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 frozen and exempt from list processing.
82 * This means that the slab is dedicated to a purpose
83 * such as satisfying allocations for a specific
84 * processor. Objects may be freed in the slab while
85 * it is frozen but slab_free will then skip the usual
86 * list operations. It is up to the processor holding
87 * the slab to integrate the slab into the slab lists
88 * when the slab is no longer needed.
90 * One use of this flag is to mark slabs that are
91 * used for allocations. Then such a slab becomes a cpu
92 * slab. The cpu slab may be equipped with an additional
93 * lockless_freelist that allows lockless access to
94 * free objects in addition to the regular freelist
95 * that requires the slab lock.
97 * PageError Slab requires special handling due to debug
98 * options set. This moves slab handling out of
99 * the fast path and disables lockless freelists.
102 #define FROZEN (1 << PG_active)
104 #ifdef CONFIG_SLUB_DEBUG
105 #define SLABDEBUG (1 << PG_error)
110 static inline int SlabFrozen(struct page *page)
112 return page->flags & FROZEN;
115 static inline void SetSlabFrozen(struct page *page)
117 page->flags |= FROZEN;
120 static inline void ClearSlabFrozen(struct page *page)
122 page->flags &= ~FROZEN;
125 static inline int SlabDebug(struct page *page)
127 return page->flags & SLABDEBUG;
130 static inline void SetSlabDebug(struct page *page)
132 page->flags |= SLABDEBUG;
135 static inline void ClearSlabDebug(struct page *page)
137 page->flags &= ~SLABDEBUG;
141 * Issues still to be resolved:
143 * - The per cpu array is updated for each new slab and and is a remote
144 * cacheline for most nodes. This could become a bouncing cacheline given
145 * enough frequent updates. There are 16 pointers in a cacheline, so at
146 * max 16 cpus could compete for the cacheline which may be okay.
148 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
150 * - Variable sizing of the per node arrays
153 /* Enable to test recovery from slab corruption on boot */
154 #undef SLUB_RESILIENCY_TEST
159 * Small page size. Make sure that we do not fragment memory
161 #define DEFAULT_MAX_ORDER 1
162 #define DEFAULT_MIN_OBJECTS 4
167 * Large page machines are customarily able to handle larger
170 #define DEFAULT_MAX_ORDER 2
171 #define DEFAULT_MIN_OBJECTS 8
176 * Mininum number of partial slabs. These will be left on the partial
177 * lists even if they are empty. kmem_cache_shrink may reclaim them.
179 #define MIN_PARTIAL 2
182 * Maximum number of desirable partial slabs.
183 * The existence of more partial slabs makes kmem_cache_shrink
184 * sort the partial list by the number of objects in the.
186 #define MAX_PARTIAL 10
188 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
189 SLAB_POISON | SLAB_STORE_USER)
192 * Set of flags that will prevent slab merging
194 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
195 SLAB_TRACE | SLAB_DESTROY_BY_RCU)
197 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
200 #ifndef ARCH_KMALLOC_MINALIGN
201 #define ARCH_KMALLOC_MINALIGN __alignof__(unsigned long long)
204 #ifndef ARCH_SLAB_MINALIGN
205 #define ARCH_SLAB_MINALIGN __alignof__(unsigned long long)
208 /* Internal SLUB flags */
209 #define __OBJECT_POISON 0x80000000 /* Poison object */
211 /* Not all arches define cache_line_size */
212 #ifndef cache_line_size
213 #define cache_line_size() L1_CACHE_BYTES
216 static int kmem_size = sizeof(struct kmem_cache);
219 static struct notifier_block slab_notifier;
223 DOWN, /* No slab functionality available */
224 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
225 UP, /* Everything works but does not show up in sysfs */
229 /* A list of all slab caches on the system */
230 static DECLARE_RWSEM(slub_lock);
231 LIST_HEAD(slab_caches);
234 * Tracking user of a slab.
237 void *addr; /* Called from address */
238 int cpu; /* Was running on cpu */
239 int pid; /* Pid context */
240 unsigned long when; /* When did the operation occur */
243 enum track_item { TRACK_ALLOC, TRACK_FREE };
245 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
246 static int sysfs_slab_add(struct kmem_cache *);
247 static int sysfs_slab_alias(struct kmem_cache *, const char *);
248 static void sysfs_slab_remove(struct kmem_cache *);
250 static int sysfs_slab_add(struct kmem_cache *s) { return 0; }
251 static int sysfs_slab_alias(struct kmem_cache *s, const char *p) { return 0; }
252 static void sysfs_slab_remove(struct kmem_cache *s) {}
255 /********************************************************************
256 * Core slab cache functions
257 *******************************************************************/
259 int slab_is_available(void)
261 return slab_state >= UP;
264 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
267 return s->node[node];
269 return &s->local_node;
273 static inline int check_valid_pointer(struct kmem_cache *s,
274 struct page *page, const void *object)
281 base = page_address(page);
282 if (object < base || object >= base + s->objects * s->size ||
283 (object - base) % s->size) {
291 * Slow version of get and set free pointer.
293 * This version requires touching the cache lines of kmem_cache which
294 * we avoid to do in the fast alloc free paths. There we obtain the offset
295 * from the page struct.
297 static inline void *get_freepointer(struct kmem_cache *s, void *object)
299 return *(void **)(object + s->offset);
302 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
304 *(void **)(object + s->offset) = fp;
307 /* Loop over all objects in a slab */
308 #define for_each_object(__p, __s, __addr) \
309 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
313 #define for_each_free_object(__p, __s, __free) \
314 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
316 /* Determine object index from a given position */
317 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
319 return (p - addr) / s->size;
322 #ifdef CONFIG_SLUB_DEBUG
326 static int slub_debug;
328 static char *slub_debug_slabs;
333 static void print_section(char *text, u8 *addr, unsigned int length)
341 for (i = 0; i < length; i++) {
343 printk(KERN_ERR "%10s 0x%p: ", text, addr + i);
346 printk(" %02x", addr[i]);
348 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
350 printk(" %s\n",ascii);
361 printk(" %s\n", ascii);
365 static struct track *get_track(struct kmem_cache *s, void *object,
366 enum track_item alloc)
371 p = object + s->offset + sizeof(void *);
373 p = object + s->inuse;
378 static void set_track(struct kmem_cache *s, void *object,
379 enum track_item alloc, void *addr)
384 p = object + s->offset + sizeof(void *);
386 p = object + s->inuse;
391 p->cpu = smp_processor_id();
392 p->pid = current ? current->pid : -1;
395 memset(p, 0, sizeof(struct track));
398 static void init_tracking(struct kmem_cache *s, void *object)
400 if (s->flags & SLAB_STORE_USER) {
401 set_track(s, object, TRACK_FREE, NULL);
402 set_track(s, object, TRACK_ALLOC, NULL);
406 static void print_track(const char *s, struct track *t)
411 printk(KERN_ERR "%s: ", s);
412 __print_symbol("%s", (unsigned long)t->addr);
413 printk(" jiffies_ago=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
416 static void print_trailer(struct kmem_cache *s, u8 *p)
418 unsigned int off; /* Offset of last byte */
420 if (s->flags & SLAB_RED_ZONE)
421 print_section("Redzone", p + s->objsize,
422 s->inuse - s->objsize);
424 printk(KERN_ERR "FreePointer 0x%p -> 0x%p\n",
426 get_freepointer(s, p));
429 off = s->offset + sizeof(void *);
433 if (s->flags & SLAB_STORE_USER) {
434 print_track("Last alloc", get_track(s, p, TRACK_ALLOC));
435 print_track("Last free ", get_track(s, p, TRACK_FREE));
436 off += 2 * sizeof(struct track);
440 /* Beginning of the filler is the free pointer */
441 print_section("Filler", p + off, s->size - off);
444 static void object_err(struct kmem_cache *s, struct page *page,
445 u8 *object, char *reason)
447 u8 *addr = page_address(page);
449 printk(KERN_ERR "*** SLUB %s: %s@0x%p slab 0x%p\n",
450 s->name, reason, object, page);
451 printk(KERN_ERR " offset=%tu flags=0x%04lx inuse=%u freelist=0x%p\n",
452 object - addr, page->flags, page->inuse, page->freelist);
453 if (object > addr + 16)
454 print_section("Bytes b4", object - 16, 16);
455 print_section("Object", object, min(s->objsize, 128));
456 print_trailer(s, object);
460 static void slab_err(struct kmem_cache *s, struct page *page, char *reason, ...)
465 va_start(args, reason);
466 vsnprintf(buf, sizeof(buf), reason, args);
468 printk(KERN_ERR "*** SLUB %s: %s in slab @0x%p\n", s->name, buf,
473 static void init_object(struct kmem_cache *s, void *object, int active)
477 if (s->flags & __OBJECT_POISON) {
478 memset(p, POISON_FREE, s->objsize - 1);
479 p[s->objsize -1] = POISON_END;
482 if (s->flags & SLAB_RED_ZONE)
483 memset(p + s->objsize,
484 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
485 s->inuse - s->objsize);
488 static int check_bytes(u8 *start, unsigned int value, unsigned int bytes)
491 if (*start != (u8)value)
503 * Bytes of the object to be managed.
504 * If the freepointer may overlay the object then the free
505 * pointer is the first word of the object.
507 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
510 * object + s->objsize
511 * Padding to reach word boundary. This is also used for Redzoning.
512 * Padding is extended by another word if Redzoning is enabled and
515 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
516 * 0xcc (RED_ACTIVE) for objects in use.
519 * Meta data starts here.
521 * A. Free pointer (if we cannot overwrite object on free)
522 * B. Tracking data for SLAB_STORE_USER
523 * C. Padding to reach required alignment boundary or at mininum
524 * one word if debuggin is on to be able to detect writes
525 * before the word boundary.
527 * Padding is done using 0x5a (POISON_INUSE)
530 * Nothing is used beyond s->size.
532 * If slabcaches are merged then the objsize and inuse boundaries are mostly
533 * ignored. And therefore no slab options that rely on these boundaries
534 * may be used with merged slabcaches.
537 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
538 void *from, void *to)
540 printk(KERN_ERR "@@@ SLUB %s: Restoring %s (0x%x) from 0x%p-0x%p\n",
541 s->name, message, data, from, to - 1);
542 memset(from, data, to - from);
545 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
547 unsigned long off = s->inuse; /* The end of info */
550 /* Freepointer is placed after the object. */
551 off += sizeof(void *);
553 if (s->flags & SLAB_STORE_USER)
554 /* We also have user information there */
555 off += 2 * sizeof(struct track);
560 if (check_bytes(p + off, POISON_INUSE, s->size - off))
563 object_err(s, page, p, "Object padding check fails");
568 restore_bytes(s, "object padding", POISON_INUSE, p + off, p + s->size);
572 static int slab_pad_check(struct kmem_cache *s, struct page *page)
575 int length, remainder;
577 if (!(s->flags & SLAB_POISON))
580 p = page_address(page);
581 length = s->objects * s->size;
582 remainder = (PAGE_SIZE << s->order) - length;
586 if (!check_bytes(p + length, POISON_INUSE, remainder)) {
587 slab_err(s, page, "Padding check failed");
588 restore_bytes(s, "slab padding", POISON_INUSE, p + length,
589 p + length + remainder);
595 static int check_object(struct kmem_cache *s, struct page *page,
596 void *object, int active)
599 u8 *endobject = object + s->objsize;
601 if (s->flags & SLAB_RED_ZONE) {
603 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
605 if (!check_bytes(endobject, red, s->inuse - s->objsize)) {
606 object_err(s, page, object,
607 active ? "Redzone Active" : "Redzone Inactive");
608 restore_bytes(s, "redzone", red,
609 endobject, object + s->inuse);
613 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse &&
614 !check_bytes(endobject, POISON_INUSE,
615 s->inuse - s->objsize)) {
616 object_err(s, page, p, "Alignment padding check fails");
618 * Fix it so that there will not be another report.
620 * Hmmm... We may be corrupting an object that now expects
621 * to be longer than allowed.
623 restore_bytes(s, "alignment padding", POISON_INUSE,
624 endobject, object + s->inuse);
628 if (s->flags & SLAB_POISON) {
629 if (!active && (s->flags & __OBJECT_POISON) &&
630 (!check_bytes(p, POISON_FREE, s->objsize - 1) ||
631 p[s->objsize - 1] != POISON_END)) {
633 object_err(s, page, p, "Poison check failed");
634 restore_bytes(s, "Poison", POISON_FREE,
635 p, p + s->objsize -1);
636 restore_bytes(s, "Poison", POISON_END,
637 p + s->objsize - 1, p + s->objsize);
641 * check_pad_bytes cleans up on its own.
643 check_pad_bytes(s, page, p);
646 if (!s->offset && active)
648 * Object and freepointer overlap. Cannot check
649 * freepointer while object is allocated.
653 /* Check free pointer validity */
654 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
655 object_err(s, page, p, "Freepointer corrupt");
657 * No choice but to zap it and thus loose the remainder
658 * of the free objects in this slab. May cause
659 * another error because the object count is now wrong.
661 set_freepointer(s, p, NULL);
667 static int check_slab(struct kmem_cache *s, struct page *page)
669 VM_BUG_ON(!irqs_disabled());
671 if (!PageSlab(page)) {
672 slab_err(s, page, "Not a valid slab page flags=%lx "
673 "mapping=0x%p count=%d", page->flags, page->mapping,
677 if (page->offset * sizeof(void *) != s->offset) {
678 slab_err(s, page, "Corrupted offset %lu flags=0x%lx "
679 "mapping=0x%p count=%d",
680 (unsigned long)(page->offset * sizeof(void *)),
686 if (page->inuse > s->objects) {
687 slab_err(s, page, "inuse %u > max %u @0x%p flags=%lx "
688 "mapping=0x%p count=%d",
689 s->name, page->inuse, s->objects, page->flags,
690 page->mapping, page_count(page));
693 /* Slab_pad_check fixes things up after itself */
694 slab_pad_check(s, page);
699 * Determine if a certain object on a page is on the freelist. Must hold the
700 * slab lock to guarantee that the chains are in a consistent state.
702 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
705 void *fp = page->freelist;
708 while (fp && nr <= s->objects) {
711 if (!check_valid_pointer(s, page, fp)) {
713 object_err(s, page, object,
714 "Freechain corrupt");
715 set_freepointer(s, object, NULL);
718 slab_err(s, page, "Freepointer 0x%p corrupt",
720 page->freelist = NULL;
721 page->inuse = s->objects;
722 printk(KERN_ERR "@@@ SLUB %s: Freelist "
723 "cleared. Slab 0x%p\n",
730 fp = get_freepointer(s, object);
734 if (page->inuse != s->objects - nr) {
735 slab_err(s, page, "Wrong object count. Counter is %d but "
736 "counted were %d", s, page, page->inuse,
738 page->inuse = s->objects - nr;
739 printk(KERN_ERR "@@@ SLUB %s: Object count adjusted. "
740 "Slab @0x%p\n", s->name, page);
742 return search == NULL;
746 * Tracking of fully allocated slabs for debugging purposes.
748 static void add_full(struct kmem_cache_node *n, struct page *page)
750 spin_lock(&n->list_lock);
751 list_add(&page->lru, &n->full);
752 spin_unlock(&n->list_lock);
755 static void remove_full(struct kmem_cache *s, struct page *page)
757 struct kmem_cache_node *n;
759 if (!(s->flags & SLAB_STORE_USER))
762 n = get_node(s, page_to_nid(page));
764 spin_lock(&n->list_lock);
765 list_del(&page->lru);
766 spin_unlock(&n->list_lock);
769 static int alloc_object_checks(struct kmem_cache *s, struct page *page,
772 if (!check_slab(s, page))
775 if (object && !on_freelist(s, page, object)) {
776 slab_err(s, page, "Object 0x%p already allocated", object);
780 if (!check_valid_pointer(s, page, object)) {
781 object_err(s, page, object, "Freelist Pointer check fails");
788 if (!check_object(s, page, object, 0))
793 if (PageSlab(page)) {
795 * If this is a slab page then lets do the best we can
796 * to avoid issues in the future. Marking all objects
797 * as used avoids touching the remaining objects.
799 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p. Marking all objects used.\n",
801 page->inuse = s->objects;
802 page->freelist = NULL;
803 /* Fix up fields that may be corrupted */
804 page->offset = s->offset / sizeof(void *);
809 static int free_object_checks(struct kmem_cache *s, struct page *page,
812 if (!check_slab(s, page))
815 if (!check_valid_pointer(s, page, object)) {
816 slab_err(s, page, "Invalid object pointer 0x%p", object);
820 if (on_freelist(s, page, object)) {
821 slab_err(s, page, "Object 0x%p already free", object);
825 if (!check_object(s, page, object, 1))
828 if (unlikely(s != page->slab)) {
830 slab_err(s, page, "Attempt to free object(0x%p) "
831 "outside of slab", object);
835 "SLUB <none>: no slab for object 0x%p.\n",
840 slab_err(s, page, "object at 0x%p belongs "
841 "to slab %s", object, page->slab->name);
846 printk(KERN_ERR "@@@ SLUB: %s slab 0x%p object at 0x%p not freed.\n",
847 s->name, page, object);
851 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
853 if (s->flags & SLAB_TRACE) {
854 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
856 alloc ? "alloc" : "free",
861 print_section("Object", (void *)object, s->objsize);
867 static int __init setup_slub_debug(char *str)
869 if (!str || *str != '=')
870 slub_debug = DEBUG_DEFAULT_FLAGS;
873 if (*str == 0 || *str == ',')
874 slub_debug = DEBUG_DEFAULT_FLAGS;
876 for( ;*str && *str != ','; str++)
878 case 'f' : case 'F' :
879 slub_debug |= SLAB_DEBUG_FREE;
881 case 'z' : case 'Z' :
882 slub_debug |= SLAB_RED_ZONE;
884 case 'p' : case 'P' :
885 slub_debug |= SLAB_POISON;
887 case 'u' : case 'U' :
888 slub_debug |= SLAB_STORE_USER;
890 case 't' : case 'T' :
891 slub_debug |= SLAB_TRACE;
894 printk(KERN_ERR "slub_debug option '%c' "
895 "unknown. skipped\n",*str);
900 slub_debug_slabs = str + 1;
904 __setup("slub_debug", setup_slub_debug);
906 static void kmem_cache_open_debug_check(struct kmem_cache *s)
909 * The page->offset field is only 16 bit wide. This is an offset
910 * in units of words from the beginning of an object. If the slab
911 * size is bigger then we cannot move the free pointer behind the
914 * On 32 bit platforms the limit is 256k. On 64bit platforms
917 * Debugging or ctor may create a need to move the free
918 * pointer. Fail if this happens.
920 if (s->size >= 65535 * sizeof(void *)) {
921 BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON |
922 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
927 * Enable debugging if selected on the kernel commandline.
929 if (slub_debug && (!slub_debug_slabs ||
930 strncmp(slub_debug_slabs, s->name,
931 strlen(slub_debug_slabs)) == 0))
932 s->flags |= slub_debug;
936 static inline int alloc_object_checks(struct kmem_cache *s,
937 struct page *page, void *object) { return 0; }
939 static inline int free_object_checks(struct kmem_cache *s,
940 struct page *page, void *object) { return 0; }
942 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
943 static inline void remove_full(struct kmem_cache *s, struct page *page) {}
944 static inline void trace(struct kmem_cache *s, struct page *page,
945 void *object, int alloc) {}
946 static inline void init_object(struct kmem_cache *s,
947 void *object, int active) {}
948 static inline void init_tracking(struct kmem_cache *s, void *object) {}
949 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
951 static inline int check_object(struct kmem_cache *s, struct page *page,
952 void *object, int active) { return 1; }
953 static inline void set_track(struct kmem_cache *s, void *object,
954 enum track_item alloc, void *addr) {}
955 static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {}
959 * Slab allocation and freeing
961 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
964 int pages = 1 << s->order;
969 if (s->flags & SLAB_CACHE_DMA)
973 page = alloc_pages(flags, s->order);
975 page = alloc_pages_node(node, flags, s->order);
980 mod_zone_page_state(page_zone(page),
981 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
982 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
988 static void setup_object(struct kmem_cache *s, struct page *page,
991 if (SlabDebug(page)) {
992 init_object(s, object, 0);
993 init_tracking(s, object);
996 if (unlikely(s->ctor))
997 s->ctor(object, s, SLAB_CTOR_CONSTRUCTOR);
1000 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1003 struct kmem_cache_node *n;
1009 BUG_ON(flags & ~(GFP_DMA | GFP_LEVEL_MASK));
1011 if (flags & __GFP_WAIT)
1014 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
1018 n = get_node(s, page_to_nid(page));
1020 atomic_long_inc(&n->nr_slabs);
1021 page->offset = s->offset / sizeof(void *);
1023 page->flags |= 1 << PG_slab;
1024 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1025 SLAB_STORE_USER | SLAB_TRACE))
1028 start = page_address(page);
1029 end = start + s->objects * s->size;
1031 if (unlikely(s->flags & SLAB_POISON))
1032 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1035 for_each_object(p, s, start) {
1036 setup_object(s, page, last);
1037 set_freepointer(s, last, p);
1040 setup_object(s, page, last);
1041 set_freepointer(s, last, NULL);
1043 page->freelist = start;
1044 page->lockless_freelist = NULL;
1047 if (flags & __GFP_WAIT)
1048 local_irq_disable();
1052 static void __free_slab(struct kmem_cache *s, struct page *page)
1054 int pages = 1 << s->order;
1056 if (unlikely(SlabDebug(page))) {
1059 slab_pad_check(s, page);
1060 for_each_object(p, s, page_address(page))
1061 check_object(s, page, p, 0);
1064 mod_zone_page_state(page_zone(page),
1065 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1066 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1069 page->mapping = NULL;
1070 __free_pages(page, s->order);
1073 static void rcu_free_slab(struct rcu_head *h)
1077 page = container_of((struct list_head *)h, struct page, lru);
1078 __free_slab(page->slab, page);
1081 static void free_slab(struct kmem_cache *s, struct page *page)
1083 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1085 * RCU free overloads the RCU head over the LRU
1087 struct rcu_head *head = (void *)&page->lru;
1089 call_rcu(head, rcu_free_slab);
1091 __free_slab(s, page);
1094 static void discard_slab(struct kmem_cache *s, struct page *page)
1096 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1098 atomic_long_dec(&n->nr_slabs);
1099 reset_page_mapcount(page);
1100 ClearSlabDebug(page);
1101 __ClearPageSlab(page);
1106 * Per slab locking using the pagelock
1108 static __always_inline void slab_lock(struct page *page)
1110 bit_spin_lock(PG_locked, &page->flags);
1113 static __always_inline void slab_unlock(struct page *page)
1115 bit_spin_unlock(PG_locked, &page->flags);
1118 static __always_inline int slab_trylock(struct page *page)
1122 rc = bit_spin_trylock(PG_locked, &page->flags);
1127 * Management of partially allocated slabs
1129 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1131 spin_lock(&n->list_lock);
1133 list_add_tail(&page->lru, &n->partial);
1134 spin_unlock(&n->list_lock);
1137 static void add_partial(struct kmem_cache_node *n, struct page *page)
1139 spin_lock(&n->list_lock);
1141 list_add(&page->lru, &n->partial);
1142 spin_unlock(&n->list_lock);
1145 static void remove_partial(struct kmem_cache *s,
1148 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1150 spin_lock(&n->list_lock);
1151 list_del(&page->lru);
1153 spin_unlock(&n->list_lock);
1157 * Lock slab and remove from the partial list.
1159 * Must hold list_lock.
1161 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1163 if (slab_trylock(page)) {
1164 list_del(&page->lru);
1166 SetSlabFrozen(page);
1173 * Try to allocate a partial slab from a specific node.
1175 static struct page *get_partial_node(struct kmem_cache_node *n)
1180 * Racy check. If we mistakenly see no partial slabs then we
1181 * just allocate an empty slab. If we mistakenly try to get a
1182 * partial slab and there is none available then get_partials()
1185 if (!n || !n->nr_partial)
1188 spin_lock(&n->list_lock);
1189 list_for_each_entry(page, &n->partial, lru)
1190 if (lock_and_freeze_slab(n, page))
1194 spin_unlock(&n->list_lock);
1199 * Get a page from somewhere. Search in increasing NUMA distances.
1201 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1204 struct zonelist *zonelist;
1209 * The defrag ratio allows a configuration of the tradeoffs between
1210 * inter node defragmentation and node local allocations. A lower
1211 * defrag_ratio increases the tendency to do local allocations
1212 * instead of attempting to obtain partial slabs from other nodes.
1214 * If the defrag_ratio is set to 0 then kmalloc() always
1215 * returns node local objects. If the ratio is higher then kmalloc()
1216 * may return off node objects because partial slabs are obtained
1217 * from other nodes and filled up.
1219 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1220 * defrag_ratio = 1000) then every (well almost) allocation will
1221 * first attempt to defrag slab caches on other nodes. This means
1222 * scanning over all nodes to look for partial slabs which may be
1223 * expensive if we do it every time we are trying to find a slab
1224 * with available objects.
1226 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1229 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1230 ->node_zonelists[gfp_zone(flags)];
1231 for (z = zonelist->zones; *z; z++) {
1232 struct kmem_cache_node *n;
1234 n = get_node(s, zone_to_nid(*z));
1236 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1237 n->nr_partial > MIN_PARTIAL) {
1238 page = get_partial_node(n);
1248 * Get a partial page, lock it and return it.
1250 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1253 int searchnode = (node == -1) ? numa_node_id() : node;
1255 page = get_partial_node(get_node(s, searchnode));
1256 if (page || (flags & __GFP_THISNODE))
1259 return get_any_partial(s, flags);
1263 * Move a page back to the lists.
1265 * Must be called with the slab lock held.
1267 * On exit the slab lock will have been dropped.
1269 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1271 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1273 ClearSlabFrozen(page);
1277 add_partial(n, page);
1278 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1283 if (n->nr_partial < MIN_PARTIAL) {
1285 * Adding an empty slab to the partial slabs in order
1286 * to avoid page allocator overhead. This slab needs
1287 * to come after the other slabs with objects in
1288 * order to fill them up. That way the size of the
1289 * partial list stays small. kmem_cache_shrink can
1290 * reclaim empty slabs from the partial list.
1292 add_partial_tail(n, page);
1296 discard_slab(s, page);
1302 * Remove the cpu slab
1304 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1307 * Merge cpu freelist into freelist. Typically we get here
1308 * because both freelists are empty. So this is unlikely
1311 while (unlikely(page->lockless_freelist)) {
1314 /* Retrieve object from cpu_freelist */
1315 object = page->lockless_freelist;
1316 page->lockless_freelist = page->lockless_freelist[page->offset];
1318 /* And put onto the regular freelist */
1319 object[page->offset] = page->freelist;
1320 page->freelist = object;
1323 s->cpu_slab[cpu] = NULL;
1324 unfreeze_slab(s, page);
1327 static void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1330 deactivate_slab(s, page, cpu);
1335 * Called from IPI handler with interrupts disabled.
1337 static void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1339 struct page *page = s->cpu_slab[cpu];
1342 flush_slab(s, page, cpu);
1345 static void flush_cpu_slab(void *d)
1347 struct kmem_cache *s = d;
1348 int cpu = smp_processor_id();
1350 __flush_cpu_slab(s, cpu);
1353 static void flush_all(struct kmem_cache *s)
1356 on_each_cpu(flush_cpu_slab, s, 1, 1);
1358 unsigned long flags;
1360 local_irq_save(flags);
1362 local_irq_restore(flags);
1367 * Slow path. The lockless freelist is empty or we need to perform
1370 * Interrupts are disabled.
1372 * Processing is still very fast if new objects have been freed to the
1373 * regular freelist. In that case we simply take over the regular freelist
1374 * as the lockless freelist and zap the regular freelist.
1376 * If that is not working then we fall back to the partial lists. We take the
1377 * first element of the freelist as the object to allocate now and move the
1378 * rest of the freelist to the lockless freelist.
1380 * And if we were unable to get a new slab from the partial slab lists then
1381 * we need to allocate a new slab. This is slowest path since we may sleep.
1383 static void *__slab_alloc(struct kmem_cache *s,
1384 gfp_t gfpflags, int node, void *addr, struct page *page)
1387 int cpu = smp_processor_id();
1393 if (unlikely(node != -1 && page_to_nid(page) != node))
1396 object = page->freelist;
1397 if (unlikely(!object))
1399 if (unlikely(SlabDebug(page)))
1402 object = page->freelist;
1403 page->lockless_freelist = object[page->offset];
1404 page->inuse = s->objects;
1405 page->freelist = NULL;
1410 deactivate_slab(s, page, cpu);
1413 page = get_partial(s, gfpflags, node);
1415 s->cpu_slab[cpu] = page;
1419 page = new_slab(s, gfpflags, node);
1421 cpu = smp_processor_id();
1422 if (s->cpu_slab[cpu]) {
1424 * Someone else populated the cpu_slab while we
1425 * enabled interrupts, or we have gotten scheduled
1426 * on another cpu. The page may not be on the
1427 * requested node even if __GFP_THISNODE was
1428 * specified. So we need to recheck.
1431 page_to_nid(s->cpu_slab[cpu]) == node) {
1433 * Current cpuslab is acceptable and we
1434 * want the current one since its cache hot
1436 discard_slab(s, page);
1437 page = s->cpu_slab[cpu];
1441 /* New slab does not fit our expectations */
1442 flush_slab(s, s->cpu_slab[cpu], cpu);
1445 SetSlabFrozen(page);
1446 s->cpu_slab[cpu] = page;
1451 object = page->freelist;
1452 if (!alloc_object_checks(s, page, object))
1454 if (s->flags & SLAB_STORE_USER)
1455 set_track(s, object, TRACK_ALLOC, addr);
1456 trace(s, page, object, 1);
1457 init_object(s, object, 1);
1460 page->freelist = object[page->offset];
1466 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1467 * have the fastpath folded into their functions. So no function call
1468 * overhead for requests that can be satisfied on the fastpath.
1470 * The fastpath works by first checking if the lockless freelist can be used.
1471 * If not then __slab_alloc is called for slow processing.
1473 * Otherwise we can simply pick the next object from the lockless free list.
1475 static void __always_inline *slab_alloc(struct kmem_cache *s,
1476 gfp_t gfpflags, int node, void *addr)
1480 unsigned long flags;
1482 local_irq_save(flags);
1483 page = s->cpu_slab[smp_processor_id()];
1484 if (unlikely(!page || !page->lockless_freelist ||
1485 (node != -1 && page_to_nid(page) != node)))
1487 object = __slab_alloc(s, gfpflags, node, addr, page);
1490 object = page->lockless_freelist;
1491 page->lockless_freelist = object[page->offset];
1493 local_irq_restore(flags);
1497 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1499 return slab_alloc(s, gfpflags, -1, __builtin_return_address(0));
1501 EXPORT_SYMBOL(kmem_cache_alloc);
1504 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1506 return slab_alloc(s, gfpflags, node, __builtin_return_address(0));
1508 EXPORT_SYMBOL(kmem_cache_alloc_node);
1512 * Slow patch handling. This may still be called frequently since objects
1513 * have a longer lifetime than the cpu slabs in most processing loads.
1515 * So we still attempt to reduce cache line usage. Just take the slab
1516 * lock and free the item. If there is no additional partial page
1517 * handling required then we can return immediately.
1519 static void __slab_free(struct kmem_cache *s, struct page *page,
1520 void *x, void *addr)
1523 void **object = (void *)x;
1527 if (unlikely(SlabDebug(page)))
1530 prior = object[page->offset] = page->freelist;
1531 page->freelist = object;
1534 if (unlikely(SlabFrozen(page)))
1537 if (unlikely(!page->inuse))
1541 * Objects left in the slab. If it
1542 * was not on the partial list before
1545 if (unlikely(!prior))
1546 add_partial(get_node(s, page_to_nid(page)), page);
1555 * Slab still on the partial list.
1557 remove_partial(s, page);
1560 discard_slab(s, page);
1564 if (!free_object_checks(s, page, x))
1566 if (!SlabFrozen(page) && !page->freelist)
1567 remove_full(s, page);
1568 if (s->flags & SLAB_STORE_USER)
1569 set_track(s, x, TRACK_FREE, addr);
1570 trace(s, page, object, 0);
1571 init_object(s, object, 0);
1576 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1577 * can perform fastpath freeing without additional function calls.
1579 * The fastpath is only possible if we are freeing to the current cpu slab
1580 * of this processor. This typically the case if we have just allocated
1583 * If fastpath is not possible then fall back to __slab_free where we deal
1584 * with all sorts of special processing.
1586 static void __always_inline slab_free(struct kmem_cache *s,
1587 struct page *page, void *x, void *addr)
1589 void **object = (void *)x;
1590 unsigned long flags;
1592 local_irq_save(flags);
1593 if (likely(page == s->cpu_slab[smp_processor_id()] &&
1594 !SlabDebug(page))) {
1595 object[page->offset] = page->lockless_freelist;
1596 page->lockless_freelist = object;
1598 __slab_free(s, page, x, addr);
1600 local_irq_restore(flags);
1603 void kmem_cache_free(struct kmem_cache *s, void *x)
1607 page = virt_to_head_page(x);
1609 slab_free(s, page, x, __builtin_return_address(0));
1611 EXPORT_SYMBOL(kmem_cache_free);
1613 /* Figure out on which slab object the object resides */
1614 static struct page *get_object_page(const void *x)
1616 struct page *page = virt_to_head_page(x);
1618 if (!PageSlab(page))
1625 * Object placement in a slab is made very easy because we always start at
1626 * offset 0. If we tune the size of the object to the alignment then we can
1627 * get the required alignment by putting one properly sized object after
1630 * Notice that the allocation order determines the sizes of the per cpu
1631 * caches. Each processor has always one slab available for allocations.
1632 * Increasing the allocation order reduces the number of times that slabs
1633 * must be moved on and off the partial lists and is therefore a factor in
1638 * Mininum / Maximum order of slab pages. This influences locking overhead
1639 * and slab fragmentation. A higher order reduces the number of partial slabs
1640 * and increases the number of allocations possible without having to
1641 * take the list_lock.
1643 static int slub_min_order;
1644 static int slub_max_order = DEFAULT_MAX_ORDER;
1645 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1648 * Merge control. If this is set then no merging of slab caches will occur.
1649 * (Could be removed. This was introduced to pacify the merge skeptics.)
1651 static int slub_nomerge;
1654 * Calculate the order of allocation given an slab object size.
1656 * The order of allocation has significant impact on performance and other
1657 * system components. Generally order 0 allocations should be preferred since
1658 * order 0 does not cause fragmentation in the page allocator. Larger objects
1659 * be problematic to put into order 0 slabs because there may be too much
1660 * unused space left. We go to a higher order if more than 1/8th of the slab
1663 * In order to reach satisfactory performance we must ensure that a minimum
1664 * number of objects is in one slab. Otherwise we may generate too much
1665 * activity on the partial lists which requires taking the list_lock. This is
1666 * less a concern for large slabs though which are rarely used.
1668 * slub_max_order specifies the order where we begin to stop considering the
1669 * number of objects in a slab as critical. If we reach slub_max_order then
1670 * we try to keep the page order as low as possible. So we accept more waste
1671 * of space in favor of a small page order.
1673 * Higher order allocations also allow the placement of more objects in a
1674 * slab and thereby reduce object handling overhead. If the user has
1675 * requested a higher mininum order then we start with that one instead of
1676 * the smallest order which will fit the object.
1678 static inline int slab_order(int size, int min_objects,
1679 int max_order, int fract_leftover)
1684 for (order = max(slub_min_order,
1685 fls(min_objects * size - 1) - PAGE_SHIFT);
1686 order <= max_order; order++) {
1688 unsigned long slab_size = PAGE_SIZE << order;
1690 if (slab_size < min_objects * size)
1693 rem = slab_size % size;
1695 if (rem <= slab_size / fract_leftover)
1703 static inline int calculate_order(int size)
1710 * Attempt to find best configuration for a slab. This
1711 * works by first attempting to generate a layout with
1712 * the best configuration and backing off gradually.
1714 * First we reduce the acceptable waste in a slab. Then
1715 * we reduce the minimum objects required in a slab.
1717 min_objects = slub_min_objects;
1718 while (min_objects > 1) {
1720 while (fraction >= 4) {
1721 order = slab_order(size, min_objects,
1722 slub_max_order, fraction);
1723 if (order <= slub_max_order)
1731 * We were unable to place multiple objects in a slab. Now
1732 * lets see if we can place a single object there.
1734 order = slab_order(size, 1, slub_max_order, 1);
1735 if (order <= slub_max_order)
1739 * Doh this slab cannot be placed using slub_max_order.
1741 order = slab_order(size, 1, MAX_ORDER, 1);
1742 if (order <= MAX_ORDER)
1748 * Figure out what the alignment of the objects will be.
1750 static unsigned long calculate_alignment(unsigned long flags,
1751 unsigned long align, unsigned long size)
1754 * If the user wants hardware cache aligned objects then
1755 * follow that suggestion if the object is sufficiently
1758 * The hardware cache alignment cannot override the
1759 * specified alignment though. If that is greater
1762 if ((flags & SLAB_HWCACHE_ALIGN) &&
1763 size > cache_line_size() / 2)
1764 return max_t(unsigned long, align, cache_line_size());
1766 if (align < ARCH_SLAB_MINALIGN)
1767 return ARCH_SLAB_MINALIGN;
1769 return ALIGN(align, sizeof(void *));
1772 static void init_kmem_cache_node(struct kmem_cache_node *n)
1775 atomic_long_set(&n->nr_slabs, 0);
1776 spin_lock_init(&n->list_lock);
1777 INIT_LIST_HEAD(&n->partial);
1778 INIT_LIST_HEAD(&n->full);
1783 * No kmalloc_node yet so do it by hand. We know that this is the first
1784 * slab on the node for this slabcache. There are no concurrent accesses
1787 * Note that this function only works on the kmalloc_node_cache
1788 * when allocating for the kmalloc_node_cache.
1790 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1794 struct kmem_cache_node *n;
1796 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1798 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1799 /* new_slab() disables interupts */
1805 page->freelist = get_freepointer(kmalloc_caches, n);
1807 kmalloc_caches->node[node] = n;
1808 init_object(kmalloc_caches, n, 1);
1809 init_kmem_cache_node(n);
1810 atomic_long_inc(&n->nr_slabs);
1811 add_partial(n, page);
1815 static void free_kmem_cache_nodes(struct kmem_cache *s)
1819 for_each_online_node(node) {
1820 struct kmem_cache_node *n = s->node[node];
1821 if (n && n != &s->local_node)
1822 kmem_cache_free(kmalloc_caches, n);
1823 s->node[node] = NULL;
1827 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1832 if (slab_state >= UP)
1833 local_node = page_to_nid(virt_to_page(s));
1837 for_each_online_node(node) {
1838 struct kmem_cache_node *n;
1840 if (local_node == node)
1843 if (slab_state == DOWN) {
1844 n = early_kmem_cache_node_alloc(gfpflags,
1848 n = kmem_cache_alloc_node(kmalloc_caches,
1852 free_kmem_cache_nodes(s);
1858 init_kmem_cache_node(n);
1863 static void free_kmem_cache_nodes(struct kmem_cache *s)
1867 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1869 init_kmem_cache_node(&s->local_node);
1875 * calculate_sizes() determines the order and the distribution of data within
1878 static int calculate_sizes(struct kmem_cache *s)
1880 unsigned long flags = s->flags;
1881 unsigned long size = s->objsize;
1882 unsigned long align = s->align;
1885 * Determine if we can poison the object itself. If the user of
1886 * the slab may touch the object after free or before allocation
1887 * then we should never poison the object itself.
1889 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1891 s->flags |= __OBJECT_POISON;
1893 s->flags &= ~__OBJECT_POISON;
1896 * Round up object size to the next word boundary. We can only
1897 * place the free pointer at word boundaries and this determines
1898 * the possible location of the free pointer.
1900 size = ALIGN(size, sizeof(void *));
1902 #ifdef CONFIG_SLUB_DEBUG
1904 * If we are Redzoning then check if there is some space between the
1905 * end of the object and the free pointer. If not then add an
1906 * additional word to have some bytes to store Redzone information.
1908 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1909 size += sizeof(void *);
1913 * With that we have determined the number of bytes in actual use
1914 * by the object. This is the potential offset to the free pointer.
1918 #ifdef CONFIG_SLUB_DEBUG
1919 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
1922 * Relocate free pointer after the object if it is not
1923 * permitted to overwrite the first word of the object on
1926 * This is the case if we do RCU, have a constructor or
1927 * destructor or are poisoning the objects.
1930 size += sizeof(void *);
1933 if (flags & SLAB_STORE_USER)
1935 * Need to store information about allocs and frees after
1938 size += 2 * sizeof(struct track);
1940 if (flags & SLAB_RED_ZONE)
1942 * Add some empty padding so that we can catch
1943 * overwrites from earlier objects rather than let
1944 * tracking information or the free pointer be
1945 * corrupted if an user writes before the start
1948 size += sizeof(void *);
1952 * Determine the alignment based on various parameters that the
1953 * user specified and the dynamic determination of cache line size
1956 align = calculate_alignment(flags, align, s->objsize);
1959 * SLUB stores one object immediately after another beginning from
1960 * offset 0. In order to align the objects we have to simply size
1961 * each object to conform to the alignment.
1963 size = ALIGN(size, align);
1966 s->order = calculate_order(size);
1971 * Determine the number of objects per slab
1973 s->objects = (PAGE_SIZE << s->order) / size;
1976 * Verify that the number of objects is within permitted limits.
1977 * The page->inuse field is only 16 bit wide! So we cannot have
1978 * more than 64k objects per slab.
1980 if (!s->objects || s->objects > 65535)
1986 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
1987 const char *name, size_t size,
1988 size_t align, unsigned long flags,
1989 void (*ctor)(void *, struct kmem_cache *, unsigned long))
1991 memset(s, 0, kmem_size);
1997 kmem_cache_open_debug_check(s);
1999 if (!calculate_sizes(s))
2004 s->defrag_ratio = 100;
2007 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2010 if (flags & SLAB_PANIC)
2011 panic("Cannot create slab %s size=%lu realsize=%u "
2012 "order=%u offset=%u flags=%lx\n",
2013 s->name, (unsigned long)size, s->size, s->order,
2017 EXPORT_SYMBOL(kmem_cache_open);
2020 * Check if a given pointer is valid
2022 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2026 page = get_object_page(object);
2028 if (!page || s != page->slab)
2029 /* No slab or wrong slab */
2032 if (!check_valid_pointer(s, page, object))
2036 * We could also check if the object is on the slabs freelist.
2037 * But this would be too expensive and it seems that the main
2038 * purpose of kmem_ptr_valid is to check if the object belongs
2039 * to a certain slab.
2043 EXPORT_SYMBOL(kmem_ptr_validate);
2046 * Determine the size of a slab object
2048 unsigned int kmem_cache_size(struct kmem_cache *s)
2052 EXPORT_SYMBOL(kmem_cache_size);
2054 const char *kmem_cache_name(struct kmem_cache *s)
2058 EXPORT_SYMBOL(kmem_cache_name);
2061 * Attempt to free all slabs on a node. Return the number of slabs we
2062 * were unable to free.
2064 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2065 struct list_head *list)
2067 int slabs_inuse = 0;
2068 unsigned long flags;
2069 struct page *page, *h;
2071 spin_lock_irqsave(&n->list_lock, flags);
2072 list_for_each_entry_safe(page, h, list, lru)
2074 list_del(&page->lru);
2075 discard_slab(s, page);
2078 spin_unlock_irqrestore(&n->list_lock, flags);
2083 * Release all resources used by a slab cache.
2085 static int kmem_cache_close(struct kmem_cache *s)
2091 /* Attempt to free all objects */
2092 for_each_online_node(node) {
2093 struct kmem_cache_node *n = get_node(s, node);
2095 n->nr_partial -= free_list(s, n, &n->partial);
2096 if (atomic_long_read(&n->nr_slabs))
2099 free_kmem_cache_nodes(s);
2104 * Close a cache and release the kmem_cache structure
2105 * (must be used for caches created using kmem_cache_create)
2107 void kmem_cache_destroy(struct kmem_cache *s)
2109 down_write(&slub_lock);
2113 if (kmem_cache_close(s))
2115 sysfs_slab_remove(s);
2118 up_write(&slub_lock);
2120 EXPORT_SYMBOL(kmem_cache_destroy);
2122 /********************************************************************
2124 *******************************************************************/
2126 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
2127 EXPORT_SYMBOL(kmalloc_caches);
2129 #ifdef CONFIG_ZONE_DMA
2130 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
2133 static int __init setup_slub_min_order(char *str)
2135 get_option (&str, &slub_min_order);
2140 __setup("slub_min_order=", setup_slub_min_order);
2142 static int __init setup_slub_max_order(char *str)
2144 get_option (&str, &slub_max_order);
2149 __setup("slub_max_order=", setup_slub_max_order);
2151 static int __init setup_slub_min_objects(char *str)
2153 get_option (&str, &slub_min_objects);
2158 __setup("slub_min_objects=", setup_slub_min_objects);
2160 static int __init setup_slub_nomerge(char *str)
2166 __setup("slub_nomerge", setup_slub_nomerge);
2168 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2169 const char *name, int size, gfp_t gfp_flags)
2171 unsigned int flags = 0;
2173 if (gfp_flags & SLUB_DMA)
2174 flags = SLAB_CACHE_DMA;
2176 down_write(&slub_lock);
2177 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2181 list_add(&s->list, &slab_caches);
2182 up_write(&slub_lock);
2183 if (sysfs_slab_add(s))
2188 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2191 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2193 int index = kmalloc_index(size);
2198 /* Allocation too large? */
2201 #ifdef CONFIG_ZONE_DMA
2202 if ((flags & SLUB_DMA)) {
2203 struct kmem_cache *s;
2204 struct kmem_cache *x;
2208 s = kmalloc_caches_dma[index];
2212 /* Dynamically create dma cache */
2213 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2215 panic("Unable to allocate memory for dma cache\n");
2217 if (index <= KMALLOC_SHIFT_HIGH)
2218 realsize = 1 << index;
2226 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2227 (unsigned int)realsize);
2228 s = create_kmalloc_cache(x, text, realsize, flags);
2229 kmalloc_caches_dma[index] = s;
2233 return &kmalloc_caches[index];
2236 void *__kmalloc(size_t size, gfp_t flags)
2238 struct kmem_cache *s = get_slab(size, flags);
2241 return slab_alloc(s, flags, -1, __builtin_return_address(0));
2244 EXPORT_SYMBOL(__kmalloc);
2247 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2249 struct kmem_cache *s = get_slab(size, flags);
2252 return slab_alloc(s, flags, node, __builtin_return_address(0));
2255 EXPORT_SYMBOL(__kmalloc_node);
2258 size_t ksize(const void *object)
2260 struct page *page = get_object_page(object);
2261 struct kmem_cache *s;
2268 * Debugging requires use of the padding between object
2269 * and whatever may come after it.
2271 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2275 * If we have the need to store the freelist pointer
2276 * back there or track user information then we can
2277 * only use the space before that information.
2279 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2283 * Else we can use all the padding etc for the allocation
2287 EXPORT_SYMBOL(ksize);
2289 void kfree(const void *x)
2291 struct kmem_cache *s;
2297 page = virt_to_head_page(x);
2300 slab_free(s, page, (void *)x, __builtin_return_address(0));
2302 EXPORT_SYMBOL(kfree);
2305 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2306 * the remaining slabs by the number of items in use. The slabs with the
2307 * most items in use come first. New allocations will then fill those up
2308 * and thus they can be removed from the partial lists.
2310 * The slabs with the least items are placed last. This results in them
2311 * being allocated from last increasing the chance that the last objects
2312 * are freed in them.
2314 int kmem_cache_shrink(struct kmem_cache *s)
2318 struct kmem_cache_node *n;
2321 struct list_head *slabs_by_inuse =
2322 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2323 unsigned long flags;
2325 if (!slabs_by_inuse)
2329 for_each_online_node(node) {
2330 n = get_node(s, node);
2335 for (i = 0; i < s->objects; i++)
2336 INIT_LIST_HEAD(slabs_by_inuse + i);
2338 spin_lock_irqsave(&n->list_lock, flags);
2341 * Build lists indexed by the items in use in each slab.
2343 * Note that concurrent frees may occur while we hold the
2344 * list_lock. page->inuse here is the upper limit.
2346 list_for_each_entry_safe(page, t, &n->partial, lru) {
2347 if (!page->inuse && slab_trylock(page)) {
2349 * Must hold slab lock here because slab_free
2350 * may have freed the last object and be
2351 * waiting to release the slab.
2353 list_del(&page->lru);
2356 discard_slab(s, page);
2358 if (n->nr_partial > MAX_PARTIAL)
2359 list_move(&page->lru,
2360 slabs_by_inuse + page->inuse);
2364 if (n->nr_partial <= MAX_PARTIAL)
2368 * Rebuild the partial list with the slabs filled up most
2369 * first and the least used slabs at the end.
2371 for (i = s->objects - 1; i >= 0; i--)
2372 list_splice(slabs_by_inuse + i, n->partial.prev);
2375 spin_unlock_irqrestore(&n->list_lock, flags);
2378 kfree(slabs_by_inuse);
2381 EXPORT_SYMBOL(kmem_cache_shrink);
2384 * krealloc - reallocate memory. The contents will remain unchanged.
2385 * @p: object to reallocate memory for.
2386 * @new_size: how many bytes of memory are required.
2387 * @flags: the type of memory to allocate.
2389 * The contents of the object pointed to are preserved up to the
2390 * lesser of the new and old sizes. If @p is %NULL, krealloc()
2391 * behaves exactly like kmalloc(). If @size is 0 and @p is not a
2392 * %NULL pointer, the object pointed to is freed.
2394 void *krealloc(const void *p, size_t new_size, gfp_t flags)
2400 return kmalloc(new_size, flags);
2402 if (unlikely(!new_size)) {
2411 ret = kmalloc(new_size, flags);
2413 memcpy(ret, p, min(new_size, ks));
2418 EXPORT_SYMBOL(krealloc);
2420 /********************************************************************
2421 * Basic setup of slabs
2422 *******************************************************************/
2424 void __init kmem_cache_init(void)
2430 * Must first have the slab cache available for the allocations of the
2431 * struct kmem_cache_node's. There is special bootstrap code in
2432 * kmem_cache_open for slab_state == DOWN.
2434 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2435 sizeof(struct kmem_cache_node), GFP_KERNEL);
2438 /* Able to allocate the per node structures */
2439 slab_state = PARTIAL;
2441 /* Caches that are not of the two-to-the-power-of size */
2442 create_kmalloc_cache(&kmalloc_caches[1],
2443 "kmalloc-96", 96, GFP_KERNEL);
2444 create_kmalloc_cache(&kmalloc_caches[2],
2445 "kmalloc-192", 192, GFP_KERNEL);
2447 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2448 create_kmalloc_cache(&kmalloc_caches[i],
2449 "kmalloc", 1 << i, GFP_KERNEL);
2453 /* Provide the correct kmalloc names now that the caches are up */
2454 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2455 kmalloc_caches[i]. name =
2456 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2459 register_cpu_notifier(&slab_notifier);
2462 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2463 nr_cpu_ids * sizeof(struct page *);
2465 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2466 " Processors=%d, Nodes=%d\n",
2467 KMALLOC_SHIFT_HIGH, cache_line_size(),
2468 slub_min_order, slub_max_order, slub_min_objects,
2469 nr_cpu_ids, nr_node_ids);
2473 * Find a mergeable slab cache
2475 static int slab_unmergeable(struct kmem_cache *s)
2477 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2486 static struct kmem_cache *find_mergeable(size_t size,
2487 size_t align, unsigned long flags,
2488 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2490 struct list_head *h;
2492 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2498 size = ALIGN(size, sizeof(void *));
2499 align = calculate_alignment(flags, align, size);
2500 size = ALIGN(size, align);
2502 list_for_each(h, &slab_caches) {
2503 struct kmem_cache *s =
2504 container_of(h, struct kmem_cache, list);
2506 if (slab_unmergeable(s))
2512 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2513 (s->flags & SLUB_MERGE_SAME))
2516 * Check if alignment is compatible.
2517 * Courtesy of Adrian Drzewiecki
2519 if ((s->size & ~(align -1)) != s->size)
2522 if (s->size - size >= sizeof(void *))
2530 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2531 size_t align, unsigned long flags,
2532 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2533 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2535 struct kmem_cache *s;
2538 down_write(&slub_lock);
2539 s = find_mergeable(size, align, flags, ctor);
2543 * Adjust the object sizes so that we clear
2544 * the complete object on kzalloc.
2546 s->objsize = max(s->objsize, (int)size);
2547 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2548 if (sysfs_slab_alias(s, name))
2551 s = kmalloc(kmem_size, GFP_KERNEL);
2552 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2553 size, align, flags, ctor)) {
2554 if (sysfs_slab_add(s)) {
2558 list_add(&s->list, &slab_caches);
2562 up_write(&slub_lock);
2566 up_write(&slub_lock);
2567 if (flags & SLAB_PANIC)
2568 panic("Cannot create slabcache %s\n", name);
2573 EXPORT_SYMBOL(kmem_cache_create);
2575 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2579 x = slab_alloc(s, flags, -1, __builtin_return_address(0));
2581 memset(x, 0, s->objsize);
2584 EXPORT_SYMBOL(kmem_cache_zalloc);
2587 static void for_all_slabs(void (*func)(struct kmem_cache *, int), int cpu)
2589 struct list_head *h;
2591 down_read(&slub_lock);
2592 list_for_each(h, &slab_caches) {
2593 struct kmem_cache *s =
2594 container_of(h, struct kmem_cache, list);
2598 up_read(&slub_lock);
2602 * Use the cpu notifier to insure that the cpu slabs are flushed when
2605 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2606 unsigned long action, void *hcpu)
2608 long cpu = (long)hcpu;
2611 case CPU_UP_CANCELED:
2612 case CPU_UP_CANCELED_FROZEN:
2614 case CPU_DEAD_FROZEN:
2615 for_all_slabs(__flush_cpu_slab, cpu);
2623 static struct notifier_block __cpuinitdata slab_notifier =
2624 { &slab_cpuup_callback, NULL, 0 };
2628 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2630 struct kmem_cache *s = get_slab(size, gfpflags);
2635 return slab_alloc(s, gfpflags, -1, caller);
2638 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2639 int node, void *caller)
2641 struct kmem_cache *s = get_slab(size, gfpflags);
2646 return slab_alloc(s, gfpflags, node, caller);
2649 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2650 static int validate_slab(struct kmem_cache *s, struct page *page)
2653 void *addr = page_address(page);
2654 DECLARE_BITMAP(map, s->objects);
2656 if (!check_slab(s, page) ||
2657 !on_freelist(s, page, NULL))
2660 /* Now we know that a valid freelist exists */
2661 bitmap_zero(map, s->objects);
2663 for_each_free_object(p, s, page->freelist) {
2664 set_bit(slab_index(p, s, addr), map);
2665 if (!check_object(s, page, p, 0))
2669 for_each_object(p, s, addr)
2670 if (!test_bit(slab_index(p, s, addr), map))
2671 if (!check_object(s, page, p, 1))
2676 static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2678 if (slab_trylock(page)) {
2679 validate_slab(s, page);
2682 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2685 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2686 if (!SlabDebug(page))
2687 printk(KERN_ERR "SLUB %s: SlabDebug not set "
2688 "on slab 0x%p\n", s->name, page);
2690 if (SlabDebug(page))
2691 printk(KERN_ERR "SLUB %s: SlabDebug set on "
2692 "slab 0x%p\n", s->name, page);
2696 static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2698 unsigned long count = 0;
2700 unsigned long flags;
2702 spin_lock_irqsave(&n->list_lock, flags);
2704 list_for_each_entry(page, &n->partial, lru) {
2705 validate_slab_slab(s, page);
2708 if (count != n->nr_partial)
2709 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2710 "counter=%ld\n", s->name, count, n->nr_partial);
2712 if (!(s->flags & SLAB_STORE_USER))
2715 list_for_each_entry(page, &n->full, lru) {
2716 validate_slab_slab(s, page);
2719 if (count != atomic_long_read(&n->nr_slabs))
2720 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2721 "counter=%ld\n", s->name, count,
2722 atomic_long_read(&n->nr_slabs));
2725 spin_unlock_irqrestore(&n->list_lock, flags);
2729 static unsigned long validate_slab_cache(struct kmem_cache *s)
2732 unsigned long count = 0;
2735 for_each_online_node(node) {
2736 struct kmem_cache_node *n = get_node(s, node);
2738 count += validate_slab_node(s, n);
2743 #ifdef SLUB_RESILIENCY_TEST
2744 static void resiliency_test(void)
2748 printk(KERN_ERR "SLUB resiliency testing\n");
2749 printk(KERN_ERR "-----------------------\n");
2750 printk(KERN_ERR "A. Corruption after allocation\n");
2752 p = kzalloc(16, GFP_KERNEL);
2754 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2755 " 0x12->0x%p\n\n", p + 16);
2757 validate_slab_cache(kmalloc_caches + 4);
2759 /* Hmmm... The next two are dangerous */
2760 p = kzalloc(32, GFP_KERNEL);
2761 p[32 + sizeof(void *)] = 0x34;
2762 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2763 " 0x34 -> -0x%p\n", p);
2764 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2766 validate_slab_cache(kmalloc_caches + 5);
2767 p = kzalloc(64, GFP_KERNEL);
2768 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2770 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2772 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2773 validate_slab_cache(kmalloc_caches + 6);
2775 printk(KERN_ERR "\nB. Corruption after free\n");
2776 p = kzalloc(128, GFP_KERNEL);
2779 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2780 validate_slab_cache(kmalloc_caches + 7);
2782 p = kzalloc(256, GFP_KERNEL);
2785 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2786 validate_slab_cache(kmalloc_caches + 8);
2788 p = kzalloc(512, GFP_KERNEL);
2791 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2792 validate_slab_cache(kmalloc_caches + 9);
2795 static void resiliency_test(void) {};
2799 * Generate lists of code addresses where slabcache objects are allocated
2804 unsigned long count;
2817 unsigned long count;
2818 struct location *loc;
2821 static void free_loc_track(struct loc_track *t)
2824 free_pages((unsigned long)t->loc,
2825 get_order(sizeof(struct location) * t->max));
2828 static int alloc_loc_track(struct loc_track *t, unsigned long max)
2834 max = PAGE_SIZE / sizeof(struct location);
2836 order = get_order(sizeof(struct location) * max);
2838 l = (void *)__get_free_pages(GFP_KERNEL, order);
2844 memcpy(l, t->loc, sizeof(struct location) * t->count);
2852 static int add_location(struct loc_track *t, struct kmem_cache *s,
2853 const struct track *track)
2855 long start, end, pos;
2858 unsigned long age = jiffies - track->when;
2864 pos = start + (end - start + 1) / 2;
2867 * There is nothing at "end". If we end up there
2868 * we need to add something to before end.
2873 caddr = t->loc[pos].addr;
2874 if (track->addr == caddr) {
2880 if (age < l->min_time)
2882 if (age > l->max_time)
2885 if (track->pid < l->min_pid)
2886 l->min_pid = track->pid;
2887 if (track->pid > l->max_pid)
2888 l->max_pid = track->pid;
2890 cpu_set(track->cpu, l->cpus);
2892 node_set(page_to_nid(virt_to_page(track)), l->nodes);
2896 if (track->addr < caddr)
2903 * Not found. Insert new tracking element.
2905 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max))
2911 (t->count - pos) * sizeof(struct location));
2914 l->addr = track->addr;
2918 l->min_pid = track->pid;
2919 l->max_pid = track->pid;
2920 cpus_clear(l->cpus);
2921 cpu_set(track->cpu, l->cpus);
2922 nodes_clear(l->nodes);
2923 node_set(page_to_nid(virt_to_page(track)), l->nodes);
2927 static void process_slab(struct loc_track *t, struct kmem_cache *s,
2928 struct page *page, enum track_item alloc)
2930 void *addr = page_address(page);
2931 DECLARE_BITMAP(map, s->objects);
2934 bitmap_zero(map, s->objects);
2935 for_each_free_object(p, s, page->freelist)
2936 set_bit(slab_index(p, s, addr), map);
2938 for_each_object(p, s, addr)
2939 if (!test_bit(slab_index(p, s, addr), map))
2940 add_location(t, s, get_track(s, p, alloc));
2943 static int list_locations(struct kmem_cache *s, char *buf,
2944 enum track_item alloc)
2954 /* Push back cpu slabs */
2957 for_each_online_node(node) {
2958 struct kmem_cache_node *n = get_node(s, node);
2959 unsigned long flags;
2962 if (!atomic_read(&n->nr_slabs))
2965 spin_lock_irqsave(&n->list_lock, flags);
2966 list_for_each_entry(page, &n->partial, lru)
2967 process_slab(&t, s, page, alloc);
2968 list_for_each_entry(page, &n->full, lru)
2969 process_slab(&t, s, page, alloc);
2970 spin_unlock_irqrestore(&n->list_lock, flags);
2973 for (i = 0; i < t.count; i++) {
2974 struct location *l = &t.loc[i];
2976 if (n > PAGE_SIZE - 100)
2978 n += sprintf(buf + n, "%7ld ", l->count);
2981 n += sprint_symbol(buf + n, (unsigned long)l->addr);
2983 n += sprintf(buf + n, "<not-available>");
2985 if (l->sum_time != l->min_time) {
2986 unsigned long remainder;
2988 n += sprintf(buf + n, " age=%ld/%ld/%ld",
2990 div_long_long_rem(l->sum_time, l->count, &remainder),
2993 n += sprintf(buf + n, " age=%ld",
2996 if (l->min_pid != l->max_pid)
2997 n += sprintf(buf + n, " pid=%ld-%ld",
2998 l->min_pid, l->max_pid);
3000 n += sprintf(buf + n, " pid=%ld",
3003 if (num_online_cpus() > 1 && !cpus_empty(l->cpus)) {
3004 n += sprintf(buf + n, " cpus=");
3005 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3009 if (num_online_nodes() > 1 && !nodes_empty(l->nodes)) {
3010 n += sprintf(buf + n, " nodes=");
3011 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3015 n += sprintf(buf + n, "\n");
3020 n += sprintf(buf, "No data\n");
3024 static unsigned long count_partial(struct kmem_cache_node *n)
3026 unsigned long flags;
3027 unsigned long x = 0;
3030 spin_lock_irqsave(&n->list_lock, flags);
3031 list_for_each_entry(page, &n->partial, lru)
3033 spin_unlock_irqrestore(&n->list_lock, flags);
3037 enum slab_stat_type {
3044 #define SO_FULL (1 << SL_FULL)
3045 #define SO_PARTIAL (1 << SL_PARTIAL)
3046 #define SO_CPU (1 << SL_CPU)
3047 #define SO_OBJECTS (1 << SL_OBJECTS)
3049 static unsigned long slab_objects(struct kmem_cache *s,
3050 char *buf, unsigned long flags)
3052 unsigned long total = 0;
3056 unsigned long *nodes;
3057 unsigned long *per_cpu;
3059 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3060 per_cpu = nodes + nr_node_ids;
3062 for_each_possible_cpu(cpu) {
3063 struct page *page = s->cpu_slab[cpu];
3067 node = page_to_nid(page);
3068 if (flags & SO_CPU) {
3071 if (flags & SO_OBJECTS)
3082 for_each_online_node(node) {
3083 struct kmem_cache_node *n = get_node(s, node);
3085 if (flags & SO_PARTIAL) {
3086 if (flags & SO_OBJECTS)
3087 x = count_partial(n);
3094 if (flags & SO_FULL) {
3095 int full_slabs = atomic_read(&n->nr_slabs)
3099 if (flags & SO_OBJECTS)
3100 x = full_slabs * s->objects;
3108 x = sprintf(buf, "%lu", total);
3110 for_each_online_node(node)
3112 x += sprintf(buf + x, " N%d=%lu",
3116 return x + sprintf(buf + x, "\n");
3119 static int any_slab_objects(struct kmem_cache *s)
3124 for_each_possible_cpu(cpu)
3125 if (s->cpu_slab[cpu])
3128 for_each_node(node) {
3129 struct kmem_cache_node *n = get_node(s, node);
3131 if (n->nr_partial || atomic_read(&n->nr_slabs))
3137 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3138 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3140 struct slab_attribute {
3141 struct attribute attr;
3142 ssize_t (*show)(struct kmem_cache *s, char *buf);
3143 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3146 #define SLAB_ATTR_RO(_name) \
3147 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3149 #define SLAB_ATTR(_name) \
3150 static struct slab_attribute _name##_attr = \
3151 __ATTR(_name, 0644, _name##_show, _name##_store)
3153 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3155 return sprintf(buf, "%d\n", s->size);
3157 SLAB_ATTR_RO(slab_size);
3159 static ssize_t align_show(struct kmem_cache *s, char *buf)
3161 return sprintf(buf, "%d\n", s->align);
3163 SLAB_ATTR_RO(align);
3165 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3167 return sprintf(buf, "%d\n", s->objsize);
3169 SLAB_ATTR_RO(object_size);
3171 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3173 return sprintf(buf, "%d\n", s->objects);
3175 SLAB_ATTR_RO(objs_per_slab);
3177 static ssize_t order_show(struct kmem_cache *s, char *buf)
3179 return sprintf(buf, "%d\n", s->order);
3181 SLAB_ATTR_RO(order);
3183 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3186 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3188 return n + sprintf(buf + n, "\n");
3194 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3196 return sprintf(buf, "%d\n", s->refcount - 1);
3198 SLAB_ATTR_RO(aliases);
3200 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3202 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3204 SLAB_ATTR_RO(slabs);
3206 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3208 return slab_objects(s, buf, SO_PARTIAL);
3210 SLAB_ATTR_RO(partial);
3212 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3214 return slab_objects(s, buf, SO_CPU);
3216 SLAB_ATTR_RO(cpu_slabs);
3218 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3220 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3222 SLAB_ATTR_RO(objects);
3224 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3226 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3229 static ssize_t sanity_checks_store(struct kmem_cache *s,
3230 const char *buf, size_t length)
3232 s->flags &= ~SLAB_DEBUG_FREE;
3234 s->flags |= SLAB_DEBUG_FREE;
3237 SLAB_ATTR(sanity_checks);
3239 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3241 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3244 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3247 s->flags &= ~SLAB_TRACE;
3249 s->flags |= SLAB_TRACE;
3254 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3256 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3259 static ssize_t reclaim_account_store(struct kmem_cache *s,
3260 const char *buf, size_t length)
3262 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3264 s->flags |= SLAB_RECLAIM_ACCOUNT;
3267 SLAB_ATTR(reclaim_account);
3269 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3271 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3273 SLAB_ATTR_RO(hwcache_align);
3275 #ifdef CONFIG_ZONE_DMA
3276 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3278 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3280 SLAB_ATTR_RO(cache_dma);
3283 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3285 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3287 SLAB_ATTR_RO(destroy_by_rcu);
3289 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3291 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3294 static ssize_t red_zone_store(struct kmem_cache *s,
3295 const char *buf, size_t length)
3297 if (any_slab_objects(s))
3300 s->flags &= ~SLAB_RED_ZONE;
3302 s->flags |= SLAB_RED_ZONE;
3306 SLAB_ATTR(red_zone);
3308 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3310 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3313 static ssize_t poison_store(struct kmem_cache *s,
3314 const char *buf, size_t length)
3316 if (any_slab_objects(s))
3319 s->flags &= ~SLAB_POISON;
3321 s->flags |= SLAB_POISON;
3327 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3329 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3332 static ssize_t store_user_store(struct kmem_cache *s,
3333 const char *buf, size_t length)
3335 if (any_slab_objects(s))
3338 s->flags &= ~SLAB_STORE_USER;
3340 s->flags |= SLAB_STORE_USER;
3344 SLAB_ATTR(store_user);
3346 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3351 static ssize_t validate_store(struct kmem_cache *s,
3352 const char *buf, size_t length)
3355 validate_slab_cache(s);
3360 SLAB_ATTR(validate);
3362 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3367 static ssize_t shrink_store(struct kmem_cache *s,
3368 const char *buf, size_t length)
3370 if (buf[0] == '1') {
3371 int rc = kmem_cache_shrink(s);
3381 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3383 if (!(s->flags & SLAB_STORE_USER))
3385 return list_locations(s, buf, TRACK_ALLOC);
3387 SLAB_ATTR_RO(alloc_calls);
3389 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3391 if (!(s->flags & SLAB_STORE_USER))
3393 return list_locations(s, buf, TRACK_FREE);
3395 SLAB_ATTR_RO(free_calls);
3398 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3400 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3403 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3404 const char *buf, size_t length)
3406 int n = simple_strtoul(buf, NULL, 10);
3409 s->defrag_ratio = n * 10;
3412 SLAB_ATTR(defrag_ratio);
3415 static struct attribute * slab_attrs[] = {
3416 &slab_size_attr.attr,
3417 &object_size_attr.attr,
3418 &objs_per_slab_attr.attr,
3423 &cpu_slabs_attr.attr,
3427 &sanity_checks_attr.attr,
3429 &hwcache_align_attr.attr,
3430 &reclaim_account_attr.attr,
3431 &destroy_by_rcu_attr.attr,
3432 &red_zone_attr.attr,
3434 &store_user_attr.attr,
3435 &validate_attr.attr,
3437 &alloc_calls_attr.attr,
3438 &free_calls_attr.attr,
3439 #ifdef CONFIG_ZONE_DMA
3440 &cache_dma_attr.attr,
3443 &defrag_ratio_attr.attr,
3448 static struct attribute_group slab_attr_group = {
3449 .attrs = slab_attrs,
3452 static ssize_t slab_attr_show(struct kobject *kobj,
3453 struct attribute *attr,
3456 struct slab_attribute *attribute;
3457 struct kmem_cache *s;
3460 attribute = to_slab_attr(attr);
3463 if (!attribute->show)
3466 err = attribute->show(s, buf);
3471 static ssize_t slab_attr_store(struct kobject *kobj,
3472 struct attribute *attr,
3473 const char *buf, size_t len)
3475 struct slab_attribute *attribute;
3476 struct kmem_cache *s;
3479 attribute = to_slab_attr(attr);
3482 if (!attribute->store)
3485 err = attribute->store(s, buf, len);
3490 static struct sysfs_ops slab_sysfs_ops = {
3491 .show = slab_attr_show,
3492 .store = slab_attr_store,
3495 static struct kobj_type slab_ktype = {
3496 .sysfs_ops = &slab_sysfs_ops,
3499 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3501 struct kobj_type *ktype = get_ktype(kobj);
3503 if (ktype == &slab_ktype)
3508 static struct kset_uevent_ops slab_uevent_ops = {
3509 .filter = uevent_filter,
3512 decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3514 #define ID_STR_LENGTH 64
3516 /* Create a unique string id for a slab cache:
3518 * :[flags-]size:[memory address of kmemcache]
3520 static char *create_unique_id(struct kmem_cache *s)
3522 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3529 * First flags affecting slabcache operations. We will only
3530 * get here for aliasable slabs so we do not need to support
3531 * too many flags. The flags here must cover all flags that
3532 * are matched during merging to guarantee that the id is
3535 if (s->flags & SLAB_CACHE_DMA)
3537 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3539 if (s->flags & SLAB_DEBUG_FREE)
3543 p += sprintf(p, "%07d", s->size);
3544 BUG_ON(p > name + ID_STR_LENGTH - 1);
3548 static int sysfs_slab_add(struct kmem_cache *s)
3554 if (slab_state < SYSFS)
3555 /* Defer until later */
3558 unmergeable = slab_unmergeable(s);
3561 * Slabcache can never be merged so we can use the name proper.
3562 * This is typically the case for debug situations. In that
3563 * case we can catch duplicate names easily.
3565 sysfs_remove_link(&slab_subsys.kobj, s->name);
3569 * Create a unique name for the slab as a target
3572 name = create_unique_id(s);
3575 kobj_set_kset_s(s, slab_subsys);
3576 kobject_set_name(&s->kobj, name);
3577 kobject_init(&s->kobj);
3578 err = kobject_add(&s->kobj);
3582 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3585 kobject_uevent(&s->kobj, KOBJ_ADD);
3587 /* Setup first alias */
3588 sysfs_slab_alias(s, s->name);
3594 static void sysfs_slab_remove(struct kmem_cache *s)
3596 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3597 kobject_del(&s->kobj);
3601 * Need to buffer aliases during bootup until sysfs becomes
3602 * available lest we loose that information.
3604 struct saved_alias {
3605 struct kmem_cache *s;
3607 struct saved_alias *next;
3610 struct saved_alias *alias_list;
3612 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3614 struct saved_alias *al;
3616 if (slab_state == SYSFS) {
3618 * If we have a leftover link then remove it.
3620 sysfs_remove_link(&slab_subsys.kobj, name);
3621 return sysfs_create_link(&slab_subsys.kobj,
3625 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3631 al->next = alias_list;
3636 static int __init slab_sysfs_init(void)
3638 struct list_head *h;
3641 err = subsystem_register(&slab_subsys);
3643 printk(KERN_ERR "Cannot register slab subsystem.\n");
3649 list_for_each(h, &slab_caches) {
3650 struct kmem_cache *s =
3651 container_of(h, struct kmem_cache, list);
3653 err = sysfs_slab_add(s);
3657 while (alias_list) {
3658 struct saved_alias *al = alias_list;
3660 alias_list = alias_list->next;
3661 err = sysfs_slab_alias(al->s, al->name);
3670 __initcall(slab_sysfs_init);