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)
209 * The page->inuse field is 16 bit thus we have this limitation
211 #define MAX_OBJECTS_PER_SLAB 65535
213 /* Internal SLUB flags */
214 #define __OBJECT_POISON 0x80000000 /* Poison object */
216 /* Not all arches define cache_line_size */
217 #ifndef cache_line_size
218 #define cache_line_size() L1_CACHE_BYTES
221 static int kmem_size = sizeof(struct kmem_cache);
224 static struct notifier_block slab_notifier;
228 DOWN, /* No slab functionality available */
229 PARTIAL, /* kmem_cache_open() works but kmalloc does not */
230 UP, /* Everything works but does not show up in sysfs */
234 /* A list of all slab caches on the system */
235 static DECLARE_RWSEM(slub_lock);
236 static LIST_HEAD(slab_caches);
239 * Tracking user of a slab.
242 void *addr; /* Called from address */
243 int cpu; /* Was running on cpu */
244 int pid; /* Pid context */
245 unsigned long when; /* When did the operation occur */
248 enum track_item { TRACK_ALLOC, TRACK_FREE };
250 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
251 static int sysfs_slab_add(struct kmem_cache *);
252 static int sysfs_slab_alias(struct kmem_cache *, const char *);
253 static void sysfs_slab_remove(struct kmem_cache *);
255 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
256 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
258 static inline void sysfs_slab_remove(struct kmem_cache *s) {}
261 /********************************************************************
262 * Core slab cache functions
263 *******************************************************************/
265 int slab_is_available(void)
267 return slab_state >= UP;
270 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
273 return s->node[node];
275 return &s->local_node;
279 static inline int check_valid_pointer(struct kmem_cache *s,
280 struct page *page, const void *object)
287 base = page_address(page);
288 if (object < base || object >= base + s->objects * s->size ||
289 (object - base) % s->size) {
297 * Slow version of get and set free pointer.
299 * This version requires touching the cache lines of kmem_cache which
300 * we avoid to do in the fast alloc free paths. There we obtain the offset
301 * from the page struct.
303 static inline void *get_freepointer(struct kmem_cache *s, void *object)
305 return *(void **)(object + s->offset);
308 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
310 *(void **)(object + s->offset) = fp;
313 /* Loop over all objects in a slab */
314 #define for_each_object(__p, __s, __addr) \
315 for (__p = (__addr); __p < (__addr) + (__s)->objects * (__s)->size;\
319 #define for_each_free_object(__p, __s, __free) \
320 for (__p = (__free); __p; __p = get_freepointer((__s), __p))
322 /* Determine object index from a given position */
323 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
325 return (p - addr) / s->size;
328 #ifdef CONFIG_SLUB_DEBUG
332 #ifdef CONFIG_SLUB_DEBUG_ON
333 static int slub_debug = DEBUG_DEFAULT_FLAGS;
335 static int slub_debug;
338 static char *slub_debug_slabs;
343 static void print_section(char *text, u8 *addr, unsigned int length)
351 for (i = 0; i < length; i++) {
353 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
356 printk(" %02x", addr[i]);
358 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
360 printk(" %s\n",ascii);
371 printk(" %s\n", ascii);
375 static struct track *get_track(struct kmem_cache *s, void *object,
376 enum track_item alloc)
381 p = object + s->offset + sizeof(void *);
383 p = object + s->inuse;
388 static void set_track(struct kmem_cache *s, void *object,
389 enum track_item alloc, void *addr)
394 p = object + s->offset + sizeof(void *);
396 p = object + s->inuse;
401 p->cpu = smp_processor_id();
402 p->pid = current ? current->pid : -1;
405 memset(p, 0, sizeof(struct track));
408 static void init_tracking(struct kmem_cache *s, void *object)
410 if (!(s->flags & SLAB_STORE_USER))
413 set_track(s, object, TRACK_FREE, NULL);
414 set_track(s, object, TRACK_ALLOC, NULL);
417 static void print_track(const char *s, struct track *t)
422 printk(KERN_ERR "INFO: %s in ", s);
423 __print_symbol("%s", (unsigned long)t->addr);
424 printk(" age=%lu cpu=%u pid=%d\n", jiffies - t->when, t->cpu, t->pid);
427 static void print_tracking(struct kmem_cache *s, void *object)
429 if (!(s->flags & SLAB_STORE_USER))
432 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
433 print_track("Freed", get_track(s, object, TRACK_FREE));
436 static void print_page_info(struct page *page)
438 printk(KERN_ERR "INFO: Slab 0x%p used=%u fp=0x%p flags=0x%04lx\n",
439 page, page->inuse, page->freelist, page->flags);
443 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
449 vsnprintf(buf, sizeof(buf), fmt, args);
451 printk(KERN_ERR "========================================"
452 "=====================================\n");
453 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
454 printk(KERN_ERR "----------------------------------------"
455 "-------------------------------------\n\n");
458 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
464 vsnprintf(buf, sizeof(buf), fmt, args);
466 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
469 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
471 unsigned int off; /* Offset of last byte */
472 u8 *addr = page_address(page);
474 print_tracking(s, p);
476 print_page_info(page);
478 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
479 p, p - addr, get_freepointer(s, p));
482 print_section("Bytes b4", p - 16, 16);
484 print_section("Object", p, min(s->objsize, 128));
486 if (s->flags & SLAB_RED_ZONE)
487 print_section("Redzone", p + s->objsize,
488 s->inuse - s->objsize);
491 off = s->offset + sizeof(void *);
495 if (s->flags & SLAB_STORE_USER)
496 off += 2 * sizeof(struct track);
499 /* Beginning of the filler is the free pointer */
500 print_section("Padding", p + off, s->size - off);
505 static void object_err(struct kmem_cache *s, struct page *page,
506 u8 *object, char *reason)
509 print_trailer(s, page, object);
512 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
518 vsnprintf(buf, sizeof(buf), fmt, args);
521 print_page_info(page);
525 static void init_object(struct kmem_cache *s, void *object, int active)
529 if (s->flags & __OBJECT_POISON) {
530 memset(p, POISON_FREE, s->objsize - 1);
531 p[s->objsize -1] = POISON_END;
534 if (s->flags & SLAB_RED_ZONE)
535 memset(p + s->objsize,
536 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE,
537 s->inuse - s->objsize);
540 static u8 *check_bytes(u8 *start, unsigned int value, unsigned int bytes)
543 if (*start != (u8)value)
551 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
552 void *from, void *to)
554 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
555 memset(from, data, to - from);
558 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
559 u8 *object, char *what,
560 u8* start, unsigned int value, unsigned int bytes)
565 fault = check_bytes(start, value, bytes);
570 while (end > fault && end[-1] == value)
573 slab_bug(s, "%s overwritten", what);
574 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
575 fault, end - 1, fault[0], value);
576 print_trailer(s, page, object);
578 restore_bytes(s, what, value, fault, end);
586 * Bytes of the object to be managed.
587 * If the freepointer may overlay the object then the free
588 * pointer is the first word of the object.
590 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
593 * object + s->objsize
594 * Padding to reach word boundary. This is also used for Redzoning.
595 * Padding is extended by another word if Redzoning is enabled and
598 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
599 * 0xcc (RED_ACTIVE) for objects in use.
602 * Meta data starts here.
604 * A. Free pointer (if we cannot overwrite object on free)
605 * B. Tracking data for SLAB_STORE_USER
606 * C. Padding to reach required alignment boundary or at mininum
607 * one word if debuggin is on to be able to detect writes
608 * before the word boundary.
610 * Padding is done using 0x5a (POISON_INUSE)
613 * Nothing is used beyond s->size.
615 * If slabcaches are merged then the objsize and inuse boundaries are mostly
616 * ignored. And therefore no slab options that rely on these boundaries
617 * may be used with merged slabcaches.
620 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
622 unsigned long off = s->inuse; /* The end of info */
625 /* Freepointer is placed after the object. */
626 off += sizeof(void *);
628 if (s->flags & SLAB_STORE_USER)
629 /* We also have user information there */
630 off += 2 * sizeof(struct track);
635 return check_bytes_and_report(s, page, p, "Object padding",
636 p + off, POISON_INUSE, s->size - off);
639 static int slab_pad_check(struct kmem_cache *s, struct page *page)
647 if (!(s->flags & SLAB_POISON))
650 start = page_address(page);
651 end = start + (PAGE_SIZE << s->order);
652 length = s->objects * s->size;
653 remainder = end - (start + length);
657 fault = check_bytes(start + length, POISON_INUSE, remainder);
660 while (end > fault && end[-1] == POISON_INUSE)
663 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
664 print_section("Padding", start, length);
666 restore_bytes(s, "slab padding", POISON_INUSE, start, end);
670 static int check_object(struct kmem_cache *s, struct page *page,
671 void *object, int active)
674 u8 *endobject = object + s->objsize;
676 if (s->flags & SLAB_RED_ZONE) {
678 active ? SLUB_RED_ACTIVE : SLUB_RED_INACTIVE;
680 if (!check_bytes_and_report(s, page, object, "Redzone",
681 endobject, red, s->inuse - s->objsize))
684 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse)
685 check_bytes_and_report(s, page, p, "Alignment padding", endobject,
686 POISON_INUSE, s->inuse - s->objsize);
689 if (s->flags & SLAB_POISON) {
690 if (!active && (s->flags & __OBJECT_POISON) &&
691 (!check_bytes_and_report(s, page, p, "Poison", p,
692 POISON_FREE, s->objsize - 1) ||
693 !check_bytes_and_report(s, page, p, "Poison",
694 p + s->objsize -1, POISON_END, 1)))
697 * check_pad_bytes cleans up on its own.
699 check_pad_bytes(s, page, p);
702 if (!s->offset && active)
704 * Object and freepointer overlap. Cannot check
705 * freepointer while object is allocated.
709 /* Check free pointer validity */
710 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
711 object_err(s, page, p, "Freepointer corrupt");
713 * No choice but to zap it and thus loose the remainder
714 * of the free objects in this slab. May cause
715 * another error because the object count is now wrong.
717 set_freepointer(s, p, NULL);
723 static int check_slab(struct kmem_cache *s, struct page *page)
725 VM_BUG_ON(!irqs_disabled());
727 if (!PageSlab(page)) {
728 slab_err(s, page, "Not a valid slab page");
731 if (page->offset * sizeof(void *) != s->offset) {
732 slab_err(s, page, "Corrupted offset %lu",
733 (unsigned long)(page->offset * sizeof(void *)));
736 if (page->inuse > s->objects) {
737 slab_err(s, page, "inuse %u > max %u",
738 s->name, page->inuse, s->objects);
741 /* Slab_pad_check fixes things up after itself */
742 slab_pad_check(s, page);
747 * Determine if a certain object on a page is on the freelist. Must hold the
748 * slab lock to guarantee that the chains are in a consistent state.
750 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
753 void *fp = page->freelist;
756 while (fp && nr <= s->objects) {
759 if (!check_valid_pointer(s, page, fp)) {
761 object_err(s, page, object,
762 "Freechain corrupt");
763 set_freepointer(s, object, NULL);
766 slab_err(s, page, "Freepointer corrupt");
767 page->freelist = NULL;
768 page->inuse = s->objects;
769 slab_fix(s, "Freelist cleared");
775 fp = get_freepointer(s, object);
779 if (page->inuse != s->objects - nr) {
780 slab_err(s, page, "Wrong object count. Counter is %d but "
781 "counted were %d", page->inuse, s->objects - nr);
782 page->inuse = s->objects - nr;
783 slab_fix(s, "Object count adjusted.");
785 return search == NULL;
788 static void trace(struct kmem_cache *s, struct page *page, void *object, int alloc)
790 if (s->flags & SLAB_TRACE) {
791 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
793 alloc ? "alloc" : "free",
798 print_section("Object", (void *)object, s->objsize);
805 * Tracking of fully allocated slabs for debugging purposes.
807 static void add_full(struct kmem_cache_node *n, struct page *page)
809 spin_lock(&n->list_lock);
810 list_add(&page->lru, &n->full);
811 spin_unlock(&n->list_lock);
814 static void remove_full(struct kmem_cache *s, struct page *page)
816 struct kmem_cache_node *n;
818 if (!(s->flags & SLAB_STORE_USER))
821 n = get_node(s, page_to_nid(page));
823 spin_lock(&n->list_lock);
824 list_del(&page->lru);
825 spin_unlock(&n->list_lock);
828 static void setup_object_debug(struct kmem_cache *s, struct page *page,
831 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
834 init_object(s, object, 0);
835 init_tracking(s, object);
838 static int alloc_debug_processing(struct kmem_cache *s, struct page *page,
839 void *object, void *addr)
841 if (!check_slab(s, page))
844 if (object && !on_freelist(s, page, object)) {
845 object_err(s, page, object, "Object already allocated");
849 if (!check_valid_pointer(s, page, object)) {
850 object_err(s, page, object, "Freelist Pointer check fails");
854 if (object && !check_object(s, page, object, 0))
857 /* Success perform special debug activities for allocs */
858 if (s->flags & SLAB_STORE_USER)
859 set_track(s, object, TRACK_ALLOC, addr);
860 trace(s, page, object, 1);
861 init_object(s, object, 1);
865 if (PageSlab(page)) {
867 * If this is a slab page then lets do the best we can
868 * to avoid issues in the future. Marking all objects
869 * as used avoids touching the remaining objects.
871 slab_fix(s, "Marking all objects used");
872 page->inuse = s->objects;
873 page->freelist = NULL;
874 /* Fix up fields that may be corrupted */
875 page->offset = s->offset / sizeof(void *);
880 static int free_debug_processing(struct kmem_cache *s, struct page *page,
881 void *object, void *addr)
883 if (!check_slab(s, page))
886 if (!check_valid_pointer(s, page, object)) {
887 slab_err(s, page, "Invalid object pointer 0x%p", object);
891 if (on_freelist(s, page, object)) {
892 object_err(s, page, object, "Object already free");
896 if (!check_object(s, page, object, 1))
899 if (unlikely(s != page->slab)) {
901 slab_err(s, page, "Attempt to free object(0x%p) "
902 "outside of slab", object);
906 "SLUB <none>: no slab for object 0x%p.\n",
911 object_err(s, page, object,
912 "page slab pointer corrupt.");
916 /* Special debug activities for freeing objects */
917 if (!SlabFrozen(page) && !page->freelist)
918 remove_full(s, page);
919 if (s->flags & SLAB_STORE_USER)
920 set_track(s, object, TRACK_FREE, addr);
921 trace(s, page, object, 0);
922 init_object(s, object, 0);
926 slab_fix(s, "Object at 0x%p not freed", object);
930 static int __init setup_slub_debug(char *str)
932 slub_debug = DEBUG_DEFAULT_FLAGS;
933 if (*str++ != '=' || !*str)
935 * No options specified. Switch on full debugging.
941 * No options but restriction on slabs. This means full
942 * debugging for slabs matching a pattern.
949 * Switch off all debugging measures.
954 * Determine which debug features should be switched on
956 for ( ;*str && *str != ','; str++) {
957 switch (tolower(*str)) {
959 slub_debug |= SLAB_DEBUG_FREE;
962 slub_debug |= SLAB_RED_ZONE;
965 slub_debug |= SLAB_POISON;
968 slub_debug |= SLAB_STORE_USER;
971 slub_debug |= SLAB_TRACE;
974 printk(KERN_ERR "slub_debug option '%c' "
975 "unknown. skipped\n",*str);
981 slub_debug_slabs = str + 1;
986 __setup("slub_debug", setup_slub_debug);
988 static void kmem_cache_open_debug_check(struct kmem_cache *s)
991 * The page->offset field is only 16 bit wide. This is an offset
992 * in units of words from the beginning of an object. If the slab
993 * size is bigger then we cannot move the free pointer behind the
996 * On 32 bit platforms the limit is 256k. On 64bit platforms
999 * Debugging or ctor may create a need to move the free
1000 * pointer. Fail if this happens.
1002 if (s->objsize >= 65535 * sizeof(void *)) {
1003 BUG_ON(s->flags & (SLAB_RED_ZONE | SLAB_POISON |
1004 SLAB_STORE_USER | SLAB_DESTROY_BY_RCU));
1009 * Enable debugging if selected on the kernel commandline.
1011 if (slub_debug && (!slub_debug_slabs ||
1012 strncmp(slub_debug_slabs, s->name,
1013 strlen(slub_debug_slabs)) == 0))
1014 s->flags |= slub_debug;
1017 static inline void setup_object_debug(struct kmem_cache *s,
1018 struct page *page, void *object) {}
1020 static inline int alloc_debug_processing(struct kmem_cache *s,
1021 struct page *page, void *object, void *addr) { return 0; }
1023 static inline int free_debug_processing(struct kmem_cache *s,
1024 struct page *page, void *object, void *addr) { return 0; }
1026 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1028 static inline int check_object(struct kmem_cache *s, struct page *page,
1029 void *object, int active) { return 1; }
1030 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1031 static inline void kmem_cache_open_debug_check(struct kmem_cache *s) {}
1032 #define slub_debug 0
1035 * Slab allocation and freeing
1037 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1040 int pages = 1 << s->order;
1043 flags |= __GFP_COMP;
1045 if (s->flags & SLAB_CACHE_DMA)
1049 page = alloc_pages(flags, s->order);
1051 page = alloc_pages_node(node, flags, s->order);
1056 mod_zone_page_state(page_zone(page),
1057 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1058 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1064 static void setup_object(struct kmem_cache *s, struct page *page,
1067 setup_object_debug(s, page, object);
1068 if (unlikely(s->ctor))
1069 s->ctor(object, s, 0);
1072 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1075 struct kmem_cache_node *n;
1081 BUG_ON(flags & ~(GFP_DMA | __GFP_ZERO | GFP_LEVEL_MASK));
1083 if (flags & __GFP_WAIT)
1086 page = allocate_slab(s, flags & GFP_LEVEL_MASK, node);
1090 n = get_node(s, page_to_nid(page));
1092 atomic_long_inc(&n->nr_slabs);
1093 page->offset = s->offset / sizeof(void *);
1095 page->flags |= 1 << PG_slab;
1096 if (s->flags & (SLAB_DEBUG_FREE | SLAB_RED_ZONE | SLAB_POISON |
1097 SLAB_STORE_USER | SLAB_TRACE))
1100 start = page_address(page);
1101 end = start + s->objects * s->size;
1103 if (unlikely(s->flags & SLAB_POISON))
1104 memset(start, POISON_INUSE, PAGE_SIZE << s->order);
1107 for_each_object(p, s, start) {
1108 setup_object(s, page, last);
1109 set_freepointer(s, last, p);
1112 setup_object(s, page, last);
1113 set_freepointer(s, last, NULL);
1115 page->freelist = start;
1116 page->lockless_freelist = NULL;
1119 if (flags & __GFP_WAIT)
1120 local_irq_disable();
1124 static void __free_slab(struct kmem_cache *s, struct page *page)
1126 int pages = 1 << s->order;
1128 if (unlikely(SlabDebug(page))) {
1131 slab_pad_check(s, page);
1132 for_each_object(p, s, page_address(page))
1133 check_object(s, page, p, 0);
1136 mod_zone_page_state(page_zone(page),
1137 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1138 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1141 page->mapping = NULL;
1142 __free_pages(page, s->order);
1145 static void rcu_free_slab(struct rcu_head *h)
1149 page = container_of((struct list_head *)h, struct page, lru);
1150 __free_slab(page->slab, page);
1153 static void free_slab(struct kmem_cache *s, struct page *page)
1155 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1157 * RCU free overloads the RCU head over the LRU
1159 struct rcu_head *head = (void *)&page->lru;
1161 call_rcu(head, rcu_free_slab);
1163 __free_slab(s, page);
1166 static void discard_slab(struct kmem_cache *s, struct page *page)
1168 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1170 atomic_long_dec(&n->nr_slabs);
1171 reset_page_mapcount(page);
1172 ClearSlabDebug(page);
1173 __ClearPageSlab(page);
1178 * Per slab locking using the pagelock
1180 static __always_inline void slab_lock(struct page *page)
1182 bit_spin_lock(PG_locked, &page->flags);
1185 static __always_inline void slab_unlock(struct page *page)
1187 bit_spin_unlock(PG_locked, &page->flags);
1190 static __always_inline int slab_trylock(struct page *page)
1194 rc = bit_spin_trylock(PG_locked, &page->flags);
1199 * Management of partially allocated slabs
1201 static void add_partial_tail(struct kmem_cache_node *n, struct page *page)
1203 spin_lock(&n->list_lock);
1205 list_add_tail(&page->lru, &n->partial);
1206 spin_unlock(&n->list_lock);
1209 static void add_partial(struct kmem_cache_node *n, struct page *page)
1211 spin_lock(&n->list_lock);
1213 list_add(&page->lru, &n->partial);
1214 spin_unlock(&n->list_lock);
1217 static void remove_partial(struct kmem_cache *s,
1220 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1222 spin_lock(&n->list_lock);
1223 list_del(&page->lru);
1225 spin_unlock(&n->list_lock);
1229 * Lock slab and remove from the partial list.
1231 * Must hold list_lock.
1233 static inline int lock_and_freeze_slab(struct kmem_cache_node *n, struct page *page)
1235 if (slab_trylock(page)) {
1236 list_del(&page->lru);
1238 SetSlabFrozen(page);
1245 * Try to allocate a partial slab from a specific node.
1247 static struct page *get_partial_node(struct kmem_cache_node *n)
1252 * Racy check. If we mistakenly see no partial slabs then we
1253 * just allocate an empty slab. If we mistakenly try to get a
1254 * partial slab and there is none available then get_partials()
1257 if (!n || !n->nr_partial)
1260 spin_lock(&n->list_lock);
1261 list_for_each_entry(page, &n->partial, lru)
1262 if (lock_and_freeze_slab(n, page))
1266 spin_unlock(&n->list_lock);
1271 * Get a page from somewhere. Search in increasing NUMA distances.
1273 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1276 struct zonelist *zonelist;
1281 * The defrag ratio allows a configuration of the tradeoffs between
1282 * inter node defragmentation and node local allocations. A lower
1283 * defrag_ratio increases the tendency to do local allocations
1284 * instead of attempting to obtain partial slabs from other nodes.
1286 * If the defrag_ratio is set to 0 then kmalloc() always
1287 * returns node local objects. If the ratio is higher then kmalloc()
1288 * may return off node objects because partial slabs are obtained
1289 * from other nodes and filled up.
1291 * If /sys/slab/xx/defrag_ratio is set to 100 (which makes
1292 * defrag_ratio = 1000) then every (well almost) allocation will
1293 * first attempt to defrag slab caches on other nodes. This means
1294 * scanning over all nodes to look for partial slabs which may be
1295 * expensive if we do it every time we are trying to find a slab
1296 * with available objects.
1298 if (!s->defrag_ratio || get_cycles() % 1024 > s->defrag_ratio)
1301 zonelist = &NODE_DATA(slab_node(current->mempolicy))
1302 ->node_zonelists[gfp_zone(flags)];
1303 for (z = zonelist->zones; *z; z++) {
1304 struct kmem_cache_node *n;
1306 n = get_node(s, zone_to_nid(*z));
1308 if (n && cpuset_zone_allowed_hardwall(*z, flags) &&
1309 n->nr_partial > MIN_PARTIAL) {
1310 page = get_partial_node(n);
1320 * Get a partial page, lock it and return it.
1322 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1325 int searchnode = (node == -1) ? numa_node_id() : node;
1327 page = get_partial_node(get_node(s, searchnode));
1328 if (page || (flags & __GFP_THISNODE))
1331 return get_any_partial(s, flags);
1335 * Move a page back to the lists.
1337 * Must be called with the slab lock held.
1339 * On exit the slab lock will have been dropped.
1341 static void unfreeze_slab(struct kmem_cache *s, struct page *page)
1343 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1345 ClearSlabFrozen(page);
1349 add_partial(n, page);
1350 else if (SlabDebug(page) && (s->flags & SLAB_STORE_USER))
1355 if (n->nr_partial < MIN_PARTIAL) {
1357 * Adding an empty slab to the partial slabs in order
1358 * to avoid page allocator overhead. This slab needs
1359 * to come after the other slabs with objects in
1360 * order to fill them up. That way the size of the
1361 * partial list stays small. kmem_cache_shrink can
1362 * reclaim empty slabs from the partial list.
1364 add_partial_tail(n, page);
1368 discard_slab(s, page);
1374 * Remove the cpu slab
1376 static void deactivate_slab(struct kmem_cache *s, struct page *page, int cpu)
1379 * Merge cpu freelist into freelist. Typically we get here
1380 * because both freelists are empty. So this is unlikely
1383 while (unlikely(page->lockless_freelist)) {
1386 /* Retrieve object from cpu_freelist */
1387 object = page->lockless_freelist;
1388 page->lockless_freelist = page->lockless_freelist[page->offset];
1390 /* And put onto the regular freelist */
1391 object[page->offset] = page->freelist;
1392 page->freelist = object;
1395 s->cpu_slab[cpu] = NULL;
1396 unfreeze_slab(s, page);
1399 static inline void flush_slab(struct kmem_cache *s, struct page *page, int cpu)
1402 deactivate_slab(s, page, cpu);
1407 * Called from IPI handler with interrupts disabled.
1409 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1411 struct page *page = s->cpu_slab[cpu];
1414 flush_slab(s, page, cpu);
1417 static void flush_cpu_slab(void *d)
1419 struct kmem_cache *s = d;
1420 int cpu = smp_processor_id();
1422 __flush_cpu_slab(s, cpu);
1425 static void flush_all(struct kmem_cache *s)
1428 on_each_cpu(flush_cpu_slab, s, 1, 1);
1430 unsigned long flags;
1432 local_irq_save(flags);
1434 local_irq_restore(flags);
1439 * Slow path. The lockless freelist is empty or we need to perform
1442 * Interrupts are disabled.
1444 * Processing is still very fast if new objects have been freed to the
1445 * regular freelist. In that case we simply take over the regular freelist
1446 * as the lockless freelist and zap the regular freelist.
1448 * If that is not working then we fall back to the partial lists. We take the
1449 * first element of the freelist as the object to allocate now and move the
1450 * rest of the freelist to the lockless freelist.
1452 * And if we were unable to get a new slab from the partial slab lists then
1453 * we need to allocate a new slab. This is slowest path since we may sleep.
1455 static void *__slab_alloc(struct kmem_cache *s,
1456 gfp_t gfpflags, int node, void *addr, struct page *page)
1459 int cpu = smp_processor_id();
1465 if (unlikely(node != -1 && page_to_nid(page) != node))
1468 object = page->freelist;
1469 if (unlikely(!object))
1471 if (unlikely(SlabDebug(page)))
1474 object = page->freelist;
1475 page->lockless_freelist = object[page->offset];
1476 page->inuse = s->objects;
1477 page->freelist = NULL;
1482 deactivate_slab(s, page, cpu);
1485 page = get_partial(s, gfpflags, node);
1487 s->cpu_slab[cpu] = page;
1491 page = new_slab(s, gfpflags, node);
1493 cpu = smp_processor_id();
1494 if (s->cpu_slab[cpu]) {
1496 * Someone else populated the cpu_slab while we
1497 * enabled interrupts, or we have gotten scheduled
1498 * on another cpu. The page may not be on the
1499 * requested node even if __GFP_THISNODE was
1500 * specified. So we need to recheck.
1503 page_to_nid(s->cpu_slab[cpu]) == node) {
1505 * Current cpuslab is acceptable and we
1506 * want the current one since its cache hot
1508 discard_slab(s, page);
1509 page = s->cpu_slab[cpu];
1513 /* New slab does not fit our expectations */
1514 flush_slab(s, s->cpu_slab[cpu], cpu);
1517 SetSlabFrozen(page);
1518 s->cpu_slab[cpu] = page;
1523 object = page->freelist;
1524 if (!alloc_debug_processing(s, page, object, addr))
1528 page->freelist = object[page->offset];
1534 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1535 * have the fastpath folded into their functions. So no function call
1536 * overhead for requests that can be satisfied on the fastpath.
1538 * The fastpath works by first checking if the lockless freelist can be used.
1539 * If not then __slab_alloc is called for slow processing.
1541 * Otherwise we can simply pick the next object from the lockless free list.
1543 static void __always_inline *slab_alloc(struct kmem_cache *s,
1544 gfp_t gfpflags, int node, void *addr, int length)
1548 unsigned long flags;
1550 local_irq_save(flags);
1551 page = s->cpu_slab[smp_processor_id()];
1552 if (unlikely(!page || !page->lockless_freelist ||
1553 (node != -1 && page_to_nid(page) != node)))
1555 object = __slab_alloc(s, gfpflags, node, addr, page);
1558 object = page->lockless_freelist;
1559 page->lockless_freelist = object[page->offset];
1561 local_irq_restore(flags);
1563 if (unlikely((gfpflags & __GFP_ZERO) && object))
1564 memset(object, 0, length);
1569 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
1571 return slab_alloc(s, gfpflags, -1,
1572 __builtin_return_address(0), s->objsize);
1574 EXPORT_SYMBOL(kmem_cache_alloc);
1577 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
1579 return slab_alloc(s, gfpflags, node,
1580 __builtin_return_address(0), s->objsize);
1582 EXPORT_SYMBOL(kmem_cache_alloc_node);
1586 * Slow patch handling. This may still be called frequently since objects
1587 * have a longer lifetime than the cpu slabs in most processing loads.
1589 * So we still attempt to reduce cache line usage. Just take the slab
1590 * lock and free the item. If there is no additional partial page
1591 * handling required then we can return immediately.
1593 static void __slab_free(struct kmem_cache *s, struct page *page,
1594 void *x, void *addr)
1597 void **object = (void *)x;
1601 if (unlikely(SlabDebug(page)))
1604 prior = object[page->offset] = page->freelist;
1605 page->freelist = object;
1608 if (unlikely(SlabFrozen(page)))
1611 if (unlikely(!page->inuse))
1615 * Objects left in the slab. If it
1616 * was not on the partial list before
1619 if (unlikely(!prior))
1620 add_partial(get_node(s, page_to_nid(page)), page);
1629 * Slab still on the partial list.
1631 remove_partial(s, page);
1634 discard_slab(s, page);
1638 if (!free_debug_processing(s, page, x, addr))
1644 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
1645 * can perform fastpath freeing without additional function calls.
1647 * The fastpath is only possible if we are freeing to the current cpu slab
1648 * of this processor. This typically the case if we have just allocated
1651 * If fastpath is not possible then fall back to __slab_free where we deal
1652 * with all sorts of special processing.
1654 static void __always_inline slab_free(struct kmem_cache *s,
1655 struct page *page, void *x, void *addr)
1657 void **object = (void *)x;
1658 unsigned long flags;
1660 local_irq_save(flags);
1661 if (likely(page == s->cpu_slab[smp_processor_id()] &&
1662 !SlabDebug(page))) {
1663 object[page->offset] = page->lockless_freelist;
1664 page->lockless_freelist = object;
1666 __slab_free(s, page, x, addr);
1668 local_irq_restore(flags);
1671 void kmem_cache_free(struct kmem_cache *s, void *x)
1675 page = virt_to_head_page(x);
1677 slab_free(s, page, x, __builtin_return_address(0));
1679 EXPORT_SYMBOL(kmem_cache_free);
1681 /* Figure out on which slab object the object resides */
1682 static struct page *get_object_page(const void *x)
1684 struct page *page = virt_to_head_page(x);
1686 if (!PageSlab(page))
1693 * Object placement in a slab is made very easy because we always start at
1694 * offset 0. If we tune the size of the object to the alignment then we can
1695 * get the required alignment by putting one properly sized object after
1698 * Notice that the allocation order determines the sizes of the per cpu
1699 * caches. Each processor has always one slab available for allocations.
1700 * Increasing the allocation order reduces the number of times that slabs
1701 * must be moved on and off the partial lists and is therefore a factor in
1706 * Mininum / Maximum order of slab pages. This influences locking overhead
1707 * and slab fragmentation. A higher order reduces the number of partial slabs
1708 * and increases the number of allocations possible without having to
1709 * take the list_lock.
1711 static int slub_min_order;
1712 static int slub_max_order = DEFAULT_MAX_ORDER;
1713 static int slub_min_objects = DEFAULT_MIN_OBJECTS;
1716 * Merge control. If this is set then no merging of slab caches will occur.
1717 * (Could be removed. This was introduced to pacify the merge skeptics.)
1719 static int slub_nomerge;
1722 * Calculate the order of allocation given an slab object size.
1724 * The order of allocation has significant impact on performance and other
1725 * system components. Generally order 0 allocations should be preferred since
1726 * order 0 does not cause fragmentation in the page allocator. Larger objects
1727 * be problematic to put into order 0 slabs because there may be too much
1728 * unused space left. We go to a higher order if more than 1/8th of the slab
1731 * In order to reach satisfactory performance we must ensure that a minimum
1732 * number of objects is in one slab. Otherwise we may generate too much
1733 * activity on the partial lists which requires taking the list_lock. This is
1734 * less a concern for large slabs though which are rarely used.
1736 * slub_max_order specifies the order where we begin to stop considering the
1737 * number of objects in a slab as critical. If we reach slub_max_order then
1738 * we try to keep the page order as low as possible. So we accept more waste
1739 * of space in favor of a small page order.
1741 * Higher order allocations also allow the placement of more objects in a
1742 * slab and thereby reduce object handling overhead. If the user has
1743 * requested a higher mininum order then we start with that one instead of
1744 * the smallest order which will fit the object.
1746 static inline int slab_order(int size, int min_objects,
1747 int max_order, int fract_leftover)
1751 int min_order = slub_min_order;
1754 * If we would create too many object per slab then reduce
1755 * the slab order even if it goes below slub_min_order.
1757 while (min_order > 0 &&
1758 (PAGE_SIZE << min_order) >= MAX_OBJECTS_PER_SLAB * size)
1761 for (order = max(min_order,
1762 fls(min_objects * size - 1) - PAGE_SHIFT);
1763 order <= max_order; order++) {
1765 unsigned long slab_size = PAGE_SIZE << order;
1767 if (slab_size < min_objects * size)
1770 rem = slab_size % size;
1772 if (rem <= slab_size / fract_leftover)
1775 /* If the next size is too high then exit now */
1776 if (slab_size * 2 >= MAX_OBJECTS_PER_SLAB * size)
1783 static inline int calculate_order(int size)
1790 * Attempt to find best configuration for a slab. This
1791 * works by first attempting to generate a layout with
1792 * the best configuration and backing off gradually.
1794 * First we reduce the acceptable waste in a slab. Then
1795 * we reduce the minimum objects required in a slab.
1797 min_objects = slub_min_objects;
1798 while (min_objects > 1) {
1800 while (fraction >= 4) {
1801 order = slab_order(size, min_objects,
1802 slub_max_order, fraction);
1803 if (order <= slub_max_order)
1811 * We were unable to place multiple objects in a slab. Now
1812 * lets see if we can place a single object there.
1814 order = slab_order(size, 1, slub_max_order, 1);
1815 if (order <= slub_max_order)
1819 * Doh this slab cannot be placed using slub_max_order.
1821 order = slab_order(size, 1, MAX_ORDER, 1);
1822 if (order <= MAX_ORDER)
1828 * Figure out what the alignment of the objects will be.
1830 static unsigned long calculate_alignment(unsigned long flags,
1831 unsigned long align, unsigned long size)
1834 * If the user wants hardware cache aligned objects then
1835 * follow that suggestion if the object is sufficiently
1838 * The hardware cache alignment cannot override the
1839 * specified alignment though. If that is greater
1842 if ((flags & SLAB_HWCACHE_ALIGN) &&
1843 size > cache_line_size() / 2)
1844 return max_t(unsigned long, align, cache_line_size());
1846 if (align < ARCH_SLAB_MINALIGN)
1847 return ARCH_SLAB_MINALIGN;
1849 return ALIGN(align, sizeof(void *));
1852 static void init_kmem_cache_node(struct kmem_cache_node *n)
1855 atomic_long_set(&n->nr_slabs, 0);
1856 spin_lock_init(&n->list_lock);
1857 INIT_LIST_HEAD(&n->partial);
1858 INIT_LIST_HEAD(&n->full);
1863 * No kmalloc_node yet so do it by hand. We know that this is the first
1864 * slab on the node for this slabcache. There are no concurrent accesses
1867 * Note that this function only works on the kmalloc_node_cache
1868 * when allocating for the kmalloc_node_cache.
1870 static struct kmem_cache_node * __init early_kmem_cache_node_alloc(gfp_t gfpflags,
1874 struct kmem_cache_node *n;
1876 BUG_ON(kmalloc_caches->size < sizeof(struct kmem_cache_node));
1878 page = new_slab(kmalloc_caches, gfpflags | GFP_THISNODE, node);
1883 page->freelist = get_freepointer(kmalloc_caches, n);
1885 kmalloc_caches->node[node] = n;
1886 init_object(kmalloc_caches, n, 1);
1887 init_tracking(kmalloc_caches, n);
1888 init_kmem_cache_node(n);
1889 atomic_long_inc(&n->nr_slabs);
1890 add_partial(n, page);
1893 * new_slab() disables interupts. If we do not reenable interrupts here
1894 * then bootup would continue with interrupts disabled.
1900 static void free_kmem_cache_nodes(struct kmem_cache *s)
1904 for_each_online_node(node) {
1905 struct kmem_cache_node *n = s->node[node];
1906 if (n && n != &s->local_node)
1907 kmem_cache_free(kmalloc_caches, n);
1908 s->node[node] = NULL;
1912 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1917 if (slab_state >= UP)
1918 local_node = page_to_nid(virt_to_page(s));
1922 for_each_online_node(node) {
1923 struct kmem_cache_node *n;
1925 if (local_node == node)
1928 if (slab_state == DOWN) {
1929 n = early_kmem_cache_node_alloc(gfpflags,
1933 n = kmem_cache_alloc_node(kmalloc_caches,
1937 free_kmem_cache_nodes(s);
1943 init_kmem_cache_node(n);
1948 static void free_kmem_cache_nodes(struct kmem_cache *s)
1952 static int init_kmem_cache_nodes(struct kmem_cache *s, gfp_t gfpflags)
1954 init_kmem_cache_node(&s->local_node);
1960 * calculate_sizes() determines the order and the distribution of data within
1963 static int calculate_sizes(struct kmem_cache *s)
1965 unsigned long flags = s->flags;
1966 unsigned long size = s->objsize;
1967 unsigned long align = s->align;
1970 * Determine if we can poison the object itself. If the user of
1971 * the slab may touch the object after free or before allocation
1972 * then we should never poison the object itself.
1974 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
1976 s->flags |= __OBJECT_POISON;
1978 s->flags &= ~__OBJECT_POISON;
1981 * Round up object size to the next word boundary. We can only
1982 * place the free pointer at word boundaries and this determines
1983 * the possible location of the free pointer.
1985 size = ALIGN(size, sizeof(void *));
1987 #ifdef CONFIG_SLUB_DEBUG
1989 * If we are Redzoning then check if there is some space between the
1990 * end of the object and the free pointer. If not then add an
1991 * additional word to have some bytes to store Redzone information.
1993 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
1994 size += sizeof(void *);
1998 * With that we have determined the number of bytes in actual use
1999 * by the object. This is the potential offset to the free pointer.
2003 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2006 * Relocate free pointer after the object if it is not
2007 * permitted to overwrite the first word of the object on
2010 * This is the case if we do RCU, have a constructor or
2011 * destructor or are poisoning the objects.
2014 size += sizeof(void *);
2017 #ifdef CONFIG_SLUB_DEBUG
2018 if (flags & SLAB_STORE_USER)
2020 * Need to store information about allocs and frees after
2023 size += 2 * sizeof(struct track);
2025 if (flags & SLAB_RED_ZONE)
2027 * Add some empty padding so that we can catch
2028 * overwrites from earlier objects rather than let
2029 * tracking information or the free pointer be
2030 * corrupted if an user writes before the start
2033 size += sizeof(void *);
2037 * Determine the alignment based on various parameters that the
2038 * user specified and the dynamic determination of cache line size
2041 align = calculate_alignment(flags, align, s->objsize);
2044 * SLUB stores one object immediately after another beginning from
2045 * offset 0. In order to align the objects we have to simply size
2046 * each object to conform to the alignment.
2048 size = ALIGN(size, align);
2051 s->order = calculate_order(size);
2056 * Determine the number of objects per slab
2058 s->objects = (PAGE_SIZE << s->order) / size;
2061 * Verify that the number of objects is within permitted limits.
2062 * The page->inuse field is only 16 bit wide! So we cannot have
2063 * more than 64k objects per slab.
2065 if (!s->objects || s->objects > MAX_OBJECTS_PER_SLAB)
2071 static int kmem_cache_open(struct kmem_cache *s, gfp_t gfpflags,
2072 const char *name, size_t size,
2073 size_t align, unsigned long flags,
2074 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2076 memset(s, 0, kmem_size);
2082 kmem_cache_open_debug_check(s);
2084 if (!calculate_sizes(s))
2089 s->defrag_ratio = 100;
2092 if (init_kmem_cache_nodes(s, gfpflags & ~SLUB_DMA))
2095 if (flags & SLAB_PANIC)
2096 panic("Cannot create slab %s size=%lu realsize=%u "
2097 "order=%u offset=%u flags=%lx\n",
2098 s->name, (unsigned long)size, s->size, s->order,
2104 * Check if a given pointer is valid
2106 int kmem_ptr_validate(struct kmem_cache *s, const void *object)
2110 page = get_object_page(object);
2112 if (!page || s != page->slab)
2113 /* No slab or wrong slab */
2116 if (!check_valid_pointer(s, page, object))
2120 * We could also check if the object is on the slabs freelist.
2121 * But this would be too expensive and it seems that the main
2122 * purpose of kmem_ptr_valid is to check if the object belongs
2123 * to a certain slab.
2127 EXPORT_SYMBOL(kmem_ptr_validate);
2130 * Determine the size of a slab object
2132 unsigned int kmem_cache_size(struct kmem_cache *s)
2136 EXPORT_SYMBOL(kmem_cache_size);
2138 const char *kmem_cache_name(struct kmem_cache *s)
2142 EXPORT_SYMBOL(kmem_cache_name);
2145 * Attempt to free all slabs on a node. Return the number of slabs we
2146 * were unable to free.
2148 static int free_list(struct kmem_cache *s, struct kmem_cache_node *n,
2149 struct list_head *list)
2151 int slabs_inuse = 0;
2152 unsigned long flags;
2153 struct page *page, *h;
2155 spin_lock_irqsave(&n->list_lock, flags);
2156 list_for_each_entry_safe(page, h, list, lru)
2158 list_del(&page->lru);
2159 discard_slab(s, page);
2162 spin_unlock_irqrestore(&n->list_lock, flags);
2167 * Release all resources used by a slab cache.
2169 static inline int kmem_cache_close(struct kmem_cache *s)
2175 /* Attempt to free all objects */
2176 for_each_online_node(node) {
2177 struct kmem_cache_node *n = get_node(s, node);
2179 n->nr_partial -= free_list(s, n, &n->partial);
2180 if (atomic_long_read(&n->nr_slabs))
2183 free_kmem_cache_nodes(s);
2188 * Close a cache and release the kmem_cache structure
2189 * (must be used for caches created using kmem_cache_create)
2191 void kmem_cache_destroy(struct kmem_cache *s)
2193 down_write(&slub_lock);
2197 if (kmem_cache_close(s))
2199 sysfs_slab_remove(s);
2202 up_write(&slub_lock);
2204 EXPORT_SYMBOL(kmem_cache_destroy);
2206 /********************************************************************
2208 *******************************************************************/
2210 struct kmem_cache kmalloc_caches[KMALLOC_SHIFT_HIGH + 1] __cacheline_aligned;
2211 EXPORT_SYMBOL(kmalloc_caches);
2213 #ifdef CONFIG_ZONE_DMA
2214 static struct kmem_cache *kmalloc_caches_dma[KMALLOC_SHIFT_HIGH + 1];
2217 static int __init setup_slub_min_order(char *str)
2219 get_option (&str, &slub_min_order);
2224 __setup("slub_min_order=", setup_slub_min_order);
2226 static int __init setup_slub_max_order(char *str)
2228 get_option (&str, &slub_max_order);
2233 __setup("slub_max_order=", setup_slub_max_order);
2235 static int __init setup_slub_min_objects(char *str)
2237 get_option (&str, &slub_min_objects);
2242 __setup("slub_min_objects=", setup_slub_min_objects);
2244 static int __init setup_slub_nomerge(char *str)
2250 __setup("slub_nomerge", setup_slub_nomerge);
2252 static struct kmem_cache *create_kmalloc_cache(struct kmem_cache *s,
2253 const char *name, int size, gfp_t gfp_flags)
2255 unsigned int flags = 0;
2257 if (gfp_flags & SLUB_DMA)
2258 flags = SLAB_CACHE_DMA;
2260 down_write(&slub_lock);
2261 if (!kmem_cache_open(s, gfp_flags, name, size, ARCH_KMALLOC_MINALIGN,
2265 list_add(&s->list, &slab_caches);
2266 up_write(&slub_lock);
2267 if (sysfs_slab_add(s))
2272 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2275 #ifdef CONFIG_ZONE_DMA
2276 static noinline struct kmem_cache *dma_kmalloc_cache(int index, gfp_t flags)
2278 struct kmem_cache *s;
2279 struct kmem_cache *x;
2283 s = kmalloc_caches_dma[index];
2287 /* Dynamically create dma cache */
2288 x = kmalloc(kmem_size, flags & ~SLUB_DMA);
2290 panic("Unable to allocate memory for dma cache\n");
2292 realsize = kmalloc_caches[index].objsize;
2293 text = kasprintf(flags & ~SLUB_DMA, "kmalloc_dma-%d",
2294 (unsigned int)realsize);
2295 s = create_kmalloc_cache(x, text, realsize, flags);
2296 down_write(&slub_lock);
2297 if (!kmalloc_caches_dma[index]) {
2298 kmalloc_caches_dma[index] = s;
2299 up_write(&slub_lock);
2302 up_write(&slub_lock);
2303 kmem_cache_destroy(s);
2304 return kmalloc_caches_dma[index];
2309 * Conversion table for small slabs sizes / 8 to the index in the
2310 * kmalloc array. This is necessary for slabs < 192 since we have non power
2311 * of two cache sizes there. The size of larger slabs can be determined using
2314 static s8 size_index[24] = {
2341 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2347 return ZERO_SIZE_PTR;
2349 index = size_index[(size - 1) / 8];
2351 if (size > KMALLOC_MAX_SIZE)
2354 index = fls(size - 1);
2357 #ifdef CONFIG_ZONE_DMA
2358 if (unlikely((flags & SLUB_DMA)))
2359 return dma_kmalloc_cache(index, flags);
2362 return &kmalloc_caches[index];
2365 void *__kmalloc(size_t size, gfp_t flags)
2367 struct kmem_cache *s = get_slab(size, flags);
2369 if (ZERO_OR_NULL_PTR(s))
2372 return slab_alloc(s, flags, -1, __builtin_return_address(0), size);
2374 EXPORT_SYMBOL(__kmalloc);
2377 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2379 struct kmem_cache *s = get_slab(size, flags);
2381 if (ZERO_OR_NULL_PTR(s))
2384 return slab_alloc(s, flags, node, __builtin_return_address(0), size);
2386 EXPORT_SYMBOL(__kmalloc_node);
2389 size_t ksize(const void *object)
2392 struct kmem_cache *s;
2394 if (object == ZERO_SIZE_PTR)
2397 page = get_object_page(object);
2403 * Debugging requires use of the padding between object
2404 * and whatever may come after it.
2406 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
2410 * If we have the need to store the freelist pointer
2411 * back there or track user information then we can
2412 * only use the space before that information.
2414 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
2418 * Else we can use all the padding etc for the allocation
2422 EXPORT_SYMBOL(ksize);
2424 void kfree(const void *x)
2426 struct kmem_cache *s;
2430 * This has to be an unsigned comparison. According to Linus
2431 * some gcc version treat a pointer as a signed entity. Then
2432 * this comparison would be true for all "negative" pointers
2433 * (which would cover the whole upper half of the address space).
2435 if (ZERO_OR_NULL_PTR(x))
2438 page = virt_to_head_page(x);
2441 slab_free(s, page, (void *)x, __builtin_return_address(0));
2443 EXPORT_SYMBOL(kfree);
2446 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
2447 * the remaining slabs by the number of items in use. The slabs with the
2448 * most items in use come first. New allocations will then fill those up
2449 * and thus they can be removed from the partial lists.
2451 * The slabs with the least items are placed last. This results in them
2452 * being allocated from last increasing the chance that the last objects
2453 * are freed in them.
2455 int kmem_cache_shrink(struct kmem_cache *s)
2459 struct kmem_cache_node *n;
2462 struct list_head *slabs_by_inuse =
2463 kmalloc(sizeof(struct list_head) * s->objects, GFP_KERNEL);
2464 unsigned long flags;
2466 if (!slabs_by_inuse)
2470 for_each_online_node(node) {
2471 n = get_node(s, node);
2476 for (i = 0; i < s->objects; i++)
2477 INIT_LIST_HEAD(slabs_by_inuse + i);
2479 spin_lock_irqsave(&n->list_lock, flags);
2482 * Build lists indexed by the items in use in each slab.
2484 * Note that concurrent frees may occur while we hold the
2485 * list_lock. page->inuse here is the upper limit.
2487 list_for_each_entry_safe(page, t, &n->partial, lru) {
2488 if (!page->inuse && slab_trylock(page)) {
2490 * Must hold slab lock here because slab_free
2491 * may have freed the last object and be
2492 * waiting to release the slab.
2494 list_del(&page->lru);
2497 discard_slab(s, page);
2499 if (n->nr_partial > MAX_PARTIAL)
2500 list_move(&page->lru,
2501 slabs_by_inuse + page->inuse);
2505 if (n->nr_partial <= MAX_PARTIAL)
2509 * Rebuild the partial list with the slabs filled up most
2510 * first and the least used slabs at the end.
2512 for (i = s->objects - 1; i >= 0; i--)
2513 list_splice(slabs_by_inuse + i, n->partial.prev);
2516 spin_unlock_irqrestore(&n->list_lock, flags);
2519 kfree(slabs_by_inuse);
2522 EXPORT_SYMBOL(kmem_cache_shrink);
2524 /********************************************************************
2525 * Basic setup of slabs
2526 *******************************************************************/
2528 void __init kmem_cache_init(void)
2535 * Must first have the slab cache available for the allocations of the
2536 * struct kmem_cache_node's. There is special bootstrap code in
2537 * kmem_cache_open for slab_state == DOWN.
2539 create_kmalloc_cache(&kmalloc_caches[0], "kmem_cache_node",
2540 sizeof(struct kmem_cache_node), GFP_KERNEL);
2541 kmalloc_caches[0].refcount = -1;
2545 /* Able to allocate the per node structures */
2546 slab_state = PARTIAL;
2548 /* Caches that are not of the two-to-the-power-of size */
2549 if (KMALLOC_MIN_SIZE <= 64) {
2550 create_kmalloc_cache(&kmalloc_caches[1],
2551 "kmalloc-96", 96, GFP_KERNEL);
2554 if (KMALLOC_MIN_SIZE <= 128) {
2555 create_kmalloc_cache(&kmalloc_caches[2],
2556 "kmalloc-192", 192, GFP_KERNEL);
2560 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++) {
2561 create_kmalloc_cache(&kmalloc_caches[i],
2562 "kmalloc", 1 << i, GFP_KERNEL);
2568 * Patch up the size_index table if we have strange large alignment
2569 * requirements for the kmalloc array. This is only the case for
2570 * mips it seems. The standard arches will not generate any code here.
2572 * Largest permitted alignment is 256 bytes due to the way we
2573 * handle the index determination for the smaller caches.
2575 * Make sure that nothing crazy happens if someone starts tinkering
2576 * around with ARCH_KMALLOC_MINALIGN
2578 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
2579 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
2581 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8)
2582 size_index[(i - 1) / 8] = KMALLOC_SHIFT_LOW;
2586 /* Provide the correct kmalloc names now that the caches are up */
2587 for (i = KMALLOC_SHIFT_LOW; i <= KMALLOC_SHIFT_HIGH; i++)
2588 kmalloc_caches[i]. name =
2589 kasprintf(GFP_KERNEL, "kmalloc-%d", 1 << i);
2592 register_cpu_notifier(&slab_notifier);
2595 kmem_size = offsetof(struct kmem_cache, cpu_slab) +
2596 nr_cpu_ids * sizeof(struct page *);
2598 printk(KERN_INFO "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
2599 " CPUs=%d, Nodes=%d\n",
2600 caches, cache_line_size(),
2601 slub_min_order, slub_max_order, slub_min_objects,
2602 nr_cpu_ids, nr_node_ids);
2606 * Find a mergeable slab cache
2608 static int slab_unmergeable(struct kmem_cache *s)
2610 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
2617 * We may have set a slab to be unmergeable during bootstrap.
2619 if (s->refcount < 0)
2625 static struct kmem_cache *find_mergeable(size_t size,
2626 size_t align, unsigned long flags,
2627 void (*ctor)(void *, struct kmem_cache *, unsigned long))
2629 struct kmem_cache *s;
2631 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
2637 size = ALIGN(size, sizeof(void *));
2638 align = calculate_alignment(flags, align, size);
2639 size = ALIGN(size, align);
2641 list_for_each_entry(s, &slab_caches, list) {
2642 if (slab_unmergeable(s))
2648 if (((flags | slub_debug) & SLUB_MERGE_SAME) !=
2649 (s->flags & SLUB_MERGE_SAME))
2652 * Check if alignment is compatible.
2653 * Courtesy of Adrian Drzewiecki
2655 if ((s->size & ~(align -1)) != s->size)
2658 if (s->size - size >= sizeof(void *))
2666 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
2667 size_t align, unsigned long flags,
2668 void (*ctor)(void *, struct kmem_cache *, unsigned long),
2669 void (*dtor)(void *, struct kmem_cache *, unsigned long))
2671 struct kmem_cache *s;
2674 down_write(&slub_lock);
2675 s = find_mergeable(size, align, flags, ctor);
2679 * Adjust the object sizes so that we clear
2680 * the complete object on kzalloc.
2682 s->objsize = max(s->objsize, (int)size);
2683 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
2684 if (sysfs_slab_alias(s, name))
2687 s = kmalloc(kmem_size, GFP_KERNEL);
2688 if (s && kmem_cache_open(s, GFP_KERNEL, name,
2689 size, align, flags, ctor)) {
2690 if (sysfs_slab_add(s)) {
2694 list_add(&s->list, &slab_caches);
2698 up_write(&slub_lock);
2702 up_write(&slub_lock);
2703 if (flags & SLAB_PANIC)
2704 panic("Cannot create slabcache %s\n", name);
2709 EXPORT_SYMBOL(kmem_cache_create);
2711 void *kmem_cache_zalloc(struct kmem_cache *s, gfp_t flags)
2715 x = slab_alloc(s, flags, -1, __builtin_return_address(0), 0);
2717 memset(x, 0, s->objsize);
2720 EXPORT_SYMBOL(kmem_cache_zalloc);
2724 * Use the cpu notifier to insure that the cpu slabs are flushed when
2727 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
2728 unsigned long action, void *hcpu)
2730 long cpu = (long)hcpu;
2731 struct kmem_cache *s;
2732 unsigned long flags;
2735 case CPU_UP_CANCELED:
2736 case CPU_UP_CANCELED_FROZEN:
2738 case CPU_DEAD_FROZEN:
2739 down_read(&slub_lock);
2740 list_for_each_entry(s, &slab_caches, list) {
2741 local_irq_save(flags);
2742 __flush_cpu_slab(s, cpu);
2743 local_irq_restore(flags);
2745 up_read(&slub_lock);
2753 static struct notifier_block __cpuinitdata slab_notifier =
2754 { &slab_cpuup_callback, NULL, 0 };
2758 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, void *caller)
2760 struct kmem_cache *s = get_slab(size, gfpflags);
2762 if (ZERO_OR_NULL_PTR(s))
2765 return slab_alloc(s, gfpflags, -1, caller, size);
2768 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
2769 int node, void *caller)
2771 struct kmem_cache *s = get_slab(size, gfpflags);
2773 if (ZERO_OR_NULL_PTR(s))
2776 return slab_alloc(s, gfpflags, node, caller, size);
2779 #if defined(CONFIG_SYSFS) && defined(CONFIG_SLUB_DEBUG)
2780 static int validate_slab(struct kmem_cache *s, struct page *page)
2783 void *addr = page_address(page);
2784 DECLARE_BITMAP(map, s->objects);
2786 if (!check_slab(s, page) ||
2787 !on_freelist(s, page, NULL))
2790 /* Now we know that a valid freelist exists */
2791 bitmap_zero(map, s->objects);
2793 for_each_free_object(p, s, page->freelist) {
2794 set_bit(slab_index(p, s, addr), map);
2795 if (!check_object(s, page, p, 0))
2799 for_each_object(p, s, addr)
2800 if (!test_bit(slab_index(p, s, addr), map))
2801 if (!check_object(s, page, p, 1))
2806 static void validate_slab_slab(struct kmem_cache *s, struct page *page)
2808 if (slab_trylock(page)) {
2809 validate_slab(s, page);
2812 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
2815 if (s->flags & DEBUG_DEFAULT_FLAGS) {
2816 if (!SlabDebug(page))
2817 printk(KERN_ERR "SLUB %s: SlabDebug not set "
2818 "on slab 0x%p\n", s->name, page);
2820 if (SlabDebug(page))
2821 printk(KERN_ERR "SLUB %s: SlabDebug set on "
2822 "slab 0x%p\n", s->name, page);
2826 static int validate_slab_node(struct kmem_cache *s, struct kmem_cache_node *n)
2828 unsigned long count = 0;
2830 unsigned long flags;
2832 spin_lock_irqsave(&n->list_lock, flags);
2834 list_for_each_entry(page, &n->partial, lru) {
2835 validate_slab_slab(s, page);
2838 if (count != n->nr_partial)
2839 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
2840 "counter=%ld\n", s->name, count, n->nr_partial);
2842 if (!(s->flags & SLAB_STORE_USER))
2845 list_for_each_entry(page, &n->full, lru) {
2846 validate_slab_slab(s, page);
2849 if (count != atomic_long_read(&n->nr_slabs))
2850 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
2851 "counter=%ld\n", s->name, count,
2852 atomic_long_read(&n->nr_slabs));
2855 spin_unlock_irqrestore(&n->list_lock, flags);
2859 static unsigned long validate_slab_cache(struct kmem_cache *s)
2862 unsigned long count = 0;
2865 for_each_online_node(node) {
2866 struct kmem_cache_node *n = get_node(s, node);
2868 count += validate_slab_node(s, n);
2873 #ifdef SLUB_RESILIENCY_TEST
2874 static void resiliency_test(void)
2878 printk(KERN_ERR "SLUB resiliency testing\n");
2879 printk(KERN_ERR "-----------------------\n");
2880 printk(KERN_ERR "A. Corruption after allocation\n");
2882 p = kzalloc(16, GFP_KERNEL);
2884 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
2885 " 0x12->0x%p\n\n", p + 16);
2887 validate_slab_cache(kmalloc_caches + 4);
2889 /* Hmmm... The next two are dangerous */
2890 p = kzalloc(32, GFP_KERNEL);
2891 p[32 + sizeof(void *)] = 0x34;
2892 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
2893 " 0x34 -> -0x%p\n", p);
2894 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2896 validate_slab_cache(kmalloc_caches + 5);
2897 p = kzalloc(64, GFP_KERNEL);
2898 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
2900 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
2902 printk(KERN_ERR "If allocated object is overwritten then not detectable\n\n");
2903 validate_slab_cache(kmalloc_caches + 6);
2905 printk(KERN_ERR "\nB. Corruption after free\n");
2906 p = kzalloc(128, GFP_KERNEL);
2909 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
2910 validate_slab_cache(kmalloc_caches + 7);
2912 p = kzalloc(256, GFP_KERNEL);
2915 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n", p);
2916 validate_slab_cache(kmalloc_caches + 8);
2918 p = kzalloc(512, GFP_KERNEL);
2921 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
2922 validate_slab_cache(kmalloc_caches + 9);
2925 static void resiliency_test(void) {};
2929 * Generate lists of code addresses where slabcache objects are allocated
2934 unsigned long count;
2947 unsigned long count;
2948 struct location *loc;
2951 static void free_loc_track(struct loc_track *t)
2954 free_pages((unsigned long)t->loc,
2955 get_order(sizeof(struct location) * t->max));
2958 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
2963 order = get_order(sizeof(struct location) * max);
2965 l = (void *)__get_free_pages(flags, order);
2970 memcpy(l, t->loc, sizeof(struct location) * t->count);
2978 static int add_location(struct loc_track *t, struct kmem_cache *s,
2979 const struct track *track)
2981 long start, end, pos;
2984 unsigned long age = jiffies - track->when;
2990 pos = start + (end - start + 1) / 2;
2993 * There is nothing at "end". If we end up there
2994 * we need to add something to before end.
2999 caddr = t->loc[pos].addr;
3000 if (track->addr == caddr) {
3006 if (age < l->min_time)
3008 if (age > l->max_time)
3011 if (track->pid < l->min_pid)
3012 l->min_pid = track->pid;
3013 if (track->pid > l->max_pid)
3014 l->max_pid = track->pid;
3016 cpu_set(track->cpu, l->cpus);
3018 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3022 if (track->addr < caddr)
3029 * Not found. Insert new tracking element.
3031 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3037 (t->count - pos) * sizeof(struct location));
3040 l->addr = track->addr;
3044 l->min_pid = track->pid;
3045 l->max_pid = track->pid;
3046 cpus_clear(l->cpus);
3047 cpu_set(track->cpu, l->cpus);
3048 nodes_clear(l->nodes);
3049 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3053 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3054 struct page *page, enum track_item alloc)
3056 void *addr = page_address(page);
3057 DECLARE_BITMAP(map, s->objects);
3060 bitmap_zero(map, s->objects);
3061 for_each_free_object(p, s, page->freelist)
3062 set_bit(slab_index(p, s, addr), map);
3064 for_each_object(p, s, addr)
3065 if (!test_bit(slab_index(p, s, addr), map))
3066 add_location(t, s, get_track(s, p, alloc));
3069 static int list_locations(struct kmem_cache *s, char *buf,
3070 enum track_item alloc)
3074 struct loc_track t = { 0, 0, NULL };
3077 if (!alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3079 return sprintf(buf, "Out of memory\n");
3081 /* Push back cpu slabs */
3084 for_each_online_node(node) {
3085 struct kmem_cache_node *n = get_node(s, node);
3086 unsigned long flags;
3089 if (!atomic_read(&n->nr_slabs))
3092 spin_lock_irqsave(&n->list_lock, flags);
3093 list_for_each_entry(page, &n->partial, lru)
3094 process_slab(&t, s, page, alloc);
3095 list_for_each_entry(page, &n->full, lru)
3096 process_slab(&t, s, page, alloc);
3097 spin_unlock_irqrestore(&n->list_lock, flags);
3100 for (i = 0; i < t.count; i++) {
3101 struct location *l = &t.loc[i];
3103 if (n > PAGE_SIZE - 100)
3105 n += sprintf(buf + n, "%7ld ", l->count);
3108 n += sprint_symbol(buf + n, (unsigned long)l->addr);
3110 n += sprintf(buf + n, "<not-available>");
3112 if (l->sum_time != l->min_time) {
3113 unsigned long remainder;
3115 n += sprintf(buf + n, " age=%ld/%ld/%ld",
3117 div_long_long_rem(l->sum_time, l->count, &remainder),
3120 n += sprintf(buf + n, " age=%ld",
3123 if (l->min_pid != l->max_pid)
3124 n += sprintf(buf + n, " pid=%ld-%ld",
3125 l->min_pid, l->max_pid);
3127 n += sprintf(buf + n, " pid=%ld",
3130 if (num_online_cpus() > 1 && !cpus_empty(l->cpus) &&
3131 n < PAGE_SIZE - 60) {
3132 n += sprintf(buf + n, " cpus=");
3133 n += cpulist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3137 if (num_online_nodes() > 1 && !nodes_empty(l->nodes) &&
3138 n < PAGE_SIZE - 60) {
3139 n += sprintf(buf + n, " nodes=");
3140 n += nodelist_scnprintf(buf + n, PAGE_SIZE - n - 50,
3144 n += sprintf(buf + n, "\n");
3149 n += sprintf(buf, "No data\n");
3153 static unsigned long count_partial(struct kmem_cache_node *n)
3155 unsigned long flags;
3156 unsigned long x = 0;
3159 spin_lock_irqsave(&n->list_lock, flags);
3160 list_for_each_entry(page, &n->partial, lru)
3162 spin_unlock_irqrestore(&n->list_lock, flags);
3166 enum slab_stat_type {
3173 #define SO_FULL (1 << SL_FULL)
3174 #define SO_PARTIAL (1 << SL_PARTIAL)
3175 #define SO_CPU (1 << SL_CPU)
3176 #define SO_OBJECTS (1 << SL_OBJECTS)
3178 static unsigned long slab_objects(struct kmem_cache *s,
3179 char *buf, unsigned long flags)
3181 unsigned long total = 0;
3185 unsigned long *nodes;
3186 unsigned long *per_cpu;
3188 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
3189 per_cpu = nodes + nr_node_ids;
3191 for_each_possible_cpu(cpu) {
3192 struct page *page = s->cpu_slab[cpu];
3196 node = page_to_nid(page);
3197 if (flags & SO_CPU) {
3200 if (flags & SO_OBJECTS)
3211 for_each_online_node(node) {
3212 struct kmem_cache_node *n = get_node(s, node);
3214 if (flags & SO_PARTIAL) {
3215 if (flags & SO_OBJECTS)
3216 x = count_partial(n);
3223 if (flags & SO_FULL) {
3224 int full_slabs = atomic_read(&n->nr_slabs)
3228 if (flags & SO_OBJECTS)
3229 x = full_slabs * s->objects;
3237 x = sprintf(buf, "%lu", total);
3239 for_each_online_node(node)
3241 x += sprintf(buf + x, " N%d=%lu",
3245 return x + sprintf(buf + x, "\n");
3248 static int any_slab_objects(struct kmem_cache *s)
3253 for_each_possible_cpu(cpu)
3254 if (s->cpu_slab[cpu])
3257 for_each_node(node) {
3258 struct kmem_cache_node *n = get_node(s, node);
3260 if (n->nr_partial || atomic_read(&n->nr_slabs))
3266 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
3267 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
3269 struct slab_attribute {
3270 struct attribute attr;
3271 ssize_t (*show)(struct kmem_cache *s, char *buf);
3272 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
3275 #define SLAB_ATTR_RO(_name) \
3276 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
3278 #define SLAB_ATTR(_name) \
3279 static struct slab_attribute _name##_attr = \
3280 __ATTR(_name, 0644, _name##_show, _name##_store)
3282 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
3284 return sprintf(buf, "%d\n", s->size);
3286 SLAB_ATTR_RO(slab_size);
3288 static ssize_t align_show(struct kmem_cache *s, char *buf)
3290 return sprintf(buf, "%d\n", s->align);
3292 SLAB_ATTR_RO(align);
3294 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
3296 return sprintf(buf, "%d\n", s->objsize);
3298 SLAB_ATTR_RO(object_size);
3300 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
3302 return sprintf(buf, "%d\n", s->objects);
3304 SLAB_ATTR_RO(objs_per_slab);
3306 static ssize_t order_show(struct kmem_cache *s, char *buf)
3308 return sprintf(buf, "%d\n", s->order);
3310 SLAB_ATTR_RO(order);
3312 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
3315 int n = sprint_symbol(buf, (unsigned long)s->ctor);
3317 return n + sprintf(buf + n, "\n");
3323 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
3325 return sprintf(buf, "%d\n", s->refcount - 1);
3327 SLAB_ATTR_RO(aliases);
3329 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
3331 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU);
3333 SLAB_ATTR_RO(slabs);
3335 static ssize_t partial_show(struct kmem_cache *s, char *buf)
3337 return slab_objects(s, buf, SO_PARTIAL);
3339 SLAB_ATTR_RO(partial);
3341 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
3343 return slab_objects(s, buf, SO_CPU);
3345 SLAB_ATTR_RO(cpu_slabs);
3347 static ssize_t objects_show(struct kmem_cache *s, char *buf)
3349 return slab_objects(s, buf, SO_FULL|SO_PARTIAL|SO_CPU|SO_OBJECTS);
3351 SLAB_ATTR_RO(objects);
3353 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
3355 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
3358 static ssize_t sanity_checks_store(struct kmem_cache *s,
3359 const char *buf, size_t length)
3361 s->flags &= ~SLAB_DEBUG_FREE;
3363 s->flags |= SLAB_DEBUG_FREE;
3366 SLAB_ATTR(sanity_checks);
3368 static ssize_t trace_show(struct kmem_cache *s, char *buf)
3370 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
3373 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
3376 s->flags &= ~SLAB_TRACE;
3378 s->flags |= SLAB_TRACE;
3383 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
3385 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
3388 static ssize_t reclaim_account_store(struct kmem_cache *s,
3389 const char *buf, size_t length)
3391 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
3393 s->flags |= SLAB_RECLAIM_ACCOUNT;
3396 SLAB_ATTR(reclaim_account);
3398 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
3400 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
3402 SLAB_ATTR_RO(hwcache_align);
3404 #ifdef CONFIG_ZONE_DMA
3405 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
3407 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
3409 SLAB_ATTR_RO(cache_dma);
3412 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
3414 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
3416 SLAB_ATTR_RO(destroy_by_rcu);
3418 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
3420 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
3423 static ssize_t red_zone_store(struct kmem_cache *s,
3424 const char *buf, size_t length)
3426 if (any_slab_objects(s))
3429 s->flags &= ~SLAB_RED_ZONE;
3431 s->flags |= SLAB_RED_ZONE;
3435 SLAB_ATTR(red_zone);
3437 static ssize_t poison_show(struct kmem_cache *s, char *buf)
3439 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
3442 static ssize_t poison_store(struct kmem_cache *s,
3443 const char *buf, size_t length)
3445 if (any_slab_objects(s))
3448 s->flags &= ~SLAB_POISON;
3450 s->flags |= SLAB_POISON;
3456 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
3458 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
3461 static ssize_t store_user_store(struct kmem_cache *s,
3462 const char *buf, size_t length)
3464 if (any_slab_objects(s))
3467 s->flags &= ~SLAB_STORE_USER;
3469 s->flags |= SLAB_STORE_USER;
3473 SLAB_ATTR(store_user);
3475 static ssize_t validate_show(struct kmem_cache *s, char *buf)
3480 static ssize_t validate_store(struct kmem_cache *s,
3481 const char *buf, size_t length)
3484 validate_slab_cache(s);
3489 SLAB_ATTR(validate);
3491 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
3496 static ssize_t shrink_store(struct kmem_cache *s,
3497 const char *buf, size_t length)
3499 if (buf[0] == '1') {
3500 int rc = kmem_cache_shrink(s);
3510 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
3512 if (!(s->flags & SLAB_STORE_USER))
3514 return list_locations(s, buf, TRACK_ALLOC);
3516 SLAB_ATTR_RO(alloc_calls);
3518 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
3520 if (!(s->flags & SLAB_STORE_USER))
3522 return list_locations(s, buf, TRACK_FREE);
3524 SLAB_ATTR_RO(free_calls);
3527 static ssize_t defrag_ratio_show(struct kmem_cache *s, char *buf)
3529 return sprintf(buf, "%d\n", s->defrag_ratio / 10);
3532 static ssize_t defrag_ratio_store(struct kmem_cache *s,
3533 const char *buf, size_t length)
3535 int n = simple_strtoul(buf, NULL, 10);
3538 s->defrag_ratio = n * 10;
3541 SLAB_ATTR(defrag_ratio);
3544 static struct attribute * slab_attrs[] = {
3545 &slab_size_attr.attr,
3546 &object_size_attr.attr,
3547 &objs_per_slab_attr.attr,
3552 &cpu_slabs_attr.attr,
3556 &sanity_checks_attr.attr,
3558 &hwcache_align_attr.attr,
3559 &reclaim_account_attr.attr,
3560 &destroy_by_rcu_attr.attr,
3561 &red_zone_attr.attr,
3563 &store_user_attr.attr,
3564 &validate_attr.attr,
3566 &alloc_calls_attr.attr,
3567 &free_calls_attr.attr,
3568 #ifdef CONFIG_ZONE_DMA
3569 &cache_dma_attr.attr,
3572 &defrag_ratio_attr.attr,
3577 static struct attribute_group slab_attr_group = {
3578 .attrs = slab_attrs,
3581 static ssize_t slab_attr_show(struct kobject *kobj,
3582 struct attribute *attr,
3585 struct slab_attribute *attribute;
3586 struct kmem_cache *s;
3589 attribute = to_slab_attr(attr);
3592 if (!attribute->show)
3595 err = attribute->show(s, buf);
3600 static ssize_t slab_attr_store(struct kobject *kobj,
3601 struct attribute *attr,
3602 const char *buf, size_t len)
3604 struct slab_attribute *attribute;
3605 struct kmem_cache *s;
3608 attribute = to_slab_attr(attr);
3611 if (!attribute->store)
3614 err = attribute->store(s, buf, len);
3619 static struct sysfs_ops slab_sysfs_ops = {
3620 .show = slab_attr_show,
3621 .store = slab_attr_store,
3624 static struct kobj_type slab_ktype = {
3625 .sysfs_ops = &slab_sysfs_ops,
3628 static int uevent_filter(struct kset *kset, struct kobject *kobj)
3630 struct kobj_type *ktype = get_ktype(kobj);
3632 if (ktype == &slab_ktype)
3637 static struct kset_uevent_ops slab_uevent_ops = {
3638 .filter = uevent_filter,
3641 static decl_subsys(slab, &slab_ktype, &slab_uevent_ops);
3643 #define ID_STR_LENGTH 64
3645 /* Create a unique string id for a slab cache:
3647 * :[flags-]size:[memory address of kmemcache]
3649 static char *create_unique_id(struct kmem_cache *s)
3651 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
3658 * First flags affecting slabcache operations. We will only
3659 * get here for aliasable slabs so we do not need to support
3660 * too many flags. The flags here must cover all flags that
3661 * are matched during merging to guarantee that the id is
3664 if (s->flags & SLAB_CACHE_DMA)
3666 if (s->flags & SLAB_RECLAIM_ACCOUNT)
3668 if (s->flags & SLAB_DEBUG_FREE)
3672 p += sprintf(p, "%07d", s->size);
3673 BUG_ON(p > name + ID_STR_LENGTH - 1);
3677 static int sysfs_slab_add(struct kmem_cache *s)
3683 if (slab_state < SYSFS)
3684 /* Defer until later */
3687 unmergeable = slab_unmergeable(s);
3690 * Slabcache can never be merged so we can use the name proper.
3691 * This is typically the case for debug situations. In that
3692 * case we can catch duplicate names easily.
3694 sysfs_remove_link(&slab_subsys.kobj, s->name);
3698 * Create a unique name for the slab as a target
3701 name = create_unique_id(s);
3704 kobj_set_kset_s(s, slab_subsys);
3705 kobject_set_name(&s->kobj, name);
3706 kobject_init(&s->kobj);
3707 err = kobject_add(&s->kobj);
3711 err = sysfs_create_group(&s->kobj, &slab_attr_group);
3714 kobject_uevent(&s->kobj, KOBJ_ADD);
3716 /* Setup first alias */
3717 sysfs_slab_alias(s, s->name);
3723 static void sysfs_slab_remove(struct kmem_cache *s)
3725 kobject_uevent(&s->kobj, KOBJ_REMOVE);
3726 kobject_del(&s->kobj);
3730 * Need to buffer aliases during bootup until sysfs becomes
3731 * available lest we loose that information.
3733 struct saved_alias {
3734 struct kmem_cache *s;
3736 struct saved_alias *next;
3739 static struct saved_alias *alias_list;
3741 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
3743 struct saved_alias *al;
3745 if (slab_state == SYSFS) {
3747 * If we have a leftover link then remove it.
3749 sysfs_remove_link(&slab_subsys.kobj, name);
3750 return sysfs_create_link(&slab_subsys.kobj,
3754 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
3760 al->next = alias_list;
3765 static int __init slab_sysfs_init(void)
3767 struct kmem_cache *s;
3770 err = subsystem_register(&slab_subsys);
3772 printk(KERN_ERR "Cannot register slab subsystem.\n");
3778 list_for_each_entry(s, &slab_caches, list) {
3779 err = sysfs_slab_add(s);
3783 while (alias_list) {
3784 struct saved_alias *al = alias_list;
3786 alias_list = alias_list->next;
3787 err = sysfs_slab_alias(al->s, al->name);
3796 __initcall(slab_sysfs_init);