blob: 03d26f7142c7755071f5bac0133dd2dbd0e2820c [file] [log] [blame]
Christoph Lameter81819f02007-05-06 14:49:36 -07001/*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
4 *
5 * The allocator synchronizes using per slab locks 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>
23
24/*
25 * Lock order:
26 * 1. slab_lock(page)
27 * 2. slab->list_lock
28 *
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
35 *
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
41 *
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
46 * the list lock.
47 *
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
60 *
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
65 *
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
68 *
69 * Slabs with free elements are kept on a partial list.
70 * There is no list for full slabs. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * Otherwise there is no need to track full slabs unless we have to
73 * track full slabs for debugging purposes.
74 *
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
78 *
79 * Overloading of page flags that are otherwise used for LRU management.
80 *
81 * PageActive The slab is used as a cpu cache. Allocations
82 * may be performed from the slab. The slab is not
83 * on any slab list and cannot be moved onto one.
84 *
85 * PageError Slab requires special handling due to debug
86 * options set. This moves slab handling out of
87 * the fast path.
88 */
89
90/*
91 * Issues still to be resolved:
92 *
93 * - The per cpu array is updated for each new slab and and is a remote
94 * cacheline for most nodes. This could become a bouncing cacheline given
95 * enough frequent updates. There are 16 pointers in a cacheline.so at
96 * max 16 cpus could compete. Likely okay.
97 *
98 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
99 *
Christoph Lameter81819f02007-05-06 14:49:36 -0700100 * - Variable sizing of the per node arrays
101 */
102
103/* Enable to test recovery from slab corruption on boot */
104#undef SLUB_RESILIENCY_TEST
105
106#if PAGE_SHIFT <= 12
107
108/*
109 * Small page size. Make sure that we do not fragment memory
110 */
111#define DEFAULT_MAX_ORDER 1
112#define DEFAULT_MIN_OBJECTS 4
113
114#else
115
116/*
117 * Large page machines are customarily able to handle larger
118 * page orders.
119 */
120#define DEFAULT_MAX_ORDER 2
121#define DEFAULT_MIN_OBJECTS 8
122
123#endif
124
125/*
Christoph Lameter2086d262007-05-06 14:49:46 -0700126 * Mininum number of partial slabs. These will be left on the partial
127 * lists even if they are empty. kmem_cache_shrink may reclaim them.
128 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700129#define MIN_PARTIAL 2
130
Christoph Lameter2086d262007-05-06 14:49:46 -0700131/*
132 * Maximum number of desirable partial slabs.
133 * The existence of more partial slabs makes kmem_cache_shrink
134 * sort the partial list by the number of objects in the.
135 */
136#define MAX_PARTIAL 10
137
Christoph Lameter81819f02007-05-06 14:49:36 -0700138#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
139 SLAB_POISON | SLAB_STORE_USER)
140/*
141 * Set of flags that will prevent slab merging
142 */
143#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
144 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
145
146#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
147 SLAB_CACHE_DMA)
148
149#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700150#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700151#endif
152
153#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700154#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700155#endif
156
157/* Internal SLUB flags */
158#define __OBJECT_POISON 0x80000000 /* Poison object */
159
Christoph Lameter65c02d42007-05-09 02:32:35 -0700160/* Not all arches define cache_line_size */
161#ifndef cache_line_size
162#define cache_line_size() L1_CACHE_BYTES
163#endif
164
Christoph Lameter81819f02007-05-06 14:49:36 -0700165static int kmem_size = sizeof(struct kmem_cache);
166
167#ifdef CONFIG_SMP
168static struct notifier_block slab_notifier;
169#endif
170
171static enum {
172 DOWN, /* No slab functionality available */
173 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
174 UP, /* Everything works */
175 SYSFS /* Sysfs up */
176} slab_state = DOWN;
177
178/* A list of all slab caches on the system */
179static DECLARE_RWSEM(slub_lock);
180LIST_HEAD(slab_caches);
181
182#ifdef CONFIG_SYSFS
183static int sysfs_slab_add(struct kmem_cache *);
184static int sysfs_slab_alias(struct kmem_cache *, const char *);
185static void sysfs_slab_remove(struct kmem_cache *);
186#else
187static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
188static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
189static void sysfs_slab_remove(struct kmem_cache *s) {}
190#endif
191
192/********************************************************************
193 * Core slab cache functions
194 *******************************************************************/
195
196int slab_is_available(void)
197{
198 return slab_state >= UP;
199}
200
201static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
202{
203#ifdef CONFIG_NUMA
204 return s->node[node];
205#else
206 return &s->local_node;
207#endif
208}
209
210/*
211 * Object debugging
212 */
213static void print_section(char *text, u8 *addr, unsigned int length)
214{
215 int i, offset;
216 int newline = 1;
217 char ascii[17];
218
219 ascii[16] = 0;
220
221 for (i = 0; i < length; i++) {
222 if (newline) {
223 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
224 newline = 0;
225 }
226 printk(" %02x", addr[i]);
227 offset = i % 16;
228 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
229 if (offset == 15) {
230 printk(" %s\n",ascii);
231 newline = 1;
232 }
233 }
234 if (!newline) {
235 i %= 16;
236 while (i < 16) {
237 printk(" ");
238 ascii[i] = ' ';
239 i++;
240 }
241 printk(" %s\n", ascii);
242 }
243}
244
245/*
246 * Slow version of get and set free pointer.
247 *
248 * This requires touching the cache lines of kmem_cache.
249 * The offset can also be obtained from the page. In that
250 * case it is in the cacheline that we already need to touch.
251 */
252static void *get_freepointer(struct kmem_cache *s, void *object)
253{
254 return *(void **)(object + s->offset);
255}
256
257static void set_freepointer(struct kmem_cache *s, void *object, void *fp)
258{
259 *(void **)(object + s->offset) = fp;
260}
261
262/*
263 * Tracking user of a slab.
264 */
265struct track {
266 void *addr; /* Called from address */
267 int cpu; /* Was running on cpu */
268 int pid; /* Pid context */
269 unsigned long when; /* When did the operation occur */
270};
271
272enum track_item { TRACK_ALLOC, TRACK_FREE };
273
274static struct track *get_track(struct kmem_cache *s, void *object,
275 enum track_item alloc)
276{
277 struct track *p;
278
279 if (s->offset)
280 p = object + s->offset + sizeof(void *);
281 else
282 p = object + s->inuse;
283
284 return p + alloc;
285}
286
287static void set_track(struct kmem_cache *s, void *object,
288 enum track_item alloc, void *addr)
289{
290 struct track *p;
291
292 if (s->offset)
293 p = object + s->offset + sizeof(void *);
294 else
295 p = object + s->inuse;
296
297 p += alloc;
298 if (addr) {
299 p->addr = addr;
300 p->cpu = smp_processor_id();
301 p->pid = current ? current->pid : -1;
302 p->when = jiffies;
303 } else
304 memset(p, 0, sizeof(struct track));
305}
306
Christoph Lameter81819f02007-05-06 14:49:36 -0700307static void init_tracking(struct kmem_cache *s, void *object)
308{
309 if (s->flags & SLAB_STORE_USER) {
310 set_track(s, object, TRACK_FREE, NULL);
311 set_track(s, object, TRACK_ALLOC, NULL);
312 }
313}
314
315static void print_track(const char *s, struct track *t)
316{
317 if (!t->addr)
318 return;
319
320 printk(KERN_ERR "%s: ", s);
321 __print_symbol("%s", (unsigned long)t->addr);
322 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
323}
324
325static void print_trailer(struct kmem_cache *s, u8 *p)
326{
327 unsigned int off; /* Offset of last byte */
328
329 if (s->flags & SLAB_RED_ZONE)
330 print_section("Redzone", p + s->objsize,
331 s->inuse - s->objsize);
332
333 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
334 p + s->offset,
335 get_freepointer(s, p));
336
337 if (s->offset)
338 off = s->offset + sizeof(void *);
339 else
340 off = s->inuse;
341
342 if (s->flags & SLAB_STORE_USER) {
343 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
344 print_track("Last free ", get_track(s, p, TRACK_FREE));
345 off += 2 * sizeof(struct track);
346 }
347
348 if (off != s->size)
349 /* Beginning of the filler is the free pointer */
350 print_section("Filler", p + off, s->size - off);
351}
352
353static void object_err(struct kmem_cache *s, struct page *page,
354 u8 *object, char *reason)
355{
356 u8 *addr = page_address(page);
357
358 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
359 s->name, reason, object, page);
360 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
361 object - addr, page->flags, page->inuse, page->freelist);
362 if (object > addr + 16)
363 print_section("Bytes b4", object - 16, 16);
364 print_section("Object", object, min(s->objsize, 128));
365 print_trailer(s, object);
366 dump_stack();
367}
368
369static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
370{
371 va_list args;
372 char buf[100];
373
374 va_start(args, reason);
375 vsnprintf(buf, sizeof(buf), reason, args);
376 va_end(args);
377 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
378 page);
379 dump_stack();
380}
381
382static void init_object(struct kmem_cache *s, void *object, int active)
383{
384 u8 *p = object;
385
386 if (s->flags & __OBJECT_POISON) {
387 memset(p, POISON_FREE, s->objsize - 1);
388 p[s->objsize -1] = POISON_END;
389 }
390
391 if (s->flags & SLAB_RED_ZONE)
392 memset(p + s->objsize,
393 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
394 s->inuse - s->objsize);
395}
396
397static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
398{
399 while (bytes) {
400 if (*start != (u8)value)
401 return 0;
402 start++;
403 bytes--;
404 }
405 return 1;
406}
407
Christoph Lameterabcd08a2007-05-09 02:32:37 -0700408static inline int check_valid_pointer(struct kmem_cache *s,
409 struct page *page, const void *object)
Christoph Lameter81819f02007-05-06 14:49:36 -0700410{
411 void *base;
412
413 if (!object)
414 return 1;
415
416 base = page_address(page);
417 if (object < base || object >= base + s->objects * s->size ||
418 (object - base) % s->size) {
419 return 0;
420 }
421
422 return 1;
423}
424
425/*
426 * Object layout:
427 *
428 * object address
429 * Bytes of the object to be managed.
430 * If the freepointer may overlay the object then the free
431 * pointer is the first word of the object.
432 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
433 * 0xa5 (POISON_END)
434 *
435 * object + s->objsize
436 * Padding to reach word boundary. This is also used for Redzoning.
437 * Padding is extended to word size if Redzoning is enabled
438 * and objsize == inuse.
439 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
440 * 0xcc (RED_ACTIVE) for objects in use.
441 *
442 * object + s->inuse
443 * A. Free pointer (if we cannot overwrite object on free)
444 * B. Tracking data for SLAB_STORE_USER
445 * C. Padding to reach required alignment boundary
446 * Padding is done using 0x5a (POISON_INUSE)
447 *
448 * object + s->size
449 *
450 * If slabcaches are merged then the objsize and inuse boundaries are to
451 * be ignored. And therefore no slab options that rely on these boundaries
452 * may be used with merged slabcaches.
453 */
454
455static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
456 void *from, void *to)
457{
Christoph Lameter70d71222007-05-06 14:49:47 -0700458 printk(KERN_ERR "@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700459 s->name, message, data, from, to - 1);
460 memset(from, data, to - from);
461}
462
463static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
464{
465 unsigned long off = s->inuse; /* The end of info */
466
467 if (s->offset)
468 /* Freepointer is placed after the object. */
469 off += sizeof(void *);
470
471 if (s->flags & SLAB_STORE_USER)
472 /* We also have user information there */
473 off += 2 * sizeof(struct track);
474
475 if (s->size == off)
476 return 1;
477
478 if (check_bytes(p + off, POISON_INUSE, s->size - off))
479 return 1;
480
481 object_err(s, page, p, "Object padding check fails");
482
483 /*
484 * Restore padding
485 */
486 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
487 return 0;
488}
489
490static int slab_pad_check(struct kmem_cache *s, struct page *page)
491{
492 u8 *p;
493 int length, remainder;
494
495 if (!(s->flags & SLAB_POISON))
496 return 1;
497
498 p = page_address(page);
499 length = s->objects * s->size;
500 remainder = (PAGE_SIZE << s->order) - length;
501 if (!remainder)
502 return 1;
503
504 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700505 slab_err(s, page, "Padding check failed");
Christoph Lameter81819f02007-05-06 14:49:36 -0700506 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
507 p + length + remainder);
508 return 0;
509 }
510 return 1;
511}
512
513static int check_object(struct kmem_cache *s, struct page *page,
514 void *object, int active)
515{
516 u8 *p = object;
517 u8 *endobject = object + s->objsize;
518
519 if (s->flags & SLAB_RED_ZONE) {
520 unsigned int red =
521 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
522
523 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
524 object_err(s, page, object,
525 active ? "Redzone Active" : "Redzone Inactive");
526 restore_bytes(s, "redzone", red,
527 endobject, object + s->inuse);
528 return 0;
529 }
530 } else {
531 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
532 !check_bytes(endobject, POISON_INUSE,
533 s->inuse - s->objsize)) {
534 object_err(s, page, p, "Alignment padding check fails");
535 /*
536 * Fix it so that there will not be another report.
537 *
538 * Hmmm... We may be corrupting an object that now expects
539 * to be longer than allowed.
540 */
541 restore_bytes(s, "alignment padding", POISON_INUSE,
542 endobject, object + s->inuse);
543 }
544 }
545
546 if (s->flags & SLAB_POISON) {
547 if (!active && (s->flags & __OBJECT_POISON) &&
548 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
549 p[s->objsize - 1] != POISON_END)) {
550
551 object_err(s, page, p, "Poison check failed");
552 restore_bytes(s, "Poison", POISON_FREE,
553 p, p + s->objsize -1);
554 restore_bytes(s, "Poison", POISON_END,
555 p + s->objsize - 1, p + s->objsize);
556 return 0;
557 }
558 /*
559 * check_pad_bytes cleans up on its own.
560 */
561 check_pad_bytes(s, page, p);
562 }
563
564 if (!s->offset && active)
565 /*
566 * Object and freepointer overlap. Cannot check
567 * freepointer while object is allocated.
568 */
569 return 1;
570
571 /* Check free pointer validity */
572 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
573 object_err(s, page, p, "Freepointer corrupt");
574 /*
575 * No choice but to zap it and thus loose the remainder
576 * of the free objects in this slab. May cause
577 * another error because the object count maybe
578 * wrong now.
579 */
580 set_freepointer(s, p, NULL);
581 return 0;
582 }
583 return 1;
584}
585
586static int check_slab(struct kmem_cache *s, struct page *page)
587{
588 VM_BUG_ON(!irqs_disabled());
589
590 if (!PageSlab(page)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700591 slab_err(s, page, "Not a valid slab page flags=%lx "
592 "mapping=0x%p count=%d", page->flags, page->mapping,
Christoph Lameter81819f02007-05-06 14:49:36 -0700593 page_count(page));
594 return 0;
595 }
596 if (page->offset * sizeof(void *) != s->offset) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700597 slab_err(s, page, "Corrupted offset %lu flags=0x%lx "
598 "mapping=0x%p count=%d",
Christoph Lameter81819f02007-05-06 14:49:36 -0700599 (unsigned long)(page->offset * sizeof(void *)),
Christoph Lameter81819f02007-05-06 14:49:36 -0700600 page->flags,
601 page->mapping,
602 page_count(page));
Christoph Lameter81819f02007-05-06 14:49:36 -0700603 return 0;
604 }
605 if (page->inuse > s->objects) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700606 slab_err(s, page, "inuse %u > max %u @0x%p flags=%lx "
607 "mapping=0x%p count=%d",
608 s->name, page->inuse, s->objects, page->flags,
Christoph Lameter81819f02007-05-06 14:49:36 -0700609 page->mapping, page_count(page));
Christoph Lameter81819f02007-05-06 14:49:36 -0700610 return 0;
611 }
612 /* Slab_pad_check fixes things up after itself */
613 slab_pad_check(s, page);
614 return 1;
615}
616
617/*
618 * Determine if a certain object on a page is on the freelist and
619 * therefore free. Must hold the slab lock for cpu slabs to
620 * guarantee that the chains are consistent.
621 */
622static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
623{
624 int nr = 0;
625 void *fp = page->freelist;
626 void *object = NULL;
627
628 while (fp && nr <= s->objects) {
629 if (fp == search)
630 return 1;
631 if (!check_valid_pointer(s, page, fp)) {
632 if (object) {
633 object_err(s, page, object,
634 "Freechain corrupt");
635 set_freepointer(s, object, NULL);
636 break;
637 } else {
Christoph Lameter70d71222007-05-06 14:49:47 -0700638 slab_err(s, page, "Freepointer 0x%p corrupt",
639 fp);
Christoph Lameter81819f02007-05-06 14:49:36 -0700640 page->freelist = NULL;
641 page->inuse = s->objects;
Christoph Lameter70d71222007-05-06 14:49:47 -0700642 printk(KERN_ERR "@@@ SLUB %s: Freelist "
643 "cleared. Slab 0x%p\n",
644 s->name, page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700645 return 0;
646 }
647 break;
648 }
649 object = fp;
650 fp = get_freepointer(s, object);
651 nr++;
652 }
653
654 if (page->inuse != s->objects - nr) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700655 slab_err(s, page, "Wrong object count. Counter is %d but "
656 "counted were %d", s, page, page->inuse,
657 s->objects - nr);
Christoph Lameter81819f02007-05-06 14:49:36 -0700658 page->inuse = s->objects - nr;
Christoph Lameter70d71222007-05-06 14:49:47 -0700659 printk(KERN_ERR "@@@ SLUB %s: Object count adjusted. "
660 "Slab @0x%p\n", s->name, page);
Christoph Lameter81819f02007-05-06 14:49:36 -0700661 }
662 return search == NULL;
663}
664
Christoph Lameter643b1132007-05-06 14:49:42 -0700665/*
666 * Tracking of fully allocated slabs for debugging
667 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700668static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700669{
Christoph Lameter643b1132007-05-06 14:49:42 -0700670 spin_lock(&n->list_lock);
671 list_add(&page->lru, &n->full);
672 spin_unlock(&n->list_lock);
673}
674
675static void remove_full(struct kmem_cache *s, struct page *page)
676{
677 struct kmem_cache_node *n;
678
679 if (!(s->flags & SLAB_STORE_USER))
680 return;
681
682 n = get_node(s, page_to_nid(page));
683
684 spin_lock(&n->list_lock);
685 list_del(&page->lru);
686 spin_unlock(&n->list_lock);
687}
688
Christoph Lameter81819f02007-05-06 14:49:36 -0700689static int alloc_object_checks(struct kmem_cache *s, struct page *page,
690 void *object)
691{
692 if (!check_slab(s, page))
693 goto bad;
694
695 if (object && !on_freelist(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700696 slab_err(s, page, "Object 0x%p already allocated", object);
697 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700698 }
699
700 if (!check_valid_pointer(s, page, object)) {
701 object_err(s, page, object, "Freelist Pointer check fails");
Christoph Lameter70d71222007-05-06 14:49:47 -0700702 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700703 }
704
705 if (!object)
706 return 1;
707
708 if (!check_object(s, page, object, 0))
709 goto bad;
Christoph Lameter81819f02007-05-06 14:49:36 -0700710
Christoph Lameter81819f02007-05-06 14:49:36 -0700711 return 1;
Christoph Lameter81819f02007-05-06 14:49:36 -0700712bad:
713 if (PageSlab(page)) {
714 /*
715 * If this is a slab page then lets do the best we can
716 * to avoid issues in the future. Marking all objects
717 * as used avoids touching the remainder.
718 */
719 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
720 s->name, page);
721 page->inuse = s->objects;
722 page->freelist = NULL;
723 /* Fix up fields that may be corrupted */
724 page->offset = s->offset / sizeof(void *);
725 }
726 return 0;
727}
728
729static int free_object_checks(struct kmem_cache *s, struct page *page,
730 void *object)
731{
732 if (!check_slab(s, page))
733 goto fail;
734
735 if (!check_valid_pointer(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700736 slab_err(s, page, "Invalid object pointer 0x%p", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700737 goto fail;
738 }
739
740 if (on_freelist(s, page, object)) {
Christoph Lameter70d71222007-05-06 14:49:47 -0700741 slab_err(s, page, "Object 0x%p already free", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700742 goto fail;
743 }
744
745 if (!check_object(s, page, object, 1))
746 return 0;
747
748 if (unlikely(s != page->slab)) {
749 if (!PageSlab(page))
Christoph Lameter70d71222007-05-06 14:49:47 -0700750 slab_err(s, page, "Attempt to free object(0x%p) "
751 "outside of slab", object);
Christoph Lameter81819f02007-05-06 14:49:36 -0700752 else
Christoph Lameter70d71222007-05-06 14:49:47 -0700753 if (!page->slab) {
Christoph Lameter81819f02007-05-06 14:49:36 -0700754 printk(KERN_ERR
Christoph Lameter70d71222007-05-06 14:49:47 -0700755 "SLUB <none>: no slab for object 0x%p.\n",
Christoph Lameter81819f02007-05-06 14:49:36 -0700756 object);
Christoph Lameter70d71222007-05-06 14:49:47 -0700757 dump_stack();
758 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700759 else
Christoph Lameter70d71222007-05-06 14:49:47 -0700760 slab_err(s, page, "object at 0x%p belongs "
761 "to slab %s", object, page->slab->name);
Christoph Lameter81819f02007-05-06 14:49:36 -0700762 goto fail;
763 }
Christoph Lameter81819f02007-05-06 14:49:36 -0700764 return 1;
765fail:
Christoph Lameter81819f02007-05-06 14:49:36 -0700766 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
767 s->name, page, object);
768 return 0;
769}
770
771/*
772 * Slab allocation and freeing
773 */
774static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
775{
776 struct page * page;
777 int pages = 1 << s->order;
778
779 if (s->order)
780 flags |= __GFP_COMP;
781
782 if (s->flags & SLAB_CACHE_DMA)
783 flags |= SLUB_DMA;
784
785 if (node == -1)
786 page = alloc_pages(flags, s->order);
787 else
788 page = alloc_pages_node(node, flags, s->order);
789
790 if (!page)
791 return NULL;
792
793 mod_zone_page_state(page_zone(page),
794 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
795 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
796 pages);
797
798 return page;
799}
800
801static void setup_object(struct kmem_cache *s, struct page *page,
802 void *object)
803{
804 if (PageError(page)) {
805 init_object(s, object, 0);
806 init_tracking(s, object);
807 }
808
Christoph Lameter4f104932007-05-06 14:50:17 -0700809 if (unlikely(s->ctor))
810 s->ctor(object, s, SLAB_CTOR_CONSTRUCTOR);
Christoph Lameter81819f02007-05-06 14:49:36 -0700811}
812
813static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
814{
815 struct page *page;
816 struct kmem_cache_node *n;
817 void *start;
818 void *end;
819 void *last;
820 void *p;
821
Christoph Lameter81819f02007-05-06 14:49:36 -0700822 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
823
824 if (flags & __GFP_WAIT)
825 local_irq_enable();
826
827 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
828 if (!page)
829 goto out;
830
831 n = get_node(s, page_to_nid(page));
832 if (n)
833 atomic_long_inc(&n->nr_slabs);
834 page->offset = s->offset / sizeof(void *);
835 page->slab = s;
836 page->flags |= 1 << PG_slab;
837 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
838 SLAB_STORE_USER | SLAB_TRACE))
839 page->flags |= 1 << PG_error;
840
841 start = page_address(page);
842 end = start + s->objects * s->size;
843
844 if (unlikely(s->flags & SLAB_POISON))
845 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
846
847 last = start;
848 for (p = start + s->size; p < end; p += s->size) {
849 setup_object(s, page, last);
850 set_freepointer(s, last, p);
851 last = p;
852 }
853 setup_object(s, page, last);
854 set_freepointer(s, last, NULL);
855
856 page->freelist = start;
857 page->inuse = 0;
858out:
859 if (flags & __GFP_WAIT)
860 local_irq_disable();
861 return page;
862}
863
864static void __free_slab(struct kmem_cache *s, struct page *page)
865{
866 int pages = 1 << s->order;
867
868 if (unlikely(PageError(page) || s->dtor)) {
869 void *start = page_address(page);
870 void *end = start + (pages << PAGE_SHIFT);
871 void *p;
872
873 slab_pad_check(s, page);
874 for (p = start; p <= end - s->size; p += s->size) {
875 if (s->dtor)
876 s->dtor(p, s, 0);
877 check_object(s, page, p, 0);
878 }
879 }
880
881 mod_zone_page_state(page_zone(page),
882 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
883 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
884 - pages);
885
886 page->mapping = NULL;
887 __free_pages(page, s->order);
888}
889
890static void rcu_free_slab(struct rcu_head *h)
891{
892 struct page *page;
893
894 page = container_of((struct list_head *)h, struct page, lru);
895 __free_slab(page->slab, page);
896}
897
898static void free_slab(struct kmem_cache *s, struct page *page)
899{
900 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
901 /*
902 * RCU free overloads the RCU head over the LRU
903 */
904 struct rcu_head *head = (void *)&page->lru;
905
906 call_rcu(head, rcu_free_slab);
907 } else
908 __free_slab(s, page);
909}
910
911static void discard_slab(struct kmem_cache *s, struct page *page)
912{
913 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
914
915 atomic_long_dec(&n->nr_slabs);
916 reset_page_mapcount(page);
917 page->flags &= ~(1 << PG_slab | 1 << PG_error);
918 free_slab(s, page);
919}
920
921/*
922 * Per slab locking using the pagelock
923 */
924static __always_inline void slab_lock(struct page *page)
925{
926 bit_spin_lock(PG_locked, &page->flags);
927}
928
929static __always_inline void slab_unlock(struct page *page)
930{
931 bit_spin_unlock(PG_locked, &page->flags);
932}
933
934static __always_inline int slab_trylock(struct page *page)
935{
936 int rc = 1;
937
938 rc = bit_spin_trylock(PG_locked, &page->flags);
939 return rc;
940}
941
942/*
943 * Management of partially allocated slabs
944 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700945static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -0700946{
Christoph Lametere95eed52007-05-06 14:49:44 -0700947 spin_lock(&n->list_lock);
948 n->nr_partial++;
949 list_add_tail(&page->lru, &n->partial);
950 spin_unlock(&n->list_lock);
951}
Christoph Lameter81819f02007-05-06 14:49:36 -0700952
Christoph Lametere95eed52007-05-06 14:49:44 -0700953static void add_partial(struct kmem_cache_node *n, struct page *page)
954{
Christoph Lameter81819f02007-05-06 14:49:36 -0700955 spin_lock(&n->list_lock);
956 n->nr_partial++;
957 list_add(&page->lru, &n->partial);
958 spin_unlock(&n->list_lock);
959}
960
961static void remove_partial(struct kmem_cache *s,
962 struct page *page)
963{
964 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
965
966 spin_lock(&n->list_lock);
967 list_del(&page->lru);
968 n->nr_partial--;
969 spin_unlock(&n->list_lock);
970}
971
972/*
973 * Lock page and remove it from the partial list
974 *
975 * Must hold list_lock
976 */
977static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page)
978{
979 if (slab_trylock(page)) {
980 list_del(&page->lru);
981 n->nr_partial--;
982 return 1;
983 }
984 return 0;
985}
986
987/*
988 * Try to get a partial slab from a specific node
989 */
990static struct page *get_partial_node(struct kmem_cache_node *n)
991{
992 struct page *page;
993
994 /*
995 * Racy check. If we mistakenly see no partial slabs then we
996 * just allocate an empty slab. If we mistakenly try to get a
997 * partial slab then get_partials() will return NULL.
998 */
999 if (!n || !n->nr_partial)
1000 return NULL;
1001
1002 spin_lock(&n->list_lock);
1003 list_for_each_entry(page, &n->partial, lru)
1004 if (lock_and_del_slab(n, page))
1005 goto out;
1006 page = NULL;
1007out:
1008 spin_unlock(&n->list_lock);
1009 return page;
1010}
1011
1012/*
1013 * Get a page from somewhere. Search in increasing NUMA
1014 * distances.
1015 */
1016static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1017{
1018#ifdef CONFIG_NUMA
1019 struct zonelist *zonelist;
1020 struct zone **z;
1021 struct page *page;
1022
1023 /*
1024 * The defrag ratio allows to configure the tradeoffs between
1025 * inter node defragmentation and node local allocations.
1026 * A lower defrag_ratio increases the tendency to do local
1027 * allocations instead of scanning throught the partial
1028 * lists on other nodes.
1029 *
1030 * If defrag_ratio is set to 0 then kmalloc() always
1031 * returns node local objects. If its higher then kmalloc()
1032 * may return off node objects in order to avoid fragmentation.
1033 *
1034 * A higher ratio means slabs may be taken from other nodes
1035 * thus reducing the number of partial slabs on those nodes.
1036 *
1037 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1038 * defrag_ratio = 1000) then every (well almost) allocation
1039 * will first attempt to defrag slab caches on other nodes. This
1040 * means scanning over all nodes to look for partial slabs which
1041 * may be a bit expensive to do on every slab allocation.
1042 */
1043 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1044 return NULL;
1045
1046 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1047 ->node_zonelists[gfp_zone(flags)];
1048 for (z = zonelist->zones; *z; z++) {
1049 struct kmem_cache_node *n;
1050
1051 n = get_node(s, zone_to_nid(*z));
1052
1053 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001054 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001055 page = get_partial_node(n);
1056 if (page)
1057 return page;
1058 }
1059 }
1060#endif
1061 return NULL;
1062}
1063
1064/*
1065 * Get a partial page, lock it and return it.
1066 */
1067static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1068{
1069 struct page *page;
1070 int searchnode = (node == -1) ? numa_node_id() : node;
1071
1072 page = get_partial_node(get_node(s, searchnode));
1073 if (page || (flags & __GFP_THISNODE))
1074 return page;
1075
1076 return get_any_partial(s, flags);
1077}
1078
1079/*
1080 * Move a page back to the lists.
1081 *
1082 * Must be called with the slab lock held.
1083 *
1084 * On exit the slab lock will have been dropped.
1085 */
1086static void putback_slab(struct kmem_cache *s, struct page *page)
1087{
Christoph Lametere95eed52007-05-06 14:49:44 -07001088 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1089
Christoph Lameter81819f02007-05-06 14:49:36 -07001090 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001091
Christoph Lameter81819f02007-05-06 14:49:36 -07001092 if (page->freelist)
Christoph Lametere95eed52007-05-06 14:49:44 -07001093 add_partial(n, page);
1094 else if (PageError(page) && (s->flags & SLAB_STORE_USER))
1095 add_full(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001096 slab_unlock(page);
Christoph Lametere95eed52007-05-06 14:49:44 -07001097
Christoph Lameter81819f02007-05-06 14:49:36 -07001098 } else {
Christoph Lametere95eed52007-05-06 14:49:44 -07001099 if (n->nr_partial < MIN_PARTIAL) {
1100 /*
1101 * Adding an empty page to the partial slabs in order
1102 * to avoid page allocator overhead. This page needs to
1103 * come after all the others that are not fully empty
1104 * in order to make sure that we do maximum
1105 * defragmentation.
1106 */
1107 add_partial_tail(n, page);
1108 slab_unlock(page);
1109 } else {
1110 slab_unlock(page);
1111 discard_slab(s, page);
1112 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001113 }
1114}
1115
1116/*
1117 * Remove the cpu slab
1118 */
1119static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1120{
1121 s->cpu_slab[cpu] = NULL;
1122 ClearPageActive(page);
1123
1124 putback_slab(s, page);
1125}
1126
1127static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1128{
1129 slab_lock(page);
1130 deactivate_slab(s, page, cpu);
1131}
1132
1133/*
1134 * Flush cpu slab.
1135 * Called from IPI handler with interrupts disabled.
1136 */
1137static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1138{
1139 struct page *page = s->cpu_slab[cpu];
1140
1141 if (likely(page))
1142 flush_slab(s, page, cpu);
1143}
1144
1145static void flush_cpu_slab(void *d)
1146{
1147 struct kmem_cache *s = d;
1148 int cpu = smp_processor_id();
1149
1150 __flush_cpu_slab(s, cpu);
1151}
1152
1153static void flush_all(struct kmem_cache *s)
1154{
1155#ifdef CONFIG_SMP
1156 on_each_cpu(flush_cpu_slab, s, 1, 1);
1157#else
1158 unsigned long flags;
1159
1160 local_irq_save(flags);
1161 flush_cpu_slab(s);
1162 local_irq_restore(flags);
1163#endif
1164}
1165
1166/*
1167 * slab_alloc is optimized to only modify two cachelines on the fast path
1168 * (aside from the stack):
1169 *
1170 * 1. The page struct
1171 * 2. The first cacheline of the object to be allocated.
1172 *
1173 * The only cache lines that are read (apart from code) is the
1174 * per cpu array in the kmem_cache struct.
1175 *
1176 * Fastpath is not possible if we need to get a new slab or have
1177 * debugging enabled (which means all slabs are marked with PageError)
1178 */
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001179static void *slab_alloc(struct kmem_cache *s,
1180 gfp_t gfpflags, int node, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001181{
1182 struct page *page;
1183 void **object;
1184 unsigned long flags;
1185 int cpu;
1186
1187 local_irq_save(flags);
1188 cpu = smp_processor_id();
1189 page = s->cpu_slab[cpu];
1190 if (!page)
1191 goto new_slab;
1192
1193 slab_lock(page);
1194 if (unlikely(node != -1 && page_to_nid(page) != node))
1195 goto another_slab;
1196redo:
1197 object = page->freelist;
1198 if (unlikely(!object))
1199 goto another_slab;
1200 if (unlikely(PageError(page)))
1201 goto debug;
1202
1203have_object:
1204 page->inuse++;
1205 page->freelist = object[page->offset];
1206 slab_unlock(page);
1207 local_irq_restore(flags);
1208 return object;
1209
1210another_slab:
1211 deactivate_slab(s, page, cpu);
1212
1213new_slab:
1214 page = get_partial(s, gfpflags, node);
1215 if (likely(page)) {
1216have_slab:
1217 s->cpu_slab[cpu] = page;
1218 SetPageActive(page);
1219 goto redo;
1220 }
1221
1222 page = new_slab(s, gfpflags, node);
1223 if (page) {
1224 cpu = smp_processor_id();
1225 if (s->cpu_slab[cpu]) {
1226 /*
1227 * Someone else populated the cpu_slab while we enabled
1228 * interrupts, or we have got scheduled on another cpu.
1229 * The page may not be on the requested node.
1230 */
1231 if (node == -1 ||
1232 page_to_nid(s->cpu_slab[cpu]) == node) {
1233 /*
1234 * Current cpuslab is acceptable and we
1235 * want the current one since its cache hot
1236 */
1237 discard_slab(s, page);
1238 page = s->cpu_slab[cpu];
1239 slab_lock(page);
1240 goto redo;
1241 }
1242 /* Dump the current slab */
1243 flush_slab(s, s->cpu_slab[cpu], cpu);
1244 }
1245 slab_lock(page);
1246 goto have_slab;
1247 }
1248 local_irq_restore(flags);
1249 return NULL;
1250debug:
1251 if (!alloc_object_checks(s, page, object))
1252 goto another_slab;
1253 if (s->flags & SLAB_STORE_USER)
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001254 set_track(s, object, TRACK_ALLOC, addr);
Christoph Lameter70d71222007-05-06 14:49:47 -07001255 if (s->flags & SLAB_TRACE) {
1256 printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
1257 s->name, object, page->inuse,
1258 page->freelist);
1259 dump_stack();
1260 }
1261 init_object(s, object, 1);
Christoph Lameter81819f02007-05-06 14:49:36 -07001262 goto have_object;
1263}
1264
1265void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1266{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001267 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001268}
1269EXPORT_SYMBOL(kmem_cache_alloc);
1270
1271#ifdef CONFIG_NUMA
1272void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1273{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001274 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001275}
1276EXPORT_SYMBOL(kmem_cache_alloc_node);
1277#endif
1278
1279/*
1280 * The fastpath only writes the cacheline of the page struct and the first
1281 * cacheline of the object.
1282 *
1283 * No special cachelines need to be read
1284 */
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001285static void slab_free(struct kmem_cache *s, struct page *page,
1286 void *x, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001287{
1288 void *prior;
1289 void **object = (void *)x;
1290 unsigned long flags;
1291
1292 local_irq_save(flags);
1293 slab_lock(page);
1294
1295 if (unlikely(PageError(page)))
1296 goto debug;
1297checks_ok:
1298 prior = object[page->offset] = page->freelist;
1299 page->freelist = object;
1300 page->inuse--;
1301
1302 if (unlikely(PageActive(page)))
1303 /*
1304 * Cpu slabs are never on partial lists and are
1305 * never freed.
1306 */
1307 goto out_unlock;
1308
1309 if (unlikely(!page->inuse))
1310 goto slab_empty;
1311
1312 /*
1313 * Objects left in the slab. If it
1314 * was not on the partial list before
1315 * then add it.
1316 */
1317 if (unlikely(!prior))
Christoph Lametere95eed52007-05-06 14:49:44 -07001318 add_partial(get_node(s, page_to_nid(page)), page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001319
1320out_unlock:
1321 slab_unlock(page);
1322 local_irq_restore(flags);
1323 return;
1324
1325slab_empty:
1326 if (prior)
1327 /*
Christoph Lameter643b1132007-05-06 14:49:42 -07001328 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001329 */
1330 remove_partial(s, page);
1331
1332 slab_unlock(page);
1333 discard_slab(s, page);
1334 local_irq_restore(flags);
1335 return;
1336
1337debug:
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001338 if (!free_object_checks(s, page, x))
1339 goto out_unlock;
Christoph Lameter643b1132007-05-06 14:49:42 -07001340 if (!PageActive(page) && !page->freelist)
1341 remove_full(s, page);
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001342 if (s->flags & SLAB_STORE_USER)
1343 set_track(s, x, TRACK_FREE, addr);
Christoph Lameter70d71222007-05-06 14:49:47 -07001344 if (s->flags & SLAB_TRACE) {
1345 printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n",
1346 s->name, object, page->inuse,
1347 page->freelist);
1348 print_section("Object", (void *)object, s->objsize);
1349 dump_stack();
1350 }
1351 init_object(s, object, 0);
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001352 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001353}
1354
1355void kmem_cache_free(struct kmem_cache *s, void *x)
1356{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001357 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001358
Christoph Lameterb49af682007-05-06 14:49:41 -07001359 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001360
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001361 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001362}
1363EXPORT_SYMBOL(kmem_cache_free);
1364
1365/* Figure out on which slab object the object resides */
1366static struct page *get_object_page(const void *x)
1367{
Christoph Lameterb49af682007-05-06 14:49:41 -07001368 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001369
1370 if (!PageSlab(page))
1371 return NULL;
1372
1373 return page;
1374}
1375
1376/*
1377 * kmem_cache_open produces objects aligned at "size" and the first object
1378 * is placed at offset 0 in the slab (We have no metainformation on the
1379 * slab, all slabs are in essence "off slab").
1380 *
1381 * In order to get the desired alignment one just needs to align the
1382 * size.
1383 *
1384 * Notice that the allocation order determines the sizes of the per cpu
1385 * caches. Each processor has always one slab available for allocations.
1386 * Increasing the allocation order reduces the number of times that slabs
1387 * must be moved on and off the partial lists and therefore may influence
1388 * locking overhead.
1389 *
1390 * The offset is used to relocate the free list link in each object. It is
1391 * therefore possible to move the free list link behind the object. This
1392 * is necessary for RCU to work properly and also useful for debugging.
1393 */
1394
1395/*
1396 * Mininum / Maximum order of slab pages. This influences locking overhead
1397 * and slab fragmentation. A higher order reduces the number of partial slabs
1398 * and increases the number of allocations possible without having to
1399 * take the list_lock.
1400 */
1401static int slub_min_order;
1402static int slub_max_order = DEFAULT_MAX_ORDER;
1403
1404/*
1405 * Minimum number of objects per slab. This is necessary in order to
1406 * reduce locking overhead. Similar to the queue size in SLAB.
1407 */
1408static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1409
1410/*
1411 * Merge control. If this is set then no merging of slab caches will occur.
1412 */
1413static int slub_nomerge;
1414
1415/*
1416 * Debug settings:
1417 */
1418static int slub_debug;
1419
1420static char *slub_debug_slabs;
1421
1422/*
1423 * Calculate the order of allocation given an slab object size.
1424 *
1425 * The order of allocation has significant impact on other elements
1426 * of the system. Generally order 0 allocations should be preferred
1427 * since they do not cause fragmentation in the page allocator. Larger
1428 * objects may have problems with order 0 because there may be too much
1429 * space left unused in a slab. We go to a higher order if more than 1/8th
1430 * of the slab would be wasted.
1431 *
1432 * In order to reach satisfactory performance we must ensure that
1433 * a minimum number of objects is in one slab. Otherwise we may
1434 * generate too much activity on the partial lists. This is less a
1435 * concern for large slabs though. slub_max_order specifies the order
1436 * where we begin to stop considering the number of objects in a slab.
1437 *
1438 * Higher order allocations also allow the placement of more objects
1439 * in a slab and thereby reduce object handling overhead. If the user
1440 * has requested a higher mininum order then we start with that one
1441 * instead of zero.
1442 */
1443static int calculate_order(int size)
1444{
1445 int order;
1446 int rem;
1447
1448 for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT);
1449 order < MAX_ORDER; order++) {
1450 unsigned long slab_size = PAGE_SIZE << order;
1451
1452 if (slub_max_order > order &&
1453 slab_size < slub_min_objects * size)
1454 continue;
1455
1456 if (slab_size < size)
1457 continue;
1458
1459 rem = slab_size % size;
1460
1461 if (rem <= (PAGE_SIZE << order) / 8)
1462 break;
1463
1464 }
1465 if (order >= MAX_ORDER)
1466 return -E2BIG;
1467 return order;
1468}
1469
1470/*
1471 * Function to figure out which alignment to use from the
1472 * various ways of specifying it.
1473 */
1474static unsigned long calculate_alignment(unsigned long flags,
1475 unsigned long align, unsigned long size)
1476{
1477 /*
1478 * If the user wants hardware cache aligned objects then
1479 * follow that suggestion if the object is sufficiently
1480 * large.
1481 *
1482 * The hardware cache alignment cannot override the
1483 * specified alignment though. If that is greater
1484 * then use it.
1485 */
Christoph Lameter5af60832007-05-06 14:49:56 -07001486 if ((flags & SLAB_HWCACHE_ALIGN) &&
Christoph Lameter65c02d42007-05-09 02:32:35 -07001487 size > cache_line_size() / 2)
1488 return max_t(unsigned long, align, cache_line_size());
Christoph Lameter81819f02007-05-06 14:49:36 -07001489
1490 if (align < ARCH_SLAB_MINALIGN)
1491 return ARCH_SLAB_MINALIGN;
1492
1493 return ALIGN(align, sizeof(void *));
1494}
1495
1496static void init_kmem_cache_node(struct kmem_cache_node *n)
1497{
1498 n->nr_partial = 0;
1499 atomic_long_set(&n->nr_slabs, 0);
1500 spin_lock_init(&n->list_lock);
1501 INIT_LIST_HEAD(&n->partial);
Christoph Lameter643b1132007-05-06 14:49:42 -07001502 INIT_LIST_HEAD(&n->full);
Christoph Lameter81819f02007-05-06 14:49:36 -07001503}
1504
1505#ifdef CONFIG_NUMA
1506/*
1507 * No kmalloc_node yet so do it by hand. We know that this is the first
1508 * slab on the node for this slabcache. There are no concurrent accesses
1509 * possible.
1510 *
1511 * Note that this function only works on the kmalloc_node_cache
1512 * when allocating for the kmalloc_node_cache.
1513 */
1514static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1515 int node)
1516{
1517 struct page *page;
1518 struct kmem_cache_node *n;
1519
1520 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1521
1522 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1523 /* new_slab() disables interupts */
1524 local_irq_enable();
1525
1526 BUG_ON(!page);
1527 n = page->freelist;
1528 BUG_ON(!n);
1529 page->freelist = get_freepointer(kmalloc_caches, n);
1530 page->inuse++;
1531 kmalloc_caches->node[node] = n;
1532 init_object(kmalloc_caches, n, 1);
1533 init_kmem_cache_node(n);
1534 atomic_long_inc(&n->nr_slabs);
Christoph Lametere95eed52007-05-06 14:49:44 -07001535 add_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001536 return n;
1537}
1538
1539static void free_kmem_cache_nodes(struct kmem_cache *s)
1540{
1541 int node;
1542
1543 for_each_online_node(node) {
1544 struct kmem_cache_node *n = s->node[node];
1545 if (n && n != &s->local_node)
1546 kmem_cache_free(kmalloc_caches, n);
1547 s->node[node] = NULL;
1548 }
1549}
1550
1551static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1552{
1553 int node;
1554 int local_node;
1555
1556 if (slab_state >= UP)
1557 local_node = page_to_nid(virt_to_page(s));
1558 else
1559 local_node = 0;
1560
1561 for_each_online_node(node) {
1562 struct kmem_cache_node *n;
1563
1564 if (local_node == node)
1565 n = &s->local_node;
1566 else {
1567 if (slab_state == DOWN) {
1568 n = early_kmem_cache_node_alloc(gfpflags,
1569 node);
1570 continue;
1571 }
1572 n = kmem_cache_alloc_node(kmalloc_caches,
1573 gfpflags, node);
1574
1575 if (!n) {
1576 free_kmem_cache_nodes(s);
1577 return 0;
1578 }
1579
1580 }
1581 s->node[node] = n;
1582 init_kmem_cache_node(n);
1583 }
1584 return 1;
1585}
1586#else
1587static void free_kmem_cache_nodes(struct kmem_cache *s)
1588{
1589}
1590
1591static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1592{
1593 init_kmem_cache_node(&s->local_node);
1594 return 1;
1595}
1596#endif
1597
1598/*
1599 * calculate_sizes() determines the order and the distribution of data within
1600 * a slab object.
1601 */
1602static int calculate_sizes(struct kmem_cache *s)
1603{
1604 unsigned long flags = s->flags;
1605 unsigned long size = s->objsize;
1606 unsigned long align = s->align;
1607
1608 /*
1609 * Determine if we can poison the object itself. If the user of
1610 * the slab may touch the object after free or before allocation
1611 * then we should never poison the object itself.
1612 */
1613 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1614 !s->ctor && !s->dtor)
1615 s->flags |= __OBJECT_POISON;
1616 else
1617 s->flags &= ~__OBJECT_POISON;
1618
1619 /*
1620 * Round up object size to the next word boundary. We can only
1621 * place the free pointer at word boundaries and this determines
1622 * the possible location of the free pointer.
1623 */
1624 size = ALIGN(size, sizeof(void *));
1625
1626 /*
1627 * If we are redzoning then check if there is some space between the
1628 * end of the object and the free pointer. If not then add an
1629 * additional word, so that we can establish a redzone between
1630 * the object and the freepointer to be able to check for overwrites.
1631 */
1632 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1633 size += sizeof(void *);
1634
1635 /*
1636 * With that we have determined how much of the slab is in actual
1637 * use by the object. This is the potential offset to the free
1638 * pointer.
1639 */
1640 s->inuse = size;
1641
1642 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1643 s->ctor || s->dtor)) {
1644 /*
1645 * Relocate free pointer after the object if it is not
1646 * permitted to overwrite the first word of the object on
1647 * kmem_cache_free.
1648 *
1649 * This is the case if we do RCU, have a constructor or
1650 * destructor or are poisoning the objects.
1651 */
1652 s->offset = size;
1653 size += sizeof(void *);
1654 }
1655
1656 if (flags & SLAB_STORE_USER)
1657 /*
1658 * Need to store information about allocs and frees after
1659 * the object.
1660 */
1661 size += 2 * sizeof(struct track);
1662
Christoph Lameterbe7b3fb2007-05-09 02:32:36 -07001663 if (flags & SLAB_RED_ZONE)
Christoph Lameter81819f02007-05-06 14:49:36 -07001664 /*
1665 * Add some empty padding so that we can catch
1666 * overwrites from earlier objects rather than let
1667 * tracking information or the free pointer be
1668 * corrupted if an user writes before the start
1669 * of the object.
1670 */
1671 size += sizeof(void *);
1672 /*
1673 * Determine the alignment based on various parameters that the
Christoph Lameter65c02d42007-05-09 02:32:35 -07001674 * user specified and the dynamic determination of cache line size
1675 * on bootup.
Christoph Lameter81819f02007-05-06 14:49:36 -07001676 */
1677 align = calculate_alignment(flags, align, s->objsize);
1678
1679 /*
1680 * SLUB stores one object immediately after another beginning from
1681 * offset 0. In order to align the objects we have to simply size
1682 * each object to conform to the alignment.
1683 */
1684 size = ALIGN(size, align);
1685 s->size = size;
1686
1687 s->order = calculate_order(size);
1688 if (s->order < 0)
1689 return 0;
1690
1691 /*
1692 * Determine the number of objects per slab
1693 */
1694 s->objects = (PAGE_SIZE << s->order) / size;
1695
1696 /*
1697 * Verify that the number of objects is within permitted limits.
1698 * The page->inuse field is only 16 bit wide! So we cannot have
1699 * more than 64k objects per slab.
1700 */
1701 if (!s->objects || s->objects > 65535)
1702 return 0;
1703 return 1;
1704
1705}
1706
Christoph Lameter81819f02007-05-06 14:49:36 -07001707static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1708 const char *name, size_t size,
1709 size_t align, unsigned long flags,
1710 void (*ctor)(void *, struct kmem_cache *, unsigned long),
1711 void (*dtor)(void *, struct kmem_cache *, unsigned long))
1712{
1713 memset(s, 0, kmem_size);
1714 s->name = name;
1715 s->ctor = ctor;
1716 s->dtor = dtor;
1717 s->objsize = size;
1718 s->flags = flags;
1719 s->align = align;
1720
Christoph Lameter81819f02007-05-06 14:49:36 -07001721 /*
1722 * The page->offset field is only 16 bit wide. This is an offset
1723 * in units of words from the beginning of an object. If the slab
1724 * size is bigger then we cannot move the free pointer behind the
1725 * object anymore.
1726 *
1727 * On 32 bit platforms the limit is 256k. On 64bit platforms
1728 * the limit is 512k.
1729 *
1730 * Debugging or ctor/dtors may create a need to move the free
1731 * pointer. Fail if this happens.
1732 */
1733 if (s->size >= 65535 * sizeof(void *)) {
1734 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1735 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1736 BUG_ON(ctor || dtor);
1737 }
1738 else
1739 /*
1740 * Enable debugging if selected on the kernel commandline.
1741 */
1742 if (slub_debug && (!slub_debug_slabs ||
1743 strncmp(slub_debug_slabs, name,
1744 strlen(slub_debug_slabs)) == 0))
1745 s->flags |= slub_debug;
1746
1747 if (!calculate_sizes(s))
1748 goto error;
1749
1750 s->refcount = 1;
1751#ifdef CONFIG_NUMA
1752 s->defrag_ratio = 100;
1753#endif
1754
1755 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
1756 return 1;
1757error:
1758 if (flags & SLAB_PANIC)
1759 panic("Cannot create slab %s size=%lu realsize=%u "
1760 "order=%u offset=%u flags=%lx\n",
1761 s->name, (unsigned long)size, s->size, s->order,
1762 s->offset, flags);
1763 return 0;
1764}
1765EXPORT_SYMBOL(kmem_cache_open);
1766
1767/*
1768 * Check if a given pointer is valid
1769 */
1770int kmem_ptr_validate(struct kmem_cache *s, const void *object)
1771{
1772 struct page * page;
1773 void *addr;
1774
1775 page = get_object_page(object);
1776
1777 if (!page || s != page->slab)
1778 /* No slab or wrong slab */
1779 return 0;
1780
Christoph Lameterabcd08a2007-05-09 02:32:37 -07001781 if (!check_valid_pointer(s, page, object))
Christoph Lameter81819f02007-05-06 14:49:36 -07001782 return 0;
1783
1784 /*
1785 * We could also check if the object is on the slabs freelist.
1786 * But this would be too expensive and it seems that the main
1787 * purpose of kmem_ptr_valid is to check if the object belongs
1788 * to a certain slab.
1789 */
1790 return 1;
1791}
1792EXPORT_SYMBOL(kmem_ptr_validate);
1793
1794/*
1795 * Determine the size of a slab object
1796 */
1797unsigned int kmem_cache_size(struct kmem_cache *s)
1798{
1799 return s->objsize;
1800}
1801EXPORT_SYMBOL(kmem_cache_size);
1802
1803const char *kmem_cache_name(struct kmem_cache *s)
1804{
1805 return s->name;
1806}
1807EXPORT_SYMBOL(kmem_cache_name);
1808
1809/*
1810 * Attempt to free all slabs on a node
1811 */
1812static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
1813 struct list_head *list)
1814{
1815 int slabs_inuse = 0;
1816 unsigned long flags;
1817 struct page *page, *h;
1818
1819 spin_lock_irqsave(&n->list_lock, flags);
1820 list_for_each_entry_safe(page, h, list, lru)
1821 if (!page->inuse) {
1822 list_del(&page->lru);
1823 discard_slab(s, page);
1824 } else
1825 slabs_inuse++;
1826 spin_unlock_irqrestore(&n->list_lock, flags);
1827 return slabs_inuse;
1828}
1829
1830/*
1831 * Release all resources used by slab cache
1832 */
1833static int kmem_cache_close(struct kmem_cache *s)
1834{
1835 int node;
1836
1837 flush_all(s);
1838
1839 /* Attempt to free all objects */
1840 for_each_online_node(node) {
1841 struct kmem_cache_node *n = get_node(s, node);
1842
Christoph Lameter2086d262007-05-06 14:49:46 -07001843 n->nr_partial -= free_list(s, n, &n->partial);
Christoph Lameter81819f02007-05-06 14:49:36 -07001844 if (atomic_long_read(&n->nr_slabs))
1845 return 1;
1846 }
1847 free_kmem_cache_nodes(s);
1848 return 0;
1849}
1850
1851/*
1852 * Close a cache and release the kmem_cache structure
1853 * (must be used for caches created using kmem_cache_create)
1854 */
1855void kmem_cache_destroy(struct kmem_cache *s)
1856{
1857 down_write(&slub_lock);
1858 s->refcount--;
1859 if (!s->refcount) {
1860 list_del(&s->list);
1861 if (kmem_cache_close(s))
1862 WARN_ON(1);
1863 sysfs_slab_remove(s);
1864 kfree(s);
1865 }
1866 up_write(&slub_lock);
1867}
1868EXPORT_SYMBOL(kmem_cache_destroy);
1869
1870/********************************************************************
1871 * Kmalloc subsystem
1872 *******************************************************************/
1873
1874struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
1875EXPORT_SYMBOL(kmalloc_caches);
1876
1877#ifdef CONFIG_ZONE_DMA
1878static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
1879#endif
1880
1881static int __init setup_slub_min_order(char *str)
1882{
1883 get_option (&str, &slub_min_order);
1884
1885 return 1;
1886}
1887
1888__setup("slub_min_order=", setup_slub_min_order);
1889
1890static int __init setup_slub_max_order(char *str)
1891{
1892 get_option (&str, &slub_max_order);
1893
1894 return 1;
1895}
1896
1897__setup("slub_max_order=", setup_slub_max_order);
1898
1899static int __init setup_slub_min_objects(char *str)
1900{
1901 get_option (&str, &slub_min_objects);
1902
1903 return 1;
1904}
1905
1906__setup("slub_min_objects=", setup_slub_min_objects);
1907
1908static int __init setup_slub_nomerge(char *str)
1909{
1910 slub_nomerge = 1;
1911 return 1;
1912}
1913
1914__setup("slub_nomerge", setup_slub_nomerge);
1915
1916static int __init setup_slub_debug(char *str)
1917{
1918 if (!str || *str != '=')
1919 slub_debug = DEBUG_DEFAULT_FLAGS;
1920 else {
1921 str++;
1922 if (*str == 0 || *str == ',')
1923 slub_debug = DEBUG_DEFAULT_FLAGS;
1924 else
1925 for( ;*str && *str != ','; str++)
1926 switch (*str) {
1927 case 'f' : case 'F' :
1928 slub_debug |= SLAB_DEBUG_FREE;
1929 break;
1930 case 'z' : case 'Z' :
1931 slub_debug |= SLAB_RED_ZONE;
1932 break;
1933 case 'p' : case 'P' :
1934 slub_debug |= SLAB_POISON;
1935 break;
1936 case 'u' : case 'U' :
1937 slub_debug |= SLAB_STORE_USER;
1938 break;
1939 case 't' : case 'T' :
1940 slub_debug |= SLAB_TRACE;
1941 break;
1942 default:
1943 printk(KERN_ERR "slub_debug option '%c' "
1944 "unknown. skipped\n",*str);
1945 }
1946 }
1947
1948 if (*str == ',')
1949 slub_debug_slabs = str + 1;
1950 return 1;
1951}
1952
1953__setup("slub_debug", setup_slub_debug);
1954
1955static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
1956 const char *name, int size, gfp_t gfp_flags)
1957{
1958 unsigned int flags = 0;
1959
1960 if (gfp_flags & SLUB_DMA)
1961 flags = SLAB_CACHE_DMA;
1962
1963 down_write(&slub_lock);
1964 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
1965 flags, NULL, NULL))
1966 goto panic;
1967
1968 list_add(&s->list, &slab_caches);
1969 up_write(&slub_lock);
1970 if (sysfs_slab_add(s))
1971 goto panic;
1972 return s;
1973
1974panic:
1975 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
1976}
1977
1978static struct kmem_cache *get_slab(size_t size, gfp_t flags)
1979{
1980 int index = kmalloc_index(size);
1981
Christoph Lameter614410d2007-05-06 14:49:38 -07001982 if (!index)
Christoph Lameter81819f02007-05-06 14:49:36 -07001983 return NULL;
1984
1985 /* Allocation too large? */
1986 BUG_ON(index < 0);
1987
1988#ifdef CONFIG_ZONE_DMA
1989 if ((flags & SLUB_DMA)) {
1990 struct kmem_cache *s;
1991 struct kmem_cache *x;
1992 char *text;
1993 size_t realsize;
1994
1995 s = kmalloc_caches_dma[index];
1996 if (s)
1997 return s;
1998
1999 /* Dynamically create dma cache */
2000 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2001 if (!x)
2002 panic("Unable to allocate memory for dma cache\n");
2003
2004 if (index <= KMALLOC_SHIFT_HIGH)
2005 realsize = 1 << index;
2006 else {
2007 if (index == 1)
2008 realsize = 96;
2009 else
2010 realsize = 192;
2011 }
2012
2013 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2014 (unsigned int)realsize);
2015 s = create_kmalloc_cache(x, text, realsize, flags);
2016 kmalloc_caches_dma[index] = s;
2017 return s;
2018 }
2019#endif
2020 return &kmalloc_caches[index];
2021}
2022
2023void *__kmalloc(size_t size, gfp_t flags)
2024{
2025 struct kmem_cache *s = get_slab(size, flags);
2026
2027 if (s)
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002028 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002029 return NULL;
2030}
2031EXPORT_SYMBOL(__kmalloc);
2032
2033#ifdef CONFIG_NUMA
2034void *__kmalloc_node(size_t size, gfp_t flags, int node)
2035{
2036 struct kmem_cache *s = get_slab(size, flags);
2037
2038 if (s)
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002039 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002040 return NULL;
2041}
2042EXPORT_SYMBOL(__kmalloc_node);
2043#endif
2044
2045size_t ksize(const void *object)
2046{
2047 struct page *page = get_object_page(object);
2048 struct kmem_cache *s;
2049
2050 BUG_ON(!page);
2051 s = page->slab;
2052 BUG_ON(!s);
2053
2054 /*
2055 * Debugging requires use of the padding between object
2056 * and whatever may come after it.
2057 */
2058 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2059 return s->objsize;
2060
2061 /*
2062 * If we have the need to store the freelist pointer
2063 * back there or track user information then we can
2064 * only use the space before that information.
2065 */
2066 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2067 return s->inuse;
2068
2069 /*
2070 * Else we can use all the padding etc for the allocation
2071 */
2072 return s->size;
2073}
2074EXPORT_SYMBOL(ksize);
2075
2076void kfree(const void *x)
2077{
2078 struct kmem_cache *s;
2079 struct page *page;
2080
2081 if (!x)
2082 return;
2083
Christoph Lameterb49af682007-05-06 14:49:41 -07002084 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002085 s = page->slab;
2086
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002087 slab_free(s, page, (void *)x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002088}
2089EXPORT_SYMBOL(kfree);
2090
Christoph Lameter2086d262007-05-06 14:49:46 -07002091/*
2092 * kmem_cache_shrink removes empty slabs from the partial lists
2093 * and then sorts the partially allocated slabs by the number
2094 * of items in use. The slabs with the most items in use
2095 * come first. New allocations will remove these from the
2096 * partial list because they are full. The slabs with the
2097 * least items are placed last. If it happens that the objects
2098 * are freed then the page can be returned to the page allocator.
2099 */
2100int kmem_cache_shrink(struct kmem_cache *s)
2101{
2102 int node;
2103 int i;
2104 struct kmem_cache_node *n;
2105 struct page *page;
2106 struct page *t;
2107 struct list_head *slabs_by_inuse =
2108 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2109 unsigned long flags;
2110
2111 if (!slabs_by_inuse)
2112 return -ENOMEM;
2113
2114 flush_all(s);
2115 for_each_online_node(node) {
2116 n = get_node(s, node);
2117
2118 if (!n->nr_partial)
2119 continue;
2120
2121 for (i = 0; i < s->objects; i++)
2122 INIT_LIST_HEAD(slabs_by_inuse + i);
2123
2124 spin_lock_irqsave(&n->list_lock, flags);
2125
2126 /*
2127 * Build lists indexed by the items in use in
2128 * each slab or free slabs if empty.
2129 *
2130 * Note that concurrent frees may occur while
2131 * we hold the list_lock. page->inuse here is
2132 * the upper limit.
2133 */
2134 list_for_each_entry_safe(page, t, &n->partial, lru) {
2135 if (!page->inuse && slab_trylock(page)) {
2136 /*
2137 * Must hold slab lock here because slab_free
2138 * may have freed the last object and be
2139 * waiting to release the slab.
2140 */
2141 list_del(&page->lru);
2142 n->nr_partial--;
2143 slab_unlock(page);
2144 discard_slab(s, page);
2145 } else {
2146 if (n->nr_partial > MAX_PARTIAL)
2147 list_move(&page->lru,
2148 slabs_by_inuse + page->inuse);
2149 }
2150 }
2151
2152 if (n->nr_partial <= MAX_PARTIAL)
2153 goto out;
2154
2155 /*
2156 * Rebuild the partial list with the slabs filled up
2157 * most first and the least used slabs at the end.
2158 */
2159 for (i = s->objects - 1; i >= 0; i--)
2160 list_splice(slabs_by_inuse + i, n->partial.prev);
2161
2162 out:
2163 spin_unlock_irqrestore(&n->list_lock, flags);
2164 }
2165
2166 kfree(slabs_by_inuse);
2167 return 0;
2168}
2169EXPORT_SYMBOL(kmem_cache_shrink);
2170
Christoph Lameter81819f02007-05-06 14:49:36 -07002171/**
2172 * krealloc - reallocate memory. The contents will remain unchanged.
2173 *
2174 * @p: object to reallocate memory for.
2175 * @new_size: how many bytes of memory are required.
2176 * @flags: the type of memory to allocate.
2177 *
2178 * The contents of the object pointed to are preserved up to the
2179 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2180 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2181 * %NULL pointer, the object pointed to is freed.
2182 */
2183void *krealloc(const void *p, size_t new_size, gfp_t flags)
2184{
Christoph Lameter81819f02007-05-06 14:49:36 -07002185 void *ret;
Christoph Lameter1f99a282007-05-09 02:32:38 -07002186 size_t ks;
Christoph Lameter81819f02007-05-06 14:49:36 -07002187
2188 if (unlikely(!p))
2189 return kmalloc(new_size, flags);
2190
2191 if (unlikely(!new_size)) {
2192 kfree(p);
2193 return NULL;
2194 }
2195
Christoph Lameter1f99a282007-05-09 02:32:38 -07002196 ks = ksize(p);
2197 if (ks >= new_size)
Christoph Lameter81819f02007-05-06 14:49:36 -07002198 return (void *)p;
2199
2200 ret = kmalloc(new_size, flags);
2201 if (ret) {
Christoph Lameter1f99a282007-05-09 02:32:38 -07002202 memcpy(ret, p, min(new_size, ks));
Christoph Lameter81819f02007-05-06 14:49:36 -07002203 kfree(p);
2204 }
2205 return ret;
2206}
2207EXPORT_SYMBOL(krealloc);
2208
2209/********************************************************************
2210 * Basic setup of slabs
2211 *******************************************************************/
2212
2213void __init kmem_cache_init(void)
2214{
2215 int i;
2216
2217#ifdef CONFIG_NUMA
2218 /*
2219 * Must first have the slab cache available for the allocations of the
2220 * struct kmalloc_cache_node's. There is special bootstrap code in
2221 * kmem_cache_open for slab_state == DOWN.
2222 */
2223 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2224 sizeof(struct kmem_cache_node), GFP_KERNEL);
2225#endif
2226
2227 /* Able to allocate the per node structures */
2228 slab_state = PARTIAL;
2229
2230 /* Caches that are not of the two-to-the-power-of size */
2231 create_kmalloc_cache(&kmalloc_caches[1],
2232 "kmalloc-96", 96, GFP_KERNEL);
2233 create_kmalloc_cache(&kmalloc_caches[2],
2234 "kmalloc-192", 192, GFP_KERNEL);
2235
2236 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2237 create_kmalloc_cache(&kmalloc_caches[i],
2238 "kmalloc", 1 << i, GFP_KERNEL);
2239
2240 slab_state = UP;
2241
2242 /* Provide the correct kmalloc names now that the caches are up */
2243 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2244 kmalloc_caches[i]. name =
2245 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2246
2247#ifdef CONFIG_SMP
2248 register_cpu_notifier(&slab_notifier);
2249#endif
2250
2251 if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */
2252 kmem_size = offsetof(struct kmem_cache, cpu_slab)
2253 + nr_cpu_ids * sizeof(struct page *);
2254
2255 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2256 " Processors=%d, Nodes=%d\n",
Christoph Lameter65c02d42007-05-09 02:32:35 -07002257 KMALLOC_SHIFT_HIGH, cache_line_size(),
Christoph Lameter81819f02007-05-06 14:49:36 -07002258 slub_min_order, slub_max_order, slub_min_objects,
2259 nr_cpu_ids, nr_node_ids);
2260}
2261
2262/*
2263 * Find a mergeable slab cache
2264 */
2265static int slab_unmergeable(struct kmem_cache *s)
2266{
2267 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2268 return 1;
2269
2270 if (s->ctor || s->dtor)
2271 return 1;
2272
2273 return 0;
2274}
2275
2276static struct kmem_cache *find_mergeable(size_t size,
2277 size_t align, unsigned long flags,
2278 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2279 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2280{
2281 struct list_head *h;
2282
2283 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2284 return NULL;
2285
2286 if (ctor || dtor)
2287 return NULL;
2288
2289 size = ALIGN(size, sizeof(void *));
2290 align = calculate_alignment(flags, align, size);
2291 size = ALIGN(size, align);
2292
2293 list_for_each(h, &slab_caches) {
2294 struct kmem_cache *s =
2295 container_of(h, struct kmem_cache, list);
2296
2297 if (slab_unmergeable(s))
2298 continue;
2299
2300 if (size > s->size)
2301 continue;
2302
2303 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2304 (s->flags & SLUB_MERGE_SAME))
2305 continue;
2306 /*
2307 * Check if alignment is compatible.
2308 * Courtesy of Adrian Drzewiecki
2309 */
2310 if ((s->size & ~(align -1)) != s->size)
2311 continue;
2312
2313 if (s->size - size >= sizeof(void *))
2314 continue;
2315
2316 return s;
2317 }
2318 return NULL;
2319}
2320
2321struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2322 size_t align, unsigned long flags,
2323 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2324 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2325{
2326 struct kmem_cache *s;
2327
2328 down_write(&slub_lock);
2329 s = find_mergeable(size, align, flags, dtor, ctor);
2330 if (s) {
2331 s->refcount++;
2332 /*
2333 * Adjust the object sizes so that we clear
2334 * the complete object on kzalloc.
2335 */
2336 s->objsize = max(s->objsize, (int)size);
2337 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2338 if (sysfs_slab_alias(s, name))
2339 goto err;
2340 } else {
2341 s = kmalloc(kmem_size, GFP_KERNEL);
2342 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2343 size, align, flags, ctor, dtor)) {
2344 if (sysfs_slab_add(s)) {
2345 kfree(s);
2346 goto err;
2347 }
2348 list_add(&s->list, &slab_caches);
2349 } else
2350 kfree(s);
2351 }
2352 up_write(&slub_lock);
2353 return s;
2354
2355err:
2356 up_write(&slub_lock);
2357 if (flags & SLAB_PANIC)
2358 panic("Cannot create slabcache %s\n", name);
2359 else
2360 s = NULL;
2361 return s;
2362}
2363EXPORT_SYMBOL(kmem_cache_create);
2364
2365void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2366{
2367 void *x;
2368
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002369 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002370 if (x)
2371 memset(x, 0, s->objsize);
2372 return x;
2373}
2374EXPORT_SYMBOL(kmem_cache_zalloc);
2375
2376#ifdef CONFIG_SMP
2377static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2378{
2379 struct list_head *h;
2380
2381 down_read(&slub_lock);
2382 list_for_each(h, &slab_caches) {
2383 struct kmem_cache *s =
2384 container_of(h, struct kmem_cache, list);
2385
2386 func(s, cpu);
2387 }
2388 up_read(&slub_lock);
2389}
2390
2391/*
2392 * Use the cpu notifier to insure that the slab are flushed
2393 * when necessary.
2394 */
2395static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2396 unsigned long action, void *hcpu)
2397{
2398 long cpu = (long)hcpu;
2399
2400 switch (action) {
2401 case CPU_UP_CANCELED:
2402 case CPU_DEAD:
2403 for_all_slabs(__flush_cpu_slab, cpu);
2404 break;
2405 default:
2406 break;
2407 }
2408 return NOTIFY_OK;
2409}
2410
2411static struct notifier_block __cpuinitdata slab_notifier =
2412 { &slab_cpuup_callback, NULL, 0 };
2413
2414#endif
2415
Christoph Lameter81819f02007-05-06 14:49:36 -07002416#ifdef CONFIG_NUMA
2417
2418/*****************************************************************
2419 * Generic reaper used to support the page allocator
2420 * (the cpu slabs are reaped by a per slab workqueue).
2421 *
2422 * Maybe move this to the page allocator?
2423 ****************************************************************/
2424
2425static DEFINE_PER_CPU(unsigned long, reap_node);
2426
2427static void init_reap_node(int cpu)
2428{
2429 int node;
2430
2431 node = next_node(cpu_to_node(cpu), node_online_map);
2432 if (node == MAX_NUMNODES)
2433 node = first_node(node_online_map);
2434
2435 __get_cpu_var(reap_node) = node;
2436}
2437
2438static void next_reap_node(void)
2439{
2440 int node = __get_cpu_var(reap_node);
2441
2442 /*
2443 * Also drain per cpu pages on remote zones
2444 */
2445 if (node != numa_node_id())
2446 drain_node_pages(node);
2447
2448 node = next_node(node, node_online_map);
2449 if (unlikely(node >= MAX_NUMNODES))
2450 node = first_node(node_online_map);
2451 __get_cpu_var(reap_node) = node;
2452}
2453#else
2454#define init_reap_node(cpu) do { } while (0)
2455#define next_reap_node(void) do { } while (0)
2456#endif
2457
2458#define REAPTIMEOUT_CPUC (2*HZ)
2459
2460#ifdef CONFIG_SMP
2461static DEFINE_PER_CPU(struct delayed_work, reap_work);
2462
2463static void cache_reap(struct work_struct *unused)
2464{
2465 next_reap_node();
2466 refresh_cpu_vm_stats(smp_processor_id());
2467 schedule_delayed_work(&__get_cpu_var(reap_work),
2468 REAPTIMEOUT_CPUC);
2469}
2470
2471static void __devinit start_cpu_timer(int cpu)
2472{
2473 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
2474
2475 /*
2476 * When this gets called from do_initcalls via cpucache_init(),
2477 * init_workqueues() has already run, so keventd will be setup
2478 * at that time.
2479 */
2480 if (keventd_up() && reap_work->work.func == NULL) {
2481 init_reap_node(cpu);
2482 INIT_DELAYED_WORK(reap_work, cache_reap);
2483 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
2484 }
2485}
2486
2487static int __init cpucache_init(void)
2488{
2489 int cpu;
2490
2491 /*
2492 * Register the timers that drain pcp pages and update vm statistics
2493 */
2494 for_each_online_cpu(cpu)
2495 start_cpu_timer(cpu);
2496 return 0;
2497}
2498__initcall(cpucache_init);
2499#endif
2500
2501#ifdef SLUB_RESILIENCY_TEST
2502static unsigned long validate_slab_cache(struct kmem_cache *s);
2503
2504static void resiliency_test(void)
2505{
2506 u8 *p;
2507
2508 printk(KERN_ERR "SLUB resiliency testing\n");
2509 printk(KERN_ERR "-----------------------\n");
2510 printk(KERN_ERR "A. Corruption after allocation\n");
2511
2512 p = kzalloc(16, GFP_KERNEL);
2513 p[16] = 0x12;
2514 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2515 " 0x12->0x%p\n\n", p + 16);
2516
2517 validate_slab_cache(kmalloc_caches + 4);
2518
2519 /* Hmmm... The next two are dangerous */
2520 p = kzalloc(32, GFP_KERNEL);
2521 p[32 + sizeof(void *)] = 0x34;
2522 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2523 " 0x34 -> -0x%p\n", p);
2524 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2525
2526 validate_slab_cache(kmalloc_caches + 5);
2527 p = kzalloc(64, GFP_KERNEL);
2528 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2529 *p = 0x56;
2530 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2531 p);
2532 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2533 validate_slab_cache(kmalloc_caches + 6);
2534
2535 printk(KERN_ERR "\nB. Corruption after free\n");
2536 p = kzalloc(128, GFP_KERNEL);
2537 kfree(p);
2538 *p = 0x78;
2539 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2540 validate_slab_cache(kmalloc_caches + 7);
2541
2542 p = kzalloc(256, GFP_KERNEL);
2543 kfree(p);
2544 p[50] = 0x9a;
2545 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2546 validate_slab_cache(kmalloc_caches + 8);
2547
2548 p = kzalloc(512, GFP_KERNEL);
2549 kfree(p);
2550 p[512] = 0xab;
2551 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2552 validate_slab_cache(kmalloc_caches + 9);
2553}
2554#else
2555static void resiliency_test(void) {};
2556#endif
2557
2558/*
2559 * These are not as efficient as kmalloc for the non debug case.
2560 * We do not have the page struct available so we have to touch one
2561 * cacheline in struct kmem_cache to check slab flags.
2562 */
2563void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2564{
2565 struct kmem_cache *s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002566
2567 if (!s)
2568 return NULL;
2569
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002570 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07002571}
2572
2573void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2574 int node, void *caller)
2575{
2576 struct kmem_cache *s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002577
2578 if (!s)
2579 return NULL;
2580
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002581 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07002582}
2583
2584#ifdef CONFIG_SYSFS
2585
Christoph Lameter53e15af2007-05-06 14:49:43 -07002586static int validate_slab(struct kmem_cache *s, struct page *page)
2587{
2588 void *p;
2589 void *addr = page_address(page);
2590 unsigned long map[BITS_TO_LONGS(s->objects)];
2591
2592 if (!check_slab(s, page) ||
2593 !on_freelist(s, page, NULL))
2594 return 0;
2595
2596 /* Now we know that a valid freelist exists */
2597 bitmap_zero(map, s->objects);
2598
2599 for(p = page->freelist; p; p = get_freepointer(s, p)) {
2600 set_bit((p - addr) / s->size, map);
2601 if (!check_object(s, page, p, 0))
2602 return 0;
2603 }
2604
2605 for(p = addr; p < addr + s->objects * s->size; p += s->size)
2606 if (!test_bit((p - addr) / s->size, map))
2607 if (!check_object(s, page, p, 1))
2608 return 0;
2609 return 1;
2610}
2611
2612static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2613{
2614 if (slab_trylock(page)) {
2615 validate_slab(s, page);
2616 slab_unlock(page);
2617 } else
2618 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2619 s->name, page);
2620
2621 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2622 if (!PageError(page))
2623 printk(KERN_ERR "SLUB %s: PageError not set "
2624 "on slab 0x%p\n", s->name, page);
2625 } else {
2626 if (PageError(page))
2627 printk(KERN_ERR "SLUB %s: PageError set on "
2628 "slab 0x%p\n", s->name, page);
2629 }
2630}
2631
2632static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2633{
2634 unsigned long count = 0;
2635 struct page *page;
2636 unsigned long flags;
2637
2638 spin_lock_irqsave(&n->list_lock, flags);
2639
2640 list_for_each_entry(page, &n->partial, lru) {
2641 validate_slab_slab(s, page);
2642 count++;
2643 }
2644 if (count != n->nr_partial)
2645 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2646 "counter=%ld\n", s->name, count, n->nr_partial);
2647
2648 if (!(s->flags & SLAB_STORE_USER))
2649 goto out;
2650
2651 list_for_each_entry(page, &n->full, lru) {
2652 validate_slab_slab(s, page);
2653 count++;
2654 }
2655 if (count != atomic_long_read(&n->nr_slabs))
2656 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2657 "counter=%ld\n", s->name, count,
2658 atomic_long_read(&n->nr_slabs));
2659
2660out:
2661 spin_unlock_irqrestore(&n->list_lock, flags);
2662 return count;
2663}
2664
2665static unsigned long validate_slab_cache(struct kmem_cache *s)
2666{
2667 int node;
2668 unsigned long count = 0;
2669
2670 flush_all(s);
2671 for_each_online_node(node) {
2672 struct kmem_cache_node *n = get_node(s, node);
2673
2674 count += validate_slab_node(s, n);
2675 }
2676 return count;
2677}
2678
Christoph Lameter88a420e2007-05-06 14:49:45 -07002679/*
2680 * Generate lists of locations where slabcache objects are allocated
2681 * and freed.
2682 */
2683
2684struct location {
2685 unsigned long count;
2686 void *addr;
2687};
2688
2689struct loc_track {
2690 unsigned long max;
2691 unsigned long count;
2692 struct location *loc;
2693};
2694
2695static void free_loc_track(struct loc_track *t)
2696{
2697 if (t->max)
2698 free_pages((unsigned long)t->loc,
2699 get_order(sizeof(struct location) * t->max));
2700}
2701
2702static int alloc_loc_track(struct loc_track *t, unsigned long max)
2703{
2704 struct location *l;
2705 int order;
2706
2707 if (!max)
2708 max = PAGE_SIZE / sizeof(struct location);
2709
2710 order = get_order(sizeof(struct location) * max);
2711
2712 l = (void *)__get_free_pages(GFP_KERNEL, order);
2713
2714 if (!l)
2715 return 0;
2716
2717 if (t->count) {
2718 memcpy(l, t->loc, sizeof(struct location) * t->count);
2719 free_loc_track(t);
2720 }
2721 t->max = max;
2722 t->loc = l;
2723 return 1;
2724}
2725
2726static int add_location(struct loc_track *t, struct kmem_cache *s,
2727 void *addr)
2728{
2729 long start, end, pos;
2730 struct location *l;
2731 void *caddr;
2732
2733 start = -1;
2734 end = t->count;
2735
2736 for ( ; ; ) {
2737 pos = start + (end - start + 1) / 2;
2738
2739 /*
2740 * There is nothing at "end". If we end up there
2741 * we need to add something to before end.
2742 */
2743 if (pos == end)
2744 break;
2745
2746 caddr = t->loc[pos].addr;
2747 if (addr == caddr) {
2748 t->loc[pos].count++;
2749 return 1;
2750 }
2751
2752 if (addr < caddr)
2753 end = pos;
2754 else
2755 start = pos;
2756 }
2757
2758 /*
2759 * Not found. Insert new tracking element
2760 */
2761 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max))
2762 return 0;
2763
2764 l = t->loc + pos;
2765 if (pos < t->count)
2766 memmove(l + 1, l,
2767 (t->count - pos) * sizeof(struct location));
2768 t->count++;
2769 l->count = 1;
2770 l->addr = addr;
2771 return 1;
2772}
2773
2774static void process_slab(struct loc_track *t, struct kmem_cache *s,
2775 struct page *page, enum track_item alloc)
2776{
2777 void *addr = page_address(page);
2778 unsigned long map[BITS_TO_LONGS(s->objects)];
2779 void *p;
2780
2781 bitmap_zero(map, s->objects);
2782 for (p = page->freelist; p; p = get_freepointer(s, p))
2783 set_bit((p - addr) / s->size, map);
2784
2785 for (p = addr; p < addr + s->objects * s->size; p += s->size)
2786 if (!test_bit((p - addr) / s->size, map)) {
2787 void *addr = get_track(s, p, alloc)->addr;
2788
2789 add_location(t, s, addr);
2790 }
2791}
2792
2793static int list_locations(struct kmem_cache *s, char *buf,
2794 enum track_item alloc)
2795{
2796 int n = 0;
2797 unsigned long i;
2798 struct loc_track t;
2799 int node;
2800
2801 t.count = 0;
2802 t.max = 0;
2803
2804 /* Push back cpu slabs */
2805 flush_all(s);
2806
2807 for_each_online_node(node) {
2808 struct kmem_cache_node *n = get_node(s, node);
2809 unsigned long flags;
2810 struct page *page;
2811
2812 if (!atomic_read(&n->nr_slabs))
2813 continue;
2814
2815 spin_lock_irqsave(&n->list_lock, flags);
2816 list_for_each_entry(page, &n->partial, lru)
2817 process_slab(&t, s, page, alloc);
2818 list_for_each_entry(page, &n->full, lru)
2819 process_slab(&t, s, page, alloc);
2820 spin_unlock_irqrestore(&n->list_lock, flags);
2821 }
2822
2823 for (i = 0; i < t.count; i++) {
2824 void *addr = t.loc[i].addr;
2825
2826 if (n > PAGE_SIZE - 100)
2827 break;
2828 n += sprintf(buf + n, "%7ld ", t.loc[i].count);
2829 if (addr)
2830 n += sprint_symbol(buf + n, (unsigned long)t.loc[i].addr);
2831 else
2832 n += sprintf(buf + n, "<not-available>");
2833 n += sprintf(buf + n, "\n");
2834 }
2835
2836 free_loc_track(&t);
2837 if (!t.count)
2838 n += sprintf(buf, "No data\n");
2839 return n;
2840}
2841
Christoph Lameter81819f02007-05-06 14:49:36 -07002842static unsigned long count_partial(struct kmem_cache_node *n)
2843{
2844 unsigned long flags;
2845 unsigned long x = 0;
2846 struct page *page;
2847
2848 spin_lock_irqsave(&n->list_lock, flags);
2849 list_for_each_entry(page, &n->partial, lru)
2850 x += page->inuse;
2851 spin_unlock_irqrestore(&n->list_lock, flags);
2852 return x;
2853}
2854
2855enum slab_stat_type {
2856 SL_FULL,
2857 SL_PARTIAL,
2858 SL_CPU,
2859 SL_OBJECTS
2860};
2861
2862#define SO_FULL (1 << SL_FULL)
2863#define SO_PARTIAL (1 << SL_PARTIAL)
2864#define SO_CPU (1 << SL_CPU)
2865#define SO_OBJECTS (1 << SL_OBJECTS)
2866
2867static unsigned long slab_objects(struct kmem_cache *s,
2868 char *buf, unsigned long flags)
2869{
2870 unsigned long total = 0;
2871 int cpu;
2872 int node;
2873 int x;
2874 unsigned long *nodes;
2875 unsigned long *per_cpu;
2876
2877 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
2878 per_cpu = nodes + nr_node_ids;
2879
2880 for_each_possible_cpu(cpu) {
2881 struct page *page = s->cpu_slab[cpu];
2882 int node;
2883
2884 if (page) {
2885 node = page_to_nid(page);
2886 if (flags & SO_CPU) {
2887 int x = 0;
2888
2889 if (flags & SO_OBJECTS)
2890 x = page->inuse;
2891 else
2892 x = 1;
2893 total += x;
2894 nodes[node] += x;
2895 }
2896 per_cpu[node]++;
2897 }
2898 }
2899
2900 for_each_online_node(node) {
2901 struct kmem_cache_node *n = get_node(s, node);
2902
2903 if (flags & SO_PARTIAL) {
2904 if (flags & SO_OBJECTS)
2905 x = count_partial(n);
2906 else
2907 x = n->nr_partial;
2908 total += x;
2909 nodes[node] += x;
2910 }
2911
2912 if (flags & SO_FULL) {
2913 int full_slabs = atomic_read(&n->nr_slabs)
2914 - per_cpu[node]
2915 - n->nr_partial;
2916
2917 if (flags & SO_OBJECTS)
2918 x = full_slabs * s->objects;
2919 else
2920 x = full_slabs;
2921 total += x;
2922 nodes[node] += x;
2923 }
2924 }
2925
2926 x = sprintf(buf, "%lu", total);
2927#ifdef CONFIG_NUMA
2928 for_each_online_node(node)
2929 if (nodes[node])
2930 x += sprintf(buf + x, " N%d=%lu",
2931 node, nodes[node]);
2932#endif
2933 kfree(nodes);
2934 return x + sprintf(buf + x, "\n");
2935}
2936
2937static int any_slab_objects(struct kmem_cache *s)
2938{
2939 int node;
2940 int cpu;
2941
2942 for_each_possible_cpu(cpu)
2943 if (s->cpu_slab[cpu])
2944 return 1;
2945
2946 for_each_node(node) {
2947 struct kmem_cache_node *n = get_node(s, node);
2948
2949 if (n->nr_partial || atomic_read(&n->nr_slabs))
2950 return 1;
2951 }
2952 return 0;
2953}
2954
2955#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2956#define to_slab(n) container_of(n, struct kmem_cache, kobj);
2957
2958struct slab_attribute {
2959 struct attribute attr;
2960 ssize_t (*show)(struct kmem_cache *s, char *buf);
2961 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
2962};
2963
2964#define SLAB_ATTR_RO(_name) \
2965 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
2966
2967#define SLAB_ATTR(_name) \
2968 static struct slab_attribute _name##_attr = \
2969 __ATTR(_name, 0644, _name##_show, _name##_store)
2970
Christoph Lameter81819f02007-05-06 14:49:36 -07002971static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
2972{
2973 return sprintf(buf, "%d\n", s->size);
2974}
2975SLAB_ATTR_RO(slab_size);
2976
2977static ssize_t align_show(struct kmem_cache *s, char *buf)
2978{
2979 return sprintf(buf, "%d\n", s->align);
2980}
2981SLAB_ATTR_RO(align);
2982
2983static ssize_t object_size_show(struct kmem_cache *s, char *buf)
2984{
2985 return sprintf(buf, "%d\n", s->objsize);
2986}
2987SLAB_ATTR_RO(object_size);
2988
2989static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
2990{
2991 return sprintf(buf, "%d\n", s->objects);
2992}
2993SLAB_ATTR_RO(objs_per_slab);
2994
2995static ssize_t order_show(struct kmem_cache *s, char *buf)
2996{
2997 return sprintf(buf, "%d\n", s->order);
2998}
2999SLAB_ATTR_RO(order);
3000
3001static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3002{
3003 if (s->ctor) {
3004 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3005
3006 return n + sprintf(buf + n, "\n");
3007 }
3008 return 0;
3009}
3010SLAB_ATTR_RO(ctor);
3011
3012static ssize_t dtor_show(struct kmem_cache *s, char *buf)
3013{
3014 if (s->dtor) {
3015 int n = sprint_symbol(buf, (unsigned long)s->dtor);
3016
3017 return n + sprintf(buf + n, "\n");
3018 }
3019 return 0;
3020}
3021SLAB_ATTR_RO(dtor);
3022
3023static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3024{
3025 return sprintf(buf, "%d\n", s->refcount - 1);
3026}
3027SLAB_ATTR_RO(aliases);
3028
3029static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3030{
3031 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3032}
3033SLAB_ATTR_RO(slabs);
3034
3035static ssize_t partial_show(struct kmem_cache *s, char *buf)
3036{
3037 return slab_objects(s, buf, SO_PARTIAL);
3038}
3039SLAB_ATTR_RO(partial);
3040
3041static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3042{
3043 return slab_objects(s, buf, SO_CPU);
3044}
3045SLAB_ATTR_RO(cpu_slabs);
3046
3047static ssize_t objects_show(struct kmem_cache *s, char *buf)
3048{
3049 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3050}
3051SLAB_ATTR_RO(objects);
3052
3053static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3054{
3055 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3056}
3057
3058static ssize_t sanity_checks_store(struct kmem_cache *s,
3059 const char *buf, size_t length)
3060{
3061 s->flags &= ~SLAB_DEBUG_FREE;
3062 if (buf[0] == '1')
3063 s->flags |= SLAB_DEBUG_FREE;
3064 return length;
3065}
3066SLAB_ATTR(sanity_checks);
3067
3068static ssize_t trace_show(struct kmem_cache *s, char *buf)
3069{
3070 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3071}
3072
3073static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3074 size_t length)
3075{
3076 s->flags &= ~SLAB_TRACE;
3077 if (buf[0] == '1')
3078 s->flags |= SLAB_TRACE;
3079 return length;
3080}
3081SLAB_ATTR(trace);
3082
3083static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3084{
3085 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3086}
3087
3088static ssize_t reclaim_account_store(struct kmem_cache *s,
3089 const char *buf, size_t length)
3090{
3091 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3092 if (buf[0] == '1')
3093 s->flags |= SLAB_RECLAIM_ACCOUNT;
3094 return length;
3095}
3096SLAB_ATTR(reclaim_account);
3097
3098static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3099{
Christoph Lameter5af60832007-05-06 14:49:56 -07003100 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
Christoph Lameter81819f02007-05-06 14:49:36 -07003101}
3102SLAB_ATTR_RO(hwcache_align);
3103
3104#ifdef CONFIG_ZONE_DMA
3105static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3106{
3107 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3108}
3109SLAB_ATTR_RO(cache_dma);
3110#endif
3111
3112static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3113{
3114 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3115}
3116SLAB_ATTR_RO(destroy_by_rcu);
3117
3118static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3119{
3120 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3121}
3122
3123static ssize_t red_zone_store(struct kmem_cache *s,
3124 const char *buf, size_t length)
3125{
3126 if (any_slab_objects(s))
3127 return -EBUSY;
3128
3129 s->flags &= ~SLAB_RED_ZONE;
3130 if (buf[0] == '1')
3131 s->flags |= SLAB_RED_ZONE;
3132 calculate_sizes(s);
3133 return length;
3134}
3135SLAB_ATTR(red_zone);
3136
3137static ssize_t poison_show(struct kmem_cache *s, char *buf)
3138{
3139 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3140}
3141
3142static ssize_t poison_store(struct kmem_cache *s,
3143 const char *buf, size_t length)
3144{
3145 if (any_slab_objects(s))
3146 return -EBUSY;
3147
3148 s->flags &= ~SLAB_POISON;
3149 if (buf[0] == '1')
3150 s->flags |= SLAB_POISON;
3151 calculate_sizes(s);
3152 return length;
3153}
3154SLAB_ATTR(poison);
3155
3156static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3157{
3158 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3159}
3160
3161static ssize_t store_user_store(struct kmem_cache *s,
3162 const char *buf, size_t length)
3163{
3164 if (any_slab_objects(s))
3165 return -EBUSY;
3166
3167 s->flags &= ~SLAB_STORE_USER;
3168 if (buf[0] == '1')
3169 s->flags |= SLAB_STORE_USER;
3170 calculate_sizes(s);
3171 return length;
3172}
3173SLAB_ATTR(store_user);
3174
Christoph Lameter53e15af2007-05-06 14:49:43 -07003175static ssize_t validate_show(struct kmem_cache *s, char *buf)
3176{
3177 return 0;
3178}
3179
3180static ssize_t validate_store(struct kmem_cache *s,
3181 const char *buf, size_t length)
3182{
3183 if (buf[0] == '1')
3184 validate_slab_cache(s);
3185 else
3186 return -EINVAL;
3187 return length;
3188}
3189SLAB_ATTR(validate);
3190
Christoph Lameter2086d262007-05-06 14:49:46 -07003191static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3192{
3193 return 0;
3194}
3195
3196static ssize_t shrink_store(struct kmem_cache *s,
3197 const char *buf, size_t length)
3198{
3199 if (buf[0] == '1') {
3200 int rc = kmem_cache_shrink(s);
3201
3202 if (rc)
3203 return rc;
3204 } else
3205 return -EINVAL;
3206 return length;
3207}
3208SLAB_ATTR(shrink);
3209
Christoph Lameter88a420e2007-05-06 14:49:45 -07003210static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3211{
3212 if (!(s->flags & SLAB_STORE_USER))
3213 return -ENOSYS;
3214 return list_locations(s, buf, TRACK_ALLOC);
3215}
3216SLAB_ATTR_RO(alloc_calls);
3217
3218static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3219{
3220 if (!(s->flags & SLAB_STORE_USER))
3221 return -ENOSYS;
3222 return list_locations(s, buf, TRACK_FREE);
3223}
3224SLAB_ATTR_RO(free_calls);
3225
Christoph Lameter81819f02007-05-06 14:49:36 -07003226#ifdef CONFIG_NUMA
3227static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3228{
3229 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3230}
3231
3232static ssize_t defrag_ratio_store(struct kmem_cache *s,
3233 const char *buf, size_t length)
3234{
3235 int n = simple_strtoul(buf, NULL, 10);
3236
3237 if (n < 100)
3238 s->defrag_ratio = n * 10;
3239 return length;
3240}
3241SLAB_ATTR(defrag_ratio);
3242#endif
3243
3244static struct attribute * slab_attrs[] = {
3245 &slab_size_attr.attr,
3246 &object_size_attr.attr,
3247 &objs_per_slab_attr.attr,
3248 &order_attr.attr,
3249 &objects_attr.attr,
3250 &slabs_attr.attr,
3251 &partial_attr.attr,
3252 &cpu_slabs_attr.attr,
3253 &ctor_attr.attr,
3254 &dtor_attr.attr,
3255 &aliases_attr.attr,
3256 &align_attr.attr,
3257 &sanity_checks_attr.attr,
3258 &trace_attr.attr,
3259 &hwcache_align_attr.attr,
3260 &reclaim_account_attr.attr,
3261 &destroy_by_rcu_attr.attr,
3262 &red_zone_attr.attr,
3263 &poison_attr.attr,
3264 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07003265 &validate_attr.attr,
Christoph Lameter2086d262007-05-06 14:49:46 -07003266 &shrink_attr.attr,
Christoph Lameter88a420e2007-05-06 14:49:45 -07003267 &alloc_calls_attr.attr,
3268 &free_calls_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003269#ifdef CONFIG_ZONE_DMA
3270 &cache_dma_attr.attr,
3271#endif
3272#ifdef CONFIG_NUMA
3273 &defrag_ratio_attr.attr,
3274#endif
3275 NULL
3276};
3277
3278static struct attribute_group slab_attr_group = {
3279 .attrs = slab_attrs,
3280};
3281
3282static ssize_t slab_attr_show(struct kobject *kobj,
3283 struct attribute *attr,
3284 char *buf)
3285{
3286 struct slab_attribute *attribute;
3287 struct kmem_cache *s;
3288 int err;
3289
3290 attribute = to_slab_attr(attr);
3291 s = to_slab(kobj);
3292
3293 if (!attribute->show)
3294 return -EIO;
3295
3296 err = attribute->show(s, buf);
3297
3298 return err;
3299}
3300
3301static ssize_t slab_attr_store(struct kobject *kobj,
3302 struct attribute *attr,
3303 const char *buf, size_t len)
3304{
3305 struct slab_attribute *attribute;
3306 struct kmem_cache *s;
3307 int err;
3308
3309 attribute = to_slab_attr(attr);
3310 s = to_slab(kobj);
3311
3312 if (!attribute->store)
3313 return -EIO;
3314
3315 err = attribute->store(s, buf, len);
3316
3317 return err;
3318}
3319
3320static struct sysfs_ops slab_sysfs_ops = {
3321 .show = slab_attr_show,
3322 .store = slab_attr_store,
3323};
3324
3325static struct kobj_type slab_ktype = {
3326 .sysfs_ops = &slab_sysfs_ops,
3327};
3328
3329static int uevent_filter(struct kset *kset, struct kobject *kobj)
3330{
3331 struct kobj_type *ktype = get_ktype(kobj);
3332
3333 if (ktype == &slab_ktype)
3334 return 1;
3335 return 0;
3336}
3337
3338static struct kset_uevent_ops slab_uevent_ops = {
3339 .filter = uevent_filter,
3340};
3341
3342decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3343
3344#define ID_STR_LENGTH 64
3345
3346/* Create a unique string id for a slab cache:
3347 * format
3348 * :[flags-]size:[memory address of kmemcache]
3349 */
3350static char *create_unique_id(struct kmem_cache *s)
3351{
3352 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3353 char *p = name;
3354
3355 BUG_ON(!name);
3356
3357 *p++ = ':';
3358 /*
3359 * First flags affecting slabcache operations. We will only
3360 * get here for aliasable slabs so we do not need to support
3361 * too many flags. The flags here must cover all flags that
3362 * are matched during merging to guarantee that the id is
3363 * unique.
3364 */
3365 if (s->flags & SLAB_CACHE_DMA)
3366 *p++ = 'd';
3367 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3368 *p++ = 'a';
3369 if (s->flags & SLAB_DEBUG_FREE)
3370 *p++ = 'F';
3371 if (p != name + 1)
3372 *p++ = '-';
3373 p += sprintf(p, "%07d", s->size);
3374 BUG_ON(p > name + ID_STR_LENGTH - 1);
3375 return name;
3376}
3377
3378static int sysfs_slab_add(struct kmem_cache *s)
3379{
3380 int err;
3381 const char *name;
3382 int unmergeable;
3383
3384 if (slab_state < SYSFS)
3385 /* Defer until later */
3386 return 0;
3387
3388 unmergeable = slab_unmergeable(s);
3389 if (unmergeable) {
3390 /*
3391 * Slabcache can never be merged so we can use the name proper.
3392 * This is typically the case for debug situations. In that
3393 * case we can catch duplicate names easily.
3394 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07003395 sysfs_remove_link(&slab_subsys.kobj, s->name);
Christoph Lameter81819f02007-05-06 14:49:36 -07003396 name = s->name;
3397 } else {
3398 /*
3399 * Create a unique name for the slab as a target
3400 * for the symlinks.
3401 */
3402 name = create_unique_id(s);
3403 }
3404
3405 kobj_set_kset_s(s, slab_subsys);
3406 kobject_set_name(&s->kobj, name);
3407 kobject_init(&s->kobj);
3408 err = kobject_add(&s->kobj);
3409 if (err)
3410 return err;
3411
3412 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3413 if (err)
3414 return err;
3415 kobject_uevent(&s->kobj, KOBJ_ADD);
3416 if (!unmergeable) {
3417 /* Setup first alias */
3418 sysfs_slab_alias(s, s->name);
3419 kfree(name);
3420 }
3421 return 0;
3422}
3423
3424static void sysfs_slab_remove(struct kmem_cache *s)
3425{
3426 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3427 kobject_del(&s->kobj);
3428}
3429
3430/*
3431 * Need to buffer aliases during bootup until sysfs becomes
3432 * available lest we loose that information.
3433 */
3434struct saved_alias {
3435 struct kmem_cache *s;
3436 const char *name;
3437 struct saved_alias *next;
3438};
3439
3440struct saved_alias *alias_list;
3441
3442static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3443{
3444 struct saved_alias *al;
3445
3446 if (slab_state == SYSFS) {
3447 /*
3448 * If we have a leftover link then remove it.
3449 */
Linus Torvalds0f9008e2007-05-07 12:31:58 -07003450 sysfs_remove_link(&slab_subsys.kobj, name);
3451 return sysfs_create_link(&slab_subsys.kobj,
Christoph Lameter81819f02007-05-06 14:49:36 -07003452 &s->kobj, name);
3453 }
3454
3455 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3456 if (!al)
3457 return -ENOMEM;
3458
3459 al->s = s;
3460 al->name = name;
3461 al->next = alias_list;
3462 alias_list = al;
3463 return 0;
3464}
3465
3466static int __init slab_sysfs_init(void)
3467{
Christoph Lameter26a7bd02007-05-09 02:32:39 -07003468 struct list_head *h;
Christoph Lameter81819f02007-05-06 14:49:36 -07003469 int err;
3470
3471 err = subsystem_register(&slab_subsys);
3472 if (err) {
3473 printk(KERN_ERR "Cannot register slab subsystem.\n");
3474 return -ENOSYS;
3475 }
3476
Christoph Lameter26a7bd02007-05-09 02:32:39 -07003477 slab_state = SYSFS;
3478
3479 list_for_each(h, &slab_caches) {
3480 struct kmem_cache *s =
3481 container_of(h, struct kmem_cache, list);
3482
3483 err = sysfs_slab_add(s);
3484 BUG_ON(err);
3485 }
Christoph Lameter81819f02007-05-06 14:49:36 -07003486
3487 while (alias_list) {
3488 struct saved_alias *al = alias_list;
3489
3490 alias_list = alias_list->next;
3491 err = sysfs_slab_alias(al->s, al->name);
3492 BUG_ON(err);
3493 kfree(al);
3494 }
3495
3496 resiliency_test();
3497 return 0;
3498}
3499
3500__initcall(slab_sysfs_init);
Christoph Lameter81819f02007-05-06 14:49:36 -07003501#endif