONFIG_SLAB_FREELIST_RANDOM */
1575
1576static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1577{
1578 struct page *page;
1579 struct kmem_cache_order_objects oo = s->oo;
1580 gfp_t alloc_gfp;
1581 void *start, *p;
1582 int idx, order;
1583 bool shuffle;
1584
1585 flags &= gfp_allowed_mask;
1586
1587 if (gfpflags_allow_blocking(flags))
1588 local_irq_enable();
1589
1590 flags |= s->allocflags;
1591
1592 /*
1593 * Let the initial higher-order allocation fail under memory pressure
1594 * so we fall-back to the minimum order allocation.
1595 */
1596 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1597 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1598 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1599
1600 page = alloc_slab_page(s, alloc_gfp, node, oo);
1601 if (unlikely(!page)) {
1602 oo = s->min;
1603 alloc_gfp = flags;
1604 /*
1605 * Allocation may have failed due to fragmentation.
1606 * Try a lower order alloc if possible
1607 */
1608 page = alloc_slab_page(s, alloc_gfp, node, oo);
1609 if (unlikely(!page))
1610 goto out;
1611 stat(s, ORDER_FALLBACK);
1612 }
1613
1614 page->objects = oo_objects(oo);
1615
1616 order = compound_order(page);
1617 page->slab_cache = s;
1618 __SetPageSlab(page);
1619 if (page_is_pfmemalloc(page))
1620 SetPageSlabPfmemalloc(page);
1621
1622 start = page_address(page);
1623
1624 if (unlikely(s->flags & SLAB_POISON))
1625 memset(start, POISON_INUSE, PAGE_SIZE << order);
1626
1627 kasan_poison_slab(page);
1628
1629 shuffle = shuffle_freelist(s, page);
1630
1631 if (!shuffle) {
1632 for_each_object_idx(p, idx, s, start, page->objects) {
1633 setup_object(s, page, p);
1634 if (likely(idx < page->objects))
1635 set_freepointer(s, p, p + s->size);
1636 else
1637 set_freepointer(s, p, NULL);
1638 }
1639 page->freelist = fixup_red_left(s, start);
1640 }
1641
1642 page->inuse = page->objects;
1643 page->frozen = 1;
1644
1645out:
1646 if (gfpflags_allow_blocking(flags))
1647 local_irq_disable();
1648 if (!page)
1649 return NULL;
1650
1651 mod_lruvec_page_state(page,
1652 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1653 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1654 1 << oo_order(oo));
1655
1656 inc_slabs_node(s, page_to_nid(page), page->objects);
1657
1658 return page;
1659}
1660
1661static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1662{
1663 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1664 gfp_t invalid_mask = flags & GFP_SLAB_BUG_MASK;
1665 flags &= ~GFP_SLAB_BUG_MASK;
1666 pr_warn("Unexpected gfp: %#x (%pGg). Fixing up to gfp: %#x (%pGg). Fix your code!\n",
1667 invalid_mask, &invalid_mask, flags, &flags);
1668 dump_stack();
1669 }
1670
1671 return allocate_slab(s,
1672 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1673}
1674
1675static void __free_slab(struct kmem_cache *s, struct page *page)
1676{
1677 int order = compound_order(page);
1678 int pages = 1 << order;
1679
1680 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1681 void *p;
1682
1683 slab_pad_check(s, page);
1684 for_each_object(p, s, page_address(page),
1685 page->objects)
1686 check_object(s, page, p, SLUB_RED_INACTIVE);
1687 }
1688
1689 mod_lruvec_page_state(page,
1690 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1691 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1692 -pages);
1693
1694 __ClearPageSlabPfmemalloc(page);
1695 __ClearPageSlab(page);
1696
1697 page_mapcount_reset(page);
1698 if (current->reclaim_state)
1699 current->reclaim_state->reclaimed_slab += pages;
1700 memcg_uncharge_slab(page, order, s);
1701 __free_pages(page, order);
1702}
1703
1704#define need_reserve_slab_rcu \
1705 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1706
1707static void rcu_free_slab(struct rcu_head *h)
1708{
1709 struct page *page;
1710
1711 if (need_reserve_slab_rcu)
1712 page = virt_to_head_page(h);
1713 else
1714 page = container_of((struct list_head *)h, struct page, lru);
1715
1716 __free_slab(page->slab_cache, page);
1717}
1718
1719static void free_slab(struct kmem_cache *s, struct page *page)
1720{
1721 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU)) {
1722 struct rcu_head *head;
1723
1724 if (need_reserve_slab_rcu) {
1725 int order = compound_order(page);
1726 int offset = (PAGE_SIZE << order) - s->reserved;
1727
1728 VM_BUG_ON(s->reserved != sizeof(*head));
1729 head = page_address(page) + offset;
1730 } else {
1731 head = &page->rcu_head;
1732 }
1733
1734 call_rcu(head, rcu_free_slab);
1735 } else
1736 __free_slab(s, page);
1737}
1738
1739static void discard_slab(struct kmem_cache *s, struct page *page)
1740{
1741 dec_slabs_node(s, page_to_nid(page), page->objects);
1742 free_slab(s, page);
1743}
1744
1745/*
1746 * Management of partially allocated slabs.
1747 */
1748static inline void
1749__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1750{
1751 n->nr_partial++;
1752 if (tail == DEACTIVATE_TO_TAIL)
1753 list_add_tail(&page->lru, &n->partial);
1754 else
1755 list_add(&page->lru, &n->partial);
1756}
1757
1758static inline void add_partial(struct kmem_cache_node *n,
1759 struct page *page, int tail)
1760{
1761 lockdep_assert_held(&n->list_lock);
1762 __add_partial(n, page, tail);
1763}
1764
1765static inline void remove_partial(struct kmem_cache_node *n,
1766 struct page *page)
1767{
1768 lockdep_assert_held(&n->list_lock);
1769 list_del(&page->lru);
1770 n->nr_partial--;
1771}
1772
1773/*
1774 * Remove slab from the partial list, freeze it and
1775 * return the pointer to the freelist.
1776 *
1777 * Returns a list of objects or NULL if it fails.
1778 */
1779static inline void *acquire_slab(struct kmem_cache *s,
1780 struct kmem_cache_node *n, struct page *page,
1781 int mode, int *objects)
1782{
1783 void *freelist;
1784 unsigned long counters;
1785 struct page new;
1786
1787 lockdep_assert_held(&n->list_lock);
1788
1789 /*
1790 * Zap the freelist and set the frozen bit.
1791 * The old freelist is the list of objects for the
1792 * per cpu allocation list.
1793 */
1794 freelist = page->freelist;
1795 counters = page->counters;
1796 new.counters = counters;
1797 *objects = new.objects - new.inuse;
1798 if (mode) {
1799 new.inuse = page->objects;
1800 new.freelist = NULL;
1801 } else {
1802 new.freelist = freelist;
1803 }
1804
1805 VM_BUG_ON(new.frozen);
1806 new.frozen = 1;
1807
1808 if (!__cmpxchg_double_slab(s, page,
1809 freelist, counters,
1810 new.freelist, new.counters,
1811 "acquire_slab"))
1812 return NULL;
1813
1814 remove_partial(n, page);
1815 WARN_ON(!freelist);
1816 return freelist;
1817}
1818
1819static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1820static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1821
1822/*
1823 * Try to allocate a partial slab from a specific node.
1824 */
1825static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1826 struct kmem_cache_cpu *c, gfp_t flags)
1827{
1828 struct page *page, *page2;
1829 void *object = NULL;
1830 unsigned int available = 0;
1831 int objects;
1832
1833 /*
1834 * Racy check. If we mistakenly see no partial slabs then we
1835 * just allocate an empty slab. If we mistakenly try to get a
1836 * partial slab and there is none available then get_partials()
1837 * will return NULL.
1838 */
1839 if (!n || !n->nr_partial)
1840 return NULL;
1841
1842 spin_lock(&n->list_lock);
1843 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1844 void *t;
1845
1846 if (!pfmemalloc_match(page, flags))
1847 continue;
1848
1849 t = acquire_slab(s, n, page, object == NULL, &objects);
1850 if (!t)
1851 break;
1852
1853 available += objects;
1854 if (!object) {
1855 c->page = page;
1856 stat(s, ALLOC_FROM_PARTIAL);
1857 object = t;
1858 } else {
1859 put_cpu_partial(s, page, 0);
1860 stat(s, CPU_PARTIAL_NODE);
1861 }
1862 if (!kmem_cache_has_cpu_partial(s)
1863 || available > slub_cpu_partial(s) / 2)
1864 break;
1865
1866 }
1867 spin_unlock(&n->list_lock);
1868 return object;
1869}
1870
1871/*
1872 * Get a page from somewhere. Search in increasing NUMA distances.
1873 */
1874static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1875 struct kmem_cache_cpu *c)
1876{
1877#ifdef CONFIG_NUMA
1878 struct zonelist *zonelist;
1879 struct zoneref *z;
1880 struct zone *zone;
1881 enum zone_type high_zoneidx = gfp_zone(flags);
1882 void *object;
1883 unsigned int cpuset_mems_cookie;
1884
1885 /*
1886 * The defrag ratio allows a configuration of the tradeoffs between
1887 * inter node defragmentation and node local allocations. A lower
1888 * defrag_ratio increases the tendency to do local allocations
1889 * instead of attempting to obtain partial slabs from other nodes.
1890 *
1891 * If the defrag_ratio is set to 0 then kmalloc() always
1892 * returns node local objects. If the ratio is higher then kmalloc()
1893 * may return off node objects because partial slabs are obtained
1894 * from other nodes and filled up.
1895 *
1896 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
1897 * (which makes defrag_ratio = 1000) then every (well almost)
1898 * allocation will first attempt to defrag slab caches on other nodes.
1899 * This means scanning over all nodes to look for partial slabs which
1900 * may be expensive if we do it every time we are trying to find a slab
1901 * with available objects.
1902 */
1903 if (!s->remote_node_defrag_ratio ||
1904 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1905 return NULL;
1906
1907 do {
1908 cpuset_mems_cookie = read_mems_allowed_begin();
1909 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1910 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1911 struct kmem_cache_node *n;
1912
1913 n = get_node(s, zone_to_nid(zone));
1914
1915 if (n && cpuset_zone_allowed(zone, flags) &&
1916 n->nr_partial > s->min_partial) {
1917 object = get_partial_node(s, n, c, flags);
1918 if (object) {
1919 /*
1920 * Don't check read_mems_allowed_retry()
1921 * here - if mems_allowed was updated in
1922 * parallel, that was a harmless race
1923 * between allocation and the cpuset
1924 * update
1925 */
1926 return object;
1927 }
1928 }
1929 }
1930 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1931#endif
1932 return NULL;
1933}
1934
1935/*
1936 * Get a partial page, lock it and return it.
1937 */
1938static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1939 struct kmem_cache_cpu *c)
1940{
1941 void *object;
1942 int searchnode = node;
1943
1944 if (node == NUMA_NO_NODE)
1945 searchnode = numa_mem_id();
1946 else if (!node_present_pages(node))
1947 searchnode = node_to_mem_node(node);
1948
1949 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1950 if (object || node != NUMA_NO_NODE)
1951 return object;
1952
1953 return get_any_partial(s, flags, c);
1954}
1955
1956#ifdef CONFIG_PREEMPT
1957/*
1958 * Calculate the next globally unique transaction for disambiguiation
1959 * during cmpxchg. The transactions start with the cpu number and are then
1960 * incremented by CONFIG_NR_CPUS.
1961 */
1962#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1963#else
1964/*
1965 * No preemption supported therefore also no need to check for
1966 * different cpus.
1967 */
1968#define TID_STEP 1
1969#endif
1970
1971static inline unsigned long next_tid(unsigned long tid)
1972{
1973 return tid + TID_STEP;
1974}
1975
1976static inline unsigned int tid_to_cpu(unsigned long tid)
1977{
1978 return tid % TID_STEP;
1979}
1980
1981static inline unsigned long tid_to_event(unsigned long tid)
1982{
1983 return tid / TID_STEP;
1984}
1985
1986static inline unsigned int init_tid(int cpu)
1987{
1988 return cpu;
1989}
1990
1991static inline void note_cmpxchg_failure(const char *n,
1992 const struct kmem_cache *s, unsigned long tid)
1993{
1994#ifdef SLUB_DEBUG_CMPXCHG
1995 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1996
1997 pr_info("%s %s: cmpxchg redo ", n, s->name);
1998
1999#ifdef CONFIG_PREEMPT
2000 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
2001 pr_warn("due to cpu change %d -> %d\n",
2002 tid_to_cpu(tid), tid_to_cpu(actual_tid));
2003 else
2004#endif
2005 if (tid_to_event(tid) != tid_to_event(actual_tid))
2006 pr_warn("due to cpu running other code. Event %ld->%ld\n",
2007 tid_to_event(tid), tid_to_event(actual_tid));
2008 else
2009 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
2010 actual_tid, tid, next_tid(tid));
2011#endif
2012 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
2013}
2014
2015static void init_kmem_cache_cpus(struct kmem_cache *s)
2016{
2017 int cpu;
2018
2019 for_each_possible_cpu(cpu)
2020 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
2021}
2022
2023/*
2024 * Remove the cpu slab
2025 */
2026static void deactivate_slab(struct kmem_cache *s, struct page *page,
2027 void *freelist, struct kmem_cache_cpu *c)
2028{
2029 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
2030 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
2031 int lock = 0;
2032 enum slab_modes l = M_NONE, m = M_NONE;
2033 void *nextfree;
2034 int tail = DEACTIVATE_TO_HEAD;
2035 struct page new;
2036 struct page old;
2037
2038 if (page->freelist) {
2039 stat(s, DEACTIVATE_REMOTE_FREES);
2040 tail = DEACTIVATE_TO_TAIL;
2041 }
2042
2043 /*
2044 * Stage one: Free all available per cpu objects back
2045 * to the page freelist while it is still frozen. Leave the
2046 * last one.
2047 *
2048 * There is no need to take the list->lock because the page
2049 * is still frozen.
2050 */
2051 while (freelist && (nextfree = get_freepointer(s, freelist))) {
2052 void *prior;
2053 unsigned long counters;
2054
2055 do {
2056 prior = page->freelist;
2057 counters = page->counters;
2058 set_freepointer(s, freelist, prior);
2059 new.counters = counters;
2060 new.inuse--;
2061 VM_BUG_ON(!new.frozen);
2062
2063 } while (!__cmpxchg_double_slab(s, page,
2064 prior, counters,
2065 freelist, new.counters,
2066 "drain percpu freelist"));
2067
2068 freelist = nextfree;
2069 }
2070
2071 /*
2072 * Stage two: Ensure that the page is unfrozen while the
2073 * list presence reflects the actual number of objects
2074 * during unfreeze.
2075 *
2076 * We setup the list membership and then perform a cmpxchg
2077 * with the count. If there is a mismatch then the page
2078 * is not unfrozen but the page is on the wrong list.
2079 *
2080 * Then we restart the process which may have to remove
2081 * the page from the list that we just put it on again
2082 * because the number of objects in the slab may have
2083 * changed.
2084 */
2085redo:
2086
2087 old.freelist = page->freelist;
2088 old.counters = page->counters;
2089 VM_BUG_ON(!old.frozen);
2090
2091 /* Determine target state of the slab */
2092 new.counters = old.counters;
2093 if (freelist) {
2094 new.inuse--;
2095 set_freepointer(s, freelist, old.freelist);
2096 new.freelist = freelist;
2097 } else
2098 new.freelist = old.freelist;
2099
2100 new.frozen = 0;
2101
2102 if (!new.inuse && n->nr_partial >= s->min_partial)
2103 m = M_FREE;
2104 else if (new.freelist) {
2105 m = M_PARTIAL;
2106 if (!lock) {
2107 lock = 1;
2108 /*
2109 * Taking the spinlock removes the possiblity
2110 * that acquire_slab() will see a slab page that
2111 * is frozen
2112 */
2113 spin_lock(&n->list_lock);
2114 }
2115 } else {
2116 m = M_FULL;
2117 if (kmem_cache_debug(s) && !lock) {
2118 lock = 1;
2119 /*
2120 * This also ensures that the scanning of full
2121 * slabs from diagnostic functions will not see
2122 * any frozen slabs.
2123 */
2124 spin_lock(&n->list_lock);
2125 }
2126 }
2127
2128 if (l != m) {
2129
2130 if (l == M_PARTIAL)
2131
2132 remove_partial(n, page);
2133
2134 else if (l == M_FULL)
2135
2136 remove_full(s, n, page);
2137
2138 if (m == M_PARTIAL) {
2139
2140 add_partial(n, page, tail);
2141 stat(s, tail);
2142
2143 } else if (m == M_FULL) {
2144
2145 stat(s, DEACTIVATE_FULL);
2146 add_full(s, n, page);
2147
2148 }
2149 }
2150
2151 l = m;
2152 if (!__cmpxchg_double_slab(s, page,
2153 old.freelist, old.counters,
2154 new.freelist, new.counters,
2155 "unfreezing slab"))
2156 goto redo;
2157
2158 if (lock)
2159 spin_unlock(&n->list_lock);
2160
2161 if (m == M_FREE) {
2162 stat(s, DEACTIVATE_EMPTY);
2163 discard_slab(s, page);
2164 stat(s, FREE_SLAB);
2165 }
2166
2167 c->page = NULL;
2168 c->freelist = NULL;
2169}
2170
2171/*
2172 * Unfreeze all the cpu partial slabs.
2173 *
2174 * This function must be called with interrupts disabled
2175 * for the cpu using c (or some other guarantee must be there
2176 * to guarantee no concurrent accesses).
2177 */
2178static void unfreeze_partials(struct kmem_cache *s,
2179 struct kmem_cache_cpu *c)
2180{
2181#ifdef CONFIG_SLUB_CPU_PARTIAL
2182 struct kmem_cache_node *n = NULL, *n2 = NULL;
2183 struct page *page, *discard_page = NULL;
2184
2185 while ((page = c->partial)) {
2186 struct page new;
2187 struct page old;
2188
2189 c->partial = page->next;
2190
2191 n2 = get_node(s, page_to_nid(page));
2192 if (n != n2) {
2193 if (n)
2194 spin_unlock(&n->list_lock);
2195
2196 n = n2;
2197 spin_lock(&n->list_lock);
2198 }
2199
2200 do {
2201
2202 old.freelist = page->freelist;
2203 old.counters = page->counters;
2204 VM_BUG_ON(!old.frozen);
2205
2206 new.counters = old.counters;
2207 new.freelist = old.freelist;
2208
2209 new.frozen = 0;
2210
2211 } while (!__cmpxchg_double_slab(s, page,
2212 old.freelist, old.counters,
2213 new.freelist, new.counters,
2214 "unfreezing slab"));
2215
2216 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2217 page->next = discard_page;
2218 discard_page = page;
2219 } else {
2220 add_partial(n, page, DEACTIVATE_TO_TAIL);
2221 stat(s, FREE_ADD_PARTIAL);
2222 }
2223 }
2224
2225 if (n)
2226 spin_unlock(&n->list_lock);
2227
2228 while (discard_page) {
2229 page = discard_page;
2230 discard_page = discard_page->next;
2231
2232 stat(s, DEACTIVATE_EMPTY);
2233 discard_slab(s, page);
2234 stat(s, FREE_SLAB);
2235 }
2236#endif
2237}
2238
2239/*
2240 * Put a page that was just frozen (in __slab_free) into a partial page
2241 * slot if available.
2242 *
2243 * If we did not find a slot then simply move all the partials to the
2244 * per node partial list.
2245 */
2246static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2247{
2248#ifdef CONFIG_SLUB_CPU_PARTIAL
2249 struct page *oldpage;
2250 int pages;
2251 int pobjects;
2252
2253 preempt_disable();
2254 do {
2255 pages = 0;
2256 pobjects = 0;
2257 oldpage = this_cpu_read(s->cpu_slab->partial);
2258
2259 if (oldpage) {
2260 pobjects = oldpage->pobjects;
2261 pages = oldpage->pages;
2262 if (drain && pobjects > s->cpu_partial) {
2263 unsigned long flags;
2264 /*
2265 * partial array is full. Move the existing
2266 * set to the per node partial list.
2267 */
2268 local_irq_save(flags);
2269 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2270 local_irq_restore(flags);
2271 oldpage = NULL;
2272 pobjects = 0;
2273 pages = 0;
2274 stat(s, CPU_PARTIAL_DRAIN);
2275 }
2276 }
2277
2278 pages++;
2279 pobjects += page->objects - page->inuse;
2280
2281 page->pages = pages;
2282 page->pobjects = pobjects;
2283 page->next = oldpage;
2284
2285 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2286 != oldpage);
2287 if (unlikely(!s->cpu_partial)) {
2288 unsigned long flags;
2289
2290 local_irq_save(flags);
2291 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2292 local_irq_restore(flags);
2293 }
2294 preempt_enable();
2295#endif
2296}
2297
2298static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2299{
2300 stat(s, CPUSLAB_FLUSH);
2301 deactivate_slab(s, c->page, c->freelist, c);
2302
2303 c->tid = next_tid(c->tid);
2304}
2305
2306/*
2307 * Flush cpu slab.
2308 *
2309 * Called from IPI handler with interrupts disabled.
2310 */
2311static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2312{
2313 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2314
2315 if (likely(c)) {
2316 if (c->page)
2317 flush_slab(s, c);
2318
2319 unfreeze_partials(s, c);
2320 }
2321}
2322
2323static void flush_cpu_slab(void *d)
2324{
2325 struct kmem_cache *s = d;
2326
2327 __flush_cpu_slab(s, smp_processor_id());
2328}
2329
2330static bool has_cpu_slab(int cpu, void *info)
2331{
2332 struct kmem_cache *s = info;
2333 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2334
2335 return c->page || slub_percpu_partial(c);
2336}
2337
2338static void flush_all(struct kmem_cache *s)
2339{
2340 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2341}
2342
2343/*
2344 * Use the cpu notifier to insure that the cpu slabs are flushed when
2345 * necessary.
2346 */
2347static int slub_cpu_dead(unsigned int cpu)
2348{
2349 struct kmem_cache *s;
2350 unsigned long flags;
2351
2352 mutex_lock(&slab_mutex);
2353 list_for_each_entry(s, &slab_caches, list) {
2354 local_irq_save(flags);
2355 __flush_cpu_slab(s, cpu);
2356 local_irq_restore(flags);
2357 }
2358 mutex_unlock(&slab_mutex);
2359 return 0;
2360}
2361
2362/*
2363 * Check if the objects in a per cpu structure fit numa
2364 * locality expectations.
2365 */
2366static inline int node_match(struct page *page, int node)
2367{
2368#ifdef CONFIG_NUMA
2369 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2370 return 0;
2371#endif
2372 return 1;
2373}
2374
2375#ifdef CONFIG_SLUB_DEBUG
2376static int count_free(struct page *page)
2377{
2378 return page->objects - page->inuse;
2379}
2380
2381static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2382{
2383 return atomic_long_read(&n->total_objects);
2384}
2385#endif /* CONFIG_SLUB_DEBUG */
2386
2387#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2388static unsigned long count_partial(struct kmem_cache_node *n,
2389 int (*get_count)(struct page *))
2390{
2391 unsigned long flags;
2392 unsigned long x = 0;
2393 struct page *page;
2394
2395 spin_lock_irqsave(&n->list_lock, flags);
2396 list_for_each_entry(page, &n->partial, lru)
2397 x += get_count(page);
2398 spin_unlock_irqrestore(&n->list_lock, flags);
2399 return x;
2400}
2401#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2402
2403static noinline void
2404slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2405{
2406#ifdef CONFIG_SLUB_DEBUG
2407 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2408 DEFAULT_RATELIMIT_BURST);
2409 int node;
2410 struct kmem_cache_node *n;
2411
2412 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2413 return;
2414
2415 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2416 nid, gfpflags, &gfpflags);
2417 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
2418 s->name, s->object_size, s->size, oo_order(s->oo),
2419 oo_order(s->min));
2420
2421 if (oo_order(s->min) > get_order(s->object_size))
2422 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2423 s->name);
2424
2425 for_each_kmem_cache_node(s, node, n) {
2426 unsigned long nr_slabs;
2427 unsigned long nr_objs;
2428 unsigned long nr_free;
2429
2430 nr_free = count_partial(n, count_free);
2431 nr_slabs = node_nr_slabs(n);
2432 nr_objs = node_nr_objs(n);
2433
2434 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2435 node, nr_slabs, nr_objs, nr_free);
2436 }
2437#endif
2438}
2439
2440static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2441 int node, struct kmem_cache_cpu **pc)
2442{
2443 void *freelist;
2444 struct kmem_cache_cpu *c = *pc;
2445 struct page *page;
2446
2447 freelist = get_partial(s, flags, node, c);
2448
2449 if (freelist)
2450 return freelist;
2451
2452 page = new_slab(s, flags, node);
2453 if (page) {
2454 c = raw_cpu_ptr(s->cpu_slab);
2455 if (c->page)
2456 flush_slab(s, c);
2457
2458 /*
2459 * No other reference to the page yet so we can
2460 * muck around with it freely without cmpxchg
2461 */
2462 freelist = page->freelist;
2463 page->freelist = NULL;
2464
2465 stat(s, ALLOC_SLAB);
2466 c->page = page;
2467 *pc = c;
2468 } else
2469 freelist = NULL;
2470
2471 return freelist;
2472}
2473
2474static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2475{
2476 if (unlikely(PageSlabPfmemalloc(page)))
2477 return gfp_pfmemalloc_allowed(gfpflags);
2478
2479 return true;
2480}
2481
2482/*
2483 * Check the page->freelist of a page and either transfer the freelist to the
2484 * per cpu freelist or deactivate the page.
2485 *
2486 * The page is still frozen if the return value is not NULL.
2487 *
2488 * If this function returns NULL then the page has been unfrozen.
2489 *
2490 * This function must be called with interrupt disabled.
2491 */
2492static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2493{
2494 struct page new;
2495 unsigned long counters;
2496 void *freelist;
2497
2498 do {
2499 freelist = page->freelist;
2500 counters = page->counters;
2501
2502 new.counters = counters;
2503 VM_BUG_ON(!new.frozen);
2504
2505 new.inuse = page->objects;
2506 new.frozen = freelist != NULL;
2507
2508 } while (!__cmpxchg_double_slab(s, page,
2509 freelist, counters,
2510 NULL, new.counters,
2511 "get_freelist"));
2512
2513 return freelist;
2514}
2515
2516/*
2517 * Slow path. The lockless freelist is empty or we need to perform
2518 * debugging duties.
2519 *
2520 * Processing is still very fast if new objects have been freed to the
2521 * regular freelist. In that case we simply take over the regular freelist
2522 * as the lockless freelist and zap the regular freelist.
2523 *
2524 * If that is not working then we fall back to the partial lists. We take the
2525 * first element of the freelist as the object to allocate now and move the
2526 * rest of the freelist to the lockless freelist.
2527 *
2528 * And if we were unable to get a new slab from the partial slab lists then
2529 * we need to allocate a new slab. This is the slowest path since it involves
2530 * a call to the page allocator and the setup of a new slab.
2531 *
2532 * Version of __slab_alloc to use when we know that interrupts are
2533 * already disabled (which is the case for bulk allocation).
2534 */
2535static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2536 unsigned long addr, struct kmem_cache_cpu *c)
2537{
2538 void *freelist;
2539 struct page *page;
2540
2541 page = c->page;
2542 if (!page)
2543 goto new_slab;
2544redo:
2545
2546 if (unlikely(!node_match(page, node))) {
2547 int searchnode = node;
2548
2549 if (node != NUMA_NO_NODE && !node_present_pages(node))
2550 searchnode = node_to_mem_node(node);
2551
2552 if (unlikely(!node_match(page, searchnode))) {
2553 stat(s, ALLOC_NODE_MISMATCH);
2554 deactivate_slab(s, page, c->freelist, c);
2555 goto new_slab;
2556 }
2557 }
2558
2559 /*
2560 * By rights, we should be searching for a slab page that was
2561 * PFMEMALLOC but right now, we are losing the pfmemalloc
2562 * information when the page leaves the per-cpu allocator
2563 */
2564 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2565 deactivate_slab(s, page, c->freelist, c);
2566 goto new_slab;
2567 }
2568
2569 /* must check again c->freelist in case of cpu migration or IRQ */
2570 freelist = c->freelist;
2571 if (freelist)
2572 goto load_freelist;
2573
2574 freelist = get_freelist(s, page);
2575
2576 if (!freelist) {
2577 c->page = NULL;
2578 stat(s, DEACTIVATE_BYPASS);
2579 goto new_slab;
2580 }
2581
2582 stat(s, ALLOC_REFILL);
2583
2584load_freelist:
2585 /*
2586 * freelist is pointing to the list of objects to be used.
2587 * page is pointing to the page from which the objects are obtained.
2588 * That page must be frozen for per cpu allocations to work.
2589 */
2590 VM_BUG_ON(!c->page->frozen);
2591 c->freelist = get_freepointer(s, freelist);
2592 c->tid = next_tid(c->tid);
2593 return freelist;
2594
2595new_slab:
2596
2597 if (slub_percpu_partial(c)) {
2598 page = c->page = slub_percpu_partial(c);
2599 slub_set_percpu_partial(c, page);
2600 stat(s, CPU_PARTIAL_ALLOC);
2601 goto redo;
2602 }
2603
2604 freelist = new_slab_objects(s, gfpflags, node, &c);
2605
2606 if (unlikely(!freelist)) {
2607 slab_out_of_memory(s, gfpflags, node);
2608 return NULL;
2609 }
2610
2611 page = c->page;
2612 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2613 goto load_freelist;
2614
2615 /* Only entered in the debug case */
2616 if (kmem_cache_debug(s) &&
2617 !alloc_debug_processing(s, page, freelist, addr))
2618 goto new_slab; /* Slab failed checks. Next slab needed */
2619
2620 deactivate_slab(s, page, get_freepointer(s, freelist), c);
2621 return freelist;
2622}
2623
2624/*
2625 * Another one that disabled interrupt and compensates for possible
2626 * cpu changes by refetching the per cpu area pointer.
2627 */
2628static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2629 unsigned long addr, struct kmem_cache_cpu *c)
2630{
2631 void *p;
2632 unsigned long flags;
2633
2634 local_irq_save(flags);
2635#ifdef CONFIG_PREEMPT
2636 /*
2637 * We may have been preempted and rescheduled on a different
2638 * cpu before disabling interrupts. Need to reload cpu area
2639 * pointer.
2640 */
2641 c = this_cpu_ptr(s->cpu_slab);
2642#endif
2643
2644 p = ___slab_alloc(s, gfpflags, node, addr, c);
2645 local_irq_restore(flags);
2646 return p;
2647}
2648
2649/*
2650 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2651 * have the fastpath folded into their functions. So no function call
2652 * overhead for requests that can be satisfied on the fastpath.
2653 *
2654 * The fastpath works by first checking if the lockless freelist can be used.
2655 * If not then __slab_alloc is called for slow processing.
2656 *
2657 * Otherwise we can simply pick the next object from the lockless free list.
2658 */
2659static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2660 gfp_t gfpflags, int node, unsigned long addr)
2661{
2662 void *object;
2663 struct kmem_cache_cpu *c;
2664 struct page *page;
2665 unsigned long tid;
2666
2667 s = slab_pre_alloc_hook(s, gfpflags);
2668 if (!s)
2669 return NULL;
2670redo:
2671 /*
2672 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2673 * enabled. We may switch back and forth between cpus while
2674 * reading from one cpu area. That does not matter as long
2675 * as we end up on the original cpu again when doing the cmpxchg.
2676 *
2677 * We should guarantee that tid and kmem_cache are retrieved on
2678 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2679 * to check if it is matched or not.
2680 */
2681 do {
2682 tid = this_cpu_read(s->cpu_slab->tid);
2683 c = raw_cpu_ptr(s->cpu_slab);
2684 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2685 unlikely(tid != READ_ONCE(c->tid)));
2686
2687 /*
2688 * Irqless object alloc/free algorithm used here depends on sequence
2689 * of fetching cpu_slab's data. tid should be fetched before anything
2690 * on c to guarantee that object and page associated with previous tid
2691 * won't be used with current tid. If we fetch tid first, object and
2692 * page could be one associated with next tid and our alloc/free
2693 * request will be failed. In this case, we will retry. So, no problem.
2694 */
2695 barrier();
2696
2697 /*
2698 * The transaction ids are globally unique per cpu and per operation on
2699 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2700 * occurs on the right processor and that there was no operation on the
2701 * linked list in between.
2702 */
2703
2704 object = c->freelist;
2705 page = c->page;
2706 if (unlikely(!object || !node_match(page, node))) {
2707 object = __slab_alloc(s, gfpflags, node, addr, c);
2708 stat(s, ALLOC_SLOWPATH);
2709 } else {
2710 void *next_object = get_freepointer_safe(s, object);
2711
2712 /*
2713 * The cmpxchg will only match if there was no additional
2714 * operation and if we are on the right processor.
2715 *
2716 * The cmpxchg does the following atomically (without lock
2717 * semantics!)
2718 * 1. Relocate first pointer to the current per cpu area.
2719 * 2. Verify that tid and freelist have not been changed
2720 * 3. If they were not changed replace tid and freelist
2721 *
2722 * Since this is without lock semantics the protection is only
2723 * against code executing on this cpu *not* from access by
2724 * other cpus.
2725 */
2726 if (unlikely(!this_cpu_cmpxchg_double(
2727 s->cpu_slab->freelist, s->cpu_slab->tid,
2728 object, tid,
2729 next_object, next_tid(tid)))) {
2730
2731 note_cmpxchg_failure("slab_alloc", s, tid);
2732 goto redo;
2733 }
2734 prefetch_freepointer(s, next_object);
2735 stat(s, ALLOC_FASTPATH);
2736 }
2737
2738 if (unlikely(gfpflags & __GFP_ZERO) && object)
2739 memset(object, 0, s->object_size);
2740
2741 slab_post_alloc_hook(s, gfpflags, 1, &object);
2742
2743 return object;
2744}
2745
2746static __always_inline void *slab_alloc(struct kmem_cache *s,
2747 gfp_t gfpflags, unsigned long addr)
2748{
2749 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2750}
2751
2752void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2753{
2754 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2755
2756 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2757 s->size, gfpflags);
2758
2759 return ret;
2760}
2761EXPORT_SYMBOL(kmem_cache_alloc);
2762
2763#ifdef CONFIG_TRACING
2764void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2765{
2766 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2767 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2768 kasan_kmalloc(s, ret, size, gfpflags);
2769 return ret;
2770}
2771EXPORT_SYMBOL(kmem_cache_alloc_trace);
2772#endif
2773
2774#ifdef CONFIG_NUMA
2775void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2776{
2777 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2778
2779 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2780 s->object_size, s->size, gfpflags, node);
2781
2782 return ret;
2783}
2784EXPORT_SYMBOL(kmem_cache_alloc_node);
2785
2786#ifdef CONFIG_TRACING
2787void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2788 gfp_t gfpflags,
2789 int node, size_t size)
2790{
2791 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2792
2793 trace_kmalloc_node(_RET_IP_, ret,
2794 size, s->size, gfpflags, node);
2795
2796 kasan_kmalloc(s, ret, size, gfpflags);
2797 return ret;
2798}
2799EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2800#endif
2801#endif
2802
2803/*
2804 * Slow path handling. This may still be called frequently since objects
2805 * have a longer lifetime than the cpu slabs in most processing loads.
2806 *
2807 * So we still attempt to reduce cache line usage. Just take the slab
2808 * lock and free the item. If there is no additional partial page
2809 * handling required then we can return immediately.
2810 */
2811static void __slab_free(struct kmem_cache *s, struct page *page,
2812 void *head, void *tail, int cnt,
2813 unsigned long addr)
2814
2815{
2816 void *prior;
2817 int was_frozen;
2818 struct page new;
2819 unsigned long counters;
2820 struct kmem_cache_node *n = NULL;
2821 unsigned long uninitialized_var(flags);
2822
2823 stat(s, FREE_SLOWPATH);
2824
2825 if (kmem_cache_debug(s) &&
2826 !free_debug_processing(s, page, head, tail, cnt, addr))
2827 return;
2828
2829 do {
2830 if (unlikely(n)) {
2831 spin_unlock_irqrestore(&n->list_lock, flags);
2832 n = NULL;
2833 }
2834 prior = page->freelist;
2835 counters = page->counters;
2836 set_freepointer(s, tail, prior);
2837 new.counters = counters;
2838 was_frozen = new.frozen;
2839 new.inuse -= cnt;
2840 if ((!new.inuse || !prior) && !was_frozen) {
2841
2842 if (kmem_cache_has_cpu_partial(s) && !prior) {
2843
2844 /*
2845 * Slab was on no list before and will be
2846 * partially empty
2847 * We can defer the list move and instead
2848 * freeze it.
2849 */
2850 new.frozen = 1;
2851
2852 } else { /* Needs to be taken off a list */
2853
2854 n = get_node(s, page_to_nid(page));
2855 /*
2856 * Speculatively acquire the list_lock.
2857 * If the cmpxchg does not succeed then we may
2858 * drop the list_lock without any processing.
2859 *
2860 * Otherwise the list_lock will synchronize with
2861 * other processors updating the list of slabs.
2862 */
2863 spin_lock_irqsave(&n->list_lock, flags);
2864
2865 }
2866 }
2867
2868 } while (!cmpxchg_double_slab(s, page,
2869 prior, counters,
2870 head, new.counters,
2871 "__slab_free"));
2872
2873 if (likely(!n)) {
2874
2875 /*
2876 * If we just froze the page then put it onto the
2877 * per cpu partial list.
2878 */
2879 if (new.frozen && !was_frozen) {
2880 put_cpu_partial(s, page, 1);
2881 stat(s, CPU_PARTIAL_FREE);
2882 }
2883 /*
2884 * The list lock was not taken therefore no list
2885 * activity can be necessary.
2886 */
2887 if (was_frozen)
2888 stat(s, FREE_FROZEN);
2889 return;
2890 }
2891
2892 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2893 goto slab_empty;
2894
2895 /*
2896 * Objects left in the slab. If it was not on the partial list before
2897 * then add it.
2898 */
2899 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2900 if (kmem_cache_debug(s))
2901 remove_full(s, n, page);
2902 add_partial(n, page, DEACTIVATE_TO_TAIL);
2903 stat(s, FREE_ADD_PARTIAL);
2904 }
2905 spin_unlock_irqrestore(&n->list_lock, flags);
2906 return;
2907
2908slab_empty:
2909 if (prior) {
2910 /*
2911 * Slab on the partial list.
2912 */
2913 remove_partial(n, page);
2914 stat(s, FREE_REMOVE_PARTIAL);
2915 } else {
2916 /* Slab must be on the full list */
2917 remove_full(s, n, page);
2918 }
2919
2920 spin_unlock_irqrestore(&n->list_lock, flags);
2921 stat(s, FREE_SLAB);
2922 discard_slab(s, page);
2923}
2924
2925/*
2926 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2927 * can perform fastpath freeing without additional function calls.
2928 *
2929 * The fastpath is only possible if we are freeing to the current cpu slab
2930 * of this processor. This typically the case if we have just allocated
2931 * the item before.
2932 *
2933 * If fastpath is not possible then fall back to __slab_free where we deal
2934 * with all sorts of special processing.
2935 *
2936 * Bulk free of a freelist with several objects (all pointing to the
2937 * same page) possible by specifying head and tail ptr, plus objects
2938 * count (cnt). Bulk free indicated by tail pointer being set.
2939 */
2940static __always_inline void do_slab_free(struct kmem_cache *s,
2941 struct page *page, void *head, void *tail,
2942 int cnt, unsigned long addr)
2943{
2944 void *tail_obj = tail ? : head;
2945 struct kmem_cache_cpu *c;
2946 unsigned long tid;
2947redo:
2948 /*
2949 * Determine the currently cpus per cpu slab.
2950 * The cpu may change afterward. However that does not matter since
2951 * data is retrieved via this pointer. If we are on the same cpu
2952 * during the cmpxchg then the free will succeed.
2953 */
2954 do {
2955 tid = this_cpu_read(s->cpu_slab->tid);
2956 c = raw_cpu_ptr(s->cpu_slab);
2957 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2958 unlikely(tid != READ_ONCE(c->tid)));
2959
2960 /* Same with comment on barrier() in slab_alloc_node() */
2961 barrier();
2962
2963 if (likely(page == c->page)) {
2964 set_freepointer(s, tail_obj, c->freelist);
2965
2966 if (unlikely(!this_cpu_cmpxchg_double(
2967 s->cpu_slab->freelist, s->cpu_slab->tid,
2968 c->freelist, tid,
2969 head, next_tid(tid)))) {
2970
2971 note_cmpxchg_failure("slab_free", s, tid);
2972 goto redo;
2973 }
2974 stat(s, FREE_FASTPATH);
2975 } else
2976 __slab_free(s, page, head, tail_obj, cnt, addr);
2977
2978}
2979
2980static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2981 void *head, void *tail, int cnt,
2982 unsigned long addr)
2983{
2984 /*
2985 * With KASAN enabled slab_free_freelist_hook modifies the freelist
2986 * to remove objects, whose reuse must be delayed.
2987 */
2988 if (slab_free_freelist_hook(s, &head, &tail))
2989 do_slab_free(s, page, head, tail, cnt, addr);
2990}
2991
2992#ifdef CONFIG_KASAN
2993void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
2994{
2995 do_slab_free(cache, virt_to_head_page(x), x, NULL, 1, addr);
2996}
2997#endif
2998
2999void kmem_cache_free(struct kmem_cache *s, void *x)
3000{
3001 s = cache_from_obj(s, x);
3002 if (!s)
3003 return;
3004 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
3005 trace_kmem_cache_free(_RET_IP_, x);
3006}
3007EXPORT_SYMBOL(kmem_cache_free);
3008
3009struct detached_freelist {
3010 struct page *page;
3011 void *tail;
3012 void *freelist;
3013 int cnt;
3014 struct kmem_cache *s;
3015};
3016
3017/*
3018 * This function progressively scans the array with free objects (with
3019 * a limited look ahead) and extract objects belonging to the same
3020 * page. It builds a detached freelist directly within the given
3021 * page/objects. This can happen without any need for
3022 * synchronization, because the objects are owned by running process.
3023 * The freelist is build up as a single linked list in the objects.
3024 * The idea is, that this detached freelist can then be bulk
3025 * transferred to the real freelist(s), but only requiring a single
3026 * synchronization primitive. Look ahead in the array is limited due
3027 * to performance reasons.
3028 */
3029static inline
3030int build_detached_freelist(struct kmem_cache *s, size_t size,
3031 void **p, struct detached_freelist *df)
3032{
3033 size_t first_skipped_index = 0;
3034 int lookahead = 3;
3035 void *object;
3036 struct page *page;
3037
3038 /* Always re-init detached_freelist */
3039 df->page = NULL;
3040
3041 do {
3042 object = p[--size];
3043 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
3044 } while (!object && size);
3045
3046 if (!object)
3047 return 0;
3048
3049 page = virt_to_head_page(object);
3050 if (!s) {
3051 /* Handle kalloc'ed objects */
3052 if (unlikely(!PageSlab(page))) {
3053 BUG_ON(!PageCompound(page));
3054 kfree_hook(object);
3055 __free_pages(page, compound_order(page));
3056 p[size] = NULL; /* mark object processed */
3057 return size;
3058 }
3059 /* Derive kmem_cache from object */
3060 df->s = page->slab_cache;
3061 } else {
3062 df->s = cache_from_obj(s, object); /* Support for memcg */
3063 }
3064
3065 /* Start new detached freelist */
3066 df->page = page;
3067 set_freepointer(df->s, object, NULL);
3068 df->tail = object;
3069 df->freelist = object;
3070 p[size] = NULL; /* mark object processed */
3071 df->cnt = 1;
3072
3073 while (size) {
3074 object = p[--size];
3075 if (!object)
3076 continue; /* Skip processed objects */
3077
3078 /* df->page is always set at this point */
3079 if (df->page == virt_to_head_page(object)) {
3080 /* Opportunity build freelist */
3081 set_freepointer(df->s, object, df->freelist);
3082 df->freelist = object;
3083 df->cnt++;
3084 p[size] = NULL; /* mark object processed */
3085
3086 continue;
3087 }
3088
3089 /* Limit look ahead search */
3090 if (!--lookahead)
3091 break;
3092
3093 if (!first_skipped_index)
3094 first_skipped_index = size + 1;
3095 }
3096
3097 return first_skipped_index;
3098}
3099
3100/* Note that interrupts must be enabled when calling this function. */
3101void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
3102{
3103 if (WARN_ON(!size))
3104 return;
3105
3106 do {
3107 struct detached_freelist df;
3108
3109 size = build_detached_freelist(s, size, p, &df);
3110 if (!df.page)
3111 continue;
3112
3113 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
3114 } while (likely(size));
3115}
3116EXPORT_SYMBOL(kmem_cache_free_bulk);
3117
3118/* Note that interrupts must be enabled when calling this function. */
3119int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
3120 void **p)
3121{
3122 struct kmem_cache_cpu *c;
3123 int i;
3124
3125 /* memcg and kmem_cache debug support */
3126 s = slab_pre_alloc_hook(s, flags);
3127 if (unlikely(!s))
3128 return false;
3129 /*
3130 * Drain objects in the per cpu slab, while disabling local
3131 * IRQs, which protects against PREEMPT and interrupts
3132 * handlers invoking normal fastpath.
3133 */
3134 local_irq_disable();
3135 c = this_cpu_ptr(s->cpu_slab);
3136
3137 for (i = 0; i < size; i++) {
3138 void *object = c->freelist;
3139
3140 if (unlikely(!object)) {
3141 /*
3142 * Invoking slow path likely have side-effect
3143 * of re-populating per CPU c->freelist
3144 */
3145 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
3146 _RET_IP_, c);
3147 if (unlikely(!p[i]))
3148 goto error;
3149
3150 c = this_cpu_ptr(s->cpu_slab);
3151 continue; /* goto for-loop */
3152 }
3153 c->freelist = get_freepointer(s, object);
3154 p[i] = object;
3155 }
3156 c->tid = next_tid(c->tid);
3157 local_irq_enable();
3158
3159 /* Clear memory outside IRQ disabled fastpath loop */
3160 if (unlikely(flags & __GFP_ZERO)) {
3161 int j;
3162
3163 for (j = 0; j < i; j++)
3164 memset(p[j], 0, s->object_size);
3165 }
3166
3167 /* memcg and kmem_cache debug support */
3168 slab_post_alloc_hook(s, flags, size, p);
3169 return i;
3170error:
3171 local_irq_enable();
3172 slab_post_alloc_hook(s, flags, i, p);
3173 __kmem_cache_free_bulk(s, i, p);
3174 return 0;
3175}
3176EXPORT_SYMBOL(kmem_cache_alloc_bulk);
3177
3178
3179/*
3180 * Object placement in a slab is made very easy because we always start at
3181 * offset 0. If we tune the size of the object to the alignment then we can
3182 * get the required alignment by putting one properly sized object after
3183 * another.
3184 *
3185 * Notice that the allocation order determines the sizes of the per cpu
3186 * caches. Each processor has always one slab available for allocations.
3187 * Increasing the allocation order reduces the number of times that slabs
3188 * must be moved on and off the partial lists and is therefore a factor in
3189 * locking overhead.
3190 */
3191
3192/*
3193 * Mininum / Maximum order of slab pages. This influences locking overhead
3194 * and slab fragmentation. A higher order reduces the number of partial slabs
3195 * and increases the number of allocations possible without having to
3196 * take the list_lock.
3197 */
3198static unsigned int slub_min_order;
3199static unsigned int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3200static unsigned int slub_min_objects;
3201
3202/*
3203 * Calculate the order of allocation given an slab object size.
3204 *
3205 * The order of allocation has significant impact on performance and other
3206 * system components. Generally order 0 allocations should be preferred since
3207 * order 0 does not cause fragmentation in the page allocator. Larger objects
3208 * be problematic to put into order 0 slabs because there may be too much
3209 * unused space left. We go to a higher order if more than 1/16th of the slab
3210 * would be wasted.
3211 *
3212 * In order to reach satisfactory performance we must ensure that a minimum
3213 * number of objects is in one slab. Otherwise we may generate too much
3214 * activity on the partial lists which requires taking the list_lock. This is
3215 * less a concern for large slabs though which are rarely used.
3216 *
3217 * slub_max_order specifies the order where we begin to stop considering the
3218 * number of objects in a slab as critical. If we reach slub_max_order then
3219 * we try to keep the page order as low as possible. So we accept more waste
3220 * of space in favor of a small page order.
3221 *
3222 * Higher order allocations also allow the placement of more objects in a
3223 * slab and thereby reduce object handling overhead. If the user has
3224 * requested a higher mininum order then we start with that one instead of
3225 * the smallest order which will fit the object.
3226 */
3227static inline unsigned int slab_order(unsigned int size,
3228 unsigned int min_objects, unsigned int max_order,
3229 unsigned int fract_leftover, unsigned int reserved)
3230{
3231 unsigned int min_order = slub_min_order;
3232 unsigned int order;
3233
3234 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3235 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3236
3237 for (order = max(min_order, (unsigned int)get_order(min_objects * size + reserved));
3238 order <= max_order; order++) {
3239
3240 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
3241 unsigned int rem;
3242
3243 rem = (slab_size - reserved) % size;
3244
3245 if (rem <= slab_size / fract_leftover)
3246 break;
3247 }
3248
3249 return order;
3250}
3251
3252static inline int calculate_order(unsigned int size, unsigned int reserved)
3253{
3254 unsigned int order;
3255 unsigned int min_objects;
3256 unsigned int max_objects;
3257
3258 /*
3259 * Attempt to find best configuration for a slab. This
3260 * works by first attempting to generate a layout with
3261 * the best configuration and backing off gradually.
3262 *
3263 * First we increase the acceptable waste in a slab. Then
3264 * we reduce the minimum objects required in a slab.
3265 */
3266 min_objects = slub_min_objects;
3267 if (!min_objects)
3268 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3269 max_objects = order_objects(slub_max_order, size, reserved);
3270 min_objects = min(min_objects, max_objects);
3271
3272 while (min_objects > 1) {
3273 unsigned int fraction;
3274
3275 fraction = 16;
3276 while (fraction >= 4) {
3277 order = slab_order(size, min_objects,
3278 slub_max_order, fraction, reserved);
3279 if (order <= slub_max_order)
3280 return order;
3281 fraction /= 2;
3282 }
3283 min_objects--;
3284 }
3285
3286 /*
3287 * We were unable to place multiple objects in a slab. Now
3288 * lets see if we can place a single object there.
3289 */
3290 order = slab_order(size, 1, slub_max_order, 1, reserved);
3291 if (order <= slub_max_order)
3292 return order;
3293
3294 /*
3295 * Doh this slab cannot be placed using slub_max_order.
3296 */
3297 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3298 if (order < MAX_ORDER)
3299 return order;
3300 return -ENOSYS;
3301}
3302
3303static void
3304init_kmem_cache_node(struct kmem_cache_node *n)
3305{
3306 n->nr_partial = 0;
3307 spin_lock_init(&n->list_lock);
3308 INIT_LIST_HEAD(&n->partial);
3309#ifdef CONFIG_SLUB_DEBUG
3310 atomic_long_set(&n->nr_slabs, 0);
3311 atomic_long_set(&n->total_objects, 0);
3312 INIT_LIST_HEAD(&n->full);
3313#endif
3314}
3315
3316static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3317{
3318 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3319 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3320
3321 /*
3322 * Must align to double word boundary for the double cmpxchg
3323 * instructions to work; see __pcpu_double_call_return_bool().
3324 */
3325 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3326 2 * sizeof(void *));
3327
3328 if (!s->cpu_slab)
3329 return 0;
3330
3331 init_kmem_cache_cpus(s);
3332
3333 return 1;
3334}
3335
3336static struct kmem_cache *kmem_cache_node;
3337
3338/*
3339 * No kmalloc_node yet so do it by hand. We know that this is the first
3340 * slab on the node for this slabcache. There are no concurrent accesses
3341 * possible.
3342 *
3343 * Note that this function only works on the kmem_cache_node
3344 * when allocating for the kmem_cache_node. This is used for bootstrapping
3345 * memory on a fresh node that has no slab structures yet.
3346 */
3347static void early_kmem_cache_node_alloc(int node)
3348{
3349 struct page *page;
3350 struct kmem_cache_node *n;
3351
3352 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3353
3354 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3355
3356 BUG_ON(!page);
3357 if (page_to_nid(page) != node) {
3358 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3359 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3360 }
3361
3362 n = page->freelist;
3363 BUG_ON(!n);
3364 page->freelist = get_freepointer(kmem_cache_node, n);
3365 page->inuse = 1;
3366 page->frozen = 0;
3367 kmem_cache_node->node[node] = n;
3368#ifdef CONFIG_SLUB_DEBUG
3369 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3370 init_tracking(kmem_cache_node, n);
3371#endif
3372 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3373 GFP_KERNEL);
3374 init_kmem_cache_node(n);
3375 inc_slabs_node(kmem_cache_node, node, page->objects);
3376
3377 /*
3378 * No locks need to be taken here as it has just been
3379 * initialized and there is no concurrent access.
3380 */
3381 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3382}
3383
3384static void free_kmem_cache_nodes(struct kmem_cache *s)
3385{
3386 int node;
3387 struct kmem_cache_node *n;
3388
3389 for_each_kmem_cache_node(s, node, n) {
3390 s->node[node] = NULL;
3391 kmem_cache_free(kmem_cache_node, n);
3392 }
3393}
3394
3395void __kmem_cache_release(struct kmem_cache *s)
3396{
3397 cache_random_seq_destroy(s);
3398 free_percpu(s->cpu_slab);
3399 free_kmem_cache_nodes(s);
3400}
3401
3402static int init_kmem_cache_nodes(struct kmem_cache *s)
3403{
3404 int node;
3405
3406 for_each_node_state(node, N_NORMAL_MEMORY) {
3407 struct kmem_cache_node *n;
3408
3409 if (slab_state == DOWN) {
3410 early_kmem_cache_node_alloc(node);
3411 continue;
3412 }
3413 n = kmem_cache_alloc_node(kmem_cache_node,
3414 GFP_KERNEL, node);
3415
3416 if (!n) {
3417 free_kmem_cache_nodes(s);
3418 return 0;
3419 }
3420
3421 init_kmem_cache_node(n);
3422 s->node[node] = n;
3423 }
3424 return 1;
3425}
3426
3427static void set_min_partial(struct kmem_cache *s, unsigned long min)
3428{
3429 if (min < MIN_PARTIAL)
3430 min = MIN_PARTIAL;
3431 else if (min > MAX_PARTIAL)
3432 min = MAX_PARTIAL;
3433 s->min_partial = min;
3434}
3435
3436static void set_cpu_partial(struct kmem_cache *s)
3437{
3438#ifdef CONFIG_SLUB_CPU_PARTIAL
3439 /*
3440 * cpu_partial determined the maximum number of objects kept in the
3441 * per cpu partial lists of a processor.
3442 *
3443 * Per cpu partial lists mainly contain slabs that just have one
3444 * object freed. If they are used for allocation then they can be
3445 * filled up again with minimal effort. The slab will never hit the
3446 * per node partial lists and therefore no locking will be required.
3447 *
3448 * This setting also determines
3449 *
3450 * A) The number of objects from per cpu partial slabs dumped to the
3451 * per node list when we reach the limit.
3452 * B) The number of objects in cpu partial slabs to extract from the
3453 * per node list when we run out of per cpu objects. We only fetch
3454 * 50% to keep some capacity around for frees.
3455 */
3456 if (!kmem_cache_has_cpu_partial(s))
3457 s->cpu_partial = 0;
3458 else if (s->size >= PAGE_SIZE)
3459 s->cpu_partial = 2;
3460 else if (s->size >= 1024)
3461 s->cpu_partial = 6;
3462 else if (s->size >= 256)
3463 s->cpu_partial = 13;
3464 else
3465 s->cpu_partial = 30;
3466#endif
3467}
3468
3469/*
3470 * calculate_sizes() determines the order and the distribution of data within
3471 * a slab object.
3472 */
3473static int calculate_sizes(struct kmem_cache *s, int forced_order)
3474{
3475 slab_flags_t flags = s->flags;
3476 unsigned int size = s->object_size;
3477 unsigned int order;
3478
3479 /*
3480 * Round up object size to the next word boundary. We can only
3481 * place the free pointer at word boundaries and this determines
3482 * the possible location of the free pointer.
3483 */
3484 size = ALIGN(size, sizeof(void *));
3485
3486#ifdef CONFIG_SLUB_DEBUG
3487 /*
3488 * Determine if we can poison the object itself. If the user of
3489 * the slab may touch the object after free or before allocation
3490 * then we should never poison the object itself.
3491 */
3492 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
3493 !s->ctor)
3494 s->flags |= __OBJECT_POISON;
3495 else
3496 s->flags &= ~__OBJECT_POISON;
3497
3498
3499 /*
3500 * If we are Redzoning then check if there is some space between the
3501 * end of the object and the free pointer. If not then add an
3502 * additional word to have some bytes to store Redzone information.
3503 */
3504 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3505 size += sizeof(void *);
3506#endif
3507
3508 /*
3509 * With that we have determined the number of bytes in actual use
3510 * by the object. This is the potential offset to the free pointer.
3511 */
3512 s->inuse = size;
3513
3514 if (((flags & (SLAB_TYPESAFE_BY_RCU | SLAB_POISON)) ||
3515 s->ctor)) {
3516 /*
3517 * Relocate free pointer after the object if it is not
3518 * permitted to overwrite the first word of the object on
3519 * kmem_cache_free.
3520 *
3521 * This is the case if we do RCU, have a constructor or
3522 * destructor or are poisoning the objects.
3523 */
3524 s->offset = size;
3525 size += sizeof(void *);
3526 }
3527
3528#ifdef CONFIG_SLUB_DEBUG
3529 if (flags & SLAB_STORE_USER)
3530 /*
3531 * Need to store information about allocs and frees after
3532 * the object.
3533 */
3534 size += 2 * sizeof(struct track);
3535#endif
3536
3537 kasan_cache_create(s, &size, &s->flags);
3538#ifdef CONFIG_SLUB_DEBUG
3539 if (flags & SLAB_RED_ZONE) {
3540 /*
3541 * Add some empty padding so that we can catch
3542 * overwrites from earlier objects rather than let
3543 * tracking information or the free pointer be
3544 * corrupted if a user writes before the start
3545 * of the object.
3546 */
3547 size += sizeof(void *);
3548
3549 s->red_left_pad = sizeof(void *);
3550 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3551 size += s->red_left_pad;
3552 }
3553#endif
3554
3555 /*
3556 * SLUB stores one object immediately after another beginning from
3557 * offset 0. In order to align the objects we have to simply size
3558 * each object to conform to the alignment.
3559 */
3560 size = ALIGN(size, s->align);
3561 s->size = size;
3562 if (forced_order >= 0)
3563 order = forced_order;
3564 else
3565 order = calculate_order(size, s->reserved);
3566
3567 if ((int)order < 0)
3568 return 0;
3569
3570 s->allocflags = 0;
3571 if (order)
3572 s->allocflags |= __GFP_COMP;
3573
3574 if (s->flags & SLAB_CACHE_DMA)
3575 s->allocflags |= GFP_DMA;
3576
3577 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3578 s->allocflags |= __GFP_RECLAIMABLE;
3579
3580 /*
3581 * Determine the number of objects per slab
3582 */
3583 s->oo = oo_make(order, size, s->reserved);
3584 s->min = oo_make(get_order(size), size, s->reserved);
3585 if (oo_objects(s->oo) > oo_objects(s->max))
3586 s->max = s->oo;
3587
3588 return !!oo_objects(s->oo);
3589}
3590
3591static int kmem_cache_open(struct kmem_cache *s, slab_flags_t flags)
3592{
3593 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3594 s->reserved = 0;
3595#ifdef CONFIG_SLAB_FREELIST_HARDENED
3596 s->random = get_random_long();
3597#endif
3598
3599 if (need_reserve_slab_rcu && (s->flags & SLAB_TYPESAFE_BY_RCU))
3600 s->reserved = sizeof(struct rcu_head);
3601
3602 if (!calculate_sizes(s, -1))
3603 goto error;
3604 if (disable_higher_order_debug) {
3605 /*
3606 * Disable debugging flags that store metadata if the min slab
3607 * order increased.
3608 */
3609 if (get_order(s->size) > get_order(s->object_size)) {
3610 s->flags &= ~DEBUG_METADATA_FLAGS;
3611 s->offset = 0;
3612 if (!calculate_sizes(s, -1))
3613 goto error;
3614 }
3615 }
3616
3617#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3618 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3619 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3620 /* Enable fast mode */
3621 s->flags |= __CMPXCHG_DOUBLE;
3622#endif
3623
3624 /*
3625 * The larger the object size is, the more pages we want on the partial
3626 * list to avoid pounding the page allocator excessively.
3627 */
3628 set_min_partial(s, ilog2(s->size) / 2);
3629
3630 set_cpu_partial(s);
3631
3632#ifdef CONFIG_NUMA
3633 s->remote_node_defrag_ratio = 1000;
3634#endif
3635
3636 /* Initialize the pre-computed randomized freelist if slab is up */
3637 if (slab_state >= UP) {
3638 if (init_cache_random_seq(s))
3639 goto error;
3640 }
3641
3642 if (!init_kmem_cache_nodes(s))
3643 goto error;
3644
3645 if (alloc_kmem_cache_cpus(s))
3646 return 0;
3647
3648 free_kmem_cache_nodes(s);
3649error:
3650 if (flags & SLAB_PANIC)
3651 panic("Cannot create slab %s size=%u realsize=%u order=%u offset=%u flags=%lx\n",
3652 s->name, s->size, s->size,
3653 oo_order(s->oo), s->offset, (unsigned long)flags);
3654 return -EINVAL;
3655}
3656
3657static void list_slab_objects(struct kmem_cache *s, struct page *page,
3658 const char *text)
3659{
3660#ifdef CONFIG_SLUB_DEBUG
3661 void *addr = page_address(page);
3662 void *p;
3663 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3664 sizeof(long), GFP_ATOMIC);
3665 if (!map)
3666 return;
3667 slab_err(s, page, text, s->name);
3668 slab_lock(page);
3669
3670 get_map(s, page, map);
3671 for_each_object(p, s, addr, page->objects) {
3672
3673 if (!test_bit(slab_index(p, s, addr), map)) {
3674 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3675 print_tracking(s, p);
3676 }
3677 }
3678 slab_unlock(page);
3679 kfree(map);
3680#endif
3681}
3682
3683/*
3684 * Attempt to free all partial slabs on a node.
3685 * This is called from __kmem_cache_shutdown(). We must take list_lock
3686 * because sysfs file might still access partial list after the shutdowning.
3687 */
3688static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3689{
3690 LIST_HEAD(discard);
3691 struct page *page, *h;
3692
3693 BUG_ON(irqs_disabled());
3694 spin_lock_irq(&n->list_lock);
3695 list_for_each_entry_safe(page, h, &n->partial, lru) {
3696 if (!page->inuse) {
3697 remove_partial(n, page);
3698 list_add(&page->lru, &discard);
3699 } else {
3700 list_slab_objects(s, page,
3701 "Objects remaining in %s on __kmem_cache_shutdown()");
3702 }
3703 }
3704 spin_unlock_irq(&n->list_lock);
3705
3706 list_for_each_entry_safe(page, h, &discard, lru)
3707 discard_slab(s, page);
3708}
3709
3710bool __kmem_cache_empty(struct kmem_cache *s)
3711{
3712 int node;
3713 struct kmem_cache_node *n;
3714
3715 for_each_kmem_cache_node(s, node, n)
3716 if (n->nr_partial || slabs_node(s, node))
3717 return false;
3718 return true;
3719}
3720
3721/*
3722 * Release all resources used by a slab cache.
3723 */
3724int __kmem_cache_shutdown(struct kmem_cache *s)
3725{
3726 int node;
3727 struct kmem_cache_node *n;
3728
3729 flush_all(s);
3730 /* Attempt to free all objects */
3731 for_each_kmem_cache_node(s, node, n) {
3732 free_partial(s, n);
3733 if (n->nr_partial || slabs_node(s, node))
3734 return 1;
3735 }
3736 sysfs_slab_remove(s);
3737 return 0;
3738}
3739
3740/********************************************************************
3741 * Kmalloc subsystem
3742 *******************************************************************/
3743
3744static int __init setup_slub_min_order(char *str)
3745{
3746 get_option(&str, (int *)&slub_min_order);
3747
3748 return 1;
3749}
3750
3751__setup("slub_min_order=", setup_slub_min_order);
3752
3753static int __init setup_slub_max_order(char *str)
3754{
3755 get_option(&str, (int *)&slub_max_order);
3756 slub_max_order = min(slub_max_order, (unsigned int)MAX_ORDER - 1);
3757
3758 return 1;
3759}
3760
3761__setup("slub_max_order=", setup_slub_max_order);
3762
3763static int __init setup_slub_min_objects(char *str)
3764{
3765 get_option(&str, (int *)&slub_min_objects);
3766
3767 return 1;
3768}
3769
3770__setup("slub_min_objects=", setup_slub_min_objects);
3771
3772void *__kmalloc(size_t size, gfp_t flags)
3773{
3774 struct kmem_cache *s;
3775 void *ret;
3776
3777 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3778 return kmalloc_large(size, flags);
3779
3780 s = kmalloc_slab(size, flags);
3781
3782 if (unlikely(ZERO_OR_NULL_PTR(s)))
3783 return s;
3784
3785 ret = slab_alloc(s, flags, _RET_IP_);
3786
3787 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3788
3789 kasan_kmalloc(s, ret, size, flags);
3790
3791 return ret;
3792}
3793EXPORT_SYMBOL(__kmalloc);
3794
3795#ifdef CONFIG_NUMA
3796static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3797{
3798 struct page *page;
3799 void *ptr = NULL;
3800
3801 flags |= __GFP_COMP;
3802 page = alloc_pages_node(node, flags, get_order(size));
3803 if (page)
3804 ptr = page_address(page);
3805
3806 kmalloc_large_node_hook(ptr, size, flags);
3807 return ptr;
3808}
3809
3810void *__kmalloc_node(size_t size, gfp_t flags, int node)
3811{
3812 struct kmem_cache *s;
3813 void *ret;
3814
3815 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3816 ret = kmalloc_large_node(size, flags, node);
3817
3818 trace_kmalloc_node(_RET_IP_, ret,
3819 size, PAGE_SIZE << get_order(size),
3820 flags, node);
3821
3822 return ret;
3823 }
3824
3825 s = kmalloc_slab(size, flags);
3826
3827 if (unlikely(ZERO_OR_NULL_PTR(s)))
3828 return s;
3829
3830 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3831
3832 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3833
3834 kasan_kmalloc(s, ret, size, flags);
3835
3836 return ret;
3837}
3838EXPORT_SYMBOL(__kmalloc_node);
3839#endif
3840
3841#ifdef CONFIG_HARDENED_USERCOPY
3842/*
3843 * Rejects incorrectly sized objects and objects that are to be copied
3844 * to/from userspace but do not fall entirely within the containing slab
3845 * cache's usercopy region.
3846 *
3847 * Returns NULL if check passes, otherwise const char * to name of cache
3848 * to indicate an error.
3849 */
3850void __check_heap_object(const void *ptr, unsigned long n, struct page *page,
3851 bool to_user)
3852{
3853 struct kmem_cache *s;
3854 unsigned int offset;
3855 size_t object_size;
3856
3857 /* Find object and usable object size. */
3858 s = page->slab_cache;
3859
3860 /* Reject impossible pointers. */
3861 if (ptr < page_address(page))
3862 usercopy_abort("SLUB object not in SLUB page?!", NULL,
3863 to_user, 0, n);
3864
3865 /* Find offset within object. */
3866 offset = (ptr - page_address(page)) % s->size;
3867
3868 /* Adjust for redzone and reject if within the redzone. */
3869 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE) {
3870 if (offset < s->red_left_pad)
3871 usercopy_abort("SLUB object in left red zone",
3872 s->name, to_user, offset, n);
3873 offset -= s->red_left_pad;
3874 }
3875
3876 /* Allow address range falling entirely within usercopy region. */
3877 if (offset >= s->useroffset &&
3878 offset - s->useroffset <= s->usersize &&
3879 n <= s->useroffset - offset + s->usersize)
3880 return;
3881
3882 /*
3883 * If the copy is still within the allocated object, produce
3884 * a warning instead of rejecting the copy. This is intended
3885 * to be a temporary method to find any missing usercopy
3886 * whitelists.
3887 */
3888 object_size = slab_ksize(s);
3889 if (usercopy_fallback &&
3890 offset <= object_size && n <= object_size - offset) {
3891 usercopy_warn("SLUB object", s->name, to_user, offset, n);
3892 return;
3893 }
3894
3895 usercopy_abort("SLUB object", s->name, to_user, offset, n);
3896}
3897#endif /* CONFIG_HARDENED_USERCOPY */
3898
3899static size_t __ksize(const void *object)
3900{
3901 struct page *page;
3902
3903 if (unlikely(object == ZERO_SIZE_PTR))
3904 return 0;
3905
3906 page = virt_to_head_page(object);
3907
3908 if (unlikely(!PageSlab(page))) {
3909 WARN_ON(!PageCompound(page));
3910 return PAGE_SIZE << compound_order(page);
3911 }
3912
3913 return slab_ksize(page->slab_cache);
3914}
3915
3916size_t ksize(const void *object)
3917{
3918 size_t size = __ksize(object);
3919 /* We assume that ksize callers could use whole allocated area,
3920 * so we need to unpoison this area.
3921 */
3922 kasan_unpoison_shadow(object, size);
3923 return size;
3924}
3925EXPORT_SYMBOL(ksize);
3926
3927void kfree(const void *x)
3928{
3929 struct page *page;
3930 void *object = (void *)x;
3931
3932 trace_kfree(_RET_IP_, x);
3933
3934 if (unlikely(ZERO_OR_NULL_PTR(x)))
3935 return;
3936
3937 page = virt_to_head_page(x);
3938 if (unlikely(!PageSlab(page))) {
3939 BUG_ON(!PageCompound(page));
3940 kfree_hook(object);
3941 __free_pages(page, compound_order(page));
3942 return;
3943 }
3944 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3945}
3946EXPORT_SYMBOL(kfree);
3947
3948#define SHRINK_PROMOTE_MAX 32
3949
3950/*
3951 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3952 * up most to the head of the partial lists. New allocations will then
3953 * fill those up and thus they can be removed from the partial lists.
3954 *
3955 * The slabs with the least items are placed last. This results in them
3956 * being allocated from last increasing the chance that the last objects
3957 * are freed in them.
3958 */
3959int __kmem_cache_shrink(struct kmem_cache *s)
3960{
3961 int node;
3962 int i;
3963 struct kmem_cache_node *n;
3964 struct page *page;
3965 struct page *t;
3966 struct list_head discard;
3967 struct list_head promote[SHRINK_PROMOTE_MAX];
3968 unsigned long flags;
3969 int ret = 0;
3970
3971 flush_all(s);
3972 for_each_kmem_cache_node(s, node, n) {
3973 INIT_LIST_HEAD(&discard);
3974 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3975 INIT_LIST_HEAD(promote + i);
3976
3977 spin_lock_irqsave(&n->list_lock, flags);
3978
3979 /*
3980 * Build lists of slabs to discard or promote.
3981 *
3982 * Note that concurrent frees may occur while we hold the
3983 * list_lock. page->inuse here is the upper limit.
3984 */
3985 list_for_each_entry_safe(page, t, &n->partial, lru) {
3986 int free = page->objects - page->inuse;
3987
3988 /* Do not reread page->inuse */
3989 barrier();
3990
3991 /* We do not keep full slabs on the list */
3992 BUG_ON(free <= 0);
3993
3994 if (free == page->objects) {
3995 list_move(&page->lru, &discard);
3996 n->nr_partial--;
3997 } else if (free <= SHRINK_PROMOTE_MAX)
3998 list_move(&page->lru, promote + free - 1);
3999 }
4000
4001 /*
4002 * Promote the slabs filled up most to the head of the
4003 * partial list.
4004 */
4005 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
4006 list_splice(promote + i, &n->partial);
4007
4008 spin_unlock_irqrestore(&n->list_lock, flags);
4009
4010 /* Release empty slabs */
4011 list_for_each_entry_safe(page, t, &discard, lru)
4012 discard_slab(s, page);
4013
4014 if (slabs_node(s, node))
4015 ret = 1;
4016 }
4017
4018 return ret;
4019}
4020
4021#ifdef CONFIG_MEMCG
4022static void kmemcg_cache_deact_after_rcu(struct kmem_cache *s)
4023{
4024 /*
4025 * Called with all the locks held after a sched RCU grace period.
4026 * Even if @s becomes empty after shrinking, we can't know that @s
4027 * doesn't have allocations already in-flight and thus can't
4028 * destroy @s until the associated memcg is released.
4029 *
4030 * However, let's remove the sysfs files for empty caches here.
4031 * Each cache has a lot of interface files which aren't
4032 * particularly useful for empty draining caches; otherwise, we can
4033 * easily end up with millions of unnecessary sysfs files on
4034 * systems which have a lot of memory and transient cgroups.
4035 */
4036 if (!__kmem_cache_shrink(s))
4037 sysfs_slab_remove(s);
4038}
4039
4040void __kmemcg_cache_deactivate(struct kmem_cache *s)
4041{
4042 /*
4043 * Disable empty slabs caching. Used to avoid pinning offline
4044 * memory cgroups by kmem pages that can be freed.
4045 */
4046 slub_set_cpu_partial(s, 0);
4047 s->min_partial = 0;
4048
4049 /*
4050 * s->cpu_partial is checked locklessly (see put_cpu_partial), so
4051 * we have to make sure the change is visible before shrinking.
4052 */
4053 slab_deactivate_memcg_cache_rcu_sched(s, kmemcg_cache_deact_after_rcu);
4054}
4055#endif
4056
4057static int slab_mem_going_offline_callback(void *arg)
4058{
4059 struct kmem_cache *s;
4060
4061 mutex_lock(&slab_mutex);
4062 list_for_each_entry(s, &slab_caches, list)
4063 __kmem_cache_shrink(s);
4064 mutex_unlock(&slab_mutex);
4065
4066 return 0;
4067}
4068
4069static void slab_mem_offline_callback(void *arg)
4070{
4071 struct kmem_cache_node *n;
4072 struct kmem_cache *s;
4073 struct memory_notify *marg = arg;
4074 int offline_node;
4075
4076 offline_node = marg->status_change_nid_normal;
4077
4078 /*
4079 * If the node still has available memory. we need kmem_cache_node
4080 * for it yet.
4081 */
4082 if (offline_node < 0)
4083 return;
4084
4085 mutex_lock(&slab_mutex);
4086 list_for_each_entry(s, &slab_caches, list) {
4087 n = get_node(s, offline_node);
4088 if (n) {
4089 /*
4090 * if n->nr_slabs > 0, slabs still exist on the node
4091 * that is going down. We were unable to free them,
4092 * and offline_pages() function shouldn't call this
4093 * callback. So, we must fail.
4094 */
4095 BUG_ON(slabs_node(s, offline_node));
4096
4097 s->node[offline_node] = NULL;
4098 kmem_cache_free(kmem_cache_node, n);
4099 }
4100 }
4101 mutex_unlock(&slab_mutex);
4102}
4103
4104static int slab_mem_going_online_callback(void *arg)
4105{
4106 struct kmem_cache_node *n;
4107 struct kmem_cache *s;
4108 struct memory_notify *marg = arg;
4109 int nid = marg->status_change_nid_normal;
4110 int ret = 0;
4111
4112 /*
4113 * If the node's memory is already available, then kmem_cache_node is
4114 * already created. Nothing to do.
4115 */
4116 if (nid < 0)
4117 return 0;
4118
4119 /*
4120 * We are bringing a node online. No memory is available yet. We must
4121 * allocate a kmem_cache_node structure in order to bring the node
4122 * online.
4123 */
4124 mutex_lock(&slab_mutex);
4125 list_for_each_entry(s, &slab_caches, list) {
4126 /*
4127 * XXX: kmem_cache_alloc_node will fallback to other nodes
4128 * since memory is not yet available from the node that
4129 * is brought up.
4130 */
4131 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
4132 if (!n) {
4133 ret = -ENOMEM;
4134 goto out;
4135 }
4136 init_kmem_cache_node(n);
4137 s->node[nid] = n;
4138 }
4139out:
4140 mutex_unlock(&slab_mutex);
4141 return ret;
4142}
4143
4144static int slab_memory_callback(struct notifier_block *self,
4145 unsigned long action, void *arg)
4146{
4147 int ret = 0;
4148
4149 switch (action) {
4150 case MEM_GOING_ONLINE:
4151 ret = slab_mem_going_online_callback(arg);
4152 break;
4153 case MEM_GOING_OFFLINE:
4154 ret = slab_mem_going_offline_callback(arg);
4155 break;
4156 case MEM_OFFLINE:
4157 case MEM_CANCEL_ONLINE:
4158 slab_mem_offline_callback(arg);
4159 break;
4160 case MEM_ONLINE:
4161 case MEM_CANCEL_OFFLINE:
4162 break;
4163 }
4164 if (ret)
4165 ret = notifier_from_errno(ret);
4166 else
4167 ret = NOTIFY_OK;
4168 return ret;
4169}
4170
4171static struct notifier_block slab_memory_callback_nb = {
4172 .notifier_call = slab_memory_callback,
4173 .priority = SLAB_CALLBACK_PRI,
4174};
4175
4176/********************************************************************
4177 * Basic setup of slabs
4178 *******************************************************************/
4179
4180/*
4181 * Used for early kmem_cache structures that were allocated using
4182 * the page allocator. Allocate them properly then fix up the pointers
4183 * that may be pointing to the wrong kmem_cache structure.
4184 */
4185
4186static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
4187{
4188 int node;
4189 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
4190 struct kmem_cache_node *n;
4191
4192 memcpy(s, static_cache, kmem_cache->object_size);
4193
4194 /*
4195 * This runs very early, and only the boot processor is supposed to be
4196 * up. Even if it weren't true, IRQs are not up so we couldn't fire
4197 * IPIs around.
4198 */
4199 __flush_cpu_slab(s, smp_processor_id());
4200 for_each_kmem_cache_node(s, node, n) {
4201 struct page *p;
4202
4203 list_for_each_entry(p, &n->partial, lru)
4204 p->slab_cache = s;
4205
4206#ifdef CONFIG_SLUB_DEBUG
4207 list_for_each_entry(p, &n->full, lru)
4208 p->slab_cache = s;
4209#endif
4210 }
4211 slab_init_memcg_params(s);
4212 list_add(&s->list, &slab_caches);
4213 memcg_link_cache(s);
4214 return s;
4215}
4216
4217void __init kmem_cache_init(void)
4218{
4219 static __initdata struct kmem_cache boot_kmem_cache,
4220 boot_kmem_cache_node;
4221
4222 if (debug_guardpage_minorder())
4223 slub_max_order = 0;
4224
4225 kmem_cache_node = &boot_kmem_cache_node;
4226 kmem_cache = &boot_kmem_cache;
4227
4228 create_boot_cache(kmem_cache_node, "kmem_cache_node",
4229 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN, 0, 0);
4230
4231 register_hotmemory_notifier(&slab_memory_callback_nb);
4232
4233 /* Able to allocate the per node structures */
4234 slab_state = PARTIAL;
4235
4236 create_boot_cache(kmem_cache, "kmem_cache",
4237 offsetof(struct kmem_cache, node) +
4238 nr_node_ids * sizeof(struct kmem_cache_node *),
4239 SLAB_HWCACHE_ALIGN, 0, 0);
4240
4241 kmem_cache = bootstrap(&boot_kmem_cache);
4242
4243 /*
4244 * Allocate kmem_cache_node properly from the kmem_cache slab.
4245 * kmem_cache_node is separately allocated so no need to
4246 * update any list pointers.
4247 */
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=%d\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, lru) {
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, lru) {
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 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4459 sizeof(unsigned long), GFP_KERNEL);
4460 struct kmem_cache_node *n;
4461
4462 if (!map)
4463 return -ENOMEM;
4464
4465 flush_all(s);
4466 for_each_kmem_cache_node(s, node, n)
4467 count += validate_slab_node(s, n, map);
4468 kfree(map);
4469 return count;
4470}
4471/*
4472 * Generate lists of code addresses where slabcache objects are allocated
4473 * and freed.
4474 */
4475
4476struct location {
4477 unsigned long count;
4478 unsigned long addr;
4479 long long sum_time;
4480 long min_time;
4481 long max_time;
4482 long min_pid;
4483 long max_pid;
4484 DECLARE_BITMAP(cpus, NR_CPUS);
4485 nodemask_t nodes;
4486};
4487
4488struct loc_track {
4489 unsigned long max;
4490 unsigned long count;
4491 struct location *loc;
4492};
4493
4494static void free_loc_track(struct loc_track *t)
4495{
4496 if (t->max)
4497 free_pages((unsigned long)t->loc,
4498 get_order(sizeof(struct location) * t->max));
4499}
4500
4501static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4502{
4503 struct location *l;
4504 int order;
4505
4506 order = get_order(sizeof(struct location) * max);
4507
4508 l = (void *)__get_free_pages(flags, order);
4509 if (!l)
4510 return 0;
4511
4512 if (t->count) {
4513 memcpy(l, t->loc, sizeof(struct location) * t->count);
4514 free_loc_track(t);
4515 }
4516 t->max = max;
4517 t->loc = l;
4518 return 1;
4519}
4520
4521static int add_location(struct loc_track *t, struct kmem_cache *s,
4522 const struct track *track)
4523{
4524 long start, end, pos;
4525 struct location *l;
4526 unsigned long caddr;
4527 unsigned long age = jiffies - track->when;
4528
4529 start = -1;
4530 end = t->count;
4531
4532 for ( ; ; ) {
4533 pos = start + (end - start + 1) / 2;
4534
4535 /*
4536 * There is nothing at "end". If we end up there
4537 * we need to add something to before end.
4538 */
4539 if (pos == end)
4540 break;
4541
4542 caddr = t->loc[pos].addr;
4543 if (track->addr == caddr) {
4544
4545 l = &t->loc[pos];
4546 l->count++;
4547 if (track->when) {
4548 l->sum_time += age;
4549 if (age < l->min_time)
4550 l->min_time = age;
4551 if (age > l->max_time)
4552 l->max_time = age;
4553
4554 if (track->pid < l->min_pid)
4555 l->min_pid = track->pid;
4556 if (track->pid > l->max_pid)
4557 l->max_pid = track->pid;
4558
4559 cpumask_set_cpu(track->cpu,
4560 to_cpumask(l->cpus));
4561 }
4562 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4563 return 1;
4564 }
4565
4566 if (track->addr < caddr)
4567 end = pos;
4568 else
4569 start = pos;
4570 }
4571
4572 /*
4573 * Not found. Insert new tracking element.
4574 */
4575 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4576 return 0;
4577
4578 l = t->loc + pos;
4579 if (pos < t->count)
4580 memmove(l + 1, l,
4581 (t->count - pos) * sizeof(struct location));
4582 t->count++;
4583 l->count = 1;
4584 l->addr = track->addr;
4585 l->sum_time = age;
4586 l->min_time = age;
4587 l->max_time = age;
4588 l->min_pid = track->pid;
4589 l->max_pid = track->pid;
4590 cpumask_clear(to_cpumask(l->cpus));
4591 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4592 nodes_clear(l->nodes);
4593 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4594 return 1;
4595}
4596
4597static void process_slab(struct loc_track *t, struct kmem_cache *s,
4598 struct page *page, enum track_item alloc,
4599 unsigned long *map)
4600{
4601 void *addr = page_address(page);
4602 void *p;
4603
4604 bitmap_zero(map, page->objects);
4605 get_map(s, page, map);
4606
4607 for_each_object(p, s, addr, page->objects)
4608 if (!test_bit(slab_index(p, s, addr), map))
4609 add_location(t, s, get_track(s, p, alloc));
4610}
4611
4612static int list_locations(struct kmem_cache *s, char *buf,
4613 enum track_item alloc)
4614{
4615 int len = 0;
4616 unsigned long i;
4617 struct loc_track t = { 0, 0, NULL };
4618 int node;
4619 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4620 sizeof(unsigned long), GFP_KERNEL);
4621 struct kmem_cache_node *n;
4622
4623 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4624 GFP_KERNEL)) {
4625 kfree(map);
4626 return sprintf(buf, "Out of memory\n");
4627 }
4628 /* Push back cpu slabs */
4629 flush_all(s);
4630
4631 for_each_kmem_cache_node(s, node, n) {
4632 unsigned long flags;
4633 struct page *page;
4634
4635 if (!atomic_long_read(&n->nr_slabs))
4636 continue;
4637
4638 spin_lock_irqsave(&n->list_lock, flags);
4639 list_for_each_entry(page, &n->partial, lru)
4640 process_slab(&t, s, page, alloc, map);
4641 list_for_each_entry(page, &n->full, lru)
4642 process_slab(&t, s, page, alloc, map);
4643 spin_unlock_irqrestore(&n->list_lock, flags);
4644 }
4645
4646 for (i = 0; i < t.count; i++) {
4647 struct location *l = &t.loc[i];
4648
4649 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4650 break;
4651 len += sprintf(buf + len, "%7ld ", l->count);
4652
4653 if (l->addr)
4654 len += sprintf(buf + len, "%pS", (void *)l->addr);
4655 else
4656 len += sprintf(buf + len, "<not-available>");
4657
4658 if (l->sum_time != l->min_time) {
4659 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4660 l->min_time,
4661 (long)div_u64(l->sum_time, l->count),
4662 l->max_time);
4663 } else
4664 len += sprintf(buf + len, " age=%ld",
4665 l->min_time);
4666
4667 if (l->min_pid != l->max_pid)
4668 len += sprintf(buf + len, " pid=%ld-%ld",
4669 l->min_pid, l->max_pid);
4670 else
4671 len += sprintf(buf + len, " pid=%ld",
4672 l->min_pid);
4673
4674 if (num_online_cpus() > 1 &&
4675 !cpumask_empty(to_cpumask(l->cpus)) &&
4676 len < PAGE_SIZE - 60)
4677 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4678 " cpus=%*pbl",
4679 cpumask_pr_args(to_cpumask(l->cpus)));
4680
4681 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4682 len < PAGE_SIZE - 60)
4683 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4684 " nodes=%*pbl",
4685 nodemask_pr_args(&l->nodes));
4686
4687 len += sprintf(buf + len, "\n");
4688 }
4689
4690 free_loc_track(&t);
4691 kfree(map);
4692 if (!t.count)
4693 len += sprintf(buf, "No data\n");
4694 return len;
4695}
4696#endif
4697
4698#ifdef SLUB_RESILIENCY_TEST
4699static void __init resiliency_test(void)
4700{
4701 u8 *p;
4702
4703 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4704
4705 pr_err("SLUB resiliency testing\n");
4706 pr_err("-----------------------\n");
4707 pr_err("A. Corruption after allocation\n");
4708
4709 p = kzalloc(16, GFP_KERNEL);
4710 p[16] = 0x12;
4711 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4712 p + 16);
4713
4714 validate_slab_cache(kmalloc_caches[4]);
4715
4716 /* Hmmm... The next two are dangerous */
4717 p = kzalloc(32, GFP_KERNEL);
4718 p[32 + sizeof(void *)] = 0x34;
4719 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4720 p);
4721 pr_err("If allocated object is overwritten then not detectable\n\n");
4722
4723 validate_slab_cache(kmalloc_caches[5]);
4724 p = kzalloc(64, GFP_KERNEL);
4725 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4726 *p = 0x56;
4727 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4728 p);
4729 pr_err("If allocated object is overwritten then not detectable\n\n");
4730 validate_slab_cache(kmalloc_caches[6]);
4731
4732 pr_err("\nB. Corruption after free\n");
4733 p = kzalloc(128, GFP_KERNEL);
4734 kfree(p);
4735 *p = 0x78;
4736 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4737 validate_slab_cache(kmalloc_caches[7]);
4738
4739 p = kzalloc(256, GFP_KERNEL);
4740 kfree(p);
4741 p[50] = 0x9a;
4742 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4743 validate_slab_cache(kmalloc_caches[8]);
4744
4745 p = kzalloc(512, GFP_KERNEL);
4746 kfree(p);
4747 p[512] = 0xab;
4748 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4749 validate_slab_cache(kmalloc_caches[9]);
4750}
4751#else
4752#ifdef CONFIG_SYSFS
4753static void resiliency_test(void) {};
4754#endif
4755#endif
4756
4757#ifdef CONFIG_SYSFS
4758enum slab_stat_type {
4759 SL_ALL, /* All slabs */
4760 SL_PARTIAL, /* Only partially allocated slabs */
4761 SL_CPU, /* Only slabs used for cpu caches */
4762 SL_OBJECTS, /* Determine allocated objects not slabs */
4763 SL_TOTAL /* Determine object capacity not slabs */
4764};
4765
4766#define SO_ALL (1 << SL_ALL)
4767#define SO_PARTIAL (1 << SL_PARTIAL)
4768#define SO_CPU (1 << SL_CPU)
4769#define SO_OBJECTS (1 << SL_OBJECTS)
4770#define SO_TOTAL (1 << SL_TOTAL)
4771
4772#ifdef CONFIG_MEMCG
4773static bool memcg_sysfs_enabled = IS_ENABLED(CONFIG_SLUB_MEMCG_SYSFS_ON);
4774
4775static int __init setup_slub_memcg_sysfs(char *str)
4776{
4777 int v;
4778
4779 if (get_option(&str, &v) > 0)
4780 memcg_sysfs_enabled = v;
4781
4782 return 1;
4783}
4784
4785__setup("slub_memcg_sysfs=", setup_slub_memcg_sysfs);
4786#endif
4787
4788static ssize_t show_slab_objects(struct kmem_cache *s,
4789 char *buf, unsigned long flags)
4790{
4791 unsigned long total = 0;
4792 int node;
4793 int x;
4794 unsigned long *nodes;
4795
4796 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4797 if (!nodes)
4798 return -ENOMEM;
4799
4800 if (flags & SO_CPU) {
4801 int cpu;
4802
4803 for_each_possible_cpu(cpu) {
4804 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4805 cpu);
4806 int node;
4807 struct page *page;
4808
4809 page = READ_ONCE(c->page);
4810 if (!page)
4811 continue;
4812
4813 node = page_to_nid(page);
4814 if (flags & SO_TOTAL)
4815 x = page->objects;
4816 else if (flags & SO_OBJECTS)
4817 x = page->inuse;
4818 else
4819 x = 1;
4820
4821 total += x;
4822 nodes[node] += x;
4823
4824 page = slub_percpu_partial_read_once(c);
4825 if (page) {
4826 node = page_to_nid(page);
4827 if (flags & SO_TOTAL)
4828 WARN_ON_ONCE(1);
4829 else if (flags & SO_OBJECTS)
4830 WARN_ON_ONCE(1);
4831 else
4832 x = page->pages;
4833 total += x;
4834 nodes[node] += x;
4835 }
4836 }
4837 }
4838
4839 get_online_mems();
4840#ifdef CONFIG_SLUB_DEBUG
4841 if (flags & SO_ALL) {
4842 struct kmem_cache_node *n;
4843
4844 for_each_kmem_cache_node(s, node, n) {
4845
4846 if (flags & SO_TOTAL)
4847 x = atomic_long_read(&n->total_objects);
4848 else if (flags & SO_OBJECTS)
4849 x = atomic_long_read(&n->total_objects) -
4850 count_partial(n, count_free);
4851 else
4852 x = atomic_long_read(&n->nr_slabs);
4853 total += x;
4854 nodes[node] += x;
4855 }
4856
4857 } else
4858#endif
4859 if (flags & SO_PARTIAL) {
4860 struct kmem_cache_node *n;
4861
4862 for_each_kmem_cache_node(s, node, n) {
4863 if (flags & SO_TOTAL)
4864 x = count_partial(n, count_total);
4865 else if (flags & SO_OBJECTS)
4866 x = count_partial(n, count_inuse);
4867 else
4868 x = n->nr_partial;
4869 total += x;
4870 nodes[node] += x;
4871 }
4872 }
4873 x = sprintf(buf, "%lu", total);
4874#ifdef CONFIG_NUMA
4875 for (node = 0; node < nr_node_ids; node++)
4876 if (nodes[node])
4877 x += sprintf(buf + x, " N%d=%lu",
4878 node, nodes[node]);
4879#endif
4880 put_online_mems();
4881 kfree(nodes);
4882 return x + sprintf(buf + x, "\n");
4883}
4884
4885#ifdef CONFIG_SLUB_DEBUG
4886static int any_slab_objects(struct kmem_cache *s)
4887{
4888 int node;
4889 struct kmem_cache_node *n;
4890
4891 for_each_kmem_cache_node(s, node, n)
4892 if (atomic_long_read(&n->total_objects))
4893 return 1;
4894
4895 return 0;
4896}
4897#endif
4898
4899#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4900#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4901
4902struct slab_attribute {
4903 struct attribute attr;
4904 ssize_t (*show)(struct kmem_cache *s, char *buf);
4905 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4906};
4907
4908#define SLAB_ATTR_RO(_name) \
4909 static struct slab_attribute _name##_attr = \
4910 __ATTR(_name, 0400, _name##_show, NULL)
4911
4912#define SLAB_ATTR(_name) \
4913 static struct slab_attribute _name##_attr = \
4914 __ATTR(_name, 0600, _name##_show, _name##_store)
4915
4916static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4917{
4918 return sprintf(buf, "%u\n", s->size);
4919}
4920SLAB_ATTR_RO(slab_size);
4921
4922static ssize_t align_show(struct kmem_cache *s, char *buf)
4923{
4924 return sprintf(buf, "%u\n", s->align);
4925}
4926SLAB_ATTR_RO(align);
4927
4928static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4929{
4930 return sprintf(buf, "%u\n", s->object_size);
4931}
4932SLAB_ATTR_RO(object_size);
4933
4934static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4935{
4936 return sprintf(buf, "%u\n", oo_objects(s->oo));
4937}
4938SLAB_ATTR_RO(objs_per_slab);
4939
4940static ssize_t order_store(struct kmem_cache *s,
4941 const char *buf, size_t length)
4942{
4943 unsigned int order;
4944 int err;
4945
4946 err = kstrtouint(buf, 10, &order);
4947 if (err)
4948 return err;
4949
4950 if (order > slub_max_order || order < slub_min_order)
4951 return -EINVAL;
4952
4953 calculate_sizes(s, order);
4954 return length;
4955}
4956
4957static ssize_t order_show(struct kmem_cache *s, char *buf)
4958{
4959 return sprintf(buf, "%u\n", oo_order(s->oo));
4960}
4961SLAB_ATTR(order);
4962
4963static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4964{
4965 return sprintf(buf, "%lu\n", s->min_partial);
4966}
4967
4968static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4969 size_t length)
4970{
4971 unsigned long min;
4972 int err;
4973
4974 err = kstrtoul(buf, 10, &min);
4975 if (err)
4976 return err;
4977
4978 set_min_partial(s, min);
4979 return length;
4980}
4981SLAB_ATTR(min_partial);
4982
4983static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4984{
4985 return sprintf(buf, "%u\n", slub_cpu_partial(s));
4986}
4987
4988static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4989 size_t length)
4990{
4991 unsigned int objects;
4992 int err;
4993
4994 err = kstrtouint(buf, 10, &objects);
4995 if (err)
4996 return err;
4997 if (objects && !kmem_cache_has_cpu_partial(s))
4998 return -EINVAL;
4999
5000 slub_set_cpu_partial(s, objects);
5001 flush_all(s);
5002 return length;
5003}
5004SLAB_ATTR(cpu_partial);
5005
5006static ssize_t ctor_show(struct kmem_cache *s, char *buf)
5007{
5008 if (!s->ctor)
5009 return 0;
5010 return sprintf(buf, "%pS\n", s->ctor);
5011}
5012SLAB_ATTR_RO(ctor);
5013
5014static ssize_t aliases_show(struct kmem_cache *s, char *buf)
5015{
5016 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
5017}
5018SLAB_ATTR_RO(aliases);
5019
5020static ssize_t partial_show(struct kmem_cache *s, char *buf)
5021{
5022 return show_slab_objects(s, buf, SO_PARTIAL);
5023}
5024SLAB_ATTR_RO(partial);
5025
5026static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
5027{
5028 return show_slab_objects(s, buf, SO_CPU);
5029}
5030SLAB_ATTR_RO(cpu_slabs);
5031
5032static ssize_t objects_show(struct kmem_cache *s, char *buf)
5033{
5034 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
5035}
5036SLAB_ATTR_RO(objects);
5037
5038static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
5039{
5040 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
5041}
5042SLAB_ATTR_RO(objects_partial);
5043
5044static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
5045{
5046 int objects = 0;
5047 int pages = 0;
5048 int cpu;
5049 int len;
5050
5051 for_each_online_cpu(cpu) {
5052 struct page *page;
5053
5054 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5055
5056 if (page) {
5057 pages += page->pages;
5058 objects += page->pobjects;
5059 }
5060 }
5061
5062 len = sprintf(buf, "%d(%d)", objects, pages);
5063
5064#ifdef CONFIG_SMP
5065 for_each_online_cpu(cpu) {
5066 struct page *page;
5067
5068 page = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
5069
5070 if (page && len < PAGE_SIZE - 20)
5071 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
5072 page->pobjects, page->pages);
5073 }
5074#endif
5075 return len + sprintf(buf + len, "\n");
5076}
5077SLAB_ATTR_RO(slabs_cpu_partial);
5078
5079static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
5080{
5081 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
5082}
5083
5084static ssize_t reclaim_account_store(struct kmem_cache *s,
5085 const char *buf, size_t length)
5086{
5087 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
5088 if (buf[0] == '1')
5089 s->flags |= SLAB_RECLAIM_ACCOUNT;
5090 return length;
5091}
5092SLAB_ATTR(reclaim_account);
5093
5094static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
5095{
5096 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
5097}
5098SLAB_ATTR_RO(hwcache_align);
5099
5100#ifdef CONFIG_ZONE_DMA
5101static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
5102{
5103 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
5104}
5105SLAB_ATTR_RO(cache_dma);
5106#endif
5107
5108static ssize_t usersize_show(struct kmem_cache *s, char *buf)
5109{
5110 return sprintf(buf, "%u\n", s->usersize);
5111}
5112SLAB_ATTR_RO(usersize);
5113
5114static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
5115{
5116 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
5117}
5118SLAB_ATTR_RO(destroy_by_rcu);
5119
5120static ssize_t reserved_show(struct kmem_cache *s, char *buf)
5121{
5122 return sprintf(buf, "%u\n", s->reserved);
5123}
5124SLAB_ATTR_RO(reserved);
5125
5126#ifdef CONFIG_SLUB_DEBUG
5127static ssize_t slabs_show(struct kmem_cache *s, char *buf)
5128{
5129 return show_slab_objects(s, buf, SO_ALL);
5130}
5131SLAB_ATTR_RO(slabs);
5132
5133static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
5134{
5135 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
5136}
5137SLAB_ATTR_RO(total_objects);
5138
5139static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
5140{
5141 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
5142}
5143
5144static ssize_t sanity_checks_store(struct kmem_cache *s,
5145 const char *buf, size_t length)
5146{
5147 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
5148 if (buf[0] == '1') {
5149 s->flags &= ~__CMPXCHG_DOUBLE;
5150 s->flags |= SLAB_CONSISTENCY_CHECKS;
5151 }
5152 return length;
5153}
5154SLAB_ATTR(sanity_checks);
5155
5156static ssize_t trace_show(struct kmem_cache *s, char *buf)
5157{
5158 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
5159}
5160
5161static ssize_t trace_store(struct kmem_cache *s, const char *buf,
5162 size_t length)
5163{
5164 /*
5165 * Tracing a merged cache is going to give confusing results
5166 * as well as cause other issues like converting a mergeable
5167 * cache into an umergeable one.
5168 */
5169 if (s->refcount > 1)
5170 return -EINVAL;
5171
5172 s->flags &= ~SLAB_TRACE;
5173 if (buf[0] == '1') {
5174 s->flags &= ~__CMPXCHG_DOUBLE;
5175 s->flags |= SLAB_TRACE;
5176 }
5177 return length;
5178}
5179SLAB_ATTR(trace);
5180
5181static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
5182{
5183 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
5184}
5185
5186static ssize_t red_zone_store(struct kmem_cache *s,
5187 const char *buf, size_t length)
5188{
5189 if (any_slab_objects(s))
5190 return -EBUSY;
5191
5192 s->flags &= ~SLAB_RED_ZONE;
5193 if (buf[0] == '1') {
5194 s->flags |= SLAB_RED_ZONE;
5195 }
5196 calculate_sizes(s, -1);
5197 return length;
5198}
5199SLAB_ATTR(red_zone);
5200
5201static ssize_t poison_show(struct kmem_cache *s, char *buf)
5202{
5203 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
5204}
5205
5206static ssize_t poison_store(struct kmem_cache *s,
5207 const char *buf, size_t length)
5208{
5209 if (any_slab_objects(s))
5210 return -EBUSY;
5211
5212 s->flags &= ~SLAB_POISON;
5213 if (buf[0] == '1') {
5214 s->flags |= SLAB_POISON;
5215 }
5216 calculate_sizes(s, -1);
5217 return length;
5218}
5219SLAB_ATTR(poison);
5220
5221static ssize_t store_user_show(struct kmem_cache *s, char *buf)
5222{
5223 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
5224}
5225
5226static ssize_t store_user_store(struct kmem_cache *s,
5227 const char *buf, size_t length)
5228{
5229 if (any_slab_objects(s))
5230 return -EBUSY;
5231
5232 s->flags &= ~SLAB_STORE_USER;
5233 if (buf[0] == '1') {
5234 s->flags &= ~__CMPXCHG_DOUBLE;
5235 s->flags |= SLAB_STORE_USER;
5236 }
5237 calculate_sizes(s, -1);
5238 return length;
5239}
5240SLAB_ATTR(store_user);
5241
5242static ssize_t validate_show(struct kmem_cache *s, char *buf)
5243{
5244 return 0;
5245}
5246
5247static ssize_t validate_store(struct kmem_cache *s,
5248 const char *buf, size_t length)
5249{
5250 int ret = -EINVAL;
5251
5252 if (buf[0] == '1') {
5253 ret = validate_slab_cache(s);
5254 if (ret >= 0)
5255 ret = length;
5256 }
5257 return ret;
5258}
5259SLAB_ATTR(validate);
5260
5261static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
5262{
5263 if (!(s->flags & SLAB_STORE_USER))
5264 return -ENOSYS;
5265 return list_locations(s, buf, TRACK_ALLOC);
5266}
5267SLAB_ATTR_RO(alloc_calls);
5268
5269static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
5270{
5271 if (!(s->flags & SLAB_STORE_USER))
5272 return -ENOSYS;
5273 return list_locations(s, buf, TRACK_FREE);
5274}
5275SLAB_ATTR_RO(free_calls);
5276#endif /* CONFIG_SLUB_DEBUG */
5277
5278#ifdef CONFIG_FAILSLAB
5279static ssize_t failslab_show(struct kmem_cache *s, char *buf)
5280{
5281 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
5282}
5283
5284static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
5285 size_t length)
5286{
5287 if (s->refcount > 1)
5288 return -EINVAL;
5289
5290 s->flags &= ~SLAB_FAILSLAB;
5291 if (buf[0] == '1')
5292 s->flags |= SLAB_FAILSLAB;
5293 return length;
5294}
5295SLAB_ATTR(failslab);
5296#endif
5297
5298static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5299{
5300 return 0;
5301}
5302
5303static ssize_t shrink_store(struct kmem_cache *s,
5304 const char *buf, size_t length)
5305{
5306 if (buf[0] == '1')
5307 kmem_cache_shrink(s);
5308 else
5309 return -EINVAL;
5310 return length;
5311}
5312SLAB_ATTR(shrink);
5313
5314#ifdef CONFIG_NUMA
5315static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5316{
5317 return sprintf(buf, "%u\n", s->remote_node_defrag_ratio / 10);
5318}
5319
5320static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5321 const char *buf, size_t length)
5322{
5323 unsigned int ratio;
5324 int err;
5325
5326 err = kstrtouint(buf, 10, &ratio);
5327 if (err)
5328 return err;
5329 if (ratio > 100)
5330 return -ERANGE;
5331
5332 s->remote_node_defrag_ratio = ratio * 10;
5333
5334 return length;
5335}
5336SLAB_ATTR(remote_node_defrag_ratio);
5337#endif
5338
5339#ifdef CONFIG_SLUB_STATS
5340static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5341{
5342 unsigned long sum = 0;
5343 int cpu;
5344 int len;
5345 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5346
5347 if (!data)
5348 return -ENOMEM;
5349
5350 for_each_online_cpu(cpu) {
5351 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5352
5353 data[cpu] = x;
5354 sum += x;
5355 }
5356
5357 len = sprintf(buf, "%lu", sum);
5358
5359#ifdef CONFIG_SMP
5360 for_each_online_cpu(cpu) {
5361 if (data[cpu] && len < PAGE_SIZE - 20)
5362 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5363 }
5364#endif
5365 kfree(data);
5366 return len + sprintf(buf + len, "\n");
5367}
5368
5369static void clear_stat(struct kmem_cache *s, enum stat_item si)
5370{
5371 int cpu;
5372
5373 for_each_online_cpu(cpu)
5374 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5375}
5376
5377#define STAT_ATTR(si, text) \
5378static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5379{ \
5380 return show_stat(s, buf, si); \
5381} \
5382static ssize_t text##_store(struct kmem_cache *s, \
5383 const char *buf, size_t length) \
5384{ \
5385 if (buf[0] != '0') \
5386 return -EINVAL; \
5387 clear_stat(s, si); \
5388 return length; \
5389} \
5390SLAB_ATTR(text); \
5391
5392STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5393STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5394STAT_ATTR(FREE_FASTPATH, free_fastpath);
5395STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5396STAT_ATTR(FREE_FROZEN, free_frozen);
5397STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5398STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5399STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5400STAT_ATTR(ALLOC_SLAB, alloc_slab);
5401STAT_ATTR(ALLOC_REFILL, alloc_refill);
5402STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5403STAT_ATTR(FREE_SLAB, free_slab);
5404STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5405STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5406STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5407STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5408STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5409STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5410STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5411STAT_ATTR(ORDER_FALLBACK, order_fallback);
5412STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5413STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5414STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5415STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5416STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5417STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5418#endif
5419
5420static struct attribute *slab_attrs[] = {
5421 &slab_size_attr.attr,
5422 &object_size_attr.attr,
5423 &objs_per_slab_attr.attr,
5424 &order_attr.attr,
5425 &min_partial_attr.attr,
5426 &cpu_partial_attr.attr,
5427 &objects_attr.attr,
5428 &objects_partial_attr.attr,
5429 &partial_attr.attr,
5430 &cpu_slabs_attr.attr,
5431 &ctor_attr.attr,
5432 &aliases_attr.attr,
5433 &align_attr.attr,
5434 &hwcache_align_attr.attr,
5435 &reclaim_account_attr.attr,
5436 &destroy_by_rcu_attr.attr,
5437 &shrink_attr.attr,
5438 &reserved_attr.attr,
5439 &slabs_cpu_partial_attr.attr,
5440#ifdef CONFIG_SLUB_DEBUG
5441 &total_objects_attr.attr,
5442 &slabs_attr.attr,
5443 &sanity_checks_attr.attr,
5444 &trace_attr.attr,
5445 &red_zone_attr.attr,
5446 &poison_attr.attr,
5447 &store_user_attr.attr,
5448 &validate_attr.attr,
5449 &alloc_calls_attr.attr,
5450 &free_calls_attr.attr,
5451#endif
5452#ifdef CONFIG_ZONE_DMA
5453 &cache_dma_attr.attr,
5454#endif
5455#ifdef CONFIG_NUMA
5456 &remote_node_defrag_ratio_attr.attr,
5457#endif
5458#ifdef CONFIG_SLUB_STATS
5459 &alloc_fastpath_attr.attr,
5460 &alloc_slowpath_attr.attr,
5461 &free_fastpath_attr.attr,
5462 &free_slowpath_attr.attr,
5463 &free_frozen_attr.attr,
5464 &free_add_partial_attr.attr,
5465 &free_remove_partial_attr.attr,
5466 &alloc_from_partial_attr.attr,
5467 &alloc_slab_attr.attr,
5468 &alloc_refill_attr.attr,
5469 &alloc_node_mismatch_attr.attr,
5470 &free_slab_attr.attr,
5471 &cpuslab_flush_attr.attr,
5472 &deactivate_full_attr.attr,
5473 &deactivate_empty_attr.attr,
5474 &deactivate_to_head_attr.attr,
5475 &deactivate_to_tail_attr.attr,
5476 &deactivate_remote_frees_attr.attr,
5477 &deactivate_bypass_attr.attr,
5478 &order_fallback_attr.attr,
5479 &cmpxchg_double_fail_attr.attr,
5480 &cmpxchg_double_cpu_fail_attr.attr,
5481 &cpu_partial_alloc_attr.attr,
5482 &cpu_partial_free_attr.attr,
5483 &cpu_partial_node_attr.attr,
5484 &cpu_partial_drain_attr.attr,
5485#endif
5486#ifdef CONFIG_FAILSLAB
5487 &failslab_attr.attr,
5488#endif
5489 &usersize_attr.attr,
5490
5491 NULL
5492};
5493
5494static const struct attribute_group slab_attr_group = {
5495 .attrs = slab_attrs,
5496};
5497
5498static ssize_t slab_attr_show(struct kobject *kobj,
5499 struct attribute *attr,
5500 char *buf)
5501{
5502 struct slab_attribute *attribute;
5503 struct kmem_cache *s;
5504 int err;
5505
5506 attribute = to_slab_attr(attr);
5507 s = to_slab(kobj);
5508
5509 if (!attribute->show)
5510 return -EIO;
5511
5512 err = attribute->show(s, buf);
5513
5514 return err;
5515}
5516
5517static ssize_t slab_attr_store(struct kobject *kobj,
5518 struct attribute *attr,
5519 const char *buf, size_t len)
5520{
5521 struct slab_attribute *attribute;
5522 struct kmem_cache *s;
5523 int err;
5524
5525 attribute = to_slab_attr(attr);
5526 s = to_slab(kobj);
5527
5528 if (!attribute->store)
5529 return -EIO;
5530
5531 err = attribute->store(s, buf, len);
5532#ifdef CONFIG_MEMCG
5533 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5534 struct kmem_cache *c;
5535
5536 mutex_lock(&slab_mutex);
5537 if (s->max_attr_size < len)
5538 s->max_attr_size = len;
5539
5540 /*
5541 * This is a best effort propagation, so this function's return
5542 * value will be determined by the parent cache only. This is
5543 * basically because not all attributes will have a well
5544 * defined semantics for rollbacks - most of the actions will
5545 * have permanent effects.
5546 *
5547 * Returning the error value of any of the children that fail
5548 * is not 100 % defined, in the sense that users seeing the
5549 * error code won't be able to know anything about the state of
5550 * the cache.
5551 *
5552 * Only returning the error code for the parent cache at least
5553 * has well defined semantics. The cache being written to
5554 * directly either failed or succeeded, in which case we loop
5555 * through the descendants with best-effort propagation.
5556 */
5557 for_each_memcg_cache(c, s)
5558 attribute->store(c, buf, len);
5559 mutex_unlock(&slab_mutex);
5560 }
5561#endif
5562 return err;
5563}
5564
5565static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5566{
5567#ifdef CONFIG_MEMCG
5568 int i;
5569 char *buffer = NULL;
5570 struct kmem_cache *root_cache;
5571
5572 if (is_root_cache(s))
5573 return;
5574
5575 root_cache = s->memcg_params.root_cache;
5576
5577 /*
5578 * This mean this cache had no attribute written. Therefore, no point
5579 * in copying default values around
5580 */
5581 if (!root_cache->max_attr_size)
5582 return;
5583
5584 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5585 char mbuf[64];
5586 char *buf;
5587 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5588 ssize_t len;
5589
5590 if (!attr || !attr->store || !attr->show)
5591 continue;
5592
5593 /*
5594 * It is really bad that we have to allocate here, so we will
5595 * do it only as a fallback. If we actually allocate, though,
5596 * we can just use the allocated buffer until the end.
5597 *
5598 * Most of the slub attributes will tend to be very small in
5599 * size, but sysfs allows buffers up to a page, so they can
5600 * theoretically happen.
5601 */
5602 if (buffer)
5603 buf = buffer;
5604 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5605 buf = mbuf;
5606 else {
5607 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5608 if (WARN_ON(!buffer))
5609 continue;
5610 buf = buffer;
5611 }
5612
5613 len = attr->show(root_cache, buf);
5614 if (len > 0)
5615 attr->store(s, buf, len);
5616 }
5617
5618 if (buffer)
5619 free_page((unsigned long)buffer);
5620#endif
5621}
5622
5623static void kmem_cache_release(struct kobject *k)
5624{
5625 slab_kmem_cache_release(to_slab(k));
5626}
5627
5628static const struct sysfs_ops slab_sysfs_ops = {
5629 .show = slab_attr_show,
5630 .store = slab_attr_store,
5631};
5632
5633static struct kobj_type slab_ktype = {
5634 .sysfs_ops = &slab_sysfs_ops,
5635 .release = kmem_cache_release,
5636};
5637
5638static int uevent_filter(struct kset *kset, struct kobject *kobj)
5639{
5640 struct kobj_type *ktype = get_ktype(kobj);
5641
5642 if (ktype == &slab_ktype)
5643 return 1;
5644 return 0;
5645}
5646
5647static const struct kset_uevent_ops slab_uevent_ops = {
5648 .filter = uevent_filter,
5649};
5650
5651static struct kset *slab_kset;
5652
5653static inline struct kset *cache_kset(struct kmem_cache *s)
5654{
5655#ifdef CONFIG_MEMCG
5656 if (!is_root_cache(s))
5657 return s->memcg_params.root_cache->memcg_kset;
5658#endif
5659 return slab_kset;
5660}
5661
5662#define ID_STR_LENGTH 64
5663
5664/* Create a unique string id for a slab cache:
5665 *
5666 * Format :[flags-]size
5667 */
5668static char *create_unique_id(struct kmem_cache *s)
5669{
5670 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5671 char *p = name;
5672
5673 BUG_ON(!name);
5674
5675 *p++ = ':';
5676 /*
5677 * First flags affecting slabcache operations. We will only
5678 * get here for aliasable slabs so we do not need to support
5679 * too many flags. The flags here must cover all flags that
5680 * are matched during merging to guarantee that the id is
5681 * unique.
5682 */
5683 if (s->flags & SLAB_CACHE_DMA)
5684 *p++ = 'd';
5685 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5686 *p++ = 'a';
5687 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5688 *p++ = 'F';
5689 if (s->flags & SLAB_ACCOUNT)
5690 *p++ = 'A';
5691 if (p != name + 1)
5692 *p++ = '-';
5693 p += sprintf(p, "%07u", s->size);
5694
5695 BUG_ON(p > name + ID_STR_LENGTH - 1);
5696 return name;
5697}
5698
5699static void sysfs_slab_remove_workfn(struct work_struct *work)
5700{
5701 struct kmem_cache *s =
5702 container_of(work, struct kmem_cache, kobj_remove_work);
5703
5704 if (!s->kobj.state_in_sysfs)
5705 /*
5706 * For a memcg cache, this may be called during
5707 * deactivation and again on shutdown. Remove only once.
5708 * A cache is never shut down before deactivation is
5709 * complete, so no need to worry about synchronization.
5710 */
5711 goto out;
5712
5713#ifdef CONFIG_MEMCG
5714 kset_unregister(s->memcg_kset);
5715#endif
5716 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5717 kobject_del(&s->kobj);
5718out:
5719 kobject_put(&s->kobj);
5720}
5721
5722static int sysfs_slab_add(struct kmem_cache *s)
5723{
5724 int err;
5725 const char *name;
5726 struct kset *kset = cache_kset(s);
5727 int unmergeable = slab_unmergeable(s);
5728
5729 INIT_WORK(&s->kobj_remove_work, sysfs_slab_remove_workfn);
5730
5731 if (!kset) {
5732 kobject_init(&s->kobj, &slab_ktype);
5733 return 0;
5734 }
5735
5736 if (!unmergeable && disable_higher_order_debug &&
5737 (slub_debug & DEBUG_METADATA_FLAGS))
5738 unmergeable = 1;
5739
5740 if (unmergeable) {
5741 /*
5742 * Slabcache can never be merged so we can use the name proper.
5743 * This is typically the case for debug situations. In that
5744 * case we can catch duplicate names easily.
5745 */
5746 sysfs_remove_link(&slab_kset->kobj, s->name);
5747 name = s->name;
5748 } else {
5749 /*
5750 * Create a unique name for the slab as a target
5751 * for the symlinks.
5752 */
5753 name = create_unique_id(s);
5754 }
5755
5756 s->kobj.kset = kset;
5757 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5758 if (err)
5759 goto out;
5760
5761 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5762 if (err)
5763 goto out_del_kobj;
5764
5765#ifdef CONFIG_MEMCG
5766 if (is_root_cache(s) && memcg_sysfs_enabled) {
5767 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5768 if (!s->memcg_kset) {
5769 err = -ENOMEM;
5770 goto out_del_kobj;
5771 }
5772 }
5773#endif
5774
5775 kobject_uevent(&s->kobj, KOBJ_ADD);
5776 if (!unmergeable) {
5777 /* Setup first alias */
5778 sysfs_slab_alias(s, s->name);
5779 }
5780out:
5781 if (!unmergeable)
5782 kfree(name);
5783 return err;
5784out_del_kobj:
5785 kobject_del(&s->kobj);
5786 goto out;
5787}
5788
5789static void sysfs_slab_remove(struct kmem_cache *s)
5790{
5791 if (slab_state < FULL)
5792 /*
5793 * Sysfs has not been setup yet so no need to remove the
5794 * cache from sysfs.
5795 */
5796 return;
5797
5798 kobject_get(&s->kobj);
5799 schedule_work(&s->kobj_remove_work);
5800}
5801
5802void sysfs_slab_release(struct kmem_cache *s)
5803{
5804 if (slab_state >= FULL)
5805 kobject_put(&s->kobj);
5806}
5807
5808/*
5809 * Need to buffer aliases during bootup until sysfs becomes
5810 * available lest we lose that information.
5811 */
5812struct saved_alias {
5813 struct kmem_cache *s;
5814 const char *name;
5815 struct saved_alias *next;
5816};
5817
5818static struct saved_alias *alias_list;
5819
5820static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5821{
5822 struct saved_alias *al;
5823
5824 if (slab_state == FULL) {
5825 /*
5826 * If we have a leftover link then remove it.
5827 */
5828 sysfs_remove_link(&slab_kset->kobj, name);
5829 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5830 }
5831
5832 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5833 if (!al)
5834 return -ENOMEM;
5835
5836 al->s = s;
5837 al->name = name;
5838 al->next = alias_list;
5839 alias_list = al;
5840 return 0;
5841}
5842
5843static int __init slab_sysfs_init(void)
5844{
5845 struct kmem_cache *s;
5846 int err;
5847
5848 mutex_lock(&slab_mutex);
5849
5850 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5851 if (!slab_kset) {
5852 mutex_unlock(&slab_mutex);
5853 pr_err("Cannot register slab subsystem.\n");
5854 return -ENOSYS;
5855 }
5856
5857 slab_state = FULL;
5858
5859 list_for_each_entry(s, &slab_caches, list) {
5860 err = sysfs_slab_add(s);
5861 if (err)
5862 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5863 s->name);
5864 }
5865
5866 while (alias_list) {
5867 struct saved_alias *al = alias_list;
5868
5869 alias_list = alias_list->next;
5870 err = sysfs_slab_alias(al->s, al->name);
5871 if (err)
5872 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5873 al->name);
5874 kfree(al);
5875 }
5876
5877 mutex_unlock(&slab_mutex);
5878 resiliency_test();
5879 return 0;
5880}
5881
5882__initcall(slab_sysfs_init);
5883#endif /* CONFIG_SYSFS */
5884
5885/*
5886 * The /proc/slabinfo ABI
5887 */
5888#ifdef CONFIG_SLUB_DEBUG
5889void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5890{
5891 unsigned long nr_slabs = 0;
5892 unsigned long nr_objs = 0;
5893 unsigned long nr_free = 0;
5894 int node;
5895 struct kmem_cache_node *n;
5896
5897 for_each_kmem_cache_node(s, node, n) {
5898 nr_slabs += node_nr_slabs(n);
5899 nr_objs += node_nr_objs(n);
5900 nr_free += count_partial(n, count_free);
5901 }
5902
5903 sinfo->active_objs = nr_objs - nr_free;
5904 sinfo->num_objs = nr_objs;
5905 sinfo->active_slabs = nr_slabs;
5906 sinfo->num_slabs = nr_slabs;
5907 sinfo->objects_per_slab = oo_objects(s->oo);
5908 sinfo->cache_order = oo_order(s->oo);
5909}
5910
5911void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5912{
5913}
5914
5915ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5916 size_t count, loff_t *ppos)
5917{
5918 return -EIO;
5919}
5920#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#ifdef CONFIG_NUMA
222static DEFINE_STATIC_KEY_FALSE(strict_numa);
223#endif
224
225/* Structure holding parameters for get_partial() call chain */
226struct partial_context {
227 gfp_t flags;
228 unsigned int orig_size;
229 void *object;
230};
231
232static inline bool kmem_cache_debug(struct kmem_cache *s)
233{
234 return kmem_cache_debug_flags(s, SLAB_DEBUG_FLAGS);
235}
236
237void *fixup_red_left(struct kmem_cache *s, void *p)
238{
239 if (kmem_cache_debug_flags(s, SLAB_RED_ZONE))
240 p += s->red_left_pad;
241
242 return p;
243}
244
245static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
246{
247#ifdef CONFIG_SLUB_CPU_PARTIAL
248 return !kmem_cache_debug(s);
249#else
250 return false;
251#endif
252}
253
254/*
255 * Issues still to be resolved:
256 *
257 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
258 *
259 * - Variable sizing of the per node arrays
260 */
261
262/* Enable to log cmpxchg failures */
263#undef SLUB_DEBUG_CMPXCHG
264
265#ifndef CONFIG_SLUB_TINY
266/*
267 * Minimum number of partial slabs. These will be left on the partial
268 * lists even if they are empty. kmem_cache_shrink may reclaim them.
269 */
270#define MIN_PARTIAL 5
271
272/*
273 * Maximum number of desirable partial slabs.
274 * The existence of more partial slabs makes kmem_cache_shrink
275 * sort the partial list by the number of objects in use.
276 */
277#define MAX_PARTIAL 10
278#else
279#define MIN_PARTIAL 0
280#define MAX_PARTIAL 0
281#endif
282
283#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
284 SLAB_POISON | SLAB_STORE_USER)
285
286/*
287 * These debug flags cannot use CMPXCHG because there might be consistency
288 * issues when checking or reading debug information
289 */
290#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
291 SLAB_TRACE)
292
293
294/*
295 * Debugging flags that require metadata to be stored in the slab. These get
296 * disabled when slab_debug=O is used and a cache's min order increases with
297 * metadata.
298 */
299#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
300
301#define OO_SHIFT 16
302#define OO_MASK ((1 << OO_SHIFT) - 1)
303#define MAX_OBJS_PER_PAGE 32767 /* since slab.objects is u15 */
304
305/* Internal SLUB flags */
306/* Poison object */
307#define __OBJECT_POISON __SLAB_FLAG_BIT(_SLAB_OBJECT_POISON)
308/* Use cmpxchg_double */
309
310#ifdef system_has_freelist_aba
311#define __CMPXCHG_DOUBLE __SLAB_FLAG_BIT(_SLAB_CMPXCHG_DOUBLE)
312#else
313#define __CMPXCHG_DOUBLE __SLAB_FLAG_UNUSED
314#endif
315
316/*
317 * Tracking user of a slab.
318 */
319#define TRACK_ADDRS_COUNT 16
320struct track {
321 unsigned long addr; /* Called from address */
322#ifdef CONFIG_STACKDEPOT
323 depot_stack_handle_t handle;
324#endif
325 int cpu; /* Was running on cpu */
326 int pid; /* Pid context */
327 unsigned long when; /* When did the operation occur */
328};
329
330enum track_item { TRACK_ALLOC, TRACK_FREE };
331
332#ifdef SLAB_SUPPORTS_SYSFS
333static int sysfs_slab_add(struct kmem_cache *);
334static int sysfs_slab_alias(struct kmem_cache *, const char *);
335#else
336static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
337static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
338 { return 0; }
339#endif
340
341#if defined(CONFIG_DEBUG_FS) && defined(CONFIG_SLUB_DEBUG)
342static void debugfs_slab_add(struct kmem_cache *);
343#else
344static inline void debugfs_slab_add(struct kmem_cache *s) { }
345#endif
346
347enum stat_item {
348 ALLOC_FASTPATH, /* Allocation from cpu slab */
349 ALLOC_SLOWPATH, /* Allocation by getting a new cpu slab */
350 FREE_FASTPATH, /* Free to cpu slab */
351 FREE_SLOWPATH, /* Freeing not to cpu slab */
352 FREE_FROZEN, /* Freeing to frozen slab */
353 FREE_ADD_PARTIAL, /* Freeing moves slab to partial list */
354 FREE_REMOVE_PARTIAL, /* Freeing removes last object */
355 ALLOC_FROM_PARTIAL, /* Cpu slab acquired from node partial list */
356 ALLOC_SLAB, /* Cpu slab acquired from page allocator */
357 ALLOC_REFILL, /* Refill cpu slab from slab freelist */
358 ALLOC_NODE_MISMATCH, /* Switching cpu slab */
359 FREE_SLAB, /* Slab freed to the page allocator */
360 CPUSLAB_FLUSH, /* Abandoning of the cpu slab */
361 DEACTIVATE_FULL, /* Cpu slab was full when deactivated */
362 DEACTIVATE_EMPTY, /* Cpu slab was empty when deactivated */
363 DEACTIVATE_TO_HEAD, /* Cpu slab was moved to the head of partials */
364 DEACTIVATE_TO_TAIL, /* Cpu slab was moved to the tail of partials */
365 DEACTIVATE_REMOTE_FREES,/* Slab contained remotely freed objects */
366 DEACTIVATE_BYPASS, /* Implicit deactivation */
367 ORDER_FALLBACK, /* Number of times fallback was necessary */
368 CMPXCHG_DOUBLE_CPU_FAIL,/* Failures of this_cpu_cmpxchg_double */
369 CMPXCHG_DOUBLE_FAIL, /* Failures of slab freelist update */
370 CPU_PARTIAL_ALLOC, /* Used cpu partial on alloc */
371 CPU_PARTIAL_FREE, /* Refill cpu partial on free */
372 CPU_PARTIAL_NODE, /* Refill cpu partial from node partial */
373 CPU_PARTIAL_DRAIN, /* Drain cpu partial to node partial */
374 NR_SLUB_STAT_ITEMS
375};
376
377#ifndef CONFIG_SLUB_TINY
378/*
379 * When changing the layout, make sure freelist and tid are still compatible
380 * with this_cpu_cmpxchg_double() alignment requirements.
381 */
382struct kmem_cache_cpu {
383 union {
384 struct {
385 void **freelist; /* Pointer to next available object */
386 unsigned long tid; /* Globally unique transaction id */
387 };
388 freelist_aba_t freelist_tid;
389 };
390 struct slab *slab; /* The slab from which we are allocating */
391#ifdef CONFIG_SLUB_CPU_PARTIAL
392 struct slab *partial; /* Partially allocated slabs */
393#endif
394 local_lock_t lock; /* Protects the fields above */
395#ifdef CONFIG_SLUB_STATS
396 unsigned int stat[NR_SLUB_STAT_ITEMS];
397#endif
398};
399#endif /* CONFIG_SLUB_TINY */
400
401static inline void stat(const struct kmem_cache *s, enum stat_item si)
402{
403#ifdef CONFIG_SLUB_STATS
404 /*
405 * The rmw is racy on a preemptible kernel but this is acceptable, so
406 * avoid this_cpu_add()'s irq-disable overhead.
407 */
408 raw_cpu_inc(s->cpu_slab->stat[si]);
409#endif
410}
411
412static inline
413void stat_add(const struct kmem_cache *s, enum stat_item si, int v)
414{
415#ifdef CONFIG_SLUB_STATS
416 raw_cpu_add(s->cpu_slab->stat[si], v);
417#endif
418}
419
420/*
421 * The slab lists for all objects.
422 */
423struct kmem_cache_node {
424 spinlock_t list_lock;
425 unsigned long nr_partial;
426 struct list_head partial;
427#ifdef CONFIG_SLUB_DEBUG
428 atomic_long_t nr_slabs;
429 atomic_long_t total_objects;
430 struct list_head full;
431#endif
432};
433
434static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
435{
436 return s->node[node];
437}
438
439/*
440 * Iterator over all nodes. The body will be executed for each node that has
441 * a kmem_cache_node structure allocated (which is true for all online nodes)
442 */
443#define for_each_kmem_cache_node(__s, __node, __n) \
444 for (__node = 0; __node < nr_node_ids; __node++) \
445 if ((__n = get_node(__s, __node)))
446
447/*
448 * Tracks for which NUMA nodes we have kmem_cache_nodes allocated.
449 * Corresponds to node_state[N_NORMAL_MEMORY], but can temporarily
450 * differ during memory hotplug/hotremove operations.
451 * Protected by slab_mutex.
452 */
453static nodemask_t slab_nodes;
454
455#ifndef CONFIG_SLUB_TINY
456/*
457 * Workqueue used for flush_cpu_slab().
458 */
459static struct workqueue_struct *flushwq;
460#endif
461
462/********************************************************************
463 * Core slab cache functions
464 *******************************************************************/
465
466/*
467 * Returns freelist pointer (ptr). With hardening, this is obfuscated
468 * with an XOR of the address where the pointer is held and a per-cache
469 * random number.
470 */
471static inline freeptr_t freelist_ptr_encode(const struct kmem_cache *s,
472 void *ptr, unsigned long ptr_addr)
473{
474 unsigned long encoded;
475
476#ifdef CONFIG_SLAB_FREELIST_HARDENED
477 encoded = (unsigned long)ptr ^ s->random ^ swab(ptr_addr);
478#else
479 encoded = (unsigned long)ptr;
480#endif
481 return (freeptr_t){.v = encoded};
482}
483
484static inline void *freelist_ptr_decode(const struct kmem_cache *s,
485 freeptr_t ptr, unsigned long ptr_addr)
486{
487 void *decoded;
488
489#ifdef CONFIG_SLAB_FREELIST_HARDENED
490 decoded = (void *)(ptr.v ^ s->random ^ swab(ptr_addr));
491#else
492 decoded = (void *)ptr.v;
493#endif
494 return decoded;
495}
496
497static inline void *get_freepointer(struct kmem_cache *s, void *object)
498{
499 unsigned long ptr_addr;
500 freeptr_t p;
501
502 object = kasan_reset_tag(object);
503 ptr_addr = (unsigned long)object + s->offset;
504 p = *(freeptr_t *)(ptr_addr);
505 return freelist_ptr_decode(s, p, ptr_addr);
506}
507
508#ifndef CONFIG_SLUB_TINY
509static void prefetch_freepointer(const struct kmem_cache *s, void *object)
510{
511 prefetchw(object + s->offset);
512}
513#endif
514
515/*
516 * When running under KMSAN, get_freepointer_safe() may return an uninitialized
517 * pointer value in the case the current thread loses the race for the next
518 * memory chunk in the freelist. In that case this_cpu_cmpxchg_double() in
519 * slab_alloc_node() will fail, so the uninitialized value won't be used, but
520 * KMSAN will still check all arguments of cmpxchg because of imperfect
521 * handling of inline assembly.
522 * To work around this problem, we apply __no_kmsan_checks to ensure that
523 * get_freepointer_safe() returns initialized memory.
524 */
525__no_kmsan_checks
526static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
527{
528 unsigned long freepointer_addr;
529 freeptr_t p;
530
531 if (!debug_pagealloc_enabled_static())
532 return get_freepointer(s, object);
533
534 object = kasan_reset_tag(object);
535 freepointer_addr = (unsigned long)object + s->offset;
536 copy_from_kernel_nofault(&p, (freeptr_t *)freepointer_addr, sizeof(p));
537 return freelist_ptr_decode(s, p, freepointer_addr);
538}
539
540static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
541{
542 unsigned long freeptr_addr = (unsigned long)object + s->offset;
543
544#ifdef CONFIG_SLAB_FREELIST_HARDENED
545 BUG_ON(object == fp); /* naive detection of double free or corruption */
546#endif
547
548 freeptr_addr = (unsigned long)kasan_reset_tag((void *)freeptr_addr);
549 *(freeptr_t *)freeptr_addr = freelist_ptr_encode(s, fp, freeptr_addr);
550}
551
552/*
553 * See comment in calculate_sizes().
554 */
555static inline bool freeptr_outside_object(struct kmem_cache *s)
556{
557 return s->offset >= s->inuse;
558}
559
560/*
561 * Return offset of the end of info block which is inuse + free pointer if
562 * not overlapping with object.
563 */
564static inline unsigned int get_info_end(struct kmem_cache *s)
565{
566 if (freeptr_outside_object(s))
567 return s->inuse + sizeof(void *);
568 else
569 return s->inuse;
570}
571
572/* Loop over all objects in a slab */
573#define for_each_object(__p, __s, __addr, __objects) \
574 for (__p = fixup_red_left(__s, __addr); \
575 __p < (__addr) + (__objects) * (__s)->size; \
576 __p += (__s)->size)
577
578static inline unsigned int order_objects(unsigned int order, unsigned int size)
579{
580 return ((unsigned int)PAGE_SIZE << order) / size;
581}
582
583static inline struct kmem_cache_order_objects oo_make(unsigned int order,
584 unsigned int size)
585{
586 struct kmem_cache_order_objects x = {
587 (order << OO_SHIFT) + order_objects(order, size)
588 };
589
590 return x;
591}
592
593static inline unsigned int oo_order(struct kmem_cache_order_objects x)
594{
595 return x.x >> OO_SHIFT;
596}
597
598static inline unsigned int oo_objects(struct kmem_cache_order_objects x)
599{
600 return x.x & OO_MASK;
601}
602
603#ifdef CONFIG_SLUB_CPU_PARTIAL
604static void slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
605{
606 unsigned int nr_slabs;
607
608 s->cpu_partial = nr_objects;
609
610 /*
611 * We take the number of objects but actually limit the number of
612 * slabs on the per cpu partial list, in order to limit excessive
613 * growth of the list. For simplicity we assume that the slabs will
614 * be half-full.
615 */
616 nr_slabs = DIV_ROUND_UP(nr_objects * 2, oo_objects(s->oo));
617 s->cpu_partial_slabs = nr_slabs;
618}
619
620static inline unsigned int slub_get_cpu_partial(struct kmem_cache *s)
621{
622 return s->cpu_partial_slabs;
623}
624#else
625static inline void
626slub_set_cpu_partial(struct kmem_cache *s, unsigned int nr_objects)
627{
628}
629
630static inline unsigned int slub_get_cpu_partial(struct kmem_cache *s)
631{
632 return 0;
633}
634#endif /* CONFIG_SLUB_CPU_PARTIAL */
635
636/*
637 * Per slab locking using the pagelock
638 */
639static __always_inline void slab_lock(struct slab *slab)
640{
641 bit_spin_lock(PG_locked, &slab->__page_flags);
642}
643
644static __always_inline void slab_unlock(struct slab *slab)
645{
646 bit_spin_unlock(PG_locked, &slab->__page_flags);
647}
648
649static inline bool
650__update_freelist_fast(struct slab *slab,
651 void *freelist_old, unsigned long counters_old,
652 void *freelist_new, unsigned long counters_new)
653{
654#ifdef system_has_freelist_aba
655 freelist_aba_t old = { .freelist = freelist_old, .counter = counters_old };
656 freelist_aba_t new = { .freelist = freelist_new, .counter = counters_new };
657
658 return try_cmpxchg_freelist(&slab->freelist_counter.full, &old.full, new.full);
659#else
660 return false;
661#endif
662}
663
664static inline bool
665__update_freelist_slow(struct slab *slab,
666 void *freelist_old, unsigned long counters_old,
667 void *freelist_new, unsigned long counters_new)
668{
669 bool ret = false;
670
671 slab_lock(slab);
672 if (slab->freelist == freelist_old &&
673 slab->counters == counters_old) {
674 slab->freelist = freelist_new;
675 slab->counters = counters_new;
676 ret = true;
677 }
678 slab_unlock(slab);
679
680 return ret;
681}
682
683/*
684 * Interrupts must be disabled (for the fallback code to work right), typically
685 * by an _irqsave() lock variant. On PREEMPT_RT the preempt_disable(), which is
686 * part of bit_spin_lock(), is sufficient because the policy is not to allow any
687 * allocation/ free operation in hardirq context. Therefore nothing can
688 * interrupt the operation.
689 */
690static inline bool __slab_update_freelist(struct kmem_cache *s, struct slab *slab,
691 void *freelist_old, unsigned long counters_old,
692 void *freelist_new, unsigned long counters_new,
693 const char *n)
694{
695 bool ret;
696
697 if (USE_LOCKLESS_FAST_PATH())
698 lockdep_assert_irqs_disabled();
699
700 if (s->flags & __CMPXCHG_DOUBLE) {
701 ret = __update_freelist_fast(slab, freelist_old, counters_old,
702 freelist_new, counters_new);
703 } else {
704 ret = __update_freelist_slow(slab, freelist_old, counters_old,
705 freelist_new, counters_new);
706 }
707 if (likely(ret))
708 return true;
709
710 cpu_relax();
711 stat(s, CMPXCHG_DOUBLE_FAIL);
712
713#ifdef SLUB_DEBUG_CMPXCHG
714 pr_info("%s %s: cmpxchg double redo ", n, s->name);
715#endif
716
717 return false;
718}
719
720static inline bool slab_update_freelist(struct kmem_cache *s, struct slab *slab,
721 void *freelist_old, unsigned long counters_old,
722 void *freelist_new, unsigned long counters_new,
723 const char *n)
724{
725 bool ret;
726
727 if (s->flags & __CMPXCHG_DOUBLE) {
728 ret = __update_freelist_fast(slab, freelist_old, counters_old,
729 freelist_new, counters_new);
730 } else {
731 unsigned long flags;
732
733 local_irq_save(flags);
734 ret = __update_freelist_slow(slab, freelist_old, counters_old,
735 freelist_new, counters_new);
736 local_irq_restore(flags);
737 }
738 if (likely(ret))
739 return true;
740
741 cpu_relax();
742 stat(s, CMPXCHG_DOUBLE_FAIL);
743
744#ifdef SLUB_DEBUG_CMPXCHG
745 pr_info("%s %s: cmpxchg double redo ", n, s->name);
746#endif
747
748 return false;
749}
750
751/*
752 * kmalloc caches has fixed sizes (mostly power of 2), and kmalloc() API
753 * family will round up the real request size to these fixed ones, so
754 * there could be an extra area than what is requested. Save the original
755 * request size in the meta data area, for better debug and sanity check.
756 */
757static inline void set_orig_size(struct kmem_cache *s,
758 void *object, unsigned int orig_size)
759{
760 void *p = kasan_reset_tag(object);
761
762 if (!slub_debug_orig_size(s))
763 return;
764
765 p += get_info_end(s);
766 p += sizeof(struct track) * 2;
767
768 *(unsigned int *)p = orig_size;
769}
770
771static inline unsigned int get_orig_size(struct kmem_cache *s, void *object)
772{
773 void *p = kasan_reset_tag(object);
774
775 if (is_kfence_address(object))
776 return kfence_ksize(object);
777
778 if (!slub_debug_orig_size(s))
779 return s->object_size;
780
781 p += get_info_end(s);
782 p += sizeof(struct track) * 2;
783
784 return *(unsigned int *)p;
785}
786
787#ifdef CONFIG_SLUB_DEBUG
788static unsigned long object_map[BITS_TO_LONGS(MAX_OBJS_PER_PAGE)];
789static DEFINE_SPINLOCK(object_map_lock);
790
791static void __fill_map(unsigned long *obj_map, struct kmem_cache *s,
792 struct slab *slab)
793{
794 void *addr = slab_address(slab);
795 void *p;
796
797 bitmap_zero(obj_map, slab->objects);
798
799 for (p = slab->freelist; p; p = get_freepointer(s, p))
800 set_bit(__obj_to_index(s, addr, p), obj_map);
801}
802
803#if IS_ENABLED(CONFIG_KUNIT)
804static bool slab_add_kunit_errors(void)
805{
806 struct kunit_resource *resource;
807
808 if (!kunit_get_current_test())
809 return false;
810
811 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
812 if (!resource)
813 return false;
814
815 (*(int *)resource->data)++;
816 kunit_put_resource(resource);
817 return true;
818}
819
820bool slab_in_kunit_test(void)
821{
822 struct kunit_resource *resource;
823
824 if (!kunit_get_current_test())
825 return false;
826
827 resource = kunit_find_named_resource(current->kunit_test, "slab_errors");
828 if (!resource)
829 return false;
830
831 kunit_put_resource(resource);
832 return true;
833}
834#else
835static inline bool slab_add_kunit_errors(void) { return false; }
836#endif
837
838static inline unsigned int size_from_object(struct kmem_cache *s)
839{
840 if (s->flags & SLAB_RED_ZONE)
841 return s->size - s->red_left_pad;
842
843 return s->size;
844}
845
846static inline void *restore_red_left(struct kmem_cache *s, void *p)
847{
848 if (s->flags & SLAB_RED_ZONE)
849 p -= s->red_left_pad;
850
851 return p;
852}
853
854/*
855 * Debug settings:
856 */
857#if defined(CONFIG_SLUB_DEBUG_ON)
858static slab_flags_t slub_debug = DEBUG_DEFAULT_FLAGS;
859#else
860static slab_flags_t slub_debug;
861#endif
862
863static char *slub_debug_string;
864static int disable_higher_order_debug;
865
866/*
867 * slub is about to manipulate internal object metadata. This memory lies
868 * outside the range of the allocated object, so accessing it would normally
869 * be reported by kasan as a bounds error. metadata_access_enable() is used
870 * to tell kasan that these accesses are OK.
871 */
872static inline void metadata_access_enable(void)
873{
874 kasan_disable_current();
875 kmsan_disable_current();
876}
877
878static inline void metadata_access_disable(void)
879{
880 kmsan_enable_current();
881 kasan_enable_current();
882}
883
884/*
885 * Object debugging
886 */
887
888/* Verify that a pointer has an address that is valid within a slab page */
889static inline int check_valid_pointer(struct kmem_cache *s,
890 struct slab *slab, void *object)
891{
892 void *base;
893
894 if (!object)
895 return 1;
896
897 base = slab_address(slab);
898 object = kasan_reset_tag(object);
899 object = restore_red_left(s, object);
900 if (object < base || object >= base + slab->objects * s->size ||
901 (object - base) % s->size) {
902 return 0;
903 }
904
905 return 1;
906}
907
908static void print_section(char *level, char *text, u8 *addr,
909 unsigned int length)
910{
911 metadata_access_enable();
912 print_hex_dump(level, text, DUMP_PREFIX_ADDRESS,
913 16, 1, kasan_reset_tag((void *)addr), length, 1);
914 metadata_access_disable();
915}
916
917static struct track *get_track(struct kmem_cache *s, void *object,
918 enum track_item alloc)
919{
920 struct track *p;
921
922 p = object + get_info_end(s);
923
924 return kasan_reset_tag(p + alloc);
925}
926
927#ifdef CONFIG_STACKDEPOT
928static noinline depot_stack_handle_t set_track_prepare(void)
929{
930 depot_stack_handle_t handle;
931 unsigned long entries[TRACK_ADDRS_COUNT];
932 unsigned int nr_entries;
933
934 nr_entries = stack_trace_save(entries, ARRAY_SIZE(entries), 3);
935 handle = stack_depot_save(entries, nr_entries, GFP_NOWAIT);
936
937 return handle;
938}
939#else
940static inline depot_stack_handle_t set_track_prepare(void)
941{
942 return 0;
943}
944#endif
945
946static void set_track_update(struct kmem_cache *s, void *object,
947 enum track_item alloc, unsigned long addr,
948 depot_stack_handle_t handle)
949{
950 struct track *p = get_track(s, object, alloc);
951
952#ifdef CONFIG_STACKDEPOT
953 p->handle = handle;
954#endif
955 p->addr = addr;
956 p->cpu = smp_processor_id();
957 p->pid = current->pid;
958 p->when = jiffies;
959}
960
961static __always_inline void set_track(struct kmem_cache *s, void *object,
962 enum track_item alloc, unsigned long addr)
963{
964 depot_stack_handle_t handle = set_track_prepare();
965
966 set_track_update(s, object, alloc, addr, handle);
967}
968
969static void init_tracking(struct kmem_cache *s, void *object)
970{
971 struct track *p;
972
973 if (!(s->flags & SLAB_STORE_USER))
974 return;
975
976 p = get_track(s, object, TRACK_ALLOC);
977 memset(p, 0, 2*sizeof(struct track));
978}
979
980static void print_track(const char *s, struct track *t, unsigned long pr_time)
981{
982 depot_stack_handle_t handle __maybe_unused;
983
984 if (!t->addr)
985 return;
986
987 pr_err("%s in %pS age=%lu cpu=%u pid=%d\n",
988 s, (void *)t->addr, pr_time - t->when, t->cpu, t->pid);
989#ifdef CONFIG_STACKDEPOT
990 handle = READ_ONCE(t->handle);
991 if (handle)
992 stack_depot_print(handle);
993 else
994 pr_err("object allocation/free stack trace missing\n");
995#endif
996}
997
998void print_tracking(struct kmem_cache *s, void *object)
999{
1000 unsigned long pr_time = jiffies;
1001 if (!(s->flags & SLAB_STORE_USER))
1002 return;
1003
1004 print_track("Allocated", get_track(s, object, TRACK_ALLOC), pr_time);
1005 print_track("Freed", get_track(s, object, TRACK_FREE), pr_time);
1006}
1007
1008static void print_slab_info(const struct slab *slab)
1009{
1010 pr_err("Slab 0x%p objects=%u used=%u fp=0x%p flags=%pGp\n",
1011 slab, slab->objects, slab->inuse, slab->freelist,
1012 &slab->__page_flags);
1013}
1014
1015void skip_orig_size_check(struct kmem_cache *s, const void *object)
1016{
1017 set_orig_size(s, (void *)object, s->object_size);
1018}
1019
1020static void slab_bug(struct kmem_cache *s, char *fmt, ...)
1021{
1022 struct va_format vaf;
1023 va_list args;
1024
1025 va_start(args, fmt);
1026 vaf.fmt = fmt;
1027 vaf.va = &args;
1028 pr_err("=============================================================================\n");
1029 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
1030 pr_err("-----------------------------------------------------------------------------\n\n");
1031 va_end(args);
1032}
1033
1034__printf(2, 3)
1035static void slab_fix(struct kmem_cache *s, char *fmt, ...)
1036{
1037 struct va_format vaf;
1038 va_list args;
1039
1040 if (slab_add_kunit_errors())
1041 return;
1042
1043 va_start(args, fmt);
1044 vaf.fmt = fmt;
1045 vaf.va = &args;
1046 pr_err("FIX %s: %pV\n", s->name, &vaf);
1047 va_end(args);
1048}
1049
1050static void print_trailer(struct kmem_cache *s, struct slab *slab, u8 *p)
1051{
1052 unsigned int off; /* Offset of last byte */
1053 u8 *addr = slab_address(slab);
1054
1055 print_tracking(s, p);
1056
1057 print_slab_info(slab);
1058
1059 pr_err("Object 0x%p @offset=%tu fp=0x%p\n\n",
1060 p, p - addr, get_freepointer(s, p));
1061
1062 if (s->flags & SLAB_RED_ZONE)
1063 print_section(KERN_ERR, "Redzone ", p - s->red_left_pad,
1064 s->red_left_pad);
1065 else if (p > addr + 16)
1066 print_section(KERN_ERR, "Bytes b4 ", p - 16, 16);
1067
1068 print_section(KERN_ERR, "Object ", p,
1069 min_t(unsigned int, s->object_size, PAGE_SIZE));
1070 if (s->flags & SLAB_RED_ZONE)
1071 print_section(KERN_ERR, "Redzone ", p + s->object_size,
1072 s->inuse - s->object_size);
1073
1074 off = get_info_end(s);
1075
1076 if (s->flags & SLAB_STORE_USER)
1077 off += 2 * sizeof(struct track);
1078
1079 if (slub_debug_orig_size(s))
1080 off += sizeof(unsigned int);
1081
1082 off += kasan_metadata_size(s, false);
1083
1084 if (off != size_from_object(s))
1085 /* Beginning of the filler is the free pointer */
1086 print_section(KERN_ERR, "Padding ", p + off,
1087 size_from_object(s) - off);
1088
1089 dump_stack();
1090}
1091
1092static void object_err(struct kmem_cache *s, struct slab *slab,
1093 u8 *object, char *reason)
1094{
1095 if (slab_add_kunit_errors())
1096 return;
1097
1098 slab_bug(s, "%s", reason);
1099 print_trailer(s, slab, object);
1100 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1101}
1102
1103static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1104 void **freelist, void *nextfree)
1105{
1106 if ((s->flags & SLAB_CONSISTENCY_CHECKS) &&
1107 !check_valid_pointer(s, slab, nextfree) && freelist) {
1108 object_err(s, slab, *freelist, "Freechain corrupt");
1109 *freelist = NULL;
1110 slab_fix(s, "Isolate corrupted freechain");
1111 return true;
1112 }
1113
1114 return false;
1115}
1116
1117static __printf(3, 4) void slab_err(struct kmem_cache *s, struct slab *slab,
1118 const char *fmt, ...)
1119{
1120 va_list args;
1121 char buf[100];
1122
1123 if (slab_add_kunit_errors())
1124 return;
1125
1126 va_start(args, fmt);
1127 vsnprintf(buf, sizeof(buf), fmt, args);
1128 va_end(args);
1129 slab_bug(s, "%s", buf);
1130 print_slab_info(slab);
1131 dump_stack();
1132 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1133}
1134
1135static void init_object(struct kmem_cache *s, void *object, u8 val)
1136{
1137 u8 *p = kasan_reset_tag(object);
1138 unsigned int poison_size = s->object_size;
1139
1140 if (s->flags & SLAB_RED_ZONE) {
1141 /*
1142 * Here and below, avoid overwriting the KMSAN shadow. Keeping
1143 * the shadow makes it possible to distinguish uninit-value
1144 * from use-after-free.
1145 */
1146 memset_no_sanitize_memory(p - s->red_left_pad, val,
1147 s->red_left_pad);
1148
1149 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1150 /*
1151 * Redzone the extra allocated space by kmalloc than
1152 * requested, and the poison size will be limited to
1153 * the original request size accordingly.
1154 */
1155 poison_size = get_orig_size(s, object);
1156 }
1157 }
1158
1159 if (s->flags & __OBJECT_POISON) {
1160 memset_no_sanitize_memory(p, POISON_FREE, poison_size - 1);
1161 memset_no_sanitize_memory(p + poison_size - 1, POISON_END, 1);
1162 }
1163
1164 if (s->flags & SLAB_RED_ZONE)
1165 memset_no_sanitize_memory(p + poison_size, val,
1166 s->inuse - poison_size);
1167}
1168
1169static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
1170 void *from, void *to)
1171{
1172 slab_fix(s, "Restoring %s 0x%p-0x%p=0x%x", message, from, to - 1, data);
1173 memset(from, data, to - from);
1174}
1175
1176#ifdef CONFIG_KMSAN
1177#define pad_check_attributes noinline __no_kmsan_checks
1178#else
1179#define pad_check_attributes
1180#endif
1181
1182static pad_check_attributes int
1183check_bytes_and_report(struct kmem_cache *s, struct slab *slab,
1184 u8 *object, char *what,
1185 u8 *start, unsigned int value, unsigned int bytes)
1186{
1187 u8 *fault;
1188 u8 *end;
1189 u8 *addr = slab_address(slab);
1190
1191 metadata_access_enable();
1192 fault = memchr_inv(kasan_reset_tag(start), value, bytes);
1193 metadata_access_disable();
1194 if (!fault)
1195 return 1;
1196
1197 end = start + bytes;
1198 while (end > fault && end[-1] == value)
1199 end--;
1200
1201 if (slab_add_kunit_errors())
1202 goto skip_bug_print;
1203
1204 slab_bug(s, "%s overwritten", what);
1205 pr_err("0x%p-0x%p @offset=%tu. First byte 0x%x instead of 0x%x\n",
1206 fault, end - 1, fault - addr,
1207 fault[0], value);
1208
1209skip_bug_print:
1210 restore_bytes(s, what, value, fault, end);
1211 return 0;
1212}
1213
1214/*
1215 * Object layout:
1216 *
1217 * object address
1218 * Bytes of the object to be managed.
1219 * If the freepointer may overlay the object then the free
1220 * pointer is at the middle of the object.
1221 *
1222 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
1223 * 0xa5 (POISON_END)
1224 *
1225 * object + s->object_size
1226 * Padding to reach word boundary. This is also used for Redzoning.
1227 * Padding is extended by another word if Redzoning is enabled and
1228 * object_size == inuse.
1229 *
1230 * We fill with 0xbb (SLUB_RED_INACTIVE) for inactive objects and with
1231 * 0xcc (SLUB_RED_ACTIVE) for objects in use.
1232 *
1233 * object + s->inuse
1234 * Meta data starts here.
1235 *
1236 * A. Free pointer (if we cannot overwrite object on free)
1237 * B. Tracking data for SLAB_STORE_USER
1238 * C. Original request size for kmalloc object (SLAB_STORE_USER enabled)
1239 * D. Padding to reach required alignment boundary or at minimum
1240 * one word if debugging is on to be able to detect writes
1241 * before the word boundary.
1242 *
1243 * Padding is done using 0x5a (POISON_INUSE)
1244 *
1245 * object + s->size
1246 * Nothing is used beyond s->size.
1247 *
1248 * If slabcaches are merged then the object_size and inuse boundaries are mostly
1249 * ignored. And therefore no slab options that rely on these boundaries
1250 * may be used with merged slabcaches.
1251 */
1252
1253static int check_pad_bytes(struct kmem_cache *s, struct slab *slab, u8 *p)
1254{
1255 unsigned long off = get_info_end(s); /* The end of info */
1256
1257 if (s->flags & SLAB_STORE_USER) {
1258 /* We also have user information there */
1259 off += 2 * sizeof(struct track);
1260
1261 if (s->flags & SLAB_KMALLOC)
1262 off += sizeof(unsigned int);
1263 }
1264
1265 off += kasan_metadata_size(s, false);
1266
1267 if (size_from_object(s) == off)
1268 return 1;
1269
1270 return check_bytes_and_report(s, slab, p, "Object padding",
1271 p + off, POISON_INUSE, size_from_object(s) - off);
1272}
1273
1274/* Check the pad bytes at the end of a slab page */
1275static pad_check_attributes void
1276slab_pad_check(struct kmem_cache *s, struct slab *slab)
1277{
1278 u8 *start;
1279 u8 *fault;
1280 u8 *end;
1281 u8 *pad;
1282 int length;
1283 int remainder;
1284
1285 if (!(s->flags & SLAB_POISON))
1286 return;
1287
1288 start = slab_address(slab);
1289 length = slab_size(slab);
1290 end = start + length;
1291 remainder = length % s->size;
1292 if (!remainder)
1293 return;
1294
1295 pad = end - remainder;
1296 metadata_access_enable();
1297 fault = memchr_inv(kasan_reset_tag(pad), POISON_INUSE, remainder);
1298 metadata_access_disable();
1299 if (!fault)
1300 return;
1301 while (end > fault && end[-1] == POISON_INUSE)
1302 end--;
1303
1304 slab_err(s, slab, "Padding overwritten. 0x%p-0x%p @offset=%tu",
1305 fault, end - 1, fault - start);
1306 print_section(KERN_ERR, "Padding ", pad, remainder);
1307
1308 restore_bytes(s, "slab padding", POISON_INUSE, fault, end);
1309}
1310
1311static int check_object(struct kmem_cache *s, struct slab *slab,
1312 void *object, u8 val)
1313{
1314 u8 *p = object;
1315 u8 *endobject = object + s->object_size;
1316 unsigned int orig_size, kasan_meta_size;
1317 int ret = 1;
1318
1319 if (s->flags & SLAB_RED_ZONE) {
1320 if (!check_bytes_and_report(s, slab, object, "Left Redzone",
1321 object - s->red_left_pad, val, s->red_left_pad))
1322 ret = 0;
1323
1324 if (!check_bytes_and_report(s, slab, object, "Right Redzone",
1325 endobject, val, s->inuse - s->object_size))
1326 ret = 0;
1327
1328 if (slub_debug_orig_size(s) && val == SLUB_RED_ACTIVE) {
1329 orig_size = get_orig_size(s, object);
1330
1331 if (s->object_size > orig_size &&
1332 !check_bytes_and_report(s, slab, object,
1333 "kmalloc Redzone", p + orig_size,
1334 val, s->object_size - orig_size)) {
1335 ret = 0;
1336 }
1337 }
1338 } else {
1339 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
1340 if (!check_bytes_and_report(s, slab, p, "Alignment padding",
1341 endobject, POISON_INUSE,
1342 s->inuse - s->object_size))
1343 ret = 0;
1344 }
1345 }
1346
1347 if (s->flags & SLAB_POISON) {
1348 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON)) {
1349 /*
1350 * KASAN can save its free meta data inside of the
1351 * object at offset 0. Thus, skip checking the part of
1352 * the redzone that overlaps with the meta data.
1353 */
1354 kasan_meta_size = kasan_metadata_size(s, true);
1355 if (kasan_meta_size < s->object_size - 1 &&
1356 !check_bytes_and_report(s, slab, p, "Poison",
1357 p + kasan_meta_size, POISON_FREE,
1358 s->object_size - kasan_meta_size - 1))
1359 ret = 0;
1360 if (kasan_meta_size < s->object_size &&
1361 !check_bytes_and_report(s, slab, p, "End Poison",
1362 p + s->object_size - 1, POISON_END, 1))
1363 ret = 0;
1364 }
1365 /*
1366 * check_pad_bytes cleans up on its own.
1367 */
1368 if (!check_pad_bytes(s, slab, p))
1369 ret = 0;
1370 }
1371
1372 /*
1373 * Cannot check freepointer while object is allocated if
1374 * object and freepointer overlap.
1375 */
1376 if ((freeptr_outside_object(s) || val != SLUB_RED_ACTIVE) &&
1377 !check_valid_pointer(s, slab, get_freepointer(s, p))) {
1378 object_err(s, slab, p, "Freepointer corrupt");
1379 /*
1380 * No choice but to zap it and thus lose the remainder
1381 * of the free objects in this slab. May cause
1382 * another error because the object count is now wrong.
1383 */
1384 set_freepointer(s, p, NULL);
1385 ret = 0;
1386 }
1387
1388 if (!ret && !slab_in_kunit_test()) {
1389 print_trailer(s, slab, object);
1390 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
1391 }
1392
1393 return ret;
1394}
1395
1396static int check_slab(struct kmem_cache *s, struct slab *slab)
1397{
1398 int maxobj;
1399
1400 if (!folio_test_slab(slab_folio(slab))) {
1401 slab_err(s, slab, "Not a valid slab page");
1402 return 0;
1403 }
1404
1405 maxobj = order_objects(slab_order(slab), s->size);
1406 if (slab->objects > maxobj) {
1407 slab_err(s, slab, "objects %u > max %u",
1408 slab->objects, maxobj);
1409 return 0;
1410 }
1411 if (slab->inuse > slab->objects) {
1412 slab_err(s, slab, "inuse %u > max %u",
1413 slab->inuse, slab->objects);
1414 return 0;
1415 }
1416 if (slab->frozen) {
1417 slab_err(s, slab, "Slab disabled since SLUB metadata consistency check failed");
1418 return 0;
1419 }
1420
1421 /* Slab_pad_check fixes things up after itself */
1422 slab_pad_check(s, slab);
1423 return 1;
1424}
1425
1426/*
1427 * Determine if a certain object in a slab is on the freelist. Must hold the
1428 * slab lock to guarantee that the chains are in a consistent state.
1429 */
1430static int on_freelist(struct kmem_cache *s, struct slab *slab, void *search)
1431{
1432 int nr = 0;
1433 void *fp;
1434 void *object = NULL;
1435 int max_objects;
1436
1437 fp = slab->freelist;
1438 while (fp && nr <= slab->objects) {
1439 if (fp == search)
1440 return 1;
1441 if (!check_valid_pointer(s, slab, fp)) {
1442 if (object) {
1443 object_err(s, slab, object,
1444 "Freechain corrupt");
1445 set_freepointer(s, object, NULL);
1446 } else {
1447 slab_err(s, slab, "Freepointer corrupt");
1448 slab->freelist = NULL;
1449 slab->inuse = slab->objects;
1450 slab_fix(s, "Freelist cleared");
1451 return 0;
1452 }
1453 break;
1454 }
1455 object = fp;
1456 fp = get_freepointer(s, object);
1457 nr++;
1458 }
1459
1460 max_objects = order_objects(slab_order(slab), s->size);
1461 if (max_objects > MAX_OBJS_PER_PAGE)
1462 max_objects = MAX_OBJS_PER_PAGE;
1463
1464 if (slab->objects != max_objects) {
1465 slab_err(s, slab, "Wrong number of objects. Found %d but should be %d",
1466 slab->objects, max_objects);
1467 slab->objects = max_objects;
1468 slab_fix(s, "Number of objects adjusted");
1469 }
1470 if (slab->inuse != slab->objects - nr) {
1471 slab_err(s, slab, "Wrong object count. Counter is %d but counted were %d",
1472 slab->inuse, slab->objects - nr);
1473 slab->inuse = slab->objects - nr;
1474 slab_fix(s, "Object count adjusted");
1475 }
1476 return search == NULL;
1477}
1478
1479static void trace(struct kmem_cache *s, struct slab *slab, void *object,
1480 int alloc)
1481{
1482 if (s->flags & SLAB_TRACE) {
1483 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
1484 s->name,
1485 alloc ? "alloc" : "free",
1486 object, slab->inuse,
1487 slab->freelist);
1488
1489 if (!alloc)
1490 print_section(KERN_INFO, "Object ", (void *)object,
1491 s->object_size);
1492
1493 dump_stack();
1494 }
1495}
1496
1497/*
1498 * Tracking of fully allocated slabs for debugging purposes.
1499 */
1500static void add_full(struct kmem_cache *s,
1501 struct kmem_cache_node *n, struct slab *slab)
1502{
1503 if (!(s->flags & SLAB_STORE_USER))
1504 return;
1505
1506 lockdep_assert_held(&n->list_lock);
1507 list_add(&slab->slab_list, &n->full);
1508}
1509
1510static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct slab *slab)
1511{
1512 if (!(s->flags & SLAB_STORE_USER))
1513 return;
1514
1515 lockdep_assert_held(&n->list_lock);
1516 list_del(&slab->slab_list);
1517}
1518
1519static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1520{
1521 return atomic_long_read(&n->nr_slabs);
1522}
1523
1524static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1525{
1526 struct kmem_cache_node *n = get_node(s, node);
1527
1528 atomic_long_inc(&n->nr_slabs);
1529 atomic_long_add(objects, &n->total_objects);
1530}
1531static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1532{
1533 struct kmem_cache_node *n = get_node(s, node);
1534
1535 atomic_long_dec(&n->nr_slabs);
1536 atomic_long_sub(objects, &n->total_objects);
1537}
1538
1539/* Object debug checks for alloc/free paths */
1540static void setup_object_debug(struct kmem_cache *s, void *object)
1541{
1542 if (!kmem_cache_debug_flags(s, SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON))
1543 return;
1544
1545 init_object(s, object, SLUB_RED_INACTIVE);
1546 init_tracking(s, object);
1547}
1548
1549static
1550void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr)
1551{
1552 if (!kmem_cache_debug_flags(s, SLAB_POISON))
1553 return;
1554
1555 metadata_access_enable();
1556 memset(kasan_reset_tag(addr), POISON_INUSE, slab_size(slab));
1557 metadata_access_disable();
1558}
1559
1560static inline int alloc_consistency_checks(struct kmem_cache *s,
1561 struct slab *slab, void *object)
1562{
1563 if (!check_slab(s, slab))
1564 return 0;
1565
1566 if (!check_valid_pointer(s, slab, object)) {
1567 object_err(s, slab, object, "Freelist Pointer check fails");
1568 return 0;
1569 }
1570
1571 if (!check_object(s, slab, object, SLUB_RED_INACTIVE))
1572 return 0;
1573
1574 return 1;
1575}
1576
1577static noinline bool alloc_debug_processing(struct kmem_cache *s,
1578 struct slab *slab, void *object, int orig_size)
1579{
1580 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1581 if (!alloc_consistency_checks(s, slab, object))
1582 goto bad;
1583 }
1584
1585 /* Success. Perform special debug activities for allocs */
1586 trace(s, slab, object, 1);
1587 set_orig_size(s, object, orig_size);
1588 init_object(s, object, SLUB_RED_ACTIVE);
1589 return true;
1590
1591bad:
1592 if (folio_test_slab(slab_folio(slab))) {
1593 /*
1594 * If this is a slab page then lets do the best we can
1595 * to avoid issues in the future. Marking all objects
1596 * as used avoids touching the remaining objects.
1597 */
1598 slab_fix(s, "Marking all objects used");
1599 slab->inuse = slab->objects;
1600 slab->freelist = NULL;
1601 slab->frozen = 1; /* mark consistency-failed slab as frozen */
1602 }
1603 return false;
1604}
1605
1606static inline int free_consistency_checks(struct kmem_cache *s,
1607 struct slab *slab, void *object, unsigned long addr)
1608{
1609 if (!check_valid_pointer(s, slab, object)) {
1610 slab_err(s, slab, "Invalid object pointer 0x%p", object);
1611 return 0;
1612 }
1613
1614 if (on_freelist(s, slab, object)) {
1615 object_err(s, slab, object, "Object already free");
1616 return 0;
1617 }
1618
1619 if (!check_object(s, slab, object, SLUB_RED_ACTIVE))
1620 return 0;
1621
1622 if (unlikely(s != slab->slab_cache)) {
1623 if (!folio_test_slab(slab_folio(slab))) {
1624 slab_err(s, slab, "Attempt to free object(0x%p) outside of slab",
1625 object);
1626 } else if (!slab->slab_cache) {
1627 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1628 object);
1629 dump_stack();
1630 } else
1631 object_err(s, slab, object,
1632 "page slab pointer corrupt.");
1633 return 0;
1634 }
1635 return 1;
1636}
1637
1638/*
1639 * Parse a block of slab_debug options. Blocks are delimited by ';'
1640 *
1641 * @str: start of block
1642 * @flags: returns parsed flags, or DEBUG_DEFAULT_FLAGS if none specified
1643 * @slabs: return start of list of slabs, or NULL when there's no list
1644 * @init: assume this is initial parsing and not per-kmem-create parsing
1645 *
1646 * returns the start of next block if there's any, or NULL
1647 */
1648static char *
1649parse_slub_debug_flags(char *str, slab_flags_t *flags, char **slabs, bool init)
1650{
1651 bool higher_order_disable = false;
1652
1653 /* Skip any completely empty blocks */
1654 while (*str && *str == ';')
1655 str++;
1656
1657 if (*str == ',') {
1658 /*
1659 * No options but restriction on slabs. This means full
1660 * debugging for slabs matching a pattern.
1661 */
1662 *flags = DEBUG_DEFAULT_FLAGS;
1663 goto check_slabs;
1664 }
1665 *flags = 0;
1666
1667 /* Determine which debug features should be switched on */
1668 for (; *str && *str != ',' && *str != ';'; str++) {
1669 switch (tolower(*str)) {
1670 case '-':
1671 *flags = 0;
1672 break;
1673 case 'f':
1674 *flags |= SLAB_CONSISTENCY_CHECKS;
1675 break;
1676 case 'z':
1677 *flags |= SLAB_RED_ZONE;
1678 break;
1679 case 'p':
1680 *flags |= SLAB_POISON;
1681 break;
1682 case 'u':
1683 *flags |= SLAB_STORE_USER;
1684 break;
1685 case 't':
1686 *flags |= SLAB_TRACE;
1687 break;
1688 case 'a':
1689 *flags |= SLAB_FAILSLAB;
1690 break;
1691 case 'o':
1692 /*
1693 * Avoid enabling debugging on caches if its minimum
1694 * order would increase as a result.
1695 */
1696 higher_order_disable = true;
1697 break;
1698 default:
1699 if (init)
1700 pr_err("slab_debug option '%c' unknown. skipped\n", *str);
1701 }
1702 }
1703check_slabs:
1704 if (*str == ',')
1705 *slabs = ++str;
1706 else
1707 *slabs = NULL;
1708
1709 /* Skip over the slab list */
1710 while (*str && *str != ';')
1711 str++;
1712
1713 /* Skip any completely empty blocks */
1714 while (*str && *str == ';')
1715 str++;
1716
1717 if (init && higher_order_disable)
1718 disable_higher_order_debug = 1;
1719
1720 if (*str)
1721 return str;
1722 else
1723 return NULL;
1724}
1725
1726static int __init setup_slub_debug(char *str)
1727{
1728 slab_flags_t flags;
1729 slab_flags_t global_flags;
1730 char *saved_str;
1731 char *slab_list;
1732 bool global_slub_debug_changed = false;
1733 bool slab_list_specified = false;
1734
1735 global_flags = DEBUG_DEFAULT_FLAGS;
1736 if (*str++ != '=' || !*str)
1737 /*
1738 * No options specified. Switch on full debugging.
1739 */
1740 goto out;
1741
1742 saved_str = str;
1743 while (str) {
1744 str = parse_slub_debug_flags(str, &flags, &slab_list, true);
1745
1746 if (!slab_list) {
1747 global_flags = flags;
1748 global_slub_debug_changed = true;
1749 } else {
1750 slab_list_specified = true;
1751 if (flags & SLAB_STORE_USER)
1752 stack_depot_request_early_init();
1753 }
1754 }
1755
1756 /*
1757 * For backwards compatibility, a single list of flags with list of
1758 * slabs means debugging is only changed for those slabs, so the global
1759 * slab_debug should be unchanged (0 or DEBUG_DEFAULT_FLAGS, depending
1760 * on CONFIG_SLUB_DEBUG_ON). We can extended that to multiple lists as
1761 * long as there is no option specifying flags without a slab list.
1762 */
1763 if (slab_list_specified) {
1764 if (!global_slub_debug_changed)
1765 global_flags = slub_debug;
1766 slub_debug_string = saved_str;
1767 }
1768out:
1769 slub_debug = global_flags;
1770 if (slub_debug & SLAB_STORE_USER)
1771 stack_depot_request_early_init();
1772 if (slub_debug != 0 || slub_debug_string)
1773 static_branch_enable(&slub_debug_enabled);
1774 else
1775 static_branch_disable(&slub_debug_enabled);
1776 if ((static_branch_unlikely(&init_on_alloc) ||
1777 static_branch_unlikely(&init_on_free)) &&
1778 (slub_debug & SLAB_POISON))
1779 pr_info("mem auto-init: SLAB_POISON will take precedence over init_on_alloc/init_on_free\n");
1780 return 1;
1781}
1782
1783__setup("slab_debug", setup_slub_debug);
1784__setup_param("slub_debug", slub_debug, setup_slub_debug, 0);
1785
1786/*
1787 * kmem_cache_flags - apply debugging options to the cache
1788 * @flags: flags to set
1789 * @name: name of the cache
1790 *
1791 * Debug option(s) are applied to @flags. In addition to the debug
1792 * option(s), if a slab name (or multiple) is specified i.e.
1793 * slab_debug=<Debug-Options>,<slab name1>,<slab name2> ...
1794 * then only the select slabs will receive the debug option(s).
1795 */
1796slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1797{
1798 char *iter;
1799 size_t len;
1800 char *next_block;
1801 slab_flags_t block_flags;
1802 slab_flags_t slub_debug_local = slub_debug;
1803
1804 if (flags & SLAB_NO_USER_FLAGS)
1805 return flags;
1806
1807 /*
1808 * If the slab cache is for debugging (e.g. kmemleak) then
1809 * don't store user (stack trace) information by default,
1810 * but let the user enable it via the command line below.
1811 */
1812 if (flags & SLAB_NOLEAKTRACE)
1813 slub_debug_local &= ~SLAB_STORE_USER;
1814
1815 len = strlen(name);
1816 next_block = slub_debug_string;
1817 /* Go through all blocks of debug options, see if any matches our slab's name */
1818 while (next_block) {
1819 next_block = parse_slub_debug_flags(next_block, &block_flags, &iter, false);
1820 if (!iter)
1821 continue;
1822 /* Found a block that has a slab list, search it */
1823 while (*iter) {
1824 char *end, *glob;
1825 size_t cmplen;
1826
1827 end = strchrnul(iter, ',');
1828 if (next_block && next_block < end)
1829 end = next_block - 1;
1830
1831 glob = strnchr(iter, end - iter, '*');
1832 if (glob)
1833 cmplen = glob - iter;
1834 else
1835 cmplen = max_t(size_t, len, (end - iter));
1836
1837 if (!strncmp(name, iter, cmplen)) {
1838 flags |= block_flags;
1839 return flags;
1840 }
1841
1842 if (!*end || *end == ';')
1843 break;
1844 iter = end + 1;
1845 }
1846 }
1847
1848 return flags | slub_debug_local;
1849}
1850#else /* !CONFIG_SLUB_DEBUG */
1851static inline void setup_object_debug(struct kmem_cache *s, void *object) {}
1852static inline
1853void setup_slab_debug(struct kmem_cache *s, struct slab *slab, void *addr) {}
1854
1855static inline bool alloc_debug_processing(struct kmem_cache *s,
1856 struct slab *slab, void *object, int orig_size) { return true; }
1857
1858static inline bool free_debug_processing(struct kmem_cache *s,
1859 struct slab *slab, void *head, void *tail, int *bulk_cnt,
1860 unsigned long addr, depot_stack_handle_t handle) { return true; }
1861
1862static inline void slab_pad_check(struct kmem_cache *s, struct slab *slab) {}
1863static inline int check_object(struct kmem_cache *s, struct slab *slab,
1864 void *object, u8 val) { return 1; }
1865static inline depot_stack_handle_t set_track_prepare(void) { return 0; }
1866static inline void set_track(struct kmem_cache *s, void *object,
1867 enum track_item alloc, unsigned long addr) {}
1868static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1869 struct slab *slab) {}
1870static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1871 struct slab *slab) {}
1872slab_flags_t kmem_cache_flags(slab_flags_t flags, const char *name)
1873{
1874 return flags;
1875}
1876#define slub_debug 0
1877
1878#define disable_higher_order_debug 0
1879
1880static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1881 { return 0; }
1882static inline void inc_slabs_node(struct kmem_cache *s, int node,
1883 int objects) {}
1884static inline void dec_slabs_node(struct kmem_cache *s, int node,
1885 int objects) {}
1886#ifndef CONFIG_SLUB_TINY
1887static bool freelist_corrupted(struct kmem_cache *s, struct slab *slab,
1888 void **freelist, void *nextfree)
1889{
1890 return false;
1891}
1892#endif
1893#endif /* CONFIG_SLUB_DEBUG */
1894
1895#ifdef CONFIG_SLAB_OBJ_EXT
1896
1897#ifdef CONFIG_MEM_ALLOC_PROFILING_DEBUG
1898
1899static inline void mark_objexts_empty(struct slabobj_ext *obj_exts)
1900{
1901 struct slabobj_ext *slab_exts;
1902 struct slab *obj_exts_slab;
1903
1904 obj_exts_slab = virt_to_slab(obj_exts);
1905 slab_exts = slab_obj_exts(obj_exts_slab);
1906 if (slab_exts) {
1907 unsigned int offs = obj_to_index(obj_exts_slab->slab_cache,
1908 obj_exts_slab, obj_exts);
1909 /* codetag should be NULL */
1910 WARN_ON(slab_exts[offs].ref.ct);
1911 set_codetag_empty(&slab_exts[offs].ref);
1912 }
1913}
1914
1915static inline void mark_failed_objexts_alloc(struct slab *slab)
1916{
1917 slab->obj_exts = OBJEXTS_ALLOC_FAIL;
1918}
1919
1920static inline void handle_failed_objexts_alloc(unsigned long obj_exts,
1921 struct slabobj_ext *vec, unsigned int objects)
1922{
1923 /*
1924 * If vector previously failed to allocate then we have live
1925 * objects with no tag reference. Mark all references in this
1926 * vector as empty to avoid warnings later on.
1927 */
1928 if (obj_exts & OBJEXTS_ALLOC_FAIL) {
1929 unsigned int i;
1930
1931 for (i = 0; i < objects; i++)
1932 set_codetag_empty(&vec[i].ref);
1933 }
1934}
1935
1936#else /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
1937
1938static inline void mark_objexts_empty(struct slabobj_ext *obj_exts) {}
1939static inline void mark_failed_objexts_alloc(struct slab *slab) {}
1940static inline void handle_failed_objexts_alloc(unsigned long obj_exts,
1941 struct slabobj_ext *vec, unsigned int objects) {}
1942
1943#endif /* CONFIG_MEM_ALLOC_PROFILING_DEBUG */
1944
1945/*
1946 * The allocated objcg pointers array is not accounted directly.
1947 * Moreover, it should not come from DMA buffer and is not readily
1948 * reclaimable. So those GFP bits should be masked off.
1949 */
1950#define OBJCGS_CLEAR_MASK (__GFP_DMA | __GFP_RECLAIMABLE | \
1951 __GFP_ACCOUNT | __GFP_NOFAIL)
1952
1953int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
1954 gfp_t gfp, bool new_slab)
1955{
1956 unsigned int objects = objs_per_slab(s, slab);
1957 unsigned long new_exts;
1958 unsigned long old_exts;
1959 struct slabobj_ext *vec;
1960
1961 gfp &= ~OBJCGS_CLEAR_MASK;
1962 /* Prevent recursive extension vector allocation */
1963 gfp |= __GFP_NO_OBJ_EXT;
1964 vec = kcalloc_node(objects, sizeof(struct slabobj_ext), gfp,
1965 slab_nid(slab));
1966 if (!vec) {
1967 /* Mark vectors which failed to allocate */
1968 if (new_slab)
1969 mark_failed_objexts_alloc(slab);
1970
1971 return -ENOMEM;
1972 }
1973
1974 new_exts = (unsigned long)vec;
1975#ifdef CONFIG_MEMCG
1976 new_exts |= MEMCG_DATA_OBJEXTS;
1977#endif
1978 old_exts = READ_ONCE(slab->obj_exts);
1979 handle_failed_objexts_alloc(old_exts, vec, objects);
1980 if (new_slab) {
1981 /*
1982 * If the slab is brand new and nobody can yet access its
1983 * obj_exts, no synchronization is required and obj_exts can
1984 * be simply assigned.
1985 */
1986 slab->obj_exts = new_exts;
1987 } else if ((old_exts & ~OBJEXTS_FLAGS_MASK) ||
1988 cmpxchg(&slab->obj_exts, old_exts, new_exts) != old_exts) {
1989 /*
1990 * If the slab is already in use, somebody can allocate and
1991 * assign slabobj_exts in parallel. In this case the existing
1992 * objcg vector should be reused.
1993 */
1994 mark_objexts_empty(vec);
1995 kfree(vec);
1996 return 0;
1997 }
1998
1999 kmemleak_not_leak(vec);
2000 return 0;
2001}
2002
2003static inline void free_slab_obj_exts(struct slab *slab)
2004{
2005 struct slabobj_ext *obj_exts;
2006
2007 obj_exts = slab_obj_exts(slab);
2008 if (!obj_exts)
2009 return;
2010
2011 /*
2012 * obj_exts was created with __GFP_NO_OBJ_EXT flag, therefore its
2013 * corresponding extension will be NULL. alloc_tag_sub() will throw a
2014 * warning if slab has extensions but the extension of an object is
2015 * NULL, therefore replace NULL with CODETAG_EMPTY to indicate that
2016 * the extension for obj_exts is expected to be NULL.
2017 */
2018 mark_objexts_empty(obj_exts);
2019 kfree(obj_exts);
2020 slab->obj_exts = 0;
2021}
2022
2023static inline bool need_slab_obj_ext(void)
2024{
2025 if (mem_alloc_profiling_enabled())
2026 return true;
2027
2028 /*
2029 * CONFIG_MEMCG creates vector of obj_cgroup objects conditionally
2030 * inside memcg_slab_post_alloc_hook. No other users for now.
2031 */
2032 return false;
2033}
2034
2035#else /* CONFIG_SLAB_OBJ_EXT */
2036
2037static int alloc_slab_obj_exts(struct slab *slab, struct kmem_cache *s,
2038 gfp_t gfp, bool new_slab)
2039{
2040 return 0;
2041}
2042
2043static inline void free_slab_obj_exts(struct slab *slab)
2044{
2045}
2046
2047static inline bool need_slab_obj_ext(void)
2048{
2049 return false;
2050}
2051
2052#endif /* CONFIG_SLAB_OBJ_EXT */
2053
2054#ifdef CONFIG_MEM_ALLOC_PROFILING
2055
2056static inline struct slabobj_ext *
2057prepare_slab_obj_exts_hook(struct kmem_cache *s, gfp_t flags, void *p)
2058{
2059 struct slab *slab;
2060
2061 if (!p)
2062 return NULL;
2063
2064 if (s->flags & (SLAB_NO_OBJ_EXT | SLAB_NOLEAKTRACE))
2065 return NULL;
2066
2067 if (flags & __GFP_NO_OBJ_EXT)
2068 return NULL;
2069
2070 slab = virt_to_slab(p);
2071 if (!slab_obj_exts(slab) &&
2072 WARN(alloc_slab_obj_exts(slab, s, flags, false),
2073 "%s, %s: Failed to create slab extension vector!\n",
2074 __func__, s->name))
2075 return NULL;
2076
2077 return slab_obj_exts(slab) + obj_to_index(s, slab, p);
2078}
2079
2080static inline void
2081alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2082{
2083 if (need_slab_obj_ext()) {
2084 struct slabobj_ext *obj_exts;
2085
2086 obj_exts = prepare_slab_obj_exts_hook(s, flags, object);
2087 /*
2088 * Currently obj_exts is used only for allocation profiling.
2089 * If other users appear then mem_alloc_profiling_enabled()
2090 * check should be added before alloc_tag_add().
2091 */
2092 if (likely(obj_exts))
2093 alloc_tag_add(&obj_exts->ref, current->alloc_tag, s->size);
2094 }
2095}
2096
2097static inline void
2098alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2099 int objects)
2100{
2101 struct slabobj_ext *obj_exts;
2102 int i;
2103
2104 if (!mem_alloc_profiling_enabled())
2105 return;
2106
2107 /* slab->obj_exts might not be NULL if it was created for MEMCG accounting. */
2108 if (s->flags & (SLAB_NO_OBJ_EXT | SLAB_NOLEAKTRACE))
2109 return;
2110
2111 obj_exts = slab_obj_exts(slab);
2112 if (!obj_exts)
2113 return;
2114
2115 for (i = 0; i < objects; i++) {
2116 unsigned int off = obj_to_index(s, slab, p[i]);
2117
2118 alloc_tag_sub(&obj_exts[off].ref, s->size);
2119 }
2120}
2121
2122#else /* CONFIG_MEM_ALLOC_PROFILING */
2123
2124static inline void
2125alloc_tagging_slab_alloc_hook(struct kmem_cache *s, void *object, gfp_t flags)
2126{
2127}
2128
2129static inline void
2130alloc_tagging_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2131 int objects)
2132{
2133}
2134
2135#endif /* CONFIG_MEM_ALLOC_PROFILING */
2136
2137
2138#ifdef CONFIG_MEMCG
2139
2140static void memcg_alloc_abort_single(struct kmem_cache *s, void *object);
2141
2142static __fastpath_inline
2143bool memcg_slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
2144 gfp_t flags, size_t size, void **p)
2145{
2146 if (likely(!memcg_kmem_online()))
2147 return true;
2148
2149 if (likely(!(flags & __GFP_ACCOUNT) && !(s->flags & SLAB_ACCOUNT)))
2150 return true;
2151
2152 if (likely(__memcg_slab_post_alloc_hook(s, lru, flags, size, p)))
2153 return true;
2154
2155 if (likely(size == 1)) {
2156 memcg_alloc_abort_single(s, *p);
2157 *p = NULL;
2158 } else {
2159 kmem_cache_free_bulk(s, size, p);
2160 }
2161
2162 return false;
2163}
2164
2165static __fastpath_inline
2166void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab, void **p,
2167 int objects)
2168{
2169 struct slabobj_ext *obj_exts;
2170
2171 if (!memcg_kmem_online())
2172 return;
2173
2174 obj_exts = slab_obj_exts(slab);
2175 if (likely(!obj_exts))
2176 return;
2177
2178 __memcg_slab_free_hook(s, slab, p, objects, obj_exts);
2179}
2180
2181static __fastpath_inline
2182bool memcg_slab_post_charge(void *p, gfp_t flags)
2183{
2184 struct slabobj_ext *slab_exts;
2185 struct kmem_cache *s;
2186 struct folio *folio;
2187 struct slab *slab;
2188 unsigned long off;
2189
2190 folio = virt_to_folio(p);
2191 if (!folio_test_slab(folio)) {
2192 int size;
2193
2194 if (folio_memcg_kmem(folio))
2195 return true;
2196
2197 if (__memcg_kmem_charge_page(folio_page(folio, 0), flags,
2198 folio_order(folio)))
2199 return false;
2200
2201 /*
2202 * This folio has already been accounted in the global stats but
2203 * not in the memcg stats. So, subtract from the global and use
2204 * the interface which adds to both global and memcg stats.
2205 */
2206 size = folio_size(folio);
2207 node_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B, -size);
2208 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B, size);
2209 return true;
2210 }
2211
2212 slab = folio_slab(folio);
2213 s = slab->slab_cache;
2214
2215 /*
2216 * Ignore KMALLOC_NORMAL cache to avoid possible circular dependency
2217 * of slab_obj_exts being allocated from the same slab and thus the slab
2218 * becoming effectively unfreeable.
2219 */
2220 if (is_kmalloc_normal(s))
2221 return true;
2222
2223 /* Ignore already charged objects. */
2224 slab_exts = slab_obj_exts(slab);
2225 if (slab_exts) {
2226 off = obj_to_index(s, slab, p);
2227 if (unlikely(slab_exts[off].objcg))
2228 return true;
2229 }
2230
2231 return __memcg_slab_post_alloc_hook(s, NULL, flags, 1, &p);
2232}
2233
2234#else /* CONFIG_MEMCG */
2235static inline bool memcg_slab_post_alloc_hook(struct kmem_cache *s,
2236 struct list_lru *lru,
2237 gfp_t flags, size_t size,
2238 void **p)
2239{
2240 return true;
2241}
2242
2243static inline void memcg_slab_free_hook(struct kmem_cache *s, struct slab *slab,
2244 void **p, int objects)
2245{
2246}
2247
2248static inline bool memcg_slab_post_charge(void *p, gfp_t flags)
2249{
2250 return true;
2251}
2252#endif /* CONFIG_MEMCG */
2253
2254#ifdef CONFIG_SLUB_RCU_DEBUG
2255static void slab_free_after_rcu_debug(struct rcu_head *rcu_head);
2256
2257struct rcu_delayed_free {
2258 struct rcu_head head;
2259 void *object;
2260};
2261#endif
2262
2263/*
2264 * Hooks for other subsystems that check memory allocations. In a typical
2265 * production configuration these hooks all should produce no code at all.
2266 *
2267 * Returns true if freeing of the object can proceed, false if its reuse
2268 * was delayed by CONFIG_SLUB_RCU_DEBUG or KASAN quarantine, or it was returned
2269 * to KFENCE.
2270 */
2271static __always_inline
2272bool slab_free_hook(struct kmem_cache *s, void *x, bool init,
2273 bool after_rcu_delay)
2274{
2275 /* Are the object contents still accessible? */
2276 bool still_accessible = (s->flags & SLAB_TYPESAFE_BY_RCU) && !after_rcu_delay;
2277
2278 kmemleak_free_recursive(x, s->flags);
2279 kmsan_slab_free(s, x);
2280
2281 debug_check_no_locks_freed(x, s->object_size);
2282
2283 if (!(s->flags & SLAB_DEBUG_OBJECTS))
2284 debug_check_no_obj_freed(x, s->object_size);
2285
2286 /* Use KCSAN to help debug racy use-after-free. */
2287 if (!still_accessible)
2288 __kcsan_check_access(x, s->object_size,
2289 KCSAN_ACCESS_WRITE | KCSAN_ACCESS_ASSERT);
2290
2291 if (kfence_free(x))
2292 return false;
2293
2294 /*
2295 * Give KASAN a chance to notice an invalid free operation before we
2296 * modify the object.
2297 */
2298 if (kasan_slab_pre_free(s, x))
2299 return false;
2300
2301#ifdef CONFIG_SLUB_RCU_DEBUG
2302 if (still_accessible) {
2303 struct rcu_delayed_free *delayed_free;
2304
2305 delayed_free = kmalloc(sizeof(*delayed_free), GFP_NOWAIT);
2306 if (delayed_free) {
2307 /*
2308 * Let KASAN track our call stack as a "related work
2309 * creation", just like if the object had been freed
2310 * normally via kfree_rcu().
2311 * We have to do this manually because the rcu_head is
2312 * not located inside the object.
2313 */
2314 kasan_record_aux_stack_noalloc(x);
2315
2316 delayed_free->object = x;
2317 call_rcu(&delayed_free->head, slab_free_after_rcu_debug);
2318 return false;
2319 }
2320 }
2321#endif /* CONFIG_SLUB_RCU_DEBUG */
2322
2323 /*
2324 * As memory initialization might be integrated into KASAN,
2325 * kasan_slab_free and initialization memset's must be
2326 * kept together to avoid discrepancies in behavior.
2327 *
2328 * The initialization memset's clear the object and the metadata,
2329 * but don't touch the SLAB redzone.
2330 *
2331 * The object's freepointer is also avoided if stored outside the
2332 * object.
2333 */
2334 if (unlikely(init)) {
2335 int rsize;
2336 unsigned int inuse, orig_size;
2337
2338 inuse = get_info_end(s);
2339 orig_size = get_orig_size(s, x);
2340 if (!kasan_has_integrated_init())
2341 memset(kasan_reset_tag(x), 0, orig_size);
2342 rsize = (s->flags & SLAB_RED_ZONE) ? s->red_left_pad : 0;
2343 memset((char *)kasan_reset_tag(x) + inuse, 0,
2344 s->size - inuse - rsize);
2345 /*
2346 * Restore orig_size, otherwize kmalloc redzone overwritten
2347 * would be reported
2348 */
2349 set_orig_size(s, x, orig_size);
2350
2351 }
2352 /* KASAN might put x into memory quarantine, delaying its reuse. */
2353 return !kasan_slab_free(s, x, init, still_accessible);
2354}
2355
2356static __fastpath_inline
2357bool slab_free_freelist_hook(struct kmem_cache *s, void **head, void **tail,
2358 int *cnt)
2359{
2360
2361 void *object;
2362 void *next = *head;
2363 void *old_tail = *tail;
2364 bool init;
2365
2366 if (is_kfence_address(next)) {
2367 slab_free_hook(s, next, false, false);
2368 return false;
2369 }
2370
2371 /* Head and tail of the reconstructed freelist */
2372 *head = NULL;
2373 *tail = NULL;
2374
2375 init = slab_want_init_on_free(s);
2376
2377 do {
2378 object = next;
2379 next = get_freepointer(s, object);
2380
2381 /* If object's reuse doesn't have to be delayed */
2382 if (likely(slab_free_hook(s, object, init, false))) {
2383 /* Move object to the new freelist */
2384 set_freepointer(s, object, *head);
2385 *head = object;
2386 if (!*tail)
2387 *tail = object;
2388 } else {
2389 /*
2390 * Adjust the reconstructed freelist depth
2391 * accordingly if object's reuse is delayed.
2392 */
2393 --(*cnt);
2394 }
2395 } while (object != old_tail);
2396
2397 return *head != NULL;
2398}
2399
2400static void *setup_object(struct kmem_cache *s, void *object)
2401{
2402 setup_object_debug(s, object);
2403 object = kasan_init_slab_obj(s, object);
2404 if (unlikely(s->ctor)) {
2405 kasan_unpoison_new_object(s, object);
2406 s->ctor(object);
2407 kasan_poison_new_object(s, object);
2408 }
2409 return object;
2410}
2411
2412/*
2413 * Slab allocation and freeing
2414 */
2415static inline struct slab *alloc_slab_page(gfp_t flags, int node,
2416 struct kmem_cache_order_objects oo)
2417{
2418 struct folio *folio;
2419 struct slab *slab;
2420 unsigned int order = oo_order(oo);
2421
2422 if (node == NUMA_NO_NODE)
2423 folio = (struct folio *)alloc_pages(flags, order);
2424 else
2425 folio = (struct folio *)__alloc_pages_node(node, flags, order);
2426
2427 if (!folio)
2428 return NULL;
2429
2430 slab = folio_slab(folio);
2431 __folio_set_slab(folio);
2432 /* Make the flag visible before any changes to folio->mapping */
2433 smp_wmb();
2434 if (folio_is_pfmemalloc(folio))
2435 slab_set_pfmemalloc(slab);
2436
2437 return slab;
2438}
2439
2440#ifdef CONFIG_SLAB_FREELIST_RANDOM
2441/* Pre-initialize the random sequence cache */
2442static int init_cache_random_seq(struct kmem_cache *s)
2443{
2444 unsigned int count = oo_objects(s->oo);
2445 int err;
2446
2447 /* Bailout if already initialised */
2448 if (s->random_seq)
2449 return 0;
2450
2451 err = cache_random_seq_create(s, count, GFP_KERNEL);
2452 if (err) {
2453 pr_err("SLUB: Unable to initialize free list for %s\n",
2454 s->name);
2455 return err;
2456 }
2457
2458 /* Transform to an offset on the set of pages */
2459 if (s->random_seq) {
2460 unsigned int i;
2461
2462 for (i = 0; i < count; i++)
2463 s->random_seq[i] *= s->size;
2464 }
2465 return 0;
2466}
2467
2468/* Initialize each random sequence freelist per cache */
2469static void __init init_freelist_randomization(void)
2470{
2471 struct kmem_cache *s;
2472
2473 mutex_lock(&slab_mutex);
2474
2475 list_for_each_entry(s, &slab_caches, list)
2476 init_cache_random_seq(s);
2477
2478 mutex_unlock(&slab_mutex);
2479}
2480
2481/* Get the next entry on the pre-computed freelist randomized */
2482static void *next_freelist_entry(struct kmem_cache *s,
2483 unsigned long *pos, void *start,
2484 unsigned long page_limit,
2485 unsigned long freelist_count)
2486{
2487 unsigned int idx;
2488
2489 /*
2490 * If the target page allocation failed, the number of objects on the
2491 * page might be smaller than the usual size defined by the cache.
2492 */
2493 do {
2494 idx = s->random_seq[*pos];
2495 *pos += 1;
2496 if (*pos >= freelist_count)
2497 *pos = 0;
2498 } while (unlikely(idx >= page_limit));
2499
2500 return (char *)start + idx;
2501}
2502
2503/* Shuffle the single linked freelist based on a random pre-computed sequence */
2504static bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2505{
2506 void *start;
2507 void *cur;
2508 void *next;
2509 unsigned long idx, pos, page_limit, freelist_count;
2510
2511 if (slab->objects < 2 || !s->random_seq)
2512 return false;
2513
2514 freelist_count = oo_objects(s->oo);
2515 pos = get_random_u32_below(freelist_count);
2516
2517 page_limit = slab->objects * s->size;
2518 start = fixup_red_left(s, slab_address(slab));
2519
2520 /* First entry is used as the base of the freelist */
2521 cur = next_freelist_entry(s, &pos, start, page_limit, freelist_count);
2522 cur = setup_object(s, cur);
2523 slab->freelist = cur;
2524
2525 for (idx = 1; idx < slab->objects; idx++) {
2526 next = next_freelist_entry(s, &pos, start, page_limit,
2527 freelist_count);
2528 next = setup_object(s, next);
2529 set_freepointer(s, cur, next);
2530 cur = next;
2531 }
2532 set_freepointer(s, cur, NULL);
2533
2534 return true;
2535}
2536#else
2537static inline int init_cache_random_seq(struct kmem_cache *s)
2538{
2539 return 0;
2540}
2541static inline void init_freelist_randomization(void) { }
2542static inline bool shuffle_freelist(struct kmem_cache *s, struct slab *slab)
2543{
2544 return false;
2545}
2546#endif /* CONFIG_SLAB_FREELIST_RANDOM */
2547
2548static __always_inline void account_slab(struct slab *slab, int order,
2549 struct kmem_cache *s, gfp_t gfp)
2550{
2551 if (memcg_kmem_online() && (s->flags & SLAB_ACCOUNT))
2552 alloc_slab_obj_exts(slab, s, gfp, true);
2553
2554 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2555 PAGE_SIZE << order);
2556}
2557
2558static __always_inline void unaccount_slab(struct slab *slab, int order,
2559 struct kmem_cache *s)
2560{
2561 if (memcg_kmem_online() || need_slab_obj_ext())
2562 free_slab_obj_exts(slab);
2563
2564 mod_node_page_state(slab_pgdat(slab), cache_vmstat_idx(s),
2565 -(PAGE_SIZE << order));
2566}
2567
2568static struct slab *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
2569{
2570 struct slab *slab;
2571 struct kmem_cache_order_objects oo = s->oo;
2572 gfp_t alloc_gfp;
2573 void *start, *p, *next;
2574 int idx;
2575 bool shuffle;
2576
2577 flags &= gfp_allowed_mask;
2578
2579 flags |= s->allocflags;
2580
2581 /*
2582 * Let the initial higher-order allocation fail under memory pressure
2583 * so we fall-back to the minimum order allocation.
2584 */
2585 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
2586 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
2587 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~__GFP_RECLAIM;
2588
2589 slab = alloc_slab_page(alloc_gfp, node, oo);
2590 if (unlikely(!slab)) {
2591 oo = s->min;
2592 alloc_gfp = flags;
2593 /*
2594 * Allocation may have failed due to fragmentation.
2595 * Try a lower order alloc if possible
2596 */
2597 slab = alloc_slab_page(alloc_gfp, node, oo);
2598 if (unlikely(!slab))
2599 return NULL;
2600 stat(s, ORDER_FALLBACK);
2601 }
2602
2603 slab->objects = oo_objects(oo);
2604 slab->inuse = 0;
2605 slab->frozen = 0;
2606
2607 account_slab(slab, oo_order(oo), s, flags);
2608
2609 slab->slab_cache = s;
2610
2611 kasan_poison_slab(slab);
2612
2613 start = slab_address(slab);
2614
2615 setup_slab_debug(s, slab, start);
2616
2617 shuffle = shuffle_freelist(s, slab);
2618
2619 if (!shuffle) {
2620 start = fixup_red_left(s, start);
2621 start = setup_object(s, start);
2622 slab->freelist = start;
2623 for (idx = 0, p = start; idx < slab->objects - 1; idx++) {
2624 next = p + s->size;
2625 next = setup_object(s, next);
2626 set_freepointer(s, p, next);
2627 p = next;
2628 }
2629 set_freepointer(s, p, NULL);
2630 }
2631
2632 return slab;
2633}
2634
2635static struct slab *new_slab(struct kmem_cache *s, gfp_t flags, int node)
2636{
2637 if (unlikely(flags & GFP_SLAB_BUG_MASK))
2638 flags = kmalloc_fix_flags(flags);
2639
2640 WARN_ON_ONCE(s->ctor && (flags & __GFP_ZERO));
2641
2642 return allocate_slab(s,
2643 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
2644}
2645
2646static void __free_slab(struct kmem_cache *s, struct slab *slab)
2647{
2648 struct folio *folio = slab_folio(slab);
2649 int order = folio_order(folio);
2650 int pages = 1 << order;
2651
2652 __slab_clear_pfmemalloc(slab);
2653 folio->mapping = NULL;
2654 /* Make the mapping reset visible before clearing the flag */
2655 smp_wmb();
2656 __folio_clear_slab(folio);
2657 mm_account_reclaimed_pages(pages);
2658 unaccount_slab(slab, order, s);
2659 __free_pages(&folio->page, order);
2660}
2661
2662static void rcu_free_slab(struct rcu_head *h)
2663{
2664 struct slab *slab = container_of(h, struct slab, rcu_head);
2665
2666 __free_slab(slab->slab_cache, slab);
2667}
2668
2669static void free_slab(struct kmem_cache *s, struct slab *slab)
2670{
2671 if (kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS)) {
2672 void *p;
2673
2674 slab_pad_check(s, slab);
2675 for_each_object(p, s, slab_address(slab), slab->objects)
2676 check_object(s, slab, p, SLUB_RED_INACTIVE);
2677 }
2678
2679 if (unlikely(s->flags & SLAB_TYPESAFE_BY_RCU))
2680 call_rcu(&slab->rcu_head, rcu_free_slab);
2681 else
2682 __free_slab(s, slab);
2683}
2684
2685static void discard_slab(struct kmem_cache *s, struct slab *slab)
2686{
2687 dec_slabs_node(s, slab_nid(slab), slab->objects);
2688 free_slab(s, slab);
2689}
2690
2691/*
2692 * SLUB reuses PG_workingset bit to keep track of whether it's on
2693 * the per-node partial list.
2694 */
2695static inline bool slab_test_node_partial(const struct slab *slab)
2696{
2697 return folio_test_workingset(slab_folio(slab));
2698}
2699
2700static inline void slab_set_node_partial(struct slab *slab)
2701{
2702 set_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2703}
2704
2705static inline void slab_clear_node_partial(struct slab *slab)
2706{
2707 clear_bit(PG_workingset, folio_flags(slab_folio(slab), 0));
2708}
2709
2710/*
2711 * Management of partially allocated slabs.
2712 */
2713static inline void
2714__add_partial(struct kmem_cache_node *n, struct slab *slab, int tail)
2715{
2716 n->nr_partial++;
2717 if (tail == DEACTIVATE_TO_TAIL)
2718 list_add_tail(&slab->slab_list, &n->partial);
2719 else
2720 list_add(&slab->slab_list, &n->partial);
2721 slab_set_node_partial(slab);
2722}
2723
2724static inline void add_partial(struct kmem_cache_node *n,
2725 struct slab *slab, int tail)
2726{
2727 lockdep_assert_held(&n->list_lock);
2728 __add_partial(n, slab, tail);
2729}
2730
2731static inline void remove_partial(struct kmem_cache_node *n,
2732 struct slab *slab)
2733{
2734 lockdep_assert_held(&n->list_lock);
2735 list_del(&slab->slab_list);
2736 slab_clear_node_partial(slab);
2737 n->nr_partial--;
2738}
2739
2740/*
2741 * Called only for kmem_cache_debug() caches instead of remove_partial(), with a
2742 * slab from the n->partial list. Remove only a single object from the slab, do
2743 * the alloc_debug_processing() checks and leave the slab on the list, or move
2744 * it to full list if it was the last free object.
2745 */
2746static void *alloc_single_from_partial(struct kmem_cache *s,
2747 struct kmem_cache_node *n, struct slab *slab, int orig_size)
2748{
2749 void *object;
2750
2751 lockdep_assert_held(&n->list_lock);
2752
2753 object = slab->freelist;
2754 slab->freelist = get_freepointer(s, object);
2755 slab->inuse++;
2756
2757 if (!alloc_debug_processing(s, slab, object, orig_size)) {
2758 if (folio_test_slab(slab_folio(slab)))
2759 remove_partial(n, slab);
2760 return NULL;
2761 }
2762
2763 if (slab->inuse == slab->objects) {
2764 remove_partial(n, slab);
2765 add_full(s, n, slab);
2766 }
2767
2768 return object;
2769}
2770
2771/*
2772 * Called only for kmem_cache_debug() caches to allocate from a freshly
2773 * allocated slab. Allocate a single object instead of whole freelist
2774 * and put the slab to the partial (or full) list.
2775 */
2776static void *alloc_single_from_new_slab(struct kmem_cache *s,
2777 struct slab *slab, int orig_size)
2778{
2779 int nid = slab_nid(slab);
2780 struct kmem_cache_node *n = get_node(s, nid);
2781 unsigned long flags;
2782 void *object;
2783
2784
2785 object = slab->freelist;
2786 slab->freelist = get_freepointer(s, object);
2787 slab->inuse = 1;
2788
2789 if (!alloc_debug_processing(s, slab, object, orig_size))
2790 /*
2791 * It's not really expected that this would fail on a
2792 * freshly allocated slab, but a concurrent memory
2793 * corruption in theory could cause that.
2794 */
2795 return NULL;
2796
2797 spin_lock_irqsave(&n->list_lock, flags);
2798
2799 if (slab->inuse == slab->objects)
2800 add_full(s, n, slab);
2801 else
2802 add_partial(n, slab, DEACTIVATE_TO_HEAD);
2803
2804 inc_slabs_node(s, nid, slab->objects);
2805 spin_unlock_irqrestore(&n->list_lock, flags);
2806
2807 return object;
2808}
2809
2810#ifdef CONFIG_SLUB_CPU_PARTIAL
2811static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain);
2812#else
2813static inline void put_cpu_partial(struct kmem_cache *s, struct slab *slab,
2814 int drain) { }
2815#endif
2816static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags);
2817
2818/*
2819 * Try to allocate a partial slab from a specific node.
2820 */
2821static struct slab *get_partial_node(struct kmem_cache *s,
2822 struct kmem_cache_node *n,
2823 struct partial_context *pc)
2824{
2825 struct slab *slab, *slab2, *partial = NULL;
2826 unsigned long flags;
2827 unsigned int partial_slabs = 0;
2828
2829 /*
2830 * Racy check. If we mistakenly see no partial slabs then we
2831 * just allocate an empty slab. If we mistakenly try to get a
2832 * partial slab and there is none available then get_partial()
2833 * will return NULL.
2834 */
2835 if (!n || !n->nr_partial)
2836 return NULL;
2837
2838 spin_lock_irqsave(&n->list_lock, flags);
2839 list_for_each_entry_safe(slab, slab2, &n->partial, slab_list) {
2840 if (!pfmemalloc_match(slab, pc->flags))
2841 continue;
2842
2843 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
2844 void *object = alloc_single_from_partial(s, n, slab,
2845 pc->orig_size);
2846 if (object) {
2847 partial = slab;
2848 pc->object = object;
2849 break;
2850 }
2851 continue;
2852 }
2853
2854 remove_partial(n, slab);
2855
2856 if (!partial) {
2857 partial = slab;
2858 stat(s, ALLOC_FROM_PARTIAL);
2859
2860 if ((slub_get_cpu_partial(s) == 0)) {
2861 break;
2862 }
2863 } else {
2864 put_cpu_partial(s, slab, 0);
2865 stat(s, CPU_PARTIAL_NODE);
2866
2867 if (++partial_slabs > slub_get_cpu_partial(s) / 2) {
2868 break;
2869 }
2870 }
2871 }
2872 spin_unlock_irqrestore(&n->list_lock, flags);
2873 return partial;
2874}
2875
2876/*
2877 * Get a slab from somewhere. Search in increasing NUMA distances.
2878 */
2879static struct slab *get_any_partial(struct kmem_cache *s,
2880 struct partial_context *pc)
2881{
2882#ifdef CONFIG_NUMA
2883 struct zonelist *zonelist;
2884 struct zoneref *z;
2885 struct zone *zone;
2886 enum zone_type highest_zoneidx = gfp_zone(pc->flags);
2887 struct slab *slab;
2888 unsigned int cpuset_mems_cookie;
2889
2890 /*
2891 * The defrag ratio allows a configuration of the tradeoffs between
2892 * inter node defragmentation and node local allocations. A lower
2893 * defrag_ratio increases the tendency to do local allocations
2894 * instead of attempting to obtain partial slabs from other nodes.
2895 *
2896 * If the defrag_ratio is set to 0 then kmalloc() always
2897 * returns node local objects. If the ratio is higher then kmalloc()
2898 * may return off node objects because partial slabs are obtained
2899 * from other nodes and filled up.
2900 *
2901 * If /sys/kernel/slab/xx/remote_node_defrag_ratio is set to 100
2902 * (which makes defrag_ratio = 1000) then every (well almost)
2903 * allocation will first attempt to defrag slab caches on other nodes.
2904 * This means scanning over all nodes to look for partial slabs which
2905 * may be expensive if we do it every time we are trying to find a slab
2906 * with available objects.
2907 */
2908 if (!s->remote_node_defrag_ratio ||
2909 get_cycles() % 1024 > s->remote_node_defrag_ratio)
2910 return NULL;
2911
2912 do {
2913 cpuset_mems_cookie = read_mems_allowed_begin();
2914 zonelist = node_zonelist(mempolicy_slab_node(), pc->flags);
2915 for_each_zone_zonelist(zone, z, zonelist, highest_zoneidx) {
2916 struct kmem_cache_node *n;
2917
2918 n = get_node(s, zone_to_nid(zone));
2919
2920 if (n && cpuset_zone_allowed(zone, pc->flags) &&
2921 n->nr_partial > s->min_partial) {
2922 slab = get_partial_node(s, n, pc);
2923 if (slab) {
2924 /*
2925 * Don't check read_mems_allowed_retry()
2926 * here - if mems_allowed was updated in
2927 * parallel, that was a harmless race
2928 * between allocation and the cpuset
2929 * update
2930 */
2931 return slab;
2932 }
2933 }
2934 }
2935 } while (read_mems_allowed_retry(cpuset_mems_cookie));
2936#endif /* CONFIG_NUMA */
2937 return NULL;
2938}
2939
2940/*
2941 * Get a partial slab, lock it and return it.
2942 */
2943static struct slab *get_partial(struct kmem_cache *s, int node,
2944 struct partial_context *pc)
2945{
2946 struct slab *slab;
2947 int searchnode = node;
2948
2949 if (node == NUMA_NO_NODE)
2950 searchnode = numa_mem_id();
2951
2952 slab = get_partial_node(s, get_node(s, searchnode), pc);
2953 if (slab || (node != NUMA_NO_NODE && (pc->flags & __GFP_THISNODE)))
2954 return slab;
2955
2956 return get_any_partial(s, pc);
2957}
2958
2959#ifndef CONFIG_SLUB_TINY
2960
2961#ifdef CONFIG_PREEMPTION
2962/*
2963 * Calculate the next globally unique transaction for disambiguation
2964 * during cmpxchg. The transactions start with the cpu number and are then
2965 * incremented by CONFIG_NR_CPUS.
2966 */
2967#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
2968#else
2969/*
2970 * No preemption supported therefore also no need to check for
2971 * different cpus.
2972 */
2973#define TID_STEP 1
2974#endif /* CONFIG_PREEMPTION */
2975
2976static inline unsigned long next_tid(unsigned long tid)
2977{
2978 return tid + TID_STEP;
2979}
2980
2981#ifdef SLUB_DEBUG_CMPXCHG
2982static inline unsigned int tid_to_cpu(unsigned long tid)
2983{
2984 return tid % TID_STEP;
2985}
2986
2987static inline unsigned long tid_to_event(unsigned long tid)
2988{
2989 return tid / TID_STEP;
2990}
2991#endif
2992
2993static inline unsigned int init_tid(int cpu)
2994{
2995 return cpu;
2996}
2997
2998static inline void note_cmpxchg_failure(const char *n,
2999 const struct kmem_cache *s, unsigned long tid)
3000{
3001#ifdef SLUB_DEBUG_CMPXCHG
3002 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
3003
3004 pr_info("%s %s: cmpxchg redo ", n, s->name);
3005
3006#ifdef CONFIG_PREEMPTION
3007 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
3008 pr_warn("due to cpu change %d -> %d\n",
3009 tid_to_cpu(tid), tid_to_cpu(actual_tid));
3010 else
3011#endif
3012 if (tid_to_event(tid) != tid_to_event(actual_tid))
3013 pr_warn("due to cpu running other code. Event %ld->%ld\n",
3014 tid_to_event(tid), tid_to_event(actual_tid));
3015 else
3016 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
3017 actual_tid, tid, next_tid(tid));
3018#endif
3019 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
3020}
3021
3022static void init_kmem_cache_cpus(struct kmem_cache *s)
3023{
3024 int cpu;
3025 struct kmem_cache_cpu *c;
3026
3027 for_each_possible_cpu(cpu) {
3028 c = per_cpu_ptr(s->cpu_slab, cpu);
3029 local_lock_init(&c->lock);
3030 c->tid = init_tid(cpu);
3031 }
3032}
3033
3034/*
3035 * Finishes removing the cpu slab. Merges cpu's freelist with slab's freelist,
3036 * unfreezes the slabs and puts it on the proper list.
3037 * Assumes the slab has been already safely taken away from kmem_cache_cpu
3038 * by the caller.
3039 */
3040static void deactivate_slab(struct kmem_cache *s, struct slab *slab,
3041 void *freelist)
3042{
3043 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
3044 int free_delta = 0;
3045 void *nextfree, *freelist_iter, *freelist_tail;
3046 int tail = DEACTIVATE_TO_HEAD;
3047 unsigned long flags = 0;
3048 struct slab new;
3049 struct slab old;
3050
3051 if (READ_ONCE(slab->freelist)) {
3052 stat(s, DEACTIVATE_REMOTE_FREES);
3053 tail = DEACTIVATE_TO_TAIL;
3054 }
3055
3056 /*
3057 * Stage one: Count the objects on cpu's freelist as free_delta and
3058 * remember the last object in freelist_tail for later splicing.
3059 */
3060 freelist_tail = NULL;
3061 freelist_iter = freelist;
3062 while (freelist_iter) {
3063 nextfree = get_freepointer(s, freelist_iter);
3064
3065 /*
3066 * If 'nextfree' is invalid, it is possible that the object at
3067 * 'freelist_iter' is already corrupted. So isolate all objects
3068 * starting at 'freelist_iter' by skipping them.
3069 */
3070 if (freelist_corrupted(s, slab, &freelist_iter, nextfree))
3071 break;
3072
3073 freelist_tail = freelist_iter;
3074 free_delta++;
3075
3076 freelist_iter = nextfree;
3077 }
3078
3079 /*
3080 * Stage two: Unfreeze the slab while splicing the per-cpu
3081 * freelist to the head of slab's freelist.
3082 */
3083 do {
3084 old.freelist = READ_ONCE(slab->freelist);
3085 old.counters = READ_ONCE(slab->counters);
3086 VM_BUG_ON(!old.frozen);
3087
3088 /* Determine target state of the slab */
3089 new.counters = old.counters;
3090 new.frozen = 0;
3091 if (freelist_tail) {
3092 new.inuse -= free_delta;
3093 set_freepointer(s, freelist_tail, old.freelist);
3094 new.freelist = freelist;
3095 } else {
3096 new.freelist = old.freelist;
3097 }
3098 } while (!slab_update_freelist(s, slab,
3099 old.freelist, old.counters,
3100 new.freelist, new.counters,
3101 "unfreezing slab"));
3102
3103 /*
3104 * Stage three: Manipulate the slab list based on the updated state.
3105 */
3106 if (!new.inuse && n->nr_partial >= s->min_partial) {
3107 stat(s, DEACTIVATE_EMPTY);
3108 discard_slab(s, slab);
3109 stat(s, FREE_SLAB);
3110 } else if (new.freelist) {
3111 spin_lock_irqsave(&n->list_lock, flags);
3112 add_partial(n, slab, tail);
3113 spin_unlock_irqrestore(&n->list_lock, flags);
3114 stat(s, tail);
3115 } else {
3116 stat(s, DEACTIVATE_FULL);
3117 }
3118}
3119
3120#ifdef CONFIG_SLUB_CPU_PARTIAL
3121static void __put_partials(struct kmem_cache *s, struct slab *partial_slab)
3122{
3123 struct kmem_cache_node *n = NULL, *n2 = NULL;
3124 struct slab *slab, *slab_to_discard = NULL;
3125 unsigned long flags = 0;
3126
3127 while (partial_slab) {
3128 slab = partial_slab;
3129 partial_slab = slab->next;
3130
3131 n2 = get_node(s, slab_nid(slab));
3132 if (n != n2) {
3133 if (n)
3134 spin_unlock_irqrestore(&n->list_lock, flags);
3135
3136 n = n2;
3137 spin_lock_irqsave(&n->list_lock, flags);
3138 }
3139
3140 if (unlikely(!slab->inuse && n->nr_partial >= s->min_partial)) {
3141 slab->next = slab_to_discard;
3142 slab_to_discard = slab;
3143 } else {
3144 add_partial(n, slab, DEACTIVATE_TO_TAIL);
3145 stat(s, FREE_ADD_PARTIAL);
3146 }
3147 }
3148
3149 if (n)
3150 spin_unlock_irqrestore(&n->list_lock, flags);
3151
3152 while (slab_to_discard) {
3153 slab = slab_to_discard;
3154 slab_to_discard = slab_to_discard->next;
3155
3156 stat(s, DEACTIVATE_EMPTY);
3157 discard_slab(s, slab);
3158 stat(s, FREE_SLAB);
3159 }
3160}
3161
3162/*
3163 * Put all the cpu partial slabs to the node partial list.
3164 */
3165static void put_partials(struct kmem_cache *s)
3166{
3167 struct slab *partial_slab;
3168 unsigned long flags;
3169
3170 local_lock_irqsave(&s->cpu_slab->lock, flags);
3171 partial_slab = this_cpu_read(s->cpu_slab->partial);
3172 this_cpu_write(s->cpu_slab->partial, NULL);
3173 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3174
3175 if (partial_slab)
3176 __put_partials(s, partial_slab);
3177}
3178
3179static void put_partials_cpu(struct kmem_cache *s,
3180 struct kmem_cache_cpu *c)
3181{
3182 struct slab *partial_slab;
3183
3184 partial_slab = slub_percpu_partial(c);
3185 c->partial = NULL;
3186
3187 if (partial_slab)
3188 __put_partials(s, partial_slab);
3189}
3190
3191/*
3192 * Put a slab into a partial slab slot if available.
3193 *
3194 * If we did not find a slot then simply move all the partials to the
3195 * per node partial list.
3196 */
3197static void put_cpu_partial(struct kmem_cache *s, struct slab *slab, int drain)
3198{
3199 struct slab *oldslab;
3200 struct slab *slab_to_put = NULL;
3201 unsigned long flags;
3202 int slabs = 0;
3203
3204 local_lock_irqsave(&s->cpu_slab->lock, flags);
3205
3206 oldslab = this_cpu_read(s->cpu_slab->partial);
3207
3208 if (oldslab) {
3209 if (drain && oldslab->slabs >= s->cpu_partial_slabs) {
3210 /*
3211 * Partial array is full. Move the existing set to the
3212 * per node partial list. Postpone the actual unfreezing
3213 * outside of the critical section.
3214 */
3215 slab_to_put = oldslab;
3216 oldslab = NULL;
3217 } else {
3218 slabs = oldslab->slabs;
3219 }
3220 }
3221
3222 slabs++;
3223
3224 slab->slabs = slabs;
3225 slab->next = oldslab;
3226
3227 this_cpu_write(s->cpu_slab->partial, slab);
3228
3229 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3230
3231 if (slab_to_put) {
3232 __put_partials(s, slab_to_put);
3233 stat(s, CPU_PARTIAL_DRAIN);
3234 }
3235}
3236
3237#else /* CONFIG_SLUB_CPU_PARTIAL */
3238
3239static inline void put_partials(struct kmem_cache *s) { }
3240static inline void put_partials_cpu(struct kmem_cache *s,
3241 struct kmem_cache_cpu *c) { }
3242
3243#endif /* CONFIG_SLUB_CPU_PARTIAL */
3244
3245static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
3246{
3247 unsigned long flags;
3248 struct slab *slab;
3249 void *freelist;
3250
3251 local_lock_irqsave(&s->cpu_slab->lock, flags);
3252
3253 slab = c->slab;
3254 freelist = c->freelist;
3255
3256 c->slab = NULL;
3257 c->freelist = NULL;
3258 c->tid = next_tid(c->tid);
3259
3260 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3261
3262 if (slab) {
3263 deactivate_slab(s, slab, freelist);
3264 stat(s, CPUSLAB_FLUSH);
3265 }
3266}
3267
3268static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
3269{
3270 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3271 void *freelist = c->freelist;
3272 struct slab *slab = c->slab;
3273
3274 c->slab = NULL;
3275 c->freelist = NULL;
3276 c->tid = next_tid(c->tid);
3277
3278 if (slab) {
3279 deactivate_slab(s, slab, freelist);
3280 stat(s, CPUSLAB_FLUSH);
3281 }
3282
3283 put_partials_cpu(s, c);
3284}
3285
3286struct slub_flush_work {
3287 struct work_struct work;
3288 struct kmem_cache *s;
3289 bool skip;
3290};
3291
3292/*
3293 * Flush cpu slab.
3294 *
3295 * Called from CPU work handler with migration disabled.
3296 */
3297static void flush_cpu_slab(struct work_struct *w)
3298{
3299 struct kmem_cache *s;
3300 struct kmem_cache_cpu *c;
3301 struct slub_flush_work *sfw;
3302
3303 sfw = container_of(w, struct slub_flush_work, work);
3304
3305 s = sfw->s;
3306 c = this_cpu_ptr(s->cpu_slab);
3307
3308 if (c->slab)
3309 flush_slab(s, c);
3310
3311 put_partials(s);
3312}
3313
3314static bool has_cpu_slab(int cpu, struct kmem_cache *s)
3315{
3316 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
3317
3318 return c->slab || slub_percpu_partial(c);
3319}
3320
3321static DEFINE_MUTEX(flush_lock);
3322static DEFINE_PER_CPU(struct slub_flush_work, slub_flush);
3323
3324static void flush_all_cpus_locked(struct kmem_cache *s)
3325{
3326 struct slub_flush_work *sfw;
3327 unsigned int cpu;
3328
3329 lockdep_assert_cpus_held();
3330 mutex_lock(&flush_lock);
3331
3332 for_each_online_cpu(cpu) {
3333 sfw = &per_cpu(slub_flush, cpu);
3334 if (!has_cpu_slab(cpu, s)) {
3335 sfw->skip = true;
3336 continue;
3337 }
3338 INIT_WORK(&sfw->work, flush_cpu_slab);
3339 sfw->skip = false;
3340 sfw->s = s;
3341 queue_work_on(cpu, flushwq, &sfw->work);
3342 }
3343
3344 for_each_online_cpu(cpu) {
3345 sfw = &per_cpu(slub_flush, cpu);
3346 if (sfw->skip)
3347 continue;
3348 flush_work(&sfw->work);
3349 }
3350
3351 mutex_unlock(&flush_lock);
3352}
3353
3354static void flush_all(struct kmem_cache *s)
3355{
3356 cpus_read_lock();
3357 flush_all_cpus_locked(s);
3358 cpus_read_unlock();
3359}
3360
3361/*
3362 * Use the cpu notifier to insure that the cpu slabs are flushed when
3363 * necessary.
3364 */
3365static int slub_cpu_dead(unsigned int cpu)
3366{
3367 struct kmem_cache *s;
3368
3369 mutex_lock(&slab_mutex);
3370 list_for_each_entry(s, &slab_caches, list)
3371 __flush_cpu_slab(s, cpu);
3372 mutex_unlock(&slab_mutex);
3373 return 0;
3374}
3375
3376#else /* CONFIG_SLUB_TINY */
3377static inline void flush_all_cpus_locked(struct kmem_cache *s) { }
3378static inline void flush_all(struct kmem_cache *s) { }
3379static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu) { }
3380static inline int slub_cpu_dead(unsigned int cpu) { return 0; }
3381#endif /* CONFIG_SLUB_TINY */
3382
3383/*
3384 * Check if the objects in a per cpu structure fit numa
3385 * locality expectations.
3386 */
3387static inline int node_match(struct slab *slab, int node)
3388{
3389#ifdef CONFIG_NUMA
3390 if (node != NUMA_NO_NODE && slab_nid(slab) != node)
3391 return 0;
3392#endif
3393 return 1;
3394}
3395
3396#ifdef CONFIG_SLUB_DEBUG
3397static int count_free(struct slab *slab)
3398{
3399 return slab->objects - slab->inuse;
3400}
3401
3402static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
3403{
3404 return atomic_long_read(&n->total_objects);
3405}
3406
3407/* Supports checking bulk free of a constructed freelist */
3408static inline bool free_debug_processing(struct kmem_cache *s,
3409 struct slab *slab, void *head, void *tail, int *bulk_cnt,
3410 unsigned long addr, depot_stack_handle_t handle)
3411{
3412 bool checks_ok = false;
3413 void *object = head;
3414 int cnt = 0;
3415
3416 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3417 if (!check_slab(s, slab))
3418 goto out;
3419 }
3420
3421 if (slab->inuse < *bulk_cnt) {
3422 slab_err(s, slab, "Slab has %d allocated objects but %d are to be freed\n",
3423 slab->inuse, *bulk_cnt);
3424 goto out;
3425 }
3426
3427next_object:
3428
3429 if (++cnt > *bulk_cnt)
3430 goto out_cnt;
3431
3432 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
3433 if (!free_consistency_checks(s, slab, object, addr))
3434 goto out;
3435 }
3436
3437 if (s->flags & SLAB_STORE_USER)
3438 set_track_update(s, object, TRACK_FREE, addr, handle);
3439 trace(s, slab, object, 0);
3440 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
3441 init_object(s, object, SLUB_RED_INACTIVE);
3442
3443 /* Reached end of constructed freelist yet? */
3444 if (object != tail) {
3445 object = get_freepointer(s, object);
3446 goto next_object;
3447 }
3448 checks_ok = true;
3449
3450out_cnt:
3451 if (cnt != *bulk_cnt) {
3452 slab_err(s, slab, "Bulk free expected %d objects but found %d\n",
3453 *bulk_cnt, cnt);
3454 *bulk_cnt = cnt;
3455 }
3456
3457out:
3458
3459 if (!checks_ok)
3460 slab_fix(s, "Object at 0x%p not freed", object);
3461
3462 return checks_ok;
3463}
3464#endif /* CONFIG_SLUB_DEBUG */
3465
3466#if defined(CONFIG_SLUB_DEBUG) || defined(SLAB_SUPPORTS_SYSFS)
3467static unsigned long count_partial(struct kmem_cache_node *n,
3468 int (*get_count)(struct slab *))
3469{
3470 unsigned long flags;
3471 unsigned long x = 0;
3472 struct slab *slab;
3473
3474 spin_lock_irqsave(&n->list_lock, flags);
3475 list_for_each_entry(slab, &n->partial, slab_list)
3476 x += get_count(slab);
3477 spin_unlock_irqrestore(&n->list_lock, flags);
3478 return x;
3479}
3480#endif /* CONFIG_SLUB_DEBUG || SLAB_SUPPORTS_SYSFS */
3481
3482#ifdef CONFIG_SLUB_DEBUG
3483#define MAX_PARTIAL_TO_SCAN 10000
3484
3485static unsigned long count_partial_free_approx(struct kmem_cache_node *n)
3486{
3487 unsigned long flags;
3488 unsigned long x = 0;
3489 struct slab *slab;
3490
3491 spin_lock_irqsave(&n->list_lock, flags);
3492 if (n->nr_partial <= MAX_PARTIAL_TO_SCAN) {
3493 list_for_each_entry(slab, &n->partial, slab_list)
3494 x += slab->objects - slab->inuse;
3495 } else {
3496 /*
3497 * For a long list, approximate the total count of objects in
3498 * it to meet the limit on the number of slabs to scan.
3499 * Scan from both the list's head and tail for better accuracy.
3500 */
3501 unsigned long scanned = 0;
3502
3503 list_for_each_entry(slab, &n->partial, slab_list) {
3504 x += slab->objects - slab->inuse;
3505 if (++scanned == MAX_PARTIAL_TO_SCAN / 2)
3506 break;
3507 }
3508 list_for_each_entry_reverse(slab, &n->partial, slab_list) {
3509 x += slab->objects - slab->inuse;
3510 if (++scanned == MAX_PARTIAL_TO_SCAN)
3511 break;
3512 }
3513 x = mult_frac(x, n->nr_partial, scanned);
3514 x = min(x, node_nr_objs(n));
3515 }
3516 spin_unlock_irqrestore(&n->list_lock, flags);
3517 return x;
3518}
3519
3520static noinline void
3521slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
3522{
3523 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
3524 DEFAULT_RATELIMIT_BURST);
3525 int cpu = raw_smp_processor_id();
3526 int node;
3527 struct kmem_cache_node *n;
3528
3529 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
3530 return;
3531
3532 pr_warn("SLUB: Unable to allocate memory on CPU %u (of node %d) on node %d, gfp=%#x(%pGg)\n",
3533 cpu, cpu_to_node(cpu), nid, gfpflags, &gfpflags);
3534 pr_warn(" cache: %s, object size: %u, buffer size: %u, default order: %u, min order: %u\n",
3535 s->name, s->object_size, s->size, oo_order(s->oo),
3536 oo_order(s->min));
3537
3538 if (oo_order(s->min) > get_order(s->object_size))
3539 pr_warn(" %s debugging increased min order, use slab_debug=O to disable.\n",
3540 s->name);
3541
3542 for_each_kmem_cache_node(s, node, n) {
3543 unsigned long nr_slabs;
3544 unsigned long nr_objs;
3545 unsigned long nr_free;
3546
3547 nr_free = count_partial_free_approx(n);
3548 nr_slabs = node_nr_slabs(n);
3549 nr_objs = node_nr_objs(n);
3550
3551 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
3552 node, nr_slabs, nr_objs, nr_free);
3553 }
3554}
3555#else /* CONFIG_SLUB_DEBUG */
3556static inline void
3557slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid) { }
3558#endif
3559
3560static inline bool pfmemalloc_match(struct slab *slab, gfp_t gfpflags)
3561{
3562 if (unlikely(slab_test_pfmemalloc(slab)))
3563 return gfp_pfmemalloc_allowed(gfpflags);
3564
3565 return true;
3566}
3567
3568#ifndef CONFIG_SLUB_TINY
3569static inline bool
3570__update_cpu_freelist_fast(struct kmem_cache *s,
3571 void *freelist_old, void *freelist_new,
3572 unsigned long tid)
3573{
3574 freelist_aba_t old = { .freelist = freelist_old, .counter = tid };
3575 freelist_aba_t new = { .freelist = freelist_new, .counter = next_tid(tid) };
3576
3577 return this_cpu_try_cmpxchg_freelist(s->cpu_slab->freelist_tid.full,
3578 &old.full, new.full);
3579}
3580
3581/*
3582 * Check the slab->freelist and either transfer the freelist to the
3583 * per cpu freelist or deactivate the slab.
3584 *
3585 * The slab is still frozen if the return value is not NULL.
3586 *
3587 * If this function returns NULL then the slab has been unfrozen.
3588 */
3589static inline void *get_freelist(struct kmem_cache *s, struct slab *slab)
3590{
3591 struct slab new;
3592 unsigned long counters;
3593 void *freelist;
3594
3595 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3596
3597 do {
3598 freelist = slab->freelist;
3599 counters = slab->counters;
3600
3601 new.counters = counters;
3602
3603 new.inuse = slab->objects;
3604 new.frozen = freelist != NULL;
3605
3606 } while (!__slab_update_freelist(s, slab,
3607 freelist, counters,
3608 NULL, new.counters,
3609 "get_freelist"));
3610
3611 return freelist;
3612}
3613
3614/*
3615 * Freeze the partial slab and return the pointer to the freelist.
3616 */
3617static inline void *freeze_slab(struct kmem_cache *s, struct slab *slab)
3618{
3619 struct slab new;
3620 unsigned long counters;
3621 void *freelist;
3622
3623 do {
3624 freelist = slab->freelist;
3625 counters = slab->counters;
3626
3627 new.counters = counters;
3628 VM_BUG_ON(new.frozen);
3629
3630 new.inuse = slab->objects;
3631 new.frozen = 1;
3632
3633 } while (!slab_update_freelist(s, slab,
3634 freelist, counters,
3635 NULL, new.counters,
3636 "freeze_slab"));
3637
3638 return freelist;
3639}
3640
3641/*
3642 * Slow path. The lockless freelist is empty or we need to perform
3643 * debugging duties.
3644 *
3645 * Processing is still very fast if new objects have been freed to the
3646 * regular freelist. In that case we simply take over the regular freelist
3647 * as the lockless freelist and zap the regular freelist.
3648 *
3649 * If that is not working then we fall back to the partial lists. We take the
3650 * first element of the freelist as the object to allocate now and move the
3651 * rest of the freelist to the lockless freelist.
3652 *
3653 * And if we were unable to get a new slab from the partial slab lists then
3654 * we need to allocate a new slab. This is the slowest path since it involves
3655 * a call to the page allocator and the setup of a new slab.
3656 *
3657 * Version of __slab_alloc to use when we know that preemption is
3658 * already disabled (which is the case for bulk allocation).
3659 */
3660static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3661 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3662{
3663 void *freelist;
3664 struct slab *slab;
3665 unsigned long flags;
3666 struct partial_context pc;
3667 bool try_thisnode = true;
3668
3669 stat(s, ALLOC_SLOWPATH);
3670
3671reread_slab:
3672
3673 slab = READ_ONCE(c->slab);
3674 if (!slab) {
3675 /*
3676 * if the node is not online or has no normal memory, just
3677 * ignore the node constraint
3678 */
3679 if (unlikely(node != NUMA_NO_NODE &&
3680 !node_isset(node, slab_nodes)))
3681 node = NUMA_NO_NODE;
3682 goto new_slab;
3683 }
3684
3685 if (unlikely(!node_match(slab, node))) {
3686 /*
3687 * same as above but node_match() being false already
3688 * implies node != NUMA_NO_NODE
3689 */
3690 if (!node_isset(node, slab_nodes)) {
3691 node = NUMA_NO_NODE;
3692 } else {
3693 stat(s, ALLOC_NODE_MISMATCH);
3694 goto deactivate_slab;
3695 }
3696 }
3697
3698 /*
3699 * By rights, we should be searching for a slab page that was
3700 * PFMEMALLOC but right now, we are losing the pfmemalloc
3701 * information when the page leaves the per-cpu allocator
3702 */
3703 if (unlikely(!pfmemalloc_match(slab, gfpflags)))
3704 goto deactivate_slab;
3705
3706 /* must check again c->slab in case we got preempted and it changed */
3707 local_lock_irqsave(&s->cpu_slab->lock, flags);
3708 if (unlikely(slab != c->slab)) {
3709 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3710 goto reread_slab;
3711 }
3712 freelist = c->freelist;
3713 if (freelist)
3714 goto load_freelist;
3715
3716 freelist = get_freelist(s, slab);
3717
3718 if (!freelist) {
3719 c->slab = NULL;
3720 c->tid = next_tid(c->tid);
3721 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3722 stat(s, DEACTIVATE_BYPASS);
3723 goto new_slab;
3724 }
3725
3726 stat(s, ALLOC_REFILL);
3727
3728load_freelist:
3729
3730 lockdep_assert_held(this_cpu_ptr(&s->cpu_slab->lock));
3731
3732 /*
3733 * freelist is pointing to the list of objects to be used.
3734 * slab is pointing to the slab from which the objects are obtained.
3735 * That slab must be frozen for per cpu allocations to work.
3736 */
3737 VM_BUG_ON(!c->slab->frozen);
3738 c->freelist = get_freepointer(s, freelist);
3739 c->tid = next_tid(c->tid);
3740 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3741 return freelist;
3742
3743deactivate_slab:
3744
3745 local_lock_irqsave(&s->cpu_slab->lock, flags);
3746 if (slab != c->slab) {
3747 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3748 goto reread_slab;
3749 }
3750 freelist = c->freelist;
3751 c->slab = NULL;
3752 c->freelist = NULL;
3753 c->tid = next_tid(c->tid);
3754 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3755 deactivate_slab(s, slab, freelist);
3756
3757new_slab:
3758
3759#ifdef CONFIG_SLUB_CPU_PARTIAL
3760 while (slub_percpu_partial(c)) {
3761 local_lock_irqsave(&s->cpu_slab->lock, flags);
3762 if (unlikely(c->slab)) {
3763 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3764 goto reread_slab;
3765 }
3766 if (unlikely(!slub_percpu_partial(c))) {
3767 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3768 /* we were preempted and partial list got empty */
3769 goto new_objects;
3770 }
3771
3772 slab = slub_percpu_partial(c);
3773 slub_set_percpu_partial(c, slab);
3774
3775 if (likely(node_match(slab, node) &&
3776 pfmemalloc_match(slab, gfpflags))) {
3777 c->slab = slab;
3778 freelist = get_freelist(s, slab);
3779 VM_BUG_ON(!freelist);
3780 stat(s, CPU_PARTIAL_ALLOC);
3781 goto load_freelist;
3782 }
3783
3784 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3785
3786 slab->next = NULL;
3787 __put_partials(s, slab);
3788 }
3789#endif
3790
3791new_objects:
3792
3793 pc.flags = gfpflags;
3794 /*
3795 * When a preferred node is indicated but no __GFP_THISNODE
3796 *
3797 * 1) try to get a partial slab from target node only by having
3798 * __GFP_THISNODE in pc.flags for get_partial()
3799 * 2) if 1) failed, try to allocate a new slab from target node with
3800 * GPF_NOWAIT | __GFP_THISNODE opportunistically
3801 * 3) if 2) failed, retry with original gfpflags which will allow
3802 * get_partial() try partial lists of other nodes before potentially
3803 * allocating new page from other nodes
3804 */
3805 if (unlikely(node != NUMA_NO_NODE && !(gfpflags & __GFP_THISNODE)
3806 && try_thisnode))
3807 pc.flags = GFP_NOWAIT | __GFP_THISNODE;
3808
3809 pc.orig_size = orig_size;
3810 slab = get_partial(s, node, &pc);
3811 if (slab) {
3812 if (kmem_cache_debug(s)) {
3813 freelist = pc.object;
3814 /*
3815 * For debug caches here we had to go through
3816 * alloc_single_from_partial() so just store the
3817 * tracking info and return the object.
3818 */
3819 if (s->flags & SLAB_STORE_USER)
3820 set_track(s, freelist, TRACK_ALLOC, addr);
3821
3822 return freelist;
3823 }
3824
3825 freelist = freeze_slab(s, slab);
3826 goto retry_load_slab;
3827 }
3828
3829 slub_put_cpu_ptr(s->cpu_slab);
3830 slab = new_slab(s, pc.flags, node);
3831 c = slub_get_cpu_ptr(s->cpu_slab);
3832
3833 if (unlikely(!slab)) {
3834 if (node != NUMA_NO_NODE && !(gfpflags & __GFP_THISNODE)
3835 && try_thisnode) {
3836 try_thisnode = false;
3837 goto new_objects;
3838 }
3839 slab_out_of_memory(s, gfpflags, node);
3840 return NULL;
3841 }
3842
3843 stat(s, ALLOC_SLAB);
3844
3845 if (kmem_cache_debug(s)) {
3846 freelist = alloc_single_from_new_slab(s, slab, orig_size);
3847
3848 if (unlikely(!freelist))
3849 goto new_objects;
3850
3851 if (s->flags & SLAB_STORE_USER)
3852 set_track(s, freelist, TRACK_ALLOC, addr);
3853
3854 return freelist;
3855 }
3856
3857 /*
3858 * No other reference to the slab yet so we can
3859 * muck around with it freely without cmpxchg
3860 */
3861 freelist = slab->freelist;
3862 slab->freelist = NULL;
3863 slab->inuse = slab->objects;
3864 slab->frozen = 1;
3865
3866 inc_slabs_node(s, slab_nid(slab), slab->objects);
3867
3868 if (unlikely(!pfmemalloc_match(slab, gfpflags))) {
3869 /*
3870 * For !pfmemalloc_match() case we don't load freelist so that
3871 * we don't make further mismatched allocations easier.
3872 */
3873 deactivate_slab(s, slab, get_freepointer(s, freelist));
3874 return freelist;
3875 }
3876
3877retry_load_slab:
3878
3879 local_lock_irqsave(&s->cpu_slab->lock, flags);
3880 if (unlikely(c->slab)) {
3881 void *flush_freelist = c->freelist;
3882 struct slab *flush_slab = c->slab;
3883
3884 c->slab = NULL;
3885 c->freelist = NULL;
3886 c->tid = next_tid(c->tid);
3887
3888 local_unlock_irqrestore(&s->cpu_slab->lock, flags);
3889
3890 deactivate_slab(s, flush_slab, flush_freelist);
3891
3892 stat(s, CPUSLAB_FLUSH);
3893
3894 goto retry_load_slab;
3895 }
3896 c->slab = slab;
3897
3898 goto load_freelist;
3899}
3900
3901/*
3902 * A wrapper for ___slab_alloc() for contexts where preemption is not yet
3903 * disabled. Compensates for possible cpu changes by refetching the per cpu area
3904 * pointer.
3905 */
3906static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
3907 unsigned long addr, struct kmem_cache_cpu *c, unsigned int orig_size)
3908{
3909 void *p;
3910
3911#ifdef CONFIG_PREEMPT_COUNT
3912 /*
3913 * We may have been preempted and rescheduled on a different
3914 * cpu before disabling preemption. Need to reload cpu area
3915 * pointer.
3916 */
3917 c = slub_get_cpu_ptr(s->cpu_slab);
3918#endif
3919
3920 p = ___slab_alloc(s, gfpflags, node, addr, c, orig_size);
3921#ifdef CONFIG_PREEMPT_COUNT
3922 slub_put_cpu_ptr(s->cpu_slab);
3923#endif
3924 return p;
3925}
3926
3927static __always_inline void *__slab_alloc_node(struct kmem_cache *s,
3928 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
3929{
3930 struct kmem_cache_cpu *c;
3931 struct slab *slab;
3932 unsigned long tid;
3933 void *object;
3934
3935redo:
3936 /*
3937 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
3938 * enabled. We may switch back and forth between cpus while
3939 * reading from one cpu area. That does not matter as long
3940 * as we end up on the original cpu again when doing the cmpxchg.
3941 *
3942 * We must guarantee that tid and kmem_cache_cpu are retrieved on the
3943 * same cpu. We read first the kmem_cache_cpu pointer and use it to read
3944 * the tid. If we are preempted and switched to another cpu between the
3945 * two reads, it's OK as the two are still associated with the same cpu
3946 * and cmpxchg later will validate the cpu.
3947 */
3948 c = raw_cpu_ptr(s->cpu_slab);
3949 tid = READ_ONCE(c->tid);
3950
3951 /*
3952 * Irqless object alloc/free algorithm used here depends on sequence
3953 * of fetching cpu_slab's data. tid should be fetched before anything
3954 * on c to guarantee that object and slab associated with previous tid
3955 * won't be used with current tid. If we fetch tid first, object and
3956 * slab could be one associated with next tid and our alloc/free
3957 * request will be failed. In this case, we will retry. So, no problem.
3958 */
3959 barrier();
3960
3961 /*
3962 * The transaction ids are globally unique per cpu and per operation on
3963 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
3964 * occurs on the right processor and that there was no operation on the
3965 * linked list in between.
3966 */
3967
3968 object = c->freelist;
3969 slab = c->slab;
3970
3971#ifdef CONFIG_NUMA
3972 if (static_branch_unlikely(&strict_numa) &&
3973 node == NUMA_NO_NODE) {
3974
3975 struct mempolicy *mpol = current->mempolicy;
3976
3977 if (mpol) {
3978 /*
3979 * Special BIND rule support. If existing slab
3980 * is in permitted set then do not redirect
3981 * to a particular node.
3982 * Otherwise we apply the memory policy to get
3983 * the node we need to allocate on.
3984 */
3985 if (mpol->mode != MPOL_BIND || !slab ||
3986 !node_isset(slab_nid(slab), mpol->nodes))
3987
3988 node = mempolicy_slab_node();
3989 }
3990 }
3991#endif
3992
3993 if (!USE_LOCKLESS_FAST_PATH() ||
3994 unlikely(!object || !slab || !node_match(slab, node))) {
3995 object = __slab_alloc(s, gfpflags, node, addr, c, orig_size);
3996 } else {
3997 void *next_object = get_freepointer_safe(s, object);
3998
3999 /*
4000 * The cmpxchg will only match if there was no additional
4001 * operation and if we are on the right processor.
4002 *
4003 * The cmpxchg does the following atomically (without lock
4004 * semantics!)
4005 * 1. Relocate first pointer to the current per cpu area.
4006 * 2. Verify that tid and freelist have not been changed
4007 * 3. If they were not changed replace tid and freelist
4008 *
4009 * Since this is without lock semantics the protection is only
4010 * against code executing on this cpu *not* from access by
4011 * other cpus.
4012 */
4013 if (unlikely(!__update_cpu_freelist_fast(s, object, next_object, tid))) {
4014 note_cmpxchg_failure("slab_alloc", s, tid);
4015 goto redo;
4016 }
4017 prefetch_freepointer(s, next_object);
4018 stat(s, ALLOC_FASTPATH);
4019 }
4020
4021 return object;
4022}
4023#else /* CONFIG_SLUB_TINY */
4024static void *__slab_alloc_node(struct kmem_cache *s,
4025 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
4026{
4027 struct partial_context pc;
4028 struct slab *slab;
4029 void *object;
4030
4031 pc.flags = gfpflags;
4032 pc.orig_size = orig_size;
4033 slab = get_partial(s, node, &pc);
4034
4035 if (slab)
4036 return pc.object;
4037
4038 slab = new_slab(s, gfpflags, node);
4039 if (unlikely(!slab)) {
4040 slab_out_of_memory(s, gfpflags, node);
4041 return NULL;
4042 }
4043
4044 object = alloc_single_from_new_slab(s, slab, orig_size);
4045
4046 return object;
4047}
4048#endif /* CONFIG_SLUB_TINY */
4049
4050/*
4051 * If the object has been wiped upon free, make sure it's fully initialized by
4052 * zeroing out freelist pointer.
4053 *
4054 * Note that we also wipe custom freelist pointers.
4055 */
4056static __always_inline void maybe_wipe_obj_freeptr(struct kmem_cache *s,
4057 void *obj)
4058{
4059 if (unlikely(slab_want_init_on_free(s)) && obj &&
4060 !freeptr_outside_object(s))
4061 memset((void *)((char *)kasan_reset_tag(obj) + s->offset),
4062 0, sizeof(void *));
4063}
4064
4065static __fastpath_inline
4066struct kmem_cache *slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
4067{
4068 flags &= gfp_allowed_mask;
4069
4070 might_alloc(flags);
4071
4072 if (unlikely(should_failslab(s, flags)))
4073 return NULL;
4074
4075 return s;
4076}
4077
4078static __fastpath_inline
4079bool slab_post_alloc_hook(struct kmem_cache *s, struct list_lru *lru,
4080 gfp_t flags, size_t size, void **p, bool init,
4081 unsigned int orig_size)
4082{
4083 unsigned int zero_size = s->object_size;
4084 bool kasan_init = init;
4085 size_t i;
4086 gfp_t init_flags = flags & gfp_allowed_mask;
4087
4088 /*
4089 * For kmalloc object, the allocated memory size(object_size) is likely
4090 * larger than the requested size(orig_size). If redzone check is
4091 * enabled for the extra space, don't zero it, as it will be redzoned
4092 * soon. The redzone operation for this extra space could be seen as a
4093 * replacement of current poisoning under certain debug option, and
4094 * won't break other sanity checks.
4095 */
4096 if (kmem_cache_debug_flags(s, SLAB_STORE_USER | SLAB_RED_ZONE) &&
4097 (s->flags & SLAB_KMALLOC))
4098 zero_size = orig_size;
4099
4100 /*
4101 * When slab_debug is enabled, avoid memory initialization integrated
4102 * into KASAN and instead zero out the memory via the memset below with
4103 * the proper size. Otherwise, KASAN might overwrite SLUB redzones and
4104 * cause false-positive reports. This does not lead to a performance
4105 * penalty on production builds, as slab_debug is not intended to be
4106 * enabled there.
4107 */
4108 if (__slub_debug_enabled())
4109 kasan_init = false;
4110
4111 /*
4112 * As memory initialization might be integrated into KASAN,
4113 * kasan_slab_alloc and initialization memset must be
4114 * kept together to avoid discrepancies in behavior.
4115 *
4116 * As p[i] might get tagged, memset and kmemleak hook come after KASAN.
4117 */
4118 for (i = 0; i < size; i++) {
4119 p[i] = kasan_slab_alloc(s, p[i], init_flags, kasan_init);
4120 if (p[i] && init && (!kasan_init ||
4121 !kasan_has_integrated_init()))
4122 memset(p[i], 0, zero_size);
4123 kmemleak_alloc_recursive(p[i], s->object_size, 1,
4124 s->flags, init_flags);
4125 kmsan_slab_alloc(s, p[i], init_flags);
4126 alloc_tagging_slab_alloc_hook(s, p[i], flags);
4127 }
4128
4129 return memcg_slab_post_alloc_hook(s, lru, flags, size, p);
4130}
4131
4132/*
4133 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
4134 * have the fastpath folded into their functions. So no function call
4135 * overhead for requests that can be satisfied on the fastpath.
4136 *
4137 * The fastpath works by first checking if the lockless freelist can be used.
4138 * If not then __slab_alloc is called for slow processing.
4139 *
4140 * Otherwise we can simply pick the next object from the lockless free list.
4141 */
4142static __fastpath_inline void *slab_alloc_node(struct kmem_cache *s, struct list_lru *lru,
4143 gfp_t gfpflags, int node, unsigned long addr, size_t orig_size)
4144{
4145 void *object;
4146 bool init = false;
4147
4148 s = slab_pre_alloc_hook(s, gfpflags);
4149 if (unlikely(!s))
4150 return NULL;
4151
4152 object = kfence_alloc(s, orig_size, gfpflags);
4153 if (unlikely(object))
4154 goto out;
4155
4156 object = __slab_alloc_node(s, gfpflags, node, addr, orig_size);
4157
4158 maybe_wipe_obj_freeptr(s, object);
4159 init = slab_want_init_on_alloc(gfpflags, s);
4160
4161out:
4162 /*
4163 * When init equals 'true', like for kzalloc() family, only
4164 * @orig_size bytes might be zeroed instead of s->object_size
4165 * In case this fails due to memcg_slab_post_alloc_hook(),
4166 * object is set to NULL
4167 */
4168 slab_post_alloc_hook(s, lru, gfpflags, 1, &object, init, orig_size);
4169
4170 return object;
4171}
4172
4173void *kmem_cache_alloc_noprof(struct kmem_cache *s, gfp_t gfpflags)
4174{
4175 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE, _RET_IP_,
4176 s->object_size);
4177
4178 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
4179
4180 return ret;
4181}
4182EXPORT_SYMBOL(kmem_cache_alloc_noprof);
4183
4184void *kmem_cache_alloc_lru_noprof(struct kmem_cache *s, struct list_lru *lru,
4185 gfp_t gfpflags)
4186{
4187 void *ret = slab_alloc_node(s, lru, gfpflags, NUMA_NO_NODE, _RET_IP_,
4188 s->object_size);
4189
4190 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, NUMA_NO_NODE);
4191
4192 return ret;
4193}
4194EXPORT_SYMBOL(kmem_cache_alloc_lru_noprof);
4195
4196bool kmem_cache_charge(void *objp, gfp_t gfpflags)
4197{
4198 if (!memcg_kmem_online())
4199 return true;
4200
4201 return memcg_slab_post_charge(objp, gfpflags);
4202}
4203EXPORT_SYMBOL(kmem_cache_charge);
4204
4205/**
4206 * kmem_cache_alloc_node - Allocate an object on the specified node
4207 * @s: The cache to allocate from.
4208 * @gfpflags: See kmalloc().
4209 * @node: node number of the target node.
4210 *
4211 * Identical to kmem_cache_alloc but it will allocate memory on the given
4212 * node, which can improve the performance for cpu bound structures.
4213 *
4214 * Fallback to other node is possible if __GFP_THISNODE is not set.
4215 *
4216 * Return: pointer to the new object or %NULL in case of error
4217 */
4218void *kmem_cache_alloc_node_noprof(struct kmem_cache *s, gfp_t gfpflags, int node)
4219{
4220 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, s->object_size);
4221
4222 trace_kmem_cache_alloc(_RET_IP_, ret, s, gfpflags, node);
4223
4224 return ret;
4225}
4226EXPORT_SYMBOL(kmem_cache_alloc_node_noprof);
4227
4228/*
4229 * To avoid unnecessary overhead, we pass through large allocation requests
4230 * directly to the page allocator. We use __GFP_COMP, because we will need to
4231 * know the allocation order to free the pages properly in kfree.
4232 */
4233static void *___kmalloc_large_node(size_t size, gfp_t flags, int node)
4234{
4235 struct folio *folio;
4236 void *ptr = NULL;
4237 unsigned int order = get_order(size);
4238
4239 if (unlikely(flags & GFP_SLAB_BUG_MASK))
4240 flags = kmalloc_fix_flags(flags);
4241
4242 flags |= __GFP_COMP;
4243 folio = (struct folio *)alloc_pages_node_noprof(node, flags, order);
4244 if (folio) {
4245 ptr = folio_address(folio);
4246 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4247 PAGE_SIZE << order);
4248 }
4249
4250 ptr = kasan_kmalloc_large(ptr, size, flags);
4251 /* As ptr might get tagged, call kmemleak hook after KASAN. */
4252 kmemleak_alloc(ptr, size, 1, flags);
4253 kmsan_kmalloc_large(ptr, size, flags);
4254
4255 return ptr;
4256}
4257
4258void *__kmalloc_large_noprof(size_t size, gfp_t flags)
4259{
4260 void *ret = ___kmalloc_large_node(size, flags, NUMA_NO_NODE);
4261
4262 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
4263 flags, NUMA_NO_NODE);
4264 return ret;
4265}
4266EXPORT_SYMBOL(__kmalloc_large_noprof);
4267
4268void *__kmalloc_large_node_noprof(size_t size, gfp_t flags, int node)
4269{
4270 void *ret = ___kmalloc_large_node(size, flags, node);
4271
4272 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << get_order(size),
4273 flags, node);
4274 return ret;
4275}
4276EXPORT_SYMBOL(__kmalloc_large_node_noprof);
4277
4278static __always_inline
4279void *__do_kmalloc_node(size_t size, kmem_buckets *b, gfp_t flags, int node,
4280 unsigned long caller)
4281{
4282 struct kmem_cache *s;
4283 void *ret;
4284
4285 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4286 ret = __kmalloc_large_node_noprof(size, flags, node);
4287 trace_kmalloc(caller, ret, size,
4288 PAGE_SIZE << get_order(size), flags, node);
4289 return ret;
4290 }
4291
4292 if (unlikely(!size))
4293 return ZERO_SIZE_PTR;
4294
4295 s = kmalloc_slab(size, b, flags, caller);
4296
4297 ret = slab_alloc_node(s, NULL, flags, node, caller, size);
4298 ret = kasan_kmalloc(s, ret, size, flags);
4299 trace_kmalloc(caller, ret, size, s->size, flags, node);
4300 return ret;
4301}
4302void *__kmalloc_node_noprof(DECL_BUCKET_PARAMS(size, b), gfp_t flags, int node)
4303{
4304 return __do_kmalloc_node(size, PASS_BUCKET_PARAM(b), flags, node, _RET_IP_);
4305}
4306EXPORT_SYMBOL(__kmalloc_node_noprof);
4307
4308void *__kmalloc_noprof(size_t size, gfp_t flags)
4309{
4310 return __do_kmalloc_node(size, NULL, flags, NUMA_NO_NODE, _RET_IP_);
4311}
4312EXPORT_SYMBOL(__kmalloc_noprof);
4313
4314void *__kmalloc_node_track_caller_noprof(DECL_BUCKET_PARAMS(size, b), gfp_t flags,
4315 int node, unsigned long caller)
4316{
4317 return __do_kmalloc_node(size, PASS_BUCKET_PARAM(b), flags, node, caller);
4318
4319}
4320EXPORT_SYMBOL(__kmalloc_node_track_caller_noprof);
4321
4322void *__kmalloc_cache_noprof(struct kmem_cache *s, gfp_t gfpflags, size_t size)
4323{
4324 void *ret = slab_alloc_node(s, NULL, gfpflags, NUMA_NO_NODE,
4325 _RET_IP_, size);
4326
4327 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, NUMA_NO_NODE);
4328
4329 ret = kasan_kmalloc(s, ret, size, gfpflags);
4330 return ret;
4331}
4332EXPORT_SYMBOL(__kmalloc_cache_noprof);
4333
4334void *__kmalloc_cache_node_noprof(struct kmem_cache *s, gfp_t gfpflags,
4335 int node, size_t size)
4336{
4337 void *ret = slab_alloc_node(s, NULL, gfpflags, node, _RET_IP_, size);
4338
4339 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags, node);
4340
4341 ret = kasan_kmalloc(s, ret, size, gfpflags);
4342 return ret;
4343}
4344EXPORT_SYMBOL(__kmalloc_cache_node_noprof);
4345
4346static noinline void free_to_partial_list(
4347 struct kmem_cache *s, struct slab *slab,
4348 void *head, void *tail, int bulk_cnt,
4349 unsigned long addr)
4350{
4351 struct kmem_cache_node *n = get_node(s, slab_nid(slab));
4352 struct slab *slab_free = NULL;
4353 int cnt = bulk_cnt;
4354 unsigned long flags;
4355 depot_stack_handle_t handle = 0;
4356
4357 if (s->flags & SLAB_STORE_USER)
4358 handle = set_track_prepare();
4359
4360 spin_lock_irqsave(&n->list_lock, flags);
4361
4362 if (free_debug_processing(s, slab, head, tail, &cnt, addr, handle)) {
4363 void *prior = slab->freelist;
4364
4365 /* Perform the actual freeing while we still hold the locks */
4366 slab->inuse -= cnt;
4367 set_freepointer(s, tail, prior);
4368 slab->freelist = head;
4369
4370 /*
4371 * If the slab is empty, and node's partial list is full,
4372 * it should be discarded anyway no matter it's on full or
4373 * partial list.
4374 */
4375 if (slab->inuse == 0 && n->nr_partial >= s->min_partial)
4376 slab_free = slab;
4377
4378 if (!prior) {
4379 /* was on full list */
4380 remove_full(s, n, slab);
4381 if (!slab_free) {
4382 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4383 stat(s, FREE_ADD_PARTIAL);
4384 }
4385 } else if (slab_free) {
4386 remove_partial(n, slab);
4387 stat(s, FREE_REMOVE_PARTIAL);
4388 }
4389 }
4390
4391 if (slab_free) {
4392 /*
4393 * Update the counters while still holding n->list_lock to
4394 * prevent spurious validation warnings
4395 */
4396 dec_slabs_node(s, slab_nid(slab_free), slab_free->objects);
4397 }
4398
4399 spin_unlock_irqrestore(&n->list_lock, flags);
4400
4401 if (slab_free) {
4402 stat(s, FREE_SLAB);
4403 free_slab(s, slab_free);
4404 }
4405}
4406
4407/*
4408 * Slow path handling. This may still be called frequently since objects
4409 * have a longer lifetime than the cpu slabs in most processing loads.
4410 *
4411 * So we still attempt to reduce cache line usage. Just take the slab
4412 * lock and free the item. If there is no additional partial slab
4413 * handling required then we can return immediately.
4414 */
4415static void __slab_free(struct kmem_cache *s, struct slab *slab,
4416 void *head, void *tail, int cnt,
4417 unsigned long addr)
4418
4419{
4420 void *prior;
4421 int was_frozen;
4422 struct slab new;
4423 unsigned long counters;
4424 struct kmem_cache_node *n = NULL;
4425 unsigned long flags;
4426 bool on_node_partial;
4427
4428 stat(s, FREE_SLOWPATH);
4429
4430 if (IS_ENABLED(CONFIG_SLUB_TINY) || kmem_cache_debug(s)) {
4431 free_to_partial_list(s, slab, head, tail, cnt, addr);
4432 return;
4433 }
4434
4435 do {
4436 if (unlikely(n)) {
4437 spin_unlock_irqrestore(&n->list_lock, flags);
4438 n = NULL;
4439 }
4440 prior = slab->freelist;
4441 counters = slab->counters;
4442 set_freepointer(s, tail, prior);
4443 new.counters = counters;
4444 was_frozen = new.frozen;
4445 new.inuse -= cnt;
4446 if ((!new.inuse || !prior) && !was_frozen) {
4447 /* Needs to be taken off a list */
4448 if (!kmem_cache_has_cpu_partial(s) || prior) {
4449
4450 n = get_node(s, slab_nid(slab));
4451 /*
4452 * Speculatively acquire the list_lock.
4453 * If the cmpxchg does not succeed then we may
4454 * drop the list_lock without any processing.
4455 *
4456 * Otherwise the list_lock will synchronize with
4457 * other processors updating the list of slabs.
4458 */
4459 spin_lock_irqsave(&n->list_lock, flags);
4460
4461 on_node_partial = slab_test_node_partial(slab);
4462 }
4463 }
4464
4465 } while (!slab_update_freelist(s, slab,
4466 prior, counters,
4467 head, new.counters,
4468 "__slab_free"));
4469
4470 if (likely(!n)) {
4471
4472 if (likely(was_frozen)) {
4473 /*
4474 * The list lock was not taken therefore no list
4475 * activity can be necessary.
4476 */
4477 stat(s, FREE_FROZEN);
4478 } else if (kmem_cache_has_cpu_partial(s) && !prior) {
4479 /*
4480 * If we started with a full slab then put it onto the
4481 * per cpu partial list.
4482 */
4483 put_cpu_partial(s, slab, 1);
4484 stat(s, CPU_PARTIAL_FREE);
4485 }
4486
4487 return;
4488 }
4489
4490 /*
4491 * This slab was partially empty but not on the per-node partial list,
4492 * in which case we shouldn't manipulate its list, just return.
4493 */
4494 if (prior && !on_node_partial) {
4495 spin_unlock_irqrestore(&n->list_lock, flags);
4496 return;
4497 }
4498
4499 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
4500 goto slab_empty;
4501
4502 /*
4503 * Objects left in the slab. If it was not on the partial list before
4504 * then add it.
4505 */
4506 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
4507 add_partial(n, slab, DEACTIVATE_TO_TAIL);
4508 stat(s, FREE_ADD_PARTIAL);
4509 }
4510 spin_unlock_irqrestore(&n->list_lock, flags);
4511 return;
4512
4513slab_empty:
4514 if (prior) {
4515 /*
4516 * Slab on the partial list.
4517 */
4518 remove_partial(n, slab);
4519 stat(s, FREE_REMOVE_PARTIAL);
4520 }
4521
4522 spin_unlock_irqrestore(&n->list_lock, flags);
4523 stat(s, FREE_SLAB);
4524 discard_slab(s, slab);
4525}
4526
4527#ifndef CONFIG_SLUB_TINY
4528/*
4529 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
4530 * can perform fastpath freeing without additional function calls.
4531 *
4532 * The fastpath is only possible if we are freeing to the current cpu slab
4533 * of this processor. This typically the case if we have just allocated
4534 * the item before.
4535 *
4536 * If fastpath is not possible then fall back to __slab_free where we deal
4537 * with all sorts of special processing.
4538 *
4539 * Bulk free of a freelist with several objects (all pointing to the
4540 * same slab) possible by specifying head and tail ptr, plus objects
4541 * count (cnt). Bulk free indicated by tail pointer being set.
4542 */
4543static __always_inline void do_slab_free(struct kmem_cache *s,
4544 struct slab *slab, void *head, void *tail,
4545 int cnt, unsigned long addr)
4546{
4547 struct kmem_cache_cpu *c;
4548 unsigned long tid;
4549 void **freelist;
4550
4551redo:
4552 /*
4553 * Determine the currently cpus per cpu slab.
4554 * The cpu may change afterward. However that does not matter since
4555 * data is retrieved via this pointer. If we are on the same cpu
4556 * during the cmpxchg then the free will succeed.
4557 */
4558 c = raw_cpu_ptr(s->cpu_slab);
4559 tid = READ_ONCE(c->tid);
4560
4561 /* Same with comment on barrier() in __slab_alloc_node() */
4562 barrier();
4563
4564 if (unlikely(slab != c->slab)) {
4565 __slab_free(s, slab, head, tail, cnt, addr);
4566 return;
4567 }
4568
4569 if (USE_LOCKLESS_FAST_PATH()) {
4570 freelist = READ_ONCE(c->freelist);
4571
4572 set_freepointer(s, tail, freelist);
4573
4574 if (unlikely(!__update_cpu_freelist_fast(s, freelist, head, tid))) {
4575 note_cmpxchg_failure("slab_free", s, tid);
4576 goto redo;
4577 }
4578 } else {
4579 /* Update the free list under the local lock */
4580 local_lock(&s->cpu_slab->lock);
4581 c = this_cpu_ptr(s->cpu_slab);
4582 if (unlikely(slab != c->slab)) {
4583 local_unlock(&s->cpu_slab->lock);
4584 goto redo;
4585 }
4586 tid = c->tid;
4587 freelist = c->freelist;
4588
4589 set_freepointer(s, tail, freelist);
4590 c->freelist = head;
4591 c->tid = next_tid(tid);
4592
4593 local_unlock(&s->cpu_slab->lock);
4594 }
4595 stat_add(s, FREE_FASTPATH, cnt);
4596}
4597#else /* CONFIG_SLUB_TINY */
4598static void do_slab_free(struct kmem_cache *s,
4599 struct slab *slab, void *head, void *tail,
4600 int cnt, unsigned long addr)
4601{
4602 __slab_free(s, slab, head, tail, cnt, addr);
4603}
4604#endif /* CONFIG_SLUB_TINY */
4605
4606static __fastpath_inline
4607void slab_free(struct kmem_cache *s, struct slab *slab, void *object,
4608 unsigned long addr)
4609{
4610 memcg_slab_free_hook(s, slab, &object, 1);
4611 alloc_tagging_slab_free_hook(s, slab, &object, 1);
4612
4613 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s), false)))
4614 do_slab_free(s, slab, object, object, 1, addr);
4615}
4616
4617#ifdef CONFIG_MEMCG
4618/* Do not inline the rare memcg charging failed path into the allocation path */
4619static noinline
4620void memcg_alloc_abort_single(struct kmem_cache *s, void *object)
4621{
4622 if (likely(slab_free_hook(s, object, slab_want_init_on_free(s), false)))
4623 do_slab_free(s, virt_to_slab(object), object, object, 1, _RET_IP_);
4624}
4625#endif
4626
4627static __fastpath_inline
4628void slab_free_bulk(struct kmem_cache *s, struct slab *slab, void *head,
4629 void *tail, void **p, int cnt, unsigned long addr)
4630{
4631 memcg_slab_free_hook(s, slab, p, cnt);
4632 alloc_tagging_slab_free_hook(s, slab, p, cnt);
4633 /*
4634 * With KASAN enabled slab_free_freelist_hook modifies the freelist
4635 * to remove objects, whose reuse must be delayed.
4636 */
4637 if (likely(slab_free_freelist_hook(s, &head, &tail, &cnt)))
4638 do_slab_free(s, slab, head, tail, cnt, addr);
4639}
4640
4641#ifdef CONFIG_SLUB_RCU_DEBUG
4642static void slab_free_after_rcu_debug(struct rcu_head *rcu_head)
4643{
4644 struct rcu_delayed_free *delayed_free =
4645 container_of(rcu_head, struct rcu_delayed_free, head);
4646 void *object = delayed_free->object;
4647 struct slab *slab = virt_to_slab(object);
4648 struct kmem_cache *s;
4649
4650 kfree(delayed_free);
4651
4652 if (WARN_ON(is_kfence_address(object)))
4653 return;
4654
4655 /* find the object and the cache again */
4656 if (WARN_ON(!slab))
4657 return;
4658 s = slab->slab_cache;
4659 if (WARN_ON(!(s->flags & SLAB_TYPESAFE_BY_RCU)))
4660 return;
4661
4662 /* resume freeing */
4663 if (slab_free_hook(s, object, slab_want_init_on_free(s), true))
4664 do_slab_free(s, slab, object, object, 1, _THIS_IP_);
4665}
4666#endif /* CONFIG_SLUB_RCU_DEBUG */
4667
4668#ifdef CONFIG_KASAN_GENERIC
4669void ___cache_free(struct kmem_cache *cache, void *x, unsigned long addr)
4670{
4671 do_slab_free(cache, virt_to_slab(x), x, x, 1, addr);
4672}
4673#endif
4674
4675static inline struct kmem_cache *virt_to_cache(const void *obj)
4676{
4677 struct slab *slab;
4678
4679 slab = virt_to_slab(obj);
4680 if (WARN_ONCE(!slab, "%s: Object is not a Slab page!\n", __func__))
4681 return NULL;
4682 return slab->slab_cache;
4683}
4684
4685static inline struct kmem_cache *cache_from_obj(struct kmem_cache *s, void *x)
4686{
4687 struct kmem_cache *cachep;
4688
4689 if (!IS_ENABLED(CONFIG_SLAB_FREELIST_HARDENED) &&
4690 !kmem_cache_debug_flags(s, SLAB_CONSISTENCY_CHECKS))
4691 return s;
4692
4693 cachep = virt_to_cache(x);
4694 if (WARN(cachep && cachep != s,
4695 "%s: Wrong slab cache. %s but object is from %s\n",
4696 __func__, s->name, cachep->name))
4697 print_tracking(cachep, x);
4698 return cachep;
4699}
4700
4701/**
4702 * kmem_cache_free - Deallocate an object
4703 * @s: The cache the allocation was from.
4704 * @x: The previously allocated object.
4705 *
4706 * Free an object which was previously allocated from this
4707 * cache.
4708 */
4709void kmem_cache_free(struct kmem_cache *s, void *x)
4710{
4711 s = cache_from_obj(s, x);
4712 if (!s)
4713 return;
4714 trace_kmem_cache_free(_RET_IP_, x, s);
4715 slab_free(s, virt_to_slab(x), x, _RET_IP_);
4716}
4717EXPORT_SYMBOL(kmem_cache_free);
4718
4719static void free_large_kmalloc(struct folio *folio, void *object)
4720{
4721 unsigned int order = folio_order(folio);
4722
4723 if (WARN_ON_ONCE(order == 0))
4724 pr_warn_once("object pointer: 0x%p\n", object);
4725
4726 kmemleak_free(object);
4727 kasan_kfree_large(object);
4728 kmsan_kfree_large(object);
4729
4730 lruvec_stat_mod_folio(folio, NR_SLAB_UNRECLAIMABLE_B,
4731 -(PAGE_SIZE << order));
4732 folio_put(folio);
4733}
4734
4735/**
4736 * kfree - free previously allocated memory
4737 * @object: pointer returned by kmalloc() or kmem_cache_alloc()
4738 *
4739 * If @object is NULL, no operation is performed.
4740 */
4741void kfree(const void *object)
4742{
4743 struct folio *folio;
4744 struct slab *slab;
4745 struct kmem_cache *s;
4746 void *x = (void *)object;
4747
4748 trace_kfree(_RET_IP_, object);
4749
4750 if (unlikely(ZERO_OR_NULL_PTR(object)))
4751 return;
4752
4753 folio = virt_to_folio(object);
4754 if (unlikely(!folio_test_slab(folio))) {
4755 free_large_kmalloc(folio, (void *)object);
4756 return;
4757 }
4758
4759 slab = folio_slab(folio);
4760 s = slab->slab_cache;
4761 slab_free(s, slab, x, _RET_IP_);
4762}
4763EXPORT_SYMBOL(kfree);
4764
4765static __always_inline __realloc_size(2) void *
4766__do_krealloc(const void *p, size_t new_size, gfp_t flags)
4767{
4768 void *ret;
4769 size_t ks = 0;
4770 int orig_size = 0;
4771 struct kmem_cache *s = NULL;
4772
4773 if (unlikely(ZERO_OR_NULL_PTR(p)))
4774 goto alloc_new;
4775
4776 /* Check for double-free. */
4777 if (!kasan_check_byte(p))
4778 return NULL;
4779
4780 if (is_kfence_address(p)) {
4781 ks = orig_size = kfence_ksize(p);
4782 } else {
4783 struct folio *folio;
4784
4785 folio = virt_to_folio(p);
4786 if (unlikely(!folio_test_slab(folio))) {
4787 /* Big kmalloc object */
4788 WARN_ON(folio_size(folio) <= KMALLOC_MAX_CACHE_SIZE);
4789 WARN_ON(p != folio_address(folio));
4790 ks = folio_size(folio);
4791 } else {
4792 s = folio_slab(folio)->slab_cache;
4793 orig_size = get_orig_size(s, (void *)p);
4794 ks = s->object_size;
4795 }
4796 }
4797
4798 /* If the old object doesn't fit, allocate a bigger one */
4799 if (new_size > ks)
4800 goto alloc_new;
4801
4802 /* Zero out spare memory. */
4803 if (want_init_on_alloc(flags)) {
4804 kasan_disable_current();
4805 if (orig_size && orig_size < new_size)
4806 memset(kasan_reset_tag(p) + orig_size, 0, new_size - orig_size);
4807 else
4808 memset(kasan_reset_tag(p) + new_size, 0, ks - new_size);
4809 kasan_enable_current();
4810 }
4811
4812 /* Setup kmalloc redzone when needed */
4813 if (s && slub_debug_orig_size(s)) {
4814 set_orig_size(s, (void *)p, new_size);
4815 if (s->flags & SLAB_RED_ZONE && new_size < ks)
4816 memset_no_sanitize_memory(kasan_reset_tag(p) + new_size,
4817 SLUB_RED_ACTIVE, ks - new_size);
4818 }
4819
4820 p = kasan_krealloc(p, new_size, flags);
4821 return (void *)p;
4822
4823alloc_new:
4824 ret = kmalloc_node_track_caller_noprof(new_size, flags, NUMA_NO_NODE, _RET_IP_);
4825 if (ret && p) {
4826 /* Disable KASAN checks as the object's redzone is accessed. */
4827 kasan_disable_current();
4828 memcpy(ret, kasan_reset_tag(p), orig_size ?: ks);
4829 kasan_enable_current();
4830 }
4831
4832 return ret;
4833}
4834
4835/**
4836 * krealloc - reallocate memory. The contents will remain unchanged.
4837 * @p: object to reallocate memory for.
4838 * @new_size: how many bytes of memory are required.
4839 * @flags: the type of memory to allocate.
4840 *
4841 * If @p is %NULL, krealloc() behaves exactly like kmalloc(). If @new_size
4842 * is 0 and @p is not a %NULL pointer, the object pointed to is freed.
4843 *
4844 * If __GFP_ZERO logic is requested, callers must ensure that, starting with the
4845 * initial memory allocation, every subsequent call to this API for the same
4846 * memory allocation is flagged with __GFP_ZERO. Otherwise, it is possible that
4847 * __GFP_ZERO is not fully honored by this API.
4848 *
4849 * When slub_debug_orig_size() is off, krealloc() only knows about the bucket
4850 * size of an allocation (but not the exact size it was allocated with) and
4851 * hence implements the following semantics for shrinking and growing buffers
4852 * with __GFP_ZERO.
4853 *
4854 * new bucket
4855 * 0 size size
4856 * |--------|----------------|
4857 * | keep | zero |
4858 *
4859 * Otherwise, the original allocation size 'orig_size' could be used to
4860 * precisely clear the requested size, and the new size will also be stored
4861 * as the new 'orig_size'.
4862 *
4863 * In any case, the contents of the object pointed to are preserved up to the
4864 * lesser of the new and old sizes.
4865 *
4866 * Return: pointer to the allocated memory or %NULL in case of error
4867 */
4868void *krealloc_noprof(const void *p, size_t new_size, gfp_t flags)
4869{
4870 void *ret;
4871
4872 if (unlikely(!new_size)) {
4873 kfree(p);
4874 return ZERO_SIZE_PTR;
4875 }
4876
4877 ret = __do_krealloc(p, new_size, flags);
4878 if (ret && kasan_reset_tag(p) != kasan_reset_tag(ret))
4879 kfree(p);
4880
4881 return ret;
4882}
4883EXPORT_SYMBOL(krealloc_noprof);
4884
4885struct detached_freelist {
4886 struct slab *slab;
4887 void *tail;
4888 void *freelist;
4889 int cnt;
4890 struct kmem_cache *s;
4891};
4892
4893/*
4894 * This function progressively scans the array with free objects (with
4895 * a limited look ahead) and extract objects belonging to the same
4896 * slab. It builds a detached freelist directly within the given
4897 * slab/objects. This can happen without any need for
4898 * synchronization, because the objects are owned by running process.
4899 * The freelist is build up as a single linked list in the objects.
4900 * The idea is, that this detached freelist can then be bulk
4901 * transferred to the real freelist(s), but only requiring a single
4902 * synchronization primitive. Look ahead in the array is limited due
4903 * to performance reasons.
4904 */
4905static inline
4906int build_detached_freelist(struct kmem_cache *s, size_t size,
4907 void **p, struct detached_freelist *df)
4908{
4909 int lookahead = 3;
4910 void *object;
4911 struct folio *folio;
4912 size_t same;
4913
4914 object = p[--size];
4915 folio = virt_to_folio(object);
4916 if (!s) {
4917 /* Handle kalloc'ed objects */
4918 if (unlikely(!folio_test_slab(folio))) {
4919 free_large_kmalloc(folio, object);
4920 df->slab = NULL;
4921 return size;
4922 }
4923 /* Derive kmem_cache from object */
4924 df->slab = folio_slab(folio);
4925 df->s = df->slab->slab_cache;
4926 } else {
4927 df->slab = folio_slab(folio);
4928 df->s = cache_from_obj(s, object); /* Support for memcg */
4929 }
4930
4931 /* Start new detached freelist */
4932 df->tail = object;
4933 df->freelist = object;
4934 df->cnt = 1;
4935
4936 if (is_kfence_address(object))
4937 return size;
4938
4939 set_freepointer(df->s, object, NULL);
4940
4941 same = size;
4942 while (size) {
4943 object = p[--size];
4944 /* df->slab is always set at this point */
4945 if (df->slab == virt_to_slab(object)) {
4946 /* Opportunity build freelist */
4947 set_freepointer(df->s, object, df->freelist);
4948 df->freelist = object;
4949 df->cnt++;
4950 same--;
4951 if (size != same)
4952 swap(p[size], p[same]);
4953 continue;
4954 }
4955
4956 /* Limit look ahead search */
4957 if (!--lookahead)
4958 break;
4959 }
4960
4961 return same;
4962}
4963
4964/*
4965 * Internal bulk free of objects that were not initialised by the post alloc
4966 * hooks and thus should not be processed by the free hooks
4967 */
4968static void __kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4969{
4970 if (!size)
4971 return;
4972
4973 do {
4974 struct detached_freelist df;
4975
4976 size = build_detached_freelist(s, size, p, &df);
4977 if (!df.slab)
4978 continue;
4979
4980 if (kfence_free(df.freelist))
4981 continue;
4982
4983 do_slab_free(df.s, df.slab, df.freelist, df.tail, df.cnt,
4984 _RET_IP_);
4985 } while (likely(size));
4986}
4987
4988/* Note that interrupts must be enabled when calling this function. */
4989void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
4990{
4991 if (!size)
4992 return;
4993
4994 do {
4995 struct detached_freelist df;
4996
4997 size = build_detached_freelist(s, size, p, &df);
4998 if (!df.slab)
4999 continue;
5000
5001 slab_free_bulk(df.s, df.slab, df.freelist, df.tail, &p[size],
5002 df.cnt, _RET_IP_);
5003 } while (likely(size));
5004}
5005EXPORT_SYMBOL(kmem_cache_free_bulk);
5006
5007#ifndef CONFIG_SLUB_TINY
5008static inline
5009int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
5010 void **p)
5011{
5012 struct kmem_cache_cpu *c;
5013 unsigned long irqflags;
5014 int i;
5015
5016 /*
5017 * Drain objects in the per cpu slab, while disabling local
5018 * IRQs, which protects against PREEMPT and interrupts
5019 * handlers invoking normal fastpath.
5020 */
5021 c = slub_get_cpu_ptr(s->cpu_slab);
5022 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
5023
5024 for (i = 0; i < size; i++) {
5025 void *object = kfence_alloc(s, s->object_size, flags);
5026
5027 if (unlikely(object)) {
5028 p[i] = object;
5029 continue;
5030 }
5031
5032 object = c->freelist;
5033 if (unlikely(!object)) {
5034 /*
5035 * We may have removed an object from c->freelist using
5036 * the fastpath in the previous iteration; in that case,
5037 * c->tid has not been bumped yet.
5038 * Since ___slab_alloc() may reenable interrupts while
5039 * allocating memory, we should bump c->tid now.
5040 */
5041 c->tid = next_tid(c->tid);
5042
5043 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
5044
5045 /*
5046 * Invoking slow path likely have side-effect
5047 * of re-populating per CPU c->freelist
5048 */
5049 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
5050 _RET_IP_, c, s->object_size);
5051 if (unlikely(!p[i]))
5052 goto error;
5053
5054 c = this_cpu_ptr(s->cpu_slab);
5055 maybe_wipe_obj_freeptr(s, p[i]);
5056
5057 local_lock_irqsave(&s->cpu_slab->lock, irqflags);
5058
5059 continue; /* goto for-loop */
5060 }
5061 c->freelist = get_freepointer(s, object);
5062 p[i] = object;
5063 maybe_wipe_obj_freeptr(s, p[i]);
5064 stat(s, ALLOC_FASTPATH);
5065 }
5066 c->tid = next_tid(c->tid);
5067 local_unlock_irqrestore(&s->cpu_slab->lock, irqflags);
5068 slub_put_cpu_ptr(s->cpu_slab);
5069
5070 return i;
5071
5072error:
5073 slub_put_cpu_ptr(s->cpu_slab);
5074 __kmem_cache_free_bulk(s, i, p);
5075 return 0;
5076
5077}
5078#else /* CONFIG_SLUB_TINY */
5079static int __kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags,
5080 size_t size, void **p)
5081{
5082 int i;
5083
5084 for (i = 0; i < size; i++) {
5085 void *object = kfence_alloc(s, s->object_size, flags);
5086
5087 if (unlikely(object)) {
5088 p[i] = object;
5089 continue;
5090 }
5091
5092 p[i] = __slab_alloc_node(s, flags, NUMA_NO_NODE,
5093 _RET_IP_, s->object_size);
5094 if (unlikely(!p[i]))
5095 goto error;
5096
5097 maybe_wipe_obj_freeptr(s, p[i]);
5098 }
5099
5100 return i;
5101
5102error:
5103 __kmem_cache_free_bulk(s, i, p);
5104 return 0;
5105}
5106#endif /* CONFIG_SLUB_TINY */
5107
5108/* Note that interrupts must be enabled when calling this function. */
5109int kmem_cache_alloc_bulk_noprof(struct kmem_cache *s, gfp_t flags, size_t size,
5110 void **p)
5111{
5112 int i;
5113
5114 if (!size)
5115 return 0;
5116
5117 s = slab_pre_alloc_hook(s, flags);
5118 if (unlikely(!s))
5119 return 0;
5120
5121 i = __kmem_cache_alloc_bulk(s, flags, size, p);
5122 if (unlikely(i == 0))
5123 return 0;
5124
5125 /*
5126 * memcg and kmem_cache debug support and memory initialization.
5127 * Done outside of the IRQ disabled fastpath loop.
5128 */
5129 if (unlikely(!slab_post_alloc_hook(s, NULL, flags, size, p,
5130 slab_want_init_on_alloc(flags, s), s->object_size))) {
5131 return 0;
5132 }
5133 return i;
5134}
5135EXPORT_SYMBOL(kmem_cache_alloc_bulk_noprof);
5136
5137
5138/*
5139 * Object placement in a slab is made very easy because we always start at
5140 * offset 0. If we tune the size of the object to the alignment then we can
5141 * get the required alignment by putting one properly sized object after
5142 * another.
5143 *
5144 * Notice that the allocation order determines the sizes of the per cpu
5145 * caches. Each processor has always one slab available for allocations.
5146 * Increasing the allocation order reduces the number of times that slabs
5147 * must be moved on and off the partial lists and is therefore a factor in
5148 * locking overhead.
5149 */
5150
5151/*
5152 * Minimum / Maximum order of slab pages. This influences locking overhead
5153 * and slab fragmentation. A higher order reduces the number of partial slabs
5154 * and increases the number of allocations possible without having to
5155 * take the list_lock.
5156 */
5157static unsigned int slub_min_order;
5158static unsigned int slub_max_order =
5159 IS_ENABLED(CONFIG_SLUB_TINY) ? 1 : PAGE_ALLOC_COSTLY_ORDER;
5160static unsigned int slub_min_objects;
5161
5162/*
5163 * Calculate the order of allocation given an slab object size.
5164 *
5165 * The order of allocation has significant impact on performance and other
5166 * system components. Generally order 0 allocations should be preferred since
5167 * order 0 does not cause fragmentation in the page allocator. Larger objects
5168 * be problematic to put into order 0 slabs because there may be too much
5169 * unused space left. We go to a higher order if more than 1/16th of the slab
5170 * would be wasted.
5171 *
5172 * In order to reach satisfactory performance we must ensure that a minimum
5173 * number of objects is in one slab. Otherwise we may generate too much
5174 * activity on the partial lists which requires taking the list_lock. This is
5175 * less a concern for large slabs though which are rarely used.
5176 *
5177 * slab_max_order specifies the order where we begin to stop considering the
5178 * number of objects in a slab as critical. If we reach slab_max_order then
5179 * we try to keep the page order as low as possible. So we accept more waste
5180 * of space in favor of a small page order.
5181 *
5182 * Higher order allocations also allow the placement of more objects in a
5183 * slab and thereby reduce object handling overhead. If the user has
5184 * requested a higher minimum order then we start with that one instead of
5185 * the smallest order which will fit the object.
5186 */
5187static inline unsigned int calc_slab_order(unsigned int size,
5188 unsigned int min_order, unsigned int max_order,
5189 unsigned int fract_leftover)
5190{
5191 unsigned int order;
5192
5193 for (order = min_order; order <= max_order; order++) {
5194
5195 unsigned int slab_size = (unsigned int)PAGE_SIZE << order;
5196 unsigned int rem;
5197
5198 rem = slab_size % size;
5199
5200 if (rem <= slab_size / fract_leftover)
5201 break;
5202 }
5203
5204 return order;
5205}
5206
5207static inline int calculate_order(unsigned int size)
5208{
5209 unsigned int order;
5210 unsigned int min_objects;
5211 unsigned int max_objects;
5212 unsigned int min_order;
5213
5214 min_objects = slub_min_objects;
5215 if (!min_objects) {
5216 /*
5217 * Some architectures will only update present cpus when
5218 * onlining them, so don't trust the number if it's just 1. But
5219 * we also don't want to use nr_cpu_ids always, as on some other
5220 * architectures, there can be many possible cpus, but never
5221 * onlined. Here we compromise between trying to avoid too high
5222 * order on systems that appear larger than they are, and too
5223 * low order on systems that appear smaller than they are.
5224 */
5225 unsigned int nr_cpus = num_present_cpus();
5226 if (nr_cpus <= 1)
5227 nr_cpus = nr_cpu_ids;
5228 min_objects = 4 * (fls(nr_cpus) + 1);
5229 }
5230 /* min_objects can't be 0 because get_order(0) is undefined */
5231 max_objects = max(order_objects(slub_max_order, size), 1U);
5232 min_objects = min(min_objects, max_objects);
5233
5234 min_order = max_t(unsigned int, slub_min_order,
5235 get_order(min_objects * size));
5236 if (order_objects(min_order, size) > MAX_OBJS_PER_PAGE)
5237 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
5238
5239 /*
5240 * Attempt to find best configuration for a slab. This works by first
5241 * attempting to generate a layout with the best possible configuration
5242 * and backing off gradually.
5243 *
5244 * We start with accepting at most 1/16 waste and try to find the
5245 * smallest order from min_objects-derived/slab_min_order up to
5246 * slab_max_order that will satisfy the constraint. Note that increasing
5247 * the order can only result in same or less fractional waste, not more.
5248 *
5249 * If that fails, we increase the acceptable fraction of waste and try
5250 * again. The last iteration with fraction of 1/2 would effectively
5251 * accept any waste and give us the order determined by min_objects, as
5252 * long as at least single object fits within slab_max_order.
5253 */
5254 for (unsigned int fraction = 16; fraction > 1; fraction /= 2) {
5255 order = calc_slab_order(size, min_order, slub_max_order,
5256 fraction);
5257 if (order <= slub_max_order)
5258 return order;
5259 }
5260
5261 /*
5262 * Doh this slab cannot be placed using slab_max_order.
5263 */
5264 order = get_order(size);
5265 if (order <= MAX_PAGE_ORDER)
5266 return order;
5267 return -ENOSYS;
5268}
5269
5270static void
5271init_kmem_cache_node(struct kmem_cache_node *n)
5272{
5273 n->nr_partial = 0;
5274 spin_lock_init(&n->list_lock);
5275 INIT_LIST_HEAD(&n->partial);
5276#ifdef CONFIG_SLUB_DEBUG
5277 atomic_long_set(&n->nr_slabs, 0);
5278 atomic_long_set(&n->total_objects, 0);
5279 INIT_LIST_HEAD(&n->full);
5280#endif
5281}
5282
5283#ifndef CONFIG_SLUB_TINY
5284static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
5285{
5286 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
5287 NR_KMALLOC_TYPES * KMALLOC_SHIFT_HIGH *
5288 sizeof(struct kmem_cache_cpu));
5289
5290 /*
5291 * Must align to double word boundary for the double cmpxchg
5292 * instructions to work; see __pcpu_double_call_return_bool().
5293 */
5294 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
5295 2 * sizeof(void *));
5296
5297 if (!s->cpu_slab)
5298 return 0;
5299
5300 init_kmem_cache_cpus(s);
5301
5302 return 1;
5303}
5304#else
5305static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
5306{
5307 return 1;
5308}
5309#endif /* CONFIG_SLUB_TINY */
5310
5311static struct kmem_cache *kmem_cache_node;
5312
5313/*
5314 * No kmalloc_node yet so do it by hand. We know that this is the first
5315 * slab on the node for this slabcache. There are no concurrent accesses
5316 * possible.
5317 *
5318 * Note that this function only works on the kmem_cache_node
5319 * when allocating for the kmem_cache_node. This is used for bootstrapping
5320 * memory on a fresh node that has no slab structures yet.
5321 */
5322static void early_kmem_cache_node_alloc(int node)
5323{
5324 struct slab *slab;
5325 struct kmem_cache_node *n;
5326
5327 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
5328
5329 slab = new_slab(kmem_cache_node, GFP_NOWAIT, node);
5330
5331 BUG_ON(!slab);
5332 if (slab_nid(slab) != node) {
5333 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
5334 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
5335 }
5336
5337 n = slab->freelist;
5338 BUG_ON(!n);
5339#ifdef CONFIG_SLUB_DEBUG
5340 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
5341#endif
5342 n = kasan_slab_alloc(kmem_cache_node, n, GFP_KERNEL, false);
5343 slab->freelist = get_freepointer(kmem_cache_node, n);
5344 slab->inuse = 1;
5345 kmem_cache_node->node[node] = n;
5346 init_kmem_cache_node(n);
5347 inc_slabs_node(kmem_cache_node, node, slab->objects);
5348
5349 /*
5350 * No locks need to be taken here as it has just been
5351 * initialized and there is no concurrent access.
5352 */
5353 __add_partial(n, slab, DEACTIVATE_TO_HEAD);
5354}
5355
5356static void free_kmem_cache_nodes(struct kmem_cache *s)
5357{
5358 int node;
5359 struct kmem_cache_node *n;
5360
5361 for_each_kmem_cache_node(s, node, n) {
5362 s->node[node] = NULL;
5363 kmem_cache_free(kmem_cache_node, n);
5364 }
5365}
5366
5367void __kmem_cache_release(struct kmem_cache *s)
5368{
5369 cache_random_seq_destroy(s);
5370#ifndef CONFIG_SLUB_TINY
5371 free_percpu(s->cpu_slab);
5372#endif
5373 free_kmem_cache_nodes(s);
5374}
5375
5376static int init_kmem_cache_nodes(struct kmem_cache *s)
5377{
5378 int node;
5379
5380 for_each_node_mask(node, slab_nodes) {
5381 struct kmem_cache_node *n;
5382
5383 if (slab_state == DOWN) {
5384 early_kmem_cache_node_alloc(node);
5385 continue;
5386 }
5387 n = kmem_cache_alloc_node(kmem_cache_node,
5388 GFP_KERNEL, node);
5389
5390 if (!n) {
5391 free_kmem_cache_nodes(s);
5392 return 0;
5393 }
5394
5395 init_kmem_cache_node(n);
5396 s->node[node] = n;
5397 }
5398 return 1;
5399}
5400
5401static void set_cpu_partial(struct kmem_cache *s)
5402{
5403#ifdef CONFIG_SLUB_CPU_PARTIAL
5404 unsigned int nr_objects;
5405
5406 /*
5407 * cpu_partial determined the maximum number of objects kept in the
5408 * per cpu partial lists of a processor.
5409 *
5410 * Per cpu partial lists mainly contain slabs that just have one
5411 * object freed. If they are used for allocation then they can be
5412 * filled up again with minimal effort. The slab will never hit the
5413 * per node partial lists and therefore no locking will be required.
5414 *
5415 * For backwards compatibility reasons, this is determined as number
5416 * of objects, even though we now limit maximum number of pages, see
5417 * slub_set_cpu_partial()
5418 */
5419 if (!kmem_cache_has_cpu_partial(s))
5420 nr_objects = 0;
5421 else if (s->size >= PAGE_SIZE)
5422 nr_objects = 6;
5423 else if (s->size >= 1024)
5424 nr_objects = 24;
5425 else if (s->size >= 256)
5426 nr_objects = 52;
5427 else
5428 nr_objects = 120;
5429
5430 slub_set_cpu_partial(s, nr_objects);
5431#endif
5432}
5433
5434/*
5435 * calculate_sizes() determines the order and the distribution of data within
5436 * a slab object.
5437 */
5438static int calculate_sizes(struct kmem_cache_args *args, struct kmem_cache *s)
5439{
5440 slab_flags_t flags = s->flags;
5441 unsigned int size = s->object_size;
5442 unsigned int order;
5443
5444 /*
5445 * Round up object size to the next word boundary. We can only
5446 * place the free pointer at word boundaries and this determines
5447 * the possible location of the free pointer.
5448 */
5449 size = ALIGN(size, sizeof(void *));
5450
5451#ifdef CONFIG_SLUB_DEBUG
5452 /*
5453 * Determine if we can poison the object itself. If the user of
5454 * the slab may touch the object after free or before allocation
5455 * then we should never poison the object itself.
5456 */
5457 if ((flags & SLAB_POISON) && !(flags & SLAB_TYPESAFE_BY_RCU) &&
5458 !s->ctor)
5459 s->flags |= __OBJECT_POISON;
5460 else
5461 s->flags &= ~__OBJECT_POISON;
5462
5463
5464 /*
5465 * If we are Redzoning then check if there is some space between the
5466 * end of the object and the free pointer. If not then add an
5467 * additional word to have some bytes to store Redzone information.
5468 */
5469 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
5470 size += sizeof(void *);
5471#endif
5472
5473 /*
5474 * With that we have determined the number of bytes in actual use
5475 * by the object and redzoning.
5476 */
5477 s->inuse = size;
5478
5479 if (((flags & SLAB_TYPESAFE_BY_RCU) && !args->use_freeptr_offset) ||
5480 (flags & SLAB_POISON) || s->ctor ||
5481 ((flags & SLAB_RED_ZONE) &&
5482 (s->object_size < sizeof(void *) || slub_debug_orig_size(s)))) {
5483 /*
5484 * Relocate free pointer after the object if it is not
5485 * permitted to overwrite the first word of the object on
5486 * kmem_cache_free.
5487 *
5488 * This is the case if we do RCU, have a constructor or
5489 * destructor, are poisoning the objects, or are
5490 * redzoning an object smaller than sizeof(void *) or are
5491 * redzoning an object with slub_debug_orig_size() enabled,
5492 * in which case the right redzone may be extended.
5493 *
5494 * The assumption that s->offset >= s->inuse means free
5495 * pointer is outside of the object is used in the
5496 * freeptr_outside_object() function. If that is no
5497 * longer true, the function needs to be modified.
5498 */
5499 s->offset = size;
5500 size += sizeof(void *);
5501 } else if ((flags & SLAB_TYPESAFE_BY_RCU) && args->use_freeptr_offset) {
5502 s->offset = args->freeptr_offset;
5503 } else {
5504 /*
5505 * Store freelist pointer near middle of object to keep
5506 * it away from the edges of the object to avoid small
5507 * sized over/underflows from neighboring allocations.
5508 */
5509 s->offset = ALIGN_DOWN(s->object_size / 2, sizeof(void *));
5510 }
5511
5512#ifdef CONFIG_SLUB_DEBUG
5513 if (flags & SLAB_STORE_USER) {
5514 /*
5515 * Need to store information about allocs and frees after
5516 * the object.
5517 */
5518 size += 2 * sizeof(struct track);
5519
5520 /* Save the original kmalloc request size */
5521 if (flags & SLAB_KMALLOC)
5522 size += sizeof(unsigned int);
5523 }
5524#endif
5525
5526 kasan_cache_create(s, &size, &s->flags);
5527#ifdef CONFIG_SLUB_DEBUG
5528 if (flags & SLAB_RED_ZONE) {
5529 /*
5530 * Add some empty padding so that we can catch
5531 * overwrites from earlier objects rather than let
5532 * tracking information or the free pointer be
5533 * corrupted if a user writes before the start
5534 * of the object.
5535 */
5536 size += sizeof(void *);
5537
5538 s->red_left_pad = sizeof(void *);
5539 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
5540 size += s->red_left_pad;
5541 }
5542#endif
5543
5544 /*
5545 * SLUB stores one object immediately after another beginning from
5546 * offset 0. In order to align the objects we have to simply size
5547 * each object to conform to the alignment.
5548 */
5549 size = ALIGN(size, s->align);
5550 s->size = size;
5551 s->reciprocal_size = reciprocal_value(size);
5552 order = calculate_order(size);
5553
5554 if ((int)order < 0)
5555 return 0;
5556
5557 s->allocflags = __GFP_COMP;
5558
5559 if (s->flags & SLAB_CACHE_DMA)
5560 s->allocflags |= GFP_DMA;
5561
5562 if (s->flags & SLAB_CACHE_DMA32)
5563 s->allocflags |= GFP_DMA32;
5564
5565 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5566 s->allocflags |= __GFP_RECLAIMABLE;
5567
5568 /*
5569 * Determine the number of objects per slab
5570 */
5571 s->oo = oo_make(order, size);
5572 s->min = oo_make(get_order(size), size);
5573
5574 return !!oo_objects(s->oo);
5575}
5576
5577static void list_slab_objects(struct kmem_cache *s, struct slab *slab,
5578 const char *text)
5579{
5580#ifdef CONFIG_SLUB_DEBUG
5581 void *addr = slab_address(slab);
5582 void *p;
5583
5584 slab_err(s, slab, text, s->name);
5585
5586 spin_lock(&object_map_lock);
5587 __fill_map(object_map, s, slab);
5588
5589 for_each_object(p, s, addr, slab->objects) {
5590
5591 if (!test_bit(__obj_to_index(s, addr, p), object_map)) {
5592 if (slab_add_kunit_errors())
5593 continue;
5594 pr_err("Object 0x%p @offset=%tu\n", p, p - addr);
5595 print_tracking(s, p);
5596 }
5597 }
5598 spin_unlock(&object_map_lock);
5599#endif
5600}
5601
5602/*
5603 * Attempt to free all partial slabs on a node.
5604 * This is called from __kmem_cache_shutdown(). We must take list_lock
5605 * because sysfs file might still access partial list after the shutdowning.
5606 */
5607static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
5608{
5609 LIST_HEAD(discard);
5610 struct slab *slab, *h;
5611
5612 BUG_ON(irqs_disabled());
5613 spin_lock_irq(&n->list_lock);
5614 list_for_each_entry_safe(slab, h, &n->partial, slab_list) {
5615 if (!slab->inuse) {
5616 remove_partial(n, slab);
5617 list_add(&slab->slab_list, &discard);
5618 } else {
5619 list_slab_objects(s, slab,
5620 "Objects remaining in %s on __kmem_cache_shutdown()");
5621 }
5622 }
5623 spin_unlock_irq(&n->list_lock);
5624
5625 list_for_each_entry_safe(slab, h, &discard, slab_list)
5626 discard_slab(s, slab);
5627}
5628
5629bool __kmem_cache_empty(struct kmem_cache *s)
5630{
5631 int node;
5632 struct kmem_cache_node *n;
5633
5634 for_each_kmem_cache_node(s, node, n)
5635 if (n->nr_partial || node_nr_slabs(n))
5636 return false;
5637 return true;
5638}
5639
5640/*
5641 * Release all resources used by a slab cache.
5642 */
5643int __kmem_cache_shutdown(struct kmem_cache *s)
5644{
5645 int node;
5646 struct kmem_cache_node *n;
5647
5648 flush_all_cpus_locked(s);
5649 /* Attempt to free all objects */
5650 for_each_kmem_cache_node(s, node, n) {
5651 free_partial(s, n);
5652 if (n->nr_partial || node_nr_slabs(n))
5653 return 1;
5654 }
5655 return 0;
5656}
5657
5658#ifdef CONFIG_PRINTK
5659void __kmem_obj_info(struct kmem_obj_info *kpp, void *object, struct slab *slab)
5660{
5661 void *base;
5662 int __maybe_unused i;
5663 unsigned int objnr;
5664 void *objp;
5665 void *objp0;
5666 struct kmem_cache *s = slab->slab_cache;
5667 struct track __maybe_unused *trackp;
5668
5669 kpp->kp_ptr = object;
5670 kpp->kp_slab = slab;
5671 kpp->kp_slab_cache = s;
5672 base = slab_address(slab);
5673 objp0 = kasan_reset_tag(object);
5674#ifdef CONFIG_SLUB_DEBUG
5675 objp = restore_red_left(s, objp0);
5676#else
5677 objp = objp0;
5678#endif
5679 objnr = obj_to_index(s, slab, objp);
5680 kpp->kp_data_offset = (unsigned long)((char *)objp0 - (char *)objp);
5681 objp = base + s->size * objnr;
5682 kpp->kp_objp = objp;
5683 if (WARN_ON_ONCE(objp < base || objp >= base + slab->objects * s->size
5684 || (objp - base) % s->size) ||
5685 !(s->flags & SLAB_STORE_USER))
5686 return;
5687#ifdef CONFIG_SLUB_DEBUG
5688 objp = fixup_red_left(s, objp);
5689 trackp = get_track(s, objp, TRACK_ALLOC);
5690 kpp->kp_ret = (void *)trackp->addr;
5691#ifdef CONFIG_STACKDEPOT
5692 {
5693 depot_stack_handle_t handle;
5694 unsigned long *entries;
5695 unsigned int nr_entries;
5696
5697 handle = READ_ONCE(trackp->handle);
5698 if (handle) {
5699 nr_entries = stack_depot_fetch(handle, &entries);
5700 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5701 kpp->kp_stack[i] = (void *)entries[i];
5702 }
5703
5704 trackp = get_track(s, objp, TRACK_FREE);
5705 handle = READ_ONCE(trackp->handle);
5706 if (handle) {
5707 nr_entries = stack_depot_fetch(handle, &entries);
5708 for (i = 0; i < KS_ADDRS_COUNT && i < nr_entries; i++)
5709 kpp->kp_free_stack[i] = (void *)entries[i];
5710 }
5711 }
5712#endif
5713#endif
5714}
5715#endif
5716
5717/********************************************************************
5718 * Kmalloc subsystem
5719 *******************************************************************/
5720
5721static int __init setup_slub_min_order(char *str)
5722{
5723 get_option(&str, (int *)&slub_min_order);
5724
5725 if (slub_min_order > slub_max_order)
5726 slub_max_order = slub_min_order;
5727
5728 return 1;
5729}
5730
5731__setup("slab_min_order=", setup_slub_min_order);
5732__setup_param("slub_min_order=", slub_min_order, setup_slub_min_order, 0);
5733
5734
5735static int __init setup_slub_max_order(char *str)
5736{
5737 get_option(&str, (int *)&slub_max_order);
5738 slub_max_order = min_t(unsigned int, slub_max_order, MAX_PAGE_ORDER);
5739
5740 if (slub_min_order > slub_max_order)
5741 slub_min_order = slub_max_order;
5742
5743 return 1;
5744}
5745
5746__setup("slab_max_order=", setup_slub_max_order);
5747__setup_param("slub_max_order=", slub_max_order, setup_slub_max_order, 0);
5748
5749static int __init setup_slub_min_objects(char *str)
5750{
5751 get_option(&str, (int *)&slub_min_objects);
5752
5753 return 1;
5754}
5755
5756__setup("slab_min_objects=", setup_slub_min_objects);
5757__setup_param("slub_min_objects=", slub_min_objects, setup_slub_min_objects, 0);
5758
5759#ifdef CONFIG_NUMA
5760static int __init setup_slab_strict_numa(char *str)
5761{
5762 if (nr_node_ids > 1) {
5763 static_branch_enable(&strict_numa);
5764 pr_info("SLUB: Strict NUMA enabled.\n");
5765 } else {
5766 pr_warn("slab_strict_numa parameter set on non NUMA system.\n");
5767 }
5768
5769 return 1;
5770}
5771
5772__setup("slab_strict_numa", setup_slab_strict_numa);
5773#endif
5774
5775
5776#ifdef CONFIG_HARDENED_USERCOPY
5777/*
5778 * Rejects incorrectly sized objects and objects that are to be copied
5779 * to/from userspace but do not fall entirely within the containing slab
5780 * cache's usercopy region.
5781 *
5782 * Returns NULL if check passes, otherwise const char * to name of cache
5783 * to indicate an error.
5784 */
5785void __check_heap_object(const void *ptr, unsigned long n,
5786 const struct slab *slab, bool to_user)
5787{
5788 struct kmem_cache *s;
5789 unsigned int offset;
5790 bool is_kfence = is_kfence_address(ptr);
5791
5792 ptr = kasan_reset_tag(ptr);
5793
5794 /* Find object and usable object size. */
5795 s = slab->slab_cache;
5796
5797 /* Reject impossible pointers. */
5798 if (ptr < slab_address(slab))
5799 usercopy_abort("SLUB object not in SLUB page?!", NULL,
5800 to_user, 0, n);
5801
5802 /* Find offset within object. */
5803 if (is_kfence)
5804 offset = ptr - kfence_object_start(ptr);
5805 else
5806 offset = (ptr - slab_address(slab)) % s->size;
5807
5808 /* Adjust for redzone and reject if within the redzone. */
5809 if (!is_kfence && kmem_cache_debug_flags(s, SLAB_RED_ZONE)) {
5810 if (offset < s->red_left_pad)
5811 usercopy_abort("SLUB object in left red zone",
5812 s->name, to_user, offset, n);
5813 offset -= s->red_left_pad;
5814 }
5815
5816 /* Allow address range falling entirely within usercopy region. */
5817 if (offset >= s->useroffset &&
5818 offset - s->useroffset <= s->usersize &&
5819 n <= s->useroffset - offset + s->usersize)
5820 return;
5821
5822 usercopy_abort("SLUB object", s->name, to_user, offset, n);
5823}
5824#endif /* CONFIG_HARDENED_USERCOPY */
5825
5826#define SHRINK_PROMOTE_MAX 32
5827
5828/*
5829 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
5830 * up most to the head of the partial lists. New allocations will then
5831 * fill those up and thus they can be removed from the partial lists.
5832 *
5833 * The slabs with the least items are placed last. This results in them
5834 * being allocated from last increasing the chance that the last objects
5835 * are freed in them.
5836 */
5837static int __kmem_cache_do_shrink(struct kmem_cache *s)
5838{
5839 int node;
5840 int i;
5841 struct kmem_cache_node *n;
5842 struct slab *slab;
5843 struct slab *t;
5844 struct list_head discard;
5845 struct list_head promote[SHRINK_PROMOTE_MAX];
5846 unsigned long flags;
5847 int ret = 0;
5848
5849 for_each_kmem_cache_node(s, node, n) {
5850 INIT_LIST_HEAD(&discard);
5851 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
5852 INIT_LIST_HEAD(promote + i);
5853
5854 spin_lock_irqsave(&n->list_lock, flags);
5855
5856 /*
5857 * Build lists of slabs to discard or promote.
5858 *
5859 * Note that concurrent frees may occur while we hold the
5860 * list_lock. slab->inuse here is the upper limit.
5861 */
5862 list_for_each_entry_safe(slab, t, &n->partial, slab_list) {
5863 int free = slab->objects - slab->inuse;
5864
5865 /* Do not reread slab->inuse */
5866 barrier();
5867
5868 /* We do not keep full slabs on the list */
5869 BUG_ON(free <= 0);
5870
5871 if (free == slab->objects) {
5872 list_move(&slab->slab_list, &discard);
5873 slab_clear_node_partial(slab);
5874 n->nr_partial--;
5875 dec_slabs_node(s, node, slab->objects);
5876 } else if (free <= SHRINK_PROMOTE_MAX)
5877 list_move(&slab->slab_list, promote + free - 1);
5878 }
5879
5880 /*
5881 * Promote the slabs filled up most to the head of the
5882 * partial list.
5883 */
5884 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
5885 list_splice(promote + i, &n->partial);
5886
5887 spin_unlock_irqrestore(&n->list_lock, flags);
5888
5889 /* Release empty slabs */
5890 list_for_each_entry_safe(slab, t, &discard, slab_list)
5891 free_slab(s, slab);
5892
5893 if (node_nr_slabs(n))
5894 ret = 1;
5895 }
5896
5897 return ret;
5898}
5899
5900int __kmem_cache_shrink(struct kmem_cache *s)
5901{
5902 flush_all(s);
5903 return __kmem_cache_do_shrink(s);
5904}
5905
5906static int slab_mem_going_offline_callback(void *arg)
5907{
5908 struct kmem_cache *s;
5909
5910 mutex_lock(&slab_mutex);
5911 list_for_each_entry(s, &slab_caches, list) {
5912 flush_all_cpus_locked(s);
5913 __kmem_cache_do_shrink(s);
5914 }
5915 mutex_unlock(&slab_mutex);
5916
5917 return 0;
5918}
5919
5920static void slab_mem_offline_callback(void *arg)
5921{
5922 struct memory_notify *marg = arg;
5923 int offline_node;
5924
5925 offline_node = marg->status_change_nid_normal;
5926
5927 /*
5928 * If the node still has available memory. we need kmem_cache_node
5929 * for it yet.
5930 */
5931 if (offline_node < 0)
5932 return;
5933
5934 mutex_lock(&slab_mutex);
5935 node_clear(offline_node, slab_nodes);
5936 /*
5937 * We no longer free kmem_cache_node structures here, as it would be
5938 * racy with all get_node() users, and infeasible to protect them with
5939 * slab_mutex.
5940 */
5941 mutex_unlock(&slab_mutex);
5942}
5943
5944static int slab_mem_going_online_callback(void *arg)
5945{
5946 struct kmem_cache_node *n;
5947 struct kmem_cache *s;
5948 struct memory_notify *marg = arg;
5949 int nid = marg->status_change_nid_normal;
5950 int ret = 0;
5951
5952 /*
5953 * If the node's memory is already available, then kmem_cache_node is
5954 * already created. Nothing to do.
5955 */
5956 if (nid < 0)
5957 return 0;
5958
5959 /*
5960 * We are bringing a node online. No memory is available yet. We must
5961 * allocate a kmem_cache_node structure in order to bring the node
5962 * online.
5963 */
5964 mutex_lock(&slab_mutex);
5965 list_for_each_entry(s, &slab_caches, list) {
5966 /*
5967 * The structure may already exist if the node was previously
5968 * onlined and offlined.
5969 */
5970 if (get_node(s, nid))
5971 continue;
5972 /*
5973 * XXX: kmem_cache_alloc_node will fallback to other nodes
5974 * since memory is not yet available from the node that
5975 * is brought up.
5976 */
5977 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
5978 if (!n) {
5979 ret = -ENOMEM;
5980 goto out;
5981 }
5982 init_kmem_cache_node(n);
5983 s->node[nid] = n;
5984 }
5985 /*
5986 * Any cache created after this point will also have kmem_cache_node
5987 * initialized for the new node.
5988 */
5989 node_set(nid, slab_nodes);
5990out:
5991 mutex_unlock(&slab_mutex);
5992 return ret;
5993}
5994
5995static int slab_memory_callback(struct notifier_block *self,
5996 unsigned long action, void *arg)
5997{
5998 int ret = 0;
5999
6000 switch (action) {
6001 case MEM_GOING_ONLINE:
6002 ret = slab_mem_going_online_callback(arg);
6003 break;
6004 case MEM_GOING_OFFLINE:
6005 ret = slab_mem_going_offline_callback(arg);
6006 break;
6007 case MEM_OFFLINE:
6008 case MEM_CANCEL_ONLINE:
6009 slab_mem_offline_callback(arg);
6010 break;
6011 case MEM_ONLINE:
6012 case MEM_CANCEL_OFFLINE:
6013 break;
6014 }
6015 if (ret)
6016 ret = notifier_from_errno(ret);
6017 else
6018 ret = NOTIFY_OK;
6019 return ret;
6020}
6021
6022/********************************************************************
6023 * Basic setup of slabs
6024 *******************************************************************/
6025
6026/*
6027 * Used for early kmem_cache structures that were allocated using
6028 * the page allocator. Allocate them properly then fix up the pointers
6029 * that may be pointing to the wrong kmem_cache structure.
6030 */
6031
6032static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
6033{
6034 int node;
6035 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
6036 struct kmem_cache_node *n;
6037
6038 memcpy(s, static_cache, kmem_cache->object_size);
6039
6040 /*
6041 * This runs very early, and only the boot processor is supposed to be
6042 * up. Even if it weren't true, IRQs are not up so we couldn't fire
6043 * IPIs around.
6044 */
6045 __flush_cpu_slab(s, smp_processor_id());
6046 for_each_kmem_cache_node(s, node, n) {
6047 struct slab *p;
6048
6049 list_for_each_entry(p, &n->partial, slab_list)
6050 p->slab_cache = s;
6051
6052#ifdef CONFIG_SLUB_DEBUG
6053 list_for_each_entry(p, &n->full, slab_list)
6054 p->slab_cache = s;
6055#endif
6056 }
6057 list_add(&s->list, &slab_caches);
6058 return s;
6059}
6060
6061void __init kmem_cache_init(void)
6062{
6063 static __initdata struct kmem_cache boot_kmem_cache,
6064 boot_kmem_cache_node;
6065 int node;
6066
6067 if (debug_guardpage_minorder())
6068 slub_max_order = 0;
6069
6070 /* Print slub debugging pointers without hashing */
6071 if (__slub_debug_enabled())
6072 no_hash_pointers_enable(NULL);
6073
6074 kmem_cache_node = &boot_kmem_cache_node;
6075 kmem_cache = &boot_kmem_cache;
6076
6077 /*
6078 * Initialize the nodemask for which we will allocate per node
6079 * structures. Here we don't need taking slab_mutex yet.
6080 */
6081 for_each_node_state(node, N_NORMAL_MEMORY)
6082 node_set(node, slab_nodes);
6083
6084 create_boot_cache(kmem_cache_node, "kmem_cache_node",
6085 sizeof(struct kmem_cache_node),
6086 SLAB_HWCACHE_ALIGN | SLAB_NO_OBJ_EXT, 0, 0);
6087
6088 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
6089
6090 /* Able to allocate the per node structures */
6091 slab_state = PARTIAL;
6092
6093 create_boot_cache(kmem_cache, "kmem_cache",
6094 offsetof(struct kmem_cache, node) +
6095 nr_node_ids * sizeof(struct kmem_cache_node *),
6096 SLAB_HWCACHE_ALIGN | SLAB_NO_OBJ_EXT, 0, 0);
6097
6098 kmem_cache = bootstrap(&boot_kmem_cache);
6099 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
6100
6101 /* Now we can use the kmem_cache to allocate kmalloc slabs */
6102 setup_kmalloc_cache_index_table();
6103 create_kmalloc_caches();
6104
6105 /* Setup random freelists for each cache */
6106 init_freelist_randomization();
6107
6108 cpuhp_setup_state_nocalls(CPUHP_SLUB_DEAD, "slub:dead", NULL,
6109 slub_cpu_dead);
6110
6111 pr_info("SLUB: HWalign=%d, Order=%u-%u, MinObjects=%u, CPUs=%u, Nodes=%u\n",
6112 cache_line_size(),
6113 slub_min_order, slub_max_order, slub_min_objects,
6114 nr_cpu_ids, nr_node_ids);
6115}
6116
6117void __init kmem_cache_init_late(void)
6118{
6119#ifndef CONFIG_SLUB_TINY
6120 flushwq = alloc_workqueue("slub_flushwq", WQ_MEM_RECLAIM, 0);
6121 WARN_ON(!flushwq);
6122#endif
6123}
6124
6125struct kmem_cache *
6126__kmem_cache_alias(const char *name, unsigned int size, unsigned int align,
6127 slab_flags_t flags, void (*ctor)(void *))
6128{
6129 struct kmem_cache *s;
6130
6131 s = find_mergeable(size, align, flags, name, ctor);
6132 if (s) {
6133 if (sysfs_slab_alias(s, name))
6134 pr_err("SLUB: Unable to add cache alias %s to sysfs\n",
6135 name);
6136
6137 s->refcount++;
6138
6139 /*
6140 * Adjust the object sizes so that we clear
6141 * the complete object on kzalloc.
6142 */
6143 s->object_size = max(s->object_size, size);
6144 s->inuse = max(s->inuse, ALIGN(size, sizeof(void *)));
6145 }
6146
6147 return s;
6148}
6149
6150int do_kmem_cache_create(struct kmem_cache *s, const char *name,
6151 unsigned int size, struct kmem_cache_args *args,
6152 slab_flags_t flags)
6153{
6154 int err = -EINVAL;
6155
6156 s->name = name;
6157 s->size = s->object_size = size;
6158
6159 s->flags = kmem_cache_flags(flags, s->name);
6160#ifdef CONFIG_SLAB_FREELIST_HARDENED
6161 s->random = get_random_long();
6162#endif
6163 s->align = args->align;
6164 s->ctor = args->ctor;
6165#ifdef CONFIG_HARDENED_USERCOPY
6166 s->useroffset = args->useroffset;
6167 s->usersize = args->usersize;
6168#endif
6169
6170 if (!calculate_sizes(args, s))
6171 goto out;
6172 if (disable_higher_order_debug) {
6173 /*
6174 * Disable debugging flags that store metadata if the min slab
6175 * order increased.
6176 */
6177 if (get_order(s->size) > get_order(s->object_size)) {
6178 s->flags &= ~DEBUG_METADATA_FLAGS;
6179 s->offset = 0;
6180 if (!calculate_sizes(args, s))
6181 goto out;
6182 }
6183 }
6184
6185#ifdef system_has_freelist_aba
6186 if (system_has_freelist_aba() && !(s->flags & SLAB_NO_CMPXCHG)) {
6187 /* Enable fast mode */
6188 s->flags |= __CMPXCHG_DOUBLE;
6189 }
6190#endif
6191
6192 /*
6193 * The larger the object size is, the more slabs we want on the partial
6194 * list to avoid pounding the page allocator excessively.
6195 */
6196 s->min_partial = min_t(unsigned long, MAX_PARTIAL, ilog2(s->size) / 2);
6197 s->min_partial = max_t(unsigned long, MIN_PARTIAL, s->min_partial);
6198
6199 set_cpu_partial(s);
6200
6201#ifdef CONFIG_NUMA
6202 s->remote_node_defrag_ratio = 1000;
6203#endif
6204
6205 /* Initialize the pre-computed randomized freelist if slab is up */
6206 if (slab_state >= UP) {
6207 if (init_cache_random_seq(s))
6208 goto out;
6209 }
6210
6211 if (!init_kmem_cache_nodes(s))
6212 goto out;
6213
6214 if (!alloc_kmem_cache_cpus(s))
6215 goto out;
6216
6217 err = 0;
6218
6219 /* Mutex is not taken during early boot */
6220 if (slab_state <= UP)
6221 goto out;
6222
6223 /*
6224 * Failing to create sysfs files is not critical to SLUB functionality.
6225 * If it fails, proceed with cache creation without these files.
6226 */
6227 if (sysfs_slab_add(s))
6228 pr_err("SLUB: Unable to add cache %s to sysfs\n", s->name);
6229
6230 if (s->flags & SLAB_STORE_USER)
6231 debugfs_slab_add(s);
6232
6233out:
6234 if (err)
6235 __kmem_cache_release(s);
6236 return err;
6237}
6238
6239#ifdef SLAB_SUPPORTS_SYSFS
6240static int count_inuse(struct slab *slab)
6241{
6242 return slab->inuse;
6243}
6244
6245static int count_total(struct slab *slab)
6246{
6247 return slab->objects;
6248}
6249#endif
6250
6251#ifdef CONFIG_SLUB_DEBUG
6252static void validate_slab(struct kmem_cache *s, struct slab *slab,
6253 unsigned long *obj_map)
6254{
6255 void *p;
6256 void *addr = slab_address(slab);
6257
6258 if (!check_slab(s, slab) || !on_freelist(s, slab, NULL))
6259 return;
6260
6261 /* Now we know that a valid freelist exists */
6262 __fill_map(obj_map, s, slab);
6263 for_each_object(p, s, addr, slab->objects) {
6264 u8 val = test_bit(__obj_to_index(s, addr, p), obj_map) ?
6265 SLUB_RED_INACTIVE : SLUB_RED_ACTIVE;
6266
6267 if (!check_object(s, slab, p, val))
6268 break;
6269 }
6270}
6271
6272static int validate_slab_node(struct kmem_cache *s,
6273 struct kmem_cache_node *n, unsigned long *obj_map)
6274{
6275 unsigned long count = 0;
6276 struct slab *slab;
6277 unsigned long flags;
6278
6279 spin_lock_irqsave(&n->list_lock, flags);
6280
6281 list_for_each_entry(slab, &n->partial, slab_list) {
6282 validate_slab(s, slab, obj_map);
6283 count++;
6284 }
6285 if (count != n->nr_partial) {
6286 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
6287 s->name, count, n->nr_partial);
6288 slab_add_kunit_errors();
6289 }
6290
6291 if (!(s->flags & SLAB_STORE_USER))
6292 goto out;
6293
6294 list_for_each_entry(slab, &n->full, slab_list) {
6295 validate_slab(s, slab, obj_map);
6296 count++;
6297 }
6298 if (count != node_nr_slabs(n)) {
6299 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
6300 s->name, count, node_nr_slabs(n));
6301 slab_add_kunit_errors();
6302 }
6303
6304out:
6305 spin_unlock_irqrestore(&n->list_lock, flags);
6306 return count;
6307}
6308
6309long validate_slab_cache(struct kmem_cache *s)
6310{
6311 int node;
6312 unsigned long count = 0;
6313 struct kmem_cache_node *n;
6314 unsigned long *obj_map;
6315
6316 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
6317 if (!obj_map)
6318 return -ENOMEM;
6319
6320 flush_all(s);
6321 for_each_kmem_cache_node(s, node, n)
6322 count += validate_slab_node(s, n, obj_map);
6323
6324 bitmap_free(obj_map);
6325
6326 return count;
6327}
6328EXPORT_SYMBOL(validate_slab_cache);
6329
6330#ifdef CONFIG_DEBUG_FS
6331/*
6332 * Generate lists of code addresses where slabcache objects are allocated
6333 * and freed.
6334 */
6335
6336struct location {
6337 depot_stack_handle_t handle;
6338 unsigned long count;
6339 unsigned long addr;
6340 unsigned long waste;
6341 long long sum_time;
6342 long min_time;
6343 long max_time;
6344 long min_pid;
6345 long max_pid;
6346 DECLARE_BITMAP(cpus, NR_CPUS);
6347 nodemask_t nodes;
6348};
6349
6350struct loc_track {
6351 unsigned long max;
6352 unsigned long count;
6353 struct location *loc;
6354 loff_t idx;
6355};
6356
6357static struct dentry *slab_debugfs_root;
6358
6359static void free_loc_track(struct loc_track *t)
6360{
6361 if (t->max)
6362 free_pages((unsigned long)t->loc,
6363 get_order(sizeof(struct location) * t->max));
6364}
6365
6366static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
6367{
6368 struct location *l;
6369 int order;
6370
6371 order = get_order(sizeof(struct location) * max);
6372
6373 l = (void *)__get_free_pages(flags, order);
6374 if (!l)
6375 return 0;
6376
6377 if (t->count) {
6378 memcpy(l, t->loc, sizeof(struct location) * t->count);
6379 free_loc_track(t);
6380 }
6381 t->max = max;
6382 t->loc = l;
6383 return 1;
6384}
6385
6386static int add_location(struct loc_track *t, struct kmem_cache *s,
6387 const struct track *track,
6388 unsigned int orig_size)
6389{
6390 long start, end, pos;
6391 struct location *l;
6392 unsigned long caddr, chandle, cwaste;
6393 unsigned long age = jiffies - track->when;
6394 depot_stack_handle_t handle = 0;
6395 unsigned int waste = s->object_size - orig_size;
6396
6397#ifdef CONFIG_STACKDEPOT
6398 handle = READ_ONCE(track->handle);
6399#endif
6400 start = -1;
6401 end = t->count;
6402
6403 for ( ; ; ) {
6404 pos = start + (end - start + 1) / 2;
6405
6406 /*
6407 * There is nothing at "end". If we end up there
6408 * we need to add something to before end.
6409 */
6410 if (pos == end)
6411 break;
6412
6413 l = &t->loc[pos];
6414 caddr = l->addr;
6415 chandle = l->handle;
6416 cwaste = l->waste;
6417 if ((track->addr == caddr) && (handle == chandle) &&
6418 (waste == cwaste)) {
6419
6420 l->count++;
6421 if (track->when) {
6422 l->sum_time += age;
6423 if (age < l->min_time)
6424 l->min_time = age;
6425 if (age > l->max_time)
6426 l->max_time = age;
6427
6428 if (track->pid < l->min_pid)
6429 l->min_pid = track->pid;
6430 if (track->pid > l->max_pid)
6431 l->max_pid = track->pid;
6432
6433 cpumask_set_cpu(track->cpu,
6434 to_cpumask(l->cpus));
6435 }
6436 node_set(page_to_nid(virt_to_page(track)), l->nodes);
6437 return 1;
6438 }
6439
6440 if (track->addr < caddr)
6441 end = pos;
6442 else if (track->addr == caddr && handle < chandle)
6443 end = pos;
6444 else if (track->addr == caddr && handle == chandle &&
6445 waste < cwaste)
6446 end = pos;
6447 else
6448 start = pos;
6449 }
6450
6451 /*
6452 * Not found. Insert new tracking element.
6453 */
6454 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
6455 return 0;
6456
6457 l = t->loc + pos;
6458 if (pos < t->count)
6459 memmove(l + 1, l,
6460 (t->count - pos) * sizeof(struct location));
6461 t->count++;
6462 l->count = 1;
6463 l->addr = track->addr;
6464 l->sum_time = age;
6465 l->min_time = age;
6466 l->max_time = age;
6467 l->min_pid = track->pid;
6468 l->max_pid = track->pid;
6469 l->handle = handle;
6470 l->waste = waste;
6471 cpumask_clear(to_cpumask(l->cpus));
6472 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
6473 nodes_clear(l->nodes);
6474 node_set(page_to_nid(virt_to_page(track)), l->nodes);
6475 return 1;
6476}
6477
6478static void process_slab(struct loc_track *t, struct kmem_cache *s,
6479 struct slab *slab, enum track_item alloc,
6480 unsigned long *obj_map)
6481{
6482 void *addr = slab_address(slab);
6483 bool is_alloc = (alloc == TRACK_ALLOC);
6484 void *p;
6485
6486 __fill_map(obj_map, s, slab);
6487
6488 for_each_object(p, s, addr, slab->objects)
6489 if (!test_bit(__obj_to_index(s, addr, p), obj_map))
6490 add_location(t, s, get_track(s, p, alloc),
6491 is_alloc ? get_orig_size(s, p) :
6492 s->object_size);
6493}
6494#endif /* CONFIG_DEBUG_FS */
6495#endif /* CONFIG_SLUB_DEBUG */
6496
6497#ifdef SLAB_SUPPORTS_SYSFS
6498enum slab_stat_type {
6499 SL_ALL, /* All slabs */
6500 SL_PARTIAL, /* Only partially allocated slabs */
6501 SL_CPU, /* Only slabs used for cpu caches */
6502 SL_OBJECTS, /* Determine allocated objects not slabs */
6503 SL_TOTAL /* Determine object capacity not slabs */
6504};
6505
6506#define SO_ALL (1 << SL_ALL)
6507#define SO_PARTIAL (1 << SL_PARTIAL)
6508#define SO_CPU (1 << SL_CPU)
6509#define SO_OBJECTS (1 << SL_OBJECTS)
6510#define SO_TOTAL (1 << SL_TOTAL)
6511
6512static ssize_t show_slab_objects(struct kmem_cache *s,
6513 char *buf, unsigned long flags)
6514{
6515 unsigned long total = 0;
6516 int node;
6517 int x;
6518 unsigned long *nodes;
6519 int len = 0;
6520
6521 nodes = kcalloc(nr_node_ids, sizeof(unsigned long), GFP_KERNEL);
6522 if (!nodes)
6523 return -ENOMEM;
6524
6525 if (flags & SO_CPU) {
6526 int cpu;
6527
6528 for_each_possible_cpu(cpu) {
6529 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
6530 cpu);
6531 int node;
6532 struct slab *slab;
6533
6534 slab = READ_ONCE(c->slab);
6535 if (!slab)
6536 continue;
6537
6538 node = slab_nid(slab);
6539 if (flags & SO_TOTAL)
6540 x = slab->objects;
6541 else if (flags & SO_OBJECTS)
6542 x = slab->inuse;
6543 else
6544 x = 1;
6545
6546 total += x;
6547 nodes[node] += x;
6548
6549#ifdef CONFIG_SLUB_CPU_PARTIAL
6550 slab = slub_percpu_partial_read_once(c);
6551 if (slab) {
6552 node = slab_nid(slab);
6553 if (flags & SO_TOTAL)
6554 WARN_ON_ONCE(1);
6555 else if (flags & SO_OBJECTS)
6556 WARN_ON_ONCE(1);
6557 else
6558 x = data_race(slab->slabs);
6559 total += x;
6560 nodes[node] += x;
6561 }
6562#endif
6563 }
6564 }
6565
6566 /*
6567 * It is impossible to take "mem_hotplug_lock" here with "kernfs_mutex"
6568 * already held which will conflict with an existing lock order:
6569 *
6570 * mem_hotplug_lock->slab_mutex->kernfs_mutex
6571 *
6572 * We don't really need mem_hotplug_lock (to hold off
6573 * slab_mem_going_offline_callback) here because slab's memory hot
6574 * unplug code doesn't destroy the kmem_cache->node[] data.
6575 */
6576
6577#ifdef CONFIG_SLUB_DEBUG
6578 if (flags & SO_ALL) {
6579 struct kmem_cache_node *n;
6580
6581 for_each_kmem_cache_node(s, node, n) {
6582
6583 if (flags & SO_TOTAL)
6584 x = node_nr_objs(n);
6585 else if (flags & SO_OBJECTS)
6586 x = node_nr_objs(n) - count_partial(n, count_free);
6587 else
6588 x = node_nr_slabs(n);
6589 total += x;
6590 nodes[node] += x;
6591 }
6592
6593 } else
6594#endif
6595 if (flags & SO_PARTIAL) {
6596 struct kmem_cache_node *n;
6597
6598 for_each_kmem_cache_node(s, node, n) {
6599 if (flags & SO_TOTAL)
6600 x = count_partial(n, count_total);
6601 else if (flags & SO_OBJECTS)
6602 x = count_partial(n, count_inuse);
6603 else
6604 x = n->nr_partial;
6605 total += x;
6606 nodes[node] += x;
6607 }
6608 }
6609
6610 len += sysfs_emit_at(buf, len, "%lu", total);
6611#ifdef CONFIG_NUMA
6612 for (node = 0; node < nr_node_ids; node++) {
6613 if (nodes[node])
6614 len += sysfs_emit_at(buf, len, " N%d=%lu",
6615 node, nodes[node]);
6616 }
6617#endif
6618 len += sysfs_emit_at(buf, len, "\n");
6619 kfree(nodes);
6620
6621 return len;
6622}
6623
6624#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
6625#define to_slab(n) container_of(n, struct kmem_cache, kobj)
6626
6627struct slab_attribute {
6628 struct attribute attr;
6629 ssize_t (*show)(struct kmem_cache *s, char *buf);
6630 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
6631};
6632
6633#define SLAB_ATTR_RO(_name) \
6634 static struct slab_attribute _name##_attr = __ATTR_RO_MODE(_name, 0400)
6635
6636#define SLAB_ATTR(_name) \
6637 static struct slab_attribute _name##_attr = __ATTR_RW_MODE(_name, 0600)
6638
6639static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
6640{
6641 return sysfs_emit(buf, "%u\n", s->size);
6642}
6643SLAB_ATTR_RO(slab_size);
6644
6645static ssize_t align_show(struct kmem_cache *s, char *buf)
6646{
6647 return sysfs_emit(buf, "%u\n", s->align);
6648}
6649SLAB_ATTR_RO(align);
6650
6651static ssize_t object_size_show(struct kmem_cache *s, char *buf)
6652{
6653 return sysfs_emit(buf, "%u\n", s->object_size);
6654}
6655SLAB_ATTR_RO(object_size);
6656
6657static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
6658{
6659 return sysfs_emit(buf, "%u\n", oo_objects(s->oo));
6660}
6661SLAB_ATTR_RO(objs_per_slab);
6662
6663static ssize_t order_show(struct kmem_cache *s, char *buf)
6664{
6665 return sysfs_emit(buf, "%u\n", oo_order(s->oo));
6666}
6667SLAB_ATTR_RO(order);
6668
6669static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
6670{
6671 return sysfs_emit(buf, "%lu\n", s->min_partial);
6672}
6673
6674static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
6675 size_t length)
6676{
6677 unsigned long min;
6678 int err;
6679
6680 err = kstrtoul(buf, 10, &min);
6681 if (err)
6682 return err;
6683
6684 s->min_partial = min;
6685 return length;
6686}
6687SLAB_ATTR(min_partial);
6688
6689static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
6690{
6691 unsigned int nr_partial = 0;
6692#ifdef CONFIG_SLUB_CPU_PARTIAL
6693 nr_partial = s->cpu_partial;
6694#endif
6695
6696 return sysfs_emit(buf, "%u\n", nr_partial);
6697}
6698
6699static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
6700 size_t length)
6701{
6702 unsigned int objects;
6703 int err;
6704
6705 err = kstrtouint(buf, 10, &objects);
6706 if (err)
6707 return err;
6708 if (objects && !kmem_cache_has_cpu_partial(s))
6709 return -EINVAL;
6710
6711 slub_set_cpu_partial(s, objects);
6712 flush_all(s);
6713 return length;
6714}
6715SLAB_ATTR(cpu_partial);
6716
6717static ssize_t ctor_show(struct kmem_cache *s, char *buf)
6718{
6719 if (!s->ctor)
6720 return 0;
6721 return sysfs_emit(buf, "%pS\n", s->ctor);
6722}
6723SLAB_ATTR_RO(ctor);
6724
6725static ssize_t aliases_show(struct kmem_cache *s, char *buf)
6726{
6727 return sysfs_emit(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
6728}
6729SLAB_ATTR_RO(aliases);
6730
6731static ssize_t partial_show(struct kmem_cache *s, char *buf)
6732{
6733 return show_slab_objects(s, buf, SO_PARTIAL);
6734}
6735SLAB_ATTR_RO(partial);
6736
6737static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
6738{
6739 return show_slab_objects(s, buf, SO_CPU);
6740}
6741SLAB_ATTR_RO(cpu_slabs);
6742
6743static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
6744{
6745 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
6746}
6747SLAB_ATTR_RO(objects_partial);
6748
6749static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
6750{
6751 int objects = 0;
6752 int slabs = 0;
6753 int cpu __maybe_unused;
6754 int len = 0;
6755
6756#ifdef CONFIG_SLUB_CPU_PARTIAL
6757 for_each_online_cpu(cpu) {
6758 struct slab *slab;
6759
6760 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6761
6762 if (slab)
6763 slabs += data_race(slab->slabs);
6764 }
6765#endif
6766
6767 /* Approximate half-full slabs, see slub_set_cpu_partial() */
6768 objects = (slabs * oo_objects(s->oo)) / 2;
6769 len += sysfs_emit_at(buf, len, "%d(%d)", objects, slabs);
6770
6771#ifdef CONFIG_SLUB_CPU_PARTIAL
6772 for_each_online_cpu(cpu) {
6773 struct slab *slab;
6774
6775 slab = slub_percpu_partial(per_cpu_ptr(s->cpu_slab, cpu));
6776 if (slab) {
6777 slabs = data_race(slab->slabs);
6778 objects = (slabs * oo_objects(s->oo)) / 2;
6779 len += sysfs_emit_at(buf, len, " C%d=%d(%d)",
6780 cpu, objects, slabs);
6781 }
6782 }
6783#endif
6784 len += sysfs_emit_at(buf, len, "\n");
6785
6786 return len;
6787}
6788SLAB_ATTR_RO(slabs_cpu_partial);
6789
6790static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
6791{
6792 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
6793}
6794SLAB_ATTR_RO(reclaim_account);
6795
6796static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
6797{
6798 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
6799}
6800SLAB_ATTR_RO(hwcache_align);
6801
6802#ifdef CONFIG_ZONE_DMA
6803static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
6804{
6805 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
6806}
6807SLAB_ATTR_RO(cache_dma);
6808#endif
6809
6810#ifdef CONFIG_HARDENED_USERCOPY
6811static ssize_t usersize_show(struct kmem_cache *s, char *buf)
6812{
6813 return sysfs_emit(buf, "%u\n", s->usersize);
6814}
6815SLAB_ATTR_RO(usersize);
6816#endif
6817
6818static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
6819{
6820 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TYPESAFE_BY_RCU));
6821}
6822SLAB_ATTR_RO(destroy_by_rcu);
6823
6824#ifdef CONFIG_SLUB_DEBUG
6825static ssize_t slabs_show(struct kmem_cache *s, char *buf)
6826{
6827 return show_slab_objects(s, buf, SO_ALL);
6828}
6829SLAB_ATTR_RO(slabs);
6830
6831static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
6832{
6833 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
6834}
6835SLAB_ATTR_RO(total_objects);
6836
6837static ssize_t objects_show(struct kmem_cache *s, char *buf)
6838{
6839 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
6840}
6841SLAB_ATTR_RO(objects);
6842
6843static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
6844{
6845 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
6846}
6847SLAB_ATTR_RO(sanity_checks);
6848
6849static ssize_t trace_show(struct kmem_cache *s, char *buf)
6850{
6851 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_TRACE));
6852}
6853SLAB_ATTR_RO(trace);
6854
6855static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
6856{
6857 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
6858}
6859
6860SLAB_ATTR_RO(red_zone);
6861
6862static ssize_t poison_show(struct kmem_cache *s, char *buf)
6863{
6864 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_POISON));
6865}
6866
6867SLAB_ATTR_RO(poison);
6868
6869static ssize_t store_user_show(struct kmem_cache *s, char *buf)
6870{
6871 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
6872}
6873
6874SLAB_ATTR_RO(store_user);
6875
6876static ssize_t validate_show(struct kmem_cache *s, char *buf)
6877{
6878 return 0;
6879}
6880
6881static ssize_t validate_store(struct kmem_cache *s,
6882 const char *buf, size_t length)
6883{
6884 int ret = -EINVAL;
6885
6886 if (buf[0] == '1' && kmem_cache_debug(s)) {
6887 ret = validate_slab_cache(s);
6888 if (ret >= 0)
6889 ret = length;
6890 }
6891 return ret;
6892}
6893SLAB_ATTR(validate);
6894
6895#endif /* CONFIG_SLUB_DEBUG */
6896
6897#ifdef CONFIG_FAILSLAB
6898static ssize_t failslab_show(struct kmem_cache *s, char *buf)
6899{
6900 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
6901}
6902
6903static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
6904 size_t length)
6905{
6906 if (s->refcount > 1)
6907 return -EINVAL;
6908
6909 if (buf[0] == '1')
6910 WRITE_ONCE(s->flags, s->flags | SLAB_FAILSLAB);
6911 else
6912 WRITE_ONCE(s->flags, s->flags & ~SLAB_FAILSLAB);
6913
6914 return length;
6915}
6916SLAB_ATTR(failslab);
6917#endif
6918
6919static ssize_t shrink_show(struct kmem_cache *s, char *buf)
6920{
6921 return 0;
6922}
6923
6924static ssize_t shrink_store(struct kmem_cache *s,
6925 const char *buf, size_t length)
6926{
6927 if (buf[0] == '1')
6928 kmem_cache_shrink(s);
6929 else
6930 return -EINVAL;
6931 return length;
6932}
6933SLAB_ATTR(shrink);
6934
6935#ifdef CONFIG_NUMA
6936static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
6937{
6938 return sysfs_emit(buf, "%u\n", s->remote_node_defrag_ratio / 10);
6939}
6940
6941static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
6942 const char *buf, size_t length)
6943{
6944 unsigned int ratio;
6945 int err;
6946
6947 err = kstrtouint(buf, 10, &ratio);
6948 if (err)
6949 return err;
6950 if (ratio > 100)
6951 return -ERANGE;
6952
6953 s->remote_node_defrag_ratio = ratio * 10;
6954
6955 return length;
6956}
6957SLAB_ATTR(remote_node_defrag_ratio);
6958#endif
6959
6960#ifdef CONFIG_SLUB_STATS
6961static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
6962{
6963 unsigned long sum = 0;
6964 int cpu;
6965 int len = 0;
6966 int *data = kmalloc_array(nr_cpu_ids, sizeof(int), GFP_KERNEL);
6967
6968 if (!data)
6969 return -ENOMEM;
6970
6971 for_each_online_cpu(cpu) {
6972 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
6973
6974 data[cpu] = x;
6975 sum += x;
6976 }
6977
6978 len += sysfs_emit_at(buf, len, "%lu", sum);
6979
6980#ifdef CONFIG_SMP
6981 for_each_online_cpu(cpu) {
6982 if (data[cpu])
6983 len += sysfs_emit_at(buf, len, " C%d=%u",
6984 cpu, data[cpu]);
6985 }
6986#endif
6987 kfree(data);
6988 len += sysfs_emit_at(buf, len, "\n");
6989
6990 return len;
6991}
6992
6993static void clear_stat(struct kmem_cache *s, enum stat_item si)
6994{
6995 int cpu;
6996
6997 for_each_online_cpu(cpu)
6998 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
6999}
7000
7001#define STAT_ATTR(si, text) \
7002static ssize_t text##_show(struct kmem_cache *s, char *buf) \
7003{ \
7004 return show_stat(s, buf, si); \
7005} \
7006static ssize_t text##_store(struct kmem_cache *s, \
7007 const char *buf, size_t length) \
7008{ \
7009 if (buf[0] != '0') \
7010 return -EINVAL; \
7011 clear_stat(s, si); \
7012 return length; \
7013} \
7014SLAB_ATTR(text); \
7015
7016STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
7017STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
7018STAT_ATTR(FREE_FASTPATH, free_fastpath);
7019STAT_ATTR(FREE_SLOWPATH, free_slowpath);
7020STAT_ATTR(FREE_FROZEN, free_frozen);
7021STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
7022STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
7023STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
7024STAT_ATTR(ALLOC_SLAB, alloc_slab);
7025STAT_ATTR(ALLOC_REFILL, alloc_refill);
7026STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
7027STAT_ATTR(FREE_SLAB, free_slab);
7028STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
7029STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
7030STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
7031STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
7032STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
7033STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
7034STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
7035STAT_ATTR(ORDER_FALLBACK, order_fallback);
7036STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
7037STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
7038STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
7039STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
7040STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
7041STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
7042#endif /* CONFIG_SLUB_STATS */
7043
7044#ifdef CONFIG_KFENCE
7045static ssize_t skip_kfence_show(struct kmem_cache *s, char *buf)
7046{
7047 return sysfs_emit(buf, "%d\n", !!(s->flags & SLAB_SKIP_KFENCE));
7048}
7049
7050static ssize_t skip_kfence_store(struct kmem_cache *s,
7051 const char *buf, size_t length)
7052{
7053 int ret = length;
7054
7055 if (buf[0] == '0')
7056 s->flags &= ~SLAB_SKIP_KFENCE;
7057 else if (buf[0] == '1')
7058 s->flags |= SLAB_SKIP_KFENCE;
7059 else
7060 ret = -EINVAL;
7061
7062 return ret;
7063}
7064SLAB_ATTR(skip_kfence);
7065#endif
7066
7067static struct attribute *slab_attrs[] = {
7068 &slab_size_attr.attr,
7069 &object_size_attr.attr,
7070 &objs_per_slab_attr.attr,
7071 &order_attr.attr,
7072 &min_partial_attr.attr,
7073 &cpu_partial_attr.attr,
7074 &objects_partial_attr.attr,
7075 &partial_attr.attr,
7076 &cpu_slabs_attr.attr,
7077 &ctor_attr.attr,
7078 &aliases_attr.attr,
7079 &align_attr.attr,
7080 &hwcache_align_attr.attr,
7081 &reclaim_account_attr.attr,
7082 &destroy_by_rcu_attr.attr,
7083 &shrink_attr.attr,
7084 &slabs_cpu_partial_attr.attr,
7085#ifdef CONFIG_SLUB_DEBUG
7086 &total_objects_attr.attr,
7087 &objects_attr.attr,
7088 &slabs_attr.attr,
7089 &sanity_checks_attr.attr,
7090 &trace_attr.attr,
7091 &red_zone_attr.attr,
7092 &poison_attr.attr,
7093 &store_user_attr.attr,
7094 &validate_attr.attr,
7095#endif
7096#ifdef CONFIG_ZONE_DMA
7097 &cache_dma_attr.attr,
7098#endif
7099#ifdef CONFIG_NUMA
7100 &remote_node_defrag_ratio_attr.attr,
7101#endif
7102#ifdef CONFIG_SLUB_STATS
7103 &alloc_fastpath_attr.attr,
7104 &alloc_slowpath_attr.attr,
7105 &free_fastpath_attr.attr,
7106 &free_slowpath_attr.attr,
7107 &free_frozen_attr.attr,
7108 &free_add_partial_attr.attr,
7109 &free_remove_partial_attr.attr,
7110 &alloc_from_partial_attr.attr,
7111 &alloc_slab_attr.attr,
7112 &alloc_refill_attr.attr,
7113 &alloc_node_mismatch_attr.attr,
7114 &free_slab_attr.attr,
7115 &cpuslab_flush_attr.attr,
7116 &deactivate_full_attr.attr,
7117 &deactivate_empty_attr.attr,
7118 &deactivate_to_head_attr.attr,
7119 &deactivate_to_tail_attr.attr,
7120 &deactivate_remote_frees_attr.attr,
7121 &deactivate_bypass_attr.attr,
7122 &order_fallback_attr.attr,
7123 &cmpxchg_double_fail_attr.attr,
7124 &cmpxchg_double_cpu_fail_attr.attr,
7125 &cpu_partial_alloc_attr.attr,
7126 &cpu_partial_free_attr.attr,
7127 &cpu_partial_node_attr.attr,
7128 &cpu_partial_drain_attr.attr,
7129#endif
7130#ifdef CONFIG_FAILSLAB
7131 &failslab_attr.attr,
7132#endif
7133#ifdef CONFIG_HARDENED_USERCOPY
7134 &usersize_attr.attr,
7135#endif
7136#ifdef CONFIG_KFENCE
7137 &skip_kfence_attr.attr,
7138#endif
7139
7140 NULL
7141};
7142
7143static const struct attribute_group slab_attr_group = {
7144 .attrs = slab_attrs,
7145};
7146
7147static ssize_t slab_attr_show(struct kobject *kobj,
7148 struct attribute *attr,
7149 char *buf)
7150{
7151 struct slab_attribute *attribute;
7152 struct kmem_cache *s;
7153
7154 attribute = to_slab_attr(attr);
7155 s = to_slab(kobj);
7156
7157 if (!attribute->show)
7158 return -EIO;
7159
7160 return attribute->show(s, buf);
7161}
7162
7163static ssize_t slab_attr_store(struct kobject *kobj,
7164 struct attribute *attr,
7165 const char *buf, size_t len)
7166{
7167 struct slab_attribute *attribute;
7168 struct kmem_cache *s;
7169
7170 attribute = to_slab_attr(attr);
7171 s = to_slab(kobj);
7172
7173 if (!attribute->store)
7174 return -EIO;
7175
7176 return attribute->store(s, buf, len);
7177}
7178
7179static void kmem_cache_release(struct kobject *k)
7180{
7181 slab_kmem_cache_release(to_slab(k));
7182}
7183
7184static const struct sysfs_ops slab_sysfs_ops = {
7185 .show = slab_attr_show,
7186 .store = slab_attr_store,
7187};
7188
7189static const struct kobj_type slab_ktype = {
7190 .sysfs_ops = &slab_sysfs_ops,
7191 .release = kmem_cache_release,
7192};
7193
7194static struct kset *slab_kset;
7195
7196static inline struct kset *cache_kset(struct kmem_cache *s)
7197{
7198 return slab_kset;
7199}
7200
7201#define ID_STR_LENGTH 32
7202
7203/* Create a unique string id for a slab cache:
7204 *
7205 * Format :[flags-]size
7206 */
7207static char *create_unique_id(struct kmem_cache *s)
7208{
7209 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
7210 char *p = name;
7211
7212 if (!name)
7213 return ERR_PTR(-ENOMEM);
7214
7215 *p++ = ':';
7216 /*
7217 * First flags affecting slabcache operations. We will only
7218 * get here for aliasable slabs so we do not need to support
7219 * too many flags. The flags here must cover all flags that
7220 * are matched during merging to guarantee that the id is
7221 * unique.
7222 */
7223 if (s->flags & SLAB_CACHE_DMA)
7224 *p++ = 'd';
7225 if (s->flags & SLAB_CACHE_DMA32)
7226 *p++ = 'D';
7227 if (s->flags & SLAB_RECLAIM_ACCOUNT)
7228 *p++ = 'a';
7229 if (s->flags & SLAB_CONSISTENCY_CHECKS)
7230 *p++ = 'F';
7231 if (s->flags & SLAB_ACCOUNT)
7232 *p++ = 'A';
7233 if (p != name + 1)
7234 *p++ = '-';
7235 p += snprintf(p, ID_STR_LENGTH - (p - name), "%07u", s->size);
7236
7237 if (WARN_ON(p > name + ID_STR_LENGTH - 1)) {
7238 kfree(name);
7239 return ERR_PTR(-EINVAL);
7240 }
7241 kmsan_unpoison_memory(name, p - name);
7242 return name;
7243}
7244
7245static int sysfs_slab_add(struct kmem_cache *s)
7246{
7247 int err;
7248 const char *name;
7249 struct kset *kset = cache_kset(s);
7250 int unmergeable = slab_unmergeable(s);
7251
7252 if (!unmergeable && disable_higher_order_debug &&
7253 (slub_debug & DEBUG_METADATA_FLAGS))
7254 unmergeable = 1;
7255
7256 if (unmergeable) {
7257 /*
7258 * Slabcache can never be merged so we can use the name proper.
7259 * This is typically the case for debug situations. In that
7260 * case we can catch duplicate names easily.
7261 */
7262 sysfs_remove_link(&slab_kset->kobj, s->name);
7263 name = s->name;
7264 } else {
7265 /*
7266 * Create a unique name for the slab as a target
7267 * for the symlinks.
7268 */
7269 name = create_unique_id(s);
7270 if (IS_ERR(name))
7271 return PTR_ERR(name);
7272 }
7273
7274 s->kobj.kset = kset;
7275 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
7276 if (err)
7277 goto out;
7278
7279 err = sysfs_create_group(&s->kobj, &slab_attr_group);
7280 if (err)
7281 goto out_del_kobj;
7282
7283 if (!unmergeable) {
7284 /* Setup first alias */
7285 sysfs_slab_alias(s, s->name);
7286 }
7287out:
7288 if (!unmergeable)
7289 kfree(name);
7290 return err;
7291out_del_kobj:
7292 kobject_del(&s->kobj);
7293 goto out;
7294}
7295
7296void sysfs_slab_unlink(struct kmem_cache *s)
7297{
7298 if (s->kobj.state_in_sysfs)
7299 kobject_del(&s->kobj);
7300}
7301
7302void sysfs_slab_release(struct kmem_cache *s)
7303{
7304 kobject_put(&s->kobj);
7305}
7306
7307/*
7308 * Need to buffer aliases during bootup until sysfs becomes
7309 * available lest we lose that information.
7310 */
7311struct saved_alias {
7312 struct kmem_cache *s;
7313 const char *name;
7314 struct saved_alias *next;
7315};
7316
7317static struct saved_alias *alias_list;
7318
7319static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
7320{
7321 struct saved_alias *al;
7322
7323 if (slab_state == FULL) {
7324 /*
7325 * If we have a leftover link then remove it.
7326 */
7327 sysfs_remove_link(&slab_kset->kobj, name);
7328 /*
7329 * The original cache may have failed to generate sysfs file.
7330 * In that case, sysfs_create_link() returns -ENOENT and
7331 * symbolic link creation is skipped.
7332 */
7333 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
7334 }
7335
7336 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
7337 if (!al)
7338 return -ENOMEM;
7339
7340 al->s = s;
7341 al->name = name;
7342 al->next = alias_list;
7343 alias_list = al;
7344 kmsan_unpoison_memory(al, sizeof(*al));
7345 return 0;
7346}
7347
7348static int __init slab_sysfs_init(void)
7349{
7350 struct kmem_cache *s;
7351 int err;
7352
7353 mutex_lock(&slab_mutex);
7354
7355 slab_kset = kset_create_and_add("slab", NULL, kernel_kobj);
7356 if (!slab_kset) {
7357 mutex_unlock(&slab_mutex);
7358 pr_err("Cannot register slab subsystem.\n");
7359 return -ENOMEM;
7360 }
7361
7362 slab_state = FULL;
7363
7364 list_for_each_entry(s, &slab_caches, list) {
7365 err = sysfs_slab_add(s);
7366 if (err)
7367 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
7368 s->name);
7369 }
7370
7371 while (alias_list) {
7372 struct saved_alias *al = alias_list;
7373
7374 alias_list = alias_list->next;
7375 err = sysfs_slab_alias(al->s, al->name);
7376 if (err)
7377 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
7378 al->name);
7379 kfree(al);
7380 }
7381
7382 mutex_unlock(&slab_mutex);
7383 return 0;
7384}
7385late_initcall(slab_sysfs_init);
7386#endif /* SLAB_SUPPORTS_SYSFS */
7387
7388#if defined(CONFIG_SLUB_DEBUG) && defined(CONFIG_DEBUG_FS)
7389static int slab_debugfs_show(struct seq_file *seq, void *v)
7390{
7391 struct loc_track *t = seq->private;
7392 struct location *l;
7393 unsigned long idx;
7394
7395 idx = (unsigned long) t->idx;
7396 if (idx < t->count) {
7397 l = &t->loc[idx];
7398
7399 seq_printf(seq, "%7ld ", l->count);
7400
7401 if (l->addr)
7402 seq_printf(seq, "%pS", (void *)l->addr);
7403 else
7404 seq_puts(seq, "<not-available>");
7405
7406 if (l->waste)
7407 seq_printf(seq, " waste=%lu/%lu",
7408 l->count * l->waste, l->waste);
7409
7410 if (l->sum_time != l->min_time) {
7411 seq_printf(seq, " age=%ld/%llu/%ld",
7412 l->min_time, div_u64(l->sum_time, l->count),
7413 l->max_time);
7414 } else
7415 seq_printf(seq, " age=%ld", l->min_time);
7416
7417 if (l->min_pid != l->max_pid)
7418 seq_printf(seq, " pid=%ld-%ld", l->min_pid, l->max_pid);
7419 else
7420 seq_printf(seq, " pid=%ld",
7421 l->min_pid);
7422
7423 if (num_online_cpus() > 1 && !cpumask_empty(to_cpumask(l->cpus)))
7424 seq_printf(seq, " cpus=%*pbl",
7425 cpumask_pr_args(to_cpumask(l->cpus)));
7426
7427 if (nr_online_nodes > 1 && !nodes_empty(l->nodes))
7428 seq_printf(seq, " nodes=%*pbl",
7429 nodemask_pr_args(&l->nodes));
7430
7431#ifdef CONFIG_STACKDEPOT
7432 {
7433 depot_stack_handle_t handle;
7434 unsigned long *entries;
7435 unsigned int nr_entries, j;
7436
7437 handle = READ_ONCE(l->handle);
7438 if (handle) {
7439 nr_entries = stack_depot_fetch(handle, &entries);
7440 seq_puts(seq, "\n");
7441 for (j = 0; j < nr_entries; j++)
7442 seq_printf(seq, " %pS\n", (void *)entries[j]);
7443 }
7444 }
7445#endif
7446 seq_puts(seq, "\n");
7447 }
7448
7449 if (!idx && !t->count)
7450 seq_puts(seq, "No data\n");
7451
7452 return 0;
7453}
7454
7455static void slab_debugfs_stop(struct seq_file *seq, void *v)
7456{
7457}
7458
7459static void *slab_debugfs_next(struct seq_file *seq, void *v, loff_t *ppos)
7460{
7461 struct loc_track *t = seq->private;
7462
7463 t->idx = ++(*ppos);
7464 if (*ppos <= t->count)
7465 return ppos;
7466
7467 return NULL;
7468}
7469
7470static int cmp_loc_by_count(const void *a, const void *b, const void *data)
7471{
7472 struct location *loc1 = (struct location *)a;
7473 struct location *loc2 = (struct location *)b;
7474
7475 if (loc1->count > loc2->count)
7476 return -1;
7477 else
7478 return 1;
7479}
7480
7481static void *slab_debugfs_start(struct seq_file *seq, loff_t *ppos)
7482{
7483 struct loc_track *t = seq->private;
7484
7485 t->idx = *ppos;
7486 return ppos;
7487}
7488
7489static const struct seq_operations slab_debugfs_sops = {
7490 .start = slab_debugfs_start,
7491 .next = slab_debugfs_next,
7492 .stop = slab_debugfs_stop,
7493 .show = slab_debugfs_show,
7494};
7495
7496static int slab_debug_trace_open(struct inode *inode, struct file *filep)
7497{
7498
7499 struct kmem_cache_node *n;
7500 enum track_item alloc;
7501 int node;
7502 struct loc_track *t = __seq_open_private(filep, &slab_debugfs_sops,
7503 sizeof(struct loc_track));
7504 struct kmem_cache *s = file_inode(filep)->i_private;
7505 unsigned long *obj_map;
7506
7507 if (!t)
7508 return -ENOMEM;
7509
7510 obj_map = bitmap_alloc(oo_objects(s->oo), GFP_KERNEL);
7511 if (!obj_map) {
7512 seq_release_private(inode, filep);
7513 return -ENOMEM;
7514 }
7515
7516 if (strcmp(filep->f_path.dentry->d_name.name, "alloc_traces") == 0)
7517 alloc = TRACK_ALLOC;
7518 else
7519 alloc = TRACK_FREE;
7520
7521 if (!alloc_loc_track(t, PAGE_SIZE / sizeof(struct location), GFP_KERNEL)) {
7522 bitmap_free(obj_map);
7523 seq_release_private(inode, filep);
7524 return -ENOMEM;
7525 }
7526
7527 for_each_kmem_cache_node(s, node, n) {
7528 unsigned long flags;
7529 struct slab *slab;
7530
7531 if (!node_nr_slabs(n))
7532 continue;
7533
7534 spin_lock_irqsave(&n->list_lock, flags);
7535 list_for_each_entry(slab, &n->partial, slab_list)
7536 process_slab(t, s, slab, alloc, obj_map);
7537 list_for_each_entry(slab, &n->full, slab_list)
7538 process_slab(t, s, slab, alloc, obj_map);
7539 spin_unlock_irqrestore(&n->list_lock, flags);
7540 }
7541
7542 /* Sort locations by count */
7543 sort_r(t->loc, t->count, sizeof(struct location),
7544 cmp_loc_by_count, NULL, NULL);
7545
7546 bitmap_free(obj_map);
7547 return 0;
7548}
7549
7550static int slab_debug_trace_release(struct inode *inode, struct file *file)
7551{
7552 struct seq_file *seq = file->private_data;
7553 struct loc_track *t = seq->private;
7554
7555 free_loc_track(t);
7556 return seq_release_private(inode, file);
7557}
7558
7559static const struct file_operations slab_debugfs_fops = {
7560 .open = slab_debug_trace_open,
7561 .read = seq_read,
7562 .llseek = seq_lseek,
7563 .release = slab_debug_trace_release,
7564};
7565
7566static void debugfs_slab_add(struct kmem_cache *s)
7567{
7568 struct dentry *slab_cache_dir;
7569
7570 if (unlikely(!slab_debugfs_root))
7571 return;
7572
7573 slab_cache_dir = debugfs_create_dir(s->name, slab_debugfs_root);
7574
7575 debugfs_create_file("alloc_traces", 0400,
7576 slab_cache_dir, s, &slab_debugfs_fops);
7577
7578 debugfs_create_file("free_traces", 0400,
7579 slab_cache_dir, s, &slab_debugfs_fops);
7580}
7581
7582void debugfs_slab_release(struct kmem_cache *s)
7583{
7584 debugfs_lookup_and_remove(s->name, slab_debugfs_root);
7585}
7586
7587static int __init slab_debugfs_init(void)
7588{
7589 struct kmem_cache *s;
7590
7591 slab_debugfs_root = debugfs_create_dir("slab", NULL);
7592
7593 list_for_each_entry(s, &slab_caches, list)
7594 if (s->flags & SLAB_STORE_USER)
7595 debugfs_slab_add(s);
7596
7597 return 0;
7598
7599}
7600__initcall(slab_debugfs_init);
7601#endif
7602/*
7603 * The /proc/slabinfo ABI
7604 */
7605#ifdef CONFIG_SLUB_DEBUG
7606void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
7607{
7608 unsigned long nr_slabs = 0;
7609 unsigned long nr_objs = 0;
7610 unsigned long nr_free = 0;
7611 int node;
7612 struct kmem_cache_node *n;
7613
7614 for_each_kmem_cache_node(s, node, n) {
7615 nr_slabs += node_nr_slabs(n);
7616 nr_objs += node_nr_objs(n);
7617 nr_free += count_partial_free_approx(n);
7618 }
7619
7620 sinfo->active_objs = nr_objs - nr_free;
7621 sinfo->num_objs = nr_objs;
7622 sinfo->active_slabs = nr_slabs;
7623 sinfo->num_slabs = nr_slabs;
7624 sinfo->objects_per_slab = oo_objects(s->oo);
7625 sinfo->cache_order = oo_order(s->oo);
7626}
7627#endif /* CONFIG_SLUB_DEBUG */