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>
12 #include <linux/module.h>
13 #include <linux/bit_spinlock.h>
14 #include <linux/interrupt.h>
15 #include <linux/bitops.h>
16 #include <linux/slab.h>
17 #include <linux/seq_file.h>
18 #include <linux/cpu.h>
19 #include <linux/cpuset.h>
20 #include <linux/mempolicy.h>
21 #include <linux/ctype.h>
22 #include <linux/kallsyms.h>
29 * The slab_lock protects operations on the object of a particular
30 * slab and its metadata in the page struct. If the slab lock
31 * has been taken then no allocations nor frees can be performed
32 * on the objects in the slab nor can the slab be added or removed
33 * from the partial or full lists since this would mean modifying
34 * the page_struct of the slab.
36 * The list_lock protects the partial and full list on each node and
37 * the partial slab counter. If taken then no new slabs may be added or
38 * removed from the lists nor make the number of partial slabs be modified.
39 * (Note that the total number of slabs is an atomic value that may be
40 * modified without taking the list lock).
42 * The list_lock is a centralized lock and thus we avoid taking it as
43 * much as possible. As long as SLUB does not have to handle partial
44 * slabs, operations can continue without any centralized lock. F.e.
45 * allocating a long series of objects that fill up slabs does not require
48 * The lock order is sometimes inverted when we are trying to get a slab
49 * off a list. We take the list_lock and then look for a page on the list
50 * to use. While we do that objects in the slabs may be freed. We can
51 * only operate on the slab if we have also taken the slab_lock. So we use
52 * a slab_trylock() on the slab. If trylock was successful then no frees
53 * can occur anymore and we can use the slab for allocations etc. If the
54 * slab_trylock() does not succeed then frees are in progress in the slab and
55 * we must stay away from it for a while since we may cause a bouncing
56 * cacheline if we try to acquire the lock. So go onto the next slab.
57 * If all pages are busy then we may allocate a new slab instead of reusing
58 * a partial slab. A new slab has noone operating on it and thus there is
59 * no danger of cacheline contention.
61 * Interrupts are disabled during allocation and deallocation in order to
62 * make the slab allocator safe to use in the context of an irq. In addition
63 * interrupts are disabled to ensure that the processor does not change
64 * while handling per_cpu slabs, due to kernel preemption.
66 * SLUB assigns one slab for allocation to each processor.
67 * Allocations only occur from these slabs called cpu slabs.
69 * Slabs with free elements are kept on a partial list and during regular
70 * operations no list for full slabs is used. If an object in a full slab is
71 * freed then the slab will show up again on the partial lists.
72 * We track full slabs for debugging purposes though because otherwise we
73 * cannot scan all objects.
75 * Slabs are freed when they become empty. Teardown and setup is
76 * minimal so we rely on the page allocators per cpu caches for
77 * fast frees and allocs.
79 * Overloading of page flags that are otherwise used for LRU management.
81 * PageActive The slab is used as a cpu cache. Allocations
82 * may be performed from the slab. The slab is not
83 * on any slab list and cannot be moved onto one.
85 * PageError Slab requires special handling due to debug
86 * options set. This moves slab handling out of
90 static inline int SlabDebug(struct page *page)
92 #ifdef CONFIG_SLUB_DEBUG
93 return PageError(page);
99 static inline void SetSlabDebug(struct page *page)
101 #ifdef CONFIG_SLUB_DEBUG
106 static inline void ClearSlabDebug(struct page *page)
108 #ifdef CONFIG_SLUB_DEBUG
109 ClearPageError(page);
114 * Issues still to be resolved:
116 * - The per cpu array is updated for each new slab and and is a remote
117 * cacheline for most nodes. This could become a bouncing cacheline given
118 * enough frequent updates. There are 16 pointers in a cacheline, so at
119 * max 16 cpus could compete for the cacheline which may be okay.
121 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
123 * - Variable sizing of the per node arrays
126 /* Enable to test recovery from slab corruption on boot */
127 #undef SLUB_RESILIENCY_TEST
132 * Small page size. Make sure that we do not fragment memory
134 #define DEFAULT_MAX_ORDER 1
135 #define DEFAULT_MIN_OBJECTS 4
140 * Large page machines are customarily able to handle larger
143 #define DEFAULT_MAX_ORDER 2
144 #define DEFAULT_MIN_OBJECTS 8
149 * Mininum number of partial slabs. These will be left on the partial
150 * lists even if they are empty. kmem_cache_shrink may reclaim them.
152 #define MIN_PARTIAL 2
155 * Maximum number of desirable partial slabs.
156 * The existence of more partial slabs makes kmem_cache_shrink
157 * sort the partial list by the number of objects in the.
159 #define MAX_PARTIAL 10
161 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
162 SLAB_POISON | SLAB_STORE_USER)
165 * Set of flags that will prevent slab merging
167 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
168 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
170 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
173 #ifndef ARCH_KMALLOC_MINALIGN
174 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
177 #ifndef ARCH_SLAB_MINALIGN
178 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
181 /* Internal SLUB flags */
182 #define __OBJECT_POISON 0x80000000 /* Poison object */
184 /* Not all arches define cache_line_size */
185 #ifndef cache_line_size
186 #define cache_line_size() L1_CACHE_BYTES
189 static int kmem_size = sizeof(struct kmem_cache);
192 static struct notifier_block slab_notifier;
196 DOWN, /* No slab functionality available */
197 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
198 UP, /* Everything works but does not show up in sysfs */
202 /* A list of all slab caches on the system */
203 static DECLARE_RWSEM(slub_lock);
204 LIST_HEAD(slab_caches);
207 * Tracking user of a slab.
210 void *addr; /* Called from address */
211 int cpu; /* Was running on cpu */
212 int pid; /* Pid context */
213 unsigned long when; /* When did the operation occur */
216 enum track_item { TRACK_ALLOC, TRACK_FREE };
218 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
219 static int sysfs_slab_add(struct kmem_cache *);
220 static int sysfs_slab_alias(struct kmem_cache *, const char *);
221 static void sysfs_slab_remove(struct kmem_cache *);
223 static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
224 static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
225 static void sysfs_slab_remove(struct kmem_cache *s) {}
228 /********************************************************************
229 * Core slab cache functions
230 *******************************************************************/
232 int slab_is_available(void)
234 return slab_state >= UP;
237 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
240 return s->node[node];
242 return &s->local_node;
246 static inline int check_valid_pointer(struct kmem_cache *s,
247 struct page *page, const void *object)
254 base = page_address(page);
255 if (object < base || object >= base + s->objects * s->size ||
256 (object - base) % s->size) {
264 * Slow version of get and set free pointer.
266 * This version requires touching the cache lines of kmem_cache which
267 * we avoid to do in the fast alloc free paths. There we obtain the offset
268 * from the page struct.
270 static inline void *get_freepointer(struct kmem_cache *s, void *object)
272 return *(void **)(object + s->offset);
275 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
277 *(void **)(object + s->offset) = fp;
280 /* Loop over all objects in a slab */
281 #define for_each_object(__p, __s, __addr) \
282 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
286 #define for_each_free_object(__p, __s, __free) \
287 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
289 /* Determine object index from a given position */
290 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
292 return (p - addr) / s->size;
295 #ifdef CONFIG_SLUB_DEBUG
299 static int slub_debug;
301 static char *slub_debug_slabs;
306 static void print_section(char *text, u8 *addr, unsigned int length)
314 for (i = 0; i < length; i++) {
316 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
319 printk(" %02x", addr[i]);
321 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
323 printk(" %s\n",ascii);
334 printk(" %s\n", ascii);
338 static struct track *get_track(struct kmem_cache *s, void *object,
339 enum track_item alloc)
344 p = object + s->offset + sizeof(void *);
346 p = object + s->inuse;
351 static void set_track(struct kmem_cache *s, void *object,
352 enum track_item alloc, void *addr)
357 p = object + s->offset + sizeof(void *);
359 p = object + s->inuse;
364 p->cpu = smp_processor_id();
365 p->pid = current ? current->pid : -1;
368 memset(p, 0, sizeof(struct track));
371 static void init_tracking(struct kmem_cache *s, void *object)
373 if (s->flags & SLAB_STORE_USER) {
374 set_track(s, object, TRACK_FREE, NULL);
375 set_track(s, object, TRACK_ALLOC, NULL);
379 static void print_track(const char *s, struct track *t)
384 printk(KERN_ERR "%s: ", s);
385 __print_symbol("%s", (unsigned long)t->addr);
386 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
389 static void print_trailer(struct kmem_cache *s, u8 *p)
391 unsigned int off; /* Offset of last byte */
393 if (s->flags & SLAB_RED_ZONE)
394 print_section("Redzone", p + s->objsize,
395 s->inuse - s->objsize);
397 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
399 get_freepointer(s, p));
402 off = s->offset + sizeof(void *);
406 if (s->flags & SLAB_STORE_USER) {
407 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
408 print_track("Last free ", get_track(s, p, TRACK_FREE));
409 off += 2 * sizeof(struct track);
413 /* Beginning of the filler is the free pointer */
414 print_section("Filler", p + off, s->size - off);
417 static void object_err(struct kmem_cache *s, struct page *page,
418 u8 *object, char *reason)
420 u8 *addr = page_address(page);
422 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
423 s->name, reason, object, page);
424 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
425 object - addr, page->flags, page->inuse, page->freelist);
426 if (object > addr + 16)
427 print_section("Bytes b4", object - 16, 16);
428 print_section("Object", object, min(s->objsize, 128));
429 print_trailer(s, object);
433 static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
438 va_start(args, reason);
439 vsnprintf(buf, sizeof(buf), reason, args);
441 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
446 static void init_object(struct kmem_cache *s, void *object, int active)
450 if (s->flags & __OBJECT_POISON) {
451 memset(p, POISON_FREE, s->objsize - 1);
452 p[s->objsize -1] = POISON_END;
455 if (s->flags & SLAB_RED_ZONE)
456 memset(p + s->objsize,
457 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
458 s->inuse - s->objsize);
461 static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
464 if (*start != (u8)value)
476 * Bytes of the object to be managed.
477 * If the freepointer may overlay the object then the free
478 * pointer is the first word of the object.
480 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
483 * object + s->objsize
484 * Padding to reach word boundary. This is also used for Redzoning.
485 * Padding is extended by another word if Redzoning is enabled and
488 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
489 * 0xcc (RED_ACTIVE) for objects in use.
492 * Meta data starts here.
494 * A. Free pointer (if we cannot overwrite object on free)
495 * B. Tracking data for SLAB_STORE_USER
496 * C. Padding to reach required alignment boundary or at mininum
497 * one word if debuggin is on to be able to detect writes
498 * before the word boundary.
500 * Padding is done using 0x5a (POISON_INUSE)
503 * Nothing is used beyond s->size.
505 * If slabcaches are merged then the objsize and inuse boundaries are mostly
506 * ignored. And therefore no slab options that rely on these boundaries
507 * may be used with merged slabcaches.
510 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
511 void *from, void *to)
513 printk(KERN_ERR "@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
514 s->name, message, data, from, to - 1);
515 memset(from, data, to - from);
518 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
520 unsigned long off = s->inuse; /* The end of info */
523 /* Freepointer is placed after the object. */
524 off += sizeof(void *);
526 if (s->flags & SLAB_STORE_USER)
527 /* We also have user information there */
528 off += 2 * sizeof(struct track);
533 if (check_bytes(p + off, POISON_INUSE, s->size - off))
536 object_err(s, page, p, "Object padding check fails");
541 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
545 static int slab_pad_check(struct kmem_cache *s, struct page *page)
548 int length, remainder;
550 if (!(s->flags & SLAB_POISON))
553 p = page_address(page);
554 length = s->objects * s->size;
555 remainder = (PAGE_SIZE << s->order) - length;
559 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
560 slab_err(s, page, "Padding check failed");
561 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
562 p + length + remainder);
568 static int check_object(struct kmem_cache *s, struct page *page,
569 void *object, int active)
572 u8 *endobject = object + s->objsize;
574 if (s->flags & SLAB_RED_ZONE) {
576 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
578 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
579 object_err(s, page, object,
580 active ? "Redzone Active" : "Redzone Inactive");
581 restore_bytes(s, "redzone", red,
582 endobject, object + s->inuse);
586 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
587 !check_bytes(endobject, POISON_INUSE,
588 s->inuse - s->objsize)) {
589 object_err(s, page, p, "Alignment padding check fails");
591 * Fix it so that there will not be another report.
593 * Hmmm... We may be corrupting an object that now expects
594 * to be longer than allowed.
596 restore_bytes(s, "alignment padding", POISON_INUSE,
597 endobject, object + s->inuse);
601 if (s->flags & SLAB_POISON) {
602 if (!active && (s->flags & __OBJECT_POISON) &&
603 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
604 p[s->objsize - 1] != POISON_END)) {
606 object_err(s, page, p, "Poison check failed");
607 restore_bytes(s, "Poison", POISON_FREE,
608 p, p + s->objsize -1);
609 restore_bytes(s, "Poison", POISON_END,
610 p + s->objsize - 1, p + s->objsize);
614 * check_pad_bytes cleans up on its own.
616 check_pad_bytes(s, page, p);
619 if (!s->offset && active)
621 * Object and freepointer overlap. Cannot check
622 * freepointer while object is allocated.
626 /* Check free pointer validity */
627 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
628 object_err(s, page, p, "Freepointer corrupt");
630 * No choice but to zap it and thus loose the remainder
631 * of the free objects in this slab. May cause
632 * another error because the object count is now wrong.
634 set_freepointer(s, p, NULL);
640 static int check_slab(struct kmem_cache *s, struct page *page)
642 VM_BUG_ON(!irqs_disabled());
644 if (!PageSlab(page)) {
645 slab_err(s, page, "Not a valid slab page flags=%lx "
646 "mapping=0x%p count=%d", page->flags, page->mapping,
650 if (page->offset * sizeof(void *) != s->offset) {
651 slab_err(s, page, "Corrupted offset %lu flags=0x%lx "
652 "mapping=0x%p count=%d",
653 (unsigned long)(page->offset * sizeof(void *)),
659 if (page->inuse > s->objects) {
660 slab_err(s, page, "inuse %u > max %u @0x%p flags=%lx "
661 "mapping=0x%p count=%d",
662 s->name, page->inuse, s->objects, page->flags,
663 page->mapping, page_count(page));
666 /* Slab_pad_check fixes things up after itself */
667 slab_pad_check(s, page);
672 * Determine if a certain object on a page is on the freelist. Must hold the
673 * slab lock to guarantee that the chains are in a consistent state.
675 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
678 void *fp = page->freelist;
681 while (fp && nr <= s->objects) {
684 if (!check_valid_pointer(s, page, fp)) {
686 object_err(s, page, object,
687 "Freechain corrupt");
688 set_freepointer(s, object, NULL);
691 slab_err(s, page, "Freepointer 0x%p corrupt",
693 page->freelist = NULL;
694 page->inuse = s->objects;
695 printk(KERN_ERR "@@@ SLUB %s: Freelist "
696 "cleared. Slab 0x%p\n",
703 fp = get_freepointer(s, object);
707 if (page->inuse != s->objects - nr) {
708 slab_err(s, page, "Wrong object count. Counter is %d but "
709 "counted were %d", s, page, page->inuse,
711 page->inuse = s->objects - nr;
712 printk(KERN_ERR "@@@ SLUB %s: Object count adjusted. "
713 "Slab @0x%p\n", s->name, page);
715 return search == NULL;
719 * Tracking of fully allocated slabs for debugging purposes.
721 static void add_full(struct kmem_cache_node *n, struct page *page)
723 spin_lock(&n->list_lock);
724 list_add(&page->lru, &n->full);
725 spin_unlock(&n->list_lock);
728 static void remove_full(struct kmem_cache *s, struct page *page)
730 struct kmem_cache_node *n;
732 if (!(s->flags & SLAB_STORE_USER))
735 n = get_node(s, page_to_nid(page));
737 spin_lock(&n->list_lock);
738 list_del(&page->lru);
739 spin_unlock(&n->list_lock);
742 static int alloc_object_checks(struct kmem_cache *s, struct page *page,
745 if (!check_slab(s, page))
748 if (object && !on_freelist(s, page, object)) {
749 slab_err(s, page, "Object 0x%p already allocated", object);
753 if (!check_valid_pointer(s, page, object)) {
754 object_err(s, page, object, "Freelist Pointer check fails");
761 if (!check_object(s, page, object, 0))
766 if (PageSlab(page)) {
768 * If this is a slab page then lets do the best we can
769 * to avoid issues in the future. Marking all objects
770 * as used avoids touching the remaining objects.
772 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
774 page->inuse = s->objects;
775 page->freelist = NULL;
776 /* Fix up fields that may be corrupted */
777 page->offset = s->offset / sizeof(void *);
782 static int free_object_checks(struct kmem_cache *s, struct page *page,
785 if (!check_slab(s, page))
788 if (!check_valid_pointer(s, page, object)) {
789 slab_err(s, page, "Invalid object pointer 0x%p", object);
793 if (on_freelist(s, page, object)) {
794 slab_err(s, page, "Object 0x%p already free", object);
798 if (!check_object(s, page, object, 1))
801 if (unlikely(s != page->slab)) {
803 slab_err(s, page, "Attempt to free object(0x%p) "
804 "outside of slab", object);
808 "SLUB <none>: no slab for object 0x%p.\n",
813 slab_err(s, page, "object at 0x%p belongs "
814 "to slab %s", object, page->slab->name);
819 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
820 s->name, page, object);
824 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
826 if (s->flags & SLAB_TRACE) {
827 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
829 alloc ? "alloc" : "free",
834 print_section("Object", (void *)object, s->objsize);
840 static int __init setup_slub_debug(char *str)
842 if (!str || *str != '=')
843 slub_debug = DEBUG_DEFAULT_FLAGS;
846 if (*str == 0 || *str == ',')
847 slub_debug = DEBUG_DEFAULT_FLAGS;
849 for( ;*str && *str != ','; str++)
851 case 'f' : case 'F' :
852 slub_debug |= SLAB_DEBUG_FREE;
854 case 'z' : case 'Z' :
855 slub_debug |= SLAB_RED_ZONE;
857 case 'p' : case 'P' :
858 slub_debug |= SLAB_POISON;
860 case 'u' : case 'U' :
861 slub_debug |= SLAB_STORE_USER;
863 case 't' : case 'T' :
864 slub_debug |= SLAB_TRACE;
867 printk(KERN_ERR "slub_debug option '%c' "
868 "unknown. skipped\n",*str);
873 slub_debug_slabs = str + 1;
877 __setup("slub_debug", setup_slub_debug);
879 static void kmem_cache_open_debug_check(struct kmem_cache *s)
882 * The page->offset field is only 16 bit wide. This is an offset
883 * in units of words from the beginning of an object. If the slab
884 * size is bigger then we cannot move the free pointer behind the
887 * On 32 bit platforms the limit is 256k. On 64bit platforms
890 * Debugging or ctor/dtors may create a need to move the free
891 * pointer. Fail if this happens.
893 if (s->size >= 65535 * sizeof(void *)) {
894 BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON |
895 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
896 BUG_ON(s->ctor || s->dtor);
900 * Enable debugging if selected on the kernel commandline.
902 if (slub_debug && (!slub_debug_slabs ||
903 strncmp(slub_debug_slabs, s->name,
904 strlen(slub_debug_slabs)) == 0))
905 s->flags |= slub_debug;
909 static inline int alloc_object_checks(struct kmem_cache *s,
910 struct page *page, void *object) { return 0; }
912 static inline int free_object_checks(struct kmem_cache *s,
913 struct page *page, void *object) { return 0; }
915 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
916 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
917 static inline void trace(struct kmem_cache *s, struct page *page,
918 void *object, int alloc) {}
919 static inline void init_object(struct kmem_cache *s,
920 void *object, int active) {}
921 static inline void init_tracking(struct kmem_cache *s, void *object) {}
922 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
924 static inline int check_object(struct kmem_cache *s, struct page *page,
925 void *object, int active) { return 1; }
926 static inline void set_track(struct kmem_cache *s, void *object,
927 enum track_item alloc, void *addr) {}
928 static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {}
932 * Slab allocation and freeing
934 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
937 int pages = 1 << s->order;
942 if (s->flags & SLAB_CACHE_DMA)
946 page = alloc_pages(flags, s->order);
948 page = alloc_pages_node(node, flags, s->order);
953 mod_zone_page_state(page_zone(page),
954 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
955 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
961 static void setup_object(struct kmem_cache *s, struct page *page,
964 if (SlabDebug(page)) {
965 init_object(s, object, 0);
966 init_tracking(s, object);
969 if (unlikely(s->ctor))
970 s->ctor(object, s, SLAB_CTOR_CONSTRUCTOR);
973 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
976 struct kmem_cache_node *n;
982 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
984 if (flags & __GFP_WAIT)
987 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
991 n = get_node(s, page_to_nid(page));
993 atomic_long_inc(&n->nr_slabs);
994 page->offset = s->offset / sizeof(void *);
996 page->flags |= 1 << PG_slab;
997 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
998 SLAB_STORE_USER | SLAB_TRACE))
1001 start = page_address(page);
1002 end = start + s->objects * s->size;
1004 if (unlikely(s->flags & SLAB_POISON))
1005 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1008 for_each_object(p, s, start) {
1009 setup_object(s, page, last);
1010 set_freepointer(s, last, p);
1013 setup_object(s, page, last);
1014 set_freepointer(s, last, NULL);
1016 page->freelist = start;
1019 if (flags & __GFP_WAIT)
1020 local_irq_disable();
1024 static void __free_slab(struct kmem_cache *s, struct page *page)
1026 int pages = 1 << s->order;
1028 if (unlikely(SlabDebug(page) || s->dtor)) {
1031 slab_pad_check(s, page);
1032 for_each_object(p, s, page_address(page)) {
1035 check_object(s, page, p, 0);
1039 mod_zone_page_state(page_zone(page),
1040 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1041 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1044 page->mapping = NULL;
1045 __free_pages(page, s->order);
1048 static void rcu_free_slab(struct rcu_head *h)
1052 page = container_of((struct list_head *)h, struct page, lru);
1053 __free_slab(page->slab, page);
1056 static void free_slab(struct kmem_cache *s, struct page *page)
1058 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1060 * RCU free overloads the RCU head over the LRU
1062 struct rcu_head *head = (void *)&page->lru;
1064 call_rcu(head, rcu_free_slab);
1066 __free_slab(s, page);
1069 static void discard_slab(struct kmem_cache *s, struct page *page)
1071 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1073 atomic_long_dec(&n->nr_slabs);
1074 reset_page_mapcount(page);
1075 ClearSlabDebug(page);
1076 __ClearPageSlab(page);
1081 * Per slab locking using the pagelock
1083 static __always_inline void slab_lock(struct page *page)
1085 bit_spin_lock(PG_locked, &page->flags);
1088 static __always_inline void slab_unlock(struct page *page)
1090 bit_spin_unlock(PG_locked, &page->flags);
1093 static __always_inline int slab_trylock(struct page *page)
1097 rc = bit_spin_trylock(PG_locked, &page->flags);
1102 * Management of partially allocated slabs
1104 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1106 spin_lock(&n->list_lock);
1108 list_add_tail(&page->lru, &n->partial);
1109 spin_unlock(&n->list_lock);
1112 static void add_partial(struct kmem_cache_node *n, struct page *page)
1114 spin_lock(&n->list_lock);
1116 list_add(&page->lru, &n->partial);
1117 spin_unlock(&n->list_lock);
1120 static void remove_partial(struct kmem_cache *s,
1123 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1125 spin_lock(&n->list_lock);
1126 list_del(&page->lru);
1128 spin_unlock(&n->list_lock);
1132 * Lock slab and remove from the partial list.
1134 * Must hold list_lock.
1136 static int lock_and_del_slab(struct kmem_cache_node *n, struct page *page)
1138 if (slab_trylock(page)) {
1139 list_del(&page->lru);
1147 * Try to allocate a partial slab from a specific node.
1149 static struct page *get_partial_node(struct kmem_cache_node *n)
1154 * Racy check. If we mistakenly see no partial slabs then we
1155 * just allocate an empty slab. If we mistakenly try to get a
1156 * partial slab and there is none available then get_partials()
1159 if (!n || !n->nr_partial)
1162 spin_lock(&n->list_lock);
1163 list_for_each_entry(page, &n->partial, lru)
1164 if (lock_and_del_slab(n, page))
1168 spin_unlock(&n->list_lock);
1173 * Get a page from somewhere. Search in increasing NUMA distances.
1175 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1178 struct zonelist *zonelist;
1183 * The defrag ratio allows a configuration of the tradeoffs between
1184 * inter node defragmentation and node local allocations. A lower
1185 * defrag_ratio increases the tendency to do local allocations
1186 * instead of attempting to obtain partial slabs from other nodes.
1188 * If the defrag_ratio is set to 0 then kmalloc() always
1189 * returns node local objects. If the ratio is higher then kmalloc()
1190 * may return off node objects because partial slabs are obtained
1191 * from other nodes and filled up.
1193 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1194 * defrag_ratio = 1000) then every (well almost) allocation will
1195 * first attempt to defrag slab caches on other nodes. This means
1196 * scanning over all nodes to look for partial slabs which may be
1197 * expensive if we do it every time we are trying to find a slab
1198 * with available objects.
1200 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1203 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1204 ->node_zonelists[gfp_zone(flags)];
1205 for (z = zonelist->zones; *z; z++) {
1206 struct kmem_cache_node *n;
1208 n = get_node(s, zone_to_nid(*z));
1210 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1211 n->nr_partial > MIN_PARTIAL) {
1212 page = get_partial_node(n);
1222 * Get a partial page, lock it and return it.
1224 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1227 int searchnode = (node == -1) ? numa_node_id() : node;
1229 page = get_partial_node(get_node(s, searchnode));
1230 if (page || (flags & __GFP_THISNODE))
1233 return get_any_partial(s, flags);
1237 * Move a page back to the lists.
1239 * Must be called with the slab lock held.
1241 * On exit the slab lock will have been dropped.
1243 static void putback_slab(struct kmem_cache *s, struct page *page)
1245 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1250 add_partial(n, page);
1251 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1256 if (n->nr_partial < MIN_PARTIAL) {
1258 * Adding an empty slab to the partial slabs in order
1259 * to avoid page allocator overhead. This slab needs
1260 * to come after the other slabs with objects in
1261 * order to fill them up. That way the size of the
1262 * partial list stays small. kmem_cache_shrink can
1263 * reclaim empty slabs from the partial list.
1265 add_partial_tail(n, page);
1269 discard_slab(s, page);
1275 * Remove the cpu slab
1277 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1279 s->cpu_slab[cpu] = NULL;
1280 ClearPageActive(page);
1282 putback_slab(s, page);
1285 static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1288 deactivate_slab(s, page, cpu);
1293 * Called from IPI handler with interrupts disabled.
1295 static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1297 struct page *page = s->cpu_slab[cpu];
1300 flush_slab(s, page, cpu);
1303 static void flush_cpu_slab(void *d)
1305 struct kmem_cache *s = d;
1306 int cpu = smp_processor_id();
1308 __flush_cpu_slab(s, cpu);
1311 static void flush_all(struct kmem_cache *s)
1314 on_each_cpu(flush_cpu_slab, s, 1, 1);
1316 unsigned long flags;
1318 local_irq_save(flags);
1320 local_irq_restore(flags);
1325 * slab_alloc is optimized to only modify two cachelines on the fast path
1326 * (aside from the stack):
1328 * 1. The page struct
1329 * 2. The first cacheline of the object to be allocated.
1331 * The only other cache lines that are read (apart from code) is the
1332 * per cpu array in the kmem_cache struct.
1334 * Fastpath is not possible if we need to get a new slab or have
1335 * debugging enabled (which means all slabs are marked with SlabDebug)
1337 static void *slab_alloc(struct kmem_cache *s,
1338 gfp_t gfpflags, int node, void *addr)
1342 unsigned long flags;
1345 local_irq_save(flags);
1346 cpu = smp_processor_id();
1347 page = s->cpu_slab[cpu];
1352 if (unlikely(node != -1 && page_to_nid(page) != node))
1355 object = page->freelist;
1356 if (unlikely(!object))
1358 if (unlikely(SlabDebug(page)))
1363 page->freelist = object[page->offset];
1365 local_irq_restore(flags);
1369 deactivate_slab(s, page, cpu);
1372 page = get_partial(s, gfpflags, node);
1375 s->cpu_slab[cpu] = page;
1376 SetPageActive(page);
1380 page = new_slab(s, gfpflags, node);
1382 cpu = smp_processor_id();
1383 if (s->cpu_slab[cpu]) {
1385 * Someone else populated the cpu_slab while we
1386 * enabled interrupts, or we have gotten scheduled
1387 * on another cpu. The page may not be on the
1388 * requested node even if __GFP_THISNODE was
1389 * specified. So we need to recheck.
1392 page_to_nid(s->cpu_slab[cpu]) == node) {
1394 * Current cpuslab is acceptable and we
1395 * want the current one since its cache hot
1397 discard_slab(s, page);
1398 page = s->cpu_slab[cpu];
1402 /* New slab does not fit our expectations */
1403 flush_slab(s, s->cpu_slab[cpu], cpu);
1408 local_irq_restore(flags);
1411 if (!alloc_object_checks(s, page, object))
1413 if (s->flags & SLAB_STORE_USER)
1414 set_track(s, object, TRACK_ALLOC, addr);
1415 trace(s, page, object, 1);
1416 init_object(s, object, 1);
1420 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1422 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1424 EXPORT_SYMBOL(kmem_cache_alloc);
1427 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1429 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1431 EXPORT_SYMBOL(kmem_cache_alloc_node);
1435 * The fastpath only writes the cacheline of the page struct and the first
1436 * cacheline of the object.
1438 * We read the cpu_slab cacheline to check if the slab is the per cpu
1439 * slab for this processor.
1441 static void slab_free(struct kmem_cache *s, struct page *page,
1442 void *x, void *addr)
1445 void **object = (void *)x;
1446 unsigned long flags;
1448 local_irq_save(flags);
1451 if (unlikely(SlabDebug(page)))
1454 prior = object[page->offset] = page->freelist;
1455 page->freelist = object;
1458 if (unlikely(PageActive(page)))
1460 * Cpu slabs are never on partial lists and are
1465 if (unlikely(!page->inuse))
1469 * Objects left in the slab. If it
1470 * was not on the partial list before
1473 if (unlikely(!prior))
1474 add_partial(get_node(s, page_to_nid(page)), page);
1478 local_irq_restore(flags);
1484 * Slab still on the partial list.
1486 remove_partial(s, page);
1489 discard_slab(s, page);
1490 local_irq_restore(flags);
1494 if (!free_object_checks(s, page, x))
1496 if (!PageActive(page) && !page->freelist)
1497 remove_full(s, page);
1498 if (s->flags & SLAB_STORE_USER)
1499 set_track(s, x, TRACK_FREE, addr);
1500 trace(s, page, object, 0);
1501 init_object(s, object, 0);
1505 void kmem_cache_free(struct kmem_cache *s, void *x)
1509 page = virt_to_head_page(x);
1511 slab_free(s, page, x, __builtin_return_address(0));
1513 EXPORT_SYMBOL(kmem_cache_free);
1515 /* Figure out on which slab object the object resides */
1516 static struct page *get_object_page(const void *x)
1518 struct page *page = virt_to_head_page(x);
1520 if (!PageSlab(page))
1527 * Object placement in a slab is made very easy because we always start at
1528 * offset 0. If we tune the size of the object to the alignment then we can
1529 * get the required alignment by putting one properly sized object after
1532 * Notice that the allocation order determines the sizes of the per cpu
1533 * caches. Each processor has always one slab available for allocations.
1534 * Increasing the allocation order reduces the number of times that slabs
1535 * must be moved on and off the partial lists and is therefore a factor in
1540 * Mininum / Maximum order of slab pages. This influences locking overhead
1541 * and slab fragmentation. A higher order reduces the number of partial slabs
1542 * and increases the number of allocations possible without having to
1543 * take the list_lock.
1545 static int slub_min_order;
1546 static int slub_max_order = DEFAULT_MAX_ORDER;
1547 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1550 * Merge control. If this is set then no merging of slab caches will occur.
1551 * (Could be removed. This was introduced to pacify the merge skeptics.)
1553 static int slub_nomerge;
1556 * Calculate the order of allocation given an slab object size.
1558 * The order of allocation has significant impact on performance and other
1559 * system components. Generally order 0 allocations should be preferred since
1560 * order 0 does not cause fragmentation in the page allocator. Larger objects
1561 * be problematic to put into order 0 slabs because there may be too much
1562 * unused space left. We go to a higher order if more than 1/8th of the slab
1565 * In order to reach satisfactory performance we must ensure that a minimum
1566 * number of objects is in one slab. Otherwise we may generate too much
1567 * activity on the partial lists which requires taking the list_lock. This is
1568 * less a concern for large slabs though which are rarely used.
1570 * slub_max_order specifies the order where we begin to stop considering the
1571 * number of objects in a slab as critical. If we reach slub_max_order then
1572 * we try to keep the page order as low as possible. So we accept more waste
1573 * of space in favor of a small page order.
1575 * Higher order allocations also allow the placement of more objects in a
1576 * slab and thereby reduce object handling overhead. If the user has
1577 * requested a higher mininum order then we start with that one instead of
1578 * the smallest order which will fit the object.
1580 static int calculate_order(int size)
1585 for (order = max(slub_min_order, fls(size - 1) - PAGE_SHIFT);
1586 order < MAX_ORDER; order++) {
1587 unsigned long slab_size = PAGE_SIZE << order;
1589 if (order < slub_max_order &&
1590 slab_size < slub_min_objects * size)
1593 if (slab_size < size)
1596 if (order >= slub_max_order)
1599 rem = slab_size % size;
1601 if (rem <= slab_size / 8)
1605 if (order >= MAX_ORDER)
1612 * Figure out what the alignment of the objects will be.
1614 static unsigned long calculate_alignment(unsigned long flags,
1615 unsigned long align, unsigned long size)
1618 * If the user wants hardware cache aligned objects then
1619 * follow that suggestion if the object is sufficiently
1622 * The hardware cache alignment cannot override the
1623 * specified alignment though. If that is greater
1626 if ((flags & SLAB_HWCACHE_ALIGN) &&
1627 size > cache_line_size() / 2)
1628 return max_t(unsigned long, align, cache_line_size());
1630 if (align < ARCH_SLAB_MINALIGN)
1631 return ARCH_SLAB_MINALIGN;
1633 return ALIGN(align, sizeof(void *));
1636 static void init_kmem_cache_node(struct kmem_cache_node *n)
1639 atomic_long_set(&n->nr_slabs, 0);
1640 spin_lock_init(&n->list_lock);
1641 INIT_LIST_HEAD(&n->partial);
1642 INIT_LIST_HEAD(&n->full);
1647 * No kmalloc_node yet so do it by hand. We know that this is the first
1648 * slab on the node for this slabcache. There are no concurrent accesses
1651 * Note that this function only works on the kmalloc_node_cache
1652 * when allocating for the kmalloc_node_cache.
1654 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1658 struct kmem_cache_node *n;
1660 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1662 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1663 /* new_slab() disables interupts */
1669 page->freelist = get_freepointer(kmalloc_caches, n);
1671 kmalloc_caches->node[node] = n;
1672 init_object(kmalloc_caches, n, 1);
1673 init_kmem_cache_node(n);
1674 atomic_long_inc(&n->nr_slabs);
1675 add_partial(n, page);
1679 static void free_kmem_cache_nodes(struct kmem_cache *s)
1683 for_each_online_node(node) {
1684 struct kmem_cache_node *n = s->node[node];
1685 if (n && n != &s->local_node)
1686 kmem_cache_free(kmalloc_caches, n);
1687 s->node[node] = NULL;
1691 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1696 if (slab_state >= UP)
1697 local_node = page_to_nid(virt_to_page(s));
1701 for_each_online_node(node) {
1702 struct kmem_cache_node *n;
1704 if (local_node == node)
1707 if (slab_state == DOWN) {
1708 n = early_kmem_cache_node_alloc(gfpflags,
1712 n = kmem_cache_alloc_node(kmalloc_caches,
1716 free_kmem_cache_nodes(s);
1722 init_kmem_cache_node(n);
1727 static void free_kmem_cache_nodes(struct kmem_cache *s)
1731 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1733 init_kmem_cache_node(&s->local_node);
1739 * calculate_sizes() determines the order and the distribution of data within
1742 static int calculate_sizes(struct kmem_cache *s)
1744 unsigned long flags = s->flags;
1745 unsigned long size = s->objsize;
1746 unsigned long align = s->align;
1749 * Determine if we can poison the object itself. If the user of
1750 * the slab may touch the object after free or before allocation
1751 * then we should never poison the object itself.
1753 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1754 !s->ctor && !s->dtor)
1755 s->flags |= __OBJECT_POISON;
1757 s->flags &= ~__OBJECT_POISON;
1760 * Round up object size to the next word boundary. We can only
1761 * place the free pointer at word boundaries and this determines
1762 * the possible location of the free pointer.
1764 size = ALIGN(size, sizeof(void *));
1766 #ifdef CONFIG_SLUB_DEBUG
1768 * If we are Redzoning then check if there is some space between the
1769 * end of the object and the free pointer. If not then add an
1770 * additional word to have some bytes to store Redzone information.
1772 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1773 size += sizeof(void *);
1777 * With that we have determined the number of bytes in actual use
1778 * by the object. This is the potential offset to the free pointer.
1782 #ifdef CONFIG_SLUB_DEBUG
1783 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1784 s->ctor || s->dtor)) {
1786 * Relocate free pointer after the object if it is not
1787 * permitted to overwrite the first word of the object on
1790 * This is the case if we do RCU, have a constructor or
1791 * destructor or are poisoning the objects.
1794 size += sizeof(void *);
1797 if (flags & SLAB_STORE_USER)
1799 * Need to store information about allocs and frees after
1802 size += 2 * sizeof(struct track);
1804 if (flags & SLAB_RED_ZONE)
1806 * Add some empty padding so that we can catch
1807 * overwrites from earlier objects rather than let
1808 * tracking information or the free pointer be
1809 * corrupted if an user writes before the start
1812 size += sizeof(void *);
1816 * Determine the alignment based on various parameters that the
1817 * user specified and the dynamic determination of cache line size
1820 align = calculate_alignment(flags, align, s->objsize);
1823 * SLUB stores one object immediately after another beginning from
1824 * offset 0. In order to align the objects we have to simply size
1825 * each object to conform to the alignment.
1827 size = ALIGN(size, align);
1830 s->order = calculate_order(size);
1835 * Determine the number of objects per slab
1837 s->objects = (PAGE_SIZE << s->order) / size;
1840 * Verify that the number of objects is within permitted limits.
1841 * The page->inuse field is only 16 bit wide! So we cannot have
1842 * more than 64k objects per slab.
1844 if (!s->objects || s->objects > 65535)
1850 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1851 const char *name, size_t size,
1852 size_t align, unsigned long flags,
1853 void (*ctor)(void *, struct kmem_cache *, unsigned long),
1854 void (*dtor)(void *, struct kmem_cache *, unsigned long))
1856 memset(s, 0, kmem_size);
1863 kmem_cache_open_debug_check(s);
1865 if (!calculate_sizes(s))
1870 s->defrag_ratio = 100;
1873 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
1876 if (flags & SLAB_PANIC)
1877 panic("Cannot create slab %s size=%lu realsize=%u "
1878 "order=%u offset=%u flags=%lx\n",
1879 s->name, (unsigned long)size, s->size, s->order,
1883 EXPORT_SYMBOL(kmem_cache_open);
1886 * Check if a given pointer is valid
1888 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
1892 page = get_object_page(object);
1894 if (!page || s != page->slab)
1895 /* No slab or wrong slab */
1898 if (!check_valid_pointer(s, page, object))
1902 * We could also check if the object is on the slabs freelist.
1903 * But this would be too expensive and it seems that the main
1904 * purpose of kmem_ptr_valid is to check if the object belongs
1905 * to a certain slab.
1909 EXPORT_SYMBOL(kmem_ptr_validate);
1912 * Determine the size of a slab object
1914 unsigned int kmem_cache_size(struct kmem_cache *s)
1918 EXPORT_SYMBOL(kmem_cache_size);
1920 const char *kmem_cache_name(struct kmem_cache *s)
1924 EXPORT_SYMBOL(kmem_cache_name);
1927 * Attempt to free all slabs on a node. Return the number of slabs we
1928 * were unable to free.
1930 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
1931 struct list_head *list)
1933 int slabs_inuse = 0;
1934 unsigned long flags;
1935 struct page *page, *h;
1937 spin_lock_irqsave(&n->list_lock, flags);
1938 list_for_each_entry_safe(page, h, list, lru)
1940 list_del(&page->lru);
1941 discard_slab(s, page);
1944 spin_unlock_irqrestore(&n->list_lock, flags);
1949 * Release all resources used by a slab cache.
1951 static int kmem_cache_close(struct kmem_cache *s)
1957 /* Attempt to free all objects */
1958 for_each_online_node(node) {
1959 struct kmem_cache_node *n = get_node(s, node);
1961 n->nr_partial -= free_list(s, n, &n->partial);
1962 if (atomic_long_read(&n->nr_slabs))
1965 free_kmem_cache_nodes(s);
1970 * Close a cache and release the kmem_cache structure
1971 * (must be used for caches created using kmem_cache_create)
1973 void kmem_cache_destroy(struct kmem_cache *s)
1975 down_write(&slub_lock);
1979 if (kmem_cache_close(s))
1981 sysfs_slab_remove(s);
1984 up_write(&slub_lock);
1986 EXPORT_SYMBOL(kmem_cache_destroy);
1988 /********************************************************************
1990 *******************************************************************/
1992 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
1993 EXPORT_SYMBOL(kmalloc_caches);
1995 #ifdef CONFIG_ZONE_DMA
1996 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
1999 static int __init setup_slub_min_order(char *str)
2001 get_option (&str, &slub_min_order);
2006 __setup("slub_min_order=", setup_slub_min_order);
2008 static int __init setup_slub_max_order(char *str)
2010 get_option (&str, &slub_max_order);
2015 __setup("slub_max_order=", setup_slub_max_order);
2017 static int __init setup_slub_min_objects(char *str)
2019 get_option (&str, &slub_min_objects);
2024 __setup("slub_min_objects=", setup_slub_min_objects);
2026 static int __init setup_slub_nomerge(char *str)
2032 __setup("slub_nomerge", setup_slub_nomerge);
2034 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2035 const char *name, int size, gfp_t gfp_flags)
2037 unsigned int flags = 0;
2039 if (gfp_flags & SLUB_DMA)
2040 flags = SLAB_CACHE_DMA;
2042 down_write(&slub_lock);
2043 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2047 list_add(&s->list, &slab_caches);
2048 up_write(&slub_lock);
2049 if (sysfs_slab_add(s))
2054 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2057 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2059 int index = kmalloc_index(size);
2064 /* Allocation too large? */
2067 #ifdef CONFIG_ZONE_DMA
2068 if ((flags & SLUB_DMA)) {
2069 struct kmem_cache *s;
2070 struct kmem_cache *x;
2074 s = kmalloc_caches_dma[index];
2078 /* Dynamically create dma cache */
2079 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2081 panic("Unable to allocate memory for dma cache\n");
2083 if (index <= KMALLOC_SHIFT_HIGH)
2084 realsize = 1 << index;
2092 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2093 (unsigned int)realsize);
2094 s = create_kmalloc_cache(x, text, realsize, flags);
2095 kmalloc_caches_dma[index] = s;
2099 return &kmalloc_caches[index];
2102 void *__kmalloc(size_t size, gfp_t flags)
2104 struct kmem_cache *s = get_slab(size, flags);
2107 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2110 EXPORT_SYMBOL(__kmalloc);
2113 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2115 struct kmem_cache *s = get_slab(size, flags);
2118 return slab_alloc(s, flags, node, __builtin_return_address(0));
2121 EXPORT_SYMBOL(__kmalloc_node);
2124 size_t ksize(const void *object)
2126 struct page *page = get_object_page(object);
2127 struct kmem_cache *s;
2134 * Debugging requires use of the padding between object
2135 * and whatever may come after it.
2137 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2141 * If we have the need to store the freelist pointer
2142 * back there or track user information then we can
2143 * only use the space before that information.
2145 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2149 * Else we can use all the padding etc for the allocation
2153 EXPORT_SYMBOL(ksize);
2155 void kfree(const void *x)
2157 struct kmem_cache *s;
2163 page = virt_to_head_page(x);
2166 slab_free(s, page, (void *)x, __builtin_return_address(0));
2168 EXPORT_SYMBOL(kfree);
2171 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2172 * the remaining slabs by the number of items in use. The slabs with the
2173 * most items in use come first. New allocations will then fill those up
2174 * and thus they can be removed from the partial lists.
2176 * The slabs with the least items are placed last. This results in them
2177 * being allocated from last increasing the chance that the last objects
2178 * are freed in them.
2180 int kmem_cache_shrink(struct kmem_cache *s)
2184 struct kmem_cache_node *n;
2187 struct list_head *slabs_by_inuse =
2188 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2189 unsigned long flags;
2191 if (!slabs_by_inuse)
2195 for_each_online_node(node) {
2196 n = get_node(s, node);
2201 for (i = 0; i < s->objects; i++)
2202 INIT_LIST_HEAD(slabs_by_inuse + i);
2204 spin_lock_irqsave(&n->list_lock, flags);
2207 * Build lists indexed by the items in use in each slab.
2209 * Note that concurrent frees may occur while we hold the
2210 * list_lock. page->inuse here is the upper limit.
2212 list_for_each_entry_safe(page, t, &n->partial, lru) {
2213 if (!page->inuse && slab_trylock(page)) {
2215 * Must hold slab lock here because slab_free
2216 * may have freed the last object and be
2217 * waiting to release the slab.
2219 list_del(&page->lru);
2222 discard_slab(s, page);
2224 if (n->nr_partial > MAX_PARTIAL)
2225 list_move(&page->lru,
2226 slabs_by_inuse + page->inuse);
2230 if (n->nr_partial <= MAX_PARTIAL)
2234 * Rebuild the partial list with the slabs filled up most
2235 * first and the least used slabs at the end.
2237 for (i = s->objects - 1; i >= 0; i--)
2238 list_splice(slabs_by_inuse + i, n->partial.prev);
2241 spin_unlock_irqrestore(&n->list_lock, flags);
2244 kfree(slabs_by_inuse);
2247 EXPORT_SYMBOL(kmem_cache_shrink);
2250 * krealloc - reallocate memory. The contents will remain unchanged.
2252 * @p: object to reallocate memory for.
2253 * @new_size: how many bytes of memory are required.
2254 * @flags: the type of memory to allocate.
2256 * The contents of the object pointed to are preserved up to the
2257 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2258 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2259 * %NULL pointer, the object pointed to is freed.
2261 void *krealloc(const void *p, size_t new_size, gfp_t flags)
2267 return kmalloc(new_size, flags);
2269 if (unlikely(!new_size)) {
2278 ret = kmalloc(new_size, flags);
2280 memcpy(ret, p, min(new_size, ks));
2285 EXPORT_SYMBOL(krealloc);
2287 /********************************************************************
2288 * Basic setup of slabs
2289 *******************************************************************/
2291 void __init kmem_cache_init(void)
2297 * Must first have the slab cache available for the allocations of the
2298 * struct kmem_cache_node's. There is special bootstrap code in
2299 * kmem_cache_open for slab_state == DOWN.
2301 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2302 sizeof(struct kmem_cache_node), GFP_KERNEL);
2305 /* Able to allocate the per node structures */
2306 slab_state = PARTIAL;
2308 /* Caches that are not of the two-to-the-power-of size */
2309 create_kmalloc_cache(&kmalloc_caches[1],
2310 "kmalloc-96", 96, GFP_KERNEL);
2311 create_kmalloc_cache(&kmalloc_caches[2],
2312 "kmalloc-192", 192, GFP_KERNEL);
2314 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2315 create_kmalloc_cache(&kmalloc_caches[i],
2316 "kmalloc", 1 << i, GFP_KERNEL);
2320 /* Provide the correct kmalloc names now that the caches are up */
2321 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2322 kmalloc_caches[i]. name =
2323 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2326 register_cpu_notifier(&slab_notifier);
2329 if (nr_cpu_ids) /* Remove when nr_cpu_ids is fixed upstream ! */
2330 kmem_size = offsetof(struct kmem_cache, cpu_slab)
2331 + nr_cpu_ids * sizeof(struct page *);
2333 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2334 " Processors=%d, Nodes=%d\n",
2335 KMALLOC_SHIFT_HIGH, cache_line_size(),
2336 slub_min_order, slub_max_order, slub_min_objects,
2337 nr_cpu_ids, nr_node_ids);
2341 * Find a mergeable slab cache
2343 static int slab_unmergeable(struct kmem_cache *s)
2345 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2348 if (s->ctor || s->dtor)
2354 static struct kmem_cache *find_mergeable(size_t size,
2355 size_t align, unsigned long flags,
2356 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2357 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2359 struct list_head *h;
2361 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2367 size = ALIGN(size, sizeof(void *));
2368 align = calculate_alignment(flags, align, size);
2369 size = ALIGN(size, align);
2371 list_for_each(h, &slab_caches) {
2372 struct kmem_cache *s =
2373 container_of(h, struct kmem_cache, list);
2375 if (slab_unmergeable(s))
2381 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2382 (s->flags & SLUB_MERGE_SAME))
2385 * Check if alignment is compatible.
2386 * Courtesy of Adrian Drzewiecki
2388 if ((s->size & ~(align -1)) != s->size)
2391 if (s->size - size >= sizeof(void *))
2399 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2400 size_t align, unsigned long flags,
2401 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2402 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2404 struct kmem_cache *s;
2406 down_write(&slub_lock);
2407 s = find_mergeable(size, align, flags, dtor, ctor);
2411 * Adjust the object sizes so that we clear
2412 * the complete object on kzalloc.
2414 s->objsize = max(s->objsize, (int)size);
2415 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2416 if (sysfs_slab_alias(s, name))
2419 s = kmalloc(kmem_size, GFP_KERNEL);
2420 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2421 size, align, flags, ctor, dtor)) {
2422 if (sysfs_slab_add(s)) {
2426 list_add(&s->list, &slab_caches);
2430 up_write(&slub_lock);
2434 up_write(&slub_lock);
2435 if (flags & SLAB_PANIC)
2436 panic("Cannot create slabcache %s\n", name);
2441 EXPORT_SYMBOL(kmem_cache_create);
2443 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2447 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
2449 memset(x, 0, s->objsize);
2452 EXPORT_SYMBOL(kmem_cache_zalloc);
2455 static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2457 struct list_head *h;
2459 down_read(&slub_lock);
2460 list_for_each(h, &slab_caches) {
2461 struct kmem_cache *s =
2462 container_of(h, struct kmem_cache, list);
2466 up_read(&slub_lock);
2470 * Use the cpu notifier to insure that the cpu slabs are flushed when
2473 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2474 unsigned long action, void *hcpu)
2476 long cpu = (long)hcpu;
2479 case CPU_UP_CANCELED:
2481 for_all_slabs(__flush_cpu_slab, cpu);
2489 static struct notifier_block __cpuinitdata slab_notifier =
2490 { &slab_cpuup_callback, NULL, 0 };
2496 /*****************************************************************
2497 * Generic reaper used to support the page allocator
2498 * (the cpu slabs are reaped by a per slab workqueue).
2500 * Maybe move this to the page allocator?
2501 ****************************************************************/
2503 static DEFINE_PER_CPU(unsigned long, reap_node);
2505 static void init_reap_node(int cpu)
2509 node = next_node(cpu_to_node(cpu), node_online_map);
2510 if (node == MAX_NUMNODES)
2511 node = first_node(node_online_map);
2513 __get_cpu_var(reap_node) = node;
2516 static void next_reap_node(void)
2518 int node = __get_cpu_var(reap_node);
2521 * Also drain per cpu pages on remote zones
2523 if (node != numa_node_id())
2524 drain_node_pages(node);
2526 node = next_node(node, node_online_map);
2527 if (unlikely(node >= MAX_NUMNODES))
2528 node = first_node(node_online_map);
2529 __get_cpu_var(reap_node) = node;
2532 #define init_reap_node(cpu) do { } while (0)
2533 #define next_reap_node(void) do { } while (0)
2536 #define REAPTIMEOUT_CPUC (2*HZ)
2539 static DEFINE_PER_CPU(struct delayed_work, reap_work);
2541 static void cache_reap(struct work_struct *unused)
2544 refresh_cpu_vm_stats(smp_processor_id());
2545 schedule_delayed_work(&__get_cpu_var(reap_work),
2549 static void __devinit start_cpu_timer(int cpu)
2551 struct delayed_work *reap_work = &per_cpu(reap_work, cpu);
2554 * When this gets called from do_initcalls via cpucache_init(),
2555 * init_workqueues() has already run, so keventd will be setup
2558 if (keventd_up() && reap_work->work.func == NULL) {
2559 init_reap_node(cpu);
2560 INIT_DELAYED_WORK(reap_work, cache_reap);
2561 schedule_delayed_work_on(cpu, reap_work, HZ + 3 * cpu);
2565 static int __init cpucache_init(void)
2570 * Register the timers that drain pcp pages and update vm statistics
2572 for_each_online_cpu(cpu)
2573 start_cpu_timer(cpu);
2576 __initcall(cpucache_init);
2579 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2581 struct kmem_cache *s = get_slab(size, gfpflags);
2586 return slab_alloc(s, gfpflags, -1, caller);
2589 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2590 int node, void *caller)
2592 struct kmem_cache *s = get_slab(size, gfpflags);
2597 return slab_alloc(s, gfpflags, node, caller);
2600 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2601 static int validate_slab(struct kmem_cache *s, struct page *page)
2604 void *addr = page_address(page);
2605 DECLARE_BITMAP(map, s->objects);
2607 if (!check_slab(s, page) ||
2608 !on_freelist(s, page, NULL))
2611 /* Now we know that a valid freelist exists */
2612 bitmap_zero(map, s->objects);
2614 for_each_free_object(p, s, page->freelist) {
2615 set_bit(slab_index(p, s, addr), map);
2616 if (!check_object(s, page, p, 0))
2620 for_each_object(p, s, addr)
2621 if (!test_bit(slab_index(p, s, addr), map))
2622 if (!check_object(s, page, p, 1))
2627 static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2629 if (slab_trylock(page)) {
2630 validate_slab(s, page);
2633 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2636 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2637 if (!SlabDebug(page))
2638 printk(KERN_ERR "SLUB %s: SlabDebug not set "
2639 "on slab 0x%p\n", s->name, page);
2641 if (SlabDebug(page))
2642 printk(KERN_ERR "SLUB %s: SlabDebug set on "
2643 "slab 0x%p\n", s->name, page);
2647 static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2649 unsigned long count = 0;
2651 unsigned long flags;
2653 spin_lock_irqsave(&n->list_lock, flags);
2655 list_for_each_entry(page, &n->partial, lru) {
2656 validate_slab_slab(s, page);
2659 if (count != n->nr_partial)
2660 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2661 "counter=%ld\n", s->name, count, n->nr_partial);
2663 if (!(s->flags & SLAB_STORE_USER))
2666 list_for_each_entry(page, &n->full, lru) {
2667 validate_slab_slab(s, page);
2670 if (count != atomic_long_read(&n->nr_slabs))
2671 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2672 "counter=%ld\n", s->name, count,
2673 atomic_long_read(&n->nr_slabs));
2676 spin_unlock_irqrestore(&n->list_lock, flags);
2680 static unsigned long validate_slab_cache(struct kmem_cache *s)
2683 unsigned long count = 0;
2686 for_each_online_node(node) {
2687 struct kmem_cache_node *n = get_node(s, node);
2689 count += validate_slab_node(s, n);
2694 #ifdef SLUB_RESILIENCY_TEST
2695 static void resiliency_test(void)
2699 printk(KERN_ERR "SLUB resiliency testing\n");
2700 printk(KERN_ERR "-----------------------\n");
2701 printk(KERN_ERR "A. Corruption after allocation\n");
2703 p = kzalloc(16, GFP_KERNEL);
2705 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2706 " 0x12->0x%p\n\n", p + 16);
2708 validate_slab_cache(kmalloc_caches + 4);
2710 /* Hmmm... The next two are dangerous */
2711 p = kzalloc(32, GFP_KERNEL);
2712 p[32 + sizeof(void *)] = 0x34;
2713 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2714 " 0x34 -> -0x%p\n", p);
2715 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2717 validate_slab_cache(kmalloc_caches + 5);
2718 p = kzalloc(64, GFP_KERNEL);
2719 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2721 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2723 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2724 validate_slab_cache(kmalloc_caches + 6);
2726 printk(KERN_ERR "\nB. Corruption after free\n");
2727 p = kzalloc(128, GFP_KERNEL);
2730 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2731 validate_slab_cache(kmalloc_caches + 7);
2733 p = kzalloc(256, GFP_KERNEL);
2736 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2737 validate_slab_cache(kmalloc_caches + 8);
2739 p = kzalloc(512, GFP_KERNEL);
2742 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2743 validate_slab_cache(kmalloc_caches + 9);
2746 static void resiliency_test(void) {};
2750 * Generate lists of code addresses where slabcache objects are allocated
2755 unsigned long count;
2768 unsigned long count;
2769 struct location *loc;
2772 static void free_loc_track(struct loc_track *t)
2775 free_pages((unsigned long)t->loc,
2776 get_order(sizeof(struct location) * t->max));
2779 static int alloc_loc_track(struct loc_track *t, unsigned long max)
2785 max = PAGE_SIZE / sizeof(struct location);
2787 order = get_order(sizeof(struct location) * max);
2789 l = (void *)__get_free_pages(GFP_KERNEL, order);
2795 memcpy(l, t->loc, sizeof(struct location) * t->count);
2803 static int add_location(struct loc_track *t, struct kmem_cache *s,
2804 const struct track *track)
2806 long start, end, pos;
2809 unsigned long age = jiffies - track->when;
2815 pos = start + (end - start + 1) / 2;
2818 * There is nothing at "end". If we end up there
2819 * we need to add something to before end.
2824 caddr = t->loc[pos].addr;
2825 if (track->addr == caddr) {
2831 if (age < l->min_time)
2833 if (age > l->max_time)
2836 if (track->pid < l->min_pid)
2837 l->min_pid = track->pid;
2838 if (track->pid > l->max_pid)
2839 l->max_pid = track->pid;
2841 cpu_set(track->cpu, l->cpus);
2843 node_set(page_to_nid(virt_to_page(track)), l->nodes);
2847 if (track->addr < caddr)
2854 * Not found. Insert new tracking element.
2856 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max))
2862 (t->count - pos) * sizeof(struct location));
2865 l->addr = track->addr;
2869 l->min_pid = track->pid;
2870 l->max_pid = track->pid;
2871 cpus_clear(l->cpus);
2872 cpu_set(track->cpu, l->cpus);
2873 nodes_clear(l->nodes);
2874 node_set(page_to_nid(virt_to_page(track)), l->nodes);
2878 static void process_slab(struct loc_track *t, struct kmem_cache *s,
2879 struct page *page, enum track_item alloc)
2881 void *addr = page_address(page);
2882 DECLARE_BITMAP(map, s->objects);
2885 bitmap_zero(map, s->objects);
2886 for_each_free_object(p, s, page->freelist)
2887 set_bit(slab_index(p, s, addr), map);
2889 for_each_object(p, s, addr)
2890 if (!test_bit(slab_index(p, s, addr), map))
2891 add_location(t, s, get_track(s, p, alloc));
2894 static int list_locations(struct kmem_cache *s, char *buf,
2895 enum track_item alloc)
2905 /* Push back cpu slabs */
2908 for_each_online_node(node) {
2909 struct kmem_cache_node *n = get_node(s, node);
2910 unsigned long flags;
2913 if (!atomic_read(&n->nr_slabs))
2916 spin_lock_irqsave(&n->list_lock, flags);
2917 list_for_each_entry(page, &n->partial, lru)
2918 process_slab(&t, s, page, alloc);
2919 list_for_each_entry(page, &n->full, lru)
2920 process_slab(&t, s, page, alloc);
2921 spin_unlock_irqrestore(&n->list_lock, flags);
2924 for (i = 0; i < t.count; i++) {
2925 struct location *l = &t.loc[i];
2927 if (n > PAGE_SIZE - 100)
2929 n += sprintf(buf + n, "%7ld ", l->count);
2932 n += sprint_symbol(buf + n, (unsigned long)l->addr);
2934 n += sprintf(buf + n, "<not-available>");
2936 if (l->sum_time != l->min_time) {
2937 unsigned long remainder;
2939 n += sprintf(buf + n, " age=%ld/%ld/%ld",
2941 div_long_long_rem(l->sum_time, l->count, &remainder),
2944 n += sprintf(buf + n, " age=%ld",
2947 if (l->min_pid != l->max_pid)
2948 n += sprintf(buf + n, " pid=%ld-%ld",
2949 l->min_pid, l->max_pid);
2951 n += sprintf(buf + n, " pid=%ld",
2954 if (num_online_cpus() > 1 && !cpus_empty(l->cpus)) {
2955 n += sprintf(buf + n, " cpus=");
2956 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
2960 if (num_online_nodes() > 1 && !nodes_empty(l->nodes)) {
2961 n += sprintf(buf + n, " nodes=");
2962 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
2966 n += sprintf(buf + n, "\n");
2971 n += sprintf(buf, "No data\n");
2975 static unsigned long count_partial(struct kmem_cache_node *n)
2977 unsigned long flags;
2978 unsigned long x = 0;
2981 spin_lock_irqsave(&n->list_lock, flags);
2982 list_for_each_entry(page, &n->partial, lru)
2984 spin_unlock_irqrestore(&n->list_lock, flags);
2988 enum slab_stat_type {
2995 #define SO_FULL (1 << SL_FULL)
2996 #define SO_PARTIAL (1 << SL_PARTIAL)
2997 #define SO_CPU (1 << SL_CPU)
2998 #define SO_OBJECTS (1 << SL_OBJECTS)
3000 static unsigned long slab_objects(struct kmem_cache *s,
3001 char *buf, unsigned long flags)
3003 unsigned long total = 0;
3007 unsigned long *nodes;
3008 unsigned long *per_cpu;
3010 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3011 per_cpu = nodes + nr_node_ids;
3013 for_each_possible_cpu(cpu) {
3014 struct page *page = s->cpu_slab[cpu];
3018 node = page_to_nid(page);
3019 if (flags & SO_CPU) {
3022 if (flags & SO_OBJECTS)
3033 for_each_online_node(node) {
3034 struct kmem_cache_node *n = get_node(s, node);
3036 if (flags & SO_PARTIAL) {
3037 if (flags & SO_OBJECTS)
3038 x = count_partial(n);
3045 if (flags & SO_FULL) {
3046 int full_slabs = atomic_read(&n->nr_slabs)
3050 if (flags & SO_OBJECTS)
3051 x = full_slabs * s->objects;
3059 x = sprintf(buf, "%lu", total);
3061 for_each_online_node(node)
3063 x += sprintf(buf + x, " N%d=%lu",
3067 return x + sprintf(buf + x, "\n");
3070 static int any_slab_objects(struct kmem_cache *s)
3075 for_each_possible_cpu(cpu)
3076 if (s->cpu_slab[cpu])
3079 for_each_node(node) {
3080 struct kmem_cache_node *n = get_node(s, node);
3082 if (n->nr_partial || atomic_read(&n->nr_slabs))
3088 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3089 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3091 struct slab_attribute {
3092 struct attribute attr;
3093 ssize_t (*show)(struct kmem_cache *s, char *buf);
3094 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3097 #define SLAB_ATTR_RO(_name) \
3098 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3100 #define SLAB_ATTR(_name) \
3101 static struct slab_attribute _name##_attr = \
3102 __ATTR(_name, 0644, _name##_show, _name##_store)
3104 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3106 return sprintf(buf, "%d\n", s->size);
3108 SLAB_ATTR_RO(slab_size);
3110 static ssize_t align_show(struct kmem_cache *s, char *buf)
3112 return sprintf(buf, "%d\n", s->align);
3114 SLAB_ATTR_RO(align);
3116 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3118 return sprintf(buf, "%d\n", s->objsize);
3120 SLAB_ATTR_RO(object_size);
3122 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3124 return sprintf(buf, "%d\n", s->objects);
3126 SLAB_ATTR_RO(objs_per_slab);
3128 static ssize_t order_show(struct kmem_cache *s, char *buf)
3130 return sprintf(buf, "%d\n", s->order);
3132 SLAB_ATTR_RO(order);
3134 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3137 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3139 return n + sprintf(buf + n, "\n");
3145 static ssize_t dtor_show(struct kmem_cache *s, char *buf)
3148 int n = sprint_symbol(buf, (unsigned long)s->dtor);
3150 return n + sprintf(buf + n, "\n");
3156 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3158 return sprintf(buf, "%d\n", s->refcount - 1);
3160 SLAB_ATTR_RO(aliases);
3162 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3164 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3166 SLAB_ATTR_RO(slabs);
3168 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3170 return slab_objects(s, buf, SO_PARTIAL);
3172 SLAB_ATTR_RO(partial);
3174 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3176 return slab_objects(s, buf, SO_CPU);
3178 SLAB_ATTR_RO(cpu_slabs);
3180 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3182 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3184 SLAB_ATTR_RO(objects);
3186 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3188 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3191 static ssize_t sanity_checks_store(struct kmem_cache *s,
3192 const char *buf, size_t length)
3194 s->flags &= ~SLAB_DEBUG_FREE;
3196 s->flags |= SLAB_DEBUG_FREE;
3199 SLAB_ATTR(sanity_checks);
3201 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3203 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3206 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3209 s->flags &= ~SLAB_TRACE;
3211 s->flags |= SLAB_TRACE;
3216 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3218 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3221 static ssize_t reclaim_account_store(struct kmem_cache *s,
3222 const char *buf, size_t length)
3224 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3226 s->flags |= SLAB_RECLAIM_ACCOUNT;
3229 SLAB_ATTR(reclaim_account);
3231 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3233 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3235 SLAB_ATTR_RO(hwcache_align);
3237 #ifdef CONFIG_ZONE_DMA
3238 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3240 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3242 SLAB_ATTR_RO(cache_dma);
3245 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3247 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3249 SLAB_ATTR_RO(destroy_by_rcu);
3251 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3253 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3256 static ssize_t red_zone_store(struct kmem_cache *s,
3257 const char *buf, size_t length)
3259 if (any_slab_objects(s))
3262 s->flags &= ~SLAB_RED_ZONE;
3264 s->flags |= SLAB_RED_ZONE;
3268 SLAB_ATTR(red_zone);
3270 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3272 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3275 static ssize_t poison_store(struct kmem_cache *s,
3276 const char *buf, size_t length)
3278 if (any_slab_objects(s))
3281 s->flags &= ~SLAB_POISON;
3283 s->flags |= SLAB_POISON;
3289 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3291 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3294 static ssize_t store_user_store(struct kmem_cache *s,
3295 const char *buf, size_t length)
3297 if (any_slab_objects(s))
3300 s->flags &= ~SLAB_STORE_USER;
3302 s->flags |= SLAB_STORE_USER;
3306 SLAB_ATTR(store_user);
3308 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3313 static ssize_t validate_store(struct kmem_cache *s,
3314 const char *buf, size_t length)
3317 validate_slab_cache(s);
3322 SLAB_ATTR(validate);
3324 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3329 static ssize_t shrink_store(struct kmem_cache *s,
3330 const char *buf, size_t length)
3332 if (buf[0] == '1') {
3333 int rc = kmem_cache_shrink(s);
3343 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3345 if (!(s->flags & SLAB_STORE_USER))
3347 return list_locations(s, buf, TRACK_ALLOC);
3349 SLAB_ATTR_RO(alloc_calls);
3351 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3353 if (!(s->flags & SLAB_STORE_USER))
3355 return list_locations(s, buf, TRACK_FREE);
3357 SLAB_ATTR_RO(free_calls);
3360 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3362 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3365 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3366 const char *buf, size_t length)
3368 int n = simple_strtoul(buf, NULL, 10);
3371 s->defrag_ratio = n * 10;
3374 SLAB_ATTR(defrag_ratio);
3377 static struct attribute * slab_attrs[] = {
3378 &slab_size_attr.attr,
3379 &object_size_attr.attr,
3380 &objs_per_slab_attr.attr,
3385 &cpu_slabs_attr.attr,
3390 &sanity_checks_attr.attr,
3392 &hwcache_align_attr.attr,
3393 &reclaim_account_attr.attr,
3394 &destroy_by_rcu_attr.attr,
3395 &red_zone_attr.attr,
3397 &store_user_attr.attr,
3398 &validate_attr.attr,
3400 &alloc_calls_attr.attr,
3401 &free_calls_attr.attr,
3402 #ifdef CONFIG_ZONE_DMA
3403 &cache_dma_attr.attr,
3406 &defrag_ratio_attr.attr,
3411 static struct attribute_group slab_attr_group = {
3412 .attrs = slab_attrs,
3415 static ssize_t slab_attr_show(struct kobject *kobj,
3416 struct attribute *attr,
3419 struct slab_attribute *attribute;
3420 struct kmem_cache *s;
3423 attribute = to_slab_attr(attr);
3426 if (!attribute->show)
3429 err = attribute->show(s, buf);
3434 static ssize_t slab_attr_store(struct kobject *kobj,
3435 struct attribute *attr,
3436 const char *buf, size_t len)
3438 struct slab_attribute *attribute;
3439 struct kmem_cache *s;
3442 attribute = to_slab_attr(attr);
3445 if (!attribute->store)
3448 err = attribute->store(s, buf, len);
3453 static struct sysfs_ops slab_sysfs_ops = {
3454 .show = slab_attr_show,
3455 .store = slab_attr_store,
3458 static struct kobj_type slab_ktype = {
3459 .sysfs_ops = &slab_sysfs_ops,
3462 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3464 struct kobj_type *ktype = get_ktype(kobj);
3466 if (ktype == &slab_ktype)
3471 static struct kset_uevent_ops slab_uevent_ops = {
3472 .filter = uevent_filter,
3475 decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3477 #define ID_STR_LENGTH 64
3479 /* Create a unique string id for a slab cache:
3481 * :[flags-]size:[memory address of kmemcache]
3483 static char *create_unique_id(struct kmem_cache *s)
3485 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3492 * First flags affecting slabcache operations. We will only
3493 * get here for aliasable slabs so we do not need to support
3494 * too many flags. The flags here must cover all flags that
3495 * are matched during merging to guarantee that the id is
3498 if (s->flags & SLAB_CACHE_DMA)
3500 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3502 if (s->flags & SLAB_DEBUG_FREE)
3506 p += sprintf(p, "%07d", s->size);
3507 BUG_ON(p > name + ID_STR_LENGTH - 1);
3511 static int sysfs_slab_add(struct kmem_cache *s)
3517 if (slab_state < SYSFS)
3518 /* Defer until later */
3521 unmergeable = slab_unmergeable(s);
3524 * Slabcache can never be merged so we can use the name proper.
3525 * This is typically the case for debug situations. In that
3526 * case we can catch duplicate names easily.
3528 sysfs_remove_link(&slab_subsys.kobj, s->name);
3532 * Create a unique name for the slab as a target
3535 name = create_unique_id(s);
3538 kobj_set_kset_s(s, slab_subsys);
3539 kobject_set_name(&s->kobj, name);
3540 kobject_init(&s->kobj);
3541 err = kobject_add(&s->kobj);
3545 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3548 kobject_uevent(&s->kobj, KOBJ_ADD);
3550 /* Setup first alias */
3551 sysfs_slab_alias(s, s->name);
3557 static void sysfs_slab_remove(struct kmem_cache *s)
3559 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3560 kobject_del(&s->kobj);
3564 * Need to buffer aliases during bootup until sysfs becomes
3565 * available lest we loose that information.
3567 struct saved_alias {
3568 struct kmem_cache *s;
3570 struct saved_alias *next;
3573 struct saved_alias *alias_list;
3575 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3577 struct saved_alias *al;
3579 if (slab_state == SYSFS) {
3581 * If we have a leftover link then remove it.
3583 sysfs_remove_link(&slab_subsys.kobj, name);
3584 return sysfs_create_link(&slab_subsys.kobj,
3588 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3594 al->next = alias_list;
3599 static int __init slab_sysfs_init(void)
3601 struct list_head *h;
3604 err = subsystem_register(&slab_subsys);
3606 printk(KERN_ERR "Cannot register slab subsystem.\n");
3612 list_for_each(h, &slab_caches) {
3613 struct kmem_cache *s =
3614 container_of(h, struct kmem_cache, list);
3616 err = sysfs_slab_add(s);
3620 while (alias_list) {
3621 struct saved_alias *al = alias_list;
3623 alias_list = alias_list->next;
3624 err = sysfs_slab_alias(al->s, al->name);
3633 __initcall(slab_sysfs_init);