Loading...
1/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
10 */
11
12#include <linux/mm.h>
13#include <linux/swap.h> /* struct reclaim_state */
14#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
19#include "slab.h"
20#include <linux/proc_fs.h>
21#include <linux/notifier.h>
22#include <linux/seq_file.h>
23#include <linux/kasan.h>
24#include <linux/kmemcheck.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
38#include <trace/events/kmem.h>
39
40#include "internal.h"
41
42/*
43 * Lock order:
44 * 1. slab_mutex (Global Mutex)
45 * 2. node->list_lock
46 * 3. slab_lock(page) (Only on some arches and for debugging)
47 *
48 * slab_mutex
49 *
50 * The role of the slab_mutex is to protect the list of all the slabs
51 * and to synchronize major metadata changes to slab cache structures.
52 *
53 * The slab_lock is only used for debugging and on arches that do not
54 * have the ability to do a cmpxchg_double. It only protects the second
55 * double word in the page struct. Meaning
56 * A. page->freelist -> List of object free in a page
57 * B. page->counters -> Counters of objects
58 * C. page->frozen -> frozen state
59 *
60 * If a slab is frozen then it is exempt from list management. It is not
61 * on any list. The processor that froze the slab is the one who can
62 * perform list operations on the page. Other processors may put objects
63 * onto the freelist but the processor that froze the slab is the only
64 * one that can retrieve the objects from the page's freelist.
65 *
66 * The list_lock protects the partial and full list on each node and
67 * the partial slab counter. If taken then no new slabs may be added or
68 * removed from the lists nor make the number of partial slabs be modified.
69 * (Note that the total number of slabs is an atomic value that may be
70 * modified without taking the list lock).
71 *
72 * The list_lock is a centralized lock and thus we avoid taking it as
73 * much as possible. As long as SLUB does not have to handle partial
74 * slabs, operations can continue without any centralized lock. F.e.
75 * allocating a long series of objects that fill up slabs does not require
76 * the list lock.
77 * Interrupts are disabled during allocation and deallocation in order to
78 * make the slab allocator safe to use in the context of an irq. In addition
79 * interrupts are disabled to ensure that the processor does not change
80 * while handling per_cpu slabs, due to kernel preemption.
81 *
82 * SLUB assigns one slab for allocation to each processor.
83 * Allocations only occur from these slabs called cpu slabs.
84 *
85 * Slabs with free elements are kept on a partial list and during regular
86 * operations no list for full slabs is used. If an object in a full slab is
87 * freed then the slab will show up again on the partial lists.
88 * We track full slabs for debugging purposes though because otherwise we
89 * cannot scan all objects.
90 *
91 * Slabs are freed when they become empty. Teardown and setup is
92 * minimal so we rely on the page allocators per cpu caches for
93 * fast frees and allocs.
94 *
95 * Overloading of page flags that are otherwise used for LRU management.
96 *
97 * PageActive The slab is frozen and exempt from list processing.
98 * This means that the slab is dedicated to a purpose
99 * such as satisfying allocations for a specific
100 * processor. Objects may be freed in the slab while
101 * it is frozen but slab_free will then skip the usual
102 * list operations. It is up to the processor holding
103 * the slab to integrate the slab into the slab lists
104 * when the slab is no longer needed.
105 *
106 * One use of this flag is to mark slabs that are
107 * used for allocations. Then such a slab becomes a cpu
108 * slab. The cpu slab may be equipped with an additional
109 * freelist that allows lockless access to
110 * free objects in addition to the regular freelist
111 * that requires the slab lock.
112 *
113 * PageError Slab requires special handling due to debug
114 * options set. This moves slab handling out of
115 * the fast path and disables lockless freelists.
116 */
117
118static inline int kmem_cache_debug(struct kmem_cache *s)
119{
120#ifdef CONFIG_SLUB_DEBUG
121 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
122#else
123 return 0;
124#endif
125}
126
127static inline void *fixup_red_left(struct kmem_cache *s, void *p)
128{
129 if (kmem_cache_debug(s) && s->flags & SLAB_RED_ZONE)
130 p += s->red_left_pad;
131
132 return p;
133}
134
135static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
136{
137#ifdef CONFIG_SLUB_CPU_PARTIAL
138 return !kmem_cache_debug(s);
139#else
140 return false;
141#endif
142}
143
144/*
145 * Issues still to be resolved:
146 *
147 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 *
149 * - Variable sizing of the per node arrays
150 */
151
152/* Enable to test recovery from slab corruption on boot */
153#undef SLUB_RESILIENCY_TEST
154
155/* Enable to log cmpxchg failures */
156#undef SLUB_DEBUG_CMPXCHG
157
158/*
159 * Mininum number of partial slabs. These will be left on the partial
160 * lists even if they are empty. kmem_cache_shrink may reclaim them.
161 */
162#define MIN_PARTIAL 5
163
164/*
165 * Maximum number of desirable partial slabs.
166 * The existence of more partial slabs makes kmem_cache_shrink
167 * sort the partial list by the number of objects in use.
168 */
169#define MAX_PARTIAL 10
170
171#define DEBUG_DEFAULT_FLAGS (SLAB_CONSISTENCY_CHECKS | SLAB_RED_ZONE | \
172 SLAB_POISON | SLAB_STORE_USER)
173
174/*
175 * These debug flags cannot use CMPXCHG because there might be consistency
176 * issues when checking or reading debug information
177 */
178#define SLAB_NO_CMPXCHG (SLAB_CONSISTENCY_CHECKS | SLAB_STORE_USER | \
179 SLAB_TRACE)
180
181
182/*
183 * Debugging flags that require metadata to be stored in the slab. These get
184 * disabled when slub_debug=O is used and a cache's min order increases with
185 * metadata.
186 */
187#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
188
189#define OO_SHIFT 16
190#define OO_MASK ((1 << OO_SHIFT) - 1)
191#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
192
193/* Internal SLUB flags */
194#define __OBJECT_POISON 0x80000000UL /* Poison object */
195#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
196
197#ifdef CONFIG_SMP
198static struct notifier_block slab_notifier;
199#endif
200
201/*
202 * Tracking user of a slab.
203 */
204#define TRACK_ADDRS_COUNT 16
205struct track {
206 unsigned long addr; /* Called from address */
207#ifdef CONFIG_STACKTRACE
208 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
209#endif
210 int cpu; /* Was running on cpu */
211 int pid; /* Pid context */
212 unsigned long when; /* When did the operation occur */
213};
214
215enum track_item { TRACK_ALLOC, TRACK_FREE };
216
217#ifdef CONFIG_SYSFS
218static int sysfs_slab_add(struct kmem_cache *);
219static int sysfs_slab_alias(struct kmem_cache *, const char *);
220static void memcg_propagate_slab_attrs(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) { }
226#endif
227
228static inline void stat(const struct kmem_cache *s, enum stat_item si)
229{
230#ifdef CONFIG_SLUB_STATS
231 /*
232 * The rmw is racy on a preemptible kernel but this is acceptable, so
233 * avoid this_cpu_add()'s irq-disable overhead.
234 */
235 raw_cpu_inc(s->cpu_slab->stat[si]);
236#endif
237}
238
239/********************************************************************
240 * Core slab cache functions
241 *******************************************************************/
242
243static inline void *get_freepointer(struct kmem_cache *s, void *object)
244{
245 return *(void **)(object + s->offset);
246}
247
248static void prefetch_freepointer(const struct kmem_cache *s, void *object)
249{
250 prefetch(object + s->offset);
251}
252
253static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
254{
255 void *p;
256
257 if (!debug_pagealloc_enabled())
258 return get_freepointer(s, object);
259
260 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
261 return p;
262}
263
264static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
265{
266 *(void **)(object + s->offset) = fp;
267}
268
269/* Loop over all objects in a slab */
270#define for_each_object(__p, __s, __addr, __objects) \
271 for (__p = fixup_red_left(__s, __addr); \
272 __p < (__addr) + (__objects) * (__s)->size; \
273 __p += (__s)->size)
274
275#define for_each_object_idx(__p, __idx, __s, __addr, __objects) \
276 for (__p = fixup_red_left(__s, __addr), __idx = 1; \
277 __idx <= __objects; \
278 __p += (__s)->size, __idx++)
279
280/* Determine object index from a given position */
281static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
282{
283 return (p - addr) / s->size;
284}
285
286static inline int order_objects(int order, unsigned long size, int reserved)
287{
288 return ((PAGE_SIZE << order) - reserved) / size;
289}
290
291static inline struct kmem_cache_order_objects oo_make(int order,
292 unsigned long size, int reserved)
293{
294 struct kmem_cache_order_objects x = {
295 (order << OO_SHIFT) + order_objects(order, size, reserved)
296 };
297
298 return x;
299}
300
301static inline int oo_order(struct kmem_cache_order_objects x)
302{
303 return x.x >> OO_SHIFT;
304}
305
306static inline int oo_objects(struct kmem_cache_order_objects x)
307{
308 return x.x & OO_MASK;
309}
310
311/*
312 * Per slab locking using the pagelock
313 */
314static __always_inline void slab_lock(struct page *page)
315{
316 VM_BUG_ON_PAGE(PageTail(page), page);
317 bit_spin_lock(PG_locked, &page->flags);
318}
319
320static __always_inline void slab_unlock(struct page *page)
321{
322 VM_BUG_ON_PAGE(PageTail(page), page);
323 __bit_spin_unlock(PG_locked, &page->flags);
324}
325
326static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
327{
328 struct page tmp;
329 tmp.counters = counters_new;
330 /*
331 * page->counters can cover frozen/inuse/objects as well
332 * as page->_count. If we assign to ->counters directly
333 * we run the risk of losing updates to page->_count, so
334 * be careful and only assign to the fields we need.
335 */
336 page->frozen = tmp.frozen;
337 page->inuse = tmp.inuse;
338 page->objects = tmp.objects;
339}
340
341/* Interrupts must be disabled (for the fallback code to work right) */
342static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
343 void *freelist_old, unsigned long counters_old,
344 void *freelist_new, unsigned long counters_new,
345 const char *n)
346{
347 VM_BUG_ON(!irqs_disabled());
348#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
349 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
350 if (s->flags & __CMPXCHG_DOUBLE) {
351 if (cmpxchg_double(&page->freelist, &page->counters,
352 freelist_old, counters_old,
353 freelist_new, counters_new))
354 return true;
355 } else
356#endif
357 {
358 slab_lock(page);
359 if (page->freelist == freelist_old &&
360 page->counters == counters_old) {
361 page->freelist = freelist_new;
362 set_page_slub_counters(page, counters_new);
363 slab_unlock(page);
364 return true;
365 }
366 slab_unlock(page);
367 }
368
369 cpu_relax();
370 stat(s, CMPXCHG_DOUBLE_FAIL);
371
372#ifdef SLUB_DEBUG_CMPXCHG
373 pr_info("%s %s: cmpxchg double redo ", n, s->name);
374#endif
375
376 return false;
377}
378
379static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
380 void *freelist_old, unsigned long counters_old,
381 void *freelist_new, unsigned long counters_new,
382 const char *n)
383{
384#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
385 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
386 if (s->flags & __CMPXCHG_DOUBLE) {
387 if (cmpxchg_double(&page->freelist, &page->counters,
388 freelist_old, counters_old,
389 freelist_new, counters_new))
390 return true;
391 } else
392#endif
393 {
394 unsigned long flags;
395
396 local_irq_save(flags);
397 slab_lock(page);
398 if (page->freelist == freelist_old &&
399 page->counters == counters_old) {
400 page->freelist = freelist_new;
401 set_page_slub_counters(page, counters_new);
402 slab_unlock(page);
403 local_irq_restore(flags);
404 return true;
405 }
406 slab_unlock(page);
407 local_irq_restore(flags);
408 }
409
410 cpu_relax();
411 stat(s, CMPXCHG_DOUBLE_FAIL);
412
413#ifdef SLUB_DEBUG_CMPXCHG
414 pr_info("%s %s: cmpxchg double redo ", n, s->name);
415#endif
416
417 return false;
418}
419
420#ifdef CONFIG_SLUB_DEBUG
421/*
422 * Determine a map of object in use on a page.
423 *
424 * Node listlock must be held to guarantee that the page does
425 * not vanish from under us.
426 */
427static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
428{
429 void *p;
430 void *addr = page_address(page);
431
432 for (p = page->freelist; p; p = get_freepointer(s, p))
433 set_bit(slab_index(p, s, addr), map);
434}
435
436static inline int size_from_object(struct kmem_cache *s)
437{
438 if (s->flags & SLAB_RED_ZONE)
439 return s->size - s->red_left_pad;
440
441 return s->size;
442}
443
444static inline void *restore_red_left(struct kmem_cache *s, void *p)
445{
446 if (s->flags & SLAB_RED_ZONE)
447 p -= s->red_left_pad;
448
449 return p;
450}
451
452/*
453 * Debug settings:
454 */
455#if defined(CONFIG_SLUB_DEBUG_ON)
456static int slub_debug = DEBUG_DEFAULT_FLAGS;
457#elif defined(CONFIG_KASAN)
458static int slub_debug = SLAB_STORE_USER;
459#else
460static int slub_debug;
461#endif
462
463static char *slub_debug_slabs;
464static int disable_higher_order_debug;
465
466/*
467 * slub is about to manipulate internal object metadata. This memory lies
468 * outside the range of the allocated object, so accessing it would normally
469 * be reported by kasan as a bounds error. metadata_access_enable() is used
470 * to tell kasan that these accesses are OK.
471 */
472static inline void metadata_access_enable(void)
473{
474 kasan_disable_current();
475}
476
477static inline void metadata_access_disable(void)
478{
479 kasan_enable_current();
480}
481
482/*
483 * Object debugging
484 */
485
486/* Verify that a pointer has an address that is valid within a slab page */
487static inline int check_valid_pointer(struct kmem_cache *s,
488 struct page *page, void *object)
489{
490 void *base;
491
492 if (!object)
493 return 1;
494
495 base = page_address(page);
496 object = restore_red_left(s, object);
497 if (object < base || object >= base + page->objects * s->size ||
498 (object - base) % s->size) {
499 return 0;
500 }
501
502 return 1;
503}
504
505static void print_section(char *text, u8 *addr, unsigned int length)
506{
507 metadata_access_enable();
508 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
509 length, 1);
510 metadata_access_disable();
511}
512
513static struct track *get_track(struct kmem_cache *s, void *object,
514 enum track_item alloc)
515{
516 struct track *p;
517
518 if (s->offset)
519 p = object + s->offset + sizeof(void *);
520 else
521 p = object + s->inuse;
522
523 return p + alloc;
524}
525
526static void set_track(struct kmem_cache *s, void *object,
527 enum track_item alloc, unsigned long addr)
528{
529 struct track *p = get_track(s, object, alloc);
530
531 if (addr) {
532#ifdef CONFIG_STACKTRACE
533 struct stack_trace trace;
534 int i;
535
536 trace.nr_entries = 0;
537 trace.max_entries = TRACK_ADDRS_COUNT;
538 trace.entries = p->addrs;
539 trace.skip = 3;
540 metadata_access_enable();
541 save_stack_trace(&trace);
542 metadata_access_disable();
543
544 /* See rant in lockdep.c */
545 if (trace.nr_entries != 0 &&
546 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
547 trace.nr_entries--;
548
549 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
550 p->addrs[i] = 0;
551#endif
552 p->addr = addr;
553 p->cpu = smp_processor_id();
554 p->pid = current->pid;
555 p->when = jiffies;
556 } else
557 memset(p, 0, sizeof(struct track));
558}
559
560static void init_tracking(struct kmem_cache *s, void *object)
561{
562 if (!(s->flags & SLAB_STORE_USER))
563 return;
564
565 set_track(s, object, TRACK_FREE, 0UL);
566 set_track(s, object, TRACK_ALLOC, 0UL);
567}
568
569static void print_track(const char *s, struct track *t)
570{
571 if (!t->addr)
572 return;
573
574 pr_err("INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
575 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
576#ifdef CONFIG_STACKTRACE
577 {
578 int i;
579 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
580 if (t->addrs[i])
581 pr_err("\t%pS\n", (void *)t->addrs[i]);
582 else
583 break;
584 }
585#endif
586}
587
588static void print_tracking(struct kmem_cache *s, void *object)
589{
590 if (!(s->flags & SLAB_STORE_USER))
591 return;
592
593 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
594 print_track("Freed", get_track(s, object, TRACK_FREE));
595}
596
597static void print_page_info(struct page *page)
598{
599 pr_err("INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
600 page, page->objects, page->inuse, page->freelist, page->flags);
601
602}
603
604static void slab_bug(struct kmem_cache *s, char *fmt, ...)
605{
606 struct va_format vaf;
607 va_list args;
608
609 va_start(args, fmt);
610 vaf.fmt = fmt;
611 vaf.va = &args;
612 pr_err("=============================================================================\n");
613 pr_err("BUG %s (%s): %pV\n", s->name, print_tainted(), &vaf);
614 pr_err("-----------------------------------------------------------------------------\n\n");
615
616 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
617 va_end(args);
618}
619
620static void slab_fix(struct kmem_cache *s, char *fmt, ...)
621{
622 struct va_format vaf;
623 va_list args;
624
625 va_start(args, fmt);
626 vaf.fmt = fmt;
627 vaf.va = &args;
628 pr_err("FIX %s: %pV\n", s->name, &vaf);
629 va_end(args);
630}
631
632static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
633{
634 unsigned int off; /* Offset of last byte */
635 u8 *addr = page_address(page);
636
637 print_tracking(s, p);
638
639 print_page_info(page);
640
641 pr_err("INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
642 p, p - addr, get_freepointer(s, p));
643
644 if (s->flags & SLAB_RED_ZONE)
645 print_section("Redzone ", p - s->red_left_pad, s->red_left_pad);
646 else if (p > addr + 16)
647 print_section("Bytes b4 ", p - 16, 16);
648
649 print_section("Object ", p, min_t(unsigned long, s->object_size,
650 PAGE_SIZE));
651 if (s->flags & SLAB_RED_ZONE)
652 print_section("Redzone ", p + s->object_size,
653 s->inuse - s->object_size);
654
655 if (s->offset)
656 off = s->offset + sizeof(void *);
657 else
658 off = s->inuse;
659
660 if (s->flags & SLAB_STORE_USER)
661 off += 2 * sizeof(struct track);
662
663 if (off != size_from_object(s))
664 /* Beginning of the filler is the free pointer */
665 print_section("Padding ", p + off, size_from_object(s) - off);
666
667 dump_stack();
668}
669
670void object_err(struct kmem_cache *s, struct page *page,
671 u8 *object, char *reason)
672{
673 slab_bug(s, "%s", reason);
674 print_trailer(s, page, object);
675}
676
677static void slab_err(struct kmem_cache *s, struct page *page,
678 const char *fmt, ...)
679{
680 va_list args;
681 char buf[100];
682
683 va_start(args, fmt);
684 vsnprintf(buf, sizeof(buf), fmt, args);
685 va_end(args);
686 slab_bug(s, "%s", buf);
687 print_page_info(page);
688 dump_stack();
689}
690
691static void init_object(struct kmem_cache *s, void *object, u8 val)
692{
693 u8 *p = object;
694
695 if (s->flags & SLAB_RED_ZONE)
696 memset(p - s->red_left_pad, val, s->red_left_pad);
697
698 if (s->flags & __OBJECT_POISON) {
699 memset(p, POISON_FREE, s->object_size - 1);
700 p[s->object_size - 1] = POISON_END;
701 }
702
703 if (s->flags & SLAB_RED_ZONE)
704 memset(p + s->object_size, val, s->inuse - s->object_size);
705}
706
707static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
708 void *from, void *to)
709{
710 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
711 memset(from, data, to - from);
712}
713
714static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
715 u8 *object, char *what,
716 u8 *start, unsigned int value, unsigned int bytes)
717{
718 u8 *fault;
719 u8 *end;
720
721 metadata_access_enable();
722 fault = memchr_inv(start, value, bytes);
723 metadata_access_disable();
724 if (!fault)
725 return 1;
726
727 end = start + bytes;
728 while (end > fault && end[-1] == value)
729 end--;
730
731 slab_bug(s, "%s overwritten", what);
732 pr_err("INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
733 fault, end - 1, fault[0], value);
734 print_trailer(s, page, object);
735
736 restore_bytes(s, what, value, fault, end);
737 return 0;
738}
739
740/*
741 * Object layout:
742 *
743 * object address
744 * Bytes of the object to be managed.
745 * If the freepointer may overlay the object then the free
746 * pointer is the first word of the object.
747 *
748 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
749 * 0xa5 (POISON_END)
750 *
751 * object + s->object_size
752 * Padding to reach word boundary. This is also used for Redzoning.
753 * Padding is extended by another word if Redzoning is enabled and
754 * object_size == inuse.
755 *
756 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
757 * 0xcc (RED_ACTIVE) for objects in use.
758 *
759 * object + s->inuse
760 * Meta data starts here.
761 *
762 * A. Free pointer (if we cannot overwrite object on free)
763 * B. Tracking data for SLAB_STORE_USER
764 * C. Padding to reach required alignment boundary or at mininum
765 * one word if debugging is on to be able to detect writes
766 * before the word boundary.
767 *
768 * Padding is done using 0x5a (POISON_INUSE)
769 *
770 * object + s->size
771 * Nothing is used beyond s->size.
772 *
773 * If slabcaches are merged then the object_size and inuse boundaries are mostly
774 * ignored. And therefore no slab options that rely on these boundaries
775 * may be used with merged slabcaches.
776 */
777
778static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
779{
780 unsigned long off = s->inuse; /* The end of info */
781
782 if (s->offset)
783 /* Freepointer is placed after the object. */
784 off += sizeof(void *);
785
786 if (s->flags & SLAB_STORE_USER)
787 /* We also have user information there */
788 off += 2 * sizeof(struct track);
789
790 if (size_from_object(s) == off)
791 return 1;
792
793 return check_bytes_and_report(s, page, p, "Object padding",
794 p + off, POISON_INUSE, size_from_object(s) - off);
795}
796
797/* Check the pad bytes at the end of a slab page */
798static int slab_pad_check(struct kmem_cache *s, struct page *page)
799{
800 u8 *start;
801 u8 *fault;
802 u8 *end;
803 int length;
804 int remainder;
805
806 if (!(s->flags & SLAB_POISON))
807 return 1;
808
809 start = page_address(page);
810 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
811 end = start + length;
812 remainder = length % s->size;
813 if (!remainder)
814 return 1;
815
816 metadata_access_enable();
817 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
818 metadata_access_disable();
819 if (!fault)
820 return 1;
821 while (end > fault && end[-1] == POISON_INUSE)
822 end--;
823
824 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
825 print_section("Padding ", end - remainder, remainder);
826
827 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
828 return 0;
829}
830
831static int check_object(struct kmem_cache *s, struct page *page,
832 void *object, u8 val)
833{
834 u8 *p = object;
835 u8 *endobject = object + s->object_size;
836
837 if (s->flags & SLAB_RED_ZONE) {
838 if (!check_bytes_and_report(s, page, object, "Redzone",
839 object - s->red_left_pad, val, s->red_left_pad))
840 return 0;
841
842 if (!check_bytes_and_report(s, page, object, "Redzone",
843 endobject, val, s->inuse - s->object_size))
844 return 0;
845 } else {
846 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
847 check_bytes_and_report(s, page, p, "Alignment padding",
848 endobject, POISON_INUSE,
849 s->inuse - s->object_size);
850 }
851 }
852
853 if (s->flags & SLAB_POISON) {
854 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
855 (!check_bytes_and_report(s, page, p, "Poison", p,
856 POISON_FREE, s->object_size - 1) ||
857 !check_bytes_and_report(s, page, p, "Poison",
858 p + s->object_size - 1, POISON_END, 1)))
859 return 0;
860 /*
861 * check_pad_bytes cleans up on its own.
862 */
863 check_pad_bytes(s, page, p);
864 }
865
866 if (!s->offset && val == SLUB_RED_ACTIVE)
867 /*
868 * Object and freepointer overlap. Cannot check
869 * freepointer while object is allocated.
870 */
871 return 1;
872
873 /* Check free pointer validity */
874 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
875 object_err(s, page, p, "Freepointer corrupt");
876 /*
877 * No choice but to zap it and thus lose the remainder
878 * of the free objects in this slab. May cause
879 * another error because the object count is now wrong.
880 */
881 set_freepointer(s, p, NULL);
882 return 0;
883 }
884 return 1;
885}
886
887static int check_slab(struct kmem_cache *s, struct page *page)
888{
889 int maxobj;
890
891 VM_BUG_ON(!irqs_disabled());
892
893 if (!PageSlab(page)) {
894 slab_err(s, page, "Not a valid slab page");
895 return 0;
896 }
897
898 maxobj = order_objects(compound_order(page), s->size, s->reserved);
899 if (page->objects > maxobj) {
900 slab_err(s, page, "objects %u > max %u",
901 page->objects, maxobj);
902 return 0;
903 }
904 if (page->inuse > page->objects) {
905 slab_err(s, page, "inuse %u > max %u",
906 page->inuse, page->objects);
907 return 0;
908 }
909 /* Slab_pad_check fixes things up after itself */
910 slab_pad_check(s, page);
911 return 1;
912}
913
914/*
915 * Determine if a certain object on a page is on the freelist. Must hold the
916 * slab lock to guarantee that the chains are in a consistent state.
917 */
918static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
919{
920 int nr = 0;
921 void *fp;
922 void *object = NULL;
923 int max_objects;
924
925 fp = page->freelist;
926 while (fp && nr <= page->objects) {
927 if (fp == search)
928 return 1;
929 if (!check_valid_pointer(s, page, fp)) {
930 if (object) {
931 object_err(s, page, object,
932 "Freechain corrupt");
933 set_freepointer(s, object, NULL);
934 } else {
935 slab_err(s, page, "Freepointer corrupt");
936 page->freelist = NULL;
937 page->inuse = page->objects;
938 slab_fix(s, "Freelist cleared");
939 return 0;
940 }
941 break;
942 }
943 object = fp;
944 fp = get_freepointer(s, object);
945 nr++;
946 }
947
948 max_objects = order_objects(compound_order(page), s->size, s->reserved);
949 if (max_objects > MAX_OBJS_PER_PAGE)
950 max_objects = MAX_OBJS_PER_PAGE;
951
952 if (page->objects != max_objects) {
953 slab_err(s, page, "Wrong number of objects. Found %d but should be %d",
954 page->objects, max_objects);
955 page->objects = max_objects;
956 slab_fix(s, "Number of objects adjusted.");
957 }
958 if (page->inuse != page->objects - nr) {
959 slab_err(s, page, "Wrong object count. Counter is %d but counted were %d",
960 page->inuse, page->objects - nr);
961 page->inuse = page->objects - nr;
962 slab_fix(s, "Object count adjusted.");
963 }
964 return search == NULL;
965}
966
967static void trace(struct kmem_cache *s, struct page *page, void *object,
968 int alloc)
969{
970 if (s->flags & SLAB_TRACE) {
971 pr_info("TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
972 s->name,
973 alloc ? "alloc" : "free",
974 object, page->inuse,
975 page->freelist);
976
977 if (!alloc)
978 print_section("Object ", (void *)object,
979 s->object_size);
980
981 dump_stack();
982 }
983}
984
985/*
986 * Tracking of fully allocated slabs for debugging purposes.
987 */
988static void add_full(struct kmem_cache *s,
989 struct kmem_cache_node *n, struct page *page)
990{
991 if (!(s->flags & SLAB_STORE_USER))
992 return;
993
994 lockdep_assert_held(&n->list_lock);
995 list_add(&page->lru, &n->full);
996}
997
998static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
999{
1000 if (!(s->flags & SLAB_STORE_USER))
1001 return;
1002
1003 lockdep_assert_held(&n->list_lock);
1004 list_del(&page->lru);
1005}
1006
1007/* Tracking of the number of slabs for debugging purposes */
1008static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1009{
1010 struct kmem_cache_node *n = get_node(s, node);
1011
1012 return atomic_long_read(&n->nr_slabs);
1013}
1014
1015static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1016{
1017 return atomic_long_read(&n->nr_slabs);
1018}
1019
1020static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1021{
1022 struct kmem_cache_node *n = get_node(s, node);
1023
1024 /*
1025 * May be called early in order to allocate a slab for the
1026 * kmem_cache_node structure. Solve the chicken-egg
1027 * dilemma by deferring the increment of the count during
1028 * bootstrap (see early_kmem_cache_node_alloc).
1029 */
1030 if (likely(n)) {
1031 atomic_long_inc(&n->nr_slabs);
1032 atomic_long_add(objects, &n->total_objects);
1033 }
1034}
1035static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1036{
1037 struct kmem_cache_node *n = get_node(s, node);
1038
1039 atomic_long_dec(&n->nr_slabs);
1040 atomic_long_sub(objects, &n->total_objects);
1041}
1042
1043/* Object debug checks for alloc/free paths */
1044static void setup_object_debug(struct kmem_cache *s, struct page *page,
1045 void *object)
1046{
1047 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1048 return;
1049
1050 init_object(s, object, SLUB_RED_INACTIVE);
1051 init_tracking(s, object);
1052}
1053
1054static inline int alloc_consistency_checks(struct kmem_cache *s,
1055 struct page *page,
1056 void *object, unsigned long addr)
1057{
1058 if (!check_slab(s, page))
1059 return 0;
1060
1061 if (!check_valid_pointer(s, page, object)) {
1062 object_err(s, page, object, "Freelist Pointer check fails");
1063 return 0;
1064 }
1065
1066 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1067 return 0;
1068
1069 return 1;
1070}
1071
1072static noinline int alloc_debug_processing(struct kmem_cache *s,
1073 struct page *page,
1074 void *object, unsigned long addr)
1075{
1076 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1077 if (!alloc_consistency_checks(s, page, object, addr))
1078 goto bad;
1079 }
1080
1081 /* Success perform special debug activities for allocs */
1082 if (s->flags & SLAB_STORE_USER)
1083 set_track(s, object, TRACK_ALLOC, addr);
1084 trace(s, page, object, 1);
1085 init_object(s, object, SLUB_RED_ACTIVE);
1086 return 1;
1087
1088bad:
1089 if (PageSlab(page)) {
1090 /*
1091 * If this is a slab page then lets do the best we can
1092 * to avoid issues in the future. Marking all objects
1093 * as used avoids touching the remaining objects.
1094 */
1095 slab_fix(s, "Marking all objects used");
1096 page->inuse = page->objects;
1097 page->freelist = NULL;
1098 }
1099 return 0;
1100}
1101
1102static inline int free_consistency_checks(struct kmem_cache *s,
1103 struct page *page, void *object, unsigned long addr)
1104{
1105 if (!check_valid_pointer(s, page, object)) {
1106 slab_err(s, page, "Invalid object pointer 0x%p", object);
1107 return 0;
1108 }
1109
1110 if (on_freelist(s, page, object)) {
1111 object_err(s, page, object, "Object already free");
1112 return 0;
1113 }
1114
1115 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1116 return 0;
1117
1118 if (unlikely(s != page->slab_cache)) {
1119 if (!PageSlab(page)) {
1120 slab_err(s, page, "Attempt to free object(0x%p) outside of slab",
1121 object);
1122 } else if (!page->slab_cache) {
1123 pr_err("SLUB <none>: no slab for object 0x%p.\n",
1124 object);
1125 dump_stack();
1126 } else
1127 object_err(s, page, object,
1128 "page slab pointer corrupt.");
1129 return 0;
1130 }
1131 return 1;
1132}
1133
1134/* Supports checking bulk free of a constructed freelist */
1135static noinline int free_debug_processing(
1136 struct kmem_cache *s, struct page *page,
1137 void *head, void *tail, int bulk_cnt,
1138 unsigned long addr)
1139{
1140 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1141 void *object = head;
1142 int cnt = 0;
1143 unsigned long uninitialized_var(flags);
1144 int ret = 0;
1145
1146 spin_lock_irqsave(&n->list_lock, flags);
1147 slab_lock(page);
1148
1149 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1150 if (!check_slab(s, page))
1151 goto out;
1152 }
1153
1154next_object:
1155 cnt++;
1156
1157 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1158 if (!free_consistency_checks(s, page, object, addr))
1159 goto out;
1160 }
1161
1162 if (s->flags & SLAB_STORE_USER)
1163 set_track(s, object, TRACK_FREE, addr);
1164 trace(s, page, object, 0);
1165 /* Freepointer not overwritten by init_object(), SLAB_POISON moved it */
1166 init_object(s, object, SLUB_RED_INACTIVE);
1167
1168 /* Reached end of constructed freelist yet? */
1169 if (object != tail) {
1170 object = get_freepointer(s, object);
1171 goto next_object;
1172 }
1173 ret = 1;
1174
1175out:
1176 if (cnt != bulk_cnt)
1177 slab_err(s, page, "Bulk freelist count(%d) invalid(%d)\n",
1178 bulk_cnt, cnt);
1179
1180 slab_unlock(page);
1181 spin_unlock_irqrestore(&n->list_lock, flags);
1182 if (!ret)
1183 slab_fix(s, "Object at 0x%p not freed", object);
1184 return ret;
1185}
1186
1187static int __init setup_slub_debug(char *str)
1188{
1189 slub_debug = DEBUG_DEFAULT_FLAGS;
1190 if (*str++ != '=' || !*str)
1191 /*
1192 * No options specified. Switch on full debugging.
1193 */
1194 goto out;
1195
1196 if (*str == ',')
1197 /*
1198 * No options but restriction on slabs. This means full
1199 * debugging for slabs matching a pattern.
1200 */
1201 goto check_slabs;
1202
1203 slub_debug = 0;
1204 if (*str == '-')
1205 /*
1206 * Switch off all debugging measures.
1207 */
1208 goto out;
1209
1210 /*
1211 * Determine which debug features should be switched on
1212 */
1213 for (; *str && *str != ','; str++) {
1214 switch (tolower(*str)) {
1215 case 'f':
1216 slub_debug |= SLAB_CONSISTENCY_CHECKS;
1217 break;
1218 case 'z':
1219 slub_debug |= SLAB_RED_ZONE;
1220 break;
1221 case 'p':
1222 slub_debug |= SLAB_POISON;
1223 break;
1224 case 'u':
1225 slub_debug |= SLAB_STORE_USER;
1226 break;
1227 case 't':
1228 slub_debug |= SLAB_TRACE;
1229 break;
1230 case 'a':
1231 slub_debug |= SLAB_FAILSLAB;
1232 break;
1233 case 'o':
1234 /*
1235 * Avoid enabling debugging on caches if its minimum
1236 * order would increase as a result.
1237 */
1238 disable_higher_order_debug = 1;
1239 break;
1240 default:
1241 pr_err("slub_debug option '%c' unknown. skipped\n",
1242 *str);
1243 }
1244 }
1245
1246check_slabs:
1247 if (*str == ',')
1248 slub_debug_slabs = str + 1;
1249out:
1250 return 1;
1251}
1252
1253__setup("slub_debug", setup_slub_debug);
1254
1255unsigned long kmem_cache_flags(unsigned long object_size,
1256 unsigned long flags, const char *name,
1257 void (*ctor)(void *))
1258{
1259 /*
1260 * Enable debugging if selected on the kernel commandline.
1261 */
1262 if (slub_debug && (!slub_debug_slabs || (name &&
1263 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1264 flags |= slub_debug;
1265
1266 return flags;
1267}
1268#else /* !CONFIG_SLUB_DEBUG */
1269static inline void setup_object_debug(struct kmem_cache *s,
1270 struct page *page, void *object) {}
1271
1272static inline int alloc_debug_processing(struct kmem_cache *s,
1273 struct page *page, void *object, unsigned long addr) { return 0; }
1274
1275static inline int free_debug_processing(
1276 struct kmem_cache *s, struct page *page,
1277 void *head, void *tail, int bulk_cnt,
1278 unsigned long addr) { return 0; }
1279
1280static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1281 { return 1; }
1282static inline int check_object(struct kmem_cache *s, struct page *page,
1283 void *object, u8 val) { return 1; }
1284static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1285 struct page *page) {}
1286static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1287 struct page *page) {}
1288unsigned long kmem_cache_flags(unsigned long object_size,
1289 unsigned long flags, const char *name,
1290 void (*ctor)(void *))
1291{
1292 return flags;
1293}
1294#define slub_debug 0
1295
1296#define disable_higher_order_debug 0
1297
1298static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1299 { return 0; }
1300static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1301 { return 0; }
1302static inline void inc_slabs_node(struct kmem_cache *s, int node,
1303 int objects) {}
1304static inline void dec_slabs_node(struct kmem_cache *s, int node,
1305 int objects) {}
1306
1307#endif /* CONFIG_SLUB_DEBUG */
1308
1309/*
1310 * Hooks for other subsystems that check memory allocations. In a typical
1311 * production configuration these hooks all should produce no code at all.
1312 */
1313static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1314{
1315 kmemleak_alloc(ptr, size, 1, flags);
1316 kasan_kmalloc_large(ptr, size, flags);
1317}
1318
1319static inline void kfree_hook(const void *x)
1320{
1321 kmemleak_free(x);
1322 kasan_kfree_large(x);
1323}
1324
1325static inline void slab_free_hook(struct kmem_cache *s, void *x)
1326{
1327 kmemleak_free_recursive(x, s->flags);
1328
1329 /*
1330 * Trouble is that we may no longer disable interrupts in the fast path
1331 * So in order to make the debug calls that expect irqs to be
1332 * disabled we need to disable interrupts temporarily.
1333 */
1334#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1335 {
1336 unsigned long flags;
1337
1338 local_irq_save(flags);
1339 kmemcheck_slab_free(s, x, s->object_size);
1340 debug_check_no_locks_freed(x, s->object_size);
1341 local_irq_restore(flags);
1342 }
1343#endif
1344 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1345 debug_check_no_obj_freed(x, s->object_size);
1346
1347 kasan_slab_free(s, x);
1348}
1349
1350static inline void slab_free_freelist_hook(struct kmem_cache *s,
1351 void *head, void *tail)
1352{
1353/*
1354 * Compiler cannot detect this function can be removed if slab_free_hook()
1355 * evaluates to nothing. Thus, catch all relevant config debug options here.
1356 */
1357#if defined(CONFIG_KMEMCHECK) || \
1358 defined(CONFIG_LOCKDEP) || \
1359 defined(CONFIG_DEBUG_KMEMLEAK) || \
1360 defined(CONFIG_DEBUG_OBJECTS_FREE) || \
1361 defined(CONFIG_KASAN)
1362
1363 void *object = head;
1364 void *tail_obj = tail ? : head;
1365
1366 do {
1367 slab_free_hook(s, object);
1368 } while ((object != tail_obj) &&
1369 (object = get_freepointer(s, object)));
1370#endif
1371}
1372
1373static void setup_object(struct kmem_cache *s, struct page *page,
1374 void *object)
1375{
1376 setup_object_debug(s, page, object);
1377 if (unlikely(s->ctor)) {
1378 kasan_unpoison_object_data(s, object);
1379 s->ctor(object);
1380 kasan_poison_object_data(s, object);
1381 }
1382}
1383
1384/*
1385 * Slab allocation and freeing
1386 */
1387static inline struct page *alloc_slab_page(struct kmem_cache *s,
1388 gfp_t flags, int node, struct kmem_cache_order_objects oo)
1389{
1390 struct page *page;
1391 int order = oo_order(oo);
1392
1393 flags |= __GFP_NOTRACK;
1394
1395 if (node == NUMA_NO_NODE)
1396 page = alloc_pages(flags, order);
1397 else
1398 page = __alloc_pages_node(node, flags, order);
1399
1400 if (page && memcg_charge_slab(page, flags, order, s)) {
1401 __free_pages(page, order);
1402 page = NULL;
1403 }
1404
1405 return page;
1406}
1407
1408static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1409{
1410 struct page *page;
1411 struct kmem_cache_order_objects oo = s->oo;
1412 gfp_t alloc_gfp;
1413 void *start, *p;
1414 int idx, order;
1415
1416 flags &= gfp_allowed_mask;
1417
1418 if (gfpflags_allow_blocking(flags))
1419 local_irq_enable();
1420
1421 flags |= s->allocflags;
1422
1423 /*
1424 * Let the initial higher-order allocation fail under memory pressure
1425 * so we fall-back to the minimum order allocation.
1426 */
1427 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1428 if ((alloc_gfp & __GFP_DIRECT_RECLAIM) && oo_order(oo) > oo_order(s->min))
1429 alloc_gfp = (alloc_gfp | __GFP_NOMEMALLOC) & ~(__GFP_RECLAIM|__GFP_NOFAIL);
1430
1431 page = alloc_slab_page(s, alloc_gfp, node, oo);
1432 if (unlikely(!page)) {
1433 oo = s->min;
1434 alloc_gfp = flags;
1435 /*
1436 * Allocation may have failed due to fragmentation.
1437 * Try a lower order alloc if possible
1438 */
1439 page = alloc_slab_page(s, alloc_gfp, node, oo);
1440 if (unlikely(!page))
1441 goto out;
1442 stat(s, ORDER_FALLBACK);
1443 }
1444
1445 if (kmemcheck_enabled &&
1446 !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1447 int pages = 1 << oo_order(oo);
1448
1449 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1450
1451 /*
1452 * Objects from caches that have a constructor don't get
1453 * cleared when they're allocated, so we need to do it here.
1454 */
1455 if (s->ctor)
1456 kmemcheck_mark_uninitialized_pages(page, pages);
1457 else
1458 kmemcheck_mark_unallocated_pages(page, pages);
1459 }
1460
1461 page->objects = oo_objects(oo);
1462
1463 order = compound_order(page);
1464 page->slab_cache = s;
1465 __SetPageSlab(page);
1466 if (page_is_pfmemalloc(page))
1467 SetPageSlabPfmemalloc(page);
1468
1469 start = page_address(page);
1470
1471 if (unlikely(s->flags & SLAB_POISON))
1472 memset(start, POISON_INUSE, PAGE_SIZE << order);
1473
1474 kasan_poison_slab(page);
1475
1476 for_each_object_idx(p, idx, s, start, page->objects) {
1477 setup_object(s, page, p);
1478 if (likely(idx < page->objects))
1479 set_freepointer(s, p, p + s->size);
1480 else
1481 set_freepointer(s, p, NULL);
1482 }
1483
1484 page->freelist = fixup_red_left(s, start);
1485 page->inuse = page->objects;
1486 page->frozen = 1;
1487
1488out:
1489 if (gfpflags_allow_blocking(flags))
1490 local_irq_disable();
1491 if (!page)
1492 return NULL;
1493
1494 mod_zone_page_state(page_zone(page),
1495 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1496 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1497 1 << oo_order(oo));
1498
1499 inc_slabs_node(s, page_to_nid(page), page->objects);
1500
1501 return page;
1502}
1503
1504static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1505{
1506 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1507 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1508 BUG();
1509 }
1510
1511 return allocate_slab(s,
1512 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1513}
1514
1515static void __free_slab(struct kmem_cache *s, struct page *page)
1516{
1517 int order = compound_order(page);
1518 int pages = 1 << order;
1519
1520 if (s->flags & SLAB_CONSISTENCY_CHECKS) {
1521 void *p;
1522
1523 slab_pad_check(s, page);
1524 for_each_object(p, s, page_address(page),
1525 page->objects)
1526 check_object(s, page, p, SLUB_RED_INACTIVE);
1527 }
1528
1529 kmemcheck_free_shadow(page, compound_order(page));
1530
1531 mod_zone_page_state(page_zone(page),
1532 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1533 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1534 -pages);
1535
1536 __ClearPageSlabPfmemalloc(page);
1537 __ClearPageSlab(page);
1538
1539 page_mapcount_reset(page);
1540 if (current->reclaim_state)
1541 current->reclaim_state->reclaimed_slab += pages;
1542 memcg_uncharge_slab(page, order, s);
1543 __free_pages(page, order);
1544}
1545
1546#define need_reserve_slab_rcu \
1547 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1548
1549static void rcu_free_slab(struct rcu_head *h)
1550{
1551 struct page *page;
1552
1553 if (need_reserve_slab_rcu)
1554 page = virt_to_head_page(h);
1555 else
1556 page = container_of((struct list_head *)h, struct page, lru);
1557
1558 __free_slab(page->slab_cache, page);
1559}
1560
1561static void free_slab(struct kmem_cache *s, struct page *page)
1562{
1563 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1564 struct rcu_head *head;
1565
1566 if (need_reserve_slab_rcu) {
1567 int order = compound_order(page);
1568 int offset = (PAGE_SIZE << order) - s->reserved;
1569
1570 VM_BUG_ON(s->reserved != sizeof(*head));
1571 head = page_address(page) + offset;
1572 } else {
1573 head = &page->rcu_head;
1574 }
1575
1576 call_rcu(head, rcu_free_slab);
1577 } else
1578 __free_slab(s, page);
1579}
1580
1581static void discard_slab(struct kmem_cache *s, struct page *page)
1582{
1583 dec_slabs_node(s, page_to_nid(page), page->objects);
1584 free_slab(s, page);
1585}
1586
1587/*
1588 * Management of partially allocated slabs.
1589 */
1590static inline void
1591__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1592{
1593 n->nr_partial++;
1594 if (tail == DEACTIVATE_TO_TAIL)
1595 list_add_tail(&page->lru, &n->partial);
1596 else
1597 list_add(&page->lru, &n->partial);
1598}
1599
1600static inline void add_partial(struct kmem_cache_node *n,
1601 struct page *page, int tail)
1602{
1603 lockdep_assert_held(&n->list_lock);
1604 __add_partial(n, page, tail);
1605}
1606
1607static inline void remove_partial(struct kmem_cache_node *n,
1608 struct page *page)
1609{
1610 lockdep_assert_held(&n->list_lock);
1611 list_del(&page->lru);
1612 n->nr_partial--;
1613}
1614
1615/*
1616 * Remove slab from the partial list, freeze it and
1617 * return the pointer to the freelist.
1618 *
1619 * Returns a list of objects or NULL if it fails.
1620 */
1621static inline void *acquire_slab(struct kmem_cache *s,
1622 struct kmem_cache_node *n, struct page *page,
1623 int mode, int *objects)
1624{
1625 void *freelist;
1626 unsigned long counters;
1627 struct page new;
1628
1629 lockdep_assert_held(&n->list_lock);
1630
1631 /*
1632 * Zap the freelist and set the frozen bit.
1633 * The old freelist is the list of objects for the
1634 * per cpu allocation list.
1635 */
1636 freelist = page->freelist;
1637 counters = page->counters;
1638 new.counters = counters;
1639 *objects = new.objects - new.inuse;
1640 if (mode) {
1641 new.inuse = page->objects;
1642 new.freelist = NULL;
1643 } else {
1644 new.freelist = freelist;
1645 }
1646
1647 VM_BUG_ON(new.frozen);
1648 new.frozen = 1;
1649
1650 if (!__cmpxchg_double_slab(s, page,
1651 freelist, counters,
1652 new.freelist, new.counters,
1653 "acquire_slab"))
1654 return NULL;
1655
1656 remove_partial(n, page);
1657 WARN_ON(!freelist);
1658 return freelist;
1659}
1660
1661static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1662static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1663
1664/*
1665 * Try to allocate a partial slab from a specific node.
1666 */
1667static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1668 struct kmem_cache_cpu *c, gfp_t flags)
1669{
1670 struct page *page, *page2;
1671 void *object = NULL;
1672 int available = 0;
1673 int objects;
1674
1675 /*
1676 * Racy check. If we mistakenly see no partial slabs then we
1677 * just allocate an empty slab. If we mistakenly try to get a
1678 * partial slab and there is none available then get_partials()
1679 * will return NULL.
1680 */
1681 if (!n || !n->nr_partial)
1682 return NULL;
1683
1684 spin_lock(&n->list_lock);
1685 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1686 void *t;
1687
1688 if (!pfmemalloc_match(page, flags))
1689 continue;
1690
1691 t = acquire_slab(s, n, page, object == NULL, &objects);
1692 if (!t)
1693 break;
1694
1695 available += objects;
1696 if (!object) {
1697 c->page = page;
1698 stat(s, ALLOC_FROM_PARTIAL);
1699 object = t;
1700 } else {
1701 put_cpu_partial(s, page, 0);
1702 stat(s, CPU_PARTIAL_NODE);
1703 }
1704 if (!kmem_cache_has_cpu_partial(s)
1705 || available > s->cpu_partial / 2)
1706 break;
1707
1708 }
1709 spin_unlock(&n->list_lock);
1710 return object;
1711}
1712
1713/*
1714 * Get a page from somewhere. Search in increasing NUMA distances.
1715 */
1716static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1717 struct kmem_cache_cpu *c)
1718{
1719#ifdef CONFIG_NUMA
1720 struct zonelist *zonelist;
1721 struct zoneref *z;
1722 struct zone *zone;
1723 enum zone_type high_zoneidx = gfp_zone(flags);
1724 void *object;
1725 unsigned int cpuset_mems_cookie;
1726
1727 /*
1728 * The defrag ratio allows a configuration of the tradeoffs between
1729 * inter node defragmentation and node local allocations. A lower
1730 * defrag_ratio increases the tendency to do local allocations
1731 * instead of attempting to obtain partial slabs from other nodes.
1732 *
1733 * If the defrag_ratio is set to 0 then kmalloc() always
1734 * returns node local objects. If the ratio is higher then kmalloc()
1735 * may return off node objects because partial slabs are obtained
1736 * from other nodes and filled up.
1737 *
1738 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1739 * defrag_ratio = 1000) then every (well almost) allocation will
1740 * first attempt to defrag slab caches on other nodes. This means
1741 * scanning over all nodes to look for partial slabs which may be
1742 * expensive if we do it every time we are trying to find a slab
1743 * with available objects.
1744 */
1745 if (!s->remote_node_defrag_ratio ||
1746 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1747 return NULL;
1748
1749 do {
1750 cpuset_mems_cookie = read_mems_allowed_begin();
1751 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1752 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1753 struct kmem_cache_node *n;
1754
1755 n = get_node(s, zone_to_nid(zone));
1756
1757 if (n && cpuset_zone_allowed(zone, flags) &&
1758 n->nr_partial > s->min_partial) {
1759 object = get_partial_node(s, n, c, flags);
1760 if (object) {
1761 /*
1762 * Don't check read_mems_allowed_retry()
1763 * here - if mems_allowed was updated in
1764 * parallel, that was a harmless race
1765 * between allocation and the cpuset
1766 * update
1767 */
1768 return object;
1769 }
1770 }
1771 }
1772 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1773#endif
1774 return NULL;
1775}
1776
1777/*
1778 * Get a partial page, lock it and return it.
1779 */
1780static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1781 struct kmem_cache_cpu *c)
1782{
1783 void *object;
1784 int searchnode = node;
1785
1786 if (node == NUMA_NO_NODE)
1787 searchnode = numa_mem_id();
1788 else if (!node_present_pages(node))
1789 searchnode = node_to_mem_node(node);
1790
1791 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1792 if (object || node != NUMA_NO_NODE)
1793 return object;
1794
1795 return get_any_partial(s, flags, c);
1796}
1797
1798#ifdef CONFIG_PREEMPT
1799/*
1800 * Calculate the next globally unique transaction for disambiguiation
1801 * during cmpxchg. The transactions start with the cpu number and are then
1802 * incremented by CONFIG_NR_CPUS.
1803 */
1804#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1805#else
1806/*
1807 * No preemption supported therefore also no need to check for
1808 * different cpus.
1809 */
1810#define TID_STEP 1
1811#endif
1812
1813static inline unsigned long next_tid(unsigned long tid)
1814{
1815 return tid + TID_STEP;
1816}
1817
1818static inline unsigned int tid_to_cpu(unsigned long tid)
1819{
1820 return tid % TID_STEP;
1821}
1822
1823static inline unsigned long tid_to_event(unsigned long tid)
1824{
1825 return tid / TID_STEP;
1826}
1827
1828static inline unsigned int init_tid(int cpu)
1829{
1830 return cpu;
1831}
1832
1833static inline void note_cmpxchg_failure(const char *n,
1834 const struct kmem_cache *s, unsigned long tid)
1835{
1836#ifdef SLUB_DEBUG_CMPXCHG
1837 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1838
1839 pr_info("%s %s: cmpxchg redo ", n, s->name);
1840
1841#ifdef CONFIG_PREEMPT
1842 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1843 pr_warn("due to cpu change %d -> %d\n",
1844 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1845 else
1846#endif
1847 if (tid_to_event(tid) != tid_to_event(actual_tid))
1848 pr_warn("due to cpu running other code. Event %ld->%ld\n",
1849 tid_to_event(tid), tid_to_event(actual_tid));
1850 else
1851 pr_warn("for unknown reason: actual=%lx was=%lx target=%lx\n",
1852 actual_tid, tid, next_tid(tid));
1853#endif
1854 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1855}
1856
1857static void init_kmem_cache_cpus(struct kmem_cache *s)
1858{
1859 int cpu;
1860
1861 for_each_possible_cpu(cpu)
1862 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1863}
1864
1865/*
1866 * Remove the cpu slab
1867 */
1868static void deactivate_slab(struct kmem_cache *s, struct page *page,
1869 void *freelist)
1870{
1871 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1872 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1873 int lock = 0;
1874 enum slab_modes l = M_NONE, m = M_NONE;
1875 void *nextfree;
1876 int tail = DEACTIVATE_TO_HEAD;
1877 struct page new;
1878 struct page old;
1879
1880 if (page->freelist) {
1881 stat(s, DEACTIVATE_REMOTE_FREES);
1882 tail = DEACTIVATE_TO_TAIL;
1883 }
1884
1885 /*
1886 * Stage one: Free all available per cpu objects back
1887 * to the page freelist while it is still frozen. Leave the
1888 * last one.
1889 *
1890 * There is no need to take the list->lock because the page
1891 * is still frozen.
1892 */
1893 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1894 void *prior;
1895 unsigned long counters;
1896
1897 do {
1898 prior = page->freelist;
1899 counters = page->counters;
1900 set_freepointer(s, freelist, prior);
1901 new.counters = counters;
1902 new.inuse--;
1903 VM_BUG_ON(!new.frozen);
1904
1905 } while (!__cmpxchg_double_slab(s, page,
1906 prior, counters,
1907 freelist, new.counters,
1908 "drain percpu freelist"));
1909
1910 freelist = nextfree;
1911 }
1912
1913 /*
1914 * Stage two: Ensure that the page is unfrozen while the
1915 * list presence reflects the actual number of objects
1916 * during unfreeze.
1917 *
1918 * We setup the list membership and then perform a cmpxchg
1919 * with the count. If there is a mismatch then the page
1920 * is not unfrozen but the page is on the wrong list.
1921 *
1922 * Then we restart the process which may have to remove
1923 * the page from the list that we just put it on again
1924 * because the number of objects in the slab may have
1925 * changed.
1926 */
1927redo:
1928
1929 old.freelist = page->freelist;
1930 old.counters = page->counters;
1931 VM_BUG_ON(!old.frozen);
1932
1933 /* Determine target state of the slab */
1934 new.counters = old.counters;
1935 if (freelist) {
1936 new.inuse--;
1937 set_freepointer(s, freelist, old.freelist);
1938 new.freelist = freelist;
1939 } else
1940 new.freelist = old.freelist;
1941
1942 new.frozen = 0;
1943
1944 if (!new.inuse && n->nr_partial >= s->min_partial)
1945 m = M_FREE;
1946 else if (new.freelist) {
1947 m = M_PARTIAL;
1948 if (!lock) {
1949 lock = 1;
1950 /*
1951 * Taking the spinlock removes the possiblity
1952 * that acquire_slab() will see a slab page that
1953 * is frozen
1954 */
1955 spin_lock(&n->list_lock);
1956 }
1957 } else {
1958 m = M_FULL;
1959 if (kmem_cache_debug(s) && !lock) {
1960 lock = 1;
1961 /*
1962 * This also ensures that the scanning of full
1963 * slabs from diagnostic functions will not see
1964 * any frozen slabs.
1965 */
1966 spin_lock(&n->list_lock);
1967 }
1968 }
1969
1970 if (l != m) {
1971
1972 if (l == M_PARTIAL)
1973
1974 remove_partial(n, page);
1975
1976 else if (l == M_FULL)
1977
1978 remove_full(s, n, page);
1979
1980 if (m == M_PARTIAL) {
1981
1982 add_partial(n, page, tail);
1983 stat(s, tail);
1984
1985 } else if (m == M_FULL) {
1986
1987 stat(s, DEACTIVATE_FULL);
1988 add_full(s, n, page);
1989
1990 }
1991 }
1992
1993 l = m;
1994 if (!__cmpxchg_double_slab(s, page,
1995 old.freelist, old.counters,
1996 new.freelist, new.counters,
1997 "unfreezing slab"))
1998 goto redo;
1999
2000 if (lock)
2001 spin_unlock(&n->list_lock);
2002
2003 if (m == M_FREE) {
2004 stat(s, DEACTIVATE_EMPTY);
2005 discard_slab(s, page);
2006 stat(s, FREE_SLAB);
2007 }
2008}
2009
2010/*
2011 * Unfreeze all the cpu partial slabs.
2012 *
2013 * This function must be called with interrupts disabled
2014 * for the cpu using c (or some other guarantee must be there
2015 * to guarantee no concurrent accesses).
2016 */
2017static void unfreeze_partials(struct kmem_cache *s,
2018 struct kmem_cache_cpu *c)
2019{
2020#ifdef CONFIG_SLUB_CPU_PARTIAL
2021 struct kmem_cache_node *n = NULL, *n2 = NULL;
2022 struct page *page, *discard_page = NULL;
2023
2024 while ((page = c->partial)) {
2025 struct page new;
2026 struct page old;
2027
2028 c->partial = page->next;
2029
2030 n2 = get_node(s, page_to_nid(page));
2031 if (n != n2) {
2032 if (n)
2033 spin_unlock(&n->list_lock);
2034
2035 n = n2;
2036 spin_lock(&n->list_lock);
2037 }
2038
2039 do {
2040
2041 old.freelist = page->freelist;
2042 old.counters = page->counters;
2043 VM_BUG_ON(!old.frozen);
2044
2045 new.counters = old.counters;
2046 new.freelist = old.freelist;
2047
2048 new.frozen = 0;
2049
2050 } while (!__cmpxchg_double_slab(s, page,
2051 old.freelist, old.counters,
2052 new.freelist, new.counters,
2053 "unfreezing slab"));
2054
2055 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial)) {
2056 page->next = discard_page;
2057 discard_page = page;
2058 } else {
2059 add_partial(n, page, DEACTIVATE_TO_TAIL);
2060 stat(s, FREE_ADD_PARTIAL);
2061 }
2062 }
2063
2064 if (n)
2065 spin_unlock(&n->list_lock);
2066
2067 while (discard_page) {
2068 page = discard_page;
2069 discard_page = discard_page->next;
2070
2071 stat(s, DEACTIVATE_EMPTY);
2072 discard_slab(s, page);
2073 stat(s, FREE_SLAB);
2074 }
2075#endif
2076}
2077
2078/*
2079 * Put a page that was just frozen (in __slab_free) into a partial page
2080 * slot if available. This is done without interrupts disabled and without
2081 * preemption disabled. The cmpxchg is racy and may put the partial page
2082 * onto a random cpus partial slot.
2083 *
2084 * If we did not find a slot then simply move all the partials to the
2085 * per node partial list.
2086 */
2087static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2088{
2089#ifdef CONFIG_SLUB_CPU_PARTIAL
2090 struct page *oldpage;
2091 int pages;
2092 int pobjects;
2093
2094 preempt_disable();
2095 do {
2096 pages = 0;
2097 pobjects = 0;
2098 oldpage = this_cpu_read(s->cpu_slab->partial);
2099
2100 if (oldpage) {
2101 pobjects = oldpage->pobjects;
2102 pages = oldpage->pages;
2103 if (drain && pobjects > s->cpu_partial) {
2104 unsigned long flags;
2105 /*
2106 * partial array is full. Move the existing
2107 * set to the per node partial list.
2108 */
2109 local_irq_save(flags);
2110 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2111 local_irq_restore(flags);
2112 oldpage = NULL;
2113 pobjects = 0;
2114 pages = 0;
2115 stat(s, CPU_PARTIAL_DRAIN);
2116 }
2117 }
2118
2119 pages++;
2120 pobjects += page->objects - page->inuse;
2121
2122 page->pages = pages;
2123 page->pobjects = pobjects;
2124 page->next = oldpage;
2125
2126 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2127 != oldpage);
2128 if (unlikely(!s->cpu_partial)) {
2129 unsigned long flags;
2130
2131 local_irq_save(flags);
2132 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2133 local_irq_restore(flags);
2134 }
2135 preempt_enable();
2136#endif
2137}
2138
2139static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2140{
2141 stat(s, CPUSLAB_FLUSH);
2142 deactivate_slab(s, c->page, c->freelist);
2143
2144 c->tid = next_tid(c->tid);
2145 c->page = NULL;
2146 c->freelist = NULL;
2147}
2148
2149/*
2150 * Flush cpu slab.
2151 *
2152 * Called from IPI handler with interrupts disabled.
2153 */
2154static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2155{
2156 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2157
2158 if (likely(c)) {
2159 if (c->page)
2160 flush_slab(s, c);
2161
2162 unfreeze_partials(s, c);
2163 }
2164}
2165
2166static void flush_cpu_slab(void *d)
2167{
2168 struct kmem_cache *s = d;
2169
2170 __flush_cpu_slab(s, smp_processor_id());
2171}
2172
2173static bool has_cpu_slab(int cpu, void *info)
2174{
2175 struct kmem_cache *s = info;
2176 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2177
2178 return c->page || c->partial;
2179}
2180
2181static void flush_all(struct kmem_cache *s)
2182{
2183 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2184}
2185
2186/*
2187 * Check if the objects in a per cpu structure fit numa
2188 * locality expectations.
2189 */
2190static inline int node_match(struct page *page, int node)
2191{
2192#ifdef CONFIG_NUMA
2193 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2194 return 0;
2195#endif
2196 return 1;
2197}
2198
2199#ifdef CONFIG_SLUB_DEBUG
2200static int count_free(struct page *page)
2201{
2202 return page->objects - page->inuse;
2203}
2204
2205static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2206{
2207 return atomic_long_read(&n->total_objects);
2208}
2209#endif /* CONFIG_SLUB_DEBUG */
2210
2211#if defined(CONFIG_SLUB_DEBUG) || defined(CONFIG_SYSFS)
2212static unsigned long count_partial(struct kmem_cache_node *n,
2213 int (*get_count)(struct page *))
2214{
2215 unsigned long flags;
2216 unsigned long x = 0;
2217 struct page *page;
2218
2219 spin_lock_irqsave(&n->list_lock, flags);
2220 list_for_each_entry(page, &n->partial, lru)
2221 x += get_count(page);
2222 spin_unlock_irqrestore(&n->list_lock, flags);
2223 return x;
2224}
2225#endif /* CONFIG_SLUB_DEBUG || CONFIG_SYSFS */
2226
2227static noinline void
2228slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2229{
2230#ifdef CONFIG_SLUB_DEBUG
2231 static DEFINE_RATELIMIT_STATE(slub_oom_rs, DEFAULT_RATELIMIT_INTERVAL,
2232 DEFAULT_RATELIMIT_BURST);
2233 int node;
2234 struct kmem_cache_node *n;
2235
2236 if ((gfpflags & __GFP_NOWARN) || !__ratelimit(&slub_oom_rs))
2237 return;
2238
2239 pr_warn("SLUB: Unable to allocate memory on node %d, gfp=%#x(%pGg)\n",
2240 nid, gfpflags, &gfpflags);
2241 pr_warn(" cache: %s, object size: %d, buffer size: %d, default order: %d, min order: %d\n",
2242 s->name, s->object_size, s->size, oo_order(s->oo),
2243 oo_order(s->min));
2244
2245 if (oo_order(s->min) > get_order(s->object_size))
2246 pr_warn(" %s debugging increased min order, use slub_debug=O to disable.\n",
2247 s->name);
2248
2249 for_each_kmem_cache_node(s, node, n) {
2250 unsigned long nr_slabs;
2251 unsigned long nr_objs;
2252 unsigned long nr_free;
2253
2254 nr_free = count_partial(n, count_free);
2255 nr_slabs = node_nr_slabs(n);
2256 nr_objs = node_nr_objs(n);
2257
2258 pr_warn(" node %d: slabs: %ld, objs: %ld, free: %ld\n",
2259 node, nr_slabs, nr_objs, nr_free);
2260 }
2261#endif
2262}
2263
2264static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2265 int node, struct kmem_cache_cpu **pc)
2266{
2267 void *freelist;
2268 struct kmem_cache_cpu *c = *pc;
2269 struct page *page;
2270
2271 freelist = get_partial(s, flags, node, c);
2272
2273 if (freelist)
2274 return freelist;
2275
2276 page = new_slab(s, flags, node);
2277 if (page) {
2278 c = raw_cpu_ptr(s->cpu_slab);
2279 if (c->page)
2280 flush_slab(s, c);
2281
2282 /*
2283 * No other reference to the page yet so we can
2284 * muck around with it freely without cmpxchg
2285 */
2286 freelist = page->freelist;
2287 page->freelist = NULL;
2288
2289 stat(s, ALLOC_SLAB);
2290 c->page = page;
2291 *pc = c;
2292 } else
2293 freelist = NULL;
2294
2295 return freelist;
2296}
2297
2298static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2299{
2300 if (unlikely(PageSlabPfmemalloc(page)))
2301 return gfp_pfmemalloc_allowed(gfpflags);
2302
2303 return true;
2304}
2305
2306/*
2307 * Check the page->freelist of a page and either transfer the freelist to the
2308 * per cpu freelist or deactivate the page.
2309 *
2310 * The page is still frozen if the return value is not NULL.
2311 *
2312 * If this function returns NULL then the page has been unfrozen.
2313 *
2314 * This function must be called with interrupt disabled.
2315 */
2316static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2317{
2318 struct page new;
2319 unsigned long counters;
2320 void *freelist;
2321
2322 do {
2323 freelist = page->freelist;
2324 counters = page->counters;
2325
2326 new.counters = counters;
2327 VM_BUG_ON(!new.frozen);
2328
2329 new.inuse = page->objects;
2330 new.frozen = freelist != NULL;
2331
2332 } while (!__cmpxchg_double_slab(s, page,
2333 freelist, counters,
2334 NULL, new.counters,
2335 "get_freelist"));
2336
2337 return freelist;
2338}
2339
2340/*
2341 * Slow path. The lockless freelist is empty or we need to perform
2342 * debugging duties.
2343 *
2344 * Processing is still very fast if new objects have been freed to the
2345 * regular freelist. In that case we simply take over the regular freelist
2346 * as the lockless freelist and zap the regular freelist.
2347 *
2348 * If that is not working then we fall back to the partial lists. We take the
2349 * first element of the freelist as the object to allocate now and move the
2350 * rest of the freelist to the lockless freelist.
2351 *
2352 * And if we were unable to get a new slab from the partial slab lists then
2353 * we need to allocate a new slab. This is the slowest path since it involves
2354 * a call to the page allocator and the setup of a new slab.
2355 *
2356 * Version of __slab_alloc to use when we know that interrupts are
2357 * already disabled (which is the case for bulk allocation).
2358 */
2359static void *___slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2360 unsigned long addr, struct kmem_cache_cpu *c)
2361{
2362 void *freelist;
2363 struct page *page;
2364
2365 page = c->page;
2366 if (!page)
2367 goto new_slab;
2368redo:
2369
2370 if (unlikely(!node_match(page, node))) {
2371 int searchnode = node;
2372
2373 if (node != NUMA_NO_NODE && !node_present_pages(node))
2374 searchnode = node_to_mem_node(node);
2375
2376 if (unlikely(!node_match(page, searchnode))) {
2377 stat(s, ALLOC_NODE_MISMATCH);
2378 deactivate_slab(s, page, c->freelist);
2379 c->page = NULL;
2380 c->freelist = NULL;
2381 goto new_slab;
2382 }
2383 }
2384
2385 /*
2386 * By rights, we should be searching for a slab page that was
2387 * PFMEMALLOC but right now, we are losing the pfmemalloc
2388 * information when the page leaves the per-cpu allocator
2389 */
2390 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2391 deactivate_slab(s, page, c->freelist);
2392 c->page = NULL;
2393 c->freelist = NULL;
2394 goto new_slab;
2395 }
2396
2397 /* must check again c->freelist in case of cpu migration or IRQ */
2398 freelist = c->freelist;
2399 if (freelist)
2400 goto load_freelist;
2401
2402 freelist = get_freelist(s, page);
2403
2404 if (!freelist) {
2405 c->page = NULL;
2406 stat(s, DEACTIVATE_BYPASS);
2407 goto new_slab;
2408 }
2409
2410 stat(s, ALLOC_REFILL);
2411
2412load_freelist:
2413 /*
2414 * freelist is pointing to the list of objects to be used.
2415 * page is pointing to the page from which the objects are obtained.
2416 * That page must be frozen for per cpu allocations to work.
2417 */
2418 VM_BUG_ON(!c->page->frozen);
2419 c->freelist = get_freepointer(s, freelist);
2420 c->tid = next_tid(c->tid);
2421 return freelist;
2422
2423new_slab:
2424
2425 if (c->partial) {
2426 page = c->page = c->partial;
2427 c->partial = page->next;
2428 stat(s, CPU_PARTIAL_ALLOC);
2429 c->freelist = NULL;
2430 goto redo;
2431 }
2432
2433 freelist = new_slab_objects(s, gfpflags, node, &c);
2434
2435 if (unlikely(!freelist)) {
2436 slab_out_of_memory(s, gfpflags, node);
2437 return NULL;
2438 }
2439
2440 page = c->page;
2441 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2442 goto load_freelist;
2443
2444 /* Only entered in the debug case */
2445 if (kmem_cache_debug(s) &&
2446 !alloc_debug_processing(s, page, freelist, addr))
2447 goto new_slab; /* Slab failed checks. Next slab needed */
2448
2449 deactivate_slab(s, page, get_freepointer(s, freelist));
2450 c->page = NULL;
2451 c->freelist = NULL;
2452 return freelist;
2453}
2454
2455/*
2456 * Another one that disabled interrupt and compensates for possible
2457 * cpu changes by refetching the per cpu area pointer.
2458 */
2459static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2460 unsigned long addr, struct kmem_cache_cpu *c)
2461{
2462 void *p;
2463 unsigned long flags;
2464
2465 local_irq_save(flags);
2466#ifdef CONFIG_PREEMPT
2467 /*
2468 * We may have been preempted and rescheduled on a different
2469 * cpu before disabling interrupts. Need to reload cpu area
2470 * pointer.
2471 */
2472 c = this_cpu_ptr(s->cpu_slab);
2473#endif
2474
2475 p = ___slab_alloc(s, gfpflags, node, addr, c);
2476 local_irq_restore(flags);
2477 return p;
2478}
2479
2480/*
2481 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2482 * have the fastpath folded into their functions. So no function call
2483 * overhead for requests that can be satisfied on the fastpath.
2484 *
2485 * The fastpath works by first checking if the lockless freelist can be used.
2486 * If not then __slab_alloc is called for slow processing.
2487 *
2488 * Otherwise we can simply pick the next object from the lockless free list.
2489 */
2490static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2491 gfp_t gfpflags, int node, unsigned long addr)
2492{
2493 void *object;
2494 struct kmem_cache_cpu *c;
2495 struct page *page;
2496 unsigned long tid;
2497
2498 s = slab_pre_alloc_hook(s, gfpflags);
2499 if (!s)
2500 return NULL;
2501redo:
2502 /*
2503 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2504 * enabled. We may switch back and forth between cpus while
2505 * reading from one cpu area. That does not matter as long
2506 * as we end up on the original cpu again when doing the cmpxchg.
2507 *
2508 * We should guarantee that tid and kmem_cache are retrieved on
2509 * the same cpu. It could be different if CONFIG_PREEMPT so we need
2510 * to check if it is matched or not.
2511 */
2512 do {
2513 tid = this_cpu_read(s->cpu_slab->tid);
2514 c = raw_cpu_ptr(s->cpu_slab);
2515 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2516 unlikely(tid != READ_ONCE(c->tid)));
2517
2518 /*
2519 * Irqless object alloc/free algorithm used here depends on sequence
2520 * of fetching cpu_slab's data. tid should be fetched before anything
2521 * on c to guarantee that object and page associated with previous tid
2522 * won't be used with current tid. If we fetch tid first, object and
2523 * page could be one associated with next tid and our alloc/free
2524 * request will be failed. In this case, we will retry. So, no problem.
2525 */
2526 barrier();
2527
2528 /*
2529 * The transaction ids are globally unique per cpu and per operation on
2530 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2531 * occurs on the right processor and that there was no operation on the
2532 * linked list in between.
2533 */
2534
2535 object = c->freelist;
2536 page = c->page;
2537 if (unlikely(!object || !node_match(page, node))) {
2538 object = __slab_alloc(s, gfpflags, node, addr, c);
2539 stat(s, ALLOC_SLOWPATH);
2540 } else {
2541 void *next_object = get_freepointer_safe(s, object);
2542
2543 /*
2544 * The cmpxchg will only match if there was no additional
2545 * operation and if we are on the right processor.
2546 *
2547 * The cmpxchg does the following atomically (without lock
2548 * semantics!)
2549 * 1. Relocate first pointer to the current per cpu area.
2550 * 2. Verify that tid and freelist have not been changed
2551 * 3. If they were not changed replace tid and freelist
2552 *
2553 * Since this is without lock semantics the protection is only
2554 * against code executing on this cpu *not* from access by
2555 * other cpus.
2556 */
2557 if (unlikely(!this_cpu_cmpxchg_double(
2558 s->cpu_slab->freelist, s->cpu_slab->tid,
2559 object, tid,
2560 next_object, next_tid(tid)))) {
2561
2562 note_cmpxchg_failure("slab_alloc", s, tid);
2563 goto redo;
2564 }
2565 prefetch_freepointer(s, next_object);
2566 stat(s, ALLOC_FASTPATH);
2567 }
2568
2569 if (unlikely(gfpflags & __GFP_ZERO) && object)
2570 memset(object, 0, s->object_size);
2571
2572 slab_post_alloc_hook(s, gfpflags, 1, &object);
2573
2574 return object;
2575}
2576
2577static __always_inline void *slab_alloc(struct kmem_cache *s,
2578 gfp_t gfpflags, unsigned long addr)
2579{
2580 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2581}
2582
2583void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2584{
2585 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2586
2587 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2588 s->size, gfpflags);
2589
2590 return ret;
2591}
2592EXPORT_SYMBOL(kmem_cache_alloc);
2593
2594#ifdef CONFIG_TRACING
2595void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2596{
2597 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2598 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2599 kasan_kmalloc(s, ret, size, gfpflags);
2600 return ret;
2601}
2602EXPORT_SYMBOL(kmem_cache_alloc_trace);
2603#endif
2604
2605#ifdef CONFIG_NUMA
2606void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2607{
2608 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2609
2610 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2611 s->object_size, s->size, gfpflags, node);
2612
2613 return ret;
2614}
2615EXPORT_SYMBOL(kmem_cache_alloc_node);
2616
2617#ifdef CONFIG_TRACING
2618void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2619 gfp_t gfpflags,
2620 int node, size_t size)
2621{
2622 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2623
2624 trace_kmalloc_node(_RET_IP_, ret,
2625 size, s->size, gfpflags, node);
2626
2627 kasan_kmalloc(s, ret, size, gfpflags);
2628 return ret;
2629}
2630EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2631#endif
2632#endif
2633
2634/*
2635 * Slow path handling. This may still be called frequently since objects
2636 * have a longer lifetime than the cpu slabs in most processing loads.
2637 *
2638 * So we still attempt to reduce cache line usage. Just take the slab
2639 * lock and free the item. If there is no additional partial page
2640 * handling required then we can return immediately.
2641 */
2642static void __slab_free(struct kmem_cache *s, struct page *page,
2643 void *head, void *tail, int cnt,
2644 unsigned long addr)
2645
2646{
2647 void *prior;
2648 int was_frozen;
2649 struct page new;
2650 unsigned long counters;
2651 struct kmem_cache_node *n = NULL;
2652 unsigned long uninitialized_var(flags);
2653
2654 stat(s, FREE_SLOWPATH);
2655
2656 if (kmem_cache_debug(s) &&
2657 !free_debug_processing(s, page, head, tail, cnt, addr))
2658 return;
2659
2660 do {
2661 if (unlikely(n)) {
2662 spin_unlock_irqrestore(&n->list_lock, flags);
2663 n = NULL;
2664 }
2665 prior = page->freelist;
2666 counters = page->counters;
2667 set_freepointer(s, tail, prior);
2668 new.counters = counters;
2669 was_frozen = new.frozen;
2670 new.inuse -= cnt;
2671 if ((!new.inuse || !prior) && !was_frozen) {
2672
2673 if (kmem_cache_has_cpu_partial(s) && !prior) {
2674
2675 /*
2676 * Slab was on no list before and will be
2677 * partially empty
2678 * We can defer the list move and instead
2679 * freeze it.
2680 */
2681 new.frozen = 1;
2682
2683 } else { /* Needs to be taken off a list */
2684
2685 n = get_node(s, page_to_nid(page));
2686 /*
2687 * Speculatively acquire the list_lock.
2688 * If the cmpxchg does not succeed then we may
2689 * drop the list_lock without any processing.
2690 *
2691 * Otherwise the list_lock will synchronize with
2692 * other processors updating the list of slabs.
2693 */
2694 spin_lock_irqsave(&n->list_lock, flags);
2695
2696 }
2697 }
2698
2699 } while (!cmpxchg_double_slab(s, page,
2700 prior, counters,
2701 head, new.counters,
2702 "__slab_free"));
2703
2704 if (likely(!n)) {
2705
2706 /*
2707 * If we just froze the page then put it onto the
2708 * per cpu partial list.
2709 */
2710 if (new.frozen && !was_frozen) {
2711 put_cpu_partial(s, page, 1);
2712 stat(s, CPU_PARTIAL_FREE);
2713 }
2714 /*
2715 * The list lock was not taken therefore no list
2716 * activity can be necessary.
2717 */
2718 if (was_frozen)
2719 stat(s, FREE_FROZEN);
2720 return;
2721 }
2722
2723 if (unlikely(!new.inuse && n->nr_partial >= s->min_partial))
2724 goto slab_empty;
2725
2726 /*
2727 * Objects left in the slab. If it was not on the partial list before
2728 * then add it.
2729 */
2730 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2731 if (kmem_cache_debug(s))
2732 remove_full(s, n, page);
2733 add_partial(n, page, DEACTIVATE_TO_TAIL);
2734 stat(s, FREE_ADD_PARTIAL);
2735 }
2736 spin_unlock_irqrestore(&n->list_lock, flags);
2737 return;
2738
2739slab_empty:
2740 if (prior) {
2741 /*
2742 * Slab on the partial list.
2743 */
2744 remove_partial(n, page);
2745 stat(s, FREE_REMOVE_PARTIAL);
2746 } else {
2747 /* Slab must be on the full list */
2748 remove_full(s, n, page);
2749 }
2750
2751 spin_unlock_irqrestore(&n->list_lock, flags);
2752 stat(s, FREE_SLAB);
2753 discard_slab(s, page);
2754}
2755
2756/*
2757 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2758 * can perform fastpath freeing without additional function calls.
2759 *
2760 * The fastpath is only possible if we are freeing to the current cpu slab
2761 * of this processor. This typically the case if we have just allocated
2762 * the item before.
2763 *
2764 * If fastpath is not possible then fall back to __slab_free where we deal
2765 * with all sorts of special processing.
2766 *
2767 * Bulk free of a freelist with several objects (all pointing to the
2768 * same page) possible by specifying head and tail ptr, plus objects
2769 * count (cnt). Bulk free indicated by tail pointer being set.
2770 */
2771static __always_inline void slab_free(struct kmem_cache *s, struct page *page,
2772 void *head, void *tail, int cnt,
2773 unsigned long addr)
2774{
2775 void *tail_obj = tail ? : head;
2776 struct kmem_cache_cpu *c;
2777 unsigned long tid;
2778
2779 slab_free_freelist_hook(s, head, tail);
2780
2781redo:
2782 /*
2783 * Determine the currently cpus per cpu slab.
2784 * The cpu may change afterward. However that does not matter since
2785 * data is retrieved via this pointer. If we are on the same cpu
2786 * during the cmpxchg then the free will succeed.
2787 */
2788 do {
2789 tid = this_cpu_read(s->cpu_slab->tid);
2790 c = raw_cpu_ptr(s->cpu_slab);
2791 } while (IS_ENABLED(CONFIG_PREEMPT) &&
2792 unlikely(tid != READ_ONCE(c->tid)));
2793
2794 /* Same with comment on barrier() in slab_alloc_node() */
2795 barrier();
2796
2797 if (likely(page == c->page)) {
2798 set_freepointer(s, tail_obj, c->freelist);
2799
2800 if (unlikely(!this_cpu_cmpxchg_double(
2801 s->cpu_slab->freelist, s->cpu_slab->tid,
2802 c->freelist, tid,
2803 head, next_tid(tid)))) {
2804
2805 note_cmpxchg_failure("slab_free", s, tid);
2806 goto redo;
2807 }
2808 stat(s, FREE_FASTPATH);
2809 } else
2810 __slab_free(s, page, head, tail_obj, cnt, addr);
2811
2812}
2813
2814void kmem_cache_free(struct kmem_cache *s, void *x)
2815{
2816 s = cache_from_obj(s, x);
2817 if (!s)
2818 return;
2819 slab_free(s, virt_to_head_page(x), x, NULL, 1, _RET_IP_);
2820 trace_kmem_cache_free(_RET_IP_, x);
2821}
2822EXPORT_SYMBOL(kmem_cache_free);
2823
2824struct detached_freelist {
2825 struct page *page;
2826 void *tail;
2827 void *freelist;
2828 int cnt;
2829 struct kmem_cache *s;
2830};
2831
2832/*
2833 * This function progressively scans the array with free objects (with
2834 * a limited look ahead) and extract objects belonging to the same
2835 * page. It builds a detached freelist directly within the given
2836 * page/objects. This can happen without any need for
2837 * synchronization, because the objects are owned by running process.
2838 * The freelist is build up as a single linked list in the objects.
2839 * The idea is, that this detached freelist can then be bulk
2840 * transferred to the real freelist(s), but only requiring a single
2841 * synchronization primitive. Look ahead in the array is limited due
2842 * to performance reasons.
2843 */
2844static inline
2845int build_detached_freelist(struct kmem_cache *s, size_t size,
2846 void **p, struct detached_freelist *df)
2847{
2848 size_t first_skipped_index = 0;
2849 int lookahead = 3;
2850 void *object;
2851 struct page *page;
2852
2853 /* Always re-init detached_freelist */
2854 df->page = NULL;
2855
2856 do {
2857 object = p[--size];
2858 /* Do we need !ZERO_OR_NULL_PTR(object) here? (for kfree) */
2859 } while (!object && size);
2860
2861 if (!object)
2862 return 0;
2863
2864 page = virt_to_head_page(object);
2865 if (!s) {
2866 /* Handle kalloc'ed objects */
2867 if (unlikely(!PageSlab(page))) {
2868 BUG_ON(!PageCompound(page));
2869 kfree_hook(object);
2870 __free_kmem_pages(page, compound_order(page));
2871 p[size] = NULL; /* mark object processed */
2872 return size;
2873 }
2874 /* Derive kmem_cache from object */
2875 df->s = page->slab_cache;
2876 } else {
2877 df->s = cache_from_obj(s, object); /* Support for memcg */
2878 }
2879
2880 /* Start new detached freelist */
2881 df->page = page;
2882 set_freepointer(df->s, object, NULL);
2883 df->tail = object;
2884 df->freelist = object;
2885 p[size] = NULL; /* mark object processed */
2886 df->cnt = 1;
2887
2888 while (size) {
2889 object = p[--size];
2890 if (!object)
2891 continue; /* Skip processed objects */
2892
2893 /* df->page is always set at this point */
2894 if (df->page == virt_to_head_page(object)) {
2895 /* Opportunity build freelist */
2896 set_freepointer(df->s, object, df->freelist);
2897 df->freelist = object;
2898 df->cnt++;
2899 p[size] = NULL; /* mark object processed */
2900
2901 continue;
2902 }
2903
2904 /* Limit look ahead search */
2905 if (!--lookahead)
2906 break;
2907
2908 if (!first_skipped_index)
2909 first_skipped_index = size + 1;
2910 }
2911
2912 return first_skipped_index;
2913}
2914
2915/* Note that interrupts must be enabled when calling this function. */
2916void kmem_cache_free_bulk(struct kmem_cache *s, size_t size, void **p)
2917{
2918 if (WARN_ON(!size))
2919 return;
2920
2921 do {
2922 struct detached_freelist df;
2923
2924 size = build_detached_freelist(s, size, p, &df);
2925 if (unlikely(!df.page))
2926 continue;
2927
2928 slab_free(df.s, df.page, df.freelist, df.tail, df.cnt,_RET_IP_);
2929 } while (likely(size));
2930}
2931EXPORT_SYMBOL(kmem_cache_free_bulk);
2932
2933/* Note that interrupts must be enabled when calling this function. */
2934int kmem_cache_alloc_bulk(struct kmem_cache *s, gfp_t flags, size_t size,
2935 void **p)
2936{
2937 struct kmem_cache_cpu *c;
2938 int i;
2939
2940 /* memcg and kmem_cache debug support */
2941 s = slab_pre_alloc_hook(s, flags);
2942 if (unlikely(!s))
2943 return false;
2944 /*
2945 * Drain objects in the per cpu slab, while disabling local
2946 * IRQs, which protects against PREEMPT and interrupts
2947 * handlers invoking normal fastpath.
2948 */
2949 local_irq_disable();
2950 c = this_cpu_ptr(s->cpu_slab);
2951
2952 for (i = 0; i < size; i++) {
2953 void *object = c->freelist;
2954
2955 if (unlikely(!object)) {
2956 /*
2957 * Invoking slow path likely have side-effect
2958 * of re-populating per CPU c->freelist
2959 */
2960 p[i] = ___slab_alloc(s, flags, NUMA_NO_NODE,
2961 _RET_IP_, c);
2962 if (unlikely(!p[i]))
2963 goto error;
2964
2965 c = this_cpu_ptr(s->cpu_slab);
2966 continue; /* goto for-loop */
2967 }
2968 c->freelist = get_freepointer(s, object);
2969 p[i] = object;
2970 }
2971 c->tid = next_tid(c->tid);
2972 local_irq_enable();
2973
2974 /* Clear memory outside IRQ disabled fastpath loop */
2975 if (unlikely(flags & __GFP_ZERO)) {
2976 int j;
2977
2978 for (j = 0; j < i; j++)
2979 memset(p[j], 0, s->object_size);
2980 }
2981
2982 /* memcg and kmem_cache debug support */
2983 slab_post_alloc_hook(s, flags, size, p);
2984 return i;
2985error:
2986 local_irq_enable();
2987 slab_post_alloc_hook(s, flags, i, p);
2988 __kmem_cache_free_bulk(s, i, p);
2989 return 0;
2990}
2991EXPORT_SYMBOL(kmem_cache_alloc_bulk);
2992
2993
2994/*
2995 * Object placement in a slab is made very easy because we always start at
2996 * offset 0. If we tune the size of the object to the alignment then we can
2997 * get the required alignment by putting one properly sized object after
2998 * another.
2999 *
3000 * Notice that the allocation order determines the sizes of the per cpu
3001 * caches. Each processor has always one slab available for allocations.
3002 * Increasing the allocation order reduces the number of times that slabs
3003 * must be moved on and off the partial lists and is therefore a factor in
3004 * locking overhead.
3005 */
3006
3007/*
3008 * Mininum / Maximum order of slab pages. This influences locking overhead
3009 * and slab fragmentation. A higher order reduces the number of partial slabs
3010 * and increases the number of allocations possible without having to
3011 * take the list_lock.
3012 */
3013static int slub_min_order;
3014static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
3015static int slub_min_objects;
3016
3017/*
3018 * Calculate the order of allocation given an slab object size.
3019 *
3020 * The order of allocation has significant impact on performance and other
3021 * system components. Generally order 0 allocations should be preferred since
3022 * order 0 does not cause fragmentation in the page allocator. Larger objects
3023 * be problematic to put into order 0 slabs because there may be too much
3024 * unused space left. We go to a higher order if more than 1/16th of the slab
3025 * would be wasted.
3026 *
3027 * In order to reach satisfactory performance we must ensure that a minimum
3028 * number of objects is in one slab. Otherwise we may generate too much
3029 * activity on the partial lists which requires taking the list_lock. This is
3030 * less a concern for large slabs though which are rarely used.
3031 *
3032 * slub_max_order specifies the order where we begin to stop considering the
3033 * number of objects in a slab as critical. If we reach slub_max_order then
3034 * we try to keep the page order as low as possible. So we accept more waste
3035 * of space in favor of a small page order.
3036 *
3037 * Higher order allocations also allow the placement of more objects in a
3038 * slab and thereby reduce object handling overhead. If the user has
3039 * requested a higher mininum order then we start with that one instead of
3040 * the smallest order which will fit the object.
3041 */
3042static inline int slab_order(int size, int min_objects,
3043 int max_order, int fract_leftover, int reserved)
3044{
3045 int order;
3046 int rem;
3047 int min_order = slub_min_order;
3048
3049 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
3050 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
3051
3052 for (order = max(min_order, get_order(min_objects * size + reserved));
3053 order <= max_order; order++) {
3054
3055 unsigned long slab_size = PAGE_SIZE << order;
3056
3057 rem = (slab_size - reserved) % size;
3058
3059 if (rem <= slab_size / fract_leftover)
3060 break;
3061 }
3062
3063 return order;
3064}
3065
3066static inline int calculate_order(int size, int reserved)
3067{
3068 int order;
3069 int min_objects;
3070 int fraction;
3071 int max_objects;
3072
3073 /*
3074 * Attempt to find best configuration for a slab. This
3075 * works by first attempting to generate a layout with
3076 * the best configuration and backing off gradually.
3077 *
3078 * First we increase the acceptable waste in a slab. Then
3079 * we reduce the minimum objects required in a slab.
3080 */
3081 min_objects = slub_min_objects;
3082 if (!min_objects)
3083 min_objects = 4 * (fls(nr_cpu_ids) + 1);
3084 max_objects = order_objects(slub_max_order, size, reserved);
3085 min_objects = min(min_objects, max_objects);
3086
3087 while (min_objects > 1) {
3088 fraction = 16;
3089 while (fraction >= 4) {
3090 order = slab_order(size, min_objects,
3091 slub_max_order, fraction, reserved);
3092 if (order <= slub_max_order)
3093 return order;
3094 fraction /= 2;
3095 }
3096 min_objects--;
3097 }
3098
3099 /*
3100 * We were unable to place multiple objects in a slab. Now
3101 * lets see if we can place a single object there.
3102 */
3103 order = slab_order(size, 1, slub_max_order, 1, reserved);
3104 if (order <= slub_max_order)
3105 return order;
3106
3107 /*
3108 * Doh this slab cannot be placed using slub_max_order.
3109 */
3110 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
3111 if (order < MAX_ORDER)
3112 return order;
3113 return -ENOSYS;
3114}
3115
3116static void
3117init_kmem_cache_node(struct kmem_cache_node *n)
3118{
3119 n->nr_partial = 0;
3120 spin_lock_init(&n->list_lock);
3121 INIT_LIST_HEAD(&n->partial);
3122#ifdef CONFIG_SLUB_DEBUG
3123 atomic_long_set(&n->nr_slabs, 0);
3124 atomic_long_set(&n->total_objects, 0);
3125 INIT_LIST_HEAD(&n->full);
3126#endif
3127}
3128
3129static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
3130{
3131 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
3132 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
3133
3134 /*
3135 * Must align to double word boundary for the double cmpxchg
3136 * instructions to work; see __pcpu_double_call_return_bool().
3137 */
3138 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
3139 2 * sizeof(void *));
3140
3141 if (!s->cpu_slab)
3142 return 0;
3143
3144 init_kmem_cache_cpus(s);
3145
3146 return 1;
3147}
3148
3149static struct kmem_cache *kmem_cache_node;
3150
3151/*
3152 * No kmalloc_node yet so do it by hand. We know that this is the first
3153 * slab on the node for this slabcache. There are no concurrent accesses
3154 * possible.
3155 *
3156 * Note that this function only works on the kmem_cache_node
3157 * when allocating for the kmem_cache_node. This is used for bootstrapping
3158 * memory on a fresh node that has no slab structures yet.
3159 */
3160static void early_kmem_cache_node_alloc(int node)
3161{
3162 struct page *page;
3163 struct kmem_cache_node *n;
3164
3165 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
3166
3167 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
3168
3169 BUG_ON(!page);
3170 if (page_to_nid(page) != node) {
3171 pr_err("SLUB: Unable to allocate memory from node %d\n", node);
3172 pr_err("SLUB: Allocating a useless per node structure in order to be able to continue\n");
3173 }
3174
3175 n = page->freelist;
3176 BUG_ON(!n);
3177 page->freelist = get_freepointer(kmem_cache_node, n);
3178 page->inuse = 1;
3179 page->frozen = 0;
3180 kmem_cache_node->node[node] = n;
3181#ifdef CONFIG_SLUB_DEBUG
3182 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
3183 init_tracking(kmem_cache_node, n);
3184#endif
3185 kasan_kmalloc(kmem_cache_node, n, sizeof(struct kmem_cache_node),
3186 GFP_KERNEL);
3187 init_kmem_cache_node(n);
3188 inc_slabs_node(kmem_cache_node, node, page->objects);
3189
3190 /*
3191 * No locks need to be taken here as it has just been
3192 * initialized and there is no concurrent access.
3193 */
3194 __add_partial(n, page, DEACTIVATE_TO_HEAD);
3195}
3196
3197static void free_kmem_cache_nodes(struct kmem_cache *s)
3198{
3199 int node;
3200 struct kmem_cache_node *n;
3201
3202 for_each_kmem_cache_node(s, node, n) {
3203 kmem_cache_free(kmem_cache_node, n);
3204 s->node[node] = NULL;
3205 }
3206}
3207
3208void __kmem_cache_release(struct kmem_cache *s)
3209{
3210 free_percpu(s->cpu_slab);
3211 free_kmem_cache_nodes(s);
3212}
3213
3214static int init_kmem_cache_nodes(struct kmem_cache *s)
3215{
3216 int node;
3217
3218 for_each_node_state(node, N_NORMAL_MEMORY) {
3219 struct kmem_cache_node *n;
3220
3221 if (slab_state == DOWN) {
3222 early_kmem_cache_node_alloc(node);
3223 continue;
3224 }
3225 n = kmem_cache_alloc_node(kmem_cache_node,
3226 GFP_KERNEL, node);
3227
3228 if (!n) {
3229 free_kmem_cache_nodes(s);
3230 return 0;
3231 }
3232
3233 s->node[node] = n;
3234 init_kmem_cache_node(n);
3235 }
3236 return 1;
3237}
3238
3239static void set_min_partial(struct kmem_cache *s, unsigned long min)
3240{
3241 if (min < MIN_PARTIAL)
3242 min = MIN_PARTIAL;
3243 else if (min > MAX_PARTIAL)
3244 min = MAX_PARTIAL;
3245 s->min_partial = min;
3246}
3247
3248/*
3249 * calculate_sizes() determines the order and the distribution of data within
3250 * a slab object.
3251 */
3252static int calculate_sizes(struct kmem_cache *s, int forced_order)
3253{
3254 unsigned long flags = s->flags;
3255 unsigned long size = s->object_size;
3256 int order;
3257
3258 /*
3259 * Round up object size to the next word boundary. We can only
3260 * place the free pointer at word boundaries and this determines
3261 * the possible location of the free pointer.
3262 */
3263 size = ALIGN(size, sizeof(void *));
3264
3265#ifdef CONFIG_SLUB_DEBUG
3266 /*
3267 * Determine if we can poison the object itself. If the user of
3268 * the slab may touch the object after free or before allocation
3269 * then we should never poison the object itself.
3270 */
3271 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
3272 !s->ctor)
3273 s->flags |= __OBJECT_POISON;
3274 else
3275 s->flags &= ~__OBJECT_POISON;
3276
3277
3278 /*
3279 * If we are Redzoning then check if there is some space between the
3280 * end of the object and the free pointer. If not then add an
3281 * additional word to have some bytes to store Redzone information.
3282 */
3283 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3284 size += sizeof(void *);
3285#endif
3286
3287 /*
3288 * With that we have determined the number of bytes in actual use
3289 * by the object. This is the potential offset to the free pointer.
3290 */
3291 s->inuse = size;
3292
3293 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3294 s->ctor)) {
3295 /*
3296 * Relocate free pointer after the object if it is not
3297 * permitted to overwrite the first word of the object on
3298 * kmem_cache_free.
3299 *
3300 * This is the case if we do RCU, have a constructor or
3301 * destructor or are poisoning the objects.
3302 */
3303 s->offset = size;
3304 size += sizeof(void *);
3305 }
3306
3307#ifdef CONFIG_SLUB_DEBUG
3308 if (flags & SLAB_STORE_USER)
3309 /*
3310 * Need to store information about allocs and frees after
3311 * the object.
3312 */
3313 size += 2 * sizeof(struct track);
3314
3315 if (flags & SLAB_RED_ZONE) {
3316 /*
3317 * Add some empty padding so that we can catch
3318 * overwrites from earlier objects rather than let
3319 * tracking information or the free pointer be
3320 * corrupted if a user writes before the start
3321 * of the object.
3322 */
3323 size += sizeof(void *);
3324
3325 s->red_left_pad = sizeof(void *);
3326 s->red_left_pad = ALIGN(s->red_left_pad, s->align);
3327 size += s->red_left_pad;
3328 }
3329#endif
3330
3331 /*
3332 * SLUB stores one object immediately after another beginning from
3333 * offset 0. In order to align the objects we have to simply size
3334 * each object to conform to the alignment.
3335 */
3336 size = ALIGN(size, s->align);
3337 s->size = size;
3338 if (forced_order >= 0)
3339 order = forced_order;
3340 else
3341 order = calculate_order(size, s->reserved);
3342
3343 if (order < 0)
3344 return 0;
3345
3346 s->allocflags = 0;
3347 if (order)
3348 s->allocflags |= __GFP_COMP;
3349
3350 if (s->flags & SLAB_CACHE_DMA)
3351 s->allocflags |= GFP_DMA;
3352
3353 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3354 s->allocflags |= __GFP_RECLAIMABLE;
3355
3356 /*
3357 * Determine the number of objects per slab
3358 */
3359 s->oo = oo_make(order, size, s->reserved);
3360 s->min = oo_make(get_order(size), size, s->reserved);
3361 if (oo_objects(s->oo) > oo_objects(s->max))
3362 s->max = s->oo;
3363
3364 return !!oo_objects(s->oo);
3365}
3366
3367static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3368{
3369 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3370 s->reserved = 0;
3371
3372 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3373 s->reserved = sizeof(struct rcu_head);
3374
3375 if (!calculate_sizes(s, -1))
3376 goto error;
3377 if (disable_higher_order_debug) {
3378 /*
3379 * Disable debugging flags that store metadata if the min slab
3380 * order increased.
3381 */
3382 if (get_order(s->size) > get_order(s->object_size)) {
3383 s->flags &= ~DEBUG_METADATA_FLAGS;
3384 s->offset = 0;
3385 if (!calculate_sizes(s, -1))
3386 goto error;
3387 }
3388 }
3389
3390#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3391 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3392 if (system_has_cmpxchg_double() && (s->flags & SLAB_NO_CMPXCHG) == 0)
3393 /* Enable fast mode */
3394 s->flags |= __CMPXCHG_DOUBLE;
3395#endif
3396
3397 /*
3398 * The larger the object size is, the more pages we want on the partial
3399 * list to avoid pounding the page allocator excessively.
3400 */
3401 set_min_partial(s, ilog2(s->size) / 2);
3402
3403 /*
3404 * cpu_partial determined the maximum number of objects kept in the
3405 * per cpu partial lists of a processor.
3406 *
3407 * Per cpu partial lists mainly contain slabs that just have one
3408 * object freed. If they are used for allocation then they can be
3409 * filled up again with minimal effort. The slab will never hit the
3410 * per node partial lists and therefore no locking will be required.
3411 *
3412 * This setting also determines
3413 *
3414 * A) The number of objects from per cpu partial slabs dumped to the
3415 * per node list when we reach the limit.
3416 * B) The number of objects in cpu partial slabs to extract from the
3417 * per node list when we run out of per cpu objects. We only fetch
3418 * 50% to keep some capacity around for frees.
3419 */
3420 if (!kmem_cache_has_cpu_partial(s))
3421 s->cpu_partial = 0;
3422 else if (s->size >= PAGE_SIZE)
3423 s->cpu_partial = 2;
3424 else if (s->size >= 1024)
3425 s->cpu_partial = 6;
3426 else if (s->size >= 256)
3427 s->cpu_partial = 13;
3428 else
3429 s->cpu_partial = 30;
3430
3431#ifdef CONFIG_NUMA
3432 s->remote_node_defrag_ratio = 1000;
3433#endif
3434 if (!init_kmem_cache_nodes(s))
3435 goto error;
3436
3437 if (alloc_kmem_cache_cpus(s))
3438 return 0;
3439
3440 free_kmem_cache_nodes(s);
3441error:
3442 if (flags & SLAB_PANIC)
3443 panic("Cannot create slab %s size=%lu realsize=%u order=%u offset=%u flags=%lx\n",
3444 s->name, (unsigned long)s->size, s->size,
3445 oo_order(s->oo), s->offset, flags);
3446 return -EINVAL;
3447}
3448
3449static void list_slab_objects(struct kmem_cache *s, struct page *page,
3450 const char *text)
3451{
3452#ifdef CONFIG_SLUB_DEBUG
3453 void *addr = page_address(page);
3454 void *p;
3455 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3456 sizeof(long), GFP_ATOMIC);
3457 if (!map)
3458 return;
3459 slab_err(s, page, text, s->name);
3460 slab_lock(page);
3461
3462 get_map(s, page, map);
3463 for_each_object(p, s, addr, page->objects) {
3464
3465 if (!test_bit(slab_index(p, s, addr), map)) {
3466 pr_err("INFO: Object 0x%p @offset=%tu\n", p, p - addr);
3467 print_tracking(s, p);
3468 }
3469 }
3470 slab_unlock(page);
3471 kfree(map);
3472#endif
3473}
3474
3475/*
3476 * Attempt to free all partial slabs on a node.
3477 * This is called from __kmem_cache_shutdown(). We must take list_lock
3478 * because sysfs file might still access partial list after the shutdowning.
3479 */
3480static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3481{
3482 struct page *page, *h;
3483
3484 BUG_ON(irqs_disabled());
3485 spin_lock_irq(&n->list_lock);
3486 list_for_each_entry_safe(page, h, &n->partial, lru) {
3487 if (!page->inuse) {
3488 remove_partial(n, page);
3489 discard_slab(s, page);
3490 } else {
3491 list_slab_objects(s, page,
3492 "Objects remaining in %s on __kmem_cache_shutdown()");
3493 }
3494 }
3495 spin_unlock_irq(&n->list_lock);
3496}
3497
3498/*
3499 * Release all resources used by a slab cache.
3500 */
3501int __kmem_cache_shutdown(struct kmem_cache *s)
3502{
3503 int node;
3504 struct kmem_cache_node *n;
3505
3506 flush_all(s);
3507 /* Attempt to free all objects */
3508 for_each_kmem_cache_node(s, node, n) {
3509 free_partial(s, n);
3510 if (n->nr_partial || slabs_node(s, node))
3511 return 1;
3512 }
3513 return 0;
3514}
3515
3516/********************************************************************
3517 * Kmalloc subsystem
3518 *******************************************************************/
3519
3520static int __init setup_slub_min_order(char *str)
3521{
3522 get_option(&str, &slub_min_order);
3523
3524 return 1;
3525}
3526
3527__setup("slub_min_order=", setup_slub_min_order);
3528
3529static int __init setup_slub_max_order(char *str)
3530{
3531 get_option(&str, &slub_max_order);
3532 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3533
3534 return 1;
3535}
3536
3537__setup("slub_max_order=", setup_slub_max_order);
3538
3539static int __init setup_slub_min_objects(char *str)
3540{
3541 get_option(&str, &slub_min_objects);
3542
3543 return 1;
3544}
3545
3546__setup("slub_min_objects=", setup_slub_min_objects);
3547
3548void *__kmalloc(size_t size, gfp_t flags)
3549{
3550 struct kmem_cache *s;
3551 void *ret;
3552
3553 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3554 return kmalloc_large(size, flags);
3555
3556 s = kmalloc_slab(size, flags);
3557
3558 if (unlikely(ZERO_OR_NULL_PTR(s)))
3559 return s;
3560
3561 ret = slab_alloc(s, flags, _RET_IP_);
3562
3563 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3564
3565 kasan_kmalloc(s, ret, size, flags);
3566
3567 return ret;
3568}
3569EXPORT_SYMBOL(__kmalloc);
3570
3571#ifdef CONFIG_NUMA
3572static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3573{
3574 struct page *page;
3575 void *ptr = NULL;
3576
3577 flags |= __GFP_COMP | __GFP_NOTRACK;
3578 page = alloc_kmem_pages_node(node, flags, get_order(size));
3579 if (page)
3580 ptr = page_address(page);
3581
3582 kmalloc_large_node_hook(ptr, size, flags);
3583 return ptr;
3584}
3585
3586void *__kmalloc_node(size_t size, gfp_t flags, int node)
3587{
3588 struct kmem_cache *s;
3589 void *ret;
3590
3591 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3592 ret = kmalloc_large_node(size, flags, node);
3593
3594 trace_kmalloc_node(_RET_IP_, ret,
3595 size, PAGE_SIZE << get_order(size),
3596 flags, node);
3597
3598 return ret;
3599 }
3600
3601 s = kmalloc_slab(size, flags);
3602
3603 if (unlikely(ZERO_OR_NULL_PTR(s)))
3604 return s;
3605
3606 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3607
3608 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3609
3610 kasan_kmalloc(s, ret, size, flags);
3611
3612 return ret;
3613}
3614EXPORT_SYMBOL(__kmalloc_node);
3615#endif
3616
3617static size_t __ksize(const void *object)
3618{
3619 struct page *page;
3620
3621 if (unlikely(object == ZERO_SIZE_PTR))
3622 return 0;
3623
3624 page = virt_to_head_page(object);
3625
3626 if (unlikely(!PageSlab(page))) {
3627 WARN_ON(!PageCompound(page));
3628 return PAGE_SIZE << compound_order(page);
3629 }
3630
3631 return slab_ksize(page->slab_cache);
3632}
3633
3634size_t ksize(const void *object)
3635{
3636 size_t size = __ksize(object);
3637 /* We assume that ksize callers could use whole allocated area,
3638 so we need unpoison this area. */
3639 kasan_krealloc(object, size, GFP_NOWAIT);
3640 return size;
3641}
3642EXPORT_SYMBOL(ksize);
3643
3644void kfree(const void *x)
3645{
3646 struct page *page;
3647 void *object = (void *)x;
3648
3649 trace_kfree(_RET_IP_, x);
3650
3651 if (unlikely(ZERO_OR_NULL_PTR(x)))
3652 return;
3653
3654 page = virt_to_head_page(x);
3655 if (unlikely(!PageSlab(page))) {
3656 BUG_ON(!PageCompound(page));
3657 kfree_hook(x);
3658 __free_kmem_pages(page, compound_order(page));
3659 return;
3660 }
3661 slab_free(page->slab_cache, page, object, NULL, 1, _RET_IP_);
3662}
3663EXPORT_SYMBOL(kfree);
3664
3665#define SHRINK_PROMOTE_MAX 32
3666
3667/*
3668 * kmem_cache_shrink discards empty slabs and promotes the slabs filled
3669 * up most to the head of the partial lists. New allocations will then
3670 * fill those up and thus they can be removed from the partial lists.
3671 *
3672 * The slabs with the least items are placed last. This results in them
3673 * being allocated from last increasing the chance that the last objects
3674 * are freed in them.
3675 */
3676int __kmem_cache_shrink(struct kmem_cache *s, bool deactivate)
3677{
3678 int node;
3679 int i;
3680 struct kmem_cache_node *n;
3681 struct page *page;
3682 struct page *t;
3683 struct list_head discard;
3684 struct list_head promote[SHRINK_PROMOTE_MAX];
3685 unsigned long flags;
3686 int ret = 0;
3687
3688 if (deactivate) {
3689 /*
3690 * Disable empty slabs caching. Used to avoid pinning offline
3691 * memory cgroups by kmem pages that can be freed.
3692 */
3693 s->cpu_partial = 0;
3694 s->min_partial = 0;
3695
3696 /*
3697 * s->cpu_partial is checked locklessly (see put_cpu_partial),
3698 * so we have to make sure the change is visible.
3699 */
3700 kick_all_cpus_sync();
3701 }
3702
3703 flush_all(s);
3704 for_each_kmem_cache_node(s, node, n) {
3705 INIT_LIST_HEAD(&discard);
3706 for (i = 0; i < SHRINK_PROMOTE_MAX; i++)
3707 INIT_LIST_HEAD(promote + i);
3708
3709 spin_lock_irqsave(&n->list_lock, flags);
3710
3711 /*
3712 * Build lists of slabs to discard or promote.
3713 *
3714 * Note that concurrent frees may occur while we hold the
3715 * list_lock. page->inuse here is the upper limit.
3716 */
3717 list_for_each_entry_safe(page, t, &n->partial, lru) {
3718 int free = page->objects - page->inuse;
3719
3720 /* Do not reread page->inuse */
3721 barrier();
3722
3723 /* We do not keep full slabs on the list */
3724 BUG_ON(free <= 0);
3725
3726 if (free == page->objects) {
3727 list_move(&page->lru, &discard);
3728 n->nr_partial--;
3729 } else if (free <= SHRINK_PROMOTE_MAX)
3730 list_move(&page->lru, promote + free - 1);
3731 }
3732
3733 /*
3734 * Promote the slabs filled up most to the head of the
3735 * partial list.
3736 */
3737 for (i = SHRINK_PROMOTE_MAX - 1; i >= 0; i--)
3738 list_splice(promote + i, &n->partial);
3739
3740 spin_unlock_irqrestore(&n->list_lock, flags);
3741
3742 /* Release empty slabs */
3743 list_for_each_entry_safe(page, t, &discard, lru)
3744 discard_slab(s, page);
3745
3746 if (slabs_node(s, node))
3747 ret = 1;
3748 }
3749
3750 return ret;
3751}
3752
3753static int slab_mem_going_offline_callback(void *arg)
3754{
3755 struct kmem_cache *s;
3756
3757 mutex_lock(&slab_mutex);
3758 list_for_each_entry(s, &slab_caches, list)
3759 __kmem_cache_shrink(s, false);
3760 mutex_unlock(&slab_mutex);
3761
3762 return 0;
3763}
3764
3765static void slab_mem_offline_callback(void *arg)
3766{
3767 struct kmem_cache_node *n;
3768 struct kmem_cache *s;
3769 struct memory_notify *marg = arg;
3770 int offline_node;
3771
3772 offline_node = marg->status_change_nid_normal;
3773
3774 /*
3775 * If the node still has available memory. we need kmem_cache_node
3776 * for it yet.
3777 */
3778 if (offline_node < 0)
3779 return;
3780
3781 mutex_lock(&slab_mutex);
3782 list_for_each_entry(s, &slab_caches, list) {
3783 n = get_node(s, offline_node);
3784 if (n) {
3785 /*
3786 * if n->nr_slabs > 0, slabs still exist on the node
3787 * that is going down. We were unable to free them,
3788 * and offline_pages() function shouldn't call this
3789 * callback. So, we must fail.
3790 */
3791 BUG_ON(slabs_node(s, offline_node));
3792
3793 s->node[offline_node] = NULL;
3794 kmem_cache_free(kmem_cache_node, n);
3795 }
3796 }
3797 mutex_unlock(&slab_mutex);
3798}
3799
3800static int slab_mem_going_online_callback(void *arg)
3801{
3802 struct kmem_cache_node *n;
3803 struct kmem_cache *s;
3804 struct memory_notify *marg = arg;
3805 int nid = marg->status_change_nid_normal;
3806 int ret = 0;
3807
3808 /*
3809 * If the node's memory is already available, then kmem_cache_node is
3810 * already created. Nothing to do.
3811 */
3812 if (nid < 0)
3813 return 0;
3814
3815 /*
3816 * We are bringing a node online. No memory is available yet. We must
3817 * allocate a kmem_cache_node structure in order to bring the node
3818 * online.
3819 */
3820 mutex_lock(&slab_mutex);
3821 list_for_each_entry(s, &slab_caches, list) {
3822 /*
3823 * XXX: kmem_cache_alloc_node will fallback to other nodes
3824 * since memory is not yet available from the node that
3825 * is brought up.
3826 */
3827 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3828 if (!n) {
3829 ret = -ENOMEM;
3830 goto out;
3831 }
3832 init_kmem_cache_node(n);
3833 s->node[nid] = n;
3834 }
3835out:
3836 mutex_unlock(&slab_mutex);
3837 return ret;
3838}
3839
3840static int slab_memory_callback(struct notifier_block *self,
3841 unsigned long action, void *arg)
3842{
3843 int ret = 0;
3844
3845 switch (action) {
3846 case MEM_GOING_ONLINE:
3847 ret = slab_mem_going_online_callback(arg);
3848 break;
3849 case MEM_GOING_OFFLINE:
3850 ret = slab_mem_going_offline_callback(arg);
3851 break;
3852 case MEM_OFFLINE:
3853 case MEM_CANCEL_ONLINE:
3854 slab_mem_offline_callback(arg);
3855 break;
3856 case MEM_ONLINE:
3857 case MEM_CANCEL_OFFLINE:
3858 break;
3859 }
3860 if (ret)
3861 ret = notifier_from_errno(ret);
3862 else
3863 ret = NOTIFY_OK;
3864 return ret;
3865}
3866
3867static struct notifier_block slab_memory_callback_nb = {
3868 .notifier_call = slab_memory_callback,
3869 .priority = SLAB_CALLBACK_PRI,
3870};
3871
3872/********************************************************************
3873 * Basic setup of slabs
3874 *******************************************************************/
3875
3876/*
3877 * Used for early kmem_cache structures that were allocated using
3878 * the page allocator. Allocate them properly then fix up the pointers
3879 * that may be pointing to the wrong kmem_cache structure.
3880 */
3881
3882static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3883{
3884 int node;
3885 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3886 struct kmem_cache_node *n;
3887
3888 memcpy(s, static_cache, kmem_cache->object_size);
3889
3890 /*
3891 * This runs very early, and only the boot processor is supposed to be
3892 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3893 * IPIs around.
3894 */
3895 __flush_cpu_slab(s, smp_processor_id());
3896 for_each_kmem_cache_node(s, node, n) {
3897 struct page *p;
3898
3899 list_for_each_entry(p, &n->partial, lru)
3900 p->slab_cache = s;
3901
3902#ifdef CONFIG_SLUB_DEBUG
3903 list_for_each_entry(p, &n->full, lru)
3904 p->slab_cache = s;
3905#endif
3906 }
3907 slab_init_memcg_params(s);
3908 list_add(&s->list, &slab_caches);
3909 return s;
3910}
3911
3912void __init kmem_cache_init(void)
3913{
3914 static __initdata struct kmem_cache boot_kmem_cache,
3915 boot_kmem_cache_node;
3916
3917 if (debug_guardpage_minorder())
3918 slub_max_order = 0;
3919
3920 kmem_cache_node = &boot_kmem_cache_node;
3921 kmem_cache = &boot_kmem_cache;
3922
3923 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3924 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3925
3926 register_hotmemory_notifier(&slab_memory_callback_nb);
3927
3928 /* Able to allocate the per node structures */
3929 slab_state = PARTIAL;
3930
3931 create_boot_cache(kmem_cache, "kmem_cache",
3932 offsetof(struct kmem_cache, node) +
3933 nr_node_ids * sizeof(struct kmem_cache_node *),
3934 SLAB_HWCACHE_ALIGN);
3935
3936 kmem_cache = bootstrap(&boot_kmem_cache);
3937
3938 /*
3939 * Allocate kmem_cache_node properly from the kmem_cache slab.
3940 * kmem_cache_node is separately allocated so no need to
3941 * update any list pointers.
3942 */
3943 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3944
3945 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3946 setup_kmalloc_cache_index_table();
3947 create_kmalloc_caches(0);
3948
3949#ifdef CONFIG_SMP
3950 register_cpu_notifier(&slab_notifier);
3951#endif
3952
3953 pr_info("SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d, CPUs=%d, Nodes=%d\n",
3954 cache_line_size(),
3955 slub_min_order, slub_max_order, slub_min_objects,
3956 nr_cpu_ids, nr_node_ids);
3957}
3958
3959void __init kmem_cache_init_late(void)
3960{
3961}
3962
3963struct kmem_cache *
3964__kmem_cache_alias(const char *name, size_t size, size_t align,
3965 unsigned long flags, void (*ctor)(void *))
3966{
3967 struct kmem_cache *s, *c;
3968
3969 s = find_mergeable(size, align, flags, name, ctor);
3970 if (s) {
3971 s->refcount++;
3972
3973 /*
3974 * Adjust the object sizes so that we clear
3975 * the complete object on kzalloc.
3976 */
3977 s->object_size = max(s->object_size, (int)size);
3978 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3979
3980 for_each_memcg_cache(c, s) {
3981 c->object_size = s->object_size;
3982 c->inuse = max_t(int, c->inuse,
3983 ALIGN(size, sizeof(void *)));
3984 }
3985
3986 if (sysfs_slab_alias(s, name)) {
3987 s->refcount--;
3988 s = NULL;
3989 }
3990 }
3991
3992 return s;
3993}
3994
3995int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3996{
3997 int err;
3998
3999 err = kmem_cache_open(s, flags);
4000 if (err)
4001 return err;
4002
4003 /* Mutex is not taken during early boot */
4004 if (slab_state <= UP)
4005 return 0;
4006
4007 memcg_propagate_slab_attrs(s);
4008 err = sysfs_slab_add(s);
4009 if (err)
4010 __kmem_cache_release(s);
4011
4012 return err;
4013}
4014
4015#ifdef CONFIG_SMP
4016/*
4017 * Use the cpu notifier to insure that the cpu slabs are flushed when
4018 * necessary.
4019 */
4020static int slab_cpuup_callback(struct notifier_block *nfb,
4021 unsigned long action, void *hcpu)
4022{
4023 long cpu = (long)hcpu;
4024 struct kmem_cache *s;
4025 unsigned long flags;
4026
4027 switch (action) {
4028 case CPU_UP_CANCELED:
4029 case CPU_UP_CANCELED_FROZEN:
4030 case CPU_DEAD:
4031 case CPU_DEAD_FROZEN:
4032 mutex_lock(&slab_mutex);
4033 list_for_each_entry(s, &slab_caches, list) {
4034 local_irq_save(flags);
4035 __flush_cpu_slab(s, cpu);
4036 local_irq_restore(flags);
4037 }
4038 mutex_unlock(&slab_mutex);
4039 break;
4040 default:
4041 break;
4042 }
4043 return NOTIFY_OK;
4044}
4045
4046static struct notifier_block slab_notifier = {
4047 .notifier_call = slab_cpuup_callback
4048};
4049
4050#endif
4051
4052void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4053{
4054 struct kmem_cache *s;
4055 void *ret;
4056
4057 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
4058 return kmalloc_large(size, gfpflags);
4059
4060 s = kmalloc_slab(size, gfpflags);
4061
4062 if (unlikely(ZERO_OR_NULL_PTR(s)))
4063 return s;
4064
4065 ret = slab_alloc(s, gfpflags, caller);
4066
4067 /* Honor the call site pointer we received. */
4068 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4069
4070 return ret;
4071}
4072
4073#ifdef CONFIG_NUMA
4074void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4075 int node, unsigned long caller)
4076{
4077 struct kmem_cache *s;
4078 void *ret;
4079
4080 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
4081 ret = kmalloc_large_node(size, gfpflags, node);
4082
4083 trace_kmalloc_node(caller, ret,
4084 size, PAGE_SIZE << get_order(size),
4085 gfpflags, node);
4086
4087 return ret;
4088 }
4089
4090 s = kmalloc_slab(size, gfpflags);
4091
4092 if (unlikely(ZERO_OR_NULL_PTR(s)))
4093 return s;
4094
4095 ret = slab_alloc_node(s, gfpflags, node, caller);
4096
4097 /* Honor the call site pointer we received. */
4098 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4099
4100 return ret;
4101}
4102#endif
4103
4104#ifdef CONFIG_SYSFS
4105static int count_inuse(struct page *page)
4106{
4107 return page->inuse;
4108}
4109
4110static int count_total(struct page *page)
4111{
4112 return page->objects;
4113}
4114#endif
4115
4116#ifdef CONFIG_SLUB_DEBUG
4117static int validate_slab(struct kmem_cache *s, struct page *page,
4118 unsigned long *map)
4119{
4120 void *p;
4121 void *addr = page_address(page);
4122
4123 if (!check_slab(s, page) ||
4124 !on_freelist(s, page, NULL))
4125 return 0;
4126
4127 /* Now we know that a valid freelist exists */
4128 bitmap_zero(map, page->objects);
4129
4130 get_map(s, page, map);
4131 for_each_object(p, s, addr, page->objects) {
4132 if (test_bit(slab_index(p, s, addr), map))
4133 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4134 return 0;
4135 }
4136
4137 for_each_object(p, s, addr, page->objects)
4138 if (!test_bit(slab_index(p, s, addr), map))
4139 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4140 return 0;
4141 return 1;
4142}
4143
4144static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4145 unsigned long *map)
4146{
4147 slab_lock(page);
4148 validate_slab(s, page, map);
4149 slab_unlock(page);
4150}
4151
4152static int validate_slab_node(struct kmem_cache *s,
4153 struct kmem_cache_node *n, unsigned long *map)
4154{
4155 unsigned long count = 0;
4156 struct page *page;
4157 unsigned long flags;
4158
4159 spin_lock_irqsave(&n->list_lock, flags);
4160
4161 list_for_each_entry(page, &n->partial, lru) {
4162 validate_slab_slab(s, page, map);
4163 count++;
4164 }
4165 if (count != n->nr_partial)
4166 pr_err("SLUB %s: %ld partial slabs counted but counter=%ld\n",
4167 s->name, count, n->nr_partial);
4168
4169 if (!(s->flags & SLAB_STORE_USER))
4170 goto out;
4171
4172 list_for_each_entry(page, &n->full, lru) {
4173 validate_slab_slab(s, page, map);
4174 count++;
4175 }
4176 if (count != atomic_long_read(&n->nr_slabs))
4177 pr_err("SLUB: %s %ld slabs counted but counter=%ld\n",
4178 s->name, count, atomic_long_read(&n->nr_slabs));
4179
4180out:
4181 spin_unlock_irqrestore(&n->list_lock, flags);
4182 return count;
4183}
4184
4185static long validate_slab_cache(struct kmem_cache *s)
4186{
4187 int node;
4188 unsigned long count = 0;
4189 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4190 sizeof(unsigned long), GFP_KERNEL);
4191 struct kmem_cache_node *n;
4192
4193 if (!map)
4194 return -ENOMEM;
4195
4196 flush_all(s);
4197 for_each_kmem_cache_node(s, node, n)
4198 count += validate_slab_node(s, n, map);
4199 kfree(map);
4200 return count;
4201}
4202/*
4203 * Generate lists of code addresses where slabcache objects are allocated
4204 * and freed.
4205 */
4206
4207struct location {
4208 unsigned long count;
4209 unsigned long addr;
4210 long long sum_time;
4211 long min_time;
4212 long max_time;
4213 long min_pid;
4214 long max_pid;
4215 DECLARE_BITMAP(cpus, NR_CPUS);
4216 nodemask_t nodes;
4217};
4218
4219struct loc_track {
4220 unsigned long max;
4221 unsigned long count;
4222 struct location *loc;
4223};
4224
4225static void free_loc_track(struct loc_track *t)
4226{
4227 if (t->max)
4228 free_pages((unsigned long)t->loc,
4229 get_order(sizeof(struct location) * t->max));
4230}
4231
4232static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4233{
4234 struct location *l;
4235 int order;
4236
4237 order = get_order(sizeof(struct location) * max);
4238
4239 l = (void *)__get_free_pages(flags, order);
4240 if (!l)
4241 return 0;
4242
4243 if (t->count) {
4244 memcpy(l, t->loc, sizeof(struct location) * t->count);
4245 free_loc_track(t);
4246 }
4247 t->max = max;
4248 t->loc = l;
4249 return 1;
4250}
4251
4252static int add_location(struct loc_track *t, struct kmem_cache *s,
4253 const struct track *track)
4254{
4255 long start, end, pos;
4256 struct location *l;
4257 unsigned long caddr;
4258 unsigned long age = jiffies - track->when;
4259
4260 start = -1;
4261 end = t->count;
4262
4263 for ( ; ; ) {
4264 pos = start + (end - start + 1) / 2;
4265
4266 /*
4267 * There is nothing at "end". If we end up there
4268 * we need to add something to before end.
4269 */
4270 if (pos == end)
4271 break;
4272
4273 caddr = t->loc[pos].addr;
4274 if (track->addr == caddr) {
4275
4276 l = &t->loc[pos];
4277 l->count++;
4278 if (track->when) {
4279 l->sum_time += age;
4280 if (age < l->min_time)
4281 l->min_time = age;
4282 if (age > l->max_time)
4283 l->max_time = age;
4284
4285 if (track->pid < l->min_pid)
4286 l->min_pid = track->pid;
4287 if (track->pid > l->max_pid)
4288 l->max_pid = track->pid;
4289
4290 cpumask_set_cpu(track->cpu,
4291 to_cpumask(l->cpus));
4292 }
4293 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4294 return 1;
4295 }
4296
4297 if (track->addr < caddr)
4298 end = pos;
4299 else
4300 start = pos;
4301 }
4302
4303 /*
4304 * Not found. Insert new tracking element.
4305 */
4306 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4307 return 0;
4308
4309 l = t->loc + pos;
4310 if (pos < t->count)
4311 memmove(l + 1, l,
4312 (t->count - pos) * sizeof(struct location));
4313 t->count++;
4314 l->count = 1;
4315 l->addr = track->addr;
4316 l->sum_time = age;
4317 l->min_time = age;
4318 l->max_time = age;
4319 l->min_pid = track->pid;
4320 l->max_pid = track->pid;
4321 cpumask_clear(to_cpumask(l->cpus));
4322 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4323 nodes_clear(l->nodes);
4324 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4325 return 1;
4326}
4327
4328static void process_slab(struct loc_track *t, struct kmem_cache *s,
4329 struct page *page, enum track_item alloc,
4330 unsigned long *map)
4331{
4332 void *addr = page_address(page);
4333 void *p;
4334
4335 bitmap_zero(map, page->objects);
4336 get_map(s, page, map);
4337
4338 for_each_object(p, s, addr, page->objects)
4339 if (!test_bit(slab_index(p, s, addr), map))
4340 add_location(t, s, get_track(s, p, alloc));
4341}
4342
4343static int list_locations(struct kmem_cache *s, char *buf,
4344 enum track_item alloc)
4345{
4346 int len = 0;
4347 unsigned long i;
4348 struct loc_track t = { 0, 0, NULL };
4349 int node;
4350 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4351 sizeof(unsigned long), GFP_KERNEL);
4352 struct kmem_cache_node *n;
4353
4354 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4355 GFP_TEMPORARY)) {
4356 kfree(map);
4357 return sprintf(buf, "Out of memory\n");
4358 }
4359 /* Push back cpu slabs */
4360 flush_all(s);
4361
4362 for_each_kmem_cache_node(s, node, n) {
4363 unsigned long flags;
4364 struct page *page;
4365
4366 if (!atomic_long_read(&n->nr_slabs))
4367 continue;
4368
4369 spin_lock_irqsave(&n->list_lock, flags);
4370 list_for_each_entry(page, &n->partial, lru)
4371 process_slab(&t, s, page, alloc, map);
4372 list_for_each_entry(page, &n->full, lru)
4373 process_slab(&t, s, page, alloc, map);
4374 spin_unlock_irqrestore(&n->list_lock, flags);
4375 }
4376
4377 for (i = 0; i < t.count; i++) {
4378 struct location *l = &t.loc[i];
4379
4380 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4381 break;
4382 len += sprintf(buf + len, "%7ld ", l->count);
4383
4384 if (l->addr)
4385 len += sprintf(buf + len, "%pS", (void *)l->addr);
4386 else
4387 len += sprintf(buf + len, "<not-available>");
4388
4389 if (l->sum_time != l->min_time) {
4390 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4391 l->min_time,
4392 (long)div_u64(l->sum_time, l->count),
4393 l->max_time);
4394 } else
4395 len += sprintf(buf + len, " age=%ld",
4396 l->min_time);
4397
4398 if (l->min_pid != l->max_pid)
4399 len += sprintf(buf + len, " pid=%ld-%ld",
4400 l->min_pid, l->max_pid);
4401 else
4402 len += sprintf(buf + len, " pid=%ld",
4403 l->min_pid);
4404
4405 if (num_online_cpus() > 1 &&
4406 !cpumask_empty(to_cpumask(l->cpus)) &&
4407 len < PAGE_SIZE - 60)
4408 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4409 " cpus=%*pbl",
4410 cpumask_pr_args(to_cpumask(l->cpus)));
4411
4412 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4413 len < PAGE_SIZE - 60)
4414 len += scnprintf(buf + len, PAGE_SIZE - len - 50,
4415 " nodes=%*pbl",
4416 nodemask_pr_args(&l->nodes));
4417
4418 len += sprintf(buf + len, "\n");
4419 }
4420
4421 free_loc_track(&t);
4422 kfree(map);
4423 if (!t.count)
4424 len += sprintf(buf, "No data\n");
4425 return len;
4426}
4427#endif
4428
4429#ifdef SLUB_RESILIENCY_TEST
4430static void __init resiliency_test(void)
4431{
4432 u8 *p;
4433
4434 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4435
4436 pr_err("SLUB resiliency testing\n");
4437 pr_err("-----------------------\n");
4438 pr_err("A. Corruption after allocation\n");
4439
4440 p = kzalloc(16, GFP_KERNEL);
4441 p[16] = 0x12;
4442 pr_err("\n1. kmalloc-16: Clobber Redzone/next pointer 0x12->0x%p\n\n",
4443 p + 16);
4444
4445 validate_slab_cache(kmalloc_caches[4]);
4446
4447 /* Hmmm... The next two are dangerous */
4448 p = kzalloc(32, GFP_KERNEL);
4449 p[32 + sizeof(void *)] = 0x34;
4450 pr_err("\n2. kmalloc-32: Clobber next pointer/next slab 0x34 -> -0x%p\n",
4451 p);
4452 pr_err("If allocated object is overwritten then not detectable\n\n");
4453
4454 validate_slab_cache(kmalloc_caches[5]);
4455 p = kzalloc(64, GFP_KERNEL);
4456 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4457 *p = 0x56;
4458 pr_err("\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4459 p);
4460 pr_err("If allocated object is overwritten then not detectable\n\n");
4461 validate_slab_cache(kmalloc_caches[6]);
4462
4463 pr_err("\nB. Corruption after free\n");
4464 p = kzalloc(128, GFP_KERNEL);
4465 kfree(p);
4466 *p = 0x78;
4467 pr_err("1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4468 validate_slab_cache(kmalloc_caches[7]);
4469
4470 p = kzalloc(256, GFP_KERNEL);
4471 kfree(p);
4472 p[50] = 0x9a;
4473 pr_err("\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
4474 validate_slab_cache(kmalloc_caches[8]);
4475
4476 p = kzalloc(512, GFP_KERNEL);
4477 kfree(p);
4478 p[512] = 0xab;
4479 pr_err("\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4480 validate_slab_cache(kmalloc_caches[9]);
4481}
4482#else
4483#ifdef CONFIG_SYSFS
4484static void resiliency_test(void) {};
4485#endif
4486#endif
4487
4488#ifdef CONFIG_SYSFS
4489enum slab_stat_type {
4490 SL_ALL, /* All slabs */
4491 SL_PARTIAL, /* Only partially allocated slabs */
4492 SL_CPU, /* Only slabs used for cpu caches */
4493 SL_OBJECTS, /* Determine allocated objects not slabs */
4494 SL_TOTAL /* Determine object capacity not slabs */
4495};
4496
4497#define SO_ALL (1 << SL_ALL)
4498#define SO_PARTIAL (1 << SL_PARTIAL)
4499#define SO_CPU (1 << SL_CPU)
4500#define SO_OBJECTS (1 << SL_OBJECTS)
4501#define SO_TOTAL (1 << SL_TOTAL)
4502
4503static ssize_t show_slab_objects(struct kmem_cache *s,
4504 char *buf, unsigned long flags)
4505{
4506 unsigned long total = 0;
4507 int node;
4508 int x;
4509 unsigned long *nodes;
4510
4511 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4512 if (!nodes)
4513 return -ENOMEM;
4514
4515 if (flags & SO_CPU) {
4516 int cpu;
4517
4518 for_each_possible_cpu(cpu) {
4519 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4520 cpu);
4521 int node;
4522 struct page *page;
4523
4524 page = READ_ONCE(c->page);
4525 if (!page)
4526 continue;
4527
4528 node = page_to_nid(page);
4529 if (flags & SO_TOTAL)
4530 x = page->objects;
4531 else if (flags & SO_OBJECTS)
4532 x = page->inuse;
4533 else
4534 x = 1;
4535
4536 total += x;
4537 nodes[node] += x;
4538
4539 page = READ_ONCE(c->partial);
4540 if (page) {
4541 node = page_to_nid(page);
4542 if (flags & SO_TOTAL)
4543 WARN_ON_ONCE(1);
4544 else if (flags & SO_OBJECTS)
4545 WARN_ON_ONCE(1);
4546 else
4547 x = page->pages;
4548 total += x;
4549 nodes[node] += x;
4550 }
4551 }
4552 }
4553
4554 get_online_mems();
4555#ifdef CONFIG_SLUB_DEBUG
4556 if (flags & SO_ALL) {
4557 struct kmem_cache_node *n;
4558
4559 for_each_kmem_cache_node(s, node, n) {
4560
4561 if (flags & SO_TOTAL)
4562 x = atomic_long_read(&n->total_objects);
4563 else if (flags & SO_OBJECTS)
4564 x = atomic_long_read(&n->total_objects) -
4565 count_partial(n, count_free);
4566 else
4567 x = atomic_long_read(&n->nr_slabs);
4568 total += x;
4569 nodes[node] += x;
4570 }
4571
4572 } else
4573#endif
4574 if (flags & SO_PARTIAL) {
4575 struct kmem_cache_node *n;
4576
4577 for_each_kmem_cache_node(s, node, n) {
4578 if (flags & SO_TOTAL)
4579 x = count_partial(n, count_total);
4580 else if (flags & SO_OBJECTS)
4581 x = count_partial(n, count_inuse);
4582 else
4583 x = n->nr_partial;
4584 total += x;
4585 nodes[node] += x;
4586 }
4587 }
4588 x = sprintf(buf, "%lu", total);
4589#ifdef CONFIG_NUMA
4590 for (node = 0; node < nr_node_ids; node++)
4591 if (nodes[node])
4592 x += sprintf(buf + x, " N%d=%lu",
4593 node, nodes[node]);
4594#endif
4595 put_online_mems();
4596 kfree(nodes);
4597 return x + sprintf(buf + x, "\n");
4598}
4599
4600#ifdef CONFIG_SLUB_DEBUG
4601static int any_slab_objects(struct kmem_cache *s)
4602{
4603 int node;
4604 struct kmem_cache_node *n;
4605
4606 for_each_kmem_cache_node(s, node, n)
4607 if (atomic_long_read(&n->total_objects))
4608 return 1;
4609
4610 return 0;
4611}
4612#endif
4613
4614#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4615#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4616
4617struct slab_attribute {
4618 struct attribute attr;
4619 ssize_t (*show)(struct kmem_cache *s, char *buf);
4620 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4621};
4622
4623#define SLAB_ATTR_RO(_name) \
4624 static struct slab_attribute _name##_attr = \
4625 __ATTR(_name, 0400, _name##_show, NULL)
4626
4627#define SLAB_ATTR(_name) \
4628 static struct slab_attribute _name##_attr = \
4629 __ATTR(_name, 0600, _name##_show, _name##_store)
4630
4631static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4632{
4633 return sprintf(buf, "%d\n", s->size);
4634}
4635SLAB_ATTR_RO(slab_size);
4636
4637static ssize_t align_show(struct kmem_cache *s, char *buf)
4638{
4639 return sprintf(buf, "%d\n", s->align);
4640}
4641SLAB_ATTR_RO(align);
4642
4643static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4644{
4645 return sprintf(buf, "%d\n", s->object_size);
4646}
4647SLAB_ATTR_RO(object_size);
4648
4649static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4650{
4651 return sprintf(buf, "%d\n", oo_objects(s->oo));
4652}
4653SLAB_ATTR_RO(objs_per_slab);
4654
4655static ssize_t order_store(struct kmem_cache *s,
4656 const char *buf, size_t length)
4657{
4658 unsigned long order;
4659 int err;
4660
4661 err = kstrtoul(buf, 10, &order);
4662 if (err)
4663 return err;
4664
4665 if (order > slub_max_order || order < slub_min_order)
4666 return -EINVAL;
4667
4668 calculate_sizes(s, order);
4669 return length;
4670}
4671
4672static ssize_t order_show(struct kmem_cache *s, char *buf)
4673{
4674 return sprintf(buf, "%d\n", oo_order(s->oo));
4675}
4676SLAB_ATTR(order);
4677
4678static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4679{
4680 return sprintf(buf, "%lu\n", s->min_partial);
4681}
4682
4683static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4684 size_t length)
4685{
4686 unsigned long min;
4687 int err;
4688
4689 err = kstrtoul(buf, 10, &min);
4690 if (err)
4691 return err;
4692
4693 set_min_partial(s, min);
4694 return length;
4695}
4696SLAB_ATTR(min_partial);
4697
4698static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4699{
4700 return sprintf(buf, "%u\n", s->cpu_partial);
4701}
4702
4703static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4704 size_t length)
4705{
4706 unsigned long objects;
4707 int err;
4708
4709 err = kstrtoul(buf, 10, &objects);
4710 if (err)
4711 return err;
4712 if (objects && !kmem_cache_has_cpu_partial(s))
4713 return -EINVAL;
4714
4715 s->cpu_partial = objects;
4716 flush_all(s);
4717 return length;
4718}
4719SLAB_ATTR(cpu_partial);
4720
4721static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4722{
4723 if (!s->ctor)
4724 return 0;
4725 return sprintf(buf, "%pS\n", s->ctor);
4726}
4727SLAB_ATTR_RO(ctor);
4728
4729static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4730{
4731 return sprintf(buf, "%d\n", s->refcount < 0 ? 0 : s->refcount - 1);
4732}
4733SLAB_ATTR_RO(aliases);
4734
4735static ssize_t partial_show(struct kmem_cache *s, char *buf)
4736{
4737 return show_slab_objects(s, buf, SO_PARTIAL);
4738}
4739SLAB_ATTR_RO(partial);
4740
4741static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4742{
4743 return show_slab_objects(s, buf, SO_CPU);
4744}
4745SLAB_ATTR_RO(cpu_slabs);
4746
4747static ssize_t objects_show(struct kmem_cache *s, char *buf)
4748{
4749 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4750}
4751SLAB_ATTR_RO(objects);
4752
4753static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4754{
4755 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4756}
4757SLAB_ATTR_RO(objects_partial);
4758
4759static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4760{
4761 int objects = 0;
4762 int pages = 0;
4763 int cpu;
4764 int len;
4765
4766 for_each_online_cpu(cpu) {
4767 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4768
4769 if (page) {
4770 pages += page->pages;
4771 objects += page->pobjects;
4772 }
4773 }
4774
4775 len = sprintf(buf, "%d(%d)", objects, pages);
4776
4777#ifdef CONFIG_SMP
4778 for_each_online_cpu(cpu) {
4779 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4780
4781 if (page && len < PAGE_SIZE - 20)
4782 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4783 page->pobjects, page->pages);
4784 }
4785#endif
4786 return len + sprintf(buf + len, "\n");
4787}
4788SLAB_ATTR_RO(slabs_cpu_partial);
4789
4790static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4791{
4792 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4793}
4794
4795static ssize_t reclaim_account_store(struct kmem_cache *s,
4796 const char *buf, size_t length)
4797{
4798 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4799 if (buf[0] == '1')
4800 s->flags |= SLAB_RECLAIM_ACCOUNT;
4801 return length;
4802}
4803SLAB_ATTR(reclaim_account);
4804
4805static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4806{
4807 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4808}
4809SLAB_ATTR_RO(hwcache_align);
4810
4811#ifdef CONFIG_ZONE_DMA
4812static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4813{
4814 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4815}
4816SLAB_ATTR_RO(cache_dma);
4817#endif
4818
4819static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4820{
4821 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4822}
4823SLAB_ATTR_RO(destroy_by_rcu);
4824
4825static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4826{
4827 return sprintf(buf, "%d\n", s->reserved);
4828}
4829SLAB_ATTR_RO(reserved);
4830
4831#ifdef CONFIG_SLUB_DEBUG
4832static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4833{
4834 return show_slab_objects(s, buf, SO_ALL);
4835}
4836SLAB_ATTR_RO(slabs);
4837
4838static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4839{
4840 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4841}
4842SLAB_ATTR_RO(total_objects);
4843
4844static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4845{
4846 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CONSISTENCY_CHECKS));
4847}
4848
4849static ssize_t sanity_checks_store(struct kmem_cache *s,
4850 const char *buf, size_t length)
4851{
4852 s->flags &= ~SLAB_CONSISTENCY_CHECKS;
4853 if (buf[0] == '1') {
4854 s->flags &= ~__CMPXCHG_DOUBLE;
4855 s->flags |= SLAB_CONSISTENCY_CHECKS;
4856 }
4857 return length;
4858}
4859SLAB_ATTR(sanity_checks);
4860
4861static ssize_t trace_show(struct kmem_cache *s, char *buf)
4862{
4863 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4864}
4865
4866static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4867 size_t length)
4868{
4869 /*
4870 * Tracing a merged cache is going to give confusing results
4871 * as well as cause other issues like converting a mergeable
4872 * cache into an umergeable one.
4873 */
4874 if (s->refcount > 1)
4875 return -EINVAL;
4876
4877 s->flags &= ~SLAB_TRACE;
4878 if (buf[0] == '1') {
4879 s->flags &= ~__CMPXCHG_DOUBLE;
4880 s->flags |= SLAB_TRACE;
4881 }
4882 return length;
4883}
4884SLAB_ATTR(trace);
4885
4886static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4887{
4888 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4889}
4890
4891static ssize_t red_zone_store(struct kmem_cache *s,
4892 const char *buf, size_t length)
4893{
4894 if (any_slab_objects(s))
4895 return -EBUSY;
4896
4897 s->flags &= ~SLAB_RED_ZONE;
4898 if (buf[0] == '1') {
4899 s->flags |= SLAB_RED_ZONE;
4900 }
4901 calculate_sizes(s, -1);
4902 return length;
4903}
4904SLAB_ATTR(red_zone);
4905
4906static ssize_t poison_show(struct kmem_cache *s, char *buf)
4907{
4908 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4909}
4910
4911static ssize_t poison_store(struct kmem_cache *s,
4912 const char *buf, size_t length)
4913{
4914 if (any_slab_objects(s))
4915 return -EBUSY;
4916
4917 s->flags &= ~SLAB_POISON;
4918 if (buf[0] == '1') {
4919 s->flags |= SLAB_POISON;
4920 }
4921 calculate_sizes(s, -1);
4922 return length;
4923}
4924SLAB_ATTR(poison);
4925
4926static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4927{
4928 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4929}
4930
4931static ssize_t store_user_store(struct kmem_cache *s,
4932 const char *buf, size_t length)
4933{
4934 if (any_slab_objects(s))
4935 return -EBUSY;
4936
4937 s->flags &= ~SLAB_STORE_USER;
4938 if (buf[0] == '1') {
4939 s->flags &= ~__CMPXCHG_DOUBLE;
4940 s->flags |= SLAB_STORE_USER;
4941 }
4942 calculate_sizes(s, -1);
4943 return length;
4944}
4945SLAB_ATTR(store_user);
4946
4947static ssize_t validate_show(struct kmem_cache *s, char *buf)
4948{
4949 return 0;
4950}
4951
4952static ssize_t validate_store(struct kmem_cache *s,
4953 const char *buf, size_t length)
4954{
4955 int ret = -EINVAL;
4956
4957 if (buf[0] == '1') {
4958 ret = validate_slab_cache(s);
4959 if (ret >= 0)
4960 ret = length;
4961 }
4962 return ret;
4963}
4964SLAB_ATTR(validate);
4965
4966static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4967{
4968 if (!(s->flags & SLAB_STORE_USER))
4969 return -ENOSYS;
4970 return list_locations(s, buf, TRACK_ALLOC);
4971}
4972SLAB_ATTR_RO(alloc_calls);
4973
4974static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4975{
4976 if (!(s->flags & SLAB_STORE_USER))
4977 return -ENOSYS;
4978 return list_locations(s, buf, TRACK_FREE);
4979}
4980SLAB_ATTR_RO(free_calls);
4981#endif /* CONFIG_SLUB_DEBUG */
4982
4983#ifdef CONFIG_FAILSLAB
4984static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4985{
4986 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4987}
4988
4989static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4990 size_t length)
4991{
4992 if (s->refcount > 1)
4993 return -EINVAL;
4994
4995 s->flags &= ~SLAB_FAILSLAB;
4996 if (buf[0] == '1')
4997 s->flags |= SLAB_FAILSLAB;
4998 return length;
4999}
5000SLAB_ATTR(failslab);
5001#endif
5002
5003static ssize_t shrink_show(struct kmem_cache *s, char *buf)
5004{
5005 return 0;
5006}
5007
5008static ssize_t shrink_store(struct kmem_cache *s,
5009 const char *buf, size_t length)
5010{
5011 if (buf[0] == '1')
5012 kmem_cache_shrink(s);
5013 else
5014 return -EINVAL;
5015 return length;
5016}
5017SLAB_ATTR(shrink);
5018
5019#ifdef CONFIG_NUMA
5020static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
5021{
5022 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
5023}
5024
5025static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5026 const char *buf, size_t length)
5027{
5028 unsigned long ratio;
5029 int err;
5030
5031 err = kstrtoul(buf, 10, &ratio);
5032 if (err)
5033 return err;
5034
5035 if (ratio <= 100)
5036 s->remote_node_defrag_ratio = ratio * 10;
5037
5038 return length;
5039}
5040SLAB_ATTR(remote_node_defrag_ratio);
5041#endif
5042
5043#ifdef CONFIG_SLUB_STATS
5044static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5045{
5046 unsigned long sum = 0;
5047 int cpu;
5048 int len;
5049 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5050
5051 if (!data)
5052 return -ENOMEM;
5053
5054 for_each_online_cpu(cpu) {
5055 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5056
5057 data[cpu] = x;
5058 sum += x;
5059 }
5060
5061 len = sprintf(buf, "%lu", sum);
5062
5063#ifdef CONFIG_SMP
5064 for_each_online_cpu(cpu) {
5065 if (data[cpu] && len < PAGE_SIZE - 20)
5066 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5067 }
5068#endif
5069 kfree(data);
5070 return len + sprintf(buf + len, "\n");
5071}
5072
5073static void clear_stat(struct kmem_cache *s, enum stat_item si)
5074{
5075 int cpu;
5076
5077 for_each_online_cpu(cpu)
5078 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5079}
5080
5081#define STAT_ATTR(si, text) \
5082static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5083{ \
5084 return show_stat(s, buf, si); \
5085} \
5086static ssize_t text##_store(struct kmem_cache *s, \
5087 const char *buf, size_t length) \
5088{ \
5089 if (buf[0] != '0') \
5090 return -EINVAL; \
5091 clear_stat(s, si); \
5092 return length; \
5093} \
5094SLAB_ATTR(text); \
5095
5096STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5097STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5098STAT_ATTR(FREE_FASTPATH, free_fastpath);
5099STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5100STAT_ATTR(FREE_FROZEN, free_frozen);
5101STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5102STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5103STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5104STAT_ATTR(ALLOC_SLAB, alloc_slab);
5105STAT_ATTR(ALLOC_REFILL, alloc_refill);
5106STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5107STAT_ATTR(FREE_SLAB, free_slab);
5108STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5109STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5110STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5111STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5112STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5113STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5114STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5115STAT_ATTR(ORDER_FALLBACK, order_fallback);
5116STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5117STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5118STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5119STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5120STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5121STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5122#endif
5123
5124static struct attribute *slab_attrs[] = {
5125 &slab_size_attr.attr,
5126 &object_size_attr.attr,
5127 &objs_per_slab_attr.attr,
5128 &order_attr.attr,
5129 &min_partial_attr.attr,
5130 &cpu_partial_attr.attr,
5131 &objects_attr.attr,
5132 &objects_partial_attr.attr,
5133 &partial_attr.attr,
5134 &cpu_slabs_attr.attr,
5135 &ctor_attr.attr,
5136 &aliases_attr.attr,
5137 &align_attr.attr,
5138 &hwcache_align_attr.attr,
5139 &reclaim_account_attr.attr,
5140 &destroy_by_rcu_attr.attr,
5141 &shrink_attr.attr,
5142 &reserved_attr.attr,
5143 &slabs_cpu_partial_attr.attr,
5144#ifdef CONFIG_SLUB_DEBUG
5145 &total_objects_attr.attr,
5146 &slabs_attr.attr,
5147 &sanity_checks_attr.attr,
5148 &trace_attr.attr,
5149 &red_zone_attr.attr,
5150 &poison_attr.attr,
5151 &store_user_attr.attr,
5152 &validate_attr.attr,
5153 &alloc_calls_attr.attr,
5154 &free_calls_attr.attr,
5155#endif
5156#ifdef CONFIG_ZONE_DMA
5157 &cache_dma_attr.attr,
5158#endif
5159#ifdef CONFIG_NUMA
5160 &remote_node_defrag_ratio_attr.attr,
5161#endif
5162#ifdef CONFIG_SLUB_STATS
5163 &alloc_fastpath_attr.attr,
5164 &alloc_slowpath_attr.attr,
5165 &free_fastpath_attr.attr,
5166 &free_slowpath_attr.attr,
5167 &free_frozen_attr.attr,
5168 &free_add_partial_attr.attr,
5169 &free_remove_partial_attr.attr,
5170 &alloc_from_partial_attr.attr,
5171 &alloc_slab_attr.attr,
5172 &alloc_refill_attr.attr,
5173 &alloc_node_mismatch_attr.attr,
5174 &free_slab_attr.attr,
5175 &cpuslab_flush_attr.attr,
5176 &deactivate_full_attr.attr,
5177 &deactivate_empty_attr.attr,
5178 &deactivate_to_head_attr.attr,
5179 &deactivate_to_tail_attr.attr,
5180 &deactivate_remote_frees_attr.attr,
5181 &deactivate_bypass_attr.attr,
5182 &order_fallback_attr.attr,
5183 &cmpxchg_double_fail_attr.attr,
5184 &cmpxchg_double_cpu_fail_attr.attr,
5185 &cpu_partial_alloc_attr.attr,
5186 &cpu_partial_free_attr.attr,
5187 &cpu_partial_node_attr.attr,
5188 &cpu_partial_drain_attr.attr,
5189#endif
5190#ifdef CONFIG_FAILSLAB
5191 &failslab_attr.attr,
5192#endif
5193
5194 NULL
5195};
5196
5197static struct attribute_group slab_attr_group = {
5198 .attrs = slab_attrs,
5199};
5200
5201static ssize_t slab_attr_show(struct kobject *kobj,
5202 struct attribute *attr,
5203 char *buf)
5204{
5205 struct slab_attribute *attribute;
5206 struct kmem_cache *s;
5207 int err;
5208
5209 attribute = to_slab_attr(attr);
5210 s = to_slab(kobj);
5211
5212 if (!attribute->show)
5213 return -EIO;
5214
5215 err = attribute->show(s, buf);
5216
5217 return err;
5218}
5219
5220static ssize_t slab_attr_store(struct kobject *kobj,
5221 struct attribute *attr,
5222 const char *buf, size_t len)
5223{
5224 struct slab_attribute *attribute;
5225 struct kmem_cache *s;
5226 int err;
5227
5228 attribute = to_slab_attr(attr);
5229 s = to_slab(kobj);
5230
5231 if (!attribute->store)
5232 return -EIO;
5233
5234 err = attribute->store(s, buf, len);
5235#ifdef CONFIG_MEMCG
5236 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5237 struct kmem_cache *c;
5238
5239 mutex_lock(&slab_mutex);
5240 if (s->max_attr_size < len)
5241 s->max_attr_size = len;
5242
5243 /*
5244 * This is a best effort propagation, so this function's return
5245 * value will be determined by the parent cache only. This is
5246 * basically because not all attributes will have a well
5247 * defined semantics for rollbacks - most of the actions will
5248 * have permanent effects.
5249 *
5250 * Returning the error value of any of the children that fail
5251 * is not 100 % defined, in the sense that users seeing the
5252 * error code won't be able to know anything about the state of
5253 * the cache.
5254 *
5255 * Only returning the error code for the parent cache at least
5256 * has well defined semantics. The cache being written to
5257 * directly either failed or succeeded, in which case we loop
5258 * through the descendants with best-effort propagation.
5259 */
5260 for_each_memcg_cache(c, s)
5261 attribute->store(c, buf, len);
5262 mutex_unlock(&slab_mutex);
5263 }
5264#endif
5265 return err;
5266}
5267
5268static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5269{
5270#ifdef CONFIG_MEMCG
5271 int i;
5272 char *buffer = NULL;
5273 struct kmem_cache *root_cache;
5274
5275 if (is_root_cache(s))
5276 return;
5277
5278 root_cache = s->memcg_params.root_cache;
5279
5280 /*
5281 * This mean this cache had no attribute written. Therefore, no point
5282 * in copying default values around
5283 */
5284 if (!root_cache->max_attr_size)
5285 return;
5286
5287 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5288 char mbuf[64];
5289 char *buf;
5290 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5291
5292 if (!attr || !attr->store || !attr->show)
5293 continue;
5294
5295 /*
5296 * It is really bad that we have to allocate here, so we will
5297 * do it only as a fallback. If we actually allocate, though,
5298 * we can just use the allocated buffer until the end.
5299 *
5300 * Most of the slub attributes will tend to be very small in
5301 * size, but sysfs allows buffers up to a page, so they can
5302 * theoretically happen.
5303 */
5304 if (buffer)
5305 buf = buffer;
5306 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5307 buf = mbuf;
5308 else {
5309 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5310 if (WARN_ON(!buffer))
5311 continue;
5312 buf = buffer;
5313 }
5314
5315 attr->show(root_cache, buf);
5316 attr->store(s, buf, strlen(buf));
5317 }
5318
5319 if (buffer)
5320 free_page((unsigned long)buffer);
5321#endif
5322}
5323
5324static void kmem_cache_release(struct kobject *k)
5325{
5326 slab_kmem_cache_release(to_slab(k));
5327}
5328
5329static const struct sysfs_ops slab_sysfs_ops = {
5330 .show = slab_attr_show,
5331 .store = slab_attr_store,
5332};
5333
5334static struct kobj_type slab_ktype = {
5335 .sysfs_ops = &slab_sysfs_ops,
5336 .release = kmem_cache_release,
5337};
5338
5339static int uevent_filter(struct kset *kset, struct kobject *kobj)
5340{
5341 struct kobj_type *ktype = get_ktype(kobj);
5342
5343 if (ktype == &slab_ktype)
5344 return 1;
5345 return 0;
5346}
5347
5348static const struct kset_uevent_ops slab_uevent_ops = {
5349 .filter = uevent_filter,
5350};
5351
5352static struct kset *slab_kset;
5353
5354static inline struct kset *cache_kset(struct kmem_cache *s)
5355{
5356#ifdef CONFIG_MEMCG
5357 if (!is_root_cache(s))
5358 return s->memcg_params.root_cache->memcg_kset;
5359#endif
5360 return slab_kset;
5361}
5362
5363#define ID_STR_LENGTH 64
5364
5365/* Create a unique string id for a slab cache:
5366 *
5367 * Format :[flags-]size
5368 */
5369static char *create_unique_id(struct kmem_cache *s)
5370{
5371 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5372 char *p = name;
5373
5374 BUG_ON(!name);
5375
5376 *p++ = ':';
5377 /*
5378 * First flags affecting slabcache operations. We will only
5379 * get here for aliasable slabs so we do not need to support
5380 * too many flags. The flags here must cover all flags that
5381 * are matched during merging to guarantee that the id is
5382 * unique.
5383 */
5384 if (s->flags & SLAB_CACHE_DMA)
5385 *p++ = 'd';
5386 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5387 *p++ = 'a';
5388 if (s->flags & SLAB_CONSISTENCY_CHECKS)
5389 *p++ = 'F';
5390 if (!(s->flags & SLAB_NOTRACK))
5391 *p++ = 't';
5392 if (s->flags & SLAB_ACCOUNT)
5393 *p++ = 'A';
5394 if (p != name + 1)
5395 *p++ = '-';
5396 p += sprintf(p, "%07d", s->size);
5397
5398 BUG_ON(p > name + ID_STR_LENGTH - 1);
5399 return name;
5400}
5401
5402static int sysfs_slab_add(struct kmem_cache *s)
5403{
5404 int err;
5405 const char *name;
5406 int unmergeable = slab_unmergeable(s);
5407
5408 if (unmergeable) {
5409 /*
5410 * Slabcache can never be merged so we can use the name proper.
5411 * This is typically the case for debug situations. In that
5412 * case we can catch duplicate names easily.
5413 */
5414 sysfs_remove_link(&slab_kset->kobj, s->name);
5415 name = s->name;
5416 } else {
5417 /*
5418 * Create a unique name for the slab as a target
5419 * for the symlinks.
5420 */
5421 name = create_unique_id(s);
5422 }
5423
5424 s->kobj.kset = cache_kset(s);
5425 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5426 if (err)
5427 goto out;
5428
5429 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5430 if (err)
5431 goto out_del_kobj;
5432
5433#ifdef CONFIG_MEMCG
5434 if (is_root_cache(s)) {
5435 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5436 if (!s->memcg_kset) {
5437 err = -ENOMEM;
5438 goto out_del_kobj;
5439 }
5440 }
5441#endif
5442
5443 kobject_uevent(&s->kobj, KOBJ_ADD);
5444 if (!unmergeable) {
5445 /* Setup first alias */
5446 sysfs_slab_alias(s, s->name);
5447 }
5448out:
5449 if (!unmergeable)
5450 kfree(name);
5451 return err;
5452out_del_kobj:
5453 kobject_del(&s->kobj);
5454 goto out;
5455}
5456
5457void sysfs_slab_remove(struct kmem_cache *s)
5458{
5459 if (slab_state < FULL)
5460 /*
5461 * Sysfs has not been setup yet so no need to remove the
5462 * cache from sysfs.
5463 */
5464 return;
5465
5466#ifdef CONFIG_MEMCG
5467 kset_unregister(s->memcg_kset);
5468#endif
5469 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5470 kobject_del(&s->kobj);
5471 kobject_put(&s->kobj);
5472}
5473
5474/*
5475 * Need to buffer aliases during bootup until sysfs becomes
5476 * available lest we lose that information.
5477 */
5478struct saved_alias {
5479 struct kmem_cache *s;
5480 const char *name;
5481 struct saved_alias *next;
5482};
5483
5484static struct saved_alias *alias_list;
5485
5486static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5487{
5488 struct saved_alias *al;
5489
5490 if (slab_state == FULL) {
5491 /*
5492 * If we have a leftover link then remove it.
5493 */
5494 sysfs_remove_link(&slab_kset->kobj, name);
5495 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5496 }
5497
5498 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5499 if (!al)
5500 return -ENOMEM;
5501
5502 al->s = s;
5503 al->name = name;
5504 al->next = alias_list;
5505 alias_list = al;
5506 return 0;
5507}
5508
5509static int __init slab_sysfs_init(void)
5510{
5511 struct kmem_cache *s;
5512 int err;
5513
5514 mutex_lock(&slab_mutex);
5515
5516 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5517 if (!slab_kset) {
5518 mutex_unlock(&slab_mutex);
5519 pr_err("Cannot register slab subsystem.\n");
5520 return -ENOSYS;
5521 }
5522
5523 slab_state = FULL;
5524
5525 list_for_each_entry(s, &slab_caches, list) {
5526 err = sysfs_slab_add(s);
5527 if (err)
5528 pr_err("SLUB: Unable to add boot slab %s to sysfs\n",
5529 s->name);
5530 }
5531
5532 while (alias_list) {
5533 struct saved_alias *al = alias_list;
5534
5535 alias_list = alias_list->next;
5536 err = sysfs_slab_alias(al->s, al->name);
5537 if (err)
5538 pr_err("SLUB: Unable to add boot slab alias %s to sysfs\n",
5539 al->name);
5540 kfree(al);
5541 }
5542
5543 mutex_unlock(&slab_mutex);
5544 resiliency_test();
5545 return 0;
5546}
5547
5548__initcall(slab_sysfs_init);
5549#endif /* CONFIG_SYSFS */
5550
5551/*
5552 * The /proc/slabinfo ABI
5553 */
5554#ifdef CONFIG_SLABINFO
5555void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5556{
5557 unsigned long nr_slabs = 0;
5558 unsigned long nr_objs = 0;
5559 unsigned long nr_free = 0;
5560 int node;
5561 struct kmem_cache_node *n;
5562
5563 for_each_kmem_cache_node(s, node, n) {
5564 nr_slabs += node_nr_slabs(n);
5565 nr_objs += node_nr_objs(n);
5566 nr_free += count_partial(n, count_free);
5567 }
5568
5569 sinfo->active_objs = nr_objs - nr_free;
5570 sinfo->num_objs = nr_objs;
5571 sinfo->active_slabs = nr_slabs;
5572 sinfo->num_slabs = nr_slabs;
5573 sinfo->objects_per_slab = oo_objects(s->oo);
5574 sinfo->cache_order = oo_order(s->oo);
5575}
5576
5577void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5578{
5579}
5580
5581ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5582 size_t count, loff_t *ppos)
5583{
5584 return -EIO;
5585}
5586#endif /* CONFIG_SLABINFO */
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 */