Loading...
1/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
10 */
11
12#include <linux/mm.h>
13#include <linux/swap.h> /* struct reclaim_state */
14#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
19#include "slab.h"
20#include <linux/proc_fs.h>
21#include <linux/notifier.h>
22#include <linux/seq_file.h>
23#include <linux/kmemcheck.h>
24#include <linux/cpu.h>
25#include <linux/cpuset.h>
26#include <linux/mempolicy.h>
27#include <linux/ctype.h>
28#include <linux/debugobjects.h>
29#include <linux/kallsyms.h>
30#include <linux/memory.h>
31#include <linux/math64.h>
32#include <linux/fault-inject.h>
33#include <linux/stacktrace.h>
34#include <linux/prefetch.h>
35#include <linux/memcontrol.h>
36
37#include <trace/events/kmem.h>
38
39#include "internal.h"
40
41/*
42 * Lock order:
43 * 1. slab_mutex (Global Mutex)
44 * 2. node->list_lock
45 * 3. slab_lock(page) (Only on some arches and for debugging)
46 *
47 * slab_mutex
48 *
49 * The role of the slab_mutex is to protect the list of all the slabs
50 * and to synchronize major metadata changes to slab cache structures.
51 *
52 * The slab_lock is only used for debugging and on arches that do not
53 * have the ability to do a cmpxchg_double. It only protects the second
54 * double word in the page struct. Meaning
55 * A. page->freelist -> List of object free in a page
56 * B. page->counters -> Counters of objects
57 * C. page->frozen -> frozen state
58 *
59 * If a slab is frozen then it is exempt from list management. It is not
60 * on any list. The processor that froze the slab is the one who can
61 * perform list operations on the page. Other processors may put objects
62 * onto the freelist but the processor that froze the slab is the only
63 * one that can retrieve the objects from the page's freelist.
64 *
65 * The list_lock protects the partial and full list on each node and
66 * the partial slab counter. If taken then no new slabs may be added or
67 * removed from the lists nor make the number of partial slabs be modified.
68 * (Note that the total number of slabs is an atomic value that may be
69 * modified without taking the list lock).
70 *
71 * The list_lock is a centralized lock and thus we avoid taking it as
72 * much as possible. As long as SLUB does not have to handle partial
73 * slabs, operations can continue without any centralized lock. F.e.
74 * allocating a long series of objects that fill up slabs does not require
75 * the list lock.
76 * Interrupts are disabled during allocation and deallocation in order to
77 * make the slab allocator safe to use in the context of an irq. In addition
78 * interrupts are disabled to ensure that the processor does not change
79 * while handling per_cpu slabs, due to kernel preemption.
80 *
81 * SLUB assigns one slab for allocation to each processor.
82 * Allocations only occur from these slabs called cpu slabs.
83 *
84 * Slabs with free elements are kept on a partial list and during regular
85 * operations no list for full slabs is used. If an object in a full slab is
86 * freed then the slab will show up again on the partial lists.
87 * We track full slabs for debugging purposes though because otherwise we
88 * cannot scan all objects.
89 *
90 * Slabs are freed when they become empty. Teardown and setup is
91 * minimal so we rely on the page allocators per cpu caches for
92 * fast frees and allocs.
93 *
94 * Overloading of page flags that are otherwise used for LRU management.
95 *
96 * PageActive The slab is frozen and exempt from list processing.
97 * This means that the slab is dedicated to a purpose
98 * such as satisfying allocations for a specific
99 * processor. Objects may be freed in the slab while
100 * it is frozen but slab_free will then skip the usual
101 * list operations. It is up to the processor holding
102 * the slab to integrate the slab into the slab lists
103 * when the slab is no longer needed.
104 *
105 * One use of this flag is to mark slabs that are
106 * used for allocations. Then such a slab becomes a cpu
107 * slab. The cpu slab may be equipped with an additional
108 * freelist that allows lockless access to
109 * free objects in addition to the regular freelist
110 * that requires the slab lock.
111 *
112 * PageError Slab requires special handling due to debug
113 * options set. This moves slab handling out of
114 * the fast path and disables lockless freelists.
115 */
116
117static inline int kmem_cache_debug(struct kmem_cache *s)
118{
119#ifdef CONFIG_SLUB_DEBUG
120 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
121#else
122 return 0;
123#endif
124}
125
126static inline bool kmem_cache_has_cpu_partial(struct kmem_cache *s)
127{
128#ifdef CONFIG_SLUB_CPU_PARTIAL
129 return !kmem_cache_debug(s);
130#else
131 return false;
132#endif
133}
134
135/*
136 * Issues still to be resolved:
137 *
138 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
139 *
140 * - Variable sizing of the per node arrays
141 */
142
143/* Enable to test recovery from slab corruption on boot */
144#undef SLUB_RESILIENCY_TEST
145
146/* Enable to log cmpxchg failures */
147#undef SLUB_DEBUG_CMPXCHG
148
149/*
150 * Mininum number of partial slabs. These will be left on the partial
151 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 */
153#define MIN_PARTIAL 5
154
155/*
156 * Maximum number of desirable partial slabs.
157 * The existence of more partial slabs makes kmem_cache_shrink
158 * sort the partial list by the number of objects in use.
159 */
160#define MAX_PARTIAL 10
161
162#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
163 SLAB_POISON | SLAB_STORE_USER)
164
165/*
166 * Debugging flags that require metadata to be stored in the slab. These get
167 * disabled when slub_debug=O is used and a cache's min order increases with
168 * metadata.
169 */
170#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
171
172/*
173 * Set of flags that will prevent slab merging
174 */
175#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
176 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
177 SLAB_FAILSLAB)
178
179#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
180 SLAB_CACHE_DMA | SLAB_NOTRACK)
181
182#define OO_SHIFT 16
183#define OO_MASK ((1 << OO_SHIFT) - 1)
184#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
185
186/* Internal SLUB flags */
187#define __OBJECT_POISON 0x80000000UL /* Poison object */
188#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
189
190#ifdef CONFIG_SMP
191static struct notifier_block slab_notifier;
192#endif
193
194/*
195 * Tracking user of a slab.
196 */
197#define TRACK_ADDRS_COUNT 16
198struct track {
199 unsigned long addr; /* Called from address */
200#ifdef CONFIG_STACKTRACE
201 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
202#endif
203 int cpu; /* Was running on cpu */
204 int pid; /* Pid context */
205 unsigned long when; /* When did the operation occur */
206};
207
208enum track_item { TRACK_ALLOC, TRACK_FREE };
209
210#ifdef CONFIG_SYSFS
211static int sysfs_slab_add(struct kmem_cache *);
212static int sysfs_slab_alias(struct kmem_cache *, const char *);
213static void memcg_propagate_slab_attrs(struct kmem_cache *s);
214#else
215static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
216static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
217 { return 0; }
218static inline void memcg_propagate_slab_attrs(struct kmem_cache *s) { }
219#endif
220
221static inline void stat(const struct kmem_cache *s, enum stat_item si)
222{
223#ifdef CONFIG_SLUB_STATS
224 /*
225 * The rmw is racy on a preemptible kernel but this is acceptable, so
226 * avoid this_cpu_add()'s irq-disable overhead.
227 */
228 raw_cpu_inc(s->cpu_slab->stat[si]);
229#endif
230}
231
232/********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
235
236static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
237{
238 return s->node[node];
239}
240
241/* Verify that a pointer has an address that is valid within a slab page */
242static inline int check_valid_pointer(struct kmem_cache *s,
243 struct page *page, const void *object)
244{
245 void *base;
246
247 if (!object)
248 return 1;
249
250 base = page_address(page);
251 if (object < base || object >= base + page->objects * s->size ||
252 (object - base) % s->size) {
253 return 0;
254 }
255
256 return 1;
257}
258
259static inline void *get_freepointer(struct kmem_cache *s, void *object)
260{
261 return *(void **)(object + s->offset);
262}
263
264static void prefetch_freepointer(const struct kmem_cache *s, void *object)
265{
266 prefetch(object + s->offset);
267}
268
269static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
270{
271 void *p;
272
273#ifdef CONFIG_DEBUG_PAGEALLOC
274 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
275#else
276 p = get_freepointer(s, object);
277#endif
278 return p;
279}
280
281static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
282{
283 *(void **)(object + s->offset) = fp;
284}
285
286/* Loop over all objects in a slab */
287#define for_each_object(__p, __s, __addr, __objects) \
288 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
289 __p += (__s)->size)
290
291/* Determine object index from a given position */
292static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
293{
294 return (p - addr) / s->size;
295}
296
297static inline size_t slab_ksize(const struct kmem_cache *s)
298{
299#ifdef CONFIG_SLUB_DEBUG
300 /*
301 * Debugging requires use of the padding between object
302 * and whatever may come after it.
303 */
304 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
305 return s->object_size;
306
307#endif
308 /*
309 * If we have the need to store the freelist pointer
310 * back there or track user information then we can
311 * only use the space before that information.
312 */
313 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
314 return s->inuse;
315 /*
316 * Else we can use all the padding etc for the allocation
317 */
318 return s->size;
319}
320
321static inline int order_objects(int order, unsigned long size, int reserved)
322{
323 return ((PAGE_SIZE << order) - reserved) / size;
324}
325
326static inline struct kmem_cache_order_objects oo_make(int order,
327 unsigned long size, int reserved)
328{
329 struct kmem_cache_order_objects x = {
330 (order << OO_SHIFT) + order_objects(order, size, reserved)
331 };
332
333 return x;
334}
335
336static inline int oo_order(struct kmem_cache_order_objects x)
337{
338 return x.x >> OO_SHIFT;
339}
340
341static inline int oo_objects(struct kmem_cache_order_objects x)
342{
343 return x.x & OO_MASK;
344}
345
346/*
347 * Per slab locking using the pagelock
348 */
349static __always_inline void slab_lock(struct page *page)
350{
351 bit_spin_lock(PG_locked, &page->flags);
352}
353
354static __always_inline void slab_unlock(struct page *page)
355{
356 __bit_spin_unlock(PG_locked, &page->flags);
357}
358
359static inline void set_page_slub_counters(struct page *page, unsigned long counters_new)
360{
361 struct page tmp;
362 tmp.counters = counters_new;
363 /*
364 * page->counters can cover frozen/inuse/objects as well
365 * as page->_count. If we assign to ->counters directly
366 * we run the risk of losing updates to page->_count, so
367 * be careful and only assign to the fields we need.
368 */
369 page->frozen = tmp.frozen;
370 page->inuse = tmp.inuse;
371 page->objects = tmp.objects;
372}
373
374/* Interrupts must be disabled (for the fallback code to work right) */
375static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
376 void *freelist_old, unsigned long counters_old,
377 void *freelist_new, unsigned long counters_new,
378 const char *n)
379{
380 VM_BUG_ON(!irqs_disabled());
381#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
382 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
383 if (s->flags & __CMPXCHG_DOUBLE) {
384 if (cmpxchg_double(&page->freelist, &page->counters,
385 freelist_old, counters_old,
386 freelist_new, counters_new))
387 return 1;
388 } else
389#endif
390 {
391 slab_lock(page);
392 if (page->freelist == freelist_old &&
393 page->counters == counters_old) {
394 page->freelist = freelist_new;
395 set_page_slub_counters(page, counters_new);
396 slab_unlock(page);
397 return 1;
398 }
399 slab_unlock(page);
400 }
401
402 cpu_relax();
403 stat(s, CMPXCHG_DOUBLE_FAIL);
404
405#ifdef SLUB_DEBUG_CMPXCHG
406 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
407#endif
408
409 return 0;
410}
411
412static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
413 void *freelist_old, unsigned long counters_old,
414 void *freelist_new, unsigned long counters_new,
415 const char *n)
416{
417#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
418 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
419 if (s->flags & __CMPXCHG_DOUBLE) {
420 if (cmpxchg_double(&page->freelist, &page->counters,
421 freelist_old, counters_old,
422 freelist_new, counters_new))
423 return 1;
424 } else
425#endif
426 {
427 unsigned long flags;
428
429 local_irq_save(flags);
430 slab_lock(page);
431 if (page->freelist == freelist_old &&
432 page->counters == counters_old) {
433 page->freelist = freelist_new;
434 set_page_slub_counters(page, counters_new);
435 slab_unlock(page);
436 local_irq_restore(flags);
437 return 1;
438 }
439 slab_unlock(page);
440 local_irq_restore(flags);
441 }
442
443 cpu_relax();
444 stat(s, CMPXCHG_DOUBLE_FAIL);
445
446#ifdef SLUB_DEBUG_CMPXCHG
447 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
448#endif
449
450 return 0;
451}
452
453#ifdef CONFIG_SLUB_DEBUG
454/*
455 * Determine a map of object in use on a page.
456 *
457 * Node listlock must be held to guarantee that the page does
458 * not vanish from under us.
459 */
460static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
461{
462 void *p;
463 void *addr = page_address(page);
464
465 for (p = page->freelist; p; p = get_freepointer(s, p))
466 set_bit(slab_index(p, s, addr), map);
467}
468
469/*
470 * Debug settings:
471 */
472#ifdef CONFIG_SLUB_DEBUG_ON
473static int slub_debug = DEBUG_DEFAULT_FLAGS;
474#else
475static int slub_debug;
476#endif
477
478static char *slub_debug_slabs;
479static int disable_higher_order_debug;
480
481/*
482 * Object debugging
483 */
484static void print_section(char *text, u8 *addr, unsigned int length)
485{
486 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
487 length, 1);
488}
489
490static struct track *get_track(struct kmem_cache *s, void *object,
491 enum track_item alloc)
492{
493 struct track *p;
494
495 if (s->offset)
496 p = object + s->offset + sizeof(void *);
497 else
498 p = object + s->inuse;
499
500 return p + alloc;
501}
502
503static void set_track(struct kmem_cache *s, void *object,
504 enum track_item alloc, unsigned long addr)
505{
506 struct track *p = get_track(s, object, alloc);
507
508 if (addr) {
509#ifdef CONFIG_STACKTRACE
510 struct stack_trace trace;
511 int i;
512
513 trace.nr_entries = 0;
514 trace.max_entries = TRACK_ADDRS_COUNT;
515 trace.entries = p->addrs;
516 trace.skip = 3;
517 save_stack_trace(&trace);
518
519 /* See rant in lockdep.c */
520 if (trace.nr_entries != 0 &&
521 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
522 trace.nr_entries--;
523
524 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
525 p->addrs[i] = 0;
526#endif
527 p->addr = addr;
528 p->cpu = smp_processor_id();
529 p->pid = current->pid;
530 p->when = jiffies;
531 } else
532 memset(p, 0, sizeof(struct track));
533}
534
535static void init_tracking(struct kmem_cache *s, void *object)
536{
537 if (!(s->flags & SLAB_STORE_USER))
538 return;
539
540 set_track(s, object, TRACK_FREE, 0UL);
541 set_track(s, object, TRACK_ALLOC, 0UL);
542}
543
544static void print_track(const char *s, struct track *t)
545{
546 if (!t->addr)
547 return;
548
549 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
550 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
551#ifdef CONFIG_STACKTRACE
552 {
553 int i;
554 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
555 if (t->addrs[i])
556 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
557 else
558 break;
559 }
560#endif
561}
562
563static void print_tracking(struct kmem_cache *s, void *object)
564{
565 if (!(s->flags & SLAB_STORE_USER))
566 return;
567
568 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
569 print_track("Freed", get_track(s, object, TRACK_FREE));
570}
571
572static void print_page_info(struct page *page)
573{
574 printk(KERN_ERR
575 "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
576 page, page->objects, page->inuse, page->freelist, page->flags);
577
578}
579
580static void slab_bug(struct kmem_cache *s, char *fmt, ...)
581{
582 va_list args;
583 char buf[100];
584
585 va_start(args, fmt);
586 vsnprintf(buf, sizeof(buf), fmt, args);
587 va_end(args);
588 printk(KERN_ERR "========================================"
589 "=====================================\n");
590 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
591 printk(KERN_ERR "----------------------------------------"
592 "-------------------------------------\n\n");
593
594 add_taint(TAINT_BAD_PAGE, LOCKDEP_NOW_UNRELIABLE);
595}
596
597static void slab_fix(struct kmem_cache *s, char *fmt, ...)
598{
599 va_list args;
600 char buf[100];
601
602 va_start(args, fmt);
603 vsnprintf(buf, sizeof(buf), fmt, args);
604 va_end(args);
605 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
606}
607
608static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
609{
610 unsigned int off; /* Offset of last byte */
611 u8 *addr = page_address(page);
612
613 print_tracking(s, p);
614
615 print_page_info(page);
616
617 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
618 p, p - addr, get_freepointer(s, p));
619
620 if (p > addr + 16)
621 print_section("Bytes b4 ", p - 16, 16);
622
623 print_section("Object ", p, min_t(unsigned long, s->object_size,
624 PAGE_SIZE));
625 if (s->flags & SLAB_RED_ZONE)
626 print_section("Redzone ", p + s->object_size,
627 s->inuse - s->object_size);
628
629 if (s->offset)
630 off = s->offset + sizeof(void *);
631 else
632 off = s->inuse;
633
634 if (s->flags & SLAB_STORE_USER)
635 off += 2 * sizeof(struct track);
636
637 if (off != s->size)
638 /* Beginning of the filler is the free pointer */
639 print_section("Padding ", p + off, s->size - off);
640
641 dump_stack();
642}
643
644static void object_err(struct kmem_cache *s, struct page *page,
645 u8 *object, char *reason)
646{
647 slab_bug(s, "%s", reason);
648 print_trailer(s, page, object);
649}
650
651static void slab_err(struct kmem_cache *s, struct page *page,
652 const char *fmt, ...)
653{
654 va_list args;
655 char buf[100];
656
657 va_start(args, fmt);
658 vsnprintf(buf, sizeof(buf), fmt, args);
659 va_end(args);
660 slab_bug(s, "%s", buf);
661 print_page_info(page);
662 dump_stack();
663}
664
665static void init_object(struct kmem_cache *s, void *object, u8 val)
666{
667 u8 *p = object;
668
669 if (s->flags & __OBJECT_POISON) {
670 memset(p, POISON_FREE, s->object_size - 1);
671 p[s->object_size - 1] = POISON_END;
672 }
673
674 if (s->flags & SLAB_RED_ZONE)
675 memset(p + s->object_size, val, s->inuse - s->object_size);
676}
677
678static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
679 void *from, void *to)
680{
681 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
682 memset(from, data, to - from);
683}
684
685static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
686 u8 *object, char *what,
687 u8 *start, unsigned int value, unsigned int bytes)
688{
689 u8 *fault;
690 u8 *end;
691
692 fault = memchr_inv(start, value, bytes);
693 if (!fault)
694 return 1;
695
696 end = start + bytes;
697 while (end > fault && end[-1] == value)
698 end--;
699
700 slab_bug(s, "%s overwritten", what);
701 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
702 fault, end - 1, fault[0], value);
703 print_trailer(s, page, object);
704
705 restore_bytes(s, what, value, fault, end);
706 return 0;
707}
708
709/*
710 * Object layout:
711 *
712 * object address
713 * Bytes of the object to be managed.
714 * If the freepointer may overlay the object then the free
715 * pointer is the first word of the object.
716 *
717 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
718 * 0xa5 (POISON_END)
719 *
720 * object + s->object_size
721 * Padding to reach word boundary. This is also used for Redzoning.
722 * Padding is extended by another word if Redzoning is enabled and
723 * object_size == inuse.
724 *
725 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
726 * 0xcc (RED_ACTIVE) for objects in use.
727 *
728 * object + s->inuse
729 * Meta data starts here.
730 *
731 * A. Free pointer (if we cannot overwrite object on free)
732 * B. Tracking data for SLAB_STORE_USER
733 * C. Padding to reach required alignment boundary or at mininum
734 * one word if debugging is on to be able to detect writes
735 * before the word boundary.
736 *
737 * Padding is done using 0x5a (POISON_INUSE)
738 *
739 * object + s->size
740 * Nothing is used beyond s->size.
741 *
742 * If slabcaches are merged then the object_size and inuse boundaries are mostly
743 * ignored. And therefore no slab options that rely on these boundaries
744 * may be used with merged slabcaches.
745 */
746
747static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
748{
749 unsigned long off = s->inuse; /* The end of info */
750
751 if (s->offset)
752 /* Freepointer is placed after the object. */
753 off += sizeof(void *);
754
755 if (s->flags & SLAB_STORE_USER)
756 /* We also have user information there */
757 off += 2 * sizeof(struct track);
758
759 if (s->size == off)
760 return 1;
761
762 return check_bytes_and_report(s, page, p, "Object padding",
763 p + off, POISON_INUSE, s->size - off);
764}
765
766/* Check the pad bytes at the end of a slab page */
767static int slab_pad_check(struct kmem_cache *s, struct page *page)
768{
769 u8 *start;
770 u8 *fault;
771 u8 *end;
772 int length;
773 int remainder;
774
775 if (!(s->flags & SLAB_POISON))
776 return 1;
777
778 start = page_address(page);
779 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
780 end = start + length;
781 remainder = length % s->size;
782 if (!remainder)
783 return 1;
784
785 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
786 if (!fault)
787 return 1;
788 while (end > fault && end[-1] == POISON_INUSE)
789 end--;
790
791 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
792 print_section("Padding ", end - remainder, remainder);
793
794 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
795 return 0;
796}
797
798static int check_object(struct kmem_cache *s, struct page *page,
799 void *object, u8 val)
800{
801 u8 *p = object;
802 u8 *endobject = object + s->object_size;
803
804 if (s->flags & SLAB_RED_ZONE) {
805 if (!check_bytes_and_report(s, page, object, "Redzone",
806 endobject, val, s->inuse - s->object_size))
807 return 0;
808 } else {
809 if ((s->flags & SLAB_POISON) && s->object_size < s->inuse) {
810 check_bytes_and_report(s, page, p, "Alignment padding",
811 endobject, POISON_INUSE,
812 s->inuse - s->object_size);
813 }
814 }
815
816 if (s->flags & SLAB_POISON) {
817 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
818 (!check_bytes_and_report(s, page, p, "Poison", p,
819 POISON_FREE, s->object_size - 1) ||
820 !check_bytes_and_report(s, page, p, "Poison",
821 p + s->object_size - 1, POISON_END, 1)))
822 return 0;
823 /*
824 * check_pad_bytes cleans up on its own.
825 */
826 check_pad_bytes(s, page, p);
827 }
828
829 if (!s->offset && val == SLUB_RED_ACTIVE)
830 /*
831 * Object and freepointer overlap. Cannot check
832 * freepointer while object is allocated.
833 */
834 return 1;
835
836 /* Check free pointer validity */
837 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
838 object_err(s, page, p, "Freepointer corrupt");
839 /*
840 * No choice but to zap it and thus lose the remainder
841 * of the free objects in this slab. May cause
842 * another error because the object count is now wrong.
843 */
844 set_freepointer(s, p, NULL);
845 return 0;
846 }
847 return 1;
848}
849
850static int check_slab(struct kmem_cache *s, struct page *page)
851{
852 int maxobj;
853
854 VM_BUG_ON(!irqs_disabled());
855
856 if (!PageSlab(page)) {
857 slab_err(s, page, "Not a valid slab page");
858 return 0;
859 }
860
861 maxobj = order_objects(compound_order(page), s->size, s->reserved);
862 if (page->objects > maxobj) {
863 slab_err(s, page, "objects %u > max %u",
864 s->name, page->objects, maxobj);
865 return 0;
866 }
867 if (page->inuse > page->objects) {
868 slab_err(s, page, "inuse %u > max %u",
869 s->name, page->inuse, page->objects);
870 return 0;
871 }
872 /* Slab_pad_check fixes things up after itself */
873 slab_pad_check(s, page);
874 return 1;
875}
876
877/*
878 * Determine if a certain object on a page is on the freelist. Must hold the
879 * slab lock to guarantee that the chains are in a consistent state.
880 */
881static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
882{
883 int nr = 0;
884 void *fp;
885 void *object = NULL;
886 unsigned long max_objects;
887
888 fp = page->freelist;
889 while (fp && nr <= page->objects) {
890 if (fp == search)
891 return 1;
892 if (!check_valid_pointer(s, page, fp)) {
893 if (object) {
894 object_err(s, page, object,
895 "Freechain corrupt");
896 set_freepointer(s, object, NULL);
897 } else {
898 slab_err(s, page, "Freepointer corrupt");
899 page->freelist = NULL;
900 page->inuse = page->objects;
901 slab_fix(s, "Freelist cleared");
902 return 0;
903 }
904 break;
905 }
906 object = fp;
907 fp = get_freepointer(s, object);
908 nr++;
909 }
910
911 max_objects = order_objects(compound_order(page), s->size, s->reserved);
912 if (max_objects > MAX_OBJS_PER_PAGE)
913 max_objects = MAX_OBJS_PER_PAGE;
914
915 if (page->objects != max_objects) {
916 slab_err(s, page, "Wrong number of objects. Found %d but "
917 "should be %d", page->objects, max_objects);
918 page->objects = max_objects;
919 slab_fix(s, "Number of objects adjusted.");
920 }
921 if (page->inuse != page->objects - nr) {
922 slab_err(s, page, "Wrong object count. Counter is %d but "
923 "counted were %d", page->inuse, page->objects - nr);
924 page->inuse = page->objects - nr;
925 slab_fix(s, "Object count adjusted.");
926 }
927 return search == NULL;
928}
929
930static void trace(struct kmem_cache *s, struct page *page, void *object,
931 int alloc)
932{
933 if (s->flags & SLAB_TRACE) {
934 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
935 s->name,
936 alloc ? "alloc" : "free",
937 object, page->inuse,
938 page->freelist);
939
940 if (!alloc)
941 print_section("Object ", (void *)object,
942 s->object_size);
943
944 dump_stack();
945 }
946}
947
948/*
949 * Hooks for other subsystems that check memory allocations. In a typical
950 * production configuration these hooks all should produce no code at all.
951 */
952static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
953{
954 kmemleak_alloc(ptr, size, 1, flags);
955}
956
957static inline void kfree_hook(const void *x)
958{
959 kmemleak_free(x);
960}
961
962static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
963{
964 flags &= gfp_allowed_mask;
965 lockdep_trace_alloc(flags);
966 might_sleep_if(flags & __GFP_WAIT);
967
968 return should_failslab(s->object_size, flags, s->flags);
969}
970
971static inline void slab_post_alloc_hook(struct kmem_cache *s,
972 gfp_t flags, void *object)
973{
974 flags &= gfp_allowed_mask;
975 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
976 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags, flags);
977}
978
979static inline void slab_free_hook(struct kmem_cache *s, void *x)
980{
981 kmemleak_free_recursive(x, s->flags);
982
983 /*
984 * Trouble is that we may no longer disable interrupts in the fast path
985 * So in order to make the debug calls that expect irqs to be
986 * disabled we need to disable interrupts temporarily.
987 */
988#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
989 {
990 unsigned long flags;
991
992 local_irq_save(flags);
993 kmemcheck_slab_free(s, x, s->object_size);
994 debug_check_no_locks_freed(x, s->object_size);
995 local_irq_restore(flags);
996 }
997#endif
998 if (!(s->flags & SLAB_DEBUG_OBJECTS))
999 debug_check_no_obj_freed(x, s->object_size);
1000}
1001
1002/*
1003 * Tracking of fully allocated slabs for debugging purposes.
1004 */
1005static void add_full(struct kmem_cache *s,
1006 struct kmem_cache_node *n, struct page *page)
1007{
1008 if (!(s->flags & SLAB_STORE_USER))
1009 return;
1010
1011 lockdep_assert_held(&n->list_lock);
1012 list_add(&page->lru, &n->full);
1013}
1014
1015static void remove_full(struct kmem_cache *s, struct kmem_cache_node *n, struct page *page)
1016{
1017 if (!(s->flags & SLAB_STORE_USER))
1018 return;
1019
1020 lockdep_assert_held(&n->list_lock);
1021 list_del(&page->lru);
1022}
1023
1024/* Tracking of the number of slabs for debugging purposes */
1025static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1026{
1027 struct kmem_cache_node *n = get_node(s, node);
1028
1029 return atomic_long_read(&n->nr_slabs);
1030}
1031
1032static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1033{
1034 return atomic_long_read(&n->nr_slabs);
1035}
1036
1037static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1038{
1039 struct kmem_cache_node *n = get_node(s, node);
1040
1041 /*
1042 * May be called early in order to allocate a slab for the
1043 * kmem_cache_node structure. Solve the chicken-egg
1044 * dilemma by deferring the increment of the count during
1045 * bootstrap (see early_kmem_cache_node_alloc).
1046 */
1047 if (likely(n)) {
1048 atomic_long_inc(&n->nr_slabs);
1049 atomic_long_add(objects, &n->total_objects);
1050 }
1051}
1052static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1053{
1054 struct kmem_cache_node *n = get_node(s, node);
1055
1056 atomic_long_dec(&n->nr_slabs);
1057 atomic_long_sub(objects, &n->total_objects);
1058}
1059
1060/* Object debug checks for alloc/free paths */
1061static void setup_object_debug(struct kmem_cache *s, struct page *page,
1062 void *object)
1063{
1064 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1065 return;
1066
1067 init_object(s, object, SLUB_RED_INACTIVE);
1068 init_tracking(s, object);
1069}
1070
1071static noinline int alloc_debug_processing(struct kmem_cache *s,
1072 struct page *page,
1073 void *object, unsigned long addr)
1074{
1075 if (!check_slab(s, page))
1076 goto bad;
1077
1078 if (!check_valid_pointer(s, page, object)) {
1079 object_err(s, page, object, "Freelist Pointer check fails");
1080 goto bad;
1081 }
1082
1083 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1084 goto bad;
1085
1086 /* Success perform special debug activities for allocs */
1087 if (s->flags & SLAB_STORE_USER)
1088 set_track(s, object, TRACK_ALLOC, addr);
1089 trace(s, page, object, 1);
1090 init_object(s, object, SLUB_RED_ACTIVE);
1091 return 1;
1092
1093bad:
1094 if (PageSlab(page)) {
1095 /*
1096 * If this is a slab page then lets do the best we can
1097 * to avoid issues in the future. Marking all objects
1098 * as used avoids touching the remaining objects.
1099 */
1100 slab_fix(s, "Marking all objects used");
1101 page->inuse = page->objects;
1102 page->freelist = NULL;
1103 }
1104 return 0;
1105}
1106
1107static noinline struct kmem_cache_node *free_debug_processing(
1108 struct kmem_cache *s, struct page *page, void *object,
1109 unsigned long addr, unsigned long *flags)
1110{
1111 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1112
1113 spin_lock_irqsave(&n->list_lock, *flags);
1114 slab_lock(page);
1115
1116 if (!check_slab(s, page))
1117 goto fail;
1118
1119 if (!check_valid_pointer(s, page, object)) {
1120 slab_err(s, page, "Invalid object pointer 0x%p", object);
1121 goto fail;
1122 }
1123
1124 if (on_freelist(s, page, object)) {
1125 object_err(s, page, object, "Object already free");
1126 goto fail;
1127 }
1128
1129 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1130 goto out;
1131
1132 if (unlikely(s != page->slab_cache)) {
1133 if (!PageSlab(page)) {
1134 slab_err(s, page, "Attempt to free object(0x%p) "
1135 "outside of slab", object);
1136 } else if (!page->slab_cache) {
1137 printk(KERN_ERR
1138 "SLUB <none>: no slab for object 0x%p.\n",
1139 object);
1140 dump_stack();
1141 } else
1142 object_err(s, page, object,
1143 "page slab pointer corrupt.");
1144 goto fail;
1145 }
1146
1147 if (s->flags & SLAB_STORE_USER)
1148 set_track(s, object, TRACK_FREE, addr);
1149 trace(s, page, object, 0);
1150 init_object(s, object, SLUB_RED_INACTIVE);
1151out:
1152 slab_unlock(page);
1153 /*
1154 * Keep node_lock to preserve integrity
1155 * until the object is actually freed
1156 */
1157 return n;
1158
1159fail:
1160 slab_unlock(page);
1161 spin_unlock_irqrestore(&n->list_lock, *flags);
1162 slab_fix(s, "Object at 0x%p not freed", object);
1163 return NULL;
1164}
1165
1166static int __init setup_slub_debug(char *str)
1167{
1168 slub_debug = DEBUG_DEFAULT_FLAGS;
1169 if (*str++ != '=' || !*str)
1170 /*
1171 * No options specified. Switch on full debugging.
1172 */
1173 goto out;
1174
1175 if (*str == ',')
1176 /*
1177 * No options but restriction on slabs. This means full
1178 * debugging for slabs matching a pattern.
1179 */
1180 goto check_slabs;
1181
1182 if (tolower(*str) == 'o') {
1183 /*
1184 * Avoid enabling debugging on caches if its minimum order
1185 * would increase as a result.
1186 */
1187 disable_higher_order_debug = 1;
1188 goto out;
1189 }
1190
1191 slub_debug = 0;
1192 if (*str == '-')
1193 /*
1194 * Switch off all debugging measures.
1195 */
1196 goto out;
1197
1198 /*
1199 * Determine which debug features should be switched on
1200 */
1201 for (; *str && *str != ','; str++) {
1202 switch (tolower(*str)) {
1203 case 'f':
1204 slub_debug |= SLAB_DEBUG_FREE;
1205 break;
1206 case 'z':
1207 slub_debug |= SLAB_RED_ZONE;
1208 break;
1209 case 'p':
1210 slub_debug |= SLAB_POISON;
1211 break;
1212 case 'u':
1213 slub_debug |= SLAB_STORE_USER;
1214 break;
1215 case 't':
1216 slub_debug |= SLAB_TRACE;
1217 break;
1218 case 'a':
1219 slub_debug |= SLAB_FAILSLAB;
1220 break;
1221 default:
1222 printk(KERN_ERR "slub_debug option '%c' "
1223 "unknown. skipped\n", *str);
1224 }
1225 }
1226
1227check_slabs:
1228 if (*str == ',')
1229 slub_debug_slabs = str + 1;
1230out:
1231 return 1;
1232}
1233
1234__setup("slub_debug", setup_slub_debug);
1235
1236static unsigned long kmem_cache_flags(unsigned long object_size,
1237 unsigned long flags, const char *name,
1238 void (*ctor)(void *))
1239{
1240 /*
1241 * Enable debugging if selected on the kernel commandline.
1242 */
1243 if (slub_debug && (!slub_debug_slabs || (name &&
1244 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)))))
1245 flags |= slub_debug;
1246
1247 return flags;
1248}
1249#else
1250static inline void setup_object_debug(struct kmem_cache *s,
1251 struct page *page, void *object) {}
1252
1253static inline int alloc_debug_processing(struct kmem_cache *s,
1254 struct page *page, void *object, unsigned long addr) { return 0; }
1255
1256static inline struct kmem_cache_node *free_debug_processing(
1257 struct kmem_cache *s, struct page *page, void *object,
1258 unsigned long addr, unsigned long *flags) { return NULL; }
1259
1260static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1261 { return 1; }
1262static inline int check_object(struct kmem_cache *s, struct page *page,
1263 void *object, u8 val) { return 1; }
1264static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1265 struct page *page) {}
1266static inline void remove_full(struct kmem_cache *s, struct kmem_cache_node *n,
1267 struct page *page) {}
1268static inline unsigned long kmem_cache_flags(unsigned long object_size,
1269 unsigned long flags, const char *name,
1270 void (*ctor)(void *))
1271{
1272 return flags;
1273}
1274#define slub_debug 0
1275
1276#define disable_higher_order_debug 0
1277
1278static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1279 { return 0; }
1280static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1281 { return 0; }
1282static inline void inc_slabs_node(struct kmem_cache *s, int node,
1283 int objects) {}
1284static inline void dec_slabs_node(struct kmem_cache *s, int node,
1285 int objects) {}
1286
1287static inline void kmalloc_large_node_hook(void *ptr, size_t size, gfp_t flags)
1288{
1289 kmemleak_alloc(ptr, size, 1, flags);
1290}
1291
1292static inline void kfree_hook(const void *x)
1293{
1294 kmemleak_free(x);
1295}
1296
1297static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1298 { return 0; }
1299
1300static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1301 void *object)
1302{
1303 kmemleak_alloc_recursive(object, s->object_size, 1, s->flags,
1304 flags & gfp_allowed_mask);
1305}
1306
1307static inline void slab_free_hook(struct kmem_cache *s, void *x)
1308{
1309 kmemleak_free_recursive(x, s->flags);
1310}
1311
1312#endif /* CONFIG_SLUB_DEBUG */
1313
1314/*
1315 * Slab allocation and freeing
1316 */
1317static inline struct page *alloc_slab_page(gfp_t flags, int node,
1318 struct kmem_cache_order_objects oo)
1319{
1320 int order = oo_order(oo);
1321
1322 flags |= __GFP_NOTRACK;
1323
1324 if (node == NUMA_NO_NODE)
1325 return alloc_pages(flags, order);
1326 else
1327 return alloc_pages_exact_node(node, flags, order);
1328}
1329
1330static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1331{
1332 struct page *page;
1333 struct kmem_cache_order_objects oo = s->oo;
1334 gfp_t alloc_gfp;
1335
1336 flags &= gfp_allowed_mask;
1337
1338 if (flags & __GFP_WAIT)
1339 local_irq_enable();
1340
1341 flags |= s->allocflags;
1342
1343 /*
1344 * Let the initial higher-order allocation fail under memory pressure
1345 * so we fall-back to the minimum order allocation.
1346 */
1347 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1348
1349 page = alloc_slab_page(alloc_gfp, node, oo);
1350 if (unlikely(!page)) {
1351 oo = s->min;
1352 alloc_gfp = flags;
1353 /*
1354 * Allocation may have failed due to fragmentation.
1355 * Try a lower order alloc if possible
1356 */
1357 page = alloc_slab_page(alloc_gfp, node, oo);
1358
1359 if (page)
1360 stat(s, ORDER_FALLBACK);
1361 }
1362
1363 if (kmemcheck_enabled && page
1364 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1365 int pages = 1 << oo_order(oo);
1366
1367 kmemcheck_alloc_shadow(page, oo_order(oo), alloc_gfp, node);
1368
1369 /*
1370 * Objects from caches that have a constructor don't get
1371 * cleared when they're allocated, so we need to do it here.
1372 */
1373 if (s->ctor)
1374 kmemcheck_mark_uninitialized_pages(page, pages);
1375 else
1376 kmemcheck_mark_unallocated_pages(page, pages);
1377 }
1378
1379 if (flags & __GFP_WAIT)
1380 local_irq_disable();
1381 if (!page)
1382 return NULL;
1383
1384 page->objects = oo_objects(oo);
1385 mod_zone_page_state(page_zone(page),
1386 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1387 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1388 1 << oo_order(oo));
1389
1390 return page;
1391}
1392
1393static void setup_object(struct kmem_cache *s, struct page *page,
1394 void *object)
1395{
1396 setup_object_debug(s, page, object);
1397 if (unlikely(s->ctor))
1398 s->ctor(object);
1399}
1400
1401static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1402{
1403 struct page *page;
1404 void *start;
1405 void *last;
1406 void *p;
1407 int order;
1408
1409 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1410
1411 page = allocate_slab(s,
1412 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1413 if (!page)
1414 goto out;
1415
1416 order = compound_order(page);
1417 inc_slabs_node(s, page_to_nid(page), page->objects);
1418 memcg_bind_pages(s, order);
1419 page->slab_cache = s;
1420 __SetPageSlab(page);
1421 if (page->pfmemalloc)
1422 SetPageSlabPfmemalloc(page);
1423
1424 start = page_address(page);
1425
1426 if (unlikely(s->flags & SLAB_POISON))
1427 memset(start, POISON_INUSE, PAGE_SIZE << order);
1428
1429 last = start;
1430 for_each_object(p, s, start, page->objects) {
1431 setup_object(s, page, last);
1432 set_freepointer(s, last, p);
1433 last = p;
1434 }
1435 setup_object(s, page, last);
1436 set_freepointer(s, last, NULL);
1437
1438 page->freelist = start;
1439 page->inuse = page->objects;
1440 page->frozen = 1;
1441out:
1442 return page;
1443}
1444
1445static void __free_slab(struct kmem_cache *s, struct page *page)
1446{
1447 int order = compound_order(page);
1448 int pages = 1 << order;
1449
1450 if (kmem_cache_debug(s)) {
1451 void *p;
1452
1453 slab_pad_check(s, page);
1454 for_each_object(p, s, page_address(page),
1455 page->objects)
1456 check_object(s, page, p, SLUB_RED_INACTIVE);
1457 }
1458
1459 kmemcheck_free_shadow(page, compound_order(page));
1460
1461 mod_zone_page_state(page_zone(page),
1462 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1463 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1464 -pages);
1465
1466 __ClearPageSlabPfmemalloc(page);
1467 __ClearPageSlab(page);
1468
1469 memcg_release_pages(s, order);
1470 page_mapcount_reset(page);
1471 if (current->reclaim_state)
1472 current->reclaim_state->reclaimed_slab += pages;
1473 __free_memcg_kmem_pages(page, order);
1474}
1475
1476#define need_reserve_slab_rcu \
1477 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1478
1479static void rcu_free_slab(struct rcu_head *h)
1480{
1481 struct page *page;
1482
1483 if (need_reserve_slab_rcu)
1484 page = virt_to_head_page(h);
1485 else
1486 page = container_of((struct list_head *)h, struct page, lru);
1487
1488 __free_slab(page->slab_cache, page);
1489}
1490
1491static void free_slab(struct kmem_cache *s, struct page *page)
1492{
1493 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1494 struct rcu_head *head;
1495
1496 if (need_reserve_slab_rcu) {
1497 int order = compound_order(page);
1498 int offset = (PAGE_SIZE << order) - s->reserved;
1499
1500 VM_BUG_ON(s->reserved != sizeof(*head));
1501 head = page_address(page) + offset;
1502 } else {
1503 /*
1504 * RCU free overloads the RCU head over the LRU
1505 */
1506 head = (void *)&page->lru;
1507 }
1508
1509 call_rcu(head, rcu_free_slab);
1510 } else
1511 __free_slab(s, page);
1512}
1513
1514static void discard_slab(struct kmem_cache *s, struct page *page)
1515{
1516 dec_slabs_node(s, page_to_nid(page), page->objects);
1517 free_slab(s, page);
1518}
1519
1520/*
1521 * Management of partially allocated slabs.
1522 */
1523static inline void
1524__add_partial(struct kmem_cache_node *n, struct page *page, int tail)
1525{
1526 n->nr_partial++;
1527 if (tail == DEACTIVATE_TO_TAIL)
1528 list_add_tail(&page->lru, &n->partial);
1529 else
1530 list_add(&page->lru, &n->partial);
1531}
1532
1533static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1535{
1536 lockdep_assert_held(&n->list_lock);
1537 __add_partial(n, page, tail);
1538}
1539
1540static inline void
1541__remove_partial(struct kmem_cache_node *n, struct page *page)
1542{
1543 list_del(&page->lru);
1544 n->nr_partial--;
1545}
1546
1547static inline void remove_partial(struct kmem_cache_node *n,
1548 struct page *page)
1549{
1550 lockdep_assert_held(&n->list_lock);
1551 __remove_partial(n, page);
1552}
1553
1554/*
1555 * Remove slab from the partial list, freeze it and
1556 * return the pointer to the freelist.
1557 *
1558 * Returns a list of objects or NULL if it fails.
1559 */
1560static inline void *acquire_slab(struct kmem_cache *s,
1561 struct kmem_cache_node *n, struct page *page,
1562 int mode, int *objects)
1563{
1564 void *freelist;
1565 unsigned long counters;
1566 struct page new;
1567
1568 lockdep_assert_held(&n->list_lock);
1569
1570 /*
1571 * Zap the freelist and set the frozen bit.
1572 * The old freelist is the list of objects for the
1573 * per cpu allocation list.
1574 */
1575 freelist = page->freelist;
1576 counters = page->counters;
1577 new.counters = counters;
1578 *objects = new.objects - new.inuse;
1579 if (mode) {
1580 new.inuse = page->objects;
1581 new.freelist = NULL;
1582 } else {
1583 new.freelist = freelist;
1584 }
1585
1586 VM_BUG_ON(new.frozen);
1587 new.frozen = 1;
1588
1589 if (!__cmpxchg_double_slab(s, page,
1590 freelist, counters,
1591 new.freelist, new.counters,
1592 "acquire_slab"))
1593 return NULL;
1594
1595 remove_partial(n, page);
1596 WARN_ON(!freelist);
1597 return freelist;
1598}
1599
1600static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1601static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags);
1602
1603/*
1604 * Try to allocate a partial slab from a specific node.
1605 */
1606static void *get_partial_node(struct kmem_cache *s, struct kmem_cache_node *n,
1607 struct kmem_cache_cpu *c, gfp_t flags)
1608{
1609 struct page *page, *page2;
1610 void *object = NULL;
1611 int available = 0;
1612 int objects;
1613
1614 /*
1615 * Racy check. If we mistakenly see no partial slabs then we
1616 * just allocate an empty slab. If we mistakenly try to get a
1617 * partial slab and there is none available then get_partials()
1618 * will return NULL.
1619 */
1620 if (!n || !n->nr_partial)
1621 return NULL;
1622
1623 spin_lock(&n->list_lock);
1624 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1625 void *t;
1626
1627 if (!pfmemalloc_match(page, flags))
1628 continue;
1629
1630 t = acquire_slab(s, n, page, object == NULL, &objects);
1631 if (!t)
1632 break;
1633
1634 available += objects;
1635 if (!object) {
1636 c->page = page;
1637 stat(s, ALLOC_FROM_PARTIAL);
1638 object = t;
1639 } else {
1640 put_cpu_partial(s, page, 0);
1641 stat(s, CPU_PARTIAL_NODE);
1642 }
1643 if (!kmem_cache_has_cpu_partial(s)
1644 || available > s->cpu_partial / 2)
1645 break;
1646
1647 }
1648 spin_unlock(&n->list_lock);
1649 return object;
1650}
1651
1652/*
1653 * Get a page from somewhere. Search in increasing NUMA distances.
1654 */
1655static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1656 struct kmem_cache_cpu *c)
1657{
1658#ifdef CONFIG_NUMA
1659 struct zonelist *zonelist;
1660 struct zoneref *z;
1661 struct zone *zone;
1662 enum zone_type high_zoneidx = gfp_zone(flags);
1663 void *object;
1664 unsigned int cpuset_mems_cookie;
1665
1666 /*
1667 * The defrag ratio allows a configuration of the tradeoffs between
1668 * inter node defragmentation and node local allocations. A lower
1669 * defrag_ratio increases the tendency to do local allocations
1670 * instead of attempting to obtain partial slabs from other nodes.
1671 *
1672 * If the defrag_ratio is set to 0 then kmalloc() always
1673 * returns node local objects. If the ratio is higher then kmalloc()
1674 * may return off node objects because partial slabs are obtained
1675 * from other nodes and filled up.
1676 *
1677 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1678 * defrag_ratio = 1000) then every (well almost) allocation will
1679 * first attempt to defrag slab caches on other nodes. This means
1680 * scanning over all nodes to look for partial slabs which may be
1681 * expensive if we do it every time we are trying to find a slab
1682 * with available objects.
1683 */
1684 if (!s->remote_node_defrag_ratio ||
1685 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1686 return NULL;
1687
1688 do {
1689 cpuset_mems_cookie = read_mems_allowed_begin();
1690 zonelist = node_zonelist(mempolicy_slab_node(), flags);
1691 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1692 struct kmem_cache_node *n;
1693
1694 n = get_node(s, zone_to_nid(zone));
1695
1696 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1697 n->nr_partial > s->min_partial) {
1698 object = get_partial_node(s, n, c, flags);
1699 if (object) {
1700 /*
1701 * Don't check read_mems_allowed_retry()
1702 * here - if mems_allowed was updated in
1703 * parallel, that was a harmless race
1704 * between allocation and the cpuset
1705 * update
1706 */
1707 return object;
1708 }
1709 }
1710 }
1711 } while (read_mems_allowed_retry(cpuset_mems_cookie));
1712#endif
1713 return NULL;
1714}
1715
1716/*
1717 * Get a partial page, lock it and return it.
1718 */
1719static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1720 struct kmem_cache_cpu *c)
1721{
1722 void *object;
1723 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1724
1725 object = get_partial_node(s, get_node(s, searchnode), c, flags);
1726 if (object || node != NUMA_NO_NODE)
1727 return object;
1728
1729 return get_any_partial(s, flags, c);
1730}
1731
1732#ifdef CONFIG_PREEMPT
1733/*
1734 * Calculate the next globally unique transaction for disambiguiation
1735 * during cmpxchg. The transactions start with the cpu number and are then
1736 * incremented by CONFIG_NR_CPUS.
1737 */
1738#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1739#else
1740/*
1741 * No preemption supported therefore also no need to check for
1742 * different cpus.
1743 */
1744#define TID_STEP 1
1745#endif
1746
1747static inline unsigned long next_tid(unsigned long tid)
1748{
1749 return tid + TID_STEP;
1750}
1751
1752static inline unsigned int tid_to_cpu(unsigned long tid)
1753{
1754 return tid % TID_STEP;
1755}
1756
1757static inline unsigned long tid_to_event(unsigned long tid)
1758{
1759 return tid / TID_STEP;
1760}
1761
1762static inline unsigned int init_tid(int cpu)
1763{
1764 return cpu;
1765}
1766
1767static inline void note_cmpxchg_failure(const char *n,
1768 const struct kmem_cache *s, unsigned long tid)
1769{
1770#ifdef SLUB_DEBUG_CMPXCHG
1771 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1772
1773 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1774
1775#ifdef CONFIG_PREEMPT
1776 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1777 printk("due to cpu change %d -> %d\n",
1778 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1779 else
1780#endif
1781 if (tid_to_event(tid) != tid_to_event(actual_tid))
1782 printk("due to cpu running other code. Event %ld->%ld\n",
1783 tid_to_event(tid), tid_to_event(actual_tid));
1784 else
1785 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1786 actual_tid, tid, next_tid(tid));
1787#endif
1788 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1789}
1790
1791static void init_kmem_cache_cpus(struct kmem_cache *s)
1792{
1793 int cpu;
1794
1795 for_each_possible_cpu(cpu)
1796 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1797}
1798
1799/*
1800 * Remove the cpu slab
1801 */
1802static void deactivate_slab(struct kmem_cache *s, struct page *page,
1803 void *freelist)
1804{
1805 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1806 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1807 int lock = 0;
1808 enum slab_modes l = M_NONE, m = M_NONE;
1809 void *nextfree;
1810 int tail = DEACTIVATE_TO_HEAD;
1811 struct page new;
1812 struct page old;
1813
1814 if (page->freelist) {
1815 stat(s, DEACTIVATE_REMOTE_FREES);
1816 tail = DEACTIVATE_TO_TAIL;
1817 }
1818
1819 /*
1820 * Stage one: Free all available per cpu objects back
1821 * to the page freelist while it is still frozen. Leave the
1822 * last one.
1823 *
1824 * There is no need to take the list->lock because the page
1825 * is still frozen.
1826 */
1827 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1828 void *prior;
1829 unsigned long counters;
1830
1831 do {
1832 prior = page->freelist;
1833 counters = page->counters;
1834 set_freepointer(s, freelist, prior);
1835 new.counters = counters;
1836 new.inuse--;
1837 VM_BUG_ON(!new.frozen);
1838
1839 } while (!__cmpxchg_double_slab(s, page,
1840 prior, counters,
1841 freelist, new.counters,
1842 "drain percpu freelist"));
1843
1844 freelist = nextfree;
1845 }
1846
1847 /*
1848 * Stage two: Ensure that the page is unfrozen while the
1849 * list presence reflects the actual number of objects
1850 * during unfreeze.
1851 *
1852 * We setup the list membership and then perform a cmpxchg
1853 * with the count. If there is a mismatch then the page
1854 * is not unfrozen but the page is on the wrong list.
1855 *
1856 * Then we restart the process which may have to remove
1857 * the page from the list that we just put it on again
1858 * because the number of objects in the slab may have
1859 * changed.
1860 */
1861redo:
1862
1863 old.freelist = page->freelist;
1864 old.counters = page->counters;
1865 VM_BUG_ON(!old.frozen);
1866
1867 /* Determine target state of the slab */
1868 new.counters = old.counters;
1869 if (freelist) {
1870 new.inuse--;
1871 set_freepointer(s, freelist, old.freelist);
1872 new.freelist = freelist;
1873 } else
1874 new.freelist = old.freelist;
1875
1876 new.frozen = 0;
1877
1878 if (!new.inuse && n->nr_partial > s->min_partial)
1879 m = M_FREE;
1880 else if (new.freelist) {
1881 m = M_PARTIAL;
1882 if (!lock) {
1883 lock = 1;
1884 /*
1885 * Taking the spinlock removes the possiblity
1886 * that acquire_slab() will see a slab page that
1887 * is frozen
1888 */
1889 spin_lock(&n->list_lock);
1890 }
1891 } else {
1892 m = M_FULL;
1893 if (kmem_cache_debug(s) && !lock) {
1894 lock = 1;
1895 /*
1896 * This also ensures that the scanning of full
1897 * slabs from diagnostic functions will not see
1898 * any frozen slabs.
1899 */
1900 spin_lock(&n->list_lock);
1901 }
1902 }
1903
1904 if (l != m) {
1905
1906 if (l == M_PARTIAL)
1907
1908 remove_partial(n, page);
1909
1910 else if (l == M_FULL)
1911
1912 remove_full(s, n, page);
1913
1914 if (m == M_PARTIAL) {
1915
1916 add_partial(n, page, tail);
1917 stat(s, tail);
1918
1919 } else if (m == M_FULL) {
1920
1921 stat(s, DEACTIVATE_FULL);
1922 add_full(s, n, page);
1923
1924 }
1925 }
1926
1927 l = m;
1928 if (!__cmpxchg_double_slab(s, page,
1929 old.freelist, old.counters,
1930 new.freelist, new.counters,
1931 "unfreezing slab"))
1932 goto redo;
1933
1934 if (lock)
1935 spin_unlock(&n->list_lock);
1936
1937 if (m == M_FREE) {
1938 stat(s, DEACTIVATE_EMPTY);
1939 discard_slab(s, page);
1940 stat(s, FREE_SLAB);
1941 }
1942}
1943
1944/*
1945 * Unfreeze all the cpu partial slabs.
1946 *
1947 * This function must be called with interrupts disabled
1948 * for the cpu using c (or some other guarantee must be there
1949 * to guarantee no concurrent accesses).
1950 */
1951static void unfreeze_partials(struct kmem_cache *s,
1952 struct kmem_cache_cpu *c)
1953{
1954#ifdef CONFIG_SLUB_CPU_PARTIAL
1955 struct kmem_cache_node *n = NULL, *n2 = NULL;
1956 struct page *page, *discard_page = NULL;
1957
1958 while ((page = c->partial)) {
1959 struct page new;
1960 struct page old;
1961
1962 c->partial = page->next;
1963
1964 n2 = get_node(s, page_to_nid(page));
1965 if (n != n2) {
1966 if (n)
1967 spin_unlock(&n->list_lock);
1968
1969 n = n2;
1970 spin_lock(&n->list_lock);
1971 }
1972
1973 do {
1974
1975 old.freelist = page->freelist;
1976 old.counters = page->counters;
1977 VM_BUG_ON(!old.frozen);
1978
1979 new.counters = old.counters;
1980 new.freelist = old.freelist;
1981
1982 new.frozen = 0;
1983
1984 } while (!__cmpxchg_double_slab(s, page,
1985 old.freelist, old.counters,
1986 new.freelist, new.counters,
1987 "unfreezing slab"));
1988
1989 if (unlikely(!new.inuse && n->nr_partial > s->min_partial)) {
1990 page->next = discard_page;
1991 discard_page = page;
1992 } else {
1993 add_partial(n, page, DEACTIVATE_TO_TAIL);
1994 stat(s, FREE_ADD_PARTIAL);
1995 }
1996 }
1997
1998 if (n)
1999 spin_unlock(&n->list_lock);
2000
2001 while (discard_page) {
2002 page = discard_page;
2003 discard_page = discard_page->next;
2004
2005 stat(s, DEACTIVATE_EMPTY);
2006 discard_slab(s, page);
2007 stat(s, FREE_SLAB);
2008 }
2009#endif
2010}
2011
2012/*
2013 * Put a page that was just frozen (in __slab_free) into a partial page
2014 * slot if available. This is done without interrupts disabled and without
2015 * preemption disabled. The cmpxchg is racy and may put the partial page
2016 * onto a random cpus partial slot.
2017 *
2018 * If we did not find a slot then simply move all the partials to the
2019 * per node partial list.
2020 */
2021static void put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
2022{
2023#ifdef CONFIG_SLUB_CPU_PARTIAL
2024 struct page *oldpage;
2025 int pages;
2026 int pobjects;
2027
2028 do {
2029 pages = 0;
2030 pobjects = 0;
2031 oldpage = this_cpu_read(s->cpu_slab->partial);
2032
2033 if (oldpage) {
2034 pobjects = oldpage->pobjects;
2035 pages = oldpage->pages;
2036 if (drain && pobjects > s->cpu_partial) {
2037 unsigned long flags;
2038 /*
2039 * partial array is full. Move the existing
2040 * set to the per node partial list.
2041 */
2042 local_irq_save(flags);
2043 unfreeze_partials(s, this_cpu_ptr(s->cpu_slab));
2044 local_irq_restore(flags);
2045 oldpage = NULL;
2046 pobjects = 0;
2047 pages = 0;
2048 stat(s, CPU_PARTIAL_DRAIN);
2049 }
2050 }
2051
2052 pages++;
2053 pobjects += page->objects - page->inuse;
2054
2055 page->pages = pages;
2056 page->pobjects = pobjects;
2057 page->next = oldpage;
2058
2059 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page)
2060 != oldpage);
2061#endif
2062}
2063
2064static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2065{
2066 stat(s, CPUSLAB_FLUSH);
2067 deactivate_slab(s, c->page, c->freelist);
2068
2069 c->tid = next_tid(c->tid);
2070 c->page = NULL;
2071 c->freelist = NULL;
2072}
2073
2074/*
2075 * Flush cpu slab.
2076 *
2077 * Called from IPI handler with interrupts disabled.
2078 */
2079static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2080{
2081 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2082
2083 if (likely(c)) {
2084 if (c->page)
2085 flush_slab(s, c);
2086
2087 unfreeze_partials(s, c);
2088 }
2089}
2090
2091static void flush_cpu_slab(void *d)
2092{
2093 struct kmem_cache *s = d;
2094
2095 __flush_cpu_slab(s, smp_processor_id());
2096}
2097
2098static bool has_cpu_slab(int cpu, void *info)
2099{
2100 struct kmem_cache *s = info;
2101 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2102
2103 return c->page || c->partial;
2104}
2105
2106static void flush_all(struct kmem_cache *s)
2107{
2108 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2109}
2110
2111/*
2112 * Check if the objects in a per cpu structure fit numa
2113 * locality expectations.
2114 */
2115static inline int node_match(struct page *page, int node)
2116{
2117#ifdef CONFIG_NUMA
2118 if (!page || (node != NUMA_NO_NODE && page_to_nid(page) != node))
2119 return 0;
2120#endif
2121 return 1;
2122}
2123
2124static int count_free(struct page *page)
2125{
2126 return page->objects - page->inuse;
2127}
2128
2129static unsigned long count_partial(struct kmem_cache_node *n,
2130 int (*get_count)(struct page *))
2131{
2132 unsigned long flags;
2133 unsigned long x = 0;
2134 struct page *page;
2135
2136 spin_lock_irqsave(&n->list_lock, flags);
2137 list_for_each_entry(page, &n->partial, lru)
2138 x += get_count(page);
2139 spin_unlock_irqrestore(&n->list_lock, flags);
2140 return x;
2141}
2142
2143static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2144{
2145#ifdef CONFIG_SLUB_DEBUG
2146 return atomic_long_read(&n->total_objects);
2147#else
2148 return 0;
2149#endif
2150}
2151
2152static noinline void
2153slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2154{
2155 int node;
2156
2157 printk(KERN_WARNING
2158 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2159 nid, gfpflags);
2160 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2161 "default order: %d, min order: %d\n", s->name, s->object_size,
2162 s->size, oo_order(s->oo), oo_order(s->min));
2163
2164 if (oo_order(s->min) > get_order(s->object_size))
2165 printk(KERN_WARNING " %s debugging increased min order, use "
2166 "slub_debug=O to disable.\n", s->name);
2167
2168 for_each_online_node(node) {
2169 struct kmem_cache_node *n = get_node(s, node);
2170 unsigned long nr_slabs;
2171 unsigned long nr_objs;
2172 unsigned long nr_free;
2173
2174 if (!n)
2175 continue;
2176
2177 nr_free = count_partial(n, count_free);
2178 nr_slabs = node_nr_slabs(n);
2179 nr_objs = node_nr_objs(n);
2180
2181 printk(KERN_WARNING
2182 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2183 node, nr_slabs, nr_objs, nr_free);
2184 }
2185}
2186
2187static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2188 int node, struct kmem_cache_cpu **pc)
2189{
2190 void *freelist;
2191 struct kmem_cache_cpu *c = *pc;
2192 struct page *page;
2193
2194 freelist = get_partial(s, flags, node, c);
2195
2196 if (freelist)
2197 return freelist;
2198
2199 page = new_slab(s, flags, node);
2200 if (page) {
2201 c = __this_cpu_ptr(s->cpu_slab);
2202 if (c->page)
2203 flush_slab(s, c);
2204
2205 /*
2206 * No other reference to the page yet so we can
2207 * muck around with it freely without cmpxchg
2208 */
2209 freelist = page->freelist;
2210 page->freelist = NULL;
2211
2212 stat(s, ALLOC_SLAB);
2213 c->page = page;
2214 *pc = c;
2215 } else
2216 freelist = NULL;
2217
2218 return freelist;
2219}
2220
2221static inline bool pfmemalloc_match(struct page *page, gfp_t gfpflags)
2222{
2223 if (unlikely(PageSlabPfmemalloc(page)))
2224 return gfp_pfmemalloc_allowed(gfpflags);
2225
2226 return true;
2227}
2228
2229/*
2230 * Check the page->freelist of a page and either transfer the freelist to the
2231 * per cpu freelist or deactivate the page.
2232 *
2233 * The page is still frozen if the return value is not NULL.
2234 *
2235 * If this function returns NULL then the page has been unfrozen.
2236 *
2237 * This function must be called with interrupt disabled.
2238 */
2239static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2240{
2241 struct page new;
2242 unsigned long counters;
2243 void *freelist;
2244
2245 do {
2246 freelist = page->freelist;
2247 counters = page->counters;
2248
2249 new.counters = counters;
2250 VM_BUG_ON(!new.frozen);
2251
2252 new.inuse = page->objects;
2253 new.frozen = freelist != NULL;
2254
2255 } while (!__cmpxchg_double_slab(s, page,
2256 freelist, counters,
2257 NULL, new.counters,
2258 "get_freelist"));
2259
2260 return freelist;
2261}
2262
2263/*
2264 * Slow path. The lockless freelist is empty or we need to perform
2265 * debugging duties.
2266 *
2267 * Processing is still very fast if new objects have been freed to the
2268 * regular freelist. In that case we simply take over the regular freelist
2269 * as the lockless freelist and zap the regular freelist.
2270 *
2271 * If that is not working then we fall back to the partial lists. We take the
2272 * first element of the freelist as the object to allocate now and move the
2273 * rest of the freelist to the lockless freelist.
2274 *
2275 * And if we were unable to get a new slab from the partial slab lists then
2276 * we need to allocate a new slab. This is the slowest path since it involves
2277 * a call to the page allocator and the setup of a new slab.
2278 */
2279static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2280 unsigned long addr, struct kmem_cache_cpu *c)
2281{
2282 void *freelist;
2283 struct page *page;
2284 unsigned long flags;
2285
2286 local_irq_save(flags);
2287#ifdef CONFIG_PREEMPT
2288 /*
2289 * We may have been preempted and rescheduled on a different
2290 * cpu before disabling interrupts. Need to reload cpu area
2291 * pointer.
2292 */
2293 c = this_cpu_ptr(s->cpu_slab);
2294#endif
2295
2296 page = c->page;
2297 if (!page)
2298 goto new_slab;
2299redo:
2300
2301 if (unlikely(!node_match(page, node))) {
2302 stat(s, ALLOC_NODE_MISMATCH);
2303 deactivate_slab(s, page, c->freelist);
2304 c->page = NULL;
2305 c->freelist = NULL;
2306 goto new_slab;
2307 }
2308
2309 /*
2310 * By rights, we should be searching for a slab page that was
2311 * PFMEMALLOC but right now, we are losing the pfmemalloc
2312 * information when the page leaves the per-cpu allocator
2313 */
2314 if (unlikely(!pfmemalloc_match(page, gfpflags))) {
2315 deactivate_slab(s, page, c->freelist);
2316 c->page = NULL;
2317 c->freelist = NULL;
2318 goto new_slab;
2319 }
2320
2321 /* must check again c->freelist in case of cpu migration or IRQ */
2322 freelist = c->freelist;
2323 if (freelist)
2324 goto load_freelist;
2325
2326 stat(s, ALLOC_SLOWPATH);
2327
2328 freelist = get_freelist(s, page);
2329
2330 if (!freelist) {
2331 c->page = NULL;
2332 stat(s, DEACTIVATE_BYPASS);
2333 goto new_slab;
2334 }
2335
2336 stat(s, ALLOC_REFILL);
2337
2338load_freelist:
2339 /*
2340 * freelist is pointing to the list of objects to be used.
2341 * page is pointing to the page from which the objects are obtained.
2342 * That page must be frozen for per cpu allocations to work.
2343 */
2344 VM_BUG_ON(!c->page->frozen);
2345 c->freelist = get_freepointer(s, freelist);
2346 c->tid = next_tid(c->tid);
2347 local_irq_restore(flags);
2348 return freelist;
2349
2350new_slab:
2351
2352 if (c->partial) {
2353 page = c->page = c->partial;
2354 c->partial = page->next;
2355 stat(s, CPU_PARTIAL_ALLOC);
2356 c->freelist = NULL;
2357 goto redo;
2358 }
2359
2360 freelist = new_slab_objects(s, gfpflags, node, &c);
2361
2362 if (unlikely(!freelist)) {
2363 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2364 slab_out_of_memory(s, gfpflags, node);
2365
2366 local_irq_restore(flags);
2367 return NULL;
2368 }
2369
2370 page = c->page;
2371 if (likely(!kmem_cache_debug(s) && pfmemalloc_match(page, gfpflags)))
2372 goto load_freelist;
2373
2374 /* Only entered in the debug case */
2375 if (kmem_cache_debug(s) &&
2376 !alloc_debug_processing(s, page, freelist, addr))
2377 goto new_slab; /* Slab failed checks. Next slab needed */
2378
2379 deactivate_slab(s, page, get_freepointer(s, freelist));
2380 c->page = NULL;
2381 c->freelist = NULL;
2382 local_irq_restore(flags);
2383 return freelist;
2384}
2385
2386/*
2387 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2388 * have the fastpath folded into their functions. So no function call
2389 * overhead for requests that can be satisfied on the fastpath.
2390 *
2391 * The fastpath works by first checking if the lockless freelist can be used.
2392 * If not then __slab_alloc is called for slow processing.
2393 *
2394 * Otherwise we can simply pick the next object from the lockless free list.
2395 */
2396static __always_inline void *slab_alloc_node(struct kmem_cache *s,
2397 gfp_t gfpflags, int node, unsigned long addr)
2398{
2399 void **object;
2400 struct kmem_cache_cpu *c;
2401 struct page *page;
2402 unsigned long tid;
2403
2404 if (slab_pre_alloc_hook(s, gfpflags))
2405 return NULL;
2406
2407 s = memcg_kmem_get_cache(s, gfpflags);
2408redo:
2409 /*
2410 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2411 * enabled. We may switch back and forth between cpus while
2412 * reading from one cpu area. That does not matter as long
2413 * as we end up on the original cpu again when doing the cmpxchg.
2414 *
2415 * Preemption is disabled for the retrieval of the tid because that
2416 * must occur from the current processor. We cannot allow rescheduling
2417 * on a different processor between the determination of the pointer
2418 * and the retrieval of the tid.
2419 */
2420 preempt_disable();
2421 c = __this_cpu_ptr(s->cpu_slab);
2422
2423 /*
2424 * The transaction ids are globally unique per cpu and per operation on
2425 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2426 * occurs on the right processor and that there was no operation on the
2427 * linked list in between.
2428 */
2429 tid = c->tid;
2430 preempt_enable();
2431
2432 object = c->freelist;
2433 page = c->page;
2434 if (unlikely(!object || !node_match(page, node)))
2435 object = __slab_alloc(s, gfpflags, node, addr, c);
2436
2437 else {
2438 void *next_object = get_freepointer_safe(s, object);
2439
2440 /*
2441 * The cmpxchg will only match if there was no additional
2442 * operation and if we are on the right processor.
2443 *
2444 * The cmpxchg does the following atomically (without lock
2445 * semantics!)
2446 * 1. Relocate first pointer to the current per cpu area.
2447 * 2. Verify that tid and freelist have not been changed
2448 * 3. If they were not changed replace tid and freelist
2449 *
2450 * Since this is without lock semantics the protection is only
2451 * against code executing on this cpu *not* from access by
2452 * other cpus.
2453 */
2454 if (unlikely(!this_cpu_cmpxchg_double(
2455 s->cpu_slab->freelist, s->cpu_slab->tid,
2456 object, tid,
2457 next_object, next_tid(tid)))) {
2458
2459 note_cmpxchg_failure("slab_alloc", s, tid);
2460 goto redo;
2461 }
2462 prefetch_freepointer(s, next_object);
2463 stat(s, ALLOC_FASTPATH);
2464 }
2465
2466 if (unlikely(gfpflags & __GFP_ZERO) && object)
2467 memset(object, 0, s->object_size);
2468
2469 slab_post_alloc_hook(s, gfpflags, object);
2470
2471 return object;
2472}
2473
2474static __always_inline void *slab_alloc(struct kmem_cache *s,
2475 gfp_t gfpflags, unsigned long addr)
2476{
2477 return slab_alloc_node(s, gfpflags, NUMA_NO_NODE, addr);
2478}
2479
2480void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2481{
2482 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2483
2484 trace_kmem_cache_alloc(_RET_IP_, ret, s->object_size,
2485 s->size, gfpflags);
2486
2487 return ret;
2488}
2489EXPORT_SYMBOL(kmem_cache_alloc);
2490
2491#ifdef CONFIG_TRACING
2492void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2493{
2494 void *ret = slab_alloc(s, gfpflags, _RET_IP_);
2495 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2496 return ret;
2497}
2498EXPORT_SYMBOL(kmem_cache_alloc_trace);
2499#endif
2500
2501#ifdef CONFIG_NUMA
2502void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2503{
2504 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2505
2506 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2507 s->object_size, s->size, gfpflags, node);
2508
2509 return ret;
2510}
2511EXPORT_SYMBOL(kmem_cache_alloc_node);
2512
2513#ifdef CONFIG_TRACING
2514void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2515 gfp_t gfpflags,
2516 int node, size_t size)
2517{
2518 void *ret = slab_alloc_node(s, gfpflags, node, _RET_IP_);
2519
2520 trace_kmalloc_node(_RET_IP_, ret,
2521 size, s->size, gfpflags, node);
2522 return ret;
2523}
2524EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2525#endif
2526#endif
2527
2528/*
2529 * Slow patch handling. This may still be called frequently since objects
2530 * have a longer lifetime than the cpu slabs in most processing loads.
2531 *
2532 * So we still attempt to reduce cache line usage. Just take the slab
2533 * lock and free the item. If there is no additional partial page
2534 * handling required then we can return immediately.
2535 */
2536static void __slab_free(struct kmem_cache *s, struct page *page,
2537 void *x, unsigned long addr)
2538{
2539 void *prior;
2540 void **object = (void *)x;
2541 int was_frozen;
2542 struct page new;
2543 unsigned long counters;
2544 struct kmem_cache_node *n = NULL;
2545 unsigned long uninitialized_var(flags);
2546
2547 stat(s, FREE_SLOWPATH);
2548
2549 if (kmem_cache_debug(s) &&
2550 !(n = free_debug_processing(s, page, x, addr, &flags)))
2551 return;
2552
2553 do {
2554 if (unlikely(n)) {
2555 spin_unlock_irqrestore(&n->list_lock, flags);
2556 n = NULL;
2557 }
2558 prior = page->freelist;
2559 counters = page->counters;
2560 set_freepointer(s, object, prior);
2561 new.counters = counters;
2562 was_frozen = new.frozen;
2563 new.inuse--;
2564 if ((!new.inuse || !prior) && !was_frozen) {
2565
2566 if (kmem_cache_has_cpu_partial(s) && !prior) {
2567
2568 /*
2569 * Slab was on no list before and will be
2570 * partially empty
2571 * We can defer the list move and instead
2572 * freeze it.
2573 */
2574 new.frozen = 1;
2575
2576 } else { /* Needs to be taken off a list */
2577
2578 n = get_node(s, page_to_nid(page));
2579 /*
2580 * Speculatively acquire the list_lock.
2581 * If the cmpxchg does not succeed then we may
2582 * drop the list_lock without any processing.
2583 *
2584 * Otherwise the list_lock will synchronize with
2585 * other processors updating the list of slabs.
2586 */
2587 spin_lock_irqsave(&n->list_lock, flags);
2588
2589 }
2590 }
2591
2592 } while (!cmpxchg_double_slab(s, page,
2593 prior, counters,
2594 object, new.counters,
2595 "__slab_free"));
2596
2597 if (likely(!n)) {
2598
2599 /*
2600 * If we just froze the page then put it onto the
2601 * per cpu partial list.
2602 */
2603 if (new.frozen && !was_frozen) {
2604 put_cpu_partial(s, page, 1);
2605 stat(s, CPU_PARTIAL_FREE);
2606 }
2607 /*
2608 * The list lock was not taken therefore no list
2609 * activity can be necessary.
2610 */
2611 if (was_frozen)
2612 stat(s, FREE_FROZEN);
2613 return;
2614 }
2615
2616 if (unlikely(!new.inuse && n->nr_partial > s->min_partial))
2617 goto slab_empty;
2618
2619 /*
2620 * Objects left in the slab. If it was not on the partial list before
2621 * then add it.
2622 */
2623 if (!kmem_cache_has_cpu_partial(s) && unlikely(!prior)) {
2624 if (kmem_cache_debug(s))
2625 remove_full(s, n, page);
2626 add_partial(n, page, DEACTIVATE_TO_TAIL);
2627 stat(s, FREE_ADD_PARTIAL);
2628 }
2629 spin_unlock_irqrestore(&n->list_lock, flags);
2630 return;
2631
2632slab_empty:
2633 if (prior) {
2634 /*
2635 * Slab on the partial list.
2636 */
2637 remove_partial(n, page);
2638 stat(s, FREE_REMOVE_PARTIAL);
2639 } else {
2640 /* Slab must be on the full list */
2641 remove_full(s, n, page);
2642 }
2643
2644 spin_unlock_irqrestore(&n->list_lock, flags);
2645 stat(s, FREE_SLAB);
2646 discard_slab(s, page);
2647}
2648
2649/*
2650 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2651 * can perform fastpath freeing without additional function calls.
2652 *
2653 * The fastpath is only possible if we are freeing to the current cpu slab
2654 * of this processor. This typically the case if we have just allocated
2655 * the item before.
2656 *
2657 * If fastpath is not possible then fall back to __slab_free where we deal
2658 * with all sorts of special processing.
2659 */
2660static __always_inline void slab_free(struct kmem_cache *s,
2661 struct page *page, void *x, unsigned long addr)
2662{
2663 void **object = (void *)x;
2664 struct kmem_cache_cpu *c;
2665 unsigned long tid;
2666
2667 slab_free_hook(s, x);
2668
2669redo:
2670 /*
2671 * Determine the currently cpus per cpu slab.
2672 * The cpu may change afterward. However that does not matter since
2673 * data is retrieved via this pointer. If we are on the same cpu
2674 * during the cmpxchg then the free will succedd.
2675 */
2676 preempt_disable();
2677 c = __this_cpu_ptr(s->cpu_slab);
2678
2679 tid = c->tid;
2680 preempt_enable();
2681
2682 if (likely(page == c->page)) {
2683 set_freepointer(s, object, c->freelist);
2684
2685 if (unlikely(!this_cpu_cmpxchg_double(
2686 s->cpu_slab->freelist, s->cpu_slab->tid,
2687 c->freelist, tid,
2688 object, next_tid(tid)))) {
2689
2690 note_cmpxchg_failure("slab_free", s, tid);
2691 goto redo;
2692 }
2693 stat(s, FREE_FASTPATH);
2694 } else
2695 __slab_free(s, page, x, addr);
2696
2697}
2698
2699void kmem_cache_free(struct kmem_cache *s, void *x)
2700{
2701 s = cache_from_obj(s, x);
2702 if (!s)
2703 return;
2704 slab_free(s, virt_to_head_page(x), x, _RET_IP_);
2705 trace_kmem_cache_free(_RET_IP_, x);
2706}
2707EXPORT_SYMBOL(kmem_cache_free);
2708
2709/*
2710 * Object placement in a slab is made very easy because we always start at
2711 * offset 0. If we tune the size of the object to the alignment then we can
2712 * get the required alignment by putting one properly sized object after
2713 * another.
2714 *
2715 * Notice that the allocation order determines the sizes of the per cpu
2716 * caches. Each processor has always one slab available for allocations.
2717 * Increasing the allocation order reduces the number of times that slabs
2718 * must be moved on and off the partial lists and is therefore a factor in
2719 * locking overhead.
2720 */
2721
2722/*
2723 * Mininum / Maximum order of slab pages. This influences locking overhead
2724 * and slab fragmentation. A higher order reduces the number of partial slabs
2725 * and increases the number of allocations possible without having to
2726 * take the list_lock.
2727 */
2728static int slub_min_order;
2729static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2730static int slub_min_objects;
2731
2732/*
2733 * Merge control. If this is set then no merging of slab caches will occur.
2734 * (Could be removed. This was introduced to pacify the merge skeptics.)
2735 */
2736static int slub_nomerge;
2737
2738/*
2739 * Calculate the order of allocation given an slab object size.
2740 *
2741 * The order of allocation has significant impact on performance and other
2742 * system components. Generally order 0 allocations should be preferred since
2743 * order 0 does not cause fragmentation in the page allocator. Larger objects
2744 * be problematic to put into order 0 slabs because there may be too much
2745 * unused space left. We go to a higher order if more than 1/16th of the slab
2746 * would be wasted.
2747 *
2748 * In order to reach satisfactory performance we must ensure that a minimum
2749 * number of objects is in one slab. Otherwise we may generate too much
2750 * activity on the partial lists which requires taking the list_lock. This is
2751 * less a concern for large slabs though which are rarely used.
2752 *
2753 * slub_max_order specifies the order where we begin to stop considering the
2754 * number of objects in a slab as critical. If we reach slub_max_order then
2755 * we try to keep the page order as low as possible. So we accept more waste
2756 * of space in favor of a small page order.
2757 *
2758 * Higher order allocations also allow the placement of more objects in a
2759 * slab and thereby reduce object handling overhead. If the user has
2760 * requested a higher mininum order then we start with that one instead of
2761 * the smallest order which will fit the object.
2762 */
2763static inline int slab_order(int size, int min_objects,
2764 int max_order, int fract_leftover, int reserved)
2765{
2766 int order;
2767 int rem;
2768 int min_order = slub_min_order;
2769
2770 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2771 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2772
2773 for (order = max(min_order,
2774 fls(min_objects * size - 1) - PAGE_SHIFT);
2775 order <= max_order; order++) {
2776
2777 unsigned long slab_size = PAGE_SIZE << order;
2778
2779 if (slab_size < min_objects * size + reserved)
2780 continue;
2781
2782 rem = (slab_size - reserved) % size;
2783
2784 if (rem <= slab_size / fract_leftover)
2785 break;
2786
2787 }
2788
2789 return order;
2790}
2791
2792static inline int calculate_order(int size, int reserved)
2793{
2794 int order;
2795 int min_objects;
2796 int fraction;
2797 int max_objects;
2798
2799 /*
2800 * Attempt to find best configuration for a slab. This
2801 * works by first attempting to generate a layout with
2802 * the best configuration and backing off gradually.
2803 *
2804 * First we reduce the acceptable waste in a slab. Then
2805 * we reduce the minimum objects required in a slab.
2806 */
2807 min_objects = slub_min_objects;
2808 if (!min_objects)
2809 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2810 max_objects = order_objects(slub_max_order, size, reserved);
2811 min_objects = min(min_objects, max_objects);
2812
2813 while (min_objects > 1) {
2814 fraction = 16;
2815 while (fraction >= 4) {
2816 order = slab_order(size, min_objects,
2817 slub_max_order, fraction, reserved);
2818 if (order <= slub_max_order)
2819 return order;
2820 fraction /= 2;
2821 }
2822 min_objects--;
2823 }
2824
2825 /*
2826 * We were unable to place multiple objects in a slab. Now
2827 * lets see if we can place a single object there.
2828 */
2829 order = slab_order(size, 1, slub_max_order, 1, reserved);
2830 if (order <= slub_max_order)
2831 return order;
2832
2833 /*
2834 * Doh this slab cannot be placed using slub_max_order.
2835 */
2836 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2837 if (order < MAX_ORDER)
2838 return order;
2839 return -ENOSYS;
2840}
2841
2842static void
2843init_kmem_cache_node(struct kmem_cache_node *n)
2844{
2845 n->nr_partial = 0;
2846 spin_lock_init(&n->list_lock);
2847 INIT_LIST_HEAD(&n->partial);
2848#ifdef CONFIG_SLUB_DEBUG
2849 atomic_long_set(&n->nr_slabs, 0);
2850 atomic_long_set(&n->total_objects, 0);
2851 INIT_LIST_HEAD(&n->full);
2852#endif
2853}
2854
2855static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2856{
2857 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2858 KMALLOC_SHIFT_HIGH * sizeof(struct kmem_cache_cpu));
2859
2860 /*
2861 * Must align to double word boundary for the double cmpxchg
2862 * instructions to work; see __pcpu_double_call_return_bool().
2863 */
2864 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2865 2 * sizeof(void *));
2866
2867 if (!s->cpu_slab)
2868 return 0;
2869
2870 init_kmem_cache_cpus(s);
2871
2872 return 1;
2873}
2874
2875static struct kmem_cache *kmem_cache_node;
2876
2877/*
2878 * No kmalloc_node yet so do it by hand. We know that this is the first
2879 * slab on the node for this slabcache. There are no concurrent accesses
2880 * possible.
2881 *
2882 * Note that this function only works on the kmem_cache_node
2883 * when allocating for the kmem_cache_node. This is used for bootstrapping
2884 * memory on a fresh node that has no slab structures yet.
2885 */
2886static void early_kmem_cache_node_alloc(int node)
2887{
2888 struct page *page;
2889 struct kmem_cache_node *n;
2890
2891 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2892
2893 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2894
2895 BUG_ON(!page);
2896 if (page_to_nid(page) != node) {
2897 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2898 "node %d\n", node);
2899 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2900 "in order to be able to continue\n");
2901 }
2902
2903 n = page->freelist;
2904 BUG_ON(!n);
2905 page->freelist = get_freepointer(kmem_cache_node, n);
2906 page->inuse = 1;
2907 page->frozen = 0;
2908 kmem_cache_node->node[node] = n;
2909#ifdef CONFIG_SLUB_DEBUG
2910 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2911 init_tracking(kmem_cache_node, n);
2912#endif
2913 init_kmem_cache_node(n);
2914 inc_slabs_node(kmem_cache_node, node, page->objects);
2915
2916 /*
2917 * No locks need to be taken here as it has just been
2918 * initialized and there is no concurrent access.
2919 */
2920 __add_partial(n, page, DEACTIVATE_TO_HEAD);
2921}
2922
2923static void free_kmem_cache_nodes(struct kmem_cache *s)
2924{
2925 int node;
2926
2927 for_each_node_state(node, N_NORMAL_MEMORY) {
2928 struct kmem_cache_node *n = s->node[node];
2929
2930 if (n)
2931 kmem_cache_free(kmem_cache_node, n);
2932
2933 s->node[node] = NULL;
2934 }
2935}
2936
2937static int init_kmem_cache_nodes(struct kmem_cache *s)
2938{
2939 int node;
2940
2941 for_each_node_state(node, N_NORMAL_MEMORY) {
2942 struct kmem_cache_node *n;
2943
2944 if (slab_state == DOWN) {
2945 early_kmem_cache_node_alloc(node);
2946 continue;
2947 }
2948 n = kmem_cache_alloc_node(kmem_cache_node,
2949 GFP_KERNEL, node);
2950
2951 if (!n) {
2952 free_kmem_cache_nodes(s);
2953 return 0;
2954 }
2955
2956 s->node[node] = n;
2957 init_kmem_cache_node(n);
2958 }
2959 return 1;
2960}
2961
2962static void set_min_partial(struct kmem_cache *s, unsigned long min)
2963{
2964 if (min < MIN_PARTIAL)
2965 min = MIN_PARTIAL;
2966 else if (min > MAX_PARTIAL)
2967 min = MAX_PARTIAL;
2968 s->min_partial = min;
2969}
2970
2971/*
2972 * calculate_sizes() determines the order and the distribution of data within
2973 * a slab object.
2974 */
2975static int calculate_sizes(struct kmem_cache *s, int forced_order)
2976{
2977 unsigned long flags = s->flags;
2978 unsigned long size = s->object_size;
2979 int order;
2980
2981 /*
2982 * Round up object size to the next word boundary. We can only
2983 * place the free pointer at word boundaries and this determines
2984 * the possible location of the free pointer.
2985 */
2986 size = ALIGN(size, sizeof(void *));
2987
2988#ifdef CONFIG_SLUB_DEBUG
2989 /*
2990 * Determine if we can poison the object itself. If the user of
2991 * the slab may touch the object after free or before allocation
2992 * then we should never poison the object itself.
2993 */
2994 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2995 !s->ctor)
2996 s->flags |= __OBJECT_POISON;
2997 else
2998 s->flags &= ~__OBJECT_POISON;
2999
3000
3001 /*
3002 * If we are Redzoning then check if there is some space between the
3003 * end of the object and the free pointer. If not then add an
3004 * additional word to have some bytes to store Redzone information.
3005 */
3006 if ((flags & SLAB_RED_ZONE) && size == s->object_size)
3007 size += sizeof(void *);
3008#endif
3009
3010 /*
3011 * With that we have determined the number of bytes in actual use
3012 * by the object. This is the potential offset to the free pointer.
3013 */
3014 s->inuse = size;
3015
3016 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
3017 s->ctor)) {
3018 /*
3019 * Relocate free pointer after the object if it is not
3020 * permitted to overwrite the first word of the object on
3021 * kmem_cache_free.
3022 *
3023 * This is the case if we do RCU, have a constructor or
3024 * destructor or are poisoning the objects.
3025 */
3026 s->offset = size;
3027 size += sizeof(void *);
3028 }
3029
3030#ifdef CONFIG_SLUB_DEBUG
3031 if (flags & SLAB_STORE_USER)
3032 /*
3033 * Need to store information about allocs and frees after
3034 * the object.
3035 */
3036 size += 2 * sizeof(struct track);
3037
3038 if (flags & SLAB_RED_ZONE)
3039 /*
3040 * Add some empty padding so that we can catch
3041 * overwrites from earlier objects rather than let
3042 * tracking information or the free pointer be
3043 * corrupted if a user writes before the start
3044 * of the object.
3045 */
3046 size += sizeof(void *);
3047#endif
3048
3049 /*
3050 * SLUB stores one object immediately after another beginning from
3051 * offset 0. In order to align the objects we have to simply size
3052 * each object to conform to the alignment.
3053 */
3054 size = ALIGN(size, s->align);
3055 s->size = size;
3056 if (forced_order >= 0)
3057 order = forced_order;
3058 else
3059 order = calculate_order(size, s->reserved);
3060
3061 if (order < 0)
3062 return 0;
3063
3064 s->allocflags = 0;
3065 if (order)
3066 s->allocflags |= __GFP_COMP;
3067
3068 if (s->flags & SLAB_CACHE_DMA)
3069 s->allocflags |= GFP_DMA;
3070
3071 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3072 s->allocflags |= __GFP_RECLAIMABLE;
3073
3074 /*
3075 * Determine the number of objects per slab
3076 */
3077 s->oo = oo_make(order, size, s->reserved);
3078 s->min = oo_make(get_order(size), size, s->reserved);
3079 if (oo_objects(s->oo) > oo_objects(s->max))
3080 s->max = s->oo;
3081
3082 return !!oo_objects(s->oo);
3083}
3084
3085static int kmem_cache_open(struct kmem_cache *s, unsigned long flags)
3086{
3087 s->flags = kmem_cache_flags(s->size, flags, s->name, s->ctor);
3088 s->reserved = 0;
3089
3090 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3091 s->reserved = sizeof(struct rcu_head);
3092
3093 if (!calculate_sizes(s, -1))
3094 goto error;
3095 if (disable_higher_order_debug) {
3096 /*
3097 * Disable debugging flags that store metadata if the min slab
3098 * order increased.
3099 */
3100 if (get_order(s->size) > get_order(s->object_size)) {
3101 s->flags &= ~DEBUG_METADATA_FLAGS;
3102 s->offset = 0;
3103 if (!calculate_sizes(s, -1))
3104 goto error;
3105 }
3106 }
3107
3108#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3109 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3110 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3111 /* Enable fast mode */
3112 s->flags |= __CMPXCHG_DOUBLE;
3113#endif
3114
3115 /*
3116 * The larger the object size is, the more pages we want on the partial
3117 * list to avoid pounding the page allocator excessively.
3118 */
3119 set_min_partial(s, ilog2(s->size) / 2);
3120
3121 /*
3122 * cpu_partial determined the maximum number of objects kept in the
3123 * per cpu partial lists of a processor.
3124 *
3125 * Per cpu partial lists mainly contain slabs that just have one
3126 * object freed. If they are used for allocation then they can be
3127 * filled up again with minimal effort. The slab will never hit the
3128 * per node partial lists and therefore no locking will be required.
3129 *
3130 * This setting also determines
3131 *
3132 * A) The number of objects from per cpu partial slabs dumped to the
3133 * per node list when we reach the limit.
3134 * B) The number of objects in cpu partial slabs to extract from the
3135 * per node list when we run out of per cpu objects. We only fetch
3136 * 50% to keep some capacity around for frees.
3137 */
3138 if (!kmem_cache_has_cpu_partial(s))
3139 s->cpu_partial = 0;
3140 else if (s->size >= PAGE_SIZE)
3141 s->cpu_partial = 2;
3142 else if (s->size >= 1024)
3143 s->cpu_partial = 6;
3144 else if (s->size >= 256)
3145 s->cpu_partial = 13;
3146 else
3147 s->cpu_partial = 30;
3148
3149#ifdef CONFIG_NUMA
3150 s->remote_node_defrag_ratio = 1000;
3151#endif
3152 if (!init_kmem_cache_nodes(s))
3153 goto error;
3154
3155 if (alloc_kmem_cache_cpus(s))
3156 return 0;
3157
3158 free_kmem_cache_nodes(s);
3159error:
3160 if (flags & SLAB_PANIC)
3161 panic("Cannot create slab %s size=%lu realsize=%u "
3162 "order=%u offset=%u flags=%lx\n",
3163 s->name, (unsigned long)s->size, s->size,
3164 oo_order(s->oo), s->offset, flags);
3165 return -EINVAL;
3166}
3167
3168static void list_slab_objects(struct kmem_cache *s, struct page *page,
3169 const char *text)
3170{
3171#ifdef CONFIG_SLUB_DEBUG
3172 void *addr = page_address(page);
3173 void *p;
3174 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3175 sizeof(long), GFP_ATOMIC);
3176 if (!map)
3177 return;
3178 slab_err(s, page, text, s->name);
3179 slab_lock(page);
3180
3181 get_map(s, page, map);
3182 for_each_object(p, s, addr, page->objects) {
3183
3184 if (!test_bit(slab_index(p, s, addr), map)) {
3185 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3186 p, p - addr);
3187 print_tracking(s, p);
3188 }
3189 }
3190 slab_unlock(page);
3191 kfree(map);
3192#endif
3193}
3194
3195/*
3196 * Attempt to free all partial slabs on a node.
3197 * This is called from kmem_cache_close(). We must be the last thread
3198 * using the cache and therefore we do not need to lock anymore.
3199 */
3200static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3201{
3202 struct page *page, *h;
3203
3204 list_for_each_entry_safe(page, h, &n->partial, lru) {
3205 if (!page->inuse) {
3206 __remove_partial(n, page);
3207 discard_slab(s, page);
3208 } else {
3209 list_slab_objects(s, page,
3210 "Objects remaining in %s on kmem_cache_close()");
3211 }
3212 }
3213}
3214
3215/*
3216 * Release all resources used by a slab cache.
3217 */
3218static inline int kmem_cache_close(struct kmem_cache *s)
3219{
3220 int node;
3221
3222 flush_all(s);
3223 /* Attempt to free all objects */
3224 for_each_node_state(node, N_NORMAL_MEMORY) {
3225 struct kmem_cache_node *n = get_node(s, node);
3226
3227 free_partial(s, n);
3228 if (n->nr_partial || slabs_node(s, node))
3229 return 1;
3230 }
3231 free_percpu(s->cpu_slab);
3232 free_kmem_cache_nodes(s);
3233 return 0;
3234}
3235
3236int __kmem_cache_shutdown(struct kmem_cache *s)
3237{
3238 return kmem_cache_close(s);
3239}
3240
3241/********************************************************************
3242 * Kmalloc subsystem
3243 *******************************************************************/
3244
3245static int __init setup_slub_min_order(char *str)
3246{
3247 get_option(&str, &slub_min_order);
3248
3249 return 1;
3250}
3251
3252__setup("slub_min_order=", setup_slub_min_order);
3253
3254static int __init setup_slub_max_order(char *str)
3255{
3256 get_option(&str, &slub_max_order);
3257 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3258
3259 return 1;
3260}
3261
3262__setup("slub_max_order=", setup_slub_max_order);
3263
3264static int __init setup_slub_min_objects(char *str)
3265{
3266 get_option(&str, &slub_min_objects);
3267
3268 return 1;
3269}
3270
3271__setup("slub_min_objects=", setup_slub_min_objects);
3272
3273static int __init setup_slub_nomerge(char *str)
3274{
3275 slub_nomerge = 1;
3276 return 1;
3277}
3278
3279__setup("slub_nomerge", setup_slub_nomerge);
3280
3281void *__kmalloc(size_t size, gfp_t flags)
3282{
3283 struct kmem_cache *s;
3284 void *ret;
3285
3286 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3287 return kmalloc_large(size, flags);
3288
3289 s = kmalloc_slab(size, flags);
3290
3291 if (unlikely(ZERO_OR_NULL_PTR(s)))
3292 return s;
3293
3294 ret = slab_alloc(s, flags, _RET_IP_);
3295
3296 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3297
3298 return ret;
3299}
3300EXPORT_SYMBOL(__kmalloc);
3301
3302#ifdef CONFIG_NUMA
3303static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3304{
3305 struct page *page;
3306 void *ptr = NULL;
3307
3308 flags |= __GFP_COMP | __GFP_NOTRACK | __GFP_KMEMCG;
3309 page = alloc_pages_node(node, flags, get_order(size));
3310 if (page)
3311 ptr = page_address(page);
3312
3313 kmalloc_large_node_hook(ptr, size, flags);
3314 return ptr;
3315}
3316
3317void *__kmalloc_node(size_t size, gfp_t flags, int node)
3318{
3319 struct kmem_cache *s;
3320 void *ret;
3321
3322 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3323 ret = kmalloc_large_node(size, flags, node);
3324
3325 trace_kmalloc_node(_RET_IP_, ret,
3326 size, PAGE_SIZE << get_order(size),
3327 flags, node);
3328
3329 return ret;
3330 }
3331
3332 s = kmalloc_slab(size, flags);
3333
3334 if (unlikely(ZERO_OR_NULL_PTR(s)))
3335 return s;
3336
3337 ret = slab_alloc_node(s, flags, node, _RET_IP_);
3338
3339 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3340
3341 return ret;
3342}
3343EXPORT_SYMBOL(__kmalloc_node);
3344#endif
3345
3346size_t ksize(const void *object)
3347{
3348 struct page *page;
3349
3350 if (unlikely(object == ZERO_SIZE_PTR))
3351 return 0;
3352
3353 page = virt_to_head_page(object);
3354
3355 if (unlikely(!PageSlab(page))) {
3356 WARN_ON(!PageCompound(page));
3357 return PAGE_SIZE << compound_order(page);
3358 }
3359
3360 return slab_ksize(page->slab_cache);
3361}
3362EXPORT_SYMBOL(ksize);
3363
3364void kfree(const void *x)
3365{
3366 struct page *page;
3367 void *object = (void *)x;
3368
3369 trace_kfree(_RET_IP_, x);
3370
3371 if (unlikely(ZERO_OR_NULL_PTR(x)))
3372 return;
3373
3374 page = virt_to_head_page(x);
3375 if (unlikely(!PageSlab(page))) {
3376 BUG_ON(!PageCompound(page));
3377 kfree_hook(x);
3378 __free_memcg_kmem_pages(page, compound_order(page));
3379 return;
3380 }
3381 slab_free(page->slab_cache, page, object, _RET_IP_);
3382}
3383EXPORT_SYMBOL(kfree);
3384
3385/*
3386 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3387 * the remaining slabs by the number of items in use. The slabs with the
3388 * most items in use come first. New allocations will then fill those up
3389 * and thus they can be removed from the partial lists.
3390 *
3391 * The slabs with the least items are placed last. This results in them
3392 * being allocated from last increasing the chance that the last objects
3393 * are freed in them.
3394 */
3395int kmem_cache_shrink(struct kmem_cache *s)
3396{
3397 int node;
3398 int i;
3399 struct kmem_cache_node *n;
3400 struct page *page;
3401 struct page *t;
3402 int objects = oo_objects(s->max);
3403 struct list_head *slabs_by_inuse =
3404 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3405 unsigned long flags;
3406
3407 if (!slabs_by_inuse)
3408 return -ENOMEM;
3409
3410 flush_all(s);
3411 for_each_node_state(node, N_NORMAL_MEMORY) {
3412 n = get_node(s, node);
3413
3414 if (!n->nr_partial)
3415 continue;
3416
3417 for (i = 0; i < objects; i++)
3418 INIT_LIST_HEAD(slabs_by_inuse + i);
3419
3420 spin_lock_irqsave(&n->list_lock, flags);
3421
3422 /*
3423 * Build lists indexed by the items in use in each slab.
3424 *
3425 * Note that concurrent frees may occur while we hold the
3426 * list_lock. page->inuse here is the upper limit.
3427 */
3428 list_for_each_entry_safe(page, t, &n->partial, lru) {
3429 list_move(&page->lru, slabs_by_inuse + page->inuse);
3430 if (!page->inuse)
3431 n->nr_partial--;
3432 }
3433
3434 /*
3435 * Rebuild the partial list with the slabs filled up most
3436 * first and the least used slabs at the end.
3437 */
3438 for (i = objects - 1; i > 0; i--)
3439 list_splice(slabs_by_inuse + i, n->partial.prev);
3440
3441 spin_unlock_irqrestore(&n->list_lock, flags);
3442
3443 /* Release empty slabs */
3444 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3445 discard_slab(s, page);
3446 }
3447
3448 kfree(slabs_by_inuse);
3449 return 0;
3450}
3451EXPORT_SYMBOL(kmem_cache_shrink);
3452
3453static int slab_mem_going_offline_callback(void *arg)
3454{
3455 struct kmem_cache *s;
3456
3457 mutex_lock(&slab_mutex);
3458 list_for_each_entry(s, &slab_caches, list)
3459 kmem_cache_shrink(s);
3460 mutex_unlock(&slab_mutex);
3461
3462 return 0;
3463}
3464
3465static void slab_mem_offline_callback(void *arg)
3466{
3467 struct kmem_cache_node *n;
3468 struct kmem_cache *s;
3469 struct memory_notify *marg = arg;
3470 int offline_node;
3471
3472 offline_node = marg->status_change_nid_normal;
3473
3474 /*
3475 * If the node still has available memory. we need kmem_cache_node
3476 * for it yet.
3477 */
3478 if (offline_node < 0)
3479 return;
3480
3481 mutex_lock(&slab_mutex);
3482 list_for_each_entry(s, &slab_caches, list) {
3483 n = get_node(s, offline_node);
3484 if (n) {
3485 /*
3486 * if n->nr_slabs > 0, slabs still exist on the node
3487 * that is going down. We were unable to free them,
3488 * and offline_pages() function shouldn't call this
3489 * callback. So, we must fail.
3490 */
3491 BUG_ON(slabs_node(s, offline_node));
3492
3493 s->node[offline_node] = NULL;
3494 kmem_cache_free(kmem_cache_node, n);
3495 }
3496 }
3497 mutex_unlock(&slab_mutex);
3498}
3499
3500static int slab_mem_going_online_callback(void *arg)
3501{
3502 struct kmem_cache_node *n;
3503 struct kmem_cache *s;
3504 struct memory_notify *marg = arg;
3505 int nid = marg->status_change_nid_normal;
3506 int ret = 0;
3507
3508 /*
3509 * If the node's memory is already available, then kmem_cache_node is
3510 * already created. Nothing to do.
3511 */
3512 if (nid < 0)
3513 return 0;
3514
3515 /*
3516 * We are bringing a node online. No memory is available yet. We must
3517 * allocate a kmem_cache_node structure in order to bring the node
3518 * online.
3519 */
3520 mutex_lock(&slab_mutex);
3521 list_for_each_entry(s, &slab_caches, list) {
3522 /*
3523 * XXX: kmem_cache_alloc_node will fallback to other nodes
3524 * since memory is not yet available from the node that
3525 * is brought up.
3526 */
3527 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3528 if (!n) {
3529 ret = -ENOMEM;
3530 goto out;
3531 }
3532 init_kmem_cache_node(n);
3533 s->node[nid] = n;
3534 }
3535out:
3536 mutex_unlock(&slab_mutex);
3537 return ret;
3538}
3539
3540static int slab_memory_callback(struct notifier_block *self,
3541 unsigned long action, void *arg)
3542{
3543 int ret = 0;
3544
3545 switch (action) {
3546 case MEM_GOING_ONLINE:
3547 ret = slab_mem_going_online_callback(arg);
3548 break;
3549 case MEM_GOING_OFFLINE:
3550 ret = slab_mem_going_offline_callback(arg);
3551 break;
3552 case MEM_OFFLINE:
3553 case MEM_CANCEL_ONLINE:
3554 slab_mem_offline_callback(arg);
3555 break;
3556 case MEM_ONLINE:
3557 case MEM_CANCEL_OFFLINE:
3558 break;
3559 }
3560 if (ret)
3561 ret = notifier_from_errno(ret);
3562 else
3563 ret = NOTIFY_OK;
3564 return ret;
3565}
3566
3567static struct notifier_block slab_memory_callback_nb = {
3568 .notifier_call = slab_memory_callback,
3569 .priority = SLAB_CALLBACK_PRI,
3570};
3571
3572/********************************************************************
3573 * Basic setup of slabs
3574 *******************************************************************/
3575
3576/*
3577 * Used for early kmem_cache structures that were allocated using
3578 * the page allocator. Allocate them properly then fix up the pointers
3579 * that may be pointing to the wrong kmem_cache structure.
3580 */
3581
3582static struct kmem_cache * __init bootstrap(struct kmem_cache *static_cache)
3583{
3584 int node;
3585 struct kmem_cache *s = kmem_cache_zalloc(kmem_cache, GFP_NOWAIT);
3586
3587 memcpy(s, static_cache, kmem_cache->object_size);
3588
3589 /*
3590 * This runs very early, and only the boot processor is supposed to be
3591 * up. Even if it weren't true, IRQs are not up so we couldn't fire
3592 * IPIs around.
3593 */
3594 __flush_cpu_slab(s, smp_processor_id());
3595 for_each_node_state(node, N_NORMAL_MEMORY) {
3596 struct kmem_cache_node *n = get_node(s, node);
3597 struct page *p;
3598
3599 if (n) {
3600 list_for_each_entry(p, &n->partial, lru)
3601 p->slab_cache = s;
3602
3603#ifdef CONFIG_SLUB_DEBUG
3604 list_for_each_entry(p, &n->full, lru)
3605 p->slab_cache = s;
3606#endif
3607 }
3608 }
3609 list_add(&s->list, &slab_caches);
3610 return s;
3611}
3612
3613void __init kmem_cache_init(void)
3614{
3615 static __initdata struct kmem_cache boot_kmem_cache,
3616 boot_kmem_cache_node;
3617
3618 if (debug_guardpage_minorder())
3619 slub_max_order = 0;
3620
3621 kmem_cache_node = &boot_kmem_cache_node;
3622 kmem_cache = &boot_kmem_cache;
3623
3624 create_boot_cache(kmem_cache_node, "kmem_cache_node",
3625 sizeof(struct kmem_cache_node), SLAB_HWCACHE_ALIGN);
3626
3627 register_hotmemory_notifier(&slab_memory_callback_nb);
3628
3629 /* Able to allocate the per node structures */
3630 slab_state = PARTIAL;
3631
3632 create_boot_cache(kmem_cache, "kmem_cache",
3633 offsetof(struct kmem_cache, node) +
3634 nr_node_ids * sizeof(struct kmem_cache_node *),
3635 SLAB_HWCACHE_ALIGN);
3636
3637 kmem_cache = bootstrap(&boot_kmem_cache);
3638
3639 /*
3640 * Allocate kmem_cache_node properly from the kmem_cache slab.
3641 * kmem_cache_node is separately allocated so no need to
3642 * update any list pointers.
3643 */
3644 kmem_cache_node = bootstrap(&boot_kmem_cache_node);
3645
3646 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3647 create_kmalloc_caches(0);
3648
3649#ifdef CONFIG_SMP
3650 register_cpu_notifier(&slab_notifier);
3651#endif
3652
3653 printk(KERN_INFO
3654 "SLUB: HWalign=%d, Order=%d-%d, MinObjects=%d,"
3655 " CPUs=%d, Nodes=%d\n",
3656 cache_line_size(),
3657 slub_min_order, slub_max_order, slub_min_objects,
3658 nr_cpu_ids, nr_node_ids);
3659}
3660
3661void __init kmem_cache_init_late(void)
3662{
3663}
3664
3665/*
3666 * Find a mergeable slab cache
3667 */
3668static int slab_unmergeable(struct kmem_cache *s)
3669{
3670 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3671 return 1;
3672
3673 if (!is_root_cache(s))
3674 return 1;
3675
3676 if (s->ctor)
3677 return 1;
3678
3679 /*
3680 * We may have set a slab to be unmergeable during bootstrap.
3681 */
3682 if (s->refcount < 0)
3683 return 1;
3684
3685 return 0;
3686}
3687
3688static struct kmem_cache *find_mergeable(size_t size, size_t align,
3689 unsigned long flags, const char *name, void (*ctor)(void *))
3690{
3691 struct kmem_cache *s;
3692
3693 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3694 return NULL;
3695
3696 if (ctor)
3697 return NULL;
3698
3699 size = ALIGN(size, sizeof(void *));
3700 align = calculate_alignment(flags, align, size);
3701 size = ALIGN(size, align);
3702 flags = kmem_cache_flags(size, flags, name, NULL);
3703
3704 list_for_each_entry(s, &slab_caches, list) {
3705 if (slab_unmergeable(s))
3706 continue;
3707
3708 if (size > s->size)
3709 continue;
3710
3711 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3712 continue;
3713 /*
3714 * Check if alignment is compatible.
3715 * Courtesy of Adrian Drzewiecki
3716 */
3717 if ((s->size & ~(align - 1)) != s->size)
3718 continue;
3719
3720 if (s->size - size >= sizeof(void *))
3721 continue;
3722
3723 return s;
3724 }
3725 return NULL;
3726}
3727
3728struct kmem_cache *
3729__kmem_cache_alias(const char *name, size_t size, size_t align,
3730 unsigned long flags, void (*ctor)(void *))
3731{
3732 struct kmem_cache *s;
3733
3734 s = find_mergeable(size, align, flags, name, ctor);
3735 if (s) {
3736 int i;
3737 struct kmem_cache *c;
3738
3739 s->refcount++;
3740
3741 /*
3742 * Adjust the object sizes so that we clear
3743 * the complete object on kzalloc.
3744 */
3745 s->object_size = max(s->object_size, (int)size);
3746 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3747
3748 for_each_memcg_cache_index(i) {
3749 c = cache_from_memcg_idx(s, i);
3750 if (!c)
3751 continue;
3752 c->object_size = s->object_size;
3753 c->inuse = max_t(int, c->inuse,
3754 ALIGN(size, sizeof(void *)));
3755 }
3756
3757 if (sysfs_slab_alias(s, name)) {
3758 s->refcount--;
3759 s = NULL;
3760 }
3761 }
3762
3763 return s;
3764}
3765
3766int __kmem_cache_create(struct kmem_cache *s, unsigned long flags)
3767{
3768 int err;
3769
3770 err = kmem_cache_open(s, flags);
3771 if (err)
3772 return err;
3773
3774 /* Mutex is not taken during early boot */
3775 if (slab_state <= UP)
3776 return 0;
3777
3778 memcg_propagate_slab_attrs(s);
3779 err = sysfs_slab_add(s);
3780 if (err)
3781 kmem_cache_close(s);
3782
3783 return err;
3784}
3785
3786#ifdef CONFIG_SMP
3787/*
3788 * Use the cpu notifier to insure that the cpu slabs are flushed when
3789 * necessary.
3790 */
3791static int slab_cpuup_callback(struct notifier_block *nfb,
3792 unsigned long action, void *hcpu)
3793{
3794 long cpu = (long)hcpu;
3795 struct kmem_cache *s;
3796 unsigned long flags;
3797
3798 switch (action) {
3799 case CPU_UP_CANCELED:
3800 case CPU_UP_CANCELED_FROZEN:
3801 case CPU_DEAD:
3802 case CPU_DEAD_FROZEN:
3803 mutex_lock(&slab_mutex);
3804 list_for_each_entry(s, &slab_caches, list) {
3805 local_irq_save(flags);
3806 __flush_cpu_slab(s, cpu);
3807 local_irq_restore(flags);
3808 }
3809 mutex_unlock(&slab_mutex);
3810 break;
3811 default:
3812 break;
3813 }
3814 return NOTIFY_OK;
3815}
3816
3817static struct notifier_block slab_notifier = {
3818 .notifier_call = slab_cpuup_callback
3819};
3820
3821#endif
3822
3823void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3824{
3825 struct kmem_cache *s;
3826 void *ret;
3827
3828 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE))
3829 return kmalloc_large(size, gfpflags);
3830
3831 s = kmalloc_slab(size, gfpflags);
3832
3833 if (unlikely(ZERO_OR_NULL_PTR(s)))
3834 return s;
3835
3836 ret = slab_alloc(s, gfpflags, caller);
3837
3838 /* Honor the call site pointer we received. */
3839 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3840
3841 return ret;
3842}
3843
3844#ifdef CONFIG_NUMA
3845void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3846 int node, unsigned long caller)
3847{
3848 struct kmem_cache *s;
3849 void *ret;
3850
3851 if (unlikely(size > KMALLOC_MAX_CACHE_SIZE)) {
3852 ret = kmalloc_large_node(size, gfpflags, node);
3853
3854 trace_kmalloc_node(caller, ret,
3855 size, PAGE_SIZE << get_order(size),
3856 gfpflags, node);
3857
3858 return ret;
3859 }
3860
3861 s = kmalloc_slab(size, gfpflags);
3862
3863 if (unlikely(ZERO_OR_NULL_PTR(s)))
3864 return s;
3865
3866 ret = slab_alloc_node(s, gfpflags, node, caller);
3867
3868 /* Honor the call site pointer we received. */
3869 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3870
3871 return ret;
3872}
3873#endif
3874
3875#ifdef CONFIG_SYSFS
3876static int count_inuse(struct page *page)
3877{
3878 return page->inuse;
3879}
3880
3881static int count_total(struct page *page)
3882{
3883 return page->objects;
3884}
3885#endif
3886
3887#ifdef CONFIG_SLUB_DEBUG
3888static int validate_slab(struct kmem_cache *s, struct page *page,
3889 unsigned long *map)
3890{
3891 void *p;
3892 void *addr = page_address(page);
3893
3894 if (!check_slab(s, page) ||
3895 !on_freelist(s, page, NULL))
3896 return 0;
3897
3898 /* Now we know that a valid freelist exists */
3899 bitmap_zero(map, page->objects);
3900
3901 get_map(s, page, map);
3902 for_each_object(p, s, addr, page->objects) {
3903 if (test_bit(slab_index(p, s, addr), map))
3904 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3905 return 0;
3906 }
3907
3908 for_each_object(p, s, addr, page->objects)
3909 if (!test_bit(slab_index(p, s, addr), map))
3910 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3911 return 0;
3912 return 1;
3913}
3914
3915static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3916 unsigned long *map)
3917{
3918 slab_lock(page);
3919 validate_slab(s, page, map);
3920 slab_unlock(page);
3921}
3922
3923static int validate_slab_node(struct kmem_cache *s,
3924 struct kmem_cache_node *n, unsigned long *map)
3925{
3926 unsigned long count = 0;
3927 struct page *page;
3928 unsigned long flags;
3929
3930 spin_lock_irqsave(&n->list_lock, flags);
3931
3932 list_for_each_entry(page, &n->partial, lru) {
3933 validate_slab_slab(s, page, map);
3934 count++;
3935 }
3936 if (count != n->nr_partial)
3937 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3938 "counter=%ld\n", s->name, count, n->nr_partial);
3939
3940 if (!(s->flags & SLAB_STORE_USER))
3941 goto out;
3942
3943 list_for_each_entry(page, &n->full, lru) {
3944 validate_slab_slab(s, page, map);
3945 count++;
3946 }
3947 if (count != atomic_long_read(&n->nr_slabs))
3948 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3949 "counter=%ld\n", s->name, count,
3950 atomic_long_read(&n->nr_slabs));
3951
3952out:
3953 spin_unlock_irqrestore(&n->list_lock, flags);
3954 return count;
3955}
3956
3957static long validate_slab_cache(struct kmem_cache *s)
3958{
3959 int node;
3960 unsigned long count = 0;
3961 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3962 sizeof(unsigned long), GFP_KERNEL);
3963
3964 if (!map)
3965 return -ENOMEM;
3966
3967 flush_all(s);
3968 for_each_node_state(node, N_NORMAL_MEMORY) {
3969 struct kmem_cache_node *n = get_node(s, node);
3970
3971 count += validate_slab_node(s, n, map);
3972 }
3973 kfree(map);
3974 return count;
3975}
3976/*
3977 * Generate lists of code addresses where slabcache objects are allocated
3978 * and freed.
3979 */
3980
3981struct location {
3982 unsigned long count;
3983 unsigned long addr;
3984 long long sum_time;
3985 long min_time;
3986 long max_time;
3987 long min_pid;
3988 long max_pid;
3989 DECLARE_BITMAP(cpus, NR_CPUS);
3990 nodemask_t nodes;
3991};
3992
3993struct loc_track {
3994 unsigned long max;
3995 unsigned long count;
3996 struct location *loc;
3997};
3998
3999static void free_loc_track(struct loc_track *t)
4000{
4001 if (t->max)
4002 free_pages((unsigned long)t->loc,
4003 get_order(sizeof(struct location) * t->max));
4004}
4005
4006static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4007{
4008 struct location *l;
4009 int order;
4010
4011 order = get_order(sizeof(struct location) * max);
4012
4013 l = (void *)__get_free_pages(flags, order);
4014 if (!l)
4015 return 0;
4016
4017 if (t->count) {
4018 memcpy(l, t->loc, sizeof(struct location) * t->count);
4019 free_loc_track(t);
4020 }
4021 t->max = max;
4022 t->loc = l;
4023 return 1;
4024}
4025
4026static int add_location(struct loc_track *t, struct kmem_cache *s,
4027 const struct track *track)
4028{
4029 long start, end, pos;
4030 struct location *l;
4031 unsigned long caddr;
4032 unsigned long age = jiffies - track->when;
4033
4034 start = -1;
4035 end = t->count;
4036
4037 for ( ; ; ) {
4038 pos = start + (end - start + 1) / 2;
4039
4040 /*
4041 * There is nothing at "end". If we end up there
4042 * we need to add something to before end.
4043 */
4044 if (pos == end)
4045 break;
4046
4047 caddr = t->loc[pos].addr;
4048 if (track->addr == caddr) {
4049
4050 l = &t->loc[pos];
4051 l->count++;
4052 if (track->when) {
4053 l->sum_time += age;
4054 if (age < l->min_time)
4055 l->min_time = age;
4056 if (age > l->max_time)
4057 l->max_time = age;
4058
4059 if (track->pid < l->min_pid)
4060 l->min_pid = track->pid;
4061 if (track->pid > l->max_pid)
4062 l->max_pid = track->pid;
4063
4064 cpumask_set_cpu(track->cpu,
4065 to_cpumask(l->cpus));
4066 }
4067 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4068 return 1;
4069 }
4070
4071 if (track->addr < caddr)
4072 end = pos;
4073 else
4074 start = pos;
4075 }
4076
4077 /*
4078 * Not found. Insert new tracking element.
4079 */
4080 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4081 return 0;
4082
4083 l = t->loc + pos;
4084 if (pos < t->count)
4085 memmove(l + 1, l,
4086 (t->count - pos) * sizeof(struct location));
4087 t->count++;
4088 l->count = 1;
4089 l->addr = track->addr;
4090 l->sum_time = age;
4091 l->min_time = age;
4092 l->max_time = age;
4093 l->min_pid = track->pid;
4094 l->max_pid = track->pid;
4095 cpumask_clear(to_cpumask(l->cpus));
4096 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4097 nodes_clear(l->nodes);
4098 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4099 return 1;
4100}
4101
4102static void process_slab(struct loc_track *t, struct kmem_cache *s,
4103 struct page *page, enum track_item alloc,
4104 unsigned long *map)
4105{
4106 void *addr = page_address(page);
4107 void *p;
4108
4109 bitmap_zero(map, page->objects);
4110 get_map(s, page, map);
4111
4112 for_each_object(p, s, addr, page->objects)
4113 if (!test_bit(slab_index(p, s, addr), map))
4114 add_location(t, s, get_track(s, p, alloc));
4115}
4116
4117static int list_locations(struct kmem_cache *s, char *buf,
4118 enum track_item alloc)
4119{
4120 int len = 0;
4121 unsigned long i;
4122 struct loc_track t = { 0, 0, NULL };
4123 int node;
4124 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4125 sizeof(unsigned long), GFP_KERNEL);
4126
4127 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4128 GFP_TEMPORARY)) {
4129 kfree(map);
4130 return sprintf(buf, "Out of memory\n");
4131 }
4132 /* Push back cpu slabs */
4133 flush_all(s);
4134
4135 for_each_node_state(node, N_NORMAL_MEMORY) {
4136 struct kmem_cache_node *n = get_node(s, node);
4137 unsigned long flags;
4138 struct page *page;
4139
4140 if (!atomic_long_read(&n->nr_slabs))
4141 continue;
4142
4143 spin_lock_irqsave(&n->list_lock, flags);
4144 list_for_each_entry(page, &n->partial, lru)
4145 process_slab(&t, s, page, alloc, map);
4146 list_for_each_entry(page, &n->full, lru)
4147 process_slab(&t, s, page, alloc, map);
4148 spin_unlock_irqrestore(&n->list_lock, flags);
4149 }
4150
4151 for (i = 0; i < t.count; i++) {
4152 struct location *l = &t.loc[i];
4153
4154 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4155 break;
4156 len += sprintf(buf + len, "%7ld ", l->count);
4157
4158 if (l->addr)
4159 len += sprintf(buf + len, "%pS", (void *)l->addr);
4160 else
4161 len += sprintf(buf + len, "<not-available>");
4162
4163 if (l->sum_time != l->min_time) {
4164 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4165 l->min_time,
4166 (long)div_u64(l->sum_time, l->count),
4167 l->max_time);
4168 } else
4169 len += sprintf(buf + len, " age=%ld",
4170 l->min_time);
4171
4172 if (l->min_pid != l->max_pid)
4173 len += sprintf(buf + len, " pid=%ld-%ld",
4174 l->min_pid, l->max_pid);
4175 else
4176 len += sprintf(buf + len, " pid=%ld",
4177 l->min_pid);
4178
4179 if (num_online_cpus() > 1 &&
4180 !cpumask_empty(to_cpumask(l->cpus)) &&
4181 len < PAGE_SIZE - 60) {
4182 len += sprintf(buf + len, " cpus=");
4183 len += cpulist_scnprintf(buf + len,
4184 PAGE_SIZE - len - 50,
4185 to_cpumask(l->cpus));
4186 }
4187
4188 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4189 len < PAGE_SIZE - 60) {
4190 len += sprintf(buf + len, " nodes=");
4191 len += nodelist_scnprintf(buf + len,
4192 PAGE_SIZE - len - 50,
4193 l->nodes);
4194 }
4195
4196 len += sprintf(buf + len, "\n");
4197 }
4198
4199 free_loc_track(&t);
4200 kfree(map);
4201 if (!t.count)
4202 len += sprintf(buf, "No data\n");
4203 return len;
4204}
4205#endif
4206
4207#ifdef SLUB_RESILIENCY_TEST
4208static void resiliency_test(void)
4209{
4210 u8 *p;
4211
4212 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || KMALLOC_SHIFT_HIGH < 10);
4213
4214 printk(KERN_ERR "SLUB resiliency testing\n");
4215 printk(KERN_ERR "-----------------------\n");
4216 printk(KERN_ERR "A. Corruption after allocation\n");
4217
4218 p = kzalloc(16, GFP_KERNEL);
4219 p[16] = 0x12;
4220 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4221 " 0x12->0x%p\n\n", p + 16);
4222
4223 validate_slab_cache(kmalloc_caches[4]);
4224
4225 /* Hmmm... The next two are dangerous */
4226 p = kzalloc(32, GFP_KERNEL);
4227 p[32 + sizeof(void *)] = 0x34;
4228 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4229 " 0x34 -> -0x%p\n", p);
4230 printk(KERN_ERR
4231 "If allocated object is overwritten then not detectable\n\n");
4232
4233 validate_slab_cache(kmalloc_caches[5]);
4234 p = kzalloc(64, GFP_KERNEL);
4235 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4236 *p = 0x56;
4237 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4238 p);
4239 printk(KERN_ERR
4240 "If allocated object is overwritten then not detectable\n\n");
4241 validate_slab_cache(kmalloc_caches[6]);
4242
4243 printk(KERN_ERR "\nB. Corruption after free\n");
4244 p = kzalloc(128, GFP_KERNEL);
4245 kfree(p);
4246 *p = 0x78;
4247 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4248 validate_slab_cache(kmalloc_caches[7]);
4249
4250 p = kzalloc(256, GFP_KERNEL);
4251 kfree(p);
4252 p[50] = 0x9a;
4253 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4254 p);
4255 validate_slab_cache(kmalloc_caches[8]);
4256
4257 p = kzalloc(512, GFP_KERNEL);
4258 kfree(p);
4259 p[512] = 0xab;
4260 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4261 validate_slab_cache(kmalloc_caches[9]);
4262}
4263#else
4264#ifdef CONFIG_SYSFS
4265static void resiliency_test(void) {};
4266#endif
4267#endif
4268
4269#ifdef CONFIG_SYSFS
4270enum slab_stat_type {
4271 SL_ALL, /* All slabs */
4272 SL_PARTIAL, /* Only partially allocated slabs */
4273 SL_CPU, /* Only slabs used for cpu caches */
4274 SL_OBJECTS, /* Determine allocated objects not slabs */
4275 SL_TOTAL /* Determine object capacity not slabs */
4276};
4277
4278#define SO_ALL (1 << SL_ALL)
4279#define SO_PARTIAL (1 << SL_PARTIAL)
4280#define SO_CPU (1 << SL_CPU)
4281#define SO_OBJECTS (1 << SL_OBJECTS)
4282#define SO_TOTAL (1 << SL_TOTAL)
4283
4284static ssize_t show_slab_objects(struct kmem_cache *s,
4285 char *buf, unsigned long flags)
4286{
4287 unsigned long total = 0;
4288 int node;
4289 int x;
4290 unsigned long *nodes;
4291
4292 nodes = kzalloc(sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4293 if (!nodes)
4294 return -ENOMEM;
4295
4296 if (flags & SO_CPU) {
4297 int cpu;
4298
4299 for_each_possible_cpu(cpu) {
4300 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab,
4301 cpu);
4302 int node;
4303 struct page *page;
4304
4305 page = ACCESS_ONCE(c->page);
4306 if (!page)
4307 continue;
4308
4309 node = page_to_nid(page);
4310 if (flags & SO_TOTAL)
4311 x = page->objects;
4312 else if (flags & SO_OBJECTS)
4313 x = page->inuse;
4314 else
4315 x = 1;
4316
4317 total += x;
4318 nodes[node] += x;
4319
4320 page = ACCESS_ONCE(c->partial);
4321 if (page) {
4322 node = page_to_nid(page);
4323 if (flags & SO_TOTAL)
4324 WARN_ON_ONCE(1);
4325 else if (flags & SO_OBJECTS)
4326 WARN_ON_ONCE(1);
4327 else
4328 x = page->pages;
4329 total += x;
4330 nodes[node] += x;
4331 }
4332 }
4333 }
4334
4335 lock_memory_hotplug();
4336#ifdef CONFIG_SLUB_DEBUG
4337 if (flags & SO_ALL) {
4338 for_each_node_state(node, N_NORMAL_MEMORY) {
4339 struct kmem_cache_node *n = get_node(s, node);
4340
4341 if (flags & SO_TOTAL)
4342 x = atomic_long_read(&n->total_objects);
4343 else if (flags & SO_OBJECTS)
4344 x = atomic_long_read(&n->total_objects) -
4345 count_partial(n, count_free);
4346 else
4347 x = atomic_long_read(&n->nr_slabs);
4348 total += x;
4349 nodes[node] += x;
4350 }
4351
4352 } else
4353#endif
4354 if (flags & SO_PARTIAL) {
4355 for_each_node_state(node, N_NORMAL_MEMORY) {
4356 struct kmem_cache_node *n = get_node(s, node);
4357
4358 if (flags & SO_TOTAL)
4359 x = count_partial(n, count_total);
4360 else if (flags & SO_OBJECTS)
4361 x = count_partial(n, count_inuse);
4362 else
4363 x = n->nr_partial;
4364 total += x;
4365 nodes[node] += x;
4366 }
4367 }
4368 x = sprintf(buf, "%lu", total);
4369#ifdef CONFIG_NUMA
4370 for_each_node_state(node, N_NORMAL_MEMORY)
4371 if (nodes[node])
4372 x += sprintf(buf + x, " N%d=%lu",
4373 node, nodes[node]);
4374#endif
4375 unlock_memory_hotplug();
4376 kfree(nodes);
4377 return x + sprintf(buf + x, "\n");
4378}
4379
4380#ifdef CONFIG_SLUB_DEBUG
4381static int any_slab_objects(struct kmem_cache *s)
4382{
4383 int node;
4384
4385 for_each_online_node(node) {
4386 struct kmem_cache_node *n = get_node(s, node);
4387
4388 if (!n)
4389 continue;
4390
4391 if (atomic_long_read(&n->total_objects))
4392 return 1;
4393 }
4394 return 0;
4395}
4396#endif
4397
4398#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4399#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4400
4401struct slab_attribute {
4402 struct attribute attr;
4403 ssize_t (*show)(struct kmem_cache *s, char *buf);
4404 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4405};
4406
4407#define SLAB_ATTR_RO(_name) \
4408 static struct slab_attribute _name##_attr = \
4409 __ATTR(_name, 0400, _name##_show, NULL)
4410
4411#define SLAB_ATTR(_name) \
4412 static struct slab_attribute _name##_attr = \
4413 __ATTR(_name, 0600, _name##_show, _name##_store)
4414
4415static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4416{
4417 return sprintf(buf, "%d\n", s->size);
4418}
4419SLAB_ATTR_RO(slab_size);
4420
4421static ssize_t align_show(struct kmem_cache *s, char *buf)
4422{
4423 return sprintf(buf, "%d\n", s->align);
4424}
4425SLAB_ATTR_RO(align);
4426
4427static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4428{
4429 return sprintf(buf, "%d\n", s->object_size);
4430}
4431SLAB_ATTR_RO(object_size);
4432
4433static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4434{
4435 return sprintf(buf, "%d\n", oo_objects(s->oo));
4436}
4437SLAB_ATTR_RO(objs_per_slab);
4438
4439static ssize_t order_store(struct kmem_cache *s,
4440 const char *buf, size_t length)
4441{
4442 unsigned long order;
4443 int err;
4444
4445 err = kstrtoul(buf, 10, &order);
4446 if (err)
4447 return err;
4448
4449 if (order > slub_max_order || order < slub_min_order)
4450 return -EINVAL;
4451
4452 calculate_sizes(s, order);
4453 return length;
4454}
4455
4456static ssize_t order_show(struct kmem_cache *s, char *buf)
4457{
4458 return sprintf(buf, "%d\n", oo_order(s->oo));
4459}
4460SLAB_ATTR(order);
4461
4462static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4463{
4464 return sprintf(buf, "%lu\n", s->min_partial);
4465}
4466
4467static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4468 size_t length)
4469{
4470 unsigned long min;
4471 int err;
4472
4473 err = kstrtoul(buf, 10, &min);
4474 if (err)
4475 return err;
4476
4477 set_min_partial(s, min);
4478 return length;
4479}
4480SLAB_ATTR(min_partial);
4481
4482static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4483{
4484 return sprintf(buf, "%u\n", s->cpu_partial);
4485}
4486
4487static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4488 size_t length)
4489{
4490 unsigned long objects;
4491 int err;
4492
4493 err = kstrtoul(buf, 10, &objects);
4494 if (err)
4495 return err;
4496 if (objects && !kmem_cache_has_cpu_partial(s))
4497 return -EINVAL;
4498
4499 s->cpu_partial = objects;
4500 flush_all(s);
4501 return length;
4502}
4503SLAB_ATTR(cpu_partial);
4504
4505static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4506{
4507 if (!s->ctor)
4508 return 0;
4509 return sprintf(buf, "%pS\n", s->ctor);
4510}
4511SLAB_ATTR_RO(ctor);
4512
4513static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4514{
4515 return sprintf(buf, "%d\n", s->refcount - 1);
4516}
4517SLAB_ATTR_RO(aliases);
4518
4519static ssize_t partial_show(struct kmem_cache *s, char *buf)
4520{
4521 return show_slab_objects(s, buf, SO_PARTIAL);
4522}
4523SLAB_ATTR_RO(partial);
4524
4525static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4526{
4527 return show_slab_objects(s, buf, SO_CPU);
4528}
4529SLAB_ATTR_RO(cpu_slabs);
4530
4531static ssize_t objects_show(struct kmem_cache *s, char *buf)
4532{
4533 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4534}
4535SLAB_ATTR_RO(objects);
4536
4537static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4538{
4539 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4540}
4541SLAB_ATTR_RO(objects_partial);
4542
4543static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4544{
4545 int objects = 0;
4546 int pages = 0;
4547 int cpu;
4548 int len;
4549
4550 for_each_online_cpu(cpu) {
4551 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4552
4553 if (page) {
4554 pages += page->pages;
4555 objects += page->pobjects;
4556 }
4557 }
4558
4559 len = sprintf(buf, "%d(%d)", objects, pages);
4560
4561#ifdef CONFIG_SMP
4562 for_each_online_cpu(cpu) {
4563 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4564
4565 if (page && len < PAGE_SIZE - 20)
4566 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4567 page->pobjects, page->pages);
4568 }
4569#endif
4570 return len + sprintf(buf + len, "\n");
4571}
4572SLAB_ATTR_RO(slabs_cpu_partial);
4573
4574static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4575{
4576 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4577}
4578
4579static ssize_t reclaim_account_store(struct kmem_cache *s,
4580 const char *buf, size_t length)
4581{
4582 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4583 if (buf[0] == '1')
4584 s->flags |= SLAB_RECLAIM_ACCOUNT;
4585 return length;
4586}
4587SLAB_ATTR(reclaim_account);
4588
4589static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4590{
4591 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4592}
4593SLAB_ATTR_RO(hwcache_align);
4594
4595#ifdef CONFIG_ZONE_DMA
4596static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4597{
4598 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4599}
4600SLAB_ATTR_RO(cache_dma);
4601#endif
4602
4603static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4604{
4605 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4606}
4607SLAB_ATTR_RO(destroy_by_rcu);
4608
4609static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4610{
4611 return sprintf(buf, "%d\n", s->reserved);
4612}
4613SLAB_ATTR_RO(reserved);
4614
4615#ifdef CONFIG_SLUB_DEBUG
4616static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4617{
4618 return show_slab_objects(s, buf, SO_ALL);
4619}
4620SLAB_ATTR_RO(slabs);
4621
4622static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4623{
4624 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4625}
4626SLAB_ATTR_RO(total_objects);
4627
4628static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4629{
4630 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4631}
4632
4633static ssize_t sanity_checks_store(struct kmem_cache *s,
4634 const char *buf, size_t length)
4635{
4636 s->flags &= ~SLAB_DEBUG_FREE;
4637 if (buf[0] == '1') {
4638 s->flags &= ~__CMPXCHG_DOUBLE;
4639 s->flags |= SLAB_DEBUG_FREE;
4640 }
4641 return length;
4642}
4643SLAB_ATTR(sanity_checks);
4644
4645static ssize_t trace_show(struct kmem_cache *s, char *buf)
4646{
4647 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4648}
4649
4650static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4651 size_t length)
4652{
4653 s->flags &= ~SLAB_TRACE;
4654 if (buf[0] == '1') {
4655 s->flags &= ~__CMPXCHG_DOUBLE;
4656 s->flags |= SLAB_TRACE;
4657 }
4658 return length;
4659}
4660SLAB_ATTR(trace);
4661
4662static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4663{
4664 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4665}
4666
4667static ssize_t red_zone_store(struct kmem_cache *s,
4668 const char *buf, size_t length)
4669{
4670 if (any_slab_objects(s))
4671 return -EBUSY;
4672
4673 s->flags &= ~SLAB_RED_ZONE;
4674 if (buf[0] == '1') {
4675 s->flags &= ~__CMPXCHG_DOUBLE;
4676 s->flags |= SLAB_RED_ZONE;
4677 }
4678 calculate_sizes(s, -1);
4679 return length;
4680}
4681SLAB_ATTR(red_zone);
4682
4683static ssize_t poison_show(struct kmem_cache *s, char *buf)
4684{
4685 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4686}
4687
4688static ssize_t poison_store(struct kmem_cache *s,
4689 const char *buf, size_t length)
4690{
4691 if (any_slab_objects(s))
4692 return -EBUSY;
4693
4694 s->flags &= ~SLAB_POISON;
4695 if (buf[0] == '1') {
4696 s->flags &= ~__CMPXCHG_DOUBLE;
4697 s->flags |= SLAB_POISON;
4698 }
4699 calculate_sizes(s, -1);
4700 return length;
4701}
4702SLAB_ATTR(poison);
4703
4704static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4705{
4706 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4707}
4708
4709static ssize_t store_user_store(struct kmem_cache *s,
4710 const char *buf, size_t length)
4711{
4712 if (any_slab_objects(s))
4713 return -EBUSY;
4714
4715 s->flags &= ~SLAB_STORE_USER;
4716 if (buf[0] == '1') {
4717 s->flags &= ~__CMPXCHG_DOUBLE;
4718 s->flags |= SLAB_STORE_USER;
4719 }
4720 calculate_sizes(s, -1);
4721 return length;
4722}
4723SLAB_ATTR(store_user);
4724
4725static ssize_t validate_show(struct kmem_cache *s, char *buf)
4726{
4727 return 0;
4728}
4729
4730static ssize_t validate_store(struct kmem_cache *s,
4731 const char *buf, size_t length)
4732{
4733 int ret = -EINVAL;
4734
4735 if (buf[0] == '1') {
4736 ret = validate_slab_cache(s);
4737 if (ret >= 0)
4738 ret = length;
4739 }
4740 return ret;
4741}
4742SLAB_ATTR(validate);
4743
4744static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4745{
4746 if (!(s->flags & SLAB_STORE_USER))
4747 return -ENOSYS;
4748 return list_locations(s, buf, TRACK_ALLOC);
4749}
4750SLAB_ATTR_RO(alloc_calls);
4751
4752static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4753{
4754 if (!(s->flags & SLAB_STORE_USER))
4755 return -ENOSYS;
4756 return list_locations(s, buf, TRACK_FREE);
4757}
4758SLAB_ATTR_RO(free_calls);
4759#endif /* CONFIG_SLUB_DEBUG */
4760
4761#ifdef CONFIG_FAILSLAB
4762static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4763{
4764 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4765}
4766
4767static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4768 size_t length)
4769{
4770 s->flags &= ~SLAB_FAILSLAB;
4771 if (buf[0] == '1')
4772 s->flags |= SLAB_FAILSLAB;
4773 return length;
4774}
4775SLAB_ATTR(failslab);
4776#endif
4777
4778static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4779{
4780 return 0;
4781}
4782
4783static ssize_t shrink_store(struct kmem_cache *s,
4784 const char *buf, size_t length)
4785{
4786 if (buf[0] == '1') {
4787 int rc = kmem_cache_shrink(s);
4788
4789 if (rc)
4790 return rc;
4791 } else
4792 return -EINVAL;
4793 return length;
4794}
4795SLAB_ATTR(shrink);
4796
4797#ifdef CONFIG_NUMA
4798static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4799{
4800 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4801}
4802
4803static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4804 const char *buf, size_t length)
4805{
4806 unsigned long ratio;
4807 int err;
4808
4809 err = kstrtoul(buf, 10, &ratio);
4810 if (err)
4811 return err;
4812
4813 if (ratio <= 100)
4814 s->remote_node_defrag_ratio = ratio * 10;
4815
4816 return length;
4817}
4818SLAB_ATTR(remote_node_defrag_ratio);
4819#endif
4820
4821#ifdef CONFIG_SLUB_STATS
4822static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4823{
4824 unsigned long sum = 0;
4825 int cpu;
4826 int len;
4827 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4828
4829 if (!data)
4830 return -ENOMEM;
4831
4832 for_each_online_cpu(cpu) {
4833 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4834
4835 data[cpu] = x;
4836 sum += x;
4837 }
4838
4839 len = sprintf(buf, "%lu", sum);
4840
4841#ifdef CONFIG_SMP
4842 for_each_online_cpu(cpu) {
4843 if (data[cpu] && len < PAGE_SIZE - 20)
4844 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4845 }
4846#endif
4847 kfree(data);
4848 return len + sprintf(buf + len, "\n");
4849}
4850
4851static void clear_stat(struct kmem_cache *s, enum stat_item si)
4852{
4853 int cpu;
4854
4855 for_each_online_cpu(cpu)
4856 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4857}
4858
4859#define STAT_ATTR(si, text) \
4860static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4861{ \
4862 return show_stat(s, buf, si); \
4863} \
4864static ssize_t text##_store(struct kmem_cache *s, \
4865 const char *buf, size_t length) \
4866{ \
4867 if (buf[0] != '0') \
4868 return -EINVAL; \
4869 clear_stat(s, si); \
4870 return length; \
4871} \
4872SLAB_ATTR(text); \
4873
4874STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4875STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4876STAT_ATTR(FREE_FASTPATH, free_fastpath);
4877STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4878STAT_ATTR(FREE_FROZEN, free_frozen);
4879STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4880STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4881STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4882STAT_ATTR(ALLOC_SLAB, alloc_slab);
4883STAT_ATTR(ALLOC_REFILL, alloc_refill);
4884STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4885STAT_ATTR(FREE_SLAB, free_slab);
4886STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4887STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4888STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4889STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4890STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4891STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4892STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4893STAT_ATTR(ORDER_FALLBACK, order_fallback);
4894STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4895STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4896STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
4897STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
4898STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
4899STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
4900#endif
4901
4902static struct attribute *slab_attrs[] = {
4903 &slab_size_attr.attr,
4904 &object_size_attr.attr,
4905 &objs_per_slab_attr.attr,
4906 &order_attr.attr,
4907 &min_partial_attr.attr,
4908 &cpu_partial_attr.attr,
4909 &objects_attr.attr,
4910 &objects_partial_attr.attr,
4911 &partial_attr.attr,
4912 &cpu_slabs_attr.attr,
4913 &ctor_attr.attr,
4914 &aliases_attr.attr,
4915 &align_attr.attr,
4916 &hwcache_align_attr.attr,
4917 &reclaim_account_attr.attr,
4918 &destroy_by_rcu_attr.attr,
4919 &shrink_attr.attr,
4920 &reserved_attr.attr,
4921 &slabs_cpu_partial_attr.attr,
4922#ifdef CONFIG_SLUB_DEBUG
4923 &total_objects_attr.attr,
4924 &slabs_attr.attr,
4925 &sanity_checks_attr.attr,
4926 &trace_attr.attr,
4927 &red_zone_attr.attr,
4928 &poison_attr.attr,
4929 &store_user_attr.attr,
4930 &validate_attr.attr,
4931 &alloc_calls_attr.attr,
4932 &free_calls_attr.attr,
4933#endif
4934#ifdef CONFIG_ZONE_DMA
4935 &cache_dma_attr.attr,
4936#endif
4937#ifdef CONFIG_NUMA
4938 &remote_node_defrag_ratio_attr.attr,
4939#endif
4940#ifdef CONFIG_SLUB_STATS
4941 &alloc_fastpath_attr.attr,
4942 &alloc_slowpath_attr.attr,
4943 &free_fastpath_attr.attr,
4944 &free_slowpath_attr.attr,
4945 &free_frozen_attr.attr,
4946 &free_add_partial_attr.attr,
4947 &free_remove_partial_attr.attr,
4948 &alloc_from_partial_attr.attr,
4949 &alloc_slab_attr.attr,
4950 &alloc_refill_attr.attr,
4951 &alloc_node_mismatch_attr.attr,
4952 &free_slab_attr.attr,
4953 &cpuslab_flush_attr.attr,
4954 &deactivate_full_attr.attr,
4955 &deactivate_empty_attr.attr,
4956 &deactivate_to_head_attr.attr,
4957 &deactivate_to_tail_attr.attr,
4958 &deactivate_remote_frees_attr.attr,
4959 &deactivate_bypass_attr.attr,
4960 &order_fallback_attr.attr,
4961 &cmpxchg_double_fail_attr.attr,
4962 &cmpxchg_double_cpu_fail_attr.attr,
4963 &cpu_partial_alloc_attr.attr,
4964 &cpu_partial_free_attr.attr,
4965 &cpu_partial_node_attr.attr,
4966 &cpu_partial_drain_attr.attr,
4967#endif
4968#ifdef CONFIG_FAILSLAB
4969 &failslab_attr.attr,
4970#endif
4971
4972 NULL
4973};
4974
4975static struct attribute_group slab_attr_group = {
4976 .attrs = slab_attrs,
4977};
4978
4979static ssize_t slab_attr_show(struct kobject *kobj,
4980 struct attribute *attr,
4981 char *buf)
4982{
4983 struct slab_attribute *attribute;
4984 struct kmem_cache *s;
4985 int err;
4986
4987 attribute = to_slab_attr(attr);
4988 s = to_slab(kobj);
4989
4990 if (!attribute->show)
4991 return -EIO;
4992
4993 err = attribute->show(s, buf);
4994
4995 return err;
4996}
4997
4998static ssize_t slab_attr_store(struct kobject *kobj,
4999 struct attribute *attr,
5000 const char *buf, size_t len)
5001{
5002 struct slab_attribute *attribute;
5003 struct kmem_cache *s;
5004 int err;
5005
5006 attribute = to_slab_attr(attr);
5007 s = to_slab(kobj);
5008
5009 if (!attribute->store)
5010 return -EIO;
5011
5012 err = attribute->store(s, buf, len);
5013#ifdef CONFIG_MEMCG_KMEM
5014 if (slab_state >= FULL && err >= 0 && is_root_cache(s)) {
5015 int i;
5016
5017 mutex_lock(&slab_mutex);
5018 if (s->max_attr_size < len)
5019 s->max_attr_size = len;
5020
5021 /*
5022 * This is a best effort propagation, so this function's return
5023 * value will be determined by the parent cache only. This is
5024 * basically because not all attributes will have a well
5025 * defined semantics for rollbacks - most of the actions will
5026 * have permanent effects.
5027 *
5028 * Returning the error value of any of the children that fail
5029 * is not 100 % defined, in the sense that users seeing the
5030 * error code won't be able to know anything about the state of
5031 * the cache.
5032 *
5033 * Only returning the error code for the parent cache at least
5034 * has well defined semantics. The cache being written to
5035 * directly either failed or succeeded, in which case we loop
5036 * through the descendants with best-effort propagation.
5037 */
5038 for_each_memcg_cache_index(i) {
5039 struct kmem_cache *c = cache_from_memcg_idx(s, i);
5040 if (c)
5041 attribute->store(c, buf, len);
5042 }
5043 mutex_unlock(&slab_mutex);
5044 }
5045#endif
5046 return err;
5047}
5048
5049static void memcg_propagate_slab_attrs(struct kmem_cache *s)
5050{
5051#ifdef CONFIG_MEMCG_KMEM
5052 int i;
5053 char *buffer = NULL;
5054 struct kmem_cache *root_cache;
5055
5056 if (is_root_cache(s))
5057 return;
5058
5059 root_cache = s->memcg_params->root_cache;
5060
5061 /*
5062 * This mean this cache had no attribute written. Therefore, no point
5063 * in copying default values around
5064 */
5065 if (!root_cache->max_attr_size)
5066 return;
5067
5068 for (i = 0; i < ARRAY_SIZE(slab_attrs); i++) {
5069 char mbuf[64];
5070 char *buf;
5071 struct slab_attribute *attr = to_slab_attr(slab_attrs[i]);
5072
5073 if (!attr || !attr->store || !attr->show)
5074 continue;
5075
5076 /*
5077 * It is really bad that we have to allocate here, so we will
5078 * do it only as a fallback. If we actually allocate, though,
5079 * we can just use the allocated buffer until the end.
5080 *
5081 * Most of the slub attributes will tend to be very small in
5082 * size, but sysfs allows buffers up to a page, so they can
5083 * theoretically happen.
5084 */
5085 if (buffer)
5086 buf = buffer;
5087 else if (root_cache->max_attr_size < ARRAY_SIZE(mbuf))
5088 buf = mbuf;
5089 else {
5090 buffer = (char *) get_zeroed_page(GFP_KERNEL);
5091 if (WARN_ON(!buffer))
5092 continue;
5093 buf = buffer;
5094 }
5095
5096 attr->show(root_cache, buf);
5097 attr->store(s, buf, strlen(buf));
5098 }
5099
5100 if (buffer)
5101 free_page((unsigned long)buffer);
5102#endif
5103}
5104
5105static void kmem_cache_release(struct kobject *k)
5106{
5107 slab_kmem_cache_release(to_slab(k));
5108}
5109
5110static const struct sysfs_ops slab_sysfs_ops = {
5111 .show = slab_attr_show,
5112 .store = slab_attr_store,
5113};
5114
5115static struct kobj_type slab_ktype = {
5116 .sysfs_ops = &slab_sysfs_ops,
5117 .release = kmem_cache_release,
5118};
5119
5120static int uevent_filter(struct kset *kset, struct kobject *kobj)
5121{
5122 struct kobj_type *ktype = get_ktype(kobj);
5123
5124 if (ktype == &slab_ktype)
5125 return 1;
5126 return 0;
5127}
5128
5129static const struct kset_uevent_ops slab_uevent_ops = {
5130 .filter = uevent_filter,
5131};
5132
5133static struct kset *slab_kset;
5134
5135static inline struct kset *cache_kset(struct kmem_cache *s)
5136{
5137#ifdef CONFIG_MEMCG_KMEM
5138 if (!is_root_cache(s))
5139 return s->memcg_params->root_cache->memcg_kset;
5140#endif
5141 return slab_kset;
5142}
5143
5144#define ID_STR_LENGTH 64
5145
5146/* Create a unique string id for a slab cache:
5147 *
5148 * Format :[flags-]size
5149 */
5150static char *create_unique_id(struct kmem_cache *s)
5151{
5152 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5153 char *p = name;
5154
5155 BUG_ON(!name);
5156
5157 *p++ = ':';
5158 /*
5159 * First flags affecting slabcache operations. We will only
5160 * get here for aliasable slabs so we do not need to support
5161 * too many flags. The flags here must cover all flags that
5162 * are matched during merging to guarantee that the id is
5163 * unique.
5164 */
5165 if (s->flags & SLAB_CACHE_DMA)
5166 *p++ = 'd';
5167 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5168 *p++ = 'a';
5169 if (s->flags & SLAB_DEBUG_FREE)
5170 *p++ = 'F';
5171 if (!(s->flags & SLAB_NOTRACK))
5172 *p++ = 't';
5173 if (p != name + 1)
5174 *p++ = '-';
5175 p += sprintf(p, "%07d", s->size);
5176
5177#ifdef CONFIG_MEMCG_KMEM
5178 if (!is_root_cache(s))
5179 p += sprintf(p, "-%08d",
5180 memcg_cache_id(s->memcg_params->memcg));
5181#endif
5182
5183 BUG_ON(p > name + ID_STR_LENGTH - 1);
5184 return name;
5185}
5186
5187static int sysfs_slab_add(struct kmem_cache *s)
5188{
5189 int err;
5190 const char *name;
5191 int unmergeable = slab_unmergeable(s);
5192
5193 if (unmergeable) {
5194 /*
5195 * Slabcache can never be merged so we can use the name proper.
5196 * This is typically the case for debug situations. In that
5197 * case we can catch duplicate names easily.
5198 */
5199 sysfs_remove_link(&slab_kset->kobj, s->name);
5200 name = s->name;
5201 } else {
5202 /*
5203 * Create a unique name for the slab as a target
5204 * for the symlinks.
5205 */
5206 name = create_unique_id(s);
5207 }
5208
5209 s->kobj.kset = cache_kset(s);
5210 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, "%s", name);
5211 if (err)
5212 goto out_put_kobj;
5213
5214 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5215 if (err)
5216 goto out_del_kobj;
5217
5218#ifdef CONFIG_MEMCG_KMEM
5219 if (is_root_cache(s)) {
5220 s->memcg_kset = kset_create_and_add("cgroup", NULL, &s->kobj);
5221 if (!s->memcg_kset) {
5222 err = -ENOMEM;
5223 goto out_del_kobj;
5224 }
5225 }
5226#endif
5227
5228 kobject_uevent(&s->kobj, KOBJ_ADD);
5229 if (!unmergeable) {
5230 /* Setup first alias */
5231 sysfs_slab_alias(s, s->name);
5232 }
5233out:
5234 if (!unmergeable)
5235 kfree(name);
5236 return err;
5237out_del_kobj:
5238 kobject_del(&s->kobj);
5239out_put_kobj:
5240 kobject_put(&s->kobj);
5241 goto out;
5242}
5243
5244void sysfs_slab_remove(struct kmem_cache *s)
5245{
5246 if (slab_state < FULL)
5247 /*
5248 * Sysfs has not been setup yet so no need to remove the
5249 * cache from sysfs.
5250 */
5251 return;
5252
5253#ifdef CONFIG_MEMCG_KMEM
5254 kset_unregister(s->memcg_kset);
5255#endif
5256 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5257 kobject_del(&s->kobj);
5258 kobject_put(&s->kobj);
5259}
5260
5261/*
5262 * Need to buffer aliases during bootup until sysfs becomes
5263 * available lest we lose that information.
5264 */
5265struct saved_alias {
5266 struct kmem_cache *s;
5267 const char *name;
5268 struct saved_alias *next;
5269};
5270
5271static struct saved_alias *alias_list;
5272
5273static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5274{
5275 struct saved_alias *al;
5276
5277 if (slab_state == FULL) {
5278 /*
5279 * If we have a leftover link then remove it.
5280 */
5281 sysfs_remove_link(&slab_kset->kobj, name);
5282 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5283 }
5284
5285 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5286 if (!al)
5287 return -ENOMEM;
5288
5289 al->s = s;
5290 al->name = name;
5291 al->next = alias_list;
5292 alias_list = al;
5293 return 0;
5294}
5295
5296static int __init slab_sysfs_init(void)
5297{
5298 struct kmem_cache *s;
5299 int err;
5300
5301 mutex_lock(&slab_mutex);
5302
5303 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5304 if (!slab_kset) {
5305 mutex_unlock(&slab_mutex);
5306 printk(KERN_ERR "Cannot register slab subsystem.\n");
5307 return -ENOSYS;
5308 }
5309
5310 slab_state = FULL;
5311
5312 list_for_each_entry(s, &slab_caches, list) {
5313 err = sysfs_slab_add(s);
5314 if (err)
5315 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5316 " to sysfs\n", s->name);
5317 }
5318
5319 while (alias_list) {
5320 struct saved_alias *al = alias_list;
5321
5322 alias_list = alias_list->next;
5323 err = sysfs_slab_alias(al->s, al->name);
5324 if (err)
5325 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5326 " %s to sysfs\n", al->name);
5327 kfree(al);
5328 }
5329
5330 mutex_unlock(&slab_mutex);
5331 resiliency_test();
5332 return 0;
5333}
5334
5335__initcall(slab_sysfs_init);
5336#endif /* CONFIG_SYSFS */
5337
5338/*
5339 * The /proc/slabinfo ABI
5340 */
5341#ifdef CONFIG_SLABINFO
5342void get_slabinfo(struct kmem_cache *s, struct slabinfo *sinfo)
5343{
5344 unsigned long nr_slabs = 0;
5345 unsigned long nr_objs = 0;
5346 unsigned long nr_free = 0;
5347 int node;
5348
5349 for_each_online_node(node) {
5350 struct kmem_cache_node *n = get_node(s, node);
5351
5352 if (!n)
5353 continue;
5354
5355 nr_slabs += node_nr_slabs(n);
5356 nr_objs += node_nr_objs(n);
5357 nr_free += count_partial(n, count_free);
5358 }
5359
5360 sinfo->active_objs = nr_objs - nr_free;
5361 sinfo->num_objs = nr_objs;
5362 sinfo->active_slabs = nr_slabs;
5363 sinfo->num_slabs = nr_slabs;
5364 sinfo->objects_per_slab = oo_objects(s->oo);
5365 sinfo->cache_order = oo_order(s->oo);
5366}
5367
5368void slabinfo_show_stats(struct seq_file *m, struct kmem_cache *s)
5369{
5370}
5371
5372ssize_t slabinfo_write(struct file *file, const char __user *buffer,
5373 size_t count, loff_t *ppos)
5374{
5375 return -EIO;
5376}
5377#endif /* CONFIG_SLABINFO */
1/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks or atomic operatios
6 * and only uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter
9 * (C) 2011 Linux Foundation, Christoph Lameter
10 */
11
12#include <linux/mm.h>
13#include <linux/swap.h> /* struct reclaim_state */
14#include <linux/module.h>
15#include <linux/bit_spinlock.h>
16#include <linux/interrupt.h>
17#include <linux/bitops.h>
18#include <linux/slab.h>
19#include <linux/proc_fs.h>
20#include <linux/seq_file.h>
21#include <linux/kmemcheck.h>
22#include <linux/cpu.h>
23#include <linux/cpuset.h>
24#include <linux/mempolicy.h>
25#include <linux/ctype.h>
26#include <linux/debugobjects.h>
27#include <linux/kallsyms.h>
28#include <linux/memory.h>
29#include <linux/math64.h>
30#include <linux/fault-inject.h>
31#include <linux/stacktrace.h>
32
33#include <trace/events/kmem.h>
34
35/*
36 * Lock order:
37 * 1. slub_lock (Global Semaphore)
38 * 2. node->list_lock
39 * 3. slab_lock(page) (Only on some arches and for debugging)
40 *
41 * slub_lock
42 *
43 * The role of the slub_lock is to protect the list of all the slabs
44 * and to synchronize major metadata changes to slab cache structures.
45 *
46 * The slab_lock is only used for debugging and on arches that do not
47 * have the ability to do a cmpxchg_double. It only protects the second
48 * double word in the page struct. Meaning
49 * A. page->freelist -> List of object free in a page
50 * B. page->counters -> Counters of objects
51 * C. page->frozen -> frozen state
52 *
53 * If a slab is frozen then it is exempt from list management. It is not
54 * on any list. The processor that froze the slab is the one who can
55 * perform list operations on the page. Other processors may put objects
56 * onto the freelist but the processor that froze the slab is the only
57 * one that can retrieve the objects from the page's freelist.
58 *
59 * The list_lock protects the partial and full list on each node and
60 * the partial slab counter. If taken then no new slabs may be added or
61 * removed from the lists nor make the number of partial slabs be modified.
62 * (Note that the total number of slabs is an atomic value that may be
63 * modified without taking the list lock).
64 *
65 * The list_lock is a centralized lock and thus we avoid taking it as
66 * much as possible. As long as SLUB does not have to handle partial
67 * slabs, operations can continue without any centralized lock. F.e.
68 * allocating a long series of objects that fill up slabs does not require
69 * the list lock.
70 * Interrupts are disabled during allocation and deallocation in order to
71 * make the slab allocator safe to use in the context of an irq. In addition
72 * interrupts are disabled to ensure that the processor does not change
73 * while handling per_cpu slabs, due to kernel preemption.
74 *
75 * SLUB assigns one slab for allocation to each processor.
76 * Allocations only occur from these slabs called cpu slabs.
77 *
78 * Slabs with free elements are kept on a partial list and during regular
79 * operations no list for full slabs is used. If an object in a full slab is
80 * freed then the slab will show up again on the partial lists.
81 * We track full slabs for debugging purposes though because otherwise we
82 * cannot scan all objects.
83 *
84 * Slabs are freed when they become empty. Teardown and setup is
85 * minimal so we rely on the page allocators per cpu caches for
86 * fast frees and allocs.
87 *
88 * Overloading of page flags that are otherwise used for LRU management.
89 *
90 * PageActive The slab is frozen and exempt from list processing.
91 * This means that the slab is dedicated to a purpose
92 * such as satisfying allocations for a specific
93 * processor. Objects may be freed in the slab while
94 * it is frozen but slab_free will then skip the usual
95 * list operations. It is up to the processor holding
96 * the slab to integrate the slab into the slab lists
97 * when the slab is no longer needed.
98 *
99 * One use of this flag is to mark slabs that are
100 * used for allocations. Then such a slab becomes a cpu
101 * slab. The cpu slab may be equipped with an additional
102 * freelist that allows lockless access to
103 * free objects in addition to the regular freelist
104 * that requires the slab lock.
105 *
106 * PageError Slab requires special handling due to debug
107 * options set. This moves slab handling out of
108 * the fast path and disables lockless freelists.
109 */
110
111#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
112 SLAB_TRACE | SLAB_DEBUG_FREE)
113
114static inline int kmem_cache_debug(struct kmem_cache *s)
115{
116#ifdef CONFIG_SLUB_DEBUG
117 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
118#else
119 return 0;
120#endif
121}
122
123/*
124 * Issues still to be resolved:
125 *
126 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
127 *
128 * - Variable sizing of the per node arrays
129 */
130
131/* Enable to test recovery from slab corruption on boot */
132#undef SLUB_RESILIENCY_TEST
133
134/* Enable to log cmpxchg failures */
135#undef SLUB_DEBUG_CMPXCHG
136
137/*
138 * Mininum number of partial slabs. These will be left on the partial
139 * lists even if they are empty. kmem_cache_shrink may reclaim them.
140 */
141#define MIN_PARTIAL 5
142
143/*
144 * Maximum number of desirable partial slabs.
145 * The existence of more partial slabs makes kmem_cache_shrink
146 * sort the partial list by the number of objects in the.
147 */
148#define MAX_PARTIAL 10
149
150#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
151 SLAB_POISON | SLAB_STORE_USER)
152
153/*
154 * Debugging flags that require metadata to be stored in the slab. These get
155 * disabled when slub_debug=O is used and a cache's min order increases with
156 * metadata.
157 */
158#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
159
160/*
161 * Set of flags that will prevent slab merging
162 */
163#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
164 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
165 SLAB_FAILSLAB)
166
167#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
168 SLAB_CACHE_DMA | SLAB_NOTRACK)
169
170#define OO_SHIFT 16
171#define OO_MASK ((1 << OO_SHIFT) - 1)
172#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
173
174/* Internal SLUB flags */
175#define __OBJECT_POISON 0x80000000UL /* Poison object */
176#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
177
178static int kmem_size = sizeof(struct kmem_cache);
179
180#ifdef CONFIG_SMP
181static struct notifier_block slab_notifier;
182#endif
183
184static enum {
185 DOWN, /* No slab functionality available */
186 PARTIAL, /* Kmem_cache_node works */
187 UP, /* Everything works but does not show up in sysfs */
188 SYSFS /* Sysfs up */
189} slab_state = DOWN;
190
191/* A list of all slab caches on the system */
192static DECLARE_RWSEM(slub_lock);
193static LIST_HEAD(slab_caches);
194
195/*
196 * Tracking user of a slab.
197 */
198#define TRACK_ADDRS_COUNT 16
199struct track {
200 unsigned long addr; /* Called from address */
201#ifdef CONFIG_STACKTRACE
202 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
203#endif
204 int cpu; /* Was running on cpu */
205 int pid; /* Pid context */
206 unsigned long when; /* When did the operation occur */
207};
208
209enum track_item { TRACK_ALLOC, TRACK_FREE };
210
211#ifdef CONFIG_SYSFS
212static int sysfs_slab_add(struct kmem_cache *);
213static int sysfs_slab_alias(struct kmem_cache *, const char *);
214static void sysfs_slab_remove(struct kmem_cache *);
215
216#else
217static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
218static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
219 { return 0; }
220static inline void sysfs_slab_remove(struct kmem_cache *s)
221{
222 kfree(s->name);
223 kfree(s);
224}
225
226#endif
227
228static inline void stat(const struct kmem_cache *s, enum stat_item si)
229{
230#ifdef CONFIG_SLUB_STATS
231 __this_cpu_inc(s->cpu_slab->stat[si]);
232#endif
233}
234
235/********************************************************************
236 * Core slab cache functions
237 *******************************************************************/
238
239int slab_is_available(void)
240{
241 return slab_state >= UP;
242}
243
244static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
245{
246 return s->node[node];
247}
248
249/* Verify that a pointer has an address that is valid within a slab page */
250static inline int check_valid_pointer(struct kmem_cache *s,
251 struct page *page, const void *object)
252{
253 void *base;
254
255 if (!object)
256 return 1;
257
258 base = page_address(page);
259 if (object < base || object >= base + page->objects * s->size ||
260 (object - base) % s->size) {
261 return 0;
262 }
263
264 return 1;
265}
266
267static inline void *get_freepointer(struct kmem_cache *s, void *object)
268{
269 return *(void **)(object + s->offset);
270}
271
272static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
273{
274 void *p;
275
276#ifdef CONFIG_DEBUG_PAGEALLOC
277 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
278#else
279 p = get_freepointer(s, object);
280#endif
281 return p;
282}
283
284static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
285{
286 *(void **)(object + s->offset) = fp;
287}
288
289/* Loop over all objects in a slab */
290#define for_each_object(__p, __s, __addr, __objects) \
291 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
292 __p += (__s)->size)
293
294/* Determine object index from a given position */
295static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
296{
297 return (p - addr) / s->size;
298}
299
300static inline size_t slab_ksize(const struct kmem_cache *s)
301{
302#ifdef CONFIG_SLUB_DEBUG
303 /*
304 * Debugging requires use of the padding between object
305 * and whatever may come after it.
306 */
307 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
308 return s->objsize;
309
310#endif
311 /*
312 * If we have the need to store the freelist pointer
313 * back there or track user information then we can
314 * only use the space before that information.
315 */
316 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
317 return s->inuse;
318 /*
319 * Else we can use all the padding etc for the allocation
320 */
321 return s->size;
322}
323
324static inline int order_objects(int order, unsigned long size, int reserved)
325{
326 return ((PAGE_SIZE << order) - reserved) / size;
327}
328
329static inline struct kmem_cache_order_objects oo_make(int order,
330 unsigned long size, int reserved)
331{
332 struct kmem_cache_order_objects x = {
333 (order << OO_SHIFT) + order_objects(order, size, reserved)
334 };
335
336 return x;
337}
338
339static inline int oo_order(struct kmem_cache_order_objects x)
340{
341 return x.x >> OO_SHIFT;
342}
343
344static inline int oo_objects(struct kmem_cache_order_objects x)
345{
346 return x.x & OO_MASK;
347}
348
349/*
350 * Per slab locking using the pagelock
351 */
352static __always_inline void slab_lock(struct page *page)
353{
354 bit_spin_lock(PG_locked, &page->flags);
355}
356
357static __always_inline void slab_unlock(struct page *page)
358{
359 __bit_spin_unlock(PG_locked, &page->flags);
360}
361
362/* Interrupts must be disabled (for the fallback code to work right) */
363static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
364 void *freelist_old, unsigned long counters_old,
365 void *freelist_new, unsigned long counters_new,
366 const char *n)
367{
368 VM_BUG_ON(!irqs_disabled());
369#ifdef CONFIG_CMPXCHG_DOUBLE
370 if (s->flags & __CMPXCHG_DOUBLE) {
371 if (cmpxchg_double(&page->freelist,
372 freelist_old, counters_old,
373 freelist_new, counters_new))
374 return 1;
375 } else
376#endif
377 {
378 slab_lock(page);
379 if (page->freelist == freelist_old && page->counters == counters_old) {
380 page->freelist = freelist_new;
381 page->counters = counters_new;
382 slab_unlock(page);
383 return 1;
384 }
385 slab_unlock(page);
386 }
387
388 cpu_relax();
389 stat(s, CMPXCHG_DOUBLE_FAIL);
390
391#ifdef SLUB_DEBUG_CMPXCHG
392 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
393#endif
394
395 return 0;
396}
397
398static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
399 void *freelist_old, unsigned long counters_old,
400 void *freelist_new, unsigned long counters_new,
401 const char *n)
402{
403#ifdef CONFIG_CMPXCHG_DOUBLE
404 if (s->flags & __CMPXCHG_DOUBLE) {
405 if (cmpxchg_double(&page->freelist,
406 freelist_old, counters_old,
407 freelist_new, counters_new))
408 return 1;
409 } else
410#endif
411 {
412 unsigned long flags;
413
414 local_irq_save(flags);
415 slab_lock(page);
416 if (page->freelist == freelist_old && page->counters == counters_old) {
417 page->freelist = freelist_new;
418 page->counters = counters_new;
419 slab_unlock(page);
420 local_irq_restore(flags);
421 return 1;
422 }
423 slab_unlock(page);
424 local_irq_restore(flags);
425 }
426
427 cpu_relax();
428 stat(s, CMPXCHG_DOUBLE_FAIL);
429
430#ifdef SLUB_DEBUG_CMPXCHG
431 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
432#endif
433
434 return 0;
435}
436
437#ifdef CONFIG_SLUB_DEBUG
438/*
439 * Determine a map of object in use on a page.
440 *
441 * Node listlock must be held to guarantee that the page does
442 * not vanish from under us.
443 */
444static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
445{
446 void *p;
447 void *addr = page_address(page);
448
449 for (p = page->freelist; p; p = get_freepointer(s, p))
450 set_bit(slab_index(p, s, addr), map);
451}
452
453/*
454 * Debug settings:
455 */
456#ifdef CONFIG_SLUB_DEBUG_ON
457static int slub_debug = DEBUG_DEFAULT_FLAGS;
458#else
459static int slub_debug;
460#endif
461
462static char *slub_debug_slabs;
463static int disable_higher_order_debug;
464
465/*
466 * Object debugging
467 */
468static void print_section(char *text, u8 *addr, unsigned int length)
469{
470 int i, offset;
471 int newline = 1;
472 char ascii[17];
473
474 ascii[16] = 0;
475
476 for (i = 0; i < length; i++) {
477 if (newline) {
478 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
479 newline = 0;
480 }
481 printk(KERN_CONT " %02x", addr[i]);
482 offset = i % 16;
483 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
484 if (offset == 15) {
485 printk(KERN_CONT " %s\n", ascii);
486 newline = 1;
487 }
488 }
489 if (!newline) {
490 i %= 16;
491 while (i < 16) {
492 printk(KERN_CONT " ");
493 ascii[i] = ' ';
494 i++;
495 }
496 printk(KERN_CONT " %s\n", ascii);
497 }
498}
499
500static struct track *get_track(struct kmem_cache *s, void *object,
501 enum track_item alloc)
502{
503 struct track *p;
504
505 if (s->offset)
506 p = object + s->offset + sizeof(void *);
507 else
508 p = object + s->inuse;
509
510 return p + alloc;
511}
512
513static void set_track(struct kmem_cache *s, void *object,
514 enum track_item alloc, unsigned long addr)
515{
516 struct track *p = get_track(s, object, alloc);
517
518 if (addr) {
519#ifdef CONFIG_STACKTRACE
520 struct stack_trace trace;
521 int i;
522
523 trace.nr_entries = 0;
524 trace.max_entries = TRACK_ADDRS_COUNT;
525 trace.entries = p->addrs;
526 trace.skip = 3;
527 save_stack_trace(&trace);
528
529 /* See rant in lockdep.c */
530 if (trace.nr_entries != 0 &&
531 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
532 trace.nr_entries--;
533
534 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
535 p->addrs[i] = 0;
536#endif
537 p->addr = addr;
538 p->cpu = smp_processor_id();
539 p->pid = current->pid;
540 p->when = jiffies;
541 } else
542 memset(p, 0, sizeof(struct track));
543}
544
545static void init_tracking(struct kmem_cache *s, void *object)
546{
547 if (!(s->flags & SLAB_STORE_USER))
548 return;
549
550 set_track(s, object, TRACK_FREE, 0UL);
551 set_track(s, object, TRACK_ALLOC, 0UL);
552}
553
554static void print_track(const char *s, struct track *t)
555{
556 if (!t->addr)
557 return;
558
559 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
560 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
561#ifdef CONFIG_STACKTRACE
562 {
563 int i;
564 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
565 if (t->addrs[i])
566 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
567 else
568 break;
569 }
570#endif
571}
572
573static void print_tracking(struct kmem_cache *s, void *object)
574{
575 if (!(s->flags & SLAB_STORE_USER))
576 return;
577
578 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
579 print_track("Freed", get_track(s, object, TRACK_FREE));
580}
581
582static void print_page_info(struct page *page)
583{
584 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
585 page, page->objects, page->inuse, page->freelist, page->flags);
586
587}
588
589static void slab_bug(struct kmem_cache *s, char *fmt, ...)
590{
591 va_list args;
592 char buf[100];
593
594 va_start(args, fmt);
595 vsnprintf(buf, sizeof(buf), fmt, args);
596 va_end(args);
597 printk(KERN_ERR "========================================"
598 "=====================================\n");
599 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
600 printk(KERN_ERR "----------------------------------------"
601 "-------------------------------------\n\n");
602}
603
604static void slab_fix(struct kmem_cache *s, char *fmt, ...)
605{
606 va_list args;
607 char buf[100];
608
609 va_start(args, fmt);
610 vsnprintf(buf, sizeof(buf), fmt, args);
611 va_end(args);
612 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
613}
614
615static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
616{
617 unsigned int off; /* Offset of last byte */
618 u8 *addr = page_address(page);
619
620 print_tracking(s, p);
621
622 print_page_info(page);
623
624 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
625 p, p - addr, get_freepointer(s, p));
626
627 if (p > addr + 16)
628 print_section("Bytes b4", p - 16, 16);
629
630 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
631
632 if (s->flags & SLAB_RED_ZONE)
633 print_section("Redzone", p + s->objsize,
634 s->inuse - s->objsize);
635
636 if (s->offset)
637 off = s->offset + sizeof(void *);
638 else
639 off = s->inuse;
640
641 if (s->flags & SLAB_STORE_USER)
642 off += 2 * sizeof(struct track);
643
644 if (off != s->size)
645 /* Beginning of the filler is the free pointer */
646 print_section("Padding", p + off, s->size - off);
647
648 dump_stack();
649}
650
651static void object_err(struct kmem_cache *s, struct page *page,
652 u8 *object, char *reason)
653{
654 slab_bug(s, "%s", reason);
655 print_trailer(s, page, object);
656}
657
658static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
659{
660 va_list args;
661 char buf[100];
662
663 va_start(args, fmt);
664 vsnprintf(buf, sizeof(buf), fmt, args);
665 va_end(args);
666 slab_bug(s, "%s", buf);
667 print_page_info(page);
668 dump_stack();
669}
670
671static void init_object(struct kmem_cache *s, void *object, u8 val)
672{
673 u8 *p = object;
674
675 if (s->flags & __OBJECT_POISON) {
676 memset(p, POISON_FREE, s->objsize - 1);
677 p[s->objsize - 1] = POISON_END;
678 }
679
680 if (s->flags & SLAB_RED_ZONE)
681 memset(p + s->objsize, val, s->inuse - s->objsize);
682}
683
684static u8 *check_bytes8(u8 *start, u8 value, unsigned int bytes)
685{
686 while (bytes) {
687 if (*start != value)
688 return start;
689 start++;
690 bytes--;
691 }
692 return NULL;
693}
694
695static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
696{
697 u64 value64;
698 unsigned int words, prefix;
699
700 if (bytes <= 16)
701 return check_bytes8(start, value, bytes);
702
703 value64 = value | value << 8 | value << 16 | value << 24;
704 value64 = (value64 & 0xffffffff) | value64 << 32;
705 prefix = 8 - ((unsigned long)start) % 8;
706
707 if (prefix) {
708 u8 *r = check_bytes8(start, value, prefix);
709 if (r)
710 return r;
711 start += prefix;
712 bytes -= prefix;
713 }
714
715 words = bytes / 8;
716
717 while (words) {
718 if (*(u64 *)start != value64)
719 return check_bytes8(start, value, 8);
720 start += 8;
721 words--;
722 }
723
724 return check_bytes8(start, value, bytes % 8);
725}
726
727static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
728 void *from, void *to)
729{
730 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
731 memset(from, data, to - from);
732}
733
734static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
735 u8 *object, char *what,
736 u8 *start, unsigned int value, unsigned int bytes)
737{
738 u8 *fault;
739 u8 *end;
740
741 fault = check_bytes(start, value, bytes);
742 if (!fault)
743 return 1;
744
745 end = start + bytes;
746 while (end > fault && end[-1] == value)
747 end--;
748
749 slab_bug(s, "%s overwritten", what);
750 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
751 fault, end - 1, fault[0], value);
752 print_trailer(s, page, object);
753
754 restore_bytes(s, what, value, fault, end);
755 return 0;
756}
757
758/*
759 * Object layout:
760 *
761 * object address
762 * Bytes of the object to be managed.
763 * If the freepointer may overlay the object then the free
764 * pointer is the first word of the object.
765 *
766 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
767 * 0xa5 (POISON_END)
768 *
769 * object + s->objsize
770 * Padding to reach word boundary. This is also used for Redzoning.
771 * Padding is extended by another word if Redzoning is enabled and
772 * objsize == inuse.
773 *
774 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
775 * 0xcc (RED_ACTIVE) for objects in use.
776 *
777 * object + s->inuse
778 * Meta data starts here.
779 *
780 * A. Free pointer (if we cannot overwrite object on free)
781 * B. Tracking data for SLAB_STORE_USER
782 * C. Padding to reach required alignment boundary or at mininum
783 * one word if debugging is on to be able to detect writes
784 * before the word boundary.
785 *
786 * Padding is done using 0x5a (POISON_INUSE)
787 *
788 * object + s->size
789 * Nothing is used beyond s->size.
790 *
791 * If slabcaches are merged then the objsize and inuse boundaries are mostly
792 * ignored. And therefore no slab options that rely on these boundaries
793 * may be used with merged slabcaches.
794 */
795
796static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
797{
798 unsigned long off = s->inuse; /* The end of info */
799
800 if (s->offset)
801 /* Freepointer is placed after the object. */
802 off += sizeof(void *);
803
804 if (s->flags & SLAB_STORE_USER)
805 /* We also have user information there */
806 off += 2 * sizeof(struct track);
807
808 if (s->size == off)
809 return 1;
810
811 return check_bytes_and_report(s, page, p, "Object padding",
812 p + off, POISON_INUSE, s->size - off);
813}
814
815/* Check the pad bytes at the end of a slab page */
816static int slab_pad_check(struct kmem_cache *s, struct page *page)
817{
818 u8 *start;
819 u8 *fault;
820 u8 *end;
821 int length;
822 int remainder;
823
824 if (!(s->flags & SLAB_POISON))
825 return 1;
826
827 start = page_address(page);
828 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
829 end = start + length;
830 remainder = length % s->size;
831 if (!remainder)
832 return 1;
833
834 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
835 if (!fault)
836 return 1;
837 while (end > fault && end[-1] == POISON_INUSE)
838 end--;
839
840 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
841 print_section("Padding", end - remainder, remainder);
842
843 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
844 return 0;
845}
846
847static int check_object(struct kmem_cache *s, struct page *page,
848 void *object, u8 val)
849{
850 u8 *p = object;
851 u8 *endobject = object + s->objsize;
852
853 if (s->flags & SLAB_RED_ZONE) {
854 if (!check_bytes_and_report(s, page, object, "Redzone",
855 endobject, val, s->inuse - s->objsize))
856 return 0;
857 } else {
858 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
859 check_bytes_and_report(s, page, p, "Alignment padding",
860 endobject, POISON_INUSE, s->inuse - s->objsize);
861 }
862 }
863
864 if (s->flags & SLAB_POISON) {
865 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
866 (!check_bytes_and_report(s, page, p, "Poison", p,
867 POISON_FREE, s->objsize - 1) ||
868 !check_bytes_and_report(s, page, p, "Poison",
869 p + s->objsize - 1, POISON_END, 1)))
870 return 0;
871 /*
872 * check_pad_bytes cleans up on its own.
873 */
874 check_pad_bytes(s, page, p);
875 }
876
877 if (!s->offset && val == SLUB_RED_ACTIVE)
878 /*
879 * Object and freepointer overlap. Cannot check
880 * freepointer while object is allocated.
881 */
882 return 1;
883
884 /* Check free pointer validity */
885 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
886 object_err(s, page, p, "Freepointer corrupt");
887 /*
888 * No choice but to zap it and thus lose the remainder
889 * of the free objects in this slab. May cause
890 * another error because the object count is now wrong.
891 */
892 set_freepointer(s, p, NULL);
893 return 0;
894 }
895 return 1;
896}
897
898static int check_slab(struct kmem_cache *s, struct page *page)
899{
900 int maxobj;
901
902 VM_BUG_ON(!irqs_disabled());
903
904 if (!PageSlab(page)) {
905 slab_err(s, page, "Not a valid slab page");
906 return 0;
907 }
908
909 maxobj = order_objects(compound_order(page), s->size, s->reserved);
910 if (page->objects > maxobj) {
911 slab_err(s, page, "objects %u > max %u",
912 s->name, page->objects, maxobj);
913 return 0;
914 }
915 if (page->inuse > page->objects) {
916 slab_err(s, page, "inuse %u > max %u",
917 s->name, page->inuse, page->objects);
918 return 0;
919 }
920 /* Slab_pad_check fixes things up after itself */
921 slab_pad_check(s, page);
922 return 1;
923}
924
925/*
926 * Determine if a certain object on a page is on the freelist. Must hold the
927 * slab lock to guarantee that the chains are in a consistent state.
928 */
929static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
930{
931 int nr = 0;
932 void *fp;
933 void *object = NULL;
934 unsigned long max_objects;
935
936 fp = page->freelist;
937 while (fp && nr <= page->objects) {
938 if (fp == search)
939 return 1;
940 if (!check_valid_pointer(s, page, fp)) {
941 if (object) {
942 object_err(s, page, object,
943 "Freechain corrupt");
944 set_freepointer(s, object, NULL);
945 break;
946 } else {
947 slab_err(s, page, "Freepointer corrupt");
948 page->freelist = NULL;
949 page->inuse = page->objects;
950 slab_fix(s, "Freelist cleared");
951 return 0;
952 }
953 break;
954 }
955 object = fp;
956 fp = get_freepointer(s, object);
957 nr++;
958 }
959
960 max_objects = order_objects(compound_order(page), s->size, s->reserved);
961 if (max_objects > MAX_OBJS_PER_PAGE)
962 max_objects = MAX_OBJS_PER_PAGE;
963
964 if (page->objects != max_objects) {
965 slab_err(s, page, "Wrong number of objects. Found %d but "
966 "should be %d", page->objects, max_objects);
967 page->objects = max_objects;
968 slab_fix(s, "Number of objects adjusted.");
969 }
970 if (page->inuse != page->objects - nr) {
971 slab_err(s, page, "Wrong object count. Counter is %d but "
972 "counted were %d", page->inuse, page->objects - nr);
973 page->inuse = page->objects - nr;
974 slab_fix(s, "Object count adjusted.");
975 }
976 return search == NULL;
977}
978
979static void trace(struct kmem_cache *s, struct page *page, void *object,
980 int alloc)
981{
982 if (s->flags & SLAB_TRACE) {
983 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
984 s->name,
985 alloc ? "alloc" : "free",
986 object, page->inuse,
987 page->freelist);
988
989 if (!alloc)
990 print_section("Object", (void *)object, s->objsize);
991
992 dump_stack();
993 }
994}
995
996/*
997 * Hooks for other subsystems that check memory allocations. In a typical
998 * production configuration these hooks all should produce no code at all.
999 */
1000static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1001{
1002 flags &= gfp_allowed_mask;
1003 lockdep_trace_alloc(flags);
1004 might_sleep_if(flags & __GFP_WAIT);
1005
1006 return should_failslab(s->objsize, flags, s->flags);
1007}
1008
1009static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
1010{
1011 flags &= gfp_allowed_mask;
1012 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
1013 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
1014}
1015
1016static inline void slab_free_hook(struct kmem_cache *s, void *x)
1017{
1018 kmemleak_free_recursive(x, s->flags);
1019
1020 /*
1021 * Trouble is that we may no longer disable interupts in the fast path
1022 * So in order to make the debug calls that expect irqs to be
1023 * disabled we need to disable interrupts temporarily.
1024 */
1025#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
1026 {
1027 unsigned long flags;
1028
1029 local_irq_save(flags);
1030 kmemcheck_slab_free(s, x, s->objsize);
1031 debug_check_no_locks_freed(x, s->objsize);
1032 local_irq_restore(flags);
1033 }
1034#endif
1035 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1036 debug_check_no_obj_freed(x, s->objsize);
1037}
1038
1039/*
1040 * Tracking of fully allocated slabs for debugging purposes.
1041 *
1042 * list_lock must be held.
1043 */
1044static void add_full(struct kmem_cache *s,
1045 struct kmem_cache_node *n, struct page *page)
1046{
1047 if (!(s->flags & SLAB_STORE_USER))
1048 return;
1049
1050 list_add(&page->lru, &n->full);
1051}
1052
1053/*
1054 * list_lock must be held.
1055 */
1056static void remove_full(struct kmem_cache *s, struct page *page)
1057{
1058 if (!(s->flags & SLAB_STORE_USER))
1059 return;
1060
1061 list_del(&page->lru);
1062}
1063
1064/* Tracking of the number of slabs for debugging purposes */
1065static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1066{
1067 struct kmem_cache_node *n = get_node(s, node);
1068
1069 return atomic_long_read(&n->nr_slabs);
1070}
1071
1072static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1073{
1074 return atomic_long_read(&n->nr_slabs);
1075}
1076
1077static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1078{
1079 struct kmem_cache_node *n = get_node(s, node);
1080
1081 /*
1082 * May be called early in order to allocate a slab for the
1083 * kmem_cache_node structure. Solve the chicken-egg
1084 * dilemma by deferring the increment of the count during
1085 * bootstrap (see early_kmem_cache_node_alloc).
1086 */
1087 if (n) {
1088 atomic_long_inc(&n->nr_slabs);
1089 atomic_long_add(objects, &n->total_objects);
1090 }
1091}
1092static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1093{
1094 struct kmem_cache_node *n = get_node(s, node);
1095
1096 atomic_long_dec(&n->nr_slabs);
1097 atomic_long_sub(objects, &n->total_objects);
1098}
1099
1100/* Object debug checks for alloc/free paths */
1101static void setup_object_debug(struct kmem_cache *s, struct page *page,
1102 void *object)
1103{
1104 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1105 return;
1106
1107 init_object(s, object, SLUB_RED_INACTIVE);
1108 init_tracking(s, object);
1109}
1110
1111static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1112 void *object, unsigned long addr)
1113{
1114 if (!check_slab(s, page))
1115 goto bad;
1116
1117 if (!check_valid_pointer(s, page, object)) {
1118 object_err(s, page, object, "Freelist Pointer check fails");
1119 goto bad;
1120 }
1121
1122 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1123 goto bad;
1124
1125 /* Success perform special debug activities for allocs */
1126 if (s->flags & SLAB_STORE_USER)
1127 set_track(s, object, TRACK_ALLOC, addr);
1128 trace(s, page, object, 1);
1129 init_object(s, object, SLUB_RED_ACTIVE);
1130 return 1;
1131
1132bad:
1133 if (PageSlab(page)) {
1134 /*
1135 * If this is a slab page then lets do the best we can
1136 * to avoid issues in the future. Marking all objects
1137 * as used avoids touching the remaining objects.
1138 */
1139 slab_fix(s, "Marking all objects used");
1140 page->inuse = page->objects;
1141 page->freelist = NULL;
1142 }
1143 return 0;
1144}
1145
1146static noinline int free_debug_processing(struct kmem_cache *s,
1147 struct page *page, void *object, unsigned long addr)
1148{
1149 unsigned long flags;
1150 int rc = 0;
1151
1152 local_irq_save(flags);
1153 slab_lock(page);
1154
1155 if (!check_slab(s, page))
1156 goto fail;
1157
1158 if (!check_valid_pointer(s, page, object)) {
1159 slab_err(s, page, "Invalid object pointer 0x%p", object);
1160 goto fail;
1161 }
1162
1163 if (on_freelist(s, page, object)) {
1164 object_err(s, page, object, "Object already free");
1165 goto fail;
1166 }
1167
1168 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1169 goto out;
1170
1171 if (unlikely(s != page->slab)) {
1172 if (!PageSlab(page)) {
1173 slab_err(s, page, "Attempt to free object(0x%p) "
1174 "outside of slab", object);
1175 } else if (!page->slab) {
1176 printk(KERN_ERR
1177 "SLUB <none>: no slab for object 0x%p.\n",
1178 object);
1179 dump_stack();
1180 } else
1181 object_err(s, page, object,
1182 "page slab pointer corrupt.");
1183 goto fail;
1184 }
1185
1186 if (s->flags & SLAB_STORE_USER)
1187 set_track(s, object, TRACK_FREE, addr);
1188 trace(s, page, object, 0);
1189 init_object(s, object, SLUB_RED_INACTIVE);
1190 rc = 1;
1191out:
1192 slab_unlock(page);
1193 local_irq_restore(flags);
1194 return rc;
1195
1196fail:
1197 slab_fix(s, "Object at 0x%p not freed", object);
1198 goto out;
1199}
1200
1201static int __init setup_slub_debug(char *str)
1202{
1203 slub_debug = DEBUG_DEFAULT_FLAGS;
1204 if (*str++ != '=' || !*str)
1205 /*
1206 * No options specified. Switch on full debugging.
1207 */
1208 goto out;
1209
1210 if (*str == ',')
1211 /*
1212 * No options but restriction on slabs. This means full
1213 * debugging for slabs matching a pattern.
1214 */
1215 goto check_slabs;
1216
1217 if (tolower(*str) == 'o') {
1218 /*
1219 * Avoid enabling debugging on caches if its minimum order
1220 * would increase as a result.
1221 */
1222 disable_higher_order_debug = 1;
1223 goto out;
1224 }
1225
1226 slub_debug = 0;
1227 if (*str == '-')
1228 /*
1229 * Switch off all debugging measures.
1230 */
1231 goto out;
1232
1233 /*
1234 * Determine which debug features should be switched on
1235 */
1236 for (; *str && *str != ','; str++) {
1237 switch (tolower(*str)) {
1238 case 'f':
1239 slub_debug |= SLAB_DEBUG_FREE;
1240 break;
1241 case 'z':
1242 slub_debug |= SLAB_RED_ZONE;
1243 break;
1244 case 'p':
1245 slub_debug |= SLAB_POISON;
1246 break;
1247 case 'u':
1248 slub_debug |= SLAB_STORE_USER;
1249 break;
1250 case 't':
1251 slub_debug |= SLAB_TRACE;
1252 break;
1253 case 'a':
1254 slub_debug |= SLAB_FAILSLAB;
1255 break;
1256 default:
1257 printk(KERN_ERR "slub_debug option '%c' "
1258 "unknown. skipped\n", *str);
1259 }
1260 }
1261
1262check_slabs:
1263 if (*str == ',')
1264 slub_debug_slabs = str + 1;
1265out:
1266 return 1;
1267}
1268
1269__setup("slub_debug", setup_slub_debug);
1270
1271static unsigned long kmem_cache_flags(unsigned long objsize,
1272 unsigned long flags, const char *name,
1273 void (*ctor)(void *))
1274{
1275 /*
1276 * Enable debugging if selected on the kernel commandline.
1277 */
1278 if (slub_debug && (!slub_debug_slabs ||
1279 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1280 flags |= slub_debug;
1281
1282 return flags;
1283}
1284#else
1285static inline void setup_object_debug(struct kmem_cache *s,
1286 struct page *page, void *object) {}
1287
1288static inline int alloc_debug_processing(struct kmem_cache *s,
1289 struct page *page, void *object, unsigned long addr) { return 0; }
1290
1291static inline int free_debug_processing(struct kmem_cache *s,
1292 struct page *page, void *object, unsigned long addr) { return 0; }
1293
1294static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1295 { return 1; }
1296static inline int check_object(struct kmem_cache *s, struct page *page,
1297 void *object, u8 val) { return 1; }
1298static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1299 struct page *page) {}
1300static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1301static inline unsigned long kmem_cache_flags(unsigned long objsize,
1302 unsigned long flags, const char *name,
1303 void (*ctor)(void *))
1304{
1305 return flags;
1306}
1307#define slub_debug 0
1308
1309#define disable_higher_order_debug 0
1310
1311static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1312 { return 0; }
1313static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1314 { return 0; }
1315static inline void inc_slabs_node(struct kmem_cache *s, int node,
1316 int objects) {}
1317static inline void dec_slabs_node(struct kmem_cache *s, int node,
1318 int objects) {}
1319
1320static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1321 { return 0; }
1322
1323static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1324 void *object) {}
1325
1326static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1327
1328#endif /* CONFIG_SLUB_DEBUG */
1329
1330/*
1331 * Slab allocation and freeing
1332 */
1333static inline struct page *alloc_slab_page(gfp_t flags, int node,
1334 struct kmem_cache_order_objects oo)
1335{
1336 int order = oo_order(oo);
1337
1338 flags |= __GFP_NOTRACK;
1339
1340 if (node == NUMA_NO_NODE)
1341 return alloc_pages(flags, order);
1342 else
1343 return alloc_pages_exact_node(node, flags, order);
1344}
1345
1346static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1347{
1348 struct page *page;
1349 struct kmem_cache_order_objects oo = s->oo;
1350 gfp_t alloc_gfp;
1351
1352 flags &= gfp_allowed_mask;
1353
1354 if (flags & __GFP_WAIT)
1355 local_irq_enable();
1356
1357 flags |= s->allocflags;
1358
1359 /*
1360 * Let the initial higher-order allocation fail under memory pressure
1361 * so we fall-back to the minimum order allocation.
1362 */
1363 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1364
1365 page = alloc_slab_page(alloc_gfp, node, oo);
1366 if (unlikely(!page)) {
1367 oo = s->min;
1368 /*
1369 * Allocation may have failed due to fragmentation.
1370 * Try a lower order alloc if possible
1371 */
1372 page = alloc_slab_page(flags, node, oo);
1373
1374 if (page)
1375 stat(s, ORDER_FALLBACK);
1376 }
1377
1378 if (flags & __GFP_WAIT)
1379 local_irq_disable();
1380
1381 if (!page)
1382 return NULL;
1383
1384 if (kmemcheck_enabled
1385 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1386 int pages = 1 << oo_order(oo);
1387
1388 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1389
1390 /*
1391 * Objects from caches that have a constructor don't get
1392 * cleared when they're allocated, so we need to do it here.
1393 */
1394 if (s->ctor)
1395 kmemcheck_mark_uninitialized_pages(page, pages);
1396 else
1397 kmemcheck_mark_unallocated_pages(page, pages);
1398 }
1399
1400 page->objects = oo_objects(oo);
1401 mod_zone_page_state(page_zone(page),
1402 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1403 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1404 1 << oo_order(oo));
1405
1406 return page;
1407}
1408
1409static void setup_object(struct kmem_cache *s, struct page *page,
1410 void *object)
1411{
1412 setup_object_debug(s, page, object);
1413 if (unlikely(s->ctor))
1414 s->ctor(object);
1415}
1416
1417static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1418{
1419 struct page *page;
1420 void *start;
1421 void *last;
1422 void *p;
1423
1424 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1425
1426 page = allocate_slab(s,
1427 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1428 if (!page)
1429 goto out;
1430
1431 inc_slabs_node(s, page_to_nid(page), page->objects);
1432 page->slab = s;
1433 page->flags |= 1 << PG_slab;
1434
1435 start = page_address(page);
1436
1437 if (unlikely(s->flags & SLAB_POISON))
1438 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1439
1440 last = start;
1441 for_each_object(p, s, start, page->objects) {
1442 setup_object(s, page, last);
1443 set_freepointer(s, last, p);
1444 last = p;
1445 }
1446 setup_object(s, page, last);
1447 set_freepointer(s, last, NULL);
1448
1449 page->freelist = start;
1450 page->inuse = 0;
1451 page->frozen = 1;
1452out:
1453 return page;
1454}
1455
1456static void __free_slab(struct kmem_cache *s, struct page *page)
1457{
1458 int order = compound_order(page);
1459 int pages = 1 << order;
1460
1461 if (kmem_cache_debug(s)) {
1462 void *p;
1463
1464 slab_pad_check(s, page);
1465 for_each_object(p, s, page_address(page),
1466 page->objects)
1467 check_object(s, page, p, SLUB_RED_INACTIVE);
1468 }
1469
1470 kmemcheck_free_shadow(page, compound_order(page));
1471
1472 mod_zone_page_state(page_zone(page),
1473 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1474 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1475 -pages);
1476
1477 __ClearPageSlab(page);
1478 reset_page_mapcount(page);
1479 if (current->reclaim_state)
1480 current->reclaim_state->reclaimed_slab += pages;
1481 __free_pages(page, order);
1482}
1483
1484#define need_reserve_slab_rcu \
1485 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1486
1487static void rcu_free_slab(struct rcu_head *h)
1488{
1489 struct page *page;
1490
1491 if (need_reserve_slab_rcu)
1492 page = virt_to_head_page(h);
1493 else
1494 page = container_of((struct list_head *)h, struct page, lru);
1495
1496 __free_slab(page->slab, page);
1497}
1498
1499static void free_slab(struct kmem_cache *s, struct page *page)
1500{
1501 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1502 struct rcu_head *head;
1503
1504 if (need_reserve_slab_rcu) {
1505 int order = compound_order(page);
1506 int offset = (PAGE_SIZE << order) - s->reserved;
1507
1508 VM_BUG_ON(s->reserved != sizeof(*head));
1509 head = page_address(page) + offset;
1510 } else {
1511 /*
1512 * RCU free overloads the RCU head over the LRU
1513 */
1514 head = (void *)&page->lru;
1515 }
1516
1517 call_rcu(head, rcu_free_slab);
1518 } else
1519 __free_slab(s, page);
1520}
1521
1522static void discard_slab(struct kmem_cache *s, struct page *page)
1523{
1524 dec_slabs_node(s, page_to_nid(page), page->objects);
1525 free_slab(s, page);
1526}
1527
1528/*
1529 * Management of partially allocated slabs.
1530 *
1531 * list_lock must be held.
1532 */
1533static inline void add_partial(struct kmem_cache_node *n,
1534 struct page *page, int tail)
1535{
1536 n->nr_partial++;
1537 if (tail)
1538 list_add_tail(&page->lru, &n->partial);
1539 else
1540 list_add(&page->lru, &n->partial);
1541}
1542
1543/*
1544 * list_lock must be held.
1545 */
1546static inline void remove_partial(struct kmem_cache_node *n,
1547 struct page *page)
1548{
1549 list_del(&page->lru);
1550 n->nr_partial--;
1551}
1552
1553/*
1554 * Lock slab, remove from the partial list and put the object into the
1555 * per cpu freelist.
1556 *
1557 * Must hold list_lock.
1558 */
1559static inline int acquire_slab(struct kmem_cache *s,
1560 struct kmem_cache_node *n, struct page *page)
1561{
1562 void *freelist;
1563 unsigned long counters;
1564 struct page new;
1565
1566 /*
1567 * Zap the freelist and set the frozen bit.
1568 * The old freelist is the list of objects for the
1569 * per cpu allocation list.
1570 */
1571 do {
1572 freelist = page->freelist;
1573 counters = page->counters;
1574 new.counters = counters;
1575 new.inuse = page->objects;
1576
1577 VM_BUG_ON(new.frozen);
1578 new.frozen = 1;
1579
1580 } while (!__cmpxchg_double_slab(s, page,
1581 freelist, counters,
1582 NULL, new.counters,
1583 "lock and freeze"));
1584
1585 remove_partial(n, page);
1586
1587 if (freelist) {
1588 /* Populate the per cpu freelist */
1589 this_cpu_write(s->cpu_slab->freelist, freelist);
1590 this_cpu_write(s->cpu_slab->page, page);
1591 this_cpu_write(s->cpu_slab->node, page_to_nid(page));
1592 return 1;
1593 } else {
1594 /*
1595 * Slab page came from the wrong list. No object to allocate
1596 * from. Put it onto the correct list and continue partial
1597 * scan.
1598 */
1599 printk(KERN_ERR "SLUB: %s : Page without available objects on"
1600 " partial list\n", s->name);
1601 return 0;
1602 }
1603}
1604
1605/*
1606 * Try to allocate a partial slab from a specific node.
1607 */
1608static struct page *get_partial_node(struct kmem_cache *s,
1609 struct kmem_cache_node *n)
1610{
1611 struct page *page;
1612
1613 /*
1614 * Racy check. If we mistakenly see no partial slabs then we
1615 * just allocate an empty slab. If we mistakenly try to get a
1616 * partial slab and there is none available then get_partials()
1617 * will return NULL.
1618 */
1619 if (!n || !n->nr_partial)
1620 return NULL;
1621
1622 spin_lock(&n->list_lock);
1623 list_for_each_entry(page, &n->partial, lru)
1624 if (acquire_slab(s, n, page))
1625 goto out;
1626 page = NULL;
1627out:
1628 spin_unlock(&n->list_lock);
1629 return page;
1630}
1631
1632/*
1633 * Get a page from somewhere. Search in increasing NUMA distances.
1634 */
1635static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1636{
1637#ifdef CONFIG_NUMA
1638 struct zonelist *zonelist;
1639 struct zoneref *z;
1640 struct zone *zone;
1641 enum zone_type high_zoneidx = gfp_zone(flags);
1642 struct page *page;
1643
1644 /*
1645 * The defrag ratio allows a configuration of the tradeoffs between
1646 * inter node defragmentation and node local allocations. A lower
1647 * defrag_ratio increases the tendency to do local allocations
1648 * instead of attempting to obtain partial slabs from other nodes.
1649 *
1650 * If the defrag_ratio is set to 0 then kmalloc() always
1651 * returns node local objects. If the ratio is higher then kmalloc()
1652 * may return off node objects because partial slabs are obtained
1653 * from other nodes and filled up.
1654 *
1655 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1656 * defrag_ratio = 1000) then every (well almost) allocation will
1657 * first attempt to defrag slab caches on other nodes. This means
1658 * scanning over all nodes to look for partial slabs which may be
1659 * expensive if we do it every time we are trying to find a slab
1660 * with available objects.
1661 */
1662 if (!s->remote_node_defrag_ratio ||
1663 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1664 return NULL;
1665
1666 get_mems_allowed();
1667 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1668 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1669 struct kmem_cache_node *n;
1670
1671 n = get_node(s, zone_to_nid(zone));
1672
1673 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1674 n->nr_partial > s->min_partial) {
1675 page = get_partial_node(s, n);
1676 if (page) {
1677 put_mems_allowed();
1678 return page;
1679 }
1680 }
1681 }
1682 put_mems_allowed();
1683#endif
1684 return NULL;
1685}
1686
1687/*
1688 * Get a partial page, lock it and return it.
1689 */
1690static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1691{
1692 struct page *page;
1693 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1694
1695 page = get_partial_node(s, get_node(s, searchnode));
1696 if (page || node != NUMA_NO_NODE)
1697 return page;
1698
1699 return get_any_partial(s, flags);
1700}
1701
1702#ifdef CONFIG_PREEMPT
1703/*
1704 * Calculate the next globally unique transaction for disambiguiation
1705 * during cmpxchg. The transactions start with the cpu number and are then
1706 * incremented by CONFIG_NR_CPUS.
1707 */
1708#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1709#else
1710/*
1711 * No preemption supported therefore also no need to check for
1712 * different cpus.
1713 */
1714#define TID_STEP 1
1715#endif
1716
1717static inline unsigned long next_tid(unsigned long tid)
1718{
1719 return tid + TID_STEP;
1720}
1721
1722static inline unsigned int tid_to_cpu(unsigned long tid)
1723{
1724 return tid % TID_STEP;
1725}
1726
1727static inline unsigned long tid_to_event(unsigned long tid)
1728{
1729 return tid / TID_STEP;
1730}
1731
1732static inline unsigned int init_tid(int cpu)
1733{
1734 return cpu;
1735}
1736
1737static inline void note_cmpxchg_failure(const char *n,
1738 const struct kmem_cache *s, unsigned long tid)
1739{
1740#ifdef SLUB_DEBUG_CMPXCHG
1741 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1742
1743 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1744
1745#ifdef CONFIG_PREEMPT
1746 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1747 printk("due to cpu change %d -> %d\n",
1748 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1749 else
1750#endif
1751 if (tid_to_event(tid) != tid_to_event(actual_tid))
1752 printk("due to cpu running other code. Event %ld->%ld\n",
1753 tid_to_event(tid), tid_to_event(actual_tid));
1754 else
1755 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1756 actual_tid, tid, next_tid(tid));
1757#endif
1758 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1759}
1760
1761void init_kmem_cache_cpus(struct kmem_cache *s)
1762{
1763 int cpu;
1764
1765 for_each_possible_cpu(cpu)
1766 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1767}
1768/*
1769 * Remove the cpu slab
1770 */
1771
1772/*
1773 * Remove the cpu slab
1774 */
1775static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1776{
1777 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1778 struct page *page = c->page;
1779 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1780 int lock = 0;
1781 enum slab_modes l = M_NONE, m = M_NONE;
1782 void *freelist;
1783 void *nextfree;
1784 int tail = 0;
1785 struct page new;
1786 struct page old;
1787
1788 if (page->freelist) {
1789 stat(s, DEACTIVATE_REMOTE_FREES);
1790 tail = 1;
1791 }
1792
1793 c->tid = next_tid(c->tid);
1794 c->page = NULL;
1795 freelist = c->freelist;
1796 c->freelist = NULL;
1797
1798 /*
1799 * Stage one: Free all available per cpu objects back
1800 * to the page freelist while it is still frozen. Leave the
1801 * last one.
1802 *
1803 * There is no need to take the list->lock because the page
1804 * is still frozen.
1805 */
1806 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1807 void *prior;
1808 unsigned long counters;
1809
1810 do {
1811 prior = page->freelist;
1812 counters = page->counters;
1813 set_freepointer(s, freelist, prior);
1814 new.counters = counters;
1815 new.inuse--;
1816 VM_BUG_ON(!new.frozen);
1817
1818 } while (!__cmpxchg_double_slab(s, page,
1819 prior, counters,
1820 freelist, new.counters,
1821 "drain percpu freelist"));
1822
1823 freelist = nextfree;
1824 }
1825
1826 /*
1827 * Stage two: Ensure that the page is unfrozen while the
1828 * list presence reflects the actual number of objects
1829 * during unfreeze.
1830 *
1831 * We setup the list membership and then perform a cmpxchg
1832 * with the count. If there is a mismatch then the page
1833 * is not unfrozen but the page is on the wrong list.
1834 *
1835 * Then we restart the process which may have to remove
1836 * the page from the list that we just put it on again
1837 * because the number of objects in the slab may have
1838 * changed.
1839 */
1840redo:
1841
1842 old.freelist = page->freelist;
1843 old.counters = page->counters;
1844 VM_BUG_ON(!old.frozen);
1845
1846 /* Determine target state of the slab */
1847 new.counters = old.counters;
1848 if (freelist) {
1849 new.inuse--;
1850 set_freepointer(s, freelist, old.freelist);
1851 new.freelist = freelist;
1852 } else
1853 new.freelist = old.freelist;
1854
1855 new.frozen = 0;
1856
1857 if (!new.inuse && n->nr_partial > s->min_partial)
1858 m = M_FREE;
1859 else if (new.freelist) {
1860 m = M_PARTIAL;
1861 if (!lock) {
1862 lock = 1;
1863 /*
1864 * Taking the spinlock removes the possiblity
1865 * that acquire_slab() will see a slab page that
1866 * is frozen
1867 */
1868 spin_lock(&n->list_lock);
1869 }
1870 } else {
1871 m = M_FULL;
1872 if (kmem_cache_debug(s) && !lock) {
1873 lock = 1;
1874 /*
1875 * This also ensures that the scanning of full
1876 * slabs from diagnostic functions will not see
1877 * any frozen slabs.
1878 */
1879 spin_lock(&n->list_lock);
1880 }
1881 }
1882
1883 if (l != m) {
1884
1885 if (l == M_PARTIAL)
1886
1887 remove_partial(n, page);
1888
1889 else if (l == M_FULL)
1890
1891 remove_full(s, page);
1892
1893 if (m == M_PARTIAL) {
1894
1895 add_partial(n, page, tail);
1896 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1897
1898 } else if (m == M_FULL) {
1899
1900 stat(s, DEACTIVATE_FULL);
1901 add_full(s, n, page);
1902
1903 }
1904 }
1905
1906 l = m;
1907 if (!__cmpxchg_double_slab(s, page,
1908 old.freelist, old.counters,
1909 new.freelist, new.counters,
1910 "unfreezing slab"))
1911 goto redo;
1912
1913 if (lock)
1914 spin_unlock(&n->list_lock);
1915
1916 if (m == M_FREE) {
1917 stat(s, DEACTIVATE_EMPTY);
1918 discard_slab(s, page);
1919 stat(s, FREE_SLAB);
1920 }
1921}
1922
1923static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1924{
1925 stat(s, CPUSLAB_FLUSH);
1926 deactivate_slab(s, c);
1927}
1928
1929/*
1930 * Flush cpu slab.
1931 *
1932 * Called from IPI handler with interrupts disabled.
1933 */
1934static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1935{
1936 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1937
1938 if (likely(c && c->page))
1939 flush_slab(s, c);
1940}
1941
1942static void flush_cpu_slab(void *d)
1943{
1944 struct kmem_cache *s = d;
1945
1946 __flush_cpu_slab(s, smp_processor_id());
1947}
1948
1949static void flush_all(struct kmem_cache *s)
1950{
1951 on_each_cpu(flush_cpu_slab, s, 1);
1952}
1953
1954/*
1955 * Check if the objects in a per cpu structure fit numa
1956 * locality expectations.
1957 */
1958static inline int node_match(struct kmem_cache_cpu *c, int node)
1959{
1960#ifdef CONFIG_NUMA
1961 if (node != NUMA_NO_NODE && c->node != node)
1962 return 0;
1963#endif
1964 return 1;
1965}
1966
1967static int count_free(struct page *page)
1968{
1969 return page->objects - page->inuse;
1970}
1971
1972static unsigned long count_partial(struct kmem_cache_node *n,
1973 int (*get_count)(struct page *))
1974{
1975 unsigned long flags;
1976 unsigned long x = 0;
1977 struct page *page;
1978
1979 spin_lock_irqsave(&n->list_lock, flags);
1980 list_for_each_entry(page, &n->partial, lru)
1981 x += get_count(page);
1982 spin_unlock_irqrestore(&n->list_lock, flags);
1983 return x;
1984}
1985
1986static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1987{
1988#ifdef CONFIG_SLUB_DEBUG
1989 return atomic_long_read(&n->total_objects);
1990#else
1991 return 0;
1992#endif
1993}
1994
1995static noinline void
1996slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1997{
1998 int node;
1999
2000 printk(KERN_WARNING
2001 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2002 nid, gfpflags);
2003 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2004 "default order: %d, min order: %d\n", s->name, s->objsize,
2005 s->size, oo_order(s->oo), oo_order(s->min));
2006
2007 if (oo_order(s->min) > get_order(s->objsize))
2008 printk(KERN_WARNING " %s debugging increased min order, use "
2009 "slub_debug=O to disable.\n", s->name);
2010
2011 for_each_online_node(node) {
2012 struct kmem_cache_node *n = get_node(s, node);
2013 unsigned long nr_slabs;
2014 unsigned long nr_objs;
2015 unsigned long nr_free;
2016
2017 if (!n)
2018 continue;
2019
2020 nr_free = count_partial(n, count_free);
2021 nr_slabs = node_nr_slabs(n);
2022 nr_objs = node_nr_objs(n);
2023
2024 printk(KERN_WARNING
2025 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2026 node, nr_slabs, nr_objs, nr_free);
2027 }
2028}
2029
2030/*
2031 * Slow path. The lockless freelist is empty or we need to perform
2032 * debugging duties.
2033 *
2034 * Interrupts are disabled.
2035 *
2036 * Processing is still very fast if new objects have been freed to the
2037 * regular freelist. In that case we simply take over the regular freelist
2038 * as the lockless freelist and zap the regular freelist.
2039 *
2040 * If that is not working then we fall back to the partial lists. We take the
2041 * first element of the freelist as the object to allocate now and move the
2042 * rest of the freelist to the lockless freelist.
2043 *
2044 * And if we were unable to get a new slab from the partial slab lists then
2045 * we need to allocate a new slab. This is the slowest path since it involves
2046 * a call to the page allocator and the setup of a new slab.
2047 */
2048static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2049 unsigned long addr, struct kmem_cache_cpu *c)
2050{
2051 void **object;
2052 struct page *page;
2053 unsigned long flags;
2054 struct page new;
2055 unsigned long counters;
2056
2057 local_irq_save(flags);
2058#ifdef CONFIG_PREEMPT
2059 /*
2060 * We may have been preempted and rescheduled on a different
2061 * cpu before disabling interrupts. Need to reload cpu area
2062 * pointer.
2063 */
2064 c = this_cpu_ptr(s->cpu_slab);
2065#endif
2066
2067 /* We handle __GFP_ZERO in the caller */
2068 gfpflags &= ~__GFP_ZERO;
2069
2070 page = c->page;
2071 if (!page)
2072 goto new_slab;
2073
2074 if (unlikely(!node_match(c, node))) {
2075 stat(s, ALLOC_NODE_MISMATCH);
2076 deactivate_slab(s, c);
2077 goto new_slab;
2078 }
2079
2080 stat(s, ALLOC_SLOWPATH);
2081
2082 do {
2083 object = page->freelist;
2084 counters = page->counters;
2085 new.counters = counters;
2086 VM_BUG_ON(!new.frozen);
2087
2088 /*
2089 * If there is no object left then we use this loop to
2090 * deactivate the slab which is simple since no objects
2091 * are left in the slab and therefore we do not need to
2092 * put the page back onto the partial list.
2093 *
2094 * If there are objects left then we retrieve them
2095 * and use them to refill the per cpu queue.
2096 */
2097
2098 new.inuse = page->objects;
2099 new.frozen = object != NULL;
2100
2101 } while (!__cmpxchg_double_slab(s, page,
2102 object, counters,
2103 NULL, new.counters,
2104 "__slab_alloc"));
2105
2106 if (unlikely(!object)) {
2107 c->page = NULL;
2108 stat(s, DEACTIVATE_BYPASS);
2109 goto new_slab;
2110 }
2111
2112 stat(s, ALLOC_REFILL);
2113
2114load_freelist:
2115 VM_BUG_ON(!page->frozen);
2116 c->freelist = get_freepointer(s, object);
2117 c->tid = next_tid(c->tid);
2118 local_irq_restore(flags);
2119 return object;
2120
2121new_slab:
2122 page = get_partial(s, gfpflags, node);
2123 if (page) {
2124 stat(s, ALLOC_FROM_PARTIAL);
2125 object = c->freelist;
2126
2127 if (kmem_cache_debug(s))
2128 goto debug;
2129 goto load_freelist;
2130 }
2131
2132 page = new_slab(s, gfpflags, node);
2133
2134 if (page) {
2135 c = __this_cpu_ptr(s->cpu_slab);
2136 if (c->page)
2137 flush_slab(s, c);
2138
2139 /*
2140 * No other reference to the page yet so we can
2141 * muck around with it freely without cmpxchg
2142 */
2143 object = page->freelist;
2144 page->freelist = NULL;
2145 page->inuse = page->objects;
2146
2147 stat(s, ALLOC_SLAB);
2148 c->node = page_to_nid(page);
2149 c->page = page;
2150
2151 if (kmem_cache_debug(s))
2152 goto debug;
2153 goto load_freelist;
2154 }
2155 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2156 slab_out_of_memory(s, gfpflags, node);
2157 local_irq_restore(flags);
2158 return NULL;
2159
2160debug:
2161 if (!object || !alloc_debug_processing(s, page, object, addr))
2162 goto new_slab;
2163
2164 c->freelist = get_freepointer(s, object);
2165 deactivate_slab(s, c);
2166 c->page = NULL;
2167 c->node = NUMA_NO_NODE;
2168 local_irq_restore(flags);
2169 return object;
2170}
2171
2172/*
2173 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2174 * have the fastpath folded into their functions. So no function call
2175 * overhead for requests that can be satisfied on the fastpath.
2176 *
2177 * The fastpath works by first checking if the lockless freelist can be used.
2178 * If not then __slab_alloc is called for slow processing.
2179 *
2180 * Otherwise we can simply pick the next object from the lockless free list.
2181 */
2182static __always_inline void *slab_alloc(struct kmem_cache *s,
2183 gfp_t gfpflags, int node, unsigned long addr)
2184{
2185 void **object;
2186 struct kmem_cache_cpu *c;
2187 unsigned long tid;
2188
2189 if (slab_pre_alloc_hook(s, gfpflags))
2190 return NULL;
2191
2192redo:
2193
2194 /*
2195 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2196 * enabled. We may switch back and forth between cpus while
2197 * reading from one cpu area. That does not matter as long
2198 * as we end up on the original cpu again when doing the cmpxchg.
2199 */
2200 c = __this_cpu_ptr(s->cpu_slab);
2201
2202 /*
2203 * The transaction ids are globally unique per cpu and per operation on
2204 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2205 * occurs on the right processor and that there was no operation on the
2206 * linked list in between.
2207 */
2208 tid = c->tid;
2209 barrier();
2210
2211 object = c->freelist;
2212 if (unlikely(!object || !node_match(c, node)))
2213
2214 object = __slab_alloc(s, gfpflags, node, addr, c);
2215
2216 else {
2217 /*
2218 * The cmpxchg will only match if there was no additional
2219 * operation and if we are on the right processor.
2220 *
2221 * The cmpxchg does the following atomically (without lock semantics!)
2222 * 1. Relocate first pointer to the current per cpu area.
2223 * 2. Verify that tid and freelist have not been changed
2224 * 3. If they were not changed replace tid and freelist
2225 *
2226 * Since this is without lock semantics the protection is only against
2227 * code executing on this cpu *not* from access by other cpus.
2228 */
2229 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2230 s->cpu_slab->freelist, s->cpu_slab->tid,
2231 object, tid,
2232 get_freepointer_safe(s, object), next_tid(tid)))) {
2233
2234 note_cmpxchg_failure("slab_alloc", s, tid);
2235 goto redo;
2236 }
2237 stat(s, ALLOC_FASTPATH);
2238 }
2239
2240 if (unlikely(gfpflags & __GFP_ZERO) && object)
2241 memset(object, 0, s->objsize);
2242
2243 slab_post_alloc_hook(s, gfpflags, object);
2244
2245 return object;
2246}
2247
2248void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2249{
2250 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2251
2252 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2253
2254 return ret;
2255}
2256EXPORT_SYMBOL(kmem_cache_alloc);
2257
2258#ifdef CONFIG_TRACING
2259void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2260{
2261 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2262 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2263 return ret;
2264}
2265EXPORT_SYMBOL(kmem_cache_alloc_trace);
2266
2267void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2268{
2269 void *ret = kmalloc_order(size, flags, order);
2270 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2271 return ret;
2272}
2273EXPORT_SYMBOL(kmalloc_order_trace);
2274#endif
2275
2276#ifdef CONFIG_NUMA
2277void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2278{
2279 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2280
2281 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2282 s->objsize, s->size, gfpflags, node);
2283
2284 return ret;
2285}
2286EXPORT_SYMBOL(kmem_cache_alloc_node);
2287
2288#ifdef CONFIG_TRACING
2289void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2290 gfp_t gfpflags,
2291 int node, size_t size)
2292{
2293 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2294
2295 trace_kmalloc_node(_RET_IP_, ret,
2296 size, s->size, gfpflags, node);
2297 return ret;
2298}
2299EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2300#endif
2301#endif
2302
2303/*
2304 * Slow patch handling. This may still be called frequently since objects
2305 * have a longer lifetime than the cpu slabs in most processing loads.
2306 *
2307 * So we still attempt to reduce cache line usage. Just take the slab
2308 * lock and free the item. If there is no additional partial page
2309 * handling required then we can return immediately.
2310 */
2311static void __slab_free(struct kmem_cache *s, struct page *page,
2312 void *x, unsigned long addr)
2313{
2314 void *prior;
2315 void **object = (void *)x;
2316 int was_frozen;
2317 int inuse;
2318 struct page new;
2319 unsigned long counters;
2320 struct kmem_cache_node *n = NULL;
2321 unsigned long uninitialized_var(flags);
2322
2323 stat(s, FREE_SLOWPATH);
2324
2325 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2326 return;
2327
2328 do {
2329 prior = page->freelist;
2330 counters = page->counters;
2331 set_freepointer(s, object, prior);
2332 new.counters = counters;
2333 was_frozen = new.frozen;
2334 new.inuse--;
2335 if ((!new.inuse || !prior) && !was_frozen && !n) {
2336 n = get_node(s, page_to_nid(page));
2337 /*
2338 * Speculatively acquire the list_lock.
2339 * If the cmpxchg does not succeed then we may
2340 * drop the list_lock without any processing.
2341 *
2342 * Otherwise the list_lock will synchronize with
2343 * other processors updating the list of slabs.
2344 */
2345 spin_lock_irqsave(&n->list_lock, flags);
2346 }
2347 inuse = new.inuse;
2348
2349 } while (!cmpxchg_double_slab(s, page,
2350 prior, counters,
2351 object, new.counters,
2352 "__slab_free"));
2353
2354 if (likely(!n)) {
2355 /*
2356 * The list lock was not taken therefore no list
2357 * activity can be necessary.
2358 */
2359 if (was_frozen)
2360 stat(s, FREE_FROZEN);
2361 return;
2362 }
2363
2364 /*
2365 * was_frozen may have been set after we acquired the list_lock in
2366 * an earlier loop. So we need to check it here again.
2367 */
2368 if (was_frozen)
2369 stat(s, FREE_FROZEN);
2370 else {
2371 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2372 goto slab_empty;
2373
2374 /*
2375 * Objects left in the slab. If it was not on the partial list before
2376 * then add it.
2377 */
2378 if (unlikely(!prior)) {
2379 remove_full(s, page);
2380 add_partial(n, page, 1);
2381 stat(s, FREE_ADD_PARTIAL);
2382 }
2383 }
2384 spin_unlock_irqrestore(&n->list_lock, flags);
2385 return;
2386
2387slab_empty:
2388 if (prior) {
2389 /*
2390 * Slab on the partial list.
2391 */
2392 remove_partial(n, page);
2393 stat(s, FREE_REMOVE_PARTIAL);
2394 } else
2395 /* Slab must be on the full list */
2396 remove_full(s, page);
2397
2398 spin_unlock_irqrestore(&n->list_lock, flags);
2399 stat(s, FREE_SLAB);
2400 discard_slab(s, page);
2401}
2402
2403/*
2404 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2405 * can perform fastpath freeing without additional function calls.
2406 *
2407 * The fastpath is only possible if we are freeing to the current cpu slab
2408 * of this processor. This typically the case if we have just allocated
2409 * the item before.
2410 *
2411 * If fastpath is not possible then fall back to __slab_free where we deal
2412 * with all sorts of special processing.
2413 */
2414static __always_inline void slab_free(struct kmem_cache *s,
2415 struct page *page, void *x, unsigned long addr)
2416{
2417 void **object = (void *)x;
2418 struct kmem_cache_cpu *c;
2419 unsigned long tid;
2420
2421 slab_free_hook(s, x);
2422
2423redo:
2424
2425 /*
2426 * Determine the currently cpus per cpu slab.
2427 * The cpu may change afterward. However that does not matter since
2428 * data is retrieved via this pointer. If we are on the same cpu
2429 * during the cmpxchg then the free will succedd.
2430 */
2431 c = __this_cpu_ptr(s->cpu_slab);
2432
2433 tid = c->tid;
2434 barrier();
2435
2436 if (likely(page == c->page)) {
2437 set_freepointer(s, object, c->freelist);
2438
2439 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2440 s->cpu_slab->freelist, s->cpu_slab->tid,
2441 c->freelist, tid,
2442 object, next_tid(tid)))) {
2443
2444 note_cmpxchg_failure("slab_free", s, tid);
2445 goto redo;
2446 }
2447 stat(s, FREE_FASTPATH);
2448 } else
2449 __slab_free(s, page, x, addr);
2450
2451}
2452
2453void kmem_cache_free(struct kmem_cache *s, void *x)
2454{
2455 struct page *page;
2456
2457 page = virt_to_head_page(x);
2458
2459 slab_free(s, page, x, _RET_IP_);
2460
2461 trace_kmem_cache_free(_RET_IP_, x);
2462}
2463EXPORT_SYMBOL(kmem_cache_free);
2464
2465/*
2466 * Object placement in a slab is made very easy because we always start at
2467 * offset 0. If we tune the size of the object to the alignment then we can
2468 * get the required alignment by putting one properly sized object after
2469 * another.
2470 *
2471 * Notice that the allocation order determines the sizes of the per cpu
2472 * caches. Each processor has always one slab available for allocations.
2473 * Increasing the allocation order reduces the number of times that slabs
2474 * must be moved on and off the partial lists and is therefore a factor in
2475 * locking overhead.
2476 */
2477
2478/*
2479 * Mininum / Maximum order of slab pages. This influences locking overhead
2480 * and slab fragmentation. A higher order reduces the number of partial slabs
2481 * and increases the number of allocations possible without having to
2482 * take the list_lock.
2483 */
2484static int slub_min_order;
2485static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2486static int slub_min_objects;
2487
2488/*
2489 * Merge control. If this is set then no merging of slab caches will occur.
2490 * (Could be removed. This was introduced to pacify the merge skeptics.)
2491 */
2492static int slub_nomerge;
2493
2494/*
2495 * Calculate the order of allocation given an slab object size.
2496 *
2497 * The order of allocation has significant impact on performance and other
2498 * system components. Generally order 0 allocations should be preferred since
2499 * order 0 does not cause fragmentation in the page allocator. Larger objects
2500 * be problematic to put into order 0 slabs because there may be too much
2501 * unused space left. We go to a higher order if more than 1/16th of the slab
2502 * would be wasted.
2503 *
2504 * In order to reach satisfactory performance we must ensure that a minimum
2505 * number of objects is in one slab. Otherwise we may generate too much
2506 * activity on the partial lists which requires taking the list_lock. This is
2507 * less a concern for large slabs though which are rarely used.
2508 *
2509 * slub_max_order specifies the order where we begin to stop considering the
2510 * number of objects in a slab as critical. If we reach slub_max_order then
2511 * we try to keep the page order as low as possible. So we accept more waste
2512 * of space in favor of a small page order.
2513 *
2514 * Higher order allocations also allow the placement of more objects in a
2515 * slab and thereby reduce object handling overhead. If the user has
2516 * requested a higher mininum order then we start with that one instead of
2517 * the smallest order which will fit the object.
2518 */
2519static inline int slab_order(int size, int min_objects,
2520 int max_order, int fract_leftover, int reserved)
2521{
2522 int order;
2523 int rem;
2524 int min_order = slub_min_order;
2525
2526 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2527 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2528
2529 for (order = max(min_order,
2530 fls(min_objects * size - 1) - PAGE_SHIFT);
2531 order <= max_order; order++) {
2532
2533 unsigned long slab_size = PAGE_SIZE << order;
2534
2535 if (slab_size < min_objects * size + reserved)
2536 continue;
2537
2538 rem = (slab_size - reserved) % size;
2539
2540 if (rem <= slab_size / fract_leftover)
2541 break;
2542
2543 }
2544
2545 return order;
2546}
2547
2548static inline int calculate_order(int size, int reserved)
2549{
2550 int order;
2551 int min_objects;
2552 int fraction;
2553 int max_objects;
2554
2555 /*
2556 * Attempt to find best configuration for a slab. This
2557 * works by first attempting to generate a layout with
2558 * the best configuration and backing off gradually.
2559 *
2560 * First we reduce the acceptable waste in a slab. Then
2561 * we reduce the minimum objects required in a slab.
2562 */
2563 min_objects = slub_min_objects;
2564 if (!min_objects)
2565 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2566 max_objects = order_objects(slub_max_order, size, reserved);
2567 min_objects = min(min_objects, max_objects);
2568
2569 while (min_objects > 1) {
2570 fraction = 16;
2571 while (fraction >= 4) {
2572 order = slab_order(size, min_objects,
2573 slub_max_order, fraction, reserved);
2574 if (order <= slub_max_order)
2575 return order;
2576 fraction /= 2;
2577 }
2578 min_objects--;
2579 }
2580
2581 /*
2582 * We were unable to place multiple objects in a slab. Now
2583 * lets see if we can place a single object there.
2584 */
2585 order = slab_order(size, 1, slub_max_order, 1, reserved);
2586 if (order <= slub_max_order)
2587 return order;
2588
2589 /*
2590 * Doh this slab cannot be placed using slub_max_order.
2591 */
2592 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2593 if (order < MAX_ORDER)
2594 return order;
2595 return -ENOSYS;
2596}
2597
2598/*
2599 * Figure out what the alignment of the objects will be.
2600 */
2601static unsigned long calculate_alignment(unsigned long flags,
2602 unsigned long align, unsigned long size)
2603{
2604 /*
2605 * If the user wants hardware cache aligned objects then follow that
2606 * suggestion if the object is sufficiently large.
2607 *
2608 * The hardware cache alignment cannot override the specified
2609 * alignment though. If that is greater then use it.
2610 */
2611 if (flags & SLAB_HWCACHE_ALIGN) {
2612 unsigned long ralign = cache_line_size();
2613 while (size <= ralign / 2)
2614 ralign /= 2;
2615 align = max(align, ralign);
2616 }
2617
2618 if (align < ARCH_SLAB_MINALIGN)
2619 align = ARCH_SLAB_MINALIGN;
2620
2621 return ALIGN(align, sizeof(void *));
2622}
2623
2624static void
2625init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2626{
2627 n->nr_partial = 0;
2628 spin_lock_init(&n->list_lock);
2629 INIT_LIST_HEAD(&n->partial);
2630#ifdef CONFIG_SLUB_DEBUG
2631 atomic_long_set(&n->nr_slabs, 0);
2632 atomic_long_set(&n->total_objects, 0);
2633 INIT_LIST_HEAD(&n->full);
2634#endif
2635}
2636
2637static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2638{
2639 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2640 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2641
2642 /*
2643 * Must align to double word boundary for the double cmpxchg
2644 * instructions to work; see __pcpu_double_call_return_bool().
2645 */
2646 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2647 2 * sizeof(void *));
2648
2649 if (!s->cpu_slab)
2650 return 0;
2651
2652 init_kmem_cache_cpus(s);
2653
2654 return 1;
2655}
2656
2657static struct kmem_cache *kmem_cache_node;
2658
2659/*
2660 * No kmalloc_node yet so do it by hand. We know that this is the first
2661 * slab on the node for this slabcache. There are no concurrent accesses
2662 * possible.
2663 *
2664 * Note that this function only works on the kmalloc_node_cache
2665 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2666 * memory on a fresh node that has no slab structures yet.
2667 */
2668static void early_kmem_cache_node_alloc(int node)
2669{
2670 struct page *page;
2671 struct kmem_cache_node *n;
2672
2673 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2674
2675 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2676
2677 BUG_ON(!page);
2678 if (page_to_nid(page) != node) {
2679 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2680 "node %d\n", node);
2681 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2682 "in order to be able to continue\n");
2683 }
2684
2685 n = page->freelist;
2686 BUG_ON(!n);
2687 page->freelist = get_freepointer(kmem_cache_node, n);
2688 page->inuse++;
2689 page->frozen = 0;
2690 kmem_cache_node->node[node] = n;
2691#ifdef CONFIG_SLUB_DEBUG
2692 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2693 init_tracking(kmem_cache_node, n);
2694#endif
2695 init_kmem_cache_node(n, kmem_cache_node);
2696 inc_slabs_node(kmem_cache_node, node, page->objects);
2697
2698 add_partial(n, page, 0);
2699}
2700
2701static void free_kmem_cache_nodes(struct kmem_cache *s)
2702{
2703 int node;
2704
2705 for_each_node_state(node, N_NORMAL_MEMORY) {
2706 struct kmem_cache_node *n = s->node[node];
2707
2708 if (n)
2709 kmem_cache_free(kmem_cache_node, n);
2710
2711 s->node[node] = NULL;
2712 }
2713}
2714
2715static int init_kmem_cache_nodes(struct kmem_cache *s)
2716{
2717 int node;
2718
2719 for_each_node_state(node, N_NORMAL_MEMORY) {
2720 struct kmem_cache_node *n;
2721
2722 if (slab_state == DOWN) {
2723 early_kmem_cache_node_alloc(node);
2724 continue;
2725 }
2726 n = kmem_cache_alloc_node(kmem_cache_node,
2727 GFP_KERNEL, node);
2728
2729 if (!n) {
2730 free_kmem_cache_nodes(s);
2731 return 0;
2732 }
2733
2734 s->node[node] = n;
2735 init_kmem_cache_node(n, s);
2736 }
2737 return 1;
2738}
2739
2740static void set_min_partial(struct kmem_cache *s, unsigned long min)
2741{
2742 if (min < MIN_PARTIAL)
2743 min = MIN_PARTIAL;
2744 else if (min > MAX_PARTIAL)
2745 min = MAX_PARTIAL;
2746 s->min_partial = min;
2747}
2748
2749/*
2750 * calculate_sizes() determines the order and the distribution of data within
2751 * a slab object.
2752 */
2753static int calculate_sizes(struct kmem_cache *s, int forced_order)
2754{
2755 unsigned long flags = s->flags;
2756 unsigned long size = s->objsize;
2757 unsigned long align = s->align;
2758 int order;
2759
2760 /*
2761 * Round up object size to the next word boundary. We can only
2762 * place the free pointer at word boundaries and this determines
2763 * the possible location of the free pointer.
2764 */
2765 size = ALIGN(size, sizeof(void *));
2766
2767#ifdef CONFIG_SLUB_DEBUG
2768 /*
2769 * Determine if we can poison the object itself. If the user of
2770 * the slab may touch the object after free or before allocation
2771 * then we should never poison the object itself.
2772 */
2773 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2774 !s->ctor)
2775 s->flags |= __OBJECT_POISON;
2776 else
2777 s->flags &= ~__OBJECT_POISON;
2778
2779
2780 /*
2781 * If we are Redzoning then check if there is some space between the
2782 * end of the object and the free pointer. If not then add an
2783 * additional word to have some bytes to store Redzone information.
2784 */
2785 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2786 size += sizeof(void *);
2787#endif
2788
2789 /*
2790 * With that we have determined the number of bytes in actual use
2791 * by the object. This is the potential offset to the free pointer.
2792 */
2793 s->inuse = size;
2794
2795 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2796 s->ctor)) {
2797 /*
2798 * Relocate free pointer after the object if it is not
2799 * permitted to overwrite the first word of the object on
2800 * kmem_cache_free.
2801 *
2802 * This is the case if we do RCU, have a constructor or
2803 * destructor or are poisoning the objects.
2804 */
2805 s->offset = size;
2806 size += sizeof(void *);
2807 }
2808
2809#ifdef CONFIG_SLUB_DEBUG
2810 if (flags & SLAB_STORE_USER)
2811 /*
2812 * Need to store information about allocs and frees after
2813 * the object.
2814 */
2815 size += 2 * sizeof(struct track);
2816
2817 if (flags & SLAB_RED_ZONE)
2818 /*
2819 * Add some empty padding so that we can catch
2820 * overwrites from earlier objects rather than let
2821 * tracking information or the free pointer be
2822 * corrupted if a user writes before the start
2823 * of the object.
2824 */
2825 size += sizeof(void *);
2826#endif
2827
2828 /*
2829 * Determine the alignment based on various parameters that the
2830 * user specified and the dynamic determination of cache line size
2831 * on bootup.
2832 */
2833 align = calculate_alignment(flags, align, s->objsize);
2834 s->align = align;
2835
2836 /*
2837 * SLUB stores one object immediately after another beginning from
2838 * offset 0. In order to align the objects we have to simply size
2839 * each object to conform to the alignment.
2840 */
2841 size = ALIGN(size, align);
2842 s->size = size;
2843 if (forced_order >= 0)
2844 order = forced_order;
2845 else
2846 order = calculate_order(size, s->reserved);
2847
2848 if (order < 0)
2849 return 0;
2850
2851 s->allocflags = 0;
2852 if (order)
2853 s->allocflags |= __GFP_COMP;
2854
2855 if (s->flags & SLAB_CACHE_DMA)
2856 s->allocflags |= SLUB_DMA;
2857
2858 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2859 s->allocflags |= __GFP_RECLAIMABLE;
2860
2861 /*
2862 * Determine the number of objects per slab
2863 */
2864 s->oo = oo_make(order, size, s->reserved);
2865 s->min = oo_make(get_order(size), size, s->reserved);
2866 if (oo_objects(s->oo) > oo_objects(s->max))
2867 s->max = s->oo;
2868
2869 return !!oo_objects(s->oo);
2870
2871}
2872
2873static int kmem_cache_open(struct kmem_cache *s,
2874 const char *name, size_t size,
2875 size_t align, unsigned long flags,
2876 void (*ctor)(void *))
2877{
2878 memset(s, 0, kmem_size);
2879 s->name = name;
2880 s->ctor = ctor;
2881 s->objsize = size;
2882 s->align = align;
2883 s->flags = kmem_cache_flags(size, flags, name, ctor);
2884 s->reserved = 0;
2885
2886 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2887 s->reserved = sizeof(struct rcu_head);
2888
2889 if (!calculate_sizes(s, -1))
2890 goto error;
2891 if (disable_higher_order_debug) {
2892 /*
2893 * Disable debugging flags that store metadata if the min slab
2894 * order increased.
2895 */
2896 if (get_order(s->size) > get_order(s->objsize)) {
2897 s->flags &= ~DEBUG_METADATA_FLAGS;
2898 s->offset = 0;
2899 if (!calculate_sizes(s, -1))
2900 goto error;
2901 }
2902 }
2903
2904#ifdef CONFIG_CMPXCHG_DOUBLE
2905 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
2906 /* Enable fast mode */
2907 s->flags |= __CMPXCHG_DOUBLE;
2908#endif
2909
2910 /*
2911 * The larger the object size is, the more pages we want on the partial
2912 * list to avoid pounding the page allocator excessively.
2913 */
2914 set_min_partial(s, ilog2(s->size));
2915 s->refcount = 1;
2916#ifdef CONFIG_NUMA
2917 s->remote_node_defrag_ratio = 1000;
2918#endif
2919 if (!init_kmem_cache_nodes(s))
2920 goto error;
2921
2922 if (alloc_kmem_cache_cpus(s))
2923 return 1;
2924
2925 free_kmem_cache_nodes(s);
2926error:
2927 if (flags & SLAB_PANIC)
2928 panic("Cannot create slab %s size=%lu realsize=%u "
2929 "order=%u offset=%u flags=%lx\n",
2930 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2931 s->offset, flags);
2932 return 0;
2933}
2934
2935/*
2936 * Determine the size of a slab object
2937 */
2938unsigned int kmem_cache_size(struct kmem_cache *s)
2939{
2940 return s->objsize;
2941}
2942EXPORT_SYMBOL(kmem_cache_size);
2943
2944static void list_slab_objects(struct kmem_cache *s, struct page *page,
2945 const char *text)
2946{
2947#ifdef CONFIG_SLUB_DEBUG
2948 void *addr = page_address(page);
2949 void *p;
2950 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2951 sizeof(long), GFP_ATOMIC);
2952 if (!map)
2953 return;
2954 slab_err(s, page, "%s", text);
2955 slab_lock(page);
2956
2957 get_map(s, page, map);
2958 for_each_object(p, s, addr, page->objects) {
2959
2960 if (!test_bit(slab_index(p, s, addr), map)) {
2961 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2962 p, p - addr);
2963 print_tracking(s, p);
2964 }
2965 }
2966 slab_unlock(page);
2967 kfree(map);
2968#endif
2969}
2970
2971/*
2972 * Attempt to free all partial slabs on a node.
2973 */
2974static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2975{
2976 unsigned long flags;
2977 struct page *page, *h;
2978
2979 spin_lock_irqsave(&n->list_lock, flags);
2980 list_for_each_entry_safe(page, h, &n->partial, lru) {
2981 if (!page->inuse) {
2982 remove_partial(n, page);
2983 discard_slab(s, page);
2984 } else {
2985 list_slab_objects(s, page,
2986 "Objects remaining on kmem_cache_close()");
2987 }
2988 }
2989 spin_unlock_irqrestore(&n->list_lock, flags);
2990}
2991
2992/*
2993 * Release all resources used by a slab cache.
2994 */
2995static inline int kmem_cache_close(struct kmem_cache *s)
2996{
2997 int node;
2998
2999 flush_all(s);
3000 free_percpu(s->cpu_slab);
3001 /* Attempt to free all objects */
3002 for_each_node_state(node, N_NORMAL_MEMORY) {
3003 struct kmem_cache_node *n = get_node(s, node);
3004
3005 free_partial(s, n);
3006 if (n->nr_partial || slabs_node(s, node))
3007 return 1;
3008 }
3009 free_kmem_cache_nodes(s);
3010 return 0;
3011}
3012
3013/*
3014 * Close a cache and release the kmem_cache structure
3015 * (must be used for caches created using kmem_cache_create)
3016 */
3017void kmem_cache_destroy(struct kmem_cache *s)
3018{
3019 down_write(&slub_lock);
3020 s->refcount--;
3021 if (!s->refcount) {
3022 list_del(&s->list);
3023 if (kmem_cache_close(s)) {
3024 printk(KERN_ERR "SLUB %s: %s called for cache that "
3025 "still has objects.\n", s->name, __func__);
3026 dump_stack();
3027 }
3028 if (s->flags & SLAB_DESTROY_BY_RCU)
3029 rcu_barrier();
3030 sysfs_slab_remove(s);
3031 }
3032 up_write(&slub_lock);
3033}
3034EXPORT_SYMBOL(kmem_cache_destroy);
3035
3036/********************************************************************
3037 * Kmalloc subsystem
3038 *******************************************************************/
3039
3040struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3041EXPORT_SYMBOL(kmalloc_caches);
3042
3043static struct kmem_cache *kmem_cache;
3044
3045#ifdef CONFIG_ZONE_DMA
3046static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3047#endif
3048
3049static int __init setup_slub_min_order(char *str)
3050{
3051 get_option(&str, &slub_min_order);
3052
3053 return 1;
3054}
3055
3056__setup("slub_min_order=", setup_slub_min_order);
3057
3058static int __init setup_slub_max_order(char *str)
3059{
3060 get_option(&str, &slub_max_order);
3061 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3062
3063 return 1;
3064}
3065
3066__setup("slub_max_order=", setup_slub_max_order);
3067
3068static int __init setup_slub_min_objects(char *str)
3069{
3070 get_option(&str, &slub_min_objects);
3071
3072 return 1;
3073}
3074
3075__setup("slub_min_objects=", setup_slub_min_objects);
3076
3077static int __init setup_slub_nomerge(char *str)
3078{
3079 slub_nomerge = 1;
3080 return 1;
3081}
3082
3083__setup("slub_nomerge", setup_slub_nomerge);
3084
3085static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3086 int size, unsigned int flags)
3087{
3088 struct kmem_cache *s;
3089
3090 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3091
3092 /*
3093 * This function is called with IRQs disabled during early-boot on
3094 * single CPU so there's no need to take slub_lock here.
3095 */
3096 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3097 flags, NULL))
3098 goto panic;
3099
3100 list_add(&s->list, &slab_caches);
3101 return s;
3102
3103panic:
3104 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3105 return NULL;
3106}
3107
3108/*
3109 * Conversion table for small slabs sizes / 8 to the index in the
3110 * kmalloc array. This is necessary for slabs < 192 since we have non power
3111 * of two cache sizes there. The size of larger slabs can be determined using
3112 * fls.
3113 */
3114static s8 size_index[24] = {
3115 3, /* 8 */
3116 4, /* 16 */
3117 5, /* 24 */
3118 5, /* 32 */
3119 6, /* 40 */
3120 6, /* 48 */
3121 6, /* 56 */
3122 6, /* 64 */
3123 1, /* 72 */
3124 1, /* 80 */
3125 1, /* 88 */
3126 1, /* 96 */
3127 7, /* 104 */
3128 7, /* 112 */
3129 7, /* 120 */
3130 7, /* 128 */
3131 2, /* 136 */
3132 2, /* 144 */
3133 2, /* 152 */
3134 2, /* 160 */
3135 2, /* 168 */
3136 2, /* 176 */
3137 2, /* 184 */
3138 2 /* 192 */
3139};
3140
3141static inline int size_index_elem(size_t bytes)
3142{
3143 return (bytes - 1) / 8;
3144}
3145
3146static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3147{
3148 int index;
3149
3150 if (size <= 192) {
3151 if (!size)
3152 return ZERO_SIZE_PTR;
3153
3154 index = size_index[size_index_elem(size)];
3155 } else
3156 index = fls(size - 1);
3157
3158#ifdef CONFIG_ZONE_DMA
3159 if (unlikely((flags & SLUB_DMA)))
3160 return kmalloc_dma_caches[index];
3161
3162#endif
3163 return kmalloc_caches[index];
3164}
3165
3166void *__kmalloc(size_t size, gfp_t flags)
3167{
3168 struct kmem_cache *s;
3169 void *ret;
3170
3171 if (unlikely(size > SLUB_MAX_SIZE))
3172 return kmalloc_large(size, flags);
3173
3174 s = get_slab(size, flags);
3175
3176 if (unlikely(ZERO_OR_NULL_PTR(s)))
3177 return s;
3178
3179 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3180
3181 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3182
3183 return ret;
3184}
3185EXPORT_SYMBOL(__kmalloc);
3186
3187#ifdef CONFIG_NUMA
3188static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3189{
3190 struct page *page;
3191 void *ptr = NULL;
3192
3193 flags |= __GFP_COMP | __GFP_NOTRACK;
3194 page = alloc_pages_node(node, flags, get_order(size));
3195 if (page)
3196 ptr = page_address(page);
3197
3198 kmemleak_alloc(ptr, size, 1, flags);
3199 return ptr;
3200}
3201
3202void *__kmalloc_node(size_t size, gfp_t flags, int node)
3203{
3204 struct kmem_cache *s;
3205 void *ret;
3206
3207 if (unlikely(size > SLUB_MAX_SIZE)) {
3208 ret = kmalloc_large_node(size, flags, node);
3209
3210 trace_kmalloc_node(_RET_IP_, ret,
3211 size, PAGE_SIZE << get_order(size),
3212 flags, node);
3213
3214 return ret;
3215 }
3216
3217 s = get_slab(size, flags);
3218
3219 if (unlikely(ZERO_OR_NULL_PTR(s)))
3220 return s;
3221
3222 ret = slab_alloc(s, flags, node, _RET_IP_);
3223
3224 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3225
3226 return ret;
3227}
3228EXPORT_SYMBOL(__kmalloc_node);
3229#endif
3230
3231size_t ksize(const void *object)
3232{
3233 struct page *page;
3234
3235 if (unlikely(object == ZERO_SIZE_PTR))
3236 return 0;
3237
3238 page = virt_to_head_page(object);
3239
3240 if (unlikely(!PageSlab(page))) {
3241 WARN_ON(!PageCompound(page));
3242 return PAGE_SIZE << compound_order(page);
3243 }
3244
3245 return slab_ksize(page->slab);
3246}
3247EXPORT_SYMBOL(ksize);
3248
3249#ifdef CONFIG_SLUB_DEBUG
3250bool verify_mem_not_deleted(const void *x)
3251{
3252 struct page *page;
3253 void *object = (void *)x;
3254 unsigned long flags;
3255 bool rv;
3256
3257 if (unlikely(ZERO_OR_NULL_PTR(x)))
3258 return false;
3259
3260 local_irq_save(flags);
3261
3262 page = virt_to_head_page(x);
3263 if (unlikely(!PageSlab(page))) {
3264 /* maybe it was from stack? */
3265 rv = true;
3266 goto out_unlock;
3267 }
3268
3269 slab_lock(page);
3270 if (on_freelist(page->slab, page, object)) {
3271 object_err(page->slab, page, object, "Object is on free-list");
3272 rv = false;
3273 } else {
3274 rv = true;
3275 }
3276 slab_unlock(page);
3277
3278out_unlock:
3279 local_irq_restore(flags);
3280 return rv;
3281}
3282EXPORT_SYMBOL(verify_mem_not_deleted);
3283#endif
3284
3285void kfree(const void *x)
3286{
3287 struct page *page;
3288 void *object = (void *)x;
3289
3290 trace_kfree(_RET_IP_, x);
3291
3292 if (unlikely(ZERO_OR_NULL_PTR(x)))
3293 return;
3294
3295 page = virt_to_head_page(x);
3296 if (unlikely(!PageSlab(page))) {
3297 BUG_ON(!PageCompound(page));
3298 kmemleak_free(x);
3299 put_page(page);
3300 return;
3301 }
3302 slab_free(page->slab, page, object, _RET_IP_);
3303}
3304EXPORT_SYMBOL(kfree);
3305
3306/*
3307 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3308 * the remaining slabs by the number of items in use. The slabs with the
3309 * most items in use come first. New allocations will then fill those up
3310 * and thus they can be removed from the partial lists.
3311 *
3312 * The slabs with the least items are placed last. This results in them
3313 * being allocated from last increasing the chance that the last objects
3314 * are freed in them.
3315 */
3316int kmem_cache_shrink(struct kmem_cache *s)
3317{
3318 int node;
3319 int i;
3320 struct kmem_cache_node *n;
3321 struct page *page;
3322 struct page *t;
3323 int objects = oo_objects(s->max);
3324 struct list_head *slabs_by_inuse =
3325 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3326 unsigned long flags;
3327
3328 if (!slabs_by_inuse)
3329 return -ENOMEM;
3330
3331 flush_all(s);
3332 for_each_node_state(node, N_NORMAL_MEMORY) {
3333 n = get_node(s, node);
3334
3335 if (!n->nr_partial)
3336 continue;
3337
3338 for (i = 0; i < objects; i++)
3339 INIT_LIST_HEAD(slabs_by_inuse + i);
3340
3341 spin_lock_irqsave(&n->list_lock, flags);
3342
3343 /*
3344 * Build lists indexed by the items in use in each slab.
3345 *
3346 * Note that concurrent frees may occur while we hold the
3347 * list_lock. page->inuse here is the upper limit.
3348 */
3349 list_for_each_entry_safe(page, t, &n->partial, lru) {
3350 if (!page->inuse) {
3351 remove_partial(n, page);
3352 discard_slab(s, page);
3353 } else {
3354 list_move(&page->lru,
3355 slabs_by_inuse + page->inuse);
3356 }
3357 }
3358
3359 /*
3360 * Rebuild the partial list with the slabs filled up most
3361 * first and the least used slabs at the end.
3362 */
3363 for (i = objects - 1; i >= 0; i--)
3364 list_splice(slabs_by_inuse + i, n->partial.prev);
3365
3366 spin_unlock_irqrestore(&n->list_lock, flags);
3367 }
3368
3369 kfree(slabs_by_inuse);
3370 return 0;
3371}
3372EXPORT_SYMBOL(kmem_cache_shrink);
3373
3374#if defined(CONFIG_MEMORY_HOTPLUG)
3375static int slab_mem_going_offline_callback(void *arg)
3376{
3377 struct kmem_cache *s;
3378
3379 down_read(&slub_lock);
3380 list_for_each_entry(s, &slab_caches, list)
3381 kmem_cache_shrink(s);
3382 up_read(&slub_lock);
3383
3384 return 0;
3385}
3386
3387static void slab_mem_offline_callback(void *arg)
3388{
3389 struct kmem_cache_node *n;
3390 struct kmem_cache *s;
3391 struct memory_notify *marg = arg;
3392 int offline_node;
3393
3394 offline_node = marg->status_change_nid;
3395
3396 /*
3397 * If the node still has available memory. we need kmem_cache_node
3398 * for it yet.
3399 */
3400 if (offline_node < 0)
3401 return;
3402
3403 down_read(&slub_lock);
3404 list_for_each_entry(s, &slab_caches, list) {
3405 n = get_node(s, offline_node);
3406 if (n) {
3407 /*
3408 * if n->nr_slabs > 0, slabs still exist on the node
3409 * that is going down. We were unable to free them,
3410 * and offline_pages() function shouldn't call this
3411 * callback. So, we must fail.
3412 */
3413 BUG_ON(slabs_node(s, offline_node));
3414
3415 s->node[offline_node] = NULL;
3416 kmem_cache_free(kmem_cache_node, n);
3417 }
3418 }
3419 up_read(&slub_lock);
3420}
3421
3422static int slab_mem_going_online_callback(void *arg)
3423{
3424 struct kmem_cache_node *n;
3425 struct kmem_cache *s;
3426 struct memory_notify *marg = arg;
3427 int nid = marg->status_change_nid;
3428 int ret = 0;
3429
3430 /*
3431 * If the node's memory is already available, then kmem_cache_node is
3432 * already created. Nothing to do.
3433 */
3434 if (nid < 0)
3435 return 0;
3436
3437 /*
3438 * We are bringing a node online. No memory is available yet. We must
3439 * allocate a kmem_cache_node structure in order to bring the node
3440 * online.
3441 */
3442 down_read(&slub_lock);
3443 list_for_each_entry(s, &slab_caches, list) {
3444 /*
3445 * XXX: kmem_cache_alloc_node will fallback to other nodes
3446 * since memory is not yet available from the node that
3447 * is brought up.
3448 */
3449 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3450 if (!n) {
3451 ret = -ENOMEM;
3452 goto out;
3453 }
3454 init_kmem_cache_node(n, s);
3455 s->node[nid] = n;
3456 }
3457out:
3458 up_read(&slub_lock);
3459 return ret;
3460}
3461
3462static int slab_memory_callback(struct notifier_block *self,
3463 unsigned long action, void *arg)
3464{
3465 int ret = 0;
3466
3467 switch (action) {
3468 case MEM_GOING_ONLINE:
3469 ret = slab_mem_going_online_callback(arg);
3470 break;
3471 case MEM_GOING_OFFLINE:
3472 ret = slab_mem_going_offline_callback(arg);
3473 break;
3474 case MEM_OFFLINE:
3475 case MEM_CANCEL_ONLINE:
3476 slab_mem_offline_callback(arg);
3477 break;
3478 case MEM_ONLINE:
3479 case MEM_CANCEL_OFFLINE:
3480 break;
3481 }
3482 if (ret)
3483 ret = notifier_from_errno(ret);
3484 else
3485 ret = NOTIFY_OK;
3486 return ret;
3487}
3488
3489#endif /* CONFIG_MEMORY_HOTPLUG */
3490
3491/********************************************************************
3492 * Basic setup of slabs
3493 *******************************************************************/
3494
3495/*
3496 * Used for early kmem_cache structures that were allocated using
3497 * the page allocator
3498 */
3499
3500static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3501{
3502 int node;
3503
3504 list_add(&s->list, &slab_caches);
3505 s->refcount = -1;
3506
3507 for_each_node_state(node, N_NORMAL_MEMORY) {
3508 struct kmem_cache_node *n = get_node(s, node);
3509 struct page *p;
3510
3511 if (n) {
3512 list_for_each_entry(p, &n->partial, lru)
3513 p->slab = s;
3514
3515#ifdef CONFIG_SLUB_DEBUG
3516 list_for_each_entry(p, &n->full, lru)
3517 p->slab = s;
3518#endif
3519 }
3520 }
3521}
3522
3523void __init kmem_cache_init(void)
3524{
3525 int i;
3526 int caches = 0;
3527 struct kmem_cache *temp_kmem_cache;
3528 int order;
3529 struct kmem_cache *temp_kmem_cache_node;
3530 unsigned long kmalloc_size;
3531
3532 kmem_size = offsetof(struct kmem_cache, node) +
3533 nr_node_ids * sizeof(struct kmem_cache_node *);
3534
3535 /* Allocate two kmem_caches from the page allocator */
3536 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3537 order = get_order(2 * kmalloc_size);
3538 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3539
3540 /*
3541 * Must first have the slab cache available for the allocations of the
3542 * struct kmem_cache_node's. There is special bootstrap code in
3543 * kmem_cache_open for slab_state == DOWN.
3544 */
3545 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3546
3547 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3548 sizeof(struct kmem_cache_node),
3549 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3550
3551 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3552
3553 /* Able to allocate the per node structures */
3554 slab_state = PARTIAL;
3555
3556 temp_kmem_cache = kmem_cache;
3557 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3558 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3559 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3560 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3561
3562 /*
3563 * Allocate kmem_cache_node properly from the kmem_cache slab.
3564 * kmem_cache_node is separately allocated so no need to
3565 * update any list pointers.
3566 */
3567 temp_kmem_cache_node = kmem_cache_node;
3568
3569 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3570 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3571
3572 kmem_cache_bootstrap_fixup(kmem_cache_node);
3573
3574 caches++;
3575 kmem_cache_bootstrap_fixup(kmem_cache);
3576 caches++;
3577 /* Free temporary boot structure */
3578 free_pages((unsigned long)temp_kmem_cache, order);
3579
3580 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3581
3582 /*
3583 * Patch up the size_index table if we have strange large alignment
3584 * requirements for the kmalloc array. This is only the case for
3585 * MIPS it seems. The standard arches will not generate any code here.
3586 *
3587 * Largest permitted alignment is 256 bytes due to the way we
3588 * handle the index determination for the smaller caches.
3589 *
3590 * Make sure that nothing crazy happens if someone starts tinkering
3591 * around with ARCH_KMALLOC_MINALIGN
3592 */
3593 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3594 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3595
3596 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3597 int elem = size_index_elem(i);
3598 if (elem >= ARRAY_SIZE(size_index))
3599 break;
3600 size_index[elem] = KMALLOC_SHIFT_LOW;
3601 }
3602
3603 if (KMALLOC_MIN_SIZE == 64) {
3604 /*
3605 * The 96 byte size cache is not used if the alignment
3606 * is 64 byte.
3607 */
3608 for (i = 64 + 8; i <= 96; i += 8)
3609 size_index[size_index_elem(i)] = 7;
3610 } else if (KMALLOC_MIN_SIZE == 128) {
3611 /*
3612 * The 192 byte sized cache is not used if the alignment
3613 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3614 * instead.
3615 */
3616 for (i = 128 + 8; i <= 192; i += 8)
3617 size_index[size_index_elem(i)] = 8;
3618 }
3619
3620 /* Caches that are not of the two-to-the-power-of size */
3621 if (KMALLOC_MIN_SIZE <= 32) {
3622 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3623 caches++;
3624 }
3625
3626 if (KMALLOC_MIN_SIZE <= 64) {
3627 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3628 caches++;
3629 }
3630
3631 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3632 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3633 caches++;
3634 }
3635
3636 slab_state = UP;
3637
3638 /* Provide the correct kmalloc names now that the caches are up */
3639 if (KMALLOC_MIN_SIZE <= 32) {
3640 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3641 BUG_ON(!kmalloc_caches[1]->name);
3642 }
3643
3644 if (KMALLOC_MIN_SIZE <= 64) {
3645 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3646 BUG_ON(!kmalloc_caches[2]->name);
3647 }
3648
3649 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3650 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3651
3652 BUG_ON(!s);
3653 kmalloc_caches[i]->name = s;
3654 }
3655
3656#ifdef CONFIG_SMP
3657 register_cpu_notifier(&slab_notifier);
3658#endif
3659
3660#ifdef CONFIG_ZONE_DMA
3661 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3662 struct kmem_cache *s = kmalloc_caches[i];
3663
3664 if (s && s->size) {
3665 char *name = kasprintf(GFP_NOWAIT,
3666 "dma-kmalloc-%d", s->objsize);
3667
3668 BUG_ON(!name);
3669 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3670 s->objsize, SLAB_CACHE_DMA);
3671 }
3672 }
3673#endif
3674 printk(KERN_INFO
3675 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3676 " CPUs=%d, Nodes=%d\n",
3677 caches, cache_line_size(),
3678 slub_min_order, slub_max_order, slub_min_objects,
3679 nr_cpu_ids, nr_node_ids);
3680}
3681
3682void __init kmem_cache_init_late(void)
3683{
3684}
3685
3686/*
3687 * Find a mergeable slab cache
3688 */
3689static int slab_unmergeable(struct kmem_cache *s)
3690{
3691 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3692 return 1;
3693
3694 if (s->ctor)
3695 return 1;
3696
3697 /*
3698 * We may have set a slab to be unmergeable during bootstrap.
3699 */
3700 if (s->refcount < 0)
3701 return 1;
3702
3703 return 0;
3704}
3705
3706static struct kmem_cache *find_mergeable(size_t size,
3707 size_t align, unsigned long flags, const char *name,
3708 void (*ctor)(void *))
3709{
3710 struct kmem_cache *s;
3711
3712 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3713 return NULL;
3714
3715 if (ctor)
3716 return NULL;
3717
3718 size = ALIGN(size, sizeof(void *));
3719 align = calculate_alignment(flags, align, size);
3720 size = ALIGN(size, align);
3721 flags = kmem_cache_flags(size, flags, name, NULL);
3722
3723 list_for_each_entry(s, &slab_caches, list) {
3724 if (slab_unmergeable(s))
3725 continue;
3726
3727 if (size > s->size)
3728 continue;
3729
3730 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3731 continue;
3732 /*
3733 * Check if alignment is compatible.
3734 * Courtesy of Adrian Drzewiecki
3735 */
3736 if ((s->size & ~(align - 1)) != s->size)
3737 continue;
3738
3739 if (s->size - size >= sizeof(void *))
3740 continue;
3741
3742 return s;
3743 }
3744 return NULL;
3745}
3746
3747struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3748 size_t align, unsigned long flags, void (*ctor)(void *))
3749{
3750 struct kmem_cache *s;
3751 char *n;
3752
3753 if (WARN_ON(!name))
3754 return NULL;
3755
3756 down_write(&slub_lock);
3757 s = find_mergeable(size, align, flags, name, ctor);
3758 if (s) {
3759 s->refcount++;
3760 /*
3761 * Adjust the object sizes so that we clear
3762 * the complete object on kzalloc.
3763 */
3764 s->objsize = max(s->objsize, (int)size);
3765 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3766
3767 if (sysfs_slab_alias(s, name)) {
3768 s->refcount--;
3769 goto err;
3770 }
3771 up_write(&slub_lock);
3772 return s;
3773 }
3774
3775 n = kstrdup(name, GFP_KERNEL);
3776 if (!n)
3777 goto err;
3778
3779 s = kmalloc(kmem_size, GFP_KERNEL);
3780 if (s) {
3781 if (kmem_cache_open(s, n,
3782 size, align, flags, ctor)) {
3783 list_add(&s->list, &slab_caches);
3784 if (sysfs_slab_add(s)) {
3785 list_del(&s->list);
3786 kfree(n);
3787 kfree(s);
3788 goto err;
3789 }
3790 up_write(&slub_lock);
3791 return s;
3792 }
3793 kfree(n);
3794 kfree(s);
3795 }
3796err:
3797 up_write(&slub_lock);
3798
3799 if (flags & SLAB_PANIC)
3800 panic("Cannot create slabcache %s\n", name);
3801 else
3802 s = NULL;
3803 return s;
3804}
3805EXPORT_SYMBOL(kmem_cache_create);
3806
3807#ifdef CONFIG_SMP
3808/*
3809 * Use the cpu notifier to insure that the cpu slabs are flushed when
3810 * necessary.
3811 */
3812static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3813 unsigned long action, void *hcpu)
3814{
3815 long cpu = (long)hcpu;
3816 struct kmem_cache *s;
3817 unsigned long flags;
3818
3819 switch (action) {
3820 case CPU_UP_CANCELED:
3821 case CPU_UP_CANCELED_FROZEN:
3822 case CPU_DEAD:
3823 case CPU_DEAD_FROZEN:
3824 down_read(&slub_lock);
3825 list_for_each_entry(s, &slab_caches, list) {
3826 local_irq_save(flags);
3827 __flush_cpu_slab(s, cpu);
3828 local_irq_restore(flags);
3829 }
3830 up_read(&slub_lock);
3831 break;
3832 default:
3833 break;
3834 }
3835 return NOTIFY_OK;
3836}
3837
3838static struct notifier_block __cpuinitdata slab_notifier = {
3839 .notifier_call = slab_cpuup_callback
3840};
3841
3842#endif
3843
3844void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3845{
3846 struct kmem_cache *s;
3847 void *ret;
3848
3849 if (unlikely(size > SLUB_MAX_SIZE))
3850 return kmalloc_large(size, gfpflags);
3851
3852 s = get_slab(size, gfpflags);
3853
3854 if (unlikely(ZERO_OR_NULL_PTR(s)))
3855 return s;
3856
3857 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3858
3859 /* Honor the call site pointer we received. */
3860 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3861
3862 return ret;
3863}
3864
3865#ifdef CONFIG_NUMA
3866void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3867 int node, unsigned long caller)
3868{
3869 struct kmem_cache *s;
3870 void *ret;
3871
3872 if (unlikely(size > SLUB_MAX_SIZE)) {
3873 ret = kmalloc_large_node(size, gfpflags, node);
3874
3875 trace_kmalloc_node(caller, ret,
3876 size, PAGE_SIZE << get_order(size),
3877 gfpflags, node);
3878
3879 return ret;
3880 }
3881
3882 s = get_slab(size, gfpflags);
3883
3884 if (unlikely(ZERO_OR_NULL_PTR(s)))
3885 return s;
3886
3887 ret = slab_alloc(s, gfpflags, node, caller);
3888
3889 /* Honor the call site pointer we received. */
3890 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3891
3892 return ret;
3893}
3894#endif
3895
3896#ifdef CONFIG_SYSFS
3897static int count_inuse(struct page *page)
3898{
3899 return page->inuse;
3900}
3901
3902static int count_total(struct page *page)
3903{
3904 return page->objects;
3905}
3906#endif
3907
3908#ifdef CONFIG_SLUB_DEBUG
3909static int validate_slab(struct kmem_cache *s, struct page *page,
3910 unsigned long *map)
3911{
3912 void *p;
3913 void *addr = page_address(page);
3914
3915 if (!check_slab(s, page) ||
3916 !on_freelist(s, page, NULL))
3917 return 0;
3918
3919 /* Now we know that a valid freelist exists */
3920 bitmap_zero(map, page->objects);
3921
3922 get_map(s, page, map);
3923 for_each_object(p, s, addr, page->objects) {
3924 if (test_bit(slab_index(p, s, addr), map))
3925 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3926 return 0;
3927 }
3928
3929 for_each_object(p, s, addr, page->objects)
3930 if (!test_bit(slab_index(p, s, addr), map))
3931 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3932 return 0;
3933 return 1;
3934}
3935
3936static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3937 unsigned long *map)
3938{
3939 slab_lock(page);
3940 validate_slab(s, page, map);
3941 slab_unlock(page);
3942}
3943
3944static int validate_slab_node(struct kmem_cache *s,
3945 struct kmem_cache_node *n, unsigned long *map)
3946{
3947 unsigned long count = 0;
3948 struct page *page;
3949 unsigned long flags;
3950
3951 spin_lock_irqsave(&n->list_lock, flags);
3952
3953 list_for_each_entry(page, &n->partial, lru) {
3954 validate_slab_slab(s, page, map);
3955 count++;
3956 }
3957 if (count != n->nr_partial)
3958 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3959 "counter=%ld\n", s->name, count, n->nr_partial);
3960
3961 if (!(s->flags & SLAB_STORE_USER))
3962 goto out;
3963
3964 list_for_each_entry(page, &n->full, lru) {
3965 validate_slab_slab(s, page, map);
3966 count++;
3967 }
3968 if (count != atomic_long_read(&n->nr_slabs))
3969 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3970 "counter=%ld\n", s->name, count,
3971 atomic_long_read(&n->nr_slabs));
3972
3973out:
3974 spin_unlock_irqrestore(&n->list_lock, flags);
3975 return count;
3976}
3977
3978static long validate_slab_cache(struct kmem_cache *s)
3979{
3980 int node;
3981 unsigned long count = 0;
3982 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3983 sizeof(unsigned long), GFP_KERNEL);
3984
3985 if (!map)
3986 return -ENOMEM;
3987
3988 flush_all(s);
3989 for_each_node_state(node, N_NORMAL_MEMORY) {
3990 struct kmem_cache_node *n = get_node(s, node);
3991
3992 count += validate_slab_node(s, n, map);
3993 }
3994 kfree(map);
3995 return count;
3996}
3997/*
3998 * Generate lists of code addresses where slabcache objects are allocated
3999 * and freed.
4000 */
4001
4002struct location {
4003 unsigned long count;
4004 unsigned long addr;
4005 long long sum_time;
4006 long min_time;
4007 long max_time;
4008 long min_pid;
4009 long max_pid;
4010 DECLARE_BITMAP(cpus, NR_CPUS);
4011 nodemask_t nodes;
4012};
4013
4014struct loc_track {
4015 unsigned long max;
4016 unsigned long count;
4017 struct location *loc;
4018};
4019
4020static void free_loc_track(struct loc_track *t)
4021{
4022 if (t->max)
4023 free_pages((unsigned long)t->loc,
4024 get_order(sizeof(struct location) * t->max));
4025}
4026
4027static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4028{
4029 struct location *l;
4030 int order;
4031
4032 order = get_order(sizeof(struct location) * max);
4033
4034 l = (void *)__get_free_pages(flags, order);
4035 if (!l)
4036 return 0;
4037
4038 if (t->count) {
4039 memcpy(l, t->loc, sizeof(struct location) * t->count);
4040 free_loc_track(t);
4041 }
4042 t->max = max;
4043 t->loc = l;
4044 return 1;
4045}
4046
4047static int add_location(struct loc_track *t, struct kmem_cache *s,
4048 const struct track *track)
4049{
4050 long start, end, pos;
4051 struct location *l;
4052 unsigned long caddr;
4053 unsigned long age = jiffies - track->when;
4054
4055 start = -1;
4056 end = t->count;
4057
4058 for ( ; ; ) {
4059 pos = start + (end - start + 1) / 2;
4060
4061 /*
4062 * There is nothing at "end". If we end up there
4063 * we need to add something to before end.
4064 */
4065 if (pos == end)
4066 break;
4067
4068 caddr = t->loc[pos].addr;
4069 if (track->addr == caddr) {
4070
4071 l = &t->loc[pos];
4072 l->count++;
4073 if (track->when) {
4074 l->sum_time += age;
4075 if (age < l->min_time)
4076 l->min_time = age;
4077 if (age > l->max_time)
4078 l->max_time = age;
4079
4080 if (track->pid < l->min_pid)
4081 l->min_pid = track->pid;
4082 if (track->pid > l->max_pid)
4083 l->max_pid = track->pid;
4084
4085 cpumask_set_cpu(track->cpu,
4086 to_cpumask(l->cpus));
4087 }
4088 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4089 return 1;
4090 }
4091
4092 if (track->addr < caddr)
4093 end = pos;
4094 else
4095 start = pos;
4096 }
4097
4098 /*
4099 * Not found. Insert new tracking element.
4100 */
4101 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4102 return 0;
4103
4104 l = t->loc + pos;
4105 if (pos < t->count)
4106 memmove(l + 1, l,
4107 (t->count - pos) * sizeof(struct location));
4108 t->count++;
4109 l->count = 1;
4110 l->addr = track->addr;
4111 l->sum_time = age;
4112 l->min_time = age;
4113 l->max_time = age;
4114 l->min_pid = track->pid;
4115 l->max_pid = track->pid;
4116 cpumask_clear(to_cpumask(l->cpus));
4117 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4118 nodes_clear(l->nodes);
4119 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4120 return 1;
4121}
4122
4123static void process_slab(struct loc_track *t, struct kmem_cache *s,
4124 struct page *page, enum track_item alloc,
4125 unsigned long *map)
4126{
4127 void *addr = page_address(page);
4128 void *p;
4129
4130 bitmap_zero(map, page->objects);
4131 get_map(s, page, map);
4132
4133 for_each_object(p, s, addr, page->objects)
4134 if (!test_bit(slab_index(p, s, addr), map))
4135 add_location(t, s, get_track(s, p, alloc));
4136}
4137
4138static int list_locations(struct kmem_cache *s, char *buf,
4139 enum track_item alloc)
4140{
4141 int len = 0;
4142 unsigned long i;
4143 struct loc_track t = { 0, 0, NULL };
4144 int node;
4145 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4146 sizeof(unsigned long), GFP_KERNEL);
4147
4148 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4149 GFP_TEMPORARY)) {
4150 kfree(map);
4151 return sprintf(buf, "Out of memory\n");
4152 }
4153 /* Push back cpu slabs */
4154 flush_all(s);
4155
4156 for_each_node_state(node, N_NORMAL_MEMORY) {
4157 struct kmem_cache_node *n = get_node(s, node);
4158 unsigned long flags;
4159 struct page *page;
4160
4161 if (!atomic_long_read(&n->nr_slabs))
4162 continue;
4163
4164 spin_lock_irqsave(&n->list_lock, flags);
4165 list_for_each_entry(page, &n->partial, lru)
4166 process_slab(&t, s, page, alloc, map);
4167 list_for_each_entry(page, &n->full, lru)
4168 process_slab(&t, s, page, alloc, map);
4169 spin_unlock_irqrestore(&n->list_lock, flags);
4170 }
4171
4172 for (i = 0; i < t.count; i++) {
4173 struct location *l = &t.loc[i];
4174
4175 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4176 break;
4177 len += sprintf(buf + len, "%7ld ", l->count);
4178
4179 if (l->addr)
4180 len += sprintf(buf + len, "%pS", (void *)l->addr);
4181 else
4182 len += sprintf(buf + len, "<not-available>");
4183
4184 if (l->sum_time != l->min_time) {
4185 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4186 l->min_time,
4187 (long)div_u64(l->sum_time, l->count),
4188 l->max_time);
4189 } else
4190 len += sprintf(buf + len, " age=%ld",
4191 l->min_time);
4192
4193 if (l->min_pid != l->max_pid)
4194 len += sprintf(buf + len, " pid=%ld-%ld",
4195 l->min_pid, l->max_pid);
4196 else
4197 len += sprintf(buf + len, " pid=%ld",
4198 l->min_pid);
4199
4200 if (num_online_cpus() > 1 &&
4201 !cpumask_empty(to_cpumask(l->cpus)) &&
4202 len < PAGE_SIZE - 60) {
4203 len += sprintf(buf + len, " cpus=");
4204 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4205 to_cpumask(l->cpus));
4206 }
4207
4208 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4209 len < PAGE_SIZE - 60) {
4210 len += sprintf(buf + len, " nodes=");
4211 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4212 l->nodes);
4213 }
4214
4215 len += sprintf(buf + len, "\n");
4216 }
4217
4218 free_loc_track(&t);
4219 kfree(map);
4220 if (!t.count)
4221 len += sprintf(buf, "No data\n");
4222 return len;
4223}
4224#endif
4225
4226#ifdef SLUB_RESILIENCY_TEST
4227static void resiliency_test(void)
4228{
4229 u8 *p;
4230
4231 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4232
4233 printk(KERN_ERR "SLUB resiliency testing\n");
4234 printk(KERN_ERR "-----------------------\n");
4235 printk(KERN_ERR "A. Corruption after allocation\n");
4236
4237 p = kzalloc(16, GFP_KERNEL);
4238 p[16] = 0x12;
4239 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4240 " 0x12->0x%p\n\n", p + 16);
4241
4242 validate_slab_cache(kmalloc_caches[4]);
4243
4244 /* Hmmm... The next two are dangerous */
4245 p = kzalloc(32, GFP_KERNEL);
4246 p[32 + sizeof(void *)] = 0x34;
4247 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4248 " 0x34 -> -0x%p\n", p);
4249 printk(KERN_ERR
4250 "If allocated object is overwritten then not detectable\n\n");
4251
4252 validate_slab_cache(kmalloc_caches[5]);
4253 p = kzalloc(64, GFP_KERNEL);
4254 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4255 *p = 0x56;
4256 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4257 p);
4258 printk(KERN_ERR
4259 "If allocated object is overwritten then not detectable\n\n");
4260 validate_slab_cache(kmalloc_caches[6]);
4261
4262 printk(KERN_ERR "\nB. Corruption after free\n");
4263 p = kzalloc(128, GFP_KERNEL);
4264 kfree(p);
4265 *p = 0x78;
4266 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4267 validate_slab_cache(kmalloc_caches[7]);
4268
4269 p = kzalloc(256, GFP_KERNEL);
4270 kfree(p);
4271 p[50] = 0x9a;
4272 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4273 p);
4274 validate_slab_cache(kmalloc_caches[8]);
4275
4276 p = kzalloc(512, GFP_KERNEL);
4277 kfree(p);
4278 p[512] = 0xab;
4279 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4280 validate_slab_cache(kmalloc_caches[9]);
4281}
4282#else
4283#ifdef CONFIG_SYSFS
4284static void resiliency_test(void) {};
4285#endif
4286#endif
4287
4288#ifdef CONFIG_SYSFS
4289enum slab_stat_type {
4290 SL_ALL, /* All slabs */
4291 SL_PARTIAL, /* Only partially allocated slabs */
4292 SL_CPU, /* Only slabs used for cpu caches */
4293 SL_OBJECTS, /* Determine allocated objects not slabs */
4294 SL_TOTAL /* Determine object capacity not slabs */
4295};
4296
4297#define SO_ALL (1 << SL_ALL)
4298#define SO_PARTIAL (1 << SL_PARTIAL)
4299#define SO_CPU (1 << SL_CPU)
4300#define SO_OBJECTS (1 << SL_OBJECTS)
4301#define SO_TOTAL (1 << SL_TOTAL)
4302
4303static ssize_t show_slab_objects(struct kmem_cache *s,
4304 char *buf, unsigned long flags)
4305{
4306 unsigned long total = 0;
4307 int node;
4308 int x;
4309 unsigned long *nodes;
4310 unsigned long *per_cpu;
4311
4312 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4313 if (!nodes)
4314 return -ENOMEM;
4315 per_cpu = nodes + nr_node_ids;
4316
4317 if (flags & SO_CPU) {
4318 int cpu;
4319
4320 for_each_possible_cpu(cpu) {
4321 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4322
4323 if (!c || c->node < 0)
4324 continue;
4325
4326 if (c->page) {
4327 if (flags & SO_TOTAL)
4328 x = c->page->objects;
4329 else if (flags & SO_OBJECTS)
4330 x = c->page->inuse;
4331 else
4332 x = 1;
4333
4334 total += x;
4335 nodes[c->node] += x;
4336 }
4337 per_cpu[c->node]++;
4338 }
4339 }
4340
4341 lock_memory_hotplug();
4342#ifdef CONFIG_SLUB_DEBUG
4343 if (flags & SO_ALL) {
4344 for_each_node_state(node, N_NORMAL_MEMORY) {
4345 struct kmem_cache_node *n = get_node(s, node);
4346
4347 if (flags & SO_TOTAL)
4348 x = atomic_long_read(&n->total_objects);
4349 else if (flags & SO_OBJECTS)
4350 x = atomic_long_read(&n->total_objects) -
4351 count_partial(n, count_free);
4352
4353 else
4354 x = atomic_long_read(&n->nr_slabs);
4355 total += x;
4356 nodes[node] += x;
4357 }
4358
4359 } else
4360#endif
4361 if (flags & SO_PARTIAL) {
4362 for_each_node_state(node, N_NORMAL_MEMORY) {
4363 struct kmem_cache_node *n = get_node(s, node);
4364
4365 if (flags & SO_TOTAL)
4366 x = count_partial(n, count_total);
4367 else if (flags & SO_OBJECTS)
4368 x = count_partial(n, count_inuse);
4369 else
4370 x = n->nr_partial;
4371 total += x;
4372 nodes[node] += x;
4373 }
4374 }
4375 x = sprintf(buf, "%lu", total);
4376#ifdef CONFIG_NUMA
4377 for_each_node_state(node, N_NORMAL_MEMORY)
4378 if (nodes[node])
4379 x += sprintf(buf + x, " N%d=%lu",
4380 node, nodes[node]);
4381#endif
4382 unlock_memory_hotplug();
4383 kfree(nodes);
4384 return x + sprintf(buf + x, "\n");
4385}
4386
4387#ifdef CONFIG_SLUB_DEBUG
4388static int any_slab_objects(struct kmem_cache *s)
4389{
4390 int node;
4391
4392 for_each_online_node(node) {
4393 struct kmem_cache_node *n = get_node(s, node);
4394
4395 if (!n)
4396 continue;
4397
4398 if (atomic_long_read(&n->total_objects))
4399 return 1;
4400 }
4401 return 0;
4402}
4403#endif
4404
4405#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4406#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4407
4408struct slab_attribute {
4409 struct attribute attr;
4410 ssize_t (*show)(struct kmem_cache *s, char *buf);
4411 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4412};
4413
4414#define SLAB_ATTR_RO(_name) \
4415 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4416
4417#define SLAB_ATTR(_name) \
4418 static struct slab_attribute _name##_attr = \
4419 __ATTR(_name, 0644, _name##_show, _name##_store)
4420
4421static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4422{
4423 return sprintf(buf, "%d\n", s->size);
4424}
4425SLAB_ATTR_RO(slab_size);
4426
4427static ssize_t align_show(struct kmem_cache *s, char *buf)
4428{
4429 return sprintf(buf, "%d\n", s->align);
4430}
4431SLAB_ATTR_RO(align);
4432
4433static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4434{
4435 return sprintf(buf, "%d\n", s->objsize);
4436}
4437SLAB_ATTR_RO(object_size);
4438
4439static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4440{
4441 return sprintf(buf, "%d\n", oo_objects(s->oo));
4442}
4443SLAB_ATTR_RO(objs_per_slab);
4444
4445static ssize_t order_store(struct kmem_cache *s,
4446 const char *buf, size_t length)
4447{
4448 unsigned long order;
4449 int err;
4450
4451 err = strict_strtoul(buf, 10, &order);
4452 if (err)
4453 return err;
4454
4455 if (order > slub_max_order || order < slub_min_order)
4456 return -EINVAL;
4457
4458 calculate_sizes(s, order);
4459 return length;
4460}
4461
4462static ssize_t order_show(struct kmem_cache *s, char *buf)
4463{
4464 return sprintf(buf, "%d\n", oo_order(s->oo));
4465}
4466SLAB_ATTR(order);
4467
4468static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4469{
4470 return sprintf(buf, "%lu\n", s->min_partial);
4471}
4472
4473static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4474 size_t length)
4475{
4476 unsigned long min;
4477 int err;
4478
4479 err = strict_strtoul(buf, 10, &min);
4480 if (err)
4481 return err;
4482
4483 set_min_partial(s, min);
4484 return length;
4485}
4486SLAB_ATTR(min_partial);
4487
4488static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4489{
4490 if (!s->ctor)
4491 return 0;
4492 return sprintf(buf, "%pS\n", s->ctor);
4493}
4494SLAB_ATTR_RO(ctor);
4495
4496static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4497{
4498 return sprintf(buf, "%d\n", s->refcount - 1);
4499}
4500SLAB_ATTR_RO(aliases);
4501
4502static ssize_t partial_show(struct kmem_cache *s, char *buf)
4503{
4504 return show_slab_objects(s, buf, SO_PARTIAL);
4505}
4506SLAB_ATTR_RO(partial);
4507
4508static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4509{
4510 return show_slab_objects(s, buf, SO_CPU);
4511}
4512SLAB_ATTR_RO(cpu_slabs);
4513
4514static ssize_t objects_show(struct kmem_cache *s, char *buf)
4515{
4516 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4517}
4518SLAB_ATTR_RO(objects);
4519
4520static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4521{
4522 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4523}
4524SLAB_ATTR_RO(objects_partial);
4525
4526static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4527{
4528 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4529}
4530
4531static ssize_t reclaim_account_store(struct kmem_cache *s,
4532 const char *buf, size_t length)
4533{
4534 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4535 if (buf[0] == '1')
4536 s->flags |= SLAB_RECLAIM_ACCOUNT;
4537 return length;
4538}
4539SLAB_ATTR(reclaim_account);
4540
4541static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4542{
4543 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4544}
4545SLAB_ATTR_RO(hwcache_align);
4546
4547#ifdef CONFIG_ZONE_DMA
4548static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4549{
4550 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4551}
4552SLAB_ATTR_RO(cache_dma);
4553#endif
4554
4555static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4556{
4557 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4558}
4559SLAB_ATTR_RO(destroy_by_rcu);
4560
4561static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4562{
4563 return sprintf(buf, "%d\n", s->reserved);
4564}
4565SLAB_ATTR_RO(reserved);
4566
4567#ifdef CONFIG_SLUB_DEBUG
4568static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4569{
4570 return show_slab_objects(s, buf, SO_ALL);
4571}
4572SLAB_ATTR_RO(slabs);
4573
4574static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4575{
4576 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4577}
4578SLAB_ATTR_RO(total_objects);
4579
4580static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4581{
4582 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4583}
4584
4585static ssize_t sanity_checks_store(struct kmem_cache *s,
4586 const char *buf, size_t length)
4587{
4588 s->flags &= ~SLAB_DEBUG_FREE;
4589 if (buf[0] == '1') {
4590 s->flags &= ~__CMPXCHG_DOUBLE;
4591 s->flags |= SLAB_DEBUG_FREE;
4592 }
4593 return length;
4594}
4595SLAB_ATTR(sanity_checks);
4596
4597static ssize_t trace_show(struct kmem_cache *s, char *buf)
4598{
4599 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4600}
4601
4602static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4603 size_t length)
4604{
4605 s->flags &= ~SLAB_TRACE;
4606 if (buf[0] == '1') {
4607 s->flags &= ~__CMPXCHG_DOUBLE;
4608 s->flags |= SLAB_TRACE;
4609 }
4610 return length;
4611}
4612SLAB_ATTR(trace);
4613
4614static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4615{
4616 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4617}
4618
4619static ssize_t red_zone_store(struct kmem_cache *s,
4620 const char *buf, size_t length)
4621{
4622 if (any_slab_objects(s))
4623 return -EBUSY;
4624
4625 s->flags &= ~SLAB_RED_ZONE;
4626 if (buf[0] == '1') {
4627 s->flags &= ~__CMPXCHG_DOUBLE;
4628 s->flags |= SLAB_RED_ZONE;
4629 }
4630 calculate_sizes(s, -1);
4631 return length;
4632}
4633SLAB_ATTR(red_zone);
4634
4635static ssize_t poison_show(struct kmem_cache *s, char *buf)
4636{
4637 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4638}
4639
4640static ssize_t poison_store(struct kmem_cache *s,
4641 const char *buf, size_t length)
4642{
4643 if (any_slab_objects(s))
4644 return -EBUSY;
4645
4646 s->flags &= ~SLAB_POISON;
4647 if (buf[0] == '1') {
4648 s->flags &= ~__CMPXCHG_DOUBLE;
4649 s->flags |= SLAB_POISON;
4650 }
4651 calculate_sizes(s, -1);
4652 return length;
4653}
4654SLAB_ATTR(poison);
4655
4656static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4657{
4658 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4659}
4660
4661static ssize_t store_user_store(struct kmem_cache *s,
4662 const char *buf, size_t length)
4663{
4664 if (any_slab_objects(s))
4665 return -EBUSY;
4666
4667 s->flags &= ~SLAB_STORE_USER;
4668 if (buf[0] == '1') {
4669 s->flags &= ~__CMPXCHG_DOUBLE;
4670 s->flags |= SLAB_STORE_USER;
4671 }
4672 calculate_sizes(s, -1);
4673 return length;
4674}
4675SLAB_ATTR(store_user);
4676
4677static ssize_t validate_show(struct kmem_cache *s, char *buf)
4678{
4679 return 0;
4680}
4681
4682static ssize_t validate_store(struct kmem_cache *s,
4683 const char *buf, size_t length)
4684{
4685 int ret = -EINVAL;
4686
4687 if (buf[0] == '1') {
4688 ret = validate_slab_cache(s);
4689 if (ret >= 0)
4690 ret = length;
4691 }
4692 return ret;
4693}
4694SLAB_ATTR(validate);
4695
4696static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4697{
4698 if (!(s->flags & SLAB_STORE_USER))
4699 return -ENOSYS;
4700 return list_locations(s, buf, TRACK_ALLOC);
4701}
4702SLAB_ATTR_RO(alloc_calls);
4703
4704static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4705{
4706 if (!(s->flags & SLAB_STORE_USER))
4707 return -ENOSYS;
4708 return list_locations(s, buf, TRACK_FREE);
4709}
4710SLAB_ATTR_RO(free_calls);
4711#endif /* CONFIG_SLUB_DEBUG */
4712
4713#ifdef CONFIG_FAILSLAB
4714static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4715{
4716 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4717}
4718
4719static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4720 size_t length)
4721{
4722 s->flags &= ~SLAB_FAILSLAB;
4723 if (buf[0] == '1')
4724 s->flags |= SLAB_FAILSLAB;
4725 return length;
4726}
4727SLAB_ATTR(failslab);
4728#endif
4729
4730static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4731{
4732 return 0;
4733}
4734
4735static ssize_t shrink_store(struct kmem_cache *s,
4736 const char *buf, size_t length)
4737{
4738 if (buf[0] == '1') {
4739 int rc = kmem_cache_shrink(s);
4740
4741 if (rc)
4742 return rc;
4743 } else
4744 return -EINVAL;
4745 return length;
4746}
4747SLAB_ATTR(shrink);
4748
4749#ifdef CONFIG_NUMA
4750static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4751{
4752 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4753}
4754
4755static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4756 const char *buf, size_t length)
4757{
4758 unsigned long ratio;
4759 int err;
4760
4761 err = strict_strtoul(buf, 10, &ratio);
4762 if (err)
4763 return err;
4764
4765 if (ratio <= 100)
4766 s->remote_node_defrag_ratio = ratio * 10;
4767
4768 return length;
4769}
4770SLAB_ATTR(remote_node_defrag_ratio);
4771#endif
4772
4773#ifdef CONFIG_SLUB_STATS
4774static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4775{
4776 unsigned long sum = 0;
4777 int cpu;
4778 int len;
4779 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4780
4781 if (!data)
4782 return -ENOMEM;
4783
4784 for_each_online_cpu(cpu) {
4785 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4786
4787 data[cpu] = x;
4788 sum += x;
4789 }
4790
4791 len = sprintf(buf, "%lu", sum);
4792
4793#ifdef CONFIG_SMP
4794 for_each_online_cpu(cpu) {
4795 if (data[cpu] && len < PAGE_SIZE - 20)
4796 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4797 }
4798#endif
4799 kfree(data);
4800 return len + sprintf(buf + len, "\n");
4801}
4802
4803static void clear_stat(struct kmem_cache *s, enum stat_item si)
4804{
4805 int cpu;
4806
4807 for_each_online_cpu(cpu)
4808 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4809}
4810
4811#define STAT_ATTR(si, text) \
4812static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4813{ \
4814 return show_stat(s, buf, si); \
4815} \
4816static ssize_t text##_store(struct kmem_cache *s, \
4817 const char *buf, size_t length) \
4818{ \
4819 if (buf[0] != '0') \
4820 return -EINVAL; \
4821 clear_stat(s, si); \
4822 return length; \
4823} \
4824SLAB_ATTR(text); \
4825
4826STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4827STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4828STAT_ATTR(FREE_FASTPATH, free_fastpath);
4829STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4830STAT_ATTR(FREE_FROZEN, free_frozen);
4831STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4832STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4833STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4834STAT_ATTR(ALLOC_SLAB, alloc_slab);
4835STAT_ATTR(ALLOC_REFILL, alloc_refill);
4836STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
4837STAT_ATTR(FREE_SLAB, free_slab);
4838STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4839STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4840STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4841STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4842STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4843STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4844STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
4845STAT_ATTR(ORDER_FALLBACK, order_fallback);
4846STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
4847STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
4848#endif
4849
4850static struct attribute *slab_attrs[] = {
4851 &slab_size_attr.attr,
4852 &object_size_attr.attr,
4853 &objs_per_slab_attr.attr,
4854 &order_attr.attr,
4855 &min_partial_attr.attr,
4856 &objects_attr.attr,
4857 &objects_partial_attr.attr,
4858 &partial_attr.attr,
4859 &cpu_slabs_attr.attr,
4860 &ctor_attr.attr,
4861 &aliases_attr.attr,
4862 &align_attr.attr,
4863 &hwcache_align_attr.attr,
4864 &reclaim_account_attr.attr,
4865 &destroy_by_rcu_attr.attr,
4866 &shrink_attr.attr,
4867 &reserved_attr.attr,
4868#ifdef CONFIG_SLUB_DEBUG
4869 &total_objects_attr.attr,
4870 &slabs_attr.attr,
4871 &sanity_checks_attr.attr,
4872 &trace_attr.attr,
4873 &red_zone_attr.attr,
4874 &poison_attr.attr,
4875 &store_user_attr.attr,
4876 &validate_attr.attr,
4877 &alloc_calls_attr.attr,
4878 &free_calls_attr.attr,
4879#endif
4880#ifdef CONFIG_ZONE_DMA
4881 &cache_dma_attr.attr,
4882#endif
4883#ifdef CONFIG_NUMA
4884 &remote_node_defrag_ratio_attr.attr,
4885#endif
4886#ifdef CONFIG_SLUB_STATS
4887 &alloc_fastpath_attr.attr,
4888 &alloc_slowpath_attr.attr,
4889 &free_fastpath_attr.attr,
4890 &free_slowpath_attr.attr,
4891 &free_frozen_attr.attr,
4892 &free_add_partial_attr.attr,
4893 &free_remove_partial_attr.attr,
4894 &alloc_from_partial_attr.attr,
4895 &alloc_slab_attr.attr,
4896 &alloc_refill_attr.attr,
4897 &alloc_node_mismatch_attr.attr,
4898 &free_slab_attr.attr,
4899 &cpuslab_flush_attr.attr,
4900 &deactivate_full_attr.attr,
4901 &deactivate_empty_attr.attr,
4902 &deactivate_to_head_attr.attr,
4903 &deactivate_to_tail_attr.attr,
4904 &deactivate_remote_frees_attr.attr,
4905 &deactivate_bypass_attr.attr,
4906 &order_fallback_attr.attr,
4907 &cmpxchg_double_fail_attr.attr,
4908 &cmpxchg_double_cpu_fail_attr.attr,
4909#endif
4910#ifdef CONFIG_FAILSLAB
4911 &failslab_attr.attr,
4912#endif
4913
4914 NULL
4915};
4916
4917static struct attribute_group slab_attr_group = {
4918 .attrs = slab_attrs,
4919};
4920
4921static ssize_t slab_attr_show(struct kobject *kobj,
4922 struct attribute *attr,
4923 char *buf)
4924{
4925 struct slab_attribute *attribute;
4926 struct kmem_cache *s;
4927 int err;
4928
4929 attribute = to_slab_attr(attr);
4930 s = to_slab(kobj);
4931
4932 if (!attribute->show)
4933 return -EIO;
4934
4935 err = attribute->show(s, buf);
4936
4937 return err;
4938}
4939
4940static ssize_t slab_attr_store(struct kobject *kobj,
4941 struct attribute *attr,
4942 const char *buf, size_t len)
4943{
4944 struct slab_attribute *attribute;
4945 struct kmem_cache *s;
4946 int err;
4947
4948 attribute = to_slab_attr(attr);
4949 s = to_slab(kobj);
4950
4951 if (!attribute->store)
4952 return -EIO;
4953
4954 err = attribute->store(s, buf, len);
4955
4956 return err;
4957}
4958
4959static void kmem_cache_release(struct kobject *kobj)
4960{
4961 struct kmem_cache *s = to_slab(kobj);
4962
4963 kfree(s->name);
4964 kfree(s);
4965}
4966
4967static const struct sysfs_ops slab_sysfs_ops = {
4968 .show = slab_attr_show,
4969 .store = slab_attr_store,
4970};
4971
4972static struct kobj_type slab_ktype = {
4973 .sysfs_ops = &slab_sysfs_ops,
4974 .release = kmem_cache_release
4975};
4976
4977static int uevent_filter(struct kset *kset, struct kobject *kobj)
4978{
4979 struct kobj_type *ktype = get_ktype(kobj);
4980
4981 if (ktype == &slab_ktype)
4982 return 1;
4983 return 0;
4984}
4985
4986static const struct kset_uevent_ops slab_uevent_ops = {
4987 .filter = uevent_filter,
4988};
4989
4990static struct kset *slab_kset;
4991
4992#define ID_STR_LENGTH 64
4993
4994/* Create a unique string id for a slab cache:
4995 *
4996 * Format :[flags-]size
4997 */
4998static char *create_unique_id(struct kmem_cache *s)
4999{
5000 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5001 char *p = name;
5002
5003 BUG_ON(!name);
5004
5005 *p++ = ':';
5006 /*
5007 * First flags affecting slabcache operations. We will only
5008 * get here for aliasable slabs so we do not need to support
5009 * too many flags. The flags here must cover all flags that
5010 * are matched during merging to guarantee that the id is
5011 * unique.
5012 */
5013 if (s->flags & SLAB_CACHE_DMA)
5014 *p++ = 'd';
5015 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5016 *p++ = 'a';
5017 if (s->flags & SLAB_DEBUG_FREE)
5018 *p++ = 'F';
5019 if (!(s->flags & SLAB_NOTRACK))
5020 *p++ = 't';
5021 if (p != name + 1)
5022 *p++ = '-';
5023 p += sprintf(p, "%07d", s->size);
5024 BUG_ON(p > name + ID_STR_LENGTH - 1);
5025 return name;
5026}
5027
5028static int sysfs_slab_add(struct kmem_cache *s)
5029{
5030 int err;
5031 const char *name;
5032 int unmergeable;
5033
5034 if (slab_state < SYSFS)
5035 /* Defer until later */
5036 return 0;
5037
5038 unmergeable = slab_unmergeable(s);
5039 if (unmergeable) {
5040 /*
5041 * Slabcache can never be merged so we can use the name proper.
5042 * This is typically the case for debug situations. In that
5043 * case we can catch duplicate names easily.
5044 */
5045 sysfs_remove_link(&slab_kset->kobj, s->name);
5046 name = s->name;
5047 } else {
5048 /*
5049 * Create a unique name for the slab as a target
5050 * for the symlinks.
5051 */
5052 name = create_unique_id(s);
5053 }
5054
5055 s->kobj.kset = slab_kset;
5056 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5057 if (err) {
5058 kobject_put(&s->kobj);
5059 return err;
5060 }
5061
5062 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5063 if (err) {
5064 kobject_del(&s->kobj);
5065 kobject_put(&s->kobj);
5066 return err;
5067 }
5068 kobject_uevent(&s->kobj, KOBJ_ADD);
5069 if (!unmergeable) {
5070 /* Setup first alias */
5071 sysfs_slab_alias(s, s->name);
5072 kfree(name);
5073 }
5074 return 0;
5075}
5076
5077static void sysfs_slab_remove(struct kmem_cache *s)
5078{
5079 if (slab_state < SYSFS)
5080 /*
5081 * Sysfs has not been setup yet so no need to remove the
5082 * cache from sysfs.
5083 */
5084 return;
5085
5086 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5087 kobject_del(&s->kobj);
5088 kobject_put(&s->kobj);
5089}
5090
5091/*
5092 * Need to buffer aliases during bootup until sysfs becomes
5093 * available lest we lose that information.
5094 */
5095struct saved_alias {
5096 struct kmem_cache *s;
5097 const char *name;
5098 struct saved_alias *next;
5099};
5100
5101static struct saved_alias *alias_list;
5102
5103static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5104{
5105 struct saved_alias *al;
5106
5107 if (slab_state == SYSFS) {
5108 /*
5109 * If we have a leftover link then remove it.
5110 */
5111 sysfs_remove_link(&slab_kset->kobj, name);
5112 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5113 }
5114
5115 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5116 if (!al)
5117 return -ENOMEM;
5118
5119 al->s = s;
5120 al->name = name;
5121 al->next = alias_list;
5122 alias_list = al;
5123 return 0;
5124}
5125
5126static int __init slab_sysfs_init(void)
5127{
5128 struct kmem_cache *s;
5129 int err;
5130
5131 down_write(&slub_lock);
5132
5133 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5134 if (!slab_kset) {
5135 up_write(&slub_lock);
5136 printk(KERN_ERR "Cannot register slab subsystem.\n");
5137 return -ENOSYS;
5138 }
5139
5140 slab_state = SYSFS;
5141
5142 list_for_each_entry(s, &slab_caches, list) {
5143 err = sysfs_slab_add(s);
5144 if (err)
5145 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5146 " to sysfs\n", s->name);
5147 }
5148
5149 while (alias_list) {
5150 struct saved_alias *al = alias_list;
5151
5152 alias_list = alias_list->next;
5153 err = sysfs_slab_alias(al->s, al->name);
5154 if (err)
5155 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5156 " %s to sysfs\n", s->name);
5157 kfree(al);
5158 }
5159
5160 up_write(&slub_lock);
5161 resiliency_test();
5162 return 0;
5163}
5164
5165__initcall(slab_sysfs_init);
5166#endif /* CONFIG_SYSFS */
5167
5168/*
5169 * The /proc/slabinfo ABI
5170 */
5171#ifdef CONFIG_SLABINFO
5172static void print_slabinfo_header(struct seq_file *m)
5173{
5174 seq_puts(m, "slabinfo - version: 2.1\n");
5175 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5176 "<objperslab> <pagesperslab>");
5177 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5178 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5179 seq_putc(m, '\n');
5180}
5181
5182static void *s_start(struct seq_file *m, loff_t *pos)
5183{
5184 loff_t n = *pos;
5185
5186 down_read(&slub_lock);
5187 if (!n)
5188 print_slabinfo_header(m);
5189
5190 return seq_list_start(&slab_caches, *pos);
5191}
5192
5193static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5194{
5195 return seq_list_next(p, &slab_caches, pos);
5196}
5197
5198static void s_stop(struct seq_file *m, void *p)
5199{
5200 up_read(&slub_lock);
5201}
5202
5203static int s_show(struct seq_file *m, void *p)
5204{
5205 unsigned long nr_partials = 0;
5206 unsigned long nr_slabs = 0;
5207 unsigned long nr_inuse = 0;
5208 unsigned long nr_objs = 0;
5209 unsigned long nr_free = 0;
5210 struct kmem_cache *s;
5211 int node;
5212
5213 s = list_entry(p, struct kmem_cache, list);
5214
5215 for_each_online_node(node) {
5216 struct kmem_cache_node *n = get_node(s, node);
5217
5218 if (!n)
5219 continue;
5220
5221 nr_partials += n->nr_partial;
5222 nr_slabs += atomic_long_read(&n->nr_slabs);
5223 nr_objs += atomic_long_read(&n->total_objects);
5224 nr_free += count_partial(n, count_free);
5225 }
5226
5227 nr_inuse = nr_objs - nr_free;
5228
5229 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5230 nr_objs, s->size, oo_objects(s->oo),
5231 (1 << oo_order(s->oo)));
5232 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5233 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5234 0UL);
5235 seq_putc(m, '\n');
5236 return 0;
5237}
5238
5239static const struct seq_operations slabinfo_op = {
5240 .start = s_start,
5241 .next = s_next,
5242 .stop = s_stop,
5243 .show = s_show,
5244};
5245
5246static int slabinfo_open(struct inode *inode, struct file *file)
5247{
5248 return seq_open(file, &slabinfo_op);
5249}
5250
5251static const struct file_operations proc_slabinfo_operations = {
5252 .open = slabinfo_open,
5253 .read = seq_read,
5254 .llseek = seq_lseek,
5255 .release = seq_release,
5256};
5257
5258static int __init slab_proc_init(void)
5259{
5260 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5261 return 0;
5262}
5263module_init(slab_proc_init);
5264#endif /* CONFIG_SLABINFO */