spi_s3c24xx signedness fix
[linux-2.6/mini2440.git] / mm / slub.c
blob32b62623846af6a7242c11f641774adeff09d45b
1 /*
2 * SLUB: A slab allocator that limits cache line use instead of queuing
3 * objects in per cpu and per node lists.
5 * The allocator synchronizes using per slab locks and only
6 * uses a centralized lock to manage a pool of partial slabs.
8 * (C) 2007 SGI, Christoph Lameter <clameter@sgi.com>
9 */
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/debugobjects.h>
23 #include <linux/kallsyms.h>
24 #include <linux/memory.h>
25 #include <linux/math64.h>
28 * Lock order:
29 * 1. slab_lock(page)
30 * 2. slab->list_lock
32 * The slab_lock protects operations on the object of a particular
33 * slab and its metadata in the page struct. If the slab lock
34 * has been taken then no allocations nor frees can be performed
35 * on the objects in the slab nor can the slab be added or removed
36 * from the partial or full lists since this would mean modifying
37 * the page_struct of the slab.
39 * The list_lock protects the partial and full list on each node and
40 * the partial slab counter. If taken then no new slabs may be added or
41 * removed from the lists nor make the number of partial slabs be modified.
42 * (Note that the total number of slabs is an atomic value that may be
43 * modified without taking the list lock).
45 * The list_lock is a centralized lock and thus we avoid taking it as
46 * much as possible. As long as SLUB does not have to handle partial
47 * slabs, operations can continue without any centralized lock. F.e.
48 * allocating a long series of objects that fill up slabs does not require
49 * the list lock.
51 * The lock order is sometimes inverted when we are trying to get a slab
52 * off a list. We take the list_lock and then look for a page on the list
53 * to use. While we do that objects in the slabs may be freed. We can
54 * only operate on the slab if we have also taken the slab_lock. So we use
55 * a slab_trylock() on the slab. If trylock was successful then no frees
56 * can occur anymore and we can use the slab for allocations etc. If the
57 * slab_trylock() does not succeed then frees are in progress in the slab and
58 * we must stay away from it for a while since we may cause a bouncing
59 * cacheline if we try to acquire the lock. So go onto the next slab.
60 * If all pages are busy then we may allocate a new slab instead of reusing
61 * a partial slab. A new slab has noone operating on it and thus there is
62 * no danger of cacheline contention.
64 * Interrupts are disabled during allocation and deallocation in order to
65 * make the slab allocator safe to use in the context of an irq. In addition
66 * interrupts are disabled to ensure that the processor does not change
67 * while handling per_cpu slabs, due to kernel preemption.
69 * SLUB assigns one slab for allocation to each processor.
70 * Allocations only occur from these slabs called cpu slabs.
72 * Slabs with free elements are kept on a partial list and during regular
73 * operations no list for full slabs is used. If an object in a full slab is
74 * freed then the slab will show up again on the partial lists.
75 * We track full slabs for debugging purposes though because otherwise we
76 * cannot scan all objects.
78 * Slabs are freed when they become empty. Teardown and setup is
79 * minimal so we rely on the page allocators per cpu caches for
80 * fast frees and allocs.
82 * Overloading of page flags that are otherwise used for LRU management.
84 * PageActive The slab is frozen and exempt from list processing.
85 * This means that the slab is dedicated to a purpose
86 * such as satisfying allocations for a specific
87 * processor. Objects may be freed in the slab while
88 * it is frozen but slab_free will then skip the usual
89 * list operations. It is up to the processor holding
90 * the slab to integrate the slab into the slab lists
91 * when the slab is no longer needed.
93 * One use of this flag is to mark slabs that are
94 * used for allocations. Then such a slab becomes a cpu
95 * slab. The cpu slab may be equipped with an additional
96 * freelist that allows lockless access to
97 * free objects in addition to the regular freelist
98 * that requires the slab lock.
100 * PageError Slab requires special handling due to debug
101 * options set. This moves slab handling out of
102 * the fast path and disables lockless freelists.
105 #define FROZEN (1 << PG_active)
107 #ifdef CONFIG_SLUB_DEBUG
108 #define SLABDEBUG (1 << PG_error)
109 #else
110 #define SLABDEBUG 0
111 #endif
113 static inline int SlabFrozen(struct page *page)
115 return page->flags & FROZEN;
118 static inline void SetSlabFrozen(struct page *page)
120 page->flags |= FROZEN;
123 static inline void ClearSlabFrozen(struct page *page)
125 page->flags &= ~FROZEN;
128 static inline int SlabDebug(struct page *page)
130 return page->flags & SLABDEBUG;
133 static inline void SetSlabDebug(struct page *page)
135 page->flags |= SLABDEBUG;
138 static inline void ClearSlabDebug(struct page *page)
140 page->flags &= ~SLABDEBUG;
144 * Issues still to be resolved:
146 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
148 * - Variable sizing of the per node arrays
151 /* Enable to test recovery from slab corruption on boot */
152 #undef SLUB_RESILIENCY_TEST
155 * Mininum number of partial slabs. These will be left on the partial
156 * lists even if they are empty. kmem_cache_shrink may reclaim them.
158 #define MIN_PARTIAL 5
161 * Maximum number of desirable partial slabs.
162 * The existence of more partial slabs makes kmem_cache_shrink
163 * sort the partial list by the number of objects in the.
165 #define MAX_PARTIAL 10
167 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
168 SLAB_POISON | SLAB_STORE_USER)
171 * Set of flags that will prevent slab merging
173 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
174 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
176 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
177 SLAB_CACHE_DMA)
179 #ifndef ARCH_KMALLOC_MINALIGN
180 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
181 #endif
183 #ifndef ARCH_SLAB_MINALIGN
184 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
185 #endif
187 /* Internal SLUB flags */
188 #define __OBJECT_POISON 0x80000000 /* Poison object */
189 #define __SYSFS_ADD_DEFERRED 0x40000000 /* Not yet visible via sysfs */
191 static int kmem_size = sizeof(struct kmem_cache);
193 #ifdef CONFIG_SMP
194 static struct notifier_block slab_notifier;
195 #endif
197 static enum {
198 DOWN, /* No slab functionality available */
199 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
200 UP, /* Everything works but does not show up in sysfs */
201 SYSFS /* Sysfs up */
202 } slab_state = DOWN;
204 /* A list of all slab caches on the system */
205 static DECLARE_RWSEM(slub_lock);
206 static LIST_HEAD(slab_caches);
209 * Tracking user of a slab.
211 struct track {
212 void *addr; /* Called from address */
213 int cpu; /* Was running on cpu */
214 int pid; /* Pid context */
215 unsigned long when; /* When did the operation occur */
218 enum track_item { TRACK_ALLOC, TRACK_FREE };
220 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
221 static int sysfs_slab_add(struct kmem_cache *);
222 static int sysfs_slab_alias(struct kmem_cache *, const char *);
223 static void sysfs_slab_remove(struct kmem_cache *);
225 #else
226 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
227 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
228 { return 0; }
229 static inline void sysfs_slab_remove(struct kmem_cache *s)
231 kfree(s);
234 #endif
236 static inline void stat(struct kmem_cache_cpu *c, enum stat_item si)
238 #ifdef CONFIG_SLUB_STATS
239 c->stat[si]++;
240 #endif
243 /********************************************************************
244 * Core slab cache functions
245 *******************************************************************/
247 int slab_is_available(void)
249 return slab_state >= UP;
252 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
254 #ifdef CONFIG_NUMA
255 return s->node[node];
256 #else
257 return &s->local_node;
258 #endif
261 static inline struct kmem_cache_cpu *get_cpu_slab(struct kmem_cache *s, int cpu)
263 #ifdef CONFIG_SMP
264 return s->cpu_slab[cpu];
265 #else
266 return &s->cpu_slab;
267 #endif
270 /* Verify that a pointer has an address that is valid within a slab page */
271 static inline int check_valid_pointer(struct kmem_cache *s,
272 struct page *page, const void *object)
274 void *base;
276 if (!object)
277 return 1;
279 base = page_address(page);
280 if (object < base || object >= base + page->objects * s->size ||
281 (object - base) % s->size) {
282 return 0;
285 return 1;
289 * Slow version of get and set free pointer.
291 * This version requires touching the cache lines of kmem_cache which
292 * we avoid to do in the fast alloc free paths. There we obtain the offset
293 * from the page struct.
295 static inline void *get_freepointer(struct kmem_cache *s, void *object)
297 return *(void **)(object + s->offset);
300 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
302 *(void **)(object + s->offset) = fp;
305 /* Loop over all objects in a slab */
306 #define for_each_object(__p, __s, __addr, __objects) \
307 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
308 __p += (__s)->size)
310 /* Scan freelist */
311 #define for_each_free_object(__p, __s, __free) \
312 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
314 /* Determine object index from a given position */
315 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
317 return (p - addr) / s->size;
320 static inline struct kmem_cache_order_objects oo_make(int order,
321 unsigned long size)
323 struct kmem_cache_order_objects x = {
324 (order << 16) + (PAGE_SIZE << order) / size
327 return x;
330 static inline int oo_order(struct kmem_cache_order_objects x)
332 return x.x >> 16;
335 static inline int oo_objects(struct kmem_cache_order_objects x)
337 return x.x & ((1 << 16) - 1);
340 #ifdef CONFIG_SLUB_DEBUG
342 * Debug settings:
344 #ifdef CONFIG_SLUB_DEBUG_ON
345 static int slub_debug = DEBUG_DEFAULT_FLAGS;
346 #else
347 static int slub_debug;
348 #endif
350 static char *slub_debug_slabs;
353 * Object debugging
355 static void print_section(char *text, u8 *addr, unsigned int length)
357 int i, offset;
358 int newline = 1;
359 char ascii[17];
361 ascii[16] = 0;
363 for (i = 0; i < length; i++) {
364 if (newline) {
365 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
366 newline = 0;
368 printk(KERN_CONT " %02x", addr[i]);
369 offset = i % 16;
370 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
371 if (offset == 15) {
372 printk(KERN_CONT " %s\n", ascii);
373 newline = 1;
376 if (!newline) {
377 i %= 16;
378 while (i < 16) {
379 printk(KERN_CONT " ");
380 ascii[i] = ' ';
381 i++;
383 printk(KERN_CONT " %s\n", ascii);
387 static struct track *get_track(struct kmem_cache *s, void *object,
388 enum track_item alloc)
390 struct track *p;
392 if (s->offset)
393 p = object + s->offset + sizeof(void *);
394 else
395 p = object + s->inuse;
397 return p + alloc;
400 static void set_track(struct kmem_cache *s, void *object,
401 enum track_item alloc, void *addr)
403 struct track *p;
405 if (s->offset)
406 p = object + s->offset + sizeof(void *);
407 else
408 p = object + s->inuse;
410 p += alloc;
411 if (addr) {
412 p->addr = addr;
413 p->cpu = smp_processor_id();
414 p->pid = current ? current->pid : -1;
415 p->when = jiffies;
416 } else
417 memset(p, 0, sizeof(struct track));
420 static void init_tracking(struct kmem_cache *s, void *object)
422 if (!(s->flags & SLAB_STORE_USER))
423 return;
425 set_track(s, object, TRACK_FREE, NULL);
426 set_track(s, object, TRACK_ALLOC, NULL);
429 static void print_track(const char *s, struct track *t)
431 if (!t->addr)
432 return;
434 printk(KERN_ERR "INFO: %s in ", s);
435 __print_symbol("%s", (unsigned long)t->addr);
436 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
439 static void print_tracking(struct kmem_cache *s, void *object)
441 if (!(s->flags & SLAB_STORE_USER))
442 return;
444 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
445 print_track("Freed", get_track(s, object, TRACK_FREE));
448 static void print_page_info(struct page *page)
450 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
451 page, page->objects, page->inuse, page->freelist, page->flags);
455 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
457 va_list args;
458 char buf[100];
460 va_start(args, fmt);
461 vsnprintf(buf, sizeof(buf), fmt, args);
462 va_end(args);
463 printk(KERN_ERR "========================================"
464 "=====================================\n");
465 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
466 printk(KERN_ERR "----------------------------------------"
467 "-------------------------------------\n\n");
470 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
472 va_list args;
473 char buf[100];
475 va_start(args, fmt);
476 vsnprintf(buf, sizeof(buf), fmt, args);
477 va_end(args);
478 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
481 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
483 unsigned int off; /* Offset of last byte */
484 u8 *addr = page_address(page);
486 print_tracking(s, p);
488 print_page_info(page);
490 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
491 p, p - addr, get_freepointer(s, p));
493 if (p > addr + 16)
494 print_section("Bytes b4", p - 16, 16);
496 print_section("Object", p, min(s->objsize, 128));
498 if (s->flags & SLAB_RED_ZONE)
499 print_section("Redzone", p + s->objsize,
500 s->inuse - s->objsize);
502 if (s->offset)
503 off = s->offset + sizeof(void *);
504 else
505 off = s->inuse;
507 if (s->flags & SLAB_STORE_USER)
508 off += 2 * sizeof(struct track);
510 if (off != s->size)
511 /* Beginning of the filler is the free pointer */
512 print_section("Padding", p + off, s->size - off);
514 dump_stack();
517 static void object_err(struct kmem_cache *s, struct page *page,
518 u8 *object, char *reason)
520 slab_bug(s, "%s", reason);
521 print_trailer(s, page, object);
524 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
526 va_list args;
527 char buf[100];
529 va_start(args, fmt);
530 vsnprintf(buf, sizeof(buf), fmt, args);
531 va_end(args);
532 slab_bug(s, "%s", buf);
533 print_page_info(page);
534 dump_stack();
537 static void init_object(struct kmem_cache *s, void *object, int active)
539 u8 *p = object;
541 if (s->flags & __OBJECT_POISON) {
542 memset(p, POISON_FREE, s->objsize - 1);
543 p[s->objsize - 1] = POISON_END;
546 if (s->flags & SLAB_RED_ZONE)
547 memset(p + s->objsize,
548 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
549 s->inuse - s->objsize);
552 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
554 while (bytes) {
555 if (*start != (u8)value)
556 return start;
557 start++;
558 bytes--;
560 return NULL;
563 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
564 void *from, void *to)
566 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
567 memset(from, data, to - from);
570 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
571 u8 *object, char *what,
572 u8 *start, unsigned int value, unsigned int bytes)
574 u8 *fault;
575 u8 *end;
577 fault = check_bytes(start, value, bytes);
578 if (!fault)
579 return 1;
581 end = start + bytes;
582 while (end > fault && end[-1] == value)
583 end--;
585 slab_bug(s, "%s overwritten", what);
586 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
587 fault, end - 1, fault[0], value);
588 print_trailer(s, page, object);
590 restore_bytes(s, what, value, fault, end);
591 return 0;
595 * Object layout:
597 * object address
598 * Bytes of the object to be managed.
599 * If the freepointer may overlay the object then the free
600 * pointer is the first word of the object.
602 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
603 * 0xa5 (POISON_END)
605 * object + s->objsize
606 * Padding to reach word boundary. This is also used for Redzoning.
607 * Padding is extended by another word if Redzoning is enabled and
608 * objsize == inuse.
610 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
611 * 0xcc (RED_ACTIVE) for objects in use.
613 * object + s->inuse
614 * Meta data starts here.
616 * A. Free pointer (if we cannot overwrite object on free)
617 * B. Tracking data for SLAB_STORE_USER
618 * C. Padding to reach required alignment boundary or at mininum
619 * one word if debugging is on to be able to detect writes
620 * before the word boundary.
622 * Padding is done using 0x5a (POISON_INUSE)
624 * object + s->size
625 * Nothing is used beyond s->size.
627 * If slabcaches are merged then the objsize and inuse boundaries are mostly
628 * ignored. And therefore no slab options that rely on these boundaries
629 * may be used with merged slabcaches.
632 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
634 unsigned long off = s->inuse; /* The end of info */
636 if (s->offset)
637 /* Freepointer is placed after the object. */
638 off += sizeof(void *);
640 if (s->flags & SLAB_STORE_USER)
641 /* We also have user information there */
642 off += 2 * sizeof(struct track);
644 if (s->size == off)
645 return 1;
647 return check_bytes_and_report(s, page, p, "Object padding",
648 p + off, POISON_INUSE, s->size - off);
651 /* Check the pad bytes at the end of a slab page */
652 static int slab_pad_check(struct kmem_cache *s, struct page *page)
654 u8 *start;
655 u8 *fault;
656 u8 *end;
657 int length;
658 int remainder;
660 if (!(s->flags & SLAB_POISON))
661 return 1;
663 start = page_address(page);
664 length = (PAGE_SIZE << compound_order(page));
665 end = start + length;
666 remainder = length % s->size;
667 if (!remainder)
668 return 1;
670 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
671 if (!fault)
672 return 1;
673 while (end > fault && end[-1] == POISON_INUSE)
674 end--;
676 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
677 print_section("Padding", end - remainder, remainder);
679 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
680 return 0;
683 static int check_object(struct kmem_cache *s, struct page *page,
684 void *object, int active)
686 u8 *p = object;
687 u8 *endobject = object + s->objsize;
689 if (s->flags & SLAB_RED_ZONE) {
690 unsigned int red =
691 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
693 if (!check_bytes_and_report(s, page, object, "Redzone",
694 endobject, red, s->inuse - s->objsize))
695 return 0;
696 } else {
697 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
698 check_bytes_and_report(s, page, p, "Alignment padding",
699 endobject, POISON_INUSE, s->inuse - s->objsize);
703 if (s->flags & SLAB_POISON) {
704 if (!active && (s->flags & __OBJECT_POISON) &&
705 (!check_bytes_and_report(s, page, p, "Poison", p,
706 POISON_FREE, s->objsize - 1) ||
707 !check_bytes_and_report(s, page, p, "Poison",
708 p + s->objsize - 1, POISON_END, 1)))
709 return 0;
711 * check_pad_bytes cleans up on its own.
713 check_pad_bytes(s, page, p);
716 if (!s->offset && active)
718 * Object and freepointer overlap. Cannot check
719 * freepointer while object is allocated.
721 return 1;
723 /* Check free pointer validity */
724 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
725 object_err(s, page, p, "Freepointer corrupt");
727 * No choice but to zap it and thus loose the remainder
728 * of the free objects in this slab. May cause
729 * another error because the object count is now wrong.
731 set_freepointer(s, p, NULL);
732 return 0;
734 return 1;
737 static int check_slab(struct kmem_cache *s, struct page *page)
739 int maxobj;
741 VM_BUG_ON(!irqs_disabled());
743 if (!PageSlab(page)) {
744 slab_err(s, page, "Not a valid slab page");
745 return 0;
748 maxobj = (PAGE_SIZE << compound_order(page)) / s->size;
749 if (page->objects > maxobj) {
750 slab_err(s, page, "objects %u > max %u",
751 s->name, page->objects, maxobj);
752 return 0;
754 if (page->inuse > page->objects) {
755 slab_err(s, page, "inuse %u > max %u",
756 s->name, page->inuse, page->objects);
757 return 0;
759 /* Slab_pad_check fixes things up after itself */
760 slab_pad_check(s, page);
761 return 1;
765 * Determine if a certain object on a page is on the freelist. Must hold the
766 * slab lock to guarantee that the chains are in a consistent state.
768 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
770 int nr = 0;
771 void *fp = page->freelist;
772 void *object = NULL;
773 unsigned long max_objects;
775 while (fp && nr <= page->objects) {
776 if (fp == search)
777 return 1;
778 if (!check_valid_pointer(s, page, fp)) {
779 if (object) {
780 object_err(s, page, object,
781 "Freechain corrupt");
782 set_freepointer(s, object, NULL);
783 break;
784 } else {
785 slab_err(s, page, "Freepointer corrupt");
786 page->freelist = NULL;
787 page->inuse = page->objects;
788 slab_fix(s, "Freelist cleared");
789 return 0;
791 break;
793 object = fp;
794 fp = get_freepointer(s, object);
795 nr++;
798 max_objects = (PAGE_SIZE << compound_order(page)) / s->size;
799 if (max_objects > 65535)
800 max_objects = 65535;
802 if (page->objects != max_objects) {
803 slab_err(s, page, "Wrong number of objects. Found %d but "
804 "should be %d", page->objects, max_objects);
805 page->objects = max_objects;
806 slab_fix(s, "Number of objects adjusted.");
808 if (page->inuse != page->objects - nr) {
809 slab_err(s, page, "Wrong object count. Counter is %d but "
810 "counted were %d", page->inuse, page->objects - nr);
811 page->inuse = page->objects - nr;
812 slab_fix(s, "Object count adjusted.");
814 return search == NULL;
817 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
819 if (s->flags & SLAB_TRACE) {
820 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
821 s->name,
822 alloc ? "alloc" : "free",
823 object, page->inuse,
824 page->freelist);
826 if (!alloc)
827 print_section("Object", (void *)object, s->objsize);
829 dump_stack();
834 * Tracking of fully allocated slabs for debugging purposes.
836 static void add_full(struct kmem_cache_node *n, struct page *page)
838 spin_lock(&n->list_lock);
839 list_add(&page->lru, &n->full);
840 spin_unlock(&n->list_lock);
843 static void remove_full(struct kmem_cache *s, struct page *page)
845 struct kmem_cache_node *n;
847 if (!(s->flags & SLAB_STORE_USER))
848 return;
850 n = get_node(s, page_to_nid(page));
852 spin_lock(&n->list_lock);
853 list_del(&page->lru);
854 spin_unlock(&n->list_lock);
857 /* Tracking of the number of slabs for debugging purposes */
858 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
860 struct kmem_cache_node *n = get_node(s, node);
862 return atomic_long_read(&n->nr_slabs);
865 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
867 struct kmem_cache_node *n = get_node(s, node);
870 * May be called early in order to allocate a slab for the
871 * kmem_cache_node structure. Solve the chicken-egg
872 * dilemma by deferring the increment of the count during
873 * bootstrap (see early_kmem_cache_node_alloc).
875 if (!NUMA_BUILD || n) {
876 atomic_long_inc(&n->nr_slabs);
877 atomic_long_add(objects, &n->total_objects);
880 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
882 struct kmem_cache_node *n = get_node(s, node);
884 atomic_long_dec(&n->nr_slabs);
885 atomic_long_sub(objects, &n->total_objects);
888 /* Object debug checks for alloc/free paths */
889 static void setup_object_debug(struct kmem_cache *s, struct page *page,
890 void *object)
892 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
893 return;
895 init_object(s, object, 0);
896 init_tracking(s, object);
899 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
900 void *object, void *addr)
902 if (!check_slab(s, page))
903 goto bad;
905 if (!on_freelist(s, page, object)) {
906 object_err(s, page, object, "Object already allocated");
907 goto bad;
910 if (!check_valid_pointer(s, page, object)) {
911 object_err(s, page, object, "Freelist Pointer check fails");
912 goto bad;
915 if (!check_object(s, page, object, 0))
916 goto bad;
918 /* Success perform special debug activities for allocs */
919 if (s->flags & SLAB_STORE_USER)
920 set_track(s, object, TRACK_ALLOC, addr);
921 trace(s, page, object, 1);
922 init_object(s, object, 1);
923 return 1;
925 bad:
926 if (PageSlab(page)) {
928 * If this is a slab page then lets do the best we can
929 * to avoid issues in the future. Marking all objects
930 * as used avoids touching the remaining objects.
932 slab_fix(s, "Marking all objects used");
933 page->inuse = page->objects;
934 page->freelist = NULL;
936 return 0;
939 static int free_debug_processing(struct kmem_cache *s, struct page *page,
940 void *object, void *addr)
942 if (!check_slab(s, page))
943 goto fail;
945 if (!check_valid_pointer(s, page, object)) {
946 slab_err(s, page, "Invalid object pointer 0x%p", object);
947 goto fail;
950 if (on_freelist(s, page, object)) {
951 object_err(s, page, object, "Object already free");
952 goto fail;
955 if (!check_object(s, page, object, 1))
956 return 0;
958 if (unlikely(s != page->slab)) {
959 if (!PageSlab(page)) {
960 slab_err(s, page, "Attempt to free object(0x%p) "
961 "outside of slab", object);
962 } else if (!page->slab) {
963 printk(KERN_ERR
964 "SLUB <none>: no slab for object 0x%p.\n",
965 object);
966 dump_stack();
967 } else
968 object_err(s, page, object,
969 "page slab pointer corrupt.");
970 goto fail;
973 /* Special debug activities for freeing objects */
974 if (!SlabFrozen(page) && !page->freelist)
975 remove_full(s, page);
976 if (s->flags & SLAB_STORE_USER)
977 set_track(s, object, TRACK_FREE, addr);
978 trace(s, page, object, 0);
979 init_object(s, object, 0);
980 return 1;
982 fail:
983 slab_fix(s, "Object at 0x%p not freed", object);
984 return 0;
987 static int __init setup_slub_debug(char *str)
989 slub_debug = DEBUG_DEFAULT_FLAGS;
990 if (*str++ != '=' || !*str)
992 * No options specified. Switch on full debugging.
994 goto out;
996 if (*str == ',')
998 * No options but restriction on slabs. This means full
999 * debugging for slabs matching a pattern.
1001 goto check_slabs;
1003 slub_debug = 0;
1004 if (*str == '-')
1006 * Switch off all debugging measures.
1008 goto out;
1011 * Determine which debug features should be switched on
1013 for (; *str && *str != ','; str++) {
1014 switch (tolower(*str)) {
1015 case 'f':
1016 slub_debug |= SLAB_DEBUG_FREE;
1017 break;
1018 case 'z':
1019 slub_debug |= SLAB_RED_ZONE;
1020 break;
1021 case 'p':
1022 slub_debug |= SLAB_POISON;
1023 break;
1024 case 'u':
1025 slub_debug |= SLAB_STORE_USER;
1026 break;
1027 case 't':
1028 slub_debug |= SLAB_TRACE;
1029 break;
1030 default:
1031 printk(KERN_ERR "slub_debug option '%c' "
1032 "unknown. skipped\n", *str);
1036 check_slabs:
1037 if (*str == ',')
1038 slub_debug_slabs = str + 1;
1039 out:
1040 return 1;
1043 __setup("slub_debug", setup_slub_debug);
1045 static unsigned long kmem_cache_flags(unsigned long objsize,
1046 unsigned long flags, const char *name,
1047 void (*ctor)(struct kmem_cache *, void *))
1050 * Enable debugging if selected on the kernel commandline.
1052 if (slub_debug && (!slub_debug_slabs ||
1053 strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs)) == 0))
1054 flags |= slub_debug;
1056 return flags;
1058 #else
1059 static inline void setup_object_debug(struct kmem_cache *s,
1060 struct page *page, void *object) {}
1062 static inline int alloc_debug_processing(struct kmem_cache *s,
1063 struct page *page, void *object, void *addr) { return 0; }
1065 static inline int free_debug_processing(struct kmem_cache *s,
1066 struct page *page, void *object, void *addr) { return 0; }
1068 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1069 { return 1; }
1070 static inline int check_object(struct kmem_cache *s, struct page *page,
1071 void *object, int active) { return 1; }
1072 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1073 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1074 unsigned long flags, const char *name,
1075 void (*ctor)(struct kmem_cache *, void *))
1077 return flags;
1079 #define slub_debug 0
1081 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1082 { return 0; }
1083 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1084 int objects) {}
1085 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1086 int objects) {}
1087 #endif
1090 * Slab allocation and freeing
1092 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1093 struct kmem_cache_order_objects oo)
1095 int order = oo_order(oo);
1097 if (node == -1)
1098 return alloc_pages(flags, order);
1099 else
1100 return alloc_pages_node(node, flags, order);
1103 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1105 struct page *page;
1106 struct kmem_cache_order_objects oo = s->oo;
1108 flags |= s->allocflags;
1110 page = alloc_slab_page(flags | __GFP_NOWARN | __GFP_NORETRY, node,
1111 oo);
1112 if (unlikely(!page)) {
1113 oo = s->min;
1115 * Allocation may have failed due to fragmentation.
1116 * Try a lower order alloc if possible
1118 page = alloc_slab_page(flags, node, oo);
1119 if (!page)
1120 return NULL;
1122 stat(get_cpu_slab(s, raw_smp_processor_id()), ORDER_FALLBACK);
1124 page->objects = oo_objects(oo);
1125 mod_zone_page_state(page_zone(page),
1126 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1127 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1128 1 << oo_order(oo));
1130 return page;
1133 static void setup_object(struct kmem_cache *s, struct page *page,
1134 void *object)
1136 setup_object_debug(s, page, object);
1137 if (unlikely(s->ctor))
1138 s->ctor(s, object);
1141 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1143 struct page *page;
1144 void *start;
1145 void *last;
1146 void *p;
1148 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1150 page = allocate_slab(s,
1151 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1152 if (!page)
1153 goto out;
1155 inc_slabs_node(s, page_to_nid(page), page->objects);
1156 page->slab = s;
1157 page->flags |= 1 << PG_slab;
1158 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1159 SLAB_STORE_USER | SLAB_TRACE))
1160 SetSlabDebug(page);
1162 start = page_address(page);
1164 if (unlikely(s->flags & SLAB_POISON))
1165 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1167 last = start;
1168 for_each_object(p, s, start, page->objects) {
1169 setup_object(s, page, last);
1170 set_freepointer(s, last, p);
1171 last = p;
1173 setup_object(s, page, last);
1174 set_freepointer(s, last, NULL);
1176 page->freelist = start;
1177 page->inuse = 0;
1178 out:
1179 return page;
1182 static void __free_slab(struct kmem_cache *s, struct page *page)
1184 int order = compound_order(page);
1185 int pages = 1 << order;
1187 if (unlikely(SlabDebug(page))) {
1188 void *p;
1190 slab_pad_check(s, page);
1191 for_each_object(p, s, page_address(page),
1192 page->objects)
1193 check_object(s, page, p, 0);
1194 ClearSlabDebug(page);
1197 mod_zone_page_state(page_zone(page),
1198 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1199 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1200 -pages);
1202 __ClearPageSlab(page);
1203 reset_page_mapcount(page);
1204 __free_pages(page, order);
1207 static void rcu_free_slab(struct rcu_head *h)
1209 struct page *page;
1211 page = container_of((struct list_head *)h, struct page, lru);
1212 __free_slab(page->slab, page);
1215 static void free_slab(struct kmem_cache *s, struct page *page)
1217 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1219 * RCU free overloads the RCU head over the LRU
1221 struct rcu_head *head = (void *)&page->lru;
1223 call_rcu(head, rcu_free_slab);
1224 } else
1225 __free_slab(s, page);
1228 static void discard_slab(struct kmem_cache *s, struct page *page)
1230 dec_slabs_node(s, page_to_nid(page), page->objects);
1231 free_slab(s, page);
1235 * Per slab locking using the pagelock
1237 static __always_inline void slab_lock(struct page *page)
1239 bit_spin_lock(PG_locked, &page->flags);
1242 static __always_inline void slab_unlock(struct page *page)
1244 __bit_spin_unlock(PG_locked, &page->flags);
1247 static __always_inline int slab_trylock(struct page *page)
1249 int rc = 1;
1251 rc = bit_spin_trylock(PG_locked, &page->flags);
1252 return rc;
1256 * Management of partially allocated slabs
1258 static void add_partial(struct kmem_cache_node *n,
1259 struct page *page, int tail)
1261 spin_lock(&n->list_lock);
1262 n->nr_partial++;
1263 if (tail)
1264 list_add_tail(&page->lru, &n->partial);
1265 else
1266 list_add(&page->lru, &n->partial);
1267 spin_unlock(&n->list_lock);
1270 static void remove_partial(struct kmem_cache *s,
1271 struct page *page)
1273 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1275 spin_lock(&n->list_lock);
1276 list_del(&page->lru);
1277 n->nr_partial--;
1278 spin_unlock(&n->list_lock);
1282 * Lock slab and remove from the partial list.
1284 * Must hold list_lock.
1286 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1288 if (slab_trylock(page)) {
1289 list_del(&page->lru);
1290 n->nr_partial--;
1291 SetSlabFrozen(page);
1292 return 1;
1294 return 0;
1298 * Try to allocate a partial slab from a specific node.
1300 static struct page *get_partial_node(struct kmem_cache_node *n)
1302 struct page *page;
1305 * Racy check. If we mistakenly see no partial slabs then we
1306 * just allocate an empty slab. If we mistakenly try to get a
1307 * partial slab and there is none available then get_partials()
1308 * will return NULL.
1310 if (!n || !n->nr_partial)
1311 return NULL;
1313 spin_lock(&n->list_lock);
1314 list_for_each_entry(page, &n->partial, lru)
1315 if (lock_and_freeze_slab(n, page))
1316 goto out;
1317 page = NULL;
1318 out:
1319 spin_unlock(&n->list_lock);
1320 return page;
1324 * Get a page from somewhere. Search in increasing NUMA distances.
1326 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1328 #ifdef CONFIG_NUMA
1329 struct zonelist *zonelist;
1330 struct zoneref *z;
1331 struct zone *zone;
1332 enum zone_type high_zoneidx = gfp_zone(flags);
1333 struct page *page;
1336 * The defrag ratio allows a configuration of the tradeoffs between
1337 * inter node defragmentation and node local allocations. A lower
1338 * defrag_ratio increases the tendency to do local allocations
1339 * instead of attempting to obtain partial slabs from other nodes.
1341 * If the defrag_ratio is set to 0 then kmalloc() always
1342 * returns node local objects. If the ratio is higher then kmalloc()
1343 * may return off node objects because partial slabs are obtained
1344 * from other nodes and filled up.
1346 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1347 * defrag_ratio = 1000) then every (well almost) allocation will
1348 * first attempt to defrag slab caches on other nodes. This means
1349 * scanning over all nodes to look for partial slabs which may be
1350 * expensive if we do it every time we are trying to find a slab
1351 * with available objects.
1353 if (!s->remote_node_defrag_ratio ||
1354 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1355 return NULL;
1357 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1358 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1359 struct kmem_cache_node *n;
1361 n = get_node(s, zone_to_nid(zone));
1363 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1364 n->nr_partial > MIN_PARTIAL) {
1365 page = get_partial_node(n);
1366 if (page)
1367 return page;
1370 #endif
1371 return NULL;
1375 * Get a partial page, lock it and return it.
1377 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1379 struct page *page;
1380 int searchnode = (node == -1) ? numa_node_id() : node;
1382 page = get_partial_node(get_node(s, searchnode));
1383 if (page || (flags & __GFP_THISNODE))
1384 return page;
1386 return get_any_partial(s, flags);
1390 * Move a page back to the lists.
1392 * Must be called with the slab lock held.
1394 * On exit the slab lock will have been dropped.
1396 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1398 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1399 struct kmem_cache_cpu *c = get_cpu_slab(s, smp_processor_id());
1401 ClearSlabFrozen(page);
1402 if (page->inuse) {
1404 if (page->freelist) {
1405 add_partial(n, page, tail);
1406 stat(c, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1407 } else {
1408 stat(c, DEACTIVATE_FULL);
1409 if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1410 add_full(n, page);
1412 slab_unlock(page);
1413 } else {
1414 stat(c, DEACTIVATE_EMPTY);
1415 if (n->nr_partial < MIN_PARTIAL) {
1417 * Adding an empty slab to the partial slabs in order
1418 * to avoid page allocator overhead. This slab needs
1419 * to come after the other slabs with objects in
1420 * so that the others get filled first. That way the
1421 * size of the partial list stays small.
1423 * kmem_cache_shrink can reclaim any empty slabs from the
1424 * partial list.
1426 add_partial(n, page, 1);
1427 slab_unlock(page);
1428 } else {
1429 slab_unlock(page);
1430 stat(get_cpu_slab(s, raw_smp_processor_id()), FREE_SLAB);
1431 discard_slab(s, page);
1437 * Remove the cpu slab
1439 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1441 struct page *page = c->page;
1442 int tail = 1;
1444 if (page->freelist)
1445 stat(c, DEACTIVATE_REMOTE_FREES);
1447 * Merge cpu freelist into slab freelist. Typically we get here
1448 * because both freelists are empty. So this is unlikely
1449 * to occur.
1451 while (unlikely(c->freelist)) {
1452 void **object;
1454 tail = 0; /* Hot objects. Put the slab first */
1456 /* Retrieve object from cpu_freelist */
1457 object = c->freelist;
1458 c->freelist = c->freelist[c->offset];
1460 /* And put onto the regular freelist */
1461 object[c->offset] = page->freelist;
1462 page->freelist = object;
1463 page->inuse--;
1465 c->page = NULL;
1466 unfreeze_slab(s, page, tail);
1469 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1471 stat(c, CPUSLAB_FLUSH);
1472 slab_lock(c->page);
1473 deactivate_slab(s, c);
1477 * Flush cpu slab.
1479 * Called from IPI handler with interrupts disabled.
1481 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1483 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
1485 if (likely(c && c->page))
1486 flush_slab(s, c);
1489 static void flush_cpu_slab(void *d)
1491 struct kmem_cache *s = d;
1493 __flush_cpu_slab(s, smp_processor_id());
1496 static void flush_all(struct kmem_cache *s)
1498 #ifdef CONFIG_SMP
1499 on_each_cpu(flush_cpu_slab, s, 1, 1);
1500 #else
1501 unsigned long flags;
1503 local_irq_save(flags);
1504 flush_cpu_slab(s);
1505 local_irq_restore(flags);
1506 #endif
1510 * Check if the objects in a per cpu structure fit numa
1511 * locality expectations.
1513 static inline int node_match(struct kmem_cache_cpu *c, int node)
1515 #ifdef CONFIG_NUMA
1516 if (node != -1 && c->node != node)
1517 return 0;
1518 #endif
1519 return 1;
1523 * Slow path. The lockless freelist is empty or we need to perform
1524 * debugging duties.
1526 * Interrupts are disabled.
1528 * Processing is still very fast if new objects have been freed to the
1529 * regular freelist. In that case we simply take over the regular freelist
1530 * as the lockless freelist and zap the regular freelist.
1532 * If that is not working then we fall back to the partial lists. We take the
1533 * first element of the freelist as the object to allocate now and move the
1534 * rest of the freelist to the lockless freelist.
1536 * And if we were unable to get a new slab from the partial slab lists then
1537 * we need to allocate a new slab. This is the slowest path since it involves
1538 * a call to the page allocator and the setup of a new slab.
1540 static void *__slab_alloc(struct kmem_cache *s,
1541 gfp_t gfpflags, int node, void *addr, struct kmem_cache_cpu *c)
1543 void **object;
1544 struct page *new;
1546 /* We handle __GFP_ZERO in the caller */
1547 gfpflags &= ~__GFP_ZERO;
1549 if (!c->page)
1550 goto new_slab;
1552 slab_lock(c->page);
1553 if (unlikely(!node_match(c, node)))
1554 goto another_slab;
1556 stat(c, ALLOC_REFILL);
1558 load_freelist:
1559 object = c->page->freelist;
1560 if (unlikely(!object))
1561 goto another_slab;
1562 if (unlikely(SlabDebug(c->page)))
1563 goto debug;
1565 c->freelist = object[c->offset];
1566 c->page->inuse = c->page->objects;
1567 c->page->freelist = NULL;
1568 c->node = page_to_nid(c->page);
1569 unlock_out:
1570 slab_unlock(c->page);
1571 stat(c, ALLOC_SLOWPATH);
1572 return object;
1574 another_slab:
1575 deactivate_slab(s, c);
1577 new_slab:
1578 new = get_partial(s, gfpflags, node);
1579 if (new) {
1580 c->page = new;
1581 stat(c, ALLOC_FROM_PARTIAL);
1582 goto load_freelist;
1585 if (gfpflags & __GFP_WAIT)
1586 local_irq_enable();
1588 new = new_slab(s, gfpflags, node);
1590 if (gfpflags & __GFP_WAIT)
1591 local_irq_disable();
1593 if (new) {
1594 c = get_cpu_slab(s, smp_processor_id());
1595 stat(c, ALLOC_SLAB);
1596 if (c->page)
1597 flush_slab(s, c);
1598 slab_lock(new);
1599 SetSlabFrozen(new);
1600 c->page = new;
1601 goto load_freelist;
1603 return NULL;
1604 debug:
1605 if (!alloc_debug_processing(s, c->page, object, addr))
1606 goto another_slab;
1608 c->page->inuse++;
1609 c->page->freelist = object[c->offset];
1610 c->node = -1;
1611 goto unlock_out;
1615 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1616 * have the fastpath folded into their functions. So no function call
1617 * overhead for requests that can be satisfied on the fastpath.
1619 * The fastpath works by first checking if the lockless freelist can be used.
1620 * If not then __slab_alloc is called for slow processing.
1622 * Otherwise we can simply pick the next object from the lockless free list.
1624 static __always_inline void *slab_alloc(struct kmem_cache *s,
1625 gfp_t gfpflags, int node, void *addr)
1627 void **object;
1628 struct kmem_cache_cpu *c;
1629 unsigned long flags;
1631 local_irq_save(flags);
1632 c = get_cpu_slab(s, smp_processor_id());
1633 if (unlikely(!c->freelist || !node_match(c, node)))
1635 object = __slab_alloc(s, gfpflags, node, addr, c);
1637 else {
1638 object = c->freelist;
1639 c->freelist = object[c->offset];
1640 stat(c, ALLOC_FASTPATH);
1642 local_irq_restore(flags);
1644 if (unlikely((gfpflags & __GFP_ZERO) && object))
1645 memset(object, 0, c->objsize);
1647 return object;
1650 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1652 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1654 EXPORT_SYMBOL(kmem_cache_alloc);
1656 #ifdef CONFIG_NUMA
1657 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1659 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1661 EXPORT_SYMBOL(kmem_cache_alloc_node);
1662 #endif
1665 * Slow patch handling. This may still be called frequently since objects
1666 * have a longer lifetime than the cpu slabs in most processing loads.
1668 * So we still attempt to reduce cache line usage. Just take the slab
1669 * lock and free the item. If there is no additional partial page
1670 * handling required then we can return immediately.
1672 static void __slab_free(struct kmem_cache *s, struct page *page,
1673 void *x, void *addr, unsigned int offset)
1675 void *prior;
1676 void **object = (void *)x;
1677 struct kmem_cache_cpu *c;
1679 c = get_cpu_slab(s, raw_smp_processor_id());
1680 stat(c, FREE_SLOWPATH);
1681 slab_lock(page);
1683 if (unlikely(SlabDebug(page)))
1684 goto debug;
1686 checks_ok:
1687 prior = object[offset] = page->freelist;
1688 page->freelist = object;
1689 page->inuse--;
1691 if (unlikely(SlabFrozen(page))) {
1692 stat(c, FREE_FROZEN);
1693 goto out_unlock;
1696 if (unlikely(!page->inuse))
1697 goto slab_empty;
1700 * Objects left in the slab. If it was not on the partial list before
1701 * then add it.
1703 if (unlikely(!prior)) {
1704 add_partial(get_node(s, page_to_nid(page)), page, 1);
1705 stat(c, FREE_ADD_PARTIAL);
1708 out_unlock:
1709 slab_unlock(page);
1710 return;
1712 slab_empty:
1713 if (prior) {
1715 * Slab still on the partial list.
1717 remove_partial(s, page);
1718 stat(c, FREE_REMOVE_PARTIAL);
1720 slab_unlock(page);
1721 stat(c, FREE_SLAB);
1722 discard_slab(s, page);
1723 return;
1725 debug:
1726 if (!free_debug_processing(s, page, x, addr))
1727 goto out_unlock;
1728 goto checks_ok;
1732 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1733 * can perform fastpath freeing without additional function calls.
1735 * The fastpath is only possible if we are freeing to the current cpu slab
1736 * of this processor. This typically the case if we have just allocated
1737 * the item before.
1739 * If fastpath is not possible then fall back to __slab_free where we deal
1740 * with all sorts of special processing.
1742 static __always_inline void slab_free(struct kmem_cache *s,
1743 struct page *page, void *x, void *addr)
1745 void **object = (void *)x;
1746 struct kmem_cache_cpu *c;
1747 unsigned long flags;
1749 local_irq_save(flags);
1750 c = get_cpu_slab(s, smp_processor_id());
1751 debug_check_no_locks_freed(object, c->objsize);
1752 if (!(s->flags & SLAB_DEBUG_OBJECTS))
1753 debug_check_no_obj_freed(object, s->objsize);
1754 if (likely(page == c->page && c->node >= 0)) {
1755 object[c->offset] = c->freelist;
1756 c->freelist = object;
1757 stat(c, FREE_FASTPATH);
1758 } else
1759 __slab_free(s, page, x, addr, c->offset);
1761 local_irq_restore(flags);
1764 void kmem_cache_free(struct kmem_cache *s, void *x)
1766 struct page *page;
1768 page = virt_to_head_page(x);
1770 slab_free(s, page, x, __builtin_return_address(0));
1772 EXPORT_SYMBOL(kmem_cache_free);
1774 /* Figure out on which slab object the object resides */
1775 static struct page *get_object_page(const void *x)
1777 struct page *page = virt_to_head_page(x);
1779 if (!PageSlab(page))
1780 return NULL;
1782 return page;
1786 * Object placement in a slab is made very easy because we always start at
1787 * offset 0. If we tune the size of the object to the alignment then we can
1788 * get the required alignment by putting one properly sized object after
1789 * another.
1791 * Notice that the allocation order determines the sizes of the per cpu
1792 * caches. Each processor has always one slab available for allocations.
1793 * Increasing the allocation order reduces the number of times that slabs
1794 * must be moved on and off the partial lists and is therefore a factor in
1795 * locking overhead.
1799 * Mininum / Maximum order of slab pages. This influences locking overhead
1800 * and slab fragmentation. A higher order reduces the number of partial slabs
1801 * and increases the number of allocations possible without having to
1802 * take the list_lock.
1804 static int slub_min_order;
1805 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
1806 static int slub_min_objects;
1809 * Merge control. If this is set then no merging of slab caches will occur.
1810 * (Could be removed. This was introduced to pacify the merge skeptics.)
1812 static int slub_nomerge;
1815 * Calculate the order of allocation given an slab object size.
1817 * The order of allocation has significant impact on performance and other
1818 * system components. Generally order 0 allocations should be preferred since
1819 * order 0 does not cause fragmentation in the page allocator. Larger objects
1820 * be problematic to put into order 0 slabs because there may be too much
1821 * unused space left. We go to a higher order if more than 1/16th of the slab
1822 * would be wasted.
1824 * In order to reach satisfactory performance we must ensure that a minimum
1825 * number of objects is in one slab. Otherwise we may generate too much
1826 * activity on the partial lists which requires taking the list_lock. This is
1827 * less a concern for large slabs though which are rarely used.
1829 * slub_max_order specifies the order where we begin to stop considering the
1830 * number of objects in a slab as critical. If we reach slub_max_order then
1831 * we try to keep the page order as low as possible. So we accept more waste
1832 * of space in favor of a small page order.
1834 * Higher order allocations also allow the placement of more objects in a
1835 * slab and thereby reduce object handling overhead. If the user has
1836 * requested a higher mininum order then we start with that one instead of
1837 * the smallest order which will fit the object.
1839 static inline int slab_order(int size, int min_objects,
1840 int max_order, int fract_leftover)
1842 int order;
1843 int rem;
1844 int min_order = slub_min_order;
1846 if ((PAGE_SIZE << min_order) / size > 65535)
1847 return get_order(size * 65535) - 1;
1849 for (order = max(min_order,
1850 fls(min_objects * size - 1) - PAGE_SHIFT);
1851 order <= max_order; order++) {
1853 unsigned long slab_size = PAGE_SIZE << order;
1855 if (slab_size < min_objects * size)
1856 continue;
1858 rem = slab_size % size;
1860 if (rem <= slab_size / fract_leftover)
1861 break;
1865 return order;
1868 static inline int calculate_order(int size)
1870 int order;
1871 int min_objects;
1872 int fraction;
1875 * Attempt to find best configuration for a slab. This
1876 * works by first attempting to generate a layout with
1877 * the best configuration and backing off gradually.
1879 * First we reduce the acceptable waste in a slab. Then
1880 * we reduce the minimum objects required in a slab.
1882 min_objects = slub_min_objects;
1883 if (!min_objects)
1884 min_objects = 4 * (fls(nr_cpu_ids) + 1);
1885 while (min_objects > 1) {
1886 fraction = 16;
1887 while (fraction >= 4) {
1888 order = slab_order(size, min_objects,
1889 slub_max_order, fraction);
1890 if (order <= slub_max_order)
1891 return order;
1892 fraction /= 2;
1894 min_objects /= 2;
1898 * We were unable to place multiple objects in a slab. Now
1899 * lets see if we can place a single object there.
1901 order = slab_order(size, 1, slub_max_order, 1);
1902 if (order <= slub_max_order)
1903 return order;
1906 * Doh this slab cannot be placed using slub_max_order.
1908 order = slab_order(size, 1, MAX_ORDER, 1);
1909 if (order <= MAX_ORDER)
1910 return order;
1911 return -ENOSYS;
1915 * Figure out what the alignment of the objects will be.
1917 static unsigned long calculate_alignment(unsigned long flags,
1918 unsigned long align, unsigned long size)
1921 * If the user wants hardware cache aligned objects then follow that
1922 * suggestion if the object is sufficiently large.
1924 * The hardware cache alignment cannot override the specified
1925 * alignment though. If that is greater then use it.
1927 if (flags & SLAB_HWCACHE_ALIGN) {
1928 unsigned long ralign = cache_line_size();
1929 while (size <= ralign / 2)
1930 ralign /= 2;
1931 align = max(align, ralign);
1934 if (align < ARCH_SLAB_MINALIGN)
1935 align = ARCH_SLAB_MINALIGN;
1937 return ALIGN(align, sizeof(void *));
1940 static void init_kmem_cache_cpu(struct kmem_cache *s,
1941 struct kmem_cache_cpu *c)
1943 c->page = NULL;
1944 c->freelist = NULL;
1945 c->node = 0;
1946 c->offset = s->offset / sizeof(void *);
1947 c->objsize = s->objsize;
1948 #ifdef CONFIG_SLUB_STATS
1949 memset(c->stat, 0, NR_SLUB_STAT_ITEMS * sizeof(unsigned));
1950 #endif
1953 static void init_kmem_cache_node(struct kmem_cache_node *n)
1955 n->nr_partial = 0;
1956 spin_lock_init(&n->list_lock);
1957 INIT_LIST_HEAD(&n->partial);
1958 #ifdef CONFIG_SLUB_DEBUG
1959 atomic_long_set(&n->nr_slabs, 0);
1960 INIT_LIST_HEAD(&n->full);
1961 #endif
1964 #ifdef CONFIG_SMP
1966 * Per cpu array for per cpu structures.
1968 * The per cpu array places all kmem_cache_cpu structures from one processor
1969 * close together meaning that it becomes possible that multiple per cpu
1970 * structures are contained in one cacheline. This may be particularly
1971 * beneficial for the kmalloc caches.
1973 * A desktop system typically has around 60-80 slabs. With 100 here we are
1974 * likely able to get per cpu structures for all caches from the array defined
1975 * here. We must be able to cover all kmalloc caches during bootstrap.
1977 * If the per cpu array is exhausted then fall back to kmalloc
1978 * of individual cachelines. No sharing is possible then.
1980 #define NR_KMEM_CACHE_CPU 100
1982 static DEFINE_PER_CPU(struct kmem_cache_cpu,
1983 kmem_cache_cpu)[NR_KMEM_CACHE_CPU];
1985 static DEFINE_PER_CPU(struct kmem_cache_cpu *, kmem_cache_cpu_free);
1986 static cpumask_t kmem_cach_cpu_free_init_once = CPU_MASK_NONE;
1988 static struct kmem_cache_cpu *alloc_kmem_cache_cpu(struct kmem_cache *s,
1989 int cpu, gfp_t flags)
1991 struct kmem_cache_cpu *c = per_cpu(kmem_cache_cpu_free, cpu);
1993 if (c)
1994 per_cpu(kmem_cache_cpu_free, cpu) =
1995 (void *)c->freelist;
1996 else {
1997 /* Table overflow: So allocate ourselves */
1998 c = kmalloc_node(
1999 ALIGN(sizeof(struct kmem_cache_cpu), cache_line_size()),
2000 flags, cpu_to_node(cpu));
2001 if (!c)
2002 return NULL;
2005 init_kmem_cache_cpu(s, c);
2006 return c;
2009 static void free_kmem_cache_cpu(struct kmem_cache_cpu *c, int cpu)
2011 if (c < per_cpu(kmem_cache_cpu, cpu) ||
2012 c > per_cpu(kmem_cache_cpu, cpu) + NR_KMEM_CACHE_CPU) {
2013 kfree(c);
2014 return;
2016 c->freelist = (void *)per_cpu(kmem_cache_cpu_free, cpu);
2017 per_cpu(kmem_cache_cpu_free, cpu) = c;
2020 static void free_kmem_cache_cpus(struct kmem_cache *s)
2022 int cpu;
2024 for_each_online_cpu(cpu) {
2025 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2027 if (c) {
2028 s->cpu_slab[cpu] = NULL;
2029 free_kmem_cache_cpu(c, cpu);
2034 static int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2036 int cpu;
2038 for_each_online_cpu(cpu) {
2039 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
2041 if (c)
2042 continue;
2044 c = alloc_kmem_cache_cpu(s, cpu, flags);
2045 if (!c) {
2046 free_kmem_cache_cpus(s);
2047 return 0;
2049 s->cpu_slab[cpu] = c;
2051 return 1;
2055 * Initialize the per cpu array.
2057 static void init_alloc_cpu_cpu(int cpu)
2059 int i;
2061 if (cpu_isset(cpu, kmem_cach_cpu_free_init_once))
2062 return;
2064 for (i = NR_KMEM_CACHE_CPU - 1; i >= 0; i--)
2065 free_kmem_cache_cpu(&per_cpu(kmem_cache_cpu, cpu)[i], cpu);
2067 cpu_set(cpu, kmem_cach_cpu_free_init_once);
2070 static void __init init_alloc_cpu(void)
2072 int cpu;
2074 for_each_online_cpu(cpu)
2075 init_alloc_cpu_cpu(cpu);
2078 #else
2079 static inline void free_kmem_cache_cpus(struct kmem_cache *s) {}
2080 static inline void init_alloc_cpu(void) {}
2082 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s, gfp_t flags)
2084 init_kmem_cache_cpu(s, &s->cpu_slab);
2085 return 1;
2087 #endif
2089 #ifdef CONFIG_NUMA
2091 * No kmalloc_node yet so do it by hand. We know that this is the first
2092 * slab on the node for this slabcache. There are no concurrent accesses
2093 * possible.
2095 * Note that this function only works on the kmalloc_node_cache
2096 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2097 * memory on a fresh node that has no slab structures yet.
2099 static struct kmem_cache_node *early_kmem_cache_node_alloc(gfp_t gfpflags,
2100 int node)
2102 struct page *page;
2103 struct kmem_cache_node *n;
2104 unsigned long flags;
2106 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
2108 page = new_slab(kmalloc_caches, gfpflags, node);
2110 BUG_ON(!page);
2111 if (page_to_nid(page) != node) {
2112 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2113 "node %d\n", node);
2114 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2115 "in order to be able to continue\n");
2118 n = page->freelist;
2119 BUG_ON(!n);
2120 page->freelist = get_freepointer(kmalloc_caches, n);
2121 page->inuse++;
2122 kmalloc_caches->node[node] = n;
2123 #ifdef CONFIG_SLUB_DEBUG
2124 init_object(kmalloc_caches, n, 1);
2125 init_tracking(kmalloc_caches, n);
2126 #endif
2127 init_kmem_cache_node(n);
2128 inc_slabs_node(kmalloc_caches, node, page->objects);
2131 * lockdep requires consistent irq usage for each lock
2132 * so even though there cannot be a race this early in
2133 * the boot sequence, we still disable irqs.
2135 local_irq_save(flags);
2136 add_partial(n, page, 0);
2137 local_irq_restore(flags);
2138 return n;
2141 static void free_kmem_cache_nodes(struct kmem_cache *s)
2143 int node;
2145 for_each_node_state(node, N_NORMAL_MEMORY) {
2146 struct kmem_cache_node *n = s->node[node];
2147 if (n && n != &s->local_node)
2148 kmem_cache_free(kmalloc_caches, n);
2149 s->node[node] = NULL;
2153 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2155 int node;
2156 int local_node;
2158 if (slab_state >= UP)
2159 local_node = page_to_nid(virt_to_page(s));
2160 else
2161 local_node = 0;
2163 for_each_node_state(node, N_NORMAL_MEMORY) {
2164 struct kmem_cache_node *n;
2166 if (local_node == node)
2167 n = &s->local_node;
2168 else {
2169 if (slab_state == DOWN) {
2170 n = early_kmem_cache_node_alloc(gfpflags,
2171 node);
2172 continue;
2174 n = kmem_cache_alloc_node(kmalloc_caches,
2175 gfpflags, node);
2177 if (!n) {
2178 free_kmem_cache_nodes(s);
2179 return 0;
2183 s->node[node] = n;
2184 init_kmem_cache_node(n);
2186 return 1;
2188 #else
2189 static void free_kmem_cache_nodes(struct kmem_cache *s)
2193 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
2195 init_kmem_cache_node(&s->local_node);
2196 return 1;
2198 #endif
2201 * calculate_sizes() determines the order and the distribution of data within
2202 * a slab object.
2204 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2206 unsigned long flags = s->flags;
2207 unsigned long size = s->objsize;
2208 unsigned long align = s->align;
2209 int order;
2212 * Round up object size to the next word boundary. We can only
2213 * place the free pointer at word boundaries and this determines
2214 * the possible location of the free pointer.
2216 size = ALIGN(size, sizeof(void *));
2218 #ifdef CONFIG_SLUB_DEBUG
2220 * Determine if we can poison the object itself. If the user of
2221 * the slab may touch the object after free or before allocation
2222 * then we should never poison the object itself.
2224 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2225 !s->ctor)
2226 s->flags |= __OBJECT_POISON;
2227 else
2228 s->flags &= ~__OBJECT_POISON;
2232 * If we are Redzoning then check if there is some space between the
2233 * end of the object and the free pointer. If not then add an
2234 * additional word to have some bytes to store Redzone information.
2236 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2237 size += sizeof(void *);
2238 #endif
2241 * With that we have determined the number of bytes in actual use
2242 * by the object. This is the potential offset to the free pointer.
2244 s->inuse = size;
2246 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2247 s->ctor)) {
2249 * Relocate free pointer after the object if it is not
2250 * permitted to overwrite the first word of the object on
2251 * kmem_cache_free.
2253 * This is the case if we do RCU, have a constructor or
2254 * destructor or are poisoning the objects.
2256 s->offset = size;
2257 size += sizeof(void *);
2260 #ifdef CONFIG_SLUB_DEBUG
2261 if (flags & SLAB_STORE_USER)
2263 * Need to store information about allocs and frees after
2264 * the object.
2266 size += 2 * sizeof(struct track);
2268 if (flags & SLAB_RED_ZONE)
2270 * Add some empty padding so that we can catch
2271 * overwrites from earlier objects rather than let
2272 * tracking information or the free pointer be
2273 * corrupted if an user writes before the start
2274 * of the object.
2276 size += sizeof(void *);
2277 #endif
2280 * Determine the alignment based on various parameters that the
2281 * user specified and the dynamic determination of cache line size
2282 * on bootup.
2284 align = calculate_alignment(flags, align, s->objsize);
2287 * SLUB stores one object immediately after another beginning from
2288 * offset 0. In order to align the objects we have to simply size
2289 * each object to conform to the alignment.
2291 size = ALIGN(size, align);
2292 s->size = size;
2293 if (forced_order >= 0)
2294 order = forced_order;
2295 else
2296 order = calculate_order(size);
2298 if (order < 0)
2299 return 0;
2301 s->allocflags = 0;
2302 if (order)
2303 s->allocflags |= __GFP_COMP;
2305 if (s->flags & SLAB_CACHE_DMA)
2306 s->allocflags |= SLUB_DMA;
2308 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2309 s->allocflags |= __GFP_RECLAIMABLE;
2312 * Determine the number of objects per slab
2314 s->oo = oo_make(order, size);
2315 s->min = oo_make(get_order(size), size);
2316 if (oo_objects(s->oo) > oo_objects(s->max))
2317 s->max = s->oo;
2319 return !!oo_objects(s->oo);
2323 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2324 const char *name, size_t size,
2325 size_t align, unsigned long flags,
2326 void (*ctor)(struct kmem_cache *, void *))
2328 memset(s, 0, kmem_size);
2329 s->name = name;
2330 s->ctor = ctor;
2331 s->objsize = size;
2332 s->align = align;
2333 s->flags = kmem_cache_flags(size, flags, name, ctor);
2335 if (!calculate_sizes(s, -1))
2336 goto error;
2338 s->refcount = 1;
2339 #ifdef CONFIG_NUMA
2340 s->remote_node_defrag_ratio = 100;
2341 #endif
2342 if (!init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2343 goto error;
2345 if (alloc_kmem_cache_cpus(s, gfpflags & ~SLUB_DMA))
2346 return 1;
2347 free_kmem_cache_nodes(s);
2348 error:
2349 if (flags & SLAB_PANIC)
2350 panic("Cannot create slab %s size=%lu realsize=%u "
2351 "order=%u offset=%u flags=%lx\n",
2352 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2353 s->offset, flags);
2354 return 0;
2358 * Check if a given pointer is valid
2360 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2362 struct page *page;
2364 page = get_object_page(object);
2366 if (!page || s != page->slab)
2367 /* No slab or wrong slab */
2368 return 0;
2370 if (!check_valid_pointer(s, page, object))
2371 return 0;
2374 * We could also check if the object is on the slabs freelist.
2375 * But this would be too expensive and it seems that the main
2376 * purpose of kmem_ptr_valid() is to check if the object belongs
2377 * to a certain slab.
2379 return 1;
2381 EXPORT_SYMBOL(kmem_ptr_validate);
2384 * Determine the size of a slab object
2386 unsigned int kmem_cache_size(struct kmem_cache *s)
2388 return s->objsize;
2390 EXPORT_SYMBOL(kmem_cache_size);
2392 const char *kmem_cache_name(struct kmem_cache *s)
2394 return s->name;
2396 EXPORT_SYMBOL(kmem_cache_name);
2398 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2399 const char *text)
2401 #ifdef CONFIG_SLUB_DEBUG
2402 void *addr = page_address(page);
2403 void *p;
2404 DECLARE_BITMAP(map, page->objects);
2406 bitmap_zero(map, page->objects);
2407 slab_err(s, page, "%s", text);
2408 slab_lock(page);
2409 for_each_free_object(p, s, page->freelist)
2410 set_bit(slab_index(p, s, addr), map);
2412 for_each_object(p, s, addr, page->objects) {
2414 if (!test_bit(slab_index(p, s, addr), map)) {
2415 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2416 p, p - addr);
2417 print_tracking(s, p);
2420 slab_unlock(page);
2421 #endif
2425 * Attempt to free all partial slabs on a node.
2427 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2429 unsigned long flags;
2430 struct page *page, *h;
2432 spin_lock_irqsave(&n->list_lock, flags);
2433 list_for_each_entry_safe(page, h, &n->partial, lru) {
2434 if (!page->inuse) {
2435 list_del(&page->lru);
2436 discard_slab(s, page);
2437 n->nr_partial--;
2438 } else {
2439 list_slab_objects(s, page,
2440 "Objects remaining on kmem_cache_close()");
2443 spin_unlock_irqrestore(&n->list_lock, flags);
2447 * Release all resources used by a slab cache.
2449 static inline int kmem_cache_close(struct kmem_cache *s)
2451 int node;
2453 flush_all(s);
2455 /* Attempt to free all objects */
2456 free_kmem_cache_cpus(s);
2457 for_each_node_state(node, N_NORMAL_MEMORY) {
2458 struct kmem_cache_node *n = get_node(s, node);
2460 free_partial(s, n);
2461 if (n->nr_partial || slabs_node(s, node))
2462 return 1;
2464 free_kmem_cache_nodes(s);
2465 return 0;
2469 * Close a cache and release the kmem_cache structure
2470 * (must be used for caches created using kmem_cache_create)
2472 void kmem_cache_destroy(struct kmem_cache *s)
2474 down_write(&slub_lock);
2475 s->refcount--;
2476 if (!s->refcount) {
2477 list_del(&s->list);
2478 up_write(&slub_lock);
2479 if (kmem_cache_close(s)) {
2480 printk(KERN_ERR "SLUB %s: %s called for cache that "
2481 "still has objects.\n", s->name, __func__);
2482 dump_stack();
2484 sysfs_slab_remove(s);
2485 } else
2486 up_write(&slub_lock);
2488 EXPORT_SYMBOL(kmem_cache_destroy);
2490 /********************************************************************
2491 * Kmalloc subsystem
2492 *******************************************************************/
2494 struct kmem_cache kmalloc_caches[PAGE_SHIFT + 1] __cacheline_aligned;
2495 EXPORT_SYMBOL(kmalloc_caches);
2497 static int __init setup_slub_min_order(char *str)
2499 get_option(&str, &slub_min_order);
2501 return 1;
2504 __setup("slub_min_order=", setup_slub_min_order);
2506 static int __init setup_slub_max_order(char *str)
2508 get_option(&str, &slub_max_order);
2510 return 1;
2513 __setup("slub_max_order=", setup_slub_max_order);
2515 static int __init setup_slub_min_objects(char *str)
2517 get_option(&str, &slub_min_objects);
2519 return 1;
2522 __setup("slub_min_objects=", setup_slub_min_objects);
2524 static int __init setup_slub_nomerge(char *str)
2526 slub_nomerge = 1;
2527 return 1;
2530 __setup("slub_nomerge", setup_slub_nomerge);
2532 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2533 const char *name, int size, gfp_t gfp_flags)
2535 unsigned int flags = 0;
2537 if (gfp_flags & SLUB_DMA)
2538 flags = SLAB_CACHE_DMA;
2540 down_write(&slub_lock);
2541 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2542 flags, NULL))
2543 goto panic;
2545 list_add(&s->list, &slab_caches);
2546 up_write(&slub_lock);
2547 if (sysfs_slab_add(s))
2548 goto panic;
2549 return s;
2551 panic:
2552 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2555 #ifdef CONFIG_ZONE_DMA
2556 static struct kmem_cache *kmalloc_caches_dma[PAGE_SHIFT + 1];
2558 static void sysfs_add_func(struct work_struct *w)
2560 struct kmem_cache *s;
2562 down_write(&slub_lock);
2563 list_for_each_entry(s, &slab_caches, list) {
2564 if (s->flags & __SYSFS_ADD_DEFERRED) {
2565 s->flags &= ~__SYSFS_ADD_DEFERRED;
2566 sysfs_slab_add(s);
2569 up_write(&slub_lock);
2572 static DECLARE_WORK(sysfs_add_work, sysfs_add_func);
2574 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2576 struct kmem_cache *s;
2577 char *text;
2578 size_t realsize;
2580 s = kmalloc_caches_dma[index];
2581 if (s)
2582 return s;
2584 /* Dynamically create dma cache */
2585 if (flags & __GFP_WAIT)
2586 down_write(&slub_lock);
2587 else {
2588 if (!down_write_trylock(&slub_lock))
2589 goto out;
2592 if (kmalloc_caches_dma[index])
2593 goto unlock_out;
2595 realsize = kmalloc_caches[index].objsize;
2596 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2597 (unsigned int)realsize);
2598 s = kmalloc(kmem_size, flags & ~SLUB_DMA);
2600 if (!s || !text || !kmem_cache_open(s, flags, text,
2601 realsize, ARCH_KMALLOC_MINALIGN,
2602 SLAB_CACHE_DMA|__SYSFS_ADD_DEFERRED, NULL)) {
2603 kfree(s);
2604 kfree(text);
2605 goto unlock_out;
2608 list_add(&s->list, &slab_caches);
2609 kmalloc_caches_dma[index] = s;
2611 schedule_work(&sysfs_add_work);
2613 unlock_out:
2614 up_write(&slub_lock);
2615 out:
2616 return kmalloc_caches_dma[index];
2618 #endif
2621 * Conversion table for small slabs sizes / 8 to the index in the
2622 * kmalloc array. This is necessary for slabs < 192 since we have non power
2623 * of two cache sizes there. The size of larger slabs can be determined using
2624 * fls.
2626 static s8 size_index[24] = {
2627 3, /* 8 */
2628 4, /* 16 */
2629 5, /* 24 */
2630 5, /* 32 */
2631 6, /* 40 */
2632 6, /* 48 */
2633 6, /* 56 */
2634 6, /* 64 */
2635 1, /* 72 */
2636 1, /* 80 */
2637 1, /* 88 */
2638 1, /* 96 */
2639 7, /* 104 */
2640 7, /* 112 */
2641 7, /* 120 */
2642 7, /* 128 */
2643 2, /* 136 */
2644 2, /* 144 */
2645 2, /* 152 */
2646 2, /* 160 */
2647 2, /* 168 */
2648 2, /* 176 */
2649 2, /* 184 */
2650 2 /* 192 */
2653 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2655 int index;
2657 if (size <= 192) {
2658 if (!size)
2659 return ZERO_SIZE_PTR;
2661 index = size_index[(size - 1) / 8];
2662 } else
2663 index = fls(size - 1);
2665 #ifdef CONFIG_ZONE_DMA
2666 if (unlikely((flags & SLUB_DMA)))
2667 return dma_kmalloc_cache(index, flags);
2669 #endif
2670 return &kmalloc_caches[index];
2673 void *__kmalloc(size_t size, gfp_t flags)
2675 struct kmem_cache *s;
2677 if (unlikely(size > PAGE_SIZE))
2678 return kmalloc_large(size, flags);
2680 s = get_slab(size, flags);
2682 if (unlikely(ZERO_OR_NULL_PTR(s)))
2683 return s;
2685 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2687 EXPORT_SYMBOL(__kmalloc);
2689 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2691 struct page *page = alloc_pages_node(node, flags | __GFP_COMP,
2692 get_order(size));
2694 if (page)
2695 return page_address(page);
2696 else
2697 return NULL;
2700 #ifdef CONFIG_NUMA
2701 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2703 struct kmem_cache *s;
2705 if (unlikely(size > PAGE_SIZE))
2706 return kmalloc_large_node(size, flags, node);
2708 s = get_slab(size, flags);
2710 if (unlikely(ZERO_OR_NULL_PTR(s)))
2711 return s;
2713 return slab_alloc(s, flags, node, __builtin_return_address(0));
2715 EXPORT_SYMBOL(__kmalloc_node);
2716 #endif
2718 size_t ksize(const void *object)
2720 struct page *page;
2721 struct kmem_cache *s;
2723 if (unlikely(object == ZERO_SIZE_PTR))
2724 return 0;
2726 page = virt_to_head_page(object);
2728 if (unlikely(!PageSlab(page)))
2729 return PAGE_SIZE << compound_order(page);
2731 s = page->slab;
2733 #ifdef CONFIG_SLUB_DEBUG
2735 * Debugging requires use of the padding between object
2736 * and whatever may come after it.
2738 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2739 return s->objsize;
2741 #endif
2743 * If we have the need to store the freelist pointer
2744 * back there or track user information then we can
2745 * only use the space before that information.
2747 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2748 return s->inuse;
2750 * Else we can use all the padding etc for the allocation
2752 return s->size;
2754 EXPORT_SYMBOL(ksize);
2756 void kfree(const void *x)
2758 struct page *page;
2759 void *object = (void *)x;
2761 if (unlikely(ZERO_OR_NULL_PTR(x)))
2762 return;
2764 page = virt_to_head_page(x);
2765 if (unlikely(!PageSlab(page))) {
2766 put_page(page);
2767 return;
2769 slab_free(page->slab, page, object, __builtin_return_address(0));
2771 EXPORT_SYMBOL(kfree);
2774 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2775 * the remaining slabs by the number of items in use. The slabs with the
2776 * most items in use come first. New allocations will then fill those up
2777 * and thus they can be removed from the partial lists.
2779 * The slabs with the least items are placed last. This results in them
2780 * being allocated from last increasing the chance that the last objects
2781 * are freed in them.
2783 int kmem_cache_shrink(struct kmem_cache *s)
2785 int node;
2786 int i;
2787 struct kmem_cache_node *n;
2788 struct page *page;
2789 struct page *t;
2790 int objects = oo_objects(s->max);
2791 struct list_head *slabs_by_inuse =
2792 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
2793 unsigned long flags;
2795 if (!slabs_by_inuse)
2796 return -ENOMEM;
2798 flush_all(s);
2799 for_each_node_state(node, N_NORMAL_MEMORY) {
2800 n = get_node(s, node);
2802 if (!n->nr_partial)
2803 continue;
2805 for (i = 0; i < objects; i++)
2806 INIT_LIST_HEAD(slabs_by_inuse + i);
2808 spin_lock_irqsave(&n->list_lock, flags);
2811 * Build lists indexed by the items in use in each slab.
2813 * Note that concurrent frees may occur while we hold the
2814 * list_lock. page->inuse here is the upper limit.
2816 list_for_each_entry_safe(page, t, &n->partial, lru) {
2817 if (!page->inuse && slab_trylock(page)) {
2819 * Must hold slab lock here because slab_free
2820 * may have freed the last object and be
2821 * waiting to release the slab.
2823 list_del(&page->lru);
2824 n->nr_partial--;
2825 slab_unlock(page);
2826 discard_slab(s, page);
2827 } else {
2828 list_move(&page->lru,
2829 slabs_by_inuse + page->inuse);
2834 * Rebuild the partial list with the slabs filled up most
2835 * first and the least used slabs at the end.
2837 for (i = objects - 1; i >= 0; i--)
2838 list_splice(slabs_by_inuse + i, n->partial.prev);
2840 spin_unlock_irqrestore(&n->list_lock, flags);
2843 kfree(slabs_by_inuse);
2844 return 0;
2846 EXPORT_SYMBOL(kmem_cache_shrink);
2848 #if defined(CONFIG_NUMA) && defined(CONFIG_MEMORY_HOTPLUG)
2849 static int slab_mem_going_offline_callback(void *arg)
2851 struct kmem_cache *s;
2853 down_read(&slub_lock);
2854 list_for_each_entry(s, &slab_caches, list)
2855 kmem_cache_shrink(s);
2856 up_read(&slub_lock);
2858 return 0;
2861 static void slab_mem_offline_callback(void *arg)
2863 struct kmem_cache_node *n;
2864 struct kmem_cache *s;
2865 struct memory_notify *marg = arg;
2866 int offline_node;
2868 offline_node = marg->status_change_nid;
2871 * If the node still has available memory. we need kmem_cache_node
2872 * for it yet.
2874 if (offline_node < 0)
2875 return;
2877 down_read(&slub_lock);
2878 list_for_each_entry(s, &slab_caches, list) {
2879 n = get_node(s, offline_node);
2880 if (n) {
2882 * if n->nr_slabs > 0, slabs still exist on the node
2883 * that is going down. We were unable to free them,
2884 * and offline_pages() function shoudn't call this
2885 * callback. So, we must fail.
2887 BUG_ON(slabs_node(s, offline_node));
2889 s->node[offline_node] = NULL;
2890 kmem_cache_free(kmalloc_caches, n);
2893 up_read(&slub_lock);
2896 static int slab_mem_going_online_callback(void *arg)
2898 struct kmem_cache_node *n;
2899 struct kmem_cache *s;
2900 struct memory_notify *marg = arg;
2901 int nid = marg->status_change_nid;
2902 int ret = 0;
2905 * If the node's memory is already available, then kmem_cache_node is
2906 * already created. Nothing to do.
2908 if (nid < 0)
2909 return 0;
2912 * We are bringing a node online. No memory is availabe yet. We must
2913 * allocate a kmem_cache_node structure in order to bring the node
2914 * online.
2916 down_read(&slub_lock);
2917 list_for_each_entry(s, &slab_caches, list) {
2919 * XXX: kmem_cache_alloc_node will fallback to other nodes
2920 * since memory is not yet available from the node that
2921 * is brought up.
2923 n = kmem_cache_alloc(kmalloc_caches, GFP_KERNEL);
2924 if (!n) {
2925 ret = -ENOMEM;
2926 goto out;
2928 init_kmem_cache_node(n);
2929 s->node[nid] = n;
2931 out:
2932 up_read(&slub_lock);
2933 return ret;
2936 static int slab_memory_callback(struct notifier_block *self,
2937 unsigned long action, void *arg)
2939 int ret = 0;
2941 switch (action) {
2942 case MEM_GOING_ONLINE:
2943 ret = slab_mem_going_online_callback(arg);
2944 break;
2945 case MEM_GOING_OFFLINE:
2946 ret = slab_mem_going_offline_callback(arg);
2947 break;
2948 case MEM_OFFLINE:
2949 case MEM_CANCEL_ONLINE:
2950 slab_mem_offline_callback(arg);
2951 break;
2952 case MEM_ONLINE:
2953 case MEM_CANCEL_OFFLINE:
2954 break;
2957 ret = notifier_from_errno(ret);
2958 return ret;
2961 #endif /* CONFIG_MEMORY_HOTPLUG */
2963 /********************************************************************
2964 * Basic setup of slabs
2965 *******************************************************************/
2967 void __init kmem_cache_init(void)
2969 int i;
2970 int caches = 0;
2972 init_alloc_cpu();
2974 #ifdef CONFIG_NUMA
2976 * Must first have the slab cache available for the allocations of the
2977 * struct kmem_cache_node's. There is special bootstrap code in
2978 * kmem_cache_open for slab_state == DOWN.
2980 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2981 sizeof(struct kmem_cache_node), GFP_KERNEL);
2982 kmalloc_caches[0].refcount = -1;
2983 caches++;
2985 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
2986 #endif
2988 /* Able to allocate the per node structures */
2989 slab_state = PARTIAL;
2991 /* Caches that are not of the two-to-the-power-of size */
2992 if (KMALLOC_MIN_SIZE <= 64) {
2993 create_kmalloc_cache(&kmalloc_caches[1],
2994 "kmalloc-96", 96, GFP_KERNEL);
2995 caches++;
2997 if (KMALLOC_MIN_SIZE <= 128) {
2998 create_kmalloc_cache(&kmalloc_caches[2],
2999 "kmalloc-192", 192, GFP_KERNEL);
3000 caches++;
3003 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++) {
3004 create_kmalloc_cache(&kmalloc_caches[i],
3005 "kmalloc", 1 << i, GFP_KERNEL);
3006 caches++;
3011 * Patch up the size_index table if we have strange large alignment
3012 * requirements for the kmalloc array. This is only the case for
3013 * MIPS it seems. The standard arches will not generate any code here.
3015 * Largest permitted alignment is 256 bytes due to the way we
3016 * handle the index determination for the smaller caches.
3018 * Make sure that nothing crazy happens if someone starts tinkering
3019 * around with ARCH_KMALLOC_MINALIGN
3021 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3022 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3024 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
3025 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
3027 slab_state = UP;
3029 /* Provide the correct kmalloc names now that the caches are up */
3030 for (i = KMALLOC_SHIFT_LOW; i <= PAGE_SHIFT; i++)
3031 kmalloc_caches[i]. name =
3032 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
3034 #ifdef CONFIG_SMP
3035 register_cpu_notifier(&slab_notifier);
3036 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
3037 nr_cpu_ids * sizeof(struct kmem_cache_cpu *);
3038 #else
3039 kmem_size = sizeof(struct kmem_cache);
3040 #endif
3042 printk(KERN_INFO
3043 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3044 " CPUs=%d, Nodes=%d\n",
3045 caches, cache_line_size(),
3046 slub_min_order, slub_max_order, slub_min_objects,
3047 nr_cpu_ids, nr_node_ids);
3051 * Find a mergeable slab cache
3053 static int slab_unmergeable(struct kmem_cache *s)
3055 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3056 return 1;
3058 if (s->ctor)
3059 return 1;
3062 * We may have set a slab to be unmergeable during bootstrap.
3064 if (s->refcount < 0)
3065 return 1;
3067 return 0;
3070 static struct kmem_cache *find_mergeable(size_t size,
3071 size_t align, unsigned long flags, const char *name,
3072 void (*ctor)(struct kmem_cache *, void *))
3074 struct kmem_cache *s;
3076 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3077 return NULL;
3079 if (ctor)
3080 return NULL;
3082 size = ALIGN(size, sizeof(void *));
3083 align = calculate_alignment(flags, align, size);
3084 size = ALIGN(size, align);
3085 flags = kmem_cache_flags(size, flags, name, NULL);
3087 list_for_each_entry(s, &slab_caches, list) {
3088 if (slab_unmergeable(s))
3089 continue;
3091 if (size > s->size)
3092 continue;
3094 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3095 continue;
3097 * Check if alignment is compatible.
3098 * Courtesy of Adrian Drzewiecki
3100 if ((s->size & ~(align - 1)) != s->size)
3101 continue;
3103 if (s->size - size >= sizeof(void *))
3104 continue;
3106 return s;
3108 return NULL;
3111 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3112 size_t align, unsigned long flags,
3113 void (*ctor)(struct kmem_cache *, void *))
3115 struct kmem_cache *s;
3117 down_write(&slub_lock);
3118 s = find_mergeable(size, align, flags, name, ctor);
3119 if (s) {
3120 int cpu;
3122 s->refcount++;
3124 * Adjust the object sizes so that we clear
3125 * the complete object on kzalloc.
3127 s->objsize = max(s->objsize, (int)size);
3130 * And then we need to update the object size in the
3131 * per cpu structures
3133 for_each_online_cpu(cpu)
3134 get_cpu_slab(s, cpu)->objsize = s->objsize;
3136 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3137 up_write(&slub_lock);
3139 if (sysfs_slab_alias(s, name))
3140 goto err;
3141 return s;
3144 s = kmalloc(kmem_size, GFP_KERNEL);
3145 if (s) {
3146 if (kmem_cache_open(s, GFP_KERNEL, name,
3147 size, align, flags, ctor)) {
3148 list_add(&s->list, &slab_caches);
3149 up_write(&slub_lock);
3150 if (sysfs_slab_add(s))
3151 goto err;
3152 return s;
3154 kfree(s);
3156 up_write(&slub_lock);
3158 err:
3159 if (flags & SLAB_PANIC)
3160 panic("Cannot create slabcache %s\n", name);
3161 else
3162 s = NULL;
3163 return s;
3165 EXPORT_SYMBOL(kmem_cache_create);
3167 #ifdef CONFIG_SMP
3169 * Use the cpu notifier to insure that the cpu slabs are flushed when
3170 * necessary.
3172 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3173 unsigned long action, void *hcpu)
3175 long cpu = (long)hcpu;
3176 struct kmem_cache *s;
3177 unsigned long flags;
3179 switch (action) {
3180 case CPU_UP_PREPARE:
3181 case CPU_UP_PREPARE_FROZEN:
3182 init_alloc_cpu_cpu(cpu);
3183 down_read(&slub_lock);
3184 list_for_each_entry(s, &slab_caches, list)
3185 s->cpu_slab[cpu] = alloc_kmem_cache_cpu(s, cpu,
3186 GFP_KERNEL);
3187 up_read(&slub_lock);
3188 break;
3190 case CPU_UP_CANCELED:
3191 case CPU_UP_CANCELED_FROZEN:
3192 case CPU_DEAD:
3193 case CPU_DEAD_FROZEN:
3194 down_read(&slub_lock);
3195 list_for_each_entry(s, &slab_caches, list) {
3196 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3198 local_irq_save(flags);
3199 __flush_cpu_slab(s, cpu);
3200 local_irq_restore(flags);
3201 free_kmem_cache_cpu(c, cpu);
3202 s->cpu_slab[cpu] = NULL;
3204 up_read(&slub_lock);
3205 break;
3206 default:
3207 break;
3209 return NOTIFY_OK;
3212 static struct notifier_block __cpuinitdata slab_notifier = {
3213 .notifier_call = slab_cpuup_callback
3216 #endif
3218 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
3220 struct kmem_cache *s;
3222 if (unlikely(size > PAGE_SIZE))
3223 return kmalloc_large(size, gfpflags);
3225 s = get_slab(size, gfpflags);
3227 if (unlikely(ZERO_OR_NULL_PTR(s)))
3228 return s;
3230 return slab_alloc(s, gfpflags, -1, caller);
3233 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3234 int node, void *caller)
3236 struct kmem_cache *s;
3238 if (unlikely(size > PAGE_SIZE))
3239 return kmalloc_large_node(size, gfpflags, node);
3241 s = get_slab(size, gfpflags);
3243 if (unlikely(ZERO_OR_NULL_PTR(s)))
3244 return s;
3246 return slab_alloc(s, gfpflags, node, caller);
3249 #if (defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)) || defined(CONFIG_SLABINFO)
3250 static unsigned long count_partial(struct kmem_cache_node *n,
3251 int (*get_count)(struct page *))
3253 unsigned long flags;
3254 unsigned long x = 0;
3255 struct page *page;
3257 spin_lock_irqsave(&n->list_lock, flags);
3258 list_for_each_entry(page, &n->partial, lru)
3259 x += get_count(page);
3260 spin_unlock_irqrestore(&n->list_lock, flags);
3261 return x;
3264 static int count_inuse(struct page *page)
3266 return page->inuse;
3269 static int count_total(struct page *page)
3271 return page->objects;
3274 static int count_free(struct page *page)
3276 return page->objects - page->inuse;
3278 #endif
3280 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
3281 static int validate_slab(struct kmem_cache *s, struct page *page,
3282 unsigned long *map)
3284 void *p;
3285 void *addr = page_address(page);
3287 if (!check_slab(s, page) ||
3288 !on_freelist(s, page, NULL))
3289 return 0;
3291 /* Now we know that a valid freelist exists */
3292 bitmap_zero(map, page->objects);
3294 for_each_free_object(p, s, page->freelist) {
3295 set_bit(slab_index(p, s, addr), map);
3296 if (!check_object(s, page, p, 0))
3297 return 0;
3300 for_each_object(p, s, addr, page->objects)
3301 if (!test_bit(slab_index(p, s, addr), map))
3302 if (!check_object(s, page, p, 1))
3303 return 0;
3304 return 1;
3307 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3308 unsigned long *map)
3310 if (slab_trylock(page)) {
3311 validate_slab(s, page, map);
3312 slab_unlock(page);
3313 } else
3314 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3315 s->name, page);
3317 if (s->flags & DEBUG_DEFAULT_FLAGS) {
3318 if (!SlabDebug(page))
3319 printk(KERN_ERR "SLUB %s: SlabDebug not set "
3320 "on slab 0x%p\n", s->name, page);
3321 } else {
3322 if (SlabDebug(page))
3323 printk(KERN_ERR "SLUB %s: SlabDebug set on "
3324 "slab 0x%p\n", s->name, page);
3328 static int validate_slab_node(struct kmem_cache *s,
3329 struct kmem_cache_node *n, unsigned long *map)
3331 unsigned long count = 0;
3332 struct page *page;
3333 unsigned long flags;
3335 spin_lock_irqsave(&n->list_lock, flags);
3337 list_for_each_entry(page, &n->partial, lru) {
3338 validate_slab_slab(s, page, map);
3339 count++;
3341 if (count != n->nr_partial)
3342 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3343 "counter=%ld\n", s->name, count, n->nr_partial);
3345 if (!(s->flags & SLAB_STORE_USER))
3346 goto out;
3348 list_for_each_entry(page, &n->full, lru) {
3349 validate_slab_slab(s, page, map);
3350 count++;
3352 if (count != atomic_long_read(&n->nr_slabs))
3353 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3354 "counter=%ld\n", s->name, count,
3355 atomic_long_read(&n->nr_slabs));
3357 out:
3358 spin_unlock_irqrestore(&n->list_lock, flags);
3359 return count;
3362 static long validate_slab_cache(struct kmem_cache *s)
3364 int node;
3365 unsigned long count = 0;
3366 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3367 sizeof(unsigned long), GFP_KERNEL);
3369 if (!map)
3370 return -ENOMEM;
3372 flush_all(s);
3373 for_each_node_state(node, N_NORMAL_MEMORY) {
3374 struct kmem_cache_node *n = get_node(s, node);
3376 count += validate_slab_node(s, n, map);
3378 kfree(map);
3379 return count;
3382 #ifdef SLUB_RESILIENCY_TEST
3383 static void resiliency_test(void)
3385 u8 *p;
3387 printk(KERN_ERR "SLUB resiliency testing\n");
3388 printk(KERN_ERR "-----------------------\n");
3389 printk(KERN_ERR "A. Corruption after allocation\n");
3391 p = kzalloc(16, GFP_KERNEL);
3392 p[16] = 0x12;
3393 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
3394 " 0x12->0x%p\n\n", p + 16);
3396 validate_slab_cache(kmalloc_caches + 4);
3398 /* Hmmm... The next two are dangerous */
3399 p = kzalloc(32, GFP_KERNEL);
3400 p[32 + sizeof(void *)] = 0x34;
3401 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
3402 " 0x34 -> -0x%p\n", p);
3403 printk(KERN_ERR
3404 "If allocated object is overwritten then not detectable\n\n");
3406 validate_slab_cache(kmalloc_caches + 5);
3407 p = kzalloc(64, GFP_KERNEL);
3408 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
3409 *p = 0x56;
3410 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
3412 printk(KERN_ERR
3413 "If allocated object is overwritten then not detectable\n\n");
3414 validate_slab_cache(kmalloc_caches + 6);
3416 printk(KERN_ERR "\nB. Corruption after free\n");
3417 p = kzalloc(128, GFP_KERNEL);
3418 kfree(p);
3419 *p = 0x78;
3420 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
3421 validate_slab_cache(kmalloc_caches + 7);
3423 p = kzalloc(256, GFP_KERNEL);
3424 kfree(p);
3425 p[50] = 0x9a;
3426 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
3428 validate_slab_cache(kmalloc_caches + 8);
3430 p = kzalloc(512, GFP_KERNEL);
3431 kfree(p);
3432 p[512] = 0xab;
3433 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
3434 validate_slab_cache(kmalloc_caches + 9);
3436 #else
3437 static void resiliency_test(void) {};
3438 #endif
3441 * Generate lists of code addresses where slabcache objects are allocated
3442 * and freed.
3445 struct location {
3446 unsigned long count;
3447 void *addr;
3448 long long sum_time;
3449 long min_time;
3450 long max_time;
3451 long min_pid;
3452 long max_pid;
3453 cpumask_t cpus;
3454 nodemask_t nodes;
3457 struct loc_track {
3458 unsigned long max;
3459 unsigned long count;
3460 struct location *loc;
3463 static void free_loc_track(struct loc_track *t)
3465 if (t->max)
3466 free_pages((unsigned long)t->loc,
3467 get_order(sizeof(struct location) * t->max));
3470 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3472 struct location *l;
3473 int order;
3475 order = get_order(sizeof(struct location) * max);
3477 l = (void *)__get_free_pages(flags, order);
3478 if (!l)
3479 return 0;
3481 if (t->count) {
3482 memcpy(l, t->loc, sizeof(struct location) * t->count);
3483 free_loc_track(t);
3485 t->max = max;
3486 t->loc = l;
3487 return 1;
3490 static int add_location(struct loc_track *t, struct kmem_cache *s,
3491 const struct track *track)
3493 long start, end, pos;
3494 struct location *l;
3495 void *caddr;
3496 unsigned long age = jiffies - track->when;
3498 start = -1;
3499 end = t->count;
3501 for ( ; ; ) {
3502 pos = start + (end - start + 1) / 2;
3505 * There is nothing at "end". If we end up there
3506 * we need to add something to before end.
3508 if (pos == end)
3509 break;
3511 caddr = t->loc[pos].addr;
3512 if (track->addr == caddr) {
3514 l = &t->loc[pos];
3515 l->count++;
3516 if (track->when) {
3517 l->sum_time += age;
3518 if (age < l->min_time)
3519 l->min_time = age;
3520 if (age > l->max_time)
3521 l->max_time = age;
3523 if (track->pid < l->min_pid)
3524 l->min_pid = track->pid;
3525 if (track->pid > l->max_pid)
3526 l->max_pid = track->pid;
3528 cpu_set(track->cpu, l->cpus);
3530 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3531 return 1;
3534 if (track->addr < caddr)
3535 end = pos;
3536 else
3537 start = pos;
3541 * Not found. Insert new tracking element.
3543 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3544 return 0;
3546 l = t->loc + pos;
3547 if (pos < t->count)
3548 memmove(l + 1, l,
3549 (t->count - pos) * sizeof(struct location));
3550 t->count++;
3551 l->count = 1;
3552 l->addr = track->addr;
3553 l->sum_time = age;
3554 l->min_time = age;
3555 l->max_time = age;
3556 l->min_pid = track->pid;
3557 l->max_pid = track->pid;
3558 cpus_clear(l->cpus);
3559 cpu_set(track->cpu, l->cpus);
3560 nodes_clear(l->nodes);
3561 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3562 return 1;
3565 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3566 struct page *page, enum track_item alloc)
3568 void *addr = page_address(page);
3569 DECLARE_BITMAP(map, page->objects);
3570 void *p;
3572 bitmap_zero(map, page->objects);
3573 for_each_free_object(p, s, page->freelist)
3574 set_bit(slab_index(p, s, addr), map);
3576 for_each_object(p, s, addr, page->objects)
3577 if (!test_bit(slab_index(p, s, addr), map))
3578 add_location(t, s, get_track(s, p, alloc));
3581 static int list_locations(struct kmem_cache *s, char *buf,
3582 enum track_item alloc)
3584 int len = 0;
3585 unsigned long i;
3586 struct loc_track t = { 0, 0, NULL };
3587 int node;
3589 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3590 GFP_TEMPORARY))
3591 return sprintf(buf, "Out of memory\n");
3593 /* Push back cpu slabs */
3594 flush_all(s);
3596 for_each_node_state(node, N_NORMAL_MEMORY) {
3597 struct kmem_cache_node *n = get_node(s, node);
3598 unsigned long flags;
3599 struct page *page;
3601 if (!atomic_long_read(&n->nr_slabs))
3602 continue;
3604 spin_lock_irqsave(&n->list_lock, flags);
3605 list_for_each_entry(page, &n->partial, lru)
3606 process_slab(&t, s, page, alloc);
3607 list_for_each_entry(page, &n->full, lru)
3608 process_slab(&t, s, page, alloc);
3609 spin_unlock_irqrestore(&n->list_lock, flags);
3612 for (i = 0; i < t.count; i++) {
3613 struct location *l = &t.loc[i];
3615 if (len > PAGE_SIZE - 100)
3616 break;
3617 len += sprintf(buf + len, "%7ld ", l->count);
3619 if (l->addr)
3620 len += sprint_symbol(buf + len, (unsigned long)l->addr);
3621 else
3622 len += sprintf(buf + len, "<not-available>");
3624 if (l->sum_time != l->min_time) {
3625 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3626 l->min_time,
3627 (long)div_u64(l->sum_time, l->count),
3628 l->max_time);
3629 } else
3630 len += sprintf(buf + len, " age=%ld",
3631 l->min_time);
3633 if (l->min_pid != l->max_pid)
3634 len += sprintf(buf + len, " pid=%ld-%ld",
3635 l->min_pid, l->max_pid);
3636 else
3637 len += sprintf(buf + len, " pid=%ld",
3638 l->min_pid);
3640 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3641 len < PAGE_SIZE - 60) {
3642 len += sprintf(buf + len, " cpus=");
3643 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3644 l->cpus);
3647 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3648 len < PAGE_SIZE - 60) {
3649 len += sprintf(buf + len, " nodes=");
3650 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3651 l->nodes);
3654 len += sprintf(buf + len, "\n");
3657 free_loc_track(&t);
3658 if (!t.count)
3659 len += sprintf(buf, "No data\n");
3660 return len;
3663 enum slab_stat_type {
3664 SL_ALL, /* All slabs */
3665 SL_PARTIAL, /* Only partially allocated slabs */
3666 SL_CPU, /* Only slabs used for cpu caches */
3667 SL_OBJECTS, /* Determine allocated objects not slabs */
3668 SL_TOTAL /* Determine object capacity not slabs */
3671 #define SO_ALL (1 << SL_ALL)
3672 #define SO_PARTIAL (1 << SL_PARTIAL)
3673 #define SO_CPU (1 << SL_CPU)
3674 #define SO_OBJECTS (1 << SL_OBJECTS)
3675 #define SO_TOTAL (1 << SL_TOTAL)
3677 static ssize_t show_slab_objects(struct kmem_cache *s,
3678 char *buf, unsigned long flags)
3680 unsigned long total = 0;
3681 int node;
3682 int x;
3683 unsigned long *nodes;
3684 unsigned long *per_cpu;
3686 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3687 if (!nodes)
3688 return -ENOMEM;
3689 per_cpu = nodes + nr_node_ids;
3691 if (flags & SO_CPU) {
3692 int cpu;
3694 for_each_possible_cpu(cpu) {
3695 struct kmem_cache_cpu *c = get_cpu_slab(s, cpu);
3697 if (!c || c->node < 0)
3698 continue;
3700 if (c->page) {
3701 if (flags & SO_TOTAL)
3702 x = c->page->objects;
3703 else if (flags & SO_OBJECTS)
3704 x = c->page->inuse;
3705 else
3706 x = 1;
3708 total += x;
3709 nodes[c->node] += x;
3711 per_cpu[c->node]++;
3715 if (flags & SO_ALL) {
3716 for_each_node_state(node, N_NORMAL_MEMORY) {
3717 struct kmem_cache_node *n = get_node(s, node);
3719 if (flags & SO_TOTAL)
3720 x = atomic_long_read(&n->total_objects);
3721 else if (flags & SO_OBJECTS)
3722 x = atomic_long_read(&n->total_objects) -
3723 count_partial(n, count_free);
3725 else
3726 x = atomic_long_read(&n->nr_slabs);
3727 total += x;
3728 nodes[node] += x;
3731 } else if (flags & SO_PARTIAL) {
3732 for_each_node_state(node, N_NORMAL_MEMORY) {
3733 struct kmem_cache_node *n = get_node(s, node);
3735 if (flags & SO_TOTAL)
3736 x = count_partial(n, count_total);
3737 else if (flags & SO_OBJECTS)
3738 x = count_partial(n, count_inuse);
3739 else
3740 x = n->nr_partial;
3741 total += x;
3742 nodes[node] += x;
3745 x = sprintf(buf, "%lu", total);
3746 #ifdef CONFIG_NUMA
3747 for_each_node_state(node, N_NORMAL_MEMORY)
3748 if (nodes[node])
3749 x += sprintf(buf + x, " N%d=%lu",
3750 node, nodes[node]);
3751 #endif
3752 kfree(nodes);
3753 return x + sprintf(buf + x, "\n");
3756 static int any_slab_objects(struct kmem_cache *s)
3758 int node;
3760 for_each_online_node(node) {
3761 struct kmem_cache_node *n = get_node(s, node);
3763 if (!n)
3764 continue;
3766 if (atomic_read(&n->total_objects))
3767 return 1;
3769 return 0;
3772 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3773 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3775 struct slab_attribute {
3776 struct attribute attr;
3777 ssize_t (*show)(struct kmem_cache *s, char *buf);
3778 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3781 #define SLAB_ATTR_RO(_name) \
3782 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3784 #define SLAB_ATTR(_name) \
3785 static struct slab_attribute _name##_attr = \
3786 __ATTR(_name, 0644, _name##_show, _name##_store)
3788 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3790 return sprintf(buf, "%d\n", s->size);
3792 SLAB_ATTR_RO(slab_size);
3794 static ssize_t align_show(struct kmem_cache *s, char *buf)
3796 return sprintf(buf, "%d\n", s->align);
3798 SLAB_ATTR_RO(align);
3800 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3802 return sprintf(buf, "%d\n", s->objsize);
3804 SLAB_ATTR_RO(object_size);
3806 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3808 return sprintf(buf, "%d\n", oo_objects(s->oo));
3810 SLAB_ATTR_RO(objs_per_slab);
3812 static ssize_t order_store(struct kmem_cache *s,
3813 const char *buf, size_t length)
3815 int order = simple_strtoul(buf, NULL, 10);
3817 if (order > slub_max_order || order < slub_min_order)
3818 return -EINVAL;
3820 calculate_sizes(s, order);
3821 return length;
3824 static ssize_t order_show(struct kmem_cache *s, char *buf)
3826 return sprintf(buf, "%d\n", oo_order(s->oo));
3828 SLAB_ATTR(order);
3830 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3832 if (s->ctor) {
3833 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3835 return n + sprintf(buf + n, "\n");
3837 return 0;
3839 SLAB_ATTR_RO(ctor);
3841 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3843 return sprintf(buf, "%d\n", s->refcount - 1);
3845 SLAB_ATTR_RO(aliases);
3847 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3849 return show_slab_objects(s, buf, SO_ALL);
3851 SLAB_ATTR_RO(slabs);
3853 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3855 return show_slab_objects(s, buf, SO_PARTIAL);
3857 SLAB_ATTR_RO(partial);
3859 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3861 return show_slab_objects(s, buf, SO_CPU);
3863 SLAB_ATTR_RO(cpu_slabs);
3865 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3867 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
3869 SLAB_ATTR_RO(objects);
3871 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
3873 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
3875 SLAB_ATTR_RO(objects_partial);
3877 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
3879 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
3881 SLAB_ATTR_RO(total_objects);
3883 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3885 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3888 static ssize_t sanity_checks_store(struct kmem_cache *s,
3889 const char *buf, size_t length)
3891 s->flags &= ~SLAB_DEBUG_FREE;
3892 if (buf[0] == '1')
3893 s->flags |= SLAB_DEBUG_FREE;
3894 return length;
3896 SLAB_ATTR(sanity_checks);
3898 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3900 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3903 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3904 size_t length)
3906 s->flags &= ~SLAB_TRACE;
3907 if (buf[0] == '1')
3908 s->flags |= SLAB_TRACE;
3909 return length;
3911 SLAB_ATTR(trace);
3913 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3915 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3918 static ssize_t reclaim_account_store(struct kmem_cache *s,
3919 const char *buf, size_t length)
3921 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3922 if (buf[0] == '1')
3923 s->flags |= SLAB_RECLAIM_ACCOUNT;
3924 return length;
3926 SLAB_ATTR(reclaim_account);
3928 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3930 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3932 SLAB_ATTR_RO(hwcache_align);
3934 #ifdef CONFIG_ZONE_DMA
3935 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3937 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3939 SLAB_ATTR_RO(cache_dma);
3940 #endif
3942 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3944 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3946 SLAB_ATTR_RO(destroy_by_rcu);
3948 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3950 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3953 static ssize_t red_zone_store(struct kmem_cache *s,
3954 const char *buf, size_t length)
3956 if (any_slab_objects(s))
3957 return -EBUSY;
3959 s->flags &= ~SLAB_RED_ZONE;
3960 if (buf[0] == '1')
3961 s->flags |= SLAB_RED_ZONE;
3962 calculate_sizes(s, -1);
3963 return length;
3965 SLAB_ATTR(red_zone);
3967 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3969 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3972 static ssize_t poison_store(struct kmem_cache *s,
3973 const char *buf, size_t length)
3975 if (any_slab_objects(s))
3976 return -EBUSY;
3978 s->flags &= ~SLAB_POISON;
3979 if (buf[0] == '1')
3980 s->flags |= SLAB_POISON;
3981 calculate_sizes(s, -1);
3982 return length;
3984 SLAB_ATTR(poison);
3986 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3988 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3991 static ssize_t store_user_store(struct kmem_cache *s,
3992 const char *buf, size_t length)
3994 if (any_slab_objects(s))
3995 return -EBUSY;
3997 s->flags &= ~SLAB_STORE_USER;
3998 if (buf[0] == '1')
3999 s->flags |= SLAB_STORE_USER;
4000 calculate_sizes(s, -1);
4001 return length;
4003 SLAB_ATTR(store_user);
4005 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4007 return 0;
4010 static ssize_t validate_store(struct kmem_cache *s,
4011 const char *buf, size_t length)
4013 int ret = -EINVAL;
4015 if (buf[0] == '1') {
4016 ret = validate_slab_cache(s);
4017 if (ret >= 0)
4018 ret = length;
4020 return ret;
4022 SLAB_ATTR(validate);
4024 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4026 return 0;
4029 static ssize_t shrink_store(struct kmem_cache *s,
4030 const char *buf, size_t length)
4032 if (buf[0] == '1') {
4033 int rc = kmem_cache_shrink(s);
4035 if (rc)
4036 return rc;
4037 } else
4038 return -EINVAL;
4039 return length;
4041 SLAB_ATTR(shrink);
4043 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4045 if (!(s->flags & SLAB_STORE_USER))
4046 return -ENOSYS;
4047 return list_locations(s, buf, TRACK_ALLOC);
4049 SLAB_ATTR_RO(alloc_calls);
4051 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4053 if (!(s->flags & SLAB_STORE_USER))
4054 return -ENOSYS;
4055 return list_locations(s, buf, TRACK_FREE);
4057 SLAB_ATTR_RO(free_calls);
4059 #ifdef CONFIG_NUMA
4060 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4062 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4065 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4066 const char *buf, size_t length)
4068 int n = simple_strtoul(buf, NULL, 10);
4070 if (n < 100)
4071 s->remote_node_defrag_ratio = n * 10;
4072 return length;
4074 SLAB_ATTR(remote_node_defrag_ratio);
4075 #endif
4077 #ifdef CONFIG_SLUB_STATS
4078 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4080 unsigned long sum = 0;
4081 int cpu;
4082 int len;
4083 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4085 if (!data)
4086 return -ENOMEM;
4088 for_each_online_cpu(cpu) {
4089 unsigned x = get_cpu_slab(s, cpu)->stat[si];
4091 data[cpu] = x;
4092 sum += x;
4095 len = sprintf(buf, "%lu", sum);
4097 #ifdef CONFIG_SMP
4098 for_each_online_cpu(cpu) {
4099 if (data[cpu] && len < PAGE_SIZE - 20)
4100 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4102 #endif
4103 kfree(data);
4104 return len + sprintf(buf + len, "\n");
4107 #define STAT_ATTR(si, text) \
4108 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4110 return show_stat(s, buf, si); \
4112 SLAB_ATTR_RO(text); \
4114 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4115 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4116 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4117 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4118 STAT_ATTR(FREE_FROZEN, free_frozen);
4119 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4120 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4121 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4122 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4123 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4124 STAT_ATTR(FREE_SLAB, free_slab);
4125 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4126 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4127 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4128 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4129 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4130 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4131 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4132 #endif
4134 static struct attribute *slab_attrs[] = {
4135 &slab_size_attr.attr,
4136 &object_size_attr.attr,
4137 &objs_per_slab_attr.attr,
4138 &order_attr.attr,
4139 &objects_attr.attr,
4140 &objects_partial_attr.attr,
4141 &total_objects_attr.attr,
4142 &slabs_attr.attr,
4143 &partial_attr.attr,
4144 &cpu_slabs_attr.attr,
4145 &ctor_attr.attr,
4146 &aliases_attr.attr,
4147 &align_attr.attr,
4148 &sanity_checks_attr.attr,
4149 &trace_attr.attr,
4150 &hwcache_align_attr.attr,
4151 &reclaim_account_attr.attr,
4152 &destroy_by_rcu_attr.attr,
4153 &red_zone_attr.attr,
4154 &poison_attr.attr,
4155 &store_user_attr.attr,
4156 &validate_attr.attr,
4157 &shrink_attr.attr,
4158 &alloc_calls_attr.attr,
4159 &free_calls_attr.attr,
4160 #ifdef CONFIG_ZONE_DMA
4161 &cache_dma_attr.attr,
4162 #endif
4163 #ifdef CONFIG_NUMA
4164 &remote_node_defrag_ratio_attr.attr,
4165 #endif
4166 #ifdef CONFIG_SLUB_STATS
4167 &alloc_fastpath_attr.attr,
4168 &alloc_slowpath_attr.attr,
4169 &free_fastpath_attr.attr,
4170 &free_slowpath_attr.attr,
4171 &free_frozen_attr.attr,
4172 &free_add_partial_attr.attr,
4173 &free_remove_partial_attr.attr,
4174 &alloc_from_partial_attr.attr,
4175 &alloc_slab_attr.attr,
4176 &alloc_refill_attr.attr,
4177 &free_slab_attr.attr,
4178 &cpuslab_flush_attr.attr,
4179 &deactivate_full_attr.attr,
4180 &deactivate_empty_attr.attr,
4181 &deactivate_to_head_attr.attr,
4182 &deactivate_to_tail_attr.attr,
4183 &deactivate_remote_frees_attr.attr,
4184 &order_fallback_attr.attr,
4185 #endif
4186 NULL
4189 static struct attribute_group slab_attr_group = {
4190 .attrs = slab_attrs,
4193 static ssize_t slab_attr_show(struct kobject *kobj,
4194 struct attribute *attr,
4195 char *buf)
4197 struct slab_attribute *attribute;
4198 struct kmem_cache *s;
4199 int err;
4201 attribute = to_slab_attr(attr);
4202 s = to_slab(kobj);
4204 if (!attribute->show)
4205 return -EIO;
4207 err = attribute->show(s, buf);
4209 return err;
4212 static ssize_t slab_attr_store(struct kobject *kobj,
4213 struct attribute *attr,
4214 const char *buf, size_t len)
4216 struct slab_attribute *attribute;
4217 struct kmem_cache *s;
4218 int err;
4220 attribute = to_slab_attr(attr);
4221 s = to_slab(kobj);
4223 if (!attribute->store)
4224 return -EIO;
4226 err = attribute->store(s, buf, len);
4228 return err;
4231 static void kmem_cache_release(struct kobject *kobj)
4233 struct kmem_cache *s = to_slab(kobj);
4235 kfree(s);
4238 static struct sysfs_ops slab_sysfs_ops = {
4239 .show = slab_attr_show,
4240 .store = slab_attr_store,
4243 static struct kobj_type slab_ktype = {
4244 .sysfs_ops = &slab_sysfs_ops,
4245 .release = kmem_cache_release
4248 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4250 struct kobj_type *ktype = get_ktype(kobj);
4252 if (ktype == &slab_ktype)
4253 return 1;
4254 return 0;
4257 static struct kset_uevent_ops slab_uevent_ops = {
4258 .filter = uevent_filter,
4261 static struct kset *slab_kset;
4263 #define ID_STR_LENGTH 64
4265 /* Create a unique string id for a slab cache:
4267 * Format :[flags-]size
4269 static char *create_unique_id(struct kmem_cache *s)
4271 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4272 char *p = name;
4274 BUG_ON(!name);
4276 *p++ = ':';
4278 * First flags affecting slabcache operations. We will only
4279 * get here for aliasable slabs so we do not need to support
4280 * too many flags. The flags here must cover all flags that
4281 * are matched during merging to guarantee that the id is
4282 * unique.
4284 if (s->flags & SLAB_CACHE_DMA)
4285 *p++ = 'd';
4286 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4287 *p++ = 'a';
4288 if (s->flags & SLAB_DEBUG_FREE)
4289 *p++ = 'F';
4290 if (p != name + 1)
4291 *p++ = '-';
4292 p += sprintf(p, "%07d", s->size);
4293 BUG_ON(p > name + ID_STR_LENGTH - 1);
4294 return name;
4297 static int sysfs_slab_add(struct kmem_cache *s)
4299 int err;
4300 const char *name;
4301 int unmergeable;
4303 if (slab_state < SYSFS)
4304 /* Defer until later */
4305 return 0;
4307 unmergeable = slab_unmergeable(s);
4308 if (unmergeable) {
4310 * Slabcache can never be merged so we can use the name proper.
4311 * This is typically the case for debug situations. In that
4312 * case we can catch duplicate names easily.
4314 sysfs_remove_link(&slab_kset->kobj, s->name);
4315 name = s->name;
4316 } else {
4318 * Create a unique name for the slab as a target
4319 * for the symlinks.
4321 name = create_unique_id(s);
4324 s->kobj.kset = slab_kset;
4325 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4326 if (err) {
4327 kobject_put(&s->kobj);
4328 return err;
4331 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4332 if (err)
4333 return err;
4334 kobject_uevent(&s->kobj, KOBJ_ADD);
4335 if (!unmergeable) {
4336 /* Setup first alias */
4337 sysfs_slab_alias(s, s->name);
4338 kfree(name);
4340 return 0;
4343 static void sysfs_slab_remove(struct kmem_cache *s)
4345 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4346 kobject_del(&s->kobj);
4347 kobject_put(&s->kobj);
4351 * Need to buffer aliases during bootup until sysfs becomes
4352 * available lest we loose that information.
4354 struct saved_alias {
4355 struct kmem_cache *s;
4356 const char *name;
4357 struct saved_alias *next;
4360 static struct saved_alias *alias_list;
4362 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4364 struct saved_alias *al;
4366 if (slab_state == SYSFS) {
4368 * If we have a leftover link then remove it.
4370 sysfs_remove_link(&slab_kset->kobj, name);
4371 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4374 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4375 if (!al)
4376 return -ENOMEM;
4378 al->s = s;
4379 al->name = name;
4380 al->next = alias_list;
4381 alias_list = al;
4382 return 0;
4385 static int __init slab_sysfs_init(void)
4387 struct kmem_cache *s;
4388 int err;
4390 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4391 if (!slab_kset) {
4392 printk(KERN_ERR "Cannot register slab subsystem.\n");
4393 return -ENOSYS;
4396 slab_state = SYSFS;
4398 list_for_each_entry(s, &slab_caches, list) {
4399 err = sysfs_slab_add(s);
4400 if (err)
4401 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4402 " to sysfs\n", s->name);
4405 while (alias_list) {
4406 struct saved_alias *al = alias_list;
4408 alias_list = alias_list->next;
4409 err = sysfs_slab_alias(al->s, al->name);
4410 if (err)
4411 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4412 " %s to sysfs\n", s->name);
4413 kfree(al);
4416 resiliency_test();
4417 return 0;
4420 __initcall(slab_sysfs_init);
4421 #endif
4424 * The /proc/slabinfo ABI
4426 #ifdef CONFIG_SLABINFO
4428 ssize_t slabinfo_write(struct file *file, const char __user * buffer,
4429 size_t count, loff_t *ppos)
4431 return -EINVAL;
4435 static void print_slabinfo_header(struct seq_file *m)
4437 seq_puts(m, "slabinfo - version: 2.1\n");
4438 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4439 "<objperslab> <pagesperslab>");
4440 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4441 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4442 seq_putc(m, '\n');
4445 static void *s_start(struct seq_file *m, loff_t *pos)
4447 loff_t n = *pos;
4449 down_read(&slub_lock);
4450 if (!n)
4451 print_slabinfo_header(m);
4453 return seq_list_start(&slab_caches, *pos);
4456 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4458 return seq_list_next(p, &slab_caches, pos);
4461 static void s_stop(struct seq_file *m, void *p)
4463 up_read(&slub_lock);
4466 static int s_show(struct seq_file *m, void *p)
4468 unsigned long nr_partials = 0;
4469 unsigned long nr_slabs = 0;
4470 unsigned long nr_inuse = 0;
4471 unsigned long nr_objs = 0;
4472 unsigned long nr_free = 0;
4473 struct kmem_cache *s;
4474 int node;
4476 s = list_entry(p, struct kmem_cache, list);
4478 for_each_online_node(node) {
4479 struct kmem_cache_node *n = get_node(s, node);
4481 if (!n)
4482 continue;
4484 nr_partials += n->nr_partial;
4485 nr_slabs += atomic_long_read(&n->nr_slabs);
4486 nr_objs += atomic_long_read(&n->total_objects);
4487 nr_free += count_partial(n, count_free);
4490 nr_inuse = nr_objs - nr_free;
4492 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4493 nr_objs, s->size, oo_objects(s->oo),
4494 (1 << oo_order(s->oo)));
4495 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4496 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4497 0UL);
4498 seq_putc(m, '\n');
4499 return 0;
4502 const struct seq_operations slabinfo_op = {
4503 .start = s_start,
4504 .next = s_next,
4505 .stop = s_stop,
4506 .show = s_show,
4509 #endif /* CONFIG_SLABINFO */