Merge branch 'upstream-linus' of git://git.kernel.org/pub/scm/linux/kernel/git/jgarzi...
[deliverable/linux.git] / mm / slub.c
CommitLineData
81819f0f
CL
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 and only
6 * uses a centralized lock to manage a pool of partial slabs.
7 *
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
9 */
10
11#include <linux/mm.h>
12#include <linux/module.h>
13#include <linux/bit_spinlock.h>
14#include <linux/interrupt.h>
15#include <linux/bitops.h>
16#include <linux/slab.h>
17#include <linux/seq_file.h>
18#include <linux/cpu.h>
19#include <linux/cpuset.h>
20#include <linux/mempolicy.h>
21#include <linux/ctype.h>
22#include <linux/kallsyms.h>
b9049e23 23#include <linux/memory.h>
81819f0f
CL
24
25/*
26 * Lock order:
27 * 1. slab_lock(page)
28 * 2. slab->list_lock
29 *
30 * The slab_lock protects operations on the object of a particular
31 * slab and its metadata in the page struct. If the slab lock
32 * has been taken then no allocations nor frees can be performed
33 * on the objects in the slab nor can the slab be added or removed
34 * from the partial or full lists since this would mean modifying
35 * the page_struct of the slab.
36 *
37 * The list_lock protects the partial and full list on each node and
38 * the partial slab counter. If taken then no new slabs may be added or
39 * removed from the lists nor make the number of partial slabs be modified.
40 * (Note that the total number of slabs is an atomic value that may be
41 * modified without taking the list lock).
42 *
43 * The list_lock is a centralized lock and thus we avoid taking it as
44 * much as possible. As long as SLUB does not have to handle partial
45 * slabs, operations can continue without any centralized lock. F.e.
46 * allocating a long series of objects that fill up slabs does not require
47 * the list lock.
48 *
49 * The lock order is sometimes inverted when we are trying to get a slab
50 * off a list. We take the list_lock and then look for a page on the list
51 * to use. While we do that objects in the slabs may be freed. We can
52 * only operate on the slab if we have also taken the slab_lock. So we use
53 * a slab_trylock() on the slab. If trylock was successful then no frees
54 * can occur anymore and we can use the slab for allocations etc. If the
55 * slab_trylock() does not succeed then frees are in progress in the slab and
56 * we must stay away from it for a while since we may cause a bouncing
57 * cacheline if we try to acquire the lock. So go onto the next slab.
58 * If all pages are busy then we may allocate a new slab instead of reusing
59 * a partial slab. A new slab has noone operating on it and thus there is
60 * no danger of cacheline contention.
61 *
62 * Interrupts are disabled during allocation and deallocation in order to
63 * make the slab allocator safe to use in the context of an irq. In addition
64 * interrupts are disabled to ensure that the processor does not change
65 * while handling per_cpu slabs, due to kernel preemption.
66 *
67 * SLUB assigns one slab for allocation to each processor.
68 * Allocations only occur from these slabs called cpu slabs.
69 *
672bba3a
CL
70 * Slabs with free elements are kept on a partial list and during regular
71 * operations no list for full slabs is used. If an object in a full slab is
81819f0f 72 * freed then the slab will show up again on the partial lists.
672bba3a
CL
73 * We track full slabs for debugging purposes though because otherwise we
74 * cannot scan all objects.
81819f0f
CL
75 *
76 * Slabs are freed when they become empty. Teardown and setup is
77 * minimal so we rely on the page allocators per cpu caches for
78 * fast frees and allocs.
79 *
80 * Overloading of page flags that are otherwise used for LRU management.
81 *
4b6f0750
CL
82 * PageActive The slab is frozen and exempt from list processing.
83 * This means that the slab is dedicated to a purpose
84 * such as satisfying allocations for a specific
85 * processor. Objects may be freed in the slab while
86 * it is frozen but slab_free will then skip the usual
87 * list operations. It is up to the processor holding
88 * the slab to integrate the slab into the slab lists
89 * when the slab is no longer needed.
90 *
91 * One use of this flag is to mark slabs that are
92 * used for allocations. Then such a slab becomes a cpu
93 * slab. The cpu slab may be equipped with an additional
dfb4f096 94 * freelist that allows lockless access to
894b8788
CL
95 * free objects in addition to the regular freelist
96 * that requires the slab lock.
81819f0f
CL
97 *
98 * PageError Slab requires special handling due to debug
99 * options set. This moves slab handling out of
894b8788 100 * the fast path and disables lockless freelists.
81819f0f
CL
101 */
102
5577bd8a
CL
103#define FROZEN (1 << PG_active)
104
105#ifdef CONFIG_SLUB_DEBUG
106#define SLABDEBUG (1 << PG_error)
107#else
108#define SLABDEBUG 0
109#endif
110
4b6f0750
CL
111static inline int SlabFrozen(struct page *page)
112{
5577bd8a 113 return page->flags & FROZEN;
4b6f0750
CL
114}
115
116static inline void SetSlabFrozen(struct page *page)
117{
5577bd8a 118 page->flags |= FROZEN;
4b6f0750
CL
119}
120
121static inline void ClearSlabFrozen(struct page *page)
122{
5577bd8a 123 page->flags &= ~FROZEN;
4b6f0750
CL
124}
125
35e5d7ee
CL
126static inline int SlabDebug(struct page *page)
127{
5577bd8a 128 return page->flags & SLABDEBUG;
35e5d7ee
CL
129}
130
131static inline void SetSlabDebug(struct page *page)
132{
5577bd8a 133 page->flags |= SLABDEBUG;
35e5d7ee
CL
134}
135
136static inline void ClearSlabDebug(struct page *page)
137{
5577bd8a 138 page->flags &= ~SLABDEBUG;
35e5d7ee
CL
139}
140
81819f0f
CL
141/*
142 * Issues still to be resolved:
143 *
81819f0f
CL
144 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
145 *
81819f0f
CL
146 * - Variable sizing of the per node arrays
147 */
148
149/* Enable to test recovery from slab corruption on boot */
150#undef SLUB_RESILIENCY_TEST
151
152#if PAGE_SHIFT <= 12
153
154/*
155 * Small page size. Make sure that we do not fragment memory
156 */
157#define DEFAULT_MAX_ORDER 1
158#define DEFAULT_MIN_OBJECTS 4
159
160#else
161
162/*
163 * Large page machines are customarily able to handle larger
164 * page orders.
165 */
166#define DEFAULT_MAX_ORDER 2
167#define DEFAULT_MIN_OBJECTS 8
168
169#endif
170
2086d26a
CL
171/*
172 * Mininum number of partial slabs. These will be left on the partial
173 * lists even if they are empty. kmem_cache_shrink may reclaim them.
174 */
76be8950 175#define MIN_PARTIAL 5
e95eed57 176
2086d26a
CL
177/*
178 * Maximum number of desirable partial slabs.
179 * The existence of more partial slabs makes kmem_cache_shrink
180 * sort the partial list by the number of objects in the.
181 */
182#define MAX_PARTIAL 10
183
81819f0f
CL
184#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
185 SLAB_POISON | SLAB_STORE_USER)
672bba3a 186
81819f0f
CL
187/*
188 * Set of flags that will prevent slab merging
189 */
190#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
191 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
192
193#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
194 SLAB_CACHE_DMA)
195
196#ifndef ARCH_KMALLOC_MINALIGN
47bfdc0d 197#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
81819f0f
CL
198#endif
199
200#ifndef ARCH_SLAB_MINALIGN
47bfdc0d 201#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
81819f0f
CL
202#endif
203
204/* Internal SLUB flags */
1ceef402
CL
205#define __OBJECT_POISON 0x80000000 /* Poison object */
206#define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
71c7a06f
CL
207#define __KMALLOC_CACHE 0x20000000 /* objects freed using kfree */
208#define __PAGE_ALLOC_FALLBACK 0x10000000 /* Allow fallback to page alloc */
81819f0f 209
65c02d4c
CL
210/* Not all arches define cache_line_size */
211#ifndef cache_line_size
212#define cache_line_size() L1_CACHE_BYTES
213#endif
214
81819f0f
CL
215static int kmem_size = sizeof(struct kmem_cache);
216
217#ifdef CONFIG_SMP
218static struct notifier_block slab_notifier;
219#endif
220
221static enum {
222 DOWN, /* No slab functionality available */
223 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
672bba3a 224 UP, /* Everything works but does not show up in sysfs */
81819f0f
CL
225 SYSFS /* Sysfs up */
226} slab_state = DOWN;
227
228/* A list of all slab caches on the system */
229static DECLARE_RWSEM(slub_lock);
5af328a5 230static LIST_HEAD(slab_caches);
81819f0f 231
02cbc874
CL
232/*
233 * Tracking user of a slab.
234 */
235struct track {
236 void *addr; /* Called from address */
237 int cpu; /* Was running on cpu */
238 int pid; /* Pid context */
239 unsigned long when; /* When did the operation occur */
240};
241
242enum track_item { TRACK_ALLOC, TRACK_FREE };
243
41ecc55b 244#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
81819f0f
CL
245static int sysfs_slab_add(struct kmem_cache *);
246static int sysfs_slab_alias(struct kmem_cache *, const char *);
247static void sysfs_slab_remove(struct kmem_cache *);
8ff12cfc 248
81819f0f 249#else
0c710013
CL
250static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
252 { return 0; }
151c602f
CL
253static inline void sysfs_slab_remove(struct kmem_cache *s)
254{
255 kfree(s);
256}
8ff12cfc 257
81819f0f
CL
258#endif
259
8ff12cfc
CL
260static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
261{
262#ifdef CONFIG_SLUB_STATS
263 c->stat[si]++;
264#endif
265}
266
81819f0f
CL
267/********************************************************************
268 * Core slab cache functions
269 *******************************************************************/
270
271int slab_is_available(void)
272{
273 return slab_state >= UP;
274}
275
276static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
277{
278#ifdef CONFIG_NUMA
279 return s->node[node];
280#else
281 return &s->local_node;
282#endif
283}
284
dfb4f096
CL
285static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
286{
4c93c355
CL
287#ifdef CONFIG_SMP
288 return s->cpu_slab[cpu];
289#else
290 return &s->cpu_slab;
291#endif
dfb4f096
CL
292}
293
6446faa2 294/* Verify that a pointer has an address that is valid within a slab page */
02cbc874
CL
295static inline int check_valid_pointer(struct kmem_cache *s,
296 struct page *page, const void *object)
297{
298 void *base;
299
a973e9dd 300 if (!object)
02cbc874
CL
301 return 1;
302
a973e9dd 303 base = page_address(page);
02cbc874
CL
304 if (object < base || object >= base + s->objects * s->size ||
305 (object - base) % s->size) {
306 return 0;
307 }
308
309 return 1;
310}
311
7656c72b
CL
312/*
313 * Slow version of get and set free pointer.
314 *
315 * This version requires touching the cache lines of kmem_cache which
316 * we avoid to do in the fast alloc free paths. There we obtain the offset
317 * from the page struct.
318 */
319static inline void *get_freepointer(struct kmem_cache *s, void *object)
320{
321 return *(void **)(object + s->offset);
322}
323
324static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
325{
326 *(void **)(object + s->offset) = fp;
327}
328
329/* Loop over all objects in a slab */
330#define for_each_object(__p, __s, __addr) \
331 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
332 __p += (__s)->size)
333
334/* Scan freelist */
335#define for_each_free_object(__p, __s, __free) \
a973e9dd 336 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
7656c72b
CL
337
338/* Determine object index from a given position */
339static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
340{
341 return (p - addr) / s->size;
342}
343
41ecc55b
CL
344#ifdef CONFIG_SLUB_DEBUG
345/*
346 * Debug settings:
347 */
f0630fff
CL
348#ifdef CONFIG_SLUB_DEBUG_ON
349static int slub_debug = DEBUG_DEFAULT_FLAGS;
350#else
41ecc55b 351static int slub_debug;
f0630fff 352#endif
41ecc55b
CL
353
354static char *slub_debug_slabs;
355
81819f0f
CL
356/*
357 * Object debugging
358 */
359static void print_section(char *text, u8 *addr, unsigned int length)
360{
361 int i, offset;
362 int newline = 1;
363 char ascii[17];
364
365 ascii[16] = 0;
366
367 for (i = 0; i < length; i++) {
368 if (newline) {
24922684 369 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
81819f0f
CL
370 newline = 0;
371 }
06428780 372 printk(KERN_CONT " %02x", addr[i]);
81819f0f
CL
373 offset = i % 16;
374 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
375 if (offset == 15) {
06428780 376 printk(KERN_CONT " %s\n", ascii);
81819f0f
CL
377 newline = 1;
378 }
379 }
380 if (!newline) {
381 i %= 16;
382 while (i < 16) {
06428780 383 printk(KERN_CONT " ");
81819f0f
CL
384 ascii[i] = ' ';
385 i++;
386 }
06428780 387 printk(KERN_CONT " %s\n", ascii);
81819f0f
CL
388 }
389}
390
81819f0f
CL
391static struct track *get_track(struct kmem_cache *s, void *object,
392 enum track_item alloc)
393{
394 struct track *p;
395
396 if (s->offset)
397 p = object + s->offset + sizeof(void *);
398 else
399 p = object + s->inuse;
400
401 return p + alloc;
402}
403
404static void set_track(struct kmem_cache *s, void *object,
405 enum track_item alloc, void *addr)
406{
407 struct track *p;
408
409 if (s->offset)
410 p = object + s->offset + sizeof(void *);
411 else
412 p = object + s->inuse;
413
414 p += alloc;
415 if (addr) {
416 p->addr = addr;
417 p->cpu = smp_processor_id();
418 p->pid = current ? current->pid : -1;
419 p->when = jiffies;
420 } else
421 memset(p, 0, sizeof(struct track));
422}
423
81819f0f
CL
424static void init_tracking(struct kmem_cache *s, void *object)
425{
24922684
CL
426 if (!(s->flags & SLAB_STORE_USER))
427 return;
428
429 set_track(s, object, TRACK_FREE, NULL);
430 set_track(s, object, TRACK_ALLOC, NULL);
81819f0f
CL
431}
432
433static void print_track(const char *s, struct track *t)
434{
435 if (!t->addr)
436 return;
437
24922684 438 printk(KERN_ERR "INFO: %s in ", s);
81819f0f 439 __print_symbol("%s", (unsigned long)t->addr);
24922684
CL
440 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
441}
442
443static void print_tracking(struct kmem_cache *s, void *object)
444{
445 if (!(s->flags & SLAB_STORE_USER))
446 return;
447
448 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
449 print_track("Freed", get_track(s, object, TRACK_FREE));
450}
451
452static void print_page_info(struct page *page)
453{
454 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
455 page, page->inuse, page->freelist, page->flags);
456
457}
458
459static void slab_bug(struct kmem_cache *s, char *fmt, ...)
460{
461 va_list args;
462 char buf[100];
463
464 va_start(args, fmt);
465 vsnprintf(buf, sizeof(buf), fmt, args);
466 va_end(args);
467 printk(KERN_ERR "========================================"
468 "=====================================\n");
469 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
470 printk(KERN_ERR "----------------------------------------"
471 "-------------------------------------\n\n");
81819f0f
CL
472}
473
24922684
CL
474static void slab_fix(struct kmem_cache *s, char *fmt, ...)
475{
476 va_list args;
477 char buf[100];
478
479 va_start(args, fmt);
480 vsnprintf(buf, sizeof(buf), fmt, args);
481 va_end(args);
482 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
483}
484
485static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
81819f0f
CL
486{
487 unsigned int off; /* Offset of last byte */
a973e9dd 488 u8 *addr = page_address(page);
24922684
CL
489
490 print_tracking(s, p);
491
492 print_page_info(page);
493
494 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
495 p, p - addr, get_freepointer(s, p));
496
497 if (p > addr + 16)
498 print_section("Bytes b4", p - 16, 16);
499
500 print_section("Object", p, min(s->objsize, 128));
81819f0f
CL
501
502 if (s->flags & SLAB_RED_ZONE)
503 print_section("Redzone", p + s->objsize,
504 s->inuse - s->objsize);
505
81819f0f
CL
506 if (s->offset)
507 off = s->offset + sizeof(void *);
508 else
509 off = s->inuse;
510
24922684 511 if (s->flags & SLAB_STORE_USER)
81819f0f 512 off += 2 * sizeof(struct track);
81819f0f
CL
513
514 if (off != s->size)
515 /* Beginning of the filler is the free pointer */
24922684
CL
516 print_section("Padding", p + off, s->size - off);
517
518 dump_stack();
81819f0f
CL
519}
520
521static void object_err(struct kmem_cache *s, struct page *page,
522 u8 *object, char *reason)
523{
24922684
CL
524 slab_bug(s, reason);
525 print_trailer(s, page, object);
81819f0f
CL
526}
527
24922684 528static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
81819f0f
CL
529{
530 va_list args;
531 char buf[100];
532
24922684
CL
533 va_start(args, fmt);
534 vsnprintf(buf, sizeof(buf), fmt, args);
81819f0f 535 va_end(args);
24922684
CL
536 slab_bug(s, fmt);
537 print_page_info(page);
81819f0f
CL
538 dump_stack();
539}
540
541static void init_object(struct kmem_cache *s, void *object, int active)
542{
543 u8 *p = object;
544
545 if (s->flags & __OBJECT_POISON) {
546 memset(p, POISON_FREE, s->objsize - 1);
06428780 547 p[s->objsize - 1] = POISON_END;
81819f0f
CL
548 }
549
550 if (s->flags & SLAB_RED_ZONE)
551 memset(p + s->objsize,
552 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
553 s->inuse - s->objsize);
554}
555
24922684 556static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
81819f0f
CL
557{
558 while (bytes) {
559 if (*start != (u8)value)
24922684 560 return start;
81819f0f
CL
561 start++;
562 bytes--;
563 }
24922684
CL
564 return NULL;
565}
566
567static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
568 void *from, void *to)
569{
570 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
571 memset(from, data, to - from);
572}
573
574static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
575 u8 *object, char *what,
06428780 576 u8 *start, unsigned int value, unsigned int bytes)
24922684
CL
577{
578 u8 *fault;
579 u8 *end;
580
581 fault = check_bytes(start, value, bytes);
582 if (!fault)
583 return 1;
584
585 end = start + bytes;
586 while (end > fault && end[-1] == value)
587 end--;
588
589 slab_bug(s, "%s overwritten", what);
590 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
591 fault, end - 1, fault[0], value);
592 print_trailer(s, page, object);
593
594 restore_bytes(s, what, value, fault, end);
595 return 0;
81819f0f
CL
596}
597
81819f0f
CL
598/*
599 * Object layout:
600 *
601 * object address
602 * Bytes of the object to be managed.
603 * If the freepointer may overlay the object then the free
604 * pointer is the first word of the object.
672bba3a 605 *
81819f0f
CL
606 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
607 * 0xa5 (POISON_END)
608 *
609 * object + s->objsize
610 * Padding to reach word boundary. This is also used for Redzoning.
672bba3a
CL
611 * Padding is extended by another word if Redzoning is enabled and
612 * objsize == inuse.
613 *
81819f0f
CL
614 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
615 * 0xcc (RED_ACTIVE) for objects in use.
616 *
617 * object + s->inuse
672bba3a
CL
618 * Meta data starts here.
619 *
81819f0f
CL
620 * A. Free pointer (if we cannot overwrite object on free)
621 * B. Tracking data for SLAB_STORE_USER
672bba3a 622 * C. Padding to reach required alignment boundary or at mininum
6446faa2 623 * one word if debugging is on to be able to detect writes
672bba3a
CL
624 * before the word boundary.
625 *
626 * Padding is done using 0x5a (POISON_INUSE)
81819f0f
CL
627 *
628 * object + s->size
672bba3a 629 * Nothing is used beyond s->size.
81819f0f 630 *
672bba3a
CL
631 * If slabcaches are merged then the objsize and inuse boundaries are mostly
632 * ignored. And therefore no slab options that rely on these boundaries
81819f0f
CL
633 * may be used with merged slabcaches.
634 */
635
81819f0f
CL
636static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
637{
638 unsigned long off = s->inuse; /* The end of info */
639
640 if (s->offset)
641 /* Freepointer is placed after the object. */
642 off += sizeof(void *);
643
644 if (s->flags & SLAB_STORE_USER)
645 /* We also have user information there */
646 off += 2 * sizeof(struct track);
647
648 if (s->size == off)
649 return 1;
650
24922684
CL
651 return check_bytes_and_report(s, page, p, "Object padding",
652 p + off, POISON_INUSE, s->size - off);
81819f0f
CL
653}
654
655static int slab_pad_check(struct kmem_cache *s, struct page *page)
656{
24922684
CL
657 u8 *start;
658 u8 *fault;
659 u8 *end;
660 int length;
661 int remainder;
81819f0f
CL
662
663 if (!(s->flags & SLAB_POISON))
664 return 1;
665
a973e9dd 666 start = page_address(page);
24922684 667 end = start + (PAGE_SIZE << s->order);
81819f0f 668 length = s->objects * s->size;
24922684 669 remainder = end - (start + length);
81819f0f
CL
670 if (!remainder)
671 return 1;
672
24922684
CL
673 fault = check_bytes(start + length, POISON_INUSE, remainder);
674 if (!fault)
675 return 1;
676 while (end > fault && end[-1] == POISON_INUSE)
677 end--;
678
679 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
680 print_section("Padding", start, length);
681
682 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
683 return 0;
81819f0f
CL
684}
685
686static int check_object(struct kmem_cache *s, struct page *page,
687 void *object, int active)
688{
689 u8 *p = object;
690 u8 *endobject = object + s->objsize;
691
692 if (s->flags & SLAB_RED_ZONE) {
693 unsigned int red =
694 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
695
24922684
CL
696 if (!check_bytes_and_report(s, page, object, "Redzone",
697 endobject, red, s->inuse - s->objsize))
81819f0f 698 return 0;
81819f0f 699 } else {
3adbefee
IM
700 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
701 check_bytes_and_report(s, page, p, "Alignment padding",
702 endobject, POISON_INUSE, s->inuse - s->objsize);
703 }
81819f0f
CL
704 }
705
706 if (s->flags & SLAB_POISON) {
707 if (!active && (s->flags & __OBJECT_POISON) &&
24922684
CL
708 (!check_bytes_and_report(s, page, p, "Poison", p,
709 POISON_FREE, s->objsize - 1) ||
710 !check_bytes_and_report(s, page, p, "Poison",
06428780 711 p + s->objsize - 1, POISON_END, 1)))
81819f0f 712 return 0;
81819f0f
CL
713 /*
714 * check_pad_bytes cleans up on its own.
715 */
716 check_pad_bytes(s, page, p);
717 }
718
719 if (!s->offset && active)
720 /*
721 * Object and freepointer overlap. Cannot check
722 * freepointer while object is allocated.
723 */
724 return 1;
725
726 /* Check free pointer validity */
727 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
728 object_err(s, page, p, "Freepointer corrupt");
729 /*
730 * No choice but to zap it and thus loose the remainder
731 * of the free objects in this slab. May cause
672bba3a 732 * another error because the object count is now wrong.
81819f0f 733 */
a973e9dd 734 set_freepointer(s, p, NULL);
81819f0f
CL
735 return 0;
736 }
737 return 1;
738}
739
740static int check_slab(struct kmem_cache *s, struct page *page)
741{
742 VM_BUG_ON(!irqs_disabled());
743
744 if (!PageSlab(page)) {
24922684 745 slab_err(s, page, "Not a valid slab page");
81819f0f
CL
746 return 0;
747 }
81819f0f 748 if (page->inuse > s->objects) {
24922684
CL
749 slab_err(s, page, "inuse %u > max %u",
750 s->name, page->inuse, s->objects);
81819f0f
CL
751 return 0;
752 }
753 /* Slab_pad_check fixes things up after itself */
754 slab_pad_check(s, page);
755 return 1;
756}
757
758/*
672bba3a
CL
759 * Determine if a certain object on a page is on the freelist. Must hold the
760 * slab lock to guarantee that the chains are in a consistent state.
81819f0f
CL
761 */
762static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
763{
764 int nr = 0;
765 void *fp = page->freelist;
766 void *object = NULL;
767
a973e9dd 768 while (fp && nr <= s->objects) {
81819f0f
CL
769 if (fp == search)
770 return 1;
771 if (!check_valid_pointer(s, page, fp)) {
772 if (object) {
773 object_err(s, page, object,
774 "Freechain corrupt");
a973e9dd 775 set_freepointer(s, object, NULL);
81819f0f
CL
776 break;
777 } else {
24922684 778 slab_err(s, page, "Freepointer corrupt");
a973e9dd 779 page->freelist = NULL;
81819f0f 780 page->inuse = s->objects;
24922684 781 slab_fix(s, "Freelist cleared");
81819f0f
CL
782 return 0;
783 }
784 break;
785 }
786 object = fp;
787 fp = get_freepointer(s, object);
788 nr++;
789 }
790
791 if (page->inuse != s->objects - nr) {
70d71228 792 slab_err(s, page, "Wrong object count. Counter is %d but "
24922684 793 "counted were %d", page->inuse, s->objects - nr);
81819f0f 794 page->inuse = s->objects - nr;
24922684 795 slab_fix(s, "Object count adjusted.");
81819f0f
CL
796 }
797 return search == NULL;
798}
799
3ec09742
CL
800static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
801{
802 if (s->flags & SLAB_TRACE) {
803 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
804 s->name,
805 alloc ? "alloc" : "free",
806 object, page->inuse,
807 page->freelist);
808
809 if (!alloc)
810 print_section("Object", (void *)object, s->objsize);
811
812 dump_stack();
813 }
814}
815
643b1138 816/*
672bba3a 817 * Tracking of fully allocated slabs for debugging purposes.
643b1138 818 */
e95eed57 819static void add_full(struct kmem_cache_node *n, struct page *page)
643b1138 820{
643b1138
CL
821 spin_lock(&n->list_lock);
822 list_add(&page->lru, &n->full);
823 spin_unlock(&n->list_lock);
824}
825
826static void remove_full(struct kmem_cache *s, struct page *page)
827{
828 struct kmem_cache_node *n;
829
830 if (!(s->flags & SLAB_STORE_USER))
831 return;
832
833 n = get_node(s, page_to_nid(page));
834
835 spin_lock(&n->list_lock);
836 list_del(&page->lru);
837 spin_unlock(&n->list_lock);
838}
839
3ec09742
CL
840static void setup_object_debug(struct kmem_cache *s, struct page *page,
841 void *object)
842{
843 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
844 return;
845
846 init_object(s, object, 0);
847 init_tracking(s, object);
848}
849
850static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
851 void *object, void *addr)
81819f0f
CL
852{
853 if (!check_slab(s, page))
854 goto bad;
855
d692ef6d 856 if (!on_freelist(s, page, object)) {
24922684 857 object_err(s, page, object, "Object already allocated");
70d71228 858 goto bad;
81819f0f
CL
859 }
860
861 if (!check_valid_pointer(s, page, object)) {
862 object_err(s, page, object, "Freelist Pointer check fails");
70d71228 863 goto bad;
81819f0f
CL
864 }
865
d692ef6d 866 if (!check_object(s, page, object, 0))
81819f0f 867 goto bad;
81819f0f 868
3ec09742
CL
869 /* Success perform special debug activities for allocs */
870 if (s->flags & SLAB_STORE_USER)
871 set_track(s, object, TRACK_ALLOC, addr);
872 trace(s, page, object, 1);
873 init_object(s, object, 1);
81819f0f 874 return 1;
3ec09742 875
81819f0f
CL
876bad:
877 if (PageSlab(page)) {
878 /*
879 * If this is a slab page then lets do the best we can
880 * to avoid issues in the future. Marking all objects
672bba3a 881 * as used avoids touching the remaining objects.
81819f0f 882 */
24922684 883 slab_fix(s, "Marking all objects used");
81819f0f 884 page->inuse = s->objects;
a973e9dd 885 page->freelist = NULL;
81819f0f
CL
886 }
887 return 0;
888}
889
3ec09742
CL
890static int free_debug_processing(struct kmem_cache *s, struct page *page,
891 void *object, void *addr)
81819f0f
CL
892{
893 if (!check_slab(s, page))
894 goto fail;
895
896 if (!check_valid_pointer(s, page, object)) {
70d71228 897 slab_err(s, page, "Invalid object pointer 0x%p", object);
81819f0f
CL
898 goto fail;
899 }
900
901 if (on_freelist(s, page, object)) {
24922684 902 object_err(s, page, object, "Object already free");
81819f0f
CL
903 goto fail;
904 }
905
906 if (!check_object(s, page, object, 1))
907 return 0;
908
909 if (unlikely(s != page->slab)) {
3adbefee 910 if (!PageSlab(page)) {
70d71228
CL
911 slab_err(s, page, "Attempt to free object(0x%p) "
912 "outside of slab", object);
3adbefee 913 } else if (!page->slab) {
81819f0f 914 printk(KERN_ERR
70d71228 915 "SLUB <none>: no slab for object 0x%p.\n",
81819f0f 916 object);
70d71228 917 dump_stack();
06428780 918 } else
24922684
CL
919 object_err(s, page, object,
920 "page slab pointer corrupt.");
81819f0f
CL
921 goto fail;
922 }
3ec09742
CL
923
924 /* Special debug activities for freeing objects */
a973e9dd 925 if (!SlabFrozen(page) && !page->freelist)
3ec09742
CL
926 remove_full(s, page);
927 if (s->flags & SLAB_STORE_USER)
928 set_track(s, object, TRACK_FREE, addr);
929 trace(s, page, object, 0);
930 init_object(s, object, 0);
81819f0f 931 return 1;
3ec09742 932
81819f0f 933fail:
24922684 934 slab_fix(s, "Object at 0x%p not freed", object);
81819f0f
CL
935 return 0;
936}
937
41ecc55b
CL
938static int __init setup_slub_debug(char *str)
939{
f0630fff
CL
940 slub_debug = DEBUG_DEFAULT_FLAGS;
941 if (*str++ != '=' || !*str)
942 /*
943 * No options specified. Switch on full debugging.
944 */
945 goto out;
946
947 if (*str == ',')
948 /*
949 * No options but restriction on slabs. This means full
950 * debugging for slabs matching a pattern.
951 */
952 goto check_slabs;
953
954 slub_debug = 0;
955 if (*str == '-')
956 /*
957 * Switch off all debugging measures.
958 */
959 goto out;
960
961 /*
962 * Determine which debug features should be switched on
963 */
06428780 964 for (; *str && *str != ','; str++) {
f0630fff
CL
965 switch (tolower(*str)) {
966 case 'f':
967 slub_debug |= SLAB_DEBUG_FREE;
968 break;
969 case 'z':
970 slub_debug |= SLAB_RED_ZONE;
971 break;
972 case 'p':
973 slub_debug |= SLAB_POISON;
974 break;
975 case 'u':
976 slub_debug |= SLAB_STORE_USER;
977 break;
978 case 't':
979 slub_debug |= SLAB_TRACE;
980 break;
981 default:
982 printk(KERN_ERR "slub_debug option '%c' "
06428780 983 "unknown. skipped\n", *str);
f0630fff 984 }
41ecc55b
CL
985 }
986
f0630fff 987check_slabs:
41ecc55b
CL
988 if (*str == ',')
989 slub_debug_slabs = str + 1;
f0630fff 990out:
41ecc55b
CL
991 return 1;
992}
993
994__setup("slub_debug", setup_slub_debug);
995
ba0268a8
CL
996static unsigned long kmem_cache_flags(unsigned long objsize,
997 unsigned long flags, const char *name,
4ba9b9d0 998 void (*ctor)(struct kmem_cache *, void *))
41ecc55b
CL
999{
1000 /*
e153362a 1001 * Enable debugging if selected on the kernel commandline.
41ecc55b 1002 */
e153362a
CL
1003 if (slub_debug && (!slub_debug_slabs ||
1004 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1005 flags |= slub_debug;
ba0268a8
CL
1006
1007 return flags;
41ecc55b
CL
1008}
1009#else
3ec09742
CL
1010static inline void setup_object_debug(struct kmem_cache *s,
1011 struct page *page, void *object) {}
41ecc55b 1012
3ec09742
CL
1013static inline int alloc_debug_processing(struct kmem_cache *s,
1014 struct page *page, void *object, void *addr) { return 0; }
41ecc55b 1015
3ec09742
CL
1016static inline int free_debug_processing(struct kmem_cache *s,
1017 struct page *page, void *object, void *addr) { return 0; }
41ecc55b 1018
41ecc55b
CL
1019static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1020 { return 1; }
1021static inline int check_object(struct kmem_cache *s, struct page *page,
1022 void *object, int active) { return 1; }
3ec09742 1023static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
ba0268a8
CL
1024static inline unsigned long kmem_cache_flags(unsigned long objsize,
1025 unsigned long flags, const char *name,
4ba9b9d0 1026 void (*ctor)(struct kmem_cache *, void *))
ba0268a8
CL
1027{
1028 return flags;
1029}
41ecc55b
CL
1030#define slub_debug 0
1031#endif
81819f0f
CL
1032/*
1033 * Slab allocation and freeing
1034 */
1035static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1036{
06428780 1037 struct page *page;
81819f0f
CL
1038 int pages = 1 << s->order;
1039
b7a49f0d 1040 flags |= s->allocflags;
e12ba74d 1041
81819f0f
CL
1042 if (node == -1)
1043 page = alloc_pages(flags, s->order);
1044 else
1045 page = alloc_pages_node(node, flags, s->order);
1046
1047 if (!page)
1048 return NULL;
1049
1050 mod_zone_page_state(page_zone(page),
1051 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1052 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1053 pages);
1054
1055 return page;
1056}
1057
1058static void setup_object(struct kmem_cache *s, struct page *page,
1059 void *object)
1060{
3ec09742 1061 setup_object_debug(s, page, object);
4f104934 1062 if (unlikely(s->ctor))
4ba9b9d0 1063 s->ctor(s, object);
81819f0f
CL
1064}
1065
1066static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1067{
1068 struct page *page;
1069 struct kmem_cache_node *n;
1070 void *start;
81819f0f
CL
1071 void *last;
1072 void *p;
1073
6cb06229 1074 BUG_ON(flags & GFP_SLAB_BUG_MASK);
81819f0f 1075
6cb06229
CL
1076 page = allocate_slab(s,
1077 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
81819f0f
CL
1078 if (!page)
1079 goto out;
1080
1081 n = get_node(s, page_to_nid(page));
1082 if (n)
1083 atomic_long_inc(&n->nr_slabs);
81819f0f
CL
1084 page->slab = s;
1085 page->flags |= 1 << PG_slab;
1086 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1087 SLAB_STORE_USER | SLAB_TRACE))
35e5d7ee 1088 SetSlabDebug(page);
81819f0f
CL
1089
1090 start = page_address(page);
81819f0f
CL
1091
1092 if (unlikely(s->flags & SLAB_POISON))
1093 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1094
1095 last = start;
7656c72b 1096 for_each_object(p, s, start) {
81819f0f
CL
1097 setup_object(s, page, last);
1098 set_freepointer(s, last, p);
1099 last = p;
1100 }
1101 setup_object(s, page, last);
a973e9dd 1102 set_freepointer(s, last, NULL);
81819f0f
CL
1103
1104 page->freelist = start;
1105 page->inuse = 0;
1106out:
81819f0f
CL
1107 return page;
1108}
1109
1110static void __free_slab(struct kmem_cache *s, struct page *page)
1111{
1112 int pages = 1 << s->order;
1113
c59def9f 1114 if (unlikely(SlabDebug(page))) {
81819f0f
CL
1115 void *p;
1116
1117 slab_pad_check(s, page);
a973e9dd 1118 for_each_object(p, s, page_address(page))
81819f0f 1119 check_object(s, page, p, 0);
2208b764 1120 ClearSlabDebug(page);
81819f0f
CL
1121 }
1122
1123 mod_zone_page_state(page_zone(page),
1124 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1125 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
06428780 1126 -pages);
81819f0f 1127
81819f0f
CL
1128 __free_pages(page, s->order);
1129}
1130
1131static void rcu_free_slab(struct rcu_head *h)
1132{
1133 struct page *page;
1134
1135 page = container_of((struct list_head *)h, struct page, lru);
1136 __free_slab(page->slab, page);
1137}
1138
1139static void free_slab(struct kmem_cache *s, struct page *page)
1140{
1141 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1142 /*
1143 * RCU free overloads the RCU head over the LRU
1144 */
1145 struct rcu_head *head = (void *)&page->lru;
1146
1147 call_rcu(head, rcu_free_slab);
1148 } else
1149 __free_slab(s, page);
1150}
1151
1152static void discard_slab(struct kmem_cache *s, struct page *page)
1153{
1154 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1155
1156 atomic_long_dec(&n->nr_slabs);
1157 reset_page_mapcount(page);
35e5d7ee 1158 __ClearPageSlab(page);
81819f0f
CL
1159 free_slab(s, page);
1160}
1161
1162/*
1163 * Per slab locking using the pagelock
1164 */
1165static __always_inline void slab_lock(struct page *page)
1166{
1167 bit_spin_lock(PG_locked, &page->flags);
1168}
1169
1170static __always_inline void slab_unlock(struct page *page)
1171{
a76d3546 1172 __bit_spin_unlock(PG_locked, &page->flags);
81819f0f
CL
1173}
1174
1175static __always_inline int slab_trylock(struct page *page)
1176{
1177 int rc = 1;
1178
1179 rc = bit_spin_trylock(PG_locked, &page->flags);
1180 return rc;
1181}
1182
1183/*
1184 * Management of partially allocated slabs
1185 */
7c2e132c
CL
1186static void add_partial(struct kmem_cache_node *n,
1187 struct page *page, int tail)
81819f0f 1188{
e95eed57
CL
1189 spin_lock(&n->list_lock);
1190 n->nr_partial++;
7c2e132c
CL
1191 if (tail)
1192 list_add_tail(&page->lru, &n->partial);
1193 else
1194 list_add(&page->lru, &n->partial);
81819f0f
CL
1195 spin_unlock(&n->list_lock);
1196}
1197
1198static void remove_partial(struct kmem_cache *s,
1199 struct page *page)
1200{
1201 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1202
1203 spin_lock(&n->list_lock);
1204 list_del(&page->lru);
1205 n->nr_partial--;
1206 spin_unlock(&n->list_lock);
1207}
1208
1209/*
672bba3a 1210 * Lock slab and remove from the partial list.
81819f0f 1211 *
672bba3a 1212 * Must hold list_lock.
81819f0f 1213 */
4b6f0750 1214static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
81819f0f
CL
1215{
1216 if (slab_trylock(page)) {
1217 list_del(&page->lru);
1218 n->nr_partial--;
4b6f0750 1219 SetSlabFrozen(page);
81819f0f
CL
1220 return 1;
1221 }
1222 return 0;
1223}
1224
1225/*
672bba3a 1226 * Try to allocate a partial slab from a specific node.
81819f0f
CL
1227 */
1228static struct page *get_partial_node(struct kmem_cache_node *n)
1229{
1230 struct page *page;
1231
1232 /*
1233 * Racy check. If we mistakenly see no partial slabs then we
1234 * just allocate an empty slab. If we mistakenly try to get a
672bba3a
CL
1235 * partial slab and there is none available then get_partials()
1236 * will return NULL.
81819f0f
CL
1237 */
1238 if (!n || !n->nr_partial)
1239 return NULL;
1240
1241 spin_lock(&n->list_lock);
1242 list_for_each_entry(page, &n->partial, lru)
4b6f0750 1243 if (lock_and_freeze_slab(n, page))
81819f0f
CL
1244 goto out;
1245 page = NULL;
1246out:
1247 spin_unlock(&n->list_lock);
1248 return page;
1249}
1250
1251/*
672bba3a 1252 * Get a page from somewhere. Search in increasing NUMA distances.
81819f0f
CL
1253 */
1254static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1255{
1256#ifdef CONFIG_NUMA
1257 struct zonelist *zonelist;
1258 struct zone **z;
1259 struct page *page;
1260
1261 /*
672bba3a
CL
1262 * The defrag ratio allows a configuration of the tradeoffs between
1263 * inter node defragmentation and node local allocations. A lower
1264 * defrag_ratio increases the tendency to do local allocations
1265 * instead of attempting to obtain partial slabs from other nodes.
81819f0f 1266 *
672bba3a
CL
1267 * If the defrag_ratio is set to 0 then kmalloc() always
1268 * returns node local objects. If the ratio is higher then kmalloc()
1269 * may return off node objects because partial slabs are obtained
1270 * from other nodes and filled up.
81819f0f 1271 *
6446faa2 1272 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
672bba3a
CL
1273 * defrag_ratio = 1000) then every (well almost) allocation will
1274 * first attempt to defrag slab caches on other nodes. This means
1275 * scanning over all nodes to look for partial slabs which may be
1276 * expensive if we do it every time we are trying to find a slab
1277 * with available objects.
81819f0f 1278 */
9824601e
CL
1279 if (!s->remote_node_defrag_ratio ||
1280 get_cycles() % 1024 > s->remote_node_defrag_ratio)
81819f0f
CL
1281 return NULL;
1282
3adbefee
IM
1283 zonelist = &NODE_DATA(
1284 slab_node(current->mempolicy))->node_zonelists[gfp_zone(flags)];
81819f0f
CL
1285 for (z = zonelist->zones; *z; z++) {
1286 struct kmem_cache_node *n;
1287
1288 n = get_node(s, zone_to_nid(*z));
1289
1290 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
e95eed57 1291 n->nr_partial > MIN_PARTIAL) {
81819f0f
CL
1292 page = get_partial_node(n);
1293 if (page)
1294 return page;
1295 }
1296 }
1297#endif
1298 return NULL;
1299}
1300
1301/*
1302 * Get a partial page, lock it and return it.
1303 */
1304static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1305{
1306 struct page *page;
1307 int searchnode = (node == -1) ? numa_node_id() : node;
1308
1309 page = get_partial_node(get_node(s, searchnode));
1310 if (page || (flags & __GFP_THISNODE))
1311 return page;
1312
1313 return get_any_partial(s, flags);
1314}
1315
1316/*
1317 * Move a page back to the lists.
1318 *
1319 * Must be called with the slab lock held.
1320 *
1321 * On exit the slab lock will have been dropped.
1322 */
7c2e132c 1323static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
81819f0f 1324{
e95eed57 1325 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
8ff12cfc 1326 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
e95eed57 1327
4b6f0750 1328 ClearSlabFrozen(page);
81819f0f 1329 if (page->inuse) {
e95eed57 1330
a973e9dd 1331 if (page->freelist) {
7c2e132c 1332 add_partial(n, page, tail);
8ff12cfc
CL
1333 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1334 } else {
1335 stat(c, DEACTIVATE_FULL);
1336 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1337 add_full(n, page);
1338 }
81819f0f
CL
1339 slab_unlock(page);
1340 } else {
8ff12cfc 1341 stat(c, DEACTIVATE_EMPTY);
e95eed57
CL
1342 if (n->nr_partial < MIN_PARTIAL) {
1343 /*
672bba3a
CL
1344 * Adding an empty slab to the partial slabs in order
1345 * to avoid page allocator overhead. This slab needs
1346 * to come after the other slabs with objects in
6446faa2
CL
1347 * so that the others get filled first. That way the
1348 * size of the partial list stays small.
1349 *
1350 * kmem_cache_shrink can reclaim any empty slabs from the
1351 * partial list.
e95eed57 1352 */
7c2e132c 1353 add_partial(n, page, 1);
e95eed57
CL
1354 slab_unlock(page);
1355 } else {
1356 slab_unlock(page);
8ff12cfc 1357 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
e95eed57
CL
1358 discard_slab(s, page);
1359 }
81819f0f
CL
1360 }
1361}
1362
1363/*
1364 * Remove the cpu slab
1365 */
dfb4f096 1366static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
81819f0f 1367{
dfb4f096 1368 struct page *page = c->page;
7c2e132c 1369 int tail = 1;
8ff12cfc 1370
b773ad73 1371 if (page->freelist)
8ff12cfc 1372 stat(c, DEACTIVATE_REMOTE_FREES);
894b8788 1373 /*
6446faa2 1374 * Merge cpu freelist into slab freelist. Typically we get here
894b8788
CL
1375 * because both freelists are empty. So this is unlikely
1376 * to occur.
1377 */
a973e9dd 1378 while (unlikely(c->freelist)) {
894b8788
CL
1379 void **object;
1380
7c2e132c
CL
1381 tail = 0; /* Hot objects. Put the slab first */
1382
894b8788 1383 /* Retrieve object from cpu_freelist */
dfb4f096 1384 object = c->freelist;
b3fba8da 1385 c->freelist = c->freelist[c->offset];
894b8788
CL
1386
1387 /* And put onto the regular freelist */
b3fba8da 1388 object[c->offset] = page->freelist;
894b8788
CL
1389 page->freelist = object;
1390 page->inuse--;
1391 }
dfb4f096 1392 c->page = NULL;
7c2e132c 1393 unfreeze_slab(s, page, tail);
81819f0f
CL
1394}
1395
dfb4f096 1396static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
81819f0f 1397{
8ff12cfc 1398 stat(c, CPUSLAB_FLUSH);
dfb4f096
CL
1399 slab_lock(c->page);
1400 deactivate_slab(s, c);
81819f0f
CL
1401}
1402
1403/*
1404 * Flush cpu slab.
6446faa2 1405 *
81819f0f
CL
1406 * Called from IPI handler with interrupts disabled.
1407 */
0c710013 1408static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
81819f0f 1409{
dfb4f096 1410 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
81819f0f 1411
dfb4f096
CL
1412 if (likely(c && c->page))
1413 flush_slab(s, c);
81819f0f
CL
1414}
1415
1416static void flush_cpu_slab(void *d)
1417{
1418 struct kmem_cache *s = d;
81819f0f 1419
dfb4f096 1420 __flush_cpu_slab(s, smp_processor_id());
81819f0f
CL
1421}
1422
1423static void flush_all(struct kmem_cache *s)
1424{
1425#ifdef CONFIG_SMP
1426 on_each_cpu(flush_cpu_slab, s, 1, 1);
1427#else
1428 unsigned long flags;
1429
1430 local_irq_save(flags);
1431 flush_cpu_slab(s);
1432 local_irq_restore(flags);
1433#endif
1434}
1435
dfb4f096
CL
1436/*
1437 * Check if the objects in a per cpu structure fit numa
1438 * locality expectations.
1439 */
1440static inline int node_match(struct kmem_cache_cpu *c, int node)
1441{
1442#ifdef CONFIG_NUMA
1443 if (node != -1 && c->node != node)
1444 return 0;
1445#endif
1446 return 1;
1447}
1448
81819f0f 1449/*
894b8788
CL
1450 * Slow path. The lockless freelist is empty or we need to perform
1451 * debugging duties.
1452 *
1453 * Interrupts are disabled.
81819f0f 1454 *
894b8788
CL
1455 * Processing is still very fast if new objects have been freed to the
1456 * regular freelist. In that case we simply take over the regular freelist
1457 * as the lockless freelist and zap the regular freelist.
81819f0f 1458 *
894b8788
CL
1459 * If that is not working then we fall back to the partial lists. We take the
1460 * first element of the freelist as the object to allocate now and move the
1461 * rest of the freelist to the lockless freelist.
81819f0f 1462 *
894b8788 1463 * And if we were unable to get a new slab from the partial slab lists then
6446faa2
CL
1464 * we need to allocate a new slab. This is the slowest path since it involves
1465 * a call to the page allocator and the setup of a new slab.
81819f0f 1466 */
894b8788 1467static void *__slab_alloc(struct kmem_cache *s,
dfb4f096 1468 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
81819f0f 1469{
81819f0f 1470 void **object;
dfb4f096 1471 struct page *new;
81819f0f 1472
dfb4f096 1473 if (!c->page)
81819f0f
CL
1474 goto new_slab;
1475
dfb4f096
CL
1476 slab_lock(c->page);
1477 if (unlikely(!node_match(c, node)))
81819f0f 1478 goto another_slab;
6446faa2 1479
8ff12cfc 1480 stat(c, ALLOC_REFILL);
6446faa2 1481
894b8788 1482load_freelist:
dfb4f096 1483 object = c->page->freelist;
a973e9dd 1484 if (unlikely(!object))
81819f0f 1485 goto another_slab;
dfb4f096 1486 if (unlikely(SlabDebug(c->page)))
81819f0f
CL
1487 goto debug;
1488
b3fba8da 1489 c->freelist = object[c->offset];
dfb4f096 1490 c->page->inuse = s->objects;
a973e9dd 1491 c->page->freelist = NULL;
dfb4f096 1492 c->node = page_to_nid(c->page);
1f84260c 1493unlock_out:
dfb4f096 1494 slab_unlock(c->page);
8ff12cfc 1495 stat(c, ALLOC_SLOWPATH);
81819f0f
CL
1496 return object;
1497
1498another_slab:
dfb4f096 1499 deactivate_slab(s, c);
81819f0f
CL
1500
1501new_slab:
dfb4f096
CL
1502 new = get_partial(s, gfpflags, node);
1503 if (new) {
1504 c->page = new;
8ff12cfc 1505 stat(c, ALLOC_FROM_PARTIAL);
894b8788 1506 goto load_freelist;
81819f0f
CL
1507 }
1508
b811c202
CL
1509 if (gfpflags & __GFP_WAIT)
1510 local_irq_enable();
1511
dfb4f096 1512 new = new_slab(s, gfpflags, node);
b811c202
CL
1513
1514 if (gfpflags & __GFP_WAIT)
1515 local_irq_disable();
1516
dfb4f096
CL
1517 if (new) {
1518 c = get_cpu_slab(s, smp_processor_id());
8ff12cfc 1519 stat(c, ALLOC_SLAB);
05aa3450 1520 if (c->page)
dfb4f096 1521 flush_slab(s, c);
dfb4f096
CL
1522 slab_lock(new);
1523 SetSlabFrozen(new);
1524 c->page = new;
4b6f0750 1525 goto load_freelist;
81819f0f 1526 }
00e962c5 1527
71c7a06f
CL
1528 /*
1529 * No memory available.
1530 *
1531 * If the slab uses higher order allocs but the object is
1532 * smaller than a page size then we can fallback in emergencies
1533 * to the page allocator via kmalloc_large. The page allocator may
1534 * have failed to obtain a higher order page and we can try to
1535 * allocate a single page if the object fits into a single page.
1536 * That is only possible if certain conditions are met that are being
1537 * checked when a slab is created.
1538 */
1539 if (!(gfpflags & __GFP_NORETRY) && (s->flags & __PAGE_ALLOC_FALLBACK))
1540 return kmalloc_large(s->objsize, gfpflags);
1541
1542 return NULL;
81819f0f 1543debug:
dfb4f096 1544 if (!alloc_debug_processing(s, c->page, object, addr))
81819f0f 1545 goto another_slab;
894b8788 1546
dfb4f096 1547 c->page->inuse++;
b3fba8da 1548 c->page->freelist = object[c->offset];
ee3c72a1 1549 c->node = -1;
1f84260c 1550 goto unlock_out;
894b8788
CL
1551}
1552
1553/*
1554 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1555 * have the fastpath folded into their functions. So no function call
1556 * overhead for requests that can be satisfied on the fastpath.
1557 *
1558 * The fastpath works by first checking if the lockless freelist can be used.
1559 * If not then __slab_alloc is called for slow processing.
1560 *
1561 * Otherwise we can simply pick the next object from the lockless free list.
1562 */
06428780 1563static __always_inline void *slab_alloc(struct kmem_cache *s,
ce15fea8 1564 gfp_t gfpflags, int node, void *addr)
894b8788 1565{
894b8788 1566 void **object;
dfb4f096 1567 struct kmem_cache_cpu *c;
1f84260c
CL
1568 unsigned long flags;
1569
894b8788 1570 local_irq_save(flags);
dfb4f096 1571 c = get_cpu_slab(s, smp_processor_id());
a973e9dd 1572 if (unlikely(!c->freelist || !node_match(c, node)))
894b8788 1573
dfb4f096 1574 object = __slab_alloc(s, gfpflags, node, addr, c);
894b8788
CL
1575
1576 else {
dfb4f096 1577 object = c->freelist;
b3fba8da 1578 c->freelist = object[c->offset];
8ff12cfc 1579 stat(c, ALLOC_FASTPATH);
894b8788
CL
1580 }
1581 local_irq_restore(flags);
d07dbea4
CL
1582
1583 if (unlikely((gfpflags & __GFP_ZERO) && object))
42a9fdbb 1584 memset(object, 0, c->objsize);
d07dbea4 1585
894b8788 1586 return object;
81819f0f
CL
1587}
1588
1589void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1590{
ce15fea8 1591 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
81819f0f
CL
1592}
1593EXPORT_SYMBOL(kmem_cache_alloc);
1594
1595#ifdef CONFIG_NUMA
1596void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1597{
ce15fea8 1598 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
81819f0f
CL
1599}
1600EXPORT_SYMBOL(kmem_cache_alloc_node);
1601#endif
1602
1603/*
894b8788
CL
1604 * Slow patch handling. This may still be called frequently since objects
1605 * have a longer lifetime than the cpu slabs in most processing loads.
81819f0f 1606 *
894b8788
CL
1607 * So we still attempt to reduce cache line usage. Just take the slab
1608 * lock and free the item. If there is no additional partial page
1609 * handling required then we can return immediately.
81819f0f 1610 */
894b8788 1611static void __slab_free(struct kmem_cache *s, struct page *page,
b3fba8da 1612 void *x, void *addr, unsigned int offset)
81819f0f
CL
1613{
1614 void *prior;
1615 void **object = (void *)x;
8ff12cfc 1616 struct kmem_cache_cpu *c;
81819f0f 1617
8ff12cfc
CL
1618 c = get_cpu_slab(s, raw_smp_processor_id());
1619 stat(c, FREE_SLOWPATH);
81819f0f
CL
1620 slab_lock(page);
1621
35e5d7ee 1622 if (unlikely(SlabDebug(page)))
81819f0f 1623 goto debug;
6446faa2 1624
81819f0f 1625checks_ok:
b3fba8da 1626 prior = object[offset] = page->freelist;
81819f0f
CL
1627 page->freelist = object;
1628 page->inuse--;
1629
8ff12cfc
CL
1630 if (unlikely(SlabFrozen(page))) {
1631 stat(c, FREE_FROZEN);
81819f0f 1632 goto out_unlock;
8ff12cfc 1633 }
81819f0f
CL
1634
1635 if (unlikely(!page->inuse))
1636 goto slab_empty;
1637
1638 /*
6446faa2 1639 * Objects left in the slab. If it was not on the partial list before
81819f0f
CL
1640 * then add it.
1641 */
a973e9dd 1642 if (unlikely(!prior)) {
7c2e132c 1643 add_partial(get_node(s, page_to_nid(page)), page, 1);
8ff12cfc
CL
1644 stat(c, FREE_ADD_PARTIAL);
1645 }
81819f0f
CL
1646
1647out_unlock:
1648 slab_unlock(page);
81819f0f
CL
1649 return;
1650
1651slab_empty:
a973e9dd 1652 if (prior) {
81819f0f 1653 /*
672bba3a 1654 * Slab still on the partial list.
81819f0f
CL
1655 */
1656 remove_partial(s, page);
8ff12cfc
CL
1657 stat(c, FREE_REMOVE_PARTIAL);
1658 }
81819f0f 1659 slab_unlock(page);
8ff12cfc 1660 stat(c, FREE_SLAB);
81819f0f 1661 discard_slab(s, page);
81819f0f
CL
1662 return;
1663
1664debug:
3ec09742 1665 if (!free_debug_processing(s, page, x, addr))
77c5e2d0 1666 goto out_unlock;
77c5e2d0 1667 goto checks_ok;
81819f0f
CL
1668}
1669
894b8788
CL
1670/*
1671 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1672 * can perform fastpath freeing without additional function calls.
1673 *
1674 * The fastpath is only possible if we are freeing to the current cpu slab
1675 * of this processor. This typically the case if we have just allocated
1676 * the item before.
1677 *
1678 * If fastpath is not possible then fall back to __slab_free where we deal
1679 * with all sorts of special processing.
1680 */
06428780 1681static __always_inline void slab_free(struct kmem_cache *s,
894b8788
CL
1682 struct page *page, void *x, void *addr)
1683{
1684 void **object = (void *)x;
dfb4f096 1685 struct kmem_cache_cpu *c;
1f84260c
CL
1686 unsigned long flags;
1687
894b8788 1688 local_irq_save(flags);
dfb4f096 1689 c = get_cpu_slab(s, smp_processor_id());
27d9e4e9 1690 debug_check_no_locks_freed(object, c->objsize);
ee3c72a1 1691 if (likely(page == c->page && c->node >= 0)) {
b3fba8da 1692 object[c->offset] = c->freelist;
dfb4f096 1693 c->freelist = object;
8ff12cfc 1694 stat(c, FREE_FASTPATH);
894b8788 1695 } else
b3fba8da 1696 __slab_free(s, page, x, addr, c->offset);
894b8788
CL
1697
1698 local_irq_restore(flags);
1699}
1700
81819f0f
CL
1701void kmem_cache_free(struct kmem_cache *s, void *x)
1702{
77c5e2d0 1703 struct page *page;
81819f0f 1704
b49af68f 1705 page = virt_to_head_page(x);
81819f0f 1706
77c5e2d0 1707 slab_free(s, page, x, __builtin_return_address(0));
81819f0f
CL
1708}
1709EXPORT_SYMBOL(kmem_cache_free);
1710
1711/* Figure out on which slab object the object resides */
1712static struct page *get_object_page(const void *x)
1713{
b49af68f 1714 struct page *page = virt_to_head_page(x);
81819f0f
CL
1715
1716 if (!PageSlab(page))
1717 return NULL;
1718
1719 return page;
1720}
1721
1722/*
672bba3a
CL
1723 * Object placement in a slab is made very easy because we always start at
1724 * offset 0. If we tune the size of the object to the alignment then we can
1725 * get the required alignment by putting one properly sized object after
1726 * another.
81819f0f
CL
1727 *
1728 * Notice that the allocation order determines the sizes of the per cpu
1729 * caches. Each processor has always one slab available for allocations.
1730 * Increasing the allocation order reduces the number of times that slabs
672bba3a 1731 * must be moved on and off the partial lists and is therefore a factor in
81819f0f 1732 * locking overhead.
81819f0f
CL
1733 */
1734
1735/*
1736 * Mininum / Maximum order of slab pages. This influences locking overhead
1737 * and slab fragmentation. A higher order reduces the number of partial slabs
1738 * and increases the number of allocations possible without having to
1739 * take the list_lock.
1740 */
1741static int slub_min_order;
1742static int slub_max_order = DEFAULT_MAX_ORDER;
81819f0f
CL
1743static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1744
1745/*
1746 * Merge control. If this is set then no merging of slab caches will occur.
672bba3a 1747 * (Could be removed. This was introduced to pacify the merge skeptics.)
81819f0f
CL
1748 */
1749static int slub_nomerge;
1750
81819f0f
CL
1751/*
1752 * Calculate the order of allocation given an slab object size.
1753 *
672bba3a
CL
1754 * The order of allocation has significant impact on performance and other
1755 * system components. Generally order 0 allocations should be preferred since
1756 * order 0 does not cause fragmentation in the page allocator. Larger objects
1757 * be problematic to put into order 0 slabs because there may be too much
1758 * unused space left. We go to a higher order if more than 1/8th of the slab
1759 * would be wasted.
1760 *
1761 * In order to reach satisfactory performance we must ensure that a minimum
1762 * number of objects is in one slab. Otherwise we may generate too much
1763 * activity on the partial lists which requires taking the list_lock. This is
1764 * less a concern for large slabs though which are rarely used.
81819f0f 1765 *
672bba3a
CL
1766 * slub_max_order specifies the order where we begin to stop considering the
1767 * number of objects in a slab as critical. If we reach slub_max_order then
1768 * we try to keep the page order as low as possible. So we accept more waste
1769 * of space in favor of a small page order.
81819f0f 1770 *
672bba3a
CL
1771 * Higher order allocations also allow the placement of more objects in a
1772 * slab and thereby reduce object handling overhead. If the user has
1773 * requested a higher mininum order then we start with that one instead of
1774 * the smallest order which will fit the object.
81819f0f 1775 */
5e6d444e
CL
1776static inline int slab_order(int size, int min_objects,
1777 int max_order, int fract_leftover)
81819f0f
CL
1778{
1779 int order;
1780 int rem;
6300ea75 1781 int min_order = slub_min_order;
81819f0f 1782
6300ea75 1783 for (order = max(min_order,
5e6d444e
CL
1784 fls(min_objects * size - 1) - PAGE_SHIFT);
1785 order <= max_order; order++) {
81819f0f 1786
5e6d444e 1787 unsigned long slab_size = PAGE_SIZE << order;
81819f0f 1788
5e6d444e 1789 if (slab_size < min_objects * size)
81819f0f
CL
1790 continue;
1791
1792 rem = slab_size % size;
1793
5e6d444e 1794 if (rem <= slab_size / fract_leftover)
81819f0f
CL
1795 break;
1796
1797 }
672bba3a 1798
81819f0f
CL
1799 return order;
1800}
1801
5e6d444e
CL
1802static inline int calculate_order(int size)
1803{
1804 int order;
1805 int min_objects;
1806 int fraction;
1807
1808 /*
1809 * Attempt to find best configuration for a slab. This
1810 * works by first attempting to generate a layout with
1811 * the best configuration and backing off gradually.
1812 *
1813 * First we reduce the acceptable waste in a slab. Then
1814 * we reduce the minimum objects required in a slab.
1815 */
1816 min_objects = slub_min_objects;
1817 while (min_objects > 1) {
1818 fraction = 8;
1819 while (fraction >= 4) {
1820 order = slab_order(size, min_objects,
1821 slub_max_order, fraction);
1822 if (order <= slub_max_order)
1823 return order;
1824 fraction /= 2;
1825 }
1826 min_objects /= 2;
1827 }
1828
1829 /*
1830 * We were unable to place multiple objects in a slab. Now
1831 * lets see if we can place a single object there.
1832 */
1833 order = slab_order(size, 1, slub_max_order, 1);
1834 if (order <= slub_max_order)
1835 return order;
1836
1837 /*
1838 * Doh this slab cannot be placed using slub_max_order.
1839 */
1840 order = slab_order(size, 1, MAX_ORDER, 1);
1841 if (order <= MAX_ORDER)
1842 return order;
1843 return -ENOSYS;
1844}
1845
81819f0f 1846/*
672bba3a 1847 * Figure out what the alignment of the objects will be.
81819f0f
CL
1848 */
1849static unsigned long calculate_alignment(unsigned long flags,
1850 unsigned long align, unsigned long size)
1851{
1852 /*
6446faa2
CL
1853 * If the user wants hardware cache aligned objects then follow that
1854 * suggestion if the object is sufficiently large.
81819f0f 1855 *
6446faa2
CL
1856 * The hardware cache alignment cannot override the specified
1857 * alignment though. If that is greater then use it.
81819f0f 1858 */
b6210386
NP
1859 if (flags & SLAB_HWCACHE_ALIGN) {
1860 unsigned long ralign = cache_line_size();
1861 while (size <= ralign / 2)
1862 ralign /= 2;
1863 align = max(align, ralign);
1864 }
81819f0f
CL
1865
1866 if (align < ARCH_SLAB_MINALIGN)
b6210386 1867 align = ARCH_SLAB_MINALIGN;
81819f0f
CL
1868
1869 return ALIGN(align, sizeof(void *));
1870}
1871
dfb4f096
CL
1872static void init_kmem_cache_cpu(struct kmem_cache *s,
1873 struct kmem_cache_cpu *c)
1874{
1875 c->page = NULL;
a973e9dd 1876 c->freelist = NULL;
dfb4f096 1877 c->node = 0;
42a9fdbb
CL
1878 c->offset = s->offset / sizeof(void *);
1879 c->objsize = s->objsize;
dfb4f096
CL
1880}
1881
81819f0f
CL
1882static void init_kmem_cache_node(struct kmem_cache_node *n)
1883{
1884 n->nr_partial = 0;
1885 atomic_long_set(&n->nr_slabs, 0);
1886 spin_lock_init(&n->list_lock);
1887 INIT_LIST_HEAD(&n->partial);
8ab1372f 1888#ifdef CONFIG_SLUB_DEBUG
643b1138 1889 INIT_LIST_HEAD(&n->full);
8ab1372f 1890#endif
81819f0f
CL
1891}
1892
4c93c355
CL
1893#ifdef CONFIG_SMP
1894/*
1895 * Per cpu array for per cpu structures.
1896 *
1897 * The per cpu array places all kmem_cache_cpu structures from one processor
1898 * close together meaning that it becomes possible that multiple per cpu
1899 * structures are contained in one cacheline. This may be particularly
1900 * beneficial for the kmalloc caches.
1901 *
1902 * A desktop system typically has around 60-80 slabs. With 100 here we are
1903 * likely able to get per cpu structures for all caches from the array defined
1904 * here. We must be able to cover all kmalloc caches during bootstrap.
1905 *
1906 * If the per cpu array is exhausted then fall back to kmalloc
1907 * of individual cachelines. No sharing is possible then.
1908 */
1909#define NR_KMEM_CACHE_CPU 100
1910
1911static DEFINE_PER_CPU(struct kmem_cache_cpu,
1912 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1913
1914static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1915static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1916
1917static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1918 int cpu, gfp_t flags)
1919{
1920 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1921
1922 if (c)
1923 per_cpu(kmem_cache_cpu_free, cpu) =
1924 (void *)c->freelist;
1925 else {
1926 /* Table overflow: So allocate ourselves */
1927 c = kmalloc_node(
1928 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
1929 flags, cpu_to_node(cpu));
1930 if (!c)
1931 return NULL;
1932 }
1933
1934 init_kmem_cache_cpu(s, c);
1935 return c;
1936}
1937
1938static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
1939{
1940 if (c < per_cpu(kmem_cache_cpu, cpu) ||
1941 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
1942 kfree(c);
1943 return;
1944 }
1945 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
1946 per_cpu(kmem_cache_cpu_free, cpu) = c;
1947}
1948
1949static void free_kmem_cache_cpus(struct kmem_cache *s)
1950{
1951 int cpu;
1952
1953 for_each_online_cpu(cpu) {
1954 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1955
1956 if (c) {
1957 s->cpu_slab[cpu] = NULL;
1958 free_kmem_cache_cpu(c, cpu);
1959 }
1960 }
1961}
1962
1963static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
1964{
1965 int cpu;
1966
1967 for_each_online_cpu(cpu) {
1968 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1969
1970 if (c)
1971 continue;
1972
1973 c = alloc_kmem_cache_cpu(s, cpu, flags);
1974 if (!c) {
1975 free_kmem_cache_cpus(s);
1976 return 0;
1977 }
1978 s->cpu_slab[cpu] = c;
1979 }
1980 return 1;
1981}
1982
1983/*
1984 * Initialize the per cpu array.
1985 */
1986static void init_alloc_cpu_cpu(int cpu)
1987{
1988 int i;
1989
1990 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
1991 return;
1992
1993 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
1994 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
1995
1996 cpu_set(cpu, kmem_cach_cpu_free_init_once);
1997}
1998
1999static void __init init_alloc_cpu(void)
2000{
2001 int cpu;
2002
2003 for_each_online_cpu(cpu)
2004 init_alloc_cpu_cpu(cpu);
2005 }
2006
2007#else
2008static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2009static inline void init_alloc_cpu(void) {}
2010
2011static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2012{
2013 init_kmem_cache_cpu(s, &s->cpu_slab);
2014 return 1;
2015}
2016#endif
2017
81819f0f
CL
2018#ifdef CONFIG_NUMA
2019/*
2020 * No kmalloc_node yet so do it by hand. We know that this is the first
2021 * slab on the node for this slabcache. There are no concurrent accesses
2022 * possible.
2023 *
2024 * Note that this function only works on the kmalloc_node_cache
4c93c355
CL
2025 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2026 * memory on a fresh node that has no slab structures yet.
81819f0f 2027 */
1cd7daa5
AB
2028static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2029 int node)
81819f0f
CL
2030{
2031 struct page *page;
2032 struct kmem_cache_node *n;
ba84c73c 2033 unsigned long flags;
81819f0f
CL
2034
2035 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2036
a2f92ee7 2037 page = new_slab(kmalloc_caches, gfpflags, node);
81819f0f
CL
2038
2039 BUG_ON(!page);
a2f92ee7
CL
2040 if (page_to_nid(page) != node) {
2041 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2042 "node %d\n", node);
2043 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2044 "in order to be able to continue\n");
2045 }
2046
81819f0f
CL
2047 n = page->freelist;
2048 BUG_ON(!n);
2049 page->freelist = get_freepointer(kmalloc_caches, n);
2050 page->inuse++;
2051 kmalloc_caches->node[node] = n;
8ab1372f 2052#ifdef CONFIG_SLUB_DEBUG
d45f39cb
CL
2053 init_object(kmalloc_caches, n, 1);
2054 init_tracking(kmalloc_caches, n);
8ab1372f 2055#endif
81819f0f
CL
2056 init_kmem_cache_node(n);
2057 atomic_long_inc(&n->nr_slabs);
6446faa2 2058
ba84c73c 2059 /*
2060 * lockdep requires consistent irq usage for each lock
2061 * so even though there cannot be a race this early in
2062 * the boot sequence, we still disable irqs.
2063 */
2064 local_irq_save(flags);
7c2e132c 2065 add_partial(n, page, 0);
ba84c73c 2066 local_irq_restore(flags);
81819f0f
CL
2067 return n;
2068}
2069
2070static void free_kmem_cache_nodes(struct kmem_cache *s)
2071{
2072 int node;
2073
f64dc58c 2074 for_each_node_state(node, N_NORMAL_MEMORY) {
81819f0f
CL
2075 struct kmem_cache_node *n = s->node[node];
2076 if (n && n != &s->local_node)
2077 kmem_cache_free(kmalloc_caches, n);
2078 s->node[node] = NULL;
2079 }
2080}
2081
2082static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2083{
2084 int node;
2085 int local_node;
2086
2087 if (slab_state >= UP)
2088 local_node = page_to_nid(virt_to_page(s));
2089 else
2090 local_node = 0;
2091
f64dc58c 2092 for_each_node_state(node, N_NORMAL_MEMORY) {
81819f0f
CL
2093 struct kmem_cache_node *n;
2094
2095 if (local_node == node)
2096 n = &s->local_node;
2097 else {
2098 if (slab_state == DOWN) {
2099 n = early_kmem_cache_node_alloc(gfpflags,
2100 node);
2101 continue;
2102 }
2103 n = kmem_cache_alloc_node(kmalloc_caches,
2104 gfpflags, node);
2105
2106 if (!n) {
2107 free_kmem_cache_nodes(s);
2108 return 0;
2109 }
2110
2111 }
2112 s->node[node] = n;
2113 init_kmem_cache_node(n);
2114 }
2115 return 1;
2116}
2117#else
2118static void free_kmem_cache_nodes(struct kmem_cache *s)
2119{
2120}
2121
2122static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2123{
2124 init_kmem_cache_node(&s->local_node);
2125 return 1;
2126}
2127#endif
2128
2129/*
2130 * calculate_sizes() determines the order and the distribution of data within
2131 * a slab object.
2132 */
2133static int calculate_sizes(struct kmem_cache *s)
2134{
2135 unsigned long flags = s->flags;
2136 unsigned long size = s->objsize;
2137 unsigned long align = s->align;
2138
d8b42bf5
CL
2139 /*
2140 * Round up object size to the next word boundary. We can only
2141 * place the free pointer at word boundaries and this determines
2142 * the possible location of the free pointer.
2143 */
2144 size = ALIGN(size, sizeof(void *));
2145
2146#ifdef CONFIG_SLUB_DEBUG
81819f0f
CL
2147 /*
2148 * Determine if we can poison the object itself. If the user of
2149 * the slab may touch the object after free or before allocation
2150 * then we should never poison the object itself.
2151 */
2152 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
c59def9f 2153 !s->ctor)
81819f0f
CL
2154 s->flags |= __OBJECT_POISON;
2155 else
2156 s->flags &= ~__OBJECT_POISON;
2157
81819f0f
CL
2158
2159 /*
672bba3a 2160 * If we are Redzoning then check if there is some space between the
81819f0f 2161 * end of the object and the free pointer. If not then add an
672bba3a 2162 * additional word to have some bytes to store Redzone information.
81819f0f
CL
2163 */
2164 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2165 size += sizeof(void *);
41ecc55b 2166#endif
81819f0f
CL
2167
2168 /*
672bba3a
CL
2169 * With that we have determined the number of bytes in actual use
2170 * by the object. This is the potential offset to the free pointer.
81819f0f
CL
2171 */
2172 s->inuse = size;
2173
2174 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
c59def9f 2175 s->ctor)) {
81819f0f
CL
2176 /*
2177 * Relocate free pointer after the object if it is not
2178 * permitted to overwrite the first word of the object on
2179 * kmem_cache_free.
2180 *
2181 * This is the case if we do RCU, have a constructor or
2182 * destructor or are poisoning the objects.
2183 */
2184 s->offset = size;
2185 size += sizeof(void *);
2186 }
2187
c12b3c62 2188#ifdef CONFIG_SLUB_DEBUG
81819f0f
CL
2189 if (flags & SLAB_STORE_USER)
2190 /*
2191 * Need to store information about allocs and frees after
2192 * the object.
2193 */
2194 size += 2 * sizeof(struct track);
2195
be7b3fbc 2196 if (flags & SLAB_RED_ZONE)
81819f0f
CL
2197 /*
2198 * Add some empty padding so that we can catch
2199 * overwrites from earlier objects rather than let
2200 * tracking information or the free pointer be
2201 * corrupted if an user writes before the start
2202 * of the object.
2203 */
2204 size += sizeof(void *);
41ecc55b 2205#endif
672bba3a 2206
81819f0f
CL
2207 /*
2208 * Determine the alignment based on various parameters that the
65c02d4c
CL
2209 * user specified and the dynamic determination of cache line size
2210 * on bootup.
81819f0f
CL
2211 */
2212 align = calculate_alignment(flags, align, s->objsize);
2213
2214 /*
2215 * SLUB stores one object immediately after another beginning from
2216 * offset 0. In order to align the objects we have to simply size
2217 * each object to conform to the alignment.
2218 */
2219 size = ALIGN(size, align);
2220 s->size = size;
2221
71c7a06f
CL
2222 if ((flags & __KMALLOC_CACHE) &&
2223 PAGE_SIZE / size < slub_min_objects) {
2224 /*
2225 * Kmalloc cache that would not have enough objects in
2226 * an order 0 page. Kmalloc slabs can fallback to
2227 * page allocator order 0 allocs so take a reasonably large
2228 * order that will allows us a good number of objects.
2229 */
2230 s->order = max(slub_max_order, PAGE_ALLOC_COSTLY_ORDER);
2231 s->flags |= __PAGE_ALLOC_FALLBACK;
2232 s->allocflags |= __GFP_NOWARN;
2233 } else
2234 s->order = calculate_order(size);
2235
81819f0f
CL
2236 if (s->order < 0)
2237 return 0;
2238
b7a49f0d
CL
2239 s->allocflags = 0;
2240 if (s->order)
2241 s->allocflags |= __GFP_COMP;
2242
2243 if (s->flags & SLAB_CACHE_DMA)
2244 s->allocflags |= SLUB_DMA;
2245
2246 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2247 s->allocflags |= __GFP_RECLAIMABLE;
2248
81819f0f
CL
2249 /*
2250 * Determine the number of objects per slab
2251 */
2252 s->objects = (PAGE_SIZE << s->order) / size;
2253
b3fba8da 2254 return !!s->objects;
81819f0f
CL
2255
2256}
2257
81819f0f
CL
2258static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2259 const char *name, size_t size,
2260 size_t align, unsigned long flags,
4ba9b9d0 2261 void (*ctor)(struct kmem_cache *, void *))
81819f0f
CL
2262{
2263 memset(s, 0, kmem_size);
2264 s->name = name;
2265 s->ctor = ctor;
81819f0f 2266 s->objsize = size;
81819f0f 2267 s->align = align;
ba0268a8 2268 s->flags = kmem_cache_flags(size, flags, name, ctor);
81819f0f
CL
2269
2270 if (!calculate_sizes(s))
2271 goto error;
2272
2273 s->refcount = 1;
2274#ifdef CONFIG_NUMA
9824601e 2275 s->remote_node_defrag_ratio = 100;
81819f0f 2276#endif
dfb4f096
CL
2277 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2278 goto error;
81819f0f 2279
dfb4f096 2280 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
81819f0f 2281 return 1;
4c93c355 2282 free_kmem_cache_nodes(s);
81819f0f
CL
2283error:
2284 if (flags & SLAB_PANIC)
2285 panic("Cannot create slab %s size=%lu realsize=%u "
2286 "order=%u offset=%u flags=%lx\n",
2287 s->name, (unsigned long)size, s->size, s->order,
2288 s->offset, flags);
2289 return 0;
2290}
81819f0f
CL
2291
2292/*
2293 * Check if a given pointer is valid
2294 */
2295int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2296{
06428780 2297 struct page *page;
81819f0f
CL
2298
2299 page = get_object_page(object);
2300
2301 if (!page || s != page->slab)
2302 /* No slab or wrong slab */
2303 return 0;
2304
abcd08a6 2305 if (!check_valid_pointer(s, page, object))
81819f0f
CL
2306 return 0;
2307
2308 /*
2309 * We could also check if the object is on the slabs freelist.
2310 * But this would be too expensive and it seems that the main
6446faa2 2311 * purpose of kmem_ptr_valid() is to check if the object belongs
81819f0f
CL
2312 * to a certain slab.
2313 */
2314 return 1;
2315}
2316EXPORT_SYMBOL(kmem_ptr_validate);
2317
2318/*
2319 * Determine the size of a slab object
2320 */
2321unsigned int kmem_cache_size(struct kmem_cache *s)
2322{
2323 return s->objsize;
2324}
2325EXPORT_SYMBOL(kmem_cache_size);
2326
2327const char *kmem_cache_name(struct kmem_cache *s)
2328{
2329 return s->name;
2330}
2331EXPORT_SYMBOL(kmem_cache_name);
2332
2333/*
672bba3a
CL
2334 * Attempt to free all slabs on a node. Return the number of slabs we
2335 * were unable to free.
81819f0f
CL
2336 */
2337static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2338 struct list_head *list)
2339{
2340 int slabs_inuse = 0;
2341 unsigned long flags;
2342 struct page *page, *h;
2343
2344 spin_lock_irqsave(&n->list_lock, flags);
2345 list_for_each_entry_safe(page, h, list, lru)
2346 if (!page->inuse) {
2347 list_del(&page->lru);
2348 discard_slab(s, page);
2349 } else
2350 slabs_inuse++;
2351 spin_unlock_irqrestore(&n->list_lock, flags);
2352 return slabs_inuse;
2353}
2354
2355/*
672bba3a 2356 * Release all resources used by a slab cache.
81819f0f 2357 */
0c710013 2358static inline int kmem_cache_close(struct kmem_cache *s)
81819f0f
CL
2359{
2360 int node;
2361
2362 flush_all(s);
2363
2364 /* Attempt to free all objects */
4c93c355 2365 free_kmem_cache_cpus(s);
f64dc58c 2366 for_each_node_state(node, N_NORMAL_MEMORY) {
81819f0f
CL
2367 struct kmem_cache_node *n = get_node(s, node);
2368
2086d26a 2369 n->nr_partial -= free_list(s, n, &n->partial);
81819f0f
CL
2370 if (atomic_long_read(&n->nr_slabs))
2371 return 1;
2372 }
2373 free_kmem_cache_nodes(s);
2374 return 0;
2375}
2376
2377/*
2378 * Close a cache and release the kmem_cache structure
2379 * (must be used for caches created using kmem_cache_create)
2380 */
2381void kmem_cache_destroy(struct kmem_cache *s)
2382{
2383 down_write(&slub_lock);
2384 s->refcount--;
2385 if (!s->refcount) {
2386 list_del(&s->list);
a0e1d1be 2387 up_write(&slub_lock);
81819f0f
CL
2388 if (kmem_cache_close(s))
2389 WARN_ON(1);
2390 sysfs_slab_remove(s);
a0e1d1be
CL
2391 } else
2392 up_write(&slub_lock);
81819f0f
CL
2393}
2394EXPORT_SYMBOL(kmem_cache_destroy);
2395
2396/********************************************************************
2397 * Kmalloc subsystem
2398 *******************************************************************/
2399
331dc558 2400struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
81819f0f
CL
2401EXPORT_SYMBOL(kmalloc_caches);
2402
2403#ifdef CONFIG_ZONE_DMA
331dc558 2404static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
81819f0f
CL
2405#endif
2406
2407static int __init setup_slub_min_order(char *str)
2408{
06428780 2409 get_option(&str, &slub_min_order);
81819f0f
CL
2410
2411 return 1;
2412}
2413
2414__setup("slub_min_order=", setup_slub_min_order);
2415
2416static int __init setup_slub_max_order(char *str)
2417{
06428780 2418 get_option(&str, &slub_max_order);
81819f0f
CL
2419
2420 return 1;
2421}
2422
2423__setup("slub_max_order=", setup_slub_max_order);
2424
2425static int __init setup_slub_min_objects(char *str)
2426{
06428780 2427 get_option(&str, &slub_min_objects);
81819f0f
CL
2428
2429 return 1;
2430}
2431
2432__setup("slub_min_objects=", setup_slub_min_objects);
2433
2434static int __init setup_slub_nomerge(char *str)
2435{
2436 slub_nomerge = 1;
2437 return 1;
2438}
2439
2440__setup("slub_nomerge", setup_slub_nomerge);
2441
81819f0f
CL
2442static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2443 const char *name, int size, gfp_t gfp_flags)
2444{
2445 unsigned int flags = 0;
2446
2447 if (gfp_flags & SLUB_DMA)
2448 flags = SLAB_CACHE_DMA;
2449
2450 down_write(&slub_lock);
2451 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
71c7a06f 2452 flags | __KMALLOC_CACHE, NULL))
81819f0f
CL
2453 goto panic;
2454
2455 list_add(&s->list, &slab_caches);
2456 up_write(&slub_lock);
2457 if (sysfs_slab_add(s))
2458 goto panic;
2459 return s;
2460
2461panic:
2462 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2463}
2464
2e443fd0 2465#ifdef CONFIG_ZONE_DMA
1ceef402
CL
2466
2467static void sysfs_add_func(struct work_struct *w)
2468{
2469 struct kmem_cache *s;
2470
2471 down_write(&slub_lock);
2472 list_for_each_entry(s, &slab_caches, list) {
2473 if (s->flags & __SYSFS_ADD_DEFERRED) {
2474 s->flags &= ~__SYSFS_ADD_DEFERRED;
2475 sysfs_slab_add(s);
2476 }
2477 }
2478 up_write(&slub_lock);
2479}
2480
2481static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2482
2e443fd0
CL
2483static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2484{
2485 struct kmem_cache *s;
2e443fd0
CL
2486 char *text;
2487 size_t realsize;
2488
2489 s = kmalloc_caches_dma[index];
2490 if (s)
2491 return s;
2492
2493 /* Dynamically create dma cache */
1ceef402
CL
2494 if (flags & __GFP_WAIT)
2495 down_write(&slub_lock);
2496 else {
2497 if (!down_write_trylock(&slub_lock))
2498 goto out;
2499 }
2500
2501 if (kmalloc_caches_dma[index])
2502 goto unlock_out;
2e443fd0 2503
7b55f620 2504 realsize = kmalloc_caches[index].objsize;
3adbefee
IM
2505 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2506 (unsigned int)realsize);
1ceef402
CL
2507 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2508
2509 if (!s || !text || !kmem_cache_open(s, flags, text,
2510 realsize, ARCH_KMALLOC_MINALIGN,
2511 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2512 kfree(s);
2513 kfree(text);
2514 goto unlock_out;
dfce8648 2515 }
1ceef402
CL
2516
2517 list_add(&s->list, &slab_caches);
2518 kmalloc_caches_dma[index] = s;
2519
2520 schedule_work(&sysfs_add_work);
2521
2522unlock_out:
dfce8648 2523 up_write(&slub_lock);
1ceef402 2524out:
dfce8648 2525 return kmalloc_caches_dma[index];
2e443fd0
CL
2526}
2527#endif
2528
f1b26339
CL
2529/*
2530 * Conversion table for small slabs sizes / 8 to the index in the
2531 * kmalloc array. This is necessary for slabs < 192 since we have non power
2532 * of two cache sizes there. The size of larger slabs can be determined using
2533 * fls.
2534 */
2535static s8 size_index[24] = {
2536 3, /* 8 */
2537 4, /* 16 */
2538 5, /* 24 */
2539 5, /* 32 */
2540 6, /* 40 */
2541 6, /* 48 */
2542 6, /* 56 */
2543 6, /* 64 */
2544 1, /* 72 */
2545 1, /* 80 */
2546 1, /* 88 */
2547 1, /* 96 */
2548 7, /* 104 */
2549 7, /* 112 */
2550 7, /* 120 */
2551 7, /* 128 */
2552 2, /* 136 */
2553 2, /* 144 */
2554 2, /* 152 */
2555 2, /* 160 */
2556 2, /* 168 */
2557 2, /* 176 */
2558 2, /* 184 */
2559 2 /* 192 */
2560};
2561
81819f0f
CL
2562static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2563{
f1b26339 2564 int index;
81819f0f 2565
f1b26339
CL
2566 if (size <= 192) {
2567 if (!size)
2568 return ZERO_SIZE_PTR;
81819f0f 2569
f1b26339 2570 index = size_index[(size - 1) / 8];
aadb4bc4 2571 } else
f1b26339 2572 index = fls(size - 1);
81819f0f
CL
2573
2574#ifdef CONFIG_ZONE_DMA
f1b26339 2575 if (unlikely((flags & SLUB_DMA)))
2e443fd0 2576 return dma_kmalloc_cache(index, flags);
f1b26339 2577
81819f0f
CL
2578#endif
2579 return &kmalloc_caches[index];
2580}
2581
2582void *__kmalloc(size_t size, gfp_t flags)
2583{
aadb4bc4 2584 struct kmem_cache *s;
81819f0f 2585
331dc558 2586 if (unlikely(size > PAGE_SIZE))
eada35ef 2587 return kmalloc_large(size, flags);
aadb4bc4
CL
2588
2589 s = get_slab(size, flags);
2590
2591 if (unlikely(ZERO_OR_NULL_PTR(s)))
6cb8f913
CL
2592 return s;
2593
ce15fea8 2594 return slab_alloc(s, flags, -1, __builtin_return_address(0));
81819f0f
CL
2595}
2596EXPORT_SYMBOL(__kmalloc);
2597
f619cfe1
CL
2598static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2599{
2600 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2601 get_order(size));
2602
2603 if (page)
2604 return page_address(page);
2605 else
2606 return NULL;
2607}
2608
81819f0f
CL
2609#ifdef CONFIG_NUMA
2610void *__kmalloc_node(size_t size, gfp_t flags, int node)
2611{
aadb4bc4 2612 struct kmem_cache *s;
81819f0f 2613
331dc558 2614 if (unlikely(size > PAGE_SIZE))
f619cfe1 2615 return kmalloc_large_node(size, flags, node);
aadb4bc4
CL
2616
2617 s = get_slab(size, flags);
2618
2619 if (unlikely(ZERO_OR_NULL_PTR(s)))
6cb8f913
CL
2620 return s;
2621
ce15fea8 2622 return slab_alloc(s, flags, node, __builtin_return_address(0));
81819f0f
CL
2623}
2624EXPORT_SYMBOL(__kmalloc_node);
2625#endif
2626
2627size_t ksize(const void *object)
2628{
272c1d21 2629 struct page *page;
81819f0f
CL
2630 struct kmem_cache *s;
2631
ef8b4520 2632 if (unlikely(object == ZERO_SIZE_PTR))
272c1d21
CL
2633 return 0;
2634
294a80a8 2635 page = virt_to_head_page(object);
294a80a8
VN
2636
2637 if (unlikely(!PageSlab(page)))
2638 return PAGE_SIZE << compound_order(page);
2639
81819f0f 2640 s = page->slab;
81819f0f 2641
ae20bfda 2642#ifdef CONFIG_SLUB_DEBUG
81819f0f
CL
2643 /*
2644 * Debugging requires use of the padding between object
2645 * and whatever may come after it.
2646 */
2647 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2648 return s->objsize;
2649
ae20bfda 2650#endif
81819f0f
CL
2651 /*
2652 * If we have the need to store the freelist pointer
2653 * back there or track user information then we can
2654 * only use the space before that information.
2655 */
2656 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2657 return s->inuse;
81819f0f
CL
2658 /*
2659 * Else we can use all the padding etc for the allocation
2660 */
2661 return s->size;
2662}
2663EXPORT_SYMBOL(ksize);
2664
2665void kfree(const void *x)
2666{
81819f0f 2667 struct page *page;
5bb983b0 2668 void *object = (void *)x;
81819f0f 2669
2408c550 2670 if (unlikely(ZERO_OR_NULL_PTR(x)))
81819f0f
CL
2671 return;
2672
b49af68f 2673 page = virt_to_head_page(x);
aadb4bc4
CL
2674 if (unlikely(!PageSlab(page))) {
2675 put_page(page);
2676 return;
2677 }
5bb983b0 2678 slab_free(page->slab, page, object, __builtin_return_address(0));
81819f0f
CL
2679}
2680EXPORT_SYMBOL(kfree);
2681
f61396ae
CL
2682static unsigned long count_partial(struct kmem_cache_node *n)
2683{
2684 unsigned long flags;
2685 unsigned long x = 0;
2686 struct page *page;
2687
2688 spin_lock_irqsave(&n->list_lock, flags);
2689 list_for_each_entry(page, &n->partial, lru)
2690 x += page->inuse;
2691 spin_unlock_irqrestore(&n->list_lock, flags);
2692 return x;
2693}
2694
2086d26a 2695/*
672bba3a
CL
2696 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2697 * the remaining slabs by the number of items in use. The slabs with the
2698 * most items in use come first. New allocations will then fill those up
2699 * and thus they can be removed from the partial lists.
2700 *
2701 * The slabs with the least items are placed last. This results in them
2702 * being allocated from last increasing the chance that the last objects
2703 * are freed in them.
2086d26a
CL
2704 */
2705int kmem_cache_shrink(struct kmem_cache *s)
2706{
2707 int node;
2708 int i;
2709 struct kmem_cache_node *n;
2710 struct page *page;
2711 struct page *t;
2712 struct list_head *slabs_by_inuse =
2713 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2714 unsigned long flags;
2715
2716 if (!slabs_by_inuse)
2717 return -ENOMEM;
2718
2719 flush_all(s);
f64dc58c 2720 for_each_node_state(node, N_NORMAL_MEMORY) {
2086d26a
CL
2721 n = get_node(s, node);
2722
2723 if (!n->nr_partial)
2724 continue;
2725
2726 for (i = 0; i < s->objects; i++)
2727 INIT_LIST_HEAD(slabs_by_inuse + i);
2728
2729 spin_lock_irqsave(&n->list_lock, flags);
2730
2731 /*
672bba3a 2732 * Build lists indexed by the items in use in each slab.
2086d26a 2733 *
672bba3a
CL
2734 * Note that concurrent frees may occur while we hold the
2735 * list_lock. page->inuse here is the upper limit.
2086d26a
CL
2736 */
2737 list_for_each_entry_safe(page, t, &n->partial, lru) {
2738 if (!page->inuse && slab_trylock(page)) {
2739 /*
2740 * Must hold slab lock here because slab_free
2741 * may have freed the last object and be
2742 * waiting to release the slab.
2743 */
2744 list_del(&page->lru);
2745 n->nr_partial--;
2746 slab_unlock(page);
2747 discard_slab(s, page);
2748 } else {
fcda3d89
CL
2749 list_move(&page->lru,
2750 slabs_by_inuse + page->inuse);
2086d26a
CL
2751 }
2752 }
2753
2086d26a 2754 /*
672bba3a
CL
2755 * Rebuild the partial list with the slabs filled up most
2756 * first and the least used slabs at the end.
2086d26a
CL
2757 */
2758 for (i = s->objects - 1; i >= 0; i--)
2759 list_splice(slabs_by_inuse + i, n->partial.prev);
2760
2086d26a
CL
2761 spin_unlock_irqrestore(&n->list_lock, flags);
2762 }
2763
2764 kfree(slabs_by_inuse);
2765 return 0;
2766}
2767EXPORT_SYMBOL(kmem_cache_shrink);
2768
b9049e23
YG
2769#if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2770static int slab_mem_going_offline_callback(void *arg)
2771{
2772 struct kmem_cache *s;
2773
2774 down_read(&slub_lock);
2775 list_for_each_entry(s, &slab_caches, list)
2776 kmem_cache_shrink(s);
2777 up_read(&slub_lock);
2778
2779 return 0;
2780}
2781
2782static void slab_mem_offline_callback(void *arg)
2783{
2784 struct kmem_cache_node *n;
2785 struct kmem_cache *s;
2786 struct memory_notify *marg = arg;
2787 int offline_node;
2788
2789 offline_node = marg->status_change_nid;
2790
2791 /*
2792 * If the node still has available memory. we need kmem_cache_node
2793 * for it yet.
2794 */
2795 if (offline_node < 0)
2796 return;
2797
2798 down_read(&slub_lock);
2799 list_for_each_entry(s, &slab_caches, list) {
2800 n = get_node(s, offline_node);
2801 if (n) {
2802 /*
2803 * if n->nr_slabs > 0, slabs still exist on the node
2804 * that is going down. We were unable to free them,
2805 * and offline_pages() function shoudn't call this
2806 * callback. So, we must fail.
2807 */
27bb628a 2808 BUG_ON(atomic_long_read(&n->nr_slabs));
b9049e23
YG
2809
2810 s->node[offline_node] = NULL;
2811 kmem_cache_free(kmalloc_caches, n);
2812 }
2813 }
2814 up_read(&slub_lock);
2815}
2816
2817static int slab_mem_going_online_callback(void *arg)
2818{
2819 struct kmem_cache_node *n;
2820 struct kmem_cache *s;
2821 struct memory_notify *marg = arg;
2822 int nid = marg->status_change_nid;
2823 int ret = 0;
2824
2825 /*
2826 * If the node's memory is already available, then kmem_cache_node is
2827 * already created. Nothing to do.
2828 */
2829 if (nid < 0)
2830 return 0;
2831
2832 /*
2833 * We are bringing a node online. No memory is availabe yet. We must
2834 * allocate a kmem_cache_node structure in order to bring the node
2835 * online.
2836 */
2837 down_read(&slub_lock);
2838 list_for_each_entry(s, &slab_caches, list) {
2839 /*
2840 * XXX: kmem_cache_alloc_node will fallback to other nodes
2841 * since memory is not yet available from the node that
2842 * is brought up.
2843 */
2844 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2845 if (!n) {
2846 ret = -ENOMEM;
2847 goto out;
2848 }
2849 init_kmem_cache_node(n);
2850 s->node[nid] = n;
2851 }
2852out:
2853 up_read(&slub_lock);
2854 return ret;
2855}
2856
2857static int slab_memory_callback(struct notifier_block *self,
2858 unsigned long action, void *arg)
2859{
2860 int ret = 0;
2861
2862 switch (action) {
2863 case MEM_GOING_ONLINE:
2864 ret = slab_mem_going_online_callback(arg);
2865 break;
2866 case MEM_GOING_OFFLINE:
2867 ret = slab_mem_going_offline_callback(arg);
2868 break;
2869 case MEM_OFFLINE:
2870 case MEM_CANCEL_ONLINE:
2871 slab_mem_offline_callback(arg);
2872 break;
2873 case MEM_ONLINE:
2874 case MEM_CANCEL_OFFLINE:
2875 break;
2876 }
2877
2878 ret = notifier_from_errno(ret);
2879 return ret;
2880}
2881
2882#endif /* CONFIG_MEMORY_HOTPLUG */
2883
81819f0f
CL
2884/********************************************************************
2885 * Basic setup of slabs
2886 *******************************************************************/
2887
2888void __init kmem_cache_init(void)
2889{
2890 int i;
4b356be0 2891 int caches = 0;
81819f0f 2892
4c93c355
CL
2893 init_alloc_cpu();
2894
81819f0f
CL
2895#ifdef CONFIG_NUMA
2896 /*
2897 * Must first have the slab cache available for the allocations of the
672bba3a 2898 * struct kmem_cache_node's. There is special bootstrap code in
81819f0f
CL
2899 * kmem_cache_open for slab_state == DOWN.
2900 */
2901 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2902 sizeof(struct kmem_cache_node), GFP_KERNEL);
8ffa6875 2903 kmalloc_caches[0].refcount = -1;
4b356be0 2904 caches++;
b9049e23
YG
2905
2906 hotplug_memory_notifier(slab_memory_callback, 1);
81819f0f
CL
2907#endif
2908
2909 /* Able to allocate the per node structures */
2910 slab_state = PARTIAL;
2911
2912 /* Caches that are not of the two-to-the-power-of size */
4b356be0
CL
2913 if (KMALLOC_MIN_SIZE <= 64) {
2914 create_kmalloc_cache(&kmalloc_caches[1],
81819f0f 2915 "kmalloc-96", 96, GFP_KERNEL);
4b356be0
CL
2916 caches++;
2917 }
2918 if (KMALLOC_MIN_SIZE <= 128) {
2919 create_kmalloc_cache(&kmalloc_caches[2],
81819f0f 2920 "kmalloc-192", 192, GFP_KERNEL);
4b356be0
CL
2921 caches++;
2922 }
81819f0f 2923
331dc558 2924 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
81819f0f
CL
2925 create_kmalloc_cache(&kmalloc_caches[i],
2926 "kmalloc", 1 << i, GFP_KERNEL);
4b356be0
CL
2927 caches++;
2928 }
81819f0f 2929
f1b26339
CL
2930
2931 /*
2932 * Patch up the size_index table if we have strange large alignment
2933 * requirements for the kmalloc array. This is only the case for
6446faa2 2934 * MIPS it seems. The standard arches will not generate any code here.
f1b26339
CL
2935 *
2936 * Largest permitted alignment is 256 bytes due to the way we
2937 * handle the index determination for the smaller caches.
2938 *
2939 * Make sure that nothing crazy happens if someone starts tinkering
2940 * around with ARCH_KMALLOC_MINALIGN
2941 */
2942 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2943 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2944
12ad6843 2945 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
f1b26339
CL
2946 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2947
81819f0f
CL
2948 slab_state = UP;
2949
2950 /* Provide the correct kmalloc names now that the caches are up */
331dc558 2951 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
81819f0f
CL
2952 kmalloc_caches[i]. name =
2953 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2954
2955#ifdef CONFIG_SMP
2956 register_cpu_notifier(&slab_notifier);
4c93c355
CL
2957 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2958 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
2959#else
2960 kmem_size = sizeof(struct kmem_cache);
81819f0f
CL
2961#endif
2962
3adbefee
IM
2963 printk(KERN_INFO
2964 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
4b356be0
CL
2965 " CPUs=%d, Nodes=%d\n",
2966 caches, cache_line_size(),
81819f0f
CL
2967 slub_min_order, slub_max_order, slub_min_objects,
2968 nr_cpu_ids, nr_node_ids);
2969}
2970
2971/*
2972 * Find a mergeable slab cache
2973 */
2974static int slab_unmergeable(struct kmem_cache *s)
2975{
2976 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2977 return 1;
2978
331dc558 2979 if ((s->flags & __PAGE_ALLOC_FALLBACK))
71c7a06f
CL
2980 return 1;
2981
c59def9f 2982 if (s->ctor)
81819f0f
CL
2983 return 1;
2984
8ffa6875
CL
2985 /*
2986 * We may have set a slab to be unmergeable during bootstrap.
2987 */
2988 if (s->refcount < 0)
2989 return 1;
2990
81819f0f
CL
2991 return 0;
2992}
2993
2994static struct kmem_cache *find_mergeable(size_t size,
ba0268a8 2995 size_t align, unsigned long flags, const char *name,
4ba9b9d0 2996 void (*ctor)(struct kmem_cache *, void *))
81819f0f 2997{
5b95a4ac 2998 struct kmem_cache *s;
81819f0f
CL
2999
3000 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3001 return NULL;
3002
c59def9f 3003 if (ctor)
81819f0f
CL
3004 return NULL;
3005
3006 size = ALIGN(size, sizeof(void *));
3007 align = calculate_alignment(flags, align, size);
3008 size = ALIGN(size, align);
ba0268a8 3009 flags = kmem_cache_flags(size, flags, name, NULL);
81819f0f 3010
5b95a4ac 3011 list_for_each_entry(s, &slab_caches, list) {
81819f0f
CL
3012 if (slab_unmergeable(s))
3013 continue;
3014
3015 if (size > s->size)
3016 continue;
3017
ba0268a8 3018 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
81819f0f
CL
3019 continue;
3020 /*
3021 * Check if alignment is compatible.
3022 * Courtesy of Adrian Drzewiecki
3023 */
06428780 3024 if ((s->size & ~(align - 1)) != s->size)
81819f0f
CL
3025 continue;
3026
3027 if (s->size - size >= sizeof(void *))
3028 continue;
3029
3030 return s;
3031 }
3032 return NULL;
3033}
3034
3035struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3036 size_t align, unsigned long flags,
4ba9b9d0 3037 void (*ctor)(struct kmem_cache *, void *))
81819f0f
CL
3038{
3039 struct kmem_cache *s;
3040
3041 down_write(&slub_lock);
ba0268a8 3042 s = find_mergeable(size, align, flags, name, ctor);
81819f0f 3043 if (s) {
42a9fdbb
CL
3044 int cpu;
3045
81819f0f
CL
3046 s->refcount++;
3047 /*
3048 * Adjust the object sizes so that we clear
3049 * the complete object on kzalloc.
3050 */
3051 s->objsize = max(s->objsize, (int)size);
42a9fdbb
CL
3052
3053 /*
3054 * And then we need to update the object size in the
3055 * per cpu structures
3056 */
3057 for_each_online_cpu(cpu)
3058 get_cpu_slab(s, cpu)->objsize = s->objsize;
6446faa2 3059
81819f0f 3060 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
a0e1d1be 3061 up_write(&slub_lock);
6446faa2 3062
81819f0f
CL
3063 if (sysfs_slab_alias(s, name))
3064 goto err;
a0e1d1be
CL
3065 return s;
3066 }
6446faa2 3067
a0e1d1be
CL
3068 s = kmalloc(kmem_size, GFP_KERNEL);
3069 if (s) {
3070 if (kmem_cache_open(s, GFP_KERNEL, name,
c59def9f 3071 size, align, flags, ctor)) {
81819f0f 3072 list_add(&s->list, &slab_caches);
a0e1d1be
CL
3073 up_write(&slub_lock);
3074 if (sysfs_slab_add(s))
3075 goto err;
3076 return s;
3077 }
3078 kfree(s);
81819f0f
CL
3079 }
3080 up_write(&slub_lock);
81819f0f
CL
3081
3082err:
81819f0f
CL
3083 if (flags & SLAB_PANIC)
3084 panic("Cannot create slabcache %s\n", name);
3085 else
3086 s = NULL;
3087 return s;
3088}
3089EXPORT_SYMBOL(kmem_cache_create);
3090
81819f0f 3091#ifdef CONFIG_SMP
81819f0f 3092/*
672bba3a
CL
3093 * Use the cpu notifier to insure that the cpu slabs are flushed when
3094 * necessary.
81819f0f
CL
3095 */
3096static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3097 unsigned long action, void *hcpu)
3098{
3099 long cpu = (long)hcpu;
5b95a4ac
CL
3100 struct kmem_cache *s;
3101 unsigned long flags;
81819f0f
CL
3102
3103 switch (action) {
4c93c355
CL
3104 case CPU_UP_PREPARE:
3105 case CPU_UP_PREPARE_FROZEN:
3106 init_alloc_cpu_cpu(cpu);
3107 down_read(&slub_lock);
3108 list_for_each_entry(s, &slab_caches, list)
3109 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3110 GFP_KERNEL);
3111 up_read(&slub_lock);
3112 break;
3113
81819f0f 3114 case CPU_UP_CANCELED:
8bb78442 3115 case CPU_UP_CANCELED_FROZEN:
81819f0f 3116 case CPU_DEAD:
8bb78442 3117 case CPU_DEAD_FROZEN:
5b95a4ac
CL
3118 down_read(&slub_lock);
3119 list_for_each_entry(s, &slab_caches, list) {
4c93c355
CL
3120 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3121
5b95a4ac
CL
3122 local_irq_save(flags);
3123 __flush_cpu_slab(s, cpu);
3124 local_irq_restore(flags);
4c93c355
CL
3125 free_kmem_cache_cpu(c, cpu);
3126 s->cpu_slab[cpu] = NULL;
5b95a4ac
CL
3127 }
3128 up_read(&slub_lock);
81819f0f
CL
3129 break;
3130 default:
3131 break;
3132 }
3133 return NOTIFY_OK;
3134}
3135
06428780 3136static struct notifier_block __cpuinitdata slab_notifier = {
3adbefee 3137 .notifier_call = slab_cpuup_callback
06428780 3138};
81819f0f
CL
3139
3140#endif
3141
81819f0f
CL
3142void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3143{
aadb4bc4
CL
3144 struct kmem_cache *s;
3145
331dc558 3146 if (unlikely(size > PAGE_SIZE))
eada35ef
PE
3147 return kmalloc_large(size, gfpflags);
3148
aadb4bc4 3149 s = get_slab(size, gfpflags);
81819f0f 3150
2408c550 3151 if (unlikely(ZERO_OR_NULL_PTR(s)))
6cb8f913 3152 return s;
81819f0f 3153
ce15fea8 3154 return slab_alloc(s, gfpflags, -1, caller);
81819f0f
CL
3155}
3156
3157void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3158 int node, void *caller)
3159{
aadb4bc4
CL
3160 struct kmem_cache *s;
3161
331dc558 3162 if (unlikely(size > PAGE_SIZE))
f619cfe1 3163 return kmalloc_large_node(size, gfpflags, node);
eada35ef 3164
aadb4bc4 3165 s = get_slab(size, gfpflags);
81819f0f 3166
2408c550 3167 if (unlikely(ZERO_OR_NULL_PTR(s)))
6cb8f913 3168 return s;
81819f0f 3169
ce15fea8 3170 return slab_alloc(s, gfpflags, node, caller);
81819f0f
CL
3171}
3172
41ecc55b 3173#if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
434e245d
CL
3174static int validate_slab(struct kmem_cache *s, struct page *page,
3175 unsigned long *map)
53e15af0
CL
3176{
3177 void *p;
a973e9dd 3178 void *addr = page_address(page);
53e15af0
CL
3179
3180 if (!check_slab(s, page) ||
3181 !on_freelist(s, page, NULL))
3182 return 0;
3183
3184 /* Now we know that a valid freelist exists */
3185 bitmap_zero(map, s->objects);
3186
7656c72b
CL
3187 for_each_free_object(p, s, page->freelist) {
3188 set_bit(slab_index(p, s, addr), map);
53e15af0
CL
3189 if (!check_object(s, page, p, 0))
3190 return 0;
3191 }
3192
7656c72b
CL
3193 for_each_object(p, s, addr)
3194 if (!test_bit(slab_index(p, s, addr), map))
53e15af0
CL
3195 if (!check_object(s, page, p, 1))
3196 return 0;
3197 return 1;
3198}
3199
434e245d
CL
3200static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3201 unsigned long *map)
53e15af0
CL
3202{
3203 if (slab_trylock(page)) {
434e245d 3204 validate_slab(s, page, map);
53e15af0
CL
3205 slab_unlock(page);
3206 } else
3207 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3208 s->name, page);
3209
3210 if (s->flags & DEBUG_DEFAULT_FLAGS) {
35e5d7ee
CL
3211 if (!SlabDebug(page))
3212 printk(KERN_ERR "SLUB %s: SlabDebug not set "
53e15af0
CL
3213 "on slab 0x%p\n", s->name, page);
3214 } else {
35e5d7ee
CL
3215 if (SlabDebug(page))
3216 printk(KERN_ERR "SLUB %s: SlabDebug set on "
53e15af0
CL
3217 "slab 0x%p\n", s->name, page);
3218 }
3219}
3220
434e245d
CL
3221static int validate_slab_node(struct kmem_cache *s,
3222 struct kmem_cache_node *n, unsigned long *map)
53e15af0
CL
3223{
3224 unsigned long count = 0;
3225 struct page *page;
3226 unsigned long flags;
3227
3228 spin_lock_irqsave(&n->list_lock, flags);
3229
3230 list_for_each_entry(page, &n->partial, lru) {
434e245d 3231 validate_slab_slab(s, page, map);
53e15af0
CL
3232 count++;
3233 }
3234 if (count != n->nr_partial)
3235 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3236 "counter=%ld\n", s->name, count, n->nr_partial);
3237
3238 if (!(s->flags & SLAB_STORE_USER))
3239 goto out;
3240
3241 list_for_each_entry(page, &n->full, lru) {
434e245d 3242 validate_slab_slab(s, page, map);
53e15af0
CL
3243 count++;
3244 }
3245 if (count != atomic_long_read(&n->nr_slabs))
3246 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3247 "counter=%ld\n", s->name, count,
3248 atomic_long_read(&n->nr_slabs));
3249
3250out:
3251 spin_unlock_irqrestore(&n->list_lock, flags);
3252 return count;
3253}
3254
434e245d 3255static long validate_slab_cache(struct kmem_cache *s)
53e15af0
CL
3256{
3257 int node;
3258 unsigned long count = 0;
434e245d
CL
3259 unsigned long *map = kmalloc(BITS_TO_LONGS(s->objects) *
3260 sizeof(unsigned long), GFP_KERNEL);
3261
3262 if (!map)
3263 return -ENOMEM;
53e15af0
CL
3264
3265 flush_all(s);
f64dc58c 3266 for_each_node_state(node, N_NORMAL_MEMORY) {
53e15af0
CL
3267 struct kmem_cache_node *n = get_node(s, node);
3268
434e245d 3269 count += validate_slab_node(s, n, map);
53e15af0 3270 }
434e245d 3271 kfree(map);
53e15af0
CL
3272 return count;
3273}
3274
b3459709
CL
3275#ifdef SLUB_RESILIENCY_TEST
3276static void resiliency_test(void)
3277{
3278 u8 *p;
3279
3280 printk(KERN_ERR "SLUB resiliency testing\n");
3281 printk(KERN_ERR "-----------------------\n");
3282 printk(KERN_ERR "A. Corruption after allocation\n");
3283
3284 p = kzalloc(16, GFP_KERNEL);
3285 p[16] = 0x12;
3286 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3287 " 0x12->0x%p\n\n", p + 16);
3288
3289 validate_slab_cache(kmalloc_caches + 4);
3290
3291 /* Hmmm... The next two are dangerous */
3292 p = kzalloc(32, GFP_KERNEL);
3293 p[32 + sizeof(void *)] = 0x34;
3294 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3adbefee
IM
3295 " 0x34 -> -0x%p\n", p);
3296 printk(KERN_ERR
3297 "If allocated object is overwritten then not detectable\n\n");
b3459709
CL
3298
3299 validate_slab_cache(kmalloc_caches + 5);
3300 p = kzalloc(64, GFP_KERNEL);
3301 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3302 *p = 0x56;
3303 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3304 p);
3adbefee
IM
3305 printk(KERN_ERR
3306 "If allocated object is overwritten then not detectable\n\n");
b3459709
CL
3307 validate_slab_cache(kmalloc_caches + 6);
3308
3309 printk(KERN_ERR "\nB. Corruption after free\n");
3310 p = kzalloc(128, GFP_KERNEL);
3311 kfree(p);
3312 *p = 0x78;
3313 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3314 validate_slab_cache(kmalloc_caches + 7);
3315
3316 p = kzalloc(256, GFP_KERNEL);
3317 kfree(p);
3318 p[50] = 0x9a;
3adbefee
IM
3319 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3320 p);
b3459709
CL
3321 validate_slab_cache(kmalloc_caches + 8);
3322
3323 p = kzalloc(512, GFP_KERNEL);
3324 kfree(p);
3325 p[512] = 0xab;
3326 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3327 validate_slab_cache(kmalloc_caches + 9);
3328}
3329#else
3330static void resiliency_test(void) {};
3331#endif
3332
88a420e4 3333/*
672bba3a 3334 * Generate lists of code addresses where slabcache objects are allocated
88a420e4
CL
3335 * and freed.
3336 */
3337
3338struct location {
3339 unsigned long count;
3340 void *addr;
45edfa58
CL
3341 long long sum_time;
3342 long min_time;
3343 long max_time;
3344 long min_pid;
3345 long max_pid;
3346 cpumask_t cpus;
3347 nodemask_t nodes;
88a420e4
CL
3348};
3349
3350struct loc_track {
3351 unsigned long max;
3352 unsigned long count;
3353 struct location *loc;
3354};
3355
3356static void free_loc_track(struct loc_track *t)
3357{
3358 if (t->max)
3359 free_pages((unsigned long)t->loc,
3360 get_order(sizeof(struct location) * t->max));
3361}
3362
68dff6a9 3363static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
88a420e4
CL
3364{
3365 struct location *l;
3366 int order;
3367
88a420e4
CL
3368 order = get_order(sizeof(struct location) * max);
3369
68dff6a9 3370 l = (void *)__get_free_pages(flags, order);
88a420e4
CL
3371 if (!l)
3372 return 0;
3373
3374 if (t->count) {
3375 memcpy(l, t->loc, sizeof(struct location) * t->count);
3376 free_loc_track(t);
3377 }
3378 t->max = max;
3379 t->loc = l;
3380 return 1;
3381}
3382
3383static int add_location(struct loc_track *t, struct kmem_cache *s,
45edfa58 3384 const struct track *track)
88a420e4
CL
3385{
3386 long start, end, pos;
3387 struct location *l;
3388 void *caddr;
45edfa58 3389 unsigned long age = jiffies - track->when;
88a420e4
CL
3390
3391 start = -1;
3392 end = t->count;
3393
3394 for ( ; ; ) {
3395 pos = start + (end - start + 1) / 2;
3396
3397 /*
3398 * There is nothing at "end". If we end up there
3399 * we need to add something to before end.
3400 */
3401 if (pos == end)
3402 break;
3403
3404 caddr = t->loc[pos].addr;
45edfa58
CL
3405 if (track->addr == caddr) {
3406
3407 l = &t->loc[pos];
3408 l->count++;
3409 if (track->when) {
3410 l->sum_time += age;
3411 if (age < l->min_time)
3412 l->min_time = age;
3413 if (age > l->max_time)
3414 l->max_time = age;
3415
3416 if (track->pid < l->min_pid)
3417 l->min_pid = track->pid;
3418 if (track->pid > l->max_pid)
3419 l->max_pid = track->pid;
3420
3421 cpu_set(track->cpu, l->cpus);
3422 }
3423 node_set(page_to_nid(virt_to_page(track)), l->nodes);
88a420e4
CL
3424 return 1;
3425 }
3426
45edfa58 3427 if (track->addr < caddr)
88a420e4
CL
3428 end = pos;
3429 else
3430 start = pos;
3431 }
3432
3433 /*
672bba3a 3434 * Not found. Insert new tracking element.
88a420e4 3435 */
68dff6a9 3436 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
88a420e4
CL
3437 return 0;
3438
3439 l = t->loc + pos;
3440 if (pos < t->count)
3441 memmove(l + 1, l,
3442 (t->count - pos) * sizeof(struct location));
3443 t->count++;
3444 l->count = 1;
45edfa58
CL
3445 l->addr = track->addr;
3446 l->sum_time = age;
3447 l->min_time = age;
3448 l->max_time = age;
3449 l->min_pid = track->pid;
3450 l->max_pid = track->pid;
3451 cpus_clear(l->cpus);
3452 cpu_set(track->cpu, l->cpus);
3453 nodes_clear(l->nodes);
3454 node_set(page_to_nid(virt_to_page(track)), l->nodes);
88a420e4
CL
3455 return 1;
3456}
3457
3458static void process_slab(struct loc_track *t, struct kmem_cache *s,
3459 struct page *page, enum track_item alloc)
3460{
a973e9dd 3461 void *addr = page_address(page);
7656c72b 3462 DECLARE_BITMAP(map, s->objects);
88a420e4
CL
3463 void *p;
3464
3465 bitmap_zero(map, s->objects);
7656c72b
CL
3466 for_each_free_object(p, s, page->freelist)
3467 set_bit(slab_index(p, s, addr), map);
88a420e4 3468
7656c72b 3469 for_each_object(p, s, addr)
45edfa58
CL
3470 if (!test_bit(slab_index(p, s, addr), map))
3471 add_location(t, s, get_track(s, p, alloc));
88a420e4
CL
3472}
3473
3474static int list_locations(struct kmem_cache *s, char *buf,
3475 enum track_item alloc)
3476{
e374d483 3477 int len = 0;
88a420e4 3478 unsigned long i;
68dff6a9 3479 struct loc_track t = { 0, 0, NULL };
88a420e4
CL
3480 int node;
3481
68dff6a9 3482 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
ea3061d2 3483 GFP_TEMPORARY))
68dff6a9 3484 return sprintf(buf, "Out of memory\n");
88a420e4
CL
3485
3486 /* Push back cpu slabs */
3487 flush_all(s);
3488
f64dc58c 3489 for_each_node_state(node, N_NORMAL_MEMORY) {
88a420e4
CL
3490 struct kmem_cache_node *n = get_node(s, node);
3491 unsigned long flags;
3492 struct page *page;
3493
9e86943b 3494 if (!atomic_long_read(&n->nr_slabs))
88a420e4
CL
3495 continue;
3496
3497 spin_lock_irqsave(&n->list_lock, flags);
3498 list_for_each_entry(page, &n->partial, lru)
3499 process_slab(&t, s, page, alloc);
3500 list_for_each_entry(page, &n->full, lru)
3501 process_slab(&t, s, page, alloc);
3502 spin_unlock_irqrestore(&n->list_lock, flags);
3503 }
3504
3505 for (i = 0; i < t.count; i++) {
45edfa58 3506 struct location *l = &t.loc[i];
88a420e4 3507
e374d483 3508 if (len > PAGE_SIZE - 100)
88a420e4 3509 break;
e374d483 3510 len += sprintf(buf + len, "%7ld ", l->count);
45edfa58
CL
3511
3512 if (l->addr)
e374d483 3513 len += sprint_symbol(buf + len, (unsigned long)l->addr);
88a420e4 3514 else
e374d483 3515 len += sprintf(buf + len, "<not-available>");
45edfa58
CL
3516
3517 if (l->sum_time != l->min_time) {
3518 unsigned long remainder;
3519
e374d483 3520 len += sprintf(buf + len, " age=%ld/%ld/%ld",
45edfa58
CL
3521 l->min_time,
3522 div_long_long_rem(l->sum_time, l->count, &remainder),
3523 l->max_time);
3524 } else
e374d483 3525 len += sprintf(buf + len, " age=%ld",
45edfa58
CL
3526 l->min_time);
3527
3528 if (l->min_pid != l->max_pid)
e374d483 3529 len += sprintf(buf + len, " pid=%ld-%ld",
45edfa58
CL
3530 l->min_pid, l->max_pid);
3531 else
e374d483 3532 len += sprintf(buf + len, " pid=%ld",
45edfa58
CL
3533 l->min_pid);
3534
84966343 3535 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
e374d483
HH
3536 len < PAGE_SIZE - 60) {
3537 len += sprintf(buf + len, " cpus=");
3538 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
45edfa58
CL
3539 l->cpus);
3540 }
3541
84966343 3542 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
e374d483
HH
3543 len < PAGE_SIZE - 60) {
3544 len += sprintf(buf + len, " nodes=");
3545 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
45edfa58
CL
3546 l->nodes);
3547 }
3548
e374d483 3549 len += sprintf(buf + len, "\n");
88a420e4
CL
3550 }
3551
3552 free_loc_track(&t);
3553 if (!t.count)
e374d483
HH
3554 len += sprintf(buf, "No data\n");
3555 return len;
88a420e4
CL
3556}
3557
81819f0f
CL
3558enum slab_stat_type {
3559 SL_FULL,
3560 SL_PARTIAL,
3561 SL_CPU,
3562 SL_OBJECTS
3563};
3564
3565#define SO_FULL (1 << SL_FULL)
3566#define SO_PARTIAL (1 << SL_PARTIAL)
3567#define SO_CPU (1 << SL_CPU)
3568#define SO_OBJECTS (1 << SL_OBJECTS)
3569
62e5c4b4
CG
3570static ssize_t show_slab_objects(struct kmem_cache *s,
3571 char *buf, unsigned long flags)
81819f0f
CL
3572{
3573 unsigned long total = 0;
3574 int cpu;
3575 int node;
3576 int x;
3577 unsigned long *nodes;
3578 unsigned long *per_cpu;
3579
3580 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
62e5c4b4
CG
3581 if (!nodes)
3582 return -ENOMEM;
81819f0f
CL
3583 per_cpu = nodes + nr_node_ids;
3584
3585 for_each_possible_cpu(cpu) {
dfb4f096
CL
3586 struct page *page;
3587 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
81819f0f 3588
dfb4f096
CL
3589 if (!c)
3590 continue;
3591
3592 page = c->page;
ee3c72a1
CL
3593 node = c->node;
3594 if (node < 0)
3595 continue;
81819f0f 3596 if (page) {
81819f0f 3597 if (flags & SO_CPU) {
81819f0f
CL
3598 if (flags & SO_OBJECTS)
3599 x = page->inuse;
3600 else
3601 x = 1;
3602 total += x;
ee3c72a1 3603 nodes[node] += x;
81819f0f 3604 }
ee3c72a1 3605 per_cpu[node]++;
81819f0f
CL
3606 }
3607 }
3608
f64dc58c 3609 for_each_node_state(node, N_NORMAL_MEMORY) {
81819f0f
CL
3610 struct kmem_cache_node *n = get_node(s, node);
3611
3612 if (flags & SO_PARTIAL) {
3613 if (flags & SO_OBJECTS)
3614 x = count_partial(n);
3615 else
3616 x = n->nr_partial;
3617 total += x;
3618 nodes[node] += x;
3619 }
3620
3621 if (flags & SO_FULL) {
9e86943b 3622 int full_slabs = atomic_long_read(&n->nr_slabs)
81819f0f
CL
3623 - per_cpu[node]
3624 - n->nr_partial;
3625
3626 if (flags & SO_OBJECTS)
3627 x = full_slabs * s->objects;
3628 else
3629 x = full_slabs;
3630 total += x;
3631 nodes[node] += x;
3632 }
3633 }
3634
3635 x = sprintf(buf, "%lu", total);
3636#ifdef CONFIG_NUMA
f64dc58c 3637 for_each_node_state(node, N_NORMAL_MEMORY)
81819f0f
CL
3638 if (nodes[node])
3639 x += sprintf(buf + x, " N%d=%lu",
3640 node, nodes[node]);
3641#endif
3642 kfree(nodes);
3643 return x + sprintf(buf + x, "\n");
3644}
3645
3646static int any_slab_objects(struct kmem_cache *s)
3647{
3648 int node;
3649 int cpu;
3650
dfb4f096
CL
3651 for_each_possible_cpu(cpu) {
3652 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3653
3654 if (c && c->page)
81819f0f 3655 return 1;
dfb4f096 3656 }
81819f0f 3657
dfb4f096 3658 for_each_online_node(node) {
81819f0f
CL
3659 struct kmem_cache_node *n = get_node(s, node);
3660
dfb4f096
CL
3661 if (!n)
3662 continue;
3663
9e86943b 3664 if (n->nr_partial || atomic_long_read(&n->nr_slabs))
81819f0f
CL
3665 return 1;
3666 }
3667 return 0;
3668}
3669
3670#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3671#define to_slab(n) container_of(n, struct kmem_cache, kobj);
3672
3673struct slab_attribute {
3674 struct attribute attr;
3675 ssize_t (*show)(struct kmem_cache *s, char *buf);
3676 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3677};
3678
3679#define SLAB_ATTR_RO(_name) \
3680 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3681
3682#define SLAB_ATTR(_name) \
3683 static struct slab_attribute _name##_attr = \
3684 __ATTR(_name, 0644, _name##_show, _name##_store)
3685
81819f0f
CL
3686static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3687{
3688 return sprintf(buf, "%d\n", s->size);
3689}
3690SLAB_ATTR_RO(slab_size);
3691
3692static ssize_t align_show(struct kmem_cache *s, char *buf)
3693{
3694 return sprintf(buf, "%d\n", s->align);
3695}
3696SLAB_ATTR_RO(align);
3697
3698static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3699{
3700 return sprintf(buf, "%d\n", s->objsize);
3701}
3702SLAB_ATTR_RO(object_size);
3703
3704static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3705{
3706 return sprintf(buf, "%d\n", s->objects);
3707}
3708SLAB_ATTR_RO(objs_per_slab);
3709
3710static ssize_t order_show(struct kmem_cache *s, char *buf)
3711{
3712 return sprintf(buf, "%d\n", s->order);
3713}
3714SLAB_ATTR_RO(order);
3715
3716static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3717{
3718 if (s->ctor) {
3719 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3720
3721 return n + sprintf(buf + n, "\n");
3722 }
3723 return 0;
3724}
3725SLAB_ATTR_RO(ctor);
3726
81819f0f
CL
3727static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3728{
3729 return sprintf(buf, "%d\n", s->refcount - 1);
3730}
3731SLAB_ATTR_RO(aliases);
3732
3733static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3734{
d9acf4b7 3735 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
81819f0f
CL
3736}
3737SLAB_ATTR_RO(slabs);
3738
3739static ssize_t partial_show(struct kmem_cache *s, char *buf)
3740{
d9acf4b7 3741 return show_slab_objects(s, buf, SO_PARTIAL);
81819f0f
CL
3742}
3743SLAB_ATTR_RO(partial);
3744
3745static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3746{
d9acf4b7 3747 return show_slab_objects(s, buf, SO_CPU);
81819f0f
CL
3748}
3749SLAB_ATTR_RO(cpu_slabs);
3750
3751static ssize_t objects_show(struct kmem_cache *s, char *buf)
3752{
d9acf4b7 3753 return show_slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
81819f0f
CL
3754}
3755SLAB_ATTR_RO(objects);
3756
3757static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3758{
3759 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3760}
3761
3762static ssize_t sanity_checks_store(struct kmem_cache *s,
3763 const char *buf, size_t length)
3764{
3765 s->flags &= ~SLAB_DEBUG_FREE;
3766 if (buf[0] == '1')
3767 s->flags |= SLAB_DEBUG_FREE;
3768 return length;
3769}
3770SLAB_ATTR(sanity_checks);
3771
3772static ssize_t trace_show(struct kmem_cache *s, char *buf)
3773{
3774 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3775}
3776
3777static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3778 size_t length)
3779{
3780 s->flags &= ~SLAB_TRACE;
3781 if (buf[0] == '1')
3782 s->flags |= SLAB_TRACE;
3783 return length;
3784}
3785SLAB_ATTR(trace);
3786
3787static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3788{
3789 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3790}
3791
3792static ssize_t reclaim_account_store(struct kmem_cache *s,
3793 const char *buf, size_t length)
3794{
3795 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3796 if (buf[0] == '1')
3797 s->flags |= SLAB_RECLAIM_ACCOUNT;
3798 return length;
3799}
3800SLAB_ATTR(reclaim_account);
3801
3802static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3803{
5af60839 3804 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
81819f0f
CL
3805}
3806SLAB_ATTR_RO(hwcache_align);
3807
3808#ifdef CONFIG_ZONE_DMA
3809static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3810{
3811 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3812}
3813SLAB_ATTR_RO(cache_dma);
3814#endif
3815
3816static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3817{
3818 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3819}
3820SLAB_ATTR_RO(destroy_by_rcu);
3821
3822static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3823{
3824 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3825}
3826
3827static ssize_t red_zone_store(struct kmem_cache *s,
3828 const char *buf, size_t length)
3829{
3830 if (any_slab_objects(s))
3831 return -EBUSY;
3832
3833 s->flags &= ~SLAB_RED_ZONE;
3834 if (buf[0] == '1')
3835 s->flags |= SLAB_RED_ZONE;
3836 calculate_sizes(s);
3837 return length;
3838}
3839SLAB_ATTR(red_zone);
3840
3841static ssize_t poison_show(struct kmem_cache *s, char *buf)
3842{
3843 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3844}
3845
3846static ssize_t poison_store(struct kmem_cache *s,
3847 const char *buf, size_t length)
3848{
3849 if (any_slab_objects(s))
3850 return -EBUSY;
3851
3852 s->flags &= ~SLAB_POISON;
3853 if (buf[0] == '1')
3854 s->flags |= SLAB_POISON;
3855 calculate_sizes(s);
3856 return length;
3857}
3858SLAB_ATTR(poison);
3859
3860static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3861{
3862 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3863}
3864
3865static ssize_t store_user_store(struct kmem_cache *s,
3866 const char *buf, size_t length)
3867{
3868 if (any_slab_objects(s))
3869 return -EBUSY;
3870
3871 s->flags &= ~SLAB_STORE_USER;
3872 if (buf[0] == '1')
3873 s->flags |= SLAB_STORE_USER;
3874 calculate_sizes(s);
3875 return length;
3876}
3877SLAB_ATTR(store_user);
3878
53e15af0
CL
3879static ssize_t validate_show(struct kmem_cache *s, char *buf)
3880{
3881 return 0;
3882}
3883
3884static ssize_t validate_store(struct kmem_cache *s,
3885 const char *buf, size_t length)
3886{
434e245d
CL
3887 int ret = -EINVAL;
3888
3889 if (buf[0] == '1') {
3890 ret = validate_slab_cache(s);
3891 if (ret >= 0)
3892 ret = length;
3893 }
3894 return ret;
53e15af0
CL
3895}
3896SLAB_ATTR(validate);
3897
2086d26a
CL
3898static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3899{
3900 return 0;
3901}
3902
3903static ssize_t shrink_store(struct kmem_cache *s,
3904 const char *buf, size_t length)
3905{
3906 if (buf[0] == '1') {
3907 int rc = kmem_cache_shrink(s);
3908
3909 if (rc)
3910 return rc;
3911 } else
3912 return -EINVAL;
3913 return length;
3914}
3915SLAB_ATTR(shrink);
3916
88a420e4
CL
3917static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3918{
3919 if (!(s->flags & SLAB_STORE_USER))
3920 return -ENOSYS;
3921 return list_locations(s, buf, TRACK_ALLOC);
3922}
3923SLAB_ATTR_RO(alloc_calls);
3924
3925static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3926{
3927 if (!(s->flags & SLAB_STORE_USER))
3928 return -ENOSYS;
3929 return list_locations(s, buf, TRACK_FREE);
3930}
3931SLAB_ATTR_RO(free_calls);
3932
81819f0f 3933#ifdef CONFIG_NUMA
9824601e 3934static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
81819f0f 3935{
9824601e 3936 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
81819f0f
CL
3937}
3938
9824601e 3939static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
81819f0f
CL
3940 const char *buf, size_t length)
3941{
3942 int n = simple_strtoul(buf, NULL, 10);
3943
3944 if (n < 100)
9824601e 3945 s->remote_node_defrag_ratio = n * 10;
81819f0f
CL
3946 return length;
3947}
9824601e 3948SLAB_ATTR(remote_node_defrag_ratio);
81819f0f
CL
3949#endif
3950
8ff12cfc 3951#ifdef CONFIG_SLUB_STATS
8ff12cfc
CL
3952static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
3953{
3954 unsigned long sum = 0;
3955 int cpu;
3956 int len;
3957 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
3958
3959 if (!data)
3960 return -ENOMEM;
3961
3962 for_each_online_cpu(cpu) {
3963 unsigned x = get_cpu_slab(s, cpu)->stat[si];
3964
3965 data[cpu] = x;
3966 sum += x;
3967 }
3968
3969 len = sprintf(buf, "%lu", sum);
3970
3971 for_each_online_cpu(cpu) {
3972 if (data[cpu] && len < PAGE_SIZE - 20)
3973 len += sprintf(buf + len, " c%d=%u", cpu, data[cpu]);
3974 }
3975 kfree(data);
3976 return len + sprintf(buf + len, "\n");
3977}
3978
3979#define STAT_ATTR(si, text) \
3980static ssize_t text##_show(struct kmem_cache *s, char *buf) \
3981{ \
3982 return show_stat(s, buf, si); \
3983} \
3984SLAB_ATTR_RO(text); \
3985
3986STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
3987STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
3988STAT_ATTR(FREE_FASTPATH, free_fastpath);
3989STAT_ATTR(FREE_SLOWPATH, free_slowpath);
3990STAT_ATTR(FREE_FROZEN, free_frozen);
3991STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
3992STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
3993STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
3994STAT_ATTR(ALLOC_SLAB, alloc_slab);
3995STAT_ATTR(ALLOC_REFILL, alloc_refill);
3996STAT_ATTR(FREE_SLAB, free_slab);
3997STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
3998STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
3999STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4000STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4001STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4002STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4003
4004#endif
4005
06428780 4006static struct attribute *slab_attrs[] = {
81819f0f
CL
4007 &slab_size_attr.attr,
4008 &object_size_attr.attr,
4009 &objs_per_slab_attr.attr,
4010 &order_attr.attr,
4011 &objects_attr.attr,
4012 &slabs_attr.attr,
4013 &partial_attr.attr,
4014 &cpu_slabs_attr.attr,
4015 &ctor_attr.attr,
81819f0f
CL
4016 &aliases_attr.attr,
4017 &align_attr.attr,
4018 &sanity_checks_attr.attr,
4019 &trace_attr.attr,
4020 &hwcache_align_attr.attr,
4021 &reclaim_account_attr.attr,
4022 &destroy_by_rcu_attr.attr,
4023 &red_zone_attr.attr,
4024 &poison_attr.attr,
4025 &store_user_attr.attr,
53e15af0 4026 &validate_attr.attr,
2086d26a 4027 &shrink_attr.attr,
88a420e4
CL
4028 &alloc_calls_attr.attr,
4029 &free_calls_attr.attr,
81819f0f
CL
4030#ifdef CONFIG_ZONE_DMA
4031 &cache_dma_attr.attr,
4032#endif
4033#ifdef CONFIG_NUMA
9824601e 4034 &remote_node_defrag_ratio_attr.attr,
8ff12cfc
CL
4035#endif
4036#ifdef CONFIG_SLUB_STATS
4037 &alloc_fastpath_attr.attr,
4038 &alloc_slowpath_attr.attr,
4039 &free_fastpath_attr.attr,
4040 &free_slowpath_attr.attr,
4041 &free_frozen_attr.attr,
4042 &free_add_partial_attr.attr,
4043 &free_remove_partial_attr.attr,
4044 &alloc_from_partial_attr.attr,
4045 &alloc_slab_attr.attr,
4046 &alloc_refill_attr.attr,
4047 &free_slab_attr.attr,
4048 &cpuslab_flush_attr.attr,
4049 &deactivate_full_attr.attr,
4050 &deactivate_empty_attr.attr,
4051 &deactivate_to_head_attr.attr,
4052 &deactivate_to_tail_attr.attr,
4053 &deactivate_remote_frees_attr.attr,
81819f0f
CL
4054#endif
4055 NULL
4056};
4057
4058static struct attribute_group slab_attr_group = {
4059 .attrs = slab_attrs,
4060};
4061
4062static ssize_t slab_attr_show(struct kobject *kobj,
4063 struct attribute *attr,
4064 char *buf)
4065{
4066 struct slab_attribute *attribute;
4067 struct kmem_cache *s;
4068 int err;
4069
4070 attribute = to_slab_attr(attr);
4071 s = to_slab(kobj);
4072
4073 if (!attribute->show)
4074 return -EIO;
4075
4076 err = attribute->show(s, buf);
4077
4078 return err;
4079}
4080
4081static ssize_t slab_attr_store(struct kobject *kobj,
4082 struct attribute *attr,
4083 const char *buf, size_t len)
4084{
4085 struct slab_attribute *attribute;
4086 struct kmem_cache *s;
4087 int err;
4088
4089 attribute = to_slab_attr(attr);
4090 s = to_slab(kobj);
4091
4092 if (!attribute->store)
4093 return -EIO;
4094
4095 err = attribute->store(s, buf, len);
4096
4097 return err;
4098}
4099
151c602f
CL
4100static void kmem_cache_release(struct kobject *kobj)
4101{
4102 struct kmem_cache *s = to_slab(kobj);
4103
4104 kfree(s);
4105}
4106
81819f0f
CL
4107static struct sysfs_ops slab_sysfs_ops = {
4108 .show = slab_attr_show,
4109 .store = slab_attr_store,
4110};
4111
4112static struct kobj_type slab_ktype = {
4113 .sysfs_ops = &slab_sysfs_ops,
151c602f 4114 .release = kmem_cache_release
81819f0f
CL
4115};
4116
4117static int uevent_filter(struct kset *kset, struct kobject *kobj)
4118{
4119 struct kobj_type *ktype = get_ktype(kobj);
4120
4121 if (ktype == &slab_ktype)
4122 return 1;
4123 return 0;
4124}
4125
4126static struct kset_uevent_ops slab_uevent_ops = {
4127 .filter = uevent_filter,
4128};
4129
27c3a314 4130static struct kset *slab_kset;
81819f0f
CL
4131
4132#define ID_STR_LENGTH 64
4133
4134/* Create a unique string id for a slab cache:
6446faa2
CL
4135 *
4136 * Format :[flags-]size
81819f0f
CL
4137 */
4138static char *create_unique_id(struct kmem_cache *s)
4139{
4140 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4141 char *p = name;
4142
4143 BUG_ON(!name);
4144
4145 *p++ = ':';
4146 /*
4147 * First flags affecting slabcache operations. We will only
4148 * get here for aliasable slabs so we do not need to support
4149 * too many flags. The flags here must cover all flags that
4150 * are matched during merging to guarantee that the id is
4151 * unique.
4152 */
4153 if (s->flags & SLAB_CACHE_DMA)
4154 *p++ = 'd';
4155 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4156 *p++ = 'a';
4157 if (s->flags & SLAB_DEBUG_FREE)
4158 *p++ = 'F';
4159 if (p != name + 1)
4160 *p++ = '-';
4161 p += sprintf(p, "%07d", s->size);
4162 BUG_ON(p > name + ID_STR_LENGTH - 1);
4163 return name;
4164}
4165
4166static int sysfs_slab_add(struct kmem_cache *s)
4167{
4168 int err;
4169 const char *name;
4170 int unmergeable;
4171
4172 if (slab_state < SYSFS)
4173 /* Defer until later */
4174 return 0;
4175
4176 unmergeable = slab_unmergeable(s);
4177 if (unmergeable) {
4178 /*
4179 * Slabcache can never be merged so we can use the name proper.
4180 * This is typically the case for debug situations. In that
4181 * case we can catch duplicate names easily.
4182 */
27c3a314 4183 sysfs_remove_link(&slab_kset->kobj, s->name);
81819f0f
CL
4184 name = s->name;
4185 } else {
4186 /*
4187 * Create a unique name for the slab as a target
4188 * for the symlinks.
4189 */
4190 name = create_unique_id(s);
4191 }
4192
27c3a314 4193 s->kobj.kset = slab_kset;
1eada11c
GKH
4194 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4195 if (err) {
4196 kobject_put(&s->kobj);
81819f0f 4197 return err;
1eada11c 4198 }
81819f0f
CL
4199
4200 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4201 if (err)
4202 return err;
4203 kobject_uevent(&s->kobj, KOBJ_ADD);
4204 if (!unmergeable) {
4205 /* Setup first alias */
4206 sysfs_slab_alias(s, s->name);
4207 kfree(name);
4208 }
4209 return 0;
4210}
4211
4212static void sysfs_slab_remove(struct kmem_cache *s)
4213{
4214 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4215 kobject_del(&s->kobj);
151c602f 4216 kobject_put(&s->kobj);
81819f0f
CL
4217}
4218
4219/*
4220 * Need to buffer aliases during bootup until sysfs becomes
4221 * available lest we loose that information.
4222 */
4223struct saved_alias {
4224 struct kmem_cache *s;
4225 const char *name;
4226 struct saved_alias *next;
4227};
4228
5af328a5 4229static struct saved_alias *alias_list;
81819f0f
CL
4230
4231static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4232{
4233 struct saved_alias *al;
4234
4235 if (slab_state == SYSFS) {
4236 /*
4237 * If we have a leftover link then remove it.
4238 */
27c3a314
GKH
4239 sysfs_remove_link(&slab_kset->kobj, name);
4240 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
81819f0f
CL
4241 }
4242
4243 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4244 if (!al)
4245 return -ENOMEM;
4246
4247 al->s = s;
4248 al->name = name;
4249 al->next = alias_list;
4250 alias_list = al;
4251 return 0;
4252}
4253
4254static int __init slab_sysfs_init(void)
4255{
5b95a4ac 4256 struct kmem_cache *s;
81819f0f
CL
4257 int err;
4258
0ff21e46 4259 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
27c3a314 4260 if (!slab_kset) {
81819f0f
CL
4261 printk(KERN_ERR "Cannot register slab subsystem.\n");
4262 return -ENOSYS;
4263 }
4264
26a7bd03
CL
4265 slab_state = SYSFS;
4266
5b95a4ac 4267 list_for_each_entry(s, &slab_caches, list) {
26a7bd03 4268 err = sysfs_slab_add(s);
5d540fb7
CL
4269 if (err)
4270 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4271 " to sysfs\n", s->name);
26a7bd03 4272 }
81819f0f
CL
4273
4274 while (alias_list) {
4275 struct saved_alias *al = alias_list;
4276
4277 alias_list = alias_list->next;
4278 err = sysfs_slab_alias(al->s, al->name);
5d540fb7
CL
4279 if (err)
4280 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4281 " %s to sysfs\n", s->name);
81819f0f
CL
4282 kfree(al);
4283 }
4284
4285 resiliency_test();
4286 return 0;
4287}
4288
4289__initcall(slab_sysfs_init);
81819f0f 4290#endif
57ed3eda
PE
4291
4292/*
4293 * The /proc/slabinfo ABI
4294 */
158a9624
LT
4295#ifdef CONFIG_SLABINFO
4296
4297ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4298 size_t count, loff_t *ppos)
4299{
4300 return -EINVAL;
4301}
4302
57ed3eda
PE
4303
4304static void print_slabinfo_header(struct seq_file *m)
4305{
4306 seq_puts(m, "slabinfo - version: 2.1\n");
4307 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4308 "<objperslab> <pagesperslab>");
4309 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4310 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4311 seq_putc(m, '\n');
4312}
4313
4314static void *s_start(struct seq_file *m, loff_t *pos)
4315{
4316 loff_t n = *pos;
4317
4318 down_read(&slub_lock);
4319 if (!n)
4320 print_slabinfo_header(m);
4321
4322 return seq_list_start(&slab_caches, *pos);
4323}
4324
4325static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4326{
4327 return seq_list_next(p, &slab_caches, pos);
4328}
4329
4330static void s_stop(struct seq_file *m, void *p)
4331{
4332 up_read(&slub_lock);
4333}
4334
4335static int s_show(struct seq_file *m, void *p)
4336{
4337 unsigned long nr_partials = 0;
4338 unsigned long nr_slabs = 0;
4339 unsigned long nr_inuse = 0;
4340 unsigned long nr_objs;
4341 struct kmem_cache *s;
4342 int node;
4343
4344 s = list_entry(p, struct kmem_cache, list);
4345
4346 for_each_online_node(node) {
4347 struct kmem_cache_node *n = get_node(s, node);
4348
4349 if (!n)
4350 continue;
4351
4352 nr_partials += n->nr_partial;
4353 nr_slabs += atomic_long_read(&n->nr_slabs);
4354 nr_inuse += count_partial(n);
4355 }
4356
4357 nr_objs = nr_slabs * s->objects;
4358 nr_inuse += (nr_slabs - nr_partials) * s->objects;
4359
4360 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4361 nr_objs, s->size, s->objects, (1 << s->order));
4362 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4363 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4364 0UL);
4365 seq_putc(m, '\n');
4366 return 0;
4367}
4368
4369const struct seq_operations slabinfo_op = {
4370 .start = s_start,
4371 .next = s_next,
4372 .stop = s_stop,
4373 .show = s_show,
4374};
4375
158a9624 4376#endif /* CONFIG_SLABINFO */
This page took 0.374294 seconds and 5 git commands to generate.