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/*
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 <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/kmemcheck.h>
22#include <linux/cpu.h>
23#include <linux/cpuset.h>
24#include <linux/mempolicy.h>
25#include <linux/ctype.h>
26#include <linux/debugobjects.h>
27#include <linux/kallsyms.h>
28#include <linux/memory.h>
29#include <linux/math64.h>
30#include <linux/fault-inject.h>
31#include <linux/stacktrace.h>
32
33#include <trace/events/kmem.h>
34
35/*
36 * Lock order:
37 * 1. slub_lock (Global Semaphore)
38 * 2. node->list_lock
39 * 3. slab_lock(page) (Only on some arches and for debugging)
40 *
41 * slub_lock
42 *
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
45 *
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
52 *
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
58 *
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
64 *
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
69 * the list lock.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
74 *
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
77 *
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
83 *
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
87 *
88 * Overloading of page flags that are otherwise used for LRU management.
89 *
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
98 *
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
105 *
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
109 */
110
111#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
113
114static inline int kmem_cache_debug(struct kmem_cache *s)
115{
116#ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
118#else
119 return 0;
120#endif
121}
122
123/*
124 * Issues still to be resolved:
125 *
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
127 *
128 * - Variable sizing of the per node arrays
129 */
130
131/* Enable to test recovery from slab corruption on boot */
132#undef SLUB_RESILIENCY_TEST
133
134/* Enable to log cmpxchg failures */
135#undef SLUB_DEBUG_CMPXCHG
136
137/*
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
140 */
141#define MIN_PARTIAL 5
142
143/*
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
147 */
148#define MAX_PARTIAL 10
149
150#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
152
153/*
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
156 * metadata.
157 */
158#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
159
160/*
161 * Set of flags that will prevent slab merging
162 */
163#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
165 SLAB_FAILSLAB)
166
167#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
169
170#define OO_SHIFT 16
171#define OO_MASK ((1 << OO_SHIFT) - 1)
172#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
173
174/* Internal SLUB flags */
175#define __OBJECT_POISON 0x80000000UL /* Poison object */
176#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
177
178static int kmem_size = sizeof(struct kmem_cache);
179
180#ifdef CONFIG_SMP
181static struct notifier_block slab_notifier;
182#endif
183
184static enum {
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
188 SYSFS /* Sysfs up */
189} slab_state = DOWN;
190
191/* A list of all slab caches on the system */
192static DECLARE_RWSEM(slub_lock);
193static LIST_HEAD(slab_caches);
194
195/*
196 * Tracking user of a slab.
197 */
198#define TRACK_ADDRS_COUNT 16
199struct track {
200 unsigned long addr; /* Called from address */
201#ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203#endif
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
207};
208
209enum track_item { TRACK_ALLOC, TRACK_FREE };
210
211#ifdef CONFIG_SYSFS
212static int sysfs_slab_add(struct kmem_cache *);
213static int sysfs_slab_alias(struct kmem_cache *, const char *);
214static void sysfs_slab_remove(struct kmem_cache *);
215
216#else
217static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
219 { return 0; }
220static inline void sysfs_slab_remove(struct kmem_cache *s)
221{
222 kfree(s->name);
223 kfree(s);
224}
225
226#endif
227
228static inline void stat(const struct kmem_cache *s, enum stat_item si)
229{
230#ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
232#endif
233}
234
235/********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
238
239int slab_is_available(void)
240{
241 return slab_state >= UP;
242}
243
244static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
245{
246 return s->node[node];
247}
248
249/* Verify that a pointer has an address that is valid within a slab page */
250static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
252{
253 void *base;
254
255 if (!object)
256 return 1;
257
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
261 return 0;
262 }
263
264 return 1;
265}
266
267static inline void *get_freepointer(struct kmem_cache *s, void *object)
268{
269 return *(void **)(object + s->offset);
270}
271
272static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
273{
274 void *p;
275
276#ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
278#else
279 p = get_freepointer(s, object);
280#endif
281 return p;
282}
283
284static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
285{
286 *(void **)(object + s->offset) = fp;
287}
288
289/* Loop over all objects in a slab */
290#define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
292 __p += (__s)->size)
293
294/* Determine object index from a given position */
295static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
296{
297 return (p - addr) / s->size;
298}
299
300static inline size_t slab_ksize(const struct kmem_cache *s)
301{
302#ifdef CONFIG_SLUB_DEBUG
303 /*
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
306 */
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 return s->objsize;
309
310#endif
311 /*
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
315 */
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
317 return s->inuse;
318 /*
319 * Else we can use all the padding etc for the allocation
320 */
321 return s->size;
322}
323
324static inline int order_objects(int order, unsigned long size, int reserved)
325{
326 return ((PAGE_SIZE << order) - reserved) / size;
327}
328
329static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
331{
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
334 };
335
336 return x;
337}
338
339static inline int oo_order(struct kmem_cache_order_objects x)
340{
341 return x.x >> OO_SHIFT;
342}
343
344static inline int oo_objects(struct kmem_cache_order_objects x)
345{
346 return x.x & OO_MASK;
347}
348
349/*
350 * Per slab locking using the pagelock
351 */
352static __always_inline void slab_lock(struct page *page)
353{
354 bit_spin_lock(PG_locked, &page->flags);
355}
356
357static __always_inline void slab_unlock(struct page *page)
358{
359 __bit_spin_unlock(PG_locked, &page->flags);
360}
361
362/* Interrupts must be disabled (for the fallback code to work right) */
363static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
366 const char *n)
367{
368 VM_BUG_ON(!irqs_disabled());
369#ifdef CONFIG_CMPXCHG_DOUBLE
370 if (s->flags & __CMPXCHG_DOUBLE) {
371 if (cmpxchg_double(&page->freelist,
372 freelist_old, counters_old,
373 freelist_new, counters_new))
374 return 1;
375 } else
376#endif
377 {
378 slab_lock(page);
379 if (page->freelist == freelist_old && page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
382 slab_unlock(page);
383 return 1;
384 }
385 slab_unlock(page);
386 }
387
388 cpu_relax();
389 stat(s, CMPXCHG_DOUBLE_FAIL);
390
391#ifdef SLUB_DEBUG_CMPXCHG
392 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
393#endif
394
395 return 0;
396}
397
398static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
401 const char *n)
402{
403#ifdef CONFIG_CMPXCHG_DOUBLE
404 if (s->flags & __CMPXCHG_DOUBLE) {
405 if (cmpxchg_double(&page->freelist,
406 freelist_old, counters_old,
407 freelist_new, counters_new))
408 return 1;
409 } else
410#endif
411 {
412 unsigned long flags;
413
414 local_irq_save(flags);
415 slab_lock(page);
416 if (page->freelist == freelist_old && page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
419 slab_unlock(page);
420 local_irq_restore(flags);
421 return 1;
422 }
423 slab_unlock(page);
424 local_irq_restore(flags);
425 }
426
427 cpu_relax();
428 stat(s, CMPXCHG_DOUBLE_FAIL);
429
430#ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432#endif
433
434 return 0;
435}
436
437#ifdef CONFIG_SLUB_DEBUG
438/*
439 * Determine a map of object in use on a page.
440 *
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
443 */
444static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
445{
446 void *p;
447 void *addr = page_address(page);
448
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
451}
452
453/*
454 * Debug settings:
455 */
456#ifdef CONFIG_SLUB_DEBUG_ON
457static int slub_debug = DEBUG_DEFAULT_FLAGS;
458#else
459static int slub_debug;
460#endif
461
462static char *slub_debug_slabs;
463static int disable_higher_order_debug;
464
465/*
466 * Object debugging
467 */
468static void print_section(char *text, u8 *addr, unsigned int length)
469{
470 int i, offset;
471 int newline = 1;
472 char ascii[17];
473
474 ascii[16] = 0;
475
476 for (i = 0; i < length; i++) {
477 if (newline) {
478 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
479 newline = 0;
480 }
481 printk(KERN_CONT " %02x", addr[i]);
482 offset = i % 16;
483 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
484 if (offset == 15) {
485 printk(KERN_CONT " %s\n", ascii);
486 newline = 1;
487 }
488 }
489 if (!newline) {
490 i %= 16;
491 while (i < 16) {
492 printk(KERN_CONT " ");
493 ascii[i] = ' ';
494 i++;
495 }
496 printk(KERN_CONT " %s\n", ascii);
497 }
498}
499
500static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
502{
503 struct track *p;
504
505 if (s->offset)
506 p = object + s->offset + sizeof(void *);
507 else
508 p = object + s->inuse;
509
510 return p + alloc;
511}
512
513static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
515{
516 struct track *p = get_track(s, object, alloc);
517
518 if (addr) {
519#ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
521 int i;
522
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
526 trace.skip = 3;
527 save_stack_trace(&trace);
528
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
532 trace.nr_entries--;
533
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
535 p->addrs[i] = 0;
536#endif
537 p->addr = addr;
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
540 p->when = jiffies;
541 } else
542 memset(p, 0, sizeof(struct track));
543}
544
545static void init_tracking(struct kmem_cache *s, void *object)
546{
547 if (!(s->flags & SLAB_STORE_USER))
548 return;
549
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
552}
553
554static void print_track(const char *s, struct track *t)
555{
556 if (!t->addr)
557 return;
558
559 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561#ifdef CONFIG_STACKTRACE
562 {
563 int i;
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
565 if (t->addrs[i])
566 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
567 else
568 break;
569 }
570#endif
571}
572
573static void print_tracking(struct kmem_cache *s, void *object)
574{
575 if (!(s->flags & SLAB_STORE_USER))
576 return;
577
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
580}
581
582static void print_page_info(struct page *page)
583{
584 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
586
587}
588
589static void slab_bug(struct kmem_cache *s, char *fmt, ...)
590{
591 va_list args;
592 char buf[100];
593
594 va_start(args, fmt);
595 vsnprintf(buf, sizeof(buf), fmt, args);
596 va_end(args);
597 printk(KERN_ERR "========================================"
598 "=====================================\n");
599 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
600 printk(KERN_ERR "----------------------------------------"
601 "-------------------------------------\n\n");
602}
603
604static void slab_fix(struct kmem_cache *s, char *fmt, ...)
605{
606 va_list args;
607 char buf[100];
608
609 va_start(args, fmt);
610 vsnprintf(buf, sizeof(buf), fmt, args);
611 va_end(args);
612 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
613}
614
615static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
616{
617 unsigned int off; /* Offset of last byte */
618 u8 *addr = page_address(page);
619
620 print_tracking(s, p);
621
622 print_page_info(page);
623
624 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
625 p, p - addr, get_freepointer(s, p));
626
627 if (p > addr + 16)
628 print_section("Bytes b4", p - 16, 16);
629
630 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
631
632 if (s->flags & SLAB_RED_ZONE)
633 print_section("Redzone", p + s->objsize,
634 s->inuse - s->objsize);
635
636 if (s->offset)
637 off = s->offset + sizeof(void *);
638 else
639 off = s->inuse;
640
641 if (s->flags & SLAB_STORE_USER)
642 off += 2 * sizeof(struct track);
643
644 if (off != s->size)
645 /* Beginning of the filler is the free pointer */
646 print_section("Padding", p + off, s->size - off);
647
648 dump_stack();
649}
650
651static void object_err(struct kmem_cache *s, struct page *page,
652 u8 *object, char *reason)
653{
654 slab_bug(s, "%s", reason);
655 print_trailer(s, page, object);
656}
657
658static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
659{
660 va_list args;
661 char buf[100];
662
663 va_start(args, fmt);
664 vsnprintf(buf, sizeof(buf), fmt, args);
665 va_end(args);
666 slab_bug(s, "%s", buf);
667 print_page_info(page);
668 dump_stack();
669}
670
671static void init_object(struct kmem_cache *s, void *object, u8 val)
672{
673 u8 *p = object;
674
675 if (s->flags & __OBJECT_POISON) {
676 memset(p, POISON_FREE, s->objsize - 1);
677 p[s->objsize - 1] = POISON_END;
678 }
679
680 if (s->flags & SLAB_RED_ZONE)
681 memset(p + s->objsize, val, s->inuse - s->objsize);
682}
683
684static u8 *check_bytes8(u8 *start, u8 value, unsigned int bytes)
685{
686 while (bytes) {
687 if (*start != value)
688 return start;
689 start++;
690 bytes--;
691 }
692 return NULL;
693}
694
695static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
696{
697 u64 value64;
698 unsigned int words, prefix;
699
700 if (bytes <= 16)
701 return check_bytes8(start, value, bytes);
702
703 value64 = value | value << 8 | value << 16 | value << 24;
704 value64 = (value64 & 0xffffffff) | value64 << 32;
705 prefix = 8 - ((unsigned long)start) % 8;
706
707 if (prefix) {
708 u8 *r = check_bytes8(start, value, prefix);
709 if (r)
710 return r;
711 start += prefix;
712 bytes -= prefix;
713 }
714
715 words = bytes / 8;
716
717 while (words) {
718 if (*(u64 *)start != value64)
719 return check_bytes8(start, value, 8);
720 start += 8;
721 words--;
722 }
723
724 return check_bytes8(start, value, bytes % 8);
725}
726
727static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
728 void *from, void *to)
729{
730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
731 memset(from, data, to - from);
732}
733
734static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
735 u8 *object, char *what,
736 u8 *start, unsigned int value, unsigned int bytes)
737{
738 u8 *fault;
739 u8 *end;
740
741 fault = check_bytes(start, value, bytes);
742 if (!fault)
743 return 1;
744
745 end = start + bytes;
746 while (end > fault && end[-1] == value)
747 end--;
748
749 slab_bug(s, "%s overwritten", what);
750 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
751 fault, end - 1, fault[0], value);
752 print_trailer(s, page, object);
753
754 restore_bytes(s, what, value, fault, end);
755 return 0;
756}
757
758/*
759 * Object layout:
760 *
761 * object address
762 * Bytes of the object to be managed.
763 * If the freepointer may overlay the object then the free
764 * pointer is the first word of the object.
765 *
766 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
767 * 0xa5 (POISON_END)
768 *
769 * object + s->objsize
770 * Padding to reach word boundary. This is also used for Redzoning.
771 * Padding is extended by another word if Redzoning is enabled and
772 * objsize == inuse.
773 *
774 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
775 * 0xcc (RED_ACTIVE) for objects in use.
776 *
777 * object + s->inuse
778 * Meta data starts here.
779 *
780 * A. Free pointer (if we cannot overwrite object on free)
781 * B. Tracking data for SLAB_STORE_USER
782 * C. Padding to reach required alignment boundary or at mininum
783 * one word if debugging is on to be able to detect writes
784 * before the word boundary.
785 *
786 * Padding is done using 0x5a (POISON_INUSE)
787 *
788 * object + s->size
789 * Nothing is used beyond s->size.
790 *
791 * If slabcaches are merged then the objsize and inuse boundaries are mostly
792 * ignored. And therefore no slab options that rely on these boundaries
793 * may be used with merged slabcaches.
794 */
795
796static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
797{
798 unsigned long off = s->inuse; /* The end of info */
799
800 if (s->offset)
801 /* Freepointer is placed after the object. */
802 off += sizeof(void *);
803
804 if (s->flags & SLAB_STORE_USER)
805 /* We also have user information there */
806 off += 2 * sizeof(struct track);
807
808 if (s->size == off)
809 return 1;
810
811 return check_bytes_and_report(s, page, p, "Object padding",
812 p + off, POISON_INUSE, s->size - off);
813}
814
815/* Check the pad bytes at the end of a slab page */
816static int slab_pad_check(struct kmem_cache *s, struct page *page)
817{
818 u8 *start;
819 u8 *fault;
820 u8 *end;
821 int length;
822 int remainder;
823
824 if (!(s->flags & SLAB_POISON))
825 return 1;
826
827 start = page_address(page);
828 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
829 end = start + length;
830 remainder = length % s->size;
831 if (!remainder)
832 return 1;
833
834 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
835 if (!fault)
836 return 1;
837 while (end > fault && end[-1] == POISON_INUSE)
838 end--;
839
840 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
841 print_section("Padding", end - remainder, remainder);
842
843 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
844 return 0;
845}
846
847static int check_object(struct kmem_cache *s, struct page *page,
848 void *object, u8 val)
849{
850 u8 *p = object;
851 u8 *endobject = object + s->objsize;
852
853 if (s->flags & SLAB_RED_ZONE) {
854 if (!check_bytes_and_report(s, page, object, "Redzone",
855 endobject, val, s->inuse - s->objsize))
856 return 0;
857 } else {
858 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
859 check_bytes_and_report(s, page, p, "Alignment padding",
860 endobject, POISON_INUSE, s->inuse - s->objsize);
861 }
862 }
863
864 if (s->flags & SLAB_POISON) {
865 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
866 (!check_bytes_and_report(s, page, p, "Poison", p,
867 POISON_FREE, s->objsize - 1) ||
868 !check_bytes_and_report(s, page, p, "Poison",
869 p + s->objsize - 1, POISON_END, 1)))
870 return 0;
871 /*
872 * check_pad_bytes cleans up on its own.
873 */
874 check_pad_bytes(s, page, p);
875 }
876
877 if (!s->offset && val == SLUB_RED_ACTIVE)
878 /*
879 * Object and freepointer overlap. Cannot check
880 * freepointer while object is allocated.
881 */
882 return 1;
883
884 /* Check free pointer validity */
885 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
886 object_err(s, page, p, "Freepointer corrupt");
887 /*
888 * No choice but to zap it and thus lose the remainder
889 * of the free objects in this slab. May cause
890 * another error because the object count is now wrong.
891 */
892 set_freepointer(s, p, NULL);
893 return 0;
894 }
895 return 1;
896}
897
898static int check_slab(struct kmem_cache *s, struct page *page)
899{
900 int maxobj;
901
902 VM_BUG_ON(!irqs_disabled());
903
904 if (!PageSlab(page)) {
905 slab_err(s, page, "Not a valid slab page");
906 return 0;
907 }
908
909 maxobj = order_objects(compound_order(page), s->size, s->reserved);
910 if (page->objects > maxobj) {
911 slab_err(s, page, "objects %u > max %u",
912 s->name, page->objects, maxobj);
913 return 0;
914 }
915 if (page->inuse > page->objects) {
916 slab_err(s, page, "inuse %u > max %u",
917 s->name, page->inuse, page->objects);
918 return 0;
919 }
920 /* Slab_pad_check fixes things up after itself */
921 slab_pad_check(s, page);
922 return 1;
923}
924
925/*
926 * Determine if a certain object on a page is on the freelist. Must hold the
927 * slab lock to guarantee that the chains are in a consistent state.
928 */
929static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
930{
931 int nr = 0;
932 void *fp;
933 void *object = NULL;
934 unsigned long max_objects;
935
936 fp = page->freelist;
937 while (fp && nr <= page->objects) {
938 if (fp == search)
939 return 1;
940 if (!check_valid_pointer(s, page, fp)) {
941 if (object) {
942 object_err(s, page, object,
943 "Freechain corrupt");
944 set_freepointer(s, object, NULL);
945 break;
946 } else {
947 slab_err(s, page, "Freepointer corrupt");
948 page->freelist = NULL;
949 page->inuse = page->objects;
950 slab_fix(s, "Freelist cleared");
951 return 0;
952 }
953 break;
954 }
955 object = fp;
956 fp = get_freepointer(s, object);
957 nr++;
958 }
959
960 max_objects = order_objects(compound_order(page), s->size, s->reserved);
961 if (max_objects > MAX_OBJS_PER_PAGE)
962 max_objects = MAX_OBJS_PER_PAGE;
963
964 if (page->objects != max_objects) {
965 slab_err(s, page, "Wrong number of objects. Found %d but "
966 "should be %d", page->objects, max_objects);
967 page->objects = max_objects;
968 slab_fix(s, "Number of objects adjusted.");
969 }
970 if (page->inuse != page->objects - nr) {
971 slab_err(s, page, "Wrong object count. Counter is %d but "
972 "counted were %d", page->inuse, page->objects - nr);
973 page->inuse = page->objects - nr;
974 slab_fix(s, "Object count adjusted.");
975 }
976 return search == NULL;
977}
978
979static void trace(struct kmem_cache *s, struct page *page, void *object,
980 int alloc)
981{
982 if (s->flags & SLAB_TRACE) {
983 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
984 s->name,
985 alloc ? "alloc" : "free",
986 object, page->inuse,
987 page->freelist);
988
989 if (!alloc)
990 print_section("Object", (void *)object, s->objsize);
991
992 dump_stack();
993 }
994}
995
996/*
997 * Hooks for other subsystems that check memory allocations. In a typical
998 * production configuration these hooks all should produce no code at all.
999 */
1000static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1001{
1002 flags &= gfp_allowed_mask;
1003 lockdep_trace_alloc(flags);
1004 might_sleep_if(flags & __GFP_WAIT);
1005
1006 return should_failslab(s->objsize, flags, s->flags);
1007}
1008
1009static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
1010{
1011 flags &= gfp_allowed_mask;
1012 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1013 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
1014}
1015
1016static inline void slab_free_hook(struct kmem_cache *s, void *x)
1017{
1018 kmemleak_free_recursive(x, s->flags);
1019
1020 /*
1021 * Trouble is that we may no longer disable interupts in the fast path
1022 * So in order to make the debug calls that expect irqs to be
1023 * disabled we need to disable interrupts temporarily.
1024 */
1025#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1026 {
1027 unsigned long flags;
1028
1029 local_irq_save(flags);
1030 kmemcheck_slab_free(s, x, s->objsize);
1031 debug_check_no_locks_freed(x, s->objsize);
1032 local_irq_restore(flags);
1033 }
1034#endif
1035 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1036 debug_check_no_obj_freed(x, s->objsize);
1037}
1038
1039/*
1040 * Tracking of fully allocated slabs for debugging purposes.
1041 *
1042 * list_lock must be held.
1043 */
1044static void add_full(struct kmem_cache *s,
1045 struct kmem_cache_node *n, struct page *page)
1046{
1047 if (!(s->flags & SLAB_STORE_USER))
1048 return;
1049
1050 list_add(&page->lru, &n->full);
1051}
1052
1053/*
1054 * list_lock must be held.
1055 */
1056static void remove_full(struct kmem_cache *s, struct page *page)
1057{
1058 if (!(s->flags & SLAB_STORE_USER))
1059 return;
1060
1061 list_del(&page->lru);
1062}
1063
1064/* Tracking of the number of slabs for debugging purposes */
1065static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1066{
1067 struct kmem_cache_node *n = get_node(s, node);
1068
1069 return atomic_long_read(&n->nr_slabs);
1070}
1071
1072static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1073{
1074 return atomic_long_read(&n->nr_slabs);
1075}
1076
1077static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1078{
1079 struct kmem_cache_node *n = get_node(s, node);
1080
1081 /*
1082 * May be called early in order to allocate a slab for the
1083 * kmem_cache_node structure. Solve the chicken-egg
1084 * dilemma by deferring the increment of the count during
1085 * bootstrap (see early_kmem_cache_node_alloc).
1086 */
1087 if (n) {
1088 atomic_long_inc(&n->nr_slabs);
1089 atomic_long_add(objects, &n->total_objects);
1090 }
1091}
1092static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1093{
1094 struct kmem_cache_node *n = get_node(s, node);
1095
1096 atomic_long_dec(&n->nr_slabs);
1097 atomic_long_sub(objects, &n->total_objects);
1098}
1099
1100/* Object debug checks for alloc/free paths */
1101static void setup_object_debug(struct kmem_cache *s, struct page *page,
1102 void *object)
1103{
1104 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1105 return;
1106
1107 init_object(s, object, SLUB_RED_INACTIVE);
1108 init_tracking(s, object);
1109}
1110
1111static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1112 void *object, unsigned long addr)
1113{
1114 if (!check_slab(s, page))
1115 goto bad;
1116
1117 if (!check_valid_pointer(s, page, object)) {
1118 object_err(s, page, object, "Freelist Pointer check fails");
1119 goto bad;
1120 }
1121
1122 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1123 goto bad;
1124
1125 /* Success perform special debug activities for allocs */
1126 if (s->flags & SLAB_STORE_USER)
1127 set_track(s, object, TRACK_ALLOC, addr);
1128 trace(s, page, object, 1);
1129 init_object(s, object, SLUB_RED_ACTIVE);
1130 return 1;
1131
1132bad:
1133 if (PageSlab(page)) {
1134 /*
1135 * If this is a slab page then lets do the best we can
1136 * to avoid issues in the future. Marking all objects
1137 * as used avoids touching the remaining objects.
1138 */
1139 slab_fix(s, "Marking all objects used");
1140 page->inuse = page->objects;
1141 page->freelist = NULL;
1142 }
1143 return 0;
1144}
1145
1146static noinline int free_debug_processing(struct kmem_cache *s,
1147 struct page *page, void *object, unsigned long addr)
1148{
1149 unsigned long flags;
1150 int rc = 0;
1151
1152 local_irq_save(flags);
1153 slab_lock(page);
1154
1155 if (!check_slab(s, page))
1156 goto fail;
1157
1158 if (!check_valid_pointer(s, page, object)) {
1159 slab_err(s, page, "Invalid object pointer 0x%p", object);
1160 goto fail;
1161 }
1162
1163 if (on_freelist(s, page, object)) {
1164 object_err(s, page, object, "Object already free");
1165 goto fail;
1166 }
1167
1168 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1169 goto out;
1170
1171 if (unlikely(s != page->slab)) {
1172 if (!PageSlab(page)) {
1173 slab_err(s, page, "Attempt to free object(0x%p) "
1174 "outside of slab", object);
1175 } else if (!page->slab) {
1176 printk(KERN_ERR
1177 "SLUB <none>: no slab for object 0x%p.\n",
1178 object);
1179 dump_stack();
1180 } else
1181 object_err(s, page, object,
1182 "page slab pointer corrupt.");
1183 goto fail;
1184 }
1185
1186 if (s->flags & SLAB_STORE_USER)
1187 set_track(s, object, TRACK_FREE, addr);
1188 trace(s, page, object, 0);
1189 init_object(s, object, SLUB_RED_INACTIVE);
1190 rc = 1;
1191out:
1192 slab_unlock(page);
1193 local_irq_restore(flags);
1194 return rc;
1195
1196fail:
1197 slab_fix(s, "Object at 0x%p not freed", object);
1198 goto out;
1199}
1200
1201static int __init setup_slub_debug(char *str)
1202{
1203 slub_debug = DEBUG_DEFAULT_FLAGS;
1204 if (*str++ != '=' || !*str)
1205 /*
1206 * No options specified. Switch on full debugging.
1207 */
1208 goto out;
1209
1210 if (*str == ',')
1211 /*
1212 * No options but restriction on slabs. This means full
1213 * debugging for slabs matching a pattern.
1214 */
1215 goto check_slabs;
1216
1217 if (tolower(*str) == 'o') {
1218 /*
1219 * Avoid enabling debugging on caches if its minimum order
1220 * would increase as a result.
1221 */
1222 disable_higher_order_debug = 1;
1223 goto out;
1224 }
1225
1226 slub_debug = 0;
1227 if (*str == '-')
1228 /*
1229 * Switch off all debugging measures.
1230 */
1231 goto out;
1232
1233 /*
1234 * Determine which debug features should be switched on
1235 */
1236 for (; *str && *str != ','; str++) {
1237 switch (tolower(*str)) {
1238 case 'f':
1239 slub_debug |= SLAB_DEBUG_FREE;
1240 break;
1241 case 'z':
1242 slub_debug |= SLAB_RED_ZONE;
1243 break;
1244 case 'p':
1245 slub_debug |= SLAB_POISON;
1246 break;
1247 case 'u':
1248 slub_debug |= SLAB_STORE_USER;
1249 break;
1250 case 't':
1251 slub_debug |= SLAB_TRACE;
1252 break;
1253 case 'a':
1254 slub_debug |= SLAB_FAILSLAB;
1255 break;
1256 default:
1257 printk(KERN_ERR "slub_debug option '%c' "
1258 "unknown. skipped\n", *str);
1259 }
1260 }
1261
1262check_slabs:
1263 if (*str == ',')
1264 slub_debug_slabs = str + 1;
1265out:
1266 return 1;
1267}
1268
1269__setup("slub_debug", setup_slub_debug);
1270
1271static unsigned long kmem_cache_flags(unsigned long objsize,
1272 unsigned long flags, const char *name,
1273 void (*ctor)(void *))
1274{
1275 /*
1276 * Enable debugging if selected on the kernel commandline.
1277 */
1278 if (slub_debug && (!slub_debug_slabs ||
1279 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1280 flags |= slub_debug;
1281
1282 return flags;
1283}
1284#else
1285static inline void setup_object_debug(struct kmem_cache *s,
1286 struct page *page, void *object) {}
1287
1288static inline int alloc_debug_processing(struct kmem_cache *s,
1289 struct page *page, void *object, unsigned long addr) { return 0; }
1290
1291static inline int free_debug_processing(struct kmem_cache *s,
1292 struct page *page, void *object, unsigned long addr) { return 0; }
1293
1294static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1295 { return 1; }
1296static inline int check_object(struct kmem_cache *s, struct page *page,
1297 void *object, u8 val) { return 1; }
1298static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1299 struct page *page) {}
1300static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1301static inline unsigned long kmem_cache_flags(unsigned long objsize,
1302 unsigned long flags, const char *name,
1303 void (*ctor)(void *))
1304{
1305 return flags;
1306}
1307#define slub_debug 0
1308
1309#define disable_higher_order_debug 0
1310
1311static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1312 { return 0; }
1313static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1314 { return 0; }
1315static inline void inc_slabs_node(struct kmem_cache *s, int node,
1316 int objects) {}
1317static inline void dec_slabs_node(struct kmem_cache *s, int node,
1318 int objects) {}
1319
1320static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1321 { return 0; }
1322
1323static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1324 void *object) {}
1325
1326static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1327
1328#endif /* CONFIG_SLUB_DEBUG */
1329
1330/*
1331 * Slab allocation and freeing
1332 */
1333static inline struct page *alloc_slab_page(gfp_t flags, int node,
1334 struct kmem_cache_order_objects oo)
1335{
1336 int order = oo_order(oo);
1337
1338 flags |= __GFP_NOTRACK;
1339
1340 if (node == NUMA_NO_NODE)
1341 return alloc_pages(flags, order);
1342 else
1343 return alloc_pages_exact_node(node, flags, order);
1344}
1345
1346static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1347{
1348 struct page *page;
1349 struct kmem_cache_order_objects oo = s->oo;
1350 gfp_t alloc_gfp;
1351
1352 flags &= gfp_allowed_mask;
1353
1354 if (flags & __GFP_WAIT)
1355 local_irq_enable();
1356
1357 flags |= s->allocflags;
1358
1359 /*
1360 * Let the initial higher-order allocation fail under memory pressure
1361 * so we fall-back to the minimum order allocation.
1362 */
1363 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1364
1365 page = alloc_slab_page(alloc_gfp, node, oo);
1366 if (unlikely(!page)) {
1367 oo = s->min;
1368 /*
1369 * Allocation may have failed due to fragmentation.
1370 * Try a lower order alloc if possible
1371 */
1372 page = alloc_slab_page(flags, node, oo);
1373
1374 if (page)
1375 stat(s, ORDER_FALLBACK);
1376 }
1377
1378 if (flags & __GFP_WAIT)
1379 local_irq_disable();
1380
1381 if (!page)
1382 return NULL;
1383
1384 if (kmemcheck_enabled
1385 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1386 int pages = 1 << oo_order(oo);
1387
1388 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1389
1390 /*
1391 * Objects from caches that have a constructor don't get
1392 * cleared when they're allocated, so we need to do it here.
1393 */
1394 if (s->ctor)
1395 kmemcheck_mark_uninitialized_pages(page, pages);
1396 else
1397 kmemcheck_mark_unallocated_pages(page, pages);
1398 }
1399
1400 page->objects = oo_objects(oo);
1401 mod_zone_page_state(page_zone(page),
1402 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1403 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1404 1 << oo_order(oo));
1405
1406 return page;
1407}
1408
1409static void setup_object(struct kmem_cache *s, struct page *page,
1410 void *object)
1411{
1412 setup_object_debug(s, page, object);
1413 if (unlikely(s->ctor))
1414 s->ctor(object);
1415}
1416
1417static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1418{
1419 struct page *page;
1420 void *start;
1421 void *last;
1422 void *p;
1423
1424 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1425
1426 page = allocate_slab(s,
1427 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1428 if (!page)
1429 goto out;
1430
1431 inc_slabs_node(s, page_to_nid(page), page->objects);
1432 page->slab = s;
1433 page->flags |= 1 << PG_slab;
1434
1435 start = page_address(page);
1436
1437 if (unlikely(s->flags & SLAB_POISON))
1438 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1439
1440 last = start;
1441 for_each_object(p, s, start, page->objects) {
1442 setup_object(s, page, last);
1443 set_freepointer(s, last, p);
1444 last = p;
1445 }
1446 setup_object(s, page, last);
1447 set_freepointer(s, last, NULL);
1448
1449 page->freelist = start;
1450 page->inuse = 0;
1451 page->frozen = 1;
1452out:
1453 return page;
1454}
1455
1456static void __free_slab(struct kmem_cache *s, struct page *page)
1457{
1458 int order = compound_order(page);
1459 int pages = 1 << order;
1460
1461 if (kmem_cache_debug(s)) {
1462 void *p;
1463
1464 slab_pad_check(s, page);
1465 for_each_object(p, s, page_address(page),
1466 page->objects)
1467 check_object(s, page, p, SLUB_RED_INACTIVE);
1468 }
1469
1470 kmemcheck_free_shadow(page, compound_order(page));
1471
1472 mod_zone_page_state(page_zone(page),
1473 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1474 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1475 -pages);
1476
1477 __ClearPageSlab(page);
1478 reset_page_mapcount(page);
1479 if (current->reclaim_state)
1480 current->reclaim_state->reclaimed_slab += pages;
1481 __free_pages(page, order);
1482}
1483
1484#define need_reserve_slab_rcu \
1485 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1486
1487static void rcu_free_slab(struct rcu_head *h)
1488{
1489 struct page *page;
1490
1491 if (need_reserve_slab_rcu)
1492 page = virt_to_head_page(h);
1493 else
1494 page = container_of((struct list_head *)h, struct page, lru);
1495
1496 __free_slab(page->slab, page);
1497}
1498
1499static void free_slab(struct kmem_cache *s, struct page *page)
1500{
1501 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1502 struct rcu_head *head;
1503
1504 if (need_reserve_slab_rcu) {
1505 int order = compound_order(page);
1506 int offset = (PAGE_SIZE << order) - s->reserved;
1507
1508 VM_BUG_ON(s->reserved != sizeof(*head));
1509 head = page_address(page) + offset;
1510 } else {
1511 /*
1512 * RCU free overloads the RCU head over the LRU
1513 */
1514 head = (void *)&page->lru;
1515 }
1516
1517 call_rcu(head, rcu_free_slab);
1518 } else
1519 __free_slab(s, page);
1520}
1521
1522static void discard_slab(struct kmem_cache *s, struct page *page)
1523{
1524 dec_slabs_node(s, page_to_nid(page), page->objects);
1525 free_slab(s, page);
1526}
1527
1528/*
1529 * Management of partially allocated slabs.
1530 *
1531 * list_lock must be held.
1532 */
1533static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1535{
1536 n->nr_partial++;
1537 if (tail)
1538 list_add_tail(&page->lru, &n->partial);
1539 else
1540 list_add(&page->lru, &n->partial);
1541}
1542
1543/*
1544 * list_lock must be held.
1545 */
1546static inline void remove_partial(struct kmem_cache_node *n,
1547 struct page *page)
1548{
1549 list_del(&page->lru);
1550 n->nr_partial--;
1551}
1552
1553/*
1554 * Lock slab, remove from the partial list and put the object into the
1555 * per cpu freelist.
1556 *
1557 * Must hold list_lock.
1558 */
1559static inline int acquire_slab(struct kmem_cache *s,
1560 struct kmem_cache_node *n, struct page *page)
1561{
1562 void *freelist;
1563 unsigned long counters;
1564 struct page new;
1565
1566 /*
1567 * Zap the freelist and set the frozen bit.
1568 * The old freelist is the list of objects for the
1569 * per cpu allocation list.
1570 */
1571 do {
1572 freelist = page->freelist;
1573 counters = page->counters;
1574 new.counters = counters;
1575 new.inuse = page->objects;
1576
1577 VM_BUG_ON(new.frozen);
1578 new.frozen = 1;
1579
1580 } while (!__cmpxchg_double_slab(s, page,
1581 freelist, counters,
1582 NULL, new.counters,
1583 "lock and freeze"));
1584
1585 remove_partial(n, page);
1586
1587 if (freelist) {
1588 /* Populate the per cpu freelist */
1589 this_cpu_write(s->cpu_slab->freelist, freelist);
1590 this_cpu_write(s->cpu_slab->page, page);
1591 this_cpu_write(s->cpu_slab->node, page_to_nid(page));
1592 return 1;
1593 } else {
1594 /*
1595 * Slab page came from the wrong list. No object to allocate
1596 * from. Put it onto the correct list and continue partial
1597 * scan.
1598 */
1599 printk(KERN_ERR "SLUB: %s : Page without available objects on"
1600 " partial list\n", s->name);
1601 return 0;
1602 }
1603}
1604
1605/*
1606 * Try to allocate a partial slab from a specific node.
1607 */
1608static struct page *get_partial_node(struct kmem_cache *s,
1609 struct kmem_cache_node *n)
1610{
1611 struct page *page;
1612
1613 /*
1614 * Racy check. If we mistakenly see no partial slabs then we
1615 * just allocate an empty slab. If we mistakenly try to get a
1616 * partial slab and there is none available then get_partials()
1617 * will return NULL.
1618 */
1619 if (!n || !n->nr_partial)
1620 return NULL;
1621
1622 spin_lock(&n->list_lock);
1623 list_for_each_entry(page, &n->partial, lru)
1624 if (acquire_slab(s, n, page))
1625 goto out;
1626 page = NULL;
1627out:
1628 spin_unlock(&n->list_lock);
1629 return page;
1630}
1631
1632/*
1633 * Get a page from somewhere. Search in increasing NUMA distances.
1634 */
1635static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1636{
1637#ifdef CONFIG_NUMA
1638 struct zonelist *zonelist;
1639 struct zoneref *z;
1640 struct zone *zone;
1641 enum zone_type high_zoneidx = gfp_zone(flags);
1642 struct page *page;
1643
1644 /*
1645 * The defrag ratio allows a configuration of the tradeoffs between
1646 * inter node defragmentation and node local allocations. A lower
1647 * defrag_ratio increases the tendency to do local allocations
1648 * instead of attempting to obtain partial slabs from other nodes.
1649 *
1650 * If the defrag_ratio is set to 0 then kmalloc() always
1651 * returns node local objects. If the ratio is higher then kmalloc()
1652 * may return off node objects because partial slabs are obtained
1653 * from other nodes and filled up.
1654 *
1655 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1656 * defrag_ratio = 1000) then every (well almost) allocation will
1657 * first attempt to defrag slab caches on other nodes. This means
1658 * scanning over all nodes to look for partial slabs which may be
1659 * expensive if we do it every time we are trying to find a slab
1660 * with available objects.
1661 */
1662 if (!s->remote_node_defrag_ratio ||
1663 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1664 return NULL;
1665
1666 get_mems_allowed();
1667 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1668 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1669 struct kmem_cache_node *n;
1670
1671 n = get_node(s, zone_to_nid(zone));
1672
1673 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1674 n->nr_partial > s->min_partial) {
1675 page = get_partial_node(s, n);
1676 if (page) {
1677 put_mems_allowed();
1678 return page;
1679 }
1680 }
1681 }
1682 put_mems_allowed();
1683#endif
1684 return NULL;
1685}
1686
1687/*
1688 * Get a partial page, lock it and return it.
1689 */
1690static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1691{
1692 struct page *page;
1693 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1694
1695 page = get_partial_node(s, get_node(s, searchnode));
1696 if (page || node != NUMA_NO_NODE)
1697 return page;
1698
1699 return get_any_partial(s, flags);
1700}
1701
1702#ifdef CONFIG_PREEMPT
1703/*
1704 * Calculate the next globally unique transaction for disambiguiation
1705 * during cmpxchg. The transactions start with the cpu number and are then
1706 * incremented by CONFIG_NR_CPUS.
1707 */
1708#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1709#else
1710/*
1711 * No preemption supported therefore also no need to check for
1712 * different cpus.
1713 */
1714#define TID_STEP 1
1715#endif
1716
1717static inline unsigned long next_tid(unsigned long tid)
1718{
1719 return tid + TID_STEP;
1720}
1721
1722static inline unsigned int tid_to_cpu(unsigned long tid)
1723{
1724 return tid % TID_STEP;
1725}
1726
1727static inline unsigned long tid_to_event(unsigned long tid)
1728{
1729 return tid / TID_STEP;
1730}
1731
1732static inline unsigned int init_tid(int cpu)
1733{
1734 return cpu;
1735}
1736
1737static inline void note_cmpxchg_failure(const char *n,
1738 const struct kmem_cache *s, unsigned long tid)
1739{
1740#ifdef SLUB_DEBUG_CMPXCHG
1741 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1742
1743 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1744
1745#ifdef CONFIG_PREEMPT
1746 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1747 printk("due to cpu change %d -> %d\n",
1748 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1749 else
1750#endif
1751 if (tid_to_event(tid) != tid_to_event(actual_tid))
1752 printk("due to cpu running other code. Event %ld->%ld\n",
1753 tid_to_event(tid), tid_to_event(actual_tid));
1754 else
1755 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1756 actual_tid, tid, next_tid(tid));
1757#endif
1758 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1759}
1760
1761void init_kmem_cache_cpus(struct kmem_cache *s)
1762{
1763 int cpu;
1764
1765 for_each_possible_cpu(cpu)
1766 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1767}
1768/*
1769 * Remove the cpu slab
1770 */
1771
1772/*
1773 * Remove the cpu slab
1774 */
1775static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1776{
1777 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1778 struct page *page = c->page;
1779 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1780 int lock = 0;
1781 enum slab_modes l = M_NONE, m = M_NONE;
1782 void *freelist;
1783 void *nextfree;
1784 int tail = 0;
1785 struct page new;
1786 struct page old;
1787
1788 if (page->freelist) {
1789 stat(s, DEACTIVATE_REMOTE_FREES);
1790 tail = 1;
1791 }
1792
1793 c->tid = next_tid(c->tid);
1794 c->page = NULL;
1795 freelist = c->freelist;
1796 c->freelist = NULL;
1797
1798 /*
1799 * Stage one: Free all available per cpu objects back
1800 * to the page freelist while it is still frozen. Leave the
1801 * last one.
1802 *
1803 * There is no need to take the list->lock because the page
1804 * is still frozen.
1805 */
1806 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1807 void *prior;
1808 unsigned long counters;
1809
1810 do {
1811 prior = page->freelist;
1812 counters = page->counters;
1813 set_freepointer(s, freelist, prior);
1814 new.counters = counters;
1815 new.inuse--;
1816 VM_BUG_ON(!new.frozen);
1817
1818 } while (!__cmpxchg_double_slab(s, page,
1819 prior, counters,
1820 freelist, new.counters,
1821 "drain percpu freelist"));
1822
1823 freelist = nextfree;
1824 }
1825
1826 /*
1827 * Stage two: Ensure that the page is unfrozen while the
1828 * list presence reflects the actual number of objects
1829 * during unfreeze.
1830 *
1831 * We setup the list membership and then perform a cmpxchg
1832 * with the count. If there is a mismatch then the page
1833 * is not unfrozen but the page is on the wrong list.
1834 *
1835 * Then we restart the process which may have to remove
1836 * the page from the list that we just put it on again
1837 * because the number of objects in the slab may have
1838 * changed.
1839 */
1840redo:
1841
1842 old.freelist = page->freelist;
1843 old.counters = page->counters;
1844 VM_BUG_ON(!old.frozen);
1845
1846 /* Determine target state of the slab */
1847 new.counters = old.counters;
1848 if (freelist) {
1849 new.inuse--;
1850 set_freepointer(s, freelist, old.freelist);
1851 new.freelist = freelist;
1852 } else
1853 new.freelist = old.freelist;
1854
1855 new.frozen = 0;
1856
1857 if (!new.inuse && n->nr_partial > s->min_partial)
1858 m = M_FREE;
1859 else if (new.freelist) {
1860 m = M_PARTIAL;
1861 if (!lock) {
1862 lock = 1;
1863 /*
1864 * Taking the spinlock removes the possiblity
1865 * that acquire_slab() will see a slab page that
1866 * is frozen
1867 */
1868 spin_lock(&n->list_lock);
1869 }
1870 } else {
1871 m = M_FULL;
1872 if (kmem_cache_debug(s) && !lock) {
1873 lock = 1;
1874 /*
1875 * This also ensures that the scanning of full
1876 * slabs from diagnostic functions will not see
1877 * any frozen slabs.
1878 */
1879 spin_lock(&n->list_lock);
1880 }
1881 }
1882
1883 if (l != m) {
1884
1885 if (l == M_PARTIAL)
1886
1887 remove_partial(n, page);
1888
1889 else if (l == M_FULL)
1890
1891 remove_full(s, page);
1892
1893 if (m == M_PARTIAL) {
1894
1895 add_partial(n, page, tail);
1896 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1897
1898 } else if (m == M_FULL) {
1899
1900 stat(s, DEACTIVATE_FULL);
1901 add_full(s, n, page);
1902
1903 }
1904 }
1905
1906 l = m;
1907 if (!__cmpxchg_double_slab(s, page,
1908 old.freelist, old.counters,
1909 new.freelist, new.counters,
1910 "unfreezing slab"))
1911 goto redo;
1912
1913 if (lock)
1914 spin_unlock(&n->list_lock);
1915
1916 if (m == M_FREE) {
1917 stat(s, DEACTIVATE_EMPTY);
1918 discard_slab(s, page);
1919 stat(s, FREE_SLAB);
1920 }
1921}
1922
1923static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1924{
1925 stat(s, CPUSLAB_FLUSH);
1926 deactivate_slab(s, c);
1927}
1928
1929/*
1930 * Flush cpu slab.
1931 *
1932 * Called from IPI handler with interrupts disabled.
1933 */
1934static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1935{
1936 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1937
1938 if (likely(c && c->page))
1939 flush_slab(s, c);
1940}
1941
1942static void flush_cpu_slab(void *d)
1943{
1944 struct kmem_cache *s = d;
1945
1946 __flush_cpu_slab(s, smp_processor_id());
1947}
1948
1949static void flush_all(struct kmem_cache *s)
1950{
1951 on_each_cpu(flush_cpu_slab, s, 1);
1952}
1953
1954/*
1955 * Check if the objects in a per cpu structure fit numa
1956 * locality expectations.
1957 */
1958static inline int node_match(struct kmem_cache_cpu *c, int node)
1959{
1960#ifdef CONFIG_NUMA
1961 if (node != NUMA_NO_NODE && c->node != node)
1962 return 0;
1963#endif
1964 return 1;
1965}
1966
1967static int count_free(struct page *page)
1968{
1969 return page->objects - page->inuse;
1970}
1971
1972static unsigned long count_partial(struct kmem_cache_node *n,
1973 int (*get_count)(struct page *))
1974{
1975 unsigned long flags;
1976 unsigned long x = 0;
1977 struct page *page;
1978
1979 spin_lock_irqsave(&n->list_lock, flags);
1980 list_for_each_entry(page, &n->partial, lru)
1981 x += get_count(page);
1982 spin_unlock_irqrestore(&n->list_lock, flags);
1983 return x;
1984}
1985
1986static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1987{
1988#ifdef CONFIG_SLUB_DEBUG
1989 return atomic_long_read(&n->total_objects);
1990#else
1991 return 0;
1992#endif
1993}
1994
1995static noinline void
1996slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1997{
1998 int node;
1999
2000 printk(KERN_WARNING
2001 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2002 nid, gfpflags);
2003 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2004 "default order: %d, min order: %d\n", s->name, s->objsize,
2005 s->size, oo_order(s->oo), oo_order(s->min));
2006
2007 if (oo_order(s->min) > get_order(s->objsize))
2008 printk(KERN_WARNING " %s debugging increased min order, use "
2009 "slub_debug=O to disable.\n", s->name);
2010
2011 for_each_online_node(node) {
2012 struct kmem_cache_node *n = get_node(s, node);
2013 unsigned long nr_slabs;
2014 unsigned long nr_objs;
2015 unsigned long nr_free;
2016
2017 if (!n)
2018 continue;
2019
2020 nr_free = count_partial(n, count_free);
2021 nr_slabs = node_nr_slabs(n);
2022 nr_objs = node_nr_objs(n);
2023
2024 printk(KERN_WARNING
2025 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2026 node, nr_slabs, nr_objs, nr_free);
2027 }
2028}
2029
2030/*
2031 * Slow path. The lockless freelist is empty or we need to perform
2032 * debugging duties.
2033 *
2034 * Interrupts are disabled.
2035 *
2036 * Processing is still very fast if new objects have been freed to the
2037 * regular freelist. In that case we simply take over the regular freelist
2038 * as the lockless freelist and zap the regular freelist.
2039 *
2040 * If that is not working then we fall back to the partial lists. We take the
2041 * first element of the freelist as the object to allocate now and move the
2042 * rest of the freelist to the lockless freelist.
2043 *
2044 * And if we were unable to get a new slab from the partial slab lists then
2045 * we need to allocate a new slab. This is the slowest path since it involves
2046 * a call to the page allocator and the setup of a new slab.
2047 */
2048static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2049 unsigned long addr, struct kmem_cache_cpu *c)
2050{
2051 void **object;
2052 struct page *page;
2053 unsigned long flags;
2054 struct page new;
2055 unsigned long counters;
2056
2057 local_irq_save(flags);
2058#ifdef CONFIG_PREEMPT
2059 /*
2060 * We may have been preempted and rescheduled on a different
2061 * cpu before disabling interrupts. Need to reload cpu area
2062 * pointer.
2063 */
2064 c = this_cpu_ptr(s->cpu_slab);
2065#endif
2066
2067 /* We handle __GFP_ZERO in the caller */
2068 gfpflags &= ~__GFP_ZERO;
2069
2070 page = c->page;
2071 if (!page)
2072 goto new_slab;
2073
2074 if (unlikely(!node_match(c, node))) {
2075 stat(s, ALLOC_NODE_MISMATCH);
2076 deactivate_slab(s, c);
2077 goto new_slab;
2078 }
2079
2080 stat(s, ALLOC_SLOWPATH);
2081
2082 do {
2083 object = page->freelist;
2084 counters = page->counters;
2085 new.counters = counters;
2086 VM_BUG_ON(!new.frozen);
2087
2088 /*
2089 * If there is no object left then we use this loop to
2090 * deactivate the slab which is simple since no objects
2091 * are left in the slab and therefore we do not need to
2092 * put the page back onto the partial list.
2093 *
2094 * If there are objects left then we retrieve them
2095 * and use them to refill the per cpu queue.
2096 */
2097
2098 new.inuse = page->objects;
2099 new.frozen = object != NULL;
2100
2101 } while (!__cmpxchg_double_slab(s, page,
2102 object, counters,
2103 NULL, new.counters,
2104 "__slab_alloc"));
2105
2106 if (unlikely(!object)) {
2107 c->page = NULL;
2108 stat(s, DEACTIVATE_BYPASS);
2109 goto new_slab;
2110 }
2111
2112 stat(s, ALLOC_REFILL);
2113
2114load_freelist:
2115 VM_BUG_ON(!page->frozen);
2116 c->freelist = get_freepointer(s, object);
2117 c->tid = next_tid(c->tid);
2118 local_irq_restore(flags);
2119 return object;
2120
2121new_slab:
2122 page = get_partial(s, gfpflags, node);
2123 if (page) {
2124 stat(s, ALLOC_FROM_PARTIAL);
2125 object = c->freelist;
2126
2127 if (kmem_cache_debug(s))
2128 goto debug;
2129 goto load_freelist;
2130 }
2131
2132 page = new_slab(s, gfpflags, node);
2133
2134 if (page) {
2135 c = __this_cpu_ptr(s->cpu_slab);
2136 if (c->page)
2137 flush_slab(s, c);
2138
2139 /*
2140 * No other reference to the page yet so we can
2141 * muck around with it freely without cmpxchg
2142 */
2143 object = page->freelist;
2144 page->freelist = NULL;
2145 page->inuse = page->objects;
2146
2147 stat(s, ALLOC_SLAB);
2148 c->node = page_to_nid(page);
2149 c->page = page;
2150
2151 if (kmem_cache_debug(s))
2152 goto debug;
2153 goto load_freelist;
2154 }
2155 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2156 slab_out_of_memory(s, gfpflags, node);
2157 local_irq_restore(flags);
2158 return NULL;
2159
2160debug:
2161 if (!object || !alloc_debug_processing(s, page, object, addr))
2162 goto new_slab;
2163
2164 c->freelist = get_freepointer(s, object);
2165 deactivate_slab(s, c);
2166 c->page = NULL;
2167 c->node = NUMA_NO_NODE;
2168 local_irq_restore(flags);
2169 return object;
2170}
2171
2172/*
2173 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2174 * have the fastpath folded into their functions. So no function call
2175 * overhead for requests that can be satisfied on the fastpath.
2176 *
2177 * The fastpath works by first checking if the lockless freelist can be used.
2178 * If not then __slab_alloc is called for slow processing.
2179 *
2180 * Otherwise we can simply pick the next object from the lockless free list.
2181 */
2182static __always_inline void *slab_alloc(struct kmem_cache *s,
2183 gfp_t gfpflags, int node, unsigned long addr)
2184{
2185 void **object;
2186 struct kmem_cache_cpu *c;
2187 unsigned long tid;
2188
2189 if (slab_pre_alloc_hook(s, gfpflags))
2190 return NULL;
2191
2192redo:
2193
2194 /*
2195 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2196 * enabled. We may switch back and forth between cpus while
2197 * reading from one cpu area. That does not matter as long
2198 * as we end up on the original cpu again when doing the cmpxchg.
2199 */
2200 c = __this_cpu_ptr(s->cpu_slab);
2201
2202 /*
2203 * The transaction ids are globally unique per cpu and per operation on
2204 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2205 * occurs on the right processor and that there was no operation on the
2206 * linked list in between.
2207 */
2208 tid = c->tid;
2209 barrier();
2210
2211 object = c->freelist;
2212 if (unlikely(!object || !node_match(c, node)))
2213
2214 object = __slab_alloc(s, gfpflags, node, addr, c);
2215
2216 else {
2217 /*
2218 * The cmpxchg will only match if there was no additional
2219 * operation and if we are on the right processor.
2220 *
2221 * The cmpxchg does the following atomically (without lock semantics!)
2222 * 1. Relocate first pointer to the current per cpu area.
2223 * 2. Verify that tid and freelist have not been changed
2224 * 3. If they were not changed replace tid and freelist
2225 *
2226 * Since this is without lock semantics the protection is only against
2227 * code executing on this cpu *not* from access by other cpus.
2228 */
2229 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2230 s->cpu_slab->freelist, s->cpu_slab->tid,
2231 object, tid,
2232 get_freepointer_safe(s, object), next_tid(tid)))) {
2233
2234 note_cmpxchg_failure("slab_alloc", s, tid);
2235 goto redo;
2236 }
2237 stat(s, ALLOC_FASTPATH);
2238 }
2239
2240 if (unlikely(gfpflags & __GFP_ZERO) && object)
2241 memset(object, 0, s->objsize);
2242
2243 slab_post_alloc_hook(s, gfpflags, object);
2244
2245 return object;
2246}
2247
2248void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2249{
2250 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2251
2252 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2253
2254 return ret;
2255}
2256EXPORT_SYMBOL(kmem_cache_alloc);
2257
2258#ifdef CONFIG_TRACING
2259void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2260{
2261 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2262 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2263 return ret;
2264}
2265EXPORT_SYMBOL(kmem_cache_alloc_trace);
2266
2267void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2268{
2269 void *ret = kmalloc_order(size, flags, order);
2270 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2271 return ret;
2272}
2273EXPORT_SYMBOL(kmalloc_order_trace);
2274#endif
2275
2276#ifdef CONFIG_NUMA
2277void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2278{
2279 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2280
2281 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2282 s->objsize, s->size, gfpflags, node);
2283
2284 return ret;
2285}
2286EXPORT_SYMBOL(kmem_cache_alloc_node);
2287
2288#ifdef CONFIG_TRACING
2289void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2290 gfp_t gfpflags,
2291 int node, size_t size)
2292{
2293 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2294
2295 trace_kmalloc_node(_RET_IP_, ret,
2296 size, s->size, gfpflags, node);
2297 return ret;
2298}
2299EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2300#endif
2301#endif
2302
2303/*
2304 * Slow patch handling. This may still be called frequently since objects
2305 * have a longer lifetime than the cpu slabs in most processing loads.
2306 *
2307 * So we still attempt to reduce cache line usage. Just take the slab
2308 * lock and free the item. If there is no additional partial page
2309 * handling required then we can return immediately.
2310 */
2311static void __slab_free(struct kmem_cache *s, struct page *page,
2312 void *x, unsigned long addr)
2313{
2314 void *prior;
2315 void **object = (void *)x;
2316 int was_frozen;
2317 int inuse;
2318 struct page new;
2319 unsigned long counters;
2320 struct kmem_cache_node *n = NULL;
2321 unsigned long uninitialized_var(flags);
2322
2323 stat(s, FREE_SLOWPATH);
2324
2325 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2326 return;
2327
2328 do {
2329 prior = page->freelist;
2330 counters = page->counters;
2331 set_freepointer(s, object, prior);
2332 new.counters = counters;
2333 was_frozen = new.frozen;
2334 new.inuse--;
2335 if ((!new.inuse || !prior) && !was_frozen && !n) {
2336 n = get_node(s, page_to_nid(page));
2337 /*
2338 * Speculatively acquire the list_lock.
2339 * If the cmpxchg does not succeed then we may
2340 * drop the list_lock without any processing.
2341 *
2342 * Otherwise the list_lock will synchronize with
2343 * other processors updating the list of slabs.
2344 */
2345 spin_lock_irqsave(&n->list_lock, flags);
2346 }
2347 inuse = new.inuse;
2348
2349 } while (!cmpxchg_double_slab(s, page,
2350 prior, counters,
2351 object, new.counters,
2352 "__slab_free"));
2353
2354 if (likely(!n)) {
2355 /*
2356 * The list lock was not taken therefore no list
2357 * activity can be necessary.
2358 */
2359 if (was_frozen)
2360 stat(s, FREE_FROZEN);
2361 return;
2362 }
2363
2364 /*
2365 * was_frozen may have been set after we acquired the list_lock in
2366 * an earlier loop. So we need to check it here again.
2367 */
2368 if (was_frozen)
2369 stat(s, FREE_FROZEN);
2370 else {
2371 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2372 goto slab_empty;
2373
2374 /*
2375 * Objects left in the slab. If it was not on the partial list before
2376 * then add it.
2377 */
2378 if (unlikely(!prior)) {
2379 remove_full(s, page);
2380 add_partial(n, page, 1);
2381 stat(s, FREE_ADD_PARTIAL);
2382 }
2383 }
2384 spin_unlock_irqrestore(&n->list_lock, flags);
2385 return;
2386
2387slab_empty:
2388 if (prior) {
2389 /*
2390 * Slab on the partial list.
2391 */
2392 remove_partial(n, page);
2393 stat(s, FREE_REMOVE_PARTIAL);
2394 } else
2395 /* Slab must be on the full list */
2396 remove_full(s, page);
2397
2398 spin_unlock_irqrestore(&n->list_lock, flags);
2399 stat(s, FREE_SLAB);
2400 discard_slab(s, page);
2401}
2402
2403/*
2404 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2405 * can perform fastpath freeing without additional function calls.
2406 *
2407 * The fastpath is only possible if we are freeing to the current cpu slab
2408 * of this processor. This typically the case if we have just allocated
2409 * the item before.
2410 *
2411 * If fastpath is not possible then fall back to __slab_free where we deal
2412 * with all sorts of special processing.
2413 */
2414static __always_inline void slab_free(struct kmem_cache *s,
2415 struct page *page, void *x, unsigned long addr)
2416{
2417 void **object = (void *)x;
2418 struct kmem_cache_cpu *c;
2419 unsigned long tid;
2420
2421 slab_free_hook(s, x);
2422
2423redo:
2424
2425 /*
2426 * Determine the currently cpus per cpu slab.
2427 * The cpu may change afterward. However that does not matter since
2428 * data is retrieved via this pointer. If we are on the same cpu
2429 * during the cmpxchg then the free will succedd.
2430 */
2431 c = __this_cpu_ptr(s->cpu_slab);
2432
2433 tid = c->tid;
2434 barrier();
2435
2436 if (likely(page == c->page)) {
2437 set_freepointer(s, object, c->freelist);
2438
2439 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2440 s->cpu_slab->freelist, s->cpu_slab->tid,
2441 c->freelist, tid,
2442 object, next_tid(tid)))) {
2443
2444 note_cmpxchg_failure("slab_free", s, tid);
2445 goto redo;
2446 }
2447 stat(s, FREE_FASTPATH);
2448 } else
2449 __slab_free(s, page, x, addr);
2450
2451}
2452
2453void kmem_cache_free(struct kmem_cache *s, void *x)
2454{
2455 struct page *page;
2456
2457 page = virt_to_head_page(x);
2458
2459 slab_free(s, page, x, _RET_IP_);
2460
2461 trace_kmem_cache_free(_RET_IP_, x);
2462}
2463EXPORT_SYMBOL(kmem_cache_free);
2464
2465/*
2466 * Object placement in a slab is made very easy because we always start at
2467 * offset 0. If we tune the size of the object to the alignment then we can
2468 * get the required alignment by putting one properly sized object after
2469 * another.
2470 *
2471 * Notice that the allocation order determines the sizes of the per cpu
2472 * caches. Each processor has always one slab available for allocations.
2473 * Increasing the allocation order reduces the number of times that slabs
2474 * must be moved on and off the partial lists and is therefore a factor in
2475 * locking overhead.
2476 */
2477
2478/*
2479 * Mininum / Maximum order of slab pages. This influences locking overhead
2480 * and slab fragmentation. A higher order reduces the number of partial slabs
2481 * and increases the number of allocations possible without having to
2482 * take the list_lock.
2483 */
2484static int slub_min_order;
2485static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2486static int slub_min_objects;
2487
2488/*
2489 * Merge control. If this is set then no merging of slab caches will occur.
2490 * (Could be removed. This was introduced to pacify the merge skeptics.)
2491 */
2492static int slub_nomerge;
2493
2494/*
2495 * Calculate the order of allocation given an slab object size.
2496 *
2497 * The order of allocation has significant impact on performance and other
2498 * system components. Generally order 0 allocations should be preferred since
2499 * order 0 does not cause fragmentation in the page allocator. Larger objects
2500 * be problematic to put into order 0 slabs because there may be too much
2501 * unused space left. We go to a higher order if more than 1/16th of the slab
2502 * would be wasted.
2503 *
2504 * In order to reach satisfactory performance we must ensure that a minimum
2505 * number of objects is in one slab. Otherwise we may generate too much
2506 * activity on the partial lists which requires taking the list_lock. This is
2507 * less a concern for large slabs though which are rarely used.
2508 *
2509 * slub_max_order specifies the order where we begin to stop considering the
2510 * number of objects in a slab as critical. If we reach slub_max_order then
2511 * we try to keep the page order as low as possible. So we accept more waste
2512 * of space in favor of a small page order.
2513 *
2514 * Higher order allocations also allow the placement of more objects in a
2515 * slab and thereby reduce object handling overhead. If the user has
2516 * requested a higher mininum order then we start with that one instead of
2517 * the smallest order which will fit the object.
2518 */
2519static inline int slab_order(int size, int min_objects,
2520 int max_order, int fract_leftover, int reserved)
2521{
2522 int order;
2523 int rem;
2524 int min_order = slub_min_order;
2525
2526 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2527 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2528
2529 for (order = max(min_order,
2530 fls(min_objects * size - 1) - PAGE_SHIFT);
2531 order <= max_order; order++) {
2532
2533 unsigned long slab_size = PAGE_SIZE << order;
2534
2535 if (slab_size < min_objects * size + reserved)
2536 continue;
2537
2538 rem = (slab_size - reserved) % size;
2539
2540 if (rem <= slab_size / fract_leftover)
2541 break;
2542
2543 }
2544
2545 return order;
2546}
2547
2548static inline int calculate_order(int size, int reserved)
2549{
2550 int order;
2551 int min_objects;
2552 int fraction;
2553 int max_objects;
2554
2555 /*
2556 * Attempt to find best configuration for a slab. This
2557 * works by first attempting to generate a layout with
2558 * the best configuration and backing off gradually.
2559 *
2560 * First we reduce the acceptable waste in a slab. Then
2561 * we reduce the minimum objects required in a slab.
2562 */
2563 min_objects = slub_min_objects;
2564 if (!min_objects)
2565 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2566 max_objects = order_objects(slub_max_order, size, reserved);
2567 min_objects = min(min_objects, max_objects);
2568
2569 while (min_objects > 1) {
2570 fraction = 16;
2571 while (fraction >= 4) {
2572 order = slab_order(size, min_objects,
2573 slub_max_order, fraction, reserved);
2574 if (order <= slub_max_order)
2575 return order;
2576 fraction /= 2;
2577 }
2578 min_objects--;
2579 }
2580
2581 /*
2582 * We were unable to place multiple objects in a slab. Now
2583 * lets see if we can place a single object there.
2584 */
2585 order = slab_order(size, 1, slub_max_order, 1, reserved);
2586 if (order <= slub_max_order)
2587 return order;
2588
2589 /*
2590 * Doh this slab cannot be placed using slub_max_order.
2591 */
2592 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2593 if (order < MAX_ORDER)
2594 return order;
2595 return -ENOSYS;
2596}
2597
2598/*
2599 * Figure out what the alignment of the objects will be.
2600 */
2601static unsigned long calculate_alignment(unsigned long flags,
2602 unsigned long align, unsigned long size)
2603{
2604 /*
2605 * If the user wants hardware cache aligned objects then follow that
2606 * suggestion if the object is sufficiently large.
2607 *
2608 * The hardware cache alignment cannot override the specified
2609 * alignment though. If that is greater then use it.
2610 */
2611 if (flags & SLAB_HWCACHE_ALIGN) {
2612 unsigned long ralign = cache_line_size();
2613 while (size <= ralign / 2)
2614 ralign /= 2;
2615 align = max(align, ralign);
2616 }
2617
2618 if (align < ARCH_SLAB_MINALIGN)
2619 align = ARCH_SLAB_MINALIGN;
2620
2621 return ALIGN(align, sizeof(void *));
2622}
2623
2624static void
2625init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2626{
2627 n->nr_partial = 0;
2628 spin_lock_init(&n->list_lock);
2629 INIT_LIST_HEAD(&n->partial);
2630#ifdef CONFIG_SLUB_DEBUG
2631 atomic_long_set(&n->nr_slabs, 0);
2632 atomic_long_set(&n->total_objects, 0);
2633 INIT_LIST_HEAD(&n->full);
2634#endif
2635}
2636
2637static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2638{
2639 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2640 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2641
2642 /*
2643 * Must align to double word boundary for the double cmpxchg
2644 * instructions to work; see __pcpu_double_call_return_bool().
2645 */
2646 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2647 2 * sizeof(void *));
2648
2649 if (!s->cpu_slab)
2650 return 0;
2651
2652 init_kmem_cache_cpus(s);
2653
2654 return 1;
2655}
2656
2657static struct kmem_cache *kmem_cache_node;
2658
2659/*
2660 * No kmalloc_node yet so do it by hand. We know that this is the first
2661 * slab on the node for this slabcache. There are no concurrent accesses
2662 * possible.
2663 *
2664 * Note that this function only works on the kmalloc_node_cache
2665 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2666 * memory on a fresh node that has no slab structures yet.
2667 */
2668static void early_kmem_cache_node_alloc(int node)
2669{
2670 struct page *page;
2671 struct kmem_cache_node *n;
2672
2673 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2674
2675 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2676
2677 BUG_ON(!page);
2678 if (page_to_nid(page) != node) {
2679 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2680 "node %d\n", node);
2681 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2682 "in order to be able to continue\n");
2683 }
2684
2685 n = page->freelist;
2686 BUG_ON(!n);
2687 page->freelist = get_freepointer(kmem_cache_node, n);
2688 page->inuse++;
2689 page->frozen = 0;
2690 kmem_cache_node->node[node] = n;
2691#ifdef CONFIG_SLUB_DEBUG
2692 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2693 init_tracking(kmem_cache_node, n);
2694#endif
2695 init_kmem_cache_node(n, kmem_cache_node);
2696 inc_slabs_node(kmem_cache_node, node, page->objects);
2697
2698 add_partial(n, page, 0);
2699}
2700
2701static void free_kmem_cache_nodes(struct kmem_cache *s)
2702{
2703 int node;
2704
2705 for_each_node_state(node, N_NORMAL_MEMORY) {
2706 struct kmem_cache_node *n = s->node[node];
2707
2708 if (n)
2709 kmem_cache_free(kmem_cache_node, n);
2710
2711 s->node[node] = NULL;
2712 }
2713}
2714
2715static int init_kmem_cache_nodes(struct kmem_cache *s)
2716{
2717 int node;
2718
2719 for_each_node_state(node, N_NORMAL_MEMORY) {
2720 struct kmem_cache_node *n;
2721
2722 if (slab_state == DOWN) {
2723 early_kmem_cache_node_alloc(node);
2724 continue;
2725 }
2726 n = kmem_cache_alloc_node(kmem_cache_node,
2727 GFP_KERNEL, node);
2728
2729 if (!n) {
2730 free_kmem_cache_nodes(s);
2731 return 0;
2732 }
2733
2734 s->node[node] = n;
2735 init_kmem_cache_node(n, s);
2736 }
2737 return 1;
2738}
2739
2740static void set_min_partial(struct kmem_cache *s, unsigned long min)
2741{
2742 if (min < MIN_PARTIAL)
2743 min = MIN_PARTIAL;
2744 else if (min > MAX_PARTIAL)
2745 min = MAX_PARTIAL;
2746 s->min_partial = min;
2747}
2748
2749/*
2750 * calculate_sizes() determines the order and the distribution of data within
2751 * a slab object.
2752 */
2753static int calculate_sizes(struct kmem_cache *s, int forced_order)
2754{
2755 unsigned long flags = s->flags;
2756 unsigned long size = s->objsize;
2757 unsigned long align = s->align;
2758 int order;
2759
2760 /*
2761 * Round up object size to the next word boundary. We can only
2762 * place the free pointer at word boundaries and this determines
2763 * the possible location of the free pointer.
2764 */
2765 size = ALIGN(size, sizeof(void *));
2766
2767#ifdef CONFIG_SLUB_DEBUG
2768 /*
2769 * Determine if we can poison the object itself. If the user of
2770 * the slab may touch the object after free or before allocation
2771 * then we should never poison the object itself.
2772 */
2773 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2774 !s->ctor)
2775 s->flags |= __OBJECT_POISON;
2776 else
2777 s->flags &= ~__OBJECT_POISON;
2778
2779
2780 /*
2781 * If we are Redzoning then check if there is some space between the
2782 * end of the object and the free pointer. If not then add an
2783 * additional word to have some bytes to store Redzone information.
2784 */
2785 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2786 size += sizeof(void *);
2787#endif
2788
2789 /*
2790 * With that we have determined the number of bytes in actual use
2791 * by the object. This is the potential offset to the free pointer.
2792 */
2793 s->inuse = size;
2794
2795 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2796 s->ctor)) {
2797 /*
2798 * Relocate free pointer after the object if it is not
2799 * permitted to overwrite the first word of the object on
2800 * kmem_cache_free.
2801 *
2802 * This is the case if we do RCU, have a constructor or
2803 * destructor or are poisoning the objects.
2804 */
2805 s->offset = size;
2806 size += sizeof(void *);
2807 }
2808
2809#ifdef CONFIG_SLUB_DEBUG
2810 if (flags & SLAB_STORE_USER)
2811 /*
2812 * Need to store information about allocs and frees after
2813 * the object.
2814 */
2815 size += 2 * sizeof(struct track);
2816
2817 if (flags & SLAB_RED_ZONE)
2818 /*
2819 * Add some empty padding so that we can catch
2820 * overwrites from earlier objects rather than let
2821 * tracking information or the free pointer be
2822 * corrupted if a user writes before the start
2823 * of the object.
2824 */
2825 size += sizeof(void *);
2826#endif
2827
2828 /*
2829 * Determine the alignment based on various parameters that the
2830 * user specified and the dynamic determination of cache line size
2831 * on bootup.
2832 */
2833 align = calculate_alignment(flags, align, s->objsize);
2834 s->align = align;
2835
2836 /*
2837 * SLUB stores one object immediately after another beginning from
2838 * offset 0. In order to align the objects we have to simply size
2839 * each object to conform to the alignment.
2840 */
2841 size = ALIGN(size, align);
2842 s->size = size;
2843 if (forced_order >= 0)
2844 order = forced_order;
2845 else
2846 order = calculate_order(size, s->reserved);
2847
2848 if (order < 0)
2849 return 0;
2850
2851 s->allocflags = 0;
2852 if (order)
2853 s->allocflags |= __GFP_COMP;
2854
2855 if (s->flags & SLAB_CACHE_DMA)
2856 s->allocflags |= SLUB_DMA;
2857
2858 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2859 s->allocflags |= __GFP_RECLAIMABLE;
2860
2861 /*
2862 * Determine the number of objects per slab
2863 */
2864 s->oo = oo_make(order, size, s->reserved);
2865 s->min = oo_make(get_order(size), size, s->reserved);
2866 if (oo_objects(s->oo) > oo_objects(s->max))
2867 s->max = s->oo;
2868
2869 return !!oo_objects(s->oo);
2870
2871}
2872
2873static int kmem_cache_open(struct kmem_cache *s,
2874 const char *name, size_t size,
2875 size_t align, unsigned long flags,
2876 void (*ctor)(void *))
2877{
2878 memset(s, 0, kmem_size);
2879 s->name = name;
2880 s->ctor = ctor;
2881 s->objsize = size;
2882 s->align = align;
2883 s->flags = kmem_cache_flags(size, flags, name, ctor);
2884 s->reserved = 0;
2885
2886 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2887 s->reserved = sizeof(struct rcu_head);
2888
2889 if (!calculate_sizes(s, -1))
2890 goto error;
2891 if (disable_higher_order_debug) {
2892 /*
2893 * Disable debugging flags that store metadata if the min slab
2894 * order increased.
2895 */
2896 if (get_order(s->size) > get_order(s->objsize)) {
2897 s->flags &= ~DEBUG_METADATA_FLAGS;
2898 s->offset = 0;
2899 if (!calculate_sizes(s, -1))
2900 goto error;
2901 }
2902 }
2903
2904#ifdef CONFIG_CMPXCHG_DOUBLE
2905 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2906 /* Enable fast mode */
2907 s->flags |= __CMPXCHG_DOUBLE;
2908#endif
2909
2910 /*
2911 * The larger the object size is, the more pages we want on the partial
2912 * list to avoid pounding the page allocator excessively.
2913 */
2914 set_min_partial(s, ilog2(s->size));
2915 s->refcount = 1;
2916#ifdef CONFIG_NUMA
2917 s->remote_node_defrag_ratio = 1000;
2918#endif
2919 if (!init_kmem_cache_nodes(s))
2920 goto error;
2921
2922 if (alloc_kmem_cache_cpus(s))
2923 return 1;
2924
2925 free_kmem_cache_nodes(s);
2926error:
2927 if (flags & SLAB_PANIC)
2928 panic("Cannot create slab %s size=%lu realsize=%u "
2929 "order=%u offset=%u flags=%lx\n",
2930 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2931 s->offset, flags);
2932 return 0;
2933}
2934
2935/*
2936 * Determine the size of a slab object
2937 */
2938unsigned int kmem_cache_size(struct kmem_cache *s)
2939{
2940 return s->objsize;
2941}
2942EXPORT_SYMBOL(kmem_cache_size);
2943
2944static void list_slab_objects(struct kmem_cache *s, struct page *page,
2945 const char *text)
2946{
2947#ifdef CONFIG_SLUB_DEBUG
2948 void *addr = page_address(page);
2949 void *p;
2950 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2951 sizeof(long), GFP_ATOMIC);
2952 if (!map)
2953 return;
2954 slab_err(s, page, "%s", text);
2955 slab_lock(page);
2956
2957 get_map(s, page, map);
2958 for_each_object(p, s, addr, page->objects) {
2959
2960 if (!test_bit(slab_index(p, s, addr), map)) {
2961 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2962 p, p - addr);
2963 print_tracking(s, p);
2964 }
2965 }
2966 slab_unlock(page);
2967 kfree(map);
2968#endif
2969}
2970
2971/*
2972 * Attempt to free all partial slabs on a node.
2973 */
2974static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2975{
2976 unsigned long flags;
2977 struct page *page, *h;
2978
2979 spin_lock_irqsave(&n->list_lock, flags);
2980 list_for_each_entry_safe(page, h, &n->partial, lru) {
2981 if (!page->inuse) {
2982 remove_partial(n, page);
2983 discard_slab(s, page);
2984 } else {
2985 list_slab_objects(s, page,
2986 "Objects remaining on kmem_cache_close()");
2987 }
2988 }
2989 spin_unlock_irqrestore(&n->list_lock, flags);
2990}
2991
2992/*
2993 * Release all resources used by a slab cache.
2994 */
2995static inline int kmem_cache_close(struct kmem_cache *s)
2996{
2997 int node;
2998
2999 flush_all(s);
3000 free_percpu(s->cpu_slab);
3001 /* Attempt to free all objects */
3002 for_each_node_state(node, N_NORMAL_MEMORY) {
3003 struct kmem_cache_node *n = get_node(s, node);
3004
3005 free_partial(s, n);
3006 if (n->nr_partial || slabs_node(s, node))
3007 return 1;
3008 }
3009 free_kmem_cache_nodes(s);
3010 return 0;
3011}
3012
3013/*
3014 * Close a cache and release the kmem_cache structure
3015 * (must be used for caches created using kmem_cache_create)
3016 */
3017void kmem_cache_destroy(struct kmem_cache *s)
3018{
3019 down_write(&slub_lock);
3020 s->refcount--;
3021 if (!s->refcount) {
3022 list_del(&s->list);
3023 if (kmem_cache_close(s)) {
3024 printk(KERN_ERR "SLUB %s: %s called for cache that "
3025 "still has objects.\n", s->name, __func__);
3026 dump_stack();
3027 }
3028 if (s->flags & SLAB_DESTROY_BY_RCU)
3029 rcu_barrier();
3030 sysfs_slab_remove(s);
3031 }
3032 up_write(&slub_lock);
3033}
3034EXPORT_SYMBOL(kmem_cache_destroy);
3035
3036/********************************************************************
3037 * Kmalloc subsystem
3038 *******************************************************************/
3039
3040struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3041EXPORT_SYMBOL(kmalloc_caches);
3042
3043static struct kmem_cache *kmem_cache;
3044
3045#ifdef CONFIG_ZONE_DMA
3046static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3047#endif
3048
3049static int __init setup_slub_min_order(char *str)
3050{
3051 get_option(&str, &slub_min_order);
3052
3053 return 1;
3054}
3055
3056__setup("slub_min_order=", setup_slub_min_order);
3057
3058static int __init setup_slub_max_order(char *str)
3059{
3060 get_option(&str, &slub_max_order);
3061 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3062
3063 return 1;
3064}
3065
3066__setup("slub_max_order=", setup_slub_max_order);
3067
3068static int __init setup_slub_min_objects(char *str)
3069{
3070 get_option(&str, &slub_min_objects);
3071
3072 return 1;
3073}
3074
3075__setup("slub_min_objects=", setup_slub_min_objects);
3076
3077static int __init setup_slub_nomerge(char *str)
3078{
3079 slub_nomerge = 1;
3080 return 1;
3081}
3082
3083__setup("slub_nomerge", setup_slub_nomerge);
3084
3085static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3086 int size, unsigned int flags)
3087{
3088 struct kmem_cache *s;
3089
3090 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3091
3092 /*
3093 * This function is called with IRQs disabled during early-boot on
3094 * single CPU so there's no need to take slub_lock here.
3095 */
3096 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3097 flags, NULL))
3098 goto panic;
3099
3100 list_add(&s->list, &slab_caches);
3101 return s;
3102
3103panic:
3104 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3105 return NULL;
3106}
3107
3108/*
3109 * Conversion table for small slabs sizes / 8 to the index in the
3110 * kmalloc array. This is necessary for slabs < 192 since we have non power
3111 * of two cache sizes there. The size of larger slabs can be determined using
3112 * fls.
3113 */
3114static s8 size_index[24] = {
3115 3, /* 8 */
3116 4, /* 16 */
3117 5, /* 24 */
3118 5, /* 32 */
3119 6, /* 40 */
3120 6, /* 48 */
3121 6, /* 56 */
3122 6, /* 64 */
3123 1, /* 72 */
3124 1, /* 80 */
3125 1, /* 88 */
3126 1, /* 96 */
3127 7, /* 104 */
3128 7, /* 112 */
3129 7, /* 120 */
3130 7, /* 128 */
3131 2, /* 136 */
3132 2, /* 144 */
3133 2, /* 152 */
3134 2, /* 160 */
3135 2, /* 168 */
3136 2, /* 176 */
3137 2, /* 184 */
3138 2 /* 192 */
3139};
3140
3141static inline int size_index_elem(size_t bytes)
3142{
3143 return (bytes - 1) / 8;
3144}
3145
3146static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3147{
3148 int index;
3149
3150 if (size <= 192) {
3151 if (!size)
3152 return ZERO_SIZE_PTR;
3153
3154 index = size_index[size_index_elem(size)];
3155 } else
3156 index = fls(size - 1);
3157
3158#ifdef CONFIG_ZONE_DMA
3159 if (unlikely((flags & SLUB_DMA)))
3160 return kmalloc_dma_caches[index];
3161
3162#endif
3163 return kmalloc_caches[index];
3164}
3165
3166void *__kmalloc(size_t size, gfp_t flags)
3167{
3168 struct kmem_cache *s;
3169 void *ret;
3170
3171 if (unlikely(size > SLUB_MAX_SIZE))
3172 return kmalloc_large(size, flags);
3173
3174 s = get_slab(size, flags);
3175
3176 if (unlikely(ZERO_OR_NULL_PTR(s)))
3177 return s;
3178
3179 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3180
3181 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3182
3183 return ret;
3184}
3185EXPORT_SYMBOL(__kmalloc);
3186
3187#ifdef CONFIG_NUMA
3188static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3189{
3190 struct page *page;
3191 void *ptr = NULL;
3192
3193 flags |= __GFP_COMP | __GFP_NOTRACK;
3194 page = alloc_pages_node(node, flags, get_order(size));
3195 if (page)
3196 ptr = page_address(page);
3197
3198 kmemleak_alloc(ptr, size, 1, flags);
3199 return ptr;
3200}
3201
3202void *__kmalloc_node(size_t size, gfp_t flags, int node)
3203{
3204 struct kmem_cache *s;
3205 void *ret;
3206
3207 if (unlikely(size > SLUB_MAX_SIZE)) {
3208 ret = kmalloc_large_node(size, flags, node);
3209
3210 trace_kmalloc_node(_RET_IP_, ret,
3211 size, PAGE_SIZE << get_order(size),
3212 flags, node);
3213
3214 return ret;
3215 }
3216
3217 s = get_slab(size, flags);
3218
3219 if (unlikely(ZERO_OR_NULL_PTR(s)))
3220 return s;
3221
3222 ret = slab_alloc(s, flags, node, _RET_IP_);
3223
3224 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3225
3226 return ret;
3227}
3228EXPORT_SYMBOL(__kmalloc_node);
3229#endif
3230
3231size_t ksize(const void *object)
3232{
3233 struct page *page;
3234
3235 if (unlikely(object == ZERO_SIZE_PTR))
3236 return 0;
3237
3238 page = virt_to_head_page(object);
3239
3240 if (unlikely(!PageSlab(page))) {
3241 WARN_ON(!PageCompound(page));
3242 return PAGE_SIZE << compound_order(page);
3243 }
3244
3245 return slab_ksize(page->slab);
3246}
3247EXPORT_SYMBOL(ksize);
3248
3249#ifdef CONFIG_SLUB_DEBUG
3250bool verify_mem_not_deleted(const void *x)
3251{
3252 struct page *page;
3253 void *object = (void *)x;
3254 unsigned long flags;
3255 bool rv;
3256
3257 if (unlikely(ZERO_OR_NULL_PTR(x)))
3258 return false;
3259
3260 local_irq_save(flags);
3261
3262 page = virt_to_head_page(x);
3263 if (unlikely(!PageSlab(page))) {
3264 /* maybe it was from stack? */
3265 rv = true;
3266 goto out_unlock;
3267 }
3268
3269 slab_lock(page);
3270 if (on_freelist(page->slab, page, object)) {
3271 object_err(page->slab, page, object, "Object is on free-list");
3272 rv = false;
3273 } else {
3274 rv = true;
3275 }
3276 slab_unlock(page);
3277
3278out_unlock:
3279 local_irq_restore(flags);
3280 return rv;
3281}
3282EXPORT_SYMBOL(verify_mem_not_deleted);
3283#endif
3284
3285void kfree(const void *x)
3286{
3287 struct page *page;
3288 void *object = (void *)x;
3289
3290 trace_kfree(_RET_IP_, x);
3291
3292 if (unlikely(ZERO_OR_NULL_PTR(x)))
3293 return;
3294
3295 page = virt_to_head_page(x);
3296 if (unlikely(!PageSlab(page))) {
3297 BUG_ON(!PageCompound(page));
3298 kmemleak_free(x);
3299 put_page(page);
3300 return;
3301 }
3302 slab_free(page->slab, page, object, _RET_IP_);
3303}
3304EXPORT_SYMBOL(kfree);
3305
3306/*
3307 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3308 * the remaining slabs by the number of items in use. The slabs with the
3309 * most items in use come first. New allocations will then fill those up
3310 * and thus they can be removed from the partial lists.
3311 *
3312 * The slabs with the least items are placed last. This results in them
3313 * being allocated from last increasing the chance that the last objects
3314 * are freed in them.
3315 */
3316int kmem_cache_shrink(struct kmem_cache *s)
3317{
3318 int node;
3319 int i;
3320 struct kmem_cache_node *n;
3321 struct page *page;
3322 struct page *t;
3323 int objects = oo_objects(s->max);
3324 struct list_head *slabs_by_inuse =
3325 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3326 unsigned long flags;
3327
3328 if (!slabs_by_inuse)
3329 return -ENOMEM;
3330
3331 flush_all(s);
3332 for_each_node_state(node, N_NORMAL_MEMORY) {
3333 n = get_node(s, node);
3334
3335 if (!n->nr_partial)
3336 continue;
3337
3338 for (i = 0; i < objects; i++)
3339 INIT_LIST_HEAD(slabs_by_inuse + i);
3340
3341 spin_lock_irqsave(&n->list_lock, flags);
3342
3343 /*
3344 * Build lists indexed by the items in use in each slab.
3345 *
3346 * Note that concurrent frees may occur while we hold the
3347 * list_lock. page->inuse here is the upper limit.
3348 */
3349 list_for_each_entry_safe(page, t, &n->partial, lru) {
3350 if (!page->inuse) {
3351 remove_partial(n, page);
3352 discard_slab(s, page);
3353 } else {
3354 list_move(&page->lru,
3355 slabs_by_inuse + page->inuse);
3356 }
3357 }
3358
3359 /*
3360 * Rebuild the partial list with the slabs filled up most
3361 * first and the least used slabs at the end.
3362 */
3363 for (i = objects - 1; i >= 0; i--)
3364 list_splice(slabs_by_inuse + i, n->partial.prev);
3365
3366 spin_unlock_irqrestore(&n->list_lock, flags);
3367 }
3368
3369 kfree(slabs_by_inuse);
3370 return 0;
3371}
3372EXPORT_SYMBOL(kmem_cache_shrink);
3373
3374#if defined(CONFIG_MEMORY_HOTPLUG)
3375static int slab_mem_going_offline_callback(void *arg)
3376{
3377 struct kmem_cache *s;
3378
3379 down_read(&slub_lock);
3380 list_for_each_entry(s, &slab_caches, list)
3381 kmem_cache_shrink(s);
3382 up_read(&slub_lock);
3383
3384 return 0;
3385}
3386
3387static void slab_mem_offline_callback(void *arg)
3388{
3389 struct kmem_cache_node *n;
3390 struct kmem_cache *s;
3391 struct memory_notify *marg = arg;
3392 int offline_node;
3393
3394 offline_node = marg->status_change_nid;
3395
3396 /*
3397 * If the node still has available memory. we need kmem_cache_node
3398 * for it yet.
3399 */
3400 if (offline_node < 0)
3401 return;
3402
3403 down_read(&slub_lock);
3404 list_for_each_entry(s, &slab_caches, list) {
3405 n = get_node(s, offline_node);
3406 if (n) {
3407 /*
3408 * if n->nr_slabs > 0, slabs still exist on the node
3409 * that is going down. We were unable to free them,
3410 * and offline_pages() function shouldn't call this
3411 * callback. So, we must fail.
3412 */
3413 BUG_ON(slabs_node(s, offline_node));
3414
3415 s->node[offline_node] = NULL;
3416 kmem_cache_free(kmem_cache_node, n);
3417 }
3418 }
3419 up_read(&slub_lock);
3420}
3421
3422static int slab_mem_going_online_callback(void *arg)
3423{
3424 struct kmem_cache_node *n;
3425 struct kmem_cache *s;
3426 struct memory_notify *marg = arg;
3427 int nid = marg->status_change_nid;
3428 int ret = 0;
3429
3430 /*
3431 * If the node's memory is already available, then kmem_cache_node is
3432 * already created. Nothing to do.
3433 */
3434 if (nid < 0)
3435 return 0;
3436
3437 /*
3438 * We are bringing a node online. No memory is available yet. We must
3439 * allocate a kmem_cache_node structure in order to bring the node
3440 * online.
3441 */
3442 down_read(&slub_lock);
3443 list_for_each_entry(s, &slab_caches, list) {
3444 /*
3445 * XXX: kmem_cache_alloc_node will fallback to other nodes
3446 * since memory is not yet available from the node that
3447 * is brought up.
3448 */
3449 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3450 if (!n) {
3451 ret = -ENOMEM;
3452 goto out;
3453 }
3454 init_kmem_cache_node(n, s);
3455 s->node[nid] = n;
3456 }
3457out:
3458 up_read(&slub_lock);
3459 return ret;
3460}
3461
3462static int slab_memory_callback(struct notifier_block *self,
3463 unsigned long action, void *arg)
3464{
3465 int ret = 0;
3466
3467 switch (action) {
3468 case MEM_GOING_ONLINE:
3469 ret = slab_mem_going_online_callback(arg);
3470 break;
3471 case MEM_GOING_OFFLINE:
3472 ret = slab_mem_going_offline_callback(arg);
3473 break;
3474 case MEM_OFFLINE:
3475 case MEM_CANCEL_ONLINE:
3476 slab_mem_offline_callback(arg);
3477 break;
3478 case MEM_ONLINE:
3479 case MEM_CANCEL_OFFLINE:
3480 break;
3481 }
3482 if (ret)
3483 ret = notifier_from_errno(ret);
3484 else
3485 ret = NOTIFY_OK;
3486 return ret;
3487}
3488
3489#endif /* CONFIG_MEMORY_HOTPLUG */
3490
3491/********************************************************************
3492 * Basic setup of slabs
3493 *******************************************************************/
3494
3495/*
3496 * Used for early kmem_cache structures that were allocated using
3497 * the page allocator
3498 */
3499
3500static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3501{
3502 int node;
3503
3504 list_add(&s->list, &slab_caches);
3505 s->refcount = -1;
3506
3507 for_each_node_state(node, N_NORMAL_MEMORY) {
3508 struct kmem_cache_node *n = get_node(s, node);
3509 struct page *p;
3510
3511 if (n) {
3512 list_for_each_entry(p, &n->partial, lru)
3513 p->slab = s;
3514
3515#ifdef CONFIG_SLUB_DEBUG
3516 list_for_each_entry(p, &n->full, lru)
3517 p->slab = s;
3518#endif
3519 }
3520 }
3521}
3522
3523void __init kmem_cache_init(void)
3524{
3525 int i;
3526 int caches = 0;
3527 struct kmem_cache *temp_kmem_cache;
3528 int order;
3529 struct kmem_cache *temp_kmem_cache_node;
3530 unsigned long kmalloc_size;
3531
3532 kmem_size = offsetof(struct kmem_cache, node) +
3533 nr_node_ids * sizeof(struct kmem_cache_node *);
3534
3535 /* Allocate two kmem_caches from the page allocator */
3536 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3537 order = get_order(2 * kmalloc_size);
3538 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3539
3540 /*
3541 * Must first have the slab cache available for the allocations of the
3542 * struct kmem_cache_node's. There is special bootstrap code in
3543 * kmem_cache_open for slab_state == DOWN.
3544 */
3545 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3546
3547 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3548 sizeof(struct kmem_cache_node),
3549 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3550
3551 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3552
3553 /* Able to allocate the per node structures */
3554 slab_state = PARTIAL;
3555
3556 temp_kmem_cache = kmem_cache;
3557 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3558 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3559 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3560 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3561
3562 /*
3563 * Allocate kmem_cache_node properly from the kmem_cache slab.
3564 * kmem_cache_node is separately allocated so no need to
3565 * update any list pointers.
3566 */
3567 temp_kmem_cache_node = kmem_cache_node;
3568
3569 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3570 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3571
3572 kmem_cache_bootstrap_fixup(kmem_cache_node);
3573
3574 caches++;
3575 kmem_cache_bootstrap_fixup(kmem_cache);
3576 caches++;
3577 /* Free temporary boot structure */
3578 free_pages((unsigned long)temp_kmem_cache, order);
3579
3580 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3581
3582 /*
3583 * Patch up the size_index table if we have strange large alignment
3584 * requirements for the kmalloc array. This is only the case for
3585 * MIPS it seems. The standard arches will not generate any code here.
3586 *
3587 * Largest permitted alignment is 256 bytes due to the way we
3588 * handle the index determination for the smaller caches.
3589 *
3590 * Make sure that nothing crazy happens if someone starts tinkering
3591 * around with ARCH_KMALLOC_MINALIGN
3592 */
3593 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3594 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3595
3596 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3597 int elem = size_index_elem(i);
3598 if (elem >= ARRAY_SIZE(size_index))
3599 break;
3600 size_index[elem] = KMALLOC_SHIFT_LOW;
3601 }
3602
3603 if (KMALLOC_MIN_SIZE == 64) {
3604 /*
3605 * The 96 byte size cache is not used if the alignment
3606 * is 64 byte.
3607 */
3608 for (i = 64 + 8; i <= 96; i += 8)
3609 size_index[size_index_elem(i)] = 7;
3610 } else if (KMALLOC_MIN_SIZE == 128) {
3611 /*
3612 * The 192 byte sized cache is not used if the alignment
3613 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3614 * instead.
3615 */
3616 for (i = 128 + 8; i <= 192; i += 8)
3617 size_index[size_index_elem(i)] = 8;
3618 }
3619
3620 /* Caches that are not of the two-to-the-power-of size */
3621 if (KMALLOC_MIN_SIZE <= 32) {
3622 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3623 caches++;
3624 }
3625
3626 if (KMALLOC_MIN_SIZE <= 64) {
3627 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3628 caches++;
3629 }
3630
3631 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3632 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3633 caches++;
3634 }
3635
3636 slab_state = UP;
3637
3638 /* Provide the correct kmalloc names now that the caches are up */
3639 if (KMALLOC_MIN_SIZE <= 32) {
3640 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3641 BUG_ON(!kmalloc_caches[1]->name);
3642 }
3643
3644 if (KMALLOC_MIN_SIZE <= 64) {
3645 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3646 BUG_ON(!kmalloc_caches[2]->name);
3647 }
3648
3649 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3650 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3651
3652 BUG_ON(!s);
3653 kmalloc_caches[i]->name = s;
3654 }
3655
3656#ifdef CONFIG_SMP
3657 register_cpu_notifier(&slab_notifier);
3658#endif
3659
3660#ifdef CONFIG_ZONE_DMA
3661 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3662 struct kmem_cache *s = kmalloc_caches[i];
3663
3664 if (s && s->size) {
3665 char *name = kasprintf(GFP_NOWAIT,
3666 "dma-kmalloc-%d", s->objsize);
3667
3668 BUG_ON(!name);
3669 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3670 s->objsize, SLAB_CACHE_DMA);
3671 }
3672 }
3673#endif
3674 printk(KERN_INFO
3675 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3676 " CPUs=%d, Nodes=%d\n",
3677 caches, cache_line_size(),
3678 slub_min_order, slub_max_order, slub_min_objects,
3679 nr_cpu_ids, nr_node_ids);
3680}
3681
3682void __init kmem_cache_init_late(void)
3683{
3684}
3685
3686/*
3687 * Find a mergeable slab cache
3688 */
3689static int slab_unmergeable(struct kmem_cache *s)
3690{
3691 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3692 return 1;
3693
3694 if (s->ctor)
3695 return 1;
3696
3697 /*
3698 * We may have set a slab to be unmergeable during bootstrap.
3699 */
3700 if (s->refcount < 0)
3701 return 1;
3702
3703 return 0;
3704}
3705
3706static struct kmem_cache *find_mergeable(size_t size,
3707 size_t align, unsigned long flags, const char *name,
3708 void (*ctor)(void *))
3709{
3710 struct kmem_cache *s;
3711
3712 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3713 return NULL;
3714
3715 if (ctor)
3716 return NULL;
3717
3718 size = ALIGN(size, sizeof(void *));
3719 align = calculate_alignment(flags, align, size);
3720 size = ALIGN(size, align);
3721 flags = kmem_cache_flags(size, flags, name, NULL);
3722
3723 list_for_each_entry(s, &slab_caches, list) {
3724 if (slab_unmergeable(s))
3725 continue;
3726
3727 if (size > s->size)
3728 continue;
3729
3730 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3731 continue;
3732 /*
3733 * Check if alignment is compatible.
3734 * Courtesy of Adrian Drzewiecki
3735 */
3736 if ((s->size & ~(align - 1)) != s->size)
3737 continue;
3738
3739 if (s->size - size >= sizeof(void *))
3740 continue;
3741
3742 return s;
3743 }
3744 return NULL;
3745}
3746
3747struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3748 size_t align, unsigned long flags, void (*ctor)(void *))
3749{
3750 struct kmem_cache *s;
3751 char *n;
3752
3753 if (WARN_ON(!name))
3754 return NULL;
3755
3756 down_write(&slub_lock);
3757 s = find_mergeable(size, align, flags, name, ctor);
3758 if (s) {
3759 s->refcount++;
3760 /*
3761 * Adjust the object sizes so that we clear
3762 * the complete object on kzalloc.
3763 */
3764 s->objsize = max(s->objsize, (int)size);
3765 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3766
3767 if (sysfs_slab_alias(s, name)) {
3768 s->refcount--;
3769 goto err;
3770 }
3771 up_write(&slub_lock);
3772 return s;
3773 }
3774
3775 n = kstrdup(name, GFP_KERNEL);
3776 if (!n)
3777 goto err;
3778
3779 s = kmalloc(kmem_size, GFP_KERNEL);
3780 if (s) {
3781 if (kmem_cache_open(s, n,
3782 size, align, flags, ctor)) {
3783 list_add(&s->list, &slab_caches);
3784 if (sysfs_slab_add(s)) {
3785 list_del(&s->list);
3786 kfree(n);
3787 kfree(s);
3788 goto err;
3789 }
3790 up_write(&slub_lock);
3791 return s;
3792 }
3793 kfree(n);
3794 kfree(s);
3795 }
3796err:
3797 up_write(&slub_lock);
3798
3799 if (flags & SLAB_PANIC)
3800 panic("Cannot create slabcache %s\n", name);
3801 else
3802 s = NULL;
3803 return s;
3804}
3805EXPORT_SYMBOL(kmem_cache_create);
3806
3807#ifdef CONFIG_SMP
3808/*
3809 * Use the cpu notifier to insure that the cpu slabs are flushed when
3810 * necessary.
3811 */
3812static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3813 unsigned long action, void *hcpu)
3814{
3815 long cpu = (long)hcpu;
3816 struct kmem_cache *s;
3817 unsigned long flags;
3818
3819 switch (action) {
3820 case CPU_UP_CANCELED:
3821 case CPU_UP_CANCELED_FROZEN:
3822 case CPU_DEAD:
3823 case CPU_DEAD_FROZEN:
3824 down_read(&slub_lock);
3825 list_for_each_entry(s, &slab_caches, list) {
3826 local_irq_save(flags);
3827 __flush_cpu_slab(s, cpu);
3828 local_irq_restore(flags);
3829 }
3830 up_read(&slub_lock);
3831 break;
3832 default:
3833 break;
3834 }
3835 return NOTIFY_OK;
3836}
3837
3838static struct notifier_block __cpuinitdata slab_notifier = {
3839 .notifier_call = slab_cpuup_callback
3840};
3841
3842#endif
3843
3844void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3845{
3846 struct kmem_cache *s;
3847 void *ret;
3848
3849 if (unlikely(size > SLUB_MAX_SIZE))
3850 return kmalloc_large(size, gfpflags);
3851
3852 s = get_slab(size, gfpflags);
3853
3854 if (unlikely(ZERO_OR_NULL_PTR(s)))
3855 return s;
3856
3857 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3858
3859 /* Honor the call site pointer we received. */
3860 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3861
3862 return ret;
3863}
3864
3865#ifdef CONFIG_NUMA
3866void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3867 int node, unsigned long caller)
3868{
3869 struct kmem_cache *s;
3870 void *ret;
3871
3872 if (unlikely(size > SLUB_MAX_SIZE)) {
3873 ret = kmalloc_large_node(size, gfpflags, node);
3874
3875 trace_kmalloc_node(caller, ret,
3876 size, PAGE_SIZE << get_order(size),
3877 gfpflags, node);
3878
3879 return ret;
3880 }
3881
3882 s = get_slab(size, gfpflags);
3883
3884 if (unlikely(ZERO_OR_NULL_PTR(s)))
3885 return s;
3886
3887 ret = slab_alloc(s, gfpflags, node, caller);
3888
3889 /* Honor the call site pointer we received. */
3890 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3891
3892 return ret;
3893}
3894#endif
3895
3896#ifdef CONFIG_SYSFS
3897static int count_inuse(struct page *page)
3898{
3899 return page->inuse;
3900}
3901
3902static int count_total(struct page *page)
3903{
3904 return page->objects;
3905}
3906#endif
3907
3908#ifdef CONFIG_SLUB_DEBUG
3909static int validate_slab(struct kmem_cache *s, struct page *page,
3910 unsigned long *map)
3911{
3912 void *p;
3913 void *addr = page_address(page);
3914
3915 if (!check_slab(s, page) ||
3916 !on_freelist(s, page, NULL))
3917 return 0;
3918
3919 /* Now we know that a valid freelist exists */
3920 bitmap_zero(map, page->objects);
3921
3922 get_map(s, page, map);
3923 for_each_object(p, s, addr, page->objects) {
3924 if (test_bit(slab_index(p, s, addr), map))
3925 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3926 return 0;
3927 }
3928
3929 for_each_object(p, s, addr, page->objects)
3930 if (!test_bit(slab_index(p, s, addr), map))
3931 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3932 return 0;
3933 return 1;
3934}
3935
3936static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3937 unsigned long *map)
3938{
3939 slab_lock(page);
3940 validate_slab(s, page, map);
3941 slab_unlock(page);
3942}
3943
3944static int validate_slab_node(struct kmem_cache *s,
3945 struct kmem_cache_node *n, unsigned long *map)
3946{
3947 unsigned long count = 0;
3948 struct page *page;
3949 unsigned long flags;
3950
3951 spin_lock_irqsave(&n->list_lock, flags);
3952
3953 list_for_each_entry(page, &n->partial, lru) {
3954 validate_slab_slab(s, page, map);
3955 count++;
3956 }
3957 if (count != n->nr_partial)
3958 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3959 "counter=%ld\n", s->name, count, n->nr_partial);
3960
3961 if (!(s->flags & SLAB_STORE_USER))
3962 goto out;
3963
3964 list_for_each_entry(page, &n->full, lru) {
3965 validate_slab_slab(s, page, map);
3966 count++;
3967 }
3968 if (count != atomic_long_read(&n->nr_slabs))
3969 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3970 "counter=%ld\n", s->name, count,
3971 atomic_long_read(&n->nr_slabs));
3972
3973out:
3974 spin_unlock_irqrestore(&n->list_lock, flags);
3975 return count;
3976}
3977
3978static long validate_slab_cache(struct kmem_cache *s)
3979{
3980 int node;
3981 unsigned long count = 0;
3982 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3983 sizeof(unsigned long), GFP_KERNEL);
3984
3985 if (!map)
3986 return -ENOMEM;
3987
3988 flush_all(s);
3989 for_each_node_state(node, N_NORMAL_MEMORY) {
3990 struct kmem_cache_node *n = get_node(s, node);
3991
3992 count += validate_slab_node(s, n, map);
3993 }
3994 kfree(map);
3995 return count;
3996}
3997/*
3998 * Generate lists of code addresses where slabcache objects are allocated
3999 * and freed.
4000 */
4001
4002struct location {
4003 unsigned long count;
4004 unsigned long addr;
4005 long long sum_time;
4006 long min_time;
4007 long max_time;
4008 long min_pid;
4009 long max_pid;
4010 DECLARE_BITMAP(cpus, NR_CPUS);
4011 nodemask_t nodes;
4012};
4013
4014struct loc_track {
4015 unsigned long max;
4016 unsigned long count;
4017 struct location *loc;
4018};
4019
4020static void free_loc_track(struct loc_track *t)
4021{
4022 if (t->max)
4023 free_pages((unsigned long)t->loc,
4024 get_order(sizeof(struct location) * t->max));
4025}
4026
4027static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4028{
4029 struct location *l;
4030 int order;
4031
4032 order = get_order(sizeof(struct location) * max);
4033
4034 l = (void *)__get_free_pages(flags, order);
4035 if (!l)
4036 return 0;
4037
4038 if (t->count) {
4039 memcpy(l, t->loc, sizeof(struct location) * t->count);
4040 free_loc_track(t);
4041 }
4042 t->max = max;
4043 t->loc = l;
4044 return 1;
4045}
4046
4047static int add_location(struct loc_track *t, struct kmem_cache *s,
4048 const struct track *track)
4049{
4050 long start, end, pos;
4051 struct location *l;
4052 unsigned long caddr;
4053 unsigned long age = jiffies - track->when;
4054
4055 start = -1;
4056 end = t->count;
4057
4058 for ( ; ; ) {
4059 pos = start + (end - start + 1) / 2;
4060
4061 /*
4062 * There is nothing at "end". If we end up there
4063 * we need to add something to before end.
4064 */
4065 if (pos == end)
4066 break;
4067
4068 caddr = t->loc[pos].addr;
4069 if (track->addr == caddr) {
4070
4071 l = &t->loc[pos];
4072 l->count++;
4073 if (track->when) {
4074 l->sum_time += age;
4075 if (age < l->min_time)
4076 l->min_time = age;
4077 if (age > l->max_time)
4078 l->max_time = age;
4079
4080 if (track->pid < l->min_pid)
4081 l->min_pid = track->pid;
4082 if (track->pid > l->max_pid)
4083 l->max_pid = track->pid;
4084
4085 cpumask_set_cpu(track->cpu,
4086 to_cpumask(l->cpus));
4087 }
4088 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4089 return 1;
4090 }
4091
4092 if (track->addr < caddr)
4093 end = pos;
4094 else
4095 start = pos;
4096 }
4097
4098 /*
4099 * Not found. Insert new tracking element.
4100 */
4101 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4102 return 0;
4103
4104 l = t->loc + pos;
4105 if (pos < t->count)
4106 memmove(l + 1, l,
4107 (t->count - pos) * sizeof(struct location));
4108 t->count++;
4109 l->count = 1;
4110 l->addr = track->addr;
4111 l->sum_time = age;
4112 l->min_time = age;
4113 l->max_time = age;
4114 l->min_pid = track->pid;
4115 l->max_pid = track->pid;
4116 cpumask_clear(to_cpumask(l->cpus));
4117 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4118 nodes_clear(l->nodes);
4119 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4120 return 1;
4121}
4122
4123static void process_slab(struct loc_track *t, struct kmem_cache *s,
4124 struct page *page, enum track_item alloc,
4125 unsigned long *map)
4126{
4127 void *addr = page_address(page);
4128 void *p;
4129
4130 bitmap_zero(map, page->objects);
4131 get_map(s, page, map);
4132
4133 for_each_object(p, s, addr, page->objects)
4134 if (!test_bit(slab_index(p, s, addr), map))
4135 add_location(t, s, get_track(s, p, alloc));
4136}
4137
4138static int list_locations(struct kmem_cache *s, char *buf,
4139 enum track_item alloc)
4140{
4141 int len = 0;
4142 unsigned long i;
4143 struct loc_track t = { 0, 0, NULL };
4144 int node;
4145 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4146 sizeof(unsigned long), GFP_KERNEL);
4147
4148 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4149 GFP_TEMPORARY)) {
4150 kfree(map);
4151 return sprintf(buf, "Out of memory\n");
4152 }
4153 /* Push back cpu slabs */
4154 flush_all(s);
4155
4156 for_each_node_state(node, N_NORMAL_MEMORY) {
4157 struct kmem_cache_node *n = get_node(s, node);
4158 unsigned long flags;
4159 struct page *page;
4160
4161 if (!atomic_long_read(&n->nr_slabs))
4162 continue;
4163
4164 spin_lock_irqsave(&n->list_lock, flags);
4165 list_for_each_entry(page, &n->partial, lru)
4166 process_slab(&t, s, page, alloc, map);
4167 list_for_each_entry(page, &n->full, lru)
4168 process_slab(&t, s, page, alloc, map);
4169 spin_unlock_irqrestore(&n->list_lock, flags);
4170 }
4171
4172 for (i = 0; i < t.count; i++) {
4173 struct location *l = &t.loc[i];
4174
4175 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4176 break;
4177 len += sprintf(buf + len, "%7ld ", l->count);
4178
4179 if (l->addr)
4180 len += sprintf(buf + len, "%pS", (void *)l->addr);
4181 else
4182 len += sprintf(buf + len, "<not-available>");
4183
4184 if (l->sum_time != l->min_time) {
4185 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4186 l->min_time,
4187 (long)div_u64(l->sum_time, l->count),
4188 l->max_time);
4189 } else
4190 len += sprintf(buf + len, " age=%ld",
4191 l->min_time);
4192
4193 if (l->min_pid != l->max_pid)
4194 len += sprintf(buf + len, " pid=%ld-%ld",
4195 l->min_pid, l->max_pid);
4196 else
4197 len += sprintf(buf + len, " pid=%ld",
4198 l->min_pid);
4199
4200 if (num_online_cpus() > 1 &&
4201 !cpumask_empty(to_cpumask(l->cpus)) &&
4202 len < PAGE_SIZE - 60) {
4203 len += sprintf(buf + len, " cpus=");
4204 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4205 to_cpumask(l->cpus));
4206 }
4207
4208 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4209 len < PAGE_SIZE - 60) {
4210 len += sprintf(buf + len, " nodes=");
4211 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4212 l->nodes);
4213 }
4214
4215 len += sprintf(buf + len, "\n");
4216 }
4217
4218 free_loc_track(&t);
4219 kfree(map);
4220 if (!t.count)
4221 len += sprintf(buf, "No data\n");
4222 return len;
4223}
4224#endif
4225
4226#ifdef SLUB_RESILIENCY_TEST
4227static void resiliency_test(void)
4228{
4229 u8 *p;
4230
4231 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4232
4233 printk(KERN_ERR "SLUB resiliency testing\n");
4234 printk(KERN_ERR "-----------------------\n");
4235 printk(KERN_ERR "A. Corruption after allocation\n");
4236
4237 p = kzalloc(16, GFP_KERNEL);
4238 p[16] = 0x12;
4239 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4240 " 0x12->0x%p\n\n", p + 16);
4241
4242 validate_slab_cache(kmalloc_caches[4]);
4243
4244 /* Hmmm... The next two are dangerous */
4245 p = kzalloc(32, GFP_KERNEL);
4246 p[32 + sizeof(void *)] = 0x34;
4247 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4248 " 0x34 -> -0x%p\n", p);
4249 printk(KERN_ERR
4250 "If allocated object is overwritten then not detectable\n\n");
4251
4252 validate_slab_cache(kmalloc_caches[5]);
4253 p = kzalloc(64, GFP_KERNEL);
4254 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4255 *p = 0x56;
4256 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4257 p);
4258 printk(KERN_ERR
4259 "If allocated object is overwritten then not detectable\n\n");
4260 validate_slab_cache(kmalloc_caches[6]);
4261
4262 printk(KERN_ERR "\nB. Corruption after free\n");
4263 p = kzalloc(128, GFP_KERNEL);
4264 kfree(p);
4265 *p = 0x78;
4266 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4267 validate_slab_cache(kmalloc_caches[7]);
4268
4269 p = kzalloc(256, GFP_KERNEL);
4270 kfree(p);
4271 p[50] = 0x9a;
4272 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4273 p);
4274 validate_slab_cache(kmalloc_caches[8]);
4275
4276 p = kzalloc(512, GFP_KERNEL);
4277 kfree(p);
4278 p[512] = 0xab;
4279 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4280 validate_slab_cache(kmalloc_caches[9]);
4281}
4282#else
4283#ifdef CONFIG_SYSFS
4284static void resiliency_test(void) {};
4285#endif
4286#endif
4287
4288#ifdef CONFIG_SYSFS
4289enum slab_stat_type {
4290 SL_ALL, /* All slabs */
4291 SL_PARTIAL, /* Only partially allocated slabs */
4292 SL_CPU, /* Only slabs used for cpu caches */
4293 SL_OBJECTS, /* Determine allocated objects not slabs */
4294 SL_TOTAL /* Determine object capacity not slabs */
4295};
4296
4297#define SO_ALL (1 << SL_ALL)
4298#define SO_PARTIAL (1 << SL_PARTIAL)
4299#define SO_CPU (1 << SL_CPU)
4300#define SO_OBJECTS (1 << SL_OBJECTS)
4301#define SO_TOTAL (1 << SL_TOTAL)
4302
4303static ssize_t show_slab_objects(struct kmem_cache *s,
4304 char *buf, unsigned long flags)
4305{
4306 unsigned long total = 0;
4307 int node;
4308 int x;
4309 unsigned long *nodes;
4310 unsigned long *per_cpu;
4311
4312 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4313 if (!nodes)
4314 return -ENOMEM;
4315 per_cpu = nodes + nr_node_ids;
4316
4317 if (flags & SO_CPU) {
4318 int cpu;
4319
4320 for_each_possible_cpu(cpu) {
4321 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4322
4323 if (!c || c->node < 0)
4324 continue;
4325
4326 if (c->page) {
4327 if (flags & SO_TOTAL)
4328 x = c->page->objects;
4329 else if (flags & SO_OBJECTS)
4330 x = c->page->inuse;
4331 else
4332 x = 1;
4333
4334 total += x;
4335 nodes[c->node] += x;
4336 }
4337 per_cpu[c->node]++;
4338 }
4339 }
4340
4341 lock_memory_hotplug();
4342#ifdef CONFIG_SLUB_DEBUG
4343 if (flags & SO_ALL) {
4344 for_each_node_state(node, N_NORMAL_MEMORY) {
4345 struct kmem_cache_node *n = get_node(s, node);
4346
4347 if (flags & SO_TOTAL)
4348 x = atomic_long_read(&n->total_objects);
4349 else if (flags & SO_OBJECTS)
4350 x = atomic_long_read(&n->total_objects) -
4351 count_partial(n, count_free);
4352
4353 else
4354 x = atomic_long_read(&n->nr_slabs);
4355 total += x;
4356 nodes[node] += x;
4357 }
4358
4359 } else
4360#endif
4361 if (flags & SO_PARTIAL) {
4362 for_each_node_state(node, N_NORMAL_MEMORY) {
4363 struct kmem_cache_node *n = get_node(s, node);
4364
4365 if (flags & SO_TOTAL)
4366 x = count_partial(n, count_total);
4367 else if (flags & SO_OBJECTS)
4368 x = count_partial(n, count_inuse);
4369 else
4370 x = n->nr_partial;
4371 total += x;
4372 nodes[node] += x;
4373 }
4374 }
4375 x = sprintf(buf, "%lu", total);
4376#ifdef CONFIG_NUMA
4377 for_each_node_state(node, N_NORMAL_MEMORY)
4378 if (nodes[node])
4379 x += sprintf(buf + x, " N%d=%lu",
4380 node, nodes[node]);
4381#endif
4382 unlock_memory_hotplug();
4383 kfree(nodes);
4384 return x + sprintf(buf + x, "\n");
4385}
4386
4387#ifdef CONFIG_SLUB_DEBUG
4388static int any_slab_objects(struct kmem_cache *s)
4389{
4390 int node;
4391
4392 for_each_online_node(node) {
4393 struct kmem_cache_node *n = get_node(s, node);
4394
4395 if (!n)
4396 continue;
4397
4398 if (atomic_long_read(&n->total_objects))
4399 return 1;
4400 }
4401 return 0;
4402}
4403#endif
4404
4405#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4406#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4407
4408struct slab_attribute {
4409 struct attribute attr;
4410 ssize_t (*show)(struct kmem_cache *s, char *buf);
4411 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4412};
4413
4414#define SLAB_ATTR_RO(_name) \
4415 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4416
4417#define SLAB_ATTR(_name) \
4418 static struct slab_attribute _name##_attr = \
4419 __ATTR(_name, 0644, _name##_show, _name##_store)
4420
4421static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4422{
4423 return sprintf(buf, "%d\n", s->size);
4424}
4425SLAB_ATTR_RO(slab_size);
4426
4427static ssize_t align_show(struct kmem_cache *s, char *buf)
4428{
4429 return sprintf(buf, "%d\n", s->align);
4430}
4431SLAB_ATTR_RO(align);
4432
4433static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4434{
4435 return sprintf(buf, "%d\n", s->objsize);
4436}
4437SLAB_ATTR_RO(object_size);
4438
4439static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4440{
4441 return sprintf(buf, "%d\n", oo_objects(s->oo));
4442}
4443SLAB_ATTR_RO(objs_per_slab);
4444
4445static ssize_t order_store(struct kmem_cache *s,
4446 const char *buf, size_t length)
4447{
4448 unsigned long order;
4449 int err;
4450
4451 err = strict_strtoul(buf, 10, &order);
4452 if (err)
4453 return err;
4454
4455 if (order > slub_max_order || order < slub_min_order)
4456 return -EINVAL;
4457
4458 calculate_sizes(s, order);
4459 return length;
4460}
4461
4462static ssize_t order_show(struct kmem_cache *s, char *buf)
4463{
4464 return sprintf(buf, "%d\n", oo_order(s->oo));
4465}
4466SLAB_ATTR(order);
4467
4468static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4469{
4470 return sprintf(buf, "%lu\n", s->min_partial);
4471}
4472
4473static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4474 size_t length)
4475{
4476 unsigned long min;
4477 int err;
4478
4479 err = strict_strtoul(buf, 10, &min);
4480 if (err)
4481 return err;
4482
4483 set_min_partial(s, min);
4484 return length;
4485}
4486SLAB_ATTR(min_partial);
4487
4488static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4489{
4490 if (!s->ctor)
4491 return 0;
4492 return sprintf(buf, "%pS\n", s->ctor);
4493}
4494SLAB_ATTR_RO(ctor);
4495
4496static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4497{
4498 return sprintf(buf, "%d\n", s->refcount - 1);
4499}
4500SLAB_ATTR_RO(aliases);
4501
4502static ssize_t partial_show(struct kmem_cache *s, char *buf)
4503{
4504 return show_slab_objects(s, buf, SO_PARTIAL);
4505}
4506SLAB_ATTR_RO(partial);
4507
4508static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4509{
4510 return show_slab_objects(s, buf, SO_CPU);
4511}
4512SLAB_ATTR_RO(cpu_slabs);
4513
4514static ssize_t objects_show(struct kmem_cache *s, char *buf)
4515{
4516 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4517}
4518SLAB_ATTR_RO(objects);
4519
4520static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4521{
4522 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4523}
4524SLAB_ATTR_RO(objects_partial);
4525
4526static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4527{
4528 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4529}
4530
4531static ssize_t reclaim_account_store(struct kmem_cache *s,
4532 const char *buf, size_t length)
4533{
4534 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4535 if (buf[0] == '1')
4536 s->flags |= SLAB_RECLAIM_ACCOUNT;
4537 return length;
4538}
4539SLAB_ATTR(reclaim_account);
4540
4541static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4542{
4543 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4544}
4545SLAB_ATTR_RO(hwcache_align);
4546
4547#ifdef CONFIG_ZONE_DMA
4548static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4549{
4550 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4551}
4552SLAB_ATTR_RO(cache_dma);
4553#endif
4554
4555static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4556{
4557 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4558}
4559SLAB_ATTR_RO(destroy_by_rcu);
4560
4561static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4562{
4563 return sprintf(buf, "%d\n", s->reserved);
4564}
4565SLAB_ATTR_RO(reserved);
4566
4567#ifdef CONFIG_SLUB_DEBUG
4568static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4569{
4570 return show_slab_objects(s, buf, SO_ALL);
4571}
4572SLAB_ATTR_RO(slabs);
4573
4574static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4575{
4576 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4577}
4578SLAB_ATTR_RO(total_objects);
4579
4580static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4581{
4582 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4583}
4584
4585static ssize_t sanity_checks_store(struct kmem_cache *s,
4586 const char *buf, size_t length)
4587{
4588 s->flags &= ~SLAB_DEBUG_FREE;
4589 if (buf[0] == '1') {
4590 s->flags &= ~__CMPXCHG_DOUBLE;
4591 s->flags |= SLAB_DEBUG_FREE;
4592 }
4593 return length;
4594}
4595SLAB_ATTR(sanity_checks);
4596
4597static ssize_t trace_show(struct kmem_cache *s, char *buf)
4598{
4599 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4600}
4601
4602static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4603 size_t length)
4604{
4605 s->flags &= ~SLAB_TRACE;
4606 if (buf[0] == '1') {
4607 s->flags &= ~__CMPXCHG_DOUBLE;
4608 s->flags |= SLAB_TRACE;
4609 }
4610 return length;
4611}
4612SLAB_ATTR(trace);
4613
4614static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4615{
4616 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4617}
4618
4619static ssize_t red_zone_store(struct kmem_cache *s,
4620 const char *buf, size_t length)
4621{
4622 if (any_slab_objects(s))
4623 return -EBUSY;
4624
4625 s->flags &= ~SLAB_RED_ZONE;
4626 if (buf[0] == '1') {
4627 s->flags &= ~__CMPXCHG_DOUBLE;
4628 s->flags |= SLAB_RED_ZONE;
4629 }
4630 calculate_sizes(s, -1);
4631 return length;
4632}
4633SLAB_ATTR(red_zone);
4634
4635static ssize_t poison_show(struct kmem_cache *s, char *buf)
4636{
4637 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4638}
4639
4640static ssize_t poison_store(struct kmem_cache *s,
4641 const char *buf, size_t length)
4642{
4643 if (any_slab_objects(s))
4644 return -EBUSY;
4645
4646 s->flags &= ~SLAB_POISON;
4647 if (buf[0] == '1') {
4648 s->flags &= ~__CMPXCHG_DOUBLE;
4649 s->flags |= SLAB_POISON;
4650 }
4651 calculate_sizes(s, -1);
4652 return length;
4653}
4654SLAB_ATTR(poison);
4655
4656static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4657{
4658 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4659}
4660
4661static ssize_t store_user_store(struct kmem_cache *s,
4662 const char *buf, size_t length)
4663{
4664 if (any_slab_objects(s))
4665 return -EBUSY;
4666
4667 s->flags &= ~SLAB_STORE_USER;
4668 if (buf[0] == '1') {
4669 s->flags &= ~__CMPXCHG_DOUBLE;
4670 s->flags |= SLAB_STORE_USER;
4671 }
4672 calculate_sizes(s, -1);
4673 return length;
4674}
4675SLAB_ATTR(store_user);
4676
4677static ssize_t validate_show(struct kmem_cache *s, char *buf)
4678{
4679 return 0;
4680}
4681
4682static ssize_t validate_store(struct kmem_cache *s,
4683 const char *buf, size_t length)
4684{
4685 int ret = -EINVAL;
4686
4687 if (buf[0] == '1') {
4688 ret = validate_slab_cache(s);
4689 if (ret >= 0)
4690 ret = length;
4691 }
4692 return ret;
4693}
4694SLAB_ATTR(validate);
4695
4696static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4697{
4698 if (!(s->flags & SLAB_STORE_USER))
4699 return -ENOSYS;
4700 return list_locations(s, buf, TRACK_ALLOC);
4701}
4702SLAB_ATTR_RO(alloc_calls);
4703
4704static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4705{
4706 if (!(s->flags & SLAB_STORE_USER))
4707 return -ENOSYS;
4708 return list_locations(s, buf, TRACK_FREE);
4709}
4710SLAB_ATTR_RO(free_calls);
4711#endif /* CONFIG_SLUB_DEBUG */
4712
4713#ifdef CONFIG_FAILSLAB
4714static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4715{
4716 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4717}
4718
4719static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4720 size_t length)
4721{
4722 s->flags &= ~SLAB_FAILSLAB;
4723 if (buf[0] == '1')
4724 s->flags |= SLAB_FAILSLAB;
4725 return length;
4726}
4727SLAB_ATTR(failslab);
4728#endif
4729
4730static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4731{
4732 return 0;
4733}
4734
4735static ssize_t shrink_store(struct kmem_cache *s,
4736 const char *buf, size_t length)
4737{
4738 if (buf[0] == '1') {
4739 int rc = kmem_cache_shrink(s);
4740
4741 if (rc)
4742 return rc;
4743 } else
4744 return -EINVAL;
4745 return length;
4746}
4747SLAB_ATTR(shrink);
4748
4749#ifdef CONFIG_NUMA
4750static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4751{
4752 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4753}
4754
4755static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4756 const char *buf, size_t length)
4757{
4758 unsigned long ratio;
4759 int err;
4760
4761 err = strict_strtoul(buf, 10, &ratio);
4762 if (err)
4763 return err;
4764
4765 if (ratio <= 100)
4766 s->remote_node_defrag_ratio = ratio * 10;
4767
4768 return length;
4769}
4770SLAB_ATTR(remote_node_defrag_ratio);
4771#endif
4772
4773#ifdef CONFIG_SLUB_STATS
4774static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4775{
4776 unsigned long sum = 0;
4777 int cpu;
4778 int len;
4779 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4780
4781 if (!data)
4782 return -ENOMEM;
4783
4784 for_each_online_cpu(cpu) {
4785 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4786
4787 data[cpu] = x;
4788 sum += x;
4789 }
4790
4791 len = sprintf(buf, "%lu", sum);
4792
4793#ifdef CONFIG_SMP
4794 for_each_online_cpu(cpu) {
4795 if (data[cpu] && len < PAGE_SIZE - 20)
4796 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4797 }
4798#endif
4799 kfree(data);
4800 return len + sprintf(buf + len, "\n");
4801}
4802
4803static void clear_stat(struct kmem_cache *s, enum stat_item si)
4804{
4805 int cpu;
4806
4807 for_each_online_cpu(cpu)
4808 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4809}
4810
4811#define STAT_ATTR(si, text) \
4812static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4813{ \
4814 return show_stat(s, buf, si); \
4815} \
4816static ssize_t text##_store(struct kmem_cache *s, \
4817 const char *buf, size_t length) \
4818{ \
4819 if (buf[0] != '0') \
4820 return -EINVAL; \
4821 clear_stat(s, si); \
4822 return length; \
4823} \
4824SLAB_ATTR(text); \
4825
4826STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4827STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4828STAT_ATTR(FREE_FASTPATH, free_fastpath);
4829STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4830STAT_ATTR(FREE_FROZEN, free_frozen);
4831STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4832STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4833STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4834STAT_ATTR(ALLOC_SLAB, alloc_slab);
4835STAT_ATTR(ALLOC_REFILL, alloc_refill);
4836STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4837STAT_ATTR(FREE_SLAB, free_slab);
4838STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4839STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4840STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4841STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4842STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4843STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4844STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4845STAT_ATTR(ORDER_FALLBACK, order_fallback);
4846STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4847STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4848#endif
4849
4850static struct attribute *slab_attrs[] = {
4851 &slab_size_attr.attr,
4852 &object_size_attr.attr,
4853 &objs_per_slab_attr.attr,
4854 &order_attr.attr,
4855 &min_partial_attr.attr,
4856 &objects_attr.attr,
4857 &objects_partial_attr.attr,
4858 &partial_attr.attr,
4859 &cpu_slabs_attr.attr,
4860 &ctor_attr.attr,
4861 &aliases_attr.attr,
4862 &align_attr.attr,
4863 &hwcache_align_attr.attr,
4864 &reclaim_account_attr.attr,
4865 &destroy_by_rcu_attr.attr,
4866 &shrink_attr.attr,
4867 &reserved_attr.attr,
4868#ifdef CONFIG_SLUB_DEBUG
4869 &total_objects_attr.attr,
4870 &slabs_attr.attr,
4871 &sanity_checks_attr.attr,
4872 &trace_attr.attr,
4873 &red_zone_attr.attr,
4874 &poison_attr.attr,
4875 &store_user_attr.attr,
4876 &validate_attr.attr,
4877 &alloc_calls_attr.attr,
4878 &free_calls_attr.attr,
4879#endif
4880#ifdef CONFIG_ZONE_DMA
4881 &cache_dma_attr.attr,
4882#endif
4883#ifdef CONFIG_NUMA
4884 &remote_node_defrag_ratio_attr.attr,
4885#endif
4886#ifdef CONFIG_SLUB_STATS
4887 &alloc_fastpath_attr.attr,
4888 &alloc_slowpath_attr.attr,
4889 &free_fastpath_attr.attr,
4890 &free_slowpath_attr.attr,
4891 &free_frozen_attr.attr,
4892 &free_add_partial_attr.attr,
4893 &free_remove_partial_attr.attr,
4894 &alloc_from_partial_attr.attr,
4895 &alloc_slab_attr.attr,
4896 &alloc_refill_attr.attr,
4897 &alloc_node_mismatch_attr.attr,
4898 &free_slab_attr.attr,
4899 &cpuslab_flush_attr.attr,
4900 &deactivate_full_attr.attr,
4901 &deactivate_empty_attr.attr,
4902 &deactivate_to_head_attr.attr,
4903 &deactivate_to_tail_attr.attr,
4904 &deactivate_remote_frees_attr.attr,
4905 &deactivate_bypass_attr.attr,
4906 &order_fallback_attr.attr,
4907 &cmpxchg_double_fail_attr.attr,
4908 &cmpxchg_double_cpu_fail_attr.attr,
4909#endif
4910#ifdef CONFIG_FAILSLAB
4911 &failslab_attr.attr,
4912#endif
4913
4914 NULL
4915};
4916
4917static struct attribute_group slab_attr_group = {
4918 .attrs = slab_attrs,
4919};
4920
4921static ssize_t slab_attr_show(struct kobject *kobj,
4922 struct attribute *attr,
4923 char *buf)
4924{
4925 struct slab_attribute *attribute;
4926 struct kmem_cache *s;
4927 int err;
4928
4929 attribute = to_slab_attr(attr);
4930 s = to_slab(kobj);
4931
4932 if (!attribute->show)
4933 return -EIO;
4934
4935 err = attribute->show(s, buf);
4936
4937 return err;
4938}
4939
4940static ssize_t slab_attr_store(struct kobject *kobj,
4941 struct attribute *attr,
4942 const char *buf, size_t len)
4943{
4944 struct slab_attribute *attribute;
4945 struct kmem_cache *s;
4946 int err;
4947
4948 attribute = to_slab_attr(attr);
4949 s = to_slab(kobj);
4950
4951 if (!attribute->store)
4952 return -EIO;
4953
4954 err = attribute->store(s, buf, len);
4955
4956 return err;
4957}
4958
4959static void kmem_cache_release(struct kobject *kobj)
4960{
4961 struct kmem_cache *s = to_slab(kobj);
4962
4963 kfree(s->name);
4964 kfree(s);
4965}
4966
4967static const struct sysfs_ops slab_sysfs_ops = {
4968 .show = slab_attr_show,
4969 .store = slab_attr_store,
4970};
4971
4972static struct kobj_type slab_ktype = {
4973 .sysfs_ops = &slab_sysfs_ops,
4974 .release = kmem_cache_release
4975};
4976
4977static int uevent_filter(struct kset *kset, struct kobject *kobj)
4978{
4979 struct kobj_type *ktype = get_ktype(kobj);
4980
4981 if (ktype == &slab_ktype)
4982 return 1;
4983 return 0;
4984}
4985
4986static const struct kset_uevent_ops slab_uevent_ops = {
4987 .filter = uevent_filter,
4988};
4989
4990static struct kset *slab_kset;
4991
4992#define ID_STR_LENGTH 64
4993
4994/* Create a unique string id for a slab cache:
4995 *
4996 * Format :[flags-]size
4997 */
4998static char *create_unique_id(struct kmem_cache *s)
4999{
5000 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5001 char *p = name;
5002
5003 BUG_ON(!name);
5004
5005 *p++ = ':';
5006 /*
5007 * First flags affecting slabcache operations. We will only
5008 * get here for aliasable slabs so we do not need to support
5009 * too many flags. The flags here must cover all flags that
5010 * are matched during merging to guarantee that the id is
5011 * unique.
5012 */
5013 if (s->flags & SLAB_CACHE_DMA)
5014 *p++ = 'd';
5015 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5016 *p++ = 'a';
5017 if (s->flags & SLAB_DEBUG_FREE)
5018 *p++ = 'F';
5019 if (!(s->flags & SLAB_NOTRACK))
5020 *p++ = 't';
5021 if (p != name + 1)
5022 *p++ = '-';
5023 p += sprintf(p, "%07d", s->size);
5024 BUG_ON(p > name + ID_STR_LENGTH - 1);
5025 return name;
5026}
5027
5028static int sysfs_slab_add(struct kmem_cache *s)
5029{
5030 int err;
5031 const char *name;
5032 int unmergeable;
5033
5034 if (slab_state < SYSFS)
5035 /* Defer until later */
5036 return 0;
5037
5038 unmergeable = slab_unmergeable(s);
5039 if (unmergeable) {
5040 /*
5041 * Slabcache can never be merged so we can use the name proper.
5042 * This is typically the case for debug situations. In that
5043 * case we can catch duplicate names easily.
5044 */
5045 sysfs_remove_link(&slab_kset->kobj, s->name);
5046 name = s->name;
5047 } else {
5048 /*
5049 * Create a unique name for the slab as a target
5050 * for the symlinks.
5051 */
5052 name = create_unique_id(s);
5053 }
5054
5055 s->kobj.kset = slab_kset;
5056 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5057 if (err) {
5058 kobject_put(&s->kobj);
5059 return err;
5060 }
5061
5062 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5063 if (err) {
5064 kobject_del(&s->kobj);
5065 kobject_put(&s->kobj);
5066 return err;
5067 }
5068 kobject_uevent(&s->kobj, KOBJ_ADD);
5069 if (!unmergeable) {
5070 /* Setup first alias */
5071 sysfs_slab_alias(s, s->name);
5072 kfree(name);
5073 }
5074 return 0;
5075}
5076
5077static void sysfs_slab_remove(struct kmem_cache *s)
5078{
5079 if (slab_state < SYSFS)
5080 /*
5081 * Sysfs has not been setup yet so no need to remove the
5082 * cache from sysfs.
5083 */
5084 return;
5085
5086 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5087 kobject_del(&s->kobj);
5088 kobject_put(&s->kobj);
5089}
5090
5091/*
5092 * Need to buffer aliases during bootup until sysfs becomes
5093 * available lest we lose that information.
5094 */
5095struct saved_alias {
5096 struct kmem_cache *s;
5097 const char *name;
5098 struct saved_alias *next;
5099};
5100
5101static struct saved_alias *alias_list;
5102
5103static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5104{
5105 struct saved_alias *al;
5106
5107 if (slab_state == SYSFS) {
5108 /*
5109 * If we have a leftover link then remove it.
5110 */
5111 sysfs_remove_link(&slab_kset->kobj, name);
5112 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5113 }
5114
5115 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5116 if (!al)
5117 return -ENOMEM;
5118
5119 al->s = s;
5120 al->name = name;
5121 al->next = alias_list;
5122 alias_list = al;
5123 return 0;
5124}
5125
5126static int __init slab_sysfs_init(void)
5127{
5128 struct kmem_cache *s;
5129 int err;
5130
5131 down_write(&slub_lock);
5132
5133 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5134 if (!slab_kset) {
5135 up_write(&slub_lock);
5136 printk(KERN_ERR "Cannot register slab subsystem.\n");
5137 return -ENOSYS;
5138 }
5139
5140 slab_state = SYSFS;
5141
5142 list_for_each_entry(s, &slab_caches, list) {
5143 err = sysfs_slab_add(s);
5144 if (err)
5145 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5146 " to sysfs\n", s->name);
5147 }
5148
5149 while (alias_list) {
5150 struct saved_alias *al = alias_list;
5151
5152 alias_list = alias_list->next;
5153 err = sysfs_slab_alias(al->s, al->name);
5154 if (err)
5155 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5156 " %s to sysfs\n", s->name);
5157 kfree(al);
5158 }
5159
5160 up_write(&slub_lock);
5161 resiliency_test();
5162 return 0;
5163}
5164
5165__initcall(slab_sysfs_init);
5166#endif /* CONFIG_SYSFS */
5167
5168/*
5169 * The /proc/slabinfo ABI
5170 */
5171#ifdef CONFIG_SLABINFO
5172static void print_slabinfo_header(struct seq_file *m)
5173{
5174 seq_puts(m, "slabinfo - version: 2.1\n");
5175 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5176 "<objperslab> <pagesperslab>");
5177 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5178 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5179 seq_putc(m, '\n');
5180}
5181
5182static void *s_start(struct seq_file *m, loff_t *pos)
5183{
5184 loff_t n = *pos;
5185
5186 down_read(&slub_lock);
5187 if (!n)
5188 print_slabinfo_header(m);
5189
5190 return seq_list_start(&slab_caches, *pos);
5191}
5192
5193static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5194{
5195 return seq_list_next(p, &slab_caches, pos);
5196}
5197
5198static void s_stop(struct seq_file *m, void *p)
5199{
5200 up_read(&slub_lock);
5201}
5202
5203static int s_show(struct seq_file *m, void *p)
5204{
5205 unsigned long nr_partials = 0;
5206 unsigned long nr_slabs = 0;
5207 unsigned long nr_inuse = 0;
5208 unsigned long nr_objs = 0;
5209 unsigned long nr_free = 0;
5210 struct kmem_cache *s;
5211 int node;
5212
5213 s = list_entry(p, struct kmem_cache, list);
5214
5215 for_each_online_node(node) {
5216 struct kmem_cache_node *n = get_node(s, node);
5217
5218 if (!n)
5219 continue;
5220
5221 nr_partials += n->nr_partial;
5222 nr_slabs += atomic_long_read(&n->nr_slabs);
5223 nr_objs += atomic_long_read(&n->total_objects);
5224 nr_free += count_partial(n, count_free);
5225 }
5226
5227 nr_inuse = nr_objs - nr_free;
5228
5229 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5230 nr_objs, s->size, oo_objects(s->oo),
5231 (1 << oo_order(s->oo)));
5232 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5233 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5234 0UL);
5235 seq_putc(m, '\n');
5236 return 0;
5237}
5238
5239static const struct seq_operations slabinfo_op = {
5240 .start = s_start,
5241 .next = s_next,
5242 .stop = s_stop,
5243 .show = s_show,
5244};
5245
5246static int slabinfo_open(struct inode *inode, struct file *file)
5247{
5248 return seq_open(file, &slabinfo_op);
5249}
5250
5251static const struct file_operations proc_slabinfo_operations = {
5252 .open = slabinfo_open,
5253 .read = seq_read,
5254 .llseek = seq_lseek,
5255 .release = seq_release,
5256};
5257
5258static int __init slab_proc_init(void)
5259{
5260 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5261 return 0;
5262}
5263module_init(slab_proc_init);
5264#endif /* CONFIG_SLABINFO */