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 <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 */
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#include <linux/prefetch.h>
33
34#include <trace/events/kmem.h>
35
36/*
37 * Lock order:
38 * 1. slub_lock (Global Semaphore)
39 * 2. node->list_lock
40 * 3. slab_lock(page) (Only on some arches and for debugging)
41 *
42 * slub_lock
43 *
44 * The role of the slub_lock is to protect the list of all the slabs
45 * and to synchronize major metadata changes to slab cache structures.
46 *
47 * The slab_lock is only used for debugging and on arches that do not
48 * have the ability to do a cmpxchg_double. It only protects the second
49 * double word in the page struct. Meaning
50 * A. page->freelist -> List of object free in a page
51 * B. page->counters -> Counters of objects
52 * C. page->frozen -> frozen state
53 *
54 * If a slab is frozen then it is exempt from list management. It is not
55 * on any list. The processor that froze the slab is the one who can
56 * perform list operations on the page. Other processors may put objects
57 * onto the freelist but the processor that froze the slab is the only
58 * one that can retrieve the objects from the page's freelist.
59 *
60 * The list_lock protects the partial and full list on each node and
61 * the partial slab counter. If taken then no new slabs may be added or
62 * removed from the lists nor make the number of partial slabs be modified.
63 * (Note that the total number of slabs is an atomic value that may be
64 * modified without taking the list lock).
65 *
66 * The list_lock is a centralized lock and thus we avoid taking it as
67 * much as possible. As long as SLUB does not have to handle partial
68 * slabs, operations can continue without any centralized lock. F.e.
69 * allocating a long series of objects that fill up slabs does not require
70 * the list lock.
71 * Interrupts are disabled during allocation and deallocation in order to
72 * make the slab allocator safe to use in the context of an irq. In addition
73 * interrupts are disabled to ensure that the processor does not change
74 * while handling per_cpu slabs, due to kernel preemption.
75 *
76 * SLUB assigns one slab for allocation to each processor.
77 * Allocations only occur from these slabs called cpu slabs.
78 *
79 * Slabs with free elements are kept on a partial list and during regular
80 * operations no list for full slabs is used. If an object in a full slab is
81 * freed then the slab will show up again on the partial lists.
82 * We track full slabs for debugging purposes though because otherwise we
83 * cannot scan all objects.
84 *
85 * Slabs are freed when they become empty. Teardown and setup is
86 * minimal so we rely on the page allocators per cpu caches for
87 * fast frees and allocs.
88 *
89 * Overloading of page flags that are otherwise used for LRU management.
90 *
91 * PageActive The slab is frozen and exempt from list processing.
92 * This means that the slab is dedicated to a purpose
93 * such as satisfying allocations for a specific
94 * processor. Objects may be freed in the slab while
95 * it is frozen but slab_free will then skip the usual
96 * list operations. It is up to the processor holding
97 * the slab to integrate the slab into the slab lists
98 * when the slab is no longer needed.
99 *
100 * One use of this flag is to mark slabs that are
101 * used for allocations. Then such a slab becomes a cpu
102 * slab. The cpu slab may be equipped with an additional
103 * freelist that allows lockless access to
104 * free objects in addition to the regular freelist
105 * that requires the slab lock.
106 *
107 * PageError Slab requires special handling due to debug
108 * options set. This moves slab handling out of
109 * the fast path and disables lockless freelists.
110 */
111
112#define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
113 SLAB_TRACE | SLAB_DEBUG_FREE)
114
115static inline int kmem_cache_debug(struct kmem_cache *s)
116{
117#ifdef CONFIG_SLUB_DEBUG
118 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
119#else
120 return 0;
121#endif
122}
123
124/*
125 * Issues still to be resolved:
126 *
127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
128 *
129 * - Variable sizing of the per node arrays
130 */
131
132/* Enable to test recovery from slab corruption on boot */
133#undef SLUB_RESILIENCY_TEST
134
135/* Enable to log cmpxchg failures */
136#undef SLUB_DEBUG_CMPXCHG
137
138/*
139 * Mininum number of partial slabs. These will be left on the partial
140 * lists even if they are empty. kmem_cache_shrink may reclaim them.
141 */
142#define MIN_PARTIAL 5
143
144/*
145 * Maximum number of desirable partial slabs.
146 * The existence of more partial slabs makes kmem_cache_shrink
147 * sort the partial list by the number of objects in the.
148 */
149#define MAX_PARTIAL 10
150
151#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
152 SLAB_POISON | SLAB_STORE_USER)
153
154/*
155 * Debugging flags that require metadata to be stored in the slab. These get
156 * disabled when slub_debug=O is used and a cache's min order increases with
157 * metadata.
158 */
159#define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
160
161/*
162 * Set of flags that will prevent slab merging
163 */
164#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
165 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
166 SLAB_FAILSLAB)
167
168#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
169 SLAB_CACHE_DMA | SLAB_NOTRACK)
170
171#define OO_SHIFT 16
172#define OO_MASK ((1 << OO_SHIFT) - 1)
173#define MAX_OBJS_PER_PAGE 32767 /* since page.objects is u15 */
174
175/* Internal SLUB flags */
176#define __OBJECT_POISON 0x80000000UL /* Poison object */
177#define __CMPXCHG_DOUBLE 0x40000000UL /* Use cmpxchg_double */
178
179static int kmem_size = sizeof(struct kmem_cache);
180
181#ifdef CONFIG_SMP
182static struct notifier_block slab_notifier;
183#endif
184
185static enum {
186 DOWN, /* No slab functionality available */
187 PARTIAL, /* Kmem_cache_node works */
188 UP, /* Everything works but does not show up in sysfs */
189 SYSFS /* Sysfs up */
190} slab_state = DOWN;
191
192/* A list of all slab caches on the system */
193static DECLARE_RWSEM(slub_lock);
194static LIST_HEAD(slab_caches);
195
196/*
197 * Tracking user of a slab.
198 */
199#define TRACK_ADDRS_COUNT 16
200struct track {
201 unsigned long addr; /* Called from address */
202#ifdef CONFIG_STACKTRACE
203 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
204#endif
205 int cpu; /* Was running on cpu */
206 int pid; /* Pid context */
207 unsigned long when; /* When did the operation occur */
208};
209
210enum track_item { TRACK_ALLOC, TRACK_FREE };
211
212#ifdef CONFIG_SYSFS
213static int sysfs_slab_add(struct kmem_cache *);
214static int sysfs_slab_alias(struct kmem_cache *, const char *);
215static void sysfs_slab_remove(struct kmem_cache *);
216
217#else
218static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
219static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
220 { return 0; }
221static inline void sysfs_slab_remove(struct kmem_cache *s)
222{
223 kfree(s->name);
224 kfree(s);
225}
226
227#endif
228
229static inline void stat(const struct kmem_cache *s, enum stat_item si)
230{
231#ifdef CONFIG_SLUB_STATS
232 __this_cpu_inc(s->cpu_slab->stat[si]);
233#endif
234}
235
236/********************************************************************
237 * Core slab cache functions
238 *******************************************************************/
239
240int slab_is_available(void)
241{
242 return slab_state >= UP;
243}
244
245static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
246{
247 return s->node[node];
248}
249
250/* Verify that a pointer has an address that is valid within a slab page */
251static inline int check_valid_pointer(struct kmem_cache *s,
252 struct page *page, const void *object)
253{
254 void *base;
255
256 if (!object)
257 return 1;
258
259 base = page_address(page);
260 if (object < base || object >= base + page->objects * s->size ||
261 (object - base) % s->size) {
262 return 0;
263 }
264
265 return 1;
266}
267
268static inline void *get_freepointer(struct kmem_cache *s, void *object)
269{
270 return *(void **)(object + s->offset);
271}
272
273static void prefetch_freepointer(const struct kmem_cache *s, void *object)
274{
275 prefetch(object + s->offset);
276}
277
278static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
279{
280 void *p;
281
282#ifdef CONFIG_DEBUG_PAGEALLOC
283 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
284#else
285 p = get_freepointer(s, object);
286#endif
287 return p;
288}
289
290static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
291{
292 *(void **)(object + s->offset) = fp;
293}
294
295/* Loop over all objects in a slab */
296#define for_each_object(__p, __s, __addr, __objects) \
297 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
298 __p += (__s)->size)
299
300/* Determine object index from a given position */
301static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
302{
303 return (p - addr) / s->size;
304}
305
306static inline size_t slab_ksize(const struct kmem_cache *s)
307{
308#ifdef CONFIG_SLUB_DEBUG
309 /*
310 * Debugging requires use of the padding between object
311 * and whatever may come after it.
312 */
313 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
314 return s->objsize;
315
316#endif
317 /*
318 * If we have the need to store the freelist pointer
319 * back there or track user information then we can
320 * only use the space before that information.
321 */
322 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
323 return s->inuse;
324 /*
325 * Else we can use all the padding etc for the allocation
326 */
327 return s->size;
328}
329
330static inline int order_objects(int order, unsigned long size, int reserved)
331{
332 return ((PAGE_SIZE << order) - reserved) / size;
333}
334
335static inline struct kmem_cache_order_objects oo_make(int order,
336 unsigned long size, int reserved)
337{
338 struct kmem_cache_order_objects x = {
339 (order << OO_SHIFT) + order_objects(order, size, reserved)
340 };
341
342 return x;
343}
344
345static inline int oo_order(struct kmem_cache_order_objects x)
346{
347 return x.x >> OO_SHIFT;
348}
349
350static inline int oo_objects(struct kmem_cache_order_objects x)
351{
352 return x.x & OO_MASK;
353}
354
355/*
356 * Per slab locking using the pagelock
357 */
358static __always_inline void slab_lock(struct page *page)
359{
360 bit_spin_lock(PG_locked, &page->flags);
361}
362
363static __always_inline void slab_unlock(struct page *page)
364{
365 __bit_spin_unlock(PG_locked, &page->flags);
366}
367
368/* Interrupts must be disabled (for the fallback code to work right) */
369static inline bool __cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
370 void *freelist_old, unsigned long counters_old,
371 void *freelist_new, unsigned long counters_new,
372 const char *n)
373{
374 VM_BUG_ON(!irqs_disabled());
375#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
376 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
377 if (s->flags & __CMPXCHG_DOUBLE) {
378 if (cmpxchg_double(&page->freelist, &page->counters,
379 freelist_old, counters_old,
380 freelist_new, counters_new))
381 return 1;
382 } else
383#endif
384 {
385 slab_lock(page);
386 if (page->freelist == freelist_old && page->counters == counters_old) {
387 page->freelist = freelist_new;
388 page->counters = counters_new;
389 slab_unlock(page);
390 return 1;
391 }
392 slab_unlock(page);
393 }
394
395 cpu_relax();
396 stat(s, CMPXCHG_DOUBLE_FAIL);
397
398#ifdef SLUB_DEBUG_CMPXCHG
399 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
400#endif
401
402 return 0;
403}
404
405static inline bool cmpxchg_double_slab(struct kmem_cache *s, struct page *page,
406 void *freelist_old, unsigned long counters_old,
407 void *freelist_new, unsigned long counters_new,
408 const char *n)
409{
410#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
411 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
412 if (s->flags & __CMPXCHG_DOUBLE) {
413 if (cmpxchg_double(&page->freelist, &page->counters,
414 freelist_old, counters_old,
415 freelist_new, counters_new))
416 return 1;
417 } else
418#endif
419 {
420 unsigned long flags;
421
422 local_irq_save(flags);
423 slab_lock(page);
424 if (page->freelist == freelist_old && page->counters == counters_old) {
425 page->freelist = freelist_new;
426 page->counters = counters_new;
427 slab_unlock(page);
428 local_irq_restore(flags);
429 return 1;
430 }
431 slab_unlock(page);
432 local_irq_restore(flags);
433 }
434
435 cpu_relax();
436 stat(s, CMPXCHG_DOUBLE_FAIL);
437
438#ifdef SLUB_DEBUG_CMPXCHG
439 printk(KERN_INFO "%s %s: cmpxchg double redo ", n, s->name);
440#endif
441
442 return 0;
443}
444
445#ifdef CONFIG_SLUB_DEBUG
446/*
447 * Determine a map of object in use on a page.
448 *
449 * Node listlock must be held to guarantee that the page does
450 * not vanish from under us.
451 */
452static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
453{
454 void *p;
455 void *addr = page_address(page);
456
457 for (p = page->freelist; p; p = get_freepointer(s, p))
458 set_bit(slab_index(p, s, addr), map);
459}
460
461/*
462 * Debug settings:
463 */
464#ifdef CONFIG_SLUB_DEBUG_ON
465static int slub_debug = DEBUG_DEFAULT_FLAGS;
466#else
467static int slub_debug;
468#endif
469
470static char *slub_debug_slabs;
471static int disable_higher_order_debug;
472
473/*
474 * Object debugging
475 */
476static void print_section(char *text, u8 *addr, unsigned int length)
477{
478 print_hex_dump(KERN_ERR, text, DUMP_PREFIX_ADDRESS, 16, 1, addr,
479 length, 1);
480}
481
482static struct track *get_track(struct kmem_cache *s, void *object,
483 enum track_item alloc)
484{
485 struct track *p;
486
487 if (s->offset)
488 p = object + s->offset + sizeof(void *);
489 else
490 p = object + s->inuse;
491
492 return p + alloc;
493}
494
495static void set_track(struct kmem_cache *s, void *object,
496 enum track_item alloc, unsigned long addr)
497{
498 struct track *p = get_track(s, object, alloc);
499
500 if (addr) {
501#ifdef CONFIG_STACKTRACE
502 struct stack_trace trace;
503 int i;
504
505 trace.nr_entries = 0;
506 trace.max_entries = TRACK_ADDRS_COUNT;
507 trace.entries = p->addrs;
508 trace.skip = 3;
509 save_stack_trace(&trace);
510
511 /* See rant in lockdep.c */
512 if (trace.nr_entries != 0 &&
513 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
514 trace.nr_entries--;
515
516 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
517 p->addrs[i] = 0;
518#endif
519 p->addr = addr;
520 p->cpu = smp_processor_id();
521 p->pid = current->pid;
522 p->when = jiffies;
523 } else
524 memset(p, 0, sizeof(struct track));
525}
526
527static void init_tracking(struct kmem_cache *s, void *object)
528{
529 if (!(s->flags & SLAB_STORE_USER))
530 return;
531
532 set_track(s, object, TRACK_FREE, 0UL);
533 set_track(s, object, TRACK_ALLOC, 0UL);
534}
535
536static void print_track(const char *s, struct track *t)
537{
538 if (!t->addr)
539 return;
540
541 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
542 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
543#ifdef CONFIG_STACKTRACE
544 {
545 int i;
546 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
547 if (t->addrs[i])
548 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
549 else
550 break;
551 }
552#endif
553}
554
555static void print_tracking(struct kmem_cache *s, void *object)
556{
557 if (!(s->flags & SLAB_STORE_USER))
558 return;
559
560 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
561 print_track("Freed", get_track(s, object, TRACK_FREE));
562}
563
564static void print_page_info(struct page *page)
565{
566 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
567 page, page->objects, page->inuse, page->freelist, page->flags);
568
569}
570
571static void slab_bug(struct kmem_cache *s, char *fmt, ...)
572{
573 va_list args;
574 char buf[100];
575
576 va_start(args, fmt);
577 vsnprintf(buf, sizeof(buf), fmt, args);
578 va_end(args);
579 printk(KERN_ERR "========================================"
580 "=====================================\n");
581 printk(KERN_ERR "BUG %s (%s): %s\n", s->name, print_tainted(), buf);
582 printk(KERN_ERR "----------------------------------------"
583 "-------------------------------------\n\n");
584}
585
586static void slab_fix(struct kmem_cache *s, char *fmt, ...)
587{
588 va_list args;
589 char buf[100];
590
591 va_start(args, fmt);
592 vsnprintf(buf, sizeof(buf), fmt, args);
593 va_end(args);
594 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
595}
596
597static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
598{
599 unsigned int off; /* Offset of last byte */
600 u8 *addr = page_address(page);
601
602 print_tracking(s, p);
603
604 print_page_info(page);
605
606 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
607 p, p - addr, get_freepointer(s, p));
608
609 if (p > addr + 16)
610 print_section("Bytes b4 ", p - 16, 16);
611
612 print_section("Object ", p, min_t(unsigned long, s->objsize,
613 PAGE_SIZE));
614 if (s->flags & SLAB_RED_ZONE)
615 print_section("Redzone ", p + s->objsize,
616 s->inuse - s->objsize);
617
618 if (s->offset)
619 off = s->offset + sizeof(void *);
620 else
621 off = s->inuse;
622
623 if (s->flags & SLAB_STORE_USER)
624 off += 2 * sizeof(struct track);
625
626 if (off != s->size)
627 /* Beginning of the filler is the free pointer */
628 print_section("Padding ", p + off, s->size - off);
629
630 dump_stack();
631}
632
633static void object_err(struct kmem_cache *s, struct page *page,
634 u8 *object, char *reason)
635{
636 slab_bug(s, "%s", reason);
637 print_trailer(s, page, object);
638}
639
640static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
641{
642 va_list args;
643 char buf[100];
644
645 va_start(args, fmt);
646 vsnprintf(buf, sizeof(buf), fmt, args);
647 va_end(args);
648 slab_bug(s, "%s", buf);
649 print_page_info(page);
650 dump_stack();
651}
652
653static void init_object(struct kmem_cache *s, void *object, u8 val)
654{
655 u8 *p = object;
656
657 if (s->flags & __OBJECT_POISON) {
658 memset(p, POISON_FREE, s->objsize - 1);
659 p[s->objsize - 1] = POISON_END;
660 }
661
662 if (s->flags & SLAB_RED_ZONE)
663 memset(p + s->objsize, val, s->inuse - s->objsize);
664}
665
666static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
667 void *from, void *to)
668{
669 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
670 memset(from, data, to - from);
671}
672
673static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
674 u8 *object, char *what,
675 u8 *start, unsigned int value, unsigned int bytes)
676{
677 u8 *fault;
678 u8 *end;
679
680 fault = memchr_inv(start, value, bytes);
681 if (!fault)
682 return 1;
683
684 end = start + bytes;
685 while (end > fault && end[-1] == value)
686 end--;
687
688 slab_bug(s, "%s overwritten", what);
689 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
690 fault, end - 1, fault[0], value);
691 print_trailer(s, page, object);
692
693 restore_bytes(s, what, value, fault, end);
694 return 0;
695}
696
697/*
698 * Object layout:
699 *
700 * object address
701 * Bytes of the object to be managed.
702 * If the freepointer may overlay the object then the free
703 * pointer is the first word of the object.
704 *
705 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
706 * 0xa5 (POISON_END)
707 *
708 * object + s->objsize
709 * Padding to reach word boundary. This is also used for Redzoning.
710 * Padding is extended by another word if Redzoning is enabled and
711 * objsize == inuse.
712 *
713 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
714 * 0xcc (RED_ACTIVE) for objects in use.
715 *
716 * object + s->inuse
717 * Meta data starts here.
718 *
719 * A. Free pointer (if we cannot overwrite object on free)
720 * B. Tracking data for SLAB_STORE_USER
721 * C. Padding to reach required alignment boundary or at mininum
722 * one word if debugging is on to be able to detect writes
723 * before the word boundary.
724 *
725 * Padding is done using 0x5a (POISON_INUSE)
726 *
727 * object + s->size
728 * Nothing is used beyond s->size.
729 *
730 * If slabcaches are merged then the objsize and inuse boundaries are mostly
731 * ignored. And therefore no slab options that rely on these boundaries
732 * may be used with merged slabcaches.
733 */
734
735static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
736{
737 unsigned long off = s->inuse; /* The end of info */
738
739 if (s->offset)
740 /* Freepointer is placed after the object. */
741 off += sizeof(void *);
742
743 if (s->flags & SLAB_STORE_USER)
744 /* We also have user information there */
745 off += 2 * sizeof(struct track);
746
747 if (s->size == off)
748 return 1;
749
750 return check_bytes_and_report(s, page, p, "Object padding",
751 p + off, POISON_INUSE, s->size - off);
752}
753
754/* Check the pad bytes at the end of a slab page */
755static int slab_pad_check(struct kmem_cache *s, struct page *page)
756{
757 u8 *start;
758 u8 *fault;
759 u8 *end;
760 int length;
761 int remainder;
762
763 if (!(s->flags & SLAB_POISON))
764 return 1;
765
766 start = page_address(page);
767 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
768 end = start + length;
769 remainder = length % s->size;
770 if (!remainder)
771 return 1;
772
773 fault = memchr_inv(end - remainder, POISON_INUSE, remainder);
774 if (!fault)
775 return 1;
776 while (end > fault && end[-1] == POISON_INUSE)
777 end--;
778
779 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
780 print_section("Padding ", end - remainder, remainder);
781
782 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
783 return 0;
784}
785
786static int check_object(struct kmem_cache *s, struct page *page,
787 void *object, u8 val)
788{
789 u8 *p = object;
790 u8 *endobject = object + s->objsize;
791
792 if (s->flags & SLAB_RED_ZONE) {
793 if (!check_bytes_and_report(s, page, object, "Redzone",
794 endobject, val, s->inuse - s->objsize))
795 return 0;
796 } else {
797 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
798 check_bytes_and_report(s, page, p, "Alignment padding",
799 endobject, POISON_INUSE, s->inuse - s->objsize);
800 }
801 }
802
803 if (s->flags & SLAB_POISON) {
804 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
805 (!check_bytes_and_report(s, page, p, "Poison", p,
806 POISON_FREE, s->objsize - 1) ||
807 !check_bytes_and_report(s, page, p, "Poison",
808 p + s->objsize - 1, POISON_END, 1)))
809 return 0;
810 /*
811 * check_pad_bytes cleans up on its own.
812 */
813 check_pad_bytes(s, page, p);
814 }
815
816 if (!s->offset && val == SLUB_RED_ACTIVE)
817 /*
818 * Object and freepointer overlap. Cannot check
819 * freepointer while object is allocated.
820 */
821 return 1;
822
823 /* Check free pointer validity */
824 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
825 object_err(s, page, p, "Freepointer corrupt");
826 /*
827 * No choice but to zap it and thus lose the remainder
828 * of the free objects in this slab. May cause
829 * another error because the object count is now wrong.
830 */
831 set_freepointer(s, p, NULL);
832 return 0;
833 }
834 return 1;
835}
836
837static int check_slab(struct kmem_cache *s, struct page *page)
838{
839 int maxobj;
840
841 VM_BUG_ON(!irqs_disabled());
842
843 if (!PageSlab(page)) {
844 slab_err(s, page, "Not a valid slab page");
845 return 0;
846 }
847
848 maxobj = order_objects(compound_order(page), s->size, s->reserved);
849 if (page->objects > maxobj) {
850 slab_err(s, page, "objects %u > max %u",
851 s->name, page->objects, maxobj);
852 return 0;
853 }
854 if (page->inuse > page->objects) {
855 slab_err(s, page, "inuse %u > max %u",
856 s->name, page->inuse, page->objects);
857 return 0;
858 }
859 /* Slab_pad_check fixes things up after itself */
860 slab_pad_check(s, page);
861 return 1;
862}
863
864/*
865 * Determine if a certain object on a page is on the freelist. Must hold the
866 * slab lock to guarantee that the chains are in a consistent state.
867 */
868static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
869{
870 int nr = 0;
871 void *fp;
872 void *object = NULL;
873 unsigned long max_objects;
874
875 fp = page->freelist;
876 while (fp && nr <= page->objects) {
877 if (fp == search)
878 return 1;
879 if (!check_valid_pointer(s, page, fp)) {
880 if (object) {
881 object_err(s, page, object,
882 "Freechain corrupt");
883 set_freepointer(s, object, NULL);
884 break;
885 } else {
886 slab_err(s, page, "Freepointer corrupt");
887 page->freelist = NULL;
888 page->inuse = page->objects;
889 slab_fix(s, "Freelist cleared");
890 return 0;
891 }
892 break;
893 }
894 object = fp;
895 fp = get_freepointer(s, object);
896 nr++;
897 }
898
899 max_objects = order_objects(compound_order(page), s->size, s->reserved);
900 if (max_objects > MAX_OBJS_PER_PAGE)
901 max_objects = MAX_OBJS_PER_PAGE;
902
903 if (page->objects != max_objects) {
904 slab_err(s, page, "Wrong number of objects. Found %d but "
905 "should be %d", page->objects, max_objects);
906 page->objects = max_objects;
907 slab_fix(s, "Number of objects adjusted.");
908 }
909 if (page->inuse != page->objects - nr) {
910 slab_err(s, page, "Wrong object count. Counter is %d but "
911 "counted were %d", page->inuse, page->objects - nr);
912 page->inuse = page->objects - nr;
913 slab_fix(s, "Object count adjusted.");
914 }
915 return search == NULL;
916}
917
918static void trace(struct kmem_cache *s, struct page *page, void *object,
919 int alloc)
920{
921 if (s->flags & SLAB_TRACE) {
922 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
923 s->name,
924 alloc ? "alloc" : "free",
925 object, page->inuse,
926 page->freelist);
927
928 if (!alloc)
929 print_section("Object ", (void *)object, s->objsize);
930
931 dump_stack();
932 }
933}
934
935/*
936 * Hooks for other subsystems that check memory allocations. In a typical
937 * production configuration these hooks all should produce no code at all.
938 */
939static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
940{
941 flags &= gfp_allowed_mask;
942 lockdep_trace_alloc(flags);
943 might_sleep_if(flags & __GFP_WAIT);
944
945 return should_failslab(s->objsize, flags, s->flags);
946}
947
948static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
949{
950 flags &= gfp_allowed_mask;
951 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
952 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
953}
954
955static inline void slab_free_hook(struct kmem_cache *s, void *x)
956{
957 kmemleak_free_recursive(x, s->flags);
958
959 /*
960 * Trouble is that we may no longer disable interupts in the fast path
961 * So in order to make the debug calls that expect irqs to be
962 * disabled we need to disable interrupts temporarily.
963 */
964#if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
965 {
966 unsigned long flags;
967
968 local_irq_save(flags);
969 kmemcheck_slab_free(s, x, s->objsize);
970 debug_check_no_locks_freed(x, s->objsize);
971 local_irq_restore(flags);
972 }
973#endif
974 if (!(s->flags & SLAB_DEBUG_OBJECTS))
975 debug_check_no_obj_freed(x, s->objsize);
976}
977
978/*
979 * Tracking of fully allocated slabs for debugging purposes.
980 *
981 * list_lock must be held.
982 */
983static void add_full(struct kmem_cache *s,
984 struct kmem_cache_node *n, struct page *page)
985{
986 if (!(s->flags & SLAB_STORE_USER))
987 return;
988
989 list_add(&page->lru, &n->full);
990}
991
992/*
993 * list_lock must be held.
994 */
995static void remove_full(struct kmem_cache *s, struct page *page)
996{
997 if (!(s->flags & SLAB_STORE_USER))
998 return;
999
1000 list_del(&page->lru);
1001}
1002
1003/* Tracking of the number of slabs for debugging purposes */
1004static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1005{
1006 struct kmem_cache_node *n = get_node(s, node);
1007
1008 return atomic_long_read(&n->nr_slabs);
1009}
1010
1011static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1012{
1013 return atomic_long_read(&n->nr_slabs);
1014}
1015
1016static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
1017{
1018 struct kmem_cache_node *n = get_node(s, node);
1019
1020 /*
1021 * May be called early in order to allocate a slab for the
1022 * kmem_cache_node structure. Solve the chicken-egg
1023 * dilemma by deferring the increment of the count during
1024 * bootstrap (see early_kmem_cache_node_alloc).
1025 */
1026 if (n) {
1027 atomic_long_inc(&n->nr_slabs);
1028 atomic_long_add(objects, &n->total_objects);
1029 }
1030}
1031static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1032{
1033 struct kmem_cache_node *n = get_node(s, node);
1034
1035 atomic_long_dec(&n->nr_slabs);
1036 atomic_long_sub(objects, &n->total_objects);
1037}
1038
1039/* Object debug checks for alloc/free paths */
1040static void setup_object_debug(struct kmem_cache *s, struct page *page,
1041 void *object)
1042{
1043 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1044 return;
1045
1046 init_object(s, object, SLUB_RED_INACTIVE);
1047 init_tracking(s, object);
1048}
1049
1050static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1051 void *object, unsigned long addr)
1052{
1053 if (!check_slab(s, page))
1054 goto bad;
1055
1056 if (!check_valid_pointer(s, page, object)) {
1057 object_err(s, page, object, "Freelist Pointer check fails");
1058 goto bad;
1059 }
1060
1061 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1062 goto bad;
1063
1064 /* Success perform special debug activities for allocs */
1065 if (s->flags & SLAB_STORE_USER)
1066 set_track(s, object, TRACK_ALLOC, addr);
1067 trace(s, page, object, 1);
1068 init_object(s, object, SLUB_RED_ACTIVE);
1069 return 1;
1070
1071bad:
1072 if (PageSlab(page)) {
1073 /*
1074 * If this is a slab page then lets do the best we can
1075 * to avoid issues in the future. Marking all objects
1076 * as used avoids touching the remaining objects.
1077 */
1078 slab_fix(s, "Marking all objects used");
1079 page->inuse = page->objects;
1080 page->freelist = NULL;
1081 }
1082 return 0;
1083}
1084
1085static noinline int free_debug_processing(struct kmem_cache *s,
1086 struct page *page, void *object, unsigned long addr)
1087{
1088 unsigned long flags;
1089 int rc = 0;
1090
1091 local_irq_save(flags);
1092 slab_lock(page);
1093
1094 if (!check_slab(s, page))
1095 goto fail;
1096
1097 if (!check_valid_pointer(s, page, object)) {
1098 slab_err(s, page, "Invalid object pointer 0x%p", object);
1099 goto fail;
1100 }
1101
1102 if (on_freelist(s, page, object)) {
1103 object_err(s, page, object, "Object already free");
1104 goto fail;
1105 }
1106
1107 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1108 goto out;
1109
1110 if (unlikely(s != page->slab)) {
1111 if (!PageSlab(page)) {
1112 slab_err(s, page, "Attempt to free object(0x%p) "
1113 "outside of slab", object);
1114 } else if (!page->slab) {
1115 printk(KERN_ERR
1116 "SLUB <none>: no slab for object 0x%p.\n",
1117 object);
1118 dump_stack();
1119 } else
1120 object_err(s, page, object,
1121 "page slab pointer corrupt.");
1122 goto fail;
1123 }
1124
1125 if (s->flags & SLAB_STORE_USER)
1126 set_track(s, object, TRACK_FREE, addr);
1127 trace(s, page, object, 0);
1128 init_object(s, object, SLUB_RED_INACTIVE);
1129 rc = 1;
1130out:
1131 slab_unlock(page);
1132 local_irq_restore(flags);
1133 return rc;
1134
1135fail:
1136 slab_fix(s, "Object at 0x%p not freed", object);
1137 goto out;
1138}
1139
1140static int __init setup_slub_debug(char *str)
1141{
1142 slub_debug = DEBUG_DEFAULT_FLAGS;
1143 if (*str++ != '=' || !*str)
1144 /*
1145 * No options specified. Switch on full debugging.
1146 */
1147 goto out;
1148
1149 if (*str == ',')
1150 /*
1151 * No options but restriction on slabs. This means full
1152 * debugging for slabs matching a pattern.
1153 */
1154 goto check_slabs;
1155
1156 if (tolower(*str) == 'o') {
1157 /*
1158 * Avoid enabling debugging on caches if its minimum order
1159 * would increase as a result.
1160 */
1161 disable_higher_order_debug = 1;
1162 goto out;
1163 }
1164
1165 slub_debug = 0;
1166 if (*str == '-')
1167 /*
1168 * Switch off all debugging measures.
1169 */
1170 goto out;
1171
1172 /*
1173 * Determine which debug features should be switched on
1174 */
1175 for (; *str && *str != ','; str++) {
1176 switch (tolower(*str)) {
1177 case 'f':
1178 slub_debug |= SLAB_DEBUG_FREE;
1179 break;
1180 case 'z':
1181 slub_debug |= SLAB_RED_ZONE;
1182 break;
1183 case 'p':
1184 slub_debug |= SLAB_POISON;
1185 break;
1186 case 'u':
1187 slub_debug |= SLAB_STORE_USER;
1188 break;
1189 case 't':
1190 slub_debug |= SLAB_TRACE;
1191 break;
1192 case 'a':
1193 slub_debug |= SLAB_FAILSLAB;
1194 break;
1195 default:
1196 printk(KERN_ERR "slub_debug option '%c' "
1197 "unknown. skipped\n", *str);
1198 }
1199 }
1200
1201check_slabs:
1202 if (*str == ',')
1203 slub_debug_slabs = str + 1;
1204out:
1205 return 1;
1206}
1207
1208__setup("slub_debug", setup_slub_debug);
1209
1210static unsigned long kmem_cache_flags(unsigned long objsize,
1211 unsigned long flags, const char *name,
1212 void (*ctor)(void *))
1213{
1214 /*
1215 * Enable debugging if selected on the kernel commandline.
1216 */
1217 if (slub_debug && (!slub_debug_slabs ||
1218 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1219 flags |= slub_debug;
1220
1221 return flags;
1222}
1223#else
1224static inline void setup_object_debug(struct kmem_cache *s,
1225 struct page *page, void *object) {}
1226
1227static inline int alloc_debug_processing(struct kmem_cache *s,
1228 struct page *page, void *object, unsigned long addr) { return 0; }
1229
1230static inline int free_debug_processing(struct kmem_cache *s,
1231 struct page *page, void *object, unsigned long addr) { return 0; }
1232
1233static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1234 { return 1; }
1235static inline int check_object(struct kmem_cache *s, struct page *page,
1236 void *object, u8 val) { return 1; }
1237static inline void add_full(struct kmem_cache *s, struct kmem_cache_node *n,
1238 struct page *page) {}
1239static inline void remove_full(struct kmem_cache *s, struct page *page) {}
1240static inline unsigned long kmem_cache_flags(unsigned long objsize,
1241 unsigned long flags, const char *name,
1242 void (*ctor)(void *))
1243{
1244 return flags;
1245}
1246#define slub_debug 0
1247
1248#define disable_higher_order_debug 0
1249
1250static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1251 { return 0; }
1252static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1253 { return 0; }
1254static inline void inc_slabs_node(struct kmem_cache *s, int node,
1255 int objects) {}
1256static inline void dec_slabs_node(struct kmem_cache *s, int node,
1257 int objects) {}
1258
1259static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1260 { return 0; }
1261
1262static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1263 void *object) {}
1264
1265static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1266
1267#endif /* CONFIG_SLUB_DEBUG */
1268
1269/*
1270 * Slab allocation and freeing
1271 */
1272static inline struct page *alloc_slab_page(gfp_t flags, int node,
1273 struct kmem_cache_order_objects oo)
1274{
1275 int order = oo_order(oo);
1276
1277 flags |= __GFP_NOTRACK;
1278
1279 if (node == NUMA_NO_NODE)
1280 return alloc_pages(flags, order);
1281 else
1282 return alloc_pages_exact_node(node, flags, order);
1283}
1284
1285static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1286{
1287 struct page *page;
1288 struct kmem_cache_order_objects oo = s->oo;
1289 gfp_t alloc_gfp;
1290
1291 flags &= gfp_allowed_mask;
1292
1293 if (flags & __GFP_WAIT)
1294 local_irq_enable();
1295
1296 flags |= s->allocflags;
1297
1298 /*
1299 * Let the initial higher-order allocation fail under memory pressure
1300 * so we fall-back to the minimum order allocation.
1301 */
1302 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1303
1304 page = alloc_slab_page(alloc_gfp, node, oo);
1305 if (unlikely(!page)) {
1306 oo = s->min;
1307 /*
1308 * Allocation may have failed due to fragmentation.
1309 * Try a lower order alloc if possible
1310 */
1311 page = alloc_slab_page(flags, node, oo);
1312
1313 if (page)
1314 stat(s, ORDER_FALLBACK);
1315 }
1316
1317 if (flags & __GFP_WAIT)
1318 local_irq_disable();
1319
1320 if (!page)
1321 return NULL;
1322
1323 if (kmemcheck_enabled
1324 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1325 int pages = 1 << oo_order(oo);
1326
1327 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1328
1329 /*
1330 * Objects from caches that have a constructor don't get
1331 * cleared when they're allocated, so we need to do it here.
1332 */
1333 if (s->ctor)
1334 kmemcheck_mark_uninitialized_pages(page, pages);
1335 else
1336 kmemcheck_mark_unallocated_pages(page, pages);
1337 }
1338
1339 page->objects = oo_objects(oo);
1340 mod_zone_page_state(page_zone(page),
1341 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1342 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1343 1 << oo_order(oo));
1344
1345 return page;
1346}
1347
1348static void setup_object(struct kmem_cache *s, struct page *page,
1349 void *object)
1350{
1351 setup_object_debug(s, page, object);
1352 if (unlikely(s->ctor))
1353 s->ctor(object);
1354}
1355
1356static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1357{
1358 struct page *page;
1359 void *start;
1360 void *last;
1361 void *p;
1362
1363 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1364
1365 page = allocate_slab(s,
1366 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1367 if (!page)
1368 goto out;
1369
1370 inc_slabs_node(s, page_to_nid(page), page->objects);
1371 page->slab = s;
1372 __SetPageSlab(page);
1373
1374 start = page_address(page);
1375
1376 if (unlikely(s->flags & SLAB_POISON))
1377 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1378
1379 last = start;
1380 for_each_object(p, s, start, page->objects) {
1381 setup_object(s, page, last);
1382 set_freepointer(s, last, p);
1383 last = p;
1384 }
1385 setup_object(s, page, last);
1386 set_freepointer(s, last, NULL);
1387
1388 page->freelist = start;
1389 page->inuse = page->objects;
1390 page->frozen = 1;
1391out:
1392 return page;
1393}
1394
1395static void __free_slab(struct kmem_cache *s, struct page *page)
1396{
1397 int order = compound_order(page);
1398 int pages = 1 << order;
1399
1400 if (kmem_cache_debug(s)) {
1401 void *p;
1402
1403 slab_pad_check(s, page);
1404 for_each_object(p, s, page_address(page),
1405 page->objects)
1406 check_object(s, page, p, SLUB_RED_INACTIVE);
1407 }
1408
1409 kmemcheck_free_shadow(page, compound_order(page));
1410
1411 mod_zone_page_state(page_zone(page),
1412 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1413 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1414 -pages);
1415
1416 __ClearPageSlab(page);
1417 reset_page_mapcount(page);
1418 if (current->reclaim_state)
1419 current->reclaim_state->reclaimed_slab += pages;
1420 __free_pages(page, order);
1421}
1422
1423#define need_reserve_slab_rcu \
1424 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1425
1426static void rcu_free_slab(struct rcu_head *h)
1427{
1428 struct page *page;
1429
1430 if (need_reserve_slab_rcu)
1431 page = virt_to_head_page(h);
1432 else
1433 page = container_of((struct list_head *)h, struct page, lru);
1434
1435 __free_slab(page->slab, page);
1436}
1437
1438static void free_slab(struct kmem_cache *s, struct page *page)
1439{
1440 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1441 struct rcu_head *head;
1442
1443 if (need_reserve_slab_rcu) {
1444 int order = compound_order(page);
1445 int offset = (PAGE_SIZE << order) - s->reserved;
1446
1447 VM_BUG_ON(s->reserved != sizeof(*head));
1448 head = page_address(page) + offset;
1449 } else {
1450 /*
1451 * RCU free overloads the RCU head over the LRU
1452 */
1453 head = (void *)&page->lru;
1454 }
1455
1456 call_rcu(head, rcu_free_slab);
1457 } else
1458 __free_slab(s, page);
1459}
1460
1461static void discard_slab(struct kmem_cache *s, struct page *page)
1462{
1463 dec_slabs_node(s, page_to_nid(page), page->objects);
1464 free_slab(s, page);
1465}
1466
1467/*
1468 * Management of partially allocated slabs.
1469 *
1470 * list_lock must be held.
1471 */
1472static inline void add_partial(struct kmem_cache_node *n,
1473 struct page *page, int tail)
1474{
1475 n->nr_partial++;
1476 if (tail == DEACTIVATE_TO_TAIL)
1477 list_add_tail(&page->lru, &n->partial);
1478 else
1479 list_add(&page->lru, &n->partial);
1480}
1481
1482/*
1483 * list_lock must be held.
1484 */
1485static inline void remove_partial(struct kmem_cache_node *n,
1486 struct page *page)
1487{
1488 list_del(&page->lru);
1489 n->nr_partial--;
1490}
1491
1492/*
1493 * Lock slab, remove from the partial list and put the object into the
1494 * per cpu freelist.
1495 *
1496 * Returns a list of objects or NULL if it fails.
1497 *
1498 * Must hold list_lock.
1499 */
1500static inline void *acquire_slab(struct kmem_cache *s,
1501 struct kmem_cache_node *n, struct page *page,
1502 int mode)
1503{
1504 void *freelist;
1505 unsigned long counters;
1506 struct page new;
1507
1508 /*
1509 * Zap the freelist and set the frozen bit.
1510 * The old freelist is the list of objects for the
1511 * per cpu allocation list.
1512 */
1513 do {
1514 freelist = page->freelist;
1515 counters = page->counters;
1516 new.counters = counters;
1517 if (mode) {
1518 new.inuse = page->objects;
1519 new.freelist = NULL;
1520 } else {
1521 new.freelist = freelist;
1522 }
1523
1524 VM_BUG_ON(new.frozen);
1525 new.frozen = 1;
1526
1527 } while (!__cmpxchg_double_slab(s, page,
1528 freelist, counters,
1529 new.freelist, new.counters,
1530 "lock and freeze"));
1531
1532 remove_partial(n, page);
1533 return freelist;
1534}
1535
1536static int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain);
1537
1538/*
1539 * Try to allocate a partial slab from a specific node.
1540 */
1541static void *get_partial_node(struct kmem_cache *s,
1542 struct kmem_cache_node *n, struct kmem_cache_cpu *c)
1543{
1544 struct page *page, *page2;
1545 void *object = NULL;
1546
1547 /*
1548 * Racy check. If we mistakenly see no partial slabs then we
1549 * just allocate an empty slab. If we mistakenly try to get a
1550 * partial slab and there is none available then get_partials()
1551 * will return NULL.
1552 */
1553 if (!n || !n->nr_partial)
1554 return NULL;
1555
1556 spin_lock(&n->list_lock);
1557 list_for_each_entry_safe(page, page2, &n->partial, lru) {
1558 void *t = acquire_slab(s, n, page, object == NULL);
1559 int available;
1560
1561 if (!t)
1562 break;
1563
1564 if (!object) {
1565 c->page = page;
1566 c->node = page_to_nid(page);
1567 stat(s, ALLOC_FROM_PARTIAL);
1568 object = t;
1569 available = page->objects - page->inuse;
1570 } else {
1571 available = put_cpu_partial(s, page, 0);
1572 stat(s, CPU_PARTIAL_NODE);
1573 }
1574 if (kmem_cache_debug(s) || available > s->cpu_partial / 2)
1575 break;
1576
1577 }
1578 spin_unlock(&n->list_lock);
1579 return object;
1580}
1581
1582/*
1583 * Get a page from somewhere. Search in increasing NUMA distances.
1584 */
1585static void *get_any_partial(struct kmem_cache *s, gfp_t flags,
1586 struct kmem_cache_cpu *c)
1587{
1588#ifdef CONFIG_NUMA
1589 struct zonelist *zonelist;
1590 struct zoneref *z;
1591 struct zone *zone;
1592 enum zone_type high_zoneidx = gfp_zone(flags);
1593 void *object;
1594 unsigned int cpuset_mems_cookie;
1595
1596 /*
1597 * The defrag ratio allows a configuration of the tradeoffs between
1598 * inter node defragmentation and node local allocations. A lower
1599 * defrag_ratio increases the tendency to do local allocations
1600 * instead of attempting to obtain partial slabs from other nodes.
1601 *
1602 * If the defrag_ratio is set to 0 then kmalloc() always
1603 * returns node local objects. If the ratio is higher then kmalloc()
1604 * may return off node objects because partial slabs are obtained
1605 * from other nodes and filled up.
1606 *
1607 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1608 * defrag_ratio = 1000) then every (well almost) allocation will
1609 * first attempt to defrag slab caches on other nodes. This means
1610 * scanning over all nodes to look for partial slabs which may be
1611 * expensive if we do it every time we are trying to find a slab
1612 * with available objects.
1613 */
1614 if (!s->remote_node_defrag_ratio ||
1615 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1616 return NULL;
1617
1618 do {
1619 cpuset_mems_cookie = get_mems_allowed();
1620 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1621 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1622 struct kmem_cache_node *n;
1623
1624 n = get_node(s, zone_to_nid(zone));
1625
1626 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1627 n->nr_partial > s->min_partial) {
1628 object = get_partial_node(s, n, c);
1629 if (object) {
1630 /*
1631 * Return the object even if
1632 * put_mems_allowed indicated that
1633 * the cpuset mems_allowed was
1634 * updated in parallel. It's a
1635 * harmless race between the alloc
1636 * and the cpuset update.
1637 */
1638 put_mems_allowed(cpuset_mems_cookie);
1639 return object;
1640 }
1641 }
1642 }
1643 } while (!put_mems_allowed(cpuset_mems_cookie));
1644#endif
1645 return NULL;
1646}
1647
1648/*
1649 * Get a partial page, lock it and return it.
1650 */
1651static void *get_partial(struct kmem_cache *s, gfp_t flags, int node,
1652 struct kmem_cache_cpu *c)
1653{
1654 void *object;
1655 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1656
1657 object = get_partial_node(s, get_node(s, searchnode), c);
1658 if (object || node != NUMA_NO_NODE)
1659 return object;
1660
1661 return get_any_partial(s, flags, c);
1662}
1663
1664#ifdef CONFIG_PREEMPT
1665/*
1666 * Calculate the next globally unique transaction for disambiguiation
1667 * during cmpxchg. The transactions start with the cpu number and are then
1668 * incremented by CONFIG_NR_CPUS.
1669 */
1670#define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1671#else
1672/*
1673 * No preemption supported therefore also no need to check for
1674 * different cpus.
1675 */
1676#define TID_STEP 1
1677#endif
1678
1679static inline unsigned long next_tid(unsigned long tid)
1680{
1681 return tid + TID_STEP;
1682}
1683
1684static inline unsigned int tid_to_cpu(unsigned long tid)
1685{
1686 return tid % TID_STEP;
1687}
1688
1689static inline unsigned long tid_to_event(unsigned long tid)
1690{
1691 return tid / TID_STEP;
1692}
1693
1694static inline unsigned int init_tid(int cpu)
1695{
1696 return cpu;
1697}
1698
1699static inline void note_cmpxchg_failure(const char *n,
1700 const struct kmem_cache *s, unsigned long tid)
1701{
1702#ifdef SLUB_DEBUG_CMPXCHG
1703 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1704
1705 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1706
1707#ifdef CONFIG_PREEMPT
1708 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1709 printk("due to cpu change %d -> %d\n",
1710 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1711 else
1712#endif
1713 if (tid_to_event(tid) != tid_to_event(actual_tid))
1714 printk("due to cpu running other code. Event %ld->%ld\n",
1715 tid_to_event(tid), tid_to_event(actual_tid));
1716 else
1717 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1718 actual_tid, tid, next_tid(tid));
1719#endif
1720 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1721}
1722
1723void init_kmem_cache_cpus(struct kmem_cache *s)
1724{
1725 int cpu;
1726
1727 for_each_possible_cpu(cpu)
1728 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1729}
1730
1731/*
1732 * Remove the cpu slab
1733 */
1734static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1735{
1736 enum slab_modes { M_NONE, M_PARTIAL, M_FULL, M_FREE };
1737 struct page *page = c->page;
1738 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1739 int lock = 0;
1740 enum slab_modes l = M_NONE, m = M_NONE;
1741 void *freelist;
1742 void *nextfree;
1743 int tail = DEACTIVATE_TO_HEAD;
1744 struct page new;
1745 struct page old;
1746
1747 if (page->freelist) {
1748 stat(s, DEACTIVATE_REMOTE_FREES);
1749 tail = DEACTIVATE_TO_TAIL;
1750 }
1751
1752 c->tid = next_tid(c->tid);
1753 c->page = NULL;
1754 freelist = c->freelist;
1755 c->freelist = NULL;
1756
1757 /*
1758 * Stage one: Free all available per cpu objects back
1759 * to the page freelist while it is still frozen. Leave the
1760 * last one.
1761 *
1762 * There is no need to take the list->lock because the page
1763 * is still frozen.
1764 */
1765 while (freelist && (nextfree = get_freepointer(s, freelist))) {
1766 void *prior;
1767 unsigned long counters;
1768
1769 do {
1770 prior = page->freelist;
1771 counters = page->counters;
1772 set_freepointer(s, freelist, prior);
1773 new.counters = counters;
1774 new.inuse--;
1775 VM_BUG_ON(!new.frozen);
1776
1777 } while (!__cmpxchg_double_slab(s, page,
1778 prior, counters,
1779 freelist, new.counters,
1780 "drain percpu freelist"));
1781
1782 freelist = nextfree;
1783 }
1784
1785 /*
1786 * Stage two: Ensure that the page is unfrozen while the
1787 * list presence reflects the actual number of objects
1788 * during unfreeze.
1789 *
1790 * We setup the list membership and then perform a cmpxchg
1791 * with the count. If there is a mismatch then the page
1792 * is not unfrozen but the page is on the wrong list.
1793 *
1794 * Then we restart the process which may have to remove
1795 * the page from the list that we just put it on again
1796 * because the number of objects in the slab may have
1797 * changed.
1798 */
1799redo:
1800
1801 old.freelist = page->freelist;
1802 old.counters = page->counters;
1803 VM_BUG_ON(!old.frozen);
1804
1805 /* Determine target state of the slab */
1806 new.counters = old.counters;
1807 if (freelist) {
1808 new.inuse--;
1809 set_freepointer(s, freelist, old.freelist);
1810 new.freelist = freelist;
1811 } else
1812 new.freelist = old.freelist;
1813
1814 new.frozen = 0;
1815
1816 if (!new.inuse && n->nr_partial > s->min_partial)
1817 m = M_FREE;
1818 else if (new.freelist) {
1819 m = M_PARTIAL;
1820 if (!lock) {
1821 lock = 1;
1822 /*
1823 * Taking the spinlock removes the possiblity
1824 * that acquire_slab() will see a slab page that
1825 * is frozen
1826 */
1827 spin_lock(&n->list_lock);
1828 }
1829 } else {
1830 m = M_FULL;
1831 if (kmem_cache_debug(s) && !lock) {
1832 lock = 1;
1833 /*
1834 * This also ensures that the scanning of full
1835 * slabs from diagnostic functions will not see
1836 * any frozen slabs.
1837 */
1838 spin_lock(&n->list_lock);
1839 }
1840 }
1841
1842 if (l != m) {
1843
1844 if (l == M_PARTIAL)
1845
1846 remove_partial(n, page);
1847
1848 else if (l == M_FULL)
1849
1850 remove_full(s, page);
1851
1852 if (m == M_PARTIAL) {
1853
1854 add_partial(n, page, tail);
1855 stat(s, tail);
1856
1857 } else if (m == M_FULL) {
1858
1859 stat(s, DEACTIVATE_FULL);
1860 add_full(s, n, page);
1861
1862 }
1863 }
1864
1865 l = m;
1866 if (!__cmpxchg_double_slab(s, page,
1867 old.freelist, old.counters,
1868 new.freelist, new.counters,
1869 "unfreezing slab"))
1870 goto redo;
1871
1872 if (lock)
1873 spin_unlock(&n->list_lock);
1874
1875 if (m == M_FREE) {
1876 stat(s, DEACTIVATE_EMPTY);
1877 discard_slab(s, page);
1878 stat(s, FREE_SLAB);
1879 }
1880}
1881
1882/* Unfreeze all the cpu partial slabs */
1883static void unfreeze_partials(struct kmem_cache *s)
1884{
1885 struct kmem_cache_node *n = NULL;
1886 struct kmem_cache_cpu *c = this_cpu_ptr(s->cpu_slab);
1887 struct page *page, *discard_page = NULL;
1888
1889 while ((page = c->partial)) {
1890 enum slab_modes { M_PARTIAL, M_FREE };
1891 enum slab_modes l, m;
1892 struct page new;
1893 struct page old;
1894
1895 c->partial = page->next;
1896 l = M_FREE;
1897
1898 do {
1899
1900 old.freelist = page->freelist;
1901 old.counters = page->counters;
1902 VM_BUG_ON(!old.frozen);
1903
1904 new.counters = old.counters;
1905 new.freelist = old.freelist;
1906
1907 new.frozen = 0;
1908
1909 if (!new.inuse && (!n || n->nr_partial > s->min_partial))
1910 m = M_FREE;
1911 else {
1912 struct kmem_cache_node *n2 = get_node(s,
1913 page_to_nid(page));
1914
1915 m = M_PARTIAL;
1916 if (n != n2) {
1917 if (n)
1918 spin_unlock(&n->list_lock);
1919
1920 n = n2;
1921 spin_lock(&n->list_lock);
1922 }
1923 }
1924
1925 if (l != m) {
1926 if (l == M_PARTIAL) {
1927 remove_partial(n, page);
1928 stat(s, FREE_REMOVE_PARTIAL);
1929 } else {
1930 add_partial(n, page,
1931 DEACTIVATE_TO_TAIL);
1932 stat(s, FREE_ADD_PARTIAL);
1933 }
1934
1935 l = m;
1936 }
1937
1938 } while (!cmpxchg_double_slab(s, page,
1939 old.freelist, old.counters,
1940 new.freelist, new.counters,
1941 "unfreezing slab"));
1942
1943 if (m == M_FREE) {
1944 page->next = discard_page;
1945 discard_page = page;
1946 }
1947 }
1948
1949 if (n)
1950 spin_unlock(&n->list_lock);
1951
1952 while (discard_page) {
1953 page = discard_page;
1954 discard_page = discard_page->next;
1955
1956 stat(s, DEACTIVATE_EMPTY);
1957 discard_slab(s, page);
1958 stat(s, FREE_SLAB);
1959 }
1960}
1961
1962/*
1963 * Put a page that was just frozen (in __slab_free) into a partial page
1964 * slot if available. This is done without interrupts disabled and without
1965 * preemption disabled. The cmpxchg is racy and may put the partial page
1966 * onto a random cpus partial slot.
1967 *
1968 * If we did not find a slot then simply move all the partials to the
1969 * per node partial list.
1970 */
1971int put_cpu_partial(struct kmem_cache *s, struct page *page, int drain)
1972{
1973 struct page *oldpage;
1974 int pages;
1975 int pobjects;
1976
1977 do {
1978 pages = 0;
1979 pobjects = 0;
1980 oldpage = this_cpu_read(s->cpu_slab->partial);
1981
1982 if (oldpage) {
1983 pobjects = oldpage->pobjects;
1984 pages = oldpage->pages;
1985 if (drain && pobjects > s->cpu_partial) {
1986 unsigned long flags;
1987 /*
1988 * partial array is full. Move the existing
1989 * set to the per node partial list.
1990 */
1991 local_irq_save(flags);
1992 unfreeze_partials(s);
1993 local_irq_restore(flags);
1994 pobjects = 0;
1995 pages = 0;
1996 stat(s, CPU_PARTIAL_DRAIN);
1997 }
1998 }
1999
2000 pages++;
2001 pobjects += page->objects - page->inuse;
2002
2003 page->pages = pages;
2004 page->pobjects = pobjects;
2005 page->next = oldpage;
2006
2007 } while (this_cpu_cmpxchg(s->cpu_slab->partial, oldpage, page) != oldpage);
2008 return pobjects;
2009}
2010
2011static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
2012{
2013 stat(s, CPUSLAB_FLUSH);
2014 deactivate_slab(s, c);
2015}
2016
2017/*
2018 * Flush cpu slab.
2019 *
2020 * Called from IPI handler with interrupts disabled.
2021 */
2022static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
2023{
2024 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2025
2026 if (likely(c)) {
2027 if (c->page)
2028 flush_slab(s, c);
2029
2030 unfreeze_partials(s);
2031 }
2032}
2033
2034static void flush_cpu_slab(void *d)
2035{
2036 struct kmem_cache *s = d;
2037
2038 __flush_cpu_slab(s, smp_processor_id());
2039}
2040
2041static bool has_cpu_slab(int cpu, void *info)
2042{
2043 struct kmem_cache *s = info;
2044 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
2045
2046 return c->page || c->partial;
2047}
2048
2049static void flush_all(struct kmem_cache *s)
2050{
2051 on_each_cpu_cond(has_cpu_slab, flush_cpu_slab, s, 1, GFP_ATOMIC);
2052}
2053
2054/*
2055 * Check if the objects in a per cpu structure fit numa
2056 * locality expectations.
2057 */
2058static inline int node_match(struct kmem_cache_cpu *c, int node)
2059{
2060#ifdef CONFIG_NUMA
2061 if (node != NUMA_NO_NODE && c->node != node)
2062 return 0;
2063#endif
2064 return 1;
2065}
2066
2067static int count_free(struct page *page)
2068{
2069 return page->objects - page->inuse;
2070}
2071
2072static unsigned long count_partial(struct kmem_cache_node *n,
2073 int (*get_count)(struct page *))
2074{
2075 unsigned long flags;
2076 unsigned long x = 0;
2077 struct page *page;
2078
2079 spin_lock_irqsave(&n->list_lock, flags);
2080 list_for_each_entry(page, &n->partial, lru)
2081 x += get_count(page);
2082 spin_unlock_irqrestore(&n->list_lock, flags);
2083 return x;
2084}
2085
2086static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
2087{
2088#ifdef CONFIG_SLUB_DEBUG
2089 return atomic_long_read(&n->total_objects);
2090#else
2091 return 0;
2092#endif
2093}
2094
2095static noinline void
2096slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
2097{
2098 int node;
2099
2100 printk(KERN_WARNING
2101 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
2102 nid, gfpflags);
2103 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
2104 "default order: %d, min order: %d\n", s->name, s->objsize,
2105 s->size, oo_order(s->oo), oo_order(s->min));
2106
2107 if (oo_order(s->min) > get_order(s->objsize))
2108 printk(KERN_WARNING " %s debugging increased min order, use "
2109 "slub_debug=O to disable.\n", s->name);
2110
2111 for_each_online_node(node) {
2112 struct kmem_cache_node *n = get_node(s, node);
2113 unsigned long nr_slabs;
2114 unsigned long nr_objs;
2115 unsigned long nr_free;
2116
2117 if (!n)
2118 continue;
2119
2120 nr_free = count_partial(n, count_free);
2121 nr_slabs = node_nr_slabs(n);
2122 nr_objs = node_nr_objs(n);
2123
2124 printk(KERN_WARNING
2125 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
2126 node, nr_slabs, nr_objs, nr_free);
2127 }
2128}
2129
2130static inline void *new_slab_objects(struct kmem_cache *s, gfp_t flags,
2131 int node, struct kmem_cache_cpu **pc)
2132{
2133 void *object;
2134 struct kmem_cache_cpu *c;
2135 struct page *page = new_slab(s, flags, node);
2136
2137 if (page) {
2138 c = __this_cpu_ptr(s->cpu_slab);
2139 if (c->page)
2140 flush_slab(s, c);
2141
2142 /*
2143 * No other reference to the page yet so we can
2144 * muck around with it freely without cmpxchg
2145 */
2146 object = page->freelist;
2147 page->freelist = NULL;
2148
2149 stat(s, ALLOC_SLAB);
2150 c->node = page_to_nid(page);
2151 c->page = page;
2152 *pc = c;
2153 } else
2154 object = NULL;
2155
2156 return object;
2157}
2158
2159/*
2160 * Check the page->freelist of a page and either transfer the freelist to the per cpu freelist
2161 * or deactivate the page.
2162 *
2163 * The page is still frozen if the return value is not NULL.
2164 *
2165 * If this function returns NULL then the page has been unfrozen.
2166 */
2167static inline void *get_freelist(struct kmem_cache *s, struct page *page)
2168{
2169 struct page new;
2170 unsigned long counters;
2171 void *freelist;
2172
2173 do {
2174 freelist = page->freelist;
2175 counters = page->counters;
2176 new.counters = counters;
2177 VM_BUG_ON(!new.frozen);
2178
2179 new.inuse = page->objects;
2180 new.frozen = freelist != NULL;
2181
2182 } while (!cmpxchg_double_slab(s, page,
2183 freelist, counters,
2184 NULL, new.counters,
2185 "get_freelist"));
2186
2187 return freelist;
2188}
2189
2190/*
2191 * Slow path. The lockless freelist is empty or we need to perform
2192 * debugging duties.
2193 *
2194 * Processing is still very fast if new objects have been freed to the
2195 * regular freelist. In that case we simply take over the regular freelist
2196 * as the lockless freelist and zap the regular freelist.
2197 *
2198 * If that is not working then we fall back to the partial lists. We take the
2199 * first element of the freelist as the object to allocate now and move the
2200 * rest of the freelist to the lockless freelist.
2201 *
2202 * And if we were unable to get a new slab from the partial slab lists then
2203 * we need to allocate a new slab. This is the slowest path since it involves
2204 * a call to the page allocator and the setup of a new slab.
2205 */
2206static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
2207 unsigned long addr, struct kmem_cache_cpu *c)
2208{
2209 void **object;
2210 unsigned long flags;
2211
2212 local_irq_save(flags);
2213#ifdef CONFIG_PREEMPT
2214 /*
2215 * We may have been preempted and rescheduled on a different
2216 * cpu before disabling interrupts. Need to reload cpu area
2217 * pointer.
2218 */
2219 c = this_cpu_ptr(s->cpu_slab);
2220#endif
2221
2222 if (!c->page)
2223 goto new_slab;
2224redo:
2225 if (unlikely(!node_match(c, node))) {
2226 stat(s, ALLOC_NODE_MISMATCH);
2227 deactivate_slab(s, c);
2228 goto new_slab;
2229 }
2230
2231 /* must check again c->freelist in case of cpu migration or IRQ */
2232 object = c->freelist;
2233 if (object)
2234 goto load_freelist;
2235
2236 stat(s, ALLOC_SLOWPATH);
2237
2238 object = get_freelist(s, c->page);
2239
2240 if (!object) {
2241 c->page = NULL;
2242 stat(s, DEACTIVATE_BYPASS);
2243 goto new_slab;
2244 }
2245
2246 stat(s, ALLOC_REFILL);
2247
2248load_freelist:
2249 c->freelist = get_freepointer(s, object);
2250 c->tid = next_tid(c->tid);
2251 local_irq_restore(flags);
2252 return object;
2253
2254new_slab:
2255
2256 if (c->partial) {
2257 c->page = c->partial;
2258 c->partial = c->page->next;
2259 c->node = page_to_nid(c->page);
2260 stat(s, CPU_PARTIAL_ALLOC);
2261 c->freelist = NULL;
2262 goto redo;
2263 }
2264
2265 /* Then do expensive stuff like retrieving pages from the partial lists */
2266 object = get_partial(s, gfpflags, node, c);
2267
2268 if (unlikely(!object)) {
2269
2270 object = new_slab_objects(s, gfpflags, node, &c);
2271
2272 if (unlikely(!object)) {
2273 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
2274 slab_out_of_memory(s, gfpflags, node);
2275
2276 local_irq_restore(flags);
2277 return NULL;
2278 }
2279 }
2280
2281 if (likely(!kmem_cache_debug(s)))
2282 goto load_freelist;
2283
2284 /* Only entered in the debug case */
2285 if (!alloc_debug_processing(s, c->page, object, addr))
2286 goto new_slab; /* Slab failed checks. Next slab needed */
2287
2288 c->freelist = get_freepointer(s, object);
2289 deactivate_slab(s, c);
2290 c->node = NUMA_NO_NODE;
2291 local_irq_restore(flags);
2292 return object;
2293}
2294
2295/*
2296 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
2297 * have the fastpath folded into their functions. So no function call
2298 * overhead for requests that can be satisfied on the fastpath.
2299 *
2300 * The fastpath works by first checking if the lockless freelist can be used.
2301 * If not then __slab_alloc is called for slow processing.
2302 *
2303 * Otherwise we can simply pick the next object from the lockless free list.
2304 */
2305static __always_inline void *slab_alloc(struct kmem_cache *s,
2306 gfp_t gfpflags, int node, unsigned long addr)
2307{
2308 void **object;
2309 struct kmem_cache_cpu *c;
2310 unsigned long tid;
2311
2312 if (slab_pre_alloc_hook(s, gfpflags))
2313 return NULL;
2314
2315redo:
2316
2317 /*
2318 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
2319 * enabled. We may switch back and forth between cpus while
2320 * reading from one cpu area. That does not matter as long
2321 * as we end up on the original cpu again when doing the cmpxchg.
2322 */
2323 c = __this_cpu_ptr(s->cpu_slab);
2324
2325 /*
2326 * The transaction ids are globally unique per cpu and per operation on
2327 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
2328 * occurs on the right processor and that there was no operation on the
2329 * linked list in between.
2330 */
2331 tid = c->tid;
2332 barrier();
2333
2334 object = c->freelist;
2335 if (unlikely(!object || !node_match(c, node)))
2336
2337 object = __slab_alloc(s, gfpflags, node, addr, c);
2338
2339 else {
2340 void *next_object = get_freepointer_safe(s, object);
2341
2342 /*
2343 * The cmpxchg will only match if there was no additional
2344 * operation and if we are on the right processor.
2345 *
2346 * The cmpxchg does the following atomically (without lock semantics!)
2347 * 1. Relocate first pointer to the current per cpu area.
2348 * 2. Verify that tid and freelist have not been changed
2349 * 3. If they were not changed replace tid and freelist
2350 *
2351 * Since this is without lock semantics the protection is only against
2352 * code executing on this cpu *not* from access by other cpus.
2353 */
2354 if (unlikely(!this_cpu_cmpxchg_double(
2355 s->cpu_slab->freelist, s->cpu_slab->tid,
2356 object, tid,
2357 next_object, next_tid(tid)))) {
2358
2359 note_cmpxchg_failure("slab_alloc", s, tid);
2360 goto redo;
2361 }
2362 prefetch_freepointer(s, next_object);
2363 stat(s, ALLOC_FASTPATH);
2364 }
2365
2366 if (unlikely(gfpflags & __GFP_ZERO) && object)
2367 memset(object, 0, s->objsize);
2368
2369 slab_post_alloc_hook(s, gfpflags, object);
2370
2371 return object;
2372}
2373
2374void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2375{
2376 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2377
2378 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2379
2380 return ret;
2381}
2382EXPORT_SYMBOL(kmem_cache_alloc);
2383
2384#ifdef CONFIG_TRACING
2385void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2386{
2387 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2388 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2389 return ret;
2390}
2391EXPORT_SYMBOL(kmem_cache_alloc_trace);
2392
2393void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2394{
2395 void *ret = kmalloc_order(size, flags, order);
2396 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2397 return ret;
2398}
2399EXPORT_SYMBOL(kmalloc_order_trace);
2400#endif
2401
2402#ifdef CONFIG_NUMA
2403void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2404{
2405 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2406
2407 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2408 s->objsize, s->size, gfpflags, node);
2409
2410 return ret;
2411}
2412EXPORT_SYMBOL(kmem_cache_alloc_node);
2413
2414#ifdef CONFIG_TRACING
2415void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2416 gfp_t gfpflags,
2417 int node, size_t size)
2418{
2419 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2420
2421 trace_kmalloc_node(_RET_IP_, ret,
2422 size, s->size, gfpflags, node);
2423 return ret;
2424}
2425EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2426#endif
2427#endif
2428
2429/*
2430 * Slow patch handling. This may still be called frequently since objects
2431 * have a longer lifetime than the cpu slabs in most processing loads.
2432 *
2433 * So we still attempt to reduce cache line usage. Just take the slab
2434 * lock and free the item. If there is no additional partial page
2435 * handling required then we can return immediately.
2436 */
2437static void __slab_free(struct kmem_cache *s, struct page *page,
2438 void *x, unsigned long addr)
2439{
2440 void *prior;
2441 void **object = (void *)x;
2442 int was_frozen;
2443 int inuse;
2444 struct page new;
2445 unsigned long counters;
2446 struct kmem_cache_node *n = NULL;
2447 unsigned long uninitialized_var(flags);
2448
2449 stat(s, FREE_SLOWPATH);
2450
2451 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2452 return;
2453
2454 do {
2455 prior = page->freelist;
2456 counters = page->counters;
2457 set_freepointer(s, object, prior);
2458 new.counters = counters;
2459 was_frozen = new.frozen;
2460 new.inuse--;
2461 if ((!new.inuse || !prior) && !was_frozen && !n) {
2462
2463 if (!kmem_cache_debug(s) && !prior)
2464
2465 /*
2466 * Slab was on no list before and will be partially empty
2467 * We can defer the list move and instead freeze it.
2468 */
2469 new.frozen = 1;
2470
2471 else { /* Needs to be taken off a list */
2472
2473 n = get_node(s, page_to_nid(page));
2474 /*
2475 * Speculatively acquire the list_lock.
2476 * If the cmpxchg does not succeed then we may
2477 * drop the list_lock without any processing.
2478 *
2479 * Otherwise the list_lock will synchronize with
2480 * other processors updating the list of slabs.
2481 */
2482 spin_lock_irqsave(&n->list_lock, flags);
2483
2484 }
2485 }
2486 inuse = new.inuse;
2487
2488 } while (!cmpxchg_double_slab(s, page,
2489 prior, counters,
2490 object, new.counters,
2491 "__slab_free"));
2492
2493 if (likely(!n)) {
2494
2495 /*
2496 * If we just froze the page then put it onto the
2497 * per cpu partial list.
2498 */
2499 if (new.frozen && !was_frozen) {
2500 put_cpu_partial(s, page, 1);
2501 stat(s, CPU_PARTIAL_FREE);
2502 }
2503 /*
2504 * The list lock was not taken therefore no list
2505 * activity can be necessary.
2506 */
2507 if (was_frozen)
2508 stat(s, FREE_FROZEN);
2509 return;
2510 }
2511
2512 /*
2513 * was_frozen may have been set after we acquired the list_lock in
2514 * an earlier loop. So we need to check it here again.
2515 */
2516 if (was_frozen)
2517 stat(s, FREE_FROZEN);
2518 else {
2519 if (unlikely(!inuse && n->nr_partial > s->min_partial))
2520 goto slab_empty;
2521
2522 /*
2523 * Objects left in the slab. If it was not on the partial list before
2524 * then add it.
2525 */
2526 if (unlikely(!prior)) {
2527 remove_full(s, page);
2528 add_partial(n, page, DEACTIVATE_TO_TAIL);
2529 stat(s, FREE_ADD_PARTIAL);
2530 }
2531 }
2532 spin_unlock_irqrestore(&n->list_lock, flags);
2533 return;
2534
2535slab_empty:
2536 if (prior) {
2537 /*
2538 * Slab on the partial list.
2539 */
2540 remove_partial(n, page);
2541 stat(s, FREE_REMOVE_PARTIAL);
2542 } else
2543 /* Slab must be on the full list */
2544 remove_full(s, page);
2545
2546 spin_unlock_irqrestore(&n->list_lock, flags);
2547 stat(s, FREE_SLAB);
2548 discard_slab(s, page);
2549}
2550
2551/*
2552 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2553 * can perform fastpath freeing without additional function calls.
2554 *
2555 * The fastpath is only possible if we are freeing to the current cpu slab
2556 * of this processor. This typically the case if we have just allocated
2557 * the item before.
2558 *
2559 * If fastpath is not possible then fall back to __slab_free where we deal
2560 * with all sorts of special processing.
2561 */
2562static __always_inline void slab_free(struct kmem_cache *s,
2563 struct page *page, void *x, unsigned long addr)
2564{
2565 void **object = (void *)x;
2566 struct kmem_cache_cpu *c;
2567 unsigned long tid;
2568
2569 slab_free_hook(s, x);
2570
2571redo:
2572 /*
2573 * Determine the currently cpus per cpu slab.
2574 * The cpu may change afterward. However that does not matter since
2575 * data is retrieved via this pointer. If we are on the same cpu
2576 * during the cmpxchg then the free will succedd.
2577 */
2578 c = __this_cpu_ptr(s->cpu_slab);
2579
2580 tid = c->tid;
2581 barrier();
2582
2583 if (likely(page == c->page)) {
2584 set_freepointer(s, object, c->freelist);
2585
2586 if (unlikely(!this_cpu_cmpxchg_double(
2587 s->cpu_slab->freelist, s->cpu_slab->tid,
2588 c->freelist, tid,
2589 object, next_tid(tid)))) {
2590
2591 note_cmpxchg_failure("slab_free", s, tid);
2592 goto redo;
2593 }
2594 stat(s, FREE_FASTPATH);
2595 } else
2596 __slab_free(s, page, x, addr);
2597
2598}
2599
2600void kmem_cache_free(struct kmem_cache *s, void *x)
2601{
2602 struct page *page;
2603
2604 page = virt_to_head_page(x);
2605
2606 slab_free(s, page, x, _RET_IP_);
2607
2608 trace_kmem_cache_free(_RET_IP_, x);
2609}
2610EXPORT_SYMBOL(kmem_cache_free);
2611
2612/*
2613 * Object placement in a slab is made very easy because we always start at
2614 * offset 0. If we tune the size of the object to the alignment then we can
2615 * get the required alignment by putting one properly sized object after
2616 * another.
2617 *
2618 * Notice that the allocation order determines the sizes of the per cpu
2619 * caches. Each processor has always one slab available for allocations.
2620 * Increasing the allocation order reduces the number of times that slabs
2621 * must be moved on and off the partial lists and is therefore a factor in
2622 * locking overhead.
2623 */
2624
2625/*
2626 * Mininum / Maximum order of slab pages. This influences locking overhead
2627 * and slab fragmentation. A higher order reduces the number of partial slabs
2628 * and increases the number of allocations possible without having to
2629 * take the list_lock.
2630 */
2631static int slub_min_order;
2632static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2633static int slub_min_objects;
2634
2635/*
2636 * Merge control. If this is set then no merging of slab caches will occur.
2637 * (Could be removed. This was introduced to pacify the merge skeptics.)
2638 */
2639static int slub_nomerge;
2640
2641/*
2642 * Calculate the order of allocation given an slab object size.
2643 *
2644 * The order of allocation has significant impact on performance and other
2645 * system components. Generally order 0 allocations should be preferred since
2646 * order 0 does not cause fragmentation in the page allocator. Larger objects
2647 * be problematic to put into order 0 slabs because there may be too much
2648 * unused space left. We go to a higher order if more than 1/16th of the slab
2649 * would be wasted.
2650 *
2651 * In order to reach satisfactory performance we must ensure that a minimum
2652 * number of objects is in one slab. Otherwise we may generate too much
2653 * activity on the partial lists which requires taking the list_lock. This is
2654 * less a concern for large slabs though which are rarely used.
2655 *
2656 * slub_max_order specifies the order where we begin to stop considering the
2657 * number of objects in a slab as critical. If we reach slub_max_order then
2658 * we try to keep the page order as low as possible. So we accept more waste
2659 * of space in favor of a small page order.
2660 *
2661 * Higher order allocations also allow the placement of more objects in a
2662 * slab and thereby reduce object handling overhead. If the user has
2663 * requested a higher mininum order then we start with that one instead of
2664 * the smallest order which will fit the object.
2665 */
2666static inline int slab_order(int size, int min_objects,
2667 int max_order, int fract_leftover, int reserved)
2668{
2669 int order;
2670 int rem;
2671 int min_order = slub_min_order;
2672
2673 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2674 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2675
2676 for (order = max(min_order,
2677 fls(min_objects * size - 1) - PAGE_SHIFT);
2678 order <= max_order; order++) {
2679
2680 unsigned long slab_size = PAGE_SIZE << order;
2681
2682 if (slab_size < min_objects * size + reserved)
2683 continue;
2684
2685 rem = (slab_size - reserved) % size;
2686
2687 if (rem <= slab_size / fract_leftover)
2688 break;
2689
2690 }
2691
2692 return order;
2693}
2694
2695static inline int calculate_order(int size, int reserved)
2696{
2697 int order;
2698 int min_objects;
2699 int fraction;
2700 int max_objects;
2701
2702 /*
2703 * Attempt to find best configuration for a slab. This
2704 * works by first attempting to generate a layout with
2705 * the best configuration and backing off gradually.
2706 *
2707 * First we reduce the acceptable waste in a slab. Then
2708 * we reduce the minimum objects required in a slab.
2709 */
2710 min_objects = slub_min_objects;
2711 if (!min_objects)
2712 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2713 max_objects = order_objects(slub_max_order, size, reserved);
2714 min_objects = min(min_objects, max_objects);
2715
2716 while (min_objects > 1) {
2717 fraction = 16;
2718 while (fraction >= 4) {
2719 order = slab_order(size, min_objects,
2720 slub_max_order, fraction, reserved);
2721 if (order <= slub_max_order)
2722 return order;
2723 fraction /= 2;
2724 }
2725 min_objects--;
2726 }
2727
2728 /*
2729 * We were unable to place multiple objects in a slab. Now
2730 * lets see if we can place a single object there.
2731 */
2732 order = slab_order(size, 1, slub_max_order, 1, reserved);
2733 if (order <= slub_max_order)
2734 return order;
2735
2736 /*
2737 * Doh this slab cannot be placed using slub_max_order.
2738 */
2739 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2740 if (order < MAX_ORDER)
2741 return order;
2742 return -ENOSYS;
2743}
2744
2745/*
2746 * Figure out what the alignment of the objects will be.
2747 */
2748static unsigned long calculate_alignment(unsigned long flags,
2749 unsigned long align, unsigned long size)
2750{
2751 /*
2752 * If the user wants hardware cache aligned objects then follow that
2753 * suggestion if the object is sufficiently large.
2754 *
2755 * The hardware cache alignment cannot override the specified
2756 * alignment though. If that is greater then use it.
2757 */
2758 if (flags & SLAB_HWCACHE_ALIGN) {
2759 unsigned long ralign = cache_line_size();
2760 while (size <= ralign / 2)
2761 ralign /= 2;
2762 align = max(align, ralign);
2763 }
2764
2765 if (align < ARCH_SLAB_MINALIGN)
2766 align = ARCH_SLAB_MINALIGN;
2767
2768 return ALIGN(align, sizeof(void *));
2769}
2770
2771static void
2772init_kmem_cache_node(struct kmem_cache_node *n)
2773{
2774 n->nr_partial = 0;
2775 spin_lock_init(&n->list_lock);
2776 INIT_LIST_HEAD(&n->partial);
2777#ifdef CONFIG_SLUB_DEBUG
2778 atomic_long_set(&n->nr_slabs, 0);
2779 atomic_long_set(&n->total_objects, 0);
2780 INIT_LIST_HEAD(&n->full);
2781#endif
2782}
2783
2784static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2785{
2786 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2787 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2788
2789 /*
2790 * Must align to double word boundary for the double cmpxchg
2791 * instructions to work; see __pcpu_double_call_return_bool().
2792 */
2793 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu),
2794 2 * sizeof(void *));
2795
2796 if (!s->cpu_slab)
2797 return 0;
2798
2799 init_kmem_cache_cpus(s);
2800
2801 return 1;
2802}
2803
2804static struct kmem_cache *kmem_cache_node;
2805
2806/*
2807 * No kmalloc_node yet so do it by hand. We know that this is the first
2808 * slab on the node for this slabcache. There are no concurrent accesses
2809 * possible.
2810 *
2811 * Note that this function only works on the kmalloc_node_cache
2812 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2813 * memory on a fresh node that has no slab structures yet.
2814 */
2815static void early_kmem_cache_node_alloc(int node)
2816{
2817 struct page *page;
2818 struct kmem_cache_node *n;
2819
2820 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2821
2822 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2823
2824 BUG_ON(!page);
2825 if (page_to_nid(page) != node) {
2826 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2827 "node %d\n", node);
2828 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2829 "in order to be able to continue\n");
2830 }
2831
2832 n = page->freelist;
2833 BUG_ON(!n);
2834 page->freelist = get_freepointer(kmem_cache_node, n);
2835 page->inuse = 1;
2836 page->frozen = 0;
2837 kmem_cache_node->node[node] = n;
2838#ifdef CONFIG_SLUB_DEBUG
2839 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2840 init_tracking(kmem_cache_node, n);
2841#endif
2842 init_kmem_cache_node(n);
2843 inc_slabs_node(kmem_cache_node, node, page->objects);
2844
2845 add_partial(n, page, DEACTIVATE_TO_HEAD);
2846}
2847
2848static void free_kmem_cache_nodes(struct kmem_cache *s)
2849{
2850 int node;
2851
2852 for_each_node_state(node, N_NORMAL_MEMORY) {
2853 struct kmem_cache_node *n = s->node[node];
2854
2855 if (n)
2856 kmem_cache_free(kmem_cache_node, n);
2857
2858 s->node[node] = NULL;
2859 }
2860}
2861
2862static int init_kmem_cache_nodes(struct kmem_cache *s)
2863{
2864 int node;
2865
2866 for_each_node_state(node, N_NORMAL_MEMORY) {
2867 struct kmem_cache_node *n;
2868
2869 if (slab_state == DOWN) {
2870 early_kmem_cache_node_alloc(node);
2871 continue;
2872 }
2873 n = kmem_cache_alloc_node(kmem_cache_node,
2874 GFP_KERNEL, node);
2875
2876 if (!n) {
2877 free_kmem_cache_nodes(s);
2878 return 0;
2879 }
2880
2881 s->node[node] = n;
2882 init_kmem_cache_node(n);
2883 }
2884 return 1;
2885}
2886
2887static void set_min_partial(struct kmem_cache *s, unsigned long min)
2888{
2889 if (min < MIN_PARTIAL)
2890 min = MIN_PARTIAL;
2891 else if (min > MAX_PARTIAL)
2892 min = MAX_PARTIAL;
2893 s->min_partial = min;
2894}
2895
2896/*
2897 * calculate_sizes() determines the order and the distribution of data within
2898 * a slab object.
2899 */
2900static int calculate_sizes(struct kmem_cache *s, int forced_order)
2901{
2902 unsigned long flags = s->flags;
2903 unsigned long size = s->objsize;
2904 unsigned long align = s->align;
2905 int order;
2906
2907 /*
2908 * Round up object size to the next word boundary. We can only
2909 * place the free pointer at word boundaries and this determines
2910 * the possible location of the free pointer.
2911 */
2912 size = ALIGN(size, sizeof(void *));
2913
2914#ifdef CONFIG_SLUB_DEBUG
2915 /*
2916 * Determine if we can poison the object itself. If the user of
2917 * the slab may touch the object after free or before allocation
2918 * then we should never poison the object itself.
2919 */
2920 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2921 !s->ctor)
2922 s->flags |= __OBJECT_POISON;
2923 else
2924 s->flags &= ~__OBJECT_POISON;
2925
2926
2927 /*
2928 * If we are Redzoning then check if there is some space between the
2929 * end of the object and the free pointer. If not then add an
2930 * additional word to have some bytes to store Redzone information.
2931 */
2932 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2933 size += sizeof(void *);
2934#endif
2935
2936 /*
2937 * With that we have determined the number of bytes in actual use
2938 * by the object. This is the potential offset to the free pointer.
2939 */
2940 s->inuse = size;
2941
2942 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2943 s->ctor)) {
2944 /*
2945 * Relocate free pointer after the object if it is not
2946 * permitted to overwrite the first word of the object on
2947 * kmem_cache_free.
2948 *
2949 * This is the case if we do RCU, have a constructor or
2950 * destructor or are poisoning the objects.
2951 */
2952 s->offset = size;
2953 size += sizeof(void *);
2954 }
2955
2956#ifdef CONFIG_SLUB_DEBUG
2957 if (flags & SLAB_STORE_USER)
2958 /*
2959 * Need to store information about allocs and frees after
2960 * the object.
2961 */
2962 size += 2 * sizeof(struct track);
2963
2964 if (flags & SLAB_RED_ZONE)
2965 /*
2966 * Add some empty padding so that we can catch
2967 * overwrites from earlier objects rather than let
2968 * tracking information or the free pointer be
2969 * corrupted if a user writes before the start
2970 * of the object.
2971 */
2972 size += sizeof(void *);
2973#endif
2974
2975 /*
2976 * Determine the alignment based on various parameters that the
2977 * user specified and the dynamic determination of cache line size
2978 * on bootup.
2979 */
2980 align = calculate_alignment(flags, align, s->objsize);
2981 s->align = align;
2982
2983 /*
2984 * SLUB stores one object immediately after another beginning from
2985 * offset 0. In order to align the objects we have to simply size
2986 * each object to conform to the alignment.
2987 */
2988 size = ALIGN(size, align);
2989 s->size = size;
2990 if (forced_order >= 0)
2991 order = forced_order;
2992 else
2993 order = calculate_order(size, s->reserved);
2994
2995 if (order < 0)
2996 return 0;
2997
2998 s->allocflags = 0;
2999 if (order)
3000 s->allocflags |= __GFP_COMP;
3001
3002 if (s->flags & SLAB_CACHE_DMA)
3003 s->allocflags |= SLUB_DMA;
3004
3005 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3006 s->allocflags |= __GFP_RECLAIMABLE;
3007
3008 /*
3009 * Determine the number of objects per slab
3010 */
3011 s->oo = oo_make(order, size, s->reserved);
3012 s->min = oo_make(get_order(size), size, s->reserved);
3013 if (oo_objects(s->oo) > oo_objects(s->max))
3014 s->max = s->oo;
3015
3016 return !!oo_objects(s->oo);
3017
3018}
3019
3020static int kmem_cache_open(struct kmem_cache *s,
3021 const char *name, size_t size,
3022 size_t align, unsigned long flags,
3023 void (*ctor)(void *))
3024{
3025 memset(s, 0, kmem_size);
3026 s->name = name;
3027 s->ctor = ctor;
3028 s->objsize = size;
3029 s->align = align;
3030 s->flags = kmem_cache_flags(size, flags, name, ctor);
3031 s->reserved = 0;
3032
3033 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
3034 s->reserved = sizeof(struct rcu_head);
3035
3036 if (!calculate_sizes(s, -1))
3037 goto error;
3038 if (disable_higher_order_debug) {
3039 /*
3040 * Disable debugging flags that store metadata if the min slab
3041 * order increased.
3042 */
3043 if (get_order(s->size) > get_order(s->objsize)) {
3044 s->flags &= ~DEBUG_METADATA_FLAGS;
3045 s->offset = 0;
3046 if (!calculate_sizes(s, -1))
3047 goto error;
3048 }
3049 }
3050
3051#if defined(CONFIG_HAVE_CMPXCHG_DOUBLE) && \
3052 defined(CONFIG_HAVE_ALIGNED_STRUCT_PAGE)
3053 if (system_has_cmpxchg_double() && (s->flags & SLAB_DEBUG_FLAGS) == 0)
3054 /* Enable fast mode */
3055 s->flags |= __CMPXCHG_DOUBLE;
3056#endif
3057
3058 /*
3059 * The larger the object size is, the more pages we want on the partial
3060 * list to avoid pounding the page allocator excessively.
3061 */
3062 set_min_partial(s, ilog2(s->size) / 2);
3063
3064 /*
3065 * cpu_partial determined the maximum number of objects kept in the
3066 * per cpu partial lists of a processor.
3067 *
3068 * Per cpu partial lists mainly contain slabs that just have one
3069 * object freed. If they are used for allocation then they can be
3070 * filled up again with minimal effort. The slab will never hit the
3071 * per node partial lists and therefore no locking will be required.
3072 *
3073 * This setting also determines
3074 *
3075 * A) The number of objects from per cpu partial slabs dumped to the
3076 * per node list when we reach the limit.
3077 * B) The number of objects in cpu partial slabs to extract from the
3078 * per node list when we run out of per cpu objects. We only fetch 50%
3079 * to keep some capacity around for frees.
3080 */
3081 if (kmem_cache_debug(s))
3082 s->cpu_partial = 0;
3083 else if (s->size >= PAGE_SIZE)
3084 s->cpu_partial = 2;
3085 else if (s->size >= 1024)
3086 s->cpu_partial = 6;
3087 else if (s->size >= 256)
3088 s->cpu_partial = 13;
3089 else
3090 s->cpu_partial = 30;
3091
3092 s->refcount = 1;
3093#ifdef CONFIG_NUMA
3094 s->remote_node_defrag_ratio = 1000;
3095#endif
3096 if (!init_kmem_cache_nodes(s))
3097 goto error;
3098
3099 if (alloc_kmem_cache_cpus(s))
3100 return 1;
3101
3102 free_kmem_cache_nodes(s);
3103error:
3104 if (flags & SLAB_PANIC)
3105 panic("Cannot create slab %s size=%lu realsize=%u "
3106 "order=%u offset=%u flags=%lx\n",
3107 s->name, (unsigned long)size, s->size, oo_order(s->oo),
3108 s->offset, flags);
3109 return 0;
3110}
3111
3112/*
3113 * Determine the size of a slab object
3114 */
3115unsigned int kmem_cache_size(struct kmem_cache *s)
3116{
3117 return s->objsize;
3118}
3119EXPORT_SYMBOL(kmem_cache_size);
3120
3121static void list_slab_objects(struct kmem_cache *s, struct page *page,
3122 const char *text)
3123{
3124#ifdef CONFIG_SLUB_DEBUG
3125 void *addr = page_address(page);
3126 void *p;
3127 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
3128 sizeof(long), GFP_ATOMIC);
3129 if (!map)
3130 return;
3131 slab_err(s, page, "%s", text);
3132 slab_lock(page);
3133
3134 get_map(s, page, map);
3135 for_each_object(p, s, addr, page->objects) {
3136
3137 if (!test_bit(slab_index(p, s, addr), map)) {
3138 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
3139 p, p - addr);
3140 print_tracking(s, p);
3141 }
3142 }
3143 slab_unlock(page);
3144 kfree(map);
3145#endif
3146}
3147
3148/*
3149 * Attempt to free all partial slabs on a node.
3150 * This is called from kmem_cache_close(). We must be the last thread
3151 * using the cache and therefore we do not need to lock anymore.
3152 */
3153static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
3154{
3155 struct page *page, *h;
3156
3157 list_for_each_entry_safe(page, h, &n->partial, lru) {
3158 if (!page->inuse) {
3159 remove_partial(n, page);
3160 discard_slab(s, page);
3161 } else {
3162 list_slab_objects(s, page,
3163 "Objects remaining on kmem_cache_close()");
3164 }
3165 }
3166}
3167
3168/*
3169 * Release all resources used by a slab cache.
3170 */
3171static inline int kmem_cache_close(struct kmem_cache *s)
3172{
3173 int node;
3174
3175 flush_all(s);
3176 free_percpu(s->cpu_slab);
3177 /* Attempt to free all objects */
3178 for_each_node_state(node, N_NORMAL_MEMORY) {
3179 struct kmem_cache_node *n = get_node(s, node);
3180
3181 free_partial(s, n);
3182 if (n->nr_partial || slabs_node(s, node))
3183 return 1;
3184 }
3185 free_kmem_cache_nodes(s);
3186 return 0;
3187}
3188
3189/*
3190 * Close a cache and release the kmem_cache structure
3191 * (must be used for caches created using kmem_cache_create)
3192 */
3193void kmem_cache_destroy(struct kmem_cache *s)
3194{
3195 down_write(&slub_lock);
3196 s->refcount--;
3197 if (!s->refcount) {
3198 list_del(&s->list);
3199 up_write(&slub_lock);
3200 if (kmem_cache_close(s)) {
3201 printk(KERN_ERR "SLUB %s: %s called for cache that "
3202 "still has objects.\n", s->name, __func__);
3203 dump_stack();
3204 }
3205 if (s->flags & SLAB_DESTROY_BY_RCU)
3206 rcu_barrier();
3207 sysfs_slab_remove(s);
3208 } else
3209 up_write(&slub_lock);
3210}
3211EXPORT_SYMBOL(kmem_cache_destroy);
3212
3213/********************************************************************
3214 * Kmalloc subsystem
3215 *******************************************************************/
3216
3217struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
3218EXPORT_SYMBOL(kmalloc_caches);
3219
3220static struct kmem_cache *kmem_cache;
3221
3222#ifdef CONFIG_ZONE_DMA
3223static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
3224#endif
3225
3226static int __init setup_slub_min_order(char *str)
3227{
3228 get_option(&str, &slub_min_order);
3229
3230 return 1;
3231}
3232
3233__setup("slub_min_order=", setup_slub_min_order);
3234
3235static int __init setup_slub_max_order(char *str)
3236{
3237 get_option(&str, &slub_max_order);
3238 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
3239
3240 return 1;
3241}
3242
3243__setup("slub_max_order=", setup_slub_max_order);
3244
3245static int __init setup_slub_min_objects(char *str)
3246{
3247 get_option(&str, &slub_min_objects);
3248
3249 return 1;
3250}
3251
3252__setup("slub_min_objects=", setup_slub_min_objects);
3253
3254static int __init setup_slub_nomerge(char *str)
3255{
3256 slub_nomerge = 1;
3257 return 1;
3258}
3259
3260__setup("slub_nomerge", setup_slub_nomerge);
3261
3262static struct kmem_cache *__init create_kmalloc_cache(const char *name,
3263 int size, unsigned int flags)
3264{
3265 struct kmem_cache *s;
3266
3267 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3268
3269 /*
3270 * This function is called with IRQs disabled during early-boot on
3271 * single CPU so there's no need to take slub_lock here.
3272 */
3273 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
3274 flags, NULL))
3275 goto panic;
3276
3277 list_add(&s->list, &slab_caches);
3278 return s;
3279
3280panic:
3281 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
3282 return NULL;
3283}
3284
3285/*
3286 * Conversion table for small slabs sizes / 8 to the index in the
3287 * kmalloc array. This is necessary for slabs < 192 since we have non power
3288 * of two cache sizes there. The size of larger slabs can be determined using
3289 * fls.
3290 */
3291static s8 size_index[24] = {
3292 3, /* 8 */
3293 4, /* 16 */
3294 5, /* 24 */
3295 5, /* 32 */
3296 6, /* 40 */
3297 6, /* 48 */
3298 6, /* 56 */
3299 6, /* 64 */
3300 1, /* 72 */
3301 1, /* 80 */
3302 1, /* 88 */
3303 1, /* 96 */
3304 7, /* 104 */
3305 7, /* 112 */
3306 7, /* 120 */
3307 7, /* 128 */
3308 2, /* 136 */
3309 2, /* 144 */
3310 2, /* 152 */
3311 2, /* 160 */
3312 2, /* 168 */
3313 2, /* 176 */
3314 2, /* 184 */
3315 2 /* 192 */
3316};
3317
3318static inline int size_index_elem(size_t bytes)
3319{
3320 return (bytes - 1) / 8;
3321}
3322
3323static struct kmem_cache *get_slab(size_t size, gfp_t flags)
3324{
3325 int index;
3326
3327 if (size <= 192) {
3328 if (!size)
3329 return ZERO_SIZE_PTR;
3330
3331 index = size_index[size_index_elem(size)];
3332 } else
3333 index = fls(size - 1);
3334
3335#ifdef CONFIG_ZONE_DMA
3336 if (unlikely((flags & SLUB_DMA)))
3337 return kmalloc_dma_caches[index];
3338
3339#endif
3340 return kmalloc_caches[index];
3341}
3342
3343void *__kmalloc(size_t size, gfp_t flags)
3344{
3345 struct kmem_cache *s;
3346 void *ret;
3347
3348 if (unlikely(size > SLUB_MAX_SIZE))
3349 return kmalloc_large(size, flags);
3350
3351 s = get_slab(size, flags);
3352
3353 if (unlikely(ZERO_OR_NULL_PTR(s)))
3354 return s;
3355
3356 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
3357
3358 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
3359
3360 return ret;
3361}
3362EXPORT_SYMBOL(__kmalloc);
3363
3364#ifdef CONFIG_NUMA
3365static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
3366{
3367 struct page *page;
3368 void *ptr = NULL;
3369
3370 flags |= __GFP_COMP | __GFP_NOTRACK;
3371 page = alloc_pages_node(node, flags, get_order(size));
3372 if (page)
3373 ptr = page_address(page);
3374
3375 kmemleak_alloc(ptr, size, 1, flags);
3376 return ptr;
3377}
3378
3379void *__kmalloc_node(size_t size, gfp_t flags, int node)
3380{
3381 struct kmem_cache *s;
3382 void *ret;
3383
3384 if (unlikely(size > SLUB_MAX_SIZE)) {
3385 ret = kmalloc_large_node(size, flags, node);
3386
3387 trace_kmalloc_node(_RET_IP_, ret,
3388 size, PAGE_SIZE << get_order(size),
3389 flags, node);
3390
3391 return ret;
3392 }
3393
3394 s = get_slab(size, flags);
3395
3396 if (unlikely(ZERO_OR_NULL_PTR(s)))
3397 return s;
3398
3399 ret = slab_alloc(s, flags, node, _RET_IP_);
3400
3401 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
3402
3403 return ret;
3404}
3405EXPORT_SYMBOL(__kmalloc_node);
3406#endif
3407
3408size_t ksize(const void *object)
3409{
3410 struct page *page;
3411
3412 if (unlikely(object == ZERO_SIZE_PTR))
3413 return 0;
3414
3415 page = virt_to_head_page(object);
3416
3417 if (unlikely(!PageSlab(page))) {
3418 WARN_ON(!PageCompound(page));
3419 return PAGE_SIZE << compound_order(page);
3420 }
3421
3422 return slab_ksize(page->slab);
3423}
3424EXPORT_SYMBOL(ksize);
3425
3426#ifdef CONFIG_SLUB_DEBUG
3427bool verify_mem_not_deleted(const void *x)
3428{
3429 struct page *page;
3430 void *object = (void *)x;
3431 unsigned long flags;
3432 bool rv;
3433
3434 if (unlikely(ZERO_OR_NULL_PTR(x)))
3435 return false;
3436
3437 local_irq_save(flags);
3438
3439 page = virt_to_head_page(x);
3440 if (unlikely(!PageSlab(page))) {
3441 /* maybe it was from stack? */
3442 rv = true;
3443 goto out_unlock;
3444 }
3445
3446 slab_lock(page);
3447 if (on_freelist(page->slab, page, object)) {
3448 object_err(page->slab, page, object, "Object is on free-list");
3449 rv = false;
3450 } else {
3451 rv = true;
3452 }
3453 slab_unlock(page);
3454
3455out_unlock:
3456 local_irq_restore(flags);
3457 return rv;
3458}
3459EXPORT_SYMBOL(verify_mem_not_deleted);
3460#endif
3461
3462void kfree(const void *x)
3463{
3464 struct page *page;
3465 void *object = (void *)x;
3466
3467 trace_kfree(_RET_IP_, x);
3468
3469 if (unlikely(ZERO_OR_NULL_PTR(x)))
3470 return;
3471
3472 page = virt_to_head_page(x);
3473 if (unlikely(!PageSlab(page))) {
3474 BUG_ON(!PageCompound(page));
3475 kmemleak_free(x);
3476 put_page(page);
3477 return;
3478 }
3479 slab_free(page->slab, page, object, _RET_IP_);
3480}
3481EXPORT_SYMBOL(kfree);
3482
3483/*
3484 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3485 * the remaining slabs by the number of items in use. The slabs with the
3486 * most items in use come first. New allocations will then fill those up
3487 * and thus they can be removed from the partial lists.
3488 *
3489 * The slabs with the least items are placed last. This results in them
3490 * being allocated from last increasing the chance that the last objects
3491 * are freed in them.
3492 */
3493int kmem_cache_shrink(struct kmem_cache *s)
3494{
3495 int node;
3496 int i;
3497 struct kmem_cache_node *n;
3498 struct page *page;
3499 struct page *t;
3500 int objects = oo_objects(s->max);
3501 struct list_head *slabs_by_inuse =
3502 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3503 unsigned long flags;
3504
3505 if (!slabs_by_inuse)
3506 return -ENOMEM;
3507
3508 flush_all(s);
3509 for_each_node_state(node, N_NORMAL_MEMORY) {
3510 n = get_node(s, node);
3511
3512 if (!n->nr_partial)
3513 continue;
3514
3515 for (i = 0; i < objects; i++)
3516 INIT_LIST_HEAD(slabs_by_inuse + i);
3517
3518 spin_lock_irqsave(&n->list_lock, flags);
3519
3520 /*
3521 * Build lists indexed by the items in use in each slab.
3522 *
3523 * Note that concurrent frees may occur while we hold the
3524 * list_lock. page->inuse here is the upper limit.
3525 */
3526 list_for_each_entry_safe(page, t, &n->partial, lru) {
3527 list_move(&page->lru, slabs_by_inuse + page->inuse);
3528 if (!page->inuse)
3529 n->nr_partial--;
3530 }
3531
3532 /*
3533 * Rebuild the partial list with the slabs filled up most
3534 * first and the least used slabs at the end.
3535 */
3536 for (i = objects - 1; i > 0; i--)
3537 list_splice(slabs_by_inuse + i, n->partial.prev);
3538
3539 spin_unlock_irqrestore(&n->list_lock, flags);
3540
3541 /* Release empty slabs */
3542 list_for_each_entry_safe(page, t, slabs_by_inuse, lru)
3543 discard_slab(s, page);
3544 }
3545
3546 kfree(slabs_by_inuse);
3547 return 0;
3548}
3549EXPORT_SYMBOL(kmem_cache_shrink);
3550
3551#if defined(CONFIG_MEMORY_HOTPLUG)
3552static int slab_mem_going_offline_callback(void *arg)
3553{
3554 struct kmem_cache *s;
3555
3556 down_read(&slub_lock);
3557 list_for_each_entry(s, &slab_caches, list)
3558 kmem_cache_shrink(s);
3559 up_read(&slub_lock);
3560
3561 return 0;
3562}
3563
3564static void slab_mem_offline_callback(void *arg)
3565{
3566 struct kmem_cache_node *n;
3567 struct kmem_cache *s;
3568 struct memory_notify *marg = arg;
3569 int offline_node;
3570
3571 offline_node = marg->status_change_nid;
3572
3573 /*
3574 * If the node still has available memory. we need kmem_cache_node
3575 * for it yet.
3576 */
3577 if (offline_node < 0)
3578 return;
3579
3580 down_read(&slub_lock);
3581 list_for_each_entry(s, &slab_caches, list) {
3582 n = get_node(s, offline_node);
3583 if (n) {
3584 /*
3585 * if n->nr_slabs > 0, slabs still exist on the node
3586 * that is going down. We were unable to free them,
3587 * and offline_pages() function shouldn't call this
3588 * callback. So, we must fail.
3589 */
3590 BUG_ON(slabs_node(s, offline_node));
3591
3592 s->node[offline_node] = NULL;
3593 kmem_cache_free(kmem_cache_node, n);
3594 }
3595 }
3596 up_read(&slub_lock);
3597}
3598
3599static int slab_mem_going_online_callback(void *arg)
3600{
3601 struct kmem_cache_node *n;
3602 struct kmem_cache *s;
3603 struct memory_notify *marg = arg;
3604 int nid = marg->status_change_nid;
3605 int ret = 0;
3606
3607 /*
3608 * If the node's memory is already available, then kmem_cache_node is
3609 * already created. Nothing to do.
3610 */
3611 if (nid < 0)
3612 return 0;
3613
3614 /*
3615 * We are bringing a node online. No memory is available yet. We must
3616 * allocate a kmem_cache_node structure in order to bring the node
3617 * online.
3618 */
3619 down_read(&slub_lock);
3620 list_for_each_entry(s, &slab_caches, list) {
3621 /*
3622 * XXX: kmem_cache_alloc_node will fallback to other nodes
3623 * since memory is not yet available from the node that
3624 * is brought up.
3625 */
3626 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3627 if (!n) {
3628 ret = -ENOMEM;
3629 goto out;
3630 }
3631 init_kmem_cache_node(n);
3632 s->node[nid] = n;
3633 }
3634out:
3635 up_read(&slub_lock);
3636 return ret;
3637}
3638
3639static int slab_memory_callback(struct notifier_block *self,
3640 unsigned long action, void *arg)
3641{
3642 int ret = 0;
3643
3644 switch (action) {
3645 case MEM_GOING_ONLINE:
3646 ret = slab_mem_going_online_callback(arg);
3647 break;
3648 case MEM_GOING_OFFLINE:
3649 ret = slab_mem_going_offline_callback(arg);
3650 break;
3651 case MEM_OFFLINE:
3652 case MEM_CANCEL_ONLINE:
3653 slab_mem_offline_callback(arg);
3654 break;
3655 case MEM_ONLINE:
3656 case MEM_CANCEL_OFFLINE:
3657 break;
3658 }
3659 if (ret)
3660 ret = notifier_from_errno(ret);
3661 else
3662 ret = NOTIFY_OK;
3663 return ret;
3664}
3665
3666#endif /* CONFIG_MEMORY_HOTPLUG */
3667
3668/********************************************************************
3669 * Basic setup of slabs
3670 *******************************************************************/
3671
3672/*
3673 * Used for early kmem_cache structures that were allocated using
3674 * the page allocator
3675 */
3676
3677static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3678{
3679 int node;
3680
3681 list_add(&s->list, &slab_caches);
3682 s->refcount = -1;
3683
3684 for_each_node_state(node, N_NORMAL_MEMORY) {
3685 struct kmem_cache_node *n = get_node(s, node);
3686 struct page *p;
3687
3688 if (n) {
3689 list_for_each_entry(p, &n->partial, lru)
3690 p->slab = s;
3691
3692#ifdef CONFIG_SLUB_DEBUG
3693 list_for_each_entry(p, &n->full, lru)
3694 p->slab = s;
3695#endif
3696 }
3697 }
3698}
3699
3700void __init kmem_cache_init(void)
3701{
3702 int i;
3703 int caches = 0;
3704 struct kmem_cache *temp_kmem_cache;
3705 int order;
3706 struct kmem_cache *temp_kmem_cache_node;
3707 unsigned long kmalloc_size;
3708
3709 if (debug_guardpage_minorder())
3710 slub_max_order = 0;
3711
3712 kmem_size = offsetof(struct kmem_cache, node) +
3713 nr_node_ids * sizeof(struct kmem_cache_node *);
3714
3715 /* Allocate two kmem_caches from the page allocator */
3716 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3717 order = get_order(2 * kmalloc_size);
3718 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3719
3720 /*
3721 * Must first have the slab cache available for the allocations of the
3722 * struct kmem_cache_node's. There is special bootstrap code in
3723 * kmem_cache_open for slab_state == DOWN.
3724 */
3725 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3726
3727 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3728 sizeof(struct kmem_cache_node),
3729 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3730
3731 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3732
3733 /* Able to allocate the per node structures */
3734 slab_state = PARTIAL;
3735
3736 temp_kmem_cache = kmem_cache;
3737 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3738 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3739 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3740 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3741
3742 /*
3743 * Allocate kmem_cache_node properly from the kmem_cache slab.
3744 * kmem_cache_node is separately allocated so no need to
3745 * update any list pointers.
3746 */
3747 temp_kmem_cache_node = kmem_cache_node;
3748
3749 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3750 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3751
3752 kmem_cache_bootstrap_fixup(kmem_cache_node);
3753
3754 caches++;
3755 kmem_cache_bootstrap_fixup(kmem_cache);
3756 caches++;
3757 /* Free temporary boot structure */
3758 free_pages((unsigned long)temp_kmem_cache, order);
3759
3760 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3761
3762 /*
3763 * Patch up the size_index table if we have strange large alignment
3764 * requirements for the kmalloc array. This is only the case for
3765 * MIPS it seems. The standard arches will not generate any code here.
3766 *
3767 * Largest permitted alignment is 256 bytes due to the way we
3768 * handle the index determination for the smaller caches.
3769 *
3770 * Make sure that nothing crazy happens if someone starts tinkering
3771 * around with ARCH_KMALLOC_MINALIGN
3772 */
3773 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3774 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3775
3776 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3777 int elem = size_index_elem(i);
3778 if (elem >= ARRAY_SIZE(size_index))
3779 break;
3780 size_index[elem] = KMALLOC_SHIFT_LOW;
3781 }
3782
3783 if (KMALLOC_MIN_SIZE == 64) {
3784 /*
3785 * The 96 byte size cache is not used if the alignment
3786 * is 64 byte.
3787 */
3788 for (i = 64 + 8; i <= 96; i += 8)
3789 size_index[size_index_elem(i)] = 7;
3790 } else if (KMALLOC_MIN_SIZE == 128) {
3791 /*
3792 * The 192 byte sized cache is not used if the alignment
3793 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3794 * instead.
3795 */
3796 for (i = 128 + 8; i <= 192; i += 8)
3797 size_index[size_index_elem(i)] = 8;
3798 }
3799
3800 /* Caches that are not of the two-to-the-power-of size */
3801 if (KMALLOC_MIN_SIZE <= 32) {
3802 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3803 caches++;
3804 }
3805
3806 if (KMALLOC_MIN_SIZE <= 64) {
3807 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3808 caches++;
3809 }
3810
3811 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3812 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3813 caches++;
3814 }
3815
3816 slab_state = UP;
3817
3818 /* Provide the correct kmalloc names now that the caches are up */
3819 if (KMALLOC_MIN_SIZE <= 32) {
3820 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3821 BUG_ON(!kmalloc_caches[1]->name);
3822 }
3823
3824 if (KMALLOC_MIN_SIZE <= 64) {
3825 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3826 BUG_ON(!kmalloc_caches[2]->name);
3827 }
3828
3829 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3830 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3831
3832 BUG_ON(!s);
3833 kmalloc_caches[i]->name = s;
3834 }
3835
3836#ifdef CONFIG_SMP
3837 register_cpu_notifier(&slab_notifier);
3838#endif
3839
3840#ifdef CONFIG_ZONE_DMA
3841 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3842 struct kmem_cache *s = kmalloc_caches[i];
3843
3844 if (s && s->size) {
3845 char *name = kasprintf(GFP_NOWAIT,
3846 "dma-kmalloc-%d", s->objsize);
3847
3848 BUG_ON(!name);
3849 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3850 s->objsize, SLAB_CACHE_DMA);
3851 }
3852 }
3853#endif
3854 printk(KERN_INFO
3855 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3856 " CPUs=%d, Nodes=%d\n",
3857 caches, cache_line_size(),
3858 slub_min_order, slub_max_order, slub_min_objects,
3859 nr_cpu_ids, nr_node_ids);
3860}
3861
3862void __init kmem_cache_init_late(void)
3863{
3864}
3865
3866/*
3867 * Find a mergeable slab cache
3868 */
3869static int slab_unmergeable(struct kmem_cache *s)
3870{
3871 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3872 return 1;
3873
3874 if (s->ctor)
3875 return 1;
3876
3877 /*
3878 * We may have set a slab to be unmergeable during bootstrap.
3879 */
3880 if (s->refcount < 0)
3881 return 1;
3882
3883 return 0;
3884}
3885
3886static struct kmem_cache *find_mergeable(size_t size,
3887 size_t align, unsigned long flags, const char *name,
3888 void (*ctor)(void *))
3889{
3890 struct kmem_cache *s;
3891
3892 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3893 return NULL;
3894
3895 if (ctor)
3896 return NULL;
3897
3898 size = ALIGN(size, sizeof(void *));
3899 align = calculate_alignment(flags, align, size);
3900 size = ALIGN(size, align);
3901 flags = kmem_cache_flags(size, flags, name, NULL);
3902
3903 list_for_each_entry(s, &slab_caches, list) {
3904 if (slab_unmergeable(s))
3905 continue;
3906
3907 if (size > s->size)
3908 continue;
3909
3910 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3911 continue;
3912 /*
3913 * Check if alignment is compatible.
3914 * Courtesy of Adrian Drzewiecki
3915 */
3916 if ((s->size & ~(align - 1)) != s->size)
3917 continue;
3918
3919 if (s->size - size >= sizeof(void *))
3920 continue;
3921
3922 return s;
3923 }
3924 return NULL;
3925}
3926
3927struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3928 size_t align, unsigned long flags, void (*ctor)(void *))
3929{
3930 struct kmem_cache *s;
3931 char *n;
3932
3933 if (WARN_ON(!name))
3934 return NULL;
3935
3936 down_write(&slub_lock);
3937 s = find_mergeable(size, align, flags, name, ctor);
3938 if (s) {
3939 s->refcount++;
3940 /*
3941 * Adjust the object sizes so that we clear
3942 * the complete object on kzalloc.
3943 */
3944 s->objsize = max(s->objsize, (int)size);
3945 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3946
3947 if (sysfs_slab_alias(s, name)) {
3948 s->refcount--;
3949 goto err;
3950 }
3951 up_write(&slub_lock);
3952 return s;
3953 }
3954
3955 n = kstrdup(name, GFP_KERNEL);
3956 if (!n)
3957 goto err;
3958
3959 s = kmalloc(kmem_size, GFP_KERNEL);
3960 if (s) {
3961 if (kmem_cache_open(s, n,
3962 size, align, flags, ctor)) {
3963 list_add(&s->list, &slab_caches);
3964 up_write(&slub_lock);
3965 if (sysfs_slab_add(s)) {
3966 down_write(&slub_lock);
3967 list_del(&s->list);
3968 kfree(n);
3969 kfree(s);
3970 goto err;
3971 }
3972 return s;
3973 }
3974 kfree(s);
3975 }
3976 kfree(n);
3977err:
3978 up_write(&slub_lock);
3979
3980 if (flags & SLAB_PANIC)
3981 panic("Cannot create slabcache %s\n", name);
3982 else
3983 s = NULL;
3984 return s;
3985}
3986EXPORT_SYMBOL(kmem_cache_create);
3987
3988#ifdef CONFIG_SMP
3989/*
3990 * Use the cpu notifier to insure that the cpu slabs are flushed when
3991 * necessary.
3992 */
3993static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3994 unsigned long action, void *hcpu)
3995{
3996 long cpu = (long)hcpu;
3997 struct kmem_cache *s;
3998 unsigned long flags;
3999
4000 switch (action) {
4001 case CPU_UP_CANCELED:
4002 case CPU_UP_CANCELED_FROZEN:
4003 case CPU_DEAD:
4004 case CPU_DEAD_FROZEN:
4005 down_read(&slub_lock);
4006 list_for_each_entry(s, &slab_caches, list) {
4007 local_irq_save(flags);
4008 __flush_cpu_slab(s, cpu);
4009 local_irq_restore(flags);
4010 }
4011 up_read(&slub_lock);
4012 break;
4013 default:
4014 break;
4015 }
4016 return NOTIFY_OK;
4017}
4018
4019static struct notifier_block __cpuinitdata slab_notifier = {
4020 .notifier_call = slab_cpuup_callback
4021};
4022
4023#endif
4024
4025void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
4026{
4027 struct kmem_cache *s;
4028 void *ret;
4029
4030 if (unlikely(size > SLUB_MAX_SIZE))
4031 return kmalloc_large(size, gfpflags);
4032
4033 s = get_slab(size, gfpflags);
4034
4035 if (unlikely(ZERO_OR_NULL_PTR(s)))
4036 return s;
4037
4038 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
4039
4040 /* Honor the call site pointer we received. */
4041 trace_kmalloc(caller, ret, size, s->size, gfpflags);
4042
4043 return ret;
4044}
4045
4046#ifdef CONFIG_NUMA
4047void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
4048 int node, unsigned long caller)
4049{
4050 struct kmem_cache *s;
4051 void *ret;
4052
4053 if (unlikely(size > SLUB_MAX_SIZE)) {
4054 ret = kmalloc_large_node(size, gfpflags, node);
4055
4056 trace_kmalloc_node(caller, ret,
4057 size, PAGE_SIZE << get_order(size),
4058 gfpflags, node);
4059
4060 return ret;
4061 }
4062
4063 s = get_slab(size, gfpflags);
4064
4065 if (unlikely(ZERO_OR_NULL_PTR(s)))
4066 return s;
4067
4068 ret = slab_alloc(s, gfpflags, node, caller);
4069
4070 /* Honor the call site pointer we received. */
4071 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
4072
4073 return ret;
4074}
4075#endif
4076
4077#ifdef CONFIG_SYSFS
4078static int count_inuse(struct page *page)
4079{
4080 return page->inuse;
4081}
4082
4083static int count_total(struct page *page)
4084{
4085 return page->objects;
4086}
4087#endif
4088
4089#ifdef CONFIG_SLUB_DEBUG
4090static int validate_slab(struct kmem_cache *s, struct page *page,
4091 unsigned long *map)
4092{
4093 void *p;
4094 void *addr = page_address(page);
4095
4096 if (!check_slab(s, page) ||
4097 !on_freelist(s, page, NULL))
4098 return 0;
4099
4100 /* Now we know that a valid freelist exists */
4101 bitmap_zero(map, page->objects);
4102
4103 get_map(s, page, map);
4104 for_each_object(p, s, addr, page->objects) {
4105 if (test_bit(slab_index(p, s, addr), map))
4106 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
4107 return 0;
4108 }
4109
4110 for_each_object(p, s, addr, page->objects)
4111 if (!test_bit(slab_index(p, s, addr), map))
4112 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
4113 return 0;
4114 return 1;
4115}
4116
4117static void validate_slab_slab(struct kmem_cache *s, struct page *page,
4118 unsigned long *map)
4119{
4120 slab_lock(page);
4121 validate_slab(s, page, map);
4122 slab_unlock(page);
4123}
4124
4125static int validate_slab_node(struct kmem_cache *s,
4126 struct kmem_cache_node *n, unsigned long *map)
4127{
4128 unsigned long count = 0;
4129 struct page *page;
4130 unsigned long flags;
4131
4132 spin_lock_irqsave(&n->list_lock, flags);
4133
4134 list_for_each_entry(page, &n->partial, lru) {
4135 validate_slab_slab(s, page, map);
4136 count++;
4137 }
4138 if (count != n->nr_partial)
4139 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
4140 "counter=%ld\n", s->name, count, n->nr_partial);
4141
4142 if (!(s->flags & SLAB_STORE_USER))
4143 goto out;
4144
4145 list_for_each_entry(page, &n->full, lru) {
4146 validate_slab_slab(s, page, map);
4147 count++;
4148 }
4149 if (count != atomic_long_read(&n->nr_slabs))
4150 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
4151 "counter=%ld\n", s->name, count,
4152 atomic_long_read(&n->nr_slabs));
4153
4154out:
4155 spin_unlock_irqrestore(&n->list_lock, flags);
4156 return count;
4157}
4158
4159static long validate_slab_cache(struct kmem_cache *s)
4160{
4161 int node;
4162 unsigned long count = 0;
4163 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4164 sizeof(unsigned long), GFP_KERNEL);
4165
4166 if (!map)
4167 return -ENOMEM;
4168
4169 flush_all(s);
4170 for_each_node_state(node, N_NORMAL_MEMORY) {
4171 struct kmem_cache_node *n = get_node(s, node);
4172
4173 count += validate_slab_node(s, n, map);
4174 }
4175 kfree(map);
4176 return count;
4177}
4178/*
4179 * Generate lists of code addresses where slabcache objects are allocated
4180 * and freed.
4181 */
4182
4183struct location {
4184 unsigned long count;
4185 unsigned long addr;
4186 long long sum_time;
4187 long min_time;
4188 long max_time;
4189 long min_pid;
4190 long max_pid;
4191 DECLARE_BITMAP(cpus, NR_CPUS);
4192 nodemask_t nodes;
4193};
4194
4195struct loc_track {
4196 unsigned long max;
4197 unsigned long count;
4198 struct location *loc;
4199};
4200
4201static void free_loc_track(struct loc_track *t)
4202{
4203 if (t->max)
4204 free_pages((unsigned long)t->loc,
4205 get_order(sizeof(struct location) * t->max));
4206}
4207
4208static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
4209{
4210 struct location *l;
4211 int order;
4212
4213 order = get_order(sizeof(struct location) * max);
4214
4215 l = (void *)__get_free_pages(flags, order);
4216 if (!l)
4217 return 0;
4218
4219 if (t->count) {
4220 memcpy(l, t->loc, sizeof(struct location) * t->count);
4221 free_loc_track(t);
4222 }
4223 t->max = max;
4224 t->loc = l;
4225 return 1;
4226}
4227
4228static int add_location(struct loc_track *t, struct kmem_cache *s,
4229 const struct track *track)
4230{
4231 long start, end, pos;
4232 struct location *l;
4233 unsigned long caddr;
4234 unsigned long age = jiffies - track->when;
4235
4236 start = -1;
4237 end = t->count;
4238
4239 for ( ; ; ) {
4240 pos = start + (end - start + 1) / 2;
4241
4242 /*
4243 * There is nothing at "end". If we end up there
4244 * we need to add something to before end.
4245 */
4246 if (pos == end)
4247 break;
4248
4249 caddr = t->loc[pos].addr;
4250 if (track->addr == caddr) {
4251
4252 l = &t->loc[pos];
4253 l->count++;
4254 if (track->when) {
4255 l->sum_time += age;
4256 if (age < l->min_time)
4257 l->min_time = age;
4258 if (age > l->max_time)
4259 l->max_time = age;
4260
4261 if (track->pid < l->min_pid)
4262 l->min_pid = track->pid;
4263 if (track->pid > l->max_pid)
4264 l->max_pid = track->pid;
4265
4266 cpumask_set_cpu(track->cpu,
4267 to_cpumask(l->cpus));
4268 }
4269 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4270 return 1;
4271 }
4272
4273 if (track->addr < caddr)
4274 end = pos;
4275 else
4276 start = pos;
4277 }
4278
4279 /*
4280 * Not found. Insert new tracking element.
4281 */
4282 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
4283 return 0;
4284
4285 l = t->loc + pos;
4286 if (pos < t->count)
4287 memmove(l + 1, l,
4288 (t->count - pos) * sizeof(struct location));
4289 t->count++;
4290 l->count = 1;
4291 l->addr = track->addr;
4292 l->sum_time = age;
4293 l->min_time = age;
4294 l->max_time = age;
4295 l->min_pid = track->pid;
4296 l->max_pid = track->pid;
4297 cpumask_clear(to_cpumask(l->cpus));
4298 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
4299 nodes_clear(l->nodes);
4300 node_set(page_to_nid(virt_to_page(track)), l->nodes);
4301 return 1;
4302}
4303
4304static void process_slab(struct loc_track *t, struct kmem_cache *s,
4305 struct page *page, enum track_item alloc,
4306 unsigned long *map)
4307{
4308 void *addr = page_address(page);
4309 void *p;
4310
4311 bitmap_zero(map, page->objects);
4312 get_map(s, page, map);
4313
4314 for_each_object(p, s, addr, page->objects)
4315 if (!test_bit(slab_index(p, s, addr), map))
4316 add_location(t, s, get_track(s, p, alloc));
4317}
4318
4319static int list_locations(struct kmem_cache *s, char *buf,
4320 enum track_item alloc)
4321{
4322 int len = 0;
4323 unsigned long i;
4324 struct loc_track t = { 0, 0, NULL };
4325 int node;
4326 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
4327 sizeof(unsigned long), GFP_KERNEL);
4328
4329 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
4330 GFP_TEMPORARY)) {
4331 kfree(map);
4332 return sprintf(buf, "Out of memory\n");
4333 }
4334 /* Push back cpu slabs */
4335 flush_all(s);
4336
4337 for_each_node_state(node, N_NORMAL_MEMORY) {
4338 struct kmem_cache_node *n = get_node(s, node);
4339 unsigned long flags;
4340 struct page *page;
4341
4342 if (!atomic_long_read(&n->nr_slabs))
4343 continue;
4344
4345 spin_lock_irqsave(&n->list_lock, flags);
4346 list_for_each_entry(page, &n->partial, lru)
4347 process_slab(&t, s, page, alloc, map);
4348 list_for_each_entry(page, &n->full, lru)
4349 process_slab(&t, s, page, alloc, map);
4350 spin_unlock_irqrestore(&n->list_lock, flags);
4351 }
4352
4353 for (i = 0; i < t.count; i++) {
4354 struct location *l = &t.loc[i];
4355
4356 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
4357 break;
4358 len += sprintf(buf + len, "%7ld ", l->count);
4359
4360 if (l->addr)
4361 len += sprintf(buf + len, "%pS", (void *)l->addr);
4362 else
4363 len += sprintf(buf + len, "<not-available>");
4364
4365 if (l->sum_time != l->min_time) {
4366 len += sprintf(buf + len, " age=%ld/%ld/%ld",
4367 l->min_time,
4368 (long)div_u64(l->sum_time, l->count),
4369 l->max_time);
4370 } else
4371 len += sprintf(buf + len, " age=%ld",
4372 l->min_time);
4373
4374 if (l->min_pid != l->max_pid)
4375 len += sprintf(buf + len, " pid=%ld-%ld",
4376 l->min_pid, l->max_pid);
4377 else
4378 len += sprintf(buf + len, " pid=%ld",
4379 l->min_pid);
4380
4381 if (num_online_cpus() > 1 &&
4382 !cpumask_empty(to_cpumask(l->cpus)) &&
4383 len < PAGE_SIZE - 60) {
4384 len += sprintf(buf + len, " cpus=");
4385 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4386 to_cpumask(l->cpus));
4387 }
4388
4389 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
4390 len < PAGE_SIZE - 60) {
4391 len += sprintf(buf + len, " nodes=");
4392 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
4393 l->nodes);
4394 }
4395
4396 len += sprintf(buf + len, "\n");
4397 }
4398
4399 free_loc_track(&t);
4400 kfree(map);
4401 if (!t.count)
4402 len += sprintf(buf, "No data\n");
4403 return len;
4404}
4405#endif
4406
4407#ifdef SLUB_RESILIENCY_TEST
4408static void resiliency_test(void)
4409{
4410 u8 *p;
4411
4412 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
4413
4414 printk(KERN_ERR "SLUB resiliency testing\n");
4415 printk(KERN_ERR "-----------------------\n");
4416 printk(KERN_ERR "A. Corruption after allocation\n");
4417
4418 p = kzalloc(16, GFP_KERNEL);
4419 p[16] = 0x12;
4420 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4421 " 0x12->0x%p\n\n", p + 16);
4422
4423 validate_slab_cache(kmalloc_caches[4]);
4424
4425 /* Hmmm... The next two are dangerous */
4426 p = kzalloc(32, GFP_KERNEL);
4427 p[32 + sizeof(void *)] = 0x34;
4428 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4429 " 0x34 -> -0x%p\n", p);
4430 printk(KERN_ERR
4431 "If allocated object is overwritten then not detectable\n\n");
4432
4433 validate_slab_cache(kmalloc_caches[5]);
4434 p = kzalloc(64, GFP_KERNEL);
4435 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4436 *p = 0x56;
4437 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4438 p);
4439 printk(KERN_ERR
4440 "If allocated object is overwritten then not detectable\n\n");
4441 validate_slab_cache(kmalloc_caches[6]);
4442
4443 printk(KERN_ERR "\nB. Corruption after free\n");
4444 p = kzalloc(128, GFP_KERNEL);
4445 kfree(p);
4446 *p = 0x78;
4447 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4448 validate_slab_cache(kmalloc_caches[7]);
4449
4450 p = kzalloc(256, GFP_KERNEL);
4451 kfree(p);
4452 p[50] = 0x9a;
4453 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4454 p);
4455 validate_slab_cache(kmalloc_caches[8]);
4456
4457 p = kzalloc(512, GFP_KERNEL);
4458 kfree(p);
4459 p[512] = 0xab;
4460 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4461 validate_slab_cache(kmalloc_caches[9]);
4462}
4463#else
4464#ifdef CONFIG_SYSFS
4465static void resiliency_test(void) {};
4466#endif
4467#endif
4468
4469#ifdef CONFIG_SYSFS
4470enum slab_stat_type {
4471 SL_ALL, /* All slabs */
4472 SL_PARTIAL, /* Only partially allocated slabs */
4473 SL_CPU, /* Only slabs used for cpu caches */
4474 SL_OBJECTS, /* Determine allocated objects not slabs */
4475 SL_TOTAL /* Determine object capacity not slabs */
4476};
4477
4478#define SO_ALL (1 << SL_ALL)
4479#define SO_PARTIAL (1 << SL_PARTIAL)
4480#define SO_CPU (1 << SL_CPU)
4481#define SO_OBJECTS (1 << SL_OBJECTS)
4482#define SO_TOTAL (1 << SL_TOTAL)
4483
4484static ssize_t show_slab_objects(struct kmem_cache *s,
4485 char *buf, unsigned long flags)
4486{
4487 unsigned long total = 0;
4488 int node;
4489 int x;
4490 unsigned long *nodes;
4491 unsigned long *per_cpu;
4492
4493 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4494 if (!nodes)
4495 return -ENOMEM;
4496 per_cpu = nodes + nr_node_ids;
4497
4498 if (flags & SO_CPU) {
4499 int cpu;
4500
4501 for_each_possible_cpu(cpu) {
4502 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4503 int node = ACCESS_ONCE(c->node);
4504 struct page *page;
4505
4506 if (node < 0)
4507 continue;
4508 page = ACCESS_ONCE(c->page);
4509 if (page) {
4510 if (flags & SO_TOTAL)
4511 x = page->objects;
4512 else if (flags & SO_OBJECTS)
4513 x = page->inuse;
4514 else
4515 x = 1;
4516
4517 total += x;
4518 nodes[node] += x;
4519 }
4520 page = c->partial;
4521
4522 if (page) {
4523 x = page->pobjects;
4524 total += x;
4525 nodes[node] += x;
4526 }
4527 per_cpu[node]++;
4528 }
4529 }
4530
4531 lock_memory_hotplug();
4532#ifdef CONFIG_SLUB_DEBUG
4533 if (flags & SO_ALL) {
4534 for_each_node_state(node, N_NORMAL_MEMORY) {
4535 struct kmem_cache_node *n = get_node(s, node);
4536
4537 if (flags & SO_TOTAL)
4538 x = atomic_long_read(&n->total_objects);
4539 else if (flags & SO_OBJECTS)
4540 x = atomic_long_read(&n->total_objects) -
4541 count_partial(n, count_free);
4542
4543 else
4544 x = atomic_long_read(&n->nr_slabs);
4545 total += x;
4546 nodes[node] += x;
4547 }
4548
4549 } else
4550#endif
4551 if (flags & SO_PARTIAL) {
4552 for_each_node_state(node, N_NORMAL_MEMORY) {
4553 struct kmem_cache_node *n = get_node(s, node);
4554
4555 if (flags & SO_TOTAL)
4556 x = count_partial(n, count_total);
4557 else if (flags & SO_OBJECTS)
4558 x = count_partial(n, count_inuse);
4559 else
4560 x = n->nr_partial;
4561 total += x;
4562 nodes[node] += x;
4563 }
4564 }
4565 x = sprintf(buf, "%lu", total);
4566#ifdef CONFIG_NUMA
4567 for_each_node_state(node, N_NORMAL_MEMORY)
4568 if (nodes[node])
4569 x += sprintf(buf + x, " N%d=%lu",
4570 node, nodes[node]);
4571#endif
4572 unlock_memory_hotplug();
4573 kfree(nodes);
4574 return x + sprintf(buf + x, "\n");
4575}
4576
4577#ifdef CONFIG_SLUB_DEBUG
4578static int any_slab_objects(struct kmem_cache *s)
4579{
4580 int node;
4581
4582 for_each_online_node(node) {
4583 struct kmem_cache_node *n = get_node(s, node);
4584
4585 if (!n)
4586 continue;
4587
4588 if (atomic_long_read(&n->total_objects))
4589 return 1;
4590 }
4591 return 0;
4592}
4593#endif
4594
4595#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4596#define to_slab(n) container_of(n, struct kmem_cache, kobj)
4597
4598struct slab_attribute {
4599 struct attribute attr;
4600 ssize_t (*show)(struct kmem_cache *s, char *buf);
4601 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4602};
4603
4604#define SLAB_ATTR_RO(_name) \
4605 static struct slab_attribute _name##_attr = \
4606 __ATTR(_name, 0400, _name##_show, NULL)
4607
4608#define SLAB_ATTR(_name) \
4609 static struct slab_attribute _name##_attr = \
4610 __ATTR(_name, 0600, _name##_show, _name##_store)
4611
4612static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4613{
4614 return sprintf(buf, "%d\n", s->size);
4615}
4616SLAB_ATTR_RO(slab_size);
4617
4618static ssize_t align_show(struct kmem_cache *s, char *buf)
4619{
4620 return sprintf(buf, "%d\n", s->align);
4621}
4622SLAB_ATTR_RO(align);
4623
4624static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4625{
4626 return sprintf(buf, "%d\n", s->objsize);
4627}
4628SLAB_ATTR_RO(object_size);
4629
4630static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4631{
4632 return sprintf(buf, "%d\n", oo_objects(s->oo));
4633}
4634SLAB_ATTR_RO(objs_per_slab);
4635
4636static ssize_t order_store(struct kmem_cache *s,
4637 const char *buf, size_t length)
4638{
4639 unsigned long order;
4640 int err;
4641
4642 err = strict_strtoul(buf, 10, &order);
4643 if (err)
4644 return err;
4645
4646 if (order > slub_max_order || order < slub_min_order)
4647 return -EINVAL;
4648
4649 calculate_sizes(s, order);
4650 return length;
4651}
4652
4653static ssize_t order_show(struct kmem_cache *s, char *buf)
4654{
4655 return sprintf(buf, "%d\n", oo_order(s->oo));
4656}
4657SLAB_ATTR(order);
4658
4659static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4660{
4661 return sprintf(buf, "%lu\n", s->min_partial);
4662}
4663
4664static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4665 size_t length)
4666{
4667 unsigned long min;
4668 int err;
4669
4670 err = strict_strtoul(buf, 10, &min);
4671 if (err)
4672 return err;
4673
4674 set_min_partial(s, min);
4675 return length;
4676}
4677SLAB_ATTR(min_partial);
4678
4679static ssize_t cpu_partial_show(struct kmem_cache *s, char *buf)
4680{
4681 return sprintf(buf, "%u\n", s->cpu_partial);
4682}
4683
4684static ssize_t cpu_partial_store(struct kmem_cache *s, const char *buf,
4685 size_t length)
4686{
4687 unsigned long objects;
4688 int err;
4689
4690 err = strict_strtoul(buf, 10, &objects);
4691 if (err)
4692 return err;
4693 if (objects && kmem_cache_debug(s))
4694 return -EINVAL;
4695
4696 s->cpu_partial = objects;
4697 flush_all(s);
4698 return length;
4699}
4700SLAB_ATTR(cpu_partial);
4701
4702static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4703{
4704 if (!s->ctor)
4705 return 0;
4706 return sprintf(buf, "%pS\n", s->ctor);
4707}
4708SLAB_ATTR_RO(ctor);
4709
4710static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4711{
4712 return sprintf(buf, "%d\n", s->refcount - 1);
4713}
4714SLAB_ATTR_RO(aliases);
4715
4716static ssize_t partial_show(struct kmem_cache *s, char *buf)
4717{
4718 return show_slab_objects(s, buf, SO_PARTIAL);
4719}
4720SLAB_ATTR_RO(partial);
4721
4722static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4723{
4724 return show_slab_objects(s, buf, SO_CPU);
4725}
4726SLAB_ATTR_RO(cpu_slabs);
4727
4728static ssize_t objects_show(struct kmem_cache *s, char *buf)
4729{
4730 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4731}
4732SLAB_ATTR_RO(objects);
4733
4734static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4735{
4736 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4737}
4738SLAB_ATTR_RO(objects_partial);
4739
4740static ssize_t slabs_cpu_partial_show(struct kmem_cache *s, char *buf)
4741{
4742 int objects = 0;
4743 int pages = 0;
4744 int cpu;
4745 int len;
4746
4747 for_each_online_cpu(cpu) {
4748 struct page *page = per_cpu_ptr(s->cpu_slab, cpu)->partial;
4749
4750 if (page) {
4751 pages += page->pages;
4752 objects += page->pobjects;
4753 }
4754 }
4755
4756 len = sprintf(buf, "%d(%d)", objects, pages);
4757
4758#ifdef CONFIG_SMP
4759 for_each_online_cpu(cpu) {
4760 struct page *page = per_cpu_ptr(s->cpu_slab, cpu) ->partial;
4761
4762 if (page && len < PAGE_SIZE - 20)
4763 len += sprintf(buf + len, " C%d=%d(%d)", cpu,
4764 page->pobjects, page->pages);
4765 }
4766#endif
4767 return len + sprintf(buf + len, "\n");
4768}
4769SLAB_ATTR_RO(slabs_cpu_partial);
4770
4771static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4772{
4773 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4774}
4775
4776static ssize_t reclaim_account_store(struct kmem_cache *s,
4777 const char *buf, size_t length)
4778{
4779 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4780 if (buf[0] == '1')
4781 s->flags |= SLAB_RECLAIM_ACCOUNT;
4782 return length;
4783}
4784SLAB_ATTR(reclaim_account);
4785
4786static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4787{
4788 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4789}
4790SLAB_ATTR_RO(hwcache_align);
4791
4792#ifdef CONFIG_ZONE_DMA
4793static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4794{
4795 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4796}
4797SLAB_ATTR_RO(cache_dma);
4798#endif
4799
4800static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4801{
4802 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4803}
4804SLAB_ATTR_RO(destroy_by_rcu);
4805
4806static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4807{
4808 return sprintf(buf, "%d\n", s->reserved);
4809}
4810SLAB_ATTR_RO(reserved);
4811
4812#ifdef CONFIG_SLUB_DEBUG
4813static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4814{
4815 return show_slab_objects(s, buf, SO_ALL);
4816}
4817SLAB_ATTR_RO(slabs);
4818
4819static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4820{
4821 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4822}
4823SLAB_ATTR_RO(total_objects);
4824
4825static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4826{
4827 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4828}
4829
4830static ssize_t sanity_checks_store(struct kmem_cache *s,
4831 const char *buf, size_t length)
4832{
4833 s->flags &= ~SLAB_DEBUG_FREE;
4834 if (buf[0] == '1') {
4835 s->flags &= ~__CMPXCHG_DOUBLE;
4836 s->flags |= SLAB_DEBUG_FREE;
4837 }
4838 return length;
4839}
4840SLAB_ATTR(sanity_checks);
4841
4842static ssize_t trace_show(struct kmem_cache *s, char *buf)
4843{
4844 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4845}
4846
4847static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4848 size_t length)
4849{
4850 s->flags &= ~SLAB_TRACE;
4851 if (buf[0] == '1') {
4852 s->flags &= ~__CMPXCHG_DOUBLE;
4853 s->flags |= SLAB_TRACE;
4854 }
4855 return length;
4856}
4857SLAB_ATTR(trace);
4858
4859static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4860{
4861 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4862}
4863
4864static ssize_t red_zone_store(struct kmem_cache *s,
4865 const char *buf, size_t length)
4866{
4867 if (any_slab_objects(s))
4868 return -EBUSY;
4869
4870 s->flags &= ~SLAB_RED_ZONE;
4871 if (buf[0] == '1') {
4872 s->flags &= ~__CMPXCHG_DOUBLE;
4873 s->flags |= SLAB_RED_ZONE;
4874 }
4875 calculate_sizes(s, -1);
4876 return length;
4877}
4878SLAB_ATTR(red_zone);
4879
4880static ssize_t poison_show(struct kmem_cache *s, char *buf)
4881{
4882 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4883}
4884
4885static ssize_t poison_store(struct kmem_cache *s,
4886 const char *buf, size_t length)
4887{
4888 if (any_slab_objects(s))
4889 return -EBUSY;
4890
4891 s->flags &= ~SLAB_POISON;
4892 if (buf[0] == '1') {
4893 s->flags &= ~__CMPXCHG_DOUBLE;
4894 s->flags |= SLAB_POISON;
4895 }
4896 calculate_sizes(s, -1);
4897 return length;
4898}
4899SLAB_ATTR(poison);
4900
4901static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4902{
4903 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4904}
4905
4906static ssize_t store_user_store(struct kmem_cache *s,
4907 const char *buf, size_t length)
4908{
4909 if (any_slab_objects(s))
4910 return -EBUSY;
4911
4912 s->flags &= ~SLAB_STORE_USER;
4913 if (buf[0] == '1') {
4914 s->flags &= ~__CMPXCHG_DOUBLE;
4915 s->flags |= SLAB_STORE_USER;
4916 }
4917 calculate_sizes(s, -1);
4918 return length;
4919}
4920SLAB_ATTR(store_user);
4921
4922static ssize_t validate_show(struct kmem_cache *s, char *buf)
4923{
4924 return 0;
4925}
4926
4927static ssize_t validate_store(struct kmem_cache *s,
4928 const char *buf, size_t length)
4929{
4930 int ret = -EINVAL;
4931
4932 if (buf[0] == '1') {
4933 ret = validate_slab_cache(s);
4934 if (ret >= 0)
4935 ret = length;
4936 }
4937 return ret;
4938}
4939SLAB_ATTR(validate);
4940
4941static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4942{
4943 if (!(s->flags & SLAB_STORE_USER))
4944 return -ENOSYS;
4945 return list_locations(s, buf, TRACK_ALLOC);
4946}
4947SLAB_ATTR_RO(alloc_calls);
4948
4949static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4950{
4951 if (!(s->flags & SLAB_STORE_USER))
4952 return -ENOSYS;
4953 return list_locations(s, buf, TRACK_FREE);
4954}
4955SLAB_ATTR_RO(free_calls);
4956#endif /* CONFIG_SLUB_DEBUG */
4957
4958#ifdef CONFIG_FAILSLAB
4959static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4960{
4961 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4962}
4963
4964static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4965 size_t length)
4966{
4967 s->flags &= ~SLAB_FAILSLAB;
4968 if (buf[0] == '1')
4969 s->flags |= SLAB_FAILSLAB;
4970 return length;
4971}
4972SLAB_ATTR(failslab);
4973#endif
4974
4975static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4976{
4977 return 0;
4978}
4979
4980static ssize_t shrink_store(struct kmem_cache *s,
4981 const char *buf, size_t length)
4982{
4983 if (buf[0] == '1') {
4984 int rc = kmem_cache_shrink(s);
4985
4986 if (rc)
4987 return rc;
4988 } else
4989 return -EINVAL;
4990 return length;
4991}
4992SLAB_ATTR(shrink);
4993
4994#ifdef CONFIG_NUMA
4995static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4996{
4997 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4998}
4999
5000static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
5001 const char *buf, size_t length)
5002{
5003 unsigned long ratio;
5004 int err;
5005
5006 err = strict_strtoul(buf, 10, &ratio);
5007 if (err)
5008 return err;
5009
5010 if (ratio <= 100)
5011 s->remote_node_defrag_ratio = ratio * 10;
5012
5013 return length;
5014}
5015SLAB_ATTR(remote_node_defrag_ratio);
5016#endif
5017
5018#ifdef CONFIG_SLUB_STATS
5019static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
5020{
5021 unsigned long sum = 0;
5022 int cpu;
5023 int len;
5024 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
5025
5026 if (!data)
5027 return -ENOMEM;
5028
5029 for_each_online_cpu(cpu) {
5030 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
5031
5032 data[cpu] = x;
5033 sum += x;
5034 }
5035
5036 len = sprintf(buf, "%lu", sum);
5037
5038#ifdef CONFIG_SMP
5039 for_each_online_cpu(cpu) {
5040 if (data[cpu] && len < PAGE_SIZE - 20)
5041 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
5042 }
5043#endif
5044 kfree(data);
5045 return len + sprintf(buf + len, "\n");
5046}
5047
5048static void clear_stat(struct kmem_cache *s, enum stat_item si)
5049{
5050 int cpu;
5051
5052 for_each_online_cpu(cpu)
5053 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
5054}
5055
5056#define STAT_ATTR(si, text) \
5057static ssize_t text##_show(struct kmem_cache *s, char *buf) \
5058{ \
5059 return show_stat(s, buf, si); \
5060} \
5061static ssize_t text##_store(struct kmem_cache *s, \
5062 const char *buf, size_t length) \
5063{ \
5064 if (buf[0] != '0') \
5065 return -EINVAL; \
5066 clear_stat(s, si); \
5067 return length; \
5068} \
5069SLAB_ATTR(text); \
5070
5071STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
5072STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
5073STAT_ATTR(FREE_FASTPATH, free_fastpath);
5074STAT_ATTR(FREE_SLOWPATH, free_slowpath);
5075STAT_ATTR(FREE_FROZEN, free_frozen);
5076STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
5077STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
5078STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
5079STAT_ATTR(ALLOC_SLAB, alloc_slab);
5080STAT_ATTR(ALLOC_REFILL, alloc_refill);
5081STAT_ATTR(ALLOC_NODE_MISMATCH, alloc_node_mismatch);
5082STAT_ATTR(FREE_SLAB, free_slab);
5083STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
5084STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
5085STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
5086STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
5087STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
5088STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
5089STAT_ATTR(DEACTIVATE_BYPASS, deactivate_bypass);
5090STAT_ATTR(ORDER_FALLBACK, order_fallback);
5091STAT_ATTR(CMPXCHG_DOUBLE_CPU_FAIL, cmpxchg_double_cpu_fail);
5092STAT_ATTR(CMPXCHG_DOUBLE_FAIL, cmpxchg_double_fail);
5093STAT_ATTR(CPU_PARTIAL_ALLOC, cpu_partial_alloc);
5094STAT_ATTR(CPU_PARTIAL_FREE, cpu_partial_free);
5095STAT_ATTR(CPU_PARTIAL_NODE, cpu_partial_node);
5096STAT_ATTR(CPU_PARTIAL_DRAIN, cpu_partial_drain);
5097#endif
5098
5099static struct attribute *slab_attrs[] = {
5100 &slab_size_attr.attr,
5101 &object_size_attr.attr,
5102 &objs_per_slab_attr.attr,
5103 &order_attr.attr,
5104 &min_partial_attr.attr,
5105 &cpu_partial_attr.attr,
5106 &objects_attr.attr,
5107 &objects_partial_attr.attr,
5108 &partial_attr.attr,
5109 &cpu_slabs_attr.attr,
5110 &ctor_attr.attr,
5111 &aliases_attr.attr,
5112 &align_attr.attr,
5113 &hwcache_align_attr.attr,
5114 &reclaim_account_attr.attr,
5115 &destroy_by_rcu_attr.attr,
5116 &shrink_attr.attr,
5117 &reserved_attr.attr,
5118 &slabs_cpu_partial_attr.attr,
5119#ifdef CONFIG_SLUB_DEBUG
5120 &total_objects_attr.attr,
5121 &slabs_attr.attr,
5122 &sanity_checks_attr.attr,
5123 &trace_attr.attr,
5124 &red_zone_attr.attr,
5125 &poison_attr.attr,
5126 &store_user_attr.attr,
5127 &validate_attr.attr,
5128 &alloc_calls_attr.attr,
5129 &free_calls_attr.attr,
5130#endif
5131#ifdef CONFIG_ZONE_DMA
5132 &cache_dma_attr.attr,
5133#endif
5134#ifdef CONFIG_NUMA
5135 &remote_node_defrag_ratio_attr.attr,
5136#endif
5137#ifdef CONFIG_SLUB_STATS
5138 &alloc_fastpath_attr.attr,
5139 &alloc_slowpath_attr.attr,
5140 &free_fastpath_attr.attr,
5141 &free_slowpath_attr.attr,
5142 &free_frozen_attr.attr,
5143 &free_add_partial_attr.attr,
5144 &free_remove_partial_attr.attr,
5145 &alloc_from_partial_attr.attr,
5146 &alloc_slab_attr.attr,
5147 &alloc_refill_attr.attr,
5148 &alloc_node_mismatch_attr.attr,
5149 &free_slab_attr.attr,
5150 &cpuslab_flush_attr.attr,
5151 &deactivate_full_attr.attr,
5152 &deactivate_empty_attr.attr,
5153 &deactivate_to_head_attr.attr,
5154 &deactivate_to_tail_attr.attr,
5155 &deactivate_remote_frees_attr.attr,
5156 &deactivate_bypass_attr.attr,
5157 &order_fallback_attr.attr,
5158 &cmpxchg_double_fail_attr.attr,
5159 &cmpxchg_double_cpu_fail_attr.attr,
5160 &cpu_partial_alloc_attr.attr,
5161 &cpu_partial_free_attr.attr,
5162 &cpu_partial_node_attr.attr,
5163 &cpu_partial_drain_attr.attr,
5164#endif
5165#ifdef CONFIG_FAILSLAB
5166 &failslab_attr.attr,
5167#endif
5168
5169 NULL
5170};
5171
5172static struct attribute_group slab_attr_group = {
5173 .attrs = slab_attrs,
5174};
5175
5176static ssize_t slab_attr_show(struct kobject *kobj,
5177 struct attribute *attr,
5178 char *buf)
5179{
5180 struct slab_attribute *attribute;
5181 struct kmem_cache *s;
5182 int err;
5183
5184 attribute = to_slab_attr(attr);
5185 s = to_slab(kobj);
5186
5187 if (!attribute->show)
5188 return -EIO;
5189
5190 err = attribute->show(s, buf);
5191
5192 return err;
5193}
5194
5195static ssize_t slab_attr_store(struct kobject *kobj,
5196 struct attribute *attr,
5197 const char *buf, size_t len)
5198{
5199 struct slab_attribute *attribute;
5200 struct kmem_cache *s;
5201 int err;
5202
5203 attribute = to_slab_attr(attr);
5204 s = to_slab(kobj);
5205
5206 if (!attribute->store)
5207 return -EIO;
5208
5209 err = attribute->store(s, buf, len);
5210
5211 return err;
5212}
5213
5214static void kmem_cache_release(struct kobject *kobj)
5215{
5216 struct kmem_cache *s = to_slab(kobj);
5217
5218 kfree(s->name);
5219 kfree(s);
5220}
5221
5222static const struct sysfs_ops slab_sysfs_ops = {
5223 .show = slab_attr_show,
5224 .store = slab_attr_store,
5225};
5226
5227static struct kobj_type slab_ktype = {
5228 .sysfs_ops = &slab_sysfs_ops,
5229 .release = kmem_cache_release
5230};
5231
5232static int uevent_filter(struct kset *kset, struct kobject *kobj)
5233{
5234 struct kobj_type *ktype = get_ktype(kobj);
5235
5236 if (ktype == &slab_ktype)
5237 return 1;
5238 return 0;
5239}
5240
5241static const struct kset_uevent_ops slab_uevent_ops = {
5242 .filter = uevent_filter,
5243};
5244
5245static struct kset *slab_kset;
5246
5247#define ID_STR_LENGTH 64
5248
5249/* Create a unique string id for a slab cache:
5250 *
5251 * Format :[flags-]size
5252 */
5253static char *create_unique_id(struct kmem_cache *s)
5254{
5255 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
5256 char *p = name;
5257
5258 BUG_ON(!name);
5259
5260 *p++ = ':';
5261 /*
5262 * First flags affecting slabcache operations. We will only
5263 * get here for aliasable slabs so we do not need to support
5264 * too many flags. The flags here must cover all flags that
5265 * are matched during merging to guarantee that the id is
5266 * unique.
5267 */
5268 if (s->flags & SLAB_CACHE_DMA)
5269 *p++ = 'd';
5270 if (s->flags & SLAB_RECLAIM_ACCOUNT)
5271 *p++ = 'a';
5272 if (s->flags & SLAB_DEBUG_FREE)
5273 *p++ = 'F';
5274 if (!(s->flags & SLAB_NOTRACK))
5275 *p++ = 't';
5276 if (p != name + 1)
5277 *p++ = '-';
5278 p += sprintf(p, "%07d", s->size);
5279 BUG_ON(p > name + ID_STR_LENGTH - 1);
5280 return name;
5281}
5282
5283static int sysfs_slab_add(struct kmem_cache *s)
5284{
5285 int err;
5286 const char *name;
5287 int unmergeable;
5288
5289 if (slab_state < SYSFS)
5290 /* Defer until later */
5291 return 0;
5292
5293 unmergeable = slab_unmergeable(s);
5294 if (unmergeable) {
5295 /*
5296 * Slabcache can never be merged so we can use the name proper.
5297 * This is typically the case for debug situations. In that
5298 * case we can catch duplicate names easily.
5299 */
5300 sysfs_remove_link(&slab_kset->kobj, s->name);
5301 name = s->name;
5302 } else {
5303 /*
5304 * Create a unique name for the slab as a target
5305 * for the symlinks.
5306 */
5307 name = create_unique_id(s);
5308 }
5309
5310 s->kobj.kset = slab_kset;
5311 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
5312 if (err) {
5313 kobject_put(&s->kobj);
5314 return err;
5315 }
5316
5317 err = sysfs_create_group(&s->kobj, &slab_attr_group);
5318 if (err) {
5319 kobject_del(&s->kobj);
5320 kobject_put(&s->kobj);
5321 return err;
5322 }
5323 kobject_uevent(&s->kobj, KOBJ_ADD);
5324 if (!unmergeable) {
5325 /* Setup first alias */
5326 sysfs_slab_alias(s, s->name);
5327 kfree(name);
5328 }
5329 return 0;
5330}
5331
5332static void sysfs_slab_remove(struct kmem_cache *s)
5333{
5334 if (slab_state < SYSFS)
5335 /*
5336 * Sysfs has not been setup yet so no need to remove the
5337 * cache from sysfs.
5338 */
5339 return;
5340
5341 kobject_uevent(&s->kobj, KOBJ_REMOVE);
5342 kobject_del(&s->kobj);
5343 kobject_put(&s->kobj);
5344}
5345
5346/*
5347 * Need to buffer aliases during bootup until sysfs becomes
5348 * available lest we lose that information.
5349 */
5350struct saved_alias {
5351 struct kmem_cache *s;
5352 const char *name;
5353 struct saved_alias *next;
5354};
5355
5356static struct saved_alias *alias_list;
5357
5358static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
5359{
5360 struct saved_alias *al;
5361
5362 if (slab_state == SYSFS) {
5363 /*
5364 * If we have a leftover link then remove it.
5365 */
5366 sysfs_remove_link(&slab_kset->kobj, name);
5367 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
5368 }
5369
5370 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
5371 if (!al)
5372 return -ENOMEM;
5373
5374 al->s = s;
5375 al->name = name;
5376 al->next = alias_list;
5377 alias_list = al;
5378 return 0;
5379}
5380
5381static int __init slab_sysfs_init(void)
5382{
5383 struct kmem_cache *s;
5384 int err;
5385
5386 down_write(&slub_lock);
5387
5388 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
5389 if (!slab_kset) {
5390 up_write(&slub_lock);
5391 printk(KERN_ERR "Cannot register slab subsystem.\n");
5392 return -ENOSYS;
5393 }
5394
5395 slab_state = SYSFS;
5396
5397 list_for_each_entry(s, &slab_caches, list) {
5398 err = sysfs_slab_add(s);
5399 if (err)
5400 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
5401 " to sysfs\n", s->name);
5402 }
5403
5404 while (alias_list) {
5405 struct saved_alias *al = alias_list;
5406
5407 alias_list = alias_list->next;
5408 err = sysfs_slab_alias(al->s, al->name);
5409 if (err)
5410 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
5411 " %s to sysfs\n", s->name);
5412 kfree(al);
5413 }
5414
5415 up_write(&slub_lock);
5416 resiliency_test();
5417 return 0;
5418}
5419
5420__initcall(slab_sysfs_init);
5421#endif /* CONFIG_SYSFS */
5422
5423/*
5424 * The /proc/slabinfo ABI
5425 */
5426#ifdef CONFIG_SLABINFO
5427static void print_slabinfo_header(struct seq_file *m)
5428{
5429 seq_puts(m, "slabinfo - version: 2.1\n");
5430 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
5431 "<objperslab> <pagesperslab>");
5432 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
5433 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
5434 seq_putc(m, '\n');
5435}
5436
5437static void *s_start(struct seq_file *m, loff_t *pos)
5438{
5439 loff_t n = *pos;
5440
5441 down_read(&slub_lock);
5442 if (!n)
5443 print_slabinfo_header(m);
5444
5445 return seq_list_start(&slab_caches, *pos);
5446}
5447
5448static void *s_next(struct seq_file *m, void *p, loff_t *pos)
5449{
5450 return seq_list_next(p, &slab_caches, pos);
5451}
5452
5453static void s_stop(struct seq_file *m, void *p)
5454{
5455 up_read(&slub_lock);
5456}
5457
5458static int s_show(struct seq_file *m, void *p)
5459{
5460 unsigned long nr_partials = 0;
5461 unsigned long nr_slabs = 0;
5462 unsigned long nr_inuse = 0;
5463 unsigned long nr_objs = 0;
5464 unsigned long nr_free = 0;
5465 struct kmem_cache *s;
5466 int node;
5467
5468 s = list_entry(p, struct kmem_cache, list);
5469
5470 for_each_online_node(node) {
5471 struct kmem_cache_node *n = get_node(s, node);
5472
5473 if (!n)
5474 continue;
5475
5476 nr_partials += n->nr_partial;
5477 nr_slabs += atomic_long_read(&n->nr_slabs);
5478 nr_objs += atomic_long_read(&n->total_objects);
5479 nr_free += count_partial(n, count_free);
5480 }
5481
5482 nr_inuse = nr_objs - nr_free;
5483
5484 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
5485 nr_objs, s->size, oo_objects(s->oo),
5486 (1 << oo_order(s->oo)));
5487 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
5488 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
5489 0UL);
5490 seq_putc(m, '\n');
5491 return 0;
5492}
5493
5494static const struct seq_operations slabinfo_op = {
5495 .start = s_start,
5496 .next = s_next,
5497 .stop = s_stop,
5498 .show = s_show,
5499};
5500
5501static int slabinfo_open(struct inode *inode, struct file *file)
5502{
5503 return seq_open(file, &slabinfo_op);
5504}
5505
5506static const struct file_operations proc_slabinfo_operations = {
5507 .open = slabinfo_open,
5508 .read = seq_read,
5509 .llseek = seq_lseek,
5510 .release = seq_release,
5511};
5512
5513static int __init slab_proc_init(void)
5514{
5515 proc_create("slabinfo", S_IRUSR, NULL, &proc_slabinfo_operations);
5516 return 0;
5517}
5518module_init(slab_proc_init);
5519#endif /* CONFIG_SLABINFO */