blob: 4251917c5da1c123a478ec86159aa32081a7140c [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 *
100 * - Support DEBUG_SLAB_LEAK. Trouble is we do not know where the full
101 * slabs are in SLUB.
102 *
103 * - SLAB_DEBUG_INITIAL is not supported but I have never seen a use of
104 * it.
105 *
106 * - Variable sizing of the per node arrays
107 */
108
109/* Enable to test recovery from slab corruption on boot */
110#undef SLUB_RESILIENCY_TEST
111
112#if PAGE_SHIFT <= 12
113
114/*
115 * Small page size. Make sure that we do not fragment memory
116 */
117#define DEFAULT_MAX_ORDER 1
118#define DEFAULT_MIN_OBJECTS 4
119
120#else
121
122/*
123 * Large page machines are customarily able to handle larger
124 * page orders.
125 */
126#define DEFAULT_MAX_ORDER 2
127#define DEFAULT_MIN_OBJECTS 8
128
129#endif
130
131/*
132 * Flags from the regular SLAB that SLUB does not support:
133 */
134#define SLUB_UNIMPLEMENTED (SLAB_DEBUG_INITIAL)
135
Christoph Lametere95eed52007-05-06 14:49:44 -0700136/* Mininum number of partial slabs */
137#define MIN_PARTIAL 2
138
Christoph Lameter81819f02007-05-06 14:49:36 -0700139#define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
140 SLAB_POISON | SLAB_STORE_USER)
141/*
142 * Set of flags that will prevent slab merging
143 */
144#define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
145 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
146
147#define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
148 SLAB_CACHE_DMA)
149
150#ifndef ARCH_KMALLOC_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700151#define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700152#endif
153
154#ifndef ARCH_SLAB_MINALIGN
Christoph Lameter47bfdc02007-05-06 14:49:37 -0700155#define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
Christoph Lameter81819f02007-05-06 14:49:36 -0700156#endif
157
158/* Internal SLUB flags */
159#define __OBJECT_POISON 0x80000000 /* Poison object */
160
161static int kmem_size = sizeof(struct kmem_cache);
162
163#ifdef CONFIG_SMP
164static struct notifier_block slab_notifier;
165#endif
166
167static enum {
168 DOWN, /* No slab functionality available */
169 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
170 UP, /* Everything works */
171 SYSFS /* Sysfs up */
172} slab_state = DOWN;
173
174/* A list of all slab caches on the system */
175static DECLARE_RWSEM(slub_lock);
176LIST_HEAD(slab_caches);
177
178#ifdef CONFIG_SYSFS
179static int sysfs_slab_add(struct kmem_cache *);
180static int sysfs_slab_alias(struct kmem_cache *, const char *);
181static void sysfs_slab_remove(struct kmem_cache *);
182#else
183static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
184static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
185static void sysfs_slab_remove(struct kmem_cache *s) {}
186#endif
187
188/********************************************************************
189 * Core slab cache functions
190 *******************************************************************/
191
192int slab_is_available(void)
193{
194 return slab_state >= UP;
195}
196
197static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
198{
199#ifdef CONFIG_NUMA
200 return s->node[node];
201#else
202 return &s->local_node;
203#endif
204}
205
206/*
207 * Object debugging
208 */
209static void print_section(char *text, u8 *addr, unsigned int length)
210{
211 int i, offset;
212 int newline = 1;
213 char ascii[17];
214
215 ascii[16] = 0;
216
217 for (i = 0; i < length; i++) {
218 if (newline) {
219 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
220 newline = 0;
221 }
222 printk(" %02x", addr[i]);
223 offset = i % 16;
224 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
225 if (offset == 15) {
226 printk(" %s\n",ascii);
227 newline = 1;
228 }
229 }
230 if (!newline) {
231 i %= 16;
232 while (i < 16) {
233 printk(" ");
234 ascii[i] = ' ';
235 i++;
236 }
237 printk(" %s\n", ascii);
238 }
239}
240
241/*
242 * Slow version of get and set free pointer.
243 *
244 * This requires touching the cache lines of kmem_cache.
245 * The offset can also be obtained from the page. In that
246 * case it is in the cacheline that we already need to touch.
247 */
248static void *get_freepointer(struct kmem_cache *s, void *object)
249{
250 return *(void **)(object + s->offset);
251}
252
253static void set_freepointer(struct kmem_cache *s, void *object, void *fp)
254{
255 *(void **)(object + s->offset) = fp;
256}
257
258/*
259 * Tracking user of a slab.
260 */
261struct track {
262 void *addr; /* Called from address */
263 int cpu; /* Was running on cpu */
264 int pid; /* Pid context */
265 unsigned long when; /* When did the operation occur */
266};
267
268enum track_item { TRACK_ALLOC, TRACK_FREE };
269
270static struct track *get_track(struct kmem_cache *s, void *object,
271 enum track_item alloc)
272{
273 struct track *p;
274
275 if (s->offset)
276 p = object + s->offset + sizeof(void *);
277 else
278 p = object + s->inuse;
279
280 return p + alloc;
281}
282
283static void set_track(struct kmem_cache *s, void *object,
284 enum track_item alloc, void *addr)
285{
286 struct track *p;
287
288 if (s->offset)
289 p = object + s->offset + sizeof(void *);
290 else
291 p = object + s->inuse;
292
293 p += alloc;
294 if (addr) {
295 p->addr = addr;
296 p->cpu = smp_processor_id();
297 p->pid = current ? current->pid : -1;
298 p->when = jiffies;
299 } else
300 memset(p, 0, sizeof(struct track));
301}
302
Christoph Lameter81819f02007-05-06 14:49:36 -0700303static void init_tracking(struct kmem_cache *s, void *object)
304{
305 if (s->flags & SLAB_STORE_USER) {
306 set_track(s, object, TRACK_FREE, NULL);
307 set_track(s, object, TRACK_ALLOC, NULL);
308 }
309}
310
311static void print_track(const char *s, struct track *t)
312{
313 if (!t->addr)
314 return;
315
316 printk(KERN_ERR "%s: ", s);
317 __print_symbol("%s", (unsigned long)t->addr);
318 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
319}
320
321static void print_trailer(struct kmem_cache *s, u8 *p)
322{
323 unsigned int off; /* Offset of last byte */
324
325 if (s->flags & SLAB_RED_ZONE)
326 print_section("Redzone", p + s->objsize,
327 s->inuse - s->objsize);
328
329 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
330 p + s->offset,
331 get_freepointer(s, p));
332
333 if (s->offset)
334 off = s->offset + sizeof(void *);
335 else
336 off = s->inuse;
337
338 if (s->flags & SLAB_STORE_USER) {
339 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
340 print_track("Last free ", get_track(s, p, TRACK_FREE));
341 off += 2 * sizeof(struct track);
342 }
343
344 if (off != s->size)
345 /* Beginning of the filler is the free pointer */
346 print_section("Filler", p + off, s->size - off);
347}
348
349static void object_err(struct kmem_cache *s, struct page *page,
350 u8 *object, char *reason)
351{
352 u8 *addr = page_address(page);
353
354 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
355 s->name, reason, object, page);
356 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
357 object - addr, page->flags, page->inuse, page->freelist);
358 if (object > addr + 16)
359 print_section("Bytes b4", object - 16, 16);
360 print_section("Object", object, min(s->objsize, 128));
361 print_trailer(s, object);
362 dump_stack();
363}
364
365static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
366{
367 va_list args;
368 char buf[100];
369
370 va_start(args, reason);
371 vsnprintf(buf, sizeof(buf), reason, args);
372 va_end(args);
373 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
374 page);
375 dump_stack();
376}
377
378static void init_object(struct kmem_cache *s, void *object, int active)
379{
380 u8 *p = object;
381
382 if (s->flags & __OBJECT_POISON) {
383 memset(p, POISON_FREE, s->objsize - 1);
384 p[s->objsize -1] = POISON_END;
385 }
386
387 if (s->flags & SLAB_RED_ZONE)
388 memset(p + s->objsize,
389 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
390 s->inuse - s->objsize);
391}
392
393static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
394{
395 while (bytes) {
396 if (*start != (u8)value)
397 return 0;
398 start++;
399 bytes--;
400 }
401 return 1;
402}
403
404
405static int check_valid_pointer(struct kmem_cache *s, struct page *page,
406 void *object)
407{
408 void *base;
409
410 if (!object)
411 return 1;
412
413 base = page_address(page);
414 if (object < base || object >= base + s->objects * s->size ||
415 (object - base) % s->size) {
416 return 0;
417 }
418
419 return 1;
420}
421
422/*
423 * Object layout:
424 *
425 * object address
426 * Bytes of the object to be managed.
427 * If the freepointer may overlay the object then the free
428 * pointer is the first word of the object.
429 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
430 * 0xa5 (POISON_END)
431 *
432 * object + s->objsize
433 * Padding to reach word boundary. This is also used for Redzoning.
434 * Padding is extended to word size if Redzoning is enabled
435 * and objsize == inuse.
436 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
437 * 0xcc (RED_ACTIVE) for objects in use.
438 *
439 * object + s->inuse
440 * A. Free pointer (if we cannot overwrite object on free)
441 * B. Tracking data for SLAB_STORE_USER
442 * C. Padding to reach required alignment boundary
443 * Padding is done using 0x5a (POISON_INUSE)
444 *
445 * object + s->size
446 *
447 * If slabcaches are merged then the objsize and inuse boundaries are to
448 * be ignored. And therefore no slab options that rely on these boundaries
449 * may be used with merged slabcaches.
450 */
451
452static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
453 void *from, void *to)
454{
455 printk(KERN_ERR "@@@ SLUB: %s Restoring %s (0x%x) from 0x%p-0x%p\n",
456 s->name, message, data, from, to - 1);
457 memset(from, data, to - from);
458}
459
460static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
461{
462 unsigned long off = s->inuse; /* The end of info */
463
464 if (s->offset)
465 /* Freepointer is placed after the object. */
466 off += sizeof(void *);
467
468 if (s->flags & SLAB_STORE_USER)
469 /* We also have user information there */
470 off += 2 * sizeof(struct track);
471
472 if (s->size == off)
473 return 1;
474
475 if (check_bytes(p + off, POISON_INUSE, s->size - off))
476 return 1;
477
478 object_err(s, page, p, "Object padding check fails");
479
480 /*
481 * Restore padding
482 */
483 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
484 return 0;
485}
486
487static int slab_pad_check(struct kmem_cache *s, struct page *page)
488{
489 u8 *p;
490 int length, remainder;
491
492 if (!(s->flags & SLAB_POISON))
493 return 1;
494
495 p = page_address(page);
496 length = s->objects * s->size;
497 remainder = (PAGE_SIZE << s->order) - length;
498 if (!remainder)
499 return 1;
500
501 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
502 printk(KERN_ERR "SLUB: %s slab 0x%p: Padding fails check\n",
503 s->name, p);
504 dump_stack();
505 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
506 p + length + remainder);
507 return 0;
508 }
509 return 1;
510}
511
512static int check_object(struct kmem_cache *s, struct page *page,
513 void *object, int active)
514{
515 u8 *p = object;
516 u8 *endobject = object + s->objsize;
517
518 if (s->flags & SLAB_RED_ZONE) {
519 unsigned int red =
520 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
521
522 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
523 object_err(s, page, object,
524 active ? "Redzone Active" : "Redzone Inactive");
525 restore_bytes(s, "redzone", red,
526 endobject, object + s->inuse);
527 return 0;
528 }
529 } else {
530 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
531 !check_bytes(endobject, POISON_INUSE,
532 s->inuse - s->objsize)) {
533 object_err(s, page, p, "Alignment padding check fails");
534 /*
535 * Fix it so that there will not be another report.
536 *
537 * Hmmm... We may be corrupting an object that now expects
538 * to be longer than allowed.
539 */
540 restore_bytes(s, "alignment padding", POISON_INUSE,
541 endobject, object + s->inuse);
542 }
543 }
544
545 if (s->flags & SLAB_POISON) {
546 if (!active && (s->flags & __OBJECT_POISON) &&
547 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
548 p[s->objsize - 1] != POISON_END)) {
549
550 object_err(s, page, p, "Poison check failed");
551 restore_bytes(s, "Poison", POISON_FREE,
552 p, p + s->objsize -1);
553 restore_bytes(s, "Poison", POISON_END,
554 p + s->objsize - 1, p + s->objsize);
555 return 0;
556 }
557 /*
558 * check_pad_bytes cleans up on its own.
559 */
560 check_pad_bytes(s, page, p);
561 }
562
563 if (!s->offset && active)
564 /*
565 * Object and freepointer overlap. Cannot check
566 * freepointer while object is allocated.
567 */
568 return 1;
569
570 /* Check free pointer validity */
571 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
572 object_err(s, page, p, "Freepointer corrupt");
573 /*
574 * No choice but to zap it and thus loose the remainder
575 * of the free objects in this slab. May cause
576 * another error because the object count maybe
577 * wrong now.
578 */
579 set_freepointer(s, p, NULL);
580 return 0;
581 }
582 return 1;
583}
584
585static int check_slab(struct kmem_cache *s, struct page *page)
586{
587 VM_BUG_ON(!irqs_disabled());
588
589 if (!PageSlab(page)) {
590 printk(KERN_ERR "SLUB: %s Not a valid slab page @0x%p "
591 "flags=%lx mapping=0x%p count=%d \n",
592 s->name, page, page->flags, page->mapping,
593 page_count(page));
594 return 0;
595 }
596 if (page->offset * sizeof(void *) != s->offset) {
597 printk(KERN_ERR "SLUB: %s Corrupted offset %lu in slab @0x%p"
598 " flags=0x%lx mapping=0x%p count=%d\n",
599 s->name,
600 (unsigned long)(page->offset * sizeof(void *)),
601 page,
602 page->flags,
603 page->mapping,
604 page_count(page));
605 dump_stack();
606 return 0;
607 }
608 if (page->inuse > s->objects) {
609 printk(KERN_ERR "SLUB: %s Inuse %u > max %u in slab "
610 "page @0x%p flags=%lx mapping=0x%p count=%d\n",
611 s->name, page->inuse, s->objects, page, page->flags,
612 page->mapping, page_count(page));
613 dump_stack();
614 return 0;
615 }
616 /* Slab_pad_check fixes things up after itself */
617 slab_pad_check(s, page);
618 return 1;
619}
620
621/*
622 * Determine if a certain object on a page is on the freelist and
623 * therefore free. Must hold the slab lock for cpu slabs to
624 * guarantee that the chains are consistent.
625 */
626static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
627{
628 int nr = 0;
629 void *fp = page->freelist;
630 void *object = NULL;
631
632 while (fp && nr <= s->objects) {
633 if (fp == search)
634 return 1;
635 if (!check_valid_pointer(s, page, fp)) {
636 if (object) {
637 object_err(s, page, object,
638 "Freechain corrupt");
639 set_freepointer(s, object, NULL);
640 break;
641 } else {
642 printk(KERN_ERR "SLUB: %s slab 0x%p "
643 "freepointer 0x%p corrupted.\n",
644 s->name, page, fp);
645 dump_stack();
646 page->freelist = NULL;
647 page->inuse = s->objects;
648 return 0;
649 }
650 break;
651 }
652 object = fp;
653 fp = get_freepointer(s, object);
654 nr++;
655 }
656
657 if (page->inuse != s->objects - nr) {
658 printk(KERN_ERR "slab %s: page 0x%p wrong object count."
659 " counter is %d but counted were %d\n",
660 s->name, page, page->inuse,
661 s->objects - nr);
662 page->inuse = s->objects - nr;
663 }
664 return search == NULL;
665}
666
Christoph Lameter643b1132007-05-06 14:49:42 -0700667/*
668 * Tracking of fully allocated slabs for debugging
669 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700670static void add_full(struct kmem_cache_node *n, struct page *page)
Christoph Lameter643b1132007-05-06 14:49:42 -0700671{
Christoph Lameter643b1132007-05-06 14:49:42 -0700672 spin_lock(&n->list_lock);
673 list_add(&page->lru, &n->full);
674 spin_unlock(&n->list_lock);
675}
676
677static void remove_full(struct kmem_cache *s, struct page *page)
678{
679 struct kmem_cache_node *n;
680
681 if (!(s->flags & SLAB_STORE_USER))
682 return;
683
684 n = get_node(s, page_to_nid(page));
685
686 spin_lock(&n->list_lock);
687 list_del(&page->lru);
688 spin_unlock(&n->list_lock);
689}
690
Christoph Lameter81819f02007-05-06 14:49:36 -0700691static int alloc_object_checks(struct kmem_cache *s, struct page *page,
692 void *object)
693{
694 if (!check_slab(s, page))
695 goto bad;
696
697 if (object && !on_freelist(s, page, object)) {
698 printk(KERN_ERR "SLUB: %s Object 0x%p@0x%p "
699 "already allocated.\n",
700 s->name, object, page);
701 goto dump;
702 }
703
704 if (!check_valid_pointer(s, page, object)) {
705 object_err(s, page, object, "Freelist Pointer check fails");
706 goto dump;
707 }
708
709 if (!object)
710 return 1;
711
712 if (!check_object(s, page, object, 0))
713 goto bad;
714 init_object(s, object, 1);
715
716 if (s->flags & SLAB_TRACE) {
717 printk(KERN_INFO "TRACE %s alloc 0x%p inuse=%d fp=0x%p\n",
718 s->name, object, page->inuse,
719 page->freelist);
720 dump_stack();
721 }
722 return 1;
723dump:
724 dump_stack();
725bad:
726 if (PageSlab(page)) {
727 /*
728 * If this is a slab page then lets do the best we can
729 * to avoid issues in the future. Marking all objects
730 * as used avoids touching the remainder.
731 */
732 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
733 s->name, page);
734 page->inuse = s->objects;
735 page->freelist = NULL;
736 /* Fix up fields that may be corrupted */
737 page->offset = s->offset / sizeof(void *);
738 }
739 return 0;
740}
741
742static int free_object_checks(struct kmem_cache *s, struct page *page,
743 void *object)
744{
745 if (!check_slab(s, page))
746 goto fail;
747
748 if (!check_valid_pointer(s, page, object)) {
749 printk(KERN_ERR "SLUB: %s slab 0x%p invalid "
750 "object pointer 0x%p\n",
751 s->name, page, object);
752 goto fail;
753 }
754
755 if (on_freelist(s, page, object)) {
756 printk(KERN_ERR "SLUB: %s slab 0x%p object "
757 "0x%p already free.\n", s->name, page, object);
758 goto fail;
759 }
760
761 if (!check_object(s, page, object, 1))
762 return 0;
763
764 if (unlikely(s != page->slab)) {
765 if (!PageSlab(page))
766 printk(KERN_ERR "slab_free %s size %d: attempt to"
767 "free object(0x%p) outside of slab.\n",
768 s->name, s->size, object);
769 else
770 if (!page->slab)
771 printk(KERN_ERR
772 "slab_free : no slab(NULL) for object 0x%p.\n",
773 object);
774 else
775 printk(KERN_ERR "slab_free %s(%d): object at 0x%p"
776 " belongs to slab %s(%d)\n",
777 s->name, s->size, object,
778 page->slab->name, page->slab->size);
779 goto fail;
780 }
781 if (s->flags & SLAB_TRACE) {
782 printk(KERN_INFO "TRACE %s free 0x%p inuse=%d fp=0x%p\n",
783 s->name, object, page->inuse,
784 page->freelist);
785 print_section("Object", object, s->objsize);
786 dump_stack();
787 }
788 init_object(s, object, 0);
789 return 1;
790fail:
791 dump_stack();
792 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
793 s->name, page, object);
794 return 0;
795}
796
797/*
798 * Slab allocation and freeing
799 */
800static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
801{
802 struct page * page;
803 int pages = 1 << s->order;
804
805 if (s->order)
806 flags |= __GFP_COMP;
807
808 if (s->flags & SLAB_CACHE_DMA)
809 flags |= SLUB_DMA;
810
811 if (node == -1)
812 page = alloc_pages(flags, s->order);
813 else
814 page = alloc_pages_node(node, flags, s->order);
815
816 if (!page)
817 return NULL;
818
819 mod_zone_page_state(page_zone(page),
820 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
821 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
822 pages);
823
824 return page;
825}
826
827static void setup_object(struct kmem_cache *s, struct page *page,
828 void *object)
829{
830 if (PageError(page)) {
831 init_object(s, object, 0);
832 init_tracking(s, object);
833 }
834
835 if (unlikely(s->ctor)) {
836 int mode = SLAB_CTOR_CONSTRUCTOR;
837
838 if (!(s->flags & __GFP_WAIT))
839 mode |= SLAB_CTOR_ATOMIC;
840
841 s->ctor(object, s, mode);
842 }
843}
844
845static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
846{
847 struct page *page;
848 struct kmem_cache_node *n;
849 void *start;
850 void *end;
851 void *last;
852 void *p;
853
854 if (flags & __GFP_NO_GROW)
855 return NULL;
856
857 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
858
859 if (flags & __GFP_WAIT)
860 local_irq_enable();
861
862 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
863 if (!page)
864 goto out;
865
866 n = get_node(s, page_to_nid(page));
867 if (n)
868 atomic_long_inc(&n->nr_slabs);
869 page->offset = s->offset / sizeof(void *);
870 page->slab = s;
871 page->flags |= 1 << PG_slab;
872 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
873 SLAB_STORE_USER | SLAB_TRACE))
874 page->flags |= 1 << PG_error;
875
876 start = page_address(page);
877 end = start + s->objects * s->size;
878
879 if (unlikely(s->flags & SLAB_POISON))
880 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
881
882 last = start;
883 for (p = start + s->size; p < end; p += s->size) {
884 setup_object(s, page, last);
885 set_freepointer(s, last, p);
886 last = p;
887 }
888 setup_object(s, page, last);
889 set_freepointer(s, last, NULL);
890
891 page->freelist = start;
892 page->inuse = 0;
893out:
894 if (flags & __GFP_WAIT)
895 local_irq_disable();
896 return page;
897}
898
899static void __free_slab(struct kmem_cache *s, struct page *page)
900{
901 int pages = 1 << s->order;
902
903 if (unlikely(PageError(page) || s->dtor)) {
904 void *start = page_address(page);
905 void *end = start + (pages << PAGE_SHIFT);
906 void *p;
907
908 slab_pad_check(s, page);
909 for (p = start; p <= end - s->size; p += s->size) {
910 if (s->dtor)
911 s->dtor(p, s, 0);
912 check_object(s, page, p, 0);
913 }
914 }
915
916 mod_zone_page_state(page_zone(page),
917 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
918 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
919 - pages);
920
921 page->mapping = NULL;
922 __free_pages(page, s->order);
923}
924
925static void rcu_free_slab(struct rcu_head *h)
926{
927 struct page *page;
928
929 page = container_of((struct list_head *)h, struct page, lru);
930 __free_slab(page->slab, page);
931}
932
933static void free_slab(struct kmem_cache *s, struct page *page)
934{
935 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
936 /*
937 * RCU free overloads the RCU head over the LRU
938 */
939 struct rcu_head *head = (void *)&page->lru;
940
941 call_rcu(head, rcu_free_slab);
942 } else
943 __free_slab(s, page);
944}
945
946static void discard_slab(struct kmem_cache *s, struct page *page)
947{
948 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
949
950 atomic_long_dec(&n->nr_slabs);
951 reset_page_mapcount(page);
952 page->flags &= ~(1 << PG_slab | 1 << PG_error);
953 free_slab(s, page);
954}
955
956/*
957 * Per slab locking using the pagelock
958 */
959static __always_inline void slab_lock(struct page *page)
960{
961 bit_spin_lock(PG_locked, &page->flags);
962}
963
964static __always_inline void slab_unlock(struct page *page)
965{
966 bit_spin_unlock(PG_locked, &page->flags);
967}
968
969static __always_inline int slab_trylock(struct page *page)
970{
971 int rc = 1;
972
973 rc = bit_spin_trylock(PG_locked, &page->flags);
974 return rc;
975}
976
977/*
978 * Management of partially allocated slabs
979 */
Christoph Lametere95eed52007-05-06 14:49:44 -0700980static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
Christoph Lameter81819f02007-05-06 14:49:36 -0700981{
Christoph Lametere95eed52007-05-06 14:49:44 -0700982 spin_lock(&n->list_lock);
983 n->nr_partial++;
984 list_add_tail(&page->lru, &n->partial);
985 spin_unlock(&n->list_lock);
986}
Christoph Lameter81819f02007-05-06 14:49:36 -0700987
Christoph Lametere95eed52007-05-06 14:49:44 -0700988static void add_partial(struct kmem_cache_node *n, struct page *page)
989{
Christoph Lameter81819f02007-05-06 14:49:36 -0700990 spin_lock(&n->list_lock);
991 n->nr_partial++;
992 list_add(&page->lru, &n->partial);
993 spin_unlock(&n->list_lock);
994}
995
996static void remove_partial(struct kmem_cache *s,
997 struct page *page)
998{
999 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1000
1001 spin_lock(&n->list_lock);
1002 list_del(&page->lru);
1003 n->nr_partial--;
1004 spin_unlock(&n->list_lock);
1005}
1006
1007/*
1008 * Lock page and remove it from the partial list
1009 *
1010 * Must hold list_lock
1011 */
1012static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page)
1013{
1014 if (slab_trylock(page)) {
1015 list_del(&page->lru);
1016 n->nr_partial--;
1017 return 1;
1018 }
1019 return 0;
1020}
1021
1022/*
1023 * Try to get a partial slab from a specific node
1024 */
1025static struct page *get_partial_node(struct kmem_cache_node *n)
1026{
1027 struct page *page;
1028
1029 /*
1030 * Racy check. If we mistakenly see no partial slabs then we
1031 * just allocate an empty slab. If we mistakenly try to get a
1032 * partial slab then get_partials() will return NULL.
1033 */
1034 if (!n || !n->nr_partial)
1035 return NULL;
1036
1037 spin_lock(&n->list_lock);
1038 list_for_each_entry(page, &n->partial, lru)
1039 if (lock_and_del_slab(n, page))
1040 goto out;
1041 page = NULL;
1042out:
1043 spin_unlock(&n->list_lock);
1044 return page;
1045}
1046
1047/*
1048 * Get a page from somewhere. Search in increasing NUMA
1049 * distances.
1050 */
1051static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1052{
1053#ifdef CONFIG_NUMA
1054 struct zonelist *zonelist;
1055 struct zone **z;
1056 struct page *page;
1057
1058 /*
1059 * The defrag ratio allows to configure the tradeoffs between
1060 * inter node defragmentation and node local allocations.
1061 * A lower defrag_ratio increases the tendency to do local
1062 * allocations instead of scanning throught the partial
1063 * lists on other nodes.
1064 *
1065 * If defrag_ratio is set to 0 then kmalloc() always
1066 * returns node local objects. If its higher then kmalloc()
1067 * may return off node objects in order to avoid fragmentation.
1068 *
1069 * A higher ratio means slabs may be taken from other nodes
1070 * thus reducing the number of partial slabs on those nodes.
1071 *
1072 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1073 * defrag_ratio = 1000) then every (well almost) allocation
1074 * will first attempt to defrag slab caches on other nodes. This
1075 * means scanning over all nodes to look for partial slabs which
1076 * may be a bit expensive to do on every slab allocation.
1077 */
1078 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1079 return NULL;
1080
1081 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1082 ->node_zonelists[gfp_zone(flags)];
1083 for (z = zonelist->zones; *z; z++) {
1084 struct kmem_cache_node *n;
1085
1086 n = get_node(s, zone_to_nid(*z));
1087
1088 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
Christoph Lametere95eed52007-05-06 14:49:44 -07001089 n->nr_partial > MIN_PARTIAL) {
Christoph Lameter81819f02007-05-06 14:49:36 -07001090 page = get_partial_node(n);
1091 if (page)
1092 return page;
1093 }
1094 }
1095#endif
1096 return NULL;
1097}
1098
1099/*
1100 * Get a partial page, lock it and return it.
1101 */
1102static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1103{
1104 struct page *page;
1105 int searchnode = (node == -1) ? numa_node_id() : node;
1106
1107 page = get_partial_node(get_node(s, searchnode));
1108 if (page || (flags & __GFP_THISNODE))
1109 return page;
1110
1111 return get_any_partial(s, flags);
1112}
1113
1114/*
1115 * Move a page back to the lists.
1116 *
1117 * Must be called with the slab lock held.
1118 *
1119 * On exit the slab lock will have been dropped.
1120 */
1121static void putback_slab(struct kmem_cache *s, struct page *page)
1122{
Christoph Lametere95eed52007-05-06 14:49:44 -07001123 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1124
Christoph Lameter81819f02007-05-06 14:49:36 -07001125 if (page->inuse) {
Christoph Lametere95eed52007-05-06 14:49:44 -07001126
Christoph Lameter81819f02007-05-06 14:49:36 -07001127 if (page->freelist)
Christoph Lametere95eed52007-05-06 14:49:44 -07001128 add_partial(n, page);
1129 else if (PageError(page) && (s->flags & SLAB_STORE_USER))
1130 add_full(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001131 slab_unlock(page);
Christoph Lametere95eed52007-05-06 14:49:44 -07001132
Christoph Lameter81819f02007-05-06 14:49:36 -07001133 } else {
Christoph Lametere95eed52007-05-06 14:49:44 -07001134 if (n->nr_partial < MIN_PARTIAL) {
1135 /*
1136 * Adding an empty page to the partial slabs in order
1137 * to avoid page allocator overhead. This page needs to
1138 * come after all the others that are not fully empty
1139 * in order to make sure that we do maximum
1140 * defragmentation.
1141 */
1142 add_partial_tail(n, page);
1143 slab_unlock(page);
1144 } else {
1145 slab_unlock(page);
1146 discard_slab(s, page);
1147 }
Christoph Lameter81819f02007-05-06 14:49:36 -07001148 }
1149}
1150
1151/*
1152 * Remove the cpu slab
1153 */
1154static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1155{
1156 s->cpu_slab[cpu] = NULL;
1157 ClearPageActive(page);
1158
1159 putback_slab(s, page);
1160}
1161
1162static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1163{
1164 slab_lock(page);
1165 deactivate_slab(s, page, cpu);
1166}
1167
1168/*
1169 * Flush cpu slab.
1170 * Called from IPI handler with interrupts disabled.
1171 */
1172static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1173{
1174 struct page *page = s->cpu_slab[cpu];
1175
1176 if (likely(page))
1177 flush_slab(s, page, cpu);
1178}
1179
1180static void flush_cpu_slab(void *d)
1181{
1182 struct kmem_cache *s = d;
1183 int cpu = smp_processor_id();
1184
1185 __flush_cpu_slab(s, cpu);
1186}
1187
1188static void flush_all(struct kmem_cache *s)
1189{
1190#ifdef CONFIG_SMP
1191 on_each_cpu(flush_cpu_slab, s, 1, 1);
1192#else
1193 unsigned long flags;
1194
1195 local_irq_save(flags);
1196 flush_cpu_slab(s);
1197 local_irq_restore(flags);
1198#endif
1199}
1200
1201/*
1202 * slab_alloc is optimized to only modify two cachelines on the fast path
1203 * (aside from the stack):
1204 *
1205 * 1. The page struct
1206 * 2. The first cacheline of the object to be allocated.
1207 *
1208 * The only cache lines that are read (apart from code) is the
1209 * per cpu array in the kmem_cache struct.
1210 *
1211 * Fastpath is not possible if we need to get a new slab or have
1212 * debugging enabled (which means all slabs are marked with PageError)
1213 */
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001214static void *slab_alloc(struct kmem_cache *s,
1215 gfp_t gfpflags, int node, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001216{
1217 struct page *page;
1218 void **object;
1219 unsigned long flags;
1220 int cpu;
1221
1222 local_irq_save(flags);
1223 cpu = smp_processor_id();
1224 page = s->cpu_slab[cpu];
1225 if (!page)
1226 goto new_slab;
1227
1228 slab_lock(page);
1229 if (unlikely(node != -1 && page_to_nid(page) != node))
1230 goto another_slab;
1231redo:
1232 object = page->freelist;
1233 if (unlikely(!object))
1234 goto another_slab;
1235 if (unlikely(PageError(page)))
1236 goto debug;
1237
1238have_object:
1239 page->inuse++;
1240 page->freelist = object[page->offset];
1241 slab_unlock(page);
1242 local_irq_restore(flags);
1243 return object;
1244
1245another_slab:
1246 deactivate_slab(s, page, cpu);
1247
1248new_slab:
1249 page = get_partial(s, gfpflags, node);
1250 if (likely(page)) {
1251have_slab:
1252 s->cpu_slab[cpu] = page;
1253 SetPageActive(page);
1254 goto redo;
1255 }
1256
1257 page = new_slab(s, gfpflags, node);
1258 if (page) {
1259 cpu = smp_processor_id();
1260 if (s->cpu_slab[cpu]) {
1261 /*
1262 * Someone else populated the cpu_slab while we enabled
1263 * interrupts, or we have got scheduled on another cpu.
1264 * The page may not be on the requested node.
1265 */
1266 if (node == -1 ||
1267 page_to_nid(s->cpu_slab[cpu]) == node) {
1268 /*
1269 * Current cpuslab is acceptable and we
1270 * want the current one since its cache hot
1271 */
1272 discard_slab(s, page);
1273 page = s->cpu_slab[cpu];
1274 slab_lock(page);
1275 goto redo;
1276 }
1277 /* Dump the current slab */
1278 flush_slab(s, s->cpu_slab[cpu], cpu);
1279 }
1280 slab_lock(page);
1281 goto have_slab;
1282 }
1283 local_irq_restore(flags);
1284 return NULL;
1285debug:
1286 if (!alloc_object_checks(s, page, object))
1287 goto another_slab;
1288 if (s->flags & SLAB_STORE_USER)
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001289 set_track(s, object, TRACK_ALLOC, addr);
Christoph Lameter81819f02007-05-06 14:49:36 -07001290 goto have_object;
1291}
1292
1293void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1294{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001295 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001296}
1297EXPORT_SYMBOL(kmem_cache_alloc);
1298
1299#ifdef CONFIG_NUMA
1300void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1301{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001302 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001303}
1304EXPORT_SYMBOL(kmem_cache_alloc_node);
1305#endif
1306
1307/*
1308 * The fastpath only writes the cacheline of the page struct and the first
1309 * cacheline of the object.
1310 *
1311 * No special cachelines need to be read
1312 */
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001313static void slab_free(struct kmem_cache *s, struct page *page,
1314 void *x, void *addr)
Christoph Lameter81819f02007-05-06 14:49:36 -07001315{
1316 void *prior;
1317 void **object = (void *)x;
1318 unsigned long flags;
1319
1320 local_irq_save(flags);
1321 slab_lock(page);
1322
1323 if (unlikely(PageError(page)))
1324 goto debug;
1325checks_ok:
1326 prior = object[page->offset] = page->freelist;
1327 page->freelist = object;
1328 page->inuse--;
1329
1330 if (unlikely(PageActive(page)))
1331 /*
1332 * Cpu slabs are never on partial lists and are
1333 * never freed.
1334 */
1335 goto out_unlock;
1336
1337 if (unlikely(!page->inuse))
1338 goto slab_empty;
1339
1340 /*
1341 * Objects left in the slab. If it
1342 * was not on the partial list before
1343 * then add it.
1344 */
1345 if (unlikely(!prior))
Christoph Lametere95eed52007-05-06 14:49:44 -07001346 add_partial(get_node(s, page_to_nid(page)), page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001347
1348out_unlock:
1349 slab_unlock(page);
1350 local_irq_restore(flags);
1351 return;
1352
1353slab_empty:
1354 if (prior)
1355 /*
Christoph Lameter643b1132007-05-06 14:49:42 -07001356 * Slab on the partial list.
Christoph Lameter81819f02007-05-06 14:49:36 -07001357 */
1358 remove_partial(s, page);
1359
1360 slab_unlock(page);
1361 discard_slab(s, page);
1362 local_irq_restore(flags);
1363 return;
1364
1365debug:
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001366 if (!free_object_checks(s, page, x))
1367 goto out_unlock;
Christoph Lameter643b1132007-05-06 14:49:42 -07001368 if (!PageActive(page) && !page->freelist)
1369 remove_full(s, page);
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001370 if (s->flags & SLAB_STORE_USER)
1371 set_track(s, x, TRACK_FREE, addr);
1372 goto checks_ok;
Christoph Lameter81819f02007-05-06 14:49:36 -07001373}
1374
1375void kmem_cache_free(struct kmem_cache *s, void *x)
1376{
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001377 struct page *page;
Christoph Lameter81819f02007-05-06 14:49:36 -07001378
Christoph Lameterb49af682007-05-06 14:49:41 -07001379 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001380
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07001381 slab_free(s, page, x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07001382}
1383EXPORT_SYMBOL(kmem_cache_free);
1384
1385/* Figure out on which slab object the object resides */
1386static struct page *get_object_page(const void *x)
1387{
Christoph Lameterb49af682007-05-06 14:49:41 -07001388 struct page *page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07001389
1390 if (!PageSlab(page))
1391 return NULL;
1392
1393 return page;
1394}
1395
1396/*
1397 * kmem_cache_open produces objects aligned at "size" and the first object
1398 * is placed at offset 0 in the slab (We have no metainformation on the
1399 * slab, all slabs are in essence "off slab").
1400 *
1401 * In order to get the desired alignment one just needs to align the
1402 * size.
1403 *
1404 * Notice that the allocation order determines the sizes of the per cpu
1405 * caches. Each processor has always one slab available for allocations.
1406 * Increasing the allocation order reduces the number of times that slabs
1407 * must be moved on and off the partial lists and therefore may influence
1408 * locking overhead.
1409 *
1410 * The offset is used to relocate the free list link in each object. It is
1411 * therefore possible to move the free list link behind the object. This
1412 * is necessary for RCU to work properly and also useful for debugging.
1413 */
1414
1415/*
1416 * Mininum / Maximum order of slab pages. This influences locking overhead
1417 * and slab fragmentation. A higher order reduces the number of partial slabs
1418 * and increases the number of allocations possible without having to
1419 * take the list_lock.
1420 */
1421static int slub_min_order;
1422static int slub_max_order = DEFAULT_MAX_ORDER;
1423
1424/*
1425 * Minimum number of objects per slab. This is necessary in order to
1426 * reduce locking overhead. Similar to the queue size in SLAB.
1427 */
1428static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1429
1430/*
1431 * Merge control. If this is set then no merging of slab caches will occur.
1432 */
1433static int slub_nomerge;
1434
1435/*
1436 * Debug settings:
1437 */
1438static int slub_debug;
1439
1440static char *slub_debug_slabs;
1441
1442/*
1443 * Calculate the order of allocation given an slab object size.
1444 *
1445 * The order of allocation has significant impact on other elements
1446 * of the system. Generally order 0 allocations should be preferred
1447 * since they do not cause fragmentation in the page allocator. Larger
1448 * objects may have problems with order 0 because there may be too much
1449 * space left unused in a slab. We go to a higher order if more than 1/8th
1450 * of the slab would be wasted.
1451 *
1452 * In order to reach satisfactory performance we must ensure that
1453 * a minimum number of objects is in one slab. Otherwise we may
1454 * generate too much activity on the partial lists. This is less a
1455 * concern for large slabs though. slub_max_order specifies the order
1456 * where we begin to stop considering the number of objects in a slab.
1457 *
1458 * Higher order allocations also allow the placement of more objects
1459 * in a slab and thereby reduce object handling overhead. If the user
1460 * has requested a higher mininum order then we start with that one
1461 * instead of zero.
1462 */
1463static int calculate_order(int size)
1464{
1465 int order;
1466 int rem;
1467
1468 for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT);
1469 order < MAX_ORDER; order++) {
1470 unsigned long slab_size = PAGE_SIZE << order;
1471
1472 if (slub_max_order > order &&
1473 slab_size < slub_min_objects * size)
1474 continue;
1475
1476 if (slab_size < size)
1477 continue;
1478
1479 rem = slab_size % size;
1480
1481 if (rem <= (PAGE_SIZE << order) / 8)
1482 break;
1483
1484 }
1485 if (order >= MAX_ORDER)
1486 return -E2BIG;
1487 return order;
1488}
1489
1490/*
1491 * Function to figure out which alignment to use from the
1492 * various ways of specifying it.
1493 */
1494static unsigned long calculate_alignment(unsigned long flags,
1495 unsigned long align, unsigned long size)
1496{
1497 /*
1498 * If the user wants hardware cache aligned objects then
1499 * follow that suggestion if the object is sufficiently
1500 * large.
1501 *
1502 * The hardware cache alignment cannot override the
1503 * specified alignment though. If that is greater
1504 * then use it.
1505 */
1506 if ((flags & (SLAB_MUST_HWCACHE_ALIGN | SLAB_HWCACHE_ALIGN)) &&
1507 size > L1_CACHE_BYTES / 2)
1508 return max_t(unsigned long, align, L1_CACHE_BYTES);
1509
1510 if (align < ARCH_SLAB_MINALIGN)
1511 return ARCH_SLAB_MINALIGN;
1512
1513 return ALIGN(align, sizeof(void *));
1514}
1515
1516static void init_kmem_cache_node(struct kmem_cache_node *n)
1517{
1518 n->nr_partial = 0;
1519 atomic_long_set(&n->nr_slabs, 0);
1520 spin_lock_init(&n->list_lock);
1521 INIT_LIST_HEAD(&n->partial);
Christoph Lameter643b1132007-05-06 14:49:42 -07001522 INIT_LIST_HEAD(&n->full);
Christoph Lameter81819f02007-05-06 14:49:36 -07001523}
1524
1525#ifdef CONFIG_NUMA
1526/*
1527 * No kmalloc_node yet so do it by hand. We know that this is the first
1528 * slab on the node for this slabcache. There are no concurrent accesses
1529 * possible.
1530 *
1531 * Note that this function only works on the kmalloc_node_cache
1532 * when allocating for the kmalloc_node_cache.
1533 */
1534static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1535 int node)
1536{
1537 struct page *page;
1538 struct kmem_cache_node *n;
1539
1540 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1541
1542 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1543 /* new_slab() disables interupts */
1544 local_irq_enable();
1545
1546 BUG_ON(!page);
1547 n = page->freelist;
1548 BUG_ON(!n);
1549 page->freelist = get_freepointer(kmalloc_caches, n);
1550 page->inuse++;
1551 kmalloc_caches->node[node] = n;
1552 init_object(kmalloc_caches, n, 1);
1553 init_kmem_cache_node(n);
1554 atomic_long_inc(&n->nr_slabs);
Christoph Lametere95eed52007-05-06 14:49:44 -07001555 add_partial(n, page);
Christoph Lameter81819f02007-05-06 14:49:36 -07001556 return n;
1557}
1558
1559static void free_kmem_cache_nodes(struct kmem_cache *s)
1560{
1561 int node;
1562
1563 for_each_online_node(node) {
1564 struct kmem_cache_node *n = s->node[node];
1565 if (n && n != &s->local_node)
1566 kmem_cache_free(kmalloc_caches, n);
1567 s->node[node] = NULL;
1568 }
1569}
1570
1571static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1572{
1573 int node;
1574 int local_node;
1575
1576 if (slab_state >= UP)
1577 local_node = page_to_nid(virt_to_page(s));
1578 else
1579 local_node = 0;
1580
1581 for_each_online_node(node) {
1582 struct kmem_cache_node *n;
1583
1584 if (local_node == node)
1585 n = &s->local_node;
1586 else {
1587 if (slab_state == DOWN) {
1588 n = early_kmem_cache_node_alloc(gfpflags,
1589 node);
1590 continue;
1591 }
1592 n = kmem_cache_alloc_node(kmalloc_caches,
1593 gfpflags, node);
1594
1595 if (!n) {
1596 free_kmem_cache_nodes(s);
1597 return 0;
1598 }
1599
1600 }
1601 s->node[node] = n;
1602 init_kmem_cache_node(n);
1603 }
1604 return 1;
1605}
1606#else
1607static void free_kmem_cache_nodes(struct kmem_cache *s)
1608{
1609}
1610
1611static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1612{
1613 init_kmem_cache_node(&s->local_node);
1614 return 1;
1615}
1616#endif
1617
1618/*
1619 * calculate_sizes() determines the order and the distribution of data within
1620 * a slab object.
1621 */
1622static int calculate_sizes(struct kmem_cache *s)
1623{
1624 unsigned long flags = s->flags;
1625 unsigned long size = s->objsize;
1626 unsigned long align = s->align;
1627
1628 /*
1629 * Determine if we can poison the object itself. If the user of
1630 * the slab may touch the object after free or before allocation
1631 * then we should never poison the object itself.
1632 */
1633 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1634 !s->ctor && !s->dtor)
1635 s->flags |= __OBJECT_POISON;
1636 else
1637 s->flags &= ~__OBJECT_POISON;
1638
1639 /*
1640 * Round up object size to the next word boundary. We can only
1641 * place the free pointer at word boundaries and this determines
1642 * the possible location of the free pointer.
1643 */
1644 size = ALIGN(size, sizeof(void *));
1645
1646 /*
1647 * If we are redzoning then check if there is some space between the
1648 * end of the object and the free pointer. If not then add an
1649 * additional word, so that we can establish a redzone between
1650 * the object and the freepointer to be able to check for overwrites.
1651 */
1652 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1653 size += sizeof(void *);
1654
1655 /*
1656 * With that we have determined how much of the slab is in actual
1657 * use by the object. This is the potential offset to the free
1658 * pointer.
1659 */
1660 s->inuse = size;
1661
1662 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1663 s->ctor || s->dtor)) {
1664 /*
1665 * Relocate free pointer after the object if it is not
1666 * permitted to overwrite the first word of the object on
1667 * kmem_cache_free.
1668 *
1669 * This is the case if we do RCU, have a constructor or
1670 * destructor or are poisoning the objects.
1671 */
1672 s->offset = size;
1673 size += sizeof(void *);
1674 }
1675
1676 if (flags & SLAB_STORE_USER)
1677 /*
1678 * Need to store information about allocs and frees after
1679 * the object.
1680 */
1681 size += 2 * sizeof(struct track);
1682
1683 if (flags & DEBUG_DEFAULT_FLAGS)
1684 /*
1685 * Add some empty padding so that we can catch
1686 * overwrites from earlier objects rather than let
1687 * tracking information or the free pointer be
1688 * corrupted if an user writes before the start
1689 * of the object.
1690 */
1691 size += sizeof(void *);
1692 /*
1693 * Determine the alignment based on various parameters that the
1694 * user specified (this is unecessarily complex due to the attempt
1695 * to be compatible with SLAB. Should be cleaned up some day).
1696 */
1697 align = calculate_alignment(flags, align, s->objsize);
1698
1699 /*
1700 * SLUB stores one object immediately after another beginning from
1701 * offset 0. In order to align the objects we have to simply size
1702 * each object to conform to the alignment.
1703 */
1704 size = ALIGN(size, align);
1705 s->size = size;
1706
1707 s->order = calculate_order(size);
1708 if (s->order < 0)
1709 return 0;
1710
1711 /*
1712 * Determine the number of objects per slab
1713 */
1714 s->objects = (PAGE_SIZE << s->order) / size;
1715
1716 /*
1717 * Verify that the number of objects is within permitted limits.
1718 * The page->inuse field is only 16 bit wide! So we cannot have
1719 * more than 64k objects per slab.
1720 */
1721 if (!s->objects || s->objects > 65535)
1722 return 0;
1723 return 1;
1724
1725}
1726
1727static int __init finish_bootstrap(void)
1728{
1729 struct list_head *h;
1730 int err;
1731
1732 slab_state = SYSFS;
1733
1734 list_for_each(h, &slab_caches) {
1735 struct kmem_cache *s =
1736 container_of(h, struct kmem_cache, list);
1737
1738 err = sysfs_slab_add(s);
1739 BUG_ON(err);
1740 }
1741 return 0;
1742}
1743
1744static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1745 const char *name, size_t size,
1746 size_t align, unsigned long flags,
1747 void (*ctor)(void *, struct kmem_cache *, unsigned long),
1748 void (*dtor)(void *, struct kmem_cache *, unsigned long))
1749{
1750 memset(s, 0, kmem_size);
1751 s->name = name;
1752 s->ctor = ctor;
1753 s->dtor = dtor;
1754 s->objsize = size;
1755 s->flags = flags;
1756 s->align = align;
1757
1758 BUG_ON(flags & SLUB_UNIMPLEMENTED);
1759
1760 /*
1761 * The page->offset field is only 16 bit wide. This is an offset
1762 * in units of words from the beginning of an object. If the slab
1763 * size is bigger then we cannot move the free pointer behind the
1764 * object anymore.
1765 *
1766 * On 32 bit platforms the limit is 256k. On 64bit platforms
1767 * the limit is 512k.
1768 *
1769 * Debugging or ctor/dtors may create a need to move the free
1770 * pointer. Fail if this happens.
1771 */
1772 if (s->size >= 65535 * sizeof(void *)) {
1773 BUG_ON(flags & (SLAB_RED_ZONE | SLAB_POISON |
1774 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1775 BUG_ON(ctor || dtor);
1776 }
1777 else
1778 /*
1779 * Enable debugging if selected on the kernel commandline.
1780 */
1781 if (slub_debug && (!slub_debug_slabs ||
1782 strncmp(slub_debug_slabs, name,
1783 strlen(slub_debug_slabs)) == 0))
1784 s->flags |= slub_debug;
1785
1786 if (!calculate_sizes(s))
1787 goto error;
1788
1789 s->refcount = 1;
1790#ifdef CONFIG_NUMA
1791 s->defrag_ratio = 100;
1792#endif
1793
1794 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
1795 return 1;
1796error:
1797 if (flags & SLAB_PANIC)
1798 panic("Cannot create slab %s size=%lu realsize=%u "
1799 "order=%u offset=%u flags=%lx\n",
1800 s->name, (unsigned long)size, s->size, s->order,
1801 s->offset, flags);
1802 return 0;
1803}
1804EXPORT_SYMBOL(kmem_cache_open);
1805
1806/*
1807 * Check if a given pointer is valid
1808 */
1809int kmem_ptr_validate(struct kmem_cache *s, const void *object)
1810{
1811 struct page * page;
1812 void *addr;
1813
1814 page = get_object_page(object);
1815
1816 if (!page || s != page->slab)
1817 /* No slab or wrong slab */
1818 return 0;
1819
1820 addr = page_address(page);
1821 if (object < addr || object >= addr + s->objects * s->size)
1822 /* Out of bounds */
1823 return 0;
1824
1825 if ((object - addr) % s->size)
1826 /* Improperly aligned */
1827 return 0;
1828
1829 /*
1830 * We could also check if the object is on the slabs freelist.
1831 * But this would be too expensive and it seems that the main
1832 * purpose of kmem_ptr_valid is to check if the object belongs
1833 * to a certain slab.
1834 */
1835 return 1;
1836}
1837EXPORT_SYMBOL(kmem_ptr_validate);
1838
1839/*
1840 * Determine the size of a slab object
1841 */
1842unsigned int kmem_cache_size(struct kmem_cache *s)
1843{
1844 return s->objsize;
1845}
1846EXPORT_SYMBOL(kmem_cache_size);
1847
1848const char *kmem_cache_name(struct kmem_cache *s)
1849{
1850 return s->name;
1851}
1852EXPORT_SYMBOL(kmem_cache_name);
1853
1854/*
1855 * Attempt to free all slabs on a node
1856 */
1857static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
1858 struct list_head *list)
1859{
1860 int slabs_inuse = 0;
1861 unsigned long flags;
1862 struct page *page, *h;
1863
1864 spin_lock_irqsave(&n->list_lock, flags);
1865 list_for_each_entry_safe(page, h, list, lru)
1866 if (!page->inuse) {
1867 list_del(&page->lru);
1868 discard_slab(s, page);
1869 } else
1870 slabs_inuse++;
1871 spin_unlock_irqrestore(&n->list_lock, flags);
1872 return slabs_inuse;
1873}
1874
1875/*
1876 * Release all resources used by slab cache
1877 */
1878static int kmem_cache_close(struct kmem_cache *s)
1879{
1880 int node;
1881
1882 flush_all(s);
1883
1884 /* Attempt to free all objects */
1885 for_each_online_node(node) {
1886 struct kmem_cache_node *n = get_node(s, node);
1887
1888 free_list(s, n, &n->partial);
1889 if (atomic_long_read(&n->nr_slabs))
1890 return 1;
1891 }
1892 free_kmem_cache_nodes(s);
1893 return 0;
1894}
1895
1896/*
1897 * Close a cache and release the kmem_cache structure
1898 * (must be used for caches created using kmem_cache_create)
1899 */
1900void kmem_cache_destroy(struct kmem_cache *s)
1901{
1902 down_write(&slub_lock);
1903 s->refcount--;
1904 if (!s->refcount) {
1905 list_del(&s->list);
1906 if (kmem_cache_close(s))
1907 WARN_ON(1);
1908 sysfs_slab_remove(s);
1909 kfree(s);
1910 }
1911 up_write(&slub_lock);
1912}
1913EXPORT_SYMBOL(kmem_cache_destroy);
1914
1915/********************************************************************
1916 * Kmalloc subsystem
1917 *******************************************************************/
1918
1919struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
1920EXPORT_SYMBOL(kmalloc_caches);
1921
1922#ifdef CONFIG_ZONE_DMA
1923static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
1924#endif
1925
1926static int __init setup_slub_min_order(char *str)
1927{
1928 get_option (&str, &slub_min_order);
1929
1930 return 1;
1931}
1932
1933__setup("slub_min_order=", setup_slub_min_order);
1934
1935static int __init setup_slub_max_order(char *str)
1936{
1937 get_option (&str, &slub_max_order);
1938
1939 return 1;
1940}
1941
1942__setup("slub_max_order=", setup_slub_max_order);
1943
1944static int __init setup_slub_min_objects(char *str)
1945{
1946 get_option (&str, &slub_min_objects);
1947
1948 return 1;
1949}
1950
1951__setup("slub_min_objects=", setup_slub_min_objects);
1952
1953static int __init setup_slub_nomerge(char *str)
1954{
1955 slub_nomerge = 1;
1956 return 1;
1957}
1958
1959__setup("slub_nomerge", setup_slub_nomerge);
1960
1961static int __init setup_slub_debug(char *str)
1962{
1963 if (!str || *str != '=')
1964 slub_debug = DEBUG_DEFAULT_FLAGS;
1965 else {
1966 str++;
1967 if (*str == 0 || *str == ',')
1968 slub_debug = DEBUG_DEFAULT_FLAGS;
1969 else
1970 for( ;*str && *str != ','; str++)
1971 switch (*str) {
1972 case 'f' : case 'F' :
1973 slub_debug |= SLAB_DEBUG_FREE;
1974 break;
1975 case 'z' : case 'Z' :
1976 slub_debug |= SLAB_RED_ZONE;
1977 break;
1978 case 'p' : case 'P' :
1979 slub_debug |= SLAB_POISON;
1980 break;
1981 case 'u' : case 'U' :
1982 slub_debug |= SLAB_STORE_USER;
1983 break;
1984 case 't' : case 'T' :
1985 slub_debug |= SLAB_TRACE;
1986 break;
1987 default:
1988 printk(KERN_ERR "slub_debug option '%c' "
1989 "unknown. skipped\n",*str);
1990 }
1991 }
1992
1993 if (*str == ',')
1994 slub_debug_slabs = str + 1;
1995 return 1;
1996}
1997
1998__setup("slub_debug", setup_slub_debug);
1999
2000static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2001 const char *name, int size, gfp_t gfp_flags)
2002{
2003 unsigned int flags = 0;
2004
2005 if (gfp_flags & SLUB_DMA)
2006 flags = SLAB_CACHE_DMA;
2007
2008 down_write(&slub_lock);
2009 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2010 flags, NULL, NULL))
2011 goto panic;
2012
2013 list_add(&s->list, &slab_caches);
2014 up_write(&slub_lock);
2015 if (sysfs_slab_add(s))
2016 goto panic;
2017 return s;
2018
2019panic:
2020 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2021}
2022
2023static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2024{
2025 int index = kmalloc_index(size);
2026
Christoph Lameter614410d2007-05-06 14:49:38 -07002027 if (!index)
Christoph Lameter81819f02007-05-06 14:49:36 -07002028 return NULL;
2029
2030 /* Allocation too large? */
2031 BUG_ON(index < 0);
2032
2033#ifdef CONFIG_ZONE_DMA
2034 if ((flags & SLUB_DMA)) {
2035 struct kmem_cache *s;
2036 struct kmem_cache *x;
2037 char *text;
2038 size_t realsize;
2039
2040 s = kmalloc_caches_dma[index];
2041 if (s)
2042 return s;
2043
2044 /* Dynamically create dma cache */
2045 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2046 if (!x)
2047 panic("Unable to allocate memory for dma cache\n");
2048
2049 if (index <= KMALLOC_SHIFT_HIGH)
2050 realsize = 1 << index;
2051 else {
2052 if (index == 1)
2053 realsize = 96;
2054 else
2055 realsize = 192;
2056 }
2057
2058 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2059 (unsigned int)realsize);
2060 s = create_kmalloc_cache(x, text, realsize, flags);
2061 kmalloc_caches_dma[index] = s;
2062 return s;
2063 }
2064#endif
2065 return &kmalloc_caches[index];
2066}
2067
2068void *__kmalloc(size_t size, gfp_t flags)
2069{
2070 struct kmem_cache *s = get_slab(size, flags);
2071
2072 if (s)
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002073 return slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002074 return NULL;
2075}
2076EXPORT_SYMBOL(__kmalloc);
2077
2078#ifdef CONFIG_NUMA
2079void *__kmalloc_node(size_t size, gfp_t flags, int node)
2080{
2081 struct kmem_cache *s = get_slab(size, flags);
2082
2083 if (s)
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002084 return slab_alloc(s, flags, node, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002085 return NULL;
2086}
2087EXPORT_SYMBOL(__kmalloc_node);
2088#endif
2089
2090size_t ksize(const void *object)
2091{
2092 struct page *page = get_object_page(object);
2093 struct kmem_cache *s;
2094
2095 BUG_ON(!page);
2096 s = page->slab;
2097 BUG_ON(!s);
2098
2099 /*
2100 * Debugging requires use of the padding between object
2101 * and whatever may come after it.
2102 */
2103 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2104 return s->objsize;
2105
2106 /*
2107 * If we have the need to store the freelist pointer
2108 * back there or track user information then we can
2109 * only use the space before that information.
2110 */
2111 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2112 return s->inuse;
2113
2114 /*
2115 * Else we can use all the padding etc for the allocation
2116 */
2117 return s->size;
2118}
2119EXPORT_SYMBOL(ksize);
2120
2121void kfree(const void *x)
2122{
2123 struct kmem_cache *s;
2124 struct page *page;
2125
2126 if (!x)
2127 return;
2128
Christoph Lameterb49af682007-05-06 14:49:41 -07002129 page = virt_to_head_page(x);
Christoph Lameter81819f02007-05-06 14:49:36 -07002130 s = page->slab;
2131
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002132 slab_free(s, page, (void *)x, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002133}
2134EXPORT_SYMBOL(kfree);
2135
2136/**
2137 * krealloc - reallocate memory. The contents will remain unchanged.
2138 *
2139 * @p: object to reallocate memory for.
2140 * @new_size: how many bytes of memory are required.
2141 * @flags: the type of memory to allocate.
2142 *
2143 * The contents of the object pointed to are preserved up to the
2144 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2145 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2146 * %NULL pointer, the object pointed to is freed.
2147 */
2148void *krealloc(const void *p, size_t new_size, gfp_t flags)
2149{
2150 struct kmem_cache *new_cache;
2151 void *ret;
2152 struct page *page;
2153
2154 if (unlikely(!p))
2155 return kmalloc(new_size, flags);
2156
2157 if (unlikely(!new_size)) {
2158 kfree(p);
2159 return NULL;
2160 }
2161
Christoph Lameterb49af682007-05-06 14:49:41 -07002162 page = virt_to_head_page(p);
Christoph Lameter81819f02007-05-06 14:49:36 -07002163
2164 new_cache = get_slab(new_size, flags);
2165
2166 /*
2167 * If new size fits in the current cache, bail out.
2168 */
2169 if (likely(page->slab == new_cache))
2170 return (void *)p;
2171
2172 ret = kmalloc(new_size, flags);
2173 if (ret) {
2174 memcpy(ret, p, min(new_size, ksize(p)));
2175 kfree(p);
2176 }
2177 return ret;
2178}
2179EXPORT_SYMBOL(krealloc);
2180
2181/********************************************************************
2182 * Basic setup of slabs
2183 *******************************************************************/
2184
2185void __init kmem_cache_init(void)
2186{
2187 int i;
2188
2189#ifdef CONFIG_NUMA
2190 /*
2191 * Must first have the slab cache available for the allocations of the
2192 * struct kmalloc_cache_node's. There is special bootstrap code in
2193 * kmem_cache_open for slab_state == DOWN.
2194 */
2195 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2196 sizeof(struct kmem_cache_node), GFP_KERNEL);
2197#endif
2198
2199 /* Able to allocate the per node structures */
2200 slab_state = PARTIAL;
2201
2202 /* Caches that are not of the two-to-the-power-of size */
2203 create_kmalloc_cache(&kmalloc_caches[1],
2204 "kmalloc-96", 96, GFP_KERNEL);
2205 create_kmalloc_cache(&kmalloc_caches[2],
2206 "kmalloc-192", 192, GFP_KERNEL);
2207
2208 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2209 create_kmalloc_cache(&kmalloc_caches[i],
2210 "kmalloc", 1 << i, GFP_KERNEL);
2211
2212 slab_state = UP;
2213
2214 /* Provide the correct kmalloc names now that the caches are up */
2215 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2216 kmalloc_caches[i]. name =
2217 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2218
2219#ifdef CONFIG_SMP
2220 register_cpu_notifier(&slab_notifier);
2221#endif
2222
2223 if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */
2224 kmem_size = offsetof(struct kmem_cache, cpu_slab)
2225 + nr_cpu_ids * sizeof(struct page *);
2226
2227 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2228 " Processors=%d, Nodes=%d\n",
2229 KMALLOC_SHIFT_HIGH, L1_CACHE_BYTES,
2230 slub_min_order, slub_max_order, slub_min_objects,
2231 nr_cpu_ids, nr_node_ids);
2232}
2233
2234/*
2235 * Find a mergeable slab cache
2236 */
2237static int slab_unmergeable(struct kmem_cache *s)
2238{
2239 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2240 return 1;
2241
2242 if (s->ctor || s->dtor)
2243 return 1;
2244
2245 return 0;
2246}
2247
2248static struct kmem_cache *find_mergeable(size_t size,
2249 size_t align, unsigned long flags,
2250 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2251 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2252{
2253 struct list_head *h;
2254
2255 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2256 return NULL;
2257
2258 if (ctor || dtor)
2259 return NULL;
2260
2261 size = ALIGN(size, sizeof(void *));
2262 align = calculate_alignment(flags, align, size);
2263 size = ALIGN(size, align);
2264
2265 list_for_each(h, &slab_caches) {
2266 struct kmem_cache *s =
2267 container_of(h, struct kmem_cache, list);
2268
2269 if (slab_unmergeable(s))
2270 continue;
2271
2272 if (size > s->size)
2273 continue;
2274
2275 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2276 (s->flags & SLUB_MERGE_SAME))
2277 continue;
2278 /*
2279 * Check if alignment is compatible.
2280 * Courtesy of Adrian Drzewiecki
2281 */
2282 if ((s->size & ~(align -1)) != s->size)
2283 continue;
2284
2285 if (s->size - size >= sizeof(void *))
2286 continue;
2287
2288 return s;
2289 }
2290 return NULL;
2291}
2292
2293struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2294 size_t align, unsigned long flags,
2295 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2296 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2297{
2298 struct kmem_cache *s;
2299
2300 down_write(&slub_lock);
2301 s = find_mergeable(size, align, flags, dtor, ctor);
2302 if (s) {
2303 s->refcount++;
2304 /*
2305 * Adjust the object sizes so that we clear
2306 * the complete object on kzalloc.
2307 */
2308 s->objsize = max(s->objsize, (int)size);
2309 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2310 if (sysfs_slab_alias(s, name))
2311 goto err;
2312 } else {
2313 s = kmalloc(kmem_size, GFP_KERNEL);
2314 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2315 size, align, flags, ctor, dtor)) {
2316 if (sysfs_slab_add(s)) {
2317 kfree(s);
2318 goto err;
2319 }
2320 list_add(&s->list, &slab_caches);
2321 } else
2322 kfree(s);
2323 }
2324 up_write(&slub_lock);
2325 return s;
2326
2327err:
2328 up_write(&slub_lock);
2329 if (flags & SLAB_PANIC)
2330 panic("Cannot create slabcache %s\n", name);
2331 else
2332 s = NULL;
2333 return s;
2334}
2335EXPORT_SYMBOL(kmem_cache_create);
2336
2337void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2338{
2339 void *x;
2340
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002341 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
Christoph Lameter81819f02007-05-06 14:49:36 -07002342 if (x)
2343 memset(x, 0, s->objsize);
2344 return x;
2345}
2346EXPORT_SYMBOL(kmem_cache_zalloc);
2347
2348#ifdef CONFIG_SMP
2349static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2350{
2351 struct list_head *h;
2352
2353 down_read(&slub_lock);
2354 list_for_each(h, &slab_caches) {
2355 struct kmem_cache *s =
2356 container_of(h, struct kmem_cache, list);
2357
2358 func(s, cpu);
2359 }
2360 up_read(&slub_lock);
2361}
2362
2363/*
2364 * Use the cpu notifier to insure that the slab are flushed
2365 * when necessary.
2366 */
2367static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2368 unsigned long action, void *hcpu)
2369{
2370 long cpu = (long)hcpu;
2371
2372 switch (action) {
2373 case CPU_UP_CANCELED:
2374 case CPU_DEAD:
2375 for_all_slabs(__flush_cpu_slab, cpu);
2376 break;
2377 default:
2378 break;
2379 }
2380 return NOTIFY_OK;
2381}
2382
2383static struct notifier_block __cpuinitdata slab_notifier =
2384 { &slab_cpuup_callback, NULL, 0 };
2385
2386#endif
2387
2388/***************************************************************
2389 * Compatiblility definitions
2390 **************************************************************/
2391
2392int kmem_cache_shrink(struct kmem_cache *s)
2393{
2394 flush_all(s);
2395 return 0;
2396}
2397EXPORT_SYMBOL(kmem_cache_shrink);
2398
2399#ifdef CONFIG_NUMA
2400
2401/*****************************************************************
2402 * Generic reaper used to support the page allocator
2403 * (the cpu slabs are reaped by a per slab workqueue).
2404 *
2405 * Maybe move this to the page allocator?
2406 ****************************************************************/
2407
2408static DEFINE_PER_CPU(unsigned long, reap_node);
2409
2410static void init_reap_node(int cpu)
2411{
2412 int node;
2413
2414 node = next_node(cpu_to_node(cpu), node_online_map);
2415 if (node == MAX_NUMNODES)
2416 node = first_node(node_online_map);
2417
2418 __get_cpu_var(reap_node) = node;
2419}
2420
2421static void next_reap_node(void)
2422{
2423 int node = __get_cpu_var(reap_node);
2424
2425 /*
2426 * Also drain per cpu pages on remote zones
2427 */
2428 if (node != numa_node_id())
2429 drain_node_pages(node);
2430
2431 node = next_node(node, node_online_map);
2432 if (unlikely(node >= MAX_NUMNODES))
2433 node = first_node(node_online_map);
2434 __get_cpu_var(reap_node) = node;
2435}
2436#else
2437#define init_reap_node(cpu) do { } while (0)
2438#define next_reap_node(void) do { } while (0)
2439#endif
2440
2441#define REAPTIMEOUT_CPUC (2*HZ)
2442
2443#ifdef CONFIG_SMP
2444static DEFINE_PER_CPU(struct delayed_work, reap_work);
2445
2446static void cache_reap(struct work_struct *unused)
2447{
2448 next_reap_node();
2449 refresh_cpu_vm_stats(smp_processor_id());
2450 schedule_delayed_work(&__get_cpu_var(reap_work),
2451 REAPTIMEOUT_CPUC);
2452}
2453
2454static void __devinit start_cpu_timer(int cpu)
2455{
2456 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
2457
2458 /*
2459 * When this gets called from do_initcalls via cpucache_init(),
2460 * init_workqueues() has already run, so keventd will be setup
2461 * at that time.
2462 */
2463 if (keventd_up() && reap_work->work.func == NULL) {
2464 init_reap_node(cpu);
2465 INIT_DELAYED_WORK(reap_work, cache_reap);
2466 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
2467 }
2468}
2469
2470static int __init cpucache_init(void)
2471{
2472 int cpu;
2473
2474 /*
2475 * Register the timers that drain pcp pages and update vm statistics
2476 */
2477 for_each_online_cpu(cpu)
2478 start_cpu_timer(cpu);
2479 return 0;
2480}
2481__initcall(cpucache_init);
2482#endif
2483
2484#ifdef SLUB_RESILIENCY_TEST
2485static unsigned long validate_slab_cache(struct kmem_cache *s);
2486
2487static void resiliency_test(void)
2488{
2489 u8 *p;
2490
2491 printk(KERN_ERR "SLUB resiliency testing\n");
2492 printk(KERN_ERR "-----------------------\n");
2493 printk(KERN_ERR "A. Corruption after allocation\n");
2494
2495 p = kzalloc(16, GFP_KERNEL);
2496 p[16] = 0x12;
2497 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2498 " 0x12->0x%p\n\n", p + 16);
2499
2500 validate_slab_cache(kmalloc_caches + 4);
2501
2502 /* Hmmm... The next two are dangerous */
2503 p = kzalloc(32, GFP_KERNEL);
2504 p[32 + sizeof(void *)] = 0x34;
2505 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2506 " 0x34 -> -0x%p\n", p);
2507 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2508
2509 validate_slab_cache(kmalloc_caches + 5);
2510 p = kzalloc(64, GFP_KERNEL);
2511 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2512 *p = 0x56;
2513 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2514 p);
2515 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2516 validate_slab_cache(kmalloc_caches + 6);
2517
2518 printk(KERN_ERR "\nB. Corruption after free\n");
2519 p = kzalloc(128, GFP_KERNEL);
2520 kfree(p);
2521 *p = 0x78;
2522 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2523 validate_slab_cache(kmalloc_caches + 7);
2524
2525 p = kzalloc(256, GFP_KERNEL);
2526 kfree(p);
2527 p[50] = 0x9a;
2528 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2529 validate_slab_cache(kmalloc_caches + 8);
2530
2531 p = kzalloc(512, GFP_KERNEL);
2532 kfree(p);
2533 p[512] = 0xab;
2534 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2535 validate_slab_cache(kmalloc_caches + 9);
2536}
2537#else
2538static void resiliency_test(void) {};
2539#endif
2540
2541/*
2542 * These are not as efficient as kmalloc for the non debug case.
2543 * We do not have the page struct available so we have to touch one
2544 * cacheline in struct kmem_cache to check slab flags.
2545 */
2546void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2547{
2548 struct kmem_cache *s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002549
2550 if (!s)
2551 return NULL;
2552
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002553 return slab_alloc(s, gfpflags, -1, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07002554}
2555
2556void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2557 int node, void *caller)
2558{
2559 struct kmem_cache *s = get_slab(size, gfpflags);
Christoph Lameter81819f02007-05-06 14:49:36 -07002560
2561 if (!s)
2562 return NULL;
2563
Christoph Lameter77c5e2d2007-05-06 14:49:42 -07002564 return slab_alloc(s, gfpflags, node, caller);
Christoph Lameter81819f02007-05-06 14:49:36 -07002565}
2566
2567#ifdef CONFIG_SYSFS
2568
Christoph Lameter53e15af2007-05-06 14:49:43 -07002569static int validate_slab(struct kmem_cache *s, struct page *page)
2570{
2571 void *p;
2572 void *addr = page_address(page);
2573 unsigned long map[BITS_TO_LONGS(s->objects)];
2574
2575 if (!check_slab(s, page) ||
2576 !on_freelist(s, page, NULL))
2577 return 0;
2578
2579 /* Now we know that a valid freelist exists */
2580 bitmap_zero(map, s->objects);
2581
2582 for(p = page->freelist; p; p = get_freepointer(s, p)) {
2583 set_bit((p - addr) / s->size, map);
2584 if (!check_object(s, page, p, 0))
2585 return 0;
2586 }
2587
2588 for(p = addr; p < addr + s->objects * s->size; p += s->size)
2589 if (!test_bit((p - addr) / s->size, map))
2590 if (!check_object(s, page, p, 1))
2591 return 0;
2592 return 1;
2593}
2594
2595static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2596{
2597 if (slab_trylock(page)) {
2598 validate_slab(s, page);
2599 slab_unlock(page);
2600 } else
2601 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2602 s->name, page);
2603
2604 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2605 if (!PageError(page))
2606 printk(KERN_ERR "SLUB %s: PageError not set "
2607 "on slab 0x%p\n", s->name, page);
2608 } else {
2609 if (PageError(page))
2610 printk(KERN_ERR "SLUB %s: PageError set on "
2611 "slab 0x%p\n", s->name, page);
2612 }
2613}
2614
2615static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2616{
2617 unsigned long count = 0;
2618 struct page *page;
2619 unsigned long flags;
2620
2621 spin_lock_irqsave(&n->list_lock, flags);
2622
2623 list_for_each_entry(page, &n->partial, lru) {
2624 validate_slab_slab(s, page);
2625 count++;
2626 }
2627 if (count != n->nr_partial)
2628 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2629 "counter=%ld\n", s->name, count, n->nr_partial);
2630
2631 if (!(s->flags & SLAB_STORE_USER))
2632 goto out;
2633
2634 list_for_each_entry(page, &n->full, lru) {
2635 validate_slab_slab(s, page);
2636 count++;
2637 }
2638 if (count != atomic_long_read(&n->nr_slabs))
2639 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2640 "counter=%ld\n", s->name, count,
2641 atomic_long_read(&n->nr_slabs));
2642
2643out:
2644 spin_unlock_irqrestore(&n->list_lock, flags);
2645 return count;
2646}
2647
2648static unsigned long validate_slab_cache(struct kmem_cache *s)
2649{
2650 int node;
2651 unsigned long count = 0;
2652
2653 flush_all(s);
2654 for_each_online_node(node) {
2655 struct kmem_cache_node *n = get_node(s, node);
2656
2657 count += validate_slab_node(s, n);
2658 }
2659 return count;
2660}
2661
Christoph Lameter81819f02007-05-06 14:49:36 -07002662static unsigned long count_partial(struct kmem_cache_node *n)
2663{
2664 unsigned long flags;
2665 unsigned long x = 0;
2666 struct page *page;
2667
2668 spin_lock_irqsave(&n->list_lock, flags);
2669 list_for_each_entry(page, &n->partial, lru)
2670 x += page->inuse;
2671 spin_unlock_irqrestore(&n->list_lock, flags);
2672 return x;
2673}
2674
2675enum slab_stat_type {
2676 SL_FULL,
2677 SL_PARTIAL,
2678 SL_CPU,
2679 SL_OBJECTS
2680};
2681
2682#define SO_FULL (1 << SL_FULL)
2683#define SO_PARTIAL (1 << SL_PARTIAL)
2684#define SO_CPU (1 << SL_CPU)
2685#define SO_OBJECTS (1 << SL_OBJECTS)
2686
2687static unsigned long slab_objects(struct kmem_cache *s,
2688 char *buf, unsigned long flags)
2689{
2690 unsigned long total = 0;
2691 int cpu;
2692 int node;
2693 int x;
2694 unsigned long *nodes;
2695 unsigned long *per_cpu;
2696
2697 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
2698 per_cpu = nodes + nr_node_ids;
2699
2700 for_each_possible_cpu(cpu) {
2701 struct page *page = s->cpu_slab[cpu];
2702 int node;
2703
2704 if (page) {
2705 node = page_to_nid(page);
2706 if (flags & SO_CPU) {
2707 int x = 0;
2708
2709 if (flags & SO_OBJECTS)
2710 x = page->inuse;
2711 else
2712 x = 1;
2713 total += x;
2714 nodes[node] += x;
2715 }
2716 per_cpu[node]++;
2717 }
2718 }
2719
2720 for_each_online_node(node) {
2721 struct kmem_cache_node *n = get_node(s, node);
2722
2723 if (flags & SO_PARTIAL) {
2724 if (flags & SO_OBJECTS)
2725 x = count_partial(n);
2726 else
2727 x = n->nr_partial;
2728 total += x;
2729 nodes[node] += x;
2730 }
2731
2732 if (flags & SO_FULL) {
2733 int full_slabs = atomic_read(&n->nr_slabs)
2734 - per_cpu[node]
2735 - n->nr_partial;
2736
2737 if (flags & SO_OBJECTS)
2738 x = full_slabs * s->objects;
2739 else
2740 x = full_slabs;
2741 total += x;
2742 nodes[node] += x;
2743 }
2744 }
2745
2746 x = sprintf(buf, "%lu", total);
2747#ifdef CONFIG_NUMA
2748 for_each_online_node(node)
2749 if (nodes[node])
2750 x += sprintf(buf + x, " N%d=%lu",
2751 node, nodes[node]);
2752#endif
2753 kfree(nodes);
2754 return x + sprintf(buf + x, "\n");
2755}
2756
2757static int any_slab_objects(struct kmem_cache *s)
2758{
2759 int node;
2760 int cpu;
2761
2762 for_each_possible_cpu(cpu)
2763 if (s->cpu_slab[cpu])
2764 return 1;
2765
2766 for_each_node(node) {
2767 struct kmem_cache_node *n = get_node(s, node);
2768
2769 if (n->nr_partial || atomic_read(&n->nr_slabs))
2770 return 1;
2771 }
2772 return 0;
2773}
2774
2775#define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
2776#define to_slab(n) container_of(n, struct kmem_cache, kobj);
2777
2778struct slab_attribute {
2779 struct attribute attr;
2780 ssize_t (*show)(struct kmem_cache *s, char *buf);
2781 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
2782};
2783
2784#define SLAB_ATTR_RO(_name) \
2785 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
2786
2787#define SLAB_ATTR(_name) \
2788 static struct slab_attribute _name##_attr = \
2789 __ATTR(_name, 0644, _name##_show, _name##_store)
2790
Christoph Lameter81819f02007-05-06 14:49:36 -07002791static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
2792{
2793 return sprintf(buf, "%d\n", s->size);
2794}
2795SLAB_ATTR_RO(slab_size);
2796
2797static ssize_t align_show(struct kmem_cache *s, char *buf)
2798{
2799 return sprintf(buf, "%d\n", s->align);
2800}
2801SLAB_ATTR_RO(align);
2802
2803static ssize_t object_size_show(struct kmem_cache *s, char *buf)
2804{
2805 return sprintf(buf, "%d\n", s->objsize);
2806}
2807SLAB_ATTR_RO(object_size);
2808
2809static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
2810{
2811 return sprintf(buf, "%d\n", s->objects);
2812}
2813SLAB_ATTR_RO(objs_per_slab);
2814
2815static ssize_t order_show(struct kmem_cache *s, char *buf)
2816{
2817 return sprintf(buf, "%d\n", s->order);
2818}
2819SLAB_ATTR_RO(order);
2820
2821static ssize_t ctor_show(struct kmem_cache *s, char *buf)
2822{
2823 if (s->ctor) {
2824 int n = sprint_symbol(buf, (unsigned long)s->ctor);
2825
2826 return n + sprintf(buf + n, "\n");
2827 }
2828 return 0;
2829}
2830SLAB_ATTR_RO(ctor);
2831
2832static ssize_t dtor_show(struct kmem_cache *s, char *buf)
2833{
2834 if (s->dtor) {
2835 int n = sprint_symbol(buf, (unsigned long)s->dtor);
2836
2837 return n + sprintf(buf + n, "\n");
2838 }
2839 return 0;
2840}
2841SLAB_ATTR_RO(dtor);
2842
2843static ssize_t aliases_show(struct kmem_cache *s, char *buf)
2844{
2845 return sprintf(buf, "%d\n", s->refcount - 1);
2846}
2847SLAB_ATTR_RO(aliases);
2848
2849static ssize_t slabs_show(struct kmem_cache *s, char *buf)
2850{
2851 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
2852}
2853SLAB_ATTR_RO(slabs);
2854
2855static ssize_t partial_show(struct kmem_cache *s, char *buf)
2856{
2857 return slab_objects(s, buf, SO_PARTIAL);
2858}
2859SLAB_ATTR_RO(partial);
2860
2861static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
2862{
2863 return slab_objects(s, buf, SO_CPU);
2864}
2865SLAB_ATTR_RO(cpu_slabs);
2866
2867static ssize_t objects_show(struct kmem_cache *s, char *buf)
2868{
2869 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
2870}
2871SLAB_ATTR_RO(objects);
2872
2873static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
2874{
2875 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
2876}
2877
2878static ssize_t sanity_checks_store(struct kmem_cache *s,
2879 const char *buf, size_t length)
2880{
2881 s->flags &= ~SLAB_DEBUG_FREE;
2882 if (buf[0] == '1')
2883 s->flags |= SLAB_DEBUG_FREE;
2884 return length;
2885}
2886SLAB_ATTR(sanity_checks);
2887
2888static ssize_t trace_show(struct kmem_cache *s, char *buf)
2889{
2890 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
2891}
2892
2893static ssize_t trace_store(struct kmem_cache *s, const char *buf,
2894 size_t length)
2895{
2896 s->flags &= ~SLAB_TRACE;
2897 if (buf[0] == '1')
2898 s->flags |= SLAB_TRACE;
2899 return length;
2900}
2901SLAB_ATTR(trace);
2902
2903static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
2904{
2905 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
2906}
2907
2908static ssize_t reclaim_account_store(struct kmem_cache *s,
2909 const char *buf, size_t length)
2910{
2911 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
2912 if (buf[0] == '1')
2913 s->flags |= SLAB_RECLAIM_ACCOUNT;
2914 return length;
2915}
2916SLAB_ATTR(reclaim_account);
2917
2918static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
2919{
2920 return sprintf(buf, "%d\n", !!(s->flags &
2921 (SLAB_HWCACHE_ALIGN|SLAB_MUST_HWCACHE_ALIGN)));
2922}
2923SLAB_ATTR_RO(hwcache_align);
2924
2925#ifdef CONFIG_ZONE_DMA
2926static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
2927{
2928 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
2929}
2930SLAB_ATTR_RO(cache_dma);
2931#endif
2932
2933static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
2934{
2935 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
2936}
2937SLAB_ATTR_RO(destroy_by_rcu);
2938
2939static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
2940{
2941 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
2942}
2943
2944static ssize_t red_zone_store(struct kmem_cache *s,
2945 const char *buf, size_t length)
2946{
2947 if (any_slab_objects(s))
2948 return -EBUSY;
2949
2950 s->flags &= ~SLAB_RED_ZONE;
2951 if (buf[0] == '1')
2952 s->flags |= SLAB_RED_ZONE;
2953 calculate_sizes(s);
2954 return length;
2955}
2956SLAB_ATTR(red_zone);
2957
2958static ssize_t poison_show(struct kmem_cache *s, char *buf)
2959{
2960 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
2961}
2962
2963static ssize_t poison_store(struct kmem_cache *s,
2964 const char *buf, size_t length)
2965{
2966 if (any_slab_objects(s))
2967 return -EBUSY;
2968
2969 s->flags &= ~SLAB_POISON;
2970 if (buf[0] == '1')
2971 s->flags |= SLAB_POISON;
2972 calculate_sizes(s);
2973 return length;
2974}
2975SLAB_ATTR(poison);
2976
2977static ssize_t store_user_show(struct kmem_cache *s, char *buf)
2978{
2979 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
2980}
2981
2982static ssize_t store_user_store(struct kmem_cache *s,
2983 const char *buf, size_t length)
2984{
2985 if (any_slab_objects(s))
2986 return -EBUSY;
2987
2988 s->flags &= ~SLAB_STORE_USER;
2989 if (buf[0] == '1')
2990 s->flags |= SLAB_STORE_USER;
2991 calculate_sizes(s);
2992 return length;
2993}
2994SLAB_ATTR(store_user);
2995
Christoph Lameter53e15af2007-05-06 14:49:43 -07002996static ssize_t validate_show(struct kmem_cache *s, char *buf)
2997{
2998 return 0;
2999}
3000
3001static ssize_t validate_store(struct kmem_cache *s,
3002 const char *buf, size_t length)
3003{
3004 if (buf[0] == '1')
3005 validate_slab_cache(s);
3006 else
3007 return -EINVAL;
3008 return length;
3009}
3010SLAB_ATTR(validate);
3011
Christoph Lameter81819f02007-05-06 14:49:36 -07003012#ifdef CONFIG_NUMA
3013static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3014{
3015 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3016}
3017
3018static ssize_t defrag_ratio_store(struct kmem_cache *s,
3019 const char *buf, size_t length)
3020{
3021 int n = simple_strtoul(buf, NULL, 10);
3022
3023 if (n < 100)
3024 s->defrag_ratio = n * 10;
3025 return length;
3026}
3027SLAB_ATTR(defrag_ratio);
3028#endif
3029
3030static struct attribute * slab_attrs[] = {
3031 &slab_size_attr.attr,
3032 &object_size_attr.attr,
3033 &objs_per_slab_attr.attr,
3034 &order_attr.attr,
3035 &objects_attr.attr,
3036 &slabs_attr.attr,
3037 &partial_attr.attr,
3038 &cpu_slabs_attr.attr,
3039 &ctor_attr.attr,
3040 &dtor_attr.attr,
3041 &aliases_attr.attr,
3042 &align_attr.attr,
3043 &sanity_checks_attr.attr,
3044 &trace_attr.attr,
3045 &hwcache_align_attr.attr,
3046 &reclaim_account_attr.attr,
3047 &destroy_by_rcu_attr.attr,
3048 &red_zone_attr.attr,
3049 &poison_attr.attr,
3050 &store_user_attr.attr,
Christoph Lameter53e15af2007-05-06 14:49:43 -07003051 &validate_attr.attr,
Christoph Lameter81819f02007-05-06 14:49:36 -07003052#ifdef CONFIG_ZONE_DMA
3053 &cache_dma_attr.attr,
3054#endif
3055#ifdef CONFIG_NUMA
3056 &defrag_ratio_attr.attr,
3057#endif
3058 NULL
3059};
3060
3061static struct attribute_group slab_attr_group = {
3062 .attrs = slab_attrs,
3063};
3064
3065static ssize_t slab_attr_show(struct kobject *kobj,
3066 struct attribute *attr,
3067 char *buf)
3068{
3069 struct slab_attribute *attribute;
3070 struct kmem_cache *s;
3071 int err;
3072
3073 attribute = to_slab_attr(attr);
3074 s = to_slab(kobj);
3075
3076 if (!attribute->show)
3077 return -EIO;
3078
3079 err = attribute->show(s, buf);
3080
3081 return err;
3082}
3083
3084static ssize_t slab_attr_store(struct kobject *kobj,
3085 struct attribute *attr,
3086 const char *buf, size_t len)
3087{
3088 struct slab_attribute *attribute;
3089 struct kmem_cache *s;
3090 int err;
3091
3092 attribute = to_slab_attr(attr);
3093 s = to_slab(kobj);
3094
3095 if (!attribute->store)
3096 return -EIO;
3097
3098 err = attribute->store(s, buf, len);
3099
3100 return err;
3101}
3102
3103static struct sysfs_ops slab_sysfs_ops = {
3104 .show = slab_attr_show,
3105 .store = slab_attr_store,
3106};
3107
3108static struct kobj_type slab_ktype = {
3109 .sysfs_ops = &slab_sysfs_ops,
3110};
3111
3112static int uevent_filter(struct kset *kset, struct kobject *kobj)
3113{
3114 struct kobj_type *ktype = get_ktype(kobj);
3115
3116 if (ktype == &slab_ktype)
3117 return 1;
3118 return 0;
3119}
3120
3121static struct kset_uevent_ops slab_uevent_ops = {
3122 .filter = uevent_filter,
3123};
3124
3125decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3126
3127#define ID_STR_LENGTH 64
3128
3129/* Create a unique string id for a slab cache:
3130 * format
3131 * :[flags-]size:[memory address of kmemcache]
3132 */
3133static char *create_unique_id(struct kmem_cache *s)
3134{
3135 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3136 char *p = name;
3137
3138 BUG_ON(!name);
3139
3140 *p++ = ':';
3141 /*
3142 * First flags affecting slabcache operations. We will only
3143 * get here for aliasable slabs so we do not need to support
3144 * too many flags. The flags here must cover all flags that
3145 * are matched during merging to guarantee that the id is
3146 * unique.
3147 */
3148 if (s->flags & SLAB_CACHE_DMA)
3149 *p++ = 'd';
3150 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3151 *p++ = 'a';
3152 if (s->flags & SLAB_DEBUG_FREE)
3153 *p++ = 'F';
3154 if (p != name + 1)
3155 *p++ = '-';
3156 p += sprintf(p, "%07d", s->size);
3157 BUG_ON(p > name + ID_STR_LENGTH - 1);
3158 return name;
3159}
3160
3161static int sysfs_slab_add(struct kmem_cache *s)
3162{
3163 int err;
3164 const char *name;
3165 int unmergeable;
3166
3167 if (slab_state < SYSFS)
3168 /* Defer until later */
3169 return 0;
3170
3171 unmergeable = slab_unmergeable(s);
3172 if (unmergeable) {
3173 /*
3174 * Slabcache can never be merged so we can use the name proper.
3175 * This is typically the case for debug situations. In that
3176 * case we can catch duplicate names easily.
3177 */
3178 sysfs_remove_link(&slab_subsys.kset.kobj, s->name);
3179 name = s->name;
3180 } else {
3181 /*
3182 * Create a unique name for the slab as a target
3183 * for the symlinks.
3184 */
3185 name = create_unique_id(s);
3186 }
3187
3188 kobj_set_kset_s(s, slab_subsys);
3189 kobject_set_name(&s->kobj, name);
3190 kobject_init(&s->kobj);
3191 err = kobject_add(&s->kobj);
3192 if (err)
3193 return err;
3194
3195 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3196 if (err)
3197 return err;
3198 kobject_uevent(&s->kobj, KOBJ_ADD);
3199 if (!unmergeable) {
3200 /* Setup first alias */
3201 sysfs_slab_alias(s, s->name);
3202 kfree(name);
3203 }
3204 return 0;
3205}
3206
3207static void sysfs_slab_remove(struct kmem_cache *s)
3208{
3209 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3210 kobject_del(&s->kobj);
3211}
3212
3213/*
3214 * Need to buffer aliases during bootup until sysfs becomes
3215 * available lest we loose that information.
3216 */
3217struct saved_alias {
3218 struct kmem_cache *s;
3219 const char *name;
3220 struct saved_alias *next;
3221};
3222
3223struct saved_alias *alias_list;
3224
3225static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3226{
3227 struct saved_alias *al;
3228
3229 if (slab_state == SYSFS) {
3230 /*
3231 * If we have a leftover link then remove it.
3232 */
3233 sysfs_remove_link(&slab_subsys.kset.kobj, name);
3234 return sysfs_create_link(&slab_subsys.kset.kobj,
3235 &s->kobj, name);
3236 }
3237
3238 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3239 if (!al)
3240 return -ENOMEM;
3241
3242 al->s = s;
3243 al->name = name;
3244 al->next = alias_list;
3245 alias_list = al;
3246 return 0;
3247}
3248
3249static int __init slab_sysfs_init(void)
3250{
3251 int err;
3252
3253 err = subsystem_register(&slab_subsys);
3254 if (err) {
3255 printk(KERN_ERR "Cannot register slab subsystem.\n");
3256 return -ENOSYS;
3257 }
3258
3259 finish_bootstrap();
3260
3261 while (alias_list) {
3262 struct saved_alias *al = alias_list;
3263
3264 alias_list = alias_list->next;
3265 err = sysfs_slab_alias(al->s, al->name);
3266 BUG_ON(err);
3267 kfree(al);
3268 }
3269
3270 resiliency_test();
3271 return 0;
3272}
3273
3274__initcall(slab_sysfs_init);
3275#else
3276__initcall(finish_bootstrap);
3277#endif