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
12 #include <linux/swap.h> /* struct reclaim_state */
13 #include <linux/module.h>
14 #include <linux/bit_spinlock.h>
15 #include <linux/interrupt.h>
16 #include <linux/bitops.h>
17 #include <linux/slab.h>
18 #include <linux/proc_fs.h>
19 #include <linux/seq_file.h>
20 #include <linux/kmemcheck.h>
21 #include <linux/cpu.h>
22 #include <linux/cpuset.h>
23 #include <linux/mempolicy.h>
24 #include <linux/ctype.h>
25 #include <linux/debugobjects.h>
26 #include <linux/kallsyms.h>
27 #include <linux/memory.h>
28 #include <linux/math64.h>
29 #include <linux/fault-inject.h>
30 #include <linux/stacktrace.h>
32 #include <trace/events/kmem.h>
39 * The slab_lock protects operations on the object of a particular
40 * slab and its metadata in the page struct. If the slab lock
41 * has been taken then no allocations nor frees can be performed
42 * on the objects in the slab nor can the slab be added or removed
43 * from the partial or full lists since this would mean modifying
44 * the page_struct of the slab.
46 * The list_lock protects the partial and full list on each node and
47 * the partial slab counter. If taken then no new slabs may be added or
48 * removed from the lists nor make the number of partial slabs be modified.
49 * (Note that the total number of slabs is an atomic value that may be
50 * modified without taking the list lock).
52 * The list_lock is a centralized lock and thus we avoid taking it as
53 * much as possible. As long as SLUB does not have to handle partial
54 * slabs, operations can continue without any centralized lock. F.e.
55 * allocating a long series of objects that fill up slabs does not require
58 * The lock order is sometimes inverted when we are trying to get a slab
59 * off a list. We take the list_lock and then look for a page on the list
60 * to use. While we do that objects in the slabs may be freed. We can
61 * only operate on the slab if we have also taken the slab_lock. So we use
62 * a slab_trylock() on the slab. If trylock was successful then no frees
63 * can occur anymore and we can use the slab for allocations etc. If the
64 * slab_trylock() does not succeed then frees are in progress in the slab and
65 * we must stay away from it for a while since we may cause a bouncing
66 * cacheline if we try to acquire the lock. So go onto the next slab.
67 * If all pages are busy then we may allocate a new slab instead of reusing
68 * a partial slab. A new slab has no one operating on it and thus there is
69 * no danger of cacheline contention.
71 * Interrupts are disabled during allocation and deallocation in order to
72 * make the slab allocator safe to use in the context of an irq. In addition
73 * interrupts are disabled to ensure that the processor does not change
74 * while handling per_cpu slabs, due to kernel preemption.
76 * SLUB assigns one slab for allocation to each processor.
77 * Allocations only occur from these slabs called cpu slabs.
79 * Slabs with free elements are kept on a partial list and during regular
80 * operations no list for full slabs is used. If an object in a full slab is
81 * freed then the slab will show up again on the partial lists.
82 * We track full slabs for debugging purposes though because otherwise we
83 * cannot scan all objects.
85 * Slabs are freed when they become empty. Teardown and setup is
86 * minimal so we rely on the page allocators per cpu caches for
87 * fast frees and allocs.
89 * Overloading of page flags that are otherwise used for LRU management.
91 * PageActive The slab is frozen and exempt from list processing.
92 * This means that the slab is dedicated to a purpose
93 * such as satisfying allocations for a specific
94 * processor. Objects may be freed in the slab while
95 * it is frozen but slab_free will then skip the usual
96 * list operations. It is up to the processor holding
97 * the slab to integrate the slab into the slab lists
98 * when the slab is no longer needed.
100 * One use of this flag is to mark slabs that are
101 * used for allocations. Then such a slab becomes a cpu
102 * slab. The cpu slab may be equipped with an additional
103 * freelist that allows lockless access to
104 * free objects in addition to the regular freelist
105 * that requires the slab lock.
107 * PageError Slab requires special handling due to debug
108 * options set. This moves slab handling out of
109 * the fast path and disables lockless freelists.
112 #define SLAB_DEBUG_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
113 SLAB_TRACE | SLAB_DEBUG_FREE)
115 static inline int kmem_cache_debug(struct kmem_cache *s)
117 #ifdef CONFIG_SLUB_DEBUG
118 return unlikely(s->flags & SLAB_DEBUG_FLAGS);
125 * Issues still to be resolved:
127 * - Support PAGE_ALLOC_DEBUG. Should be easy to do.
129 * - Variable sizing of the per node arrays
132 /* Enable to test recovery from slab corruption on boot */
133 #undef SLUB_RESILIENCY_TEST
136 * Mininum number of partial slabs. These will be left on the partial
137 * lists even if they are empty. kmem_cache_shrink may reclaim them.
139 #define MIN_PARTIAL 5
142 * Maximum number of desirable partial slabs.
143 * The existence of more partial slabs makes kmem_cache_shrink
144 * sort the partial list by the number of objects in the.
146 #define MAX_PARTIAL 10
148 #define DEBUG_DEFAULT_FLAGS (SLAB_DEBUG_FREE | SLAB_RED_ZONE | \
149 SLAB_POISON | SLAB_STORE_USER)
152 * Debugging flags that require metadata to be stored in the slab. These get
153 * disabled when slub_debug=O is used and a cache's min order increases with
156 #define DEBUG_METADATA_FLAGS (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER)
159 * Set of flags that will prevent slab merging
161 #define SLUB_NEVER_MERGE (SLAB_RED_ZONE | SLAB_POISON | SLAB_STORE_USER | \
162 SLAB_TRACE | SLAB_DESTROY_BY_RCU | SLAB_NOLEAKTRACE | \
165 #define SLUB_MERGE_SAME (SLAB_DEBUG_FREE | SLAB_RECLAIM_ACCOUNT | \
166 SLAB_CACHE_DMA | SLAB_NOTRACK)
169 #define OO_MASK ((1 << OO_SHIFT) - 1)
170 #define MAX_OBJS_PER_PAGE 65535 /* since page.objects is u16 */
172 /* Internal SLUB flags */
173 #define __OBJECT_POISON 0x80000000UL /* Poison object */
175 static int kmem_size = sizeof(struct kmem_cache);
178 static struct notifier_block slab_notifier;
182 DOWN, /* No slab functionality available */
183 PARTIAL, /* Kmem_cache_node works */
184 UP, /* Everything works but does not show up in sysfs */
188 /* A list of all slab caches on the system */
189 static DECLARE_RWSEM(slub_lock);
190 static LIST_HEAD(slab_caches);
193 * Tracking user of a slab.
195 #define TRACK_ADDRS_COUNT 16
197 unsigned long addr; /* Called from address */
198 #ifdef CONFIG_STACKTRACE
199 unsigned long addrs[TRACK_ADDRS_COUNT]; /* Called from address */
201 int cpu; /* Was running on cpu */
202 int pid; /* Pid context */
203 unsigned long when; /* When did the operation occur */
206 enum track_item { TRACK_ALLOC, TRACK_FREE };
209 static int sysfs_slab_add(struct kmem_cache *);
210 static int sysfs_slab_alias(struct kmem_cache *, const char *);
211 static void sysfs_slab_remove(struct kmem_cache *);
214 static inline int sysfs_slab_add(struct kmem_cache *s) { return 0; }
215 static inline int sysfs_slab_alias(struct kmem_cache *s, const char *p)
217 static inline void sysfs_slab_remove(struct kmem_cache *s)
225 static inline void stat(const struct kmem_cache *s, enum stat_item si)
227 #ifdef CONFIG_SLUB_STATS
228 __this_cpu_inc(s->cpu_slab->stat[si]);
232 /********************************************************************
233 * Core slab cache functions
234 *******************************************************************/
236 int slab_is_available(void)
238 return slab_state >= UP;
241 static inline struct kmem_cache_node *get_node(struct kmem_cache *s, int node)
243 return s->node[node];
246 /* Verify that a pointer has an address that is valid within a slab page */
247 static inline int check_valid_pointer(struct kmem_cache *s,
248 struct page *page, const void *object)
255 base = page_address(page);
256 if (object < base || object >= base + page->objects * s->size ||
257 (object - base) % s->size) {
264 static inline void *get_freepointer(struct kmem_cache *s, void *object)
266 return *(void **)(object + s->offset);
269 static inline void *get_freepointer_safe(struct kmem_cache *s, void *object)
273 #ifdef CONFIG_DEBUG_PAGEALLOC
274 probe_kernel_read(&p, (void **)(object + s->offset), sizeof(p));
276 p = get_freepointer(s, object);
281 static inline void set_freepointer(struct kmem_cache *s, void *object, void *fp)
283 *(void **)(object + s->offset) = fp;
286 /* Loop over all objects in a slab */
287 #define for_each_object(__p, __s, __addr, __objects) \
288 for (__p = (__addr); __p < (__addr) + (__objects) * (__s)->size;\
291 /* Determine object index from a given position */
292 static inline int slab_index(void *p, struct kmem_cache *s, void *addr)
294 return (p - addr) / s->size;
297 static inline size_t slab_ksize(const struct kmem_cache *s)
299 #ifdef CONFIG_SLUB_DEBUG
301 * Debugging requires use of the padding between object
302 * and whatever may come after it.
304 if (s->flags & (SLAB_RED_ZONE | SLAB_POISON))
309 * If we have the need to store the freelist pointer
310 * back there or track user information then we can
311 * only use the space before that information.
313 if (s->flags & (SLAB_DESTROY_BY_RCU | SLAB_STORE_USER))
316 * Else we can use all the padding etc for the allocation
321 static inline int order_objects(int order, unsigned long size, int reserved)
323 return ((PAGE_SIZE << order) - reserved) / size;
326 static inline struct kmem_cache_order_objects oo_make(int order,
327 unsigned long size, int reserved)
329 struct kmem_cache_order_objects x = {
330 (order << OO_SHIFT) + order_objects(order, size, reserved)
336 static inline int oo_order(struct kmem_cache_order_objects x)
338 return x.x >> OO_SHIFT;
341 static inline int oo_objects(struct kmem_cache_order_objects x)
343 return x.x & OO_MASK;
346 #ifdef CONFIG_SLUB_DEBUG
348 * Determine a map of object in use on a page.
350 * Slab lock or node listlock must be held to guarantee that the page does
351 * not vanish from under us.
353 static void get_map(struct kmem_cache *s, struct page *page, unsigned long *map)
356 void *addr = page_address(page);
358 for (p = page->freelist; p; p = get_freepointer(s, p))
359 set_bit(slab_index(p, s, addr), map);
365 #ifdef CONFIG_SLUB_DEBUG_ON
366 static int slub_debug = DEBUG_DEFAULT_FLAGS;
368 static int slub_debug;
371 static char *slub_debug_slabs;
372 static int disable_higher_order_debug;
377 static void print_section(char *text, u8 *addr, unsigned int length)
385 for (i = 0; i < length; i++) {
387 printk(KERN_ERR "%8s 0x%p: ", text, addr + i);
390 printk(KERN_CONT " %02x", addr[i]);
392 ascii[offset] = isgraph(addr[i]) ? addr[i] : '.';
394 printk(KERN_CONT " %s\n", ascii);
401 printk(KERN_CONT " ");
405 printk(KERN_CONT " %s\n", ascii);
409 static struct track *get_track(struct kmem_cache *s, void *object,
410 enum track_item alloc)
415 p = object + s->offset + sizeof(void *);
417 p = object + s->inuse;
422 static void set_track(struct kmem_cache *s, void *object,
423 enum track_item alloc, unsigned long addr)
425 struct track *p = get_track(s, object, alloc);
428 #ifdef CONFIG_STACKTRACE
429 struct stack_trace trace;
432 trace.nr_entries = 0;
433 trace.max_entries = TRACK_ADDRS_COUNT;
434 trace.entries = p->addrs;
436 save_stack_trace(&trace);
438 /* See rant in lockdep.c */
439 if (trace.nr_entries != 0 &&
440 trace.entries[trace.nr_entries - 1] == ULONG_MAX)
443 for (i = trace.nr_entries; i < TRACK_ADDRS_COUNT; i++)
447 p->cpu = smp_processor_id();
448 p->pid = current->pid;
451 memset(p, 0, sizeof(struct track));
454 static void init_tracking(struct kmem_cache *s, void *object)
456 if (!(s->flags & SLAB_STORE_USER))
459 set_track(s, object, TRACK_FREE, 0UL);
460 set_track(s, object, TRACK_ALLOC, 0UL);
463 static void print_track(const char *s, struct track *t)
468 printk(KERN_ERR "INFO: %s in %pS age=%lu cpu=%u pid=%d\n",
469 s, (void *)t->addr, jiffies - t->when, t->cpu, t->pid);
470 #ifdef CONFIG_STACKTRACE
473 for (i = 0; i < TRACK_ADDRS_COUNT; i++)
475 printk(KERN_ERR "\t%pS\n", (void *)t->addrs[i]);
482 static void print_tracking(struct kmem_cache *s, void *object)
484 if (!(s->flags & SLAB_STORE_USER))
487 print_track("Allocated", get_track(s, object, TRACK_ALLOC));
488 print_track("Freed", get_track(s, object, TRACK_FREE));
491 static void print_page_info(struct page *page)
493 printk(KERN_ERR "INFO: Slab 0x%p objects=%u used=%u fp=0x%p flags=0x%04lx\n",
494 page, page->objects, page->inuse, page->freelist, page->flags);
498 static void slab_bug(struct kmem_cache *s, char *fmt, ...)
504 vsnprintf(buf, sizeof(buf), fmt, args);
506 printk(KERN_ERR "========================================"
507 "=====================================\n");
508 printk(KERN_ERR "BUG %s: %s\n", s->name, buf);
509 printk(KERN_ERR "----------------------------------------"
510 "-------------------------------------\n\n");
513 static void slab_fix(struct kmem_cache *s, char *fmt, ...)
519 vsnprintf(buf, sizeof(buf), fmt, args);
521 printk(KERN_ERR "FIX %s: %s\n", s->name, buf);
524 static void print_trailer(struct kmem_cache *s, struct page *page, u8 *p)
526 unsigned int off; /* Offset of last byte */
527 u8 *addr = page_address(page);
529 print_tracking(s, p);
531 print_page_info(page);
533 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu fp=0x%p\n\n",
534 p, p - addr, get_freepointer(s, p));
537 print_section("Bytes b4", p - 16, 16);
539 print_section("Object", p, min_t(unsigned long, s->objsize, PAGE_SIZE));
541 if (s->flags & SLAB_RED_ZONE)
542 print_section("Redzone", p + s->objsize,
543 s->inuse - s->objsize);
546 off = s->offset + sizeof(void *);
550 if (s->flags & SLAB_STORE_USER)
551 off += 2 * sizeof(struct track);
554 /* Beginning of the filler is the free pointer */
555 print_section("Padding", p + off, s->size - off);
560 static void object_err(struct kmem_cache *s, struct page *page,
561 u8 *object, char *reason)
563 slab_bug(s, "%s", reason);
564 print_trailer(s, page, object);
567 static void slab_err(struct kmem_cache *s, struct page *page, char *fmt, ...)
573 vsnprintf(buf, sizeof(buf), fmt, args);
575 slab_bug(s, "%s", buf);
576 print_page_info(page);
580 static void init_object(struct kmem_cache *s, void *object, u8 val)
584 if (s->flags & __OBJECT_POISON) {
585 memset(p, POISON_FREE, s->objsize - 1);
586 p[s->objsize - 1] = POISON_END;
589 if (s->flags & SLAB_RED_ZONE)
590 memset(p + s->objsize, val, s->inuse - s->objsize);
593 static u8 *check_bytes8(u8 *start, u8 value, unsigned int bytes)
604 static u8 *check_bytes(u8 *start, u8 value, unsigned int bytes)
607 unsigned int words, prefix;
610 return check_bytes8(start, value, bytes);
612 value64 = value | value << 8 | value << 16 | value << 24;
613 value64 = value64 | value64 << 32;
614 prefix = 8 - ((unsigned long)start) % 8;
617 u8 *r = check_bytes8(start, value, prefix);
627 if (*(u64 *)start != value64)
628 return check_bytes8(start, value, 8);
633 return check_bytes8(start, value, bytes % 8);
636 static void restore_bytes(struct kmem_cache *s, char *message, u8 data,
637 void *from, void *to)
639 slab_fix(s, "Restoring 0x%p-0x%p=0x%x\n", from, to - 1, data);
640 memset(from, data, to - from);
643 static int check_bytes_and_report(struct kmem_cache *s, struct page *page,
644 u8 *object, char *what,
645 u8 *start, unsigned int value, unsigned int bytes)
650 fault = check_bytes(start, value, bytes);
655 while (end > fault && end[-1] == value)
658 slab_bug(s, "%s overwritten", what);
659 printk(KERN_ERR "INFO: 0x%p-0x%p. First byte 0x%x instead of 0x%x\n",
660 fault, end - 1, fault[0], value);
661 print_trailer(s, page, object);
663 restore_bytes(s, what, value, fault, end);
671 * Bytes of the object to be managed.
672 * If the freepointer may overlay the object then the free
673 * pointer is the first word of the object.
675 * Poisoning uses 0x6b (POISON_FREE) and the last byte is
678 * object + s->objsize
679 * Padding to reach word boundary. This is also used for Redzoning.
680 * Padding is extended by another word if Redzoning is enabled and
683 * We fill with 0xbb (RED_INACTIVE) for inactive objects and with
684 * 0xcc (RED_ACTIVE) for objects in use.
687 * Meta data starts here.
689 * A. Free pointer (if we cannot overwrite object on free)
690 * B. Tracking data for SLAB_STORE_USER
691 * C. Padding to reach required alignment boundary or at mininum
692 * one word if debugging is on to be able to detect writes
693 * before the word boundary.
695 * Padding is done using 0x5a (POISON_INUSE)
698 * Nothing is used beyond s->size.
700 * If slabcaches are merged then the objsize and inuse boundaries are mostly
701 * ignored. And therefore no slab options that rely on these boundaries
702 * may be used with merged slabcaches.
705 static int check_pad_bytes(struct kmem_cache *s, struct page *page, u8 *p)
707 unsigned long off = s->inuse; /* The end of info */
710 /* Freepointer is placed after the object. */
711 off += sizeof(void *);
713 if (s->flags & SLAB_STORE_USER)
714 /* We also have user information there */
715 off += 2 * sizeof(struct track);
720 return check_bytes_and_report(s, page, p, "Object padding",
721 p + off, POISON_INUSE, s->size - off);
724 /* Check the pad bytes at the end of a slab page */
725 static int slab_pad_check(struct kmem_cache *s, struct page *page)
733 if (!(s->flags & SLAB_POISON))
736 start = page_address(page);
737 length = (PAGE_SIZE << compound_order(page)) - s->reserved;
738 end = start + length;
739 remainder = length % s->size;
743 fault = check_bytes(end - remainder, POISON_INUSE, remainder);
746 while (end > fault && end[-1] == POISON_INUSE)
749 slab_err(s, page, "Padding overwritten. 0x%p-0x%p", fault, end - 1);
750 print_section("Padding", end - remainder, remainder);
752 restore_bytes(s, "slab padding", POISON_INUSE, end - remainder, end);
756 static int check_object(struct kmem_cache *s, struct page *page,
757 void *object, u8 val)
760 u8 *endobject = object + s->objsize;
762 if (s->flags & SLAB_RED_ZONE) {
763 if (!check_bytes_and_report(s, page, object, "Redzone",
764 endobject, val, s->inuse - s->objsize))
767 if ((s->flags & SLAB_POISON) && s->objsize < s->inuse) {
768 check_bytes_and_report(s, page, p, "Alignment padding",
769 endobject, POISON_INUSE, s->inuse - s->objsize);
773 if (s->flags & SLAB_POISON) {
774 if (val != SLUB_RED_ACTIVE && (s->flags & __OBJECT_POISON) &&
775 (!check_bytes_and_report(s, page, p, "Poison", p,
776 POISON_FREE, s->objsize - 1) ||
777 !check_bytes_and_report(s, page, p, "Poison",
778 p + s->objsize - 1, POISON_END, 1)))
781 * check_pad_bytes cleans up on its own.
783 check_pad_bytes(s, page, p);
786 if (!s->offset && val == SLUB_RED_ACTIVE)
788 * Object and freepointer overlap. Cannot check
789 * freepointer while object is allocated.
793 /* Check free pointer validity */
794 if (!check_valid_pointer(s, page, get_freepointer(s, p))) {
795 object_err(s, page, p, "Freepointer corrupt");
797 * No choice but to zap it and thus lose the remainder
798 * of the free objects in this slab. May cause
799 * another error because the object count is now wrong.
801 set_freepointer(s, p, NULL);
807 static int check_slab(struct kmem_cache *s, struct page *page)
811 VM_BUG_ON(!irqs_disabled());
813 if (!PageSlab(page)) {
814 slab_err(s, page, "Not a valid slab page");
818 maxobj = order_objects(compound_order(page), s->size, s->reserved);
819 if (page->objects > maxobj) {
820 slab_err(s, page, "objects %u > max %u",
821 s->name, page->objects, maxobj);
824 if (page->inuse > page->objects) {
825 slab_err(s, page, "inuse %u > max %u",
826 s->name, page->inuse, page->objects);
829 /* Slab_pad_check fixes things up after itself */
830 slab_pad_check(s, page);
835 * Determine if a certain object on a page is on the freelist. Must hold the
836 * slab lock to guarantee that the chains are in a consistent state.
838 static int on_freelist(struct kmem_cache *s, struct page *page, void *search)
841 void *fp = page->freelist;
843 unsigned long max_objects;
845 while (fp && nr <= page->objects) {
848 if (!check_valid_pointer(s, page, fp)) {
850 object_err(s, page, object,
851 "Freechain corrupt");
852 set_freepointer(s, object, NULL);
855 slab_err(s, page, "Freepointer corrupt");
856 page->freelist = NULL;
857 page->inuse = page->objects;
858 slab_fix(s, "Freelist cleared");
864 fp = get_freepointer(s, object);
868 max_objects = order_objects(compound_order(page), s->size, s->reserved);
869 if (max_objects > MAX_OBJS_PER_PAGE)
870 max_objects = MAX_OBJS_PER_PAGE;
872 if (page->objects != max_objects) {
873 slab_err(s, page, "Wrong number of objects. Found %d but "
874 "should be %d", page->objects, max_objects);
875 page->objects = max_objects;
876 slab_fix(s, "Number of objects adjusted.");
878 if (page->inuse != page->objects - nr) {
879 slab_err(s, page, "Wrong object count. Counter is %d but "
880 "counted were %d", page->inuse, page->objects - nr);
881 page->inuse = page->objects - nr;
882 slab_fix(s, "Object count adjusted.");
884 return search == NULL;
887 static void trace(struct kmem_cache *s, struct page *page, void *object,
890 if (s->flags & SLAB_TRACE) {
891 printk(KERN_INFO "TRACE %s %s 0x%p inuse=%d fp=0x%p\n",
893 alloc ? "alloc" : "free",
898 print_section("Object", (void *)object, s->objsize);
905 * Hooks for other subsystems that check memory allocations. In a typical
906 * production configuration these hooks all should produce no code at all.
908 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
910 flags &= gfp_allowed_mask;
911 lockdep_trace_alloc(flags);
912 might_sleep_if(flags & __GFP_WAIT);
914 return should_failslab(s->objsize, flags, s->flags);
917 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags, void *object)
919 flags &= gfp_allowed_mask;
920 kmemcheck_slab_alloc(s, flags, object, slab_ksize(s));
921 kmemleak_alloc_recursive(object, s->objsize, 1, s->flags, flags);
924 static inline void slab_free_hook(struct kmem_cache *s, void *x)
926 kmemleak_free_recursive(x, s->flags);
929 * Trouble is that we may no longer disable interupts in the fast path
930 * So in order to make the debug calls that expect irqs to be
931 * disabled we need to disable interrupts temporarily.
933 #if defined(CONFIG_KMEMCHECK) || defined(CONFIG_LOCKDEP)
937 local_irq_save(flags);
938 kmemcheck_slab_free(s, x, s->objsize);
939 debug_check_no_locks_freed(x, s->objsize);
940 local_irq_restore(flags);
943 if (!(s->flags & SLAB_DEBUG_OBJECTS))
944 debug_check_no_obj_freed(x, s->objsize);
948 * Tracking of fully allocated slabs for debugging purposes.
950 static void add_full(struct kmem_cache_node *n, struct page *page)
952 spin_lock(&n->list_lock);
953 list_add(&page->lru, &n->full);
954 spin_unlock(&n->list_lock);
957 static void remove_full(struct kmem_cache *s, struct page *page)
959 struct kmem_cache_node *n;
961 if (!(s->flags & SLAB_STORE_USER))
964 n = get_node(s, page_to_nid(page));
966 spin_lock(&n->list_lock);
967 list_del(&page->lru);
968 spin_unlock(&n->list_lock);
971 /* Tracking of the number of slabs for debugging purposes */
972 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
974 struct kmem_cache_node *n = get_node(s, node);
976 return atomic_long_read(&n->nr_slabs);
979 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
981 return atomic_long_read(&n->nr_slabs);
984 static inline void inc_slabs_node(struct kmem_cache *s, int node, int objects)
986 struct kmem_cache_node *n = get_node(s, node);
989 * May be called early in order to allocate a slab for the
990 * kmem_cache_node structure. Solve the chicken-egg
991 * dilemma by deferring the increment of the count during
992 * bootstrap (see early_kmem_cache_node_alloc).
995 atomic_long_inc(&n->nr_slabs);
996 atomic_long_add(objects, &n->total_objects);
999 static inline void dec_slabs_node(struct kmem_cache *s, int node, int objects)
1001 struct kmem_cache_node *n = get_node(s, node);
1003 atomic_long_dec(&n->nr_slabs);
1004 atomic_long_sub(objects, &n->total_objects);
1007 /* Object debug checks for alloc/free paths */
1008 static void setup_object_debug(struct kmem_cache *s, struct page *page,
1011 if (!(s->flags & (SLAB_STORE_USER|SLAB_RED_ZONE|__OBJECT_POISON)))
1014 init_object(s, object, SLUB_RED_INACTIVE);
1015 init_tracking(s, object);
1018 static noinline int alloc_debug_processing(struct kmem_cache *s, struct page *page,
1019 void *object, unsigned long addr)
1021 if (!check_slab(s, page))
1024 if (!on_freelist(s, page, object)) {
1025 object_err(s, page, object, "Object already allocated");
1029 if (!check_valid_pointer(s, page, object)) {
1030 object_err(s, page, object, "Freelist Pointer check fails");
1034 if (!check_object(s, page, object, SLUB_RED_INACTIVE))
1037 /* Success perform special debug activities for allocs */
1038 if (s->flags & SLAB_STORE_USER)
1039 set_track(s, object, TRACK_ALLOC, addr);
1040 trace(s, page, object, 1);
1041 init_object(s, object, SLUB_RED_ACTIVE);
1045 if (PageSlab(page)) {
1047 * If this is a slab page then lets do the best we can
1048 * to avoid issues in the future. Marking all objects
1049 * as used avoids touching the remaining objects.
1051 slab_fix(s, "Marking all objects used");
1052 page->inuse = page->objects;
1053 page->freelist = NULL;
1058 static noinline int free_debug_processing(struct kmem_cache *s,
1059 struct page *page, void *object, unsigned long addr)
1061 if (!check_slab(s, page))
1064 if (!check_valid_pointer(s, page, object)) {
1065 slab_err(s, page, "Invalid object pointer 0x%p", object);
1069 if (on_freelist(s, page, object)) {
1070 object_err(s, page, object, "Object already free");
1074 if (!check_object(s, page, object, SLUB_RED_ACTIVE))
1077 if (unlikely(s != page->slab)) {
1078 if (!PageSlab(page)) {
1079 slab_err(s, page, "Attempt to free object(0x%p) "
1080 "outside of slab", object);
1081 } else if (!page->slab) {
1083 "SLUB <none>: no slab for object 0x%p.\n",
1087 object_err(s, page, object,
1088 "page slab pointer corrupt.");
1092 /* Special debug activities for freeing objects */
1093 if (!PageSlubFrozen(page) && !page->freelist)
1094 remove_full(s, page);
1095 if (s->flags & SLAB_STORE_USER)
1096 set_track(s, object, TRACK_FREE, addr);
1097 trace(s, page, object, 0);
1098 init_object(s, object, SLUB_RED_INACTIVE);
1102 slab_fix(s, "Object at 0x%p not freed", object);
1106 static int __init setup_slub_debug(char *str)
1108 slub_debug = DEBUG_DEFAULT_FLAGS;
1109 if (*str++ != '=' || !*str)
1111 * No options specified. Switch on full debugging.
1117 * No options but restriction on slabs. This means full
1118 * debugging for slabs matching a pattern.
1122 if (tolower(*str) == 'o') {
1124 * Avoid enabling debugging on caches if its minimum order
1125 * would increase as a result.
1127 disable_higher_order_debug = 1;
1134 * Switch off all debugging measures.
1139 * Determine which debug features should be switched on
1141 for (; *str && *str != ','; str++) {
1142 switch (tolower(*str)) {
1144 slub_debug |= SLAB_DEBUG_FREE;
1147 slub_debug |= SLAB_RED_ZONE;
1150 slub_debug |= SLAB_POISON;
1153 slub_debug |= SLAB_STORE_USER;
1156 slub_debug |= SLAB_TRACE;
1159 slub_debug |= SLAB_FAILSLAB;
1162 printk(KERN_ERR "slub_debug option '%c' "
1163 "unknown. skipped\n", *str);
1169 slub_debug_slabs = str + 1;
1174 __setup("slub_debug", setup_slub_debug);
1176 static unsigned long kmem_cache_flags(unsigned long objsize,
1177 unsigned long flags, const char *name,
1178 void (*ctor)(void *))
1181 * Enable debugging if selected on the kernel commandline.
1183 if (slub_debug && (!slub_debug_slabs ||
1184 !strncmp(slub_debug_slabs, name, strlen(slub_debug_slabs))))
1185 flags |= slub_debug;
1190 static inline void setup_object_debug(struct kmem_cache *s,
1191 struct page *page, void *object) {}
1193 static inline int alloc_debug_processing(struct kmem_cache *s,
1194 struct page *page, void *object, unsigned long addr) { return 0; }
1196 static inline int free_debug_processing(struct kmem_cache *s,
1197 struct page *page, void *object, unsigned long addr) { return 0; }
1199 static inline int slab_pad_check(struct kmem_cache *s, struct page *page)
1201 static inline int check_object(struct kmem_cache *s, struct page *page,
1202 void *object, u8 val) { return 1; }
1203 static inline void add_full(struct kmem_cache_node *n, struct page *page) {}
1204 static inline unsigned long kmem_cache_flags(unsigned long objsize,
1205 unsigned long flags, const char *name,
1206 void (*ctor)(void *))
1210 #define slub_debug 0
1212 #define disable_higher_order_debug 0
1214 static inline unsigned long slabs_node(struct kmem_cache *s, int node)
1216 static inline unsigned long node_nr_slabs(struct kmem_cache_node *n)
1218 static inline void inc_slabs_node(struct kmem_cache *s, int node,
1220 static inline void dec_slabs_node(struct kmem_cache *s, int node,
1223 static inline int slab_pre_alloc_hook(struct kmem_cache *s, gfp_t flags)
1226 static inline void slab_post_alloc_hook(struct kmem_cache *s, gfp_t flags,
1229 static inline void slab_free_hook(struct kmem_cache *s, void *x) {}
1231 #endif /* CONFIG_SLUB_DEBUG */
1234 * Slab allocation and freeing
1236 static inline struct page *alloc_slab_page(gfp_t flags, int node,
1237 struct kmem_cache_order_objects oo)
1239 int order = oo_order(oo);
1241 flags |= __GFP_NOTRACK;
1243 if (node == NUMA_NO_NODE)
1244 return alloc_pages(flags, order);
1246 return alloc_pages_exact_node(node, flags, order);
1249 static struct page *allocate_slab(struct kmem_cache *s, gfp_t flags, int node)
1252 struct kmem_cache_order_objects oo = s->oo;
1255 flags |= s->allocflags;
1258 * Let the initial higher-order allocation fail under memory pressure
1259 * so we fall-back to the minimum order allocation.
1261 alloc_gfp = (flags | __GFP_NOWARN | __GFP_NORETRY) & ~__GFP_NOFAIL;
1263 page = alloc_slab_page(alloc_gfp, node, oo);
1264 if (unlikely(!page)) {
1267 * Allocation may have failed due to fragmentation.
1268 * Try a lower order alloc if possible
1270 page = alloc_slab_page(flags, node, oo);
1274 stat(s, ORDER_FALLBACK);
1277 if (kmemcheck_enabled
1278 && !(s->flags & (SLAB_NOTRACK | DEBUG_DEFAULT_FLAGS))) {
1279 int pages = 1 << oo_order(oo);
1281 kmemcheck_alloc_shadow(page, oo_order(oo), flags, node);
1284 * Objects from caches that have a constructor don't get
1285 * cleared when they're allocated, so we need to do it here.
1288 kmemcheck_mark_uninitialized_pages(page, pages);
1290 kmemcheck_mark_unallocated_pages(page, pages);
1293 page->objects = oo_objects(oo);
1294 mod_zone_page_state(page_zone(page),
1295 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1296 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1302 static void setup_object(struct kmem_cache *s, struct page *page,
1305 setup_object_debug(s, page, object);
1306 if (unlikely(s->ctor))
1310 static struct page *new_slab(struct kmem_cache *s, gfp_t flags, int node)
1317 BUG_ON(flags & GFP_SLAB_BUG_MASK);
1319 page = allocate_slab(s,
1320 flags & (GFP_RECLAIM_MASK | GFP_CONSTRAINT_MASK), node);
1324 inc_slabs_node(s, page_to_nid(page), page->objects);
1326 page->flags |= 1 << PG_slab;
1328 start = page_address(page);
1330 if (unlikely(s->flags & SLAB_POISON))
1331 memset(start, POISON_INUSE, PAGE_SIZE << compound_order(page));
1334 for_each_object(p, s, start, page->objects) {
1335 setup_object(s, page, last);
1336 set_freepointer(s, last, p);
1339 setup_object(s, page, last);
1340 set_freepointer(s, last, NULL);
1342 page->freelist = start;
1348 static void __free_slab(struct kmem_cache *s, struct page *page)
1350 int order = compound_order(page);
1351 int pages = 1 << order;
1353 if (kmem_cache_debug(s)) {
1356 slab_pad_check(s, page);
1357 for_each_object(p, s, page_address(page),
1359 check_object(s, page, p, SLUB_RED_INACTIVE);
1362 kmemcheck_free_shadow(page, compound_order(page));
1364 mod_zone_page_state(page_zone(page),
1365 (s->flags & SLAB_RECLAIM_ACCOUNT) ?
1366 NR_SLAB_RECLAIMABLE : NR_SLAB_UNRECLAIMABLE,
1369 __ClearPageSlab(page);
1370 reset_page_mapcount(page);
1371 if (current->reclaim_state)
1372 current->reclaim_state->reclaimed_slab += pages;
1373 __free_pages(page, order);
1376 #define need_reserve_slab_rcu \
1377 (sizeof(((struct page *)NULL)->lru) < sizeof(struct rcu_head))
1379 static void rcu_free_slab(struct rcu_head *h)
1383 if (need_reserve_slab_rcu)
1384 page = virt_to_head_page(h);
1386 page = container_of((struct list_head *)h, struct page, lru);
1388 __free_slab(page->slab, page);
1391 static void free_slab(struct kmem_cache *s, struct page *page)
1393 if (unlikely(s->flags & SLAB_DESTROY_BY_RCU)) {
1394 struct rcu_head *head;
1396 if (need_reserve_slab_rcu) {
1397 int order = compound_order(page);
1398 int offset = (PAGE_SIZE << order) - s->reserved;
1400 VM_BUG_ON(s->reserved != sizeof(*head));
1401 head = page_address(page) + offset;
1404 * RCU free overloads the RCU head over the LRU
1406 head = (void *)&page->lru;
1409 call_rcu(head, rcu_free_slab);
1411 __free_slab(s, page);
1414 static void discard_slab(struct kmem_cache *s, struct page *page)
1416 dec_slabs_node(s, page_to_nid(page), page->objects);
1421 * Per slab locking using the pagelock
1423 static __always_inline void slab_lock(struct page *page)
1425 bit_spin_lock(PG_locked, &page->flags);
1428 static __always_inline void slab_unlock(struct page *page)
1430 __bit_spin_unlock(PG_locked, &page->flags);
1433 static __always_inline int slab_trylock(struct page *page)
1437 rc = bit_spin_trylock(PG_locked, &page->flags);
1442 * Management of partially allocated slabs
1444 static void add_partial(struct kmem_cache_node *n,
1445 struct page *page, int tail)
1447 spin_lock(&n->list_lock);
1450 list_add_tail(&page->lru, &n->partial);
1452 list_add(&page->lru, &n->partial);
1453 spin_unlock(&n->list_lock);
1456 static inline void __remove_partial(struct kmem_cache_node *n,
1459 list_del(&page->lru);
1463 static void remove_partial(struct kmem_cache *s, struct page *page)
1465 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1467 spin_lock(&n->list_lock);
1468 __remove_partial(n, page);
1469 spin_unlock(&n->list_lock);
1473 * Lock slab and remove from the partial list.
1475 * Must hold list_lock.
1477 static inline int lock_and_freeze_slab(struct kmem_cache_node *n,
1480 if (slab_trylock(page)) {
1481 __remove_partial(n, page);
1482 __SetPageSlubFrozen(page);
1489 * Try to allocate a partial slab from a specific node.
1491 static struct page *get_partial_node(struct kmem_cache_node *n)
1496 * Racy check. If we mistakenly see no partial slabs then we
1497 * just allocate an empty slab. If we mistakenly try to get a
1498 * partial slab and there is none available then get_partials()
1501 if (!n || !n->nr_partial)
1504 spin_lock(&n->list_lock);
1505 list_for_each_entry(page, &n->partial, lru)
1506 if (lock_and_freeze_slab(n, page))
1510 spin_unlock(&n->list_lock);
1515 * Get a page from somewhere. Search in increasing NUMA distances.
1517 static struct page *get_any_partial(struct kmem_cache *s, gfp_t flags)
1520 struct zonelist *zonelist;
1523 enum zone_type high_zoneidx = gfp_zone(flags);
1527 * The defrag ratio allows a configuration of the tradeoffs between
1528 * inter node defragmentation and node local allocations. A lower
1529 * defrag_ratio increases the tendency to do local allocations
1530 * instead of attempting to obtain partial slabs from other nodes.
1532 * If the defrag_ratio is set to 0 then kmalloc() always
1533 * returns node local objects. If the ratio is higher then kmalloc()
1534 * may return off node objects because partial slabs are obtained
1535 * from other nodes and filled up.
1537 * If /sys/kernel/slab/xx/defrag_ratio is set to 100 (which makes
1538 * defrag_ratio = 1000) then every (well almost) allocation will
1539 * first attempt to defrag slab caches on other nodes. This means
1540 * scanning over all nodes to look for partial slabs which may be
1541 * expensive if we do it every time we are trying to find a slab
1542 * with available objects.
1544 if (!s->remote_node_defrag_ratio ||
1545 get_cycles() % 1024 > s->remote_node_defrag_ratio)
1549 zonelist = node_zonelist(slab_node(current->mempolicy), flags);
1550 for_each_zone_zonelist(zone, z, zonelist, high_zoneidx) {
1551 struct kmem_cache_node *n;
1553 n = get_node(s, zone_to_nid(zone));
1555 if (n && cpuset_zone_allowed_hardwall(zone, flags) &&
1556 n->nr_partial > s->min_partial) {
1557 page = get_partial_node(n);
1570 * Get a partial page, lock it and return it.
1572 static struct page *get_partial(struct kmem_cache *s, gfp_t flags, int node)
1575 int searchnode = (node == NUMA_NO_NODE) ? numa_node_id() : node;
1577 page = get_partial_node(get_node(s, searchnode));
1578 if (page || node != NUMA_NO_NODE)
1581 return get_any_partial(s, flags);
1585 * Move a page back to the lists.
1587 * Must be called with the slab lock held.
1589 * On exit the slab lock will have been dropped.
1591 static void unfreeze_slab(struct kmem_cache *s, struct page *page, int tail)
1594 struct kmem_cache_node *n = get_node(s, page_to_nid(page));
1596 __ClearPageSlubFrozen(page);
1599 if (page->freelist) {
1600 add_partial(n, page, tail);
1601 stat(s, tail ? DEACTIVATE_TO_TAIL : DEACTIVATE_TO_HEAD);
1603 stat(s, DEACTIVATE_FULL);
1604 if (kmem_cache_debug(s) && (s->flags & SLAB_STORE_USER))
1609 stat(s, DEACTIVATE_EMPTY);
1610 if (n->nr_partial < s->min_partial) {
1612 * Adding an empty slab to the partial slabs in order
1613 * to avoid page allocator overhead. This slab needs
1614 * to come after the other slabs with objects in
1615 * so that the others get filled first. That way the
1616 * size of the partial list stays small.
1618 * kmem_cache_shrink can reclaim any empty slabs from
1621 add_partial(n, page, 1);
1626 discard_slab(s, page);
1631 #ifdef CONFIG_PREEMPT
1633 * Calculate the next globally unique transaction for disambiguiation
1634 * during cmpxchg. The transactions start with the cpu number and are then
1635 * incremented by CONFIG_NR_CPUS.
1637 #define TID_STEP roundup_pow_of_two(CONFIG_NR_CPUS)
1640 * No preemption supported therefore also no need to check for
1646 static inline unsigned long next_tid(unsigned long tid)
1648 return tid + TID_STEP;
1651 static inline unsigned int tid_to_cpu(unsigned long tid)
1653 return tid % TID_STEP;
1656 static inline unsigned long tid_to_event(unsigned long tid)
1658 return tid / TID_STEP;
1661 static inline unsigned int init_tid(int cpu)
1666 static inline void note_cmpxchg_failure(const char *n,
1667 const struct kmem_cache *s, unsigned long tid)
1669 #ifdef SLUB_DEBUG_CMPXCHG
1670 unsigned long actual_tid = __this_cpu_read(s->cpu_slab->tid);
1672 printk(KERN_INFO "%s %s: cmpxchg redo ", n, s->name);
1674 #ifdef CONFIG_PREEMPT
1675 if (tid_to_cpu(tid) != tid_to_cpu(actual_tid))
1676 printk("due to cpu change %d -> %d\n",
1677 tid_to_cpu(tid), tid_to_cpu(actual_tid));
1680 if (tid_to_event(tid) != tid_to_event(actual_tid))
1681 printk("due to cpu running other code. Event %ld->%ld\n",
1682 tid_to_event(tid), tid_to_event(actual_tid));
1684 printk("for unknown reason: actual=%lx was=%lx target=%lx\n",
1685 actual_tid, tid, next_tid(tid));
1687 stat(s, CMPXCHG_DOUBLE_CPU_FAIL);
1690 void init_kmem_cache_cpus(struct kmem_cache *s)
1694 for_each_possible_cpu(cpu)
1695 per_cpu_ptr(s->cpu_slab, cpu)->tid = init_tid(cpu);
1698 * Remove the cpu slab
1700 static void deactivate_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1703 struct page *page = c->page;
1707 stat(s, DEACTIVATE_REMOTE_FREES);
1709 * Merge cpu freelist into slab freelist. Typically we get here
1710 * because both freelists are empty. So this is unlikely
1713 while (unlikely(c->freelist)) {
1716 tail = 0; /* Hot objects. Put the slab first */
1718 /* Retrieve object from cpu_freelist */
1719 object = c->freelist;
1720 c->freelist = get_freepointer(s, c->freelist);
1722 /* And put onto the regular freelist */
1723 set_freepointer(s, object, page->freelist);
1724 page->freelist = object;
1728 c->tid = next_tid(c->tid);
1729 unfreeze_slab(s, page, tail);
1732 static inline void flush_slab(struct kmem_cache *s, struct kmem_cache_cpu *c)
1734 stat(s, CPUSLAB_FLUSH);
1736 deactivate_slab(s, c);
1742 * Called from IPI handler with interrupts disabled.
1744 static inline void __flush_cpu_slab(struct kmem_cache *s, int cpu)
1746 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
1748 if (likely(c && c->page))
1752 static void flush_cpu_slab(void *d)
1754 struct kmem_cache *s = d;
1756 __flush_cpu_slab(s, smp_processor_id());
1759 static void flush_all(struct kmem_cache *s)
1761 on_each_cpu(flush_cpu_slab, s, 1);
1765 * Check if the objects in a per cpu structure fit numa
1766 * locality expectations.
1768 static inline int node_match(struct kmem_cache_cpu *c, int node)
1771 if (node != NUMA_NO_NODE && c->node != node)
1777 static int count_free(struct page *page)
1779 return page->objects - page->inuse;
1782 static unsigned long count_partial(struct kmem_cache_node *n,
1783 int (*get_count)(struct page *))
1785 unsigned long flags;
1786 unsigned long x = 0;
1789 spin_lock_irqsave(&n->list_lock, flags);
1790 list_for_each_entry(page, &n->partial, lru)
1791 x += get_count(page);
1792 spin_unlock_irqrestore(&n->list_lock, flags);
1796 static inline unsigned long node_nr_objs(struct kmem_cache_node *n)
1798 #ifdef CONFIG_SLUB_DEBUG
1799 return atomic_long_read(&n->total_objects);
1805 static noinline void
1806 slab_out_of_memory(struct kmem_cache *s, gfp_t gfpflags, int nid)
1811 "SLUB: Unable to allocate memory on node %d (gfp=0x%x)\n",
1813 printk(KERN_WARNING " cache: %s, object size: %d, buffer size: %d, "
1814 "default order: %d, min order: %d\n", s->name, s->objsize,
1815 s->size, oo_order(s->oo), oo_order(s->min));
1817 if (oo_order(s->min) > get_order(s->objsize))
1818 printk(KERN_WARNING " %s debugging increased min order, use "
1819 "slub_debug=O to disable.\n", s->name);
1821 for_each_online_node(node) {
1822 struct kmem_cache_node *n = get_node(s, node);
1823 unsigned long nr_slabs;
1824 unsigned long nr_objs;
1825 unsigned long nr_free;
1830 nr_free = count_partial(n, count_free);
1831 nr_slabs = node_nr_slabs(n);
1832 nr_objs = node_nr_objs(n);
1835 " node %d: slabs: %ld, objs: %ld, free: %ld\n",
1836 node, nr_slabs, nr_objs, nr_free);
1841 * Slow path. The lockless freelist is empty or we need to perform
1844 * Interrupts are disabled.
1846 * Processing is still very fast if new objects have been freed to the
1847 * regular freelist. In that case we simply take over the regular freelist
1848 * as the lockless freelist and zap the regular freelist.
1850 * If that is not working then we fall back to the partial lists. We take the
1851 * first element of the freelist as the object to allocate now and move the
1852 * rest of the freelist to the lockless freelist.
1854 * And if we were unable to get a new slab from the partial slab lists then
1855 * we need to allocate a new slab. This is the slowest path since it involves
1856 * a call to the page allocator and the setup of a new slab.
1858 static void *__slab_alloc(struct kmem_cache *s, gfp_t gfpflags, int node,
1859 unsigned long addr, struct kmem_cache_cpu *c)
1863 unsigned long flags;
1865 local_irq_save(flags);
1866 #ifdef CONFIG_PREEMPT
1868 * We may have been preempted and rescheduled on a different
1869 * cpu before disabling interrupts. Need to reload cpu area
1872 c = this_cpu_ptr(s->cpu_slab);
1875 /* We handle __GFP_ZERO in the caller */
1876 gfpflags &= ~__GFP_ZERO;
1883 if (unlikely(!node_match(c, node)))
1886 stat(s, ALLOC_REFILL);
1889 object = page->freelist;
1890 if (unlikely(!object))
1892 if (kmem_cache_debug(s))
1895 c->freelist = get_freepointer(s, object);
1896 page->inuse = page->objects;
1897 page->freelist = NULL;
1900 c->tid = next_tid(c->tid);
1901 local_irq_restore(flags);
1902 stat(s, ALLOC_SLOWPATH);
1906 deactivate_slab(s, c);
1909 page = get_partial(s, gfpflags, node);
1911 stat(s, ALLOC_FROM_PARTIAL);
1912 c->node = page_to_nid(page);
1917 gfpflags &= gfp_allowed_mask;
1918 if (gfpflags & __GFP_WAIT)
1921 page = new_slab(s, gfpflags, node);
1923 if (gfpflags & __GFP_WAIT)
1924 local_irq_disable();
1927 c = __this_cpu_ptr(s->cpu_slab);
1928 stat(s, ALLOC_SLAB);
1933 __SetPageSlubFrozen(page);
1934 c->node = page_to_nid(page);
1938 if (!(gfpflags & __GFP_NOWARN) && printk_ratelimit())
1939 slab_out_of_memory(s, gfpflags, node);
1940 local_irq_restore(flags);
1943 if (!alloc_debug_processing(s, page, object, addr))
1947 page->freelist = get_freepointer(s, object);
1948 deactivate_slab(s, c);
1950 c->node = NUMA_NO_NODE;
1951 local_irq_restore(flags);
1956 * Inlined fastpath so that allocation functions (kmalloc, kmem_cache_alloc)
1957 * have the fastpath folded into their functions. So no function call
1958 * overhead for requests that can be satisfied on the fastpath.
1960 * The fastpath works by first checking if the lockless freelist can be used.
1961 * If not then __slab_alloc is called for slow processing.
1963 * Otherwise we can simply pick the next object from the lockless free list.
1965 static __always_inline void *slab_alloc(struct kmem_cache *s,
1966 gfp_t gfpflags, int node, unsigned long addr)
1969 struct kmem_cache_cpu *c;
1972 if (slab_pre_alloc_hook(s, gfpflags))
1978 * Must read kmem_cache cpu data via this cpu ptr. Preemption is
1979 * enabled. We may switch back and forth between cpus while
1980 * reading from one cpu area. That does not matter as long
1981 * as we end up on the original cpu again when doing the cmpxchg.
1983 c = __this_cpu_ptr(s->cpu_slab);
1986 * The transaction ids are globally unique per cpu and per operation on
1987 * a per cpu queue. Thus they can be guarantee that the cmpxchg_double
1988 * occurs on the right processor and that there was no operation on the
1989 * linked list in between.
1994 object = c->freelist;
1995 if (unlikely(!object || !node_match(c, node)))
1997 object = __slab_alloc(s, gfpflags, node, addr, c);
2001 * The cmpxchg will only match if there was no additional
2002 * operation and if we are on the right processor.
2004 * The cmpxchg does the following atomically (without lock semantics!)
2005 * 1. Relocate first pointer to the current per cpu area.
2006 * 2. Verify that tid and freelist have not been changed
2007 * 3. If they were not changed replace tid and freelist
2009 * Since this is without lock semantics the protection is only against
2010 * code executing on this cpu *not* from access by other cpus.
2012 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2013 s->cpu_slab->freelist, s->cpu_slab->tid,
2015 get_freepointer_safe(s, object), next_tid(tid)))) {
2017 note_cmpxchg_failure("slab_alloc", s, tid);
2020 stat(s, ALLOC_FASTPATH);
2023 if (unlikely(gfpflags & __GFP_ZERO) && object)
2024 memset(object, 0, s->objsize);
2026 slab_post_alloc_hook(s, gfpflags, object);
2031 void *kmem_cache_alloc(struct kmem_cache *s, gfp_t gfpflags)
2033 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2035 trace_kmem_cache_alloc(_RET_IP_, ret, s->objsize, s->size, gfpflags);
2039 EXPORT_SYMBOL(kmem_cache_alloc);
2041 #ifdef CONFIG_TRACING
2042 void *kmem_cache_alloc_trace(struct kmem_cache *s, gfp_t gfpflags, size_t size)
2044 void *ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, _RET_IP_);
2045 trace_kmalloc(_RET_IP_, ret, size, s->size, gfpflags);
2048 EXPORT_SYMBOL(kmem_cache_alloc_trace);
2050 void *kmalloc_order_trace(size_t size, gfp_t flags, unsigned int order)
2052 void *ret = kmalloc_order(size, flags, order);
2053 trace_kmalloc(_RET_IP_, ret, size, PAGE_SIZE << order, flags);
2056 EXPORT_SYMBOL(kmalloc_order_trace);
2060 void *kmem_cache_alloc_node(struct kmem_cache *s, gfp_t gfpflags, int node)
2062 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2064 trace_kmem_cache_alloc_node(_RET_IP_, ret,
2065 s->objsize, s->size, gfpflags, node);
2069 EXPORT_SYMBOL(kmem_cache_alloc_node);
2071 #ifdef CONFIG_TRACING
2072 void *kmem_cache_alloc_node_trace(struct kmem_cache *s,
2074 int node, size_t size)
2076 void *ret = slab_alloc(s, gfpflags, node, _RET_IP_);
2078 trace_kmalloc_node(_RET_IP_, ret,
2079 size, s->size, gfpflags, node);
2082 EXPORT_SYMBOL(kmem_cache_alloc_node_trace);
2087 * Slow patch handling. This may still be called frequently since objects
2088 * have a longer lifetime than the cpu slabs in most processing loads.
2090 * So we still attempt to reduce cache line usage. Just take the slab
2091 * lock and free the item. If there is no additional partial page
2092 * handling required then we can return immediately.
2094 static void __slab_free(struct kmem_cache *s, struct page *page,
2095 void *x, unsigned long addr)
2098 void **object = (void *)x;
2099 unsigned long flags;
2101 local_irq_save(flags);
2103 stat(s, FREE_SLOWPATH);
2105 if (kmem_cache_debug(s) && !free_debug_processing(s, page, x, addr))
2108 prior = page->freelist;
2109 set_freepointer(s, object, prior);
2110 page->freelist = object;
2113 if (unlikely(PageSlubFrozen(page))) {
2114 stat(s, FREE_FROZEN);
2118 if (unlikely(!page->inuse))
2122 * Objects left in the slab. If it was not on the partial list before
2125 if (unlikely(!prior)) {
2126 add_partial(get_node(s, page_to_nid(page)), page, 1);
2127 stat(s, FREE_ADD_PARTIAL);
2132 local_irq_restore(flags);
2138 * Slab still on the partial list.
2140 remove_partial(s, page);
2141 stat(s, FREE_REMOVE_PARTIAL);
2144 local_irq_restore(flags);
2146 discard_slab(s, page);
2150 * Fastpath with forced inlining to produce a kfree and kmem_cache_free that
2151 * can perform fastpath freeing without additional function calls.
2153 * The fastpath is only possible if we are freeing to the current cpu slab
2154 * of this processor. This typically the case if we have just allocated
2157 * If fastpath is not possible then fall back to __slab_free where we deal
2158 * with all sorts of special processing.
2160 static __always_inline void slab_free(struct kmem_cache *s,
2161 struct page *page, void *x, unsigned long addr)
2163 void **object = (void *)x;
2164 struct kmem_cache_cpu *c;
2167 slab_free_hook(s, x);
2172 * Determine the currently cpus per cpu slab.
2173 * The cpu may change afterward. However that does not matter since
2174 * data is retrieved via this pointer. If we are on the same cpu
2175 * during the cmpxchg then the free will succedd.
2177 c = __this_cpu_ptr(s->cpu_slab);
2182 if (likely(page == c->page)) {
2183 set_freepointer(s, object, c->freelist);
2185 if (unlikely(!irqsafe_cpu_cmpxchg_double(
2186 s->cpu_slab->freelist, s->cpu_slab->tid,
2188 object, next_tid(tid)))) {
2190 note_cmpxchg_failure("slab_free", s, tid);
2193 stat(s, FREE_FASTPATH);
2195 __slab_free(s, page, x, addr);
2199 void kmem_cache_free(struct kmem_cache *s, void *x)
2203 page = virt_to_head_page(x);
2205 slab_free(s, page, x, _RET_IP_);
2207 trace_kmem_cache_free(_RET_IP_, x);
2209 EXPORT_SYMBOL(kmem_cache_free);
2212 * Object placement in a slab is made very easy because we always start at
2213 * offset 0. If we tune the size of the object to the alignment then we can
2214 * get the required alignment by putting one properly sized object after
2217 * Notice that the allocation order determines the sizes of the per cpu
2218 * caches. Each processor has always one slab available for allocations.
2219 * Increasing the allocation order reduces the number of times that slabs
2220 * must be moved on and off the partial lists and is therefore a factor in
2225 * Mininum / Maximum order of slab pages. This influences locking overhead
2226 * and slab fragmentation. A higher order reduces the number of partial slabs
2227 * and increases the number of allocations possible without having to
2228 * take the list_lock.
2230 static int slub_min_order;
2231 static int slub_max_order = PAGE_ALLOC_COSTLY_ORDER;
2232 static int slub_min_objects;
2235 * Merge control. If this is set then no merging of slab caches will occur.
2236 * (Could be removed. This was introduced to pacify the merge skeptics.)
2238 static int slub_nomerge;
2241 * Calculate the order of allocation given an slab object size.
2243 * The order of allocation has significant impact on performance and other
2244 * system components. Generally order 0 allocations should be preferred since
2245 * order 0 does not cause fragmentation in the page allocator. Larger objects
2246 * be problematic to put into order 0 slabs because there may be too much
2247 * unused space left. We go to a higher order if more than 1/16th of the slab
2250 * In order to reach satisfactory performance we must ensure that a minimum
2251 * number of objects is in one slab. Otherwise we may generate too much
2252 * activity on the partial lists which requires taking the list_lock. This is
2253 * less a concern for large slabs though which are rarely used.
2255 * slub_max_order specifies the order where we begin to stop considering the
2256 * number of objects in a slab as critical. If we reach slub_max_order then
2257 * we try to keep the page order as low as possible. So we accept more waste
2258 * of space in favor of a small page order.
2260 * Higher order allocations also allow the placement of more objects in a
2261 * slab and thereby reduce object handling overhead. If the user has
2262 * requested a higher mininum order then we start with that one instead of
2263 * the smallest order which will fit the object.
2265 static inline int slab_order(int size, int min_objects,
2266 int max_order, int fract_leftover, int reserved)
2270 int min_order = slub_min_order;
2272 if (order_objects(min_order, size, reserved) > MAX_OBJS_PER_PAGE)
2273 return get_order(size * MAX_OBJS_PER_PAGE) - 1;
2275 for (order = max(min_order,
2276 fls(min_objects * size - 1) - PAGE_SHIFT);
2277 order <= max_order; order++) {
2279 unsigned long slab_size = PAGE_SIZE << order;
2281 if (slab_size < min_objects * size + reserved)
2284 rem = (slab_size - reserved) % size;
2286 if (rem <= slab_size / fract_leftover)
2294 static inline int calculate_order(int size, int reserved)
2302 * Attempt to find best configuration for a slab. This
2303 * works by first attempting to generate a layout with
2304 * the best configuration and backing off gradually.
2306 * First we reduce the acceptable waste in a slab. Then
2307 * we reduce the minimum objects required in a slab.
2309 min_objects = slub_min_objects;
2311 min_objects = 4 * (fls(nr_cpu_ids) + 1);
2312 max_objects = order_objects(slub_max_order, size, reserved);
2313 min_objects = min(min_objects, max_objects);
2315 while (min_objects > 1) {
2317 while (fraction >= 4) {
2318 order = slab_order(size, min_objects,
2319 slub_max_order, fraction, reserved);
2320 if (order <= slub_max_order)
2328 * We were unable to place multiple objects in a slab. Now
2329 * lets see if we can place a single object there.
2331 order = slab_order(size, 1, slub_max_order, 1, reserved);
2332 if (order <= slub_max_order)
2336 * Doh this slab cannot be placed using slub_max_order.
2338 order = slab_order(size, 1, MAX_ORDER, 1, reserved);
2339 if (order < MAX_ORDER)
2345 * Figure out what the alignment of the objects will be.
2347 static unsigned long calculate_alignment(unsigned long flags,
2348 unsigned long align, unsigned long size)
2351 * If the user wants hardware cache aligned objects then follow that
2352 * suggestion if the object is sufficiently large.
2354 * The hardware cache alignment cannot override the specified
2355 * alignment though. If that is greater then use it.
2357 if (flags & SLAB_HWCACHE_ALIGN) {
2358 unsigned long ralign = cache_line_size();
2359 while (size <= ralign / 2)
2361 align = max(align, ralign);
2364 if (align < ARCH_SLAB_MINALIGN)
2365 align = ARCH_SLAB_MINALIGN;
2367 return ALIGN(align, sizeof(void *));
2371 init_kmem_cache_node(struct kmem_cache_node *n, struct kmem_cache *s)
2374 spin_lock_init(&n->list_lock);
2375 INIT_LIST_HEAD(&n->partial);
2376 #ifdef CONFIG_SLUB_DEBUG
2377 atomic_long_set(&n->nr_slabs, 0);
2378 atomic_long_set(&n->total_objects, 0);
2379 INIT_LIST_HEAD(&n->full);
2383 static inline int alloc_kmem_cache_cpus(struct kmem_cache *s)
2385 BUILD_BUG_ON(PERCPU_DYNAMIC_EARLY_SIZE <
2386 SLUB_PAGE_SHIFT * sizeof(struct kmem_cache_cpu));
2388 #ifdef CONFIG_CMPXCHG_LOCAL
2390 * Must align to double word boundary for the double cmpxchg instructions
2393 s->cpu_slab = __alloc_percpu(sizeof(struct kmem_cache_cpu), 2 * sizeof(void *));
2395 /* Regular alignment is sufficient */
2396 s->cpu_slab = alloc_percpu(struct kmem_cache_cpu);
2402 init_kmem_cache_cpus(s);
2407 static struct kmem_cache *kmem_cache_node;
2410 * No kmalloc_node yet so do it by hand. We know that this is the first
2411 * slab on the node for this slabcache. There are no concurrent accesses
2414 * Note that this function only works on the kmalloc_node_cache
2415 * when allocating for the kmalloc_node_cache. This is used for bootstrapping
2416 * memory on a fresh node that has no slab structures yet.
2418 static void early_kmem_cache_node_alloc(int node)
2421 struct kmem_cache_node *n;
2422 unsigned long flags;
2424 BUG_ON(kmem_cache_node->size < sizeof(struct kmem_cache_node));
2426 page = new_slab(kmem_cache_node, GFP_NOWAIT, node);
2429 if (page_to_nid(page) != node) {
2430 printk(KERN_ERR "SLUB: Unable to allocate memory from "
2432 printk(KERN_ERR "SLUB: Allocating a useless per node structure "
2433 "in order to be able to continue\n");
2438 page->freelist = get_freepointer(kmem_cache_node, n);
2440 kmem_cache_node->node[node] = n;
2441 #ifdef CONFIG_SLUB_DEBUG
2442 init_object(kmem_cache_node, n, SLUB_RED_ACTIVE);
2443 init_tracking(kmem_cache_node, n);
2445 init_kmem_cache_node(n, kmem_cache_node);
2446 inc_slabs_node(kmem_cache_node, node, page->objects);
2449 * lockdep requires consistent irq usage for each lock
2450 * so even though there cannot be a race this early in
2451 * the boot sequence, we still disable irqs.
2453 local_irq_save(flags);
2454 add_partial(n, page, 0);
2455 local_irq_restore(flags);
2458 static void free_kmem_cache_nodes(struct kmem_cache *s)
2462 for_each_node_state(node, N_NORMAL_MEMORY) {
2463 struct kmem_cache_node *n = s->node[node];
2466 kmem_cache_free(kmem_cache_node, n);
2468 s->node[node] = NULL;
2472 static int init_kmem_cache_nodes(struct kmem_cache *s)
2476 for_each_node_state(node, N_NORMAL_MEMORY) {
2477 struct kmem_cache_node *n;
2479 if (slab_state == DOWN) {
2480 early_kmem_cache_node_alloc(node);
2483 n = kmem_cache_alloc_node(kmem_cache_node,
2487 free_kmem_cache_nodes(s);
2492 init_kmem_cache_node(n, s);
2497 static void set_min_partial(struct kmem_cache *s, unsigned long min)
2499 if (min < MIN_PARTIAL)
2501 else if (min > MAX_PARTIAL)
2503 s->min_partial = min;
2507 * calculate_sizes() determines the order and the distribution of data within
2510 static int calculate_sizes(struct kmem_cache *s, int forced_order)
2512 unsigned long flags = s->flags;
2513 unsigned long size = s->objsize;
2514 unsigned long align = s->align;
2518 * Round up object size to the next word boundary. We can only
2519 * place the free pointer at word boundaries and this determines
2520 * the possible location of the free pointer.
2522 size = ALIGN(size, sizeof(void *));
2524 #ifdef CONFIG_SLUB_DEBUG
2526 * Determine if we can poison the object itself. If the user of
2527 * the slab may touch the object after free or before allocation
2528 * then we should never poison the object itself.
2530 if ((flags & SLAB_POISON) && !(flags & SLAB_DESTROY_BY_RCU) &&
2532 s->flags |= __OBJECT_POISON;
2534 s->flags &= ~__OBJECT_POISON;
2538 * If we are Redzoning then check if there is some space between the
2539 * end of the object and the free pointer. If not then add an
2540 * additional word to have some bytes to store Redzone information.
2542 if ((flags & SLAB_RED_ZONE) && size == s->objsize)
2543 size += sizeof(void *);
2547 * With that we have determined the number of bytes in actual use
2548 * by the object. This is the potential offset to the free pointer.
2552 if (((flags & (SLAB_DESTROY_BY_RCU | SLAB_POISON)) ||
2555 * Relocate free pointer after the object if it is not
2556 * permitted to overwrite the first word of the object on
2559 * This is the case if we do RCU, have a constructor or
2560 * destructor or are poisoning the objects.
2563 size += sizeof(void *);
2566 #ifdef CONFIG_SLUB_DEBUG
2567 if (flags & SLAB_STORE_USER)
2569 * Need to store information about allocs and frees after
2572 size += 2 * sizeof(struct track);
2574 if (flags & SLAB_RED_ZONE)
2576 * Add some empty padding so that we can catch
2577 * overwrites from earlier objects rather than let
2578 * tracking information or the free pointer be
2579 * corrupted if a user writes before the start
2582 size += sizeof(void *);
2586 * Determine the alignment based on various parameters that the
2587 * user specified and the dynamic determination of cache line size
2590 align = calculate_alignment(flags, align, s->objsize);
2594 * SLUB stores one object immediately after another beginning from
2595 * offset 0. In order to align the objects we have to simply size
2596 * each object to conform to the alignment.
2598 size = ALIGN(size, align);
2600 if (forced_order >= 0)
2601 order = forced_order;
2603 order = calculate_order(size, s->reserved);
2610 s->allocflags |= __GFP_COMP;
2612 if (s->flags & SLAB_CACHE_DMA)
2613 s->allocflags |= SLUB_DMA;
2615 if (s->flags & SLAB_RECLAIM_ACCOUNT)
2616 s->allocflags |= __GFP_RECLAIMABLE;
2619 * Determine the number of objects per slab
2621 s->oo = oo_make(order, size, s->reserved);
2622 s->min = oo_make(get_order(size), size, s->reserved);
2623 if (oo_objects(s->oo) > oo_objects(s->max))
2626 return !!oo_objects(s->oo);
2630 static int kmem_cache_open(struct kmem_cache *s,
2631 const char *name, size_t size,
2632 size_t align, unsigned long flags,
2633 void (*ctor)(void *))
2635 memset(s, 0, kmem_size);
2640 s->flags = kmem_cache_flags(size, flags, name, ctor);
2643 if (need_reserve_slab_rcu && (s->flags & SLAB_DESTROY_BY_RCU))
2644 s->reserved = sizeof(struct rcu_head);
2646 if (!calculate_sizes(s, -1))
2648 if (disable_higher_order_debug) {
2650 * Disable debugging flags that store metadata if the min slab
2653 if (get_order(s->size) > get_order(s->objsize)) {
2654 s->flags &= ~DEBUG_METADATA_FLAGS;
2656 if (!calculate_sizes(s, -1))
2662 * The larger the object size is, the more pages we want on the partial
2663 * list to avoid pounding the page allocator excessively.
2665 set_min_partial(s, ilog2(s->size));
2668 s->remote_node_defrag_ratio = 1000;
2670 if (!init_kmem_cache_nodes(s))
2673 if (alloc_kmem_cache_cpus(s))
2676 free_kmem_cache_nodes(s);
2678 if (flags & SLAB_PANIC)
2679 panic("Cannot create slab %s size=%lu realsize=%u "
2680 "order=%u offset=%u flags=%lx\n",
2681 s->name, (unsigned long)size, s->size, oo_order(s->oo),
2687 * Determine the size of a slab object
2689 unsigned int kmem_cache_size(struct kmem_cache *s)
2693 EXPORT_SYMBOL(kmem_cache_size);
2695 static void list_slab_objects(struct kmem_cache *s, struct page *page,
2698 #ifdef CONFIG_SLUB_DEBUG
2699 void *addr = page_address(page);
2701 unsigned long *map = kzalloc(BITS_TO_LONGS(page->objects) *
2702 sizeof(long), GFP_ATOMIC);
2705 slab_err(s, page, "%s", text);
2708 get_map(s, page, map);
2709 for_each_object(p, s, addr, page->objects) {
2711 if (!test_bit(slab_index(p, s, addr), map)) {
2712 printk(KERN_ERR "INFO: Object 0x%p @offset=%tu\n",
2714 print_tracking(s, p);
2723 * Attempt to free all partial slabs on a node.
2725 static void free_partial(struct kmem_cache *s, struct kmem_cache_node *n)
2727 unsigned long flags;
2728 struct page *page, *h;
2730 spin_lock_irqsave(&n->list_lock, flags);
2731 list_for_each_entry_safe(page, h, &n->partial, lru) {
2733 __remove_partial(n, page);
2734 discard_slab(s, page);
2736 list_slab_objects(s, page,
2737 "Objects remaining on kmem_cache_close()");
2740 spin_unlock_irqrestore(&n->list_lock, flags);
2744 * Release all resources used by a slab cache.
2746 static inline int kmem_cache_close(struct kmem_cache *s)
2751 free_percpu(s->cpu_slab);
2752 /* Attempt to free all objects */
2753 for_each_node_state(node, N_NORMAL_MEMORY) {
2754 struct kmem_cache_node *n = get_node(s, node);
2757 if (n->nr_partial || slabs_node(s, node))
2760 free_kmem_cache_nodes(s);
2765 * Close a cache and release the kmem_cache structure
2766 * (must be used for caches created using kmem_cache_create)
2768 void kmem_cache_destroy(struct kmem_cache *s)
2770 down_write(&slub_lock);
2774 if (kmem_cache_close(s)) {
2775 printk(KERN_ERR "SLUB %s: %s called for cache that "
2776 "still has objects.\n", s->name, __func__);
2779 if (s->flags & SLAB_DESTROY_BY_RCU)
2781 sysfs_slab_remove(s);
2783 up_write(&slub_lock);
2785 EXPORT_SYMBOL(kmem_cache_destroy);
2787 /********************************************************************
2789 *******************************************************************/
2791 struct kmem_cache *kmalloc_caches[SLUB_PAGE_SHIFT];
2792 EXPORT_SYMBOL(kmalloc_caches);
2794 static struct kmem_cache *kmem_cache;
2796 #ifdef CONFIG_ZONE_DMA
2797 static struct kmem_cache *kmalloc_dma_caches[SLUB_PAGE_SHIFT];
2800 static int __init setup_slub_min_order(char *str)
2802 get_option(&str, &slub_min_order);
2807 __setup("slub_min_order=", setup_slub_min_order);
2809 static int __init setup_slub_max_order(char *str)
2811 get_option(&str, &slub_max_order);
2812 slub_max_order = min(slub_max_order, MAX_ORDER - 1);
2817 __setup("slub_max_order=", setup_slub_max_order);
2819 static int __init setup_slub_min_objects(char *str)
2821 get_option(&str, &slub_min_objects);
2826 __setup("slub_min_objects=", setup_slub_min_objects);
2828 static int __init setup_slub_nomerge(char *str)
2834 __setup("slub_nomerge", setup_slub_nomerge);
2836 static struct kmem_cache *__init create_kmalloc_cache(const char *name,
2837 int size, unsigned int flags)
2839 struct kmem_cache *s;
2841 s = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
2844 * This function is called with IRQs disabled during early-boot on
2845 * single CPU so there's no need to take slub_lock here.
2847 if (!kmem_cache_open(s, name, size, ARCH_KMALLOC_MINALIGN,
2851 list_add(&s->list, &slab_caches);
2855 panic("Creation of kmalloc slab %s size=%d failed.\n", name, size);
2860 * Conversion table for small slabs sizes / 8 to the index in the
2861 * kmalloc array. This is necessary for slabs < 192 since we have non power
2862 * of two cache sizes there. The size of larger slabs can be determined using
2865 static s8 size_index[24] = {
2892 static inline int size_index_elem(size_t bytes)
2894 return (bytes - 1) / 8;
2897 static struct kmem_cache *get_slab(size_t size, gfp_t flags)
2903 return ZERO_SIZE_PTR;
2905 index = size_index[size_index_elem(size)];
2907 index = fls(size - 1);
2909 #ifdef CONFIG_ZONE_DMA
2910 if (unlikely((flags & SLUB_DMA)))
2911 return kmalloc_dma_caches[index];
2914 return kmalloc_caches[index];
2917 void *__kmalloc(size_t size, gfp_t flags)
2919 struct kmem_cache *s;
2922 if (unlikely(size > SLUB_MAX_SIZE))
2923 return kmalloc_large(size, flags);
2925 s = get_slab(size, flags);
2927 if (unlikely(ZERO_OR_NULL_PTR(s)))
2930 ret = slab_alloc(s, flags, NUMA_NO_NODE, _RET_IP_);
2932 trace_kmalloc(_RET_IP_, ret, size, s->size, flags);
2936 EXPORT_SYMBOL(__kmalloc);
2939 static void *kmalloc_large_node(size_t size, gfp_t flags, int node)
2944 flags |= __GFP_COMP | __GFP_NOTRACK;
2945 page = alloc_pages_node(node, flags, get_order(size));
2947 ptr = page_address(page);
2949 kmemleak_alloc(ptr, size, 1, flags);
2953 void *__kmalloc_node(size_t size, gfp_t flags, int node)
2955 struct kmem_cache *s;
2958 if (unlikely(size > SLUB_MAX_SIZE)) {
2959 ret = kmalloc_large_node(size, flags, node);
2961 trace_kmalloc_node(_RET_IP_, ret,
2962 size, PAGE_SIZE << get_order(size),
2968 s = get_slab(size, flags);
2970 if (unlikely(ZERO_OR_NULL_PTR(s)))
2973 ret = slab_alloc(s, flags, node, _RET_IP_);
2975 trace_kmalloc_node(_RET_IP_, ret, size, s->size, flags, node);
2979 EXPORT_SYMBOL(__kmalloc_node);
2982 size_t ksize(const void *object)
2986 if (unlikely(object == ZERO_SIZE_PTR))
2989 page = virt_to_head_page(object);
2991 if (unlikely(!PageSlab(page))) {
2992 WARN_ON(!PageCompound(page));
2993 return PAGE_SIZE << compound_order(page);
2996 return slab_ksize(page->slab);
2998 EXPORT_SYMBOL(ksize);
3000 #ifdef CONFIG_SLUB_DEBUG
3001 bool verify_mem_not_deleted(const void *x)
3004 void *object = (void *)x;
3005 unsigned long flags;
3008 if (unlikely(ZERO_OR_NULL_PTR(x)))
3011 local_irq_save(flags);
3013 page = virt_to_head_page(x);
3014 if (unlikely(!PageSlab(page))) {
3015 /* maybe it was from stack? */
3021 if (on_freelist(page->slab, page, object)) {
3022 object_err(page->slab, page, object, "Object is on free-list");
3030 local_irq_restore(flags);
3033 EXPORT_SYMBOL(verify_mem_not_deleted);
3036 void kfree(const void *x)
3039 void *object = (void *)x;
3041 trace_kfree(_RET_IP_, x);
3043 if (unlikely(ZERO_OR_NULL_PTR(x)))
3046 page = virt_to_head_page(x);
3047 if (unlikely(!PageSlab(page))) {
3048 BUG_ON(!PageCompound(page));
3053 slab_free(page->slab, page, object, _RET_IP_);
3055 EXPORT_SYMBOL(kfree);
3058 * kmem_cache_shrink removes empty slabs from the partial lists and sorts
3059 * the remaining slabs by the number of items in use. The slabs with the
3060 * most items in use come first. New allocations will then fill those up
3061 * and thus they can be removed from the partial lists.
3063 * The slabs with the least items are placed last. This results in them
3064 * being allocated from last increasing the chance that the last objects
3065 * are freed in them.
3067 int kmem_cache_shrink(struct kmem_cache *s)
3071 struct kmem_cache_node *n;
3074 int objects = oo_objects(s->max);
3075 struct list_head *slabs_by_inuse =
3076 kmalloc(sizeof(struct list_head) * objects, GFP_KERNEL);
3077 unsigned long flags;
3079 if (!slabs_by_inuse)
3083 for_each_node_state(node, N_NORMAL_MEMORY) {
3084 n = get_node(s, node);
3089 for (i = 0; i < objects; i++)
3090 INIT_LIST_HEAD(slabs_by_inuse + i);
3092 spin_lock_irqsave(&n->list_lock, flags);
3095 * Build lists indexed by the items in use in each slab.
3097 * Note that concurrent frees may occur while we hold the
3098 * list_lock. page->inuse here is the upper limit.
3100 list_for_each_entry_safe(page, t, &n->partial, lru) {
3101 if (!page->inuse && slab_trylock(page)) {
3103 * Must hold slab lock here because slab_free
3104 * may have freed the last object and be
3105 * waiting to release the slab.
3107 __remove_partial(n, page);
3109 discard_slab(s, page);
3111 list_move(&page->lru,
3112 slabs_by_inuse + page->inuse);
3117 * Rebuild the partial list with the slabs filled up most
3118 * first and the least used slabs at the end.
3120 for (i = objects - 1; i >= 0; i--)
3121 list_splice(slabs_by_inuse + i, n->partial.prev);
3123 spin_unlock_irqrestore(&n->list_lock, flags);
3126 kfree(slabs_by_inuse);
3129 EXPORT_SYMBOL(kmem_cache_shrink);
3131 #if defined(CONFIG_MEMORY_HOTPLUG)
3132 static int slab_mem_going_offline_callback(void *arg)
3134 struct kmem_cache *s;
3136 down_read(&slub_lock);
3137 list_for_each_entry(s, &slab_caches, list)
3138 kmem_cache_shrink(s);
3139 up_read(&slub_lock);
3144 static void slab_mem_offline_callback(void *arg)
3146 struct kmem_cache_node *n;
3147 struct kmem_cache *s;
3148 struct memory_notify *marg = arg;
3151 offline_node = marg->status_change_nid;
3154 * If the node still has available memory. we need kmem_cache_node
3157 if (offline_node < 0)
3160 down_read(&slub_lock);
3161 list_for_each_entry(s, &slab_caches, list) {
3162 n = get_node(s, offline_node);
3165 * if n->nr_slabs > 0, slabs still exist on the node
3166 * that is going down. We were unable to free them,
3167 * and offline_pages() function shouldn't call this
3168 * callback. So, we must fail.
3170 BUG_ON(slabs_node(s, offline_node));
3172 s->node[offline_node] = NULL;
3173 kmem_cache_free(kmem_cache_node, n);
3176 up_read(&slub_lock);
3179 static int slab_mem_going_online_callback(void *arg)
3181 struct kmem_cache_node *n;
3182 struct kmem_cache *s;
3183 struct memory_notify *marg = arg;
3184 int nid = marg->status_change_nid;
3188 * If the node's memory is already available, then kmem_cache_node is
3189 * already created. Nothing to do.
3195 * We are bringing a node online. No memory is available yet. We must
3196 * allocate a kmem_cache_node structure in order to bring the node
3199 down_read(&slub_lock);
3200 list_for_each_entry(s, &slab_caches, list) {
3202 * XXX: kmem_cache_alloc_node will fallback to other nodes
3203 * since memory is not yet available from the node that
3206 n = kmem_cache_alloc(kmem_cache_node, GFP_KERNEL);
3211 init_kmem_cache_node(n, s);
3215 up_read(&slub_lock);
3219 static int slab_memory_callback(struct notifier_block *self,
3220 unsigned long action, void *arg)
3225 case MEM_GOING_ONLINE:
3226 ret = slab_mem_going_online_callback(arg);
3228 case MEM_GOING_OFFLINE:
3229 ret = slab_mem_going_offline_callback(arg);
3232 case MEM_CANCEL_ONLINE:
3233 slab_mem_offline_callback(arg);
3236 case MEM_CANCEL_OFFLINE:
3240 ret = notifier_from_errno(ret);
3246 #endif /* CONFIG_MEMORY_HOTPLUG */
3248 /********************************************************************
3249 * Basic setup of slabs
3250 *******************************************************************/
3253 * Used for early kmem_cache structures that were allocated using
3254 * the page allocator
3257 static void __init kmem_cache_bootstrap_fixup(struct kmem_cache *s)
3261 list_add(&s->list, &slab_caches);
3264 for_each_node_state(node, N_NORMAL_MEMORY) {
3265 struct kmem_cache_node *n = get_node(s, node);
3269 list_for_each_entry(p, &n->partial, lru)
3272 #ifdef CONFIG_SLUB_DEBUG
3273 list_for_each_entry(p, &n->full, lru)
3280 void __init kmem_cache_init(void)
3284 struct kmem_cache *temp_kmem_cache;
3286 struct kmem_cache *temp_kmem_cache_node;
3287 unsigned long kmalloc_size;
3289 kmem_size = offsetof(struct kmem_cache, node) +
3290 nr_node_ids * sizeof(struct kmem_cache_node *);
3292 /* Allocate two kmem_caches from the page allocator */
3293 kmalloc_size = ALIGN(kmem_size, cache_line_size());
3294 order = get_order(2 * kmalloc_size);
3295 kmem_cache = (void *)__get_free_pages(GFP_NOWAIT, order);
3298 * Must first have the slab cache available for the allocations of the
3299 * struct kmem_cache_node's. There is special bootstrap code in
3300 * kmem_cache_open for slab_state == DOWN.
3302 kmem_cache_node = (void *)kmem_cache + kmalloc_size;
3304 kmem_cache_open(kmem_cache_node, "kmem_cache_node",
3305 sizeof(struct kmem_cache_node),
3306 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3308 hotplug_memory_notifier(slab_memory_callback, SLAB_CALLBACK_PRI);
3310 /* Able to allocate the per node structures */
3311 slab_state = PARTIAL;
3313 temp_kmem_cache = kmem_cache;
3314 kmem_cache_open(kmem_cache, "kmem_cache", kmem_size,
3315 0, SLAB_HWCACHE_ALIGN | SLAB_PANIC, NULL);
3316 kmem_cache = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3317 memcpy(kmem_cache, temp_kmem_cache, kmem_size);
3320 * Allocate kmem_cache_node properly from the kmem_cache slab.
3321 * kmem_cache_node is separately allocated so no need to
3322 * update any list pointers.
3324 temp_kmem_cache_node = kmem_cache_node;
3326 kmem_cache_node = kmem_cache_alloc(kmem_cache, GFP_NOWAIT);
3327 memcpy(kmem_cache_node, temp_kmem_cache_node, kmem_size);
3329 kmem_cache_bootstrap_fixup(kmem_cache_node);
3332 kmem_cache_bootstrap_fixup(kmem_cache);
3334 /* Free temporary boot structure */
3335 free_pages((unsigned long)temp_kmem_cache, order);
3337 /* Now we can use the kmem_cache to allocate kmalloc slabs */
3340 * Patch up the size_index table if we have strange large alignment
3341 * requirements for the kmalloc array. This is only the case for
3342 * MIPS it seems. The standard arches will not generate any code here.
3344 * Largest permitted alignment is 256 bytes due to the way we
3345 * handle the index determination for the smaller caches.
3347 * Make sure that nothing crazy happens if someone starts tinkering
3348 * around with ARCH_KMALLOC_MINALIGN
3350 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 256 ||
3351 (KMALLOC_MIN_SIZE & (KMALLOC_MIN_SIZE - 1)));
3353 for (i = 8; i < KMALLOC_MIN_SIZE; i += 8) {
3354 int elem = size_index_elem(i);
3355 if (elem >= ARRAY_SIZE(size_index))
3357 size_index[elem] = KMALLOC_SHIFT_LOW;
3360 if (KMALLOC_MIN_SIZE == 64) {
3362 * The 96 byte size cache is not used if the alignment
3365 for (i = 64 + 8; i <= 96; i += 8)
3366 size_index[size_index_elem(i)] = 7;
3367 } else if (KMALLOC_MIN_SIZE == 128) {
3369 * The 192 byte sized cache is not used if the alignment
3370 * is 128 byte. Redirect kmalloc to use the 256 byte cache
3373 for (i = 128 + 8; i <= 192; i += 8)
3374 size_index[size_index_elem(i)] = 8;
3377 /* Caches that are not of the two-to-the-power-of size */
3378 if (KMALLOC_MIN_SIZE <= 32) {
3379 kmalloc_caches[1] = create_kmalloc_cache("kmalloc-96", 96, 0);
3383 if (KMALLOC_MIN_SIZE <= 64) {
3384 kmalloc_caches[2] = create_kmalloc_cache("kmalloc-192", 192, 0);
3388 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3389 kmalloc_caches[i] = create_kmalloc_cache("kmalloc", 1 << i, 0);
3395 /* Provide the correct kmalloc names now that the caches are up */
3396 if (KMALLOC_MIN_SIZE <= 32) {
3397 kmalloc_caches[1]->name = kstrdup(kmalloc_caches[1]->name, GFP_NOWAIT);
3398 BUG_ON(!kmalloc_caches[1]->name);
3401 if (KMALLOC_MIN_SIZE <= 64) {
3402 kmalloc_caches[2]->name = kstrdup(kmalloc_caches[2]->name, GFP_NOWAIT);
3403 BUG_ON(!kmalloc_caches[2]->name);
3406 for (i = KMALLOC_SHIFT_LOW; i < SLUB_PAGE_SHIFT; i++) {
3407 char *s = kasprintf(GFP_NOWAIT, "kmalloc-%d", 1 << i);
3410 kmalloc_caches[i]->name = s;
3414 register_cpu_notifier(&slab_notifier);
3417 #ifdef CONFIG_ZONE_DMA
3418 for (i = 0; i < SLUB_PAGE_SHIFT; i++) {
3419 struct kmem_cache *s = kmalloc_caches[i];
3422 char *name = kasprintf(GFP_NOWAIT,
3423 "dma-kmalloc-%d", s->objsize);
3426 kmalloc_dma_caches[i] = create_kmalloc_cache(name,
3427 s->objsize, SLAB_CACHE_DMA);
3432 "SLUB: Genslabs=%d, HWalign=%d, Order=%d-%d, MinObjects=%d,"
3433 " CPUs=%d, Nodes=%d\n",
3434 caches, cache_line_size(),
3435 slub_min_order, slub_max_order, slub_min_objects,
3436 nr_cpu_ids, nr_node_ids);
3439 void __init kmem_cache_init_late(void)
3444 * Find a mergeable slab cache
3446 static int slab_unmergeable(struct kmem_cache *s)
3448 if (slub_nomerge || (s->flags & SLUB_NEVER_MERGE))
3455 * We may have set a slab to be unmergeable during bootstrap.
3457 if (s->refcount < 0)
3463 static struct kmem_cache *find_mergeable(size_t size,
3464 size_t align, unsigned long flags, const char *name,
3465 void (*ctor)(void *))
3467 struct kmem_cache *s;
3469 if (slub_nomerge || (flags & SLUB_NEVER_MERGE))
3475 size = ALIGN(size, sizeof(void *));
3476 align = calculate_alignment(flags, align, size);
3477 size = ALIGN(size, align);
3478 flags = kmem_cache_flags(size, flags, name, NULL);
3480 list_for_each_entry(s, &slab_caches, list) {
3481 if (slab_unmergeable(s))
3487 if ((flags & SLUB_MERGE_SAME) != (s->flags & SLUB_MERGE_SAME))
3490 * Check if alignment is compatible.
3491 * Courtesy of Adrian Drzewiecki
3493 if ((s->size & ~(align - 1)) != s->size)
3496 if (s->size - size >= sizeof(void *))
3504 struct kmem_cache *kmem_cache_create(const char *name, size_t size,
3505 size_t align, unsigned long flags, void (*ctor)(void *))
3507 struct kmem_cache *s;
3513 down_write(&slub_lock);
3514 s = find_mergeable(size, align, flags, name, ctor);
3518 * Adjust the object sizes so that we clear
3519 * the complete object on kzalloc.
3521 s->objsize = max(s->objsize, (int)size);
3522 s->inuse = max_t(int, s->inuse, ALIGN(size, sizeof(void *)));
3524 if (sysfs_slab_alias(s, name)) {
3528 up_write(&slub_lock);
3532 n = kstrdup(name, GFP_KERNEL);
3536 s = kmalloc(kmem_size, GFP_KERNEL);
3538 if (kmem_cache_open(s, n,
3539 size, align, flags, ctor)) {
3540 list_add(&s->list, &slab_caches);
3541 if (sysfs_slab_add(s)) {
3547 up_write(&slub_lock);
3554 up_write(&slub_lock);
3556 if (flags & SLAB_PANIC)
3557 panic("Cannot create slabcache %s\n", name);
3562 EXPORT_SYMBOL(kmem_cache_create);
3566 * Use the cpu notifier to insure that the cpu slabs are flushed when
3569 static int __cpuinit slab_cpuup_callback(struct notifier_block *nfb,
3570 unsigned long action, void *hcpu)
3572 long cpu = (long)hcpu;
3573 struct kmem_cache *s;
3574 unsigned long flags;
3577 case CPU_UP_CANCELED:
3578 case CPU_UP_CANCELED_FROZEN:
3580 case CPU_DEAD_FROZEN:
3581 down_read(&slub_lock);
3582 list_for_each_entry(s, &slab_caches, list) {
3583 local_irq_save(flags);
3584 __flush_cpu_slab(s, cpu);
3585 local_irq_restore(flags);
3587 up_read(&slub_lock);
3595 static struct notifier_block __cpuinitdata slab_notifier = {
3596 .notifier_call = slab_cpuup_callback
3601 void *__kmalloc_track_caller(size_t size, gfp_t gfpflags, unsigned long caller)
3603 struct kmem_cache *s;
3606 if (unlikely(size > SLUB_MAX_SIZE))
3607 return kmalloc_large(size, gfpflags);
3609 s = get_slab(size, gfpflags);
3611 if (unlikely(ZERO_OR_NULL_PTR(s)))
3614 ret = slab_alloc(s, gfpflags, NUMA_NO_NODE, caller);
3616 /* Honor the call site pointer we received. */
3617 trace_kmalloc(caller, ret, size, s->size, gfpflags);
3623 void *__kmalloc_node_track_caller(size_t size, gfp_t gfpflags,
3624 int node, unsigned long caller)
3626 struct kmem_cache *s;
3629 if (unlikely(size > SLUB_MAX_SIZE)) {
3630 ret = kmalloc_large_node(size, gfpflags, node);
3632 trace_kmalloc_node(caller, ret,
3633 size, PAGE_SIZE << get_order(size),
3639 s = get_slab(size, gfpflags);
3641 if (unlikely(ZERO_OR_NULL_PTR(s)))
3644 ret = slab_alloc(s, gfpflags, node, caller);
3646 /* Honor the call site pointer we received. */
3647 trace_kmalloc_node(caller, ret, size, s->size, gfpflags, node);
3654 static int count_inuse(struct page *page)
3659 static int count_total(struct page *page)
3661 return page->objects;
3665 #ifdef CONFIG_SLUB_DEBUG
3666 static int validate_slab(struct kmem_cache *s, struct page *page,
3670 void *addr = page_address(page);
3672 if (!check_slab(s, page) ||
3673 !on_freelist(s, page, NULL))
3676 /* Now we know that a valid freelist exists */
3677 bitmap_zero(map, page->objects);
3679 get_map(s, page, map);
3680 for_each_object(p, s, addr, page->objects) {
3681 if (test_bit(slab_index(p, s, addr), map))
3682 if (!check_object(s, page, p, SLUB_RED_INACTIVE))
3686 for_each_object(p, s, addr, page->objects)
3687 if (!test_bit(slab_index(p, s, addr), map))
3688 if (!check_object(s, page, p, SLUB_RED_ACTIVE))
3693 static void validate_slab_slab(struct kmem_cache *s, struct page *page,
3696 if (slab_trylock(page)) {
3697 validate_slab(s, page, map);
3700 printk(KERN_INFO "SLUB %s: Skipped busy slab 0x%p\n",
3704 static int validate_slab_node(struct kmem_cache *s,
3705 struct kmem_cache_node *n, unsigned long *map)
3707 unsigned long count = 0;
3709 unsigned long flags;
3711 spin_lock_irqsave(&n->list_lock, flags);
3713 list_for_each_entry(page, &n->partial, lru) {
3714 validate_slab_slab(s, page, map);
3717 if (count != n->nr_partial)
3718 printk(KERN_ERR "SLUB %s: %ld partial slabs counted but "
3719 "counter=%ld\n", s->name, count, n->nr_partial);
3721 if (!(s->flags & SLAB_STORE_USER))
3724 list_for_each_entry(page, &n->full, lru) {
3725 validate_slab_slab(s, page, map);
3728 if (count != atomic_long_read(&n->nr_slabs))
3729 printk(KERN_ERR "SLUB: %s %ld slabs counted but "
3730 "counter=%ld\n", s->name, count,
3731 atomic_long_read(&n->nr_slabs));
3734 spin_unlock_irqrestore(&n->list_lock, flags);
3738 static long validate_slab_cache(struct kmem_cache *s)
3741 unsigned long count = 0;
3742 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3743 sizeof(unsigned long), GFP_KERNEL);
3749 for_each_node_state(node, N_NORMAL_MEMORY) {
3750 struct kmem_cache_node *n = get_node(s, node);
3752 count += validate_slab_node(s, n, map);
3758 * Generate lists of code addresses where slabcache objects are allocated
3763 unsigned long count;
3770 DECLARE_BITMAP(cpus, NR_CPUS);
3776 unsigned long count;
3777 struct location *loc;
3780 static void free_loc_track(struct loc_track *t)
3783 free_pages((unsigned long)t->loc,
3784 get_order(sizeof(struct location) * t->max));
3787 static int alloc_loc_track(struct loc_track *t, unsigned long max, gfp_t flags)
3792 order = get_order(sizeof(struct location) * max);
3794 l = (void *)__get_free_pages(flags, order);
3799 memcpy(l, t->loc, sizeof(struct location) * t->count);
3807 static int add_location(struct loc_track *t, struct kmem_cache *s,
3808 const struct track *track)
3810 long start, end, pos;
3812 unsigned long caddr;
3813 unsigned long age = jiffies - track->when;
3819 pos = start + (end - start + 1) / 2;
3822 * There is nothing at "end". If we end up there
3823 * we need to add something to before end.
3828 caddr = t->loc[pos].addr;
3829 if (track->addr == caddr) {
3835 if (age < l->min_time)
3837 if (age > l->max_time)
3840 if (track->pid < l->min_pid)
3841 l->min_pid = track->pid;
3842 if (track->pid > l->max_pid)
3843 l->max_pid = track->pid;
3845 cpumask_set_cpu(track->cpu,
3846 to_cpumask(l->cpus));
3848 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3852 if (track->addr < caddr)
3859 * Not found. Insert new tracking element.
3861 if (t->count >= t->max && !alloc_loc_track(t, 2 * t->max, GFP_ATOMIC))
3867 (t->count - pos) * sizeof(struct location));
3870 l->addr = track->addr;
3874 l->min_pid = track->pid;
3875 l->max_pid = track->pid;
3876 cpumask_clear(to_cpumask(l->cpus));
3877 cpumask_set_cpu(track->cpu, to_cpumask(l->cpus));
3878 nodes_clear(l->nodes);
3879 node_set(page_to_nid(virt_to_page(track)), l->nodes);
3883 static void process_slab(struct loc_track *t, struct kmem_cache *s,
3884 struct page *page, enum track_item alloc,
3887 void *addr = page_address(page);
3890 bitmap_zero(map, page->objects);
3891 get_map(s, page, map);
3893 for_each_object(p, s, addr, page->objects)
3894 if (!test_bit(slab_index(p, s, addr), map))
3895 add_location(t, s, get_track(s, p, alloc));
3898 static int list_locations(struct kmem_cache *s, char *buf,
3899 enum track_item alloc)
3903 struct loc_track t = { 0, 0, NULL };
3905 unsigned long *map = kmalloc(BITS_TO_LONGS(oo_objects(s->max)) *
3906 sizeof(unsigned long), GFP_KERNEL);
3908 if (!map || !alloc_loc_track(&t, PAGE_SIZE / sizeof(struct location),
3911 return sprintf(buf, "Out of memory\n");
3913 /* Push back cpu slabs */
3916 for_each_node_state(node, N_NORMAL_MEMORY) {
3917 struct kmem_cache_node *n = get_node(s, node);
3918 unsigned long flags;
3921 if (!atomic_long_read(&n->nr_slabs))
3924 spin_lock_irqsave(&n->list_lock, flags);
3925 list_for_each_entry(page, &n->partial, lru)
3926 process_slab(&t, s, page, alloc, map);
3927 list_for_each_entry(page, &n->full, lru)
3928 process_slab(&t, s, page, alloc, map);
3929 spin_unlock_irqrestore(&n->list_lock, flags);
3932 for (i = 0; i < t.count; i++) {
3933 struct location *l = &t.loc[i];
3935 if (len > PAGE_SIZE - KSYM_SYMBOL_LEN - 100)
3937 len += sprintf(buf + len, "%7ld ", l->count);
3940 len += sprintf(buf + len, "%pS", (void *)l->addr);
3942 len += sprintf(buf + len, "<not-available>");
3944 if (l->sum_time != l->min_time) {
3945 len += sprintf(buf + len, " age=%ld/%ld/%ld",
3947 (long)div_u64(l->sum_time, l->count),
3950 len += sprintf(buf + len, " age=%ld",
3953 if (l->min_pid != l->max_pid)
3954 len += sprintf(buf + len, " pid=%ld-%ld",
3955 l->min_pid, l->max_pid);
3957 len += sprintf(buf + len, " pid=%ld",
3960 if (num_online_cpus() > 1 &&
3961 !cpumask_empty(to_cpumask(l->cpus)) &&
3962 len < PAGE_SIZE - 60) {
3963 len += sprintf(buf + len, " cpus=");
3964 len += cpulist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3965 to_cpumask(l->cpus));
3968 if (nr_online_nodes > 1 && !nodes_empty(l->nodes) &&
3969 len < PAGE_SIZE - 60) {
3970 len += sprintf(buf + len, " nodes=");
3971 len += nodelist_scnprintf(buf + len, PAGE_SIZE - len - 50,
3975 len += sprintf(buf + len, "\n");
3981 len += sprintf(buf, "No data\n");
3986 #ifdef SLUB_RESILIENCY_TEST
3987 static void resiliency_test(void)
3991 BUILD_BUG_ON(KMALLOC_MIN_SIZE > 16 || SLUB_PAGE_SHIFT < 10);
3993 printk(KERN_ERR "SLUB resiliency testing\n");
3994 printk(KERN_ERR "-----------------------\n");
3995 printk(KERN_ERR "A. Corruption after allocation\n");
3997 p = kzalloc(16, GFP_KERNEL);
3999 printk(KERN_ERR "\n1. kmalloc-16: Clobber Redzone/next pointer"
4000 " 0x12->0x%p\n\n", p + 16);
4002 validate_slab_cache(kmalloc_caches[4]);
4004 /* Hmmm... The next two are dangerous */
4005 p = kzalloc(32, GFP_KERNEL);
4006 p[32 + sizeof(void *)] = 0x34;
4007 printk(KERN_ERR "\n2. kmalloc-32: Clobber next pointer/next slab"
4008 " 0x34 -> -0x%p\n", p);
4010 "If allocated object is overwritten then not detectable\n\n");
4012 validate_slab_cache(kmalloc_caches[5]);
4013 p = kzalloc(64, GFP_KERNEL);
4014 p += 64 + (get_cycles() & 0xff) * sizeof(void *);
4016 printk(KERN_ERR "\n3. kmalloc-64: corrupting random byte 0x56->0x%p\n",
4019 "If allocated object is overwritten then not detectable\n\n");
4020 validate_slab_cache(kmalloc_caches[6]);
4022 printk(KERN_ERR "\nB. Corruption after free\n");
4023 p = kzalloc(128, GFP_KERNEL);
4026 printk(KERN_ERR "1. kmalloc-128: Clobber first word 0x78->0x%p\n\n", p);
4027 validate_slab_cache(kmalloc_caches[7]);
4029 p = kzalloc(256, GFP_KERNEL);
4032 printk(KERN_ERR "\n2. kmalloc-256: Clobber 50th byte 0x9a->0x%p\n\n",
4034 validate_slab_cache(kmalloc_caches[8]);
4036 p = kzalloc(512, GFP_KERNEL);
4039 printk(KERN_ERR "\n3. kmalloc-512: Clobber redzone 0xab->0x%p\n\n", p);
4040 validate_slab_cache(kmalloc_caches[9]);
4044 static void resiliency_test(void) {};
4049 enum slab_stat_type {
4050 SL_ALL, /* All slabs */
4051 SL_PARTIAL, /* Only partially allocated slabs */
4052 SL_CPU, /* Only slabs used for cpu caches */
4053 SL_OBJECTS, /* Determine allocated objects not slabs */
4054 SL_TOTAL /* Determine object capacity not slabs */
4057 #define SO_ALL (1 << SL_ALL)
4058 #define SO_PARTIAL (1 << SL_PARTIAL)
4059 #define SO_CPU (1 << SL_CPU)
4060 #define SO_OBJECTS (1 << SL_OBJECTS)
4061 #define SO_TOTAL (1 << SL_TOTAL)
4063 static ssize_t show_slab_objects(struct kmem_cache *s,
4064 char *buf, unsigned long flags)
4066 unsigned long total = 0;
4069 unsigned long *nodes;
4070 unsigned long *per_cpu;
4072 nodes = kzalloc(2 * sizeof(unsigned long) * nr_node_ids, GFP_KERNEL);
4075 per_cpu = nodes + nr_node_ids;
4077 if (flags & SO_CPU) {
4080 for_each_possible_cpu(cpu) {
4081 struct kmem_cache_cpu *c = per_cpu_ptr(s->cpu_slab, cpu);
4083 if (!c || c->node < 0)
4087 if (flags & SO_TOTAL)
4088 x = c->page->objects;
4089 else if (flags & SO_OBJECTS)
4095 nodes[c->node] += x;
4101 lock_memory_hotplug();
4102 #ifdef CONFIG_SLUB_DEBUG
4103 if (flags & SO_ALL) {
4104 for_each_node_state(node, N_NORMAL_MEMORY) {
4105 struct kmem_cache_node *n = get_node(s, node);
4107 if (flags & SO_TOTAL)
4108 x = atomic_long_read(&n->total_objects);
4109 else if (flags & SO_OBJECTS)
4110 x = atomic_long_read(&n->total_objects) -
4111 count_partial(n, count_free);
4114 x = atomic_long_read(&n->nr_slabs);
4121 if (flags & SO_PARTIAL) {
4122 for_each_node_state(node, N_NORMAL_MEMORY) {
4123 struct kmem_cache_node *n = get_node(s, node);
4125 if (flags & SO_TOTAL)
4126 x = count_partial(n, count_total);
4127 else if (flags & SO_OBJECTS)
4128 x = count_partial(n, count_inuse);
4135 x = sprintf(buf, "%lu", total);
4137 for_each_node_state(node, N_NORMAL_MEMORY)
4139 x += sprintf(buf + x, " N%d=%lu",
4142 unlock_memory_hotplug();
4144 return x + sprintf(buf + x, "\n");
4147 #ifdef CONFIG_SLUB_DEBUG
4148 static int any_slab_objects(struct kmem_cache *s)
4152 for_each_online_node(node) {
4153 struct kmem_cache_node *n = get_node(s, node);
4158 if (atomic_long_read(&n->total_objects))
4165 #define to_slab_attr(n) container_of(n, struct slab_attribute, attr)
4166 #define to_slab(n) container_of(n, struct kmem_cache, kobj);
4168 struct slab_attribute {
4169 struct attribute attr;
4170 ssize_t (*show)(struct kmem_cache *s, char *buf);
4171 ssize_t (*store)(struct kmem_cache *s, const char *x, size_t count);
4174 #define SLAB_ATTR_RO(_name) \
4175 static struct slab_attribute _name##_attr = __ATTR_RO(_name)
4177 #define SLAB_ATTR(_name) \
4178 static struct slab_attribute _name##_attr = \
4179 __ATTR(_name, 0644, _name##_show, _name##_store)
4181 static ssize_t slab_size_show(struct kmem_cache *s, char *buf)
4183 return sprintf(buf, "%d\n", s->size);
4185 SLAB_ATTR_RO(slab_size);
4187 static ssize_t align_show(struct kmem_cache *s, char *buf)
4189 return sprintf(buf, "%d\n", s->align);
4191 SLAB_ATTR_RO(align);
4193 static ssize_t object_size_show(struct kmem_cache *s, char *buf)
4195 return sprintf(buf, "%d\n", s->objsize);
4197 SLAB_ATTR_RO(object_size);
4199 static ssize_t objs_per_slab_show(struct kmem_cache *s, char *buf)
4201 return sprintf(buf, "%d\n", oo_objects(s->oo));
4203 SLAB_ATTR_RO(objs_per_slab);
4205 static ssize_t order_store(struct kmem_cache *s,
4206 const char *buf, size_t length)
4208 unsigned long order;
4211 err = strict_strtoul(buf, 10, &order);
4215 if (order > slub_max_order || order < slub_min_order)
4218 calculate_sizes(s, order);
4222 static ssize_t order_show(struct kmem_cache *s, char *buf)
4224 return sprintf(buf, "%d\n", oo_order(s->oo));
4228 static ssize_t min_partial_show(struct kmem_cache *s, char *buf)
4230 return sprintf(buf, "%lu\n", s->min_partial);
4233 static ssize_t min_partial_store(struct kmem_cache *s, const char *buf,
4239 err = strict_strtoul(buf, 10, &min);
4243 set_min_partial(s, min);
4246 SLAB_ATTR(min_partial);
4248 static ssize_t ctor_show(struct kmem_cache *s, char *buf)
4252 return sprintf(buf, "%pS\n", s->ctor);
4256 static ssize_t aliases_show(struct kmem_cache *s, char *buf)
4258 return sprintf(buf, "%d\n", s->refcount - 1);
4260 SLAB_ATTR_RO(aliases);
4262 static ssize_t partial_show(struct kmem_cache *s, char *buf)
4264 return show_slab_objects(s, buf, SO_PARTIAL);
4266 SLAB_ATTR_RO(partial);
4268 static ssize_t cpu_slabs_show(struct kmem_cache *s, char *buf)
4270 return show_slab_objects(s, buf, SO_CPU);
4272 SLAB_ATTR_RO(cpu_slabs);
4274 static ssize_t objects_show(struct kmem_cache *s, char *buf)
4276 return show_slab_objects(s, buf, SO_ALL|SO_OBJECTS);
4278 SLAB_ATTR_RO(objects);
4280 static ssize_t objects_partial_show(struct kmem_cache *s, char *buf)
4282 return show_slab_objects(s, buf, SO_PARTIAL|SO_OBJECTS);
4284 SLAB_ATTR_RO(objects_partial);
4286 static ssize_t reclaim_account_show(struct kmem_cache *s, char *buf)
4288 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RECLAIM_ACCOUNT));
4291 static ssize_t reclaim_account_store(struct kmem_cache *s,
4292 const char *buf, size_t length)
4294 s->flags &= ~SLAB_RECLAIM_ACCOUNT;
4296 s->flags |= SLAB_RECLAIM_ACCOUNT;
4299 SLAB_ATTR(reclaim_account);
4301 static ssize_t hwcache_align_show(struct kmem_cache *s, char *buf)
4303 return sprintf(buf, "%d\n", !!(s->flags & SLAB_HWCACHE_ALIGN));
4305 SLAB_ATTR_RO(hwcache_align);
4307 #ifdef CONFIG_ZONE_DMA
4308 static ssize_t cache_dma_show(struct kmem_cache *s, char *buf)
4310 return sprintf(buf, "%d\n", !!(s->flags & SLAB_CACHE_DMA));
4312 SLAB_ATTR_RO(cache_dma);
4315 static ssize_t destroy_by_rcu_show(struct kmem_cache *s, char *buf)
4317 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DESTROY_BY_RCU));
4319 SLAB_ATTR_RO(destroy_by_rcu);
4321 static ssize_t reserved_show(struct kmem_cache *s, char *buf)
4323 return sprintf(buf, "%d\n", s->reserved);
4325 SLAB_ATTR_RO(reserved);
4327 #ifdef CONFIG_SLUB_DEBUG
4328 static ssize_t slabs_show(struct kmem_cache *s, char *buf)
4330 return show_slab_objects(s, buf, SO_ALL);
4332 SLAB_ATTR_RO(slabs);
4334 static ssize_t total_objects_show(struct kmem_cache *s, char *buf)
4336 return show_slab_objects(s, buf, SO_ALL|SO_TOTAL);
4338 SLAB_ATTR_RO(total_objects);
4340 static ssize_t sanity_checks_show(struct kmem_cache *s, char *buf)
4342 return sprintf(buf, "%d\n", !!(s->flags & SLAB_DEBUG_FREE));
4345 static ssize_t sanity_checks_store(struct kmem_cache *s,
4346 const char *buf, size_t length)
4348 s->flags &= ~SLAB_DEBUG_FREE;
4350 s->flags |= SLAB_DEBUG_FREE;
4353 SLAB_ATTR(sanity_checks);
4355 static ssize_t trace_show(struct kmem_cache *s, char *buf)
4357 return sprintf(buf, "%d\n", !!(s->flags & SLAB_TRACE));
4360 static ssize_t trace_store(struct kmem_cache *s, const char *buf,
4363 s->flags &= ~SLAB_TRACE;
4365 s->flags |= SLAB_TRACE;
4370 static ssize_t red_zone_show(struct kmem_cache *s, char *buf)
4372 return sprintf(buf, "%d\n", !!(s->flags & SLAB_RED_ZONE));
4375 static ssize_t red_zone_store(struct kmem_cache *s,
4376 const char *buf, size_t length)
4378 if (any_slab_objects(s))
4381 s->flags &= ~SLAB_RED_ZONE;
4383 s->flags |= SLAB_RED_ZONE;
4384 calculate_sizes(s, -1);
4387 SLAB_ATTR(red_zone);
4389 static ssize_t poison_show(struct kmem_cache *s, char *buf)
4391 return sprintf(buf, "%d\n", !!(s->flags & SLAB_POISON));
4394 static ssize_t poison_store(struct kmem_cache *s,
4395 const char *buf, size_t length)
4397 if (any_slab_objects(s))
4400 s->flags &= ~SLAB_POISON;
4402 s->flags |= SLAB_POISON;
4403 calculate_sizes(s, -1);
4408 static ssize_t store_user_show(struct kmem_cache *s, char *buf)
4410 return sprintf(buf, "%d\n", !!(s->flags & SLAB_STORE_USER));
4413 static ssize_t store_user_store(struct kmem_cache *s,
4414 const char *buf, size_t length)
4416 if (any_slab_objects(s))
4419 s->flags &= ~SLAB_STORE_USER;
4421 s->flags |= SLAB_STORE_USER;
4422 calculate_sizes(s, -1);
4425 SLAB_ATTR(store_user);
4427 static ssize_t validate_show(struct kmem_cache *s, char *buf)
4432 static ssize_t validate_store(struct kmem_cache *s,
4433 const char *buf, size_t length)
4437 if (buf[0] == '1') {
4438 ret = validate_slab_cache(s);
4444 SLAB_ATTR(validate);
4446 static ssize_t alloc_calls_show(struct kmem_cache *s, char *buf)
4448 if (!(s->flags & SLAB_STORE_USER))
4450 return list_locations(s, buf, TRACK_ALLOC);
4452 SLAB_ATTR_RO(alloc_calls);
4454 static ssize_t free_calls_show(struct kmem_cache *s, char *buf)
4456 if (!(s->flags & SLAB_STORE_USER))
4458 return list_locations(s, buf, TRACK_FREE);
4460 SLAB_ATTR_RO(free_calls);
4461 #endif /* CONFIG_SLUB_DEBUG */
4463 #ifdef CONFIG_FAILSLAB
4464 static ssize_t failslab_show(struct kmem_cache *s, char *buf)
4466 return sprintf(buf, "%d\n", !!(s->flags & SLAB_FAILSLAB));
4469 static ssize_t failslab_store(struct kmem_cache *s, const char *buf,
4472 s->flags &= ~SLAB_FAILSLAB;
4474 s->flags |= SLAB_FAILSLAB;
4477 SLAB_ATTR(failslab);
4480 static ssize_t shrink_show(struct kmem_cache *s, char *buf)
4485 static ssize_t shrink_store(struct kmem_cache *s,
4486 const char *buf, size_t length)
4488 if (buf[0] == '1') {
4489 int rc = kmem_cache_shrink(s);
4500 static ssize_t remote_node_defrag_ratio_show(struct kmem_cache *s, char *buf)
4502 return sprintf(buf, "%d\n", s->remote_node_defrag_ratio / 10);
4505 static ssize_t remote_node_defrag_ratio_store(struct kmem_cache *s,
4506 const char *buf, size_t length)
4508 unsigned long ratio;
4511 err = strict_strtoul(buf, 10, &ratio);
4516 s->remote_node_defrag_ratio = ratio * 10;
4520 SLAB_ATTR(remote_node_defrag_ratio);
4523 #ifdef CONFIG_SLUB_STATS
4524 static int show_stat(struct kmem_cache *s, char *buf, enum stat_item si)
4526 unsigned long sum = 0;
4529 int *data = kmalloc(nr_cpu_ids * sizeof(int), GFP_KERNEL);
4534 for_each_online_cpu(cpu) {
4535 unsigned x = per_cpu_ptr(s->cpu_slab, cpu)->stat[si];
4541 len = sprintf(buf, "%lu", sum);
4544 for_each_online_cpu(cpu) {
4545 if (data[cpu] && len < PAGE_SIZE - 20)
4546 len += sprintf(buf + len, " C%d=%u", cpu, data[cpu]);
4550 return len + sprintf(buf + len, "\n");
4553 static void clear_stat(struct kmem_cache *s, enum stat_item si)
4557 for_each_online_cpu(cpu)
4558 per_cpu_ptr(s->cpu_slab, cpu)->stat[si] = 0;
4561 #define STAT_ATTR(si, text) \
4562 static ssize_t text##_show(struct kmem_cache *s, char *buf) \
4564 return show_stat(s, buf, si); \
4566 static ssize_t text##_store(struct kmem_cache *s, \
4567 const char *buf, size_t length) \
4569 if (buf[0] != '0') \
4571 clear_stat(s, si); \
4576 STAT_ATTR(ALLOC_FASTPATH, alloc_fastpath);
4577 STAT_ATTR(ALLOC_SLOWPATH, alloc_slowpath);
4578 STAT_ATTR(FREE_FASTPATH, free_fastpath);
4579 STAT_ATTR(FREE_SLOWPATH, free_slowpath);
4580 STAT_ATTR(FREE_FROZEN, free_frozen);
4581 STAT_ATTR(FREE_ADD_PARTIAL, free_add_partial);
4582 STAT_ATTR(FREE_REMOVE_PARTIAL, free_remove_partial);
4583 STAT_ATTR(ALLOC_FROM_PARTIAL, alloc_from_partial);
4584 STAT_ATTR(ALLOC_SLAB, alloc_slab);
4585 STAT_ATTR(ALLOC_REFILL, alloc_refill);
4586 STAT_ATTR(FREE_SLAB, free_slab);
4587 STAT_ATTR(CPUSLAB_FLUSH, cpuslab_flush);
4588 STAT_ATTR(DEACTIVATE_FULL, deactivate_full);
4589 STAT_ATTR(DEACTIVATE_EMPTY, deactivate_empty);
4590 STAT_ATTR(DEACTIVATE_TO_HEAD, deactivate_to_head);
4591 STAT_ATTR(DEACTIVATE_TO_TAIL, deactivate_to_tail);
4592 STAT_ATTR(DEACTIVATE_REMOTE_FREES, deactivate_remote_frees);
4593 STAT_ATTR(ORDER_FALLBACK, order_fallback);
4596 static struct attribute *slab_attrs[] = {
4597 &slab_size_attr.attr,
4598 &object_size_attr.attr,
4599 &objs_per_slab_attr.attr,
4601 &min_partial_attr.attr,
4603 &objects_partial_attr.attr,
4605 &cpu_slabs_attr.attr,
4609 &hwcache_align_attr.attr,
4610 &reclaim_account_attr.attr,
4611 &destroy_by_rcu_attr.attr,
4613 &reserved_attr.attr,
4614 #ifdef CONFIG_SLUB_DEBUG
4615 &total_objects_attr.attr,
4617 &sanity_checks_attr.attr,
4619 &red_zone_attr.attr,
4621 &store_user_attr.attr,
4622 &validate_attr.attr,
4623 &alloc_calls_attr.attr,
4624 &free_calls_attr.attr,
4626 #ifdef CONFIG_ZONE_DMA
4627 &cache_dma_attr.attr,
4630 &remote_node_defrag_ratio_attr.attr,
4632 #ifdef CONFIG_SLUB_STATS
4633 &alloc_fastpath_attr.attr,
4634 &alloc_slowpath_attr.attr,
4635 &free_fastpath_attr.attr,
4636 &free_slowpath_attr.attr,
4637 &free_frozen_attr.attr,
4638 &free_add_partial_attr.attr,
4639 &free_remove_partial_attr.attr,
4640 &alloc_from_partial_attr.attr,
4641 &alloc_slab_attr.attr,
4642 &alloc_refill_attr.attr,
4643 &free_slab_attr.attr,
4644 &cpuslab_flush_attr.attr,
4645 &deactivate_full_attr.attr,
4646 &deactivate_empty_attr.attr,
4647 &deactivate_to_head_attr.attr,
4648 &deactivate_to_tail_attr.attr,
4649 &deactivate_remote_frees_attr.attr,
4650 &order_fallback_attr.attr,
4652 #ifdef CONFIG_FAILSLAB
4653 &failslab_attr.attr,
4659 static struct attribute_group slab_attr_group = {
4660 .attrs = slab_attrs,
4663 static ssize_t slab_attr_show(struct kobject *kobj,
4664 struct attribute *attr,
4667 struct slab_attribute *attribute;
4668 struct kmem_cache *s;
4671 attribute = to_slab_attr(attr);
4674 if (!attribute->show)
4677 err = attribute->show(s, buf);
4682 static ssize_t slab_attr_store(struct kobject *kobj,
4683 struct attribute *attr,
4684 const char *buf, size_t len)
4686 struct slab_attribute *attribute;
4687 struct kmem_cache *s;
4690 attribute = to_slab_attr(attr);
4693 if (!attribute->store)
4696 err = attribute->store(s, buf, len);
4701 static void kmem_cache_release(struct kobject *kobj)
4703 struct kmem_cache *s = to_slab(kobj);
4709 static const struct sysfs_ops slab_sysfs_ops = {
4710 .show = slab_attr_show,
4711 .store = slab_attr_store,
4714 static struct kobj_type slab_ktype = {
4715 .sysfs_ops = &slab_sysfs_ops,
4716 .release = kmem_cache_release
4719 static int uevent_filter(struct kset *kset, struct kobject *kobj)
4721 struct kobj_type *ktype = get_ktype(kobj);
4723 if (ktype == &slab_ktype)
4728 static const struct kset_uevent_ops slab_uevent_ops = {
4729 .filter = uevent_filter,
4732 static struct kset *slab_kset;
4734 #define ID_STR_LENGTH 64
4736 /* Create a unique string id for a slab cache:
4738 * Format :[flags-]size
4740 static char *create_unique_id(struct kmem_cache *s)
4742 char *name = kmalloc(ID_STR_LENGTH, GFP_KERNEL);
4749 * First flags affecting slabcache operations. We will only
4750 * get here for aliasable slabs so we do not need to support
4751 * too many flags. The flags here must cover all flags that
4752 * are matched during merging to guarantee that the id is
4755 if (s->flags & SLAB_CACHE_DMA)
4757 if (s->flags & SLAB_RECLAIM_ACCOUNT)
4759 if (s->flags & SLAB_DEBUG_FREE)
4761 if (!(s->flags & SLAB_NOTRACK))
4765 p += sprintf(p, "%07d", s->size);
4766 BUG_ON(p > name + ID_STR_LENGTH - 1);
4770 static int sysfs_slab_add(struct kmem_cache *s)
4776 if (slab_state < SYSFS)
4777 /* Defer until later */
4780 unmergeable = slab_unmergeable(s);
4783 * Slabcache can never be merged so we can use the name proper.
4784 * This is typically the case for debug situations. In that
4785 * case we can catch duplicate names easily.
4787 sysfs_remove_link(&slab_kset->kobj, s->name);
4791 * Create a unique name for the slab as a target
4794 name = create_unique_id(s);
4797 s->kobj.kset = slab_kset;
4798 err = kobject_init_and_add(&s->kobj, &slab_ktype, NULL, name);
4800 kobject_put(&s->kobj);
4804 err = sysfs_create_group(&s->kobj, &slab_attr_group);
4806 kobject_del(&s->kobj);
4807 kobject_put(&s->kobj);
4810 kobject_uevent(&s->kobj, KOBJ_ADD);
4812 /* Setup first alias */
4813 sysfs_slab_alias(s, s->name);
4819 static void sysfs_slab_remove(struct kmem_cache *s)
4821 if (slab_state < SYSFS)
4823 * Sysfs has not been setup yet so no need to remove the
4828 kobject_uevent(&s->kobj, KOBJ_REMOVE);
4829 kobject_del(&s->kobj);
4830 kobject_put(&s->kobj);
4834 * Need to buffer aliases during bootup until sysfs becomes
4835 * available lest we lose that information.
4837 struct saved_alias {
4838 struct kmem_cache *s;
4840 struct saved_alias *next;
4843 static struct saved_alias *alias_list;
4845 static int sysfs_slab_alias(struct kmem_cache *s, const char *name)
4847 struct saved_alias *al;
4849 if (slab_state == SYSFS) {
4851 * If we have a leftover link then remove it.
4853 sysfs_remove_link(&slab_kset->kobj, name);
4854 return sysfs_create_link(&slab_kset->kobj, &s->kobj, name);
4857 al = kmalloc(sizeof(struct saved_alias), GFP_KERNEL);
4863 al->next = alias_list;
4868 static int __init slab_sysfs_init(void)
4870 struct kmem_cache *s;
4873 down_write(&slub_lock);
4875 slab_kset = kset_create_and_add("slab", &slab_uevent_ops, kernel_kobj);
4877 up_write(&slub_lock);
4878 printk(KERN_ERR "Cannot register slab subsystem.\n");
4884 list_for_each_entry(s, &slab_caches, list) {
4885 err = sysfs_slab_add(s);
4887 printk(KERN_ERR "SLUB: Unable to add boot slab %s"
4888 " to sysfs\n", s->name);
4891 while (alias_list) {
4892 struct saved_alias *al = alias_list;
4894 alias_list = alias_list->next;
4895 err = sysfs_slab_alias(al->s, al->name);
4897 printk(KERN_ERR "SLUB: Unable to add boot slab alias"
4898 " %s to sysfs\n", s->name);
4902 up_write(&slub_lock);
4907 __initcall(slab_sysfs_init);
4908 #endif /* CONFIG_SYSFS */
4911 * The /proc/slabinfo ABI
4913 #ifdef CONFIG_SLABINFO
4914 static void print_slabinfo_header(struct seq_file *m)
4916 seq_puts(m, "slabinfo - version: 2.1\n");
4917 seq_puts(m, "# name <active_objs> <num_objs> <objsize> "
4918 "<objperslab> <pagesperslab>");
4919 seq_puts(m, " : tunables <limit> <batchcount> <sharedfactor>");
4920 seq_puts(m, " : slabdata <active_slabs> <num_slabs> <sharedavail>");
4924 static void *s_start(struct seq_file *m, loff_t *pos)
4928 down_read(&slub_lock);
4930 print_slabinfo_header(m);
4932 return seq_list_start(&slab_caches, *pos);
4935 static void *s_next(struct seq_file *m, void *p, loff_t *pos)
4937 return seq_list_next(p, &slab_caches, pos);
4940 static void s_stop(struct seq_file *m, void *p)
4942 up_read(&slub_lock);
4945 static int s_show(struct seq_file *m, void *p)
4947 unsigned long nr_partials = 0;
4948 unsigned long nr_slabs = 0;
4949 unsigned long nr_inuse = 0;
4950 unsigned long nr_objs = 0;
4951 unsigned long nr_free = 0;
4952 struct kmem_cache *s;
4955 s = list_entry(p, struct kmem_cache, list);
4957 for_each_online_node(node) {
4958 struct kmem_cache_node *n = get_node(s, node);
4963 nr_partials += n->nr_partial;
4964 nr_slabs += atomic_long_read(&n->nr_slabs);
4965 nr_objs += atomic_long_read(&n->total_objects);
4966 nr_free += count_partial(n, count_free);
4969 nr_inuse = nr_objs - nr_free;
4971 seq_printf(m, "%-17s %6lu %6lu %6u %4u %4d", s->name, nr_inuse,
4972 nr_objs, s->size, oo_objects(s->oo),
4973 (1 << oo_order(s->oo)));
4974 seq_printf(m, " : tunables %4u %4u %4u", 0, 0, 0);
4975 seq_printf(m, " : slabdata %6lu %6lu %6lu", nr_slabs, nr_slabs,
4981 static const struct seq_operations slabinfo_op = {
4988 static int slabinfo_open(struct inode *inode, struct file *file)
4990 return seq_open(file, &slabinfo_op);
4993 static const struct file_operations proc_slabinfo_operations = {
4994 .open = slabinfo_open,
4996 .llseek = seq_lseek,
4997 .release = seq_release,
5000 static int __init slab_proc_init(void)
5002 proc_create("slabinfo", S_IRUGO, NULL, &proc_slabinfo_operations);
5005 module_init(slab_proc_init);
5006 #endif /* CONFIG_SLABINFO */