blob: 9308c8a2865bab2ac10fae15a56cfbb3af2f62ce [file] [log] [blame]
lh9ed821d2023-04-07 01:36:19 -07001/*
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#include <linux/prefetch.h>
33
34#include <trace/events/kmem.h>
35
36/*
37 * Lock order:
38 * 1. slub_lock (Global Semaphore)
39 * 2. node->list_lock
40 * 3. slab_lock(page) (Only on some arches and for debugging)
41 *
42 * slub_lock
43 *
44 * The role of the slub_lock is to protect the list of all the slabs
45 * and to synchronize major metadata changes to slab cache structures.
46 *
47 * The slab_lock is only used for debugging and on arches that do not
48 * have the ability to do a cmpxchg_double. It only protects the second
49 * double word in the page struct. Meaning
50 * A. page->freelist -> List of object free in a page
51 * B. page->counters -> Counters of objects
52 * C. page->frozen -> frozen state
53 *
54 * If a slab is frozen then it is exempt from list management. It is not
55 * on any list. The processor that froze the slab is the one who can
56 * perform list operations on the page. Other processors may put objects
57 * onto the freelist but the processor that froze the slab is the only
58 * one that can retrieve the objects from the page's freelist.
59 *
60 * The list_lock protects the partial and full list on each node and
61 * the partial slab counter. If taken then no new slabs may be added or
62 * removed from the lists nor make the number of partial slabs be modified.
63 * (Note that the total number of slabs is an atomic value that may be
64 * modified without taking the list lock).
65 *
66 * The list_lock is a centralized lock and thus we avoid taking it as
67 * much as possible. As long as SLUB does not have to handle partial
68 * slabs, operations can continue without any centralized lock. F.e.
69 * allocating a long series of objects that fill up slabs does not require
70 * the list lock.
71 * Interrupts are disabled during allocation and deallocation in order to
72 * make the slab allocator safe to use in the context of an irq. In addition
73 * interrupts are disabled to ensure that the processor does not change
74 * while handling per_cpu slabs, due to kernel preemption.
75 *
76 * SLUB assigns one slab for allocation to each processor.
77 * Allocations only occur from these slabs called cpu slabs.
78 *
79 * Slabs with free elements are kept on a partial list and during regular
80 * operations no list for full slabs is used. If an object in a full slab is
81 * freed then the slab will show up again on the partial lists.
82 * We track full slabs for debugging purposes though because otherwise we
83 * cannot scan all objects.
84 *
85 * Slabs are freed when they become empty. Teardown and setup is
86 * minimal so we rely on the page allocators per cpu caches for
87 * fast frees and allocs.
88 *
89 * Overloading of page flags that are otherwise used for LRU management.
90 *
91 * PageActive The slab is frozen and exempt from list processing.
92 * This means that the slab is dedicated to a purpose
93 * such as satisfying allocations for a specific
94 * processor. Objects may be freed in the slab while
95 * it is frozen but slab_free will then skip the usual
96 * list operations. It is up to the processor holding
97 * the slab to integrate the slab into the slab lists
98 * when the slab is no longer needed.
99 *
100 * One use of this flag is to mark slabs that are
101 * used for allocations. Then such a slab becomes a cpu
102 * slab. The cpu slab may be equipped with an additional
103 * freelist that allows lockless access to
104 * free objects in addition to the regular freelist
105 * that requires the slab lock.
106 *
107 * PageError Slab requires special handling due to debug
108 * options set. This moves slab handling out of
109 * the fast path and disables lockless freelists.
110 */
111
112#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
113 SLAB_TRACE | SLAB_DEBUG_FREE)
114
115static inline int kmem_cache_debug(struct kmem_cache *s)
116{
117#ifdef CONFIG_SLUB_DEBUG
118 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
119#else
120 return 0;
121#endif
122}
123
124/*
125 * Issues still to be resolved:
126 *
127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 *
129 * - Variable sizing of the per node arrays
130 */
131
132/* Enable to test recovery from slab corruption on boot */
133#undef SLUB_RESILIENCY_TEST
134
135/* Enable to log cmpxchg failures */
136#undef SLUB_DEBUG_CMPXCHG
137
138/*
139 * Mininum number of partial slabs. These will be left on the partial
140 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 */
142#define MIN_PARTIAL 5
143
144/*
145 * Maximum number of desirable partial slabs.
146 * The existence of more partial slabs makes kmem_cache_shrink
147 * sort the partial list by the number of objects in the.
148 */
149#define MAX_PARTIAL 10
150
151#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
152 SLAB_POISON | SLAB_STORE_USER)
153
154/*
155 * Debugging flags that require metadata to be stored in the slab. These get
156 * disabled when slub_debug=O is used and a cache's min order increases with
157 * metadata.
158 */
159#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
160
161/*
162 * Set of flags that will prevent slab merging
163 */
164#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
165 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
166 SLAB_FAILSLAB)
167
168#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
169 SLAB_CACHE_DMA | SLAB_NOTRACK)
170
171#define OO_SHIFT 16
172#define OO_MASK ((1 << OO_SHIFT) - 1)
173#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174
175/* Internal SLUB flags */
176#define __OBJECT_POISON 0x80000000UL /* Poison object */
177#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178
179static int kmem_size = sizeof(struct kmem_cache);
180
181#ifdef CONFIG_SMP
182static struct notifier_block slab_notifier;
183#endif
184
185static enum {
186 DOWN, /* No slab functionality available */
187 PARTIAL, /* Kmem_cache_node works */
188 UP, /* Everything works but does not show up in sysfs */
189 SYSFS /* Sysfs up */
190} slab_state = DOWN;
191
192/* A list of all slab caches on the system */
193static DECLARE_RWSEM(slub_lock);
194static LIST_HEAD(slab_caches);
195
196/*
197 * Tracking user of a slab.
198 */
199#define TRACK_ADDRS_COUNT 16
200struct track {
201 unsigned long addr; /* Called from address */
202#ifdef CONFIG_STACKTRACE
203 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
204#endif
205 int cpu; /* Was running on cpu */
206 int pid; /* Pid context */
207 unsigned long when; /* When did the operation occur */
208};
209
210enum track_item { TRACK_ALLOC, TRACK_FREE };
211
212#ifdef CONFIG_SYSFS
213static int sysfs_slab_add(struct kmem_cache *);
214static int sysfs_slab_alias(struct kmem_cache *, const char *);
215static void sysfs_slab_remove(struct kmem_cache *);
216
217#else
218static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
220 { return 0; }
221static inline void sysfs_slab_remove(struct kmem_cache *s)
222{
223 kfree(s->name);
224 kfree(s);
225}
226
227#endif
228
229static inline void stat(const struct kmem_cache *s, enum stat_item si)
230{
231#ifdef CONFIG_SLUB_STATS
232 __this_cpu_inc(s->cpu_slab->stat[si]);
233#endif
234}
235
236/********************************************************************
237 * Core slab cache functions
238 *******************************************************************/
239
240int slab_is_available(void)
241{
242 return slab_state >= UP;
243}
244
245static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246{
247 return s->node[node];
248}
249
250/* Verify that a pointer has an address that is valid within a slab page */
251static inline int check_valid_pointer(struct kmem_cache *s,
252 struct page *page, const void *object)
253{
254 void *base;
255
256 if (!object)
257 return 1;
258
259 base = page_address(page);
260 if (object < base || object >= base + page->objects * s->size ||
261 (object - base) % s->size) {
262 return 0;
263 }
264
265 return 1;
266}
267
268static inline void *get_freepointer(struct kmem_cache *s, void *object)
269{
270 return *(void **)(object + s->offset);
271}
272
273static void prefetch_freepointer(const struct kmem_cache *s, void *object)
274{
275 prefetch(object + s->offset);
276}
277
278static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
279{
280 void *p;
281
282#ifdef CONFIG_DEBUG_PAGEALLOC
283 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
284#else
285 p = get_freepointer(s, object);
286#endif
287 return p;
288}
289
290static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291{
292 *(void **)(object + s->offset) = fp;
293}
294
295/* Loop over all objects in a slab */
296#define for_each_object(__p, __s, __addr, __objects) \
297 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
298 __p += (__s)->size)
299
300/* Determine object index from a given position */
301static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
302{
303 return (p - addr) / s->size;
304}
305
306static inline size_t slab_ksize(const struct kmem_cache *s)
307{
308#ifdef CONFIG_SLUB_DEBUG
309 /*
310 * Debugging requires use of the padding between object
311 * and whatever may come after it.
312 */
313 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
314 return s->objsize;
315
316#endif
317 /*
318 * If we have the need to store the freelist pointer
319 * back there or track user information then we can
320 * only use the space before that information.
321 */
322 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
323 return s->inuse;
324 /*
325 * Else we can use all the padding etc for the allocation
326 */
327 return s->size;
328}
329
330static inline int order_objects(int order, unsigned long size, int reserved)
331{
332 return ((PAGE_SIZE << order) - reserved) / size;
333}
334
335static inline struct kmem_cache_order_objects oo_make(int order,
336 unsigned long size, int reserved)
337{
338 struct kmem_cache_order_objects x = {
339 (order << OO_SHIFT) + order_objects(order, size, reserved)
340 };
341
342 return x;
343}
344
345static inline int oo_order(struct kmem_cache_order_objects x)
346{
347 return x.x >> OO_SHIFT;
348}
349
350static inline int oo_objects(struct kmem_cache_order_objects x)
351{
352 return x.x & OO_MASK;
353}
354
355/*
356 * Per slab locking using the pagelock
357 */
358static __always_inline void slab_lock(struct page *page)
359{
360 bit_spin_lock(PG_locked, &page->flags);
361}
362
363static __always_inline void slab_unlock(struct page *page)
364{
365 __bit_spin_unlock(PG_locked, &page->flags);
366}
367
368/* Interrupts must be disabled (for the fallback code to work right) */
369static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
370 void *freelist_old, unsigned long counters_old,
371 void *freelist_new, unsigned long counters_new,
372 const char *n)
373{
374 VM_BUG_ON(!irqs_disabled());
375#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377 if (s->flags & __CMPXCHG_DOUBLE) {
378 if (cmpxchg_double(&page->freelist, &page->counters,
379 freelist_old, counters_old,
380 freelist_new, counters_new))
381 return 1;
382 } else
383#endif
384 {
385 slab_lock(page);
386 if (page->freelist == freelist_old && page->counters == counters_old) {
387 page->freelist = freelist_new;
388 page->counters = counters_new;
389 slab_unlock(page);
390 return 1;
391 }
392 slab_unlock(page);
393 }
394
395 cpu_relax();
396 stat(s, CMPXCHG_DOUBLE_FAIL);
397
398#ifdef SLUB_DEBUG_CMPXCHG
399 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
400#endif
401
402 return 0;
403}
404
405static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
406 void *freelist_old, unsigned long counters_old,
407 void *freelist_new, unsigned long counters_new,
408 const char *n)
409{
410#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
411 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
412 if (s->flags & __CMPXCHG_DOUBLE) {
413 if (cmpxchg_double(&page->freelist, &page->counters,
414 freelist_old, counters_old,
415 freelist_new, counters_new))
416 return 1;
417 } else
418#endif
419 {
420 unsigned long flags;
421
422 local_irq_save(flags);
423 slab_lock(page);
424 if (page->freelist == freelist_old && page->counters == counters_old) {
425 page->freelist = freelist_new;
426 page->counters = counters_new;
427 slab_unlock(page);
428 local_irq_restore(flags);
429 return 1;
430 }
431 slab_unlock(page);
432 local_irq_restore(flags);
433 }
434
435 cpu_relax();
436 stat(s, CMPXCHG_DOUBLE_FAIL);
437
438#ifdef SLUB_DEBUG_CMPXCHG
439 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
440#endif
441
442 return 0;
443}
444
445#ifdef CONFIG_SLUB_DEBUG
446/*
447 * Determine a map of object in use on a page.
448 *
449 * Node listlock must be held to guarantee that the page does
450 * not vanish from under us.
451 */
452static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453{
454 void *p;
455 void *addr = page_address(page);
456
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(slab_index(p, s, addr), map);
459}
460
461/*
462 * Debug settings:
463 */
464#ifdef CONFIG_SLUB_DEBUG_ON
465static int slub_debug = DEBUG_DEFAULT_FLAGS;
466#else
467static int slub_debug;
468#endif
469
470static char *slub_debug_slabs;
471static int disable_higher_order_debug;
472
473/*
474 * Object debugging
475 */
476static void print_section(char *text, u8 *addr, unsigned int length)
477{
478 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
479 length, 1);
480}
481
482static struct track *get_track(struct kmem_cache *s, void *object,
483 enum track_item alloc)
484{
485 struct track *p;
486
487 if (s->offset)
488 p = object + s->offset + sizeof(void *);
489 else
490 p = object + s->inuse;
491
492 return p + alloc;
493}
494
495static void set_track(struct kmem_cache *s, void *object,
496 enum track_item alloc, unsigned long addr)
497{
498 struct track *p = get_track(s, object, alloc);
499
500 if (addr) {
501#ifdef CONFIG_STACKTRACE
502 struct stack_trace trace;
503 int i;
504
505 trace.nr_entries = 0;
506 trace.max_entries = TRACK_ADDRS_COUNT;
507 trace.entries = p->addrs;
508 trace.skip = 3;
509 save_stack_trace(&trace);
510
511 /* See rant in lockdep.c */
512 if (trace.nr_entries != 0 &&
513 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
514 trace.nr_entries--;
515
516 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
517 p->addrs[i] = 0;
518#endif
519 p->addr = addr;
520 p->cpu = smp_processor_id();
521 p->pid = current->pid;
522 p->when = jiffies;
523 } else
524 memset(p, 0, sizeof(struct track));
525}
526
527static void init_tracking(struct kmem_cache *s, void *object)
528{
529 if (!(s->flags & SLAB_STORE_USER))
530 return;
531
532 set_track(s, object, TRACK_FREE, 0UL);
533 set_track(s, object, TRACK_ALLOC, 0UL);
534}
535
536static void print_track(const char *s, struct track *t)
537{
538 if (!t->addr)
539 return;
540
541 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
542 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
543#ifdef CONFIG_STACKTRACE
544 {
545 int i;
546 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
547 if (t->addrs[i])
548 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
549 else
550 break;
551 }
552#endif
553}
554
555static void print_tracking(struct kmem_cache *s, void *object)
556{
557 if (!(s->flags & SLAB_STORE_USER))
558 return;
559
560 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
561 print_track("Freed", get_track(s, object, TRACK_FREE));
562}
563
564static void print_page_info(struct page *page)
565{
566 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
567 page, page->objects, page->inuse, page->freelist, page->flags);
568
569}
570
571static void slab_bug(struct kmem_cache *s, char *fmt, ...)
572{
573 va_list args;
574 char buf[100];
575
576 va_start(args, fmt);
577 vsnprintf(buf, sizeof(buf), fmt, args);
578 va_end(args);
579 printk(KERN_ERR "========================================"
580 "=====================================\n");
581 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
582 printk(KERN_ERR "----------------------------------------"
583 "-------------------------------------\n\n");
584}
585
586static void slab_fix(struct kmem_cache *s, char *fmt, ...)
587{
588 va_list args;
589 char buf[100];
590
591 va_start(args, fmt);
592 vsnprintf(buf, sizeof(buf), fmt, args);
593 va_end(args);
594 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
595}
596
597static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
598{
599 unsigned int off; /* Offset of last byte */
600 u8 *addr = page_address(page);
601
602 print_tracking(s, p);
603
604 print_page_info(page);
605
606 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
608
609 if (p > addr + 16)
610 print_section("Bytes b4 ", p - 16, 16);
611
612 print_section("Object ", p, min_t(unsigned long, s->objsize,
613 PAGE_SIZE));
614 if (s->flags & SLAB_RED_ZONE)
615 print_section("Redzone ", p + s->objsize,
616 s->inuse - s->objsize);
617
618 if (s->offset)
619 off = s->offset + sizeof(void *);
620 else
621 off = s->inuse;
622
623 if (s->flags & SLAB_STORE_USER)
624 off += 2 * sizeof(struct track);
625
626 if (off != s->size)
627 /* Beginning of the filler is the free pointer */
628 print_section("Padding ", p + off, s->size - off);
629
630 dump_stack();
631}
632
633static void object_err(struct kmem_cache *s, struct page *page,
634 u8 *object, char *reason)
635{
636 slab_bug(s, "%s", reason);
637 print_trailer(s, page, object);
638}
639
640static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
641{
642 va_list args;
643 char buf[100];
644
645 va_start(args, fmt);
646 vsnprintf(buf, sizeof(buf), fmt, args);
647 va_end(args);
648 slab_bug(s, "%s", buf);
649 print_page_info(page);
650 dump_stack();
651}
652
653static void init_object(struct kmem_cache *s, void *object, u8 val)
654{
655 u8 *p = object;
656
657 if (s->flags & __OBJECT_POISON) {
658 memset(p, POISON_FREE, s->objsize - 1);
659 p[s->objsize - 1] = POISON_END;
660 }
661
662 if (s->flags & SLAB_RED_ZONE)
663 memset(p + s->objsize, val, s->inuse - s->objsize);
664}
665
666static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 void *from, void *to)
668{
669 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 memset(from, data, to - from);
671}
672
673static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 u8 *object, char *what,
675 u8 *start, unsigned int value, unsigned int bytes)
676{
677 u8 *fault;
678 u8 *end;
679
680 fault = memchr_inv(start, value, bytes);
681 if (!fault)
682 return 1;
683
684 end = start + bytes;
685 while (end > fault && end[-1] == value)
686 end--;
687
688 slab_bug(s, "%s overwritten", what);
689 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 fault, end - 1, fault[0], value);
691 print_trailer(s, page, object);
692
693 restore_bytes(s, what, value, fault, end);
694 return 0;
695}
696
697/*
698 * Object layout:
699 *
700 * object address
701 * Bytes of the object to be managed.
702 * If the freepointer may overlay the object then the free
703 * pointer is the first word of the object.
704 *
705 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
706 * 0xa5 (POISON_END)
707 *
708 * object + s->objsize
709 * Padding to reach word boundary. This is also used for Redzoning.
710 * Padding is extended by another word if Redzoning is enabled and
711 * objsize == inuse.
712 *
713 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714 * 0xcc (RED_ACTIVE) for objects in use.
715 *
716 * object + s->inuse
717 * Meta data starts here.
718 *
719 * A. Free pointer (if we cannot overwrite object on free)
720 * B. Tracking data for SLAB_STORE_USER
721 * C. Padding to reach required alignment boundary or at mininum
722 * one word if debugging is on to be able to detect writes
723 * before the word boundary.
724 *
725 * Padding is done using 0x5a (POISON_INUSE)
726 *
727 * object + s->size
728 * Nothing is used beyond s->size.
729 *
730 * If slabcaches are merged then the objsize and inuse boundaries are mostly
731 * ignored. And therefore no slab options that rely on these boundaries
732 * may be used with merged slabcaches.
733 */
734
735static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
736{
737 unsigned long off = s->inuse; /* The end of info */
738
739 if (s->offset)
740 /* Freepointer is placed after the object. */
741 off += sizeof(void *);
742
743 if (s->flags & SLAB_STORE_USER)
744 /* We also have user information there */
745 off += 2 * sizeof(struct track);
746
747 if (s->size == off)
748 return 1;
749
750 return check_bytes_and_report(s, page, p, "Object padding",
751 p + off, POISON_INUSE, s->size - off);
752}
753
754/* Check the pad bytes at the end of a slab page */
755static int slab_pad_check(struct kmem_cache *s, struct page *page)
756{
757 u8 *start;
758 u8 *fault;
759 u8 *end;
760 int length;
761 int remainder;
762
763 if (!(s->flags & SLAB_POISON))
764 return 1;
765
766 start = page_address(page);
767 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 end = start + length;
769 remainder = length % s->size;
770 if (!remainder)
771 return 1;
772
773 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
774 if (!fault)
775 return 1;
776 while (end > fault && end[-1] == POISON_INUSE)
777 end--;
778
779 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 print_section("Padding ", end - remainder, remainder);
781
782 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
783 return 0;
784}
785
786static int check_object(struct kmem_cache *s, struct page *page,
787 void *object, u8 val)
788{
789 u8 *p = object;
790 u8 *endobject = object + s->objsize;
791
792 if (s->flags & SLAB_RED_ZONE) {
793 if (!check_bytes_and_report(s, page, object, "Redzone",
794 endobject, val, s->inuse - s->objsize))
795 return 0;
796 } else {
797 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
798 check_bytes_and_report(s, page, p, "Alignment padding",
799 endobject, POISON_INUSE, s->inuse - s->objsize);
800 }
801 }
802
803 if (s->flags & SLAB_POISON) {
804 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
805 (!check_bytes_and_report(s, page, p, "Poison", p,
806 POISON_FREE, s->objsize - 1) ||
807 !check_bytes_and_report(s, page, p, "Poison",
808 p + s->objsize - 1, POISON_END, 1)))
809 return 0;
810 /*
811 * check_pad_bytes cleans up on its own.
812 */
813 check_pad_bytes(s, page, p);
814 }
815
816 if (!s->offset && val == SLUB_RED_ACTIVE)
817 /*
818 * Object and freepointer overlap. Cannot check
819 * freepointer while object is allocated.
820 */
821 return 1;
822
823 /* Check free pointer validity */
824 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
825 object_err(s, page, p, "Freepointer corrupt");
826 /*
827 * No choice but to zap it and thus lose the remainder
828 * of the free objects in this slab. May cause
829 * another error because the object count is now wrong.
830 */
831 set_freepointer(s, p, NULL);
832 return 0;
833 }
834 return 1;
835}
836
837static int check_slab(struct kmem_cache *s, struct page *page)
838{
839 int maxobj;
840
841 VM_BUG_ON(!irqs_disabled());
842
843 if (!PageSlab(page)) {
844 slab_err(s, page, "Not a valid slab page");
845 return 0;
846 }
847
848 maxobj = order_objects(compound_order(page), s->size, s->reserved);
849 if (page->objects > maxobj) {
850 slab_err(s, page, "objects %u > max %u",
851 s->name, page->objects, maxobj);
852 return 0;
853 }
854 if (page->inuse > page->objects) {
855 slab_err(s, page, "inuse %u > max %u",
856 s->name, page->inuse, page->objects);
857 return 0;
858 }
859 /* Slab_pad_check fixes things up after itself */
860 slab_pad_check(s, page);
861 return 1;
862}
863
864/*
865 * Determine if a certain object on a page is on the freelist. Must hold the
866 * slab lock to guarantee that the chains are in a consistent state.
867 */
868static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
869{
870 int nr = 0;
871 void *fp;
872 void *object = NULL;
873 unsigned long max_objects;
874
875 fp = page->freelist;
876 while (fp && nr <= page->objects) {
877 if (fp == search)
878 return 1;
879 if (!check_valid_pointer(s, page, fp)) {
880 if (object) {
881 object_err(s, page, object,
882 "Freechain corrupt");
883 set_freepointer(s, object, NULL);
884 break;
885 } else {
886 slab_err(s, page, "Freepointer corrupt");
887 page->freelist = NULL;
888 page->inuse = page->objects;
889 slab_fix(s, "Freelist cleared");
890 return 0;
891 }
892 break;
893 }
894 object = fp;
895 fp = get_freepointer(s, object);
896 nr++;
897 }
898
899 max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 if (max_objects > MAX_OBJS_PER_PAGE)
901 max_objects = MAX_OBJS_PER_PAGE;
902
903 if (page->objects != max_objects) {
904 slab_err(s, page, "Wrong number of objects. Found %d but "
905 "should be %d", page->objects, max_objects);
906 page->objects = max_objects;
907 slab_fix(s, "Number of objects adjusted.");
908 }
909 if (page->inuse != page->objects - nr) {
910 slab_err(s, page, "Wrong object count. Counter is %d but "
911 "counted were %d", page->inuse, page->objects - nr);
912 page->inuse = page->objects - nr;
913 slab_fix(s, "Object count adjusted.");
914 }
915 return search == NULL;
916}
917
918static void trace(struct kmem_cache *s, struct page *page, void *object,
919 int alloc)
920{
921 if (s->flags & SLAB_TRACE) {
922 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
923 s->name,
924 alloc ? "alloc" : "free",
925 object, page->inuse,
926 page->freelist);
927
928 if (!alloc)
929 print_section("Object ", (void *)object, s->objsize);
930
931 dump_stack();
932 }
933}
934
935/*
936 * Hooks for other subsystems that check memory allocations. In a typical
937 * production configuration these hooks all should produce no code at all.
938 */
939static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
940{
941 flags &= gfp_allowed_mask;
942 lockdep_trace_alloc(flags);
943 might_sleep_if(flags & __GFP_WAIT);
944
945 return should_failslab(s->objsize, flags, s->flags);
946}
947
948static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
949{
950 flags &= gfp_allowed_mask;
951 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
952 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
953}
954
955static inline void slab_free_hook(struct kmem_cache *s, void *x)
956{
957 kmemleak_free_recursive(x, s->flags);
958
959 /*
960 * Trouble is that we may no longer disable interupts in the fast path
961 * So in order to make the debug calls that expect irqs to be
962 * disabled we need to disable interrupts temporarily.
963 */
964#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
965 {
966 unsigned long flags;
967
968 local_irq_save(flags);
969 kmemcheck_slab_free(s, x, s->objsize);
970 debug_check_no_locks_freed(x, s->objsize);
971 local_irq_restore(flags);
972 }
973#endif
974 if (!(s->flags & SLAB_DEBUG_OBJECTS))
975 debug_check_no_obj_freed(x, s->objsize);
976}
977
978/*
979 * Tracking of fully allocated slabs for debugging purposes.
980 *
981 * list_lock must be held.
982 */
983static void add_full(struct kmem_cache *s,
984 struct kmem_cache_node *n, struct page *page)
985{
986 if (!(s->flags & SLAB_STORE_USER))
987 return;
988
989 list_add(&page->lru, &n->full);
990}
991
992/*
993 * list_lock must be held.
994 */
995static void remove_full(struct kmem_cache *s, struct page *page)
996{
997 if (!(s->flags & SLAB_STORE_USER))
998 return;
999
1000 list_del(&page->lru);
1001}
1002
1003/* Tracking of the number of slabs for debugging purposes */
1004static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1005{
1006 struct kmem_cache_node *n = get_node(s, node);
1007
1008 return atomic_long_read(&n->nr_slabs);
1009}
1010
1011static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1012{
1013 return atomic_long_read(&n->nr_slabs);
1014}
1015
1016static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1017{
1018 struct kmem_cache_node *n = get_node(s, node);
1019
1020 /*
1021 * May be called early in order to allocate a slab for the
1022 * kmem_cache_node structure. Solve the chicken-egg
1023 * dilemma by deferring the increment of the count during
1024 * bootstrap (see early_kmem_cache_node_alloc).
1025 */
1026 if (n) {
1027 atomic_long_inc(&n->nr_slabs);
1028 atomic_long_add(objects, &n->total_objects);
1029 }
1030}
1031static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1032{
1033 struct kmem_cache_node *n = get_node(s, node);
1034
1035 atomic_long_dec(&n->nr_slabs);
1036 atomic_long_sub(objects, &n->total_objects);
1037}
1038
1039/* Object debug checks for alloc/free paths */
1040static void setup_object_debug(struct kmem_cache *s, struct page *page,
1041 void *object)
1042{
1043 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1044 return;
1045
1046 init_object(s, object, SLUB_RED_INACTIVE);
1047 init_tracking(s, object);
1048}
1049
1050static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1051 void *object, unsigned long addr)
1052{
1053 if (!check_slab(s, page))
1054 goto bad;
1055
1056 if (!check_valid_pointer(s, page, object)) {
1057 object_err(s, page, object, "Freelist Pointer check fails");
1058 goto bad;
1059 }
1060
1061 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1062 goto bad;
1063
1064 /* Success perform special debug activities for allocs */
1065 if (s->flags & SLAB_STORE_USER)
1066 set_track(s, object, TRACK_ALLOC, addr);
1067 trace(s, page, object, 1);
1068 init_object(s, object, SLUB_RED_ACTIVE);
1069 return 1;
1070
1071bad:
1072 if (PageSlab(page)) {
1073 /*
1074 * If this is a slab page then lets do the best we can
1075 * to avoid issues in the future. Marking all objects
1076 * as used avoids touching the remaining objects.
1077 */
1078 slab_fix(s, "Marking all objects used");
1079 page->inuse = page->objects;
1080 page->freelist = NULL;
1081 }
1082 return 0;
1083}
1084
1085static noinline int free_debug_processing(struct kmem_cache *s,
1086 struct page *page, void *object, unsigned long addr)
1087{
1088 unsigned long flags;
1089 int rc = 0;
1090
1091 local_irq_save(flags);
1092 slab_lock(page);
1093
1094 if (!check_slab(s, page))
1095 goto fail;
1096
1097 if (!check_valid_pointer(s, page, object)) {
1098 slab_err(s, page, "Invalid object pointer 0x%p", object);
1099 goto fail;
1100 }
1101
1102 if (on_freelist(s, page, object)) {
1103 object_err(s, page, object, "Object already free");
1104 goto fail;
1105 }
1106
1107 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1108 goto out;
1109
1110 if (unlikely(s != page->slab)) {
1111 if (!PageSlab(page)) {
1112 slab_err(s, page, "Attempt to free object(0x%p) "
1113 "outside of slab", object);
1114 } else if (!page->slab) {
1115 printk(KERN_ERR
1116 "SLUB <none>: no slab for object 0x%p.\n",
1117 object);
1118 dump_stack();
1119 } else
1120 object_err(s, page, object,
1121 "page slab pointer corrupt.");
1122 goto fail;
1123 }
1124
1125 if (s->flags & SLAB_STORE_USER)
1126 set_track(s, object, TRACK_FREE, addr);
1127 trace(s, page, object, 0);
1128 init_object(s, object, SLUB_RED_INACTIVE);
1129 rc = 1;
1130out:
1131 slab_unlock(page);
1132 local_irq_restore(flags);
1133 return rc;
1134
1135fail:
1136 slab_fix(s, "Object at 0x%p not freed", object);
1137 goto out;
1138}
1139
1140static int __init setup_slub_debug(char *str)
1141{
1142 slub_debug = DEBUG_DEFAULT_FLAGS;
1143 if (*str++ != '=' || !*str)
1144 /*
1145 * No options specified. Switch on full debugging.
1146 */
1147 goto out;
1148
1149 if (*str == ',')
1150 /*
1151 * No options but restriction on slabs. This means full
1152 * debugging for slabs matching a pattern.
1153 */
1154 goto check_slabs;
1155
1156 if (tolower(*str) == 'o') {
1157 /*
1158 * Avoid enabling debugging on caches if its minimum order
1159 * would increase as a result.
1160 */
1161 disable_higher_order_debug = 1;
1162 goto out;
1163 }
1164
1165 slub_debug = 0;
1166 if (*str == '-')
1167 /*
1168 * Switch off all debugging measures.
1169 */
1170 goto out;
1171
1172 /*
1173 * Determine which debug features should be switched on
1174 */
1175 for (; *str && *str != ','; str++) {
1176 switch (tolower(*str)) {
1177 case 'f':
1178 slub_debug |= SLAB_DEBUG_FREE;
1179 break;
1180 case 'z':
1181 slub_debug |= SLAB_RED_ZONE;
1182 break;
1183 case 'p':
1184 slub_debug |= SLAB_POISON;
1185 break;
1186 case 'u':
1187 slub_debug |= SLAB_STORE_USER;
1188 break;
1189 case 't':
1190 slub_debug |= SLAB_TRACE;
1191 break;
1192 case 'a':
1193 slub_debug |= SLAB_FAILSLAB;
1194 break;
1195 default:
1196 printk(KERN_ERR "slub_debug option '%c' "
1197 "unknown. skipped\n", *str);
1198 }
1199 }
1200
1201check_slabs:
1202 if (*str == ',')
1203 slub_debug_slabs = str + 1;
1204out:
1205 return 1;
1206}
1207
1208__setup("slub_debug", setup_slub_debug);
1209
1210static unsigned long kmem_cache_flags(unsigned long objsize,
1211 unsigned long flags, const char *name,
1212 void (*ctor)(void *))
1213{
1214 /*
1215 * Enable debugging if selected on the kernel commandline.
1216 */
1217 if (slub_debug && (!slub_debug_slabs ||
1218 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1219 flags |= slub_debug;
1220
1221 return flags;
1222}
1223#else
1224static inline void setup_object_debug(struct kmem_cache *s,
1225 struct page *page, void *object) {}
1226
1227static inline int alloc_debug_processing(struct kmem_cache *s,
1228 struct page *page, void *object, unsigned long addr) { return 0; }
1229
1230static inline int free_debug_processing(struct kmem_cache *s,
1231 struct page *page, void *object, unsigned long addr) { return 0; }
1232
1233static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1234 { return 1; }
1235static inline int check_object(struct kmem_cache *s, struct page *page,
1236 void *object, u8 val) { return 1; }
1237static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1238 struct page *page) {}
1239static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1240static inline unsigned long kmem_cache_flags(unsigned long objsize,
1241 unsigned long flags, const char *name,
1242 void (*ctor)(void *))
1243{
1244 return flags;
1245}
1246#define slub_debug 0
1247
1248#define disable_higher_order_debug 0
1249
1250static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1251 { return 0; }
1252static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1253 { return 0; }
1254static inline void inc_slabs_node(struct kmem_cache *s, int node,
1255 int objects) {}
1256static inline void dec_slabs_node(struct kmem_cache *s, int node,
1257 int objects) {}
1258
1259static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1260 { return 0; }
1261
1262static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1263 void *object) {}
1264
1265static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1266
1267#endif /* CONFIG_SLUB_DEBUG */
1268
1269static void setup_object(struct kmem_cache *s, struct page *page,
1270 void *object)
1271{
1272 setup_object_debug(s, page, object);
1273 if (unlikely(s->ctor))
1274 s->ctor(object);
1275}
1276
1277/*
1278 * Slab allocation and freeing
1279 */
1280static inline struct page *alloc_slab_page(gfp_t flags, int node,
1281 struct kmem_cache_order_objects oo)
1282{
1283 int order = oo_order(oo);
1284
1285 flags |= __GFP_NOTRACK;
1286
1287 if (node == NUMA_NO_NODE)
1288 return alloc_pages(flags, order);
1289 else
1290 return alloc_pages_exact_node(node, flags, order);
1291}
1292
1293static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1294{
1295 struct page *page;
1296 struct kmem_cache_order_objects oo = s->oo;
1297 gfp_t alloc_gfp;
1298 void *start, *last, *p;
1299 int idx, order;
1300
1301 flags &= gfp_allowed_mask;
1302
1303 if (flags & __GFP_WAIT)
1304 local_irq_enable();
1305
1306 flags |= s->allocflags;
1307
1308 /*
1309 * Let the initial higher-order allocation fail under memory pressure
1310 * so we fall-back to the minimum order allocation.
1311 */
1312 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1313
1314 page = alloc_slab_page(alloc_gfp, node, oo);
1315 if (unlikely(!page)) {
1316 oo = s->min;
1317 /*
1318 * Allocation may have failed due to fragmentation.
1319 * Try a lower order alloc if possible
1320 */
1321 page = alloc_slab_page(flags, node, oo);
1322 if (unlikely(!page))
1323 goto out;
1324 stat(s, ORDER_FALLBACK);
1325 }
1326
1327 if (kmemcheck_enabled
1328 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1329 int pages = 1 << oo_order(oo);
1330
1331 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1332
1333 /*
1334 * Objects from caches that have a constructor don't get
1335 * cleared when they're allocated, so we need to do it here.
1336 */
1337 if (s->ctor)
1338 kmemcheck_mark_uninitialized_pages(page, pages);
1339 else
1340 kmemcheck_mark_unallocated_pages(page, pages);
1341 }
1342
1343 page->objects = oo_objects(oo);
1344 page->slab = s;
1345 page->flags |= 1 << PG_slab;
1346
1347 start = page_address(page);
1348
1349 if (unlikely(s->flags & SLAB_POISON))
1350 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1351
1352 last = start;
1353 for_each_object(p, s, start, page->objects) {
1354 setup_object(s, page, last);
1355 set_freepointer(s, last, p);
1356 last = p;
1357 }
1358 setup_object(s, page, last);
1359 set_freepointer(s, last, NULL);
1360
1361 page->freelist = start;
1362 page->inuse = page->objects;
1363 page->frozen = 1;
1364
1365out:
1366 if (flags & __GFP_WAIT)
1367 local_irq_disable();
1368 if (!page)
1369 return NULL;
1370
1371 mod_zone_page_state(page_zone(page),
1372 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1373 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1374 1 << oo_order(oo));
1375
1376 inc_slabs_node(s, page_to_nid(page), page->objects);
1377
1378 return page;
1379}
1380
1381static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1382{
1383 if (unlikely(flags & GFP_SLAB_BUG_MASK)) {
1384 pr_emerg("gfp: %u\n", flags & GFP_SLAB_BUG_MASK);
1385 BUG();
1386 }
1387
1388 return allocate_slab(s,
1389 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1390}
1391
1392static void __free_slab(struct kmem_cache *s, struct page *page)
1393{
1394 int order = compound_order(page);
1395 int pages = 1 << order;
1396
1397 if (kmem_cache_debug(s)) {
1398 void *p;
1399
1400 slab_pad_check(s, page);
1401 for_each_object(p, s, page_address(page),
1402 page->objects)
1403 check_object(s, page, p, SLUB_RED_INACTIVE);
1404 }
1405
1406 kmemcheck_free_shadow(page, compound_order(page));
1407
1408 mod_zone_page_state(page_zone(page),
1409 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1410 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1411 -pages);
1412
1413 __ClearPageSlab(page);
1414 reset_page_mapcount(page);
1415 if (current->reclaim_state)
1416 current->reclaim_state->reclaimed_slab += pages;
1417 __free_pages(page, order);
1418}
1419
1420#define need_reserve_slab_rcu \
1421 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1422
1423static void rcu_free_slab(struct rcu_head *h)
1424{
1425 struct page *page;
1426
1427 if (need_reserve_slab_rcu)
1428 page = virt_to_head_page(h);
1429 else
1430 page = container_of((struct list_head *)h, struct page, lru);
1431
1432 __free_slab(page->slab, page);
1433}
1434
1435static void free_slab(struct kmem_cache *s, struct page *page)
1436{
1437 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1438 struct rcu_head *head;
1439
1440 if (need_reserve_slab_rcu) {
1441 int order = compound_order(page);
1442 int offset = (PAGE_SIZE << order) - s->reserved;
1443
1444 VM_BUG_ON(s->reserved != sizeof(*head));
1445 head = page_address(page) + offset;
1446 } else {
1447 /*
1448 * RCU free overloads the RCU head over the LRU
1449 */
1450 head = (void *)&page->lru;
1451 }
1452
1453 call_rcu(head, rcu_free_slab);
1454 } else
1455 __free_slab(s, page);
1456}
1457
1458static void discard_slab(struct kmem_cache *s, struct page *page)
1459{
1460 dec_slabs_node(s, page_to_nid(page), page->objects);
1461 free_slab(s, page);
1462}
1463
1464/*
1465 * Management of partially allocated slabs.
1466 *
1467 * list_lock must be held.
1468 */
1469static inline void add_partial(struct kmem_cache_node *n,
1470 struct page *page, int tail)
1471{
1472 n->nr_partial++;
1473 if (tail == DEACTIVATE_TO_TAIL)
1474 list_add_tail(&page->lru, &n->partial);
1475 else
1476 list_add(&page->lru, &n->partial);
1477}
1478
1479/*
1480 * list_lock must be held.
1481 */
1482static inline void remove_partial(struct kmem_cache_node *n,
1483 struct page *page)
1484{
1485 list_del(&page->lru);
1486 n->nr_partial--;
1487}
1488
1489/*
1490 * Lock slab, remove from the partial list and put the object into the
1491 * per cpu freelist.
1492 *
1493 * Returns a list of objects or NULL if it fails.
1494 *
1495 * Must hold list_lock.
1496 */
1497static inline void *acquire_slab(struct kmem_cache *s,
1498 struct kmem_cache_node *n, struct page *page,
1499 int mode)
1500{
1501 void *freelist;
1502 unsigned long counters;
1503 struct page new;
1504
1505 /*
1506 * Zap the freelist and set the frozen bit.
1507 * The old freelist is the list of objects for the
1508 * per cpu allocation list.
1509 */
1510 do {
1511 freelist = page->freelist;
1512 counters = page->counters;
1513 new.counters = counters;
1514 if (mode) {
1515 new.inuse = page->objects;
1516 new.freelist = NULL;
1517 } else {
1518 new.freelist = freelist;
1519 }
1520
1521 VM_BUG_ON(new.frozen);
1522 new.frozen = 1;
1523
1524 } while (!__cmpxchg_double_slab(s, page,
1525 freelist, counters,
1526 new.freelist, new.counters,
1527 "lock and freeze"));
1528
1529 remove_partial(n, page);
1530 return freelist;
1531}
1532
1533static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1534
1535/*
1536 * Try to allocate a partial slab from a specific node.
1537 */
1538static void *get_partial_node(struct kmem_cache *s,
1539 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1540{
1541 struct page *page, *page2;
1542 void *object = NULL;
1543
1544 /*
1545 * Racy check. If we mistakenly see no partial slabs then we
1546 * just allocate an empty slab. If we mistakenly try to get a
1547 * partial slab and there is none available then get_partials()
1548 * will return NULL.
1549 */
1550 if (!n || !n->nr_partial)
1551 return NULL;
1552
1553 spin_lock(&n->list_lock);
1554 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1555 void *t = acquire_slab(s, n, page, object == NULL);
1556 int available;
1557
1558 if (!t)
1559 break;
1560
1561 if (!object) {
1562 c->page = page;
1563 c->node = page_to_nid(page);
1564 stat(s, ALLOC_FROM_PARTIAL);
1565 object = t;
1566 available = page->objects - page->inuse;
1567 } else {
1568 available = put_cpu_partial(s, page, 0);
1569 stat(s, CPU_PARTIAL_NODE);
1570 }
1571 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1572 break;
1573
1574 }
1575 spin_unlock(&n->list_lock);
1576 return object;
1577}
1578
1579/*
1580 * Get a page from somewhere. Search in increasing NUMA distances.
1581 */
1582static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags,
1583 struct kmem_cache_cpu *c)
1584{
1585#ifdef CONFIG_NUMA
1586 struct zonelist *zonelist;
1587 struct zoneref *z;
1588 struct zone *zone;
1589 enum zone_type high_zoneidx = gfp_zone(flags);
1590 void *object;
1591 unsigned int cpuset_mems_cookie;
1592
1593 /*
1594 * The defrag ratio allows a configuration of the tradeoffs between
1595 * inter node defragmentation and node local allocations. A lower
1596 * defrag_ratio increases the tendency to do local allocations
1597 * instead of attempting to obtain partial slabs from other nodes.
1598 *
1599 * If the defrag_ratio is set to 0 then kmalloc() always
1600 * returns node local objects. If the ratio is higher then kmalloc()
1601 * may return off node objects because partial slabs are obtained
1602 * from other nodes and filled up.
1603 *
1604 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1605 * defrag_ratio = 1000) then every (well almost) allocation will
1606 * first attempt to defrag slab caches on other nodes. This means
1607 * scanning over all nodes to look for partial slabs which may be
1608 * expensive if we do it every time we are trying to find a slab
1609 * with available objects.
1610 */
1611 if (!s->remote_node_defrag_ratio ||
1612 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1613 return NULL;
1614
1615 do {
1616 cpuset_mems_cookie = get_mems_allowed();
1617 zonelist = node_zonelist(slab_node(), flags);
1618 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1619 struct kmem_cache_node *n;
1620
1621 n = get_node(s, zone_to_nid(zone));
1622
1623 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1624 n->nr_partial > s->min_partial) {
1625 object = get_partial_node(s, n, c);
1626 if (object) {
1627 /*
1628 * Return the object even if
1629 * put_mems_allowed indicated that
1630 * the cpuset mems_allowed was
1631 * updated in parallel. It's a
1632 * harmless race between the alloc
1633 * and the cpuset update.
1634 */
1635 put_mems_allowed(cpuset_mems_cookie);
1636 return object;
1637 }
1638 }
1639 }
1640 } while (!put_mems_allowed(cpuset_mems_cookie));
1641#endif
1642 return NULL;
1643}
1644
1645/*
1646 * Get a partial page, lock it and return it.
1647 */
1648static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1649 struct kmem_cache_cpu *c)
1650{
1651 void *object;
1652 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1653
1654 object = get_partial_node(s, get_node(s, searchnode), c);
1655 if (object || node != NUMA_NO_NODE)
1656 return object;
1657
1658 return get_any_partial(s, flags, c);
1659}
1660
1661#ifdef CONFIG_PREEMPT
1662/*
1663 * Calculate the next globally unique transaction for disambiguiation
1664 * during cmpxchg. The transactions start with the cpu number and are then
1665 * incremented by CONFIG_NR_CPUS.
1666 */
1667#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1668#else
1669/*
1670 * No preemption supported therefore also no need to check for
1671 * different cpus.
1672 */
1673#define TID_STEP 1
1674#endif
1675
1676static inline unsigned long next_tid(unsigned long tid)
1677{
1678 return tid + TID_STEP;
1679}
1680
1681static inline unsigned int tid_to_cpu(unsigned long tid)
1682{
1683 return tid % TID_STEP;
1684}
1685
1686static inline unsigned long tid_to_event(unsigned long tid)
1687{
1688 return tid / TID_STEP;
1689}
1690
1691static inline unsigned int init_tid(int cpu)
1692{
1693 return cpu;
1694}
1695
1696static inline void note_cmpxchg_failure(const char *n,
1697 const struct kmem_cache *s, unsigned long tid)
1698{
1699#ifdef SLUB_DEBUG_CMPXCHG
1700 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1701
1702 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1703
1704#ifdef CONFIG_PREEMPT
1705 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1706 printk("due to cpu change %d -> %d\n",
1707 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1708 else
1709#endif
1710 if (tid_to_event(tid) != tid_to_event(actual_tid))
1711 printk("due to cpu running other code. Event %ld->%ld\n",
1712 tid_to_event(tid), tid_to_event(actual_tid));
1713 else
1714 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1715 actual_tid, tid, next_tid(tid));
1716#endif
1717 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1718}
1719
1720void init_kmem_cache_cpus(struct kmem_cache *s)
1721{
1722 int cpu;
1723
1724 for_each_possible_cpu(cpu)
1725 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1726}
1727
1728/*
1729 * Remove the cpu slab
1730 */
1731static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1732{
1733 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1734 struct page *page = c->page;
1735 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1736 int lock = 0;
1737 enum slab_modes l = M_NONE, m = M_NONE;
1738 void *freelist;
1739 void *nextfree;
1740 int tail = DEACTIVATE_TO_HEAD;
1741 struct page new;
1742 struct page old;
1743
1744 if (page->freelist) {
1745 stat(s, DEACTIVATE_REMOTE_FREES);
1746 tail = DEACTIVATE_TO_TAIL;
1747 }
1748
1749 c->tid = next_tid(c->tid);
1750 c->page = NULL;
1751 freelist = c->freelist;
1752 c->freelist = NULL;
1753
1754 /*
1755 * Stage one: Free all available per cpu objects back
1756 * to the page freelist while it is still frozen. Leave the
1757 * last one.
1758 *
1759 * There is no need to take the list->lock because the page
1760 * is still frozen.
1761 */
1762 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1763 void *prior;
1764 unsigned long counters;
1765
1766 do {
1767 prior = page->freelist;
1768 counters = page->counters;
1769 set_freepointer(s, freelist, prior);
1770 new.counters = counters;
1771 new.inuse--;
1772 VM_BUG_ON(!new.frozen);
1773
1774 } while (!__cmpxchg_double_slab(s, page,
1775 prior, counters,
1776 freelist, new.counters,
1777 "drain percpu freelist"));
1778
1779 freelist = nextfree;
1780 }
1781
1782 /*
1783 * Stage two: Ensure that the page is unfrozen while the
1784 * list presence reflects the actual number of objects
1785 * during unfreeze.
1786 *
1787 * We setup the list membership and then perform a cmpxchg
1788 * with the count. If there is a mismatch then the page
1789 * is not unfrozen but the page is on the wrong list.
1790 *
1791 * Then we restart the process which may have to remove
1792 * the page from the list that we just put it on again
1793 * because the number of objects in the slab may have
1794 * changed.
1795 */
1796redo:
1797
1798 old.freelist = page->freelist;
1799 old.counters = page->counters;
1800 VM_BUG_ON(!old.frozen);
1801
1802 /* Determine target state of the slab */
1803 new.counters = old.counters;
1804 if (freelist) {
1805 new.inuse--;
1806 set_freepointer(s, freelist, old.freelist);
1807 new.freelist = freelist;
1808 } else
1809 new.freelist = old.freelist;
1810
1811 new.frozen = 0;
1812
1813 if (!new.inuse && n->nr_partial > s->min_partial)
1814 m = M_FREE;
1815 else if (new.freelist) {
1816 m = M_PARTIAL;
1817 if (!lock) {
1818 lock = 1;
1819 /*
1820 * Taking the spinlock removes the possiblity
1821 * that acquire_slab() will see a slab page that
1822 * is frozen
1823 */
1824 spin_lock(&n->list_lock);
1825 }
1826 } else {
1827 m = M_FULL;
1828 if (kmem_cache_debug(s) && !lock) {
1829 lock = 1;
1830 /*
1831 * This also ensures that the scanning of full
1832 * slabs from diagnostic functions will not see
1833 * any frozen slabs.
1834 */
1835 spin_lock(&n->list_lock);
1836 }
1837 }
1838
1839 if (l != m) {
1840
1841 if (l == M_PARTIAL)
1842
1843 remove_partial(n, page);
1844
1845 else if (l == M_FULL)
1846
1847 remove_full(s, page);
1848
1849 if (m == M_PARTIAL) {
1850
1851 add_partial(n, page, tail);
1852 stat(s, tail);
1853
1854 } else if (m == M_FULL) {
1855
1856 stat(s, DEACTIVATE_FULL);
1857 add_full(s, n, page);
1858
1859 }
1860 }
1861
1862 l = m;
1863 if (!__cmpxchg_double_slab(s, page,
1864 old.freelist, old.counters,
1865 new.freelist, new.counters,
1866 "unfreezing slab"))
1867 goto redo;
1868
1869 if (lock)
1870 spin_unlock(&n->list_lock);
1871
1872 if (m == M_FREE) {
1873 stat(s, DEACTIVATE_EMPTY);
1874 discard_slab(s, page);
1875 stat(s, FREE_SLAB);
1876 }
1877}
1878
1879/* Unfreeze all the cpu partial slabs */
1880static void unfreeze_partials(struct kmem_cache *s)
1881{
1882 struct kmem_cache_node *n = NULL, *n2 = NULL;
1883 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1884 struct page *page, *discard_page = NULL;
1885
1886 while ((page = c->partial)) {
1887 struct page new;
1888 struct page old;
1889
1890 c->partial = page->next;
1891
1892 n2 = get_node(s, page_to_nid(page));
1893 if (n != n2) {
1894 if (n)
1895 spin_unlock(&n->list_lock);
1896
1897 n = n2;
1898 spin_lock(&n->list_lock);
1899 }
1900
1901 do {
1902
1903 old.freelist = page->freelist;
1904 old.counters = page->counters;
1905 VM_BUG_ON(!old.frozen);
1906
1907 new.counters = old.counters;
1908 new.freelist = old.freelist;
1909
1910 new.frozen = 0;
1911
1912 } while (!cmpxchg_double_slab(s, page,
1913 old.freelist, old.counters,
1914 new.freelist, new.counters,
1915 "unfreezing slab"));
1916
1917 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1918 page->next = discard_page;
1919 discard_page = page;
1920 } else {
1921 add_partial(n, page, DEACTIVATE_TO_TAIL);
1922 stat(s, FREE_ADD_PARTIAL);
1923 }
1924 }
1925
1926 if (n)
1927 spin_unlock(&n->list_lock);
1928
1929 while (discard_page) {
1930 page = discard_page;
1931 discard_page = discard_page->next;
1932
1933 stat(s, DEACTIVATE_EMPTY);
1934 discard_slab(s, page);
1935 stat(s, FREE_SLAB);
1936 }
1937}
1938
1939/*
1940 * Put a page that was just frozen (in __slab_free) into a partial page
1941 * slot if available. This is done without interrupts disabled and without
1942 * preemption disabled. The cmpxchg is racy and may put the partial page
1943 * onto a random cpus partial slot.
1944 *
1945 * If we did not find a slot then simply move all the partials to the
1946 * per node partial list.
1947 */
1948int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1949{
1950 struct page *oldpage;
1951 int pages;
1952 int pobjects;
1953
1954 do {
1955 pages = 0;
1956 pobjects = 0;
1957 oldpage = this_cpu_read(s->cpu_slab->partial);
1958
1959 if (oldpage) {
1960 pobjects = oldpage->pobjects;
1961 pages = oldpage->pages;
1962 if (drain && pobjects > s->cpu_partial) {
1963 unsigned long flags;
1964 /*
1965 * partial array is full. Move the existing
1966 * set to the per node partial list.
1967 */
1968 local_irq_save(flags);
1969 unfreeze_partials(s);
1970 local_irq_restore(flags);
1971 pobjects = 0;
1972 pages = 0;
1973 stat(s, CPU_PARTIAL_DRAIN);
1974 }
1975 }
1976
1977 pages++;
1978 pobjects += page->objects - page->inuse;
1979
1980 page->pages = pages;
1981 page->pobjects = pobjects;
1982 page->next = oldpage;
1983
1984 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
1985 return pobjects;
1986}
1987
1988static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1989{
1990 stat(s, CPUSLAB_FLUSH);
1991 deactivate_slab(s, c);
1992}
1993
1994/*
1995 * Flush cpu slab.
1996 *
1997 * Called from IPI handler with interrupts disabled.
1998 */
1999static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2000{
2001 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2002
2003 if (likely(c)) {
2004 if (c->page)
2005 flush_slab(s, c);
2006
2007 unfreeze_partials(s);
2008 }
2009}
2010
2011static void flush_cpu_slab(void *d)
2012{
2013 struct kmem_cache *s = d;
2014
2015 __flush_cpu_slab(s, smp_processor_id());
2016}
2017
2018static bool has_cpu_slab(int cpu, void *info)
2019{
2020 struct kmem_cache *s = info;
2021 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2022
2023 return c->page || c->partial;
2024}
2025
2026static void flush_all(struct kmem_cache *s)
2027{
2028 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2029}
2030
2031/*
2032 * Check if the objects in a per cpu structure fit numa
2033 * locality expectations.
2034 */
2035static inline int node_match(struct kmem_cache_cpu *c, int node)
2036{
2037#ifdef CONFIG_NUMA
2038 if (node != NUMA_NO_NODE && c->node != node)
2039 return 0;
2040#endif
2041 return 1;
2042}
2043
2044static int count_free(struct page *page)
2045{
2046 return page->objects - page->inuse;
2047}
2048
2049static unsigned long count_partial(struct kmem_cache_node *n,
2050 int (*get_count)(struct page *))
2051{
2052 unsigned long flags;
2053 unsigned long x = 0;
2054 struct page *page;
2055
2056 spin_lock_irqsave(&n->list_lock, flags);
2057 list_for_each_entry(page, &n->partial, lru)
2058 x += get_count(page);
2059 spin_unlock_irqrestore(&n->list_lock, flags);
2060 return x;
2061}
2062
2063static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2064{
2065#ifdef CONFIG_SLUB_DEBUG
2066 return atomic_long_read(&n->total_objects);
2067#else
2068 return 0;
2069#endif
2070}
2071
2072static noinline void
2073slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2074{
2075 int node;
2076
2077 printk(KERN_WARNING
2078 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2079 nid, gfpflags);
2080 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2081 "default order: %d, min order: %d\n", s->name, s->objsize,
2082 s->size, oo_order(s->oo), oo_order(s->min));
2083
2084 if (oo_order(s->min) > get_order(s->objsize))
2085 printk(KERN_WARNING " %s debugging increased min order, use "
2086 "slub_debug=O to disable.\n", s->name);
2087
2088 for_each_online_node(node) {
2089 struct kmem_cache_node *n = get_node(s, node);
2090 unsigned long nr_slabs;
2091 unsigned long nr_objs;
2092 unsigned long nr_free;
2093
2094 if (!n)
2095 continue;
2096
2097 nr_free = count_partial(n, count_free);
2098 nr_slabs = node_nr_slabs(n);
2099 nr_objs = node_nr_objs(n);
2100
2101 printk(KERN_WARNING
2102 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2103 node, nr_slabs, nr_objs, nr_free);
2104 }
2105}
2106
2107static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2108 int node, struct kmem_cache_cpu **pc)
2109{
2110 void *object;
2111 struct kmem_cache_cpu *c;
2112 struct page *page = new_slab(s, flags, node);
2113
2114 if (page) {
2115 c = __this_cpu_ptr(s->cpu_slab);
2116 if (c->page)
2117 flush_slab(s, c);
2118
2119 /*
2120 * No other reference to the page yet so we can
2121 * muck around with it freely without cmpxchg
2122 */
2123 object = page->freelist;
2124 page->freelist = NULL;
2125
2126 stat(s, ALLOC_SLAB);
2127 c->node = page_to_nid(page);
2128 c->page = page;
2129 *pc = c;
2130 } else
2131 object = NULL;
2132
2133 return object;
2134}
2135
2136/*
2137 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2138 * or deactivate the page.
2139 *
2140 * The page is still frozen if the return value is not NULL.
2141 *
2142 * If this function returns NULL then the page has been unfrozen.
2143 */
2144static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2145{
2146 struct page new;
2147 unsigned long counters;
2148 void *freelist;
2149
2150 do {
2151 freelist = page->freelist;
2152 counters = page->counters;
2153 new.counters = counters;
2154 VM_BUG_ON(!new.frozen);
2155
2156 new.inuse = page->objects;
2157 new.frozen = freelist != NULL;
2158
2159 } while (!cmpxchg_double_slab(s, page,
2160 freelist, counters,
2161 NULL, new.counters,
2162 "get_freelist"));
2163
2164 return freelist;
2165}
2166
2167/*
2168 * Slow path. The lockless freelist is empty or we need to perform
2169 * debugging duties.
2170 *
2171 * Processing is still very fast if new objects have been freed to the
2172 * regular freelist. In that case we simply take over the regular freelist
2173 * as the lockless freelist and zap the regular freelist.
2174 *
2175 * If that is not working then we fall back to the partial lists. We take the
2176 * first element of the freelist as the object to allocate now and move the
2177 * rest of the freelist to the lockless freelist.
2178 *
2179 * And if we were unable to get a new slab from the partial slab lists then
2180 * we need to allocate a new slab. This is the slowest path since it involves
2181 * a call to the page allocator and the setup of a new slab.
2182 */
2183static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2184 unsigned long addr, struct kmem_cache_cpu *c)
2185{
2186 void **object;
2187 unsigned long flags;
2188
2189 local_irq_save(flags);
2190#ifdef CONFIG_PREEMPT
2191 /*
2192 * We may have been preempted and rescheduled on a different
2193 * cpu before disabling interrupts. Need to reload cpu area
2194 * pointer.
2195 */
2196 c = this_cpu_ptr(s->cpu_slab);
2197#endif
2198
2199 if (!c->page)
2200 goto new_slab;
2201redo:
2202 if (unlikely(!node_match(c, node))) {
2203 stat(s, ALLOC_NODE_MISMATCH);
2204 deactivate_slab(s, c);
2205 goto new_slab;
2206 }
2207
2208 /* must check again c->freelist in case of cpu migration or IRQ */
2209 object = c->freelist;
2210 if (object)
2211 goto load_freelist;
2212
2213 stat(s, ALLOC_SLOWPATH);
2214
2215 object = get_freelist(s, c->page);
2216
2217 if (!object) {
2218 c->page = NULL;
2219 stat(s, DEACTIVATE_BYPASS);
2220 goto new_slab;
2221 }
2222
2223 stat(s, ALLOC_REFILL);
2224
2225load_freelist:
2226 c->freelist = get_freepointer(s, object);
2227 c->tid = next_tid(c->tid);
2228 local_irq_restore(flags);
2229 return object;
2230
2231new_slab:
2232
2233 if (c->partial) {
2234 c->page = c->partial;
2235 c->partial = c->page->next;
2236 c->node = page_to_nid(c->page);
2237 stat(s, CPU_PARTIAL_ALLOC);
2238 c->freelist = NULL;
2239 goto redo;
2240 }
2241
2242 /* Then do expensive stuff like retrieving pages from the partial lists */
2243 object = get_partial(s, gfpflags, node, c);
2244
2245 if (unlikely(!object)) {
2246
2247 object = new_slab_objects(s, gfpflags, node, &c);
2248
2249 if (unlikely(!object)) {
2250 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2251 slab_out_of_memory(s, gfpflags, node);
2252
2253 local_irq_restore(flags);
2254 return NULL;
2255 }
2256 }
2257
2258 if (likely(!kmem_cache_debug(s)))
2259 goto load_freelist;
2260
2261 /* Only entered in the debug case */
2262 if (!alloc_debug_processing(s, c->page, object, addr))
2263 goto new_slab; /* Slab failed checks. Next slab needed */
2264
2265 c->freelist = get_freepointer(s, object);
2266 deactivate_slab(s, c);
2267 c->node = NUMA_NO_NODE;
2268 local_irq_restore(flags);
2269 return object;
2270}
2271
2272/*
2273 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2274 * have the fastpath folded into their functions. So no function call
2275 * overhead for requests that can be satisfied on the fastpath.
2276 *
2277 * The fastpath works by first checking if the lockless freelist can be used.
2278 * If not then __slab_alloc is called for slow processing.
2279 *
2280 * Otherwise we can simply pick the next object from the lockless free list.
2281 */
2282static __always_inline void *slab_alloc(struct kmem_cache *s,
2283 gfp_t gfpflags, int node, unsigned long addr)
2284{
2285 void **object;
2286 struct kmem_cache_cpu *c;
2287 unsigned long tid;
2288
2289 if (slab_pre_alloc_hook(s, gfpflags))
2290 return NULL;
2291
2292redo:
2293
2294 /*
2295 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2296 * enabled. We may switch back and forth between cpus while
2297 * reading from one cpu area. That does not matter as long
2298 * as we end up on the original cpu again when doing the cmpxchg.
2299 */
2300 c = __this_cpu_ptr(s->cpu_slab);
2301
2302 /*
2303 * The transaction ids are globally unique per cpu and per operation on
2304 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2305 * occurs on the right processor and that there was no operation on the
2306 * linked list in between.
2307 */
2308 tid = c->tid;
2309 barrier();
2310
2311 object = c->freelist;
2312 if (unlikely(!object || !node_match(c, node)))
2313
2314 object = __slab_alloc(s, gfpflags, node, addr, c);
2315
2316 else {
2317 void *next_object = get_freepointer_safe(s, object);
2318
2319 /*
2320 * The cmpxchg will only match if there was no additional
2321 * operation and if we are on the right processor.
2322 *
2323 * The cmpxchg does the following atomically (without lock semantics!)
2324 * 1. Relocate first pointer to the current per cpu area.
2325 * 2. Verify that tid and freelist have not been changed
2326 * 3. If they were not changed replace tid and freelist
2327 *
2328 * Since this is without lock semantics the protection is only against
2329 * code executing on this cpu *not* from access by other cpus.
2330 */
2331 if (unlikely(!this_cpu_cmpxchg_double(
2332 s->cpu_slab->freelist, s->cpu_slab->tid,
2333 object, tid,
2334 next_object, next_tid(tid)))) {
2335
2336 note_cmpxchg_failure("slab_alloc", s, tid);
2337 goto redo;
2338 }
2339 prefetch_freepointer(s, next_object);
2340 stat(s, ALLOC_FASTPATH);
2341 }
2342
2343 if (unlikely(gfpflags & __GFP_ZERO) && object)
2344 memset(object, 0, s->objsize);
2345
2346 slab_post_alloc_hook(s, gfpflags, object);
2347
2348 return object;
2349}
2350
2351void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2352{
2353 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2354
2355 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2356
2357 return ret;
2358}
2359EXPORT_SYMBOL(kmem_cache_alloc);
2360
2361#ifdef CONFIG_TRACING
2362void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2363{
2364 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2365 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2366 return ret;
2367}
2368EXPORT_SYMBOL(kmem_cache_alloc_trace);
2369
2370void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2371{
2372 void *ret = kmalloc_order(size, flags, order);
2373 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2374 return ret;
2375}
2376EXPORT_SYMBOL(kmalloc_order_trace);
2377#endif
2378
2379#ifdef CONFIG_NUMA
2380void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2381{
2382 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2383
2384 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2385 s->objsize, s->size, gfpflags, node);
2386
2387 return ret;
2388}
2389EXPORT_SYMBOL(kmem_cache_alloc_node);
2390
2391#ifdef CONFIG_TRACING
2392void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2393 gfp_t gfpflags,
2394 int node, size_t size)
2395{
2396 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2397
2398 trace_kmalloc_node(_RET_IP_, ret,
2399 size, s->size, gfpflags, node);
2400 return ret;
2401}
2402EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2403#endif
2404#endif
2405
2406/*
2407 * Slow patch handling. This may still be called frequently since objects
2408 * have a longer lifetime than the cpu slabs in most processing loads.
2409 *
2410 * So we still attempt to reduce cache line usage. Just take the slab
2411 * lock and free the item. If there is no additional partial page
2412 * handling required then we can return immediately.
2413 */
2414static void __slab_free(struct kmem_cache *s, struct page *page,
2415 void *x, unsigned long addr)
2416{
2417 void *prior;
2418 void **object = (void *)x;
2419 int was_frozen;
2420 int inuse;
2421 struct page new;
2422 unsigned long counters;
2423 struct kmem_cache_node *n = NULL;
2424 unsigned long uninitialized_var(flags);
2425
2426 stat(s, FREE_SLOWPATH);
2427
2428 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2429 return;
2430
2431 do {
2432 prior = page->freelist;
2433 counters = page->counters;
2434 set_freepointer(s, object, prior);
2435 new.counters = counters;
2436 was_frozen = new.frozen;
2437 new.inuse--;
2438 if ((!new.inuse || !prior) && !was_frozen && !n) {
2439
2440 if (!kmem_cache_debug(s) && !prior)
2441
2442 /*
2443 * Slab was on no list before and will be partially empty
2444 * We can defer the list move and instead freeze it.
2445 */
2446 new.frozen = 1;
2447
2448 else { /* Needs to be taken off a list */
2449
2450 n = get_node(s, page_to_nid(page));
2451 /*
2452 * Speculatively acquire the list_lock.
2453 * If the cmpxchg does not succeed then we may
2454 * drop the list_lock without any processing.
2455 *
2456 * Otherwise the list_lock will synchronize with
2457 * other processors updating the list of slabs.
2458 */
2459 spin_lock_irqsave(&n->list_lock, flags);
2460
2461 }
2462 }
2463 inuse = new.inuse;
2464
2465 } while (!cmpxchg_double_slab(s, page,
2466 prior, counters,
2467 object, new.counters,
2468 "__slab_free"));
2469
2470 if (likely(!n)) {
2471
2472 /*
2473 * If we just froze the page then put it onto the
2474 * per cpu partial list.
2475 */
2476 if (new.frozen && !was_frozen) {
2477 put_cpu_partial(s, page, 1);
2478 stat(s, CPU_PARTIAL_FREE);
2479 }
2480 /*
2481 * The list lock was not taken therefore no list
2482 * activity can be necessary.
2483 */
2484 if (was_frozen)
2485 stat(s, FREE_FROZEN);
2486 return;
2487 }
2488
2489 /*
2490 * was_frozen may have been set after we acquired the list_lock in
2491 * an earlier loop. So we need to check it here again.
2492 */
2493 if (was_frozen)
2494 stat(s, FREE_FROZEN);
2495 else {
2496 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2497 goto slab_empty;
2498
2499 /*
2500 * Objects left in the slab. If it was not on the partial list before
2501 * then add it.
2502 */
2503 if (unlikely(!prior)) {
2504 remove_full(s, page);
2505 add_partial(n, page, DEACTIVATE_TO_TAIL);
2506 stat(s, FREE_ADD_PARTIAL);
2507 }
2508 }
2509 spin_unlock_irqrestore(&n->list_lock, flags);
2510 return;
2511
2512slab_empty:
2513 if (prior) {
2514 /*
2515 * Slab on the partial list.
2516 */
2517 remove_partial(n, page);
2518 stat(s, FREE_REMOVE_PARTIAL);
2519 } else
2520 /* Slab must be on the full list */
2521 remove_full(s, page);
2522
2523 spin_unlock_irqrestore(&n->list_lock, flags);
2524 stat(s, FREE_SLAB);
2525 discard_slab(s, page);
2526}
2527
2528/*
2529 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2530 * can perform fastpath freeing without additional function calls.
2531 *
2532 * The fastpath is only possible if we are freeing to the current cpu slab
2533 * of this processor. This typically the case if we have just allocated
2534 * the item before.
2535 *
2536 * If fastpath is not possible then fall back to __slab_free where we deal
2537 * with all sorts of special processing.
2538 */
2539static __always_inline void slab_free(struct kmem_cache *s,
2540 struct page *page, void *x, unsigned long addr)
2541{
2542 void **object = (void *)x;
2543 struct kmem_cache_cpu *c;
2544 unsigned long tid;
2545
2546 slab_free_hook(s, x);
2547
2548redo:
2549 /*
2550 * Determine the currently cpus per cpu slab.
2551 * The cpu may change afterward. However that does not matter since
2552 * data is retrieved via this pointer. If we are on the same cpu
2553 * during the cmpxchg then the free will succedd.
2554 */
2555 c = __this_cpu_ptr(s->cpu_slab);
2556
2557 tid = c->tid;
2558 barrier();
2559
2560 if (likely(page == c->page)) {
2561 set_freepointer(s, object, c->freelist);
2562
2563 if (unlikely(!this_cpu_cmpxchg_double(
2564 s->cpu_slab->freelist, s->cpu_slab->tid,
2565 c->freelist, tid,
2566 object, next_tid(tid)))) {
2567
2568 note_cmpxchg_failure("slab_free", s, tid);
2569 goto redo;
2570 }
2571 stat(s, FREE_FASTPATH);
2572 } else
2573 __slab_free(s, page, x, addr);
2574
2575}
2576
2577void kmem_cache_free(struct kmem_cache *s, void *x)
2578{
2579 struct page *page;
2580
2581 page = virt_to_head_page(x);
2582
2583 slab_free(s, page, x, _RET_IP_);
2584
2585 trace_kmem_cache_free(_RET_IP_, x);
2586}
2587EXPORT_SYMBOL(kmem_cache_free);
2588
2589/*
2590 * Object placement in a slab is made very easy because we always start at
2591 * offset 0. If we tune the size of the object to the alignment then we can
2592 * get the required alignment by putting one properly sized object after
2593 * another.
2594 *
2595 * Notice that the allocation order determines the sizes of the per cpu
2596 * caches. Each processor has always one slab available for allocations.
2597 * Increasing the allocation order reduces the number of times that slabs
2598 * must be moved on and off the partial lists and is therefore a factor in
2599 * locking overhead.
2600 */
2601
2602/*
2603 * Mininum / Maximum order of slab pages. This influences locking overhead
2604 * and slab fragmentation. A higher order reduces the number of partial slabs
2605 * and increases the number of allocations possible without having to
2606 * take the list_lock.
2607 */
2608static int slub_min_order;
2609static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2610static int slub_min_objects;
2611
2612/*
2613 * Merge control. If this is set then no merging of slab caches will occur.
2614 * (Could be removed. This was introduced to pacify the merge skeptics.)
2615 */
2616static int slub_nomerge;
2617
2618/*
2619 * Calculate the order of allocation given an slab object size.
2620 *
2621 * The order of allocation has significant impact on performance and other
2622 * system components. Generally order 0 allocations should be preferred since
2623 * order 0 does not cause fragmentation in the page allocator. Larger objects
2624 * be problematic to put into order 0 slabs because there may be too much
2625 * unused space left. We go to a higher order if more than 1/16th of the slab
2626 * would be wasted.
2627 *
2628 * In order to reach satisfactory performance we must ensure that a minimum
2629 * number of objects is in one slab. Otherwise we may generate too much
2630 * activity on the partial lists which requires taking the list_lock. This is
2631 * less a concern for large slabs though which are rarely used.
2632 *
2633 * slub_max_order specifies the order where we begin to stop considering the
2634 * number of objects in a slab as critical. If we reach slub_max_order then
2635 * we try to keep the page order as low as possible. So we accept more waste
2636 * of space in favor of a small page order.
2637 *
2638 * Higher order allocations also allow the placement of more objects in a
2639 * slab and thereby reduce object handling overhead. If the user has
2640 * requested a higher mininum order then we start with that one instead of
2641 * the smallest order which will fit the object.
2642 */
2643static inline int slab_order(int size, int min_objects,
2644 int max_order, int fract_leftover, int reserved)
2645{
2646 int order;
2647 int rem;
2648 int min_order = slub_min_order;
2649
2650 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2651 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2652
2653 for (order = max(min_order,
2654 fls(min_objects * size - 1) - PAGE_SHIFT);
2655 order <= max_order; order++) {
2656
2657 unsigned long slab_size = PAGE_SIZE << order;
2658
2659 if (slab_size < min_objects * size + reserved)
2660 continue;
2661
2662 rem = (slab_size - reserved) % size;
2663
2664 if (rem <= slab_size / fract_leftover)
2665 break;
2666
2667 }
2668
2669 return order;
2670}
2671
2672static inline int calculate_order(int size, int reserved)
2673{
2674 int order;
2675 int min_objects;
2676 int fraction;
2677 int max_objects;
2678
2679 /*
2680 * Attempt to find best configuration for a slab. This
2681 * works by first attempting to generate a layout with
2682 * the best configuration and backing off gradually.
2683 *
2684 * First we reduce the acceptable waste in a slab. Then
2685 * we reduce the minimum objects required in a slab.
2686 */
2687 min_objects = slub_min_objects;
2688 if (!min_objects)
2689 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2690 max_objects = order_objects(slub_max_order, size, reserved);
2691 min_objects = min(min_objects, max_objects);
2692
2693 while (min_objects > 1) {
2694 fraction = 16;
2695 while (fraction >= 4) {
2696 order = slab_order(size, min_objects,
2697 slub_max_order, fraction, reserved);
2698 if (order <= slub_max_order)
2699 return order;
2700 fraction /= 2;
2701 }
2702 min_objects--;
2703 }
2704
2705 /*
2706 * We were unable to place multiple objects in a slab. Now
2707 * lets see if we can place a single object there.
2708 */
2709 order = slab_order(size, 1, slub_max_order, 1, reserved);
2710 if (order <= slub_max_order)
2711 return order;
2712
2713 /*
2714 * Doh this slab cannot be placed using slub_max_order.
2715 */
2716 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2717 if (order < MAX_ORDER)
2718 return order;
2719 return -ENOSYS;
2720}
2721
2722/*
2723 * Figure out what the alignment of the objects will be.
2724 */
2725static unsigned long calculate_alignment(unsigned long flags,
2726 unsigned long align, unsigned long size)
2727{
2728 /*
2729 * If the user wants hardware cache aligned objects then follow that
2730 * suggestion if the object is sufficiently large.
2731 *
2732 * The hardware cache alignment cannot override the specified
2733 * alignment though. If that is greater then use it.
2734 */
2735 if (flags & SLAB_HWCACHE_ALIGN) {
2736 unsigned long ralign = cache_line_size();
2737 while (size <= ralign / 2)
2738 ralign /= 2;
2739 align = max(align, ralign);
2740 }
2741
2742 if (align < ARCH_SLAB_MINALIGN)
2743 align = ARCH_SLAB_MINALIGN;
2744
2745 return ALIGN(align, sizeof(void *));
2746}
2747
2748static void
2749init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2750{
2751 n->nr_partial = 0;
2752 spin_lock_init(&n->list_lock);
2753 INIT_LIST_HEAD(&n->partial);
2754#ifdef CONFIG_SLUB_DEBUG
2755 atomic_long_set(&n->nr_slabs, 0);
2756 atomic_long_set(&n->total_objects, 0);
2757 INIT_LIST_HEAD(&n->full);
2758#endif
2759}
2760
2761static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2762{
2763 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2764 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2765
2766 /*
2767 * Must align to double word boundary for the double cmpxchg
2768 * instructions to work; see __pcpu_double_call_return_bool().
2769 */
2770 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2771 2 * sizeof(void *));
2772
2773 if (!s->cpu_slab)
2774 return 0;
2775
2776 init_kmem_cache_cpus(s);
2777
2778 return 1;
2779}
2780
2781static struct kmem_cache *kmem_cache_node;
2782
2783/*
2784 * No kmalloc_node yet so do it by hand. We know that this is the first
2785 * slab on the node for this slabcache. There are no concurrent accesses
2786 * possible.
2787 *
2788 * Note that this function only works on the kmalloc_node_cache
2789 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2790 * memory on a fresh node that has no slab structures yet.
2791 */
2792static void early_kmem_cache_node_alloc(int node)
2793{
2794 struct page *page;
2795 struct kmem_cache_node *n;
2796
2797 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2798
2799 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2800
2801 BUG_ON(!page);
2802 if (page_to_nid(page) != node) {
2803 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2804 "node %d\n", node);
2805 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2806 "in order to be able to continue\n");
2807 }
2808
2809 n = page->freelist;
2810 BUG_ON(!n);
2811 page->freelist = get_freepointer(kmem_cache_node, n);
2812 page->inuse = 1;
2813 page->frozen = 0;
2814 kmem_cache_node->node[node] = n;
2815#ifdef CONFIG_SLUB_DEBUG
2816 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2817 init_tracking(kmem_cache_node, n);
2818#endif
2819 init_kmem_cache_node(n, kmem_cache_node);
2820 inc_slabs_node(kmem_cache_node, node, page->objects);
2821
2822 add_partial(n, page, DEACTIVATE_TO_HEAD);
2823}
2824
2825static void free_kmem_cache_nodes(struct kmem_cache *s)
2826{
2827 int node;
2828
2829 for_each_node_state(node, N_NORMAL_MEMORY) {
2830 struct kmem_cache_node *n = s->node[node];
2831
2832 if (n)
2833 kmem_cache_free(kmem_cache_node, n);
2834
2835 s->node[node] = NULL;
2836 }
2837}
2838
2839static int init_kmem_cache_nodes(struct kmem_cache *s)
2840{
2841 int node;
2842
2843 for_each_node_state(node, N_NORMAL_MEMORY) {
2844 struct kmem_cache_node *n;
2845
2846 if (slab_state == DOWN) {
2847 early_kmem_cache_node_alloc(node);
2848 continue;
2849 }
2850 n = kmem_cache_alloc_node(kmem_cache_node,
2851 GFP_KERNEL, node);
2852
2853 if (!n) {
2854 free_kmem_cache_nodes(s);
2855 return 0;
2856 }
2857
2858 s->node[node] = n;
2859 init_kmem_cache_node(n, s);
2860 }
2861 return 1;
2862}
2863
2864static void set_min_partial(struct kmem_cache *s, unsigned long min)
2865{
2866 if (min < MIN_PARTIAL)
2867 min = MIN_PARTIAL;
2868 else if (min > MAX_PARTIAL)
2869 min = MAX_PARTIAL;
2870 s->min_partial = min;
2871}
2872
2873/*
2874 * calculate_sizes() determines the order and the distribution of data within
2875 * a slab object.
2876 */
2877static int calculate_sizes(struct kmem_cache *s, int forced_order)
2878{
2879 unsigned long flags = s->flags;
2880 unsigned long size = s->objsize;
2881 unsigned long align = s->align;
2882 int order;
2883
2884 /*
2885 * Round up object size to the next word boundary. We can only
2886 * place the free pointer at word boundaries and this determines
2887 * the possible location of the free pointer.
2888 */
2889 size = ALIGN(size, sizeof(void *));
2890
2891#ifdef CONFIG_SLUB_DEBUG
2892 /*
2893 * Determine if we can poison the object itself. If the user of
2894 * the slab may touch the object after free or before allocation
2895 * then we should never poison the object itself.
2896 */
2897 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2898 !s->ctor)
2899 s->flags |= __OBJECT_POISON;
2900 else
2901 s->flags &= ~__OBJECT_POISON;
2902
2903
2904 /*
2905 * If we are Redzoning then check if there is some space between the
2906 * end of the object and the free pointer. If not then add an
2907 * additional word to have some bytes to store Redzone information.
2908 */
2909 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2910 size += sizeof(void *);
2911#endif
2912
2913 /*
2914 * With that we have determined the number of bytes in actual use
2915 * by the object. This is the potential offset to the free pointer.
2916 */
2917 s->inuse = size;
2918
2919 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2920 s->ctor)) {
2921 /*
2922 * Relocate free pointer after the object if it is not
2923 * permitted to overwrite the first word of the object on
2924 * kmem_cache_free.
2925 *
2926 * This is the case if we do RCU, have a constructor or
2927 * destructor or are poisoning the objects.
2928 */
2929 s->offset = size;
2930 size += sizeof(void *);
2931 }
2932
2933#ifdef CONFIG_SLUB_DEBUG
2934 if (flags & SLAB_STORE_USER)
2935 /*
2936 * Need to store information about allocs and frees after
2937 * the object.
2938 */
2939 size += 2 * sizeof(struct track);
2940
2941 if (flags & SLAB_RED_ZONE)
2942 /*
2943 * Add some empty padding so that we can catch
2944 * overwrites from earlier objects rather than let
2945 * tracking information or the free pointer be
2946 * corrupted if a user writes before the start
2947 * of the object.
2948 */
2949 size += sizeof(void *);
2950#endif
2951
2952 /*
2953 * Determine the alignment based on various parameters that the
2954 * user specified and the dynamic determination of cache line size
2955 * on bootup.
2956 */
2957 align = calculate_alignment(flags, align, s->objsize);
2958 s->align = align;
2959
2960 /*
2961 * SLUB stores one object immediately after another beginning from
2962 * offset 0. In order to align the objects we have to simply size
2963 * each object to conform to the alignment.
2964 */
2965 size = ALIGN(size, align);
2966 s->size = size;
2967 if (forced_order >= 0)
2968 order = forced_order;
2969 else
2970 order = calculate_order(size, s->reserved);
2971
2972 if (order < 0)
2973 return 0;
2974
2975 s->allocflags = 0;
2976 if (order)
2977 s->allocflags |= __GFP_COMP;
2978
2979 if (s->flags & SLAB_CACHE_DMA)
2980 s->allocflags |= SLUB_DMA;
2981
2982 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2983 s->allocflags |= __GFP_RECLAIMABLE;
2984
2985 /*
2986 * Determine the number of objects per slab
2987 */
2988 s->oo = oo_make(order, size, s->reserved);
2989 s->min = oo_make(get_order(size), size, s->reserved);
2990 if (oo_objects(s->oo) > oo_objects(s->max))
2991 s->max = s->oo;
2992
2993 return !!oo_objects(s->oo);
2994
2995}
2996
2997static int kmem_cache_open(struct kmem_cache *s,
2998 const char *name, size_t size,
2999 size_t align, unsigned long flags,
3000 void (*ctor)(void *))
3001{
3002 memset(s, 0, kmem_size);
3003 s->name = name;
3004 s->ctor = ctor;
3005 s->objsize = size;
3006 s->align = align;
3007 s->flags = kmem_cache_flags(size, flags, name, ctor);
3008 s->reserved = 0;
3009
3010 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3011 s->reserved = sizeof(struct rcu_head);
3012
3013 if (!calculate_sizes(s, -1))
3014 goto error;
3015 if (disable_higher_order_debug) {
3016 /*
3017 * Disable debugging flags that store metadata if the min slab
3018 * order increased.
3019 */
3020 if (get_order(s->size) > get_order(s->objsize)) {
3021 s->flags &= ~DEBUG_METADATA_FLAGS;
3022 s->offset = 0;
3023 if (!calculate_sizes(s, -1))
3024 goto error;
3025 }
3026 }
3027
3028#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3029 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3030 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3031 /* Enable fast mode */
3032 s->flags |= __CMPXCHG_DOUBLE;
3033#endif
3034
3035 /*
3036 * The larger the object size is, the more pages we want on the partial
3037 * list to avoid pounding the page allocator excessively.
3038 */
3039 set_min_partial(s, ilog2(s->size) / 2);
3040
3041 /*
3042 * cpu_partial determined the maximum number of objects kept in the
3043 * per cpu partial lists of a processor.
3044 *
3045 * Per cpu partial lists mainly contain slabs that just have one
3046 * object freed. If they are used for allocation then they can be
3047 * filled up again with minimal effort. The slab will never hit the
3048 * per node partial lists and therefore no locking will be required.
3049 *
3050 * This setting also determines
3051 *
3052 * A) The number of objects from per cpu partial slabs dumped to the
3053 * per node list when we reach the limit.
3054 * B) The number of objects in cpu partial slabs to extract from the
3055 * per node list when we run out of per cpu objects. We only fetch 50%
3056 * to keep some capacity around for frees.
3057 */
3058 if (kmem_cache_debug(s))
3059 s->cpu_partial = 0;
3060 else if (s->size >= PAGE_SIZE)
3061 s->cpu_partial = 2;
3062 else if (s->size >= 1024)
3063 s->cpu_partial = 6;
3064 else if (s->size >= 256)
3065 s->cpu_partial = 13;
3066 else
3067 s->cpu_partial = 30;
3068
3069 s->refcount = 1;
3070#ifdef CONFIG_NUMA
3071 s->remote_node_defrag_ratio = 1000;
3072#endif
3073 if (!init_kmem_cache_nodes(s))
3074 goto error;
3075
3076 if (alloc_kmem_cache_cpus(s))
3077 return 1;
3078
3079 free_kmem_cache_nodes(s);
3080error:
3081 if (flags & SLAB_PANIC)
3082 panic("Cannot create slab %s size=%lu realsize=%u "
3083 "order=%u offset=%u flags=%lx\n",
3084 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3085 s->offset, flags);
3086 return 0;
3087}
3088
3089/*
3090 * Determine the size of a slab object
3091 */
3092unsigned int kmem_cache_size(struct kmem_cache *s)
3093{
3094 return s->objsize;
3095}
3096EXPORT_SYMBOL(kmem_cache_size);
3097
3098static void list_slab_objects(struct kmem_cache *s, struct page *page,
3099 const char *text)
3100{
3101#ifdef CONFIG_SLUB_DEBUG
3102 void *addr = page_address(page);
3103 void *p;
3104 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3105 sizeof(long), GFP_ATOMIC);
3106 if (!map)
3107 return;
3108 slab_err(s, page, "%s", text);
3109 slab_lock(page);
3110
3111 get_map(s, page, map);
3112 for_each_object(p, s, addr, page->objects) {
3113
3114 if (!test_bit(slab_index(p, s, addr), map)) {
3115 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3116 p, p - addr);
3117 print_tracking(s, p);
3118 }
3119 }
3120 slab_unlock(page);
3121 kfree(map);
3122#endif
3123}
3124
3125/*
3126 * Attempt to free all partial slabs on a node.
3127 * This is called from kmem_cache_close(). We must be the last thread
3128 * using the cache and therefore we do not need to lock anymore.
3129 */
3130static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3131{
3132 struct page *page, *h;
3133
3134 list_for_each_entry_safe(page, h, &n->partial, lru) {
3135 if (!page->inuse) {
3136 remove_partial(n, page);
3137 discard_slab(s, page);
3138 } else {
3139 list_slab_objects(s, page,
3140 "Objects remaining on kmem_cache_close()");
3141 }
3142 }
3143}
3144
3145/*
3146 * Release all resources used by a slab cache.
3147 */
3148static inline int kmem_cache_close(struct kmem_cache *s)
3149{
3150 int node;
3151
3152 flush_all(s);
3153 free_percpu(s->cpu_slab);
3154 /* Attempt to free all objects */
3155 for_each_node_state(node, N_NORMAL_MEMORY) {
3156 struct kmem_cache_node *n = get_node(s, node);
3157
3158 free_partial(s, n);
3159 if (n->nr_partial || slabs_node(s, node))
3160 return 1;
3161 }
3162 free_kmem_cache_nodes(s);
3163 return 0;
3164}
3165
3166/*
3167 * Close a cache and release the kmem_cache structure
3168 * (must be used for caches created using kmem_cache_create)
3169 */
3170void kmem_cache_destroy(struct kmem_cache *s)
3171{
3172 down_write(&slub_lock);
3173 s->refcount--;
3174 if (!s->refcount) {
3175 list_del(&s->list);
3176 up_write(&slub_lock);
3177 if (kmem_cache_close(s)) {
3178 printk(KERN_ERR "SLUB %s: %s called for cache that "
3179 "still has objects.\n", s->name, __func__);
3180 dump_stack();
3181 }
3182 if (s->flags & SLAB_DESTROY_BY_RCU)
3183 rcu_barrier();
3184 sysfs_slab_remove(s);
3185 } else
3186 up_write(&slub_lock);
3187}
3188EXPORT_SYMBOL(kmem_cache_destroy);
3189
3190/********************************************************************
3191 * Kmalloc subsystem
3192 *******************************************************************/
3193
3194struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3195EXPORT_SYMBOL(kmalloc_caches);
3196
3197static struct kmem_cache *kmem_cache;
3198
3199#ifdef CONFIG_ZONE_DMA
3200static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3201#endif
3202
3203static int __init setup_slub_min_order(char *str)
3204{
3205 get_option(&str, &slub_min_order);
3206
3207 return 1;
3208}
3209
3210__setup("slub_min_order=", setup_slub_min_order);
3211
3212static int __init setup_slub_max_order(char *str)
3213{
3214 get_option(&str, &slub_max_order);
3215 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3216
3217 return 1;
3218}
3219
3220__setup("slub_max_order=", setup_slub_max_order);
3221
3222static int __init setup_slub_min_objects(char *str)
3223{
3224 get_option(&str, &slub_min_objects);
3225
3226 return 1;
3227}
3228
3229__setup("slub_min_objects=", setup_slub_min_objects);
3230
3231static int __init setup_slub_nomerge(char *str)
3232{
3233 slub_nomerge = 1;
3234 return 1;
3235}
3236
3237__setup("slub_nomerge", setup_slub_nomerge);
3238
3239static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3240 int size, unsigned int flags)
3241{
3242 struct kmem_cache *s;
3243
3244 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3245
3246 /*
3247 * This function is called with IRQs disabled during early-boot on
3248 * single CPU so there's no need to take slub_lock here.
3249 */
3250 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3251 flags, NULL))
3252 goto panic;
3253
3254 list_add(&s->list, &slab_caches);
3255 return s;
3256
3257panic:
3258 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3259 return NULL;
3260}
3261
3262/*
3263 * Conversion table for small slabs sizes / 8 to the index in the
3264 * kmalloc array. This is necessary for slabs < 192 since we have non power
3265 * of two cache sizes there. The size of larger slabs can be determined using
3266 * fls.
3267 */
3268static s8 size_index[24] = {
3269 3, /* 8 */
3270 4, /* 16 */
3271 5, /* 24 */
3272 5, /* 32 */
3273 6, /* 40 */
3274 6, /* 48 */
3275 6, /* 56 */
3276 6, /* 64 */
3277 1, /* 72 */
3278 1, /* 80 */
3279 1, /* 88 */
3280 1, /* 96 */
3281 7, /* 104 */
3282 7, /* 112 */
3283 7, /* 120 */
3284 7, /* 128 */
3285 2, /* 136 */
3286 2, /* 144 */
3287 2, /* 152 */
3288 2, /* 160 */
3289 2, /* 168 */
3290 2, /* 176 */
3291 2, /* 184 */
3292 2 /* 192 */
3293};
3294
3295static inline int size_index_elem(size_t bytes)
3296{
3297 return (bytes - 1) / 8;
3298}
3299
3300static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3301{
3302 int index;
3303
3304 if (size <= 192) {
3305 if (!size)
3306 return ZERO_SIZE_PTR;
3307
3308 index = size_index[size_index_elem(size)];
3309 } else
3310 index = fls(size - 1);
3311
3312#ifdef CONFIG_ZONE_DMA
3313 if (unlikely((flags & SLUB_DMA)))
3314 return kmalloc_dma_caches[index];
3315
3316#endif
3317 return kmalloc_caches[index];
3318}
3319
3320void *__kmalloc(size_t size, gfp_t flags)
3321{
3322 struct kmem_cache *s;
3323 void *ret;
3324
3325 if (unlikely(size > SLUB_MAX_SIZE))
3326 return kmalloc_large(size, flags);
3327
3328 s = get_slab(size, flags);
3329
3330 if (unlikely(ZERO_OR_NULL_PTR(s)))
3331 return s;
3332
3333 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3334
3335 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3336
3337 return ret;
3338}
3339EXPORT_SYMBOL(__kmalloc);
3340
3341#ifdef CONFIG_NUMA
3342static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3343{
3344 struct page *page;
3345 void *ptr = NULL;
3346
3347 flags |= __GFP_COMP | __GFP_NOTRACK;
3348 page = alloc_pages_node(node, flags, get_order(size));
3349 if (page)
3350 ptr = page_address(page);
3351
3352 kmemleak_alloc(ptr, size, 1, flags);
3353 return ptr;
3354}
3355
3356void *__kmalloc_node(size_t size, gfp_t flags, int node)
3357{
3358 struct kmem_cache *s;
3359 void *ret;
3360
3361 if (unlikely(size > SLUB_MAX_SIZE)) {
3362 ret = kmalloc_large_node(size, flags, node);
3363
3364 trace_kmalloc_node(_RET_IP_, ret,
3365 size, PAGE_SIZE << get_order(size),
3366 flags, node);
3367
3368 return ret;
3369 }
3370
3371 s = get_slab(size, flags);
3372
3373 if (unlikely(ZERO_OR_NULL_PTR(s)))
3374 return s;
3375
3376 ret = slab_alloc(s, flags, node, _RET_IP_);
3377
3378 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3379
3380 return ret;
3381}
3382EXPORT_SYMBOL(__kmalloc_node);
3383#endif
3384
3385size_t ksize(const void *object)
3386{
3387 struct page *page;
3388
3389 if (unlikely(object == ZERO_SIZE_PTR))
3390 return 0;
3391
3392 page = virt_to_head_page(object);
3393
3394 if (unlikely(!PageSlab(page))) {
3395 WARN_ON(!PageCompound(page));
3396 return PAGE_SIZE << compound_order(page);
3397 }
3398
3399 return slab_ksize(page->slab);
3400}
3401EXPORT_SYMBOL(ksize);
3402
3403#ifdef CONFIG_SLUB_DEBUG
3404bool verify_mem_not_deleted(const void *x)
3405{
3406 struct page *page;
3407 void *object = (void *)x;
3408 unsigned long flags;
3409 bool rv;
3410
3411 if (unlikely(ZERO_OR_NULL_PTR(x)))
3412 return false;
3413
3414 local_irq_save(flags);
3415
3416 page = virt_to_head_page(x);
3417 if (unlikely(!PageSlab(page))) {
3418 /* maybe it was from stack? */
3419 rv = true;
3420 goto out_unlock;
3421 }
3422
3423 slab_lock(page);
3424 if (on_freelist(page->slab, page, object)) {
3425 object_err(page->slab, page, object, "Object is on free-list");
3426 rv = false;
3427 } else {
3428 rv = true;
3429 }
3430 slab_unlock(page);
3431
3432out_unlock:
3433 local_irq_restore(flags);
3434 return rv;
3435}
3436EXPORT_SYMBOL(verify_mem_not_deleted);
3437#endif
3438
3439void kfree(const void *x)
3440{
3441 struct page *page;
3442 void *object = (void *)x;
3443
3444 trace_kfree(_RET_IP_, x);
3445
3446 if (unlikely(ZERO_OR_NULL_PTR(x)))
3447 return;
3448
3449 page = virt_to_head_page(x);
3450 if (unlikely(!PageSlab(page))) {
3451 BUG_ON(!PageCompound(page));
3452 kmemleak_free(x);
3453 put_page(page);
3454 return;
3455 }
3456 slab_free(page->slab, page, object, _RET_IP_);
3457}
3458EXPORT_SYMBOL(kfree);
3459
3460/*
3461 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3462 * the remaining slabs by the number of items in use. The slabs with the
3463 * most items in use come first. New allocations will then fill those up
3464 * and thus they can be removed from the partial lists.
3465 *
3466 * The slabs with the least items are placed last. This results in them
3467 * being allocated from last increasing the chance that the last objects
3468 * are freed in them.
3469 */
3470int kmem_cache_shrink(struct kmem_cache *s)
3471{
3472 int node;
3473 int i;
3474 struct kmem_cache_node *n;
3475 struct page *page;
3476 struct page *t;
3477 int objects = oo_objects(s->max);
3478 struct list_head *slabs_by_inuse =
3479 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3480 unsigned long flags;
3481
3482 if (!slabs_by_inuse)
3483 return -ENOMEM;
3484
3485 flush_all(s);
3486 for_each_node_state(node, N_NORMAL_MEMORY) {
3487 n = get_node(s, node);
3488
3489 if (!n->nr_partial)
3490 continue;
3491
3492 for (i = 0; i < objects; i++)
3493 INIT_LIST_HEAD(slabs_by_inuse + i);
3494
3495 spin_lock_irqsave(&n->list_lock, flags);
3496
3497 /*
3498 * Build lists indexed by the items in use in each slab.
3499 *
3500 * Note that concurrent frees may occur while we hold the
3501 * list_lock. page->inuse here is the upper limit.
3502 */
3503 list_for_each_entry_safe(page, t, &n->partial, lru) {
3504 list_move(&page->lru, slabs_by_inuse + page->inuse);
3505 if (!page->inuse)
3506 n->nr_partial--;
3507 }
3508
3509 /*
3510 * Rebuild the partial list with the slabs filled up most
3511 * first and the least used slabs at the end.
3512 */
3513 for (i = objects - 1; i > 0; i--)
3514 list_splice(slabs_by_inuse + i, n->partial.prev);
3515
3516 spin_unlock_irqrestore(&n->list_lock, flags);
3517
3518 /* Release empty slabs */
3519 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3520 discard_slab(s, page);
3521 }
3522
3523 kfree(slabs_by_inuse);
3524 return 0;
3525}
3526EXPORT_SYMBOL(kmem_cache_shrink);
3527
3528#if defined(CONFIG_MEMORY_HOTPLUG)
3529static int slab_mem_going_offline_callback(void *arg)
3530{
3531 struct kmem_cache *s;
3532
3533 down_read(&slub_lock);
3534 list_for_each_entry(s, &slab_caches, list)
3535 kmem_cache_shrink(s);
3536 up_read(&slub_lock);
3537
3538 return 0;
3539}
3540
3541static void slab_mem_offline_callback(void *arg)
3542{
3543 struct kmem_cache_node *n;
3544 struct kmem_cache *s;
3545 struct memory_notify *marg = arg;
3546 int offline_node;
3547
3548 offline_node = marg->status_change_nid;
3549
3550 /*
3551 * If the node still has available memory. we need kmem_cache_node
3552 * for it yet.
3553 */
3554 if (offline_node < 0)
3555 return;
3556
3557 down_read(&slub_lock);
3558 list_for_each_entry(s, &slab_caches, list) {
3559 n = get_node(s, offline_node);
3560 if (n) {
3561 /*
3562 * if n->nr_slabs > 0, slabs still exist on the node
3563 * that is going down. We were unable to free them,
3564 * and offline_pages() function shouldn't call this
3565 * callback. So, we must fail.
3566 */
3567 BUG_ON(slabs_node(s, offline_node));
3568
3569 s->node[offline_node] = NULL;
3570 kmem_cache_free(kmem_cache_node, n);
3571 }
3572 }
3573 up_read(&slub_lock);
3574}
3575
3576static int slab_mem_going_online_callback(void *arg)
3577{
3578 struct kmem_cache_node *n;
3579 struct kmem_cache *s;
3580 struct memory_notify *marg = arg;
3581 int nid = marg->status_change_nid;
3582 int ret = 0;
3583
3584 /*
3585 * If the node's memory is already available, then kmem_cache_node is
3586 * already created. Nothing to do.
3587 */
3588 if (nid < 0)
3589 return 0;
3590
3591 /*
3592 * We are bringing a node online. No memory is available yet. We must
3593 * allocate a kmem_cache_node structure in order to bring the node
3594 * online.
3595 */
3596 down_read(&slub_lock);
3597 list_for_each_entry(s, &slab_caches, list) {
3598 /*
3599 * XXX: kmem_cache_alloc_node will fallback to other nodes
3600 * since memory is not yet available from the node that
3601 * is brought up.
3602 */
3603 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3604 if (!n) {
3605 ret = -ENOMEM;
3606 goto out;
3607 }
3608 init_kmem_cache_node(n, s);
3609 s->node[nid] = n;
3610 }
3611out:
3612 up_read(&slub_lock);
3613 return ret;
3614}
3615
3616static int slab_memory_callback(struct notifier_block *self,
3617 unsigned long action, void *arg)
3618{
3619 int ret = 0;
3620
3621 switch (action) {
3622 case MEM_GOING_ONLINE:
3623 ret = slab_mem_going_online_callback(arg);
3624 break;
3625 case MEM_GOING_OFFLINE:
3626 ret = slab_mem_going_offline_callback(arg);
3627 break;
3628 case MEM_OFFLINE:
3629 case MEM_CANCEL_ONLINE:
3630 slab_mem_offline_callback(arg);
3631 break;
3632 case MEM_ONLINE:
3633 case MEM_CANCEL_OFFLINE:
3634 break;
3635 }
3636 if (ret)
3637 ret = notifier_from_errno(ret);
3638 else
3639 ret = NOTIFY_OK;
3640 return ret;
3641}
3642
3643#endif /* CONFIG_MEMORY_HOTPLUG */
3644
3645/********************************************************************
3646 * Basic setup of slabs
3647 *******************************************************************/
3648
3649/*
3650 * Used for early kmem_cache structures that were allocated using
3651 * the page allocator
3652 */
3653
3654static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3655{
3656 int node;
3657
3658 list_add(&s->list, &slab_caches);
3659 s->refcount = -1;
3660
3661 for_each_node_state(node, N_NORMAL_MEMORY) {
3662 struct kmem_cache_node *n = get_node(s, node);
3663 struct page *p;
3664
3665 if (n) {
3666 list_for_each_entry(p, &n->partial, lru)
3667 p->slab = s;
3668
3669#ifdef CONFIG_SLUB_DEBUG
3670 list_for_each_entry(p, &n->full, lru)
3671 p->slab = s;
3672#endif
3673 }
3674 }
3675}
3676
3677void __init kmem_cache_init(void)
3678{
3679 int i;
3680 int caches = 0;
3681 struct kmem_cache *temp_kmem_cache;
3682 int order;
3683 struct kmem_cache *temp_kmem_cache_node;
3684 unsigned long kmalloc_size;
3685
3686 if (debug_guardpage_minorder())
3687 slub_max_order = 0;
3688
3689 kmem_size = offsetof(struct kmem_cache, node) +
3690 nr_node_ids * sizeof(struct kmem_cache_node *);
3691
3692 /* Allocate two kmem_caches from the page allocator */
3693 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3694 order = get_order(2 * kmalloc_size);
3695 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3696
3697 /*
3698 * Must first have the slab cache available for the allocations of the
3699 * struct kmem_cache_node's. There is special bootstrap code in
3700 * kmem_cache_open for slab_state == DOWN.
3701 */
3702 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3703
3704 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3705 sizeof(struct kmem_cache_node),
3706 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3707
3708 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3709
3710 /* Able to allocate the per node structures */
3711 slab_state = PARTIAL;
3712
3713 temp_kmem_cache = kmem_cache;
3714 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3715 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3716 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3717 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3718
3719 /*
3720 * Allocate kmem_cache_node properly from the kmem_cache slab.
3721 * kmem_cache_node is separately allocated so no need to
3722 * update any list pointers.
3723 */
3724 temp_kmem_cache_node = kmem_cache_node;
3725
3726 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3727 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3728
3729 kmem_cache_bootstrap_fixup(kmem_cache_node);
3730
3731 caches++;
3732 kmem_cache_bootstrap_fixup(kmem_cache);
3733 caches++;
3734 /* Free temporary boot structure */
3735 free_pages((unsigned long)temp_kmem_cache, order);
3736
3737 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3738
3739 /*
3740 * Patch up the size_index table if we have strange large alignment
3741 * requirements for the kmalloc array. This is only the case for
3742 * MIPS it seems. The standard arches will not generate any code here.
3743 *
3744 * Largest permitted alignment is 256 bytes due to the way we
3745 * handle the index determination for the smaller caches.
3746 *
3747 * Make sure that nothing crazy happens if someone starts tinkering
3748 * around with ARCH_KMALLOC_MINALIGN
3749 */
3750 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3751 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3752
3753 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3754 int elem = size_index_elem(i);
3755 if (elem >= ARRAY_SIZE(size_index))
3756 break;
3757 size_index[elem] = KMALLOC_SHIFT_LOW;
3758 }
3759
3760 if (KMALLOC_MIN_SIZE == 64) {
3761 /*
3762 * The 96 byte size cache is not used if the alignment
3763 * is 64 byte.
3764 */
3765 for (i = 64 + 8; i <= 96; i += 8)
3766 size_index[size_index_elem(i)] = 7;
3767 } else if (KMALLOC_MIN_SIZE == 128) {
3768 /*
3769 * The 192 byte sized cache is not used if the alignment
3770 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3771 * instead.
3772 */
3773 for (i = 128 + 8; i <= 192; i += 8)
3774 size_index[size_index_elem(i)] = 8;
3775 }
3776
3777 /* Caches that are not of the two-to-the-power-of size */
3778 if (KMALLOC_MIN_SIZE <= 32) {
3779 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3780 caches++;
3781 }
3782
3783 if (KMALLOC_MIN_SIZE <= 64) {
3784 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3785 caches++;
3786 }
3787
3788 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3789 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3790 caches++;
3791 }
3792
3793 slab_state = UP;
3794
3795 /* Provide the correct kmalloc names now that the caches are up */
3796 if (KMALLOC_MIN_SIZE <= 32) {
3797 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3798 BUG_ON(!kmalloc_caches[1]->name);
3799 }
3800
3801 if (KMALLOC_MIN_SIZE <= 64) {
3802 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3803 BUG_ON(!kmalloc_caches[2]->name);
3804 }
3805
3806 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3807 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3808
3809 BUG_ON(!s);
3810 kmalloc_caches[i]->name = s;
3811 }
3812
3813#ifdef CONFIG_SMP
3814 register_cpu_notifier(&slab_notifier);
3815#endif
3816
3817#ifdef CONFIG_ZONE_DMA
3818 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3819 struct kmem_cache *s = kmalloc_caches[i];
3820
3821 if (s && s->size) {
3822 char *name = kasprintf(GFP_NOWAIT,
3823 "dma-kmalloc-%d", s->objsize);
3824
3825 BUG_ON(!name);
3826 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3827 s->objsize, SLAB_CACHE_DMA);
3828 }
3829 }
3830#endif
3831 printk(KERN_INFO
3832 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3833 " CPUs=%d, Nodes=%d\n",
3834 caches, cache_line_size(),
3835 slub_min_order, slub_max_order, slub_min_objects,
3836 nr_cpu_ids, nr_node_ids);
3837}
3838
3839void __init kmem_cache_init_late(void)
3840{
3841}
3842
3843/*
3844 * Find a mergeable slab cache
3845 */
3846static int slab_unmergeable(struct kmem_cache *s)
3847{
3848 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3849 return 1;
3850
3851 if (s->ctor)
3852 return 1;
3853
3854 /*
3855 * We may have set a slab to be unmergeable during bootstrap.
3856 */
3857 if (s->refcount < 0)
3858 return 1;
3859
3860 return 0;
3861}
3862
3863static struct kmem_cache *find_mergeable(size_t size,
3864 size_t align, unsigned long flags, const char *name,
3865 void (*ctor)(void *))
3866{
3867 struct kmem_cache *s;
3868
3869 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3870 return NULL;
3871
3872 if (ctor)
3873 return NULL;
3874
3875 size = ALIGN(size, sizeof(void *));
3876 align = calculate_alignment(flags, align, size);
3877 size = ALIGN(size, align);
3878 flags = kmem_cache_flags(size, flags, name, NULL);
3879
3880 list_for_each_entry(s, &slab_caches, list) {
3881 if (slab_unmergeable(s))
3882 continue;
3883
3884 if (size > s->size)
3885 continue;
3886
3887 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3888 continue;
3889 /*
3890 * Check if alignment is compatible.
3891 * Courtesy of Adrian Drzewiecki
3892 */
3893 if ((s->size & ~(align - 1)) != s->size)
3894 continue;
3895
3896 if (s->size - size >= sizeof(void *))
3897 continue;
3898
3899 return s;
3900 }
3901 return NULL;
3902}
3903
3904struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3905 size_t align, unsigned long flags, void (*ctor)(void *))
3906{
3907 struct kmem_cache *s;
3908 char *n;
3909
3910 if (WARN_ON(!name))
3911 return NULL;
3912
3913 down_write(&slub_lock);
3914 s = find_mergeable(size, align, flags, name, ctor);
3915 if (s) {
3916 s->refcount++;
3917 /*
3918 * Adjust the object sizes so that we clear
3919 * the complete object on kzalloc.
3920 */
3921 s->objsize = max(s->objsize, (int)size);
3922 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3923
3924 if (sysfs_slab_alias(s, name)) {
3925 s->refcount--;
3926 goto err;
3927 }
3928 up_write(&slub_lock);
3929 return s;
3930 }
3931
3932 n = kstrdup(name, GFP_KERNEL);
3933 if (!n)
3934 goto err;
3935
3936 s = kmalloc(kmem_size, GFP_KERNEL);
3937 if (s) {
3938 if (kmem_cache_open(s, n,
3939 size, align, flags, ctor)) {
3940 list_add(&s->list, &slab_caches);
3941 up_write(&slub_lock);
3942 if (sysfs_slab_add(s)) {
3943 down_write(&slub_lock);
3944 list_del(&s->list);
3945 kfree(n);
3946 kfree(s);
3947 goto err;
3948 }
3949 return s;
3950 }
3951 kfree(n);
3952 kfree(s);
3953 }
3954err:
3955 up_write(&slub_lock);
3956
3957 if (flags & SLAB_PANIC)
3958 panic("Cannot create slabcache %s\n", name);
3959 else
3960 s = NULL;
3961 return s;
3962}
3963EXPORT_SYMBOL(kmem_cache_create);
3964
3965#ifdef CONFIG_SMP
3966/*
3967 * Use the cpu notifier to insure that the cpu slabs are flushed when
3968 * necessary.
3969 */
3970static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3971 unsigned long action, void *hcpu)
3972{
3973 long cpu = (long)hcpu;
3974 struct kmem_cache *s;
3975 unsigned long flags;
3976
3977 switch (action) {
3978 case CPU_UP_CANCELED:
3979 case CPU_UP_CANCELED_FROZEN:
3980 case CPU_DEAD:
3981 case CPU_DEAD_FROZEN:
3982 down_read(&slub_lock);
3983 list_for_each_entry(s, &slab_caches, list) {
3984 local_irq_save(flags);
3985 __flush_cpu_slab(s, cpu);
3986 local_irq_restore(flags);
3987 }
3988 up_read(&slub_lock);
3989 break;
3990 default:
3991 break;
3992 }
3993 return NOTIFY_OK;
3994}
3995
3996static struct notifier_block __cpuinitdata slab_notifier = {
3997 .notifier_call = slab_cpuup_callback
3998};
3999
4000#endif
4001
4002void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4003{
4004 struct kmem_cache *s;
4005 void *ret;
4006
4007 if (unlikely(size > SLUB_MAX_SIZE))
4008 return kmalloc_large(size, gfpflags);
4009
4010 s = get_slab(size, gfpflags);
4011
4012 if (unlikely(ZERO_OR_NULL_PTR(s)))
4013 return s;
4014
4015 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4016
4017 /* Honor the call site pointer we received. */
4018 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4019
4020 return ret;
4021}
4022
4023#ifdef CONFIG_NUMA
4024void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4025 int node, unsigned long caller)
4026{
4027 struct kmem_cache *s;
4028 void *ret;
4029
4030 if (unlikely(size > SLUB_MAX_SIZE)) {
4031 ret = kmalloc_large_node(size, gfpflags, node);
4032
4033 trace_kmalloc_node(caller, ret,
4034 size, PAGE_SIZE << get_order(size),
4035 gfpflags, node);
4036
4037 return ret;
4038 }
4039
4040 s = get_slab(size, gfpflags);
4041
4042 if (unlikely(ZERO_OR_NULL_PTR(s)))
4043 return s;
4044
4045 ret = slab_alloc(s, gfpflags, node, caller);
4046
4047 /* Honor the call site pointer we received. */
4048 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4049
4050 return ret;
4051}
4052#endif
4053
4054#ifdef CONFIG_SYSFS
4055static int count_inuse(struct page *page)
4056{
4057 return page->inuse;
4058}
4059
4060static int count_total(struct page *page)
4061{
4062 return page->objects;
4063}
4064#endif
4065
4066#ifdef CONFIG_SLUB_DEBUG
4067static int validate_slab(struct kmem_cache *s, struct page *page,
4068 unsigned long *map)
4069{
4070 void *p;
4071 void *addr = page_address(page);
4072
4073 if (!check_slab(s, page) ||
4074 !on_freelist(s, page, NULL))
4075 return 0;
4076
4077 /* Now we know that a valid freelist exists */
4078 bitmap_zero(map, page->objects);
4079
4080 get_map(s, page, map);
4081 for_each_object(p, s, addr, page->objects) {
4082 if (test_bit(slab_index(p, s, addr), map))
4083 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4084 return 0;
4085 }
4086
4087 for_each_object(p, s, addr, page->objects)
4088 if (!test_bit(slab_index(p, s, addr), map))
4089 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4090 return 0;
4091 return 1;
4092}
4093
4094static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4095 unsigned long *map)
4096{
4097 slab_lock(page);
4098 validate_slab(s, page, map);
4099 slab_unlock(page);
4100}
4101
4102static int validate_slab_node(struct kmem_cache *s,
4103 struct kmem_cache_node *n, unsigned long *map)
4104{
4105 unsigned long count = 0;
4106 struct page *page;
4107 unsigned long flags;
4108
4109 spin_lock_irqsave(&n->list_lock, flags);
4110
4111 list_for_each_entry(page, &n->partial, lru) {
4112 validate_slab_slab(s, page, map);
4113 count++;
4114 }
4115 if (count != n->nr_partial)
4116 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4117 "counter=%ld\n", s->name, count, n->nr_partial);
4118
4119 if (!(s->flags & SLAB_STORE_USER))
4120 goto out;
4121
4122 list_for_each_entry(page, &n->full, lru) {
4123 validate_slab_slab(s, page, map);
4124 count++;
4125 }
4126 if (count != atomic_long_read(&n->nr_slabs))
4127 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4128 "counter=%ld\n", s->name, count,
4129 atomic_long_read(&n->nr_slabs));
4130
4131out:
4132 spin_unlock_irqrestore(&n->list_lock, flags);
4133 return count;
4134}
4135
4136static long validate_slab_cache(struct kmem_cache *s)
4137{
4138 int node;
4139 unsigned long count = 0;
4140 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4141 sizeof(unsigned long), GFP_KERNEL);
4142
4143 if (!map)
4144 return -ENOMEM;
4145
4146 flush_all(s);
4147 for_each_node_state(node, N_NORMAL_MEMORY) {
4148 struct kmem_cache_node *n = get_node(s, node);
4149
4150 count += validate_slab_node(s, n, map);
4151 }
4152 kfree(map);
4153 return count;
4154}
4155/*
4156 * Generate lists of code addresses where slabcache objects are allocated
4157 * and freed.
4158 */
4159
4160struct location {
4161 unsigned long count;
4162 unsigned long addr;
4163 long long sum_time;
4164 long min_time;
4165 long max_time;
4166 long min_pid;
4167 long max_pid;
4168 DECLARE_BITMAP(cpus, NR_CPUS);
4169 nodemask_t nodes;
4170};
4171
4172struct loc_track {
4173 unsigned long max;
4174 unsigned long count;
4175 struct location *loc;
4176};
4177
4178static void free_loc_track(struct loc_track *t)
4179{
4180 if (t->max)
4181 free_pages((unsigned long)t->loc,
4182 get_order(sizeof(struct location) * t->max));
4183}
4184
4185static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4186{
4187 struct location *l;
4188 int order;
4189
4190 order = get_order(sizeof(struct location) * max);
4191
4192 l = (void *)__get_free_pages(flags, order);
4193 if (!l)
4194 return 0;
4195
4196 if (t->count) {
4197 memcpy(l, t->loc, sizeof(struct location) * t->count);
4198 free_loc_track(t);
4199 }
4200 t->max = max;
4201 t->loc = l;
4202 return 1;
4203}
4204
4205static int add_location(struct loc_track *t, struct kmem_cache *s,
4206 const struct track *track)
4207{
4208 long start, end, pos;
4209 struct location *l;
4210 unsigned long caddr;
4211 unsigned long age = jiffies - track->when;
4212
4213 start = -1;
4214 end = t->count;
4215
4216 for ( ; ; ) {
4217 pos = start + (end - start + 1) / 2;
4218
4219 /*
4220 * There is nothing at "end". If we end up there
4221 * we need to add something to before end.
4222 */
4223 if (pos == end)
4224 break;
4225
4226 caddr = t->loc[pos].addr;
4227 if (track->addr == caddr) {
4228
4229 l = &t->loc[pos];
4230 l->count++;
4231 if (track->when) {
4232 l->sum_time += age;
4233 if (age < l->min_time)
4234 l->min_time = age;
4235 if (age > l->max_time)
4236 l->max_time = age;
4237
4238 if (track->pid < l->min_pid)
4239 l->min_pid = track->pid;
4240 if (track->pid > l->max_pid)
4241 l->max_pid = track->pid;
4242
4243 cpumask_set_cpu(track->cpu,
4244 to_cpumask(l->cpus));
4245 }
4246 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4247 return 1;
4248 }
4249
4250 if (track->addr < caddr)
4251 end = pos;
4252 else
4253 start = pos;
4254 }
4255
4256 /*
4257 * Not found. Insert new tracking element.
4258 */
4259 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4260 return 0;
4261
4262 l = t->loc + pos;
4263 if (pos < t->count)
4264 memmove(l + 1, l,
4265 (t->count - pos) * sizeof(struct location));
4266 t->count++;
4267 l->count = 1;
4268 l->addr = track->addr;
4269 l->sum_time = age;
4270 l->min_time = age;
4271 l->max_time = age;
4272 l->min_pid = track->pid;
4273 l->max_pid = track->pid;
4274 cpumask_clear(to_cpumask(l->cpus));
4275 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4276 nodes_clear(l->nodes);
4277 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4278 return 1;
4279}
4280
4281static void process_slab(struct loc_track *t, struct kmem_cache *s,
4282 struct page *page, enum track_item alloc,
4283 unsigned long *map)
4284{
4285 void *addr = page_address(page);
4286 void *p;
4287
4288 bitmap_zero(map, page->objects);
4289 get_map(s, page, map);
4290
4291 for_each_object(p, s, addr, page->objects)
4292 if (!test_bit(slab_index(p, s, addr), map))
4293 add_location(t, s, get_track(s, p, alloc));
4294}
4295
4296static int list_locations(struct kmem_cache *s, char *buf,
4297 enum track_item alloc)
4298{
4299 int len = 0;
4300 unsigned long i;
4301 struct loc_track t = { 0, 0, NULL };
4302 int node;
4303 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4304 sizeof(unsigned long), GFP_KERNEL);
4305
4306 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4307 GFP_TEMPORARY)) {
4308 kfree(map);
4309 return sprintf(buf, "Out of memory\n");
4310 }
4311 /* Push back cpu slabs */
4312 flush_all(s);
4313
4314 for_each_node_state(node, N_NORMAL_MEMORY) {
4315 struct kmem_cache_node *n = get_node(s, node);
4316 unsigned long flags;
4317 struct page *page;
4318
4319 if (!atomic_long_read(&n->nr_slabs))
4320 continue;
4321
4322 spin_lock_irqsave(&n->list_lock, flags);
4323 list_for_each_entry(page, &n->partial, lru)
4324 process_slab(&t, s, page, alloc, map);
4325 list_for_each_entry(page, &n->full, lru)
4326 process_slab(&t, s, page, alloc, map);
4327 spin_unlock_irqrestore(&n->list_lock, flags);
4328 }
4329
4330 for (i = 0; i < t.count; i++) {
4331 struct location *l = &t.loc[i];
4332
4333 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4334 break;
4335 len += sprintf(buf + len, "%7ld ", l->count);
4336
4337 if (l->addr)
4338 len += sprintf(buf + len, "%pS", (void *)l->addr);
4339 else
4340 len += sprintf(buf + len, "<not-available>");
4341
4342 if (l->sum_time != l->min_time) {
4343 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4344 l->min_time,
4345 (long)div_u64(l->sum_time, l->count),
4346 l->max_time);
4347 } else
4348 len += sprintf(buf + len, " age=%ld",
4349 l->min_time);
4350
4351 if (l->min_pid != l->max_pid)
4352 len += sprintf(buf + len, " pid=%ld-%ld",
4353 l->min_pid, l->max_pid);
4354 else
4355 len += sprintf(buf + len, " pid=%ld",
4356 l->min_pid);
4357
4358 if (num_online_cpus() > 1 &&
4359 !cpumask_empty(to_cpumask(l->cpus)) &&
4360 len < PAGE_SIZE - 60) {
4361 len += sprintf(buf + len, " cpus=");
4362 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4363 to_cpumask(l->cpus));
4364 }
4365
4366 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4367 len < PAGE_SIZE - 60) {
4368 len += sprintf(buf + len, " nodes=");
4369 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4370 l->nodes);
4371 }
4372
4373 len += sprintf(buf + len, "\n");
4374 }
4375
4376 free_loc_track(&t);
4377 kfree(map);
4378 if (!t.count)
4379 len += sprintf(buf, "No data\n");
4380 return len;
4381}
4382#endif
4383
4384#ifdef SLUB_RESILIENCY_TEST
4385static void resiliency_test(void)
4386{
4387 u8 *p;
4388
4389 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4390
4391 printk(KERN_ERR "SLUB resiliency testing\n");
4392 printk(KERN_ERR "-----------------------\n");
4393 printk(KERN_ERR "A. Corruption after allocation\n");
4394
4395 p = kzalloc(16, GFP_KERNEL);
4396 p[16] = 0x12;
4397 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4398 " 0x12->0x%p\n\n", p + 16);
4399
4400 validate_slab_cache(kmalloc_caches[4]);
4401
4402 /* Hmmm... The next two are dangerous */
4403 p = kzalloc(32, GFP_KERNEL);
4404 p[32 + sizeof(void *)] = 0x34;
4405 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4406 " 0x34 -> -0x%p\n", p);
4407 printk(KERN_ERR
4408 "If allocated object is overwritten then not detectable\n\n");
4409
4410 validate_slab_cache(kmalloc_caches[5]);
4411 p = kzalloc(64, GFP_KERNEL);
4412 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4413 *p = 0x56;
4414 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4415 p);
4416 printk(KERN_ERR
4417 "If allocated object is overwritten then not detectable\n\n");
4418 validate_slab_cache(kmalloc_caches[6]);
4419
4420 printk(KERN_ERR "\nB. Corruption after free\n");
4421 p = kzalloc(128, GFP_KERNEL);
4422 kfree(p);
4423 *p = 0x78;
4424 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4425 validate_slab_cache(kmalloc_caches[7]);
4426
4427 p = kzalloc(256, GFP_KERNEL);
4428 kfree(p);
4429 p[50] = 0x9a;
4430 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4431 p);
4432 validate_slab_cache(kmalloc_caches[8]);
4433
4434 p = kzalloc(512, GFP_KERNEL);
4435 kfree(p);
4436 p[512] = 0xab;
4437 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4438 validate_slab_cache(kmalloc_caches[9]);
4439}
4440#else
4441#ifdef CONFIG_SYSFS
4442static void resiliency_test(void) {};
4443#endif
4444#endif
4445
4446#ifdef CONFIG_SYSFS
4447enum slab_stat_type {
4448 SL_ALL, /* All slabs */
4449 SL_PARTIAL, /* Only partially allocated slabs */
4450 SL_CPU, /* Only slabs used for cpu caches */
4451 SL_OBJECTS, /* Determine allocated objects not slabs */
4452 SL_TOTAL /* Determine object capacity not slabs */
4453};
4454
4455#define SO_ALL (1 << SL_ALL)
4456#define SO_PARTIAL (1 << SL_PARTIAL)
4457#define SO_CPU (1 << SL_CPU)
4458#define SO_OBJECTS (1 << SL_OBJECTS)
4459#define SO_TOTAL (1 << SL_TOTAL)
4460
4461static ssize_t show_slab_objects(struct kmem_cache *s,
4462 char *buf, unsigned long flags)
4463{
4464 unsigned long total = 0;
4465 int node;
4466 int x;
4467 unsigned long *nodes;
4468 unsigned long *per_cpu;
4469
4470 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4471 if (!nodes)
4472 return -ENOMEM;
4473 per_cpu = nodes + nr_node_ids;
4474
4475 if (flags & SO_CPU) {
4476 int cpu;
4477
4478 for_each_possible_cpu(cpu) {
4479 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4480 int node = ACCESS_ONCE(c->node);
4481 struct page *page;
4482
4483 if (node < 0)
4484 continue;
4485 page = ACCESS_ONCE(c->page);
4486 if (page) {
4487 if (flags & SO_TOTAL)
4488 x = page->objects;
4489 else if (flags & SO_OBJECTS)
4490 x = page->inuse;
4491 else
4492 x = 1;
4493
4494 total += x;
4495 nodes[node] += x;
4496 }
4497 page = c->partial;
4498
4499 if (page) {
4500 node = page_to_nid(page);
4501 if (flags & SO_TOTAL)
4502 WARN_ON_ONCE(1);
4503 else if (flags & SO_OBJECTS)
4504 WARN_ON_ONCE(1);
4505 else
4506 x = page->pages;
4507 total += x;
4508 nodes[node] += x;
4509 }
4510 per_cpu[node]++;
4511 }
4512 }
4513
4514 lock_memory_hotplug();
4515#ifdef CONFIG_SLUB_DEBUG
4516 if (flags & SO_ALL) {
4517 for_each_node_state(node, N_NORMAL_MEMORY) {
4518 struct kmem_cache_node *n = get_node(s, node);
4519
4520 if (flags & SO_TOTAL)
4521 x = atomic_long_read(&n->total_objects);
4522 else if (flags & SO_OBJECTS)
4523 x = atomic_long_read(&n->total_objects) -
4524 count_partial(n, count_free);
4525
4526 else
4527 x = atomic_long_read(&n->nr_slabs);
4528 total += x;
4529 nodes[node] += x;
4530 }
4531
4532 } else
4533#endif
4534 if (flags & SO_PARTIAL) {
4535 for_each_node_state(node, N_NORMAL_MEMORY) {
4536 struct kmem_cache_node *n = get_node(s, node);
4537
4538 if (flags & SO_TOTAL)
4539 x = count_partial(n, count_total);
4540 else if (flags & SO_OBJECTS)
4541 x = count_partial(n, count_inuse);
4542 else
4543 x = n->nr_partial;
4544 total += x;
4545 nodes[node] += x;
4546 }
4547 }
4548 x = sprintf(buf, "%lu", total);
4549#ifdef CONFIG_NUMA
4550 for_each_node_state(node, N_NORMAL_MEMORY)
4551 if (nodes[node])
4552 x += sprintf(buf + x, " N%d=%lu",
4553 node, nodes[node]);
4554#endif
4555 unlock_memory_hotplug();
4556 kfree(nodes);
4557 return x + sprintf(buf + x, "\n");
4558}
4559
4560#ifdef CONFIG_SLUB_DEBUG
4561static int any_slab_objects(struct kmem_cache *s)
4562{
4563 int node;
4564
4565 for_each_online_node(node) {
4566 struct kmem_cache_node *n = get_node(s, node);
4567
4568 if (!n)
4569 continue;
4570
4571 if (atomic_long_read(&n->total_objects))
4572 return 1;
4573 }
4574 return 0;
4575}
4576#endif
4577
4578#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4579#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4580
4581struct slab_attribute {
4582 struct attribute attr;
4583 ssize_t (*show)(struct kmem_cache *s, char *buf);
4584 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4585};
4586
4587#define SLAB_ATTR_RO(_name) \
4588 static struct slab_attribute _name##_attr = \
4589 __ATTR(_name, 0400, _name##_show, NULL)
4590
4591#define SLAB_ATTR(_name) \
4592 static struct slab_attribute _name##_attr = \
4593 __ATTR(_name, 0600, _name##_show, _name##_store)
4594
4595static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4596{
4597 return sprintf(buf, "%d\n", s->size);
4598}
4599SLAB_ATTR_RO(slab_size);
4600
4601static ssize_t align_show(struct kmem_cache *s, char *buf)
4602{
4603 return sprintf(buf, "%d\n", s->align);
4604}
4605SLAB_ATTR_RO(align);
4606
4607static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4608{
4609 return sprintf(buf, "%d\n", s->objsize);
4610}
4611SLAB_ATTR_RO(object_size);
4612
4613static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4614{
4615 return sprintf(buf, "%d\n", oo_objects(s->oo));
4616}
4617SLAB_ATTR_RO(objs_per_slab);
4618
4619static ssize_t order_store(struct kmem_cache *s,
4620 const char *buf, size_t length)
4621{
4622 unsigned long order;
4623 int err;
4624
4625 err = strict_strtoul(buf, 10, &order);
4626 if (err)
4627 return err;
4628
4629 if (order > slub_max_order || order < slub_min_order)
4630 return -EINVAL;
4631
4632 calculate_sizes(s, order);
4633 return length;
4634}
4635
4636static ssize_t order_show(struct kmem_cache *s, char *buf)
4637{
4638 return sprintf(buf, "%d\n", oo_order(s->oo));
4639}
4640SLAB_ATTR(order);
4641
4642static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4643{
4644 return sprintf(buf, "%lu\n", s->min_partial);
4645}
4646
4647static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4648 size_t length)
4649{
4650 unsigned long min;
4651 int err;
4652
4653 err = strict_strtoul(buf, 10, &min);
4654 if (err)
4655 return err;
4656
4657 set_min_partial(s, min);
4658 return length;
4659}
4660SLAB_ATTR(min_partial);
4661
4662static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4663{
4664 return sprintf(buf, "%u\n", s->cpu_partial);
4665}
4666
4667static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4668 size_t length)
4669{
4670 unsigned long objects;
4671 int err;
4672
4673 err = strict_strtoul(buf, 10, &objects);
4674 if (err)
4675 return err;
4676 if (objects && kmem_cache_debug(s))
4677 return -EINVAL;
4678
4679 s->cpu_partial = objects;
4680 flush_all(s);
4681 return length;
4682}
4683SLAB_ATTR(cpu_partial);
4684
4685static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4686{
4687 if (!s->ctor)
4688 return 0;
4689 return sprintf(buf, "%pS\n", s->ctor);
4690}
4691SLAB_ATTR_RO(ctor);
4692
4693static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4694{
4695 return sprintf(buf, "%d\n", s->refcount - 1);
4696}
4697SLAB_ATTR_RO(aliases);
4698
4699static ssize_t partial_show(struct kmem_cache *s, char *buf)
4700{
4701 return show_slab_objects(s, buf, SO_PARTIAL);
4702}
4703SLAB_ATTR_RO(partial);
4704
4705static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4706{
4707 return show_slab_objects(s, buf, SO_CPU);
4708}
4709SLAB_ATTR_RO(cpu_slabs);
4710
4711static ssize_t objects_show(struct kmem_cache *s, char *buf)
4712{
4713 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4714}
4715SLAB_ATTR_RO(objects);
4716
4717static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4718{
4719 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4720}
4721SLAB_ATTR_RO(objects_partial);
4722
4723static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4724{
4725 int objects = 0;
4726 int pages = 0;
4727 int cpu;
4728 int len;
4729
4730 for_each_online_cpu(cpu) {
4731 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4732
4733 if (page) {
4734 pages += page->pages;
4735 objects += page->pobjects;
4736 }
4737 }
4738
4739 len = sprintf(buf, "%d(%d)", objects, pages);
4740
4741#ifdef CONFIG_SMP
4742 for_each_online_cpu(cpu) {
4743 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4744
4745 if (page && len < PAGE_SIZE - 20)
4746 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4747 page->pobjects, page->pages);
4748 }
4749#endif
4750 return len + sprintf(buf + len, "\n");
4751}
4752SLAB_ATTR_RO(slabs_cpu_partial);
4753
4754static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4755{
4756 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4757}
4758
4759static ssize_t reclaim_account_store(struct kmem_cache *s,
4760 const char *buf, size_t length)
4761{
4762 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4763 if (buf[0] == '1')
4764 s->flags |= SLAB_RECLAIM_ACCOUNT;
4765 return length;
4766}
4767SLAB_ATTR(reclaim_account);
4768
4769static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4770{
4771 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4772}
4773SLAB_ATTR_RO(hwcache_align);
4774
4775#ifdef CONFIG_ZONE_DMA
4776static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4777{
4778 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4779}
4780SLAB_ATTR_RO(cache_dma);
4781#endif
4782
4783static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4784{
4785 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4786}
4787SLAB_ATTR_RO(destroy_by_rcu);
4788
4789static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4790{
4791 return sprintf(buf, "%d\n", s->reserved);
4792}
4793SLAB_ATTR_RO(reserved);
4794
4795#ifdef CONFIG_SLUB_DEBUG
4796static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4797{
4798 return show_slab_objects(s, buf, SO_ALL);
4799}
4800SLAB_ATTR_RO(slabs);
4801
4802static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4803{
4804 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4805}
4806SLAB_ATTR_RO(total_objects);
4807
4808static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4809{
4810 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4811}
4812
4813static ssize_t sanity_checks_store(struct kmem_cache *s,
4814 const char *buf, size_t length)
4815{
4816 s->flags &= ~SLAB_DEBUG_FREE;
4817 if (buf[0] == '1') {
4818 s->flags &= ~__CMPXCHG_DOUBLE;
4819 s->flags |= SLAB_DEBUG_FREE;
4820 }
4821 return length;
4822}
4823SLAB_ATTR(sanity_checks);
4824
4825static ssize_t trace_show(struct kmem_cache *s, char *buf)
4826{
4827 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4828}
4829
4830static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4831 size_t length)
4832{
4833 s->flags &= ~SLAB_TRACE;
4834 if (buf[0] == '1') {
4835 s->flags &= ~__CMPXCHG_DOUBLE;
4836 s->flags |= SLAB_TRACE;
4837 }
4838 return length;
4839}
4840SLAB_ATTR(trace);
4841
4842static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4843{
4844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4845}
4846
4847static ssize_t red_zone_store(struct kmem_cache *s,
4848 const char *buf, size_t length)
4849{
4850 if (any_slab_objects(s))
4851 return -EBUSY;
4852
4853 s->flags &= ~SLAB_RED_ZONE;
4854 if (buf[0] == '1') {
4855 s->flags &= ~__CMPXCHG_DOUBLE;
4856 s->flags |= SLAB_RED_ZONE;
4857 }
4858 calculate_sizes(s, -1);
4859 return length;
4860}
4861SLAB_ATTR(red_zone);
4862
4863static ssize_t poison_show(struct kmem_cache *s, char *buf)
4864{
4865 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4866}
4867
4868static ssize_t poison_store(struct kmem_cache *s,
4869 const char *buf, size_t length)
4870{
4871 if (any_slab_objects(s))
4872 return -EBUSY;
4873
4874 s->flags &= ~SLAB_POISON;
4875 if (buf[0] == '1') {
4876 s->flags &= ~__CMPXCHG_DOUBLE;
4877 s->flags |= SLAB_POISON;
4878 }
4879 calculate_sizes(s, -1);
4880 return length;
4881}
4882SLAB_ATTR(poison);
4883
4884static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4885{
4886 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4887}
4888
4889static ssize_t store_user_store(struct kmem_cache *s,
4890 const char *buf, size_t length)
4891{
4892 if (any_slab_objects(s))
4893 return -EBUSY;
4894
4895 s->flags &= ~SLAB_STORE_USER;
4896 if (buf[0] == '1') {
4897 s->flags &= ~__CMPXCHG_DOUBLE;
4898 s->flags |= SLAB_STORE_USER;
4899 }
4900 calculate_sizes(s, -1);
4901 return length;
4902}
4903SLAB_ATTR(store_user);
4904
4905static ssize_t validate_show(struct kmem_cache *s, char *buf)
4906{
4907 return 0;
4908}
4909
4910static ssize_t validate_store(struct kmem_cache *s,
4911 const char *buf, size_t length)
4912{
4913 int ret = -EINVAL;
4914
4915 if (buf[0] == '1') {
4916 ret = validate_slab_cache(s);
4917 if (ret >= 0)
4918 ret = length;
4919 }
4920 return ret;
4921}
4922SLAB_ATTR(validate);
4923
4924static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4925{
4926 if (!(s->flags & SLAB_STORE_USER))
4927 return -ENOSYS;
4928 return list_locations(s, buf, TRACK_ALLOC);
4929}
4930SLAB_ATTR_RO(alloc_calls);
4931
4932static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4933{
4934 if (!(s->flags & SLAB_STORE_USER))
4935 return -ENOSYS;
4936 return list_locations(s, buf, TRACK_FREE);
4937}
4938SLAB_ATTR_RO(free_calls);
4939#endif /* CONFIG_SLUB_DEBUG */
4940
4941#ifdef CONFIG_FAILSLAB
4942static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4943{
4944 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4945}
4946
4947static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4948 size_t length)
4949{
4950 s->flags &= ~SLAB_FAILSLAB;
4951 if (buf[0] == '1')
4952 s->flags |= SLAB_FAILSLAB;
4953 return length;
4954}
4955SLAB_ATTR(failslab);
4956#endif
4957
4958static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4959{
4960 return 0;
4961}
4962
4963static ssize_t shrink_store(struct kmem_cache *s,
4964 const char *buf, size_t length)
4965{
4966 if (buf[0] == '1') {
4967 int rc = kmem_cache_shrink(s);
4968
4969 if (rc)
4970 return rc;
4971 } else
4972 return -EINVAL;
4973 return length;
4974}
4975SLAB_ATTR(shrink);
4976
4977#ifdef CONFIG_NUMA
4978static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4979{
4980 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4981}
4982
4983static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4984 const char *buf, size_t length)
4985{
4986 unsigned long ratio;
4987 int err;
4988
4989 err = strict_strtoul(buf, 10, &ratio);
4990 if (err)
4991 return err;
4992
4993 if (ratio <= 100)
4994 s->remote_node_defrag_ratio = ratio * 10;
4995
4996 return length;
4997}
4998SLAB_ATTR(remote_node_defrag_ratio);
4999#endif
5000
5001#ifdef CONFIG_SLUB_STATS
5002static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5003{
5004 unsigned long sum = 0;
5005 int cpu;
5006 int len;
5007 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5008
5009 if (!data)
5010 return -ENOMEM;
5011
5012 for_each_online_cpu(cpu) {
5013 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5014
5015 data[cpu] = x;
5016 sum += x;
5017 }
5018
5019 len = sprintf(buf, "%lu", sum);
5020
5021#ifdef CONFIG_SMP
5022 for_each_online_cpu(cpu) {
5023 if (data[cpu] && len < PAGE_SIZE - 20)
5024 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5025 }
5026#endif
5027 kfree(data);
5028 return len + sprintf(buf + len, "\n");
5029}
5030
5031static void clear_stat(struct kmem_cache *s, enum stat_item si)
5032{
5033 int cpu;
5034
5035 for_each_online_cpu(cpu)
5036 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5037}
5038
5039#define STAT_ATTR(si, text) \
5040static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5041{ \
5042 return show_stat(s, buf, si); \
5043} \
5044static ssize_t text##_store(struct kmem_cache *s, \
5045 const char *buf, size_t length) \
5046{ \
5047 if (buf[0] != '0') \
5048 return -EINVAL; \
5049 clear_stat(s, si); \
5050 return length; \
5051} \
5052SLAB_ATTR(text); \
5053
5054STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5055STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5056STAT_ATTR(FREE_FASTPATH, free_fastpath);
5057STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5058STAT_ATTR(FREE_FROZEN, free_frozen);
5059STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5060STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5061STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5062STAT_ATTR(ALLOC_SLAB, alloc_slab);
5063STAT_ATTR(ALLOC_REFILL, alloc_refill);
5064STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5065STAT_ATTR(FREE_SLAB, free_slab);
5066STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5067STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5068STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5069STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5070STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5071STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5072STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5073STAT_ATTR(ORDER_FALLBACK, order_fallback);
5074STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5075STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5076STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5077STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5078STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5079STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5080#endif
5081
5082static struct attribute *slab_attrs[] = {
5083 &slab_size_attr.attr,
5084 &object_size_attr.attr,
5085 &objs_per_slab_attr.attr,
5086 &order_attr.attr,
5087 &min_partial_attr.attr,
5088 &cpu_partial_attr.attr,
5089 &objects_attr.attr,
5090 &objects_partial_attr.attr,
5091 &partial_attr.attr,
5092 &cpu_slabs_attr.attr,
5093 &ctor_attr.attr,
5094 &aliases_attr.attr,
5095 &align_attr.attr,
5096 &hwcache_align_attr.attr,
5097 &reclaim_account_attr.attr,
5098 &destroy_by_rcu_attr.attr,
5099 &shrink_attr.attr,
5100 &reserved_attr.attr,
5101 &slabs_cpu_partial_attr.attr,
5102#ifdef CONFIG_SLUB_DEBUG
5103 &total_objects_attr.attr,
5104 &slabs_attr.attr,
5105 &sanity_checks_attr.attr,
5106 &trace_attr.attr,
5107 &red_zone_attr.attr,
5108 &poison_attr.attr,
5109 &store_user_attr.attr,
5110 &validate_attr.attr,
5111 &alloc_calls_attr.attr,
5112 &free_calls_attr.attr,
5113#endif
5114#ifdef CONFIG_ZONE_DMA
5115 &cache_dma_attr.attr,
5116#endif
5117#ifdef CONFIG_NUMA
5118 &remote_node_defrag_ratio_attr.attr,
5119#endif
5120#ifdef CONFIG_SLUB_STATS
5121 &alloc_fastpath_attr.attr,
5122 &alloc_slowpath_attr.attr,
5123 &free_fastpath_attr.attr,
5124 &free_slowpath_attr.attr,
5125 &free_frozen_attr.attr,
5126 &free_add_partial_attr.attr,
5127 &free_remove_partial_attr.attr,
5128 &alloc_from_partial_attr.attr,
5129 &alloc_slab_attr.attr,
5130 &alloc_refill_attr.attr,
5131 &alloc_node_mismatch_attr.attr,
5132 &free_slab_attr.attr,
5133 &cpuslab_flush_attr.attr,
5134 &deactivate_full_attr.attr,
5135 &deactivate_empty_attr.attr,
5136 &deactivate_to_head_attr.attr,
5137 &deactivate_to_tail_attr.attr,
5138 &deactivate_remote_frees_attr.attr,
5139 &deactivate_bypass_attr.attr,
5140 &order_fallback_attr.attr,
5141 &cmpxchg_double_fail_attr.attr,
5142 &cmpxchg_double_cpu_fail_attr.attr,
5143 &cpu_partial_alloc_attr.attr,
5144 &cpu_partial_free_attr.attr,
5145 &cpu_partial_node_attr.attr,
5146 &cpu_partial_drain_attr.attr,
5147#endif
5148#ifdef CONFIG_FAILSLAB
5149 &failslab_attr.attr,
5150#endif
5151
5152 NULL
5153};
5154
5155static struct attribute_group slab_attr_group = {
5156 .attrs = slab_attrs,
5157};
5158
5159static ssize_t slab_attr_show(struct kobject *kobj,
5160 struct attribute *attr,
5161 char *buf)
5162{
5163 struct slab_attribute *attribute;
5164 struct kmem_cache *s;
5165 int err;
5166
5167 attribute = to_slab_attr(attr);
5168 s = to_slab(kobj);
5169
5170 if (!attribute->show)
5171 return -EIO;
5172
5173 err = attribute->show(s, buf);
5174
5175 return err;
5176}
5177
5178static ssize_t slab_attr_store(struct kobject *kobj,
5179 struct attribute *attr,
5180 const char *buf, size_t len)
5181{
5182 struct slab_attribute *attribute;
5183 struct kmem_cache *s;
5184 int err;
5185
5186 attribute = to_slab_attr(attr);
5187 s = to_slab(kobj);
5188
5189 if (!attribute->store)
5190 return -EIO;
5191
5192 err = attribute->store(s, buf, len);
5193
5194 return err;
5195}
5196
5197static void kmem_cache_release(struct kobject *kobj)
5198{
5199 struct kmem_cache *s = to_slab(kobj);
5200
5201 kfree(s->name);
5202 kfree(s);
5203}
5204
5205static const struct sysfs_ops slab_sysfs_ops = {
5206 .show = slab_attr_show,
5207 .store = slab_attr_store,
5208};
5209
5210static struct kobj_type slab_ktype = {
5211 .sysfs_ops = &slab_sysfs_ops,
5212 .release = kmem_cache_release
5213};
5214
5215static int uevent_filter(struct kset *kset, struct kobject *kobj)
5216{
5217 struct kobj_type *ktype = get_ktype(kobj);
5218
5219 if (ktype == &slab_ktype)
5220 return 1;
5221 return 0;
5222}
5223
5224static const struct kset_uevent_ops slab_uevent_ops = {
5225 .filter = uevent_filter,
5226};
5227
5228static struct kset *slab_kset;
5229
5230#define ID_STR_LENGTH 64
5231
5232/* Create a unique string id for a slab cache:
5233 *
5234 * Format :[flags-]size
5235 */
5236static char *create_unique_id(struct kmem_cache *s)
5237{
5238 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5239 char *p = name;
5240
5241 BUG_ON(!name);
5242
5243 *p++ = ':';
5244 /*
5245 * First flags affecting slabcache operations. We will only
5246 * get here for aliasable slabs so we do not need to support
5247 * too many flags. The flags here must cover all flags that
5248 * are matched during merging to guarantee that the id is
5249 * unique.
5250 */
5251 if (s->flags & SLAB_CACHE_DMA)
5252 *p++ = 'd';
5253 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5254 *p++ = 'a';
5255 if (s->flags & SLAB_DEBUG_FREE)
5256 *p++ = 'F';
5257 if (!(s->flags & SLAB_NOTRACK))
5258 *p++ = 't';
5259 if (p != name + 1)
5260 *p++ = '-';
5261 p += sprintf(p, "%07d", s->size);
5262 BUG_ON(p > name + ID_STR_LENGTH - 1);
5263 return name;
5264}
5265
5266static int sysfs_slab_add(struct kmem_cache *s)
5267{
5268 int err;
5269 const char *name;
5270 int unmergeable;
5271
5272 if (slab_state < SYSFS)
5273 /* Defer until later */
5274 return 0;
5275
5276 unmergeable = slab_unmergeable(s);
5277 if (unmergeable) {
5278 /*
5279 * Slabcache can never be merged so we can use the name proper.
5280 * This is typically the case for debug situations. In that
5281 * case we can catch duplicate names easily.
5282 */
5283 sysfs_remove_link(&slab_kset->kobj, s->name);
5284 name = s->name;
5285 } else {
5286 /*
5287 * Create a unique name for the slab as a target
5288 * for the symlinks.
5289 */
5290 name = create_unique_id(s);
5291 }
5292
5293 s->kobj.kset = slab_kset;
5294 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5295 if (err) {
5296 kobject_put(&s->kobj);
5297 return err;
5298 }
5299
5300 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5301 if (err) {
5302 kobject_del(&s->kobj);
5303 kobject_put(&s->kobj);
5304 return err;
5305 }
5306 kobject_uevent(&s->kobj, KOBJ_ADD);
5307 if (!unmergeable) {
5308 /* Setup first alias */
5309 sysfs_slab_alias(s, s->name);
5310 kfree(name);
5311 }
5312 return 0;
5313}
5314
5315static void sysfs_slab_remove(struct kmem_cache *s)
5316{
5317 if (slab_state < SYSFS)
5318 /*
5319 * Sysfs has not been setup yet so no need to remove the
5320 * cache from sysfs.
5321 */
5322 return;
5323
5324 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5325 kobject_del(&s->kobj);
5326 kobject_put(&s->kobj);
5327}
5328
5329/*
5330 * Need to buffer aliases during bootup until sysfs becomes
5331 * available lest we lose that information.
5332 */
5333struct saved_alias {
5334 struct kmem_cache *s;
5335 const char *name;
5336 struct saved_alias *next;
5337};
5338
5339static struct saved_alias *alias_list;
5340
5341static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5342{
5343 struct saved_alias *al;
5344
5345 if (slab_state == SYSFS) {
5346 /*
5347 * If we have a leftover link then remove it.
5348 */
5349 sysfs_remove_link(&slab_kset->kobj, name);
5350 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5351 }
5352
5353 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5354 if (!al)
5355 return -ENOMEM;
5356
5357 al->s = s;
5358 al->name = name;
5359 al->next = alias_list;
5360 alias_list = al;
5361 return 0;
5362}
5363
5364static int __init slab_sysfs_init(void)
5365{
5366 struct kmem_cache *s;
5367 int err;
5368
5369 down_write(&slub_lock);
5370
5371 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5372 if (!slab_kset) {
5373 up_write(&slub_lock);
5374 printk(KERN_ERR "Cannot register slab subsystem.\n");
5375 return -ENOSYS;
5376 }
5377
5378 slab_state = SYSFS;
5379
5380 list_for_each_entry(s, &slab_caches, list) {
5381 err = sysfs_slab_add(s);
5382 if (err)
5383 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5384 " to sysfs\n", s->name);
5385 }
5386
5387 while (alias_list) {
5388 struct saved_alias *al = alias_list;
5389
5390 alias_list = alias_list->next;
5391 err = sysfs_slab_alias(al->s, al->name);
5392 if (err)
5393 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5394 " %s to sysfs\n", s->name);
5395 kfree(al);
5396 }
5397
5398 up_write(&slub_lock);
5399 resiliency_test();
5400 return 0;
5401}
5402
5403__initcall(slab_sysfs_init);
5404#endif /* CONFIG_SYSFS */
5405
5406/*
5407 * The /proc/slabinfo ABI
5408 */
5409#ifdef CONFIG_SLABINFO
5410static void print_slabinfo_header(struct seq_file *m)
5411{
5412 seq_puts(m, "slabinfo - version: 2.1\n");
5413 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5414 "<objperslab> <pagesperslab>");
5415 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5416 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5417 seq_putc(m, '\n');
5418}
5419
5420static void *s_start(struct seq_file *m, loff_t *pos)
5421{
5422 loff_t n = *pos;
5423
5424 down_read(&slub_lock);
5425 if (!n)
5426 print_slabinfo_header(m);
5427
5428 return seq_list_start(&slab_caches, *pos);
5429}
5430
5431static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5432{
5433 return seq_list_next(p, &slab_caches, pos);
5434}
5435
5436static void s_stop(struct seq_file *m, void *p)
5437{
5438 up_read(&slub_lock);
5439}
5440
5441static int s_show(struct seq_file *m, void *p)
5442{
5443 unsigned long nr_partials = 0;
5444 unsigned long nr_slabs = 0;
5445 unsigned long nr_inuse = 0;
5446 unsigned long nr_objs = 0;
5447 unsigned long nr_free = 0;
5448 struct kmem_cache *s;
5449 int node;
5450
5451 s = list_entry(p, struct kmem_cache, list);
5452
5453 for_each_online_node(node) {
5454 struct kmem_cache_node *n = get_node(s, node);
5455
5456 if (!n)
5457 continue;
5458
5459 nr_partials += n->nr_partial;
5460 nr_slabs += atomic_long_read(&n->nr_slabs);
5461 nr_objs += atomic_long_read(&n->total_objects);
5462 nr_free += count_partial(n, count_free);
5463 }
5464
5465 nr_inuse = nr_objs - nr_free;
5466
5467 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5468 nr_objs, s->size, oo_objects(s->oo),
5469 (1 << oo_order(s->oo)));
5470 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5471 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5472 0UL);
5473 seq_putc(m, '\n');
5474 return 0;
5475}
5476
5477static const struct seq_operations slabinfo_op = {
5478 .start = s_start,
5479 .next = s_next,
5480 .stop = s_stop,
5481 .show = s_show,
5482};
5483
5484static int slabinfo_open(struct inode *inode, struct file *file)
5485{
5486 return seq_open(file, &slabinfo_op);
5487}
5488
5489static const struct file_operations proc_slabinfo_operations = {
5490 .open = slabinfo_open,
5491 .read = seq_read,
5492 .llseek = seq_lseek,
5493 .release = seq_release,
5494};
5495
5496static int __init slab_proc_init(void)
5497{
5498 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5499 return 0;
5500}
5501module_init(slab_proc_init);
5502#endif /* CONFIG_SLABINFO */